docs(depth stats): updating docs

test(depth stats): updating tests
feat(depth stats): enforcing all depth stats to be in millimeters (default unit) for consistency
2026-07-06 09:37:06 +00:00 · 2026-07-01 18:14:01 +02:00 · 2026-07-01 18:14:01 +02:00 · 2026-07-01 18:14:01 +02:00
97 changed files with 6835 additions and 18019 deletions
@@ -22,10 +22,6 @@ outputs
 rl
 media
 # Local virtualenvs (the image provides its own)
 .venv
 venv
 # Logging
 logs
@@ -105,7 +105,7 @@ lerobot-train \
 | -------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
 | **Imitation Learning**     | [ACT](./docs/source/policy_act_README.md), [Diffusion](./docs/source/policy_diffusion_README.md), [VQ-BeT](./docs/source/policy_vqbet_README.md), [Multitask DiT Policy](./docs/source/policy_multi_task_dit_README.md)                                                                                                                                    |
 | **Reinforcement Learning** | [HIL-SERL](./docs/source/hilserl.mdx), [TDMPC](./docs/source/policy_tdmpc_README.md) & QC-FQL (coming soon)                                                                                                                                                                                                                                                |
-| **VLAs Models**            | [Pi0](./docs/source/pi0.mdx), [Pi0Fast](./docs/source/pi0fast.mdx), [Pi0.5](./docs/source/pi05.mdx), [GR00T N1.7](./docs/source/policy_groot_README.md), [SmolVLA](./docs/source/policy_smolvla_README.md), [XVLA](./docs/source/xvla.mdx), [EO-1](./docs/source/eo1.mdx), [MolmoAct2](./docs/source/molmoact2.mdx), [WALL-OSS](./docs/source/walloss.mdx) |
+| **VLAs Models**            | [Pi0](./docs/source/pi0.mdx), [Pi0Fast](./docs/source/pi0fast.mdx), [Pi0.5](./docs/source/pi05.mdx), [GR00T N1.5](./docs/source/policy_groot_README.md), [SmolVLA](./docs/source/policy_smolvla_README.md), [XVLA](./docs/source/xvla.mdx), [EO-1](./docs/source/eo1.mdx), [MolmoAct2](./docs/source/molmoact2.mdx), [WALL-OSS](./docs/source/walloss.mdx) |
 | **World Models**           | [VLA-JEPA](./docs/source/vla_jepa.mdx) (more coming soon)                                                                                                                                                                                                                                                                                                  |
 | **Reward Models**          | [SARM](./docs/source/sarm.mdx), [TOPReward](./docs/source/topreward.mdx), [Robometer](./docs/source/robometer.mdx)                                                                                                                                                                                                                                         |
@@ -69,14 +69,10 @@
    title: VLA-JEPA
  - local: eo1
    title: EO-1
  - local: lingbot_va
    title: LingBot-VA
  - local: fastwam
    title: FastWAM
  - local: evo1
    title: EVO1
  - local: groot
-    title: NVIDIA GR00T
+    title: NVIDIA GR00T N1.5
  - local: xvla
    title: X-VLA
  - local: multi_task_dit
@@ -193,7 +193,7 @@ To learn more about training policies with LeRobot, please refer to the training
 - [SmolVLA](./smolvla)
 - [Pi0.5](./pi05)
- [GR00T N1.7](./groot)
+- [GR00T N1.5](./groot)
 Sample IsaacLab Arena datasets are available on HuggingFace Hub for experimentation:
@@ -1,191 +0,0 @@
 # EVO1
 EVO1 is a Vision-Language-Action policy for robot control built around an InternVL3 backbone and a continuous flow-matching action head. This LeRobot integration exposes EVO1 as a standard policy type so it can be trained and evaluated with the usual LeRobot dataset, checkpoint, and processor APIs.
 ## Model Overview
 The policy embeds one or more camera images and the language task prompt with InternVL3, pads robot state/action vectors to fixed maximum dimensions, and predicts future action chunks with a flow-matching action head. During inference, the policy samples an action chunk and returns `n_action_steps` actions from that chunk before sampling again.
 ### What the LeRobot Integration Covers
 - Standard `policy.type=evo1` configuration through LeRobot
 - InternVL3 image/text embedding with optional FlashAttention fallback
 - Stage-based finetuning controls for action-head-only and VLM finetuning runs
 - Continuous flow-matching action prediction
 - Checkpoint save/load through LeRobot policy APIs
 - Training with `lerobot-train` and evaluation with standard policy inference APIs
 The broader EVO1 project may include additional training scripts and dataset tooling. This page focuses on the LeRobot robot-control policy path.
 ## Installation Requirements
 1. Install LeRobot by following the [Installation Guide](./installation).
 2. Install EVO1 dependencies:
   ```bash
   pip install -e ".[evo1]"
   ```
   For LIBERO evaluation, install the LIBERO extra as well:
   ```bash
   pip install -e ".[evo1,libero]"
   ```
 3. Install a `flash-attn` wheel only if it is compatible with your Python, PyTorch, CUDA, and GPU stack. EVO1 falls back to standard attention when `flash_attn` is not available.
 EVO1 uses the native Hugging Face `transformers` InternVL implementation, so `policy.vlm_model_name` must point to a natively converted checkpoint such as `OpenGVLab/InternVL3-1B-hf` (note the `-hf` suffix). The first run may download the configured VLM checkpoint unless `policy.vlm_model_name` points to a local model directory.
 ## Data Requirements
 EVO1 expects a LeRobot dataset with:
 - One to `policy.max_views` visual observations, for example `observation.images.image`
 - `observation.state`
 - `action`
 - A language task instruction in the dataset `task` field, or another field configured with `policy.task_field`
 State and action vectors are padded to `policy.max_state_dim` and `policy.max_action_dim`. Predictions are cropped back to the dataset action dimension before being returned.
 ## Usage
 To use EVO1 in a LeRobot configuration, specify:
 ```python
 policy.type=evo1
 ```
 By default, a new EVO1 policy initializes its VLM from:
 ```python
 policy.vlm_model_name=OpenGVLab/InternVL3-1B-hf
 ```
 Once a LeRobot-format EVO1 checkpoint is available, load it with:
 ```python
 policy.path=your-org/your-evo1-checkpoint
 ```
 ## Training
 ### Stage 1
 Stage 1 freezes the VLM and trains the action head:
 ```bash
 lerobot-train \
  --dataset.repo_id=your_org/your_dataset \
  --policy.type=evo1 \
  --policy.training_stage=stage1 \
  --policy.vlm_model_name=OpenGVLab/InternVL3-1B-hf \
  --policy.device=cuda \
  --policy.chunk_size=50 \
  --policy.n_action_steps=50 \
  --policy.max_state_dim=24 \
  --policy.max_action_dim=24 \
  --policy.optimizer_lr=1e-5 \
  --batch_size=4 \
  --steps=5000 \
  --output_dir=./outputs/evo1_stage1
 ```
 ### Stage 2
 Stage 2 finetunes the VLM branches and action head. A common workflow starts from a Stage 1 checkpoint:
 ```bash
 lerobot-train \
  --dataset.repo_id=your_org/your_dataset \
  --policy.path=./outputs/evo1_stage1/checkpoints/005000/pretrained_model \
  --policy.training_stage=stage2 \
  --policy.vlm_model_name=OpenGVLab/InternVL3-1B-hf \
  --policy.device=cuda \
  --policy.chunk_size=50 \
  --policy.n_action_steps=50 \
  --policy.max_state_dim=24 \
  --policy.max_action_dim=24 \
  --policy.optimizer_lr=1e-5 \
  --batch_size=4 \
  --steps=80000 \
  --output_dir=./outputs/evo1_stage2
 ```
 By default, `policy.training_stage` reapplies the finetuning defaults for that stage. This is important when
 starting Stage 2 from a Stage 1 checkpoint, because the Stage 1 checkpoint config stores the VLM finetuning
 flags as disabled. These stage defaults take precedence over saved or manually supplied `policy.finetune_*`
 flags unless `policy.apply_training_stage_defaults=false`, so set that flag only when manually controlling
 every finetuning flag.
 ### Key Training Parameters
 | Parameter                                     | Default                     | Description                                                       |
 | --------------------------------------------- | --------------------------- | ----------------------------------------------------------------- |
 | `policy.vlm_model_name`                       | `OpenGVLab/InternVL3-1B-hf` | Natively converted InternVL3 checkpoint or local model directory  |
 | `policy.training_stage`                       | `stage1`                    | `stage1` trains the action head; `stage2` finetunes VLM branches  |
 | `policy.apply_training_stage_defaults`        | `true`                      | Reapplies stage finetuning defaults after loading a checkpoint    |
 | `policy.vlm_num_layers`                       | `14`                        | Number of InternVL3 language layers kept for the policy           |
 | `policy.vlm_dtype`                            | `bfloat16`                  | Requested VLM dtype                                               |
 | `policy.use_flash_attn`                       | `true`                      | Requests FlashAttention when installed; otherwise falls back      |
 | `policy.enable_gradient_checkpointing`        | `true`                      | Enables checkpointing on supported InternVL3 modules              |
 | `policy.gradient_checkpointing_use_reentrant` | `false`                     | Reentrant setting passed to gradient checkpointing when supported |
 | `policy.chunk_size`                           | `50`                        | Number of future actions predicted per chunk                      |
 | `policy.n_action_steps`                       | `50`                        | Number of actions consumed from a sampled chunk                   |
 | `policy.max_state_dim`                        | `24`                        | State padding dimension                                           |
 | `policy.max_action_dim`                       | `24`                        | Action padding dimension                                          |
 | `policy.postprocess_action_dim`               | `null`                      | Optional action dimension returned after EVO1 postprocessing      |
 | `policy.binarize_gripper`                     | `false`                     | Binarizes the postprocessed gripper channel for LIBERO-style eval |
 | `policy.task_field`                           | `task`                      | Batch field used as the language prompt                           |
 ## Inference
 Try it out with a trained EVO1 checkpoint:
 ```bash
 lerobot-rollout \
  --policy.path=your-org/your-evo1-checkpoint \
  --inference.type=rtc \ # optional
  ...
 ```
 ## Results
 ### LIBERO Evaluation
 > [!NOTE]
 > Benchmark results for a `lerobot`-hosted LIBERO checkpoint trained with this implementation
 > will be added once training completes.
 The official EVO1 LIBERO rollout protocol uses the raw LIBERO camera feature names
 (`observation.images.agentview_image` and `observation.images.robot0_eye_in_hand_image`), replans every
 14 actions, and binarizes the gripper command before stepping the simulator. The EVO1 policy postprocessor
 can crop the padded 24D action back to the 7D LIBERO action space and apply that gripper binarization. To
 evaluate a LIBERO checkpoint under the same one-episode-per-task setting, keep the raw camera names instead
 of the default `image`/`image2` mapping and set the LIBERO action postprocessing flags:
 ```bash
 lerobot-eval \
  --policy.path=your-org/your-evo1-libero-checkpoint \
  --policy.vlm_model_name=OpenGVLab/InternVL3-1B-hf \
  --policy.device=cuda \
  --policy.use_flash_attn=true \
  --policy.n_action_steps=14 \
  --policy.postprocess_action_dim=7 \
  --policy.binarize_gripper=true \
  --env.type=libero \
  --env.task=libero_object \
  --env.camera_name_mapping="{agentview_image: agentview_image, robot0_eye_in_hand_image: robot0_eye_in_hand_image}" \
  --env.observation_height=448 \
  --env.observation_width=448 \
  --eval.batch_size=1 \
  --eval.n_episodes=1
 ```
 ## References
 - [EVO1 repository](https://github.com/MINT-SJTU/Evo-1)
 - [InternVL3-1B-hf](https://huggingface.co/OpenGVLab/InternVL3-1B-hf)
 ## License
 This LeRobot integration follows the Apache 2.0 License used by LeRobot. Check the upstream EVO1 and InternVL3 model pages for the licenses of released checkpoints and data.
@@ -1,19 +1,16 @@
-# GR00T Policy
+# GR00T N1.5 Policy
-GR00T is an NVIDIA foundation model family for generalized humanoid robot reasoning and skills. It is a cross-embodiment policy that accepts multimodal input, including language, images, and proprioception, to perform manipulation tasks in diverse environments.
+GR00T N1.5 is an open foundation model from NVIDIA designed for generalized humanoid robot reasoning and skills. It is a cross-embodiment model that accepts multimodal input, including language and images, to perform manipulation tasks in diverse environments.
-LeRobot integrates GR00T N1.7 through the `groot` policy type.
+This document outlines the specifics of its integration and usage within the LeRobot framework.
 > [!WARNING]
 > **Breaking change:** GR00T N1.5 support was removed from LeRobot, and current releases support GR00T N1.7 only. N1.5 checkpoints and configs are rejected with a migration note. To keep using an N1.5 checkpoint, pin the last release that supports it: `pip install 'lerobot==0.5.1'`. To use the current release, migrate to GR00T N1.7 (base model [`nvidia/GR00T-N1.7-3B`](https://huggingface.co/nvidia/GR00T-N1.7-3B)).
 ## Model Overview
-GR00T N1.7 uses a Cosmos-Reason2/Qwen3-VL backbone and provides checkpoints for SimplerEnv, DROID, and LIBERO.
+NVIDIA Isaac GR00T N1.5 is an upgraded version of the GR00T N1 foundation model. It is built to improve generalization and language-following abilities for humanoid robots.
-Developers and researchers can post-train GR00T with their own real or synthetic data to adapt it for specific humanoid robots or tasks.
+Developers and researchers can post-train GR00T N1.5 with their own real or synthetic data to adapt it for specific humanoid robots or tasks.
-GR00T uses pre-trained vision and language encoders with a flow matching action transformer to model a chunk of actions conditioned on vision, language, and proprioception.
+GR00T N1.5 (specifically the GR00T-N1.5-3B model) is built using pre-trained vision and language encoders. It utilizes a flow matching action transformer to model a chunk of actions, conditioned on vision, language, and proprioception.
 <img
  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/lerobot-groot-paper1%20(1).png"
@@ -31,24 +28,33 @@ This approach allows the model to be highly adaptable through post-training for
 ## Installation Requirements
-GR00T is intended for NVIDIA GPU-accelerated systems. Install LeRobot with the GR00T extra:
+As of today, GR00T N1.5 requires flash attention for it's internal working.
 We are working on making this optional, but in the meantime that means that we require an extra installation step and it can only be used in CUDA enabled devices.
 1. Following the Environment Setup of our [Installation Guide](./installation). **Attention** don't install `lerobot` in this step.
 2. Install [Flash Attention](https://github.com/Dao-AILab/flash-attention) by running:
 ```bash
-pip install "lerobot[groot]"
+# Check https://pytorch.org/get-started/locally/ for your system
 pip install "torch>=2.2.1,<2.8.0" "torchvision>=0.21.0,<0.23.0" # --index-url https://download.pytorch.org/whl/cu1XX
 pip install ninja "packaging>=24.2,<26.0" # flash attention dependencies
 pip install "flash-attn>=2.5.9,<3.0.0" --no-build-isolation
 python -c "import flash_attn; print(f'Flash Attention {flash_attn.__version__} imported successfully')"
 ```
-For a source checkout:
+3. Install LeRobot by running:
 ```bash
-pip install -e ".[groot]"
+pip install lerobot[groot]
 ```
 ## Usage
-To use GR00T N1.7:
+To use GR00T in your LeRobot configuration, specify the policy type as:
-```bash
+```python
--policy.type=groot
+policy.type=groot
 ```
 ## Training
@@ -57,171 +63,72 @@ To use GR00T N1.7:
 Here's a complete training command for finetuning the base GR00T model on your own dataset:
 This command is using the `new_embodiment` flag, which is used for the SO-101 robot, [read more about how GR00T handles different embodiments.](https://github.com/NVIDIA/Isaac-GR00T/blob/main/getting_started/policy.md#--embodiment-tag).
 ```bash
-# install extra deps for training
+# Using a multi-GPU setup
-pip install "lerobot[training]"
+accelerate launch \
-
+  --multi_gpu \
-hf auth login
+  --num_processes=$NUM_GPUS \
-wandb login
+  $(which lerobot-train) \
 export DATASET_NAME=your_data_set
 export HF_USER=your_hf_username
 export DATASET=$HF_USER/$DATASET_NAME
 export REPO_ID="${DATASET}_GR00T17" #this is the model that will be uploaded to huggingface
 export OUTPUT_DIR=outputs/train/$REPO_ID
 lerobot-train \
  --dataset.repo_id=$DATASET \
  --dataset.image_transforms.enable=true \
  --policy.type=groot \
  --policy.device=cuda \
  --policy.base_model_path=nvidia/GR00T-N1.7-3B \
  --policy.embodiment_tag=new_embodiment \
  --policy.chunk_size=16 \
  --policy.n_action_steps=16 \
  --policy.use_relative_actions=true \
  --policy.relative_exclude_joints='["gripper"]' \
  --policy.use_bf16=true \
  --policy.push_to_hub=true \
  --policy.repo_id=$REPO_ID \
  --seed=42 \
  --batch_size=64 \
  --steps=20000 \
  --save_checkpoint=true \
  --save_freq=5000 \
  --use_policy_training_preset=true \
  --env_eval_freq=0 \
  --eval_steps=0 \
  --log_freq=10 \
  --output_dir=$OUTPUT_DIR \
-  --job_name=$DATASET \
+  --save_checkpoint=true \
  --batch_size=$BATCH_SIZE \
  --steps=$NUM_STEPS \
  --save_freq=$SAVE_FREQ \
  --log_freq=$LOG_FREQ \
  --policy.push_to_hub=true \
  --policy.type=groot \
  --policy.repo_id=$REPO_ID \
  --policy.tune_diffusion_model=false \
  --dataset.repo_id=$DATASET_ID \
  --wandb.enable=true \
-  --wandb.disable_artifact=true
+  --wandb.disable_artifact=true \
-
+  --job_name=$JOB_NAME
 ```
 ## Performance Results
-### LIBERO Benchmark Results
+### Libero Benchmark Results
 > [!NOTE]
-> Follow the [LIBERO](./libero) setup instructions before running `lerobot-eval`.
+> Follow our instructions for Libero usage: [Libero](./libero)
-GR00T N1.7 has demonstrated strong performance on the LIBERO benchmark suite. To reproduce LeRobot results, follow the instructions in the [LIBERO](./libero) section.
+GR00T has demonstrated strong performance on the Libero benchmark suite. To compare and test its LeRobot implementation, we finetuned the GR00T N1.5 model for 30k steps on the Libero dataset and compared the results to the GR00T reference results.
-### Train on LIBERO
+| Benchmark          | LeRobot Implementation | GR00T Reference |
 | ------------------ | ---------------------- | --------------- |
 | **Libero Spatial** | 82.0%                  | 92.0%           |
 | **Libero Object**  | 99.0%                  | 92.0%           |
 | **Libero Long**    | 82.0%                  | 76.0%           |
 | **Average**        | 87.0%                  | 87.0%           |
-Example training command for a LIBERO suite (here `libero_spatial`):
+These results demonstrate GR00T's strong generalization capabilities across diverse robotic manipulation tasks. To reproduce these results, you can follow the instructions in the [Libero](https://huggingface.co/docs/lerobot/libero) section.
 ```bash
 IMAGE_TRANSFORMS='{
  "brightness": {"weight": 1.0, "type": "ColorJitter", "kwargs": {"brightness": [0.7, 1.3]}},
  "contrast":   {"weight": 1.0, "type": "ColorJitter", "kwargs": {"contrast":   [0.6, 1.4]}},
  "saturation": {"weight": 1.0, "type": "ColorJitter", "kwargs": {"saturation": [0.5, 1.5]}},
  "hue":        {"weight": 1.0, "type": "ColorJitter", "kwargs": {"hue":        [-0.08, 0.08]}}
 }'
 lerobot-train \
  --dataset.repo_id=IPEC-COMMUNITY/libero_spatial_no_noops_1.0.0_lerobot \
  --dataset.root=/datasets/libero_spatial \
  --dataset.revision=main \
  --dataset.video_backend=pyav \
  --dataset.image_transforms.enable=true \
  --dataset.image_transforms.max_num_transforms=4 \
  --dataset.image_transforms.tfs="$IMAGE_TRANSFORMS" \
  --policy.type=groot \
  --policy.base_model_path=nvidia/GR00T-N1.7-3B \
  --policy.embodiment_tag=libero_sim \
  --policy.push_to_hub=false \
  --policy.use_relative_actions=false \
  --policy.max_steps=20000 \
  --batch_size=320 \
  --steps=20000 \
  --save_freq=2000 \
  --env_eval_freq=0 \
  --eval_steps=0 \
  --log_freq=10 \
  --wandb.enable=true \
  --wandb.project=lerobot \
  --wandb.mode=online \
  --wandb.disable_artifact=true \
  --num_workers=4 \
  --prefetch_factor=2 \
  --persistent_workers=true \
  --output_dir=$OUTPUT_DIR \
  --job_name=$JOB_NAME
 ```
 This will follow the recipe found [here](https://github.com/NVIDIA/Isaac-GR00T/blob/main/examples/LIBERO/README.md).
 ### GR00T N1.7 LIBERO Results
 Preliminary LeRobot integration results (GR00T-LeRobot, `eval.n_episodes >= 50` per suite):
 | Suite            | Success rate |
 | ---------------- | -----------: |
 | LIBERO Spatial   |          94% |
 | LIBERO Object    |          98% |
 | LIBERO Goal      |          93% |
 | LIBERO 10 (Long) |          90% |
 | **Average**      |   **93.75%** |
 ```bash
 export MODEL_ID=your_trained_model_on_huggingface
 lerobot-eval \
  --policy.type=groot \
  --policy.base_model_path=$MODEL_ID \
  --policy.embodiment_tag=libero_sim \
  --env.type=libero \
  --env.task=libero_spatial \
  --eval.n_episodes=50
 ```
 Use `eval.n_episodes >= 50` per suite when reporting success rates.
 ### Evaluate in your hardware setup
 Once you have trained your model using your parameters you can run inference in your downstream task. Follow the instructions in [Policy Deployment (lerobot-rollout)](./inference). For example:
 ```bash
-# install extra deps for roullout and real hardware
+lerobot-rollout\
-pip install "lerobot[feetech,viz]"
+  --strategy.type=sentry \
-
+  --strategy.upload_every_n_episodes=5 \
-export MODEL_ID=your_trained_model_on_huggingface
+  --robot.type=bi_so_follower \
-
+  --robot.left_arm_port=/dev/ttyACM1 \
-# make sure that camera index matches your setup!
+  --robot.right_arm_port=/dev/ttyACM0 \
-# find index using `uv run lerobot-find-cameras opencv`
+  --robot.id=bimanual_follower \
-WRIST_CAM='wrist: {type: opencv, index_or_path: 2, width: 640, height: 480, fps: 30, fourcc: "MJPG"}'
+  --robot.cameras='{ right: {"type": "opencv", "index_or_path": 0, "width": 640, "height": 480, "fps": 30},
-FRONT_CAM='front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30, fourcc: "MJPG"}'
+    left: {"type": "opencv", "index_or_path": 2, "width": 640, "height": 480, "fps": 30},
-export ROBOT_CAMERAS="{ $WRIST_CAM, $FRONT_CAM }"
+    top: {"type": "opencv", "index_or_path": 4, "width": 640, "height": 480, "fps": 30},
-export ROBOT_ID=follower_robot
+  }' \
 export ROBOT_PORT=/dev/ttyACM0
 uv run lerobot-rollout \
  --strategy.type=base \
  --policy.path=$MODEL_ID \
  --policy.base_model_path=nvidia/GR00T-N1.7-3B \
  --policy.n_action_steps=8 \
  --robot.type=so101_follower \
  --robot.port=$ROBOT_PORT \
  --robot.id=$ROBOT_ID \
  --robot.cameras="$ROBOT_CAMERAS" \
  --task="place the vial in the rack" \
  --duration=60 \
  --device=cuda \
  --display_data=true \
-  --inference.type=rtc \
+  --dataset.repo_id=<user>/eval_groot-bimanual  \
-  --inference.rtc.enabled=True \ # set to False if it causes inference instability
+  --dataset.single_task="Grab and handover the red cube to the other arm" \
-  --inference.rtc.execution_horizon=8 \
+  --dataset.streaming_encoding=true \
-  --inference.queue_threshold=0
+  --dataset.encoder_threads=2 \
  # --dataset.rgb_encoder.vcodec=auto \
  --policy.path=<user>/groot-bimanual \ # your trained model
  --duration=600
 ```
 > [!NOTE]
 > Value of `inference.queue_threshold` should not exceed 5 to ensure stable inference.
 ## License
-GR00T N1.7 is released under the [NVIDIA Open Model License Agreement](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/).
+This model follows NVIDIA's proprietary license, consistent with the original [GR00T repository](https://github.com/NVIDIA/Isaac-GR00T). Future versions (starting from N1.7) will follow **Apache 2.0 License**.
@@ -82,18 +82,18 @@ VRAM is the first filter. Within a tier, pick by budget and availability — the
 ### Hugging Face Jobs
-[Hugging Face Jobs](https://huggingface.co/docs/hub/jobs) lets you run training on managed HF infrastructure, billed by the second, without owning a GPU. `lerobot-train` submits and streams the job for you — just add `--job.target=<flavor>` to a normal training command:
+[Hugging Face Jobs](https://huggingface.co/docs/hub/jobs) lets you run training on managed HF infrastructure, billed by the second. The repo publishes a ready-to-use image: **`huggingface/lerobot-gpu:latest`**, rebuilt **every night at 02:00 UTC from `main`** ([`docker_publish.yml`](https://github.com/huggingface/lerobot/blob/main/.github/workflows/docker_publish.yml)) — so it tracks the current state of the repo, not a tagged release.
 ```bash
-lerobot-train \
+hf jobs run --flavor a10g-large huggingface/lerobot-gpu:latest \
-  --policy.type=act --dataset.repo_id=<USER>/<DATASET> \
+  bash -c "nvidia-smi && lerobot-train \
-  --policy.repo_id=<USER>/act_<task> \
+    --policy.type=act --dataset.repo_id=<USER>/<DATASET> \
-  --job.target=a10g-large
+    --policy.repo_id=<USER>/act_<task> --batch_size=8 --steps=50000"
 ```
 Notes:
- Run `hf auth login` once before submitting, the job runs under your token.
+- The leading `nvidia-smi` is a quick sanity check that CUDA is visible inside the container — useful to fail fast if the flavor or driver mismatched.
- `--job.target` maps onto the table above: `t4-small`/`t4-medium` (T4, ACT only), `l4x1`/`l4x4` (L4 24 GB), `a10g-small/large/largex2/largex4` (A10G 24 GB scaled out), `a100-large` (A100). List the current catalogue with pricing via `hf jobs hardware`, or see [https://huggingface.co/docs/hub/jobs](https://huggingface.co/docs/hub/jobs).
+- The default Job timeout is 30 minutes; pass `--timeout 4h` (or longer) for real training.
- The job defaults to a `2d` (48h) timeout. Override it with `--job.timeout=4h` (or any other valid duration string) to shorten or extend the timeout. The job automatically stops when the command completes.
+- `--flavor` maps onto the table above: `t4-small`/`t4-medium` (T4, ACT only), `l4x1`/`l4x4` (L4 24 GB), `a10g-small/large/largex2/largex4` (A10G 24 GB scaled out), `a100-large` (A100). For the current full catalogue + pricing see [https://huggingface.co/docs/hub/jobs](https://huggingface.co/docs/hub/jobs).
- For the full walkthrough — dataset upload, checkpoint streaming, resuming a run on a job — see the [imitation-learning training guide](./il_robots#train-using-hugging-face-jobs).
+- Prefer not to write the `hf jobs run` wrapper yourself? `lerobot-train` can submit the job for you: just add `--job.target=<flavor>` to a normal training command and it handles dataset upload, log streaming, and the final model push. See the [imitation-learning training guide](./il_robots).
@@ -126,7 +126,7 @@ import time
 from lerobot.teleoperators.so_leader import SO101Leader, SO101LeaderConfig
 from lerobot.robots.so_follower import SO101Follower, SO101FollowerConfig
 from lerobot.cameras.opencv import OpenCVCameraConfig
-from lerobot.utils.visualization_utils import init_visualization, log_visualization_data, shutdown_visualization
+from lerobot.utils.visualization_utils import init_rerun, log_rerun_data, shutdown_rerun
 robot_config = SO101FollowerConfig(
    port="/dev/tty.usbmodem5AB90687491",
@@ -142,7 +142,7 @@ teleop_config = SO101LeaderConfig(
    id="my_leader_arm",
 )
-init_visualization("rerun", session_name="teleoperation")  # pass "foxglove" to stream to Foxglove instead
+init_rerun(session_name="teleoperation")
 robot = SO101Follower(robot_config)
 teleop_device = SO101Leader(teleop_config)
@@ -158,7 +158,7 @@ while True:
    observation = robot.get_observation()
    action = teleop_device.get_action()
    robot.send_action(action)
-    log_visualization_data("rerun", observation=observation, action=action)
+    log_rerun_data(observation=observation, action=action)
    elapsed_time = time.perf_counter() - start_time
    sleep_time = TIME_PER_FRAME - elapsed_time
@@ -223,7 +223,7 @@ from lerobot.teleoperators.so_leader.config_so_leader import SO101LeaderConfig
 from lerobot.teleoperators.so_leader.so_leader import SO101Leader
 from lerobot.common.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
-from lerobot.utils.visualization_utils import init_visualization
+from lerobot.utils.visualization_utils import init_rerun
 from lerobot.scripts.lerobot_record import record_loop
 from lerobot.processor import make_default_processors
@@ -270,7 +270,7 @@ def main():
    # Initialize the keyboard listener and rerun visualization
    _, events = init_keyboard_listener()
-    init_visualization("rerun", session_name="recording")
+    init_rerun(session_name="recording")
    # Connect the robot and teleoperator
    robot.connect()
@@ -532,7 +532,84 @@ If your local computer doesn't have a powerful GPU you could utilize Google Cola
 Hugging Face jobs let's you easily select hardware and run the training in the cloud. So if you don't have a powerful GPU or you need more VRAM or just want to train a model much faster use HF Jobs! It's pay as you go and you simply pay for each second of use, you can see the pricing and additional information [here](https://huggingface.co/docs/hub/jobs).
-`lerobot-train` runs locally by default. To run on a HuggingFace GPU, pass `--job.target` with a hardware flavor name:
+> **Tip:** if you just want to launch a standard training run, you can skip building the command below and use the integrated **Train on HF Jobs via `--job.target`** flow described further down — `lerobot-train` then submits the job, uploads a local-only dataset for you, and streams the logs.
 To run the training manually use this command:
 <hfoptions id="train_with_hf_jobs">
 <hfoption id="Command">
 ```bash
 hf jobs run \
  --flavor a10g-small \
  --timeout 4h \
  --secrets HF_TOKEN \
  huggingface/lerobot-gpu:latest \
  -- \
  python -m lerobot.scripts.lerobot_train \
    --dataset.repo_id=username/dataset \
    --policy.type=act \
    --steps=5000 \
    --batch_size=16 \
    --policy.device=cuda \
    --policy.repo_id=username/your_policy \
    --log_freq=100
 ```
 </hfoption>
 <hfoption id="API example">
 <!-- prettier-ignore-start -->
 ```python
 from huggingface_hub import run_job, get_token
 run_name = "act_so101_hf_jobs"
 dataset_id = "username/dataset"
 user_hub_id = "username"
 command_args = [
    "python", "-m", "lerobot.scripts.lerobot_train",
    "--dataset.repo_id", dataset_id,
    "--policy.type", "act",
    "--steps", "5000",
    "--batch_size", "16",
    "--num_workers", "4",
    "--policy.device", "cuda",
    "--log_freq", "100",
    "--save_freq", "1000",
    "--save_checkpoint", "true",
    "--wandb.enable", "false",
    "--policy.repo_id", f"{user_hub_id}/{run_name}"
 ]
 print(f"Submitting job '{run_name}' to Hugging Face Infrastructure...")
 job_info = run_job(
    image="huggingface/lerobot-gpu:latest",
    command=command_args,
    flavor="a10g-small",
    timeout="4h",
    secrets={"HF_TOKEN": get_token()}
 )
 print("\n🚀 Job successfully launched!")
 print(f"🔹 Job ID: {job_info.id}")
 print(f"🔗 Live UI Dashboard & Logs: {job_info.url}")
 ```
 <!-- prettier-ignore-end -->
 </hfoption>
 </hfoptions>
 You can modify the `--flavor` to use different hardware, for example: `t4-small`, `a100-large`, `h200`. Use `hf jobs hardware` to see the full list with pricing.
 Depending on the model you want to train and the hardware you selected you can also modify the `--batch_size` and `--number_of_workers`.
 For longer training sessions increase the timeout.
 Once the training is started you can go to [Jobs](https://huggingface.co/settings/jobs) and see if your jobs is running as well as all the outputs. Sometimes it takes a few minutes to schedule your job so be patient.
 After training the model will be pushed to hub and you can use it as any other model with LeRobot.
 #### Train on HF Jobs via `--job.target` (integrated CLI)
 `lerobot-train` runs locally by default. To run on a HuggingFace GPU without constructing the Docker command yourself, pass `--job.target` with a hardware flavor name:
 ```bash
 lerobot-train \
@@ -1,187 +0,0 @@
 # LingBot-VA
 LingBot-VA is an **autoregressive video-action world-model policy** built on the **Wan2.2**
 video-diffusion stack. It interleaves, in one autoregressive sequence, the prediction of
 future **video latents** and **robot actions** ("VA" = Video-Action). The LeRobot
 integration wires LingBot-VA into the standard training, evaluation and processor
 interfaces.
 ## Model Overview
 LingBot-VA is a **dual-stream "mixture-of-transformers"**: a video/latent stream
 (`patch_embedding_mlp → blocks → proj_out`) and an action stream
 (`action_embedder → blocks → action_proj_out`) share the same 30 transformer blocks and
 text conditioning.
 | Component                | Class                   | Role                                                        |
 | ------------------------ | ----------------------- | ----------------------------------------------------------- |
 | DiT backbone (trainable) | `WanTransformer3DModel` | ~5B-param dual-stream transformer.                          |
 | VAE (frozen)             | `AutoencoderKLWan`      | Wan2.2 VAE, `z_dim=48`. Lazy-pulled from the source repo.   |
 | Text encoder (frozen)    | `UMT5EncoderModel`      | UMT5-XXL, `d_model=4096`. Lazy-pulled from the source repo. |
 At inference the policy runs an autoregressive loop per chunk: it denoises the video-latent
 stream (CFG, ~20 steps) and the action stream (~50 steps) with two independent
 flow-matching schedulers, maintaining a KV cache across chunks. Real observed keyframes are
 fed back into the KV cache as the chunk is executed (closed-loop world modeling).
 ### What the LeRobot Integration Covers
 - Standard `policy.type=lingbot_va` configuration through LeRobot.
 - Ready-to-use LeRobot-format checkpoints on the Hub (converted from the released upstream ones).
 - Autoregressive dual-stream inference behind the standard `select_action` interface
  (single-environment eval, `--eval.batch_size=1`).
 - Opt-in saving of the policy's **predicted (imagined) videos** during eval / training.
 - Evaluation with `lerobot-eval` on LIBERO and RoboTwin.
 - Training / fine-tuning via the dual-stream flow-matching loss (`policy.forward`), see below.
 ## Installation
 1. Install LeRobot by following the [Installation Guide](./installation).
 2. Install the LingBot-VA extra:
 ```bash
 pip install -e ".[lingbot_va]"
 ```
 ## Checkpoints
 The released upstream checkpoints have been converted to LeRobot format and pushed to the Hub:
 | Variant                | LeRobot checkpoint               |
 | ---------------------- | -------------------------------- |
 | LIBERO-Long post-train | `lerobot/lingbot_va_libero_long` |
 | RoboTwin post-train    | `lerobot/lingbot_va_robotwin`    |
 | Pretrained base        | `lerobot/lingbot_va_base`        |
 Only the trainable ~5B transformer is stored in the LeRobot
 `model.safetensors`. The frozen VAE + UMT5 + tokenizer (~20 GB) are pulled from
 `config.wan_pretrained_path` at load time (defaults to the source `robbyant/*` repo). The
 UMT5-XXL text encoder runs on CPU by default (`config.text_encoder_device`) so the 5B
 transformer + VAE fit on a single 24–32 GB GPU.
 ## Evaluation (LIBERO)
 ```bash
 lerobot-eval \
    --policy.path=lerobot/lingbot_va_libero_long \
    --policy.device=cuda \
    --env.type=libero --env.task=libero_10 \
    --env.observation_height=128 --env.observation_width=128 \
    --eval.n_episodes=50 --eval.batch_size=1 \
    --output_dir=outputs/eval/lingbot_va_libero
 ```
 LingBot-VA's streaming inference (KV cache + observed-keyframe feedback) is implemented for
 single-environment eval; use `--eval.batch_size=1`.
 ## Evaluation (RoboTwin)
 RoboTwin 2.0 needs the SAPIEN + CuRobo simulator stack. You can use the benchmark Docker image
 (`docker/Dockerfile.benchmark.robotwin`, which also needs `warp-lang==1.3.1` and CuRobo built
 with the GPU's compute capability in `TORCH_CUDA_ARCH_LIST`). RoboTwin uses **end-effector-pose
 control**, so run with `--env.action_mode=ee`: the policy predicts per-arm `xyz+quaternion+gripper`
 deltas (`robotwin_tshape` latent layout) that are composed onto the episode's initial eef pose and
 executed via CuRobo IK.
 ```bash
 lerobot-eval \
    --policy.path=lerobot/lingbot_va_robotwin \
    --policy.device=cuda \
    --env.type=robotwin --env.task=beat_block_hammer --env.action_mode=ee \
    --eval.n_episodes=10 --eval.batch_size=1 \
    --output_dir=outputs/eval/lingbot_va_robotwin
 ```
 ### Saving predicted (imagined) videos
 Set `--policy.save_predicted_video=true` to additionally VAE-decode the predicted video
 latents and write `pred_episode_*.mp4` next to the env-rendered `eval_episode_*.mp4` videos.
 The same flag works for the periodic eval during `lerobot-train`.
 ## Training / fine-tuning
 `LingBotVAPolicy.forward(batch)` implements the dual-stream **flow-matching** loss
 (`latent_loss + action_loss`, timestep-weighted, action-masked) from the paper: it VAE-encodes
 the camera clips into video latents, UMT5-encodes the task, noises both streams, runs the
 transformer's block-causal training pass and returns `(loss, metrics)`. Optimizer preset is AdamW
 with a linear-warmup-then-constant schedule (matching upstream).
 Requirements:
 - The block-causal masks use PyTorch **flex-attention**, so build the policy with
  `--policy.attn_mode=flex` for training (the default `torch` SDPA is inference-only).
 - The full 5B DiT does not fit a single 24–32 GB GPU under AdamW; fine-tune with **LoRA**
  (`--policy.use_peft=true`) and/or optimizer offload. `get_optim_params` returns only the
  trainable (e.g. adapter) parameters; the VAE + UMT5 text encoder stay frozen.
 ```bash
 lerobot-train \
  --policy.path=lerobot/lingbot_va_libero_long --policy.attn_mode=flex \
  --policy.use_peft=true \
  --dataset.repo_id=<your LeRobot-format dataset> \
  --batch_size=1 --steps=... --output_dir=outputs/train/lingbot_va
 ```
 The dataset must provide camera clips (a temporal window per camera, VAE-encoded to
 `frame_chunk_size` latent frames) and `frame_chunk_size * action_per_frame` action steps per item.
 ## Data format (action channels & camera order)
 LingBot-VA is an **end-effector (Cartesian) pose** policy, it predicts EEF poses + gripper, not
 joint positions. Actions live in a fixed multi-embodiment **30-dim** layout; map your robot's
 action dimensions into these channels and pad the rest with `0` (`used_action_channel_ids` selects
 the channels a given checkpoint actually uses):
 | channels | meaning                                               |
 | -------- | ----------------------------------------------------- |
 | 0–6      | Left-arm end-effector pose                            |
 | 7–13     | Right-arm end-effector pose                           |
 | 14–20    | Left-arm joints (unused by the released checkpoints)  |
 | 21–27    | Right-arm joints (unused by the released checkpoints) |
 | 28       | Left gripper                                          |
 | 29       | Right gripper                                         |
 - **LIBERO** uses channels `0–6`: a 6-DoF EEF delta (xyz + rotation) + gripper (single arm).
 - **RoboTwin** uses channels `[0–6, 28, 7–13, 29]`: left EEF (xyz + quaternion) + left gripper +
  right EEF + right gripper (16 dims). The env converts these poses to joint trajectories via
  CuRobo IK — joints are never predicted.
 Joint-space datasets (or a different EEF convention) must be remapped into this schema before
 fine-tuning these checkpoints.
 **Camera order is fixed and order-sensitive**, per-camera latents are concatenated spatially in
 `obs_cam_keys` order, so the physical camera→slot mapping must match training:
 | benchmark | `obs_cam_keys` (in order)                                                                             | `camera_layout`                                                     |
 | --------- | ----------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------- |
 | LIBERO    | `observation.images.image` (agentview / 3rd-person), `observation.images.image2` (eye-in-hand wrist)  | `width_concat` (latents concatenated on width)                      |
 | RoboTwin  | `observation.images.head_camera`, `observation.images.left_camera`, `observation.images.right_camera` | `robotwin_tshape` (full-res head below, two half-res wrists on top) |
 The first camera is the exterior/head view and the rest are wrist views.
 ## Inference Hyperparameters (LIBERO)
 | Key                                    | Value                                                                             |
 | -------------------------------------- | --------------------------------------------------------------------------------- |
 | height × width                         | 128 × 128                                                                         |
 | cameras                                | `observation.images.image` (agentview), `observation.images.image2` (eye-in-hand) |
 | action channels used                   | 0–6 (7-DoF arm + gripper)                                                         |
 | action_per_frame / frame_chunk_size    | 4 / 4                                                                             |
 | attn_window                            | 30                                                                                |
 | video / action denoising steps         | 20 / 50                                                                           |
 | guidance_scale / action_guidance_scale | 5 / 1                                                                             |
 | snr_shift / action_snr_shift           | 5.0 / 0.05                                                                        |
 These are the defaults of `LingBotVAConfig`; override any of them via `--policy.<name>=...`.
 ## Notes
 - **Attention backend:** inference uses the `torch` SDPA backend (always available). The
  `flashattn` and `flex` backends are optional; `flex` is only needed for training.
 - **Model size:** the DiT is ~5B params and the frozen VAE+UMT5 add ~20 GB; inference needs
  roughly 18–24 GB of VRAM.
 ## License
 LingBot-VA is released under Apache-2.0. See the
 [upstream repository](https://github.com/Robbyant/lingbot-va).
@@ -1,18 +0,0 @@
 # EVO1
 EVO1 is a Vision-Language-Action policy for robot control. The LeRobot
 integration uses an InternVL3 vision-language backbone with a flow-matching
 action head, and supports staged training through the standard LeRobot policy
 APIs.
 The upstream EVO1 project is available at
 [MINT-SJTU/Evo-1](https://github.com/MINT-SJTU/Evo-1).
 ```bibtex
@misc{evo1,
  title = {EVO1},
  author = {{MINT-SJTU}},
  year = {2025},
  howpublished = {\url{https://github.com/MINT-SJTU/Evo-1}},
 }
 ```
@@ -1,13 +1,6 @@
 ## Research Paper
-GR00T N1 technical report (covers the GR00T N1.x family, including N1.7): https://arxiv.org/abs/2503.14734
+Paper: https://research.nvidia.com/labs/gear/gr00t-n1_5/
 GR00T N1.7 model card: https://huggingface.co/nvidia/GR00T-N1.7-3B
 GR00T N1.5 research page (earlier version): https://research.nvidia.com/labs/gear/gr00t-n1_5/
 > GR00T N1.5 support was removed from LeRobot; the last release supporting it is `lerobot==0.5.1`.
 > Current releases support GR00T N1.7 only.
 ## Repository
@@ -31,108 +24,4 @@ Code: https://github.com/NVIDIA/Isaac-GR00T
 Blog: https://developer.nvidia.com/isaac/gr00t
-Hugging Face Models:
+Hugging Face Model: https://huggingface.co/nvidia/GR00T-N1.5-3B
 - GR00T N1.7: https://huggingface.co/nvidia/GR00T-N1.7-3B
 - GR00T N1.7 LIBERO checkpoints: https://huggingface.co/nvidia/GR00T-N1.7-LIBERO
 <details>
 <summary><b>Original-vs-LeRobot parity test</b></summary>
 ## Original-vs-LeRobot parity test
 `tests/policies/groot/test_groot_vs_original.py` verifies this LeRobot
 reimplementation of GR00T N1.7 (Qwen3-VL backbone + flow-matching action head)
 against NVIDIA's original `gr00t` package with two comparisons, each parametrized
 over every embodiment tag present in the checkpoint:
 1. **Model parity** — given byte-identical pre-processed inputs and the same
   flow-matching seed (recorded in each artifact), both implementations must produce
   the **same raw model output** (`get_action(...)["action_pred"]`, the normalized
   flow-matching prediction). Output shapes must match exactly; any action-horizon
   or action-dim mismatch fails the test.
 2. **Preprocessor parity** — given the identical raw observations (per-camera
   frames, state vectors, language instruction), LeRobot's own preprocessor pipeline
   (real Qwen3-VL chat template / tokenizer / image packing + checkpoint-driven
   state normalization, no mocks) must produce the **same collated model inputs**
   (`input_ids`, `attention_mask`, `pixel_values`, `image_grid_thw`, `state`,
   `embodiment_id`) as the original package's processor.
 ### Why two environments
 The original `gr00t` package pins `transformers==4.57.3` (Python 3.10); this
 integration requires `transformers>=5.x` (Qwen3-VL). Under 5.x, `PretrainedConfig`
 is itself a defaulted dataclass, so the original config dataclasses fail to import
 (`non-default argument follows default argument`). The two implementations therefore
 **cannot be imported in the same Python process**.
 So the test uses a **producer / consumer** split across two venvs:
 1. **Producer** — `tests/policies/groot/utils/dump_original_n1_7.py`, run in the _original_
   gr00t venv. For each embodiment it builds dummy inputs generically from the
   checkpoint metadata (state dims from `statistics.json`; camera/language keys from
   the processor modality configs), runs the original model, and saves to one `.npz`
   per tag: the raw observations (`raw::` keys), the exact collated inputs
   (`in::` keys), the seed, and the raw `action_pred`.
 2. **Consumer** — the pytest above, run in the _LeRobot_ venv. It discovers every
   `.npz`; the model-parity case replays the byte-identical collated inputs through
   the LeRobot model with the recorded seed and asserts the outputs match, and the
   preprocessor-parity case replays the raw observations through LeRobot's full
   preprocessor pipeline and asserts the collated tensors match.
 > Artifacts generated by older versions of the dump script contain no `raw::`
 > fields; the preprocessor-parity case then **skips** with a regeneration hint.
 > Re-run the producer to refresh them.
 ### Fairness controls
 - **Same pre-processed inputs (model parity)** — the original processor's `input_ids`,
  `pixel_values`, `image_grid_thw`, `attention_mask`, `state`, `embodiment_id` are
  fed verbatim to the LeRobot model (no re-tokenization / re-normalization), so the
  model comparison isolates the model. LeRobot's own tokenization / image packing is
  covered separately by the preprocessor-parity case, which compares its output
  against those same collated tensors from identical raw observations.
 - **Same precision + attention kernel** — both sides run **fp32 + SDPA**. The
  original defaults to `use_flash_attention=True` (flash_attention_2 + bf16); the
  producer forces SDPA + fp32. (With the defaults the gap is ~3e-2 — pure
  kernel/rounding noise, not an implementation difference.)
 - **Same flow-matching seed** — fixed right before sampling on both sides; the
  producer records it in each artifact (`--seed`, default 42) and the consumer
  replays the recorded value.
 ### How to run
 ```bash
 # Resolve a local checkpoint (GR00T-N1.7-LIBERO / libero_10)
 CKPT=$(python - <<'PY'
 import os
 from huggingface_hub import snapshot_download
 print(os.path.join(snapshot_download("nvidia/GR00T-N1.7-LIBERO",
      allow_patterns=["libero_10/*"]), "libero_10"))
 PY
 )
 # 1) Produce the original-side artifacts for all embodiments (original gr00t venv, CUDA)
 CUDA_VISIBLE_DEVICES=0 /path/to/Isaac-GR00T/.venv-original/bin/python \
    tests/policies/groot/utils/dump_original_n1_7.py \
    --ckpt "$CKPT" --out-dir tests/policies/groot/artifacts --device cuda --seed 42
 # 2) Run the parity test (LeRobot venv) — one parametrized case per embodiment
 CUDA_VISIBLE_DEVICES=0 GROOT_PARITY_DEVICE=cuda \
    uv run pytest tests/policies/groot/test_groot_vs_original.py -v -s
 ```
 The `.npz` artifacts are local-only (gitignored, ~6–10 MB each) and are regenerated by
 the producer; they are never committed. The tests **skip** (do not fail) on CI or
 when the checkpoint / artifacts are absent.
 #### Env knobs (all optional)
 | Var                                       | Default                          | Purpose                               |
 | ----------------------------------------- | -------------------------------- | ------------------------------------- |
 | `GROOT_N1_7_PARITY_DIR`                   | `tests/policies/groot/artifacts` | directory of per-tag `.npz` artifacts |
 | `GROOT_N1_7_LIBERO_CKPT`                  | auto (HF cache)                  | override checkpoint dir               |
 | `GROOT_PARITY_DEVICE`                     | `cuda` if available              | `cpu` or `cuda`                       |
 | `GROOT_PARITY_ATOL` / `GROOT_PARITY_RTOL` | `1e-3`                           | comparison tolerance                  |
 </details>
@@ -265,8 +265,6 @@ lerobot-dataset-viz \
 Once executed, the tool opens `rerun.io` and displays the camera streams, robot states, and actions for the selected episode.
 To use [Foxglove](https://foxglove.dev) instead of Rerun, install the extra add `--display-mode foxglove`. This starts a WebSocket server (connect the Foxglove app to `ws://127.0.0.1:8765`) that serves the episode as a seekable timeline you can play/pause and scrub.
 For advanced usage—including visualizing datasets stored on a remote server—run:
 ```bash
@@ -134,6 +134,9 @@ lerobot-train \
 > [!TIP]
 > This is purely a decode-time presentation choice — it does **not** alter the stored video or its metadata, so the same dataset can be read as `mm` or `m` without re-encoding. It has no effect on datasets without depth cameras.
 > [!IMPORTANT]
 > Depth statistics in `meta/stats.json` are always computed in **millimetres**, regardless of the raw frame dtype.
 ---
 ## Persistence in dataset metadata
@@ -125,7 +125,6 @@ hardware = [
 ]
 viz = [
    "rerun-sdk>=0.24.0,<0.34.0",
    "foxglove-sdk>=0.25.1,<0.26.0",
 ]
 # ── User-facing composite extras (map to CLI scripts) ─────
 # lerobot-record, lerobot-replay, lerobot-calibrate, lerobot-teleoperate, etc.
@@ -164,7 +163,6 @@ pynput-dep = ["pynput>=1.7.8,<1.9.0"]
 pyzmq-dep = ["pyzmq>=26.2.1,<28.0.0"]
 motorbridge-dep = ["motorbridge>=0.3.2,<0.4.0"]
 motorbridge-smart-servo-dep = ["motorbridge-smart-servo>=0.0.4,<0.1.0"]
 timm-dep = ["timm>=1.0.0,<1.1.0"]
 # Motors
 feetech = ["feetech-servo-sdk>=1.0.0,<2.0.0", "lerobot[pyserial-dep]", "lerobot[deepdiff-dep]"]
@@ -220,10 +218,11 @@ groot = [
    "lerobot[transformers-dep]",
    "lerobot[peft-dep]",
    "lerobot[diffusers-dep]",
    "lerobot[dataset]", # NOTE: processor_groot builds a LeRobotDataset for relative-action training stats
    "dm-tree>=0.1.8,<1.0.0",
-    "lerobot[timm-dep]",
+    "timm>=1.0.0,<1.1.0",
    "decord>=0.6.0,<1.0.0; (platform_machine == 'AMD64' or platform_machine == 'x86_64')",
    "ninja>=1.11.1,<2.0.0",
    "flash-attn>=2.5.9,<3.0.0 ; sys_platform != 'darwin'"
 ]
 sarm = ["lerobot[transformers-dep]", "pydantic>=2.0.0,<3.0.0", "faker>=33.0.0,<35.0.0", "lerobot[matplotlib-dep]", "lerobot[qwen-vl-utils-dep]"]
 robometer = ["lerobot[transformers-dep]", "lerobot[qwen-vl-utils-dep]", "lerobot[peft-dep]"]
@@ -234,10 +233,8 @@ fastwam = [
    "lerobot[transformers-dep]",
    "lerobot[diffusers-dep]",
 ]
 evo1 = ["lerobot[transformers-dep]"]
 hilserl = ["lerobot[transformers-dep]", "lerobot[dataset]", "gym-hil>=0.1.14,<0.2.0", "lerobot[grpcio-dep]", "lerobot[placo-dep]"]
 vla_jepa = ["lerobot[transformers-dep]", "lerobot[diffusers-dep]", "lerobot[qwen-vl-utils-dep]"]
 lingbot_va = ["lerobot[transformers-dep]", "lerobot[diffusers-dep]", "lerobot[accelerate-dep]"]
 # Features
 async = ["lerobot[grpcio-dep]", "lerobot[matplotlib-dep]"]
@@ -316,12 +313,10 @@ all = [
    "lerobot[molmoact2]",
    "lerobot[smolvla]",
    "lerobot[fastwam]",
-    "lerobot[groot]",
+    # "lerobot[groot]", TODO(Steven): Gr00t requires specific installation instructions for flash-attn
    "lerobot[xvla]",
    "lerobot[evo1]",
    "lerobot[hilserl]",
    "lerobot[vla_jepa]",
    "lerobot[lingbot_va]",
    "lerobot[async]",
    "lerobot[dev]",
    "lerobot[test]",
@@ -0,0 +1,729 @@
 #
 # This file is autogenerated by pip-compile with Python 3.12
 # by the following command:
 #
 #    pip-compile --output-file=requirements-macos.txt requirements.in
 #
 -e .[all]
    # via -[all]
 absl-py==2.4.0
    # via
    #   dm-control
    #   dm-env
    #   dm-tree
    #   labmaze
    #   mujoco
 accelerate==1.13.0
    # via
    #   lerobot
    #   peft
 aiohappyeyeballs==2.6.1
    # via aiohttp
 aiohttp==3.13.3
    # via fsspec
 aiosignal==1.4.0
    # via aiohttp
 annotated-doc==0.0.4
    # via
    #   fastapi
    #   typer
 annotated-types==0.7.0
    # via pydantic
 anyio==4.12.1
    # via
    #   httpx
    #   starlette
    #   watchfiles
 asttokens==3.0.1
    # via stack-data
 attrs==25.4.0
    # via
    #   aiohttp
    #   dm-tree
    #   jsonlines
    #   rerun-sdk
 av==15.1.0
    # via
    #   lerobot
    #   qwen-vl-utils
 certifi==2026.2.25
    # via
    #   httpcore
    #   httpx
    #   requests
    #   sentry-sdk
 cffi==2.0.0
    # via pymunk
 cfgv==3.5.0
    # via pre-commit
 charset-normalizer==3.4.5
    # via requests
 click==8.3.1
    # via
    #   typer
    #   uvicorn
    #   wandb
 cloudpickle==3.1.2
    # via gymnasium
 cmake==4.1.3
    # via lerobot
 cmeel==0.59.0
    # via
    #   cmeel-assimp
    #   cmeel-boost
    #   cmeel-console-bridge
    #   cmeel-octomap
    #   cmeel-qhull
    #   cmeel-tinyxml2
    #   cmeel-urdfdom
    #   cmeel-zlib
    #   coal-library
    #   eigenpy
    #   eiquadprog
    #   pin
    #   placo
    #   rhoban-cmeel-jsoncpp
 cmeel-assimp==5.4.3.1
    # via coal-library
 cmeel-boost==1.87.0.1
    # via
    #   coal-library
    #   eigenpy
    #   eiquadprog
    #   pin
 cmeel-console-bridge==1.0.2.3
    # via cmeel-urdfdom
 cmeel-octomap==1.10.0
    # via coal-library
 cmeel-qhull==8.0.2.1
    # via coal-library
 cmeel-tinyxml2==10.0.0
    # via cmeel-urdfdom
 cmeel-urdfdom==4.0.1
    # via pin
 cmeel-zlib==1.3.1
    # via cmeel-assimp
 coal-library==3.0.1
    # via pin
 contourpy==1.3.3
    # via
    #   lerobot
    #   matplotlib
 coverage[toml]==7.13.4
    # via pytest-cov
 cycler==0.12.1
    # via matplotlib
 datasets==4.6.1
    # via lerobot
 debugpy==1.8.20
    # via lerobot
 decorator==5.2.1
    # via ipython
 deepdiff==8.6.1
    # via lerobot
 diffusers==0.35.2
    # via lerobot
 dill==0.4.0
    # via
    #   datasets
    #   multiprocess
 distlib==0.4.0
    # via virtualenv
 dm-control==1.0.37
    # via gym-aloha
 dm-env==1.6
    # via dm-control
 dm-tree==0.1.9
    # via
    #   dm-control
    #   dm-env
 docopt==0.6.2
    # via num2words
 draccus==0.10.0
    # via lerobot
 dynamixel-sdk==3.8.4
    # via lerobot
 eigenpy==3.10.3
    # via coal-library
 einops==0.8.2
    # via lerobot
 eiquadprog==1.2.9
    # via placo
 etils[epath,epy]==1.14.0
    # via mujoco
 executing==2.2.1
    # via stack-data
 faker==34.0.2
    # via lerobot
 farama-notifications==0.0.4
    # via gymnasium
 fastapi==0.135.1
    # via
    #   lerobot
    #   teleop
 feetech-servo-sdk==1.0.0
    # via lerobot
 filelock==3.25.0
    # via
    #   datasets
    #   diffusers
    #   huggingface-hub
    #   python-discovery
    #   torch
    #   virtualenv
 fonttools==4.61.1
    # via matplotlib
 frozenlist==1.8.0
    # via
    #   aiohttp
    #   aiosignal
 fsspec[http]==2026.2.0
    # via
    #   datasets
    #   etils
    #   huggingface-hub
    #   torch
 gitdb==4.0.12
    # via gitpython
 gitpython==3.1.46
    # via wandb
 glfw==2.10.0
    # via
    #   dm-control
    #   mujoco
 grpcio==1.73.1
    # via
    #   grpcio-tools
    #   lerobot
    #   reachy2-sdk
    #   reachy2-sdk-api
 grpcio-tools==1.73.1
    # via
    #   lerobot
    #   reachy2-sdk-api
 gym-aloha==0.1.3
    # via lerobot
 gym-hil==0.1.13
    # via lerobot
 gym-pusht==0.1.6
    # via lerobot
 gymnasium==1.2.3
    # via
    #   gym-aloha
    #   gym-hil
    #   gym-pusht
    #   lerobot
    #   metaworld
 h11==0.16.0
    # via
    #   httpcore
    #   uvicorn
 hebi-py==2.11.0
    # via lerobot
 hf-xet==1.3.2
    # via huggingface-hub
 hidapi==0.14.0.post4
    # via
    #   gym-hil
    #   lerobot
 httpcore==1.0.9
    # via httpx
 httptools==0.7.1
    # via uvicorn
 httpx==0.28.1
    # via
    #   datasets
    #   huggingface-hub
 huggingface-hub==1.6.0
    # via
    #   accelerate
    #   datasets
    #   diffusers
    #   lerobot
    #   peft
    #   tokenizers
    #   transformers
 identify==2.6.17
    # via pre-commit
 idna==3.11
    # via
    #   anyio
    #   httpx
    #   requests
    #   yarl
 imageio[ffmpeg]==2.37.2
    # via
    #   gym-aloha
    #   gym-hil
    #   lerobot
    #   metaworld
    #   scikit-image
 imageio-ffmpeg==0.6.0
    # via imageio
 importlib-metadata==8.7.1
    # via diffusers
 iniconfig==2.3.0
    # via pytest
 ipython==9.11.0
    # via meshcat
 ipython-pygments-lexers==1.1.1
    # via ipython
 ischedule==1.2.7
    # via placo
 jedi==0.19.2
    # via ipython
 jinja2==3.1.6
    # via torch
 jsonlines==4.0.0
    # via lerobot
 kiwisolver==1.4.9
    # via matplotlib
 labmaze==1.0.6
    # via dm-control
 lazy-loader==0.5
    # via scikit-image
 librt==0.8.1
    # via mypy
 lxml==6.0.2
    # via dm-control
 markdown-it-py==4.0.0
    # via rich
 markupsafe==3.0.3
    # via jinja2
 matplotlib==3.10.8
    # via lerobot
 matplotlib-inline==0.2.1
    # via ipython
 mdurl==0.1.2
    # via markdown-it-py
 mergedeep==1.3.4
    # via draccus
 meshcat==0.3.2
    # via placo
 metaworld==3.0.0
    # via lerobot
 mock-serial==0.0.1
    # via lerobot
 mpmath==1.3.0
    # via sympy
 mujoco==3.5.0
    # via
    #   dm-control
    #   gym-aloha
    #   gym-hil
    #   metaworld
 multidict==6.7.1
    # via
    #   aiohttp
    #   yarl
 multiprocess==0.70.18
    # via datasets
 mypy==1.19.1
    # via lerobot
 mypy-extensions==1.1.0
    # via
    #   mypy
    #   typing-inspect
 networkx==3.6.1
    # via
    #   scikit-image
    #   torch
 nodeenv==1.10.0
    # via pre-commit
 num2words==0.5.14
    # via lerobot
 numpy==2.2.6
    # via
    #   accelerate
    #   cmeel-boost
    #   contourpy
    #   datasets
    #   diffusers
    #   dm-control
    #   dm-env
    #   dm-tree
    #   gymnasium
    #   hebi-py
    #   imageio
    #   labmaze
    #   lerobot
    #   matplotlib
    #   meshcat
    #   metaworld
    #   mujoco
    #   opencv-python
    #   opencv-python-headless
    #   pandas
    #   peft
    #   pyquaternion
    #   reachy2-sdk
    #   rerun-sdk
    #   scikit-image
    #   scipy
    #   shapely
    #   teleop
    #   tifffile
    #   torchvision
    #   transformers
    #   transforms3d
 opencv-python==4.13.0.92
    # via
    #   gym-pusht
    #   reachy2-sdk
 opencv-python-headless==4.12.0.88
    # via lerobot
 orderly-set==5.5.0
    # via deepdiff
 packaging==25.0
    # via
    #   accelerate
    #   datasets
    #   huggingface-hub
    #   lazy-loader
    #   lerobot
    #   matplotlib
    #   peft
    #   pytest
    #   qwen-vl-utils
    #   reachy2-sdk
    #   scikit-image
    #   transformers
    #   wandb
 pandas==2.3.3
    # via
    #   datasets
    #   lerobot
 parso==0.8.6
    # via jedi
 pathspec==1.0.4
    # via mypy
 peft==0.18.1
    # via lerobot
 pexpect==4.9.0
    # via ipython
 pillow==12.1.1
    # via
    #   diffusers
    #   imageio
    #   matplotlib
    #   meshcat
    #   qwen-vl-utils
    #   rerun-sdk
    #   scikit-image
    #   torchvision
 pin==3.4.0
    # via placo
 placo==0.9.16
    # via lerobot
 platformdirs==4.9.4
    # via
    #   python-discovery
    #   virtualenv
    #   wandb
 pluggy==1.6.0
    # via
    #   pytest
    #   pytest-cov
 pre-commit==4.5.1
    # via lerobot
 prompt-toolkit==3.0.52
    # via ipython
 propcache==0.4.1
    # via
    #   aiohttp
    #   yarl
 protobuf==6.31.1
    # via
    #   dm-control
    #   grpcio-tools
    #   lerobot
    #   reachy2-sdk
    #   reachy2-sdk-api
    #   wandb
 psutil==7.2.2
    # via
    #   accelerate
    #   imageio
    #   peft
 ptyprocess==0.7.0
    # via pexpect
 pure-eval==0.2.3
    # via stack-data
 pyarrow==23.0.1
    # via
    #   datasets
    #   rerun-sdk
 pycparser==3.0
    # via cffi
 pydantic==2.12.5
    # via
    #   fastapi
    #   wandb
 pydantic-core==2.41.5
    # via pydantic
 pygame==2.6.1
    # via
    #   gym-hil
    #   gym-pusht
    #   lerobot
 pygments==2.19.2
    # via
    #   ipython
    #   ipython-pygments-lexers
    #   pytest
    #   rich
 pymunk==6.11.1
    # via
    #   gym-pusht
    #   lerobot
 pyngrok==7.5.1
    # via meshcat
 pynput==1.8.1
    # via
    #   gym-hil
    #   lerobot
 pyobjc-core==12.1
    # via
    #   pyobjc-framework-applicationservices
    #   pyobjc-framework-cocoa
    #   pyobjc-framework-coretext
    #   pyobjc-framework-quartz
 pyobjc-framework-applicationservices==12.1
    # via pynput
 pyobjc-framework-cocoa==12.1
    # via
    #   pyobjc-framework-applicationservices
    #   pyobjc-framework-coretext
    #   pyobjc-framework-quartz
 pyobjc-framework-coretext==12.1
    # via pyobjc-framework-applicationservices
 pyobjc-framework-quartz==12.1
    # via
    #   pynput
    #   pyobjc-framework-applicationservices
    #   pyobjc-framework-coretext
 pyopengl==3.1.10
    # via
    #   dm-control
    #   mujoco
 pyparsing==3.3.2
    # via
    #   dm-control
    #   matplotlib
 pyquaternion==0.9.9
    # via reachy2-sdk
 pyrealsense2-macosx==2.56.5
    # via lerobot
 pyserial==3.5
    # via
    #   dynamixel-sdk
    #   feetech-servo-sdk
    #   lerobot
 pytest==8.4.2
    # via
    #   lerobot
    #   pytest-cov
    #   pytest-timeout
    #   teleop
 pytest-cov==7.0.0
    # via lerobot
 pytest-timeout==2.4.0
    # via lerobot
 python-dateutil==2.9.0.post0
    # via
    #   faker
    #   matplotlib
    #   pandas
 python-discovery==1.1.1
    # via virtualenv
 python-dotenv==1.2.2
    # via uvicorn
 pytz==2026.1.post1
    # via pandas
 pyyaml==6.0.3
    # via
    #   accelerate
    #   datasets
    #   draccus
    #   hebi-py
    #   huggingface-hub
    #   peft
    #   pre-commit
    #   pyngrok
    #   pyyaml-include
    #   transformers
    #   uvicorn
    #   wandb
 pyyaml-include==1.4.1
    # via draccus
 pyzmq==27.1.0
    # via
    #   lerobot
    #   meshcat
 qwen-vl-utils==0.0.14
    # via lerobot
 reachy2-sdk==1.0.15
    # via lerobot
 reachy2-sdk-api==1.0.21
    # via reachy2-sdk
 regex==2026.2.28
    # via
    #   diffusers
    #   transformers
 requests==2.32.5
    # via
    #   datasets
    #   diffusers
    #   dm-control
    #   qwen-vl-utils
    #   teleop
    #   wandb
 rerun-sdk==0.26.2
    # via lerobot
 rhoban-cmeel-jsoncpp==1.9.4.9
    # via placo
 rich==14.3.3
    # via typer
 safetensors==0.7.0
    # via
    #   accelerate
    #   diffusers
    #   lerobot
    #   peft
    #   transformers
 scikit-image==0.25.2
    # via
    #   gym-pusht
    #   lerobot
 scipy==1.17.1
    # via
    #   dm-control
    #   lerobot
    #   metaworld
    #   scikit-image
    #   torchdiffeq
 sentry-sdk==2.54.0
    # via wandb
 shapely==2.1.2
    # via gym-pusht
 shellingham==1.5.4
    # via typer
 six==1.17.0
    # via
    #   pynput
    #   python-dateutil
 smmap==5.0.3
    # via gitdb
 stack-data==0.6.3
    # via ipython
 starlette==0.52.1
    # via fastapi
 sympy==1.14.0
    # via torch
 teleop==0.1.4
    # via lerobot
 termcolor==3.3.0
    # via lerobot
 tifffile==2026.3.3
    # via scikit-image
 tokenizers==0.22.2
    # via transformers
 toml==0.10.2
    # via draccus
 torch==2.10.0
    # via
    #   accelerate
    #   lerobot
    #   peft
    #   torchdiffeq
    #   torchvision
 torchcodec==0.10.0
    # via lerobot
 torchdiffeq==0.2.5
    # via lerobot
 torchvision==0.25.0
    # via lerobot
 tornado==6.5.4
    # via meshcat
 tqdm==4.67.3
    # via
    #   datasets
    #   dm-control
    #   huggingface-hub
    #   peft
    #   transformers
 traitlets==5.14.3
    # via
    #   ipython
    #   matplotlib-inline
 transformers==5.3.0
    # via
    #   lerobot
    #   peft
 transforms3d==0.4.2
    # via teleop
 typer==0.24.1
    # via
    #   huggingface-hub
    #   transformers
 typing-extensions==4.15.0
    # via
    #   aiosignal
    #   anyio
    #   etils
    #   faker
    #   fastapi
    #   gymnasium
    #   huggingface-hub
    #   mypy
    #   pydantic
    #   pydantic-core
    #   rerun-sdk
    #   starlette
    #   torch
    #   typing-inspect
    #   typing-inspection
    #   wandb
 typing-inspect==0.9.0
    # via draccus
 typing-inspection==0.4.2
    # via
    #   fastapi
    #   pydantic
 tzdata==2025.3
    # via pandas
 u-msgpack-python==2.8.0
    # via meshcat
 urllib3==2.6.3
    # via
    #   requests
    #   sentry-sdk
 uvicorn[standard]==0.41.0
    # via teleop
 uvloop==0.22.1
    # via uvicorn
 virtualenv==21.1.0
    # via pre-commit
 wandb==0.24.2
    # via lerobot
 watchfiles==1.1.1
    # via uvicorn
 wcwidth==0.6.0
    # via prompt-toolkit
 websocket-client==1.9.0
    # via teleop
 websockets==16.0
    # via uvicorn
 wrapt==2.1.2
    # via dm-tree
 xxhash==3.6.0
    # via datasets
 yarl==1.23.0
    # via aiohttp
 zipp==3.23.0
    # via
    #   etils
    #   importlib-metadata
 # The following packages are considered to be unsafe in a requirements file:
 # setuptools
@@ -0,0 +1,882 @@
 #
 # This file is autogenerated by pip-compile with Python 3.12
 # by the following command:
 #
 #    pip-compile --output-file=requirements-ubuntu.txt requirements.in
 #
 -e .[all]
    # via -[all]
 absl-py==2.4.0
    # via
    #   dm-control
    #   dm-env
    #   dm-tree
    #   labmaze
    #   mujoco
    #   tensorboard
 accelerate==1.13.0
    # via
    #   lerobot
    #   peft
 aiohappyeyeballs==2.6.1
    # via aiohttp
 aiohttp==3.13.3
    # via fsspec
 aiosignal==1.4.0
    # via aiohttp
 annotated-doc==0.0.4
    # via
    #   fastapi
    #   typer
 annotated-types==0.7.0
    # via pydantic
 antlr4-python3-runtime==4.9.3
    # via
    #   hydra-core
    #   omegaconf
 anyio==4.12.1
    # via
    #   httpx
    #   starlette
    #   watchfiles
 asttokens==3.0.1
    # via stack-data
 attrs==25.4.0
    # via
    #   aiohttp
    #   dm-tree
    #   jsonlines
    #   jsonschema
    #   referencing
    #   rerun-sdk
 av==15.1.0
    # via
    #   lerobot
    #   qwen-vl-utils
 bddl==1.0.1
    # via hf-libero
 certifi==2026.2.25
    # via
    #   httpcore
    #   httpx
    #   requests
    #   sentry-sdk
 cffi==2.0.0
    # via pymunk
 cfgv==3.5.0
    # via pre-commit
 charset-normalizer==3.4.5
    # via requests
 click==8.3.1
    # via
    #   typer
    #   uvicorn
    #   wandb
 cloudpickle==3.1.2
    # via
    #   gymnasium
    #   hf-libero
 cmake==4.1.3
    # via lerobot
 cmeel==0.59.0
    # via
    #   cmeel-assimp
    #   cmeel-boost
    #   cmeel-console-bridge
    #   cmeel-octomap
    #   cmeel-qhull
    #   cmeel-tinyxml2
    #   cmeel-urdfdom
    #   cmeel-zlib
    #   coal-library
    #   eigenpy
    #   eiquadprog
    #   pin
    #   placo
    #   rhoban-cmeel-jsoncpp
 cmeel-assimp==5.4.3.1
    # via coal-library
 cmeel-boost==1.87.0.1
    # via
    #   coal-library
    #   eigenpy
    #   eiquadprog
    #   pin
 cmeel-console-bridge==1.0.2.3
    # via cmeel-urdfdom
 cmeel-octomap==1.10.0
    # via coal-library
 cmeel-qhull==8.0.2.1
    # via coal-library
 cmeel-tinyxml2==10.0.0
    # via cmeel-urdfdom
 cmeel-urdfdom==4.0.1
    # via pin
 cmeel-zlib==1.3.1
    # via cmeel-assimp
 coal-library==3.0.1
    # via pin
 contourpy==1.3.3
    # via
    #   lerobot
    #   matplotlib
 coverage[toml]==7.13.4
    # via pytest-cov
 cuda-bindings==12.9.4
    # via torch
 cuda-pathfinder==1.4.1
    # via cuda-bindings
 cycler==0.12.1
    # via matplotlib
 datasets==4.6.1
    # via lerobot
 debugpy==1.8.20
    # via lerobot
 decorator==5.2.1
    # via ipython
 deepdiff==8.6.1
    # via lerobot
 diffusers==0.35.2
    # via lerobot
 dill==0.4.0
    # via
    #   datasets
    #   multiprocess
 distlib==0.4.0
    # via virtualenv
 dm-control==1.0.37
    # via gym-aloha
 dm-env==1.6
    # via dm-control
 dm-tree==0.1.9
    # via
    #   dm-control
    #   dm-env
 docopt==0.6.2
    # via num2words
 draccus==0.10.0
    # via lerobot
 dynamixel-sdk==3.8.4
    # via lerobot
 easydict==1.13
    # via hf-libero
 egl-probe==1.0.2
    # via robomimic
 eigenpy==3.10.3
    # via coal-library
 einops==0.8.2
    # via
    #   hf-libero
    #   lerobot
 eiquadprog==1.2.9
    # via placo
 etils[epath,epy]==1.14.0
    # via mujoco
 evdev==1.9.3
    # via pynput
 executing==2.2.1
    # via stack-data
 faker==34.0.2
    # via lerobot
 farama-notifications==0.0.4
    # via gymnasium
 fastapi==0.135.1
    # via
    #   lerobot
    #   teleop
 fastjsonschema==2.21.2
    # via nbformat
 feetech-servo-sdk==1.0.0
    # via lerobot
 filelock==3.25.0
    # via
    #   datasets
    #   diffusers
    #   huggingface-hub
    #   python-discovery
    #   torch
    #   virtualenv
 fonttools==4.61.1
    # via matplotlib
 frozenlist==1.8.0
    # via
    #   aiohttp
    #   aiosignal
 fsspec[http]==2026.2.0
    # via
    #   datasets
    #   etils
    #   huggingface-hub
    #   torch
 future==1.0.0
    # via hf-libero
 gitdb==4.0.12
    # via gitpython
 gitpython==3.1.46
    # via wandb
 glfw==2.10.0
    # via
    #   dm-control
    #   mujoco
 grpcio==1.73.1
    # via
    #   grpcio-tools
    #   lerobot
    #   reachy2-sdk
    #   reachy2-sdk-api
    #   tensorboard
 grpcio-tools==1.73.1
    # via
    #   lerobot
    #   reachy2-sdk-api
 gym-aloha==0.1.3
    # via lerobot
 gym-hil==0.1.13
    # via lerobot
 gym-pusht==0.1.6
    # via lerobot
 gymnasium==1.2.3
    # via
    #   gym-aloha
    #   gym-hil
    #   gym-pusht
    #   hf-libero
    #   lerobot
    #   metaworld
 h11==0.16.0
    # via
    #   httpcore
    #   uvicorn
 h5py==3.16.0
    # via robomimic
 hebi-py==2.11.0
    # via lerobot
 hf-egl-probe==1.0.2
    # via hf-libero
 hf-libero==0.1.3
    # via lerobot
 hf-xet==1.3.2
    # via huggingface-hub
 hidapi==0.14.0.post4
    # via
    #   gym-hil
    #   lerobot
 httpcore==1.0.9
    # via httpx
 httptools==0.7.1
    # via uvicorn
 httpx==0.28.1
    # via
    #   datasets
    #   huggingface-hub
 huggingface-hub==1.6.0
    # via
    #   accelerate
    #   datasets
    #   diffusers
    #   lerobot
    #   peft
    #   tokenizers
    #   transformers
 hydra-core==1.3.2
    # via hf-libero
 identify==2.6.17
    # via pre-commit
 idna==3.11
    # via
    #   anyio
    #   httpx
    #   requests
    #   yarl
 imageio[ffmpeg]==2.37.2
    # via
    #   gym-aloha
    #   gym-hil
    #   lerobot
    #   metaworld
    #   robomimic
    #   scikit-image
 imageio-ffmpeg==0.6.0
    # via
    #   imageio
    #   robomimic
 importlib-metadata==8.7.1
    # via diffusers
 iniconfig==2.3.0
    # via pytest
 ipython==9.11.0
    # via meshcat
 ipython-pygments-lexers==1.1.1
    # via ipython
 ischedule==1.2.7
    # via placo
 jedi==0.19.2
    # via ipython
 jinja2==3.1.6
    # via torch
 jsonlines==4.0.0
    # via lerobot
 jsonschema==4.26.0
    # via nbformat
 jsonschema-specifications==2025.9.1
    # via jsonschema
 jupyter-core==5.9.1
    # via nbformat
 jupytext==1.19.1
    # via bddl
 kiwisolver==1.4.9
    # via matplotlib
 labmaze==1.0.6
    # via dm-control
 lazy-loader==0.5
    # via scikit-image
 librt==0.8.1
    # via mypy
 llvmlite==0.46.0
    # via numba
 lxml==6.0.2
    # via dm-control
 markdown==3.10.2
    # via tensorboard
 markdown-it-py==4.0.0
    # via
    #   jupytext
    #   mdit-py-plugins
    #   rich
 markupsafe==3.0.3
    # via
    #   jinja2
    #   werkzeug
 matplotlib==3.10.8
    # via
    #   hf-libero
    #   lerobot
 matplotlib-inline==0.2.1
    # via ipython
 mdit-py-plugins==0.5.0
    # via jupytext
 mdurl==0.1.2
    # via markdown-it-py
 mergedeep==1.3.4
    # via draccus
 meshcat==0.3.2
    # via placo
 metaworld==3.0.0
    # via lerobot
 mock-serial==0.0.1
    # via lerobot
 mpmath==1.3.0
    # via sympy
 mujoco==3.5.0
    # via
    #   dm-control
    #   gym-aloha
    #   gym-hil
    #   hf-libero
    #   metaworld
    #   robosuite
 multidict==6.7.1
    # via
    #   aiohttp
    #   yarl
 multiprocess==0.70.18
    # via datasets
 mypy==1.19.1
    # via lerobot
 mypy-extensions==1.1.0
    # via
    #   mypy
    #   typing-inspect
 nbformat==5.10.4
    # via jupytext
 networkx==3.6.1
    # via
    #   bddl
    #   scikit-image
    #   torch
 nodeenv==1.10.0
    # via pre-commit
 num2words==0.5.14
    # via lerobot
 numba==0.64.0
    # via robosuite
 numpy==2.2.6
    # via
    #   accelerate
    #   bddl
    #   cmeel-boost
    #   contourpy
    #   datasets
    #   diffusers
    #   dm-control
    #   dm-env
    #   dm-tree
    #   gymnasium
    #   h5py
    #   hebi-py
    #   hf-libero
    #   imageio
    #   labmaze
    #   lerobot
    #   matplotlib
    #   meshcat
    #   metaworld
    #   mujoco
    #   numba
    #   opencv-python
    #   opencv-python-headless
    #   pandas
    #   peft
    #   pyquaternion
    #   reachy2-sdk
    #   rerun-sdk
    #   robomimic
    #   robosuite
    #   scikit-image
    #   scipy
    #   shapely
    #   teleop
    #   tensorboard
    #   tensorboardx
    #   tifffile
    #   torchvision
    #   transformers
    #   transforms3d
 nvidia-cublas-cu12==12.8.4.1
    # via
    #   nvidia-cudnn-cu12
    #   nvidia-cusolver-cu12
    #   torch
 nvidia-cuda-cupti-cu12==12.8.90
    # via torch
 nvidia-cuda-nvrtc-cu12==12.8.93
    # via torch
 nvidia-cuda-runtime-cu12==12.8.90
    # via torch
 nvidia-cudnn-cu12==9.10.2.21
    # via torch
 nvidia-cufft-cu12==11.3.3.83
    # via torch
 nvidia-cufile-cu12==1.13.1.3
    # via torch
 nvidia-curand-cu12==10.3.9.90
    # via torch
 nvidia-cusolver-cu12==11.7.3.90
    # via torch
 nvidia-cusparse-cu12==12.5.8.93
    # via
    #   nvidia-cusolver-cu12
    #   torch
 nvidia-cusparselt-cu12==0.7.1
    # via torch
 nvidia-nccl-cu12==2.27.5
    # via torch
 nvidia-nvjitlink-cu12==12.8.93
    # via
    #   nvidia-cufft-cu12
    #   nvidia-cusolver-cu12
    #   nvidia-cusparse-cu12
    #   torch
 nvidia-nvshmem-cu12==3.4.5
    # via torch
 nvidia-nvtx-cu12==12.8.90
    # via torch
 omegaconf==2.3.0
    # via hydra-core
 opencv-python==4.13.0.92
    # via
    #   gym-pusht
    #   hf-libero
    #   reachy2-sdk
    #   robosuite
 opencv-python-headless==4.12.0.88
    # via lerobot
 orderly-set==5.5.0
    # via deepdiff
 packaging==25.0
    # via
    #   accelerate
    #   datasets
    #   huggingface-hub
    #   hydra-core
    #   jupytext
    #   lazy-loader
    #   lerobot
    #   matplotlib
    #   peft
    #   pytest
    #   qwen-vl-utils
    #   reachy2-sdk
    #   scikit-image
    #   tensorboard
    #   tensorboardx
    #   transformers
    #   wandb
 pandas==2.3.3
    # via
    #   datasets
    #   lerobot
 parso==0.8.6
    # via jedi
 pathspec==1.0.4
    # via mypy
 peft==0.18.1
    # via lerobot
 pexpect==4.9.0
    # via ipython
 pillow==12.1.1
    # via
    #   diffusers
    #   imageio
    #   matplotlib
    #   meshcat
    #   qwen-vl-utils
    #   rerun-sdk
    #   robosuite
    #   scikit-image
    #   tensorboard
    #   torchvision
 pin==3.4.0
    # via placo
 placo==0.9.16
    # via lerobot
 platformdirs==4.9.4
    # via
    #   jupyter-core
    #   python-discovery
    #   virtualenv
    #   wandb
 pluggy==1.6.0
    # via
    #   pytest
    #   pytest-cov
 pre-commit==4.5.1
    # via lerobot
 prompt-toolkit==3.0.52
    # via ipython
 propcache==0.4.1
    # via
    #   aiohttp
    #   yarl
 protobuf==6.31.1
    # via
    #   dm-control
    #   grpcio-tools
    #   lerobot
    #   reachy2-sdk
    #   reachy2-sdk-api
    #   tensorboard
    #   tensorboardx
    #   wandb
 psutil==7.2.2
    # via
    #   accelerate
    #   imageio
    #   peft
    #   robomimic
 ptyprocess==0.7.0
    # via pexpect
 pure-eval==0.2.3
    # via stack-data
 pyarrow==23.0.1
    # via
    #   datasets
    #   rerun-sdk
 pycparser==3.0
    # via cffi
 pydantic==2.12.5
    # via
    #   fastapi
    #   wandb
 pydantic-core==2.41.5
    # via pydantic
 pygame==2.6.1
    # via
    #   gym-hil
    #   gym-pusht
    #   lerobot
 pygments==2.19.2
    # via
    #   ipython
    #   ipython-pygments-lexers
    #   pytest
    #   rich
 pymunk==6.11.1
    # via
    #   gym-pusht
    #   lerobot
 pyngrok==7.5.1
    # via meshcat
 pynput==1.8.1
    # via
    #   gym-hil
    #   lerobot
 pyopengl==3.1.10
    # via
    #   dm-control
    #   mujoco
 pyparsing==3.3.2
    # via
    #   dm-control
    #   matplotlib
 pyquaternion==0.9.9
    # via reachy2-sdk
 pyrealsense2==2.56.5.9235
    # via lerobot
 pyserial==3.5
    # via
    #   dynamixel-sdk
    #   feetech-servo-sdk
    #   lerobot
 pytest==8.4.2
    # via
    #   bddl
    #   lerobot
    #   pytest-cov
    #   pytest-timeout
    #   teleop
 pytest-cov==7.0.0
    # via lerobot
 pytest-timeout==2.4.0
    # via lerobot
 python-dateutil==2.9.0.post0
    # via
    #   faker
    #   matplotlib
    #   pandas
 python-discovery==1.1.1
    # via virtualenv
 python-dotenv==1.2.2
    # via uvicorn
 python-xlib==0.33
    # via pynput
 pytz==2026.1.post1
    # via pandas
 pyyaml==6.0.3
    # via
    #   accelerate
    #   datasets
    #   draccus
    #   hebi-py
    #   huggingface-hub
    #   jupytext
    #   omegaconf
    #   peft
    #   pre-commit
    #   pyngrok
    #   pyyaml-include
    #   transformers
    #   uvicorn
    #   wandb
 pyyaml-include==1.4.1
    # via draccus
 pyzmq==27.1.0
    # via
    #   lerobot
    #   meshcat
 qwen-vl-utils==0.0.14
    # via lerobot
 reachy2-sdk==1.0.15
    # via lerobot
 reachy2-sdk-api==1.0.21
    # via reachy2-sdk
 referencing==0.37.0
    # via
    #   jsonschema
    #   jsonschema-specifications
 regex==2026.2.28
    # via
    #   diffusers
    #   transformers
 requests==2.32.5
    # via
    #   datasets
    #   diffusers
    #   dm-control
    #   qwen-vl-utils
    #   teleop
    #   wandb
 rerun-sdk==0.26.2
    # via lerobot
 rhoban-cmeel-jsoncpp==1.9.4.9
    # via placo
 rich==14.3.3
    # via typer
 robomimic==0.2.0
    # via hf-libero
 robosuite==1.4.0
    # via hf-libero
 rpds-py==0.30.0
    # via
    #   jsonschema
    #   referencing
 safetensors==0.7.0
    # via
    #   accelerate
    #   diffusers
    #   lerobot
    #   peft
    #   transformers
 scikit-image==0.25.2
    # via
    #   gym-pusht
    #   lerobot
 scipy==1.17.1
    # via
    #   dm-control
    #   lerobot
    #   metaworld
    #   robosuite
    #   scikit-image
    #   torchdiffeq
 sentry-sdk==2.54.0
    # via wandb
 shapely==2.1.2
    # via gym-pusht
 shellingham==1.5.4
    # via typer
 six==1.17.0
    # via
    #   pynput
    #   python-dateutil
    #   python-xlib
 smmap==5.0.3
    # via gitdb
 stack-data==0.6.3
    # via ipython
 starlette==0.52.1
    # via fastapi
 sympy==1.14.0
    # via torch
 teleop==0.1.4
    # via lerobot
 tensorboard==2.20.0
    # via robomimic
 tensorboard-data-server==0.7.2
    # via tensorboard
 tensorboardx==2.6.4
    # via robomimic
 termcolor==3.3.0
    # via
    #   lerobot
    #   robomimic
 thop==0.1.1.post2209072238
    # via hf-libero
 tifffile==2026.3.3
    # via scikit-image
 tokenizers==0.22.2
    # via transformers
 toml==0.10.2
    # via draccus
 torch==2.10.0
    # via
    #   accelerate
    #   lerobot
    #   peft
    #   robomimic
    #   thop
    #   torchdiffeq
    #   torchvision
 torchcodec==0.10.0
    # via lerobot
 torchdiffeq==0.2.5
    # via lerobot
 torchvision==0.25.0
    # via
    #   lerobot
    #   robomimic
 tornado==6.5.4
    # via meshcat
 tqdm==4.67.3
    # via
    #   datasets
    #   dm-control
    #   huggingface-hub
    #   peft
    #   robomimic
    #   transformers
 traitlets==5.14.3
    # via
    #   ipython
    #   jupyter-core
    #   matplotlib-inline
    #   nbformat
 transformers==5.3.0
    # via
    #   hf-libero
    #   lerobot
    #   peft
 transforms3d==0.4.2
    # via teleop
 triton==3.6.0
    # via torch
 typer==0.24.1
    # via
    #   huggingface-hub
    #   transformers
 typing-extensions==4.15.0
    # via
    #   aiosignal
    #   anyio
    #   etils
    #   faker
    #   fastapi
    #   gymnasium
    #   huggingface-hub
    #   mypy
    #   pydantic
    #   pydantic-core
    #   referencing
    #   rerun-sdk
    #   starlette
    #   torch
    #   typing-inspect
    #   typing-inspection
    #   wandb
 typing-inspect==0.9.0
    # via draccus
 typing-inspection==0.4.2
    # via
    #   fastapi
    #   pydantic
 tzdata==2025.3
    # via pandas
 u-msgpack-python==2.8.0
    # via meshcat
 urllib3==2.6.3
    # via
    #   requests
    #   sentry-sdk
 uvicorn[standard]==0.41.0
    # via teleop
 uvloop==0.22.1
    # via uvicorn
 virtualenv==21.1.0
    # via pre-commit
 wandb==0.24.2
    # via
    #   hf-libero
    #   lerobot
 watchfiles==1.1.1
    # via uvicorn
 wcwidth==0.6.0
    # via prompt-toolkit
 websocket-client==1.9.0
    # via teleop
 websockets==16.0
    # via uvicorn
 werkzeug==3.1.6
    # via tensorboard
 wrapt==2.1.2
    # via dm-tree
 xxhash==3.6.0
    # via datasets
 yarl==1.23.0
    # via aiohttp
 zipp==3.23.0
    # via
    #   etils
    #   importlib-metadata
 # The following packages are considered to be unsafe in a requirements file:
 # setuptools
@@ -0,0 +1,9 @@
 # requirements.in
 # requirements-macos.txt was generated on macOS and is platform-specific (macOS 26.3.1 25D2128 arm64).
 # Darwin MacBook-Pro.local 25.3.0 Darwin Kernel Version 25.3.0: Wed Jan 28 20:54:55 PST 2026; root:xnu-12377.91.3~2/RELEASE_ARM64_T8132 arm64
 # requirements-ubuntu.txt was generated on Linux and is platform-specific (Ubuntu 24.04.4 LTS x86_64).
 # Linux lerobot-linux 6.17.0-14-generic #14~24.04.1-Ubuntu SMP PREEMPT_DYNAMIC Thu Jan 15 15:52:10 UTC 2 x86_64 x86_64 x86_64 GNU/Linux
 -e .[all]
@@ -34,8 +34,6 @@ from .types import (
 )
 from .video import (
    DEFAULT_DEPTH_UNIT,
    DEPTH_METER_UNIT,
    DEPTH_MILLIMETER_UNIT,
    VALID_VIDEO_CODECS,
    VIDEO_ENCODER_INFO_KEYS,
    DepthEncoderConfig,
@@ -43,7 +41,6 @@ from .video import (
    VideoEncoderConfig,
    depth_encoder_defaults,
    encoder_config_from_video_info,
    infer_depth_unit,
    rgb_encoder_defaults,
 )
@@ -73,11 +70,8 @@ __all__ = [
    "depth_encoder_defaults",
    # Factories
    "encoder_config_from_video_info",
    "infer_depth_unit",
    # Constants
    "DEFAULT_DEPTH_UNIT",
    "DEPTH_METER_UNIT",
    "DEPTH_MILLIMETER_UNIT",
    "VALID_VIDEO_CODECS",
    "VIDEO_ENCODER_INFO_KEYS",
 ]
@@ -22,8 +22,6 @@ import logging
 from dataclasses import dataclass, field
 from typing import Any, ClassVar, Self
 import numpy as np
 from lerobot.utils.import_utils import require_package
 logger = logging.getLogger(__name__)
@@ -38,9 +36,7 @@ HW_VIDEO_CODECS = [
    "h264_vaapi",  # Linux Intel/AMD
    "h264_qsv",  # Intel Quick Sync
 ]
-VALID_VIDEO_CODECS: frozenset[str] = frozenset(
+VALID_VIDEO_CODECS: frozenset[str] = frozenset({"h264", "hevc", "libsvtav1", "auto", *HW_VIDEO_CODECS})
    {"h264", "hevc", "libsvtav1", "libaom-av1", "auto", *HW_VIDEO_CODECS}
 )
 # Aliases for legacy video codec names.
 VIDEO_CODECS_ALIASES: dict[str, str] = {"av1": "libsvtav1"}
@@ -69,15 +65,6 @@ DEPTH_METER_UNIT: str = "m"
 DEPTH_MILLIMETER_UNIT: str = "mm"
 DEFAULT_DEPTH_UNIT: str = DEPTH_MILLIMETER_UNIT
 def infer_depth_unit(dtype: np.dtype | type) -> str:
    """Infer the physical unit of raw depth frames from their dtype.
    Floating-point frames are assumed to be in metres, integer frames in millimetres.
    """
    return DEPTH_METER_UNIT if np.issubdtype(np.dtype(dtype), np.floating) else DEPTH_MILLIMETER_UNIT
 # Depth-specific tuning fields persisted under ``features[*]["info"]`` as ``video.<name>``.
 DEPTH_ENCODER_INFO_FIELD_NAMES: frozenset[str] = frozenset({"depth_min", "depth_max", "shift", "use_log"})
@@ -226,24 +213,18 @@ class VideoEncoderConfig:
            if encoder_threads is not None:
                svtav1_parts.append(f"lp={encoder_threads}")
            if svtav1_parts:
-                set_if("svtav1-params", ":".join(svtav1_parts))
+                opts["svtav1-params"] = ":".join(svtav1_parts)
        elif self.vcodec in ("h264", "hevc"):
            set_if("crf", self.crf)
            set_if("preset", self.preset)
            if self.fast_decode:
-                set_if("tune", "fastdecode")
+                opts["tune"] = "fastdecode"
            set_if("threads", encoder_threads)
        elif self.vcodec == "libaom-av1":
            set_if("crf", self.crf)
            set_if("preset", self.preset)
            if encoder_threads is not None:
                set_if("threads", encoder_threads)
                set_if("row-mt", 1)
        elif self.vcodec in ("h264_videotoolbox", "hevc_videotoolbox"):
            if self.crf is not None:
-                set_if("q:v", max(1, min(100, 100 - self.crf * 2)))
+                opts["q:v"] = max(1, min(100, 100 - self.crf * 2))
        elif self.vcodec in ("h264_nvenc", "hevc_nvenc"):
-            set_if("rc", 0)
+            opts["rc"] = 0
            set_if("qp", self.crf)
            set_if("preset", self.preset)
        elif self.vcodec == "h264_vaapi":
@@ -22,6 +22,7 @@ import numpy as np
 from lerobot.processor import RelativeActionsProcessorStep
 from lerobot.utils.constants import ACTION, OBS_STATE
 from .depth_utils import MM_PER_METRE
 from .io_utils import load_image_as_numpy
 DEFAULT_QUANTILES = [0.01, 0.10, 0.50, 0.90, 0.99]
@@ -508,8 +509,8 @@ def compute_episode_stats(
    Note:
        For 'image'/'video' features, stats are computed per channel and kept with a
        leading channel axis (e.g. shape (3, 1, 1) for RGB). RGB stats are divided by
-        255 to land in [0, 1]; depth maps (features flagged with ``is_depth_map``) skip
+        255 to land in [0, 1]; depth maps (features flagged with ``is_depth_map``) are
-        this rescaling and remain in their stored units (stored in ``depth_unit``).
+        instead canonicalized to millimetres regardless of the raw frame unit.
    """
    if quantile_list is None:
        quantile_list = DEFAULT_QUANTILES
@@ -533,9 +534,14 @@ def compute_episode_stats(
        )
        if features[key]["dtype"] in ["image", "video"]:
-            normalization_factor = (
+            if (features[key].get("info") or {}).get("is_depth_map", False):
-                255.0 if not (features[key].get("info") or {}).get("is_depth_map", False) else 1.0
+                # Depth stats are canonically stored in millimetres; metre (float) depth is
-            )
+                # scaled up, integer (millimetre) depth is left as-is.
                normalization_factor = (
                    1.0 / MM_PER_METRE if np.issubdtype(ep_ft_array.dtype, np.floating) else 1.0
                )
            else:
                normalization_factor = 255.0
            ep_stats[key] = {
                k: v if k == "count" else np.squeeze(v / normalization_factor, axis=0)
                for k, v in ep_stats[key].items()
@@ -26,13 +26,12 @@ import pyarrow as pa
 import pyarrow.parquet as pq
 from huggingface_hub import snapshot_download
-from lerobot.configs import DEPTH_METER_UNIT, VideoEncoderConfig
+from lerobot.configs import VideoEncoderConfig
 from lerobot.utils.constants import DEFAULT_FEATURES, HF_LEROBOT_HOME, HF_LEROBOT_HUB_CACHE
 from lerobot.utils.feature_utils import _validate_feature_names
 from lerobot.utils.utils import flatten_dict
 from .compute_stats import aggregate_stats
 from .depth_utils import MM_PER_METRE
 from .feature_utils import create_empty_dataset_info
 from .io_utils import (
    get_file_size_in_mb,
@@ -359,35 +358,6 @@ class LeRobotDatasetMetadata:
        return [key for key, ft in self.features.items() if _is_depth(ft)]
    def rescale_depth_stats(self, output_unit: str) -> None:
        """Rescale depth feature stats in place from their recorded unit to ``output_unit``.
        Depth stats are stored in the unit the frames were recorded in
        (``features[key]["info"]["depth_unit"]``), while frames are returned in
        ``output_unit`` on read. This converts the unit-bearing stat entries so
        stats match the frames consumers see.
        """
        missing_unit_keys = [
            key for key in self.depth_keys if (self.features[key].get("info") or {}).get("depth_unit") is None
        ]
        if missing_unit_keys:
            logging.warning(
                f"Depth feature(s) {missing_unit_keys} have no recorded 'depth_unit' in their info. "
                f"Depth maps and stats for these keys will be returned AS IS, with no unit conversion "
                f"to the requested output unit {output_unit!r}. Re-record the dataset or set 'depth_unit' "
                f"in the feature info (meta/info.json) to enable conversion."
            )
        if self.stats is None:
            return
        for key in self.depth_keys:
            stored_unit = (self.features[key].get("info") or {}).get("depth_unit")
            if stored_unit is None or stored_unit == output_unit or key not in self.stats:
                continue
            factor = MM_PER_METRE if stored_unit == DEPTH_METER_UNIT else 1.0 / MM_PER_METRE
            self.stats[key] = {
                stat: value if stat == "count" else value * factor for stat, value in self.stats[key].items()
            }
    @property
    def camera_keys(self) -> list[str]:
        """Keys to access visual modalities (regardless of their storage method)."""
@@ -22,14 +22,10 @@ from pathlib import Path
 import datasets
 import torch
-from lerobot.configs import (
+from lerobot.configs import DEFAULT_DEPTH_UNIT, DepthEncoderConfig
    DEFAULT_DEPTH_UNIT,
    DEPTH_METER_UNIT,
    DepthEncoderConfig,
 )
 from .dataset_metadata import LeRobotDatasetMetadata
-from .depth_utils import MM_PER_METRE, dequantize_depth
+from .depth_utils import dequantize_depth
 from .feature_utils import (
    check_delta_timestamps,
    get_delta_indices,
@@ -106,13 +102,6 @@ class DatasetReader:
            for vid_key in self._meta.depth_keys
        }
        # Get the input unit of each depth feature stored as raw images.
        self._image_depth_units: dict[str, str | None] = {
            key: (self._meta.features[key].get("info") or {}).get("depth_unit")
            for key in self._meta.depth_keys
            if key in self._meta.image_keys
        }
    def set_image_transforms(self, image_transforms: Callable | None) -> None:
        """Replace the transform applied to visual observations."""
        if image_transforms is not None and not callable(image_transforms):
@@ -340,13 +329,6 @@ class DatasetReader:
                    continue
                item[cam] = self._image_transforms(item[cam])
        # Convert depth features to the output unit.
        for key, stored_unit in self._image_depth_units.items():
            if key in item and stored_unit is not None and stored_unit != self._depth_output_unit:
                item[key] = (
                    item[key] * MM_PER_METRE if stored_unit == DEPTH_METER_UNIT else item[key] / MM_PER_METRE
                )
        # Add task as a string
        task_idx = item["task_index"].item()
        item["task"] = self._meta.tasks.iloc[task_idx].name
@@ -36,7 +36,6 @@ from lerobot.configs import (
    RGBEncoderConfig,
    VideoEncoderConfig,
    depth_encoder_defaults,
    infer_depth_unit,
    rgb_encoder_defaults,
 )
@@ -210,15 +209,6 @@ class DatasetWriter:
        self.episode_buffer["timestamp"].append(timestamp)
        self.episode_buffer["task"].append(frame.pop("task"))
        # Record each depth feature's input unit once, inferred from the first frame's dtype.
        if frame_index == 0:
            for depth_key in self._meta.depth_keys:
                if depth_key not in frame:
                    continue
                info = self._meta.features[depth_key].setdefault("info", {})
                if info.get("depth_unit") is None:
                    info["depth_unit"] = infer_depth_unit(np.asarray(frame[depth_key]).dtype)
        # Start streaming encoder on first frame of episode
        if frame_index == 0 and self._streaming_encoder is not None:
            self._streaming_encoder.start_episode(
@@ -34,7 +34,6 @@ from lerobot.configs.video import (
    DEPTH_METER_UNIT,
    DEPTH_MILLIMETER_UNIT,
    DEPTH_QMAX,
    infer_depth_unit,
 )
 from .image_writer import squeeze_single_channel
@@ -58,7 +57,11 @@ def _depth_input_to_float32_and_unit(
    input_unit: Literal["auto", DEPTH_METER_UNIT, DEPTH_MILLIMETER_UNIT],
 ) -> tuple[NDArray[np.float32], Literal[DEPTH_METER_UNIT, DEPTH_MILLIMETER_UNIT]]:
    """Convert depth to float32 in the chosen unit, and return the resolved unit."""
-    resolved_unit = infer_depth_unit(depth.dtype) if input_unit == "auto" else input_unit
+    resolved_unit = (
        (DEPTH_METER_UNIT if np.issubdtype(depth.dtype, np.floating) else DEPTH_MILLIMETER_UNIT)
        if input_unit == "auto"
        else input_unit
    )
    return depth.astype(np.float32, order="K"), resolved_unit
@@ -224,7 +224,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
        )
        self.root = self.meta.root
        self.revision = self.meta.revision
        self.meta.rescale_depth_stats(self._depth_output_unit)
        if episodes is not None and any(
            episode >= self.meta.total_episodes or episode < 0 for episode in episodes
@@ -351,11 +350,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
        """Frames per second used during data collection."""
        return self.meta.fps
    @property
    def depth_output_unit(self) -> str:
        """Physical unit (``"m"`` or ``"mm"``) depth maps and statistics are returned in on read."""
        return self._depth_output_unit
    @property
    def num_frames(self) -> int:
        """Number of frames in selected episodes."""
@@ -22,11 +22,11 @@ import numpy as np
 import torch
 from datasets import load_dataset
-from lerobot.configs import DEFAULT_DEPTH_UNIT, DEPTH_METER_UNIT, DepthEncoderConfig
+from lerobot.configs import DEFAULT_DEPTH_UNIT, DepthEncoderConfig
 from lerobot.utils.constants import HF_LEROBOT_HOME, LOOKAHEAD_BACKTRACKTABLE, LOOKBACK_BACKTRACKTABLE
 from .dataset_metadata import CODEBASE_VERSION, LeRobotDatasetMetadata
-from .depth_utils import MM_PER_METRE, dequantize_depth
+from .depth_utils import dequantize_depth
 from .feature_utils import get_delta_indices
 from .io_utils import item_to_torch
 from .utils import (
@@ -310,7 +310,6 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
        )
        self.root = self.meta.root
        self.revision = self.meta.revision
        self.meta.rescale_depth_stats(self._depth_output_unit)
        # Check version
        check_version_compatibility(self.repo_id, self.meta._version, CODEBASE_VERSION)
@@ -319,13 +318,6 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
            for vid_key in self.meta.depth_keys
        }
        # Input unit of each depth feature stored as raw images (dequantized separately from videos).
        self._image_depth_units: dict[str, str | None] = {
            key: (self.meta.features[key].get("info") or {}).get("depth_unit")
            for key in self.meta.depth_keys
            if key in self.meta.image_keys
        }
        self.delta_timestamps = None
        self.delta_indices = None
@@ -356,11 +348,6 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
    def fps(self):
        return self.meta.fps
    @property
    def depth_output_unit(self) -> str:
        """Physical unit (``"m"`` or ``"mm"``) depth maps are returned in on read."""
        return self._depth_output_unit
    @staticmethod
    def _iter_random_indices(
        rng: np.random.Generator, buffer_size: int, random_batch_size=100
@@ -543,15 +530,6 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
        for update in updates:
            result.update(update)
        # Convert raw-image depth features to the output unit (video depth is already converted).
        for key, stored_unit in self._image_depth_units.items():
            if key in result and stored_unit is not None and stored_unit != self._depth_output_unit:
                result[key] = (
                    result[key] * MM_PER_METRE
                    if stored_unit == DEPTH_METER_UNIT
                    else result[key] / MM_PER_METRE
                )
        result["task"] = self.meta.tasks.iloc[item["task_index"]].name
        yield result
@@ -47,7 +47,7 @@ from lerobot.configs import (
 )
 from lerobot.utils.import_utils import get_safe_default_video_backend
-from .depth_utils import quantize_depth
+from .depth_utils import MM_PER_METRE, quantize_depth
 from .pyav_utils import get_pix_fmt_channels
 logger = logging.getLogger(__name__)
@@ -848,6 +848,9 @@ class _CameraEncoderThread(threading.Thread):
                # Reshape CHW to (H*W, C) for per-channel stats
                channels = img_downsampled.shape[0]
                img_for_stats = img_downsampled.transpose(1, 2, 0).reshape(-1, channels)
                # Depth stats are canonically stored in millimetres; metre (float) depth is scaled up.
                if self.is_depth and np.issubdtype(frame_data.dtype, np.floating):
                    img_for_stats = img_for_stats * MM_PER_METRE
                stats_tracker.update(img_for_stats)
                frame_count += 1
@@ -757,7 +757,7 @@ class RoboTwinEnvConfig(EnvConfig):
    task: str = "beat_block_hammer"  # single task or comma-separated list
    fps: int = 25
-    episode_length: int = 1200
+    episode_length: int = 300
    obs_type: str = "pixels_agent_pos"
    render_mode: str = "rgb_array"
    # Available cameras from RoboTwin's aloha-agilex embodiment: head_camera
@@ -768,9 +768,6 @@ class RoboTwinEnvConfig(EnvConfig):
    # must equal what SAPIEN actually renders.
    observation_height: int = 240
    observation_width: int = 320
    # "joint": 14-d joint-space control. "ee": 16-d end-effector-pose deltas executed via CuRobo IK
    # (for world-model policies like LingBot-VA that predict per-arm xyz+quaternion+gripper poses).
    action_mode: str = "joint"
    features: dict[str, PolicyFeature] = field(
        default_factory=lambda: {
            ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(14,)),
@@ -787,8 +784,6 @@ class RoboTwinEnvConfig(EnvConfig):
    )
    def __post_init__(self):
        if self.action_mode == "ee":
            self.features[ACTION] = PolicyFeature(type=FeatureType.ACTION, shape=(16,))
        cam_list = [c.strip() for c in self.camera_names.split(",") if c.strip()]
        for cam in cam_list:
            self.features[f"pixels/{cam}"] = PolicyFeature(
@@ -831,7 +826,6 @@ class RoboTwinEnvConfig(EnvConfig):
            observation_height=self.observation_height,
            observation_width=self.observation_width,
            episode_length=self.episode_length,
            action_mode=self.action_mode,
        )
@@ -17,7 +17,6 @@ from __future__ import annotations
 import importlib
 import logging
 import os
 from collections import defaultdict
 from collections.abc import Callable, Sequence
 from functools import partial
@@ -29,17 +28,9 @@ import torch
 from gymnasium import spaces
 from lerobot.types import RobotObservation
 from lerobot.utils.import_utils import _scipy_available
 from .utils import _LazyAsyncVectorEnv
 # scipy is only used for end-effector-pose composition (``--env.action_mode=ee``); guard it so this
 # module (and its base-env unit tests, which mock the RoboTwin runtime) imports without scipy installed.
 if _scipy_available:
    from scipy.spatial.transform import Rotation
 else:
    Rotation = None
 logger = logging.getLogger(__name__)
 # Camera names as used by RoboTwin 2.0. The wrapper appends "_rgb" when looking
@@ -50,124 +41,10 @@ ROBOTWIN_CAMERA_NAMES: tuple[str, ...] = (
    "right_camera",
 )
-ACTION_DIM = 14  # 7 DOF × 2 arms (joint-space control mode)
+ACTION_DIM = 14  # 7 DOF × 2 arms
 # End-effector-pose control mode: per arm [x, y, z, qx, qy, qz, qw, gripper] = 8, dual-arm = 16.
 # Used by world-model policies (e.g. LingBot-VA) that predict eef-pose deltas executed via CuRobo IK.
 EEF_ACTION_DIM = 16
 ACTION_LOW = -1.0
 ACTION_HIGH = 1.0
-DEFAULT_EPISODE_LENGTH = 1200
+DEFAULT_EPISODE_LENGTH = 300
 OFFICIAL_INSTRUCTION_ENV = "LEROBOT_ROBOTWIN_OFFICIAL_INSTRUCTION"
 OFFICIAL_INSTRUCTION_TYPE_ENV = "LEROBOT_ROBOTWIN_INSTRUCTION_TYPE"
 OFFICIAL_INSTRUCTION_MAX_ENV = "LEROBOT_ROBOTWIN_INSTRUCTION_MAX"
 def _compose_eef_pose(new_pose: np.ndarray, init_pose: np.ndarray) -> np.ndarray:
    """Compose a single-arm predicted delta pose onto the initial pose.
    ``new_pose`` / ``init_pose`` are 8-vectors ``[x, y, z, qx, qy, qz, qw, gripper]``. Translation
    is added, rotation is composed (``init_R * new_R``), and the gripper is taken from the
    prediction. Mirrors ``add_eef_pose`` in the upstream LingBot-VA RoboTwin client.
    """
    new_r = Rotation.from_quat(new_pose[3:7])
    init_r = Rotation.from_quat(init_pose[3:7])
    out_rot = (init_r * new_r).as_quat()
    out_trans = new_pose[:3] + init_pose[:3]
    return np.concatenate([out_trans, out_rot, new_pose[7:8]])
 def _add_init_eef_pose(delta_pose: np.ndarray, init_pose: np.ndarray) -> np.ndarray:
    """Compose a dual-arm (16-d) predicted delta pose onto the initial eef pose, normalizing quats."""
    left = _compose_eef_pose(delta_pose[:8], init_pose[:8])
    right = _compose_eef_pose(delta_pose[8:], init_pose[8:])
    out = np.concatenate([left, right])
    # Normalize the two quaternions (indices 3:7 and 11:15) as the upstream client does.
    out[3:7] = out[3:7] / (np.linalg.norm(out[3:7]) + 1e-8)
    out[11:15] = out[11:15] / (np.linalg.norm(out[11:15]) + 1e-8)
    return out
 def _env_flag(name: str, default: bool = False) -> bool:
    raw = os.environ.get(name)
    if raw is None:
        return default
    return raw.strip().lower() in {"1", "true", "yes", "on"}
 def _arm_for_block(block: Any) -> str:
    return "left" if float(block.get_pose().p[0]) < 0 else "right"
 def _robotwin_blocks_episode_info(task_name: str, env: Any) -> dict[str, str] | None:
    """Infer the episode-info dict used by RoboTwin's official instruction generator for block ranking."""
    if task_name == "blocks_ranking_rgb":
        return {
            "{A}": "red block",
            "{B}": "green block",
            "{C}": "blue block",
            "{a}": _arm_for_block(env.block1),
            "{b}": _arm_for_block(env.block2),
            "{c}": _arm_for_block(env.block3),
        }
    if task_name == "blocks_ranking_size":
        return {
            "{A}": "large block",
            "{B}": "medium block",
            "{C}": "small block",
            "{a}": _arm_for_block(env.block1),
            "{b}": _arm_for_block(env.block2),
            "{c}": _arm_for_block(env.block3),
        }
    return None
 def _generate_robotwin_official_instruction(task_name: str, env: Any) -> str:
    """Generate language with RoboTwin's official task templates, matching its eval client."""
    fallback = task_name.replace("_", " ")
    episode_info = _robotwin_blocks_episode_info(task_name, env)
    if episode_info is None:
        logger.warning(
            "Official RoboTwin instruction is not implemented for task=%s; using %r.", task_name, fallback
        )
        return fallback
    try:
        # Part of the robotwin simulator repo, this is being pulled by the docker image running robotwin
        # see https://github.com/RoboTwin-Platform/RoboTwin/tree/main/description
        # Used to generate the official instructions
        from description.utils.generate_episode_instructions import generate_episode_descriptions
    except Exception:
        logger.warning(
            "Failed to import RoboTwin official instruction generator; using %r.", fallback, exc_info=True
        )
        return fallback
    instruction_type = os.environ.get(OFFICIAL_INSTRUCTION_TYPE_ENV, "seen")
    try:
        max_descriptions = int(os.environ.get(OFFICIAL_INSTRUCTION_MAX_ENV, "1000000"))
    except ValueError:
        max_descriptions = 1000000
    results = generate_episode_descriptions(task_name, [episode_info], max_descriptions=max_descriptions)
    if not results:
        logger.warning(
            "RoboTwin generated no official instructions for task=%s; using %r.", task_name, fallback
        )
        return fallback
    options = results[0].get(instruction_type) or results[0].get("seen") or results[0].get("unseen")
    if not options:
        logger.warning(
            "RoboTwin generated no %s official instructions for task=%s; using %r.",
            instruction_type,
            task_name,
            fallback,
        )
        return fallback
    return str(np.random.choice(options))
 # D435 dims from task_config/_camera_config.yml (what demo_clean.yml selects).
 DEFAULT_CAMERA_H = 240
 DEFAULT_CAMERA_W = 320
@@ -357,7 +234,6 @@ class RoboTwinEnv(gym.Env):
        observation_width: int | None = None,
        episode_length: int = DEFAULT_EPISODE_LENGTH,
        render_mode: str = "rgb_array",
        action_mode: str = "joint",
    ):
        super().__init__()
        self.task_name = task_name
@@ -365,13 +241,6 @@ class RoboTwinEnv(gym.Env):
        self.task_description = task_name.replace("_", " ")
        self.episode_index = episode_index
        self._reset_stride = n_envs
        # "joint": 14-d joint-space actions via take_action(action). "ee": 16-d end-effector-pose
        # deltas (added onto the episode's initial eef pose) executed via take_action(.., "ee") + IK.
        if action_mode not in ("joint", "ee"):
            raise ValueError(f"action_mode must be 'joint' or 'ee'; got {action_mode!r}")
        self.action_mode = action_mode
        self._action_dim = EEF_ACTION_DIM if action_mode == "ee" else ACTION_DIM
        self._init_eef_pose: np.ndarray | None = None
        self.camera_names = list(camera_names)
        # Default to D435 dims (the camera type baked into task_config/demo_clean.yml).
        # The YAML-driven lookup is deferred to reset() so construction doesn't
@@ -402,7 +271,7 @@ class RoboTwinEnv(gym.Env):
            }
        )
        self.action_space = spaces.Box(
-            low=ACTION_LOW, high=ACTION_HIGH, shape=(self._action_dim,), dtype=np.float32
+            low=ACTION_LOW, high=ACTION_HIGH, shape=(ACTION_DIM,), dtype=np.float32
        )
    def _ensure_env(self) -> None:
@@ -448,18 +317,6 @@ class RoboTwinEnv(gym.Env):
        return {"pixels": images, "agent_pos": joint_state}
    def _read_eef_pose(self) -> np.ndarray:
        """Read the current 16-d dual-arm eef pose [left(xyz+quat)+grip, right(xyz+quat)+grip]."""
        assert self._env is not None, "_read_eef_pose called before _ensure_env()"
        ep = self._env.get_obs()["endpose"]
        pose = (
            list(ep["left_endpose"])
            + [ep["left_gripper"]]
            + list(ep["right_endpose"])
            + [ep["right_gripper"]]
        )
        return np.asarray(pose, dtype=np.float64)
    def reset(self, seed: int | None = None, **kwargs) -> tuple[RobotObservation, dict]:
        self._ensure_env()
        super().reset(seed=seed)
@@ -473,32 +330,16 @@ class RoboTwinEnv(gym.Env):
        self.episode_index += self._reset_stride
        self._step_count = 0
        use_official_instruction = self.task_name in {"blocks_ranking_rgb", "blocks_ranking_size"}
        if _env_flag(OFFICIAL_INSTRUCTION_ENV, default=use_official_instruction):
            self.task_description = _generate_robotwin_official_instruction(self.task_name, self._env)
            if hasattr(self._env, "set_instruction"):
                self._env.set_instruction(instruction=self.task_description)
            logger.info("RoboTwin official instruction | task=%s | %s", self.task_name, self.task_description)
        else:
            self.task_description = self.task_name.replace("_", " ")
        # In eef mode the policy predicts pose deltas relative to the initial eef pose.
        if self.action_mode == "ee":
            self._init_eef_pose = self._read_eef_pose()
        obs = self._get_obs()
        return obs, {"is_success": False, "task": self.task_name}
    def step(self, action: np.ndarray) -> tuple[RobotObservation, float, bool, bool, dict[str, Any]]:
        assert self._env is not None, "step() called before reset()"
-        if action.ndim != 1 or action.shape[0] != self._action_dim:
+        if action.ndim != 1 or action.shape[0] != ACTION_DIM:
-            raise ValueError(f"Expected 1-D action of shape ({self._action_dim},), got {action.shape}")
+            raise ValueError(f"Expected 1-D action of shape ({ACTION_DIM},), got {action.shape}")
        with torch.enable_grad():
-            if self.action_mode == "ee":
+            if hasattr(self._env, "take_action"):
                ee_action = _add_init_eef_pose(np.asarray(action, dtype=np.float64), self._init_eef_pose)
                self._env.take_action(ee_action, action_type="ee")
            elif hasattr(self._env, "take_action"):
                self._env.take_action(action)
            else:
                self._env.step(action)
@@ -557,7 +398,6 @@ def _make_env_fns(
    observation_height: int,
    observation_width: int,
    episode_length: int,
    action_mode: str = "joint",
 ) -> list[Callable[[], RoboTwinEnv]]:
    """Return n_envs factory callables for a single task."""
@@ -570,7 +410,6 @@ def _make_env_fns(
            observation_height=observation_height,
            observation_width=observation_width,
            episode_length=episode_length,
            action_mode=action_mode,
        )
    return [partial(_make_one, i) for i in range(n_envs)]
@@ -584,7 +423,6 @@ def create_robotwin_envs(
    observation_height: int = DEFAULT_CAMERA_H,
    observation_width: int = DEFAULT_CAMERA_W,
    episode_length: int = DEFAULT_EPISODE_LENGTH,
    action_mode: str = "joint",
 ) -> dict[str, dict[int, Any]]:
    """Create vectorized RoboTwin 2.0 environments.
@@ -635,7 +473,6 @@ def create_robotwin_envs(
            observation_height=observation_height,
            observation_width=observation_width,
            episode_length=episode_length,
            action_mode=action_mode,
        )
        if is_async:
            lazy = _LazyAsyncVectorEnv(fns, cached_obs_space, cached_act_space, cached_metadata)
@@ -113,13 +113,11 @@ class DynamixelMotorsBus(SerialMotorsBus):
        port: str,
        motors: dict[str, Motor],
        calibration: dict[str, MotorCalibration] | None = None,
        protocol_version: int = PROTOCOL_VERSION,
    ):
        require_package("dynamixel-sdk", extra="dynamixel", import_name="dynamixel_sdk")
        super().__init__(port, motors, calibration)
        self.port_handler = dxl.PortHandler(self.port)
-        self.packet_handler = dxl.PacketHandler(protocol_version)
+        self.packet_handler = dxl.PacketHandler(PROTOCOL_VERSION)
        print(f"Using protocol version {protocol_version}")
        self.sync_reader = dxl.GroupSyncRead(self.port_handler, self.packet_handler, 0, 0)
        self.sync_writer = dxl.GroupSyncWrite(self.port_handler, self.packet_handler, 0, 0)
        self._comm_success = dxl.COMM_SUCCESS
@@ -33,58 +33,6 @@
 # 2. We can change the value of the MyControlTableKey enums without impacting the client code
 # {data_name: (address, size_byte)}
 # https://emanual.robotis.com/docs/en/dxl/ax/{MODEL}/#control-table
 AX_SERIES_CONTROL_TABLE = {
    # EEPROM Area
    "Model_Number": (0, 2),
    "Firmware_Version": (2, 1),
    "ID": (3, 1),
    "Baud_Rate": (4, 1),
    "Return_Delay_Time": (5, 1),
    "CW_Angle_Limit": (6, 2),
    "CCW_Angle_Limit": (8, 2),
    "Temperature_Limit": (11, 1),
    "Min_Voltage_Limit": (12, 1),
    "Max_Voltage_Limit": (13, 1),
    "Max_Torque": (14, 2),
    "Status_Return_Level": (16, 1),
    "Alarm_LED": (17, 1),
    "Shutdown": (18, 1),
    # RAM Area
    "Torque_Enable": (24, 1),
    "LED": (25, 1),
    "CW_Compliance_Margin": (26, 1),
    "CCW_Compliance_Margin": (27, 1),
    "CW_Compliance_Slope": (28, 1),
    "CCW_Compliance_Slope": (29, 1),
    "Goal_Position": (30, 2),
    "Moving_Speed": (32, 2),
    "Torque_Limit": (34, 2),
    "Present_Position": (36, 2),
    "Present_Speed": (38, 2),
    "Present_Load": (40, 2),
    "Present_Voltage": (42, 1),
    "Present_Temperature": (43, 1),
    "Registered": (44, 1),
    "Moving": (46, 1),
    "Lock": (47, 1),
    "Punch": (48, 2),
 }
 # https://emanual.robotis.com/docs/en/dxl/ax/{MODEL}/#baud-rate4
 AX_SERIES_BAUDRATE_TABLE = {
    9_600: 207,
    19_200: 103,
    57_600: 34,
    115_200: 16,
    200_000: 9,
    250_000: 7,
    400_000: 4,
    500_000: 3,
    1_000_000: 1,
 }
 # {data_name: (address, size_byte)}
 # https://emanual.robotis.com/docs/en/dxl/x/{MODEL}/#control-table
 X_SERIES_CONTROL_TABLE = {
@@ -166,14 +114,6 @@ X_SERIES_ENCODINGS_TABLE = {
    "Present_Velocity": X_SERIES_CONTROL_TABLE["Present_Velocity"][1],
 }
 # {data_name: size_byte}
 AX_SERIES_ENCODINGS_TABLE = {
    "Goal_Position": AX_SERIES_CONTROL_TABLE["Goal_Position"][1],
    "Moving_Speed": AX_SERIES_CONTROL_TABLE["Moving_Speed"][1],
    "Present_Position": AX_SERIES_CONTROL_TABLE["Present_Position"][1],
    "Present_Speed": AX_SERIES_CONTROL_TABLE["Present_Speed"][1],
 }
 MODEL_ENCODING_TABLE = {
    "x_series": X_SERIES_ENCODINGS_TABLE,
    "xl330-m077": X_SERIES_ENCODINGS_TABLE,
@@ -182,8 +122,6 @@ MODEL_ENCODING_TABLE = {
    "xm430-w350": X_SERIES_ENCODINGS_TABLE,
    "xm540-w270": X_SERIES_ENCODINGS_TABLE,
    "xc430-w150": X_SERIES_ENCODINGS_TABLE,
    "ax_series": AX_SERIES_ENCODINGS_TABLE,
    "ax-12a": AX_SERIES_ENCODINGS_TABLE,
 }
 # {model: model_resolution}
@@ -196,8 +134,6 @@ MODEL_RESOLUTION = {
    "xm430-w350": 4096,
    "xm540-w270": 4096,
    "xc430-w150": 4096,
    "ax_series": 1024,
    "ax-12a": 1024,
 }
 # {model: model_number}
@@ -209,7 +145,6 @@ MODEL_NUMBER_TABLE = {
    "xm430-w350": 1020,
    "xm540-w270": 1120,
    "xc430-w150": 1070,
    "ax-12a": 12,
 }
 # {model: available_operating_modes}
@@ -231,8 +166,6 @@ MODEL_CONTROL_TABLE = {
    "xm430-w350": X_SERIES_CONTROL_TABLE,
    "xm540-w270": X_SERIES_CONTROL_TABLE,
    "xc430-w150": X_SERIES_CONTROL_TABLE,
    "ax_series": AX_SERIES_CONTROL_TABLE,
    "ax-12a": AX_SERIES_CONTROL_TABLE,
 }
 MODEL_BAUDRATE_TABLE = {
@@ -243,8 +176,6 @@ MODEL_BAUDRATE_TABLE = {
    "xm430-w350": X_SERIES_BAUDRATE_TABLE,
    "xm540-w270": X_SERIES_BAUDRATE_TABLE,
    "xc430-w150": X_SERIES_BAUDRATE_TABLE,
    "ax_series": AX_SERIES_BAUDRATE_TABLE,
    "ax-12a": AX_SERIES_BAUDRATE_TABLE,
 }
 AVAILABLE_BAUDRATES = [
@@ -83,50 +83,6 @@ class VQBeTSchedulerConfig(LRSchedulerConfig):
        return LambdaLR(optimizer, lr_lambda, -1)
@LRSchedulerConfig.register_subclass("constant_with_warmup")
@dataclass
 class ConstantWithWarmupSchedulerConfig(LRSchedulerConfig):
    """Linear warmup followed by a constant learning rate.
    Mirrors the ``warmup_constant_lambda`` used by LingBot-VA (upstream ``wan_va/train.py``):
    the LR ramps linearly from 0 to the peak over ``num_warmup_steps`` steps, then stays flat.
    """
    num_warmup_steps: int = 1000
    def build(self, optimizer: Optimizer, num_training_steps: int) -> LambdaLR:
        warmup_steps = self.num_warmup_steps or 0
        def lr_lambda(current_step):
            if current_step < warmup_steps:
                return float(current_step) / float(max(1, warmup_steps))
            return 1.0
        return LambdaLR(optimizer, lr_lambda, -1)
@LRSchedulerConfig.register_subclass("cosine_annealing_with_warmup")
@dataclass
 class CosineAnnealingWithWarmupSchedulerConfig(LRSchedulerConfig):
    """Linear warmup followed by cosine annealing from the peak LR to zero.
    Used by EVO1; the annealing phase always spans the remaining training steps.
    """
    num_warmup_steps: int
    def build(self, optimizer: Optimizer, num_training_steps: int) -> LambdaLR:
        def lr_lambda(current_step: int) -> float:
            if current_step < self.num_warmup_steps:
                return current_step / max(1, self.num_warmup_steps)
            progress = (current_step - self.num_warmup_steps) / max(
                1, num_training_steps - self.num_warmup_steps
            )
            return max(0.0, 0.5 * (1.0 + math.cos(math.pi * progress)))
        return LambdaLR(optimizer, lr_lambda, -1)
@LRSchedulerConfig.register_subclass("cosine_decay_with_warmup")
@dataclass
 class CosineDecayWithWarmupSchedulerConfig(LRSchedulerConfig):
@@ -17,12 +17,10 @@ from lerobot.utils.action_interpolator import ActionInterpolator as ActionInterp
 from .act.configuration_act import ACTConfig as ACTConfig
 from .diffusion.configuration_diffusion import DiffusionConfig as DiffusionConfig
 from .eo1.configuration_eo1 import EO1Config as EO1Config
 from .evo1.configuration_evo1 import Evo1Config as Evo1Config
 from .factory import get_policy_class, make_policy, make_policy_config, make_pre_post_processors
 from .fastwam.configuration_fastwam import FastWAMConfig as FastWAMConfig
 from .gaussian_actor.configuration_gaussian_actor import GaussianActorConfig as GaussianActorConfig
 from .groot.configuration_groot import GrootConfig as GrootConfig
 from .lingbot_va.configuration_lingbot_va import LingBotVAConfig as LingBotVAConfig
 from .molmoact2.configuration_molmoact2 import MolmoAct2Config as MolmoAct2Config
 from .multi_task_dit.configuration_multi_task_dit import MultiTaskDiTConfig as MultiTaskDiTConfig
 from .pi0.configuration_pi0 import PI0Config as PI0Config
@@ -47,9 +45,7 @@ __all__ = [
    "EO1Config",
    "FastWAMConfig",
    "GaussianActorConfig",
    "Evo1Config",
    "GrootConfig",
    "LingBotVAConfig",
    "MolmoAct2Config",
    "MultiTaskDiTConfig",
    "PI0Config",
@@ -1 +0,0 @@
 ../../../../docs/source/policy_evo1_README.md
@@ -1,19 +0,0 @@
 # Copyright 2026 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
 #     http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from .configuration_evo1 import Evo1Config
 from .modeling_evo1 import Evo1Policy
 from .processor_evo1 import make_evo1_pre_post_processors
 __all__ = ["Evo1Config", "Evo1Policy", "make_evo1_pre_post_processors"]
@@ -1,252 +0,0 @@
 # Copyright 2026 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
 #     http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from __future__ import annotations
 import logging
 from dataclasses import dataclass, field
 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
 from lerobot.optim.optimizers import AdamWConfig
 from lerobot.optim.schedulers import CosineAnnealingWithWarmupSchedulerConfig
 from lerobot.utils.constants import ACTION, OBS_IMAGES, OBS_STATE
 from ..rtc.configuration_rtc import RTCConfig
 logger = logging.getLogger(__name__)
@PreTrainedConfig.register_subclass("evo1")
@dataclass
 class Evo1Config(PreTrainedConfig):
    training_stage: str = "stage1"
    # When True and the policy runs on CUDA, EVO1 wraps its own forward passes (training and
    # inference) in a bfloat16 autocast block, so its numerics do not depend on the dtype of any
    # outer autocast context opened by lerobot-train/lerobot-eval.
    use_amp: bool = True
    n_obs_steps: int = 1
    chunk_size: int = 50
    n_action_steps: int = 50
    max_state_dim: int = 24
    max_action_dim: int = 24
    max_views: int = 3
    image_resolution: tuple[int, int] = (448, 448)
    empty_cameras: int = 0
    postprocess_action_dim: int | None = None
    binarize_gripper: bool = False
    gripper_index: int = 6
    gripper_threshold: float = 0.5
    gripper_below_threshold_value: float = 1.0
    gripper_above_threshold_value: float = -1.0
    normalization_mapping: dict[str, NormalizationMode] = field(
        default_factory=lambda: {
            "VISUAL": NormalizationMode.IDENTITY,
            "STATE": NormalizationMode.MIN_MAX,
            "ACTION": NormalizationMode.MIN_MAX,
        }
    )
    vlm_model_name: str = "OpenGVLab/InternVL3-1B-hf"
    vlm_num_layers: int | None = 14
    vlm_dtype: str = "bfloat16"
    # Max token length for tokenizing the (image placeholders + instruction) prompt. Prompts longer
    # than this are right-truncated, so raise it for tasks with long language instructions or many views.
    max_text_length: int = 1024
    use_flash_attn: bool = True
    action_head: str = "flowmatching"
    embed_dim: int = 896
    hidden_dim: int = 1024
    state_hidden_dim: int = 1024
    num_heads: int = 8
    num_layers: int = 8
    dropout: float = 0.0
    num_inference_timesteps: int = 32
    num_categories: int = 1
    # When True, the action head is conditioned on a single pooled VL token (the last non-padding
    # token of the causal decoder) instead of the full fused token sequence.
    return_cls_only: bool = False
    enable_gradient_checkpointing: bool = True
    gradient_checkpointing_use_reentrant: bool = False
    finetune_vlm: bool | None = None
    finetune_language_model: bool | None = None
    finetune_vision_model: bool | None = None
    finetune_action_head: bool | None = None
    # Reapply stage defaults after loading checkpoint configs so stage2 cannot
    # accidentally inherit the frozen VLM flags stored by a stage1 checkpoint.
    apply_training_stage_defaults: bool = True
    task_field: str = "task"
    embodiment_id_field: str | None = None
    default_embodiment_id: int = 0
    # Real-Time Chunking guidance for asynchronous inference (lerobot-rollout --inference.type=rtc
    # sets this and calls init_rtc_processor()); None disables RTC.
    rtc_config: RTCConfig | None = None
    optimizer_lr: float = 1e-5
    optimizer_betas: tuple[float, float] = (0.9, 0.999)
    optimizer_eps: float = 1e-8
    optimizer_weight_decay: float = 1e-5
    optimizer_grad_clip_norm: float = 1.0
    scheduler_warmup_steps: int = 300
    def __post_init__(self):
        super().__post_init__()
        if self.training_stage not in {"stage1", "stage2"}:
            raise ValueError(
                f"Unsupported EVO1 training_stage '{self.training_stage}', expected 'stage1' or 'stage2'"
            )
        if self.apply_training_stage_defaults:
            stage_defaults = {
                "stage1": {
                    "finetune_vlm": False,
                    "finetune_language_model": False,
                    "finetune_vision_model": False,
                    "finetune_action_head": True,
                },
                "stage2": {
                    "finetune_vlm": True,
                    "finetune_language_model": True,
                    "finetune_vision_model": True,
                    "finetune_action_head": True,
                },
            }[self.training_stage]
            for flag_name, default_value in stage_defaults.items():
                current_value = getattr(self, flag_name)
                if current_value is not None and current_value != default_value:
                    logger.warning(
                        "EVO1 %s=%s is overridden by training_stage=%s default %s. "
                        "Set apply_training_stage_defaults=false to keep explicit finetuning flags.",
                        flag_name,
                        current_value,
                        self.training_stage,
                        default_value,
                    )
                setattr(self, flag_name, default_value)
        elif self.training_stage == "stage1":
            if self.finetune_vlm is None:
                self.finetune_vlm = False
            if self.finetune_language_model is None:
                self.finetune_language_model = False
            if self.finetune_vision_model is None:
                self.finetune_vision_model = False
            if self.finetune_action_head is None:
                self.finetune_action_head = True
        elif self.training_stage == "stage2":
            has_explicit_branch_flags = any(
                flag is not None for flag in (self.finetune_language_model, self.finetune_vision_model)
            )
            if not has_explicit_branch_flags:
                # An explicit finetune_vlm decides both branches; otherwise stage2 defaults to a
                # full-VLM finetune.
                vlm_finetune = self.finetune_vlm if self.finetune_vlm is not None else True
                self.finetune_vlm = vlm_finetune
                self.finetune_language_model = vlm_finetune
                self.finetune_vision_model = vlm_finetune
            elif self.finetune_vlm is None:
                self.finetune_vlm = bool(self.finetune_language_model or self.finetune_vision_model)
            if self.finetune_action_head is None:
                self.finetune_action_head = True
        if self.finetune_vlm is None:
            self.finetune_vlm = False
        if self.finetune_language_model is None:
            self.finetune_language_model = False
        if self.finetune_vision_model is None:
            self.finetune_vision_model = False
        if self.finetune_action_head is None:
            self.finetune_action_head = False
        branch_vlm = self.finetune_language_model or self.finetune_vision_model
        if self.finetune_vlm != branch_vlm:
            raise ValueError(
                "Inconsistent EVO1 finetune config: "
                f"finetune_vlm={self.finetune_vlm} but "
                f"(finetune_language_model or finetune_vision_model)={branch_vlm}. "
                "When branch-level flags are used, finetune_vlm must match their effective union."
            )
        if self.n_action_steps > self.chunk_size:
            raise ValueError(
                f"n_action_steps ({self.n_action_steps}) must be <= chunk_size ({self.chunk_size})"
            )
        if len(self.image_resolution) != 2 or self.image_resolution[0] != self.image_resolution[1]:
            raise ValueError(
                "EVO1 currently expects a square image_resolution because InternVL3 preprocessing "
                f"uses a scalar image_size, got {self.image_resolution}."
            )
        if not 0 <= self.default_embodiment_id < self.num_categories:
            raise ValueError(
                f"default_embodiment_id ({self.default_embodiment_id}) must be in "
                f"[0, num_categories={self.num_categories})"
            )
    def validate_features(self) -> None:
        if self.input_features is None:
            self.input_features = {}
        if self.output_features is None:
            self.output_features = {}
        for i in range(self.empty_cameras):
            key = OBS_IMAGES + f".empty_camera_{i}"
            if key not in self.input_features:
                self.input_features[key] = PolicyFeature(
                    type=FeatureType.VISUAL,
                    shape=(3, *self.image_resolution),
                )
        if OBS_STATE not in self.input_features:
            self.input_features[OBS_STATE] = PolicyFeature(
                type=FeatureType.STATE,
                shape=(self.max_state_dim,),
            )
        if ACTION not in self.output_features:
            self.output_features[ACTION] = PolicyFeature(
                type=FeatureType.ACTION,
                shape=(self.max_action_dim,),
            )
    def get_optimizer_preset(self) -> AdamWConfig:
        return AdamWConfig(
            lr=self.optimizer_lr,
            betas=self.optimizer_betas,
            eps=self.optimizer_eps,
            weight_decay=self.optimizer_weight_decay,
            grad_clip_norm=self.optimizer_grad_clip_norm,
        )
    def get_scheduler_preset(self):
        return CosineAnnealingWithWarmupSchedulerConfig(
            num_warmup_steps=self.scheduler_warmup_steps,
        )
    @property
    def observation_delta_indices(self) -> list[int]:
        return [0]
    @property
    def action_delta_indices(self) -> list[int]:
        return list(range(self.chunk_size))
    @property
    def reward_delta_indices(self) -> None:
        return None
@@ -1,210 +0,0 @@
 # Copyright 2026 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
 #     http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from __future__ import annotations
 import torch
 import torch.nn as nn
 from .configuration_evo1 import Evo1Config
 from .flow_matching import FlowmatchingActionHead
 from .internvl3_embedder import InternVL3Embedder
 class Evo1Model(nn.Module):
    def __init__(self, config: Evo1Config, vlm_hub_kwargs: dict | None = None):
        super().__init__()
        self.config = config
        self._device = config.device
        self.return_cls_only = config.return_cls_only
        # Set by Evo1Policy.init_rtc_processor() when config.rtc_config is provided.
        self.rtc_processor = None
        # Gradient checkpointing only pays off when the VLM is actually being trained; keep it off
        # whenever every VLM branch is frozen so the frozen forward stays cheap.
        tracks_vlm_gradients = bool(
            config.finetune_vlm or config.finetune_language_model or config.finetune_vision_model
        )
        enable_gradient_checkpointing = config.enable_gradient_checkpointing and tracks_vlm_gradients
        self.embedder = InternVL3Embedder(
            model_name=config.vlm_model_name,
            image_size=int(config.image_resolution[0]),
            device=self._device,
            num_language_layers=config.vlm_num_layers,
            model_dtype=config.vlm_dtype,
            use_flash_attn=config.use_flash_attn,
            max_text_length=config.max_text_length,
            enable_gradient_checkpointing=enable_gradient_checkpointing,
            gradient_checkpointing_use_reentrant=config.gradient_checkpointing_use_reentrant,
            hub_kwargs=vlm_hub_kwargs,
        )
        action_head_type = config.action_head.lower()
        if action_head_type != "flowmatching":
            raise NotImplementedError(f"Unknown action_head: {action_head_type}")
        horizon = config.chunk_size
        per_action_dim = config.max_action_dim
        action_dim = horizon * per_action_dim
        self.horizon = horizon
        self.per_action_dim = per_action_dim
        self.action_head = FlowmatchingActionHead(
            embed_dim=config.embed_dim,
            hidden_dim=config.hidden_dim,
            action_dim=action_dim,
            horizon=horizon,
            per_action_dim=per_action_dim,
            num_heads=config.num_heads,
            num_layers=config.num_layers,
            dropout=config.dropout,
            num_inference_timesteps=config.num_inference_timesteps,
            num_categories=config.num_categories,
            state_dim=config.max_state_dim,
            state_hidden_dim=config.state_hidden_dim,
        ).to(self._device)
    def get_vl_embeddings(
        self,
        images: list[torch.Tensor],
        image_mask: torch.Tensor,
        prompt: str | list[str] | None = None,
        return_cls_only: bool | None = None,
    ) -> tuple[torch.Tensor, torch.Tensor | None]:
        """Fused VL embeddings from per-camera image batches.
        Args:
            images: list of per-camera tensors, each shaped ``(B, C, H, W)`` with values in ``[0, 1]``.
            image_mask: bool tensor ``(B, max_views)`` marking present views.
        Returns:
            ``(embeddings, valid_mask)``: the fused tokens and the bool mask of attendable context
            positions (None when a single pooled token is returned).
        """
        if return_cls_only is None:
            return_cls_only = self.return_cls_only
        if not images:
            raise ValueError("EVO1 expects at least one image per sample.")
        batch_size = images[0].shape[0]
        if prompt is None:
            prompts = [""] * batch_size
        elif isinstance(prompt, str):
            prompts = [prompt] * batch_size
        else:
            prompts = [str(p) for p in prompt]
            if len(prompts) != batch_size:
                raise ValueError(
                    f"Prompt batch size {len(prompts)} does not match image batch size {batch_size}"
                )
        if image_mask.dim() == 1:
            image_mask = image_mask.unsqueeze(0)
        if image_mask.shape[0] != batch_size:
            raise ValueError(
                f"image_mask batch size {image_mask.shape[0]} does not match image batch size {batch_size}"
            )
        return self.embedder.get_fused_image_text_embedding_batched(
            camera_images=images,
            image_masks=image_mask,
            text_prompts=prompts,
            return_cls_only=return_cls_only,
        )
    def predict_action(
        self,
        fused_tokens: torch.Tensor,
        state: torch.Tensor,
        actions_gt: torch.Tensor | None = None,
        action_mask: torch.Tensor | None = None,
        embodiment_ids: torch.Tensor | None = None,
        context_mask: torch.Tensor | None = None,
        inference_delay: int | None = None,
        prev_chunk_left_over: torch.Tensor | None = None,
        execution_horizon: int | None = None,
    ):
        if actions_gt is None:
            return self.action_head.get_action(
                fused_tokens,
                state=state,
                action_mask=action_mask,
                embodiment_id=embodiment_ids,
                context_mask=context_mask,
                inference_delay=inference_delay,
                prev_chunk_left_over=prev_chunk_left_over,
                execution_horizon=execution_horizon,
                rtc_processor=self.rtc_processor,
            )
        return self.action_head(
            fused_tokens,
            state=state,
            actions_gt=actions_gt,
            action_mask=action_mask,
            embodiment_id=embodiment_ids,
            context_mask=context_mask,
        )
    def forward(
        self,
        fused_tokens: torch.Tensor,
        state: torch.Tensor | None = None,
        actions_gt: torch.Tensor | None = None,
        action_mask: torch.Tensor | None = None,
        embodiment_ids: torch.Tensor | None = None,
        context_mask: torch.Tensor | None = None,
        inference_delay: int | None = None,
        prev_chunk_left_over: torch.Tensor | None = None,
        execution_horizon: int | None = None,
    ):
        return self.predict_action(
            fused_tokens,
            state,
            actions_gt,
            action_mask,
            embodiment_ids,
            context_mask,
            inference_delay,
            prev_chunk_left_over,
            execution_horizon,
        )
    def _set_module_trainable(self, module: nn.Module, trainable: bool):
        for param in module.parameters():
            param.requires_grad = trainable
    def _vlm_submodule(self, name: str) -> nn.Module:
        module = getattr(self.embedder.model, name, None)
        if not isinstance(module, nn.Module):
            raise AttributeError(
                f"InternVL model {type(self.embedder.model).__name__} has no '{name}' submodule; "
                "the native HF InternVL layout (language_model / vision_tower / "
                "multi_modal_projector) is required to apply the EVO1 finetune flags."
            )
        return module
    def set_finetune_flags(self):
        # __post_init__ resolves every finetune flag to a concrete boolean, so branch-level flags
        # are authoritative here. Freeze everything first, then re-enable the requested branches.
        self._set_module_trainable(self.embedder, False)
        self._set_module_trainable(
            self._vlm_submodule("language_model"), bool(self.config.finetune_language_model)
        )
        finetune_vision = bool(self.config.finetune_vision_model)
        self._set_module_trainable(self._vlm_submodule("vision_tower"), finetune_vision)
        self._set_module_trainable(self._vlm_submodule("multi_modal_projector"), finetune_vision)
        if not self.config.finetune_action_head:
            self._set_module_trainable(self.action_head, False)
@@ -1,483 +0,0 @@
 # Copyright 2026 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
 #     http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from __future__ import annotations
 import logging
 import math
 import torch
 import torch.nn as nn
 logger = logging.getLogger(__name__)
 class SinusoidalPositionalEncoding(nn.Module):
    def __init__(self, dim: int, max_len: int = 1000):
        super().__init__()
        pe = torch.zeros(max_len, dim)
        position = torch.arange(0, max_len).unsqueeze(1)
        div_term = torch.exp(torch.arange(0, dim, 2) * -(math.log(10000.0) / dim))
        pe[:, 0::2] = torch.sin(position * div_term)
        pe[:, 1::2] = torch.cos(position * div_term)
        pe = pe.unsqueeze(0)
        self.register_buffer("pe", pe)
    def forward(self, seq_len: int):
        if seq_len > self.pe.size(1):
            self._extend_pe(seq_len)
        return self.pe[:, :seq_len, :]
    def _extend_pe(self, new_max_len):
        old_max_len, dim = self.pe.size(1), self.pe.size(2)
        if new_max_len <= old_max_len:
            return
        extra_positions = torch.arange(old_max_len, new_max_len, dtype=torch.float).unsqueeze(1)
        div_term = torch.exp(torch.arange(0, dim, 2, dtype=torch.float) * -(math.log(10000.0) / dim))
        extra_pe = torch.zeros(new_max_len - old_max_len, dim)
        extra_pe[:, 0::2] = torch.sin(extra_positions * div_term)
        extra_pe[:, 1::2] = torch.cos(extra_positions * div_term)
        extra_pe = extra_pe.unsqueeze(0)
        new_pe = torch.cat([self.pe, extra_pe.to(self.pe.device)], dim=1)
        self.pe = new_pe
 class CategorySpecificLinear(nn.Module):
    def __init__(self, in_dim: int, out_dim: int, num_categories: int = 1):
        super().__init__()
        self.num_categories = num_categories
        if num_categories <= 1:
            self.linear = nn.Linear(in_dim, out_dim)
        else:
            self.weight = nn.Parameter(torch.empty(num_categories, in_dim, out_dim))
            self.bias = nn.Parameter(torch.zeros(num_categories, out_dim))
            # Initialize each per-category (in_dim, out_dim) matrix separately: xavier on the full
            # 3D tensor would compute fan_in = in_dim * out_dim and badly under-scale the weights.
            for category in range(num_categories):
                nn.init.xavier_uniform_(self.weight[category])
    def forward(self, x: torch.Tensor, category_id: torch.LongTensor):
        if self.num_categories <= 1:
            if x.dtype != self.linear.weight.dtype:
                x = x.to(dtype=self.linear.weight.dtype)
            return self.linear(x)
        if x.dtype != self.weight.dtype:
            x = x.to(dtype=self.weight.dtype)
        orig_shape = x.shape
        x_flat = x.reshape(-1, orig_shape[-1])
        if category_id.dim() == 0:
            cid = category_id.item()
            out = x_flat @ self.weight[cid] + self.bias[cid]
        else:
            category_id = category_id.reshape(-1)
            if category_id.numel() != x_flat.size(0):
                raise ValueError(
                    f"category_id length {category_id.numel()} does not match flattened batch {x_flat.size(0)}"
                )
            weight_selected = self.weight[category_id]
            bias_selected = self.bias[category_id]
            out = torch.bmm(x_flat.unsqueeze(1), weight_selected).squeeze(1) + bias_selected
        out_shape = orig_shape[:-1] + (out.shape[-1],)
        return out.view(out_shape)
 class CategorySpecificMLP(nn.Module):
    def __init__(self, input_dim: int, hidden_dim: int, output_dim: int, num_categories: int = 1):
        super().__init__()
        self.fc1 = CategorySpecificLinear(input_dim, hidden_dim, num_categories)
        self.fc2 = CategorySpecificLinear(hidden_dim, output_dim, num_categories)
        self.activation = nn.ReLU(inplace=True)
    def forward(self, x: torch.Tensor, category_id: torch.LongTensor):
        out = self.activation(self.fc1(x, category_id))
        out = self.fc2(out, category_id)
        return out
 class MultiEmbodimentActionEncoder(nn.Module):
    def __init__(
        self, action_dim: int, embed_dim: int, hidden_dim: int, horizon: int, num_categories: int = 1
    ):
        super().__init__()
        self.horizon = horizon
        self.embed_dim = embed_dim
        self.num_categories = num_categories
        self.W1 = CategorySpecificLinear(action_dim, hidden_dim, num_categories)
        self.W2 = CategorySpecificLinear(hidden_dim, hidden_dim, num_categories)
        self.W3 = CategorySpecificLinear(hidden_dim, embed_dim, num_categories)
        self.pos_encoding = SinusoidalPositionalEncoding(hidden_dim, max_len=horizon)
        self.activation = nn.ReLU(inplace=True)
    def forward(self, action_seq: torch.Tensor, category_id: torch.LongTensor):
        batch_size, horizon, action_dim = action_seq.shape
        if self.horizon != horizon:
            raise ValueError(
                f"Action sequence length must match horizon: got {horizon}, expected {self.horizon}."
            )
        x = action_seq.reshape(batch_size * horizon, action_dim)
        if category_id.dim() == 0:
            cat_ids = category_id.expand(horizon * batch_size)
        else:
            cat_ids = category_id.unsqueeze(1).expand(batch_size, horizon).reshape(batch_size * horizon)
        out = self.activation(self.W1(x, cat_ids))
        pos_enc = self.pos_encoding(horizon).to(device=out.device, dtype=out.dtype)
        out = out.view(batch_size, horizon, -1) + pos_enc
        out = out.view(batch_size * horizon, -1)
        out = self.activation(self.W2(out, cat_ids))
        out = self.W3(out, cat_ids)
        return out.view(batch_size, horizon, self.embed_dim)
 class BasicTransformerBlock(nn.Module):
    def __init__(self, embed_dim: int, num_heads: int, hidden_dim: int, dropout: float = 0.0):
        super().__init__()
        self.attn = nn.MultiheadAttention(embed_dim, num_heads, dropout=dropout, batch_first=True)
        self.norm1 = nn.LayerNorm(embed_dim)
        self.norm2 = nn.LayerNorm(embed_dim)
        self.ff = nn.Sequential(nn.Linear(embed_dim, hidden_dim), nn.GELU(), nn.Linear(hidden_dim, embed_dim))
    def forward(
        self,
        action_tokens: torch.Tensor,
        context_tokens: torch.Tensor,
        time_emb: torch.Tensor,
        context_key_padding_mask: torch.Tensor | None = None,
    ):
        x = self.norm1(action_tokens)
        attn_out, _ = self.attn(x, context_tokens, context_tokens, key_padding_mask=context_key_padding_mask)
        x = action_tokens + attn_out
        x2 = self.norm2(x)
        if time_emb is not None:
            x2 = x2 + time_emb.unsqueeze(1)
        ff_out = self.ff(x2)
        return x + ff_out
 class FlowmatchingActionHead(nn.Module):
    def __init__(
        self,
        embed_dim: int = 896,
        hidden_dim: int = 1024,
        action_dim: int = 16 * 7,
        horizon: int = 16,
        per_action_dim: int = 7,
        num_heads: int = 8,
        num_layers: int = 8,
        dropout: float = 0.0,
        num_inference_timesteps: int = 20,
        num_categories: int = 1,
        state_dim: int | None = None,
        state_hidden_dim: int | None = None,
    ):
        super().__init__()
        logger.info("FlowmatchingActionHead num_inference_timesteps=%s", num_inference_timesteps)
        self.embed_dim = embed_dim
        self.horizon = horizon
        self.per_action_dim = per_action_dim
        self.action_dim = action_dim
        self.num_inference_timesteps = num_inference_timesteps
        self.num_categories = num_categories
        self.time_pos_enc = SinusoidalPositionalEncoding(embed_dim, max_len=1000)
        self.transformer_blocks = nn.ModuleList(
            [
                BasicTransformerBlock(
                    embed_dim=embed_dim,
                    num_heads=num_heads,
                    hidden_dim=embed_dim * 4,
                    dropout=dropout,
                )
                for _ in range(num_layers)
            ]
        )
        self.norm_out = nn.LayerNorm(embed_dim)
        self.seq_pool_proj = nn.Linear(self.horizon * self.embed_dim, self.embed_dim)
        self.mlp_head = CategorySpecificMLP(
            input_dim=embed_dim,
            hidden_dim=hidden_dim,
            output_dim=action_dim,
            num_categories=num_categories,
        )
        self.state_encoder = None
        if state_dim is not None:
            state_hidden = state_hidden_dim if state_hidden_dim is not None else embed_dim
            self.state_encoder = CategorySpecificMLP(
                input_dim=state_dim,
                hidden_dim=state_hidden,
                output_dim=embed_dim,
                num_categories=num_categories,
            )
        if horizon > 1:
            self.action_encoder = MultiEmbodimentActionEncoder(
                action_dim=self.per_action_dim,
                embed_dim=embed_dim,
                hidden_dim=embed_dim,
                horizon=horizon,
                num_categories=num_categories,
            )
            self.single_action_proj = None
        else:
            self.action_encoder = None
            self.single_action_proj = nn.Linear(self.per_action_dim, self.embed_dim)
    def _project_actions(self, action_seq: torch.Tensor, embodiment_id: torch.LongTensor) -> torch.Tensor:
        if self.horizon > 1 and self.action_encoder is not None:
            return self.action_encoder(action_seq, embodiment_id)
        if self.single_action_proj is None:
            raise RuntimeError("single_action_proj is not initialized for horizon <= 1.")
        return self.single_action_proj(action_seq)
    def _expand_action_mask(
        self,
        action_mask: torch.Tensor,
        batch_size: int,
        per_action_dim: int,
        device: torch.device,
        dtype: torch.dtype,
    ) -> torch.Tensor:
        if action_mask is None:
            raise ValueError("action_mask must be provided for flow matching inference.")
        if action_mask.dim() == 2:
            expected_last_dim = self.horizon * per_action_dim
            if action_mask.shape == (batch_size, expected_last_dim):
                expanded_mask = action_mask.reshape(batch_size, self.horizon, per_action_dim)
            elif action_mask.shape == (batch_size, per_action_dim):
                expanded_mask = action_mask.unsqueeze(1).expand(batch_size, self.horizon, per_action_dim)
            else:
                raise ValueError(
                    f"Expected action_mask shape {(batch_size, expected_last_dim)} or "
                    f"{(batch_size, per_action_dim)}, got {tuple(action_mask.shape)}"
                )
        elif action_mask.dim() == 3:
            expected_shape = (batch_size, self.horizon, per_action_dim)
            if tuple(action_mask.shape) != expected_shape:
                raise ValueError(
                    f"Expected action_mask shape {expected_shape}, got {tuple(action_mask.shape)}"
                )
            expanded_mask = action_mask
        else:
            raise ValueError(f"Unsupported action_mask rank: {action_mask.dim()}")
        return expanded_mask.to(device=device, dtype=dtype)
    def _prepare_context(
        self,
        fused_tokens: torch.Tensor,
        state: torch.Tensor | None,
        embodiment_id: torch.LongTensor | None,
        context_mask: torch.Tensor | None,
    ) -> tuple[torch.Tensor, torch.Tensor | None, torch.LongTensor]:
        """Normalize the VL context and embodiment ids shared by training and inference.
        Returns the context tokens ``(B, S, E)``, a key_padding_mask for
        ``nn.MultiheadAttention`` (True = ignore) or None, and the resolved embodiment ids.
        """
        batch_size = fused_tokens.size(0)
        device = fused_tokens.device
        if embodiment_id is None:
            embodiment_id = torch.zeros(batch_size, dtype=torch.long, device=device)
        elif self.num_categories > 1 and (
            int(embodiment_id.min()) < 0 or int(embodiment_id.max()) >= self.num_categories
        ):
            raise ValueError(
                f"embodiment ids must be in [0, num_categories={self.num_categories}), "
                f"got range [{int(embodiment_id.min())}, {int(embodiment_id.max())}]"
            )
        context_tokens = fused_tokens
        if context_tokens.dim() == 2:
            # A single pooled VL token (return_cls_only): give it a sequence dim of 1.
            context_tokens = context_tokens.unsqueeze(1)
            context_mask = None
        if state is not None and self.state_encoder is not None:
            state_emb = self.state_encoder(state, embodiment_id).unsqueeze(1)
            context_tokens = torch.cat([context_tokens, state_emb], dim=1)
            if context_mask is not None:
                state_valid = torch.ones(batch_size, 1, dtype=torch.bool, device=context_mask.device)
                context_mask = torch.cat([context_mask.to(torch.bool), state_valid], dim=1)
        key_padding_mask = None if context_mask is None else ~context_mask.to(torch.bool)
        return context_tokens, key_padding_mask, embodiment_id
    def forward(
        self,
        fused_tokens: torch.Tensor,
        state: torch.Tensor = None,
        actions_gt: torch.Tensor = None,
        embodiment_id: torch.LongTensor = None,
        action_mask: torch.Tensor = None,
        context_mask: torch.Tensor = None,
    ):
        if actions_gt is None:
            return self.get_action(
                fused_tokens,
                state=state,
                embodiment_id=embodiment_id,
                action_mask=action_mask,
                context_mask=context_mask,
            )
        batch_size = fused_tokens.size(0)
        device = fused_tokens.device
        context_tokens, key_padding_mask, embodiment_id = self._prepare_context(
            fused_tokens, state, embodiment_id, context_mask
        )
        t = (
            torch.distributions.Beta(2, 2)
            .sample((batch_size,))
            .clamp(0.02, 0.98)
            .to(device)
            .to(dtype=self.dtype)
        )
        time_index = (t * 999).long().clamp_(0, 999)
        time_emb = self.time_pos_enc(1000)[:, time_index, :].squeeze(0).to(dtype=context_tokens.dtype)
        actions_gt_seq = actions_gt
        noise = torch.rand_like(actions_gt) * 2 - 1
        if action_mask is not None:
            action_mask = action_mask.to(dtype=noise.dtype, device=noise.device)
            if action_mask.shape != noise.shape:
                raise ValueError(f"action_mask shape {action_mask.shape} != noise shape {noise.shape}")
            actions_gt_seq = actions_gt_seq * action_mask
            noise = noise * action_mask
        if self.horizon > 1:
            noise_seq = noise.view(batch_size, self.horizon, self.per_action_dim)
        else:
            noise_seq = noise if noise.dim() == 3 else noise.unsqueeze(1)
        t_broadcast = t.view(batch_size, 1, 1)
        action_intermediate_seq = (1 - t_broadcast) * noise_seq + t_broadcast * actions_gt_seq
        action_tokens = self._project_actions(action_intermediate_seq, embodiment_id)
        target_dtype = self.dtype
        action_tokens = action_tokens.to(dtype=target_dtype)
        context_tokens = context_tokens.to(dtype=target_dtype)
        time_emb = time_emb.to(dtype=target_dtype)
        x = action_tokens
        for block in self.transformer_blocks:
            x = block(x, context_tokens, time_emb, key_padding_mask)
        x = self.norm_out(x)
        if self.horizon > 1:
            x_flat = x.reshape(batch_size, -1)
            x_pooled = self.seq_pool_proj(x_flat)
        else:
            x_pooled = x.squeeze(1)
        pred_velocity = self.mlp_head(x_pooled, embodiment_id)
        return pred_velocity, noise
    def get_action(
        self,
        fused_tokens: torch.Tensor,
        state: torch.Tensor = None,
        embodiment_id: torch.LongTensor = None,
        action_mask: torch.Tensor = None,
        context_mask: torch.Tensor = None,
        inference_delay: int | None = None,
        prev_chunk_left_over: torch.Tensor | None = None,
        execution_horizon: int | None = None,
        rtc_processor=None,
    ):
        batch_size = fused_tokens.size(0)
        device = fused_tokens.device
        context_tokens, key_padding_mask, embodiment_id = self._prepare_context(
            fused_tokens, state, embodiment_id, context_mask
        )
        action_dim_total = self.action_dim
        per_action_dim = self.per_action_dim
        action = torch.rand(batch_size, action_dim_total, device=device, dtype=context_tokens.dtype) * 2 - 1
        action_seq = action.view(batch_size, self.horizon, per_action_dim)
        action_mask = self._expand_action_mask(
            action_mask,
            batch_size=batch_size,
            per_action_dim=per_action_dim,
            device=action_seq.device,
            dtype=action_seq.dtype,
        )
        action_seq = action_seq * action_mask
        target_dtype = self.dtype
        context_tokens = context_tokens.to(dtype=target_dtype)
        num_steps = int(self.num_inference_timesteps)
        if num_steps <= 0:
            raise ValueError(f"num_inference_timesteps must be positive, got {num_steps}")
        dt = 1.0 / num_steps
        use_rtc = rtc_processor is not None and (
            inference_delay is not None or prev_chunk_left_over is not None
        )
        def predict_velocity(seq: torch.Tensor, step_time_emb: torch.Tensor) -> torch.Tensor:
            """Predict the masked flow velocity (x1 - x0 convention) for one integration step."""
            seq = seq * action_mask
            action_tokens = self._project_actions(seq, embodiment_id).to(dtype=target_dtype)
            x = action_tokens
            for block in self.transformer_blocks:
                x = block(x, context_tokens, step_time_emb, key_padding_mask)
            x = self.norm_out(x)
            x_pooled = self.seq_pool_proj(x.reshape(batch_size, -1)) if self.horizon > 1 else x.squeeze(1)
            pred = self.mlp_head(x_pooled, embodiment_id)
            return pred.view(batch_size, self.horizon, per_action_dim) * action_mask
        for i in range(num_steps):
            t = i / num_steps
            time_index = min(int(t * 999), 999)
            time_emb = self.time_pos_enc(1000)[:, time_index, :].to(device).squeeze(0).to(dtype=target_dtype)
            time_emb = time_emb.unsqueeze(0).repeat(batch_size, 1)
            if use_rtc:
                # RTCProcessor assumes the pi0 flow convention: its `time` runs 1 -> 0 and the
                # clean-action estimate is x1 = x_t - time * v. EVO1 integrates t: 0 -> 1 with
                # velocity v = x1 - x0 (so x1 = x_t + (1 - t) * v); passing time = 1 - t and
                # flipping the velocity sign in both directions maps one convention onto the other.
                guided = rtc_processor.denoise_step(
                    x_t=action_seq,
                    prev_chunk_left_over=prev_chunk_left_over,
                    inference_delay=inference_delay,
                    time=1.0 - t,
                    original_denoise_step_partial=lambda seq, emb=time_emb: -predict_velocity(seq, emb),
                    execution_horizon=execution_horizon,
                )
                velocity = -guided
            else:
                velocity = predict_velocity(action_seq, time_emb)
            action_seq = action_seq + dt * velocity
        action_seq = action_seq * action_mask
        return action_seq.reshape(batch_size, -1)
    @property
    def device(self):
        return next(self.parameters()).device
    @property
    def dtype(self):
        return next(self.parameters()).dtype
@@ -1,369 +0,0 @@
 # Copyright 2026 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
 #     http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from __future__ import annotations
 import logging
 from collections.abc import Sequence
 from typing import TYPE_CHECKING
 import torch
 import torch.nn as nn
 import torchvision.transforms.functional as tvf
 from torchvision.transforms.functional import InterpolationMode
 from lerobot.utils.import_utils import _transformers_available, require_package
 if TYPE_CHECKING or _transformers_available:
    from transformers import AutoModel, AutoTokenizer
 else:
    AutoModel = None
    AutoTokenizer = None
 IMAGENET_MEAN = (0.485, 0.456, 0.406)
 IMAGENET_STD = (0.229, 0.224, 0.225)
 IMG_CONTEXT_TOKEN = "<IMG_CONTEXT>"  # nosec B105
 IMG_START_TOKEN = "<img>"  # nosec B105
 IMG_END_TOKEN = "</img>"  # nosec B105
 logger = logging.getLogger(__name__)
 def _batched_resize_01(images: torch.Tensor, image_size: int) -> torch.Tensor:
    """Resize a batch of ``[0, 1]`` images to ``(image_size, image_size)`` on-device.
    Numerically mirrors InternVL3's reference PIL preprocessing
    (``to_pil_image`` -> ``Image.resize`` -> ``to_tensor``): the float input is quantized to uint8
    exactly as ``to_pil_image`` does, then resized with bicubic interpolation and antialiasing,
    which matches PIL's default resampler. Matching the reference pixel-for-pixel keeps the policy
    interchangeable with checkpoints produced by the upstream EVO1 preprocessing.
    Args:
        images: float tensor of shape ``(N, C, H, W)`` with values in ``[0, 1]``.
    Returns:
        float32 tensor of shape ``(N, C, image_size, image_size)`` with values in ``[0, 1]``.
    """
    # to_pil_image() quantizes float [0, 1] to uint8 (x * 255, truncated); replicate that so the
    # bicubic resample sees the same integer pixels PIL would.
    pixels_u8 = (images * 255.0).clamp(0, 255).to(torch.uint8)
    resized = tvf.resize(
        pixels_u8, [image_size, image_size], interpolation=InterpolationMode.BICUBIC, antialias=True
    )
    return resized.to(torch.float32) / 255.0
 def _batched_pixel_values(
    camera_images: Sequence[torch.Tensor],
    max_views: int,
    image_size: int,
    mean: torch.Tensor,
    std: torch.Tensor,
    dtype: torch.dtype,
    device: torch.device | str,
 ) -> torch.Tensor:
    """Build InternVL3 ``pixel_values`` from per-camera ``[0, 1]`` image batches without leaving the device.
    Each image is resized, converted to ``dtype``, and ImageNet-normalized (a single tile per
    image), batched across the whole minibatch. Absent views (fewer cameras than ``max_views``)
    are filled with zero images; their placeholder tokens are masked out of attention downstream
    via ``_mask_absent_image_tokens``.
    Returns:
        ``pixel_values`` of shape ``(B * max_views, C, image_size, image_size)``, ordered row-major
        over ``(sample, view)`` to line up with the per-view image placeholders in the prompt.
    """
    resized: list[torch.Tensor] = []
    for image in camera_images:
        resized.append(_batched_resize_01(image.to(device=device), image_size).to(dtype))
    batch_size = resized[0].shape[0]
    channels = resized[0].shape[1]
    while len(resized) < max_views:
        resized.append(torch.zeros(batch_size, channels, image_size, image_size, dtype=dtype, device=device))
    stacked = torch.stack(resized[:max_views], dim=1)  # (B, V, C, H, W)
    mean = mean.to(device=device, dtype=dtype).view(1, 1, -1, 1, 1)
    std = std.to(device=device, dtype=dtype).view(1, 1, -1, 1, 1)
    normalized = (stacked - mean) / std
    return normalized.reshape(batch_size * max_views, channels, image_size, image_size)
 class InternVL3Embedder(nn.Module):
    """Vision-language embedder using the native HF InternVL3 model (no trust_remote_code)."""
    def __init__(
        self,
        model_name="OpenGVLab/InternVL3-1B-hf",
        image_size=448,
        device="cuda",
        num_language_layers: int | None = 14,
        model_dtype: str | torch.dtype = "bfloat16",
        use_flash_attn: bool = True,
        max_text_length: int = 1024,
        enable_gradient_checkpointing: bool = True,
        gradient_checkpointing_use_reentrant: bool = False,
        hub_kwargs: dict | None = None,
    ):
        super().__init__()
        self._requested_device = device
        self.image_size = image_size
        self.num_language_layers = num_language_layers
        self.max_text_length = max_text_length
        self.enable_gradient_checkpointing = bool(enable_gradient_checkpointing)
        self.gradient_checkpointing_use_reentrant = bool(gradient_checkpointing_use_reentrant)
        hub_kwargs = hub_kwargs or {}
        require_package("transformers", extra="evo1")
        self.tokenizer = AutoTokenizer.from_pretrained(model_name, **hub_kwargs)
        if isinstance(model_dtype, str):
            try:
                model_dtype = getattr(torch, model_dtype)
            except AttributeError as exc:
                raise ValueError(f"Unsupported EVO1 vlm_dtype '{model_dtype}'") from exc
        self.model_dtype = model_dtype
        attn_implementation = "flash_attention_2" if (use_flash_attn and _flash_attn_available()) else "eager"
        if use_flash_attn and attn_implementation == "eager":
            logger.warning("flash_attn is not installed. Falling back to eager attention.")
        self.model = AutoModel.from_pretrained(
            model_name,
            torch_dtype=model_dtype,
            attn_implementation=attn_implementation,
            low_cpu_mem_usage=True,
            **hub_kwargs,
        ).to(self._requested_device)
        checkpoint_image_size = getattr(self.model.config.vision_config, "image_size", None)
        if isinstance(checkpoint_image_size, (list, tuple)):
            checkpoint_image_size = checkpoint_image_size[0]
        if checkpoint_image_size is not None and int(checkpoint_image_size) != int(image_size):
            raise ValueError(
                f"EVO1 image_resolution ({image_size}) must match the InternVL checkpoint's native "
                f"image size ({checkpoint_image_size}): the checkpoint's image_seq_length assumes "
                "its native resolution, so other sizes would desync the image placeholder tokens "
                "from the vision features."
            )
        self.num_image_token = self.model.config.image_seq_length
        # Truncate language model to the requested number of layers
        layers = self.model.language_model.layers
        if self.num_language_layers is not None:
            layers = layers[: self.num_language_layers]
        self.model.language_model.layers = torch.nn.ModuleList(layers)
        self._configure_memory_features()
        self.img_context_token_id = self.tokenizer.convert_tokens_to_ids(IMG_CONTEXT_TOKEN)
    def _configure_memory_features(self) -> None:
        checkpoint_kwargs = {"use_reentrant": self.gradient_checkpointing_use_reentrant}
        if not self.enable_gradient_checkpointing:
            language_model = self.model.language_model
            if hasattr(language_model, "gradient_checkpointing_disable"):
                language_model.gradient_checkpointing_disable()
            vision_tower = getattr(self.model, "vision_tower", None)
            if vision_tower is not None and hasattr(vision_tower, "encoder"):
                vision_tower.encoder.gradient_checkpointing = False
            return
        def _enable_ckpt(module: nn.Module | None) -> bool:
            if module is None:
                return False
            if hasattr(module, "gradient_checkpointing_enable"):
                try:
                    module.gradient_checkpointing_enable(gradient_checkpointing_kwargs=checkpoint_kwargs)
                except TypeError:
                    module.gradient_checkpointing_enable()
                return True
            if hasattr(module, "gradient_checkpointing"):
                module.gradient_checkpointing = True
                return True
            return False
        enabled_any = _enable_ckpt(self.model)
        vision_tower = getattr(self.model, "vision_tower", None)
        if vision_tower is not None:
            enabled_any = _enable_ckpt(vision_tower) or enabled_any
        language_model = self.model.language_model
        enabled_any = _enable_ckpt(language_model) or enabled_any
        if hasattr(language_model, "config"):
            language_model.config.use_cache = False
        if hasattr(self.model, "config"):
            self.model.config.use_cache = False
        if hasattr(self.model, "enable_input_require_grads"):
            self.model.enable_input_require_grads()
        if enabled_any:
            logger.info("Gradient checkpointing enabled for InternVL3 embedder.")
        else:
            logger.warning(
                "Requested gradient checkpointing, but model does not expose checkpointing controls."
            )
    def _build_multimodal_prompts(
        self,
        batch_num_tiles_list: list[list[int]],
        text_prompts: Sequence[str],
    ) -> list[str]:
        prompts = []
        for num_tiles_list, text_prompt in zip(batch_num_tiles_list, text_prompts, strict=True):
            prompt_segments = []
            for i, tile_count in enumerate(num_tiles_list):
                token_count = self.num_image_token * tile_count
                image_tokens = IMG_START_TOKEN + IMG_CONTEXT_TOKEN * token_count + IMG_END_TOKEN
                prompt_segments.append(f"Image-{i + 1}: {image_tokens}\n")
            prompts.append("".join(prompt_segments) + text_prompt.strip())
        return prompts
    def get_fused_image_text_embedding_batched(
        self,
        camera_images: Sequence[torch.Tensor],
        image_masks: torch.Tensor,
        text_prompts: Sequence[str],
        return_cls_only: bool = True,
    ):
        """Fused VL embedding from per-camera ``[0, 1]`` image batches (no PIL, no host round-trip).
        Args:
            camera_images: list of per-camera tensors, each shaped ``(B, C, H, W)`` in ``[0, 1]``.
            image_masks: bool tensor ``(B, max_views)`` marking present views.
        Returns:
            A ``(embeddings, valid_mask)`` tuple. With ``return_cls_only=False``, ``embeddings`` is
            ``(B, L, H)`` and ``valid_mask`` is a ``(B, L)`` bool tensor marking tokens downstream
            attention may attend to (padding and absent-view tokens are False). With
            ``return_cls_only=True``, ``embeddings`` is the pooled ``(B, H)`` last-valid-token state
            and ``valid_mask`` is None.
        """
        max_views = int(image_masks.shape[1])
        batch_size = int(image_masks.shape[0])
        mean = torch.tensor(IMAGENET_MEAN, device=self.device, dtype=self.model_dtype)
        std = torch.tensor(IMAGENET_STD, device=self.device, dtype=self.model_dtype)
        pixel_values = _batched_pixel_values(
            camera_images, max_views, self.image_size, mean, std, self.model_dtype, self.device
        )
        # InternVL3 preprocessing uses a single tile per image (max_num=1).
        batch_num_tiles_list = [[1] * max_views for _ in range(batch_size)]
        return self._forward_vlm(
            pixel_values, batch_num_tiles_list, image_masks, text_prompts, return_cls_only
        )
    def _mask_absent_image_tokens(
        self,
        input_ids: torch.Tensor,
        attention_mask: torch.Tensor,
        image_masks: torch.Tensor,
        batch_num_tiles_list: list[list[int]],
    ) -> torch.Tensor:
        """Zero attention over the image-context tokens of absent (zero-padded) views.
        Fully vectorized: runs without any host<->device synchronization.
        """
        # A single tile per image (max_num=1), so every image occupies the same number of
        # context tokens.
        tiles_per_image = (
            batch_num_tiles_list[0][0] if batch_num_tiles_list and batch_num_tiles_list[0] else 1
        )
        tokens_per_image = self.num_image_token * tiles_per_image
        image_masks = image_masks.to(device=input_ids.device).bool()
        img_token_mask = input_ids == self.img_context_token_id  # (B, L)
        # keep[b, k] tells whether the k-th image-context token (ordered view0, view1, ...) survives.
        per_token_keep = image_masks.repeat_interleave(tokens_per_image, dim=1)  # (B, V * tokens_per_image)
        # Rank each context token by its running position among the row's context tokens.
        ctx_index = img_token_mask.to(torch.long).cumsum(dim=1) - 1
        ctx_index = ctx_index.clamp(min=0, max=per_token_keep.shape[1] - 1)
        keep_here = torch.gather(per_token_keep, 1, ctx_index)  # (B, L)
        drop = img_token_mask & ~keep_here
        return attention_mask.masked_fill(drop, 0)
    def _forward_vlm(
        self,
        pixel_values: torch.Tensor,
        batch_num_tiles_list: list[list[int]],
        image_masks: torch.Tensor,
        text_prompts: Sequence[str],
        return_cls_only: bool,
    ):
        if pixel_values.shape[0] == 0:
            logger.warning("InternVL3 received an empty image batch after preprocessing.")
            hidden_size = getattr(self.model.config, "hidden_size", None)
            if hidden_size is None:
                hidden_size = getattr(self.model.config.text_config, "hidden_size", None)
            if hidden_size is None:
                raise RuntimeError("Unable to infer hidden size for empty InternVL3 batch.")
            return torch.empty(0, hidden_size, device=self.device, dtype=torch.float32), None
        prompts = self._build_multimodal_prompts(batch_num_tiles_list, text_prompts)
        model_inputs = self.tokenizer(
            list(prompts),
            return_tensors="pt",
            padding=True,
            truncation=True,
            max_length=self.max_text_length,
        ).to(self.device)
        input_ids = model_inputs["input_ids"]
        if input_ids.shape[1] >= self.max_text_length:
            # Truncation cuts from the right, so text is dropped before image placeholders — but a
            # large max_views * image_seq_length budget can still eat into them. Fail loudly instead
            # of letting the VLM crash on a placeholder/vision-feature count mismatch.
            expected_image_tokens = self.num_image_token * sum(batch_num_tiles_list[0])
            image_token_counts = (input_ids == self.img_context_token_id).sum(dim=1)
            if not bool((image_token_counts == expected_image_tokens).all()):
                raise ValueError(
                    f"Prompt truncation at max_text_length={self.max_text_length} cut into the "
                    f"image placeholder tokens ({expected_image_tokens} expected per sample). "
                    "Increase max_text_length or reduce max_views."
                )
        attention_mask = self._mask_absent_image_tokens(
            input_ids, model_inputs["attention_mask"], image_masks, batch_num_tiles_list
        )
        outputs = self.model(
            input_ids=input_ids,
            pixel_values=pixel_values,
            attention_mask=attention_mask,
            output_hidden_states=True,
            return_dict=True,
        )
        fused_hidden = outputs.hidden_states[-1].to(torch.float32)
        valid_mask = attention_mask.to(torch.bool)
        if return_cls_only:
            # Right-padded causal decoder: the last valid token is the only one that has attended
            # to the full image + text prompt.
            positions = torch.arange(valid_mask.shape[1], device=valid_mask.device)
            last_valid = (valid_mask.long() * positions).argmax(dim=1)
            batch_index = torch.arange(fused_hidden.shape[0], device=fused_hidden.device)
            return fused_hidden[batch_index, last_valid], None
        return fused_hidden, valid_mask
    @property
    def device(self) -> torch.device:
        return next(self.model.parameters()).device
 def _flash_attn_available() -> bool:
    try:
        import flash_attn  # noqa: F401
    except ModuleNotFoundError:
        return False
    return True
@@ -1,532 +0,0 @@
 # Copyright 2026 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
 #     http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from __future__ import annotations
 import builtins
 from collections import deque
 from contextlib import nullcontext
 from pathlib import Path
 from typing import TypedDict, Unpack
 import torch
 from torch import Tensor
 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.policies.pretrained import PreTrainedPolicy, T
 from lerobot.utils.constants import ACTION, OBS_IMAGES, OBS_STATE
 from ..rtc.modeling_rtc import RTCProcessor
 from .configuration_evo1 import Evo1Config
 from .evo1_model import Evo1Model
 class ActionSelectKwargs(TypedDict, total=False):
    inference_delay: int | None
    prev_chunk_left_over: Tensor | None
    execution_horizon: int | None
 class Evo1Policy(PreTrainedPolicy):
    config_class = Evo1Config
    name = "evo1"
    def __init__(self, config: Evo1Config, *, vlm_hub_kwargs: dict | None = None, **kwargs):
        super().__init__(config)
        config.validate_features()
        if len(config.image_features) > config.max_views:
            raise ValueError(
                f"EVO1 supports at most {config.max_views} camera streams, got {len(config.image_features)}"
            )
        self.config = config
        self.model = Evo1Model(config, vlm_hub_kwargs=vlm_hub_kwargs)
        self.model.set_finetune_flags()
        self._keep_frozen_embedder_eval()
        self.init_rtc_processor()
        self.reset()
    def init_rtc_processor(self):
        """Create the RTC processor when config.rtc_config is set.
        The RTC rollout backend assigns config.rtc_config after loading the policy and re-invokes
        this method.
        """
        self.rtc_processor = None
        if self.config.rtc_config is not None:
            self.rtc_processor = RTCProcessor(self.config.rtc_config)
        model = getattr(self, "model", None)
        if model is not None:
            model.rtc_processor = self.rtc_processor
    def _rtc_enabled(self) -> bool:
        return self.config.rtc_config is not None and self.config.rtc_config.enabled
    @classmethod
    def from_pretrained(
        cls: builtins.type[T],
        pretrained_name_or_path: str | Path,
        *,
        config: PreTrainedConfig | None = None,
        force_download: bool = False,
        resume_download: bool | None = None,
        proxies: dict | None = None,
        token: str | bool | None = None,
        cache_dir: str | Path | None = None,
        local_files_only: bool = False,
        revision: str | None = None,
        strict: bool | None = None,
        **kwargs,
    ) -> T:
        if strict is None:
            strict = True
        vlm_hub_kwargs = kwargs.pop("vlm_hub_kwargs", None)
        if config is None:
            config = PreTrainedConfig.from_pretrained(
                pretrained_name_or_path=pretrained_name_or_path,
                force_download=force_download,
                resume_download=resume_download,
                proxies=proxies,
                token=token,
                cache_dir=cache_dir,
                local_files_only=local_files_only,
                revision=revision,
                **kwargs,
            )
        if vlm_hub_kwargs is None:
            # Forward the hub download options to the base-VLM download as well; `revision` is not
            # forwarded because it identifies the policy repo, not the VLM repo.
            vlm_hub_kwargs = {
                key: value
                for key, value in (
                    ("token", token),
                    ("cache_dir", cache_dir),
                    ("local_files_only", local_files_only),
                    ("proxies", proxies),
                )
                if value not in (None, False)
            }
        kwargs["vlm_hub_kwargs"] = vlm_hub_kwargs
        return super().from_pretrained(
            pretrained_name_or_path=pretrained_name_or_path,
            config=config,
            force_download=force_download,
            resume_download=resume_download,
            proxies=proxies,
            token=token,
            cache_dir=cache_dir,
            local_files_only=local_files_only,
            revision=revision,
            strict=strict,
            **kwargs,
        )
    @property
    def _camera_keys(self) -> list[str]:
        return list(self.config.image_features)
    @property
    def _env_action_dim(self) -> int:
        action_feature = self.config.action_feature
        if action_feature is None:
            return self.config.max_action_dim
        return int(action_feature.shape[0])
    @property
    def _compute_dtype(self) -> torch.dtype:
        return next(self.model.action_head.parameters()).dtype
    @property
    def _device(self) -> torch.device:
        # The device the policy actually lives on. Derived from the parameters rather than
        # config.device so the policy keeps working after accelerate (or a plain .to()) moves it.
        return next(self.model.action_head.parameters()).device
    @property
    def _amp_enabled(self) -> bool:
        return bool(self.config.use_amp) and self._device.type == "cuda"
    def _maybe_autocast(self):
        # EVO1 manages its own mixed precision: an explicit bf16 autocast that also overrides any
        # outer autocast context (e.g. lerobot-eval's fp16 default), keeping train and eval
        # numerics identical.
        if self._amp_enabled:
            return torch.autocast(device_type="cuda", dtype=torch.bfloat16)
        return nullcontext()
    def get_optim_params(self) -> list[dict]:
        decay, no_decay = [], []
        for name, param in self.named_parameters():
            if not param.requires_grad:
                continue
            is_bias = name.endswith("bias") or ".bias" in name
            is_norm = param.dim() == 1 or "norm" in name.lower()
            if is_bias or is_norm:
                no_decay.append(param)
            else:
                decay.append(param)
        return [
            {"params": decay, "weight_decay": self.config.optimizer_weight_decay},
            {"params": no_decay, "weight_decay": 0.0},
        ]
    def reset(self):
        self._action_queue = deque([], maxlen=self.config.n_action_steps)
    def _normalize_task_batch(self, batch: dict[str, Tensor | list[str] | str]) -> list[str]:
        prompts = batch.get(self.config.task_field)
        if prompts is None and self.config.task_field != "task":
            prompts = batch.get("task")
        if prompts is None:
            raise ValueError(f"EVO1 expects a '{self.config.task_field}' text field in the batch.")
        if isinstance(prompts, str):
            return [prompts]
        if isinstance(prompts, (list, tuple)):
            return [str(prompt) for prompt in prompts]
        raise TypeError(f"Unsupported prompt batch type: {type(prompts)}")
    def _prepare_state(self, batch: dict[str, Tensor]) -> tuple[Tensor, Tensor]:
        if OBS_STATE not in batch:
            raise ValueError(f"EVO1 requires '{OBS_STATE}' in the batch.")
        state = batch[OBS_STATE]
        if state.dim() == 1:
            state = state.unsqueeze(0)
        elif state.dim() == 3:
            state = state[:, -1]
        elif state.dim() != 2:
            raise ValueError(f"Unsupported state tensor shape for EVO1: {tuple(state.shape)}")
        batch_size, state_dim = state.shape
        if state_dim > self.config.max_state_dim:
            raise ValueError(
                f"State dim {state_dim} exceeds configured max_state_dim {self.config.max_state_dim}"
            )
        explicit_mask = batch.get("state_mask")
        if explicit_mask is not None:
            if explicit_mask.dim() == 1:
                explicit_mask = explicit_mask.unsqueeze(0)
            elif explicit_mask.dim() == 3:
                explicit_mask = explicit_mask[:, -1]
            elif explicit_mask.dim() != 2:
                raise ValueError(
                    f"Unsupported state_mask tensor shape for EVO1: {tuple(explicit_mask.shape)}"
                )
            if explicit_mask.shape != (batch_size, state_dim):
                raise ValueError(
                    f"state_mask shape {tuple(explicit_mask.shape)} does not match state shape {(batch_size, state_dim)}"
                )
        device = self._device
        padded = torch.zeros(
            batch_size,
            self.config.max_state_dim,
            dtype=state.dtype,
            device=device,
        )
        padded[:, :state_dim] = state.to(device=device)
        mask = torch.zeros(
            batch_size,
            self.config.max_state_dim,
            dtype=torch.bool,
            device=device,
        )
        if explicit_mask is None:
            mask[:, :state_dim] = True
        else:
            mask[:, :state_dim] = explicit_mask.to(device=device, dtype=torch.bool)
        # Zero out masked state dims so an explicit state_mask actually affects the model input
        # (the state encoder has no mask argument of its own).
        padded = padded * mask.to(dtype=padded.dtype)
        return padded.to(dtype=self._compute_dtype), mask
    def _prepare_actions(self, batch: dict[str, Tensor]) -> tuple[Tensor, Tensor]:
        if ACTION not in batch:
            raise ValueError(f"EVO1 requires '{ACTION}' in the batch for training.")
        action = batch[ACTION]
        if action.dim() == 2:
            action = action.unsqueeze(1)
        batch_size, horizon, action_dim = action.shape
        if horizon != self.config.chunk_size:
            raise ValueError(
                f"EVO1 expects chunk_size={self.config.chunk_size}, got action horizon {horizon}"
            )
        if action_dim > self.config.max_action_dim:
            raise ValueError(
                f"Action dim {action_dim} exceeds configured max_action_dim {self.config.max_action_dim}"
            )
        explicit_mask = batch.get("action_mask")
        if explicit_mask is not None:
            if explicit_mask.dim() == 2:
                if horizon == 1:
                    explicit_mask = explicit_mask.unsqueeze(1)
                else:
                    raise ValueError(
                        f"2D action_mask is only supported when chunk_size=1, got action horizon {horizon}"
                    )
            elif explicit_mask.dim() != 3:
                raise ValueError(
                    f"Unsupported action_mask tensor shape for EVO1: {tuple(explicit_mask.shape)}"
                )
            if explicit_mask.shape != (batch_size, horizon, action_dim):
                raise ValueError(
                    "action_mask shape "
                    f"{tuple(explicit_mask.shape)} does not match action shape {(batch_size, horizon, action_dim)}"
                )
        device = self._device
        padded = torch.zeros(
            batch_size,
            horizon,
            self.config.max_action_dim,
            dtype=action.dtype,
            device=device,
        )
        padded[:, :, :action_dim] = action.to(device=device)
        mask = torch.zeros(
            batch_size,
            horizon,
            self.config.max_action_dim,
            dtype=torch.bool,
            device=device,
        )
        if explicit_mask is None:
            mask[:, :, :action_dim] = True
        else:
            mask[:, :, :action_dim] = explicit_mask.to(device=device, dtype=torch.bool)
        # Timesteps beyond the episode end hold fabricated (repeated) actions; exclude them from
        # the loss like the other chunked policies do.
        action_is_pad = batch.get("action_is_pad")
        if action_is_pad is not None:
            if action_is_pad.shape != (batch_size, horizon):
                raise ValueError(
                    f"action_is_pad shape {tuple(action_is_pad.shape)} does not match "
                    f"(batch_size, chunk_size)={(batch_size, horizon)}"
                )
            in_episode = ~action_is_pad.to(device=device, dtype=torch.bool)
            mask = mask & in_episode.unsqueeze(-1)
        return padded.to(dtype=self._compute_dtype), mask
    def _prepare_inference_action_mask(self, batch_size: int) -> Tensor:
        mask = torch.zeros(
            batch_size,
            self.config.max_action_dim,
            dtype=torch.bool,
            device=self._device,
        )
        mask[:, : self._env_action_dim] = True
        return mask
    def _get_embodiment_ids(self, batch: dict[str, Tensor], batch_size: int) -> Tensor:
        embodiment_ids = batch.get("embodiment_id")
        if embodiment_ids is None and self.config.embodiment_id_field:
            embodiment_ids = batch.get(self.config.embodiment_id_field)
        if embodiment_ids is None:
            return torch.full(
                (batch_size,),
                self.config.default_embodiment_id,
                dtype=torch.long,
                device=self._device,
            )
        if embodiment_ids.dim() == 0:
            embodiment_ids = embodiment_ids.unsqueeze(0)
        elif embodiment_ids.dim() > 1:
            embodiment_ids = embodiment_ids[:, -1]
        return embodiment_ids.to(device=self._device, dtype=torch.long)
    @property
    def _tracks_vlm_gradients(self) -> bool:
        return bool(
            self.config.finetune_vlm
            or self.config.finetune_language_model
            or self.config.finetune_vision_model
        )
    def _keep_frozen_embedder_eval(self) -> None:
        if self._tracks_vlm_gradients:
            return
        embedder = getattr(self.model, "embedder", None)
        if embedder is not None:
            embedder.eval()
    def train(self, mode: bool = True):
        super().train(mode)
        self._keep_frozen_embedder_eval()
        return self
    def _collect_image_batches(self, batch: dict[str, Tensor]) -> tuple[list[Tensor], Tensor]:
        camera_keys = self._camera_keys or sorted(key for key in batch if key.startswith(f"{OBS_IMAGES}."))
        if not camera_keys:
            raise ValueError("EVO1 requires at least one visual observation feature.")
        camera_keys = list(camera_keys)[: self.config.max_views]
        # Configured cameras may be absent from the batch up to the empty_cameras budget (e.g. the
        # placeholder features added by validate_features); they become masked-out views that the
        # embedder zero-pads. Any other absent camera is an error.
        present_keys = [key for key in camera_keys if key in batch]
        missing_keys = [key for key in camera_keys if key not in batch]
        if len(missing_keys) > self.config.empty_cameras:
            raise ValueError(
                f"Missing camera features {missing_keys} in batch; at most "
                f"empty_cameras={self.config.empty_cameras} may be absent."
            )
        if not present_keys:
            raise ValueError("EVO1 requires at least one visual observation in the batch.")
        # Keep each present camera as a batched (B, C, H, W) tensor on its current (GPU) device.
        # Resizing/normalization and zero-padding of absent views happen batched inside the
        # embedder, so images never leave the device here.
        camera_images: list[Tensor] = []
        for camera_key in present_keys:
            image = batch[camera_key]
            if image.dim() == 3:
                # Promote an unbatched (C, H, W) frame so batch_size is read from a real batch dim.
                image = image.unsqueeze(0)
            elif image.dim() == 5:
                image = image[:, -1]
            elif image.dim() != 4:
                raise ValueError(
                    f"Unsupported image tensor shape for EVO1: key={camera_key} shape={tuple(image.shape)}"
                )
            camera_images.append(image)
        batch_size = camera_images[0].shape[0]
        n_present = len(camera_images)
        image_masks = torch.zeros(
            batch_size, self.config.max_views, dtype=torch.bool, device=camera_images[0].device
        )
        image_masks[:, :n_present] = True
        return camera_images, image_masks
    def _compute_fused_tokens(
        self,
        prompts: list[str],
        image_batches: list[Tensor],
        image_masks: Tensor,
    ) -> tuple[Tensor, Tensor | None]:
        track_vlm_gradients = self._tracks_vlm_gradients
        grad_context = nullcontext() if track_vlm_gradients else torch.no_grad()
        with grad_context:
            fused_tokens, context_mask = self.model.get_vl_embeddings(
                images=image_batches,
                image_mask=image_masks,
                prompt=prompts,
                return_cls_only=self.config.return_cls_only,
            )
        if not track_vlm_gradients:
            fused_tokens = fused_tokens.detach()
        fused_tokens = fused_tokens.to(device=self._device, dtype=self._compute_dtype)
        if context_mask is not None:
            context_mask = context_mask.to(device=self._device)
        return fused_tokens, context_mask
    def _compute_masked_loss(
        self,
        pred_velocity: Tensor,
        target_velocity: Tensor,
        action_mask: Tensor,
        reduction: str,
    ) -> Tensor:
        flat_mask = action_mask.view(action_mask.shape[0], -1).to(dtype=pred_velocity.dtype)
        sq_error = ((pred_velocity - target_velocity) * flat_mask).pow(2)
        active = flat_mask.sum(dim=1).clamp_min(1.0)
        per_sample_loss = sq_error.sum(dim=1) / active
        if reduction == "none":
            return per_sample_loss
        if reduction != "mean":
            raise ValueError(f"Unsupported reduction '{reduction}'")
        return sq_error.sum() / active.sum()
    def forward(self, batch: dict[str, Tensor], reduction: str = "mean") -> tuple[Tensor, dict]:
        prompts = self._normalize_task_batch(batch)
        image_batches, image_masks = self._collect_image_batches(batch)
        states, _state_mask = self._prepare_state(batch)
        actions_gt, action_mask = self._prepare_actions(batch)
        embodiment_ids = self._get_embodiment_ids(batch, states.shape[0])
        with self._maybe_autocast():
            fused_tokens, context_mask = self._compute_fused_tokens(prompts, image_batches, image_masks)
            pred_velocity, noise = self.model(
                fused_tokens,
                state=states,
                actions_gt=actions_gt,
                action_mask=action_mask.to(device=self._device, dtype=self._compute_dtype),
                embodiment_ids=embodiment_ids,
                context_mask=context_mask,
            )
        # Compute the flow-matching regression loss in fp32, outside the autocast block.
        pred_velocity = pred_velocity.float()
        noise = noise.float()
        flat_action_mask = action_mask.view(action_mask.shape[0], -1).to(dtype=torch.float32)
        # Flow-matching velocity target. Padded (masked-out) action dims are already zero on both sides
        # here (`actions_gt` is zero-padded in `_prepare_actions`, and `noise` is masked inside the head),
        # and the whole difference is multiplied by `flat_action_mask`, so padded dims contribute nothing.
        target_velocity = (actions_gt.float() - noise).view(actions_gt.shape[0], -1) * flat_action_mask
        loss = self._compute_masked_loss(pred_velocity, target_velocity, action_mask, reduction)
        loss_mean = loss.mean().item() if loss.ndim > 0 else loss.item()
        return loss, {
            "loss": loss_mean,
            "active_action_dims": float(action_mask.sum(dim=(1, 2)).float().mean().item()),
        }
    @torch.no_grad()
    def predict_action_chunk(self, batch: dict[str, Tensor], **kwargs: Unpack[ActionSelectKwargs]) -> Tensor:
        inference_delay = kwargs.get("inference_delay")
        prev_chunk_left_over = kwargs.get("prev_chunk_left_over")
        execution_horizon = kwargs.get("execution_horizon")
        if (inference_delay is not None or prev_chunk_left_over is not None) and not self._rtc_enabled():
            raise RuntimeError(
                "Received RTC arguments but RTC is not configured for this EVO1 policy: set "
                "config.rtc_config and call init_rtc_processor() (lerobot-rollout does this for "
                "--inference.type=rtc)."
            )
        self.eval()
        prompts = self._normalize_task_batch(batch)
        image_batches, image_masks = self._collect_image_batches(batch)
        states, _state_mask = self._prepare_state(batch)
        embodiment_ids = self._get_embodiment_ids(batch, states.shape[0])
        action_mask = self._prepare_inference_action_mask(states.shape[0])
        if prev_chunk_left_over is not None:
            prev_chunk_left_over = prev_chunk_left_over.to(device=self._device)
        with self._maybe_autocast():
            fused_tokens, context_mask = self._compute_fused_tokens(prompts, image_batches, image_masks)
            actions = self.model(
                fused_tokens,
                state=states,
                action_mask=action_mask,
                embodiment_ids=embodiment_ids,
                context_mask=context_mask,
                inference_delay=inference_delay,
                prev_chunk_left_over=prev_chunk_left_over,
                execution_horizon=execution_horizon,
            )
        actions = actions.view(states.shape[0], self.config.chunk_size, self.config.max_action_dim)
        return actions.to(dtype=torch.float32)
    @torch.no_grad()
    def select_action(self, batch: dict[str, Tensor], **kwargs) -> Tensor:
        assert not self._rtc_enabled(), (
            "RTC is not supported for select_action, use it with predict_action_chunk"
        )
        self.eval()
        if len(self._action_queue) == 0:
            action_chunk = self.predict_action_chunk(batch)[:, : self.config.n_action_steps]
            self._action_queue.extend(action_chunk.transpose(0, 1))
        # Returns one step of shape (B, max_action_dim): actions are emitted at the padded max_action_dim
        # width and cropped to the real action dim downstream by the postprocessor (Evo1ActionProcessorStep).
        # Callers that bypass the postprocessor receive the padded width.
        return self._action_queue.popleft()
@@ -1,400 +0,0 @@
 # Copyright 2026 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
 #     http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from __future__ import annotations
 from copy import deepcopy
 from dataclasses import dataclass
 from typing import Any
 import torch
 from lerobot.configs import FeatureType, PipelineFeatureType, PolicyFeature
 from lerobot.processor import (
    AddBatchDimensionProcessorStep,
    DeviceProcessorStep,
    NormalizerProcessorStep,
    ObservationProcessorStep,
    PolicyAction,
    PolicyActionProcessorStep,
    PolicyProcessorPipeline,
    ProcessorStep,
    ProcessorStepRegistry,
    RenameObservationsProcessorStep,
    UnnormalizerProcessorStep,
 )
 from lerobot.processor.converters import (
    batch_to_transition,
    create_transition,
    policy_action_to_transition,
    transition_to_policy_action,
 )
 from lerobot.types import EnvTransition, TransitionKey
 from lerobot.utils.constants import (
    ACTION,
    DONE,
    INFO,
    OBS_PREFIX,
    OBS_STATE,
    POLICY_POSTPROCESSOR_DEFAULT_NAME,
    POLICY_PREPROCESSOR_DEFAULT_NAME,
    REWARD,
    TRUNCATED,
 )
 from .configuration_evo1 import Evo1Config
 def evo1_batch_to_transition(batch: dict[str, Any]):
    transition = batch_to_transition(batch)
    complementary_data = dict(transition.get("complementary_data") or {})
    reserved = {ACTION, REWARD, DONE, TRUNCATED, INFO}
    for key, value in batch.items():
        if key in reserved or key.startswith(OBS_PREFIX):
            continue
        complementary_data.setdefault(key, value)
    return create_transition(
        observation=transition.get("observation"),
        action=transition.get("action"),
        reward=transition.get("reward", 0.0),
        done=transition.get("done", False),
        truncated=transition.get("truncated", False),
        info=transition.get("info", {}),
        complementary_data=complementary_data,
    )
@dataclass
@ProcessorStepRegistry.register(name="evo1_pad_state_processor")
 class Evo1PadStateProcessorStep(ObservationProcessorStep):
    """Pad policy observations to EVO1's fixed state width before normalization."""
    max_state_dim: int = 24
    def observation(self, observation: dict[str, Any]) -> dict[str, Any]:
        if OBS_STATE not in observation:
            return observation
        state = observation[OBS_STATE]
        state_dim = state.shape[-1]
        if state_dim > self.max_state_dim:
            raise ValueError(
                f"EVO1 state has {state_dim} dims, which exceeds max_state_dim={self.max_state_dim}."
            )
        if state_dim < self.max_state_dim:
            observation = observation.copy()
            observation[OBS_STATE] = torch.nn.functional.pad(state, (0, self.max_state_dim - state_dim))
        return observation
    def transform_features(
        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
        new_features = {ft: feats.copy() for ft, feats in features.items()}
        obs_feats = new_features.setdefault(PipelineFeatureType.OBSERVATION, {})
        if OBS_STATE in obs_feats:
            obs_feats[OBS_STATE] = PolicyFeature(type=FeatureType.STATE, shape=(self.max_state_dim,))
        return new_features
    def get_config(self) -> dict[str, Any]:
        return {"max_state_dim": self.max_state_dim}
@dataclass
@ProcessorStepRegistry.register(name="evo1_pad_action_processor")
 class Evo1PadActionProcessorStep(ProcessorStep):
    """Pad training actions and preserve the active action dimensions with action_mask."""
    max_action_dim: int = 24
    def __call__(self, transition: EnvTransition) -> EnvTransition:
        action = transition.get(TransitionKey.ACTION)
        if action is None:
            return transition
        if not isinstance(action, PolicyAction):
            raise ValueError(f"EVO1 action should be a PolicyAction tensor, but got {type(action)}.")
        action_dim = action.shape[-1]
        if action_dim > self.max_action_dim:
            raise ValueError(
                f"EVO1 action has {action_dim} dims, which exceeds max_action_dim={self.max_action_dim}."
            )
        new_transition = transition.copy()
        new_action = action
        if action_dim < self.max_action_dim:
            new_action = torch.nn.functional.pad(action, (0, self.max_action_dim - action_dim))
        complementary_data = dict(new_transition.get(TransitionKey.COMPLEMENTARY_DATA) or {})
        action_mask = complementary_data.get("action_mask")
        if action_mask is None:
            action_mask = torch.ones(action.shape, dtype=torch.bool, device=action.device)
        else:
            action_mask = torch.as_tensor(action_mask, dtype=torch.bool, device=action.device)
            if action_mask.shape != action.shape:
                raise ValueError(
                    f"action_mask shape {tuple(action_mask.shape)} does not match action shape {tuple(action.shape)}."
                )
        if action_dim < self.max_action_dim:
            action_mask = torch.nn.functional.pad(action_mask, (0, self.max_action_dim - action_dim))
        complementary_data["action_mask"] = action_mask
        new_transition[TransitionKey.ACTION] = new_action
        new_transition[TransitionKey.COMPLEMENTARY_DATA] = complementary_data
        return new_transition
    def transform_features(
        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
        new_features = {ft: feats.copy() for ft, feats in features.items()}
        action_feats = new_features.setdefault(PipelineFeatureType.ACTION, {})
        action_feats[ACTION] = PolicyFeature(type=FeatureType.ACTION, shape=(self.max_action_dim,))
        return new_features
    def get_config(self) -> dict[str, Any]:
        return {"max_action_dim": self.max_action_dim}
@dataclass
@ProcessorStepRegistry.register(name="evo1_action_processor")
 class Evo1ActionProcessorStep(PolicyActionProcessorStep):
    """Crop padded EVO1 actions and optionally binarize the LIBERO gripper channel."""
    action_dim: int
    binarize_gripper: bool = False
    gripper_index: int = 6
    gripper_threshold: float = 0.5
    gripper_below_threshold_value: float = 1.0
    gripper_above_threshold_value: float = -1.0
    def action(self, action: PolicyAction) -> PolicyAction:
        if action.shape[-1] < self.action_dim:
            raise ValueError(
                f"EVO1 action has {action.shape[-1]} dims, which is smaller than action_dim={self.action_dim}."
            )
        action = action[..., : self.action_dim]
        if not self.binarize_gripper:
            return action
        if not 0 <= self.gripper_index < self.action_dim:
            raise ValueError(
                f"gripper_index={self.gripper_index} must be within action_dim={self.action_dim}."
            )
        action = action.clone()
        below = torch.as_tensor(
            self.gripper_below_threshold_value,
            dtype=action.dtype,
            device=action.device,
        )
        above = torch.as_tensor(
            self.gripper_above_threshold_value,
            dtype=action.dtype,
            device=action.device,
        )
        action[..., self.gripper_index] = torch.where(
            action[..., self.gripper_index] > self.gripper_threshold,
            above,
            below,
        )
        return action
    def transform_features(
        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
        new_features = {ft: feats.copy() for ft, feats in features.items()}
        action_feats = new_features.setdefault(PipelineFeatureType.ACTION, {})
        action_feats[ACTION] = PolicyFeature(type=FeatureType.ACTION, shape=(self.action_dim,))
        return new_features
    def get_config(self) -> dict[str, Any]:
        return {
            "action_dim": self.action_dim,
            "binarize_gripper": self.binarize_gripper,
            "gripper_index": self.gripper_index,
            "gripper_threshold": self.gripper_threshold,
            "gripper_below_threshold_value": self.gripper_below_threshold_value,
            "gripper_above_threshold_value": self.gripper_above_threshold_value,
        }
 def _evo1_action_dim(config: Evo1Config) -> int:
    if config.postprocess_action_dim is not None:
        return config.postprocess_action_dim
    action_feature = config.action_feature
    if action_feature is None:
        return config.max_action_dim
    return int(action_feature.shape[0])
 def _evo1_normalization_features(config: Evo1Config) -> dict[str, PolicyFeature]:
    features = {**config.input_features, **config.output_features}
    features[OBS_STATE] = PolicyFeature(type=FeatureType.STATE, shape=(config.max_state_dim,))
    features[ACTION] = PolicyFeature(type=FeatureType.ACTION, shape=(config.max_action_dim,))
    return features
 def _evo1_action_features(config: Evo1Config) -> dict[str, PolicyFeature]:
    return {ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(config.max_action_dim,))}
 _STAT_PAD_VALUES = {
    "mean": 0.0,
    "std": 1.0,
    "min": -1.0,
    "max": 1.0,
    "q01": -1.0,
    "q99": 1.0,
    "q10": -1.0,
    "q90": 1.0,
 }
 def _pad_stat_value(value: Any, target_dim: int, stat_name: str) -> torch.Tensor:
    tensor = torch.as_tensor(value)
    if not tensor.is_floating_point():
        tensor = tensor.to(dtype=torch.float32)
    if tensor.ndim == 0 or tensor.shape[-1] >= target_dim:
        return tensor
    pad_shape = (*tensor.shape[:-1], target_dim - tensor.shape[-1])
    pad_value = _STAT_PAD_VALUES.get(stat_name, 0.0)
    padding = torch.full(pad_shape, pad_value, dtype=tensor.dtype, device=tensor.device)
    return torch.cat([tensor, padding], dim=-1)
 def _pad_feature_stats(
    stats: dict[str, dict[str, Any]],
    feature_key: str,
    target_dim: int,
 ) -> None:
    if feature_key not in stats:
        return
    stats[feature_key] = {
        stat_name: _pad_stat_value(stat_value, target_dim, stat_name)
        for stat_name, stat_value in stats[feature_key].items()
    }
 def _pad_evo1_stats(
    config: Evo1Config,
    stats: dict[str, dict[str, Any]] | None,
 ) -> dict[str, dict[str, Any]] | None:
    if stats is None:
        return None
    padded_stats = deepcopy(stats)
    # Added dimensions represent zero-padding inside EVO1. These neutral stats keep
    # padded observations at normalized zero and only provide shape compatibility.
    _pad_feature_stats(padded_stats, OBS_STATE, config.max_state_dim)
    _pad_feature_stats(padded_stats, ACTION, config.max_action_dim)
    return padded_stats
 def reconcile_evo1_processors(
    config: Evo1Config,
    preprocessor: PolicyProcessorPipeline,
    postprocessor: PolicyProcessorPipeline,
 ) -> tuple[PolicyProcessorPipeline, PolicyProcessorPipeline]:
    """Reconcile checkpoint-loaded pipelines with the current EVO1 config.
    Two things cannot be restored from a serialized pipeline alone: the EVO1 batch converter
    (converters are plain functions and are never serialized), and eval-time CLI overrides of the
    action postprocessing flags (`postprocess_action_dim`, `binarize_gripper`, `gripper_*`). This
    restores the converter and rebuilds the action step from the current config so those overrides
    take effect.
    """
    # Pipelines reloaded from a checkpoint come back with the default batch converter, which drops
    # non-observation extras (embodiment_id, state_mask, custom task fields) needed by EVO1.
    preprocessor.to_transition = evo1_batch_to_transition
    action_step = Evo1ActionProcessorStep(
        action_dim=_evo1_action_dim(config),
        binarize_gripper=config.binarize_gripper,
        gripper_index=config.gripper_index,
        gripper_threshold=config.gripper_threshold,
        gripper_below_threshold_value=config.gripper_below_threshold_value,
        gripper_above_threshold_value=config.gripper_above_threshold_value,
    )
    steps = list(postprocessor.steps)
    action_step_idx = next(
        (idx for idx, step in enumerate(steps) if isinstance(step, Evo1ActionProcessorStep)), None
    )
    if action_step_idx is None:
        insert_idx = next(
            (idx + 1 for idx, step in enumerate(steps) if isinstance(step, UnnormalizerProcessorStep)),
            0,
        )
        steps.insert(insert_idx, action_step)
    else:
        steps[action_step_idx] = action_step
    postprocessor.steps = steps
    return preprocessor, postprocessor
 def make_evo1_pre_post_processors(
    config: Evo1Config,
    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
 ) -> tuple[
    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
    PolicyProcessorPipeline[PolicyAction, PolicyAction],
 ]:
    normalization_features = _evo1_normalization_features(config)
    action_features = _evo1_action_features(config)
    normalization_stats = _pad_evo1_stats(config, dataset_stats)
    input_steps = [
        RenameObservationsProcessorStep(rename_map={}),
        AddBatchDimensionProcessorStep(),
        Evo1PadStateProcessorStep(max_state_dim=config.max_state_dim),
        Evo1PadActionProcessorStep(max_action_dim=config.max_action_dim),
        NormalizerProcessorStep(
            features=normalization_features,
            norm_map=config.normalization_mapping,
            stats=normalization_stats,
        ),
        DeviceProcessorStep(device=config.device),
    ]
    output_steps = [
        UnnormalizerProcessorStep(
            features=action_features,
            norm_map=config.normalization_mapping,
            stats=normalization_stats,
        ),
        Evo1ActionProcessorStep(
            action_dim=_evo1_action_dim(config),
            binarize_gripper=config.binarize_gripper,
            gripper_index=config.gripper_index,
            gripper_threshold=config.gripper_threshold,
            gripper_below_threshold_value=config.gripper_below_threshold_value,
            gripper_above_threshold_value=config.gripper_above_threshold_value,
        ),
        # float32 so downstream numpy conversion works even when the policy computes in bf16.
        DeviceProcessorStep(device="cpu", float_dtype="float32"),
    ]
    return (
        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
            steps=input_steps,
            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
            to_transition=evo1_batch_to_transition,
        ),
        PolicyProcessorPipeline[PolicyAction, PolicyAction](
            steps=output_steps,
            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
            to_transition=policy_action_to_transition,
            to_output=transition_to_policy_action,
        ),
    )
@@ -47,11 +47,9 @@ from lerobot.utils.feature_utils import dataset_to_policy_features
 from .act.configuration_act import ACTConfig
 from .diffusion.configuration_diffusion import DiffusionConfig
 from .eo1.configuration_eo1 import EO1Config
 from .evo1.configuration_evo1 import Evo1Config
 from .fastwam.configuration_fastwam import FastWAMConfig
 from .gaussian_actor.configuration_gaussian_actor import GaussianActorConfig
 from .groot.configuration_groot import GrootConfig
 from .lingbot_va.configuration_lingbot_va import LingBotVAConfig
 from .molmoact2.configuration_molmoact2 import MolmoAct2Config
 from .multi_task_dit.configuration_multi_task_dit import MultiTaskDiTConfig
 from .pi0.configuration_pi0 import PI0Config
@@ -94,7 +92,7 @@ def get_policy_class(name: str) -> type[PreTrainedPolicy]:
    Args:
        name: The name of the policy. Supported names are "tdmpc", "diffusion", "act",
            "multi_task_dit", "vqbet", "pi0", "pi05", "gaussian_actor", "smolvla", "wall_x",
-            "molmoact2", "eo1", "evo1".
+            "molmoact2".
    Returns:
        The policy class corresponding to the given name.
@@ -165,18 +163,10 @@ def get_policy_class(name: str) -> type[PreTrainedPolicy]:
        from .vla_jepa.modeling_vla_jepa import VLAJEPAPolicy
        return VLAJEPAPolicy
    elif name == "lingbot_va":
        from .lingbot_va.modeling_lingbot_va import LingBotVAPolicy
        return LingBotVAPolicy
    elif name == "fastwam":
        from .fastwam.modeling_fastwam import FastWAMPolicy
        return FastWAMPolicy
    elif name == "evo1":
        from .evo1.modeling_evo1 import Evo1Policy
        return Evo1Policy
    else:
        try:
            return _get_policy_cls_from_policy_name(name=name)
@@ -194,7 +184,7 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
    Args:
        policy_type: The type of the policy. Supported types include "tdmpc",
                     "multi_task_dit", "diffusion", "act", "vqbet", "pi0", "pi05", "gaussian_actor",
-                     "smolvla", "wall_x", "molmoact2", "eo1", "evo1".
+                     "smolvla", "wall_x", "molmoact2".
        **kwargs: Keyword arguments to be passed to the configuration class constructor.
    Returns:
@@ -233,12 +223,8 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
        return MolmoAct2Config(**kwargs)
    elif policy_type == "vla_jepa":
        return VLAJEPAConfig(**kwargs)
    elif policy_type == "lingbot_va":
        return LingBotVAConfig(**kwargs)
    elif policy_type == "fastwam":
        return FastWAMConfig(**kwargs)
    elif policy_type == "evo1":
        return Evo1Config(**kwargs)
    else:
        try:
            config_cls = PreTrainedConfig.get_choice_class(policy_type)
@@ -302,23 +288,26 @@ def make_pre_post_processors(
            policy configuration type.
    """
    if pretrained_path:
        # TODO(Steven): Temporary patch, implement correctly the processors for Gr00t
        if isinstance(policy_cfg, GrootConfig):
-            from .groot.processor_groot import make_groot_pre_post_processors_from_pretrained
+            # GROOT handles normalization in groot_pack_inputs_v3 step
            # Need to override both stats AND normalize_min_max since saved config might be empty
            preprocessor_overrides = {}
            postprocessor_overrides = {}
            preprocessor_overrides["groot_pack_inputs_v3"] = {
                "stats": kwargs.get("dataset_stats"),
                "normalize_min_max": True,
            }
-            return make_groot_pre_post_processors_from_pretrained(
+            # Also ensure postprocessing slices to env action dim and unnormalizes with dataset stats
-                config=policy_cfg,
+            env_action_dim = policy_cfg.output_features[ACTION].shape[0]
-                pretrained_path=pretrained_path,
+            postprocessor_overrides["groot_action_unpack_unnormalize_v1"] = {
-                dataset_stats=kwargs.get("dataset_stats"),
+                "stats": kwargs.get("dataset_stats"),
-                dataset_meta=kwargs.get("dataset_meta"),
+                "normalize_min_max": True,
-                preprocessor_overrides=kwargs.get("preprocessor_overrides"),
+                "env_action_dim": env_action_dim,
-                postprocessor_overrides=kwargs.get("postprocessor_overrides"),
+            }
-                preprocessor_config_filename=kwargs.get(
+            kwargs["preprocessor_overrides"] = preprocessor_overrides
-                    "preprocessor_config_filename", f"{POLICY_PREPROCESSOR_DEFAULT_NAME}.json"
+            kwargs["postprocessor_overrides"] = postprocessor_overrides
                ),
                postprocessor_config_filename=kwargs.get(
                    "postprocessor_config_filename", f"{POLICY_POSTPROCESSOR_DEFAULT_NAME}.json"
                ),
            )
        preprocessor = PolicyProcessorPipeline.from_pretrained(
            pretrained_model_name_or_path=pretrained_path,
@@ -341,14 +330,6 @@ def make_pre_post_processors(
            revision=pretrained_revision,
        )
        _reconnect_relative_absolute_steps(preprocessor, postprocessor)
        if isinstance(policy_cfg, Evo1Config):
            from .evo1.processor_evo1 import reconcile_evo1_processors
            preprocessor, postprocessor = reconcile_evo1_processors(
                policy_cfg,
                preprocessor,
                postprocessor,
            )
        return preprocessor, postprocessor
    # Create a new processor based on policy type
@@ -432,7 +413,6 @@ def make_pre_post_processors(
        processors = make_groot_pre_post_processors(
            config=policy_cfg,
            dataset_stats=kwargs.get("dataset_stats"),
            dataset_meta=kwargs.get("dataset_meta"),
        )
    elif isinstance(policy_cfg, XVLAConfig):
@@ -460,13 +440,6 @@ def make_pre_post_processors(
            config=policy_cfg,
            dataset_stats=kwargs.get("dataset_stats"),
        )
    elif isinstance(policy_cfg, Evo1Config):
        from .evo1.processor_evo1 import make_evo1_pre_post_processors
        processors = make_evo1_pre_post_processors(
            config=policy_cfg,
            dataset_stats=kwargs.get("dataset_stats"),
        )
    elif isinstance(policy_cfg, MolmoAct2Config):
        from .molmoact2.processor_molmoact2 import make_molmoact2_pre_post_processors
@@ -485,14 +458,6 @@ def make_pre_post_processors(
            dataset_stats=kwargs.get("dataset_stats"),
        )
    elif isinstance(policy_cfg, LingBotVAConfig):
        from .lingbot_va.processor_lingbot_va import make_lingbot_va_pre_post_processors
        processors = make_lingbot_va_pre_post_processors(
            config=policy_cfg,
            dataset_stats=kwargs.get("dataset_stats"),
        )
    elif isinstance(policy_cfg, FastWAMConfig):
        from .fastwam.processor_fastwam import make_fastwam_pre_post_processors
@@ -590,7 +555,6 @@ def make_policy(
        set_dataset_feature_metadata = getattr(cfg, "set_dataset_feature_metadata", None)
        if callable(set_dataset_feature_metadata):
            set_dataset_feature_metadata(ds_meta.features)
        cfg._runtime_dataset_meta = ds_meta
    kwargs["config"] = cfg
@@ -0,0 +1,54 @@
 # SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
 # http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import torch
 import torch.nn as nn
 def swish(x):
    return x * torch.sigmoid(x)
 class SinusoidalPositionalEncoding(nn.Module):
    """
    Produces a sinusoidal encoding of shape (B, T, w)
    given timesteps of shape (B, T).
    """
    def __init__(self, embedding_dim):
        super().__init__()
        self.embedding_dim = embedding_dim
    def forward(self, timesteps):
        # timesteps: shape (B, T)
        # We'll compute sin/cos frequencies across dim T
        timesteps = timesteps.float()  # ensure float
        b, t = timesteps.shape
        device = timesteps.device
        half_dim = self.embedding_dim // 2
        # typical log space frequencies for sinusoidal encoding
        exponent = -torch.arange(half_dim, dtype=torch.float, device=device) * (
            torch.log(torch.tensor(10000.0)) / half_dim
        )
        # Expand timesteps to (B, T, 1) then multiply
        freqs = timesteps.unsqueeze(-1) * exponent.exp()  # (B, T, half_dim)
        sin = torch.sin(freqs)
        cos = torch.cos(freqs)
        enc = torch.cat([sin, cos], dim=-1)  # (B, T, w)
        return enc
@@ -1,12 +1,11 @@
-#!/usr/bin/env python
+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
-
+# SPDX-License-Identifier: Apache-2.0
 # Copyright 2025 NVIDIA Corporation and The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
-#     http://www.apache.org/licenses/LICENSE-2.0
+# http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
@@ -15,7 +14,6 @@
 # limitations under the License.
 import logging
 from typing import TYPE_CHECKING
 import torch
@@ -44,9 +42,6 @@ else:
    Timesteps = None
 logger = logging.getLogger(__name__)
 class TimestepEncoder(nn.Module):
    def __init__(self, embedding_dim, compute_dtype=torch.float32):
        require_package("diffusers", extra="groot")
@@ -186,7 +181,8 @@ class BasicTransformerBlock(nn.Module):
        attn_output = self.attn1(
            norm_hidden_states,
            encoder_hidden_states=encoder_hidden_states,
-            attention_mask=encoder_attention_mask if encoder_hidden_states is not None else attention_mask,
+            attention_mask=attention_mask,
            # encoder_attention_mask=encoder_attention_mask,
        )
        if self.final_dropout:
            attn_output = self.final_dropout(attn_output)
@@ -270,8 +266,8 @@ class DiT(ModelMixin, ConfigMixin):
        self.norm_out = nn.LayerNorm(self.inner_dim, elementwise_affine=False, eps=1e-6)
        self.proj_out_1 = nn.Linear(self.inner_dim, 2 * self.inner_dim)
        self.proj_out_2 = nn.Linear(self.inner_dim, self.config.output_dim)
-        logger.debug(
+        print(
-            "Total number of DiT parameters: %d",
+            "Total number of DiT parameters: ",
            sum(p.numel() for p in self.parameters() if p.requires_grad),
        )
@@ -322,71 +318,6 @@ class DiT(ModelMixin, ConfigMixin):
            return self.proj_out_2(hidden_states)
 class AlternateVLDiT(DiT):
    """N1.7 DiT variant that alternates cross-attention over image and text tokens."""
    def __init__(self, *args, attend_text_every_n_blocks: int = 2, **kwargs):
        super().__init__(*args, **kwargs)
        self.attend_text_every_n_blocks = attend_text_every_n_blocks
    def forward(
        self,
        hidden_states: torch.Tensor,
        encoder_hidden_states: torch.Tensor,
        timestep: torch.LongTensor | None = None,
        encoder_attention_mask: torch.Tensor | None = None,
        return_all_hidden_states: bool = False,
        image_mask: torch.Tensor | None = None,
        backbone_attention_mask: torch.Tensor | None = None,
    ):
        if image_mask is None:
            raise ValueError("image_mask is required for AlternateVLDiT.")
        if backbone_attention_mask is None:
            raise ValueError("backbone_attention_mask is required for AlternateVLDiT.")
        temb = self.timestep_encoder(timestep)
        hidden_states = hidden_states.contiguous()
        encoder_hidden_states = encoder_hidden_states.contiguous()
        image_attention_mask = image_mask & backbone_attention_mask
        non_image_attention_mask = (~image_mask) & backbone_attention_mask
        all_hidden_states = [hidden_states]
        if not self.config.interleave_self_attention:
            raise ValueError("AlternateVLDiT requires interleave_self_attention=True.")
        for idx, block in enumerate(self.transformer_blocks):
            if idx % 2 == 1:
                hidden_states = block(
                    hidden_states,
                    attention_mask=None,
                    encoder_hidden_states=None,
                    encoder_attention_mask=None,
                    temb=temb,
                )
            else:
                curr_encoder_attention_mask = (
                    non_image_attention_mask
                    if idx % (2 * self.attend_text_every_n_blocks) == 0
                    else image_attention_mask
                )
                hidden_states = block(
                    hidden_states,
                    attention_mask=None,
                    encoder_hidden_states=encoder_hidden_states,
                    encoder_attention_mask=curr_encoder_attention_mask,
                    temb=temb,
                )
            all_hidden_states.append(hidden_states)
        conditioning = temb
        shift, scale = self.proj_out_1(F.silu(conditioning)).chunk(2, dim=1)
        hidden_states = self.norm_out(hidden_states) * (1 + scale[:, None]) + shift[:, None]
        if return_all_hidden_states:
            return self.proj_out_2(hidden_states), all_hidden_states
        return self.proj_out_2(hidden_states)
 class SelfAttentionTransformer(ModelMixin, ConfigMixin):
    _supports_gradient_checkpointing = True
@@ -431,8 +362,8 @@ class SelfAttentionTransformer(ModelMixin, ConfigMixin):
                for _ in range(self.config.num_layers)
            ]
        )
-        logger.debug(
+        print(
-            "Total number of SelfAttentionTransformer parameters: %d",
+            "Total number of SelfAttentionTransformer parameters: ",
            sum(p.numel() for p in self.parameters() if p.requires_grad),
        )
@@ -0,0 +1,408 @@
 # SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
 # http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from dataclasses import field
 from typing import TYPE_CHECKING
 import torch
 import torch.nn.functional as F  # noqa: N812
 from torch import nn
 from torch.distributions import Beta
 from lerobot.utils.import_utils import _transformers_available
 # Conditional import for type checking and lazy loading
 if TYPE_CHECKING or _transformers_available:
    from transformers import PretrainedConfig
    from transformers.feature_extraction_utils import BatchFeature
 else:
    PretrainedConfig = object
    BatchFeature = None
 from .action_encoder import (
    SinusoidalPositionalEncoding,
    swish,
 )
 from .cross_attention_dit import DiT, SelfAttentionTransformer
 class CategorySpecificLinear(nn.Module):
    def __init__(self, num_categories, input_dim, hidden_dim):
        super().__init__()
        self.num_categories = num_categories
        # For each category, we have separate weights and biases.
        self.W = nn.Parameter(0.02 * torch.randn(num_categories, input_dim, hidden_dim))
        self.b = nn.Parameter(torch.zeros(num_categories, hidden_dim))
    def forward(self, x, cat_ids):
        selected_w = self.W[cat_ids]
        selected_b = self.b[cat_ids]
        return torch.bmm(x, selected_w) + selected_b.unsqueeze(1)
 class CategorySpecificMLP(nn.Module):
    def __init__(self, num_categories, input_dim, hidden_dim, output_dim):
        super().__init__()
        self.num_categories = num_categories
        self.layer1 = CategorySpecificLinear(num_categories, input_dim, hidden_dim)
        self.layer2 = CategorySpecificLinear(num_categories, hidden_dim, output_dim)
    def forward(self, x, cat_ids):
        hidden = F.relu(self.layer1(x, cat_ids))
        return self.layer2(hidden, cat_ids)
 class MultiEmbodimentActionEncoder(nn.Module):
    def __init__(self, action_dim, hidden_size, num_embodiments):
        super().__init__()
        self.hidden_size = hidden_size
        self.num_embodiments = num_embodiments
        # W1: R^{w x d}, W2: R^{w x 2w}, W3: R^{w x w}
        self.W1 = CategorySpecificLinear(num_embodiments, action_dim, hidden_size)  # (d -> w)
        self.W2 = CategorySpecificLinear(num_embodiments, 2 * hidden_size, hidden_size)  # (2w -> w)
        self.W3 = CategorySpecificLinear(num_embodiments, hidden_size, hidden_size)  # (w -> w)
        self.pos_encoding = SinusoidalPositionalEncoding(hidden_size)
    def forward(self, actions, timesteps, cat_ids):
        """
        actions:   shape (B, T, action_dim)
        timesteps: shape (B,)  -- a single scalar per batch item
        cat_ids:   shape (B,)
        returns:   shape (B, T, hidden_size)
        """
        b, t, _ = actions.shape
        # 1) Expand each batch's single scalar time 'tau' across all T steps
        #    so that shape => (B, T)
        #    e.g. if timesteps is (B,), replicate across T
        if timesteps.dim() == 1 and timesteps.shape[0] == b:
            # shape (B,) => (B,T)
            timesteps = timesteps.unsqueeze(1).expand(-1, t)
        else:
            raise ValueError("Expected `timesteps` to have shape (B,) so we can replicate across T.")
        # 2) Standard action MLP step for shape => (B, T, w)
        a_emb = self.W1(actions, cat_ids)
        # 3) Get the sinusoidal encoding (B, T, w)
        tau_emb = self.pos_encoding(timesteps).to(dtype=a_emb.dtype)
        # 4) Concat along last dim => (B, T, 2w), then W2 => (B, T, w), swish
        x = torch.cat([a_emb, tau_emb], dim=-1)
        x = swish(self.W2(x, cat_ids))
        # 5) Finally W3 => (B, T, w)
        x = self.W3(x, cat_ids)
        return x
 class FlowmatchingActionHeadConfig(PretrainedConfig):
    """NOTE: N1.5 uses XEmbFlowmatchingPolicyHeadConfig as action head"""
    add_pos_embed: bool = field(default=True, metadata={"help": "Whether to add positional embedding"})
    model_dtype: str = field(default="float32", metadata={"help": "Model data type."})
    diffusion_model_cfg: dict = field(default=None, metadata={"help": "Diffusion model configuration."})
    input_embedding_dim: int = field(default=1536, metadata={"help": "Input embedding channel dimension."})
    backbone_embedding_dim: int = field(
        default=1536, metadata={"help": "Backbone embedding channel dimension."}
    )
    hidden_size: int = field(default=1024, metadata={"help": "Input embedding dimension."})
    max_seq_len: int = field(default=1024, metadata={"help": "Maximum Sequence Length"})
    action_dim: int = field(default=None, metadata={"help": "Action dimension."})
    action_horizon: int = field(default=None, metadata={"help": "Action horizon."})
    noise_beta_alpha: float = field(default=1.5, metadata={"help": ""})
    noise_beta_beta: float = field(default=1.0, metadata={"help": ""})
    noise_s: float = field(default=0.999, metadata={"help": "Flow matching noise Beta distribution s."})
    num_timestep_buckets: int = field(
        default=1000, metadata={"help": "Number of timestep discretization buckets."}
    )
    num_inference_timesteps: int = field(
        default=None,
        metadata={"help": "Number of inference steps for noise diffusion."},
    )
    max_num_embodiments: int = field(default=32, metadata={"help": "Number of embodiments."})
    tune_projector: bool = field(default=True, metadata={"help": "Whether to tune the projector."})
    tune_diffusion_model: bool = field(
        default=True, metadata={"help": "Whether to tune the diffusion model."}
    )
    load_pretrained_det_decode_layer_path: str = field(
        default=None, metadata={"help": "Path to pretrained detection model."}
    )
    detection_coeff: float = field(default=1.0, metadata={"help": "Detection coefficient."})
    freeze_decode_layer: bool = field(default=False)
    expand_batch: int = field(default=None)
    use_vlln: bool = field(default=True)
    vl_self_attention_cfg: dict = field(default=None)
    num_target_vision_tokens: int = field(default=32, metadata={"help": "Number of target vision tokens."})
    def __init__(self, **kwargs):
        super().__init__(**kwargs)
        for key, value in kwargs.items():
            setattr(self, key, value)
 class FlowmatchingActionHead(nn.Module):
    config_class = FlowmatchingActionHeadConfig
    supports_gradient_checkpointing = True
    def __init__(
        self,
        config: FlowmatchingActionHeadConfig,
    ):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.input_embedding_dim = config.input_embedding_dim
        self.model = DiT(**config.diffusion_model_cfg)
        self.action_dim = config.action_dim
        self.action_horizon = config.action_horizon
        self.num_inference_timesteps = config.num_inference_timesteps
        self.state_encoder = CategorySpecificMLP(
            num_categories=config.max_num_embodiments,
            input_dim=config.max_state_dim,
            hidden_dim=self.hidden_size,
            output_dim=self.input_embedding_dim,
        )
        self.action_encoder = MultiEmbodimentActionEncoder(
            action_dim=config.action_dim,
            hidden_size=self.input_embedding_dim,
            num_embodiments=config.max_num_embodiments,
        )
        self.action_decoder = CategorySpecificMLP(
            num_categories=config.max_num_embodiments,
            input_dim=self.hidden_size,
            hidden_dim=self.hidden_size,
            output_dim=self.action_dim,
        )
        self.future_tokens = nn.Embedding(config.num_target_vision_tokens, self.input_embedding_dim)
        nn.init.normal_(self.future_tokens.weight, mean=0.0, std=0.02)
        self.vlln = nn.LayerNorm(config.backbone_embedding_dim) if config.use_vlln else nn.Identity()
        self.vl_self_attention = (
            SelfAttentionTransformer(**config.vl_self_attention_cfg) if config.use_vlln else nn.Identity()
        )
        if config.add_pos_embed:
            self.position_embedding = nn.Embedding(config.max_seq_len, self.input_embedding_dim)
            nn.init.normal_(self.position_embedding.weight, mean=0.0, std=0.02)
        self._noise_beta_alpha = config.noise_beta_alpha
        self._noise_beta_beta = config.noise_beta_beta
        self._beta_dist = None
        self.num_timestep_buckets = config.num_timestep_buckets
        self.config = config
        self.set_trainable_parameters(config.tune_projector, config.tune_diffusion_model)
    def set_trainable_parameters(self, tune_projector: bool, tune_diffusion_model: bool):
        self.tune_projector = tune_projector
        self.tune_diffusion_model = tune_diffusion_model
        for p in self.parameters():
            p.requires_grad = True
        if not tune_projector:
            self.state_encoder.requires_grad_(False)
            self.action_encoder.requires_grad_(False)
            self.action_decoder.requires_grad_(False)
            if self.config.add_pos_embed:
                self.position_embedding.requires_grad_(False)
        if not tune_diffusion_model:
            self.model.requires_grad_(False)
        print(f"Tune action head projector: {self.tune_projector}")
        print(f"Tune action head diffusion model: {self.tune_diffusion_model}")
        # Check if any parameters are still trainable. If not, print a warning.
        if not tune_projector and not tune_diffusion_model:
            for name, p in self.named_parameters():
                if p.requires_grad:
                    print(f"Action head trainable parameter: {name}")
        if not any(p.requires_grad for p in self.parameters()):
            print("Warning: No action head trainable parameters found.")
    def set_frozen_modules_to_eval_mode(self):
        """
        Huggingface will call model.train() at each training_step. To ensure
        the expected behaviors for modules like dropout, batchnorm, etc., we
        need to call model.eval() for the frozen modules.
        """
        if self.training:
            if not self.tune_projector:
                self.state_encoder.eval()
                self.action_encoder.eval()
                self.action_decoder.eval()
                if self.config.add_pos_embed:
                    self.position_embedding.eval()
            if not self.tune_diffusion_model:
                self.model.eval()
    def sample_time(self, batch_size, device, dtype):
        if self._beta_dist is None:
            self._beta_dist = Beta(self._noise_beta_alpha, self._noise_beta_beta, validate_args=False)
        sample = self._beta_dist.sample([batch_size]).to(device, dtype=dtype)
        return (self.config.noise_s - sample) / self.config.noise_s
    def prepare_input(self, batch: dict) -> BatchFeature:
        return BatchFeature(data=batch)
    def process_backbone_output(self, backbone_output: BatchFeature) -> BatchFeature:
        backbone_features = backbone_output["backbone_features"]
        backbone_features = self.vlln(backbone_features)
        backbone_features = self.vl_self_attention(backbone_features)
        backbone_output["backbone_features"] = backbone_features
        return backbone_output
    def forward(self, backbone_output: BatchFeature, action_input: BatchFeature) -> BatchFeature:
        # Set frozen modules to eval
        self.set_frozen_modules_to_eval_mode()
        backbone_output = self.process_backbone_output(backbone_output)
        if self.config.expand_batch is not None:
            for k, v in backbone_output.items():
                ndim = len(v.shape)
                factors = [self.config.expand_batch]
                while len(factors) < ndim:
                    factors.append(1)
                factors = tuple(factors)
                expanded = v.repeat(*factors)
                backbone_output[k] = expanded
            for k, v in action_input.items():
                ndim = len(v.shape)
                factors = [self.config.expand_batch]
                while len(factors) < ndim:
                    factors.append(1)
                factors = tuple(factors)
                expanded = v.repeat(*factors)
                action_input[k] = expanded
        # Get vision and language embeddings.
        vl_embs = backbone_output.backbone_features
        device = vl_embs.device
        # Get embodiment ID.
        embodiment_id = action_input.embodiment_id
        # Embed state.
        state_features = self.state_encoder(action_input.state, embodiment_id)
        # Embed noised action trajectory.
        actions = action_input.action
        noise = torch.randn(actions.shape, device=actions.device, dtype=actions.dtype)
        t = self.sample_time(actions.shape[0], device=actions.device, dtype=actions.dtype)
        t = t[:, None, None]  # shape (B,1,1) for broadcast
        noisy_trajectory = (1 - t) * noise + t * actions
        velocity = actions - noise
        # Convert (continuous) t -> discrete if needed
        t_discretized = (t[:, 0, 0] * self.num_timestep_buckets).long()
        action_features = self.action_encoder(noisy_trajectory, t_discretized, embodiment_id)
        # Maybe add position embedding.
        if self.config.add_pos_embed:
            pos_ids = torch.arange(action_features.shape[1], dtype=torch.long, device=device)
            pos_embs = self.position_embedding(pos_ids).unsqueeze(0)
            action_features = action_features + pos_embs
        # Join vision, language, state and action embedding along sequence dimension.
        future_tokens = self.future_tokens.weight.unsqueeze(0).expand(vl_embs.shape[0], -1, -1)
        sa_embs = torch.cat((state_features, future_tokens, action_features), dim=1)
        vl_attn_mask = backbone_output.backbone_attention_mask
        model_output = self.model(
            hidden_states=sa_embs,
            encoder_hidden_states=vl_embs,
            encoder_attention_mask=vl_attn_mask,
            timestep=t_discretized,
            return_all_hidden_states=False,  # NOTE (YL): not using flare now
        )
        pred = self.action_decoder(model_output, embodiment_id)
        pred_actions = pred[:, -actions.shape[1] :]
        # Slice out only the action portion of pred and target.
        action_mask = action_input.action_mask
        loss = F.mse_loss(pred_actions, velocity, reduction="none") * action_mask
        loss = loss.sum() / action_mask.sum()
        output_dict = {
            "loss": loss,
        }
        return BatchFeature(data=output_dict)
    @torch.no_grad()
    def get_action(self, backbone_output: BatchFeature, action_input: BatchFeature) -> BatchFeature:
        backbone_output = self.process_backbone_output(backbone_output)
        # Get vision and language embeddings.
        vl_embs = backbone_output.backbone_features
        embodiment_id = action_input.embodiment_id
        # Embed state.
        state_features = self.state_encoder(action_input.state, embodiment_id)
        # Set initial actions as the sampled noise.
        batch_size = vl_embs.shape[0]
        device = vl_embs.device
        actions = torch.randn(
            size=(batch_size, self.config.action_horizon, self.config.action_dim),
            dtype=vl_embs.dtype,
            device=device,
        )
        num_steps = self.num_inference_timesteps
        dt = 1.0 / num_steps
        # Run denoising steps.
        for t in range(num_steps):
            t_cont = t / float(num_steps)  # e.g. goes 0, 1/N, 2/N, ...
            t_discretized = int(t_cont * self.num_timestep_buckets)
            # Embed noised action trajectory.
            timesteps_tensor = torch.full(size=(batch_size,), fill_value=t_discretized, device=device)
            action_features = self.action_encoder(actions, timesteps_tensor, embodiment_id)
            # Maybe add position embedding.
            if self.config.add_pos_embed:
                pos_ids = torch.arange(action_features.shape[1], dtype=torch.long, device=device)
                pos_embs = self.position_embedding(pos_ids).unsqueeze(0)
                action_features = action_features + pos_embs
            # Join vision, language, state and action embedding along sequence dimension.
            future_tokens = self.future_tokens.weight.unsqueeze(0).expand(vl_embs.shape[0], -1, -1)
            sa_embs = torch.cat((state_features, future_tokens, action_features), dim=1)
            # Run model forward.
            model_output = self.model(
                hidden_states=sa_embs,
                encoder_hidden_states=vl_embs,
                timestep=timesteps_tensor,
            )
            pred = self.action_decoder(model_output, embodiment_id)
            pred_velocity = pred[:, -self.action_horizon :]
            # Update actions using euler integration.
            actions = actions + dt * pred_velocity
        return BatchFeature(data={"action_pred": actions})
    @property
    def device(self):
        return next(iter(self.parameters())).device
    @property
    def dtype(self):
        return next(iter(self.parameters())).dtype
@@ -14,229 +14,12 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import logging
 import math
 from dataclasses import dataclass, field
 from pathlib import Path
 from lerobot.configs import FeatureType, NormalizationMode, PolicyFeature, PreTrainedConfig
-from lerobot.optim import AdamWConfig, DiffuserSchedulerConfig
+from lerobot.optim import AdamWConfig, CosineDecayWithWarmupSchedulerConfig
 from lerobot.utils.constants import ACTION, OBS_STATE
 from .utils import read_json
 logger = logging.getLogger(__name__)
 GROOT_N1_7 = "n1.7"
 # Legacy GR00T N1.5 identifier. N1.5 is NOT a supported model_version (it is
 # intentionally absent from _GROOT_MODEL_VERSION_ALIASES so normalize_groot_model_version
 # still rejects it). It is retained only so that infer_groot_model_version can recognise
 # an N1.5 base path/checkpoint and the N1.7 config/loader can reject the mismatch.
 GROOT_N1_5 = "n1.5"
 # Canonical guidance appended to every error raised when an N1.5 checkpoint, config,
 # or processor pipeline is detected. Keep this message in sync with docs/source/groot.mdx.
 GROOT_N1_5_REMOVAL_GUIDANCE = (
    "GR00T N1.5 support was removed from LeRobot. "
    "To keep using an N1.5 checkpoint, pin the last release that supports it: "
    "`pip install 'lerobot==0.5.1'`. To use the current release, migrate to GR00T N1.7 "
    "(model_version='n1.7', base model nvidia/GR00T-N1.7-3B)."
 )
 GROOT_N1_7_BASE_MODEL = "nvidia/GR00T-N1.7-3B"
 GROOT_N1_7_BACKBONE_MODEL = "nvidia/Cosmos-Reason2-2B"
 # Default GR00T N1.7 training resolution. Fallback if processor_config lacks sizing. Prevents mismatched
 # full-res patchification by forcing a resize. Mirrored by GR00T_N1_7_DEFAULTS in groot_n1_7.py.
 N1_7_DEFAULT_IMAGE_TARGET_SIZE = (256, 256)
 N1_7_DEFAULT_IMAGE_CROP_SIZE = (230, 230)
 GROOT_ACTION_DECODE_TRANSFORM_LIBERO = "libero"
 # Sentinel meaning "the user did not pick an action decode transform": __post_init__ resolves it
 # to the embodiment default ('libero' for 'libero_sim', otherwise None). It is distinct from an
 # explicit 'none' (resolved to None) so an opt-out survives a draccus save/load round-trip.
 GROOT_ACTION_DECODE_TRANSFORM_AUTO = "auto"
 _GROOT_MODEL_VERSION_ALIASES = {
    "n1.7": GROOT_N1_7,
    "n1_7": GROOT_N1_7,
    "n1d7": GROOT_N1_7,
    "n17": GROOT_N1_7,
    "1.7": GROOT_N1_7,
 }
 # Legacy N1.5 spellings, kept ONLY so they can be detected and rejected with
 # GROOT_N1_5_REMOVAL_GUIDANCE (see GROOT_N1_5 above). Never map these to a supported version.
 _GROOT_N1_5_VERSION_ALIASES = {"n1.5", "n1_5", "n1d5", "n15", "1.5"}
 _GROOT_ACTION_DECODE_TRANSFORM_ALIASES = {
    GROOT_ACTION_DECODE_TRANSFORM_AUTO: GROOT_ACTION_DECODE_TRANSFORM_AUTO,
    "none": None,
    "": None,
    GROOT_ACTION_DECODE_TRANSFORM_LIBERO: GROOT_ACTION_DECODE_TRANSFORM_LIBERO,
 }
 def normalize_groot_model_version(model_version: str) -> str:
    normalized = _GROOT_MODEL_VERSION_ALIASES.get(model_version.lower())
    if normalized is None:
        supported = GROOT_N1_7
        message = f"Unsupported GR00T model_version '{model_version}'. Supported versions: {supported}."
        if model_version.lower() in _GROOT_N1_5_VERSION_ALIASES:
            message = f"{message} {GROOT_N1_5_REMOVAL_GUIDANCE}"
        raise ValueError(message)
    return normalized
 def normalize_groot_action_decode_transform(transform: str | None) -> str | None:
    if transform is None:
        return None
    normalized = _GROOT_ACTION_DECODE_TRANSFORM_ALIASES.get(transform.lower())
    if normalized is None and transform.lower() not in _GROOT_ACTION_DECODE_TRANSFORM_ALIASES:
        supported = ", ".join(
            sorted(key for key, value in _GROOT_ACTION_DECODE_TRANSFORM_ALIASES.items() if value is not None)
        )
        raise ValueError(
            f"Unsupported GR00T N1.7 action decode transform '{transform}'. "
            f"Supported transforms: none, {supported}."
        )
    return normalized
 def infer_groot_model_version(model_path: str | None) -> str | None:
    if not model_path:
        return None
    model_path_lower = model_path.lower()
    if "gr00t-n1.7" in model_path_lower or "gr00t_n1.7" in model_path_lower:
        return GROOT_N1_7
    # Detect legacy N1.5 paths so the N1.7 config/loader can reject the mismatch.
    # N1.5 is unsupported, but it must still be recognised here to fail loudly
    # rather than silently treating an N1.5 checkpoint as N1.7.
    if "gr00t-n1.5" in model_path_lower or "gr00t_n1.5" in model_path_lower:
        return GROOT_N1_5
    config_version = _infer_groot_model_version_from_local_config(model_path)
    if config_version is not None:
        return config_version
    return None
 def is_raw_groot_n1_7_checkpoint(model_path: str | Path | None) -> bool:
    if model_path is None:
        return False
    path = Path(model_path).expanduser()
    if path.is_dir():
        config_path = path / "config.json"
    elif path.name == "config.json":
        config_path = path
    else:
        return False
    config = read_json(config_path)
    return "type" not in config and _infer_groot_model_version_from_config(config) == GROOT_N1_7
 def infer_groot_n1_7_embodiment_tag(model_path: str | Path | None) -> str | None:
    if model_path is None:
        return None
    processor_config_path = Path(model_path).expanduser() / "processor_config.json"
    processor_config = read_json(processor_config_path)
    modality_configs = processor_config.get("processor_kwargs", {}).get("modality_configs", {})
    if not isinstance(modality_configs, dict):
        return None
    if "libero_sim" in modality_configs:
        return "libero_sim"
    if len(modality_configs) == 1:
        return next(iter(modality_configs))
    return None
 def infer_groot_n1_7_action_horizon(
    model_path: str | Path | None, embodiment_tag: str | None = None
 ) -> int | None:
    if model_path is None:
        return None
    processor_config_path = Path(model_path).expanduser() / "processor_config.json"
    processor_config = read_json(processor_config_path)
    processor_kwargs = processor_config.get("processor_kwargs", {})
    if not isinstance(processor_kwargs, dict):
        return None
    modality_configs = processor_kwargs.get("modality_configs", {})
    if not isinstance(modality_configs, dict):
        return None
    if embodiment_tag is None:
        embodiment_tag = infer_groot_n1_7_embodiment_tag(model_path)
    if embodiment_tag is None:
        return None
    embodiment_config = modality_configs.get(embodiment_tag, {})
    if not isinstance(embodiment_config, dict):
        return None
    action_config = embodiment_config.get("action", {})
    if not isinstance(action_config, dict):
        return None
    delta_indices = action_config.get("delta_indices", [])
    if not isinstance(delta_indices, list):
        return None
    return len(delta_indices) or None
 def infer_groot_n1_7_action_execution_horizon(
    model_path: str | Path | None, embodiment_tag: str | None = None
 ) -> int | None:
    action_horizon = infer_groot_n1_7_action_horizon(model_path, embodiment_tag)
    if action_horizon is None:
        return None
    if embodiment_tag is None:
        embodiment_tag = infer_groot_n1_7_embodiment_tag(model_path)
    if embodiment_tag == "libero_sim":
        # NVIDIA's N1.7 LIBERO rollout wrapper replans after 8 of the 16 decoded
        # actions. Keeping that execution cadence avoids stale open-loop chunks.
        return min(action_horizon, 8)
    return action_horizon
 def _infer_groot_model_version_from_local_config(model_path: str) -> str | None:
    path = Path(model_path).expanduser()
    if path.is_dir():
        config_path = path / "config.json"
    elif path.name == "config.json":
        config_path = path
    else:
        return None
    return _infer_groot_model_version_from_config(read_json(config_path))
 def _infer_groot_model_version_from_config(config: dict) -> str | None:
    model_version = config.get("model_version")
    if isinstance(model_version, str):
        if model_version.lower() in _GROOT_N1_5_VERSION_ALIASES:
            return GROOT_N1_5
        try:
            return normalize_groot_model_version(model_version)
        except ValueError:
            return None
    candidates = [config.get("model_type"), *(config.get("architectures") or [])]
    for candidate in candidates:
        if not isinstance(candidate, str):
            continue
        normalized = candidate.lower().replace("-", "_")
        if normalized in {"gr00tn1d7", "gr00t_n1d7", "gr00t_n1_7"}:
            return GROOT_N1_7
        if normalized in {"gr00t_n1_5", "gr00tn1_5", "gr00t_n15", "gr00t_n1d5", "gr00tn1d5"}:
            return GROOT_N1_5
    if config.get("model_name") == GROOT_N1_7_BACKBONE_MODEL:
        return GROOT_N1_7
    # The Eagle VLM backbone is specific to pre-N1.7 GR00T checkpoints (N1.7 uses Cosmos/Qwen3-VL).
    backbone_cfg = config.get("backbone_cfg")
    if isinstance(backbone_cfg, dict) and "eagle_path" in backbone_cfg:
        return GROOT_N1_5
    return None
@PreTrainedConfig.register_subclass("groot")
@dataclass
@@ -245,44 +28,35 @@ class GrootConfig(PreTrainedConfig):
    # Basic policy settings
    n_obs_steps: int = 1
-    chunk_size: int = 40
+    chunk_size: int = 50
-    n_action_steps: int = 40
+    n_action_steps: int = 50
    # Dimension settings (must match pretrained GR00T model expectations)
    # Maximum state dimension. Shorter states will be zero-padded.
-    max_state_dim: int = 132
+    max_state_dim: int = 64
    # Maximum action dimension. Shorter actions will be zero-padded.
-    max_action_dim: int = 132
+    max_action_dim: int = 32
-    # GR00T normalizes state/action internally in its processor steps (min/max with
+    # Normalization (start with identity, adjust as needed)
    # q01/q99 percentiles, per embodiment), and the Qwen3-VL backbone's image processor
    # handles image normalization. The policy therefore does NOT use LeRobot's
    # NormalizerProcessorStep/UnnormalizerProcessorStep, so this mapping is intentionally
    # IDENTITY for every feature and is not consulted by make_groot_pre_post_processors.
    normalization_mapping: dict[str, NormalizationMode] = field(
        default_factory=lambda: {
            "VISUAL": NormalizationMode.IDENTITY,
-            "STATE": NormalizationMode.IDENTITY,
+            "STATE": NormalizationMode.MEAN_STD,
-            "ACTION": NormalizationMode.IDENTITY,
+            "ACTION": NormalizationMode.MEAN_STD,
        }
    )
-    # Groot-specific model parameters
+    # Image preprocessing (adjust to match Groot's expected input)
    image_size: tuple[int, int] = (224, 224)
-    # Path or HuggingFace model ID for the base GR00T N1.7 model whose backbone weights and
+    # Groot-specific model parameters (from groot_finetune_script.py)
    # checkpoint sidecars (statistics.json, processor_config.json, ...) are loaded. This is the
    # model *source*, and is intentionally distinct from the inherited `pretrained_path`:
    # `pretrained_path` (`--policy.path`) points at a saved LeRobot checkpoint directory whose
    # `config.json` carries a `type` field, whereas a raw NVIDIA GR00T checkpoint has no such
    # field and so can only be loaded through `base_model_path` (`--policy.base_model_path`).
    # Defaults to GROOT_N1_7_BASE_MODEL when unset (resolved in __post_init__).
    base_model_path: str | None = None
-    # Optional named action transform applied after raw N1.7 checkpoint decoding and before env.step().
+    # Path or HuggingFace model ID for the base Groot model
-    # 'auto' (default) resolves to the embodiment default ('libero' for 'libero_sim', otherwise no
+    base_model_path: str = "nvidia/GR00T-N1.5-3B"
-    # transform). Pass 'none' to explicitly disable the transform, including for 'libero_sim'.
+
-    action_decode_transform: str | None = GROOT_ACTION_DECODE_TRANSFORM_AUTO
+    # HF repo ID (or local path) that hosts vocab.json and merges.txt for Eagle tokenizer.
    tokenizer_assets_repo: str = "lerobot/eagle2hg-processor-groot-n1p5"
    # Embodiment tag to use for training (e.g. 'new_embodiment', 'gr1')
    embodiment_tag: str = "new_embodiment"
@@ -301,67 +75,38 @@ class GrootConfig(PreTrainedConfig):
    # Whether to fine-tune the diffusion model
    tune_diffusion_model: bool = True
-    # Whether to fine-tune the VL LayerNorm + VL self-attention projector in the action head.
+    # LoRA parameters (from groot_finetune_script.py)
-    tune_vlln: bool = True
+    # Rank for the LORA model. If 0, no LORA will be used.
    lora_rank: int = 0
-    # Number of top LLM backbone layers to fine-tune (0 = none). Lets you adapt just the final
+    # Alpha value for the LORA model
-    # language layers without unfreezing the whole backbone; independent of `tune_llm`, which tunes
+    lora_alpha: int = 16
    # the entire LLM.
    tune_top_llm_layers: int = 0
-    # Inference-time knob: Number of flow-matching denoising steps used to decode an action chunk.
+    # Dropout rate for the LORA model
-    # Trades inference latency for action quality.
+    lora_dropout: float = 0.1
    # None keeps the checkpoint value (GR00T N1.7 default: 4).
    num_inference_timesteps: int | None = None
-    # Inference-time knob: Real-Time Chunking (RTC) overlap-blend ramp rate, used when the RTC engine
+    # Whether to use the full model for LORA
-    # supplies a previous-chunk prefix. Higher values blend the overlapping prefix more aggressively.
+    lora_full_model: bool = False
    # None keeps the checkpoint value (GR00T N1.7 default: 6.0).
    rtc_ramp_rate: float | None = None
-    # Inference-time knob: Whether to request the flash-attention-2 kernel for the Qwen3-VL backbone.
+    # Training parameters (matching groot_finetune_script.py)
    # flash-attn is an optional, user-managed optimization; when it is absent (the default),
    # the backbone transparently falls back to SDPA, which is numerically equivalent.
    # Set to True only after installing a flash-attn build matching your torch/CUDA env.
    use_flash_attention: bool = False
    # Enable GR00T-style state-relative action chunks (action chunk expressed relative to the current
    # observation state).
    use_relative_actions: bool = False
    # relative_exclude_joints names the action dimensions that stay absolute; the
    # match is substring/case-insensitive against the dataset action feature names. With the empty
    # default every dimension is treated as relative, including the gripper -- set e.g. ["gripper"] to
    # keep the gripper absolute, matching the Isaac-GR00T single-arm + absolute-gripper convention.
    relative_exclude_joints: list[str] = field(default_factory=list)
    # Training parameters
    optimizer_lr: float = 1e-4
-    # Isaac-GR00T N1.7 fine-tunes with AdamW betas (0.9, 0.999).
+    optimizer_betas: tuple[float, float] = (0.95, 0.999)
    optimizer_betas: tuple[float, float] = (0.9, 0.999)
    optimizer_eps: float = 1e-8
    optimizer_weight_decay: float = 1e-5
    warmup_ratio: float = 0.05
    use_bf16: bool = True
    # The native N1.7 fine-tuning recipe keeps model parameters in FP32 and computes under BF16 autocast.
    model_params_fp32: bool = True
-    # TODO(Steven): Remove these deprecated fields in a future release.
+    # Dataset parameters
-    # Deprecated Isaac-GR00T runner / GR00T N1.5 fields, plus the (never-wired) LoRA fields — all
+    # Video backend to use for training ('decord' or 'torchvision_av')
    # unused by the LeRobot N1.7 implementation except the `tokenizer_assets_repo` N1.5 tripwire and
    # the `image_size` legacy remap in __post_init__. They are kept ONLY so a config.json saved by an
    # earlier lerobot release (notably a GR00T N1.5 checkpoint) still parses under draccus — which
    # rejects unknown fields — and is then rejected with a clear N1.5 removal message rather than an
    # opaque draccus decoding error.
    image_size: tuple[int, int] = (256, 256)  # image sizing is handled by the backbone's image processor.
    tokenizer_assets_repo: str | None = None
    lora_rank: int = 0
    lora_alpha: int = 16
    lora_dropout: float = 0.1
    lora_full_model: bool = False
    video_backend: str = "decord"
    # Whether to balance dataset weights in mixture datasets
    balance_dataset_weights: bool = True
    # Whether to sample trajectories weighted by their length
    balance_trajectory_weights: bool = True
    # Optional dataset paths for delegating training to Isaac-GR00T runner
    dataset_paths: list[str] | None = None
    output_dir: str = "./tmp/gr00t"
    save_steps: int = 1000
@@ -372,65 +117,6 @@ class GrootConfig(PreTrainedConfig):
    resume: bool = False
    def __post_init__(self):
        if self.tokenizer_assets_repo is not None:
            raise ValueError(
                "Config sets 'tokenizer_assets_repo', which only existed for GR00T N1.5; this looks "
                f"like a legacy GR00T N1.5 checkpoint or config. {GROOT_N1_5_REMOVAL_GUIDANCE}"
            )
        self.action_decode_transform = normalize_groot_action_decode_transform(self.action_decode_transform)
        if self.base_model_path is None:
            self.base_model_path = GROOT_N1_7_BASE_MODEL
        # The N1.7 LIBERO checkpoints emit a [0, 1] gripper action, but the LIBERO
        # simulator expects the OpenVLA/[-1, 1] sign convention. NVIDIA's rollout
        # wrapper applies this conversion; mirror it here so eval on the
        # 'libero_sim' embodiment grasps correctly instead of scoring 0% success.
        # This matches the embodiment-specific handling already done for the
        # action execution horizon (see infer_groot_n1_7_action_execution_horizon).
        # Only the 'auto' sentinel resolves to the embodiment default; an explicit
        # 'none' (normalized to None above) keeps the transform disabled.
        if self.action_decode_transform == GROOT_ACTION_DECODE_TRANSFORM_AUTO:
            self.action_decode_transform = (
                GROOT_ACTION_DECODE_TRANSFORM_LIBERO if self.embodiment_tag == "libero_sim" else None
            )
        # GR00T N1.5-era default values (e.g. --policy.chunk_size=50 from old commands or
        # stale configs) are migrated to the values the N1.7 checkpoints expect, with a
        # warning. The dataclass defaults are already the N1.7 values, so a plain
        # GrootConfig() never triggers this.
        legacy_default_remaps = (
            ("max_state_dim", 64, 132),
            ("max_action_dim", 32, 132),
            ("chunk_size", 50, 40),
            ("n_action_steps", 50, 40),
            ("image_size", (224, 224), (256, 256)),
        )
        for field_name, legacy_value, n1_7_value in legacy_default_remaps:
            current_value = getattr(self, field_name)
            if isinstance(legacy_value, tuple):
                current_value = tuple(current_value)
            if current_value == legacy_value:
                logger.warning(
                    "GrootConfig.%s=%s matches a legacy GR00T N1.5-era default; remapping it to %s, "
                    "the value expected by GR00T N1.7 checkpoints. Set a different value explicitly "
                    "if this is not what you want.",
                    field_name,
                    legacy_value,
                    n1_7_value,
                )
                setattr(self, field_name, n1_7_value)
        inferred_version = infer_groot_model_version(self.base_model_path)
        if inferred_version is not None and inferred_version != GROOT_N1_7:
            message = (
                f"GR00T model_version '{GROOT_N1_7}' does not match base_model_path "
                f"'{self.base_model_path}', which looks like '{inferred_version}'."
            )
            if inferred_version == GROOT_N1_5:
                message = f"{message} {GROOT_N1_5_REMOVAL_GUIDANCE}"
            raise ValueError(message)
        super().__post_init__()
        if self.n_action_steps > self.chunk_size:
@@ -438,6 +124,9 @@ class GrootConfig(PreTrainedConfig):
                f"n_action_steps ({self.n_action_steps}) cannot exceed chunk_size ({self.chunk_size})"
            )
        # groot_repo_path is now optional since we ported the components
        # No validation needed
    def validate_features(self) -> None:
        """Validate and set up input/output features for Groot."""
        image_features = [key for key, feat in self.input_features.items() if feat.type == FeatureType.VISUAL]
@@ -484,20 +173,15 @@ class GrootConfig(PreTrainedConfig):
            betas=self.optimizer_betas,
            eps=self.optimizer_eps,
            weight_decay=self.optimizer_weight_decay,
            grad_clip_norm=1.0,
        )
-    def get_scheduler_preset(self) -> DiffuserSchedulerConfig:
+    def get_scheduler_preset(self) -> CosineDecayWithWarmupSchedulerConfig:
-        """Return scheduler configuration.
+        """Return scheduler configuration."""
-
+        return CosineDecayWithWarmupSchedulerConfig(
-        Isaac-GR00T uses the HF Trainer cosine schedule with ~5% warmup over the
+            num_warmup_steps=int(10000 * self.warmup_ratio),  # 5% warmup by default
-        actual training update count; DiffuserSchedulerConfig wraps the same
+            num_decay_steps=10000,  # Adjust based on training steps
-        diffusers/transformers `get_scheduler("cosine")` implementation and
+            peak_lr=self.optimizer_lr,
-        derives num_training_steps from the outer --steps value at runtime.
+            decay_lr=self.optimizer_lr * 0.1,
        """
        return DiffuserSchedulerConfig(
            name="cosine",
            num_warmup_steps=math.ceil(self.max_steps * self.warmup_ratio),
        )
    @property
@@ -508,15 +192,7 @@ class GrootConfig(PreTrainedConfig):
    @property
    def action_delta_indices(self) -> list[int]:
        """Return indices for delta actions."""
-        model_action_horizon = (
+        return list(range(min(self.chunk_size, 16)))
            infer_groot_n1_7_action_horizon(self.base_model_path, self.embodiment_tag) or 40
        )
        return list(range(min(self.chunk_size, model_action_horizon)))
    @property
    def drop_n_last_frames(self) -> int:
        """Exclude episode tails that cannot supply a complete N1.7 action chunk."""
        return max(0, len(self.action_delta_indices) - 1)
    @property
    def reward_delta_indices(self) -> None:
@@ -0,0 +1,135 @@
 # SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
 # http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import copy
 from transformers.configuration_utils import PretrainedConfig
 from transformers.models.llama.configuration_llama import LlamaConfig
 from transformers.models.qwen2.configuration_qwen2 import Qwen2Config
 from transformers.models.qwen3.configuration_qwen3 import Qwen3Config
 from transformers.models.siglip.configuration_siglip import SiglipVisionConfig
 from transformers.utils import logging
 logger = logging.get_logger(__name__)
 class Eagle25VLConfig(PretrainedConfig):
    model_type = "eagle_2_5_vl"
    is_composition = True
    sub_configs = {"vision_config": SiglipVisionConfig, "text_config": Qwen2Config}
    def __init__(
        self,
        vision_config=None,
        text_config=None,
        use_backbone_lora=0,
        use_llm_lora=0,
        pad2square=False,
        select_layer=-4,
        force_image_size=None,
        downsample_ratio=0.5,
        template=None,
        dynamic_image_size=False,
        use_thumbnail=False,
        loss_version="v1",
        min_dynamic_tiles=1,
        max_dynamic_tiles=6,
        mlp_checkpoint=False,
        initializer_range=0.02,
        _attn_implementation="flash_attention_2",
        _attn_implementation_autoset=False,
        llm_config=None,
        image_token_index=None,
        use_pixel_shuffle=True,
        mlp_connector_layers=2,
        **kwargs,
    ):
        super().__init__(**kwargs)
        if vision_config is None:
            vision_config = {"model_type": "siglip_vision_model"}
            logger.info("vision_config is None. Initializing the InternVisionConfig with default values.")
        if text_config is None:
            text_config = {"architectures": ["Qwen2ForCausalLM"]}
            logger.info(
                "text_config is None. Initializing the LlamaConfig config with default values (`LlamaConfig`)."
            )
        if vision_config["model_type"] == "siglip_vision_model":
            self.vision_config = SiglipVisionConfig(**vision_config)
        else:
            raise ValueError("Unsupported model_type: {}".format(vision_config["model_type"]))
        if text_config["architectures"][0] == "LlamaForCausalLM":
            self.text_config = LlamaConfig(**text_config)
        elif text_config["architectures"][0] == "Qwen2ForCausalLM":
            self.text_config = Qwen2Config(**text_config)
        elif text_config["architectures"][0] == "Qwen3ForCausalLM":
            self.text_config = Qwen3Config(**text_config)
        else:
            raise ValueError("Unsupported architecture: {}".format(text_config["architectures"][0]))
        self.use_backbone_lora = use_backbone_lora
        self.use_llm_lora = use_llm_lora
        self.mlp_checkpoint = mlp_checkpoint
        self.pad2square = pad2square
        self.select_layer = select_layer
        self.force_image_size = force_image_size
        self.downsample_ratio = downsample_ratio
        self.template = template
        self.dynamic_image_size = dynamic_image_size
        self.use_thumbnail = use_thumbnail
        self.loss_version = loss_version
        self.initializer_range = initializer_range
        self.min_dynamic_tiles = min_dynamic_tiles
        self.max_dynamic_tiles = max_dynamic_tiles
        self.tie_word_embeddings = self.text_config.tie_word_embeddings
        self._attn_implementation = _attn_implementation
        self._attn_implementation_autoset = _attn_implementation_autoset
        self.image_token_index = image_token_index
        self.use_pixel_shuffle = use_pixel_shuffle
        self.mlp_connector_layers = mlp_connector_layers
        logger.info(f"min_dynamic_tiles: {self.min_dynamic_tiles}")
        logger.info(f"max_dynamic_tiles: {self.max_dynamic_tiles}")
    def to_dict(self):
        """
        Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`].
        Returns:
            `Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance,
        """
        output = copy.deepcopy(self.__dict__)
        output["vision_config"] = self.vision_config.to_dict()
        output["text_config"] = self.text_config.to_dict()
        output["model_type"] = self.__class__.model_type
        output["use_backbone_lora"] = self.use_backbone_lora
        output["use_llm_lora"] = self.use_llm_lora
        output["pad2square"] = self.pad2square
        output["select_layer"] = self.select_layer
        output["force_image_size"] = self.force_image_size
        output["downsample_ratio"] = self.downsample_ratio
        output["template"] = self.template
        output["dynamic_image_size"] = self.dynamic_image_size
        output["use_thumbnail"] = self.use_thumbnail
        output["min_dynamic_tiles"] = self.min_dynamic_tiles
        output["max_dynamic_tiles"] = self.max_dynamic_tiles
        output["tie_word_embeddings"] = self.tie_word_embeddings
        output["_attn_implementation"] = self._attn_implementation
        output["_attn_implementation_autoset"] = self._attn_implementation_autoset
        output["use_pixel_shuffle"] = self.use_pixel_shuffle
        output["mlp_connector_layers"] = self.mlp_connector_layers
        return output
@@ -0,0 +1,503 @@
 # --------------------------------------------------------
 # NVIDIA
 # Copyright (c) 2025 NVIDIA
 # Licensed under The MIT License [see LICENSE for details]
 # --------------------------------------------------------
 from __future__ import annotations
 # copy from https://github.com/huggingface/transformers/blob/main/src/transformers/models/llava_onevision/image_processing_llava_onevision_fast.py
 from transformers.image_processing_utils import (
    BatchFeature,
    get_patch_output_size,
 )
 from transformers.image_processing_utils_fast import (
    BaseImageProcessorFast,
    ImagesKwargs,
    group_images_by_shape,
    reorder_images,
 )
 from transformers.image_utils import (
    IMAGENET_STANDARD_MEAN,  # 0.5, 0.5, 0.5
    IMAGENET_STANDARD_STD,  # 0.5, 0.5, 0.5
    ChannelDimension,
    ImageInput,
    PILImageResampling,
    SizeDict,
    get_image_size,
    make_flat_list_of_images,
    validate_kwargs,
 )
 from transformers.processing_utils import Unpack
 from transformers.utils import (
    TensorType,
    add_start_docstrings,
    is_torch_available,
    is_torchvision_v2_available,
 )
 from transformers.video_utils import VideoInput
 if is_torch_available():
    import torch
 if is_torchvision_v2_available():
    from torchvision.transforms.v2 import functional as F  # noqa: N812
    from transformers.image_utils import pil_torch_interpolation_mapping
 else:
    from torchvision.transforms import functional as F  # noqa: N812
 def crop(img: torch.Tensor, left: int, top: int, right: int, bottom: int) -> torch.Tensor:
    """Crop the given numpy array.
    Args:
        img (torch.Tensor): Image to be cropped. Format should be (C, H, W).
        left (int): The left coordinate of the crop box.
        top (int): The top coordinate of the crop box.
        right (int): The right coordinate of the crop box.
        bottom (int): The bottom coordinate of the crop box.
    Returns:
        torch.Tensor: Cropped image.
    """
    if not isinstance(img, torch.Tensor):
        raise TypeError(f"img should be torch.Tensor. Got {type(img)}")
    if img.ndim not in [2, 3]:
        raise ValueError(f"Image should have 2 or 3 dimensions. Got {img.ndim}")
    img_height = img.shape[1]
    img_width = img.shape[2]
    if top < 0 or left < 0 or bottom > img_height or right > img_width:
        raise ValueError("Crop coordinates out of bounds")
    if top >= bottom or left >= right:
        raise ValueError("Invalid crop coordinates")
    return img[:, top:bottom, left:right]
 class Eagle25VLFastImageProcessorKwargs(ImagesKwargs):
    max_dynamic_tiles: int | None
    min_dynamic_tiles: int | None
    use_thumbnail: bool | None
    pad_during_tiling: bool | None
    do_pad: bool | None
@add_start_docstrings(
    "Constructs a fast ConvNeXT image processor. Based on [`SiglipImageProcessor`] with incorporation of processing each video frame.",
    # BASE_IMAGE_PROCESSOR_FAST_DOCSTRING, TODO: this was depreciated from transformers remove!
    """
        image_grid_pinpoints (`List[List[int]]`, *optional*):
            A list of possible resolutions to use for processing high resolution images. The best resolution is selected
            based on the original size of the image. Can be overridden by `image_grid_pinpoints` in the `preprocess`
            method. Not used for processing videos.
        do_pad (`bool`, *optional*):
            Whether to pad the image. If `True`, will pad the patch dimension of the images in the batch to the largest
            number of patches in the batch. Padding will be applied to the bottom and right with zeros.
    """,
 )
 class Eagle25VLImageProcessorFast(BaseImageProcessorFast):
    resample = PILImageResampling.BICUBIC
    image_mean = IMAGENET_STANDARD_MEAN
    image_std = IMAGENET_STANDARD_STD
    size = {"height": 448, "width": 448}
    default_to_square = False
    crop_size = None
    do_resize = True
    do_center_crop = None
    do_rescale = True
    do_normalize = True
    do_convert_rgb = True
    do_pad = True
    max_dynamic_tiles = 12
    min_dynamic_tiles = 1
    use_thumbnail = True
    pad_during_tiling = False
    valid_kwargs = Eagle25VLFastImageProcessorKwargs
    model_input_names = ["pixel_values_videos"]
    def __init__(self, **kwargs: Unpack[Eagle25VLFastImageProcessorKwargs]):
        super().__init__(**kwargs)
    @add_start_docstrings(
        # BASE_IMAGE_PROCESSOR_FAST_DOCSTRING_PREPROCESS, TODO: this was depreciated from transformers remove!
        """
            max_dynamic_tiles (`int`, *optional*):
                The maximum number of dynamic tiles to use for processing high resolution images.
            min_dynamic_tiles (`int`, *optional*):
                The minimum number of dynamic tiles to use for processing high resolution images.
            use_thumbnail (`bool`, *optional*):
                Whether to use a thumbnail for processing high resolution images.
            pad_during_tiling (`bool`, *optional*):
                Whether to pad the image during tiling.
            do_pad (`bool`, *optional*):
                    Whether to pad the image. If `True`, will pad the patch dimension of the images in the batch to the largest
                    number of patches in the batch. Padding will be applied to the bottom and right with zeros.
        """,
    )
    # NOTE(YL): we will overload the preprocess method to add the image_flags
    # def preprocess(
    #     self, images: ImageInput, **kwargs: Unpack[Eagle25VLFastImageProcessorKwargs]
    # ) -> BatchFeature:
    #     return super().preprocess(images, **kwargs)
    def _prepare_images_structure(
        self,
        images: ImageInput,
        expected_ndims: int = 3,
    ) -> ImageInput:
        """
        Prepare the images structure for processing.
        Args:
            images (`ImageInput`):
                The input images to process.
            expected_ndims (`int`, *optional*, defaults to 3):
                Expected number of dimensions for the images (added for transformers >=4.53.0 compatibility).
        Returns:
            `ImageInput`: The images with a valid nesting.
        """
        return make_flat_list_of_images(images)
    def _resize_for_patching(
        self,
        image: torch.Tensor,
        target_resolution: tuple,
        interpolation: F.InterpolationMode,
        input_data_format: ChannelDimension,
    ) -> torch.Tensor:
        """
        Resizes an image to a target resolution while maintaining aspect ratio.
        Args:
            image ("torch.Tensor"):
                The input image.
            target_resolution (tuple):
                The target resolution (height, width) of the image.
            interpolation (`InterpolationMode`):
                Resampling filter to use if resizing the image.
            input_data_format (`ChannelDimension` or `str`):
                The channel dimension format of the input image.
        Returns:
            "torch.Tensor": The resized and padded image.
        """
        new_height, new_width = get_patch_output_size(image, target_resolution, input_data_format)
        # Resize the image
        resized_image = F.resize(image, (new_height, new_width), interpolation=interpolation)
        return resized_image
    def find_closest_aspect_ratio(self, aspect_ratio, target_ratios, width, height, image_size):
        """
        previous version mainly focus on ratio.
        We also consider area ratio here.
        """
        best_factor = float("-inf")
        best_ratio = (1, 1)
        area = width * height
        for ratio in target_ratios:
            target_aspect_ratio = ratio[0] / ratio[1]
            # ratio_diff = abs(aspect_ratio - target_aspect_ratio)
            # area_ratio = (ratio[0] * ratio[1] * image_size * image_size) / area
            """
            new area > 60% of original image area is enough.
            """
            factor_based_on_area_n_ratio = min(
                (ratio[0] * ratio[1] * image_size * image_size) / area, 0.6
            ) * min(target_aspect_ratio / aspect_ratio, aspect_ratio / target_aspect_ratio)
            if factor_based_on_area_n_ratio > best_factor:
                best_factor = factor_based_on_area_n_ratio
                best_ratio = ratio
        return best_ratio
    def _pad_for_patching(
        self, image: torch.Tensor, target_resolution: tuple, input_data_format: ChannelDimension
    ) -> torch.Tensor:
        """
        Pad an image to a target resolution while maintaining aspect ratio.
        """
        target_height, target_width = target_resolution
        new_height, new_width = get_patch_output_size(image, target_resolution, input_data_format)
        paste_x = (target_width - new_width) // 2
        paste_y = (target_height - new_height) // 2
        padded_image = F.pad(image, padding=[paste_x, paste_y, paste_x, paste_y])
        return padded_image
    def _get_image_patches(
        self,
        image: torch.Tensor,
        min_num: int,
        max_num: int,
        size: tuple,
        tile_size: int,
        use_thumbnail: bool,
        interpolation: F.InterpolationMode,
        pad_during_tiling: bool,
    ) -> list[torch.Tensor]:
        image_size = get_image_size(image, channel_dim=ChannelDimension.FIRST)
        orig_height, orig_width = image_size
        aspect_ratio = orig_width / orig_height
        # calculate the existing image aspect ratio
        target_ratios = {
            (i, j)
            for n in range(min_num, max_num + 1)
            for i in range(1, n + 1)
            for j in range(1, n + 1)
            if i * j <= max_num and i * j >= min_num
        }
        target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
        # find the closest aspect ratio to the target
        target_aspect_ratio = self.find_closest_aspect_ratio(
            aspect_ratio, target_ratios, orig_width, orig_height, tile_size
        )
        # calculate the target width and height
        target_width = tile_size * target_aspect_ratio[0]
        target_height = tile_size * target_aspect_ratio[1]
        blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
        if pad_during_tiling:
            resized_image = self._resize_for_patching(
                image,
                (target_height, target_width),
                interpolation=interpolation,
                input_data_format=ChannelDimension.FIRST,
            )
            padded_image = self._pad_for_patching(
                resized_image,
                (target_height, target_width),
                input_data_format=ChannelDimension.FIRST,
            )
            image_used_to_split = padded_image
        else:
            image_used_to_split = F.resize(image, (target_height, target_width), interpolation=interpolation)
        processed_tiles = []
        for i in range(blocks):
            box = (
                (i % (target_width // tile_size)) * tile_size,
                (i // (target_width // tile_size)) * tile_size,
                ((i % (target_width // tile_size)) + 1) * tile_size,
                ((i // (target_width // tile_size)) + 1) * tile_size,
            )
            # split the image
            split_img = crop(image_used_to_split, box[0], box[1], box[2], box[3])
            processed_tiles.append(split_img)
        assert len(processed_tiles) == blocks
        if use_thumbnail and len(processed_tiles) != 1:
            thumbnail_img = F.resize(image, (tile_size, tile_size), interpolation=interpolation)
            processed_tiles.append(thumbnail_img)
        return processed_tiles
    def _pad_for_batching(
        self,
        pixel_values: list[torch.Tensor],
    ) -> list[torch.Tensor]:
        """
        Pads images on the `num_of_patches` dimension with zeros to form a batch of same number of patches.
        Args:
            pixel_values (`List[torch.Tensor]`):
                An array of pixel values of each images of shape (`batch_size`, `num_patches`, `image_in_3D`)
        Returns:
            List[`torch.Tensor`]: The padded images.
        """
        max_patch = max(len(x) for x in pixel_values)
        pixel_values = [
            torch.nn.functional.pad(image, pad=[0, 0, 0, 0, 0, 0, 0, max_patch - image.shape[0]])
            for image in pixel_values
        ]
        return pixel_values
    def _preprocess(
        self,
        images: list[torch.Tensor],
        do_resize: bool,
        size: SizeDict,
        max_dynamic_tiles: int,
        min_dynamic_tiles: int,
        use_thumbnail: bool,
        pad_during_tiling: bool,
        interpolation: F.InterpolationMode | None,
        do_center_crop: bool,
        crop_size: SizeDict,
        do_rescale: bool,
        rescale_factor: float,
        do_normalize: bool,
        image_mean: float | list[float] | None,
        image_std: float | list[float] | None,
        do_pad: bool,
        return_tensors: str | TensorType | None,
        pad_size: SizeDict | None = None,  # Added for transformers >=4.53.0 compatibility
        disable_grouping: bool | None = None,  # Added for transformers >=4.53.0 compatibility
    ) -> BatchFeature:
        processed_images = []
        image_sizes = []
        # Determine the size tuple
        if size and size.height and size.width:
            size_tuple = (size.height, size.width)
        else:
            size_tuple = (size.shortest_edge, size.shortest_edge)
        # Determine the patch size
        if crop_size and crop_size.height:
            tile_size = crop_size.height
        elif size and size.height:
            tile_size = size.height
        else:
            tile_size = size.shortest_edge
        for image in images:
            image_patches = self._get_image_patches(
                image,
                min_num=min_dynamic_tiles,
                max_num=max_dynamic_tiles,
                size=size_tuple,
                tile_size=tile_size,
                use_thumbnail=use_thumbnail,
                interpolation=interpolation,
                pad_during_tiling=pad_during_tiling,
            )
            # Group images by size for batched processing
            processed_image_patches_grouped = {}
            # Added for transformers >=4.53.0 compatibility
            grouped_image_patches, grouped_image_patches_index = group_images_by_shape(
                image_patches,
                disable_grouping=disable_grouping,
            )
            for shape, stacked_image_patches in grouped_image_patches.items():
                if do_resize:
                    stacked_image_patches = self.resize(
                        image=stacked_image_patches,
                        size=size,
                        interpolation=interpolation,
                    )
                if do_center_crop:
                    stacked_image_patches = self.center_crop(stacked_image_patches, crop_size)
                # Fused rescale and normalize
                stacked_image_patches = self.rescale_and_normalize(
                    stacked_image_patches,
                    do_rescale,
                    rescale_factor,
                    do_normalize,
                    image_mean,
                    image_std,
                )
                processed_image_patches_grouped[shape] = stacked_image_patches
            processed_image_patches = reorder_images(
                processed_image_patches_grouped, grouped_image_patches_index
            )
            processed_image_patches = (
                torch.stack(processed_image_patches, dim=0) if return_tensors else processed_image_patches
            )
            processed_images.append(processed_image_patches)
            image_sizes.append(get_image_size(image, ChannelDimension.FIRST))
        if do_pad:
            processed_images = self._pad_for_batching(processed_images)
        # processed_images = torch.stack(processed_images, dim=0) if return_tensors else processed_images
        processed_images = torch.cat(processed_images, dim=0) if return_tensors else processed_images
        return BatchFeature(
            data={"pixel_values": processed_images, "image_sizes": image_sizes},
            tensor_type=return_tensors,
        )
    def preprocess(
        self,
        images: ImageInput,
        videos: VideoInput = None,
        **kwargs: Unpack[Eagle25VLFastImageProcessorKwargs],
    ) -> BatchFeature:
        validate_kwargs(
            captured_kwargs=kwargs.keys(),
            valid_processor_keys=self.valid_kwargs.__annotations__.keys(),
        )
        # Set default kwargs from self. This ensures that if a kwarg is not provided
        # by the user, it gets its default value from the instance, or is set to None.
        for kwarg_name in self.valid_kwargs.__annotations__:
            kwargs.setdefault(kwarg_name, getattr(self, kwarg_name, None))
        # Extract parameters that are only used for preparing the input images
        do_convert_rgb = kwargs.pop("do_convert_rgb")
        input_data_format = kwargs.pop("input_data_format")
        device = kwargs.pop("device")
        # Prepare input images
        # transformers >= 4.53.0: uses _prepare_image_like_inputs instead of _prepare_input_images
        if images is not None:
            images = self._prepare_image_like_inputs(
                images=images,
                do_convert_rgb=do_convert_rgb,
                input_data_format=input_data_format,
                device=device,
            )
        if videos is not None:
            videos = self._prepare_image_like_inputs(
                images=videos,
                do_convert_rgb=do_convert_rgb,
                input_data_format=input_data_format,
                device=device,
            )
        # Update kwargs that need further processing before being validated
        kwargs = self._further_process_kwargs(**kwargs)
        # Validate kwargs
        self._validate_preprocess_kwargs(**kwargs)
        # torch resize uses interpolation instead of resample
        # Added for transformers >=4.53.0 compatibility
        resample = kwargs.pop("resample", self.resample)
        kwargs["interpolation"] = (
            pil_torch_interpolation_mapping[resample]
            if isinstance(resample, PILImageResampling | int)
            else resample
        )
        # Filter kwargs to only include those accepted by _preprocess
        valid_preprocess_kwargs = {
            "do_resize",
            "size",
            "max_dynamic_tiles",
            "min_dynamic_tiles",
            "use_thumbnail",
            "pad_during_tiling",
            "interpolation",
            "do_center_crop",
            "crop_size",
            "do_rescale",
            "rescale_factor",
            "do_normalize",
            "image_mean",
            "image_std",
            "do_pad",
            "return_tensors",
            "pad_size",
            "disable_grouping",
        }
        filtered_kwargs = {k: v for k, v in kwargs.items() if k in valid_preprocess_kwargs}
        if images is not None:
            return self._preprocess(images, **filtered_kwargs)
        elif videos is not None:
            return self._preprocess(videos, **filtered_kwargs)
 __all__ = ["Eagle25VLImageProcessorFast"]
@@ -0,0 +1,396 @@
 # --------------------------------------------------------
 # NVIDIA
 # Copyright (c) 2025 NVIDIA
 # Licensed under The MIT License [see LICENSE for details]
 # --------------------------------------------------------
 import inspect
 import torch
 import torch.utils.checkpoint as cp
 from peft import LoraConfig, get_peft_model
 from torch import nn
 from torch.nn import CrossEntropyLoss
 from transformers import GenerationConfig
 from transformers.generation import GenerationMixin
 from transformers.modeling_outputs import CausalLMOutputWithPast
 from transformers.modeling_utils import PreTrainedModel
 from transformers.models.llama.modeling_llama import LlamaForCausalLM
 from transformers.models.qwen2.modeling_qwen2 import Qwen2ForCausalLM
 from transformers.models.qwen3.modeling_qwen3 import Qwen3ForCausalLM
 from transformers.models.siglip.modeling_siglip import SiglipVisionModel
 from transformers.utils import add_start_docstrings, logging
 from .configuration_eagle2_5_vl import Eagle25VLConfig
 logger = logging.get_logger(__name__)
 # copy from https://github.com/huggingface/transformers/blob/main/src/transformers/models/llava_onevision/modeling_llava_onevision.py#L241C1-L280C1
 EAGLE2_5_VL_START_DOCSTRING = r"""
    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
    etc.)
    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
    and behavior.
    Parameters:
        config ([`Eagle25VLConfig`]):
            Model configuration class with all the parameters of the model. Initializing with a config file does not
            load the weights associated with the model, only the configuration. Check out the
            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
 """
@add_start_docstrings(
    "The bare Eagle2_5_VL Model outputting raw hidden-states without any specific head on top.",
    EAGLE2_5_VL_START_DOCSTRING,
 )
 class Eagle25VLPreTrainedModel(PreTrainedModel):
    config_class = Eagle25VLConfig
    base_model_prefix = "model"
    main_input_name = "input_ids"
    supports_gradient_checkpointing = True
    _no_split_modules = [
        "Qwen2DecoderLayer",
        "LlamaDecoderLayer",
        "Siglip2EncoderLayer",
        "SiglipEncoderLayer",
    ]
    _skip_keys_device_placement = "past_key_values"
    _supports_flash_attn = True
    _supports_flash_attn_2 = True
    _supports_cache_class = True
    _supports_static_cache = True
    _supports_quantized_cache = True
    _supports_sdpa = True
    def _init_weights(self, module):
        std = self.config.initializer_range
        if isinstance(module, nn.Linear | nn.Conv2d):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()
 class Eagle25VLForConditionalGeneration(Eagle25VLPreTrainedModel, GenerationMixin):
    config_class = Eagle25VLConfig
    def __init__(self, config: Eagle25VLConfig, vision_model=None, language_model=None):
        super().__init__(config)
        image_size = config.force_image_size or config.vision_config.image_size
        patch_size = config.vision_config.patch_size
        self.patch_size = patch_size
        if config.use_pixel_shuffle:
            self.num_image_token = int((image_size // patch_size) ** 2 * (config.downsample_ratio**2))
        else:
            self.num_image_token = int((image_size // patch_size) ** 2)
        self.select_layer = config.select_layer
        self.downsample_ratio = config.downsample_ratio
        self.loss_version = config.loss_version
        self.mlp_checkpoint = config.mlp_checkpoint
        self.use_pixel_shuffle = config.use_pixel_shuffle
        self.mlp_connector_layers = config.mlp_connector_layers
        logger.info(f"num_image_token: {self.num_image_token}")
        logger.info(f"mlp_checkpoint: {self.mlp_checkpoint}")
        if vision_model is not None:
            self.vision_model = vision_model
        else:
            if config.vision_config.model_type == "siglip_vision_model":
                config.vision_config._attn_implementation = "flash_attention_2"
                self.vision_model = SiglipVisionModel(config.vision_config)
            else:
                raise NotImplementedError(f"{config.vision_config.model_type} is not implemented.")
        if language_model is not None:
            self.language_model = language_model
        else:
            if config.text_config.architectures[0] == "LlamaForCausalLM":
                self.language_model = LlamaForCausalLM(config.text_config)
            elif config.text_config.architectures[0] == "Phi3ForCausalLM":
                raise NotImplementedError("Phi3 is not implemented.")
                # self.language_model = Phi3ForCausalLM(config.text_config)
            elif config.text_config.architectures[0] == "Qwen2ForCausalLM":
                assert config.text_config._attn_implementation == "flash_attention_2", (
                    f"Qwen2 must use flash_attention_2 but got {config.text_config._attn_implementation}"
                )
                self.language_model = Qwen2ForCausalLM(config.text_config)
            elif config.text_config.architectures[0] == "Qwen3ForCausalLM":
                self.language_model = Qwen3ForCausalLM(config.text_config)
            else:
                raise NotImplementedError(f"{config.text_config.architectures[0]} is not implemented.")
        vit_hidden_size = config.vision_config.hidden_size
        llm_hidden_size = config.text_config.hidden_size
        if config.mlp_connector_layers == 2:
            self.mlp1 = nn.Sequential(
                nn.LayerNorm(vit_hidden_size * int(1 / self.downsample_ratio) ** 2),
                nn.Linear(vit_hidden_size * int(1 / self.downsample_ratio) ** 2, llm_hidden_size),
                nn.GELU(),
                nn.Linear(llm_hidden_size, llm_hidden_size),
            )
        elif config.mlp_connector_layers == 1 and config.use_pixel_shuffle:
            self.mlp1 = nn.Sequential(
                nn.Linear(vit_hidden_size * int(1 / self.downsample_ratio) ** 2, llm_hidden_size),
            )
        elif config.mlp_connector_layers == 1 and not config.use_pixel_shuffle:
            self.mlp1 = nn.Sequential(
                nn.Linear(vit_hidden_size, llm_hidden_size),
            )
        else:
            raise NotImplementedError(f"{config.mlp_connector_layers} is not implemented.")
        self.image_token_index = config.image_token_index
        self.neftune_alpha = None
        if config.use_backbone_lora:
            self.wrap_backbone_lora(r=config.use_backbone_lora, lora_alpha=2 * config.use_backbone_lora)
        self.use_llm_lora = config.use_llm_lora
        if config.use_llm_lora:
            self.wrap_llm_lora(r=config.use_llm_lora, lora_alpha=2 * config.use_llm_lora)
        self.check_forward_kwargs()
    def check_forward_kwargs(self):
        # We intentionally avoid using **kwargs in forward because Hugging Face Transformers
        # has special handling for functions with **kwargs parameters that would affect
        # how our model is processed during training and inference.
        forward_params = inspect.signature(self.forward).parameters
        assert not any(k.kind == inspect.Parameter.VAR_KEYWORD for k in forward_params.values())
    def wrap_backbone_lora(self, r=128, lora_alpha=256, lora_dropout=0.05):
        lora_config = LoraConfig(
            r=r,
            target_modules=[
                "self_attn.q_proj",
                "self_attn.k_proj",
                "self_attn.v_proj",
                "self_attn.out_proj",
                "mlp.fc1",
                "mlp.fc2",
            ],
            lora_alpha=lora_alpha,
            lora_dropout=lora_dropout,
        )
        self.vision_model = get_peft_model(self.vision_model, lora_config)
        self.vision_model.print_trainable_parameters()
    def wrap_llm_lora(self, r=128, lora_alpha=256, lora_dropout=0.05):
        lora_config = LoraConfig(
            r=r,
            target_modules=[
                "self_attn.q_proj",
                "self_attn.k_proj",
                "self_attn.v_proj",
                "self_attn.o_proj",
                "mlp.gate_proj",
                "mlp.down_proj",
                "mlp.up_proj",
            ],
            lora_alpha=lora_alpha,
            lora_dropout=lora_dropout,
            task_type="CAUSAL_LM",
        )
        self.language_model = get_peft_model(self.language_model, lora_config)
        self.language_model.enable_input_require_grads()
        self.language_model.print_trainable_parameters()
        self.use_llm_lora = True
    def forward(
        self,
        pixel_values: torch.FloatTensor,
        input_ids: torch.LongTensor = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        image_flags: torch.LongTensor | None = None,
        past_key_values: list[torch.FloatTensor] | None = None,
        labels: torch.LongTensor | None = None,
        use_cache: bool | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        num_tiles_list: list[torch.Tensor] | None = None,
    ) -> tuple | CausalLMOutputWithPast:
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        input_embeds = self.language_model.get_input_embeddings()(input_ids)
        vit_embeds = self.extract_feature(pixel_values)
        if image_flags is not None:
            image_flags = image_flags.view(-1)
            vit_embeds = vit_embeds[image_flags == 1]
        b, n, c = input_embeds.shape
        input_embeds = input_embeds.reshape(b * n, c)
        input_ids = input_ids.reshape(b * n)
        selected = input_ids == self.image_token_index
        try:
            input_embeds[selected] = input_embeds[selected] * 0.0 + vit_embeds.reshape(-1, c)
        except Exception as e:
            vit_embeds = vit_embeds.reshape(-1, c)
            print(
                f"warning: {e}, input_embeds[selected].shape={input_embeds[selected].shape}, "
                f"vit_embeds.shape={vit_embeds.shape}"
            )
            n_token = selected.sum()
            input_embeds[selected] = input_embeds[selected] * 0.0 + vit_embeds[:n_token]
        input_embeds = input_embeds.reshape(b, n, c)
        outputs = self.language_model(
            inputs_embeds=input_embeds,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
        )
        logits = outputs.logits
        loss = None
        if labels is not None:
            # Shift so that tokens < n predict n
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            # Flatten the tokens
            loss_fct = CrossEntropyLoss()
            shift_logits = shift_logits.view(-1, self.language_model.config.vocab_size)
            shift_labels = shift_labels.view(-1)
            # Enable model parallelism
            shift_labels = shift_labels.to(shift_logits.device)
            loss = loss_fct(shift_logits, shift_labels)
        if not return_dict:
            output = (logits,) + outputs[1:]
            return (loss,) + output if loss is not None else output
        return CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )
    def pixel_shuffle(self, x, scale_factor=0.5):
        n, w, h, c = x.size()
        # N, W, H, C --> N, W, H * scale, C // scale
        x = x.view(n, w, int(h * scale_factor), int(c / scale_factor))
        # N, W, H * scale, C // scale --> N, H * scale, W, C // scale
        x = x.permute(0, 2, 1, 3).contiguous()
        # N, H * scale, W, C // scale --> N, H * scale, W * scale, C // (scale ** 2)
        x = x.view(n, int(h * scale_factor), int(w * scale_factor), int(c / (scale_factor * scale_factor)))
        x = x.permute(0, 2, 1, 3).contiguous()
        return x
    def extract_feature(self, pixel_values):
        if self.select_layer == -1:
            vit_embeds = self.vision_model(
                pixel_values=pixel_values, output_hidden_states=False, return_dict=True
            )
            if hasattr(vit_embeds, "last_hidden_state"):
                vit_embeds = vit_embeds.last_hidden_state
        else:
            vit_embeds = self.vision_model(
                pixel_values=pixel_values, output_hidden_states=True, return_dict=True
            ).hidden_states[self.select_layer]
        if self.use_pixel_shuffle:
            h = w = int(vit_embeds.shape[1] ** 0.5)
            vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], h, w, -1)
            vit_embeds = self.pixel_shuffle(
                vit_embeds, scale_factor=self.downsample_ratio
            )  # torch.Size([B, 1024, 1024]) -> torch.Size([B, 16, 16, 4096])
            vit_embeds = vit_embeds.reshape(
                vit_embeds.shape[0], -1, vit_embeds.shape[-1]
            )  # torch.Size([B, 16, 16, 4096]) -> torch.Size([B, 256, 4096])
        if self.mlp_checkpoint and vit_embeds.requires_grad:
            vit_embeds = cp.checkpoint(self.mlp1, vit_embeds)
        else:
            vit_embeds = self.mlp1(vit_embeds)
        return vit_embeds
    @torch.no_grad()
    def generate(
        self,
        pixel_values: torch.FloatTensor | None = None,
        input_ids: torch.FloatTensor | None = None,
        attention_mask: torch.LongTensor | None = None,
        visual_features: torch.FloatTensor | None = None,
        generation_config: GenerationConfig | None = None,
        output_hidden_states: bool | None = None,
        image_sizes: list[tuple[int, int]] | None = None,
        **generate_kwargs,
    ) -> torch.LongTensor:
        if pixel_values is not None:
            if visual_features is not None:
                vit_embeds = visual_features
            else:
                vit_embeds = self.extract_feature(pixel_values)
            input_embeds = self.language_model.get_input_embeddings()(input_ids)
            b, n, c = input_embeds.shape
            input_embeds = input_embeds.reshape(b * n, c)
            input_ids = input_ids.reshape(b * n)
            selected = input_ids == self.config.image_token_index
            assert selected.sum() != 0
            input_embeds[selected] = vit_embeds.reshape(-1, c).to(input_embeds.device)
            input_embeds = input_embeds.reshape(b, n, c)
        else:
            input_embeds = self.language_model.get_input_embeddings()(input_ids)
        if "use_cache" not in generate_kwargs:
            generate_kwargs["use_cache"] = True
        outputs = self.language_model.generate(
            inputs_embeds=input_embeds,
            attention_mask=attention_mask,
            generation_config=generation_config,
            output_hidden_states=output_hidden_states,
            **generate_kwargs,
        )
        return outputs
    # Copied from transformers.models.llava_next.modeling_llava_next.LlavaNextForConditionalGeneration.get_input_embeddings
    def get_input_embeddings(self):
        return self.language_model.get_input_embeddings()
    # Copied from transformers.models.llava_next.modeling_llava_next.LlavaNextForConditionalGeneration.set_input_embeddings
    def set_input_embeddings(self, value):
        self.language_model.set_input_embeddings(value)
    # Copied from transformers.models.llava_next.modeling_llava_next.LlavaNextForConditionalGeneration.get_output_embeddings
    def get_output_embeddings(self):
        return self.language_model.get_output_embeddings()
    # Copied from transformers.models.llava_next.modeling_llava_next.LlavaNextForConditionalGeneration.set_output_embeddings
    def set_output_embeddings(self, new_embeddings):
        self.language_model.set_output_embeddings(new_embeddings)
    # Copied from transformers.models.llava_next.modeling_llava_next.LlavaNextForConditionalGeneration.set_decoder
    def set_decoder(self, decoder):
        self.language_model.set_decoder(decoder)
    # Copied from transformers.models.llava_next.modeling_llava_next.LlavaNextForConditionalGeneration.get_decoder
    def get_decoder(self):
        return self.language_model.get_decoder()
@@ -0,0 +1,541 @@
 # Copyright 2024 The HuggingFace Inc. team.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
 #     http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """
 Processor class for Eagle25VL.
 copy from https://github.com/huggingface/transformers/blob/main/src/transformers/models/llava_onevision/processing_llava_onevision.py
 """
 import base64
 import os
 import re
 from io import BytesIO
 import requests
 import torch
 from PIL import Image
 from transformers.feature_extraction_utils import BatchFeature
 from transformers.image_utils import ImageInput
 from transformers.processing_utils import ProcessingKwargs, ProcessorMixin, Unpack
 from transformers.tokenization_utils_base import PreTokenizedInput, TextInput
 from transformers.utils import logging
 from transformers.video_utils import VideoInput
 logger = logging.get_logger(__name__)
 FRAME_FACTOR = 2
 FPS = 2.0
 FPS_MIN_FRAMES = 4
 FPS_MAX_FRAMES = 256
 def to_rgb(pil_image: Image.Image) -> Image.Image:
    if pil_image.mode == "RGBA":
        white_background = Image.new("RGB", pil_image.size, (255, 255, 255))
        white_background.paste(pil_image, mask=pil_image.split()[3])  # Use alpha channel as mask
        return white_background
    else:
        return pil_image.convert("RGB")
 def fetch_image(ele: dict[str, str | Image.Image]) -> Image.Image:
    image = ele["image"] if "image" in ele else ele["image_url"]
    image_obj = None
    if isinstance(image, Image.Image):
        image_obj = image
    elif image.startswith("http://") or image.startswith("https://"):
        response = requests.get(image, stream=True, timeout=10)
        image_obj = Image.open(BytesIO(response.content))
    elif image.startswith("file://"):
        image_obj = Image.open(image[7:])
    elif image.startswith("data:image"):
        if "base64," in image:
            _, base64_data = image.split("base64,", 1)
            data = base64.b64decode(base64_data)
            image_obj = Image.open(BytesIO(data))
    else:
        image_obj = Image.open(image)
    if image_obj is None:
        raise ValueError(
            f"Unrecognized image input, support local path, http url, base64 and PIL.Image, got {image}"
        )
    image = to_rgb(image_obj)
    if "scale_factor" in ele:
        scale_factor = ele["scale_factor"]
        image = image.resize((image.width * scale_factor, image.height * scale_factor), Image.BILINEAR)
    return image
 class Eagle25VLProcessorKwargs(ProcessingKwargs, total=False):
    # see processing_utils.ProcessingKwargs documentation for usage.
    _defaults = {
        "text_kwargs": {
            "padding": False,
        },
        "images_kwargs": {},
        "videos_kwargs": {"max_dynamic_tiles": 1},
    }
 class Eagle25VLProcessor(ProcessorMixin):
    r"""
    Constructs a Eagle25VL processor which wraps a Eagle25VL video processor, Eagle25VL image processor and a Eagle25VL tokenizer into a single processor.
    [`Eagle25VLProcessor`] offers all the functionalities of [`Eagle25VLVideoProcessor`], [`Eagle25VLImageProcessor`] and [`Eagle25VLTokenizer`]. See the
    [`~Eagle25VLVideoProcessor.__call__`], [`~Eagle25VLProcessor.__call__`] and [`~Eagle25VLProcessor.decode`] for more information.
    Args:
        image_processor ([`LlavaOnevisionImageProcessor`], *optional*):
            The image processor is a required input.
        tokenizer ([`LlamaTokenizerFast`], *optional*):
            The tokenizer is a required input.
        num_image_tokens (`int`, *optional*):
            Number of image tokens for one imagethat will be returned by vision tower.
        vision_feature_select_strategy (`str`, *optional*):
            The feature selection strategy used to select the vision feature from the vision backbone.
            Should be same as in model's config
        chat_template (`str`, *optional*): A Jinja template which will be used to convert lists of messages
            in a chat into a tokenizable string.
        image_token (`str`, *optional*, defaults to `"<image>"`):
            Special token used to denote image location.
        video_token (`str`, *optional*, defaults to `"<video>"`):
            Special token used to denote video location.
    """
    attributes = ["image_processor", "tokenizer"]
    valid_kwargs = [
        "chat_template",
        "num_image_tokens",
        "vision_feature_select_strategy",
        "image_token",
        "video_token",
        "images_kwargs",
        "videos_kwargs",
        "text_kwargs",
    ]
    tokenizer_class = "AutoTokenizer"
    def __init__(
        self,
        image_processor=None,
        tokenizer=None,
        vision_feature_select_strategy=None,
        chat_template=None,
        image_token="<IMG_CONTEXT>",  # nosec: B107
        video_token="<IMG_CONTEXT>",  # nosec: B107
        tokens_per_tile=256,
        image_placeholder="image",
        video_placeholder="video",
        image_start_token="<img>",
        image_end_token="</img>",
        **kwargs,
    ):
        self.vision_feature_select_strategy = vision_feature_select_strategy
        self.image_token = tokenizer.image_token if hasattr(tokenizer, "image_token") else image_token
        self.video_token = tokenizer.video_token if hasattr(tokenizer, "video_token") else video_token
        self.image_token_id = (
            tokenizer.image_token_id
            if getattr(tokenizer, "image_token_id", None)
            else tokenizer.convert_tokens_to_ids(self.image_token)
        )
        self.video_token_id = (
            tokenizer.video_token_id
            if getattr(tokenizer, "video_token_id", None)
            else tokenizer.convert_tokens_to_ids(self.video_token)
        )
        self.image_placeholder = image_placeholder
        self.video_placeholder = video_placeholder
        self.tokens_per_tile = tokens_per_tile
        self.image_start_token = image_start_token
        self.image_end_token = image_end_token
        if "auto_map" in kwargs:
            self.auto_map = kwargs["auto_map"]
        super().__init__(image_processor, tokenizer, chat_template=chat_template)
    def replace_media_placeholder(
        self, text, image_list, video_list, timestamps_list, fps_list, **output_kwargs
    ):
        num_of_images_in_this_sample = 0
        num_of_videos_in_this_sample = 0
        # Regular expression pattern to match formats like <image-1> or <video-2>
        pattern = re.compile(rf"<({self.image_placeholder}|{self.video_placeholder})-(\d+)>")
        unified_frame_list = []
        # image_min_dynamic_tiles = output_kwargs["images_kwargs"].get(
        #     "min_dynamic_tiles", self.image_processor.min_dynamic_tiles
        # )
        # image_max_dynamic_tiles = output_kwargs["images_kwargs"].get(
        #     "max_dynamic_tiles", self.image_processor.max_dynamic_tiles
        # )
        # image_use_thumbnail = output_kwargs["images_kwargs"].get(
        #     "use_thumbnail", self.image_processor.use_thumbnail
        # )
        video_min_dynamic_tiles = output_kwargs["videos_kwargs"].get(
            "min_dynamic_tiles", self.image_processor.min_dynamic_tiles
        )
        video_max_dynamic_tiles = output_kwargs["videos_kwargs"].get(
            "max_dynamic_tiles", self.image_processor.max_dynamic_tiles
        )
        video_use_thumbnail = output_kwargs["videos_kwargs"].get(
            "use_thumbnail", self.image_processor.use_thumbnail
        )
        tile_size = self.image_processor.size.get("height", 448)
        # Function to replace tags in a single text
        def replace_in_text(text):
            # repl callback function for each match replacement operation
            def repl(match):
                nonlocal unified_frame_list
                nonlocal num_of_images_in_this_sample
                nonlocal num_of_videos_in_this_sample
                media_type = match.group(1)  # 'image' or 'video'
                idx_in_list = int(match.group(2)) - 1  # Convert to list index (0-based)
                # Select the corresponding path based on media type
                idx_mapper = {
                    0: "first",
                    1: "second",
                    2: "third",
                    3: "fourth",
                    4: "fifth",
                    5: "sixth",
                    6: "seventh",
                    7: "eighth",
                    8: "ninth",
                    9: "tenth",
                }
                if media_type == "image":
                    image_inputs = self.image_processor(
                        images=[image_list[idx_in_list]],
                        videos=None,
                        **output_kwargs["images_kwargs"],
                    )
                    if isinstance(image_inputs["pixel_values"], list):
                        _pv = image_inputs["pixel_values"]
                        if _pv and isinstance(_pv[0], list):
                            _pv = [t for sub in _pv for t in sub]
                        image_inputs["pixel_values"] = torch.stack(
                            [t if isinstance(t, torch.Tensor) else torch.as_tensor(t) for t in _pv]
                        )
                    num_all_tiles = image_inputs["pixel_values"].shape[0]
                    special_placeholder = f"<image {idx_in_list + 1}>{self.image_start_token}{self.image_token * num_all_tiles * self.tokens_per_tile}{self.image_end_token}"
                    unified_frame_list.append(image_inputs)
                    num_of_images_in_this_sample += 1
                elif media_type == "video":
                    video_inputs = self.image_processor(
                        images=None,
                        videos=[video_list[idx_in_list]],
                        **output_kwargs["videos_kwargs"],
                    )
                    if isinstance(video_inputs["pixel_values"], list):
                        _pv = video_inputs["pixel_values"]
                        if _pv and isinstance(_pv[0], list):
                            _pv = [t for sub in _pv for t in sub]
                        video_inputs["pixel_values"] = torch.stack(
                            [t if isinstance(t, torch.Tensor) else torch.as_tensor(t) for t in _pv]
                        )
                    num_all_tiles = video_inputs["pixel_values"].shape[0]
                    image_sizes = video_inputs["image_sizes"]
                    if timestamps_list is not None and -1 not in timestamps_list:
                        frame_timestamps = timestamps_list[idx_in_list]
                    else:
                        frame_timestamps = None
                    sampled_fps = fps_list[idx_in_list] if fps_list is not None else None
                    num_of_tiles_each_frame = [
                        self.get_number_tiles_based_on_image_size(
                            image_size,
                            video_min_dynamic_tiles,
                            video_max_dynamic_tiles,
                            video_use_thumbnail,
                            tile_size,
                        )
                        for image_size in image_sizes
                    ]
                    assert sum(num_of_tiles_each_frame) == num_all_tiles, (
                        f"The number of tiles in each frame is not equal to the total number of tiles: {sum(num_of_tiles_each_frame)} != {num_all_tiles}"
                    )
                    if frame_timestamps is not None:
                        assert len(frame_timestamps) == len(num_of_tiles_each_frame), (
                            f"The number of timestamps is not equal to the number of frames: {len(frame_timestamps)} != {len(num_of_tiles_each_frame)}"
                        )
                        special_placeholder = [
                            f"Frame {i + 1} sample at {frame_timestamps[i]:.2f}s: {self.image_start_token}{self.image_token * num_of_tiles * self.tokens_per_tile}{self.image_end_token}"
                            for i, num_of_tiles in enumerate(num_of_tiles_each_frame)
                        ]
                    else:
                        special_placeholder = [
                            f"Frame {i + 1}: {self.image_start_token}{self.image_token * num_of_tiles * self.tokens_per_tile}{self.image_end_token}"
                            for i, num_of_tiles in enumerate(num_of_tiles_each_frame)
                        ]
                    if sampled_fps is not None:
                        special_placeholder = (
                            f"The {idx_mapper[idx_in_list]} video sampled with {sampled_fps:.2f} fps: "
                            + "".join(special_placeholder)
                        )
                    else:
                        special_placeholder = f"The {idx_mapper[idx_in_list]} video: " + "".join(
                            special_placeholder
                        )
                    unified_frame_list.append(video_inputs)
                    num_of_videos_in_this_sample += 1
                else:
                    raise ValueError(f"Unknown media type: {media_type}")
                return special_placeholder
            return pattern.sub(repl, text)
        text = replace_in_text(text)
        if len(unified_frame_list) > 0:
            def _to_tensor(v):
                if isinstance(v, torch.Tensor):
                    return v
                if isinstance(v, list):
                    if v and isinstance(v[0], list):
                        v = [t for sub in v for t in sub]
                    return torch.stack([t if isinstance(t, torch.Tensor) else torch.as_tensor(t) for t in v])
                return torch.as_tensor(v)
            pixel_values = torch.cat([_to_tensor(frame["pixel_values"]) for frame in unified_frame_list])
            image_sizes = torch.cat([_to_tensor(frame["image_sizes"]) for frame in unified_frame_list])
        else:
            pixel_values = None
            image_sizes = None
        return (
            text,
            pixel_values,
            image_sizes,
            num_of_images_in_this_sample,
            num_of_videos_in_this_sample,
        )
    def __call__(
        self,
        images: ImageInput = None,
        text: TextInput | PreTokenizedInput | list[TextInput] | list[PreTokenizedInput] = None,
        audio=None,
        videos: VideoInput = None,
        **kwargs: Unpack[Eagle25VLProcessorKwargs],
    ) -> BatchFeature:
        """
        Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`
        and `kwargs` arguments to LlamaTokenizerFast's [`~LlamaTokenizerFast.__call__`] if `text` is not `None` to encode
        the text. To prepare the image(s), this method forwards the `images` and `kwrags` arguments to
        LlavaNextImageProcessor's [`~LlavaNextImageProcessor.__call__`] if `images` is not `None`. Please refer to the docstring
        of the above two methods for more information.
        Args:
            images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `List[PIL.Image.Image]`, `List[np.ndarray]`, `List[torch.Tensor]`):
                The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
                tensor. Both channels-first and channels-last formats are supported.
            text (`str`, `List[str]`, `List[List[str]]`):
                The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
                (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
                `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
            videos (`np.ndarray`, `torch.Tensor`, `List[np.ndarray]`, `List[torch.Tensor]`):
                The image or batch of videos to be prepared. Each video can be a 4D NumPy array or PyTorch
        Returns:
            [`BatchFeature`]: A [`BatchFeature`] with the following fields:
            - **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`.
            - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
              `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not
              `None`).
            - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
            - **pixel_values_videos** -- Pixel values of a video input to be fed to a model. Returned when `videos` is not `None`.
            - **image_sizes** -- Size of each image that will be used to unpad an image. Returned when `images` is not `None`.
        """
        output_kwargs = self._merge_kwargs(
            Eagle25VLProcessorKwargs,
            tokenizer_init_kwargs=self.tokenizer.init_kwargs,
            **kwargs,
        )
        if isinstance(text, str):
            text_list = [text]
        elif not isinstance(text, list) and not isinstance(text[0], str):
            raise ValueError("Invalid input text. Please provide a string, or a list of strings")
        elif isinstance(text, list) and isinstance(text[0], str):
            text_list = text
        if images is None:
            images = []
        if videos is None:
            videos = []
        pixel_values_list = []
        image_sizes_list = []
        new_sample_list = []
        image_start_idx = 0
        video_start_idx = 0
        timestamps_batch = output_kwargs["videos_kwargs"].pop("timestamps", None)
        fps_batch = output_kwargs["videos_kwargs"].pop("fps", None)
        for sample in text_list:
            timestamps_list = timestamps_batch[video_start_idx:] if timestamps_batch is not None else None
            fps_list = fps_batch[video_start_idx:] if fps_batch is not None else None
            (
                sample,
                pixel_values,
                image_sizes,
                num_of_images_in_this_sample,
                num_of_videos_in_this_sample,
            ) = self.replace_media_placeholder(
                sample,
                images[image_start_idx:],
                videos[video_start_idx:],
                timestamps_list,
                fps_list,
                **output_kwargs,
            )
            new_sample_list.append(sample)
            if pixel_values is not None:
                pixel_values_list.append(pixel_values)
                image_sizes_list.append(image_sizes)
            image_start_idx += num_of_images_in_this_sample
            video_start_idx += num_of_videos_in_this_sample
        if len(pixel_values_list) > 0:
            image_inputs = {
                "pixel_values": torch.cat(pixel_values_list),
                "image_sizes": torch.cat(image_sizes_list),
            }
        else:
            image_inputs = {}
        video_inputs = {}
        text_inputs = self.tokenizer(new_sample_list, **output_kwargs["text_kwargs"])
        return BatchFeature(data={**text_inputs, **image_inputs, **video_inputs})
    def get_number_tiles_based_on_image_size(
        self, image_size: tuple, min_num: int, max_num: int, use_thumbnail: bool, tile_size: int
    ) -> int:
        """
        Get the number of tiles based on the image size.
        """
        orig_height, orig_width = image_size
        aspect_ratio = orig_width / orig_height
        # calculate the existing image aspect ratio
        target_ratios = {
            (i, j)
            for n in range(min_num, max_num + 1)
            for i in range(1, n + 1)
            for j in range(1, n + 1)
            if i * j <= max_num and i * j >= min_num
        }
        target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
        # find the closest aspect ratio to the target
        target_aspect_ratio = self.image_processor.find_closest_aspect_ratio(
            aspect_ratio, target_ratios, orig_width, orig_height, tile_size
        )
        tiles_num = target_aspect_ratio[0] * target_aspect_ratio[1]
        if use_thumbnail and tiles_num > 1:
            tiles_num += 1
        return tiles_num
    # Copied from transformers.models.clip.processing_clip.CLIPProcessor.batch_decode with CLIP->Llama
    def batch_decode(self, *args, **kwargs):
        """
        This method forwards all its arguments to LlamaTokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please
        refer to the docstring of this method for more information.
        """
        return self.tokenizer.batch_decode(*args, **kwargs)
    # Copied from transformers.models.clip.processing_clip.CLIPProcessor.decode with CLIP->Llama
    def decode(self, *args, **kwargs):
        """
        This method forwards all its arguments to LlamaTokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer to
        the docstring of this method for more information.
        """
        return self.tokenizer.decode(*args, **kwargs)
    @property
    # Copied from transformers.models.clip.processing_clip.CLIPProcessor.model_input_names
    def model_input_names(self):
        tokenizer_input_names = self.tokenizer.model_input_names
        image_processor_input_names = self.image_processor.model_input_names
        return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))
    # override to save video-config in a separate config file
    def save_pretrained(self, save_directory, **kwargs):
        if os.path.isfile(save_directory):
            raise ValueError(f"Provided path ({save_directory}) should be a directory, not a file")
        os.makedirs(save_directory, exist_ok=True)
        outputs = super().save_pretrained(save_directory, **kwargs)
        return outputs
    # override to load video-config from a separate config file
    @classmethod
    def from_pretrained(cls, pretrained_model_name_or_path, **kwargs):
        processor = super().from_pretrained(pretrained_model_name_or_path, **kwargs)
        # if return_unused_kwargs a tuple is returned where the second element is 'unused_kwargs'
        if isinstance(processor, tuple):
            processor = processor[0]
        return processor
    # Copy from https://github.com/QwenLM/Qwen2.5-VL/blob/main/qwen-vl-utils/src/qwen_vl_utils/vision_process.py
    def process_vision_info(
        self,
        conversations: list[dict] | list[list[dict]],
        return_video_kwargs: bool = False,
    ) -> tuple[list[Image.Image] | None, list[torch.Tensor | list[Image.Image]] | None, dict | None]:
        vision_infos = self.extract_vision_info(conversations)
        ## Read images or videos
        image_inputs = []
        video_inputs = []
        video_sample_fps_list = []
        video_timestamps_list = []
        for vision_info in vision_infos:
            if "image" in vision_info or "image_url" in vision_info:
                image_inputs.append(fetch_image(vision_info))
            else:
                raise ValueError("image, image_url or video should in content.")
        if len(image_inputs) == 0:
            image_inputs = None
        if len(video_inputs) == 0:
            video_inputs = None
        if return_video_kwargs:
            return (
                image_inputs,
                video_inputs,
                {"fps": video_sample_fps_list, "timestamps": video_timestamps_list},
            )
        return image_inputs, video_inputs
    def extract_vision_info(self, conversations: list[dict] | list[list[dict]]) -> list[dict]:
        vision_infos = []
        if isinstance(conversations[0], dict):
            conversations = [conversations]
        for conversation in conversations:
            for message in conversation:
                if isinstance(message["content"], list):
                    for ele in message["content"]:
                        if (
                            "image" in ele
                            or "image_url" in ele
                            or "video" in ele
                            or ele["type"] in ("image", "image_url", "video")
                        ):
                            vision_infos.append(ele)
        return vision_infos
 __all__ = ["Eagle25VLProcessor"]
@@ -0,0 +1,380 @@
 # SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
 # http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from pathlib import Path
 from typing import TYPE_CHECKING, Any
 import numpy as np
 import torch
 import torch.nn as nn
 from huggingface_hub import snapshot_download
 from huggingface_hub.errors import HFValidationError, RepositoryNotFoundError
 from lerobot.utils.import_utils import _transformers_available
 # Conditional import for type checking and lazy loading
 if TYPE_CHECKING or _transformers_available:
    from huggingface_hub.dataclasses import strict
    from transformers import AutoConfig, AutoModel, PretrainedConfig, PreTrainedModel
    from transformers.feature_extraction_utils import BatchFeature
 else:
    def strict(cls):
        return cls
    AutoConfig = None
    AutoModel = None
    PretrainedConfig = object
    PreTrainedModel = object
    BatchFeature = None
 try:
    import tree
 except ImportError:
    tree = None
 from lerobot.utils.constants import ACTION, HF_LEROBOT_HOME
 from .action_head.flow_matching_action_head import (
    FlowmatchingActionHead,
    FlowmatchingActionHeadConfig,
 )
 from .utils import ensure_eagle_cache_ready
 DEFAULT_VENDOR_EAGLE_PATH = str((Path(__file__).resolve().parent / "eagle2_hg_model").resolve())
 DEFAULT_TOKENIZER_ASSETS_REPO = "lerobot/eagle2hg-processor-groot-n1p5"
 class EagleBackbone(nn.Module):
    def __init__(
        self,
        tune_llm: bool = False,
        tune_visual: bool = False,
        select_layer: int = -1,
        reproject_vision: bool = False,
        use_flash_attention: bool = False,
        load_bf16: bool = False,
        eagle_path: str = DEFAULT_VENDOR_EAGLE_PATH,
        tokenizer_assets_repo: str = DEFAULT_TOKENIZER_ASSETS_REPO,
        project_to_dim: int = 1536,
    ):
        """
        Args:
            tune_llm: whether to tune the LLM model (default: True)
            tune_visual: whether to tune the visual model (default: False)
        """
        super().__init__()
        assert not reproject_vision, "Reproject vision is not implemented here, set to False"
        # Prefer loading Eagle model config from the cache directory where vendor files were copied.
        vendor_dir = DEFAULT_VENDOR_EAGLE_PATH
        cache_dir = HF_LEROBOT_HOME / tokenizer_assets_repo
        try:
            ensure_eagle_cache_ready(vendor_dir, cache_dir, tokenizer_assets_repo)
        except Exception as exc:  # nosec: B110
            print(f"[GROOT] Warning: failed to prepare Eagle cache for backbone: {exc}")
        config = AutoConfig.from_pretrained(str(cache_dir), trust_remote_code=True)
        self.eagle_model = AutoModel.from_config(config, trust_remote_code=True)
        if project_to_dim is not None:
            self.eagle_linear = torch.nn.Linear(2048, project_to_dim)
        else:
            self.eagle_linear = torch.nn.Identity()
        # needed since we don't use these layers. Also saves compute
        while len(self.eagle_model.language_model.model.layers) > select_layer:
            self.eagle_model.language_model.model.layers.pop(-1)
        self.select_layer = select_layer
        self.set_trainable_parameters(tune_llm, tune_visual)
    def set_trainable_parameters(self, tune_llm: bool, tune_visual: bool):
        self.tune_llm = tune_llm
        self.tune_visual = tune_visual
        for p in self.parameters():
            p.requires_grad = True
        if not tune_llm:
            self.eagle_model.language_model.requires_grad_(False)
        if not tune_visual:
            self.eagle_model.vision_model.requires_grad_(False)
            self.eagle_model.mlp1.requires_grad_(False)
        print(f"Tune backbone llm: {self.tune_llm}")
        print(f"Tune backbone visual: {self.tune_visual}")
        # Check if any parameters are still trainable. If not, print a warning.
        if not tune_llm and not tune_visual:
            for name, p in self.named_parameters():
                if p.requires_grad:
                    print(f"Backbone trainable parameter: {name}")
        if not any(p.requires_grad for p in self.parameters()):
            print("Warning: No backbone trainable parameters found.")
    def set_frozen_modules_to_eval_mode(self):
        """
        Huggingface will call model.train() at each training_step. To ensure
        the expected behaviors for modules like dropout, batchnorm, etc., we
        need to call model.eval() for the frozen modules.
        """
        if self.training:
            if self.eagle_model.language_model and not self.tune_llm:
                self.eagle_model.language_model.eval()
            if self.eagle_model.vision_model and not self.tune_visual:
                self.eagle_model.vision_model.eval()
    def prepare_input(self, batch: dict) -> BatchFeature:
        return BatchFeature(data=batch)
    def forward_eagle(self, vl_input: BatchFeature) -> BatchFeature:
        eagle_prefix = "eagle_"
        eagle_input = {
            k.removeprefix(eagle_prefix): v for k, v in vl_input.items() if k.startswith(eagle_prefix)
        }
        del eagle_input["image_sizes"]
        eagle_output = self.eagle_model(**eagle_input, output_hidden_states=True, return_dict=True)
        eagle_features = eagle_output.hidden_states[self.select_layer]
        eagle_features = self.eagle_linear(eagle_features)
        return eagle_features, eagle_input["attention_mask"]
    def forward(self, vl_input: BatchFeature) -> BatchFeature:
        self.set_frozen_modules_to_eval_mode()
        eagle_embeds, eagle_mask = self.forward_eagle(vl_input)
        # YL (TODO HACK): to resolve DDP issue when tune_visual=True
        # Ensure all trainable parameters in vision_model are used in the forward pass for DDP compatibility
        if self.training and self.tune_visual:
            dummy_term = torch.tensor(
                0.0, device=eagle_embeds.device, dtype=eagle_embeds.dtype, requires_grad=True
            )
            for param in self.eagle_model.vision_model.parameters():
                if param.requires_grad:
                    dummy_term = dummy_term + 0.0 * param.sum()
            eagle_embeds = eagle_embeds + dummy_term
        return BatchFeature(
            data={"backbone_features": eagle_embeds, "backbone_attention_mask": eagle_mask}
        )  # [B, T2, hidden_size]
 BACKBONE_FEATURE_KEY = "backbone_features"
 ACTION_KEY = "action_pred"
 LOSS_KEY = "loss"
 ERROR_MSG = "Error: unexpected input/output"
 N_COLOR_CHANNELS = 3
 # config
@strict
 class GR00TN15Config(PretrainedConfig):
    model_type = "gr00t_n1_5"
    backbone_cfg: dict[str, Any] | None = None
    action_head_cfg: dict[str, Any] | None = None
    action_horizon: int = 0
    action_dim: int = 0
    compute_dtype: str = "float32"
    def __post_init__(self, **kwargs):
        self.backbone_cfg = {} if self.backbone_cfg is None else self.backbone_cfg
        self.action_head_cfg = {} if self.action_head_cfg is None else self.action_head_cfg
        super().__post_init__(**kwargs)
 # real model
 class GR00TN15(PreTrainedModel):
    supports_gradient_checkpointing = True
    config_class = GR00TN15Config
    """
    we expect the backbone output to have a key 'backbone_features' with shape (batch_size, n, hidden_size)
    here n is variable and can be e.g. time, 1 or user specified
    we expect the action head output to have a key 'action_pred' with shape (batch_size, time, action_dim) during inference time
    we expect these to have type BatchFeature, and they can of course have many other user specified keys too
    """
    def __init__(
        self,
        config: GR00TN15Config,
        local_model_path: str,
    ):
        assert isinstance(config.backbone_cfg, dict)
        assert isinstance(config.action_head_cfg, dict)
        super().__init__(config)
        self.local_model_path = local_model_path
        self.backbone = EagleBackbone(**config.backbone_cfg)
        action_head_cfg = FlowmatchingActionHeadConfig(**config.action_head_cfg)
        self.action_head = FlowmatchingActionHead(action_head_cfg)
        self.action_horizon = config.action_horizon
        self.action_dim = config.action_dim
        self.compute_dtype = config.compute_dtype
        self.post_init()
    def validate_inputs(self, inputs):
        # NOTE -- this should be handled internally by the model
        # however, doing that will likely be breaking changes -- so we'll need to do it after the deadline
        detected_error = False
        error_msg = ERROR_MSG
        if ACTION in inputs:
            action = inputs[ACTION]
            # In inference, action may be omitted or None; validate only when it's a tensor.
            if action is None:
                pass  # allow None during inference
            elif isinstance(action, torch.Tensor):
                shape_ok = (
                    len(action.shape) == 3
                    and action.shape[1] == self.action_horizon
                    and action.shape[2] == self.action_dim
                )
                if not shape_ok:
                    error_msg += f"\n{action.shape=}"
                    detected_error = True
            else:
                # Unexpected non-tensor type provided for action
                error_msg += f"\nInvalid type for action: {type(action)}"
                detected_error = True
        if "video" in inputs:
            video = inputs["video"]
            type_ok = isinstance(video, np.ndarray)
            dtype_ok = video.dtype == np.uint8
            shape_ok = len(video.shape) == 6 and video.shape[3] == N_COLOR_CHANNELS
            if not type_ok:
                error_msg += f"\n{type(video)=}"
                detected_error = True
            if not dtype_ok:
                error_msg += f"\n{video.dtype=}"
                detected_error = True
            if not shape_ok:
                error_msg += f"\n{video.shape=}"
                detected_error = True
        if detected_error:
            raise ValueError(error_msg)
    def validate_data(self, action_head_outputs, backbone_outputs, is_training):
        fail_backbone = (
            not isinstance(backbone_outputs, BatchFeature) or BACKBONE_FEATURE_KEY not in backbone_outputs
        )
        if fail_backbone:
            error_msg = ERROR_MSG
            error_msg += f"\n{isinstance(backbone_outputs, BatchFeature)=}"
            error_msg += f"\n{BACKBONE_FEATURE_KEY in backbone_outputs=}"
            error_msg += f"\n{backbone_outputs[BACKBONE_FEATURE_KEY].shape=}"
            raise ValueError(error_msg)
        fail_action_head = (not isinstance(action_head_outputs, BatchFeature)) or not (
            (
                LOSS_KEY in action_head_outputs and is_training
            )  # there might not be an action prediction during training
            or (
                ACTION_KEY in action_head_outputs
                and action_head_outputs[ACTION_KEY].shape[1] == self.action_horizon
                and action_head_outputs[ACTION_KEY].shape[2] == self.action_dim
            )
        )
        if fail_action_head:
            error_msg = ERROR_MSG
            error_msg += f"\n{isinstance(action_head_outputs, BatchFeature)=}"
            error_msg += f"\n{LOSS_KEY in action_head_outputs=}"
            error_msg += f"\n{action_head_outputs[ACTION_KEY].shape=}"
            error_msg += f"\n{self.action_horizon=}"
            error_msg += f"\n{self.action_dim=}"
            raise ValueError(error_msg)
    def forward(
        self,
        inputs: dict,
    ) -> BatchFeature:
        backbone_inputs, action_inputs = self.prepare_input(inputs)
        backbone_outputs = self.backbone(backbone_inputs)
        action_head_outputs = self.action_head(backbone_outputs, action_inputs)
        self.validate_data(action_head_outputs, backbone_outputs, is_training=True)
        return action_head_outputs
    def get_action(
        self,
        inputs: dict,
    ) -> BatchFeature:
        backbone_inputs, action_inputs = self.prepare_input(inputs)
        # Because the behavior of backbones remains the same for training and inference, we can use `forward` for backbones.
        backbone_outputs = self.backbone(backbone_inputs)
        action_head_outputs = self.action_head.get_action(backbone_outputs, action_inputs)
        self.validate_data(action_head_outputs, backbone_outputs, is_training=False)
        return action_head_outputs
    def prepare_input(self, inputs) -> tuple[BatchFeature, BatchFeature]:
        self.validate_inputs(inputs)
        backbone_inputs = self.backbone.prepare_input(inputs)
        action_inputs = self.action_head.prepare_input(inputs)
        def to_device_with_maybe_dtype(x):
            # Cast floating tensors to a memory-efficient compute dtype when requested.
            # Rationale: Upcasting backbone activations to fp32 significantly increases VRAM.
            # When compute_dtype is bfloat16, prefer bf16 for activations to match AMP behavior.
            if not isinstance(x, torch.Tensor):
                return x
            if torch.is_floating_point(x):
                if getattr(self, "compute_dtype", None) == "bfloat16":
                    return x.to(self.device, dtype=torch.bfloat16)
                # Fallback: preserve previous behavior if not using bf16 compute
                return x.to(self.device, dtype=self.action_head.dtype)
            # Non-floating tensors: move device only
            return x.to(self.device)
        backbone_inputs = tree.map_structure(to_device_with_maybe_dtype, backbone_inputs)
        action_inputs = tree.map_structure(to_device_with_maybe_dtype, action_inputs)
        return backbone_inputs, action_inputs
    @classmethod
    def from_pretrained(cls, pretrained_model_name_or_path: str, **kwargs):
        tune_visual = kwargs.pop("tune_visual", True)
        tune_llm = kwargs.pop("tune_llm", False)
        tune_projector = kwargs.pop("tune_projector", True)
        tune_diffusion_model = kwargs.pop("tune_diffusion_model", True)
        print(f"Loading pretrained dual brain from {pretrained_model_name_or_path}")
        print(f"Tune backbone vision tower: {tune_visual}")
        print(f"Tune backbone LLM: {tune_llm}")
        print(f"Tune action head projector: {tune_projector}")
        print(f"Tune action head DiT: {tune_diffusion_model}")
        # get the current model path being downloaded
        try:
            # NOTE(YL) This downloads the model to the local cache and returns the local path to the model
            # saved in ~/.cache/huggingface/hub/
            local_model_path = snapshot_download(pretrained_model_name_or_path, repo_type="model")
            # HFValidationError, RepositoryNotFoundError
        except (HFValidationError, RepositoryNotFoundError):
            print(
                f"Model not found or avail in the huggingface hub. Loading from local path: {pretrained_model_name_or_path}"
            )
            local_model_path = pretrained_model_name_or_path
        pretrained_model = super().from_pretrained(
            local_model_path, local_model_path=local_model_path, **kwargs
        )
        pretrained_model.backbone.set_trainable_parameters(tune_visual=tune_visual, tune_llm=tune_llm)
        pretrained_model.action_head.set_trainable_parameters(
            tune_projector=tune_projector, tune_diffusion_model=tune_diffusion_model
        )
        return pretrained_model
@@ -1,951 +0,0 @@
 #!/usr/bin/env python
 # Copyright 2026 NVIDIA Corporation and The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
 #     http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from __future__ import annotations
 import logging
 from contextlib import suppress
 from copy import deepcopy
 from typing import TYPE_CHECKING, Any
 import torch
 import torch.nn.functional as F  # noqa: N812
 from huggingface_hub import snapshot_download
 from huggingface_hub.errors import HFValidationError, RepositoryNotFoundError
 from torch import nn
 from torch.distributions import Beta
 from lerobot.utils.import_utils import _transformers_available, require_package
 from .action_head.cross_attention_dit import AlternateVLDiT, DiT, SelfAttentionTransformer
 from .configuration_groot import N1_7_DEFAULT_IMAGE_CROP_SIZE, N1_7_DEFAULT_IMAGE_TARGET_SIZE
 if TYPE_CHECKING or _transformers_available:
    from transformers import (
        AutoConfig,
        AutoModel,
        PretrainedConfig,
        PreTrainedModel,
        Qwen3VLConfig,
        Qwen3VLForConditionalGeneration,
    )
    from transformers.feature_extraction_utils import BatchFeature
 else:
    AutoConfig = None
    AutoModel = None
    PretrainedConfig = object
    PreTrainedModel = object
    BatchFeature = None
    Qwen3VLConfig = None
    Qwen3VLForConditionalGeneration = None
 try:
    import tree
 except ImportError:
    tree = None
 logger = logging.getLogger(__name__)
 def _tie_unused_qwen_lm_head(model: nn.Module) -> None:
    """Restore the TF4 weight tie so the unused LM head stays frozen and is omitted on save."""
    lm_head = getattr(model, "lm_head", None)
    get_input_embeddings = getattr(model, "get_input_embeddings", None)
    if lm_head is None or not callable(get_input_embeddings):
        return
    input_embeddings = get_input_embeddings()
    embedding_weight = getattr(input_embeddings, "weight", None)
    if embedding_weight is None:
        return
    lm_head.weight = embedding_weight
 GR00T_N1_7_DEFAULTS: dict[str, Any] = {
    "model_dtype": "bfloat16",
    "dtype": "bfloat16",
    "model_name": "nvidia/Cosmos-Reason2-2B",
    "backbone_model_type": "qwen",
    "model_revision": None,
    "tune_top_llm_layers": 0,
    "backbone_embedding_dim": 2048,
    "tune_llm": False,
    "tune_visual": False,
    "select_layer": 16,
    "reproject_vision": False,
    "use_flash_attention": False,
    "load_bf16": False,
    "backbone_trainable_params_fp32": True,
    "image_crop_size": N1_7_DEFAULT_IMAGE_CROP_SIZE,
    "image_target_size": N1_7_DEFAULT_IMAGE_TARGET_SIZE,
    "shortest_image_edge": None,
    "crop_fraction": None,
    "random_rotation_angle": None,
    "color_jitter_params": None,
    "use_albumentations_transforms": True,
    "extra_augmentation_config": None,
    "formalize_language": True,
    "apply_sincos_state_encoding": False,
    "use_percentiles": True,
    "use_relative_action": False,
    "max_state_dim": 132,
    "max_action_dim": 132,
    "action_horizon": 40,
    "hidden_size": 1024,
    "input_embedding_dim": 1536,
    "state_history_length": 1,
    "add_pos_embed": True,
    "attn_dropout": 0.2,
    "use_vlln": True,
    "max_seq_len": 1024,
    "use_alternate_vl_dit": True,
    "attend_text_every_n_blocks": 2,
    "diffusion_model_cfg": {
        "positional_embeddings": None,
        "num_layers": 32,
        "num_attention_heads": 32,
        "attention_head_dim": 48,
        "norm_type": "ada_norm",
        "dropout": 0.2,
        "final_dropout": True,
        "output_dim": 1024,
        "interleave_self_attention": True,
    },
    "vl_self_attention_cfg": {
        "positional_embeddings": None,
        "num_layers": 4,
        "num_attention_heads": 32,
        "attention_head_dim": 64,
        "dropout": 0.2,
        "final_dropout": True,
    },
    "num_inference_timesteps": 4,
    "noise_beta_alpha": 1.5,
    "noise_beta_beta": 1.0,
    "noise_s": 0.999,
    "num_timestep_buckets": 1000,
    "tune_projector": True,
    "tune_diffusion_model": True,
    "tune_vlln": True,
    "state_dropout_prob": 0.2,
    "exclude_state": False,
    "use_mean_std": False,
    "max_num_embodiments": 32,
    "rtc_ramp_rate": 6.0,
 }
 class GR00TN17Config(PretrainedConfig):
    """Configuration for NVIDIA GR00T N1.7.
    N1.7 uses the Cosmos-Reason2-2B / Qwen3-VL backbone and a multi-embodiment
    flow-matching action head. This mirrors the public N1.7 checkpoint config
    while keeping it local to LeRobot and independent from the external
    Isaac-GR00T ``gr00t`` Python package.
    """
    model_type = "Gr00tN1d7"
    _defaults = GR00T_N1_7_DEFAULTS
    def __init__(self, **kwargs):
        super().__init__(**kwargs)
        for key, value in GR00T_N1_7_DEFAULTS.items():
            setattr(self, key, deepcopy(kwargs.pop(key, value)))
        for key, value in kwargs.items():
            setattr(self, key, value)
 class CategorySpecificLinear(nn.Module):
    """Linear layer with category-specific weights for multi-embodiment support."""
    def __init__(self, num_categories: int, input_dim: int, hidden_dim: int):
        super().__init__()
        self.num_categories = num_categories
        self.W = nn.Parameter(0.02 * torch.randn(num_categories, input_dim, hidden_dim))
        self.b = nn.Parameter(torch.zeros(num_categories, hidden_dim))
    def forward(self, x: torch.Tensor, cat_ids: torch.Tensor) -> torch.Tensor:
        selected_w = self.W[cat_ids]
        selected_b = self.b[cat_ids]
        return torch.bmm(x, selected_w) + selected_b.unsqueeze(1)
 class CategorySpecificMLP(nn.Module):
    """Two-layer MLP with category-specific weights."""
    def __init__(self, num_categories: int, input_dim: int, hidden_dim: int, output_dim: int):
        super().__init__()
        self.layer1 = CategorySpecificLinear(num_categories, input_dim, hidden_dim)
        self.layer2 = CategorySpecificLinear(num_categories, hidden_dim, output_dim)
    def forward(self, x: torch.Tensor, cat_ids: torch.Tensor) -> torch.Tensor:
        hidden = F.relu(self.layer1(x, cat_ids))
        return self.layer2(hidden, cat_ids)
 class SinusoidalPositionalEncoding(nn.Module):
    """Sinusoidal encoding of shape ``(B, T, D)`` for timestep tensors ``(B, T)``.
    The frequency scalar is intentionally created on CPU and then broadcast with
    the device-local arange result. That mirrors Isaac-GR00T's N1.7 timestep
    embedding and avoids tiny dtype/device construction differences in parity
    tests.
    """
    def __init__(self, embedding_dim: int):
        super().__init__()
        self.embedding_dim = embedding_dim
    def forward(self, timesteps: torch.Tensor) -> torch.Tensor:
        timesteps = timesteps.float()
        half_dim = self.embedding_dim // 2
        exponent = -torch.arange(half_dim, dtype=torch.float, device=timesteps.device) * (
            torch.log(torch.tensor(10000.0)) / half_dim
        )
        freqs = timesteps.unsqueeze(-1) * exponent.exp()
        return torch.cat([torch.sin(freqs), torch.cos(freqs)], dim=-1)
 def swish(x: torch.Tensor) -> torch.Tensor:
    return x * torch.sigmoid(x)
 class MultiEmbodimentActionEncoder(nn.Module):
    """Action encoder with category-specific projections and sinusoidal time encoding."""
    def __init__(self, action_dim: int, hidden_size: int, num_embodiments: int):
        super().__init__()
        self.W1 = CategorySpecificLinear(num_embodiments, action_dim, hidden_size)
        self.W2 = CategorySpecificLinear(num_embodiments, 2 * hidden_size, hidden_size)
        self.W3 = CategorySpecificLinear(num_embodiments, hidden_size, hidden_size)
        self.pos_encoding = SinusoidalPositionalEncoding(hidden_size)
    def forward(self, actions: torch.Tensor, timesteps: torch.Tensor, cat_ids: torch.Tensor) -> torch.Tensor:
        batch_size, horizon, _ = actions.shape
        if timesteps.dim() != 1 or timesteps.shape[0] != batch_size:
            raise ValueError("Expected `timesteps` to have shape (B,).")
        timesteps = timesteps.unsqueeze(1).expand(-1, horizon)
        action_emb = self.W1(actions, cat_ids)
        time_emb = self.pos_encoding(timesteps).to(dtype=action_emb.dtype)
        x = swish(self.W2(torch.cat([action_emb, time_emb], dim=-1), cat_ids))
        return self.W3(x, cat_ids)
 class Qwen3Backbone(nn.Module):
    """Cosmos-Reason2/Qwen3-VL backbone used by GR00T N1.7.
    The public checkpoint stores the action head in the GR00T checkpoint but
    uses a Hugging Face Qwen3-VL-compatible backbone interface. This wrapper
    keeps the nested HF module layout compatible across transformer versions
    and exposes the hidden states consumed by the action head.
    """
    def __init__(
        self,
        model_name: str = "nvidia/Cosmos-Reason2-2B",
        tune_llm: bool = False,
        tune_visual: bool = False,
        select_layer: int = -1,
        reproject_vision: bool = False,
        use_flash_attention: bool = False,
        load_bf16: bool = False,
        tune_top_llm_layers: int = 0,
        trainable_params_fp32: bool = False,
        transformers_loading_kwargs: dict[str, Any] | None = None,
        load_pretrained_weights: bool = True,
    ):
        require_package("transformers", extra="groot")
        if Qwen3VLForConditionalGeneration is None:
            raise ImportError(
                "Qwen3VLForConditionalGeneration is required for GR00T N1.7. "
                "Install a transformers version with Qwen3-VL support."
            )
        super().__init__()
        transformers_loading_kwargs = transformers_loading_kwargs or {"trust_remote_code": True}
        extra_kwargs: dict[str, Any] = {}
        if use_flash_attention:
            try:
                import flash_attn  # noqa: F401
                extra_kwargs["attn_implementation"] = "flash_attention_2"
            except ImportError:
                logger.warning("flash_attn is not installed. Falling back to SDPA attention.")
                extra_kwargs["attn_implementation"] = "sdpa"
        if load_bf16:
            extra_kwargs["torch_dtype"] = torch.bfloat16
        if load_pretrained_weights:
            self.model = Qwen3VLForConditionalGeneration.from_pretrained(
                model_name,
                **extra_kwargs,
                **transformers_loading_kwargs,
            ).eval()
        else:
            self.model = self._from_backbone_config(
                model_name=model_name,
                model_kwargs=extra_kwargs,
                config_kwargs=transformers_loading_kwargs,
            ).eval()
        _tie_unused_qwen_lm_head(self.model)
        while len(self.language_model.layers) > select_layer:
            self.language_model.layers.pop(-1)
        self.select_layer = select_layer
        self.set_trainable_parameters(tune_llm, tune_visual, tune_top_llm_layers)
        if load_bf16 and trainable_params_fp32:
            for parameter in self.parameters():
                if parameter.requires_grad:
                    parameter.data = parameter.data.to(torch.float32)
    def set_trainable_parameters(
        self, tune_llm: bool, tune_visual: bool, tune_top_llm_layers: int = 0
    ) -> None:
        self.tune_llm = tune_llm
        self.tune_visual = tune_visual
        for parameter in self.parameters():
            parameter.requires_grad = True
        if not tune_llm:
            self.language_model.requires_grad_(False)
        if not tune_visual:
            self.visual.requires_grad_(False)
        if tune_top_llm_layers > 0:
            for layer in self.language_model.layers[-tune_top_llm_layers:]:
                for parameter in layer.parameters():
                    parameter.requires_grad = True
    def set_frozen_modules_to_eval_mode(self) -> None:
        if self.training:
            if self.language_model and not self.tune_llm:
                self.language_model.eval()
            if self.visual and not self.tune_visual:
                self.visual.eval()
    @property
    def language_model(self) -> nn.Module:
        return getattr(self.model, "model", self.model).language_model
    @property
    def visual(self) -> nn.Module:
        return getattr(self.model, "model", self.model).visual
    def _from_backbone_config(
        self,
        *,
        model_name: str,
        model_kwargs: dict[str, Any],
        config_kwargs: dict[str, Any],
    ) -> nn.Module:
        if _is_cosmos_reason2_backbone(model_name):
            backbone_config = _cosmos_reason2_qwen3_vl_config()
        else:
            backbone_config = AutoConfig.from_pretrained(model_name, **config_kwargs)
        return Qwen3VLForConditionalGeneration._from_config(backbone_config, **model_kwargs)
    def prepare_input(self, batch: dict[str, Any]) -> BatchFeature:
        return BatchFeature(data=batch)
    def _ensure_mm_token_type_ids(self, model_input: dict[str, torch.Tensor]) -> None:
        if "mm_token_type_ids" in model_input:
            return
        if "image_grid_thw" not in model_input and "video_grid_thw" not in model_input:
            return
        input_ids = model_input.get("input_ids")
        if input_ids is None:
            return
        mm_token_type_ids = torch.zeros(input_ids.shape, dtype=torch.int32, device=input_ids.device)
        image_token_id = getattr(self.model.config, "image_token_id", None)
        video_token_id = getattr(self.model.config, "video_token_id", None)
        if image_token_id is not None:
            mm_token_type_ids[input_ids == image_token_id] = 1
        if video_token_id is not None:
            mm_token_type_ids[input_ids == video_token_id] = 2
        model_input["mm_token_type_ids"] = mm_token_type_ids
    def _ensure_legacy_qwen3_position_ids(self, model_input: dict[str, torch.Tensor]) -> None:
        """Restore the Qwen3-VL text position ids used by older Transformers releases.
        Transformers 5.x computes 3-row multimodal RoPE ids for Qwen3-VL and then
        drops text position ids before calling text-layer flash attention. GR00T
        N1.7 was aligned against the older Transformers path, where a fourth text
        position row is forwarded alongside the temporal/height/width rows. Adding
        the row here preserves the newer multimodal position computation while
        keeping flash attention on the legacy code path.
        """
        if "position_ids" in model_input:
            return
        qwen3_model = getattr(self.model, "model", self.model)
        compute_3d_position_ids = getattr(qwen3_model, "compute_3d_position_ids", None)
        if compute_3d_position_ids is None:
            return
        position_ids = compute_3d_position_ids(
            input_ids=model_input.get("input_ids"),
            image_grid_thw=model_input.get("image_grid_thw"),
            video_grid_thw=model_input.get("video_grid_thw"),
            inputs_embeds=None,
            attention_mask=model_input.get("attention_mask"),
            past_key_values=None,
            mm_token_type_ids=model_input.get("mm_token_type_ids"),
        )
        if position_ids.ndim == 3 and position_ids.shape[0] == 3:
            position_ids = torch.cat([position_ids[:1], position_ids], dim=0)
        model_input["position_ids"] = position_ids
    def _last_decoder_layer_output(self, model_input: dict[str, torch.Tensor]) -> torch.Tensor:
        """Return the pre-final-norm decoder output consumed by the N1.7 action head.
        Older Transformers releases exposed this tensor as ``hidden_states[-1]``.
        Newer releases expose the post-final-norm tensor there instead. Capturing
        the last decoder layer output directly keeps the N1.7 action head input
        stable across Transformers versions.
        """
        captured: dict[str, torch.Tensor] = {}
        def capture_output(_module: nn.Module, _inputs: tuple[Any, ...], output: Any) -> None:
            if isinstance(output, torch.Tensor):
                captured["features"] = output
            elif isinstance(output, (tuple, list)) and output:
                captured["features"] = output[0]
            elif hasattr(output, "last_hidden_state"):
                captured["features"] = output.last_hidden_state
        hook = self.language_model.layers[-1].register_forward_hook(capture_output)
        try:
            outputs = self.model(**model_input, output_hidden_states=True)
        finally:
            hook.remove()
        return captured.get("features", outputs.hidden_states[-1])
    def forward(self, vl_input: BatchFeature) -> BatchFeature:
        self.set_frozen_modules_to_eval_mode()
        keys_to_use = ["input_ids", "attention_mask", "pixel_values", "image_grid_thw"]
        optional_keys = ["mm_token_type_ids", "pixel_values_videos", "video_grid_thw"]
        model_input = {key: vl_input[key] for key in keys_to_use}
        model_input.update({key: vl_input[key] for key in optional_keys if key in vl_input})
        self._ensure_mm_token_type_ids(model_input)
        self._ensure_legacy_qwen3_position_ids(model_input)
        features = self._last_decoder_layer_output(model_input)
        image_mask = model_input["input_ids"] == self.model.config.image_token_id
        attention_mask = model_input["attention_mask"] == 1
        return BatchFeature(
            data={
                "backbone_features": features,
                "backbone_attention_mask": attention_mask,
                "image_mask": image_mask,
            }
        )
 class GR00TN17ActionHead(nn.Module):
    supports_gradient_checkpointing = True
    def __init__(self, config: GR00TN17Config):
        require_package("diffusers", extra="groot")
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.input_embedding_dim = config.input_embedding_dim
        if config.use_alternate_vl_dit:
            self.model = AlternateVLDiT(
                **config.diffusion_model_cfg,
                cross_attention_dim=config.backbone_embedding_dim,
                attend_text_every_n_blocks=config.attend_text_every_n_blocks,
            )
        else:
            self.model = DiT(
                **config.diffusion_model_cfg,
                cross_attention_dim=config.backbone_embedding_dim,
            )
        self.action_dim = config.max_action_dim
        self.action_horizon = config.action_horizon
        self.num_inference_timesteps = config.num_inference_timesteps
        self.state_encoder = CategorySpecificMLP(
            num_categories=config.max_num_embodiments,
            input_dim=config.max_state_dim * config.state_history_length,
            hidden_dim=self.hidden_size,
            output_dim=self.input_embedding_dim,
        )
        self.action_encoder = MultiEmbodimentActionEncoder(
            action_dim=self.action_dim,
            hidden_size=self.input_embedding_dim,
            num_embodiments=config.max_num_embodiments,
        )
        self.action_decoder = CategorySpecificMLP(
            num_categories=config.max_num_embodiments,
            input_dim=self.hidden_size,
            hidden_dim=self.hidden_size,
            output_dim=self.action_dim,
        )
        self.vlln = nn.LayerNorm(config.backbone_embedding_dim) if config.use_vlln else nn.Identity()
        vl_self_attention_cfg = getattr(config, "vl_self_attention_cfg", None)
        if vl_self_attention_cfg and vl_self_attention_cfg.get("num_layers", 0) > 0:
            self.vl_self_attention = SelfAttentionTransformer(**vl_self_attention_cfg)
        else:
            self.vl_self_attention = nn.Identity()
        if config.add_pos_embed:
            self.position_embedding = nn.Embedding(config.max_seq_len, self.input_embedding_dim)
            nn.init.normal_(self.position_embedding.weight, mean=0.0, std=0.02)
        self.state_dropout_prob = config.state_dropout_prob
        self._noise_beta_alpha = config.noise_beta_alpha
        self._noise_beta_beta = config.noise_beta_beta
        self._beta_dist = None
        self.num_timestep_buckets = config.num_timestep_buckets
        self.set_trainable_parameters(config.tune_projector, config.tune_diffusion_model, config.tune_vlln)
    def set_trainable_parameters(
        self, tune_projector: bool, tune_diffusion_model: bool, tune_vlln: bool
    ) -> None:
        self.tune_projector = tune_projector
        self.tune_diffusion_model = tune_diffusion_model
        self.tune_vlln = tune_vlln
        for parameter in self.parameters():
            parameter.requires_grad = True
        if not tune_projector:
            self.state_encoder.requires_grad_(False)
            self.action_encoder.requires_grad_(False)
            self.action_decoder.requires_grad_(False)
            if self.config.add_pos_embed:
                self.position_embedding.requires_grad_(False)
        if not tune_diffusion_model:
            self.model.requires_grad_(False)
        if not tune_vlln:
            self.vlln.requires_grad_(False)
            self.vl_self_attention.requires_grad_(False)
    def set_frozen_modules_to_eval_mode(self) -> None:
        if self.training:
            if not self.tune_projector:
                self.state_encoder.eval()
                self.action_encoder.eval()
                self.action_decoder.eval()
                if self.config.add_pos_embed:
                    self.position_embedding.eval()
            if not self.tune_diffusion_model:
                self.model.eval()
            if not self.tune_vlln:
                self.vlln.eval()
                self.vl_self_attention.eval()
    def sample_time(self, batch_size: int, device: torch.device, dtype: torch.dtype) -> torch.Tensor:
        if self._beta_dist is None:
            beta_alpha = torch.tensor(self._noise_beta_alpha, device="cpu", dtype=torch.float32)
            beta_beta = torch.tensor(self._noise_beta_beta, device="cpu", dtype=torch.float32)
            self._beta_dist = Beta(beta_alpha, beta_beta, validate_args=False)
        sample = self._beta_dist.sample([batch_size]).to(device, dtype=dtype)
        return (1 - sample) * self.config.noise_s
    def process_backbone_output(self, backbone_output: BatchFeature) -> BatchFeature:
        backbone_features = self.vlln(backbone_output["backbone_features"])
        backbone_output["backbone_features"] = self.vl_self_attention(backbone_features)
        return backbone_output
    def forward(self, backbone_output: BatchFeature, action_input: BatchFeature) -> BatchFeature:
        self.set_frozen_modules_to_eval_mode()
        backbone_output = self.process_backbone_output(backbone_output)
        vl_embeds = backbone_output.backbone_features
        device = vl_embeds.device
        embodiment_id = action_input.embodiment_id
        if action_input.state.shape[1] != self.config.state_history_length:
            raise ValueError("state history length does not match GR00T N1.7 config.")
        state = action_input.state.view(action_input.state.shape[0], 1, -1)
        state_features = self.state_encoder(state, embodiment_id)
        if self.training and self.state_dropout_prob > 0:
            do_dropout = (
                torch.rand(state_features.shape[0], device=state_features.device) < self.state_dropout_prob
            )
            state_features = state_features * (1 - do_dropout[:, None, None].to(dtype=state_features.dtype))
        actions = action_input.action
        noise = torch.randn(actions.shape, device=actions.device, dtype=actions.dtype)
        t = self.sample_time(actions.shape[0], device=actions.device, dtype=actions.dtype)
        t = t[:, None, None]
        noisy_trajectory = (1 - t) * noise + t * actions
        velocity = actions - noise
        t_discretized = (t[:, 0, 0] * self.num_timestep_buckets).long()
        action_features = self.action_encoder(noisy_trajectory, t_discretized, embodiment_id)
        if self.config.add_pos_embed:
            pos_ids = torch.arange(action_features.shape[1], dtype=torch.long, device=device)
            action_features = action_features + self.position_embedding(pos_ids).unsqueeze(0)
        sa_embs = torch.cat((state_features, action_features), dim=1)
        if self.config.use_alternate_vl_dit:
            model_output, _ = self.model(
                hidden_states=sa_embs,
                encoder_hidden_states=vl_embeds,
                encoder_attention_mask=backbone_output.backbone_attention_mask,
                timestep=t_discretized,
                return_all_hidden_states=True,
                image_mask=backbone_output.image_mask,
                backbone_attention_mask=backbone_output.backbone_attention_mask,
            )
        else:
            model_output, _ = self.model(
                hidden_states=sa_embs,
                encoder_hidden_states=vl_embeds,
                encoder_attention_mask=backbone_output.backbone_attention_mask,
                timestep=t_discretized,
                return_all_hidden_states=True,
            )
        pred = self.action_decoder(model_output, embodiment_id)
        pred_actions = pred[:, -actions.shape[1] :]
        action_mask = action_input.action_mask
        action_loss = F.mse_loss(pred_actions, velocity, reduction="none") * action_mask
        loss = action_loss.sum() / (action_mask.sum() + 1e-6)
        return BatchFeature(
            data={
                "loss": loss,
                "action_loss": action_loss,
                "action_mask": action_mask,
                "backbone_features": vl_embeds,
                "state_features": state_features,
            }
        )
    def _encode_features(self, backbone_output: BatchFeature, action_input: BatchFeature) -> BatchFeature:
        backbone_output = self.process_backbone_output(backbone_output)
        state = action_input.state
        if state.shape[1] != self.config.state_history_length:
            raise ValueError("state history length does not match GR00T N1.7 config.")
        state = state.view(state.shape[0], 1, -1)
        state_features = self.state_encoder(state, action_input.embodiment_id)
        return BatchFeature(
            data={"backbone_features": backbone_output.backbone_features, "state_features": state_features}
        )
    @torch.no_grad()
    def get_action_with_features(
        self,
        backbone_features: torch.Tensor,
        state_features: torch.Tensor,
        embodiment_id: torch.Tensor,
        backbone_output: BatchFeature,
        action_input: BatchFeature,
        options: dict[str, Any] | None = None,
    ) -> BatchFeature:
        vl_embeds = backbone_features
        batch_size = vl_embeds.shape[0]
        device = vl_embeds.device
        actions = torch.randn(
            size=(batch_size, self.config.action_horizon, self.action_dim),
            dtype=vl_embeds.dtype,
            device=device,
        )
        dt = 1.0 / self.num_inference_timesteps
        vel_strength = torch.ones_like(actions)
        if "action" in action_input:
            if options is None:
                raise ValueError("RTC options are required when action is provided to get_action.")
            action_horizon_before_padding = options["action_horizon"]
            actions[:, : options["rtc_overlap_steps"], :] = action_input["action"][
                :,
                action_horizon_before_padding - options["rtc_overlap_steps"] : action_horizon_before_padding,
                :,
            ]
            vel_strength[:, : options["rtc_frozen_steps"], :] = 0.0
            intermediate_steps = options["rtc_overlap_steps"] - options["rtc_frozen_steps"]
            t = torch.linspace(0.0, 1.0, intermediate_steps + 2, device=device)
            ramp = 1 - torch.exp(-options["rtc_ramp_rate"] * t)
            ramp = ramp / ramp[-1].clamp_min(1e-8)
            vel_strength[:, options["rtc_frozen_steps"] : options["rtc_overlap_steps"], :] = ramp[1:-1][
                None, :, None
            ].to(device)
        for t_step in range(self.num_inference_timesteps):
            t_cont = t_step / float(self.num_inference_timesteps)
            t_discretized = int(t_cont * self.num_timestep_buckets)
            timesteps_tensor = torch.full(size=(batch_size,), fill_value=t_discretized, device=device)
            action_features = self.action_encoder(actions, timesteps_tensor, embodiment_id)
            if self.config.add_pos_embed:
                pos_ids = torch.arange(action_features.shape[1], dtype=torch.long, device=device)
                action_features = action_features + self.position_embedding(pos_ids).unsqueeze(0)
            sa_embs = torch.cat((state_features, action_features), dim=1)
            if self.config.use_alternate_vl_dit:
                model_output = self.model(
                    hidden_states=sa_embs,
                    encoder_hidden_states=vl_embeds,
                    timestep=timesteps_tensor,
                    image_mask=backbone_output.image_mask,
                    backbone_attention_mask=backbone_output.backbone_attention_mask,
                )
            else:
                model_output = self.model(
                    hidden_states=sa_embs,
                    encoder_hidden_states=vl_embeds,
                    timestep=timesteps_tensor,
                )
            pred = self.action_decoder(model_output, embodiment_id)
            actions = actions + dt * pred[:, -self.action_horizon :] * vel_strength
        return BatchFeature(
            data={
                "action_pred": actions,
                "backbone_features": vl_embeds,
                "state_features": state_features,
            }
        )
    @torch.no_grad()
    def get_action(
        self,
        backbone_output: BatchFeature,
        action_input: BatchFeature,
        options: dict[str, Any] | None = None,
    ) -> BatchFeature:
        features = self._encode_features(backbone_output, action_input)
        return self.get_action_with_features(
            backbone_features=features.backbone_features,
            state_features=features.state_features,
            embodiment_id=action_input.embodiment_id,
            backbone_output=backbone_output,
            action_input=action_input,
            options=options,
        )
    @property
    def device(self) -> torch.device:
        return next(iter(self.parameters())).device
    @property
    def dtype(self) -> torch.dtype:
        return next(iter(self.parameters())).dtype
    def prepare_input(self, batch: dict[str, Any]) -> BatchFeature:
        return BatchFeature(data=batch)
 def _is_cosmos_reason2_backbone(model_name: str) -> bool:
    return str(model_name).rstrip("/") == "nvidia/Cosmos-Reason2-2B"
 def _cosmos_reason2_qwen3_vl_config() -> PretrainedConfig:
    """Hard-coded copy of the nvidia/Cosmos-Reason2-2B config.json (a Qwen3-VL-2B-Instruct layout)."""
    return Qwen3VLConfig(
        image_token_id=151655,
        video_token_id=151656,
        vision_start_token_id=151652,
        vision_end_token_id=151653,
        tie_word_embeddings=True,
        text_config={
            "attention_bias": False,
            "attention_dropout": 0.0,
            "bos_token_id": 151643,
            "dtype": "bfloat16",
            "eos_token_id": 151645,
            "head_dim": 128,
            "hidden_act": "silu",
            "hidden_size": 2048,
            "initializer_range": 0.02,
            "intermediate_size": 6144,
            "max_position_embeddings": 262144,
            "model_type": "qwen3_vl_text",
            "num_attention_heads": 16,
            "num_hidden_layers": 28,
            "num_key_value_heads": 8,
            "rms_norm_eps": 1e-6,
            "rope_scaling": {
                "mrope_interleaved": True,
                "mrope_section": [24, 20, 20],
                "rope_type": "default",
            },
            "rope_theta": 5000000,
            "tie_word_embeddings": True,
            "use_cache": True,
            "vocab_size": 151936,
        },
        vision_config={
            "deepstack_visual_indexes": [5, 11, 17],
            "depth": 24,
            "hidden_act": "gelu_pytorch_tanh",
            "hidden_size": 1024,
            "in_channels": 3,
            "initializer_range": 0.02,
            "intermediate_size": 4096,
            "model_type": "qwen3_vl",
            "num_heads": 16,
            "num_position_embeddings": 2304,
            "out_hidden_size": 2048,
            "patch_size": 16,
            "spatial_merge_size": 2,
            "temporal_patch_size": 2,
        },
    )
 def get_backbone_cls(config: GR00TN17Config):
    if "nvidia/Cosmos-Reason2" in config.model_name or "Qwen/Qwen3-VL" in config.model_name:
        return Qwen3Backbone
    if config.backbone_model_type == "qwen":
        logger.warning(
            "Unrecognized GR00T N1.7 backbone model name '%s'; assuming a Qwen3-VL-compatible "
            "backbone because backbone_model_type='qwen'.",
            config.model_name,
        )
        return Qwen3Backbone
    raise ValueError(f"Unsupported GR00T N1.7 backbone model: {config.model_name}")
 class GR00TN17(PreTrainedModel):
    """GR00T N1.7 model with a Cosmos-Reason2/Qwen3-VL backbone."""
    config_class = GR00TN17Config
    supports_gradient_checkpointing = True
    def __init__(
        self,
        config: GR00TN17Config,
        transformers_loading_kwargs: dict[str, Any] | None = None,
        load_backbone_weights: bool = True,
    ):
        _register_with_transformers()
        super().__init__(config)
        transformers_loading_kwargs = transformers_loading_kwargs or {"trust_remote_code": True}
        self.config = config
        backbone_cls = get_backbone_cls(config)
        self.backbone = backbone_cls(
            model_name=config.model_name,
            tune_llm=config.tune_llm,
            tune_visual=config.tune_visual,
            select_layer=config.select_layer,
            reproject_vision=config.reproject_vision,
            use_flash_attention=config.use_flash_attention,
            load_bf16=config.load_bf16,
            tune_top_llm_layers=config.tune_top_llm_layers,
            trainable_params_fp32=config.backbone_trainable_params_fp32,
            transformers_loading_kwargs=transformers_loading_kwargs,
            load_pretrained_weights=load_backbone_weights,
        )
        self.action_head = GR00TN17ActionHead(config)
        self.post_init()
    def prepare_input(self, inputs: dict[str, Any]) -> tuple[BatchFeature, BatchFeature]:
        require_package("dm-tree", extra="groot", import_name="tree")
        backbone_inputs = self.backbone.prepare_input(inputs)
        action_inputs = self.action_head.prepare_input(inputs)
        def to_device_with_dtype(x):
            if not isinstance(x, torch.Tensor):
                return x
            if torch.is_floating_point(x):
                return x.to(self.device, dtype=self.dtype)
            return x.to(self.device)
        return (
            tree.map_structure(to_device_with_dtype, backbone_inputs),
            tree.map_structure(to_device_with_dtype, action_inputs),
        )
    def forward(self, inputs: dict[str, Any]) -> BatchFeature:
        backbone_inputs, action_inputs = self.prepare_input(inputs)
        backbone_outputs = self.backbone(backbone_inputs)
        return self.action_head(backbone_outputs, action_inputs)
    def get_action(self, inputs: dict[str, Any], options: dict[str, Any] | None = None) -> BatchFeature:
        backbone_inputs, action_inputs = self.prepare_input(inputs)
        backbone_outputs = self.backbone(backbone_inputs)
        return self.action_head.get_action(backbone_outputs, action_inputs, options)
    @property
    def device(self) -> torch.device:
        return next(iter(self.parameters())).device
    @property
    def dtype(self) -> torch.dtype:
        return next(iter(self.parameters())).dtype
    @classmethod
    def from_pretrained(cls, pretrained_model_name_or_path: str, **kwargs):
        tune_visual = kwargs.pop("tune_visual", True)
        tune_llm = kwargs.pop("tune_llm", False)
        tune_projector = kwargs.pop("tune_projector", True)
        tune_diffusion_model = kwargs.pop("tune_diffusion_model", True)
        tune_vlln = kwargs.pop("tune_vlln", True)
        transformers_loading_kwargs = kwargs.pop("transformers_loading_kwargs", None) or {
            "trust_remote_code": True
        }
        load_backbone_weights = kwargs.pop("load_backbone_weights", False)
        for key in ("cache_dir", "local_files_only", "token"):
            if key in kwargs:
                transformers_loading_kwargs.setdefault(key, kwargs[key])
        try:
            local_model_path = snapshot_download(
                pretrained_model_name_or_path,
                repo_type="model",
                revision=kwargs.get("revision"),
                cache_dir=kwargs.get("cache_dir"),
                local_files_only=kwargs.get("local_files_only", False),
                token=kwargs.get("token"),
            )
        except (HFValidationError, RepositoryNotFoundError):
            local_model_path = pretrained_model_name_or_path
        pretrained_model = super().from_pretrained(
            local_model_path,
            transformers_loading_kwargs=transformers_loading_kwargs,
            load_backbone_weights=load_backbone_weights,
            **kwargs,
        )
        pretrained_model.backbone.set_trainable_parameters(
            tune_visual=tune_visual,
            tune_llm=tune_llm,
            tune_top_llm_layers=pretrained_model.config.tune_top_llm_layers,
        )
        pretrained_model.action_head.set_trainable_parameters(
            tune_projector=tune_projector,
            tune_diffusion_model=tune_diffusion_model,
            tune_vlln=tune_vlln,
        )
        return pretrained_model
 def _register_with_transformers() -> None:
    """Register GR00T N1.7 with transformers' Auto* factories.
    Idempotent: ``register(..., exist_ok=True)`` makes repeat calls no-ops (with a fallback that
    suppresses the already-registered error on transformers builds whose ``register()`` predates
    ``exist_ok``), so no run-once guard is needed.
    """
    if AutoConfig is None or AutoModel is None:
        return
    try:
        AutoConfig.register(GR00TN17Config.model_type, GR00TN17Config, exist_ok=True)
    except TypeError:
        with suppress(ValueError):
            AutoConfig.register(GR00TN17Config.model_type, GR00TN17Config)
    try:
        AutoModel.register(GR00TN17Config, GR00TN17, exist_ok=True)
    except TypeError:
        with suppress(ValueError):
            AutoModel.register(GR00TN17Config, GR00TN17)
@@ -17,47 +17,37 @@
 """
 Groot Policy Wrapper for LeRobot Integration
-Minimal integration that delegates to Isaac-GR00T N1.7 components where
+Minimal integration that delegates to Isaac-GR00T components where possible
-possible without porting their code. Dataset loading and training
+without porting their code. The intent is to:
-orchestration are handled by LeRobot's standard training stack.
+
 - Download and load the pretrained GR00T model via GR00TN15.from_pretrained
 - Optionally align action horizon similar to gr00t_finetune.py
 - Expose predict_action via GR00T model.get_action
 - Provide a training forward that can call the GR00T model forward if batch
  structure matches.
 Notes:
 - Dataset loading and full training orchestration is handled by Isaac-GR00T
  TrainRunner in their codebase. If you want to invoke that flow end-to-end
  from LeRobot, see `GrootPolicy.finetune_with_groot_runner` below.
 """
 import builtins
 import logging
 import os
 from collections import deque
 from pathlib import Path
-from typing import TYPE_CHECKING, TypeVar
+from typing import TypeVar
 import torch
 from huggingface_hub import hf_hub_download
 from huggingface_hub.constants import SAFETENSORS_SINGLE_FILE
 from huggingface_hub.errors import HfHubHTTPError
 from torch import Tensor
 from lerobot.configs import FeatureType, PolicyFeature
 from lerobot.utils.constants import ACTION, OBS_IMAGES
-from lerobot.utils.import_utils import _transformers_available, require_package
+from lerobot.utils.import_utils import require_package
 from ..pretrained import PreTrainedPolicy
-from ..utils import get_device_from_parameters
+from .configuration_groot import GrootConfig
-from .configuration_groot import (
+from .groot_n1 import GR00TN15
    GROOT_N1_5,
    GROOT_N1_5_REMOVAL_GUIDANCE,
    GROOT_N1_7,
    GrootConfig,
    infer_groot_model_version,
    infer_groot_n1_7_action_execution_horizon,
    infer_groot_n1_7_action_horizon,
 )
 from .groot_n1_7 import GR00TN17, _tie_unused_qwen_lm_head
 if TYPE_CHECKING or _transformers_available:
    from transformers.trainer_pt_utils import get_parameter_names
 else:
    get_parameter_names = None  # type: ignore[assignment]
 logger = logging.getLogger(__name__)
 T = TypeVar("T", bound="GrootPolicy")
@@ -77,77 +67,37 @@ class GrootPolicy(PreTrainedPolicy):
        # Initialize GR00T model using ported components
        self._groot_model = self._create_groot_model()
        self._action_queue_steps = self._resolve_action_queue_steps()
        self._warned_native_relative_rtc_prefix_disabled = False
        self.reset()
    def _create_groot_model(self):
-        """Create and initialize the GR00T N1.7 model using the ported components."""
+        """Create and initialize the GR00T model using Isaac-GR00T API.
        model_kwargs = {
            "pretrained_model_name_or_path": self.config.base_model_path,
            "tune_llm": self.config.tune_llm,
            "tune_visual": self.config.tune_visual,
            "tune_projector": self.config.tune_projector,
            "tune_diffusion_model": self.config.tune_diffusion_model,
            # Forwarded as a GR00TN17Config override; read back by set_trainable_parameters.
            "tune_top_llm_layers": self.config.tune_top_llm_layers,
            "use_flash_attention": self.config.use_flash_attention,
        }
        # Surface the inference-time knobs onto the model config only when the user set them; None
        # leaves the value baked into the checkpoint untouched.
        if self.config.num_inference_timesteps is not None:
            model_kwargs["num_inference_timesteps"] = self.config.num_inference_timesteps
        if self.config.rtc_ramp_rate is not None:
            model_kwargs["rtc_ramp_rate"] = self.config.rtc_ramp_rate
-        model = GR00TN17.from_pretrained(
+        This is only called when creating a NEW policy (not when loading from checkpoint).
-            **model_kwargs,
+
-            tune_vlln=self.config.tune_vlln,
+        Steps (delegating to Isaac-GR00T):
-            transformers_loading_kwargs={"trust_remote_code": True},
+        1) Download and load pretrained model via GR00TN15.from_pretrained
        2) Align action horizon with data_config if provided
        """
        # Handle Flash Attention compatibility issues
        self._handle_flash_attention_compatibility()
        model = GR00TN15.from_pretrained(
            pretrained_model_name_or_path=self.config.base_model_path,
            tune_llm=self.config.tune_llm,
            tune_visual=self.config.tune_visual,
            tune_projector=self.config.tune_projector,
            tune_diffusion_model=self.config.tune_diffusion_model,
        )
-        backbone = getattr(model, "backbone", None)
+
-        qwen_model = getattr(backbone, "model", None)
+        model.compute_dtype = "bfloat16" if self.config.use_bf16 else model.compute_dtype
-        if qwen_model is not None:
+        model.config.compute_dtype = model.compute_dtype
-            _tie_unused_qwen_lm_head(qwen_model)
+
        if self.config.model_params_fp32:
            self._cast_model_parameters_to_fp32(model)
        return model
    @staticmethod
    def _cast_model_parameters_to_fp32(model: torch.nn.Module) -> None:
        for parameter in model.parameters():
            if parameter.is_floating_point():
                parameter.data = parameter.data.to(torch.float32)
    @staticmethod
    def _build_weight_decay_parameter_groups(model: torch.nn.Module) -> list[dict[str, object]]:
        forbidden_name_patterns = [
            r"bias",
            r"layernorm",
            r"rmsnorm",
            r"(?:^|\.)norm(?:$|\.)",
            r"_norm(?:$|\.)",
        ]
        decay_names = set(get_parameter_names(model, [torch.nn.LayerNorm], forbidden_name_patterns))
        decay_params = [
            parameter
            for name, parameter in model.named_parameters()
            if parameter.requires_grad and name in decay_names
        ]
        no_decay_params = [
            parameter
            for name, parameter in model.named_parameters()
            if parameter.requires_grad and name not in decay_names
        ]
        return [
            {"params": decay_params},
            {"params": no_decay_params, "weight_decay": 0.0},
        ]
    def reset(self):
        """Reset policy state when environment resets."""
-        self._action_queue = deque([], maxlen=self._action_queue_steps)
+        self._action_queue = deque([], maxlen=self.config.n_action_steps)
    @classmethod
    def from_pretrained(
@@ -168,7 +118,7 @@ class GrootPolicy(PreTrainedPolicy):
        """Load Groot policy from pretrained model.
        Handles two cases:
-        1. Base GR00T N1.7 models - loads the raw model
+        1. Base GR00T models (e.g., 'nvidia/GR00T-N1.5-3B') - loads the raw model
        2. Fine-tuned LeRobot checkpoints - loads config and weights from safetensors
        Args:
@@ -187,11 +137,13 @@ class GrootPolicy(PreTrainedPolicy):
        Returns:
            Initialized GrootPolicy instance with loaded model
        """
-        requested_version = infer_groot_model_version(str(pretrained_name_or_path)) or GROOT_N1_7
+        from huggingface_hub import hf_hub_download
-        logger.info(
+        from huggingface_hub.constants import SAFETENSORS_SINGLE_FILE
-            "The Groot policy wraps NVIDIA's GR00T %s model. Loading pretrained model from: %s",
+        from huggingface_hub.errors import HfHubHTTPError
-            requested_version,
+
-            pretrained_name_or_path,
+        print(
            "The Groot policy is a wrapper around Nvidia's GR00T N1.5 model.\n"
            f"Loading pretrained model from: {pretrained_name_or_path}"
        )
        model_id = str(pretrained_name_or_path)
@@ -222,7 +174,7 @@ class GrootPolicy(PreTrainedPolicy):
        if is_finetuned_checkpoint:
            # This is a fine-tuned LeRobot checkpoint - use parent class loading
-            logger.info("Detected fine-tuned LeRobot checkpoint, loading with state dict...")
+            print("Detected fine-tuned LeRobot checkpoint, loading with state dict...")
            return super().from_pretrained(
                pretrained_name_or_path=pretrained_name_or_path,
                config=config,
@@ -238,13 +190,11 @@ class GrootPolicy(PreTrainedPolicy):
            )
        # This is a base GR00T model - load it fresh
-        logger.info("Detected base GR00T model, loading from HuggingFace...")
+        print("Detected base GR00T model, loading from HuggingFace...")
        if config is None:
            # Create default config with the pretrained path
-            config = GrootConfig(
+            config = GrootConfig(base_model_path=str(pretrained_name_or_path))
                base_model_path=str(pretrained_name_or_path),
            )
            # Add minimal visual feature required for validation
            # validate_features() will automatically add state and action features
@@ -265,15 +215,6 @@ class GrootPolicy(PreTrainedPolicy):
            if hasattr(config, key):
                setattr(config, key, value)
        inferred_version = infer_groot_model_version(config.base_model_path)
        if inferred_version is not None and inferred_version != GROOT_N1_7:
            message = (
                f"GR00T model_version '{GROOT_N1_7}' does not match base_model_path "
                f"'{config.base_model_path}', which looks like '{inferred_version}'."
            )
            if inferred_version == GROOT_N1_5:
                message = f"{message} {GROOT_N1_5_REMOVAL_GUIDANCE}"
            raise ValueError(message)
        # Create a fresh policy instance - this will automatically load the GR00T model
        # in __init__ via _create_groot_model()
        policy = cls(config)
@@ -281,174 +222,24 @@ class GrootPolicy(PreTrainedPolicy):
        policy.eval()
        return policy
-    def get_optim_params(self):  # type: ignore[override]
+    def get_optim_params(self) -> dict:
-        """Isaac-GR00T excludes biases and normalization parameters from weight decay."""
+        return self.parameters()
        return self._build_weight_decay_parameter_groups(self)
    def _resolve_action_queue_steps(self) -> int:
        n_action_steps = int(self.config.n_action_steps)
        checkpoint_action_horizon = infer_groot_n1_7_action_horizon(
            self.config.base_model_path,
            self.config.embodiment_tag,
        )
        execution_horizon = infer_groot_n1_7_action_execution_horizon(
            self.config.base_model_path,
            self.config.embodiment_tag,
        )
        horizons = [n_action_steps]
        if checkpoint_action_horizon is not None:
            horizons.append(checkpoint_action_horizon)
        if execution_horizon is not None:
            horizons.append(execution_horizon)
        return min(horizons)
    def _resolve_prediction_horizon(self, actions: Tensor) -> int:
        """Return the policy-facing action horizon for a native GR00T prediction."""
        horizons = [actions.shape[1]]
        checkpoint_action_horizon = infer_groot_n1_7_action_horizon(
            self.config.base_model_path,
            self.config.embodiment_tag,
        )
        if checkpoint_action_horizon is not None:
            horizons.append(checkpoint_action_horizon)
        for horizon in (self.config.chunk_size, self.config.n_action_steps):
            horizon = int(horizon)
            if horizon > 0:
                horizons.append(horizon)
        return max(1, min(horizons))
    def _filter_groot_inputs(self, batch: dict[str, Tensor], *, include_action: bool) -> dict[str, Tensor]:
        allowed_base = {"state", "state_mask", "action_mask", "embodiment_id"}
        if include_action:
            allowed_base.add("action")
        allowed_base.update(
            {
                "input_ids",
                "attention_mask",
                "pixel_values",
                "image_grid_thw",
                "mm_token_type_ids",
                "pixel_values_videos",
                "video_grid_thw",
            }
        )
        return {
            k: v for k, v in batch.items() if k in allowed_base and not (k.startswith("next.") or k == "info")
        }
    def _prepare_n1_7_rtc_inputs(
        self,
        inputs: dict[str, Tensor],
        *,
        inference_delay: object,
        prev_chunk_left_over: object,
    ) -> tuple[dict[str, Tensor], dict[str, object] | None]:
        if prev_chunk_left_over is None:
            return inputs, None
        if getattr(self.config, "use_relative_actions", False):
            # Generic RTC only provides normalized leftovers from the previous chunk. For
            # native relative-action N1.7 checkpoints those rows are tied to the old
            # observation state and old per-horizon stats row, so using them as the next
            # prefix can push the policy in the wrong direction. Run without native RTC
            # overlap guidance until a GROOT-specific RTC path can pass re-anchored
            # absolute leftovers through.
            if not getattr(self, "_warned_native_relative_rtc_prefix_disabled", False):
                logger.info("Disabling native GR00T RTC prefix for relative-action policy")
                self._warned_native_relative_rtc_prefix_disabled = True
            return inputs, None
        if not isinstance(prev_chunk_left_over, torch.Tensor):
            raise TypeError("prev_chunk_left_over must be a torch.Tensor for GR00T N1.7 RTC.")
        if prev_chunk_left_over.numel() == 0:
            return inputs, None
        prev_actions = prev_chunk_left_over
        if prev_actions.ndim == 2:
            prev_actions = prev_actions.unsqueeze(0)
        elif prev_actions.ndim != 3:
            raise ValueError("prev_chunk_left_over must have shape (T, A) or (B, T, A) for GR00T N1.7 RTC.")
        state = inputs.get("state")
        if state is None:
            raise ValueError("GR00T N1.7 RTC requires `state` in the preprocessed batch.")
        batch_size = state.shape[0]
        if prev_actions.shape[0] == 1 and batch_size > 1:
            prev_actions = prev_actions.expand(batch_size, -1, -1).clone()
        elif prev_actions.shape[0] != batch_size:
            raise ValueError("prev_chunk_left_over batch size must match the current GR00T N1.7 batch size.")
        # The generic LeRobot RTC engine pads short leftovers with exact zero
        # rows for fixed-shape policy calls. Native GR00T N1.7 RTC treats every
        # provided prefix row as a real action constraint, so strip that padding
        # before constructing the native overlap options.
        valid_prefix_rows = prev_actions.detach().abs().sum(dim=(0, 2)) > 0
        if valid_prefix_rows.any():
            valid_prefix_steps = int(valid_prefix_rows.nonzero()[-1].item()) + 1
            prev_actions = prev_actions[:, :valid_prefix_steps, :]
        else:
            return inputs, None
        model_action_horizon = int(
            getattr(self._groot_model.config, "action_horizon", self.config.chunk_size)
        )
        max_action_dim = int(getattr(self._groot_model.config, "max_action_dim", self.config.max_action_dim))
        if prev_actions.shape[1] > model_action_horizon:
            prev_actions = prev_actions[:, -model_action_horizon:, :]
        action_horizon = int(prev_actions.shape[1])
        if action_horizon <= 0:
            return inputs, None
        if prev_actions.shape[2] > max_action_dim:
            prev_actions = prev_actions[:, :, :max_action_dim]
        elif prev_actions.shape[2] < max_action_dim:
            pad = torch.zeros(
                prev_actions.shape[0],
                prev_actions.shape[1],
                max_action_dim - prev_actions.shape[2],
                dtype=prev_actions.dtype,
                device=prev_actions.device,
            )
            prev_actions = torch.cat([prev_actions, pad], dim=2)
        prev_actions = prev_actions.to(device=state.device, dtype=state.dtype)
        rtc_config = getattr(self.config, "rtc_config", None)
        execution_horizon = int(getattr(rtc_config, "execution_horizon", action_horizon))
        overlap_steps = max(0, min(action_horizon, execution_horizon))
        if overlap_steps == 0:
            return inputs, None
        try:
            frozen_steps = int(inference_delay or 0)
        except (TypeError, ValueError):
            frozen_steps = 0
        frozen_steps = max(0, min(frozen_steps, overlap_steps))
        options = {
            "action_horizon": action_horizon,
            "rtc_overlap_steps": overlap_steps,
            "rtc_frozen_steps": frozen_steps,
            "rtc_ramp_rate": float(getattr(self._groot_model.config, "rtc_ramp_rate", 6.0)),
        }
        inputs = dict(inputs)
        inputs["action"] = prev_actions
        return inputs, options
    def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict]:
        """Training forward pass.
        Delegates to Isaac-GR00T model.forward when inputs are compatible.
        """
-        groot_inputs = self._filter_groot_inputs(batch, include_action=True)
+        # Build a clean input dict for GR00T: keep only tensors GR00T consumes
        allowed_base = {"state", "state_mask", "action", "action_mask", "embodiment_id"}
        groot_inputs = {
            k: v
            for k, v in batch.items()
            if (k in allowed_base or k.startswith("eagle_")) and not (k.startswith("next.") or k == "info")
        }
        # Get device from model parameters
-        device = get_device_from_parameters(self)
+        device = next(self.parameters()).device
        # Run GR00T forward under bf16 autocast when enabled to reduce activation memory
        # Rationale: Matches original GR00T finetuning (bf16 compute, fp32 params) and avoids fp32 upcasts.
@@ -457,52 +248,38 @@ class GrootPolicy(PreTrainedPolicy):
        # Isaac-GR00T returns a BatchFeature; loss key is typically 'loss'
        loss = outputs.get("loss")
        if loss is None:
            raise RuntimeError(
                "GR00T model.forward did not return a 'loss'. Training batches must include "
                "'action' and 'action_mask'; check the preprocessor output."
            )
        loss_dict = {"loss": loss.item()}
        return loss, loss_dict
    @torch.no_grad()
-    def predict_action_chunk(self, batch: dict[str, Tensor], **kwargs: object) -> Tensor:
+    def predict_action_chunk(self, batch: dict[str, Tensor]) -> Tensor:
        """Predict a chunk of actions for inference by delegating to Isaac-GR00T.
        Returns a tensor of shape (B, n_action_steps, action_dim).
        For N1.7, LeRobot's RTC leftovers are converted into the native GR00T
        action-overlap options before calling the underlying model.
        """
        self.eval()
-        # Preprocessing is handled by the processor pipeline, so we just filter the batch.
+        # Build a clean input dict for GR00T: keep only tensors GR00T consumes
-        # During inference, we do not pass action because it is predicted.
+        # Preprocessing is handled by the processor pipeline, so we just filter the batch
-        # N1.7 still carries a 2-D action horizon mask from its checkpoint processor.
+        # NOTE: During inference, we should NOT pass action/action_mask (that's what we're predicting)
-        groot_inputs = self._filter_groot_inputs(batch, include_action=False)
+        allowed_base = {"state", "state_mask", "embodiment_id"}
-        groot_inputs, groot_options = self._prepare_n1_7_rtc_inputs(
+        groot_inputs = {
-            groot_inputs,
+            k: v
-            inference_delay=kwargs.get("inference_delay"),
+            for k, v in batch.items()
-            prev_chunk_left_over=kwargs.get("prev_chunk_left_over"),
+            if (k in allowed_base or k.startswith("eagle_")) and not (k.startswith("next.") or k == "info")
-        )
+        }
        # Get device from model parameters
-        device = get_device_from_parameters(self)
+        device = next(self.parameters()).device
        # Use bf16 autocast for inference to keep memory low and match backbone dtype
        with torch.autocast(device_type=device.type, dtype=torch.bfloat16, enabled=self.config.use_bf16):
-            if groot_options is not None:
+            outputs = self._groot_model.get_action(groot_inputs)
                outputs = self._groot_model.get_action(groot_inputs, options=groot_options)
            else:
                outputs = self._groot_model.get_action(groot_inputs)
        actions = outputs.get("action_pred")
        prediction_horizon = self._resolve_prediction_horizon(actions)
        actions = actions[:, :prediction_horizon]
        original_action_dim = self.config.output_features[ACTION].shape[0]
        actions = actions[:, :, :original_action_dim]
@@ -511,17 +288,44 @@ class GrootPolicy(PreTrainedPolicy):
    @torch.no_grad()
    def select_action(self, batch: dict[str, Tensor]) -> Tensor:
        """Select single action from action queue."""
        if getattr(self.config, "use_relative_actions", False):
            raise NotImplementedError(
                "GrootPolicy.select_action does not support relative-action policies because cached "
                "relative chunk actions can be decoded against newer observation states. Use "
                "predict_action_chunk and postprocess the full chunk before queuing actions, or use "
                "the RTC/chunked rollout inference path."
            )
        self.eval()
        if len(self._action_queue) == 0:
            actions = self.predict_action_chunk(batch)
-            self._action_queue.extend(actions[:, : self._action_queue_steps].transpose(0, 1))
+            self._action_queue.extend(actions.transpose(0, 1))
        return self._action_queue.popleft()
    # -------------------------
    # Internal helpers
    # -------------------------
    def _handle_flash_attention_compatibility(self) -> None:
        """Handle Flash Attention compatibility issues by setting environment variables.
        This addresses the common 'undefined symbol' error that occurs when Flash Attention
        is compiled against a different PyTorch version than what's currently installed.
        """
        # Set environment variables to handle Flash Attention compatibility
        # These help with symbol resolution issues
        os.environ.setdefault("FLASH_ATTENTION_FORCE_BUILD", "0")
        os.environ.setdefault("FLASH_ATTENTION_SKIP_CUDA_BUILD", "0")
        # Try to import flash_attn and handle failures gracefully
        try:
            import flash_attn
            print(f"[GROOT] Flash Attention version: {flash_attn.__version__}")
        except ImportError as e:
            print(f"[GROOT] Flash Attention not available: {e}")
            print("[GROOT] Will use fallback attention mechanism")
        except Exception as e:
            if "undefined symbol" in str(e):
                print(f"[GROOT] Flash Attention compatibility issue detected: {e}")
                print("[GROOT] This is likely due to PyTorch/Flash Attention version mismatch")
                print("[GROOT] Consider reinstalling Flash Attention with compatible version:")
                print("  pip uninstall flash-attn")
                print("  pip install --no-build-isolation flash-attn==2.6.3")
                print("[GROOT] Continuing with fallback attention mechanism")
            else:
                print(f"[GROOT] Flash Attention error: {e}")
                print("[GROOT] Continuing with fallback attention mechanism")
@@ -1,264 +1,47 @@
 #!/usr/bin/env python
 # Copyright 2025 NVIDIA Corporation and The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
 #     http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """Shared, side-effect-free utilities for the GR00T N1.7 policy.
 These helpers are consumed by both the config layer (checkpoint sidecar
 inspection) and the processor layer (stat flattening, action decoding, language
 and image packing). They are pure functions with no GR00T-specific state so they
 can be unit-tested in isolation and reused without importing the heavier
 config/processor modules.
 """
 from __future__ import annotations
 import json
 from pathlib import Path
-from typing import Any
+from shutil import copytree
-import numpy as np
+from huggingface_hub import hf_hub_download
 import torch
-def read_json(path: Path) -> dict[str, Any]:
+def ensure_eagle_cache_ready(vendor_dir: Path, cache_dir: Path, assets_repo: str) -> None:
-    """Read a JSON object from ``path``, returning ``{}`` on any read/parse error."""
+    """Populate the Eagle processor directory in cache and ensure tokenizer assets exist.
    try:
        with path.open() as f:
            data = json.load(f)
    except (OSError, json.JSONDecodeError):
        return {}
    return data if isinstance(data, dict) else {}
-
+    - Copies the vendored Eagle files into cache_dir (overwriting when needed).
-def as_int_pair(value: Any) -> list[int] | None:
+    - Downloads vocab.json and merges.txt into the same cache_dir if missing.
    if not isinstance(value, (list, tuple)) or len(value) != 2:
        return None
    try:
        return [int(value[0]), int(value[1])]
    except (TypeError, ValueError):
        return None
 def as_optional_int(value: Any) -> int | None:
    if value is None:
        return None
    try:
        return int(value)
    except (TypeError, ValueError):
        return None
 def as_optional_float(value: Any) -> float | None:
    if value is None:
        return None
    try:
        return float(value)
    except (TypeError, ValueError):
        return None
 def as_float_list(values: Any) -> list[float]:
    if values is None:
        return []
    if isinstance(values, torch.Tensor):
        return values.detach().cpu().reshape(-1).float().tolist()
    if isinstance(values, np.ndarray):
        return values.reshape(-1).astype(np.float32).tolist()
    if isinstance(values, (list, tuple)):
        flattened: list[float] = []
        for value in values:
            flattened.extend(as_float_list(value))
        return flattened
    return [float(values)]
 def config_value(value: Any) -> str:
    if hasattr(value, "value"):
        value = value.value
    text = str(value).lower()
    return {
        "relative": "relative",
        "absolute": "absolute",
        "delta": "delta",
        "eef": "eef",
        "non_eef": "non_eef",
        "default": "default",
        "xyz_rot6d": "xyz+rot6d",
        "xyz+rot6d": "xyz+rot6d",
        "xyz_rotvec": "xyz+rotvec",
        "xyz+rotvec": "xyz+rotvec",
    }.get(text, text)
 def has_modality_stats(stats: dict[str, dict[str, Any]] | None) -> bool:
    if not stats:
        return False
    return any(bool(modality_stats) for modality_stats in stats.values())
 def stat_dim_from_entry(entry: dict[str, Any]) -> int:
    for stat_name in ("mean", "q01", "min", "max", "std"):
        value = entry.get(stat_name)
        if isinstance(value, torch.Tensor):
            return int(value.shape[-1]) if value.ndim > 0 else 1
        if isinstance(value, np.ndarray):
            return int(value.shape[-1]) if value.ndim > 0 else 1
        if isinstance(value, list) and len(value) > 0:
            first = value[0]
            if isinstance(first, (list, tuple)) and len(first) > 0:
                return len(first)
            return len(value)
    return 0
 def flatten_n1_7_modality_stats(
    *,
    embodiment_stats: dict[str, Any],
    embodiment_config: dict[str, Any],
    modality: str,
    use_percentiles: bool,
    use_relative_action: bool,
 ) -> dict[str, list[float]]:
    """Flatten one N1.7 modality's grouped statistics in checkpoint order.
    When checkpoints request percentile normalization, q01/q99 replace min/max
    for regular groups. Relative action groups read from ``relative_action``
    stats and keep min/max, matching Isaac-GR00T's processor override.
    """
    cache_dir = Path(cache_dir)
    vendor_dir = Path(vendor_dir)
-    source_stats = embodiment_stats.get(modality, {})
+    try:
-    modality_config = embodiment_config.get(modality, {})
+        # Populate/refresh cache with vendor files to ensure a complete processor directory
-    if not isinstance(source_stats, dict) or not isinstance(modality_config, dict):
+        print(f"[GROOT] Copying vendor Eagle files to cache: {vendor_dir} -> {cache_dir}")
-        return {}
+        copytree(vendor_dir, cache_dir, dirs_exist_ok=True)
-    modality_keys = modality_config.get("modality_keys", [])
+    except Exception as exc:  # nosec: B110
-    if not isinstance(modality_keys, list):
+        print(f"[GROOT] Warning: Failed to copy vendor Eagle files to cache: {exc}")
        return {}
-    flattened: dict[str, list[float]] = {}
+    required_assets = [
-    action_configs = modality_config.get("action_configs", []) if modality == "action" else []
+        "vocab.json",
-    if not isinstance(action_configs, list):
+        "merges.txt",
-        action_configs = []
+        "added_tokens.json",
-    relative_stats = embodiment_stats.get("relative_action", {})
+        "chat_template.json",
-    if not isinstance(relative_stats, dict):
+        "special_tokens_map.json",
-        relative_stats = {}
+        "config.json",
        "generation_config.json",
        "preprocessor_config.json",
        "processor_config.json",
        "tokenizer_config.json",
    ]
-    for stat_name in ("min", "max", "mean", "std"):
+    print(f"[GROOT] Assets repo: {assets_repo} \n Cache dir: {cache_dir}")
        values: list[float] = []
        source_stat_name = stat_name
        if use_percentiles and stat_name == "min":
            source_stat_name = "q01"
        elif use_percentiles and stat_name == "max":
            source_stat_name = "q99"
-        for idx, modality_key in enumerate(modality_keys):
+    for fname in required_assets:
-            if not isinstance(modality_key, str):
+        dst = cache_dir / fname
-                continue
+        if not dst.exists():
-            key_source_stats = source_stats
+            print(f"[GROOT] Fetching {fname}")
-            key_stat_name = source_stat_name
+            hf_hub_download(
-            if modality == "action" and use_relative_action and idx < len(action_configs):
+                repo_id=assets_repo,
-                action_config = action_configs[idx]
+                filename=fname,
-                if isinstance(action_config, dict) and config_value(action_config.get("rep")) == "relative":
+                repo_type="model",
-                    key_source_stats = relative_stats
+                local_dir=str(cache_dir),
                    key_stat_name = stat_name
            key_stats = key_source_stats.get(modality_key, {})
            if not isinstance(key_stats, dict):
                raise KeyError(f"Missing statistics for {modality}.{modality_key}")
            raw_values = key_stats.get(key_stat_name)
            if raw_values is None:
                raise KeyError(f"Missing '{key_stat_name}' statistics for {modality}.{modality_key}")
            values.extend(as_float_list(raw_values))
        if values:
            flattened[stat_name] = values
    return flattened
 def rot6d_to_matrix(rot6d: np.ndarray) -> np.ndarray:
    rows = rot6d.reshape(2, 3).astype(np.float64)
    row1 = rows[0] / np.linalg.norm(rows[0])
    row2 = rows[1] - np.dot(row1, rows[1]) * row1
    row2 = row2 / np.linalg.norm(row2)
    row3 = np.cross(row1, row2)
    return np.vstack([row1, row2, row3])
 def xyz_rot6d_to_homogeneous(xyz_rot6d: np.ndarray) -> np.ndarray:
    transform = np.eye(4, dtype=np.float64)
    transform[:3, :3] = rot6d_to_matrix(xyz_rot6d[3:])
    transform[:3, 3] = xyz_rot6d[:3]
    return transform
 def homogeneous_to_xyz_rot6d(transform: np.ndarray) -> np.ndarray:
    return np.concatenate([transform[:3, 3], transform[:2, :3].reshape(-1)], axis=0)
 def relative_eef_to_absolute(action: np.ndarray, reference_state: np.ndarray) -> np.ndarray:
    """Convert relative EEF deltas in xyz+rot6d format to absolute EEF poses."""
    out = np.empty_like(action, dtype=np.float64)
    for batch_idx in range(action.shape[0]):
        reference = xyz_rot6d_to_homogeneous(reference_state[batch_idx])
        for timestep in range(action.shape[1]):
            relative = xyz_rot6d_to_homogeneous(action[batch_idx, timestep])
            out[batch_idx, timestep] = homogeneous_to_xyz_rot6d(reference @ relative)
    return out.astype(np.float32)
 def infer_n1_7_batch_size_and_device(
    obs: dict[str, Any], action: torch.Tensor | None
 ) -> tuple[int, torch.device]:
    for value in list(obs.values()) + [action]:
        if isinstance(value, torch.Tensor):
            return value.shape[0], value.device
    video = obs.get("video")
    if isinstance(video, np.ndarray):
        return video.shape[0], torch.device("cpu")
    return 1, torch.device("cpu")
 def prepare_n1_7_language_batch(
    language: Any,
    batch_size: int,
    *,
    formalize_language: bool,
 ) -> list[str]:
    default_language = "Perform the task."
    if language is None or (isinstance(language, str) and language == ""):
        languages = [default_language] * batch_size
    elif isinstance(language, str):
        languages = [language] * batch_size
    elif isinstance(language, (list, tuple)):
        languages = list(language)
        if len(languages) == 0:
            languages = [default_language] * batch_size
        elif len(languages) == 1 and batch_size > 1:
            languages = languages * batch_size
        elif len(languages) != batch_size:
            raise ValueError(
                f"language batch has {len(languages)} entries, but GR00T N1.7 input batch has {batch_size}."
            )
    else:
        languages = [str(language)] * batch_size
    formatted = []
    for item in languages:
        text = str(item) if item else default_language
        if formalize_language:
            text = text.lower()
            text = "".join(ch for ch in text if ch.isalnum() or ch.isspace() or ch == "_")
        formatted.append(text)
    return formatted
@@ -1 +0,0 @@
 ../../../../docs/source/lingbot_va.mdx
@@ -1,21 +0,0 @@
 #!/usr/bin/env python
 # Copyright 2026 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
 #     http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from .configuration_lingbot_va import LingBotVAConfig
 from .modeling_lingbot_va import LingBotVAPolicy
 from .processor_lingbot_va import make_lingbot_va_pre_post_processors
 __all__ = ["LingBotVAConfig", "LingBotVAPolicy", "make_lingbot_va_pre_post_processors"]
@@ -1,168 +0,0 @@
 # Copyright 2026 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
 #     http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """Configuration for the LingBot-VA policy.
 LingBot-VA is an autoregressive video-action world-model policy built on the Wan2.2
 video-diffusion stack. It interleaves prediction of future video latents and robot
 actions in a single dual-stream transformer. See ``docs/source/lingbot_va.mdx`` and the
 upstream repository (https://github.com/Robbyant/lingbot-va).
 Defaults below match the upstream LIBERO configuration (``wan_va/configs/va_libero_cfg.py``)
 and the ``transformer/config.json`` of the released checkpoints.
 """
 from dataclasses import dataclass, field
 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
 from lerobot.optim.optimizers import AdamWConfig
 from lerobot.optim.schedulers import ConstantWithWarmupSchedulerConfig, LRSchedulerConfig
 from lerobot.utils.constants import ACTION
@PreTrainedConfig.register_subclass("lingbot_va")
@dataclass
 class LingBotVAConfig(PreTrainedConfig):
    """Configuration for the native LingBot-VA policy integration in LeRobot."""
    # Wan transformer architecture
    patch_size: tuple[int, int, int] = (1, 2, 2)
    num_attention_heads: int = 24
    attention_head_dim: int = 128
    in_channels: int = 48
    out_channels: int = 48
    action_dim: int = 30
    text_dim: int = 4096
    freq_dim: int = 256
    ffn_dim: int = 14336
    num_layers: int = 30
    cross_attn_norm: bool = True
    eps: float = 1e-6
    rope_max_seq_len: int = 1024
    # "flex" = training only (needs recent torch); inference uses "torch" SDPA or "flashattn".
    attn_mode: str = "torch"
    # Frozen sub-models (VAE + UMT5 text encoder + tokenizer)
    # ~20 GB of frozen weights, NOT bundled in the checkpoint; lazily pulled from this HF repo /
    # local dir (must hold diffusers-style ``vae/``, ``text_encoder/``, ``tokenizer/`` sub-folders).
    wan_pretrained_path: str = "robbyant/lingbot-va-base"
    dtype: str = "bfloat16"  # transformer / VAE / text-encoder dtype: "bfloat16", "float16", "float32"
    # Frozen UMT5-XXL encoder device; "cpu" frees ~11 GB VRAM (it runs once per episode).
    text_encoder_device: str = "cpu"
    # Observation cameras (order matters: latents are concatenated on width; LIBERO defaults)
    obs_cam_keys: list[str] = field(
        default_factory=lambda: ["observation.images.image", "observation.images.image2"]
    )
    # Undo the LIBERO env processor's extra horizontal flip to match the model's training orientation.
    image_hflip: bool = False
    # Camera latent layout: "width_concat" (cameras concatenated on width; LIBERO) or
    # "robotwin_tshape" (full-res head + half-res wrists in a "T"; RoboTwin).
    camera_layout: str = "width_concat"
    # Inference hyperparameters (LIBERO defaults)
    n_obs_steps: int = 1
    height: int = 128
    width: int = 128
    action_per_frame: int = 4
    frame_chunk_size: int = 4
    attn_window: int = 30
    num_inference_steps: int = 20
    video_exec_step: int = -1
    action_num_inference_steps: int = 50
    guidance_scale: float = 5.0
    action_guidance_scale: float = 1.0
    snr_shift: float = 5.0
    action_snr_shift: float = 0.05
    max_sequence_length: int = 512  # UMT5 prompt length
    # Subset of the 30-d action space used by the benchmark (LIBERO = 7-DoF). The action
    # (un)normalization quantiles live in the checkpoint's ``policy_postprocessor.json``, not here.
    used_action_channel_ids: list[int] = field(default_factory=lambda: list(range(7)))
    # Opt-in: VAE-decode predicted video latents to ``self.last_predicted_frames`` for saving MP4s.
    save_predicted_video: bool = False
    # Normalization: IDENTITY here; images are scaled + VAE-encoded and actions are
    # quantile-(un)normalized inside the policy / dedicated processor steps.
    normalization_mapping: dict[str, NormalizationMode] = field(
        default_factory=lambda: {
            "VISUAL": NormalizationMode.IDENTITY,
            "STATE": NormalizationMode.IDENTITY,
            "ACTION": NormalizationMode.IDENTITY,
        }
    )
    # Optimizer / scheduler (training; AdamW + warmup-constant per upstream train.py)
    optimizer_lr: float = 1e-5
    optimizer_betas: tuple[float, float] = (0.9, 0.95)
    optimizer_eps: float = 1e-8
    optimizer_weight_decay: float = 1e-4
    optimizer_grad_clip_norm: float = 1.0
    scheduler_warmup_steps: int = 1000
    def __post_init__(self):
        super().__post_init__()
        if self.attn_mode not in ("torch", "flashattn", "flex"):
            raise ValueError(f"attn_mode must be one of 'torch', 'flashattn', 'flex'; got {self.attn_mode!r}")
    @property
    def chunk_size(self) -> int:
        """Number of single-step actions produced per autoregressive chunk."""
        return self.frame_chunk_size * self.action_per_frame
    @property
    def n_action_steps(self) -> int:
        """Number of actions executed before refilling (the whole chunk)."""
        return self.chunk_size
    def validate_features(self) -> None:
        image_features = [key for key, feat in self.input_features.items() if feat.type == FeatureType.VISUAL]
        if not image_features:
            raise ValueError(
                "LingBot-VA requires at least one visual input feature. "
                "No features of type FeatureType.VISUAL found in input_features."
            )
        if ACTION not in self.output_features:
            self.output_features[ACTION] = PolicyFeature(
                type=FeatureType.ACTION, shape=(len(self.used_action_channel_ids),)
            )
    def get_optimizer_preset(self) -> AdamWConfig:
        return AdamWConfig(
            lr=self.optimizer_lr,
            betas=self.optimizer_betas,
            eps=self.optimizer_eps,
            weight_decay=self.optimizer_weight_decay,
            grad_clip_norm=self.optimizer_grad_clip_norm,
        )
    def get_scheduler_preset(self) -> LRSchedulerConfig | None:
        # Upstream uses a linear warmup followed by a constant LR (warmup_constant_lambda).
        return ConstantWithWarmupSchedulerConfig(num_warmup_steps=self.scheduler_warmup_steps)
    @property
    def observation_delta_indices(self) -> list[int]:
        temporal_downsample = 4
        stride = max(1, self.action_per_frame // temporal_downsample)
        return list(range(0, self.frame_chunk_size * temporal_downsample * stride, stride))
    @property
    def action_delta_indices(self) -> list[int]:
        return list(range(self.chunk_size))
    @property
    def reward_delta_indices(self) -> None:
        return None
@@ -1,853 +0,0 @@
 # Copyright 2024-2025 The Robbyant Team Authors. All rights reserved.
 # Copyright 2026 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
 #     http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """LingBot-VA policy: an autoregressive video-action world model on the Wan2.2 stack.
 The sampling loop is a faithful re-implementation of the upstream streaming server
 (``wan_va/wan_va_server.py``) and LIBERO client (``evaluation/libero/client.py``), adapted
 to LeRobot's ``select_action`` interface:
  * the trainable dual-stream transformer is owned as a sub-module and round-trips in the
    single ``model.safetensors`` checkpoint;
  * the frozen Wan VAE + UMT5 text encoder + tokenizer are *lazily pulled* from
    ``config.wan_pretrained_path`` (not bundled), so the LeRobot checkpoint stays small;
  * ``predict_action_chunk`` runs one autoregressive chunk (video stream then action
    stream, each with CFG and its own flow-matching scheduler) and updates the KV cache;
  * ``select_action`` drains a per-step action queue and records the real observed
    keyframes that are fed back into the KV cache when the queue is refilled.
 NOTE: The streaming path is written for single-environment eval (``--eval.batch_size=1``).
 """
 from collections import deque
 import torch
 import torch.nn.functional as F  # noqa: N812
 from einops import rearrange
 from torch import Tensor
 from lerobot.policies.pretrained import PreTrainedPolicy
 from lerobot.utils.constants import ACTION
 from lerobot.utils.import_utils import require_package
 from .configuration_lingbot_va import LingBotVAConfig
 from .utils import (
    FlowMatchScheduler,
    WanTransformer3DModel,
    WanVAEStreamingWrapper,
    _sample_timestep_id,
    _torch_dtype,
    clean_prompt,
    data_seq_to_patch,
    denormalize_latents,
    get_mesh_id,
    load_text_encoder,
    load_tokenizer,
    load_vae,
 )
 class LingBotVAPolicy(PreTrainedPolicy):
    """LeRobot wrapper for the LingBot-VA autoregressive video-action world model."""
    config_class = LingBotVAConfig
    name = "lingbot_va"
    def __init__(self, config: LingBotVAConfig, **kwargs):
        require_package("diffusers", extra="lingbot_va")
        require_package("transformers", extra="lingbot_va")
        super().__init__(config)
        config.validate_features()
        self.config = config
        self.dtype = _torch_dtype(config.dtype)
        # Trainable dual-stream transformer (the only sub-module saved in the LeRobot checkpoint).
        self.transformer = WanTransformer3DModel(
            patch_size=tuple(config.patch_size),
            num_attention_heads=config.num_attention_heads,
            attention_head_dim=config.attention_head_dim,
            in_channels=config.in_channels,
            out_channels=config.out_channels,
            action_dim=config.action_dim,
            text_dim=config.text_dim,
            freq_dim=config.freq_dim,
            ffn_dim=config.ffn_dim,
            num_layers=config.num_layers,
            cross_attn_norm=config.cross_attn_norm,
            eps=config.eps,
            rope_max_seq_len=config.rope_max_seq_len,
            attn_mode=config.attn_mode,
        )
        # Run the transformer in config.dtype (bf16); norm/modulation paths upcast to fp32 internally.
        self.transformer = self.transformer.to(self.dtype)
        # Frozen modules are stored OUTSIDE the nn.Module registry (plain dict) so they are
        # neither saved into model.safetensors nor moved by ``.to()``. They are lazily loaded
        # from ``config.wan_pretrained_path`` the first time inference runs.
        self._frozen: dict = {}
        self.last_predicted_frames: Tensor | None = None
        self.last_predicted_latents: Tensor | None = None
        self.reset()
    # Frozen-module lazy loading (VAE + UMT5 + tokenizer)
    def _ensure_frozen_modules(self):
        if self._frozen:
            return
        path = self.config.wan_pretrained_path
        device = self.config.device
        # The frozen modules always live in ``vae/``, ``text_encoder/`` and ``tokenizer/``
        # sub-folders -- both in the released diffusers-style HF repos and in the local
        # ``--bundle-frozen`` output dir. ``from_pretrained(path, subfolder=...)`` resolves
        # them for either a HF repo id or a local directory.
        vae = load_vae(path, torch_dtype=self.dtype, torch_device=device, subfolder="vae")
        # The UMT5-XXL text encoder (~11 GB) runs once per episode; keep it on its own
        # (CPU by default) device so the 5B transformer + VAE fit on a single GPU.
        text_encoder = load_text_encoder(
            path,
            torch_dtype=self.dtype,
            torch_device=self.config.text_encoder_device,
            subfolder="text_encoder",
        )
        tokenizer = load_tokenizer(path, subfolder="tokenizer")
        self._frozen = {
            "vae": vae.eval(),
            "streaming_vae": WanVAEStreamingWrapper(vae),
            "text_encoder": text_encoder.eval(),
            "tokenizer": tokenizer,
        }
        # RoboTwin's T-shape layout encodes the half-resolution wrist cameras through a second
        # streaming VAE (separate causal cache) alongside the full-res head camera.
        if self.config.camera_layout == "robotwin_tshape":
            vae_half = load_vae(path, torch_dtype=self.dtype, torch_device=device, subfolder="vae")
            self._frozen["streaming_vae_half"] = WanVAEStreamingWrapper(vae_half.eval())
    @property
    def _vae(self):
        return self._frozen["vae"]
    @property
    def _streaming_vae(self):
        return self._frozen["streaming_vae"]
    # PreTrainedPolicy API
    def get_optim_params(self) -> dict:
        # Only the transformer is trainable; the VAE / text encoder stay frozen (kept outside the
        # nn.Module registry). With PEFT/LoRA this naturally returns just the adapter params.
        return [p for p in self.transformer.parameters() if p.requires_grad]
    def reset(self):
        """Reset all per-episode streaming state (KV cache, queues, frame counter)."""
        cfg = self.config
        self._action_queue: deque = deque(maxlen=cfg.n_action_steps)
        self._obs_buffer: list = []  # raw keyframe obs (one per env substep) observed this chunk
        self._executed_actions: Tensor | None = (
            None  # last chunk's actions (model-normalized) for KV feedback
        )
        self._started = False  # first select_action call uses the obs as the conditioning frame
        self._exec_step = 0  # index of the action being executed within the current chunk
        self._prev_j = 0  # sub-step index (within a predicted frame) of the last executed action
        # Sample one keyframe every ``action_per_frame / temporal_downsample`` executed sub-steps so
        # that exactly ``frame_chunk_size * temporal_downsample`` frames are VAE-encoded per chunk
        # (the Wan2.2 VAE temporal downsample is 4 -> ``frame_chunk_size`` latent frames).
        self._keyframe_stride = max(1, cfg.action_per_frame // 4)
        self._frame_st_id = 0
        self._first_chunk = True
        self._prompt: str | None = None
        self._prompt_embeds = None
        self._negative_prompt_embeds = None
        self.last_predicted_frames = None
        self.last_predicted_latents = None
        self._use_cfg = (cfg.guidance_scale > 1) or (cfg.action_guidance_scale > 1)
        # Two independent flow-matching schedulers (video latent + action streams).
        self._scheduler = FlowMatchScheduler(shift=cfg.snr_shift, sigma_min=0.0, extra_one_step=True)
        self._action_scheduler = FlowMatchScheduler(
            shift=cfg.action_snr_shift, sigma_min=0.0, extra_one_step=True
        )
        self._scheduler.set_timesteps(1000, training=True)
        self._action_scheduler.set_timesteps(1000, training=True)
        self._cache_initialised = False
        # Clear KV cache on the (already-built) transformer, if present.
        if hasattr(self, "transformer"):
            self.transformer.clear_cache("pos")
        # Reset the causal streaming-VAE feat cache between episodes (mirrors upstream ``_reset``).
        # Without this the encoder carries over the previous episode's temporal state, corrupting the
        # latent frame counts on the next episode's first encode.
        if self._frozen:
            self._frozen["streaming_vae"].clear_cache()
            if "streaming_vae_half" in self._frozen:
                self._frozen["streaming_vae_half"].clear_cache()
    # Training (flow-matching dual-stream loss). Requires attn_mode="flex".
    def _ensure_train_schedulers(self):
        if getattr(self, "_train_sched_latent", None) is None:
            cfg = self.config
            self._train_sched_latent = FlowMatchScheduler(
                shift=cfg.snr_shift, sigma_min=0.0, extra_one_step=True
            )
            self._train_sched_latent.set_timesteps(1000, training=True)
            self._train_sched_action = FlowMatchScheduler(
                shift=cfg.action_snr_shift, sigma_min=0.0, extra_one_step=True
            )
            self._train_sched_action.set_timesteps(1000, training=True)
    @torch.no_grad()
    def _add_noise_stream(self, latent, scheduler, action_mask, action_mode, noisy_cond_prob):
        """Flow-matching noising of one stream (port of upstream ``Trainer._add_noise``)."""
        device = latent.device
        b, _c, f, _h, _w = latent.shape
        p = self.config.patch_size
        patch_f, patch_h, patch_w = (1, 1, 1) if action_mode else (p[0], p[1], p[2])
        ts_ids = _sample_timestep_id(f, num_train_timesteps=scheduler.num_train_timesteps)
        noise = torch.zeros_like(latent).normal_()
        timesteps = scheduler.timesteps[ts_ids].to(device)
        noisy_latents = scheduler.add_noise(latent, noise, timesteps, t_dim=2)
        targets = scheduler.training_target(latent, noise, timesteps)
        grid_id = (
            get_mesh_id(
                latent.shape[-3] // patch_f,
                latent.shape[-2] // patch_h,
                latent.shape[-1] // patch_w,
                t=1 if action_mode else 0,
                f_w=1,
                f_shift=0,
                action=action_mode,
            )
            .to(device)[None]
            .repeat(b, 1, 1)
        )
        if torch.rand(1).item() < noisy_cond_prob:
            cond_ids = _sample_timestep_id(
                f, min_timestep_bd=0.5, max_timestep_bd=1.0, num_train_timesteps=scheduler.num_train_timesteps
            )
            cond_noise = torch.zeros_like(latent).normal_()
            cond_timesteps = scheduler.timesteps[cond_ids].to(device)
            latent = scheduler.add_noise(latent, cond_noise, cond_timesteps, t_dim=2)
        else:
            cond_timesteps = torch.zeros_like(timesteps)
        if action_mask is not None:
            noisy_latents = noisy_latents * action_mask.float()
            targets = targets * action_mask.float()
            latent = latent * action_mask.float()
        return {
            "timesteps": timesteps[None].repeat(b, 1),
            "noisy_latents": noisy_latents,
            "targets": targets,
            "latent": latent,
            "cond_timesteps": cond_timesteps[None].repeat(b, 1),
            "grid_id": grid_id,
        }
    def _flow_matching_loss(self, input_dict, pred):
        """Dual-stream flow-matching loss (port of upstream ``Trainer.compute_loss``)."""
        latent_pred, action_pred = pred
        ld, ad = input_dict["latent_dict"], input_dict["action_dict"]
        action_pred = rearrange(action_pred, "b (f n) c -> b c f n 1", f=ad["targets"].shape[-3])
        latent_pred = data_seq_to_patch(
            self.config.patch_size,
            latent_pred,
            ld["targets"].shape[-3],
            ld["targets"].shape[-2],
            ld["targets"].shape[-1],
            batch_size=latent_pred.shape[0],
        )
        bn, fn = ld["timesteps"].shape
        lw = self._train_sched_latent.training_weight(ld["timesteps"].flatten()).reshape(bn, fn)
        aw = self._train_sched_action.training_weight(ad["timesteps"].flatten()).reshape(bn, fn)
        latent_loss = F.mse_loss(latent_pred.float(), ld["targets"].float().detach(), reduction="none")
        latent_loss = (
            (latent_loss * lw[:, None, :, None, None]).permute(0, 2, 3, 4, 1).flatten(0, 1).flatten(1)
        )
        latent_loss = (latent_loss.sum(dim=1) / (torch.ones_like(latent_loss).sum(dim=1) + 1e-6)).mean()
        amask = ad["actions_mask"].float()
        action_loss = F.mse_loss(action_pred.float(), ad["targets"].float().detach(), reduction="none")
        action_loss = (
            (action_loss * aw[:, None, :, None, None] * amask).permute(0, 2, 3, 4, 1).flatten(0, 1).flatten(1)
        )
        amask_f = amask.permute(0, 2, 3, 4, 1).flatten(0, 1).flatten(1)
        action_loss = (action_loss.sum(dim=1) / (amask_f.sum(dim=1) + 1e-6)).mean()
        return latent_loss, action_loss
    def training_loss_from_streams(self, latents, actions, actions_mask, text_emb):
        """Core dual-stream training loss given prepared latents / actions / text embeddings.
        ``latents``: ``[B, in_channels, F, h, w]`` (normalized video latents).
        ``actions`` / ``actions_mask``: ``[B, action_dim, F, action_per_frame, 1]``.
        ``text_emb``: ``[B, seq_len, text_dim]``. Returns ``(loss, {latent_loss, action_loss})``.
        """
        if self.config.attn_mode != "flex":
            raise ValueError(
                "LingBot-VA training requires attn_mode='flex' (block-causal flow-matching masks). "
                "Load/convert the policy with --policy.attn_mode=flex for training/fine-tuning."
            )
        self._ensure_train_schedulers()
        latent_dict = self._add_noise_stream(
            latents, self._train_sched_latent, action_mask=None, action_mode=False, noisy_cond_prob=0.5
        )
        action_dict = self._add_noise_stream(
            actions, self._train_sched_action, action_mask=actions_mask, action_mode=True, noisy_cond_prob=0.0
        )
        latent_dict["text_emb"] = text_emb
        action_dict["text_emb"] = text_emb
        action_dict["actions_mask"] = actions_mask
        input_dict = {
            "latent_dict": latent_dict,
            "action_dict": action_dict,
            "chunk_size": int(torch.randint(1, 5, (1,)).item()),
            "window_size": int(torch.randint(4, 65, (1,)).item()),
        }
        pred = self.transformer(input_dict, train_mode=True)
        latent_loss, action_loss = self._flow_matching_loss(input_dict, pred)
        loss = latent_loss + action_loss
        return loss, {"latent_loss": latent_loss.item(), "action_loss": action_loss.item()}
    def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict | None]:
        """Training forward: dual-stream flow-matching loss.
        Builds the (video-latent, action, text) training streams from a LeRobot batch
        (VAE-encoding the camera frames and UMT5-encoding the task), then runs the flow-matching
        dual-stream loss. Requires the policy to be built with ``attn_mode='flex'``.
        """
        self._ensure_frozen_modules()
        latents, actions, actions_mask, text_emb = self._build_training_streams(batch)
        return self.training_loss_from_streams(latents, actions, actions_mask, text_emb)
    @torch.no_grad()
    def _build_training_streams(self, batch):
        """Build (latents, actions, actions_mask, text_emb) from a LeRobot training batch.
        Camera frames per ``obs_cam_keys`` are expected as a temporal clip ``[B, C, T, H, W]`` (or
        ``[B, T, C, H, W]``); they are VAE-encoded into ``F = T / temporal_downsample`` latent frames.
        Actions ``[B, F*action_per_frame, n_used]`` are scattered into the model's ``action_dim`` space.
        """
        cfg = self.config
        device = cfg.device
        # text embeddings
        task = batch.get("task")
        if isinstance(task, str):
            task = [task]
        text_emb = self._get_t5_prompt_embeds(list(task), cfg.max_sequence_length)
        # video latents (VAE-encode the camera clips)
        latents = self._encode_training_latents(batch)
        # actions -> [B, action_dim, F, action_per_frame, 1]
        act = batch[ACTION].to(device)  # [B, F*apf, n_used]
        b = act.shape[0]
        used = cfg.used_action_channel_ids
        apf, fc = cfg.action_per_frame, cfg.frame_chunk_size
        act = act[:, : fc * apf].reshape(b, fc, apf, len(used)).permute(0, 3, 1, 2)  # [B, n_used, F, apf]
        full = act.new_zeros(b, cfg.action_dim, fc, apf)
        idx = torch.as_tensor(used, device=device)
        full[:, idx] = act
        actions = full.unsqueeze(-1).to(self.dtype)  # [B, action_dim, F, apf, 1]
        mask = torch.zeros(cfg.action_dim, device=device, dtype=self.dtype)
        mask[idx] = 1.0
        actions_mask = mask.view(1, -1, 1, 1, 1).expand_as(actions)
        return latents, actions, actions_mask, text_emb
    @torch.no_grad()
    def _encode_training_latents(self, batch) -> Tensor:
        """VAE-encode the per-camera training clips into normalized video latents [B, C, F, h, w]."""
        vae_device = next(self._vae.parameters()).device
        def _clip(key):
            x = batch[key].to(vae_device)
            if x.dim() == 4:  # [B, C, H, W] -> single frame clip
                x = x.unsqueeze(2)
            elif x.shape[1] not in (1, 3) and x.shape[2] in (1, 3):  # [B, T, C, H, W] -> [B, C, T, H, W]
                x = x.permute(0, 2, 1, 3, 4)
            return x.contiguous()
        def _encode(x, size):
            b, c, t = x.shape[:3]
            x = F.interpolate(x.flatten(0, 1).float(), size=size, mode="bilinear", align_corners=False)
            x = (x.view(b, c, t, *size) * 2.0 - 1.0).to(self.dtype)
            mu = self._vae.encode(x).latent_dist.mode()  # [B, z_dim, F, h, w]
            mean = torch.tensor(self._vae.config.latents_mean).view(1, -1, 1, 1, 1).to(mu.device)
            inv_std = (1.0 / torch.tensor(self._vae.config.latents_std)).view(1, -1, 1, 1, 1).to(mu.device)
            return ((mu.float() - mean) * inv_std).to(mu)
        keys = self.config.obs_cam_keys
        if self.config.camera_layout == "robotwin_tshape":
            h, w = self.config.height, self.config.width
            head = _encode(_clip(keys[0]), (h, w))
            left = _encode(_clip(keys[1]), (h // 2, w // 2))
            right = _encode(_clip(keys[2]), (h // 2, w // 2))
            return torch.cat([torch.cat([left, right], dim=-1), head], dim=-2).to(self.config.device)
        per_cam = [_encode(_clip(k), (self.config.height, self.config.width)) for k in keys]
        return torch.cat(per_cam, dim=-1).to(self.config.device)
    @torch.no_grad()
    def select_action(self, batch: dict[str, Tensor], **kwargs) -> Tensor:
        """Return one action, refilling the chunk (and feeding back observed keyframes) as needed.
        Mirrors the upstream LIBERO client loop (``evaluation/libero/client.py``): the first obs is
        the conditioning frame; every observation produced afterwards is buffered as a keyframe and,
        once the chunk's actions are exhausted, the buffered frames + executed actions are fed back
        into the KV cache before the next chunk is predicted.
        """
        self.eval()
        self._ensure_frozen_modules()
        self._maybe_init_prompt(batch)
        if not self._started:
            # First call: this observation conditions the first chunk (it is *not* a keyframe).
            self._started = True
            actions = self.predict_action_chunk(batch)  # [B, chunk_size, n_used]
            self._action_queue.extend(actions.transpose(0, 1))  # [chunk_size, B, n_used]
            self._obs_buffer = []
            self._exec_step = 0
        else:
            # This observation is the result of the previously executed action -> a candidate
            # keyframe. Buffer it on the sub-step boundary the upstream client samples on.
            if (self._prev_j + 1) % self._keyframe_stride == 0:
                self._obs_buffer.append(self._extract_raw_obs(batch))
            if len(self._action_queue) == 0:
                # All actions for the current chunk have been executed; feed the observed
                # keyframes + executed actions back and predict the next chunk.
                actions = self.predict_action_chunk(None)
                self._action_queue.extend(actions.transpose(0, 1))
                self._exec_step = 0
        self._prev_j = self._exec_step % self.config.action_per_frame
        self._exec_step += 1
        return self._action_queue.popleft()
    @torch.no_grad()
    def predict_action_chunk(self, batch: dict[str, Tensor], **kwargs) -> Tensor:
        """Run one autoregressive chunk and return actions ``[B, chunk_size, n_used]`` (normalized)."""
        self.eval()
        self._ensure_frozen_modules()
        self._maybe_init_prompt(batch)
        is_first = self._first_chunk
        if is_first:
            init_latent = self._encode_frames([self._extract_raw_obs(batch)])
            self._init_latent = init_latent
            self._init_streaming_cache(init_latent)
            self._obs_buffer = []  # frame 0 (the init obs) conditions the chunk; it is not fed back
            actions, latents = self._infer(init_latent, frame_st_id=0)
            self._first_chunk = False
        else:
            # Feed the real observed keyframes + the executed actions back into the KV cache.
            self._compute_kv_cache(self._obs_buffer, self._executed_actions)
            self._obs_buffer = []
            actions, latents = self._infer(None, frame_st_id=self._frame_st_id)
        # actions: [B, action_dim, F, action_per_frame, 1] (model-normalized). Keep for KV feedback.
        self._executed_actions = actions
        if self.config.save_predicted_video:
            # Match upstream LingBot-VA visualization: collect chunk latents and decode the
            # concatenated latent sequence once after the rollout finishes.
            self.last_predicted_frames = None
            self.last_predicted_latents = latents.detach().to("cpu")
        # On the first chunk, frame 0 is the conditioning frame (already "known"): the upstream
        # LIBERO client skips it (start_idx=1), so we drop the first frame's actions here.
        used = self.config.used_action_channel_ids
        a = actions[:, used]  # [B, n_used, F, action_per_frame, 1]
        if is_first:
            a = a[:, :, 1:]  # drop frame 0 -> (F-1) frames of actions
        a = a.squeeze(-1).flatten(2)  # [B, n_used, n_steps]
        a = a.transpose(1, 2).contiguous()  # [B, n_steps, n_used]
        return a.to(torch.float32)
    # Prompt / text encoding
    def _maybe_init_prompt(self, batch):
        if self._prompt_embeds is not None or batch is None:
            return
        task = batch.get("task")
        prompt = task[0] if isinstance(task, list | tuple) else task
        self._prompt = prompt or ""
        self._prompt_embeds, self._negative_prompt_embeds = self._encode_prompt(self._prompt)
    def _get_t5_prompt_embeds(self, prompt, max_sequence_length):
        tokenizer = self._frozen["tokenizer"]
        text_encoder = self._frozen["text_encoder"]
        device = self.config.device
        prompt = [prompt] if isinstance(prompt, str) else prompt
        prompt = [clean_prompt(u) for u in prompt]
        text_inputs = tokenizer(
            prompt,
            padding="max_length",
            max_length=max_sequence_length,
            truncation=True,
            add_special_tokens=True,
            return_attention_mask=True,
            return_tensors="pt",
        )
        text_input_ids, mask = text_inputs.input_ids, text_inputs.attention_mask
        seq_lens = mask.gt(0).sum(dim=1).long()
        te_device = next(text_encoder.parameters()).device
        prompt_embeds = text_encoder(text_input_ids.to(te_device), mask.to(te_device)).last_hidden_state
        prompt_embeds = prompt_embeds.to(dtype=self.dtype, device=device)
        prompt_embeds = [u[:v] for u, v in zip(prompt_embeds, seq_lens, strict=False)]
        prompt_embeds = torch.stack(
            [torch.cat([u, u.new_zeros(max_sequence_length - u.size(0), u.size(1))]) for u in prompt_embeds],
            dim=0,
        )
        return prompt_embeds.to(device)
    def _encode_prompt(self, prompt):
        max_len = self.config.max_sequence_length
        prompt_embeds = self._get_t5_prompt_embeds(prompt, max_len)
        negative_prompt_embeds = None
        if self._use_cfg:
            negative_prompt_embeds = self._get_t5_prompt_embeds("", max_len)
        return prompt_embeds, negative_prompt_embeds
    # Observation (image) encoding -> normalized video latents
    def _extract_raw_obs(self, batch) -> dict[str, Tensor]:
        """Snapshot the configured camera images from a batch (kept raw for later VAE encoding)."""
        return {k: batch[k].detach() for k in self.config.obs_cam_keys}
    def _camera_frame(self, raw_obs, key, size=None) -> Tensor:
        """Return a single-frame camera tensor [1, C, 1, H, W] resized + scaled to [-1, 1]."""
        img = raw_obs[key]
        if img.dim() == 3:  # [C, H, W]
            img = img.unsqueeze(0)
        # LeRobot images arrive as float in [0, 1], shape [B, C, H, W].
        img = img.to(self.config.device, torch.float32)
        if self.config.image_hflip:
            img = torch.flip(img, dims=[-1])  # undo the env processor's horizontal flip
        if size is None:
            size = (self.config.height, self.config.width)
        img = F.interpolate(img, size=size, mode="bilinear", align_corners=False)
        img = img * 2.0 - 1.0
        return img.unsqueeze(2).to(self.dtype)  # [1, C, F=1, H, W]
    def _normalize_vae_latent(self, enc_out: Tensor) -> Tensor:
        """Take the mean of a VAE encoder output and channel-normalize it (matches upstream)."""
        mu, _logvar = torch.chunk(enc_out, 2, dim=1)
        latents_mean = torch.tensor(self._vae.config.latents_mean).to(mu.device)
        latents_std = torch.tensor(self._vae.config.latents_std).to(mu.device)
        mean = latents_mean.view(1, -1, 1, 1, 1)
        inv_std = (1.0 / latents_std).view(1, -1, 1, 1, 1)
        return ((mu.float() - mean) * inv_std).to(mu)
    @torch.no_grad()
    def _encode_frames(self, raw_frames: list) -> Tensor:
        """VAE-encode a temporal clip of observed frames and concat the per-camera latents on width.
        ``raw_frames`` is a list of per-frame obs dicts (one per env sub-step). Each configured
        camera is stacked along the temporal axis into a ``[1, C, F, H, W]`` clip and encoded in a
        single streaming ``encode_chunk`` call so the VAE temporal downsample (x4) collapses the F
        input frames into ``F / 4`` latent frames, with the causal ``feat_cache`` carried across
        chunks (mirrors upstream ``_encode_obs``).
        """
        vae_device = next(self._vae.parameters()).device
        if self.config.camera_layout == "robotwin_tshape":
            return self._encode_frames_tshape(raw_frames, vae_device)
        per_cam_videos = []
        for k in self.config.obs_cam_keys:
            frames = [self._camera_frame(fb, k) for fb in raw_frames]
            per_cam_videos.append(torch.cat(frames, dim=2))  # [1, C, F, H, W]
        videos = torch.cat(per_cam_videos, dim=0)  # [num_cam, C, F, H, W]
        enc_out = self._streaming_vae.encode_chunk(videos.to(vae_device).to(self.dtype))
        mu_norm = self._normalize_vae_latent(enc_out)
        # Concatenate the per-camera latents along width.
        video_latent = torch.cat(mu_norm.split(1, dim=0), dim=-1)
        return video_latent.to(self.config.device)
    @torch.no_grad()
    def _encode_frames_tshape(self, raw_frames: list, vae_device) -> Tensor:
        """RoboTwin T-shape latent assembly: full-res head + half-res wrists (second streaming VAE).
        The two wrist latents are concatenated on width and stacked (on the height axis) on top of
        the head latent, mirroring upstream ``_encode_obs`` for ``env_type='robotwin_tshape'``.
        """
        cfg = self.config
        h, w = cfg.height, cfg.width
        head_key, left_key, right_key = cfg.obs_cam_keys[0], cfg.obs_cam_keys[1], cfg.obs_cam_keys[2]
        head = torch.cat([self._camera_frame(fb, head_key, size=(h, w)) for fb in raw_frames], dim=2)
        left = torch.cat(
            [self._camera_frame(fb, left_key, size=(h // 2, w // 2)) for fb in raw_frames], dim=2
        )
        right = torch.cat(
            [self._camera_frame(fb, right_key, size=(h // 2, w // 2)) for fb in raw_frames], dim=2
        )
        wrists = torch.cat([left, right], dim=0)  # [2, C, F, H/2, W/2]
        enc_high = self._streaming_vae.encode_chunk(head.to(vae_device).to(self.dtype))
        enc_lr = self._frozen["streaming_vae_half"].encode_chunk(wrists.to(vae_device).to(self.dtype))
        # wrists side-by-side on width, then stacked on top of the head latent on the height axis.
        enc_out = torch.cat([torch.cat(enc_lr.split(1, dim=0), dim=-1), enc_high], dim=-2)
        video_latent = self._normalize_vae_latent(enc_out)
        return video_latent.to(self.config.device)
    # KV cache management
    @property
    def _latent_hw(self):
        if self.config.camera_layout == "robotwin_tshape":
            # head (full) on the bottom, two half-res wrists side-by-side on top -> 1.5x height.
            return ((self.config.height // 16) * 3) // 2, self.config.width // 16
        h = self.config.height // 16
        w = (self.config.width // 16) * len(self.config.obs_cam_keys)
        return h, w
    def _init_streaming_cache(self, init_latent):
        cfg = self.config
        latent_h, latent_w = self._latent_hw
        p = cfg.patch_size
        latent_token_per_chunk = (cfg.frame_chunk_size * latent_h * latent_w) // (p[0] * p[1] * p[2])
        action_token_per_chunk = cfg.frame_chunk_size * cfg.action_per_frame
        self.transformer.create_empty_cache(
            "pos",
            cfg.attn_window,
            latent_token_per_chunk,
            action_token_per_chunk,
            device=self.config.device,
            dtype=self.dtype,
            batch_size=2 if self._use_cfg else 1,
        )
        self._cache_initialised = True
    def _repeat_input_for_cfg(self, input_dict):
        if self._use_cfg:
            input_dict["noisy_latents"] = input_dict["noisy_latents"].repeat(2, 1, 1, 1, 1)
            input_dict["text_emb"] = torch.cat(
                [
                    self._prompt_embeds.to(self.dtype).clone(),
                    self._negative_prompt_embeds.to(self.dtype).clone(),
                ],
                dim=0,
            )
            input_dict["grid_id"] = input_dict["grid_id"][None].repeat(2, 1, 1)
            input_dict["timesteps"] = input_dict["timesteps"][None].repeat(2, 1)
        else:
            input_dict["grid_id"] = input_dict["grid_id"][None]
            input_dict["timesteps"] = input_dict["timesteps"][None]
        return input_dict
    def _prepare_latent_input(
        self,
        latent_model_input,
        action_model_input,
        latent_t=0,
        action_t=0,
        latent_cond=None,
        action_cond=None,
        frame_st_id=0,
    ):
        cfg = self.config
        device = self.config.device
        p = cfg.patch_size
        out = {}
        if latent_model_input is not None:
            out["latent_res_lst"] = {
                "noisy_latents": latent_model_input,
                "timesteps": torch.ones([latent_model_input.shape[2]], dtype=torch.float32, device=device)
                * latent_t,
                "grid_id": get_mesh_id(
                    latent_model_input.shape[-3] // p[0],
                    latent_model_input.shape[-2] // p[1],
                    latent_model_input.shape[-1] // p[2],
                    0,
                    1,
                    frame_st_id,
                ).to(device),
                "text_emb": self._prompt_embeds.to(self.dtype).clone(),
            }
            if latent_cond is not None:
                out["latent_res_lst"]["noisy_latents"][:, :, 0:1] = latent_cond[:, :, 0:1]
                out["latent_res_lst"]["timesteps"][0:1] *= 0
        if action_model_input is not None:
            out["action_res_lst"] = {
                "noisy_latents": action_model_input,
                "timesteps": torch.ones([action_model_input.shape[2]], dtype=torch.float32, device=device)
                * action_t,
                "grid_id": get_mesh_id(
                    action_model_input.shape[-3],
                    action_model_input.shape[-2],
                    action_model_input.shape[-1],
                    1,
                    1,
                    frame_st_id,
                    action=True,
                ).to(device),
                "text_emb": self._prompt_embeds.to(self.dtype).clone(),
            }
            if action_cond is not None:
                out["action_res_lst"]["noisy_latents"][:, :, 0:1] = action_cond[:, :, 0:1]
                out["action_res_lst"]["timesteps"][0:1] *= 0
            out["action_res_lst"]["noisy_latents"][:, ~self._action_mask] *= 0
        return out
    @property
    def _action_mask(self):
        mask = torch.zeros([self.config.action_dim], dtype=torch.bool)
        mask[self.config.used_action_channel_ids] = True
        return mask
    # Action conditioning (executed action history) (de)normalization
    def _preprocess_action_state(self, action_norm: Tensor) -> Tensor:
        """Build the action-conditioning tensor from the already-normalized executed actions.
        ``action_norm`` is the model-space action chunk ``[B, action_dim, F, action_per_frame, 1]``.
        Upstream re-derives the conditioning from the raw executed action via quantile norm; here
        the executed actions are already in the model-normalized space, so we pass them through.
        """
        return action_norm.to(self.config.device, self.dtype)
    def _compute_kv_cache(self, obs_buffer, executed_actions):
        """Feed real observed keyframes + executed actions back into the KV cache."""
        if not obs_buffer or executed_actions is None:
            return
        self.transformer.clear_pred_cache("pos")
        # Encode the buffered keyframe clip in one streaming call (carries the causal VAE cache).
        latent_model_input = self._encode_frames(obs_buffer)
        # On the first feedback, prepend the init latent so the latent/action frame counts align
        # (upstream prepends ``init_latent`` to the observed keyframes when frame_st_id == 0).
        if self._frame_st_id == 0 and getattr(self, "_init_latent", None) is not None:
            latent_model_input = torch.cat([self._init_latent, latent_model_input], dim=2)
        action_model_input = self._preprocess_action_state(executed_actions)
        action_model_input = action_model_input.to(latent_model_input)
        input_dict = self._prepare_latent_input(
            latent_model_input, action_model_input, frame_st_id=self._frame_st_id
        )
        with torch.no_grad():
            self.transformer(
                self._repeat_input_for_cfg(input_dict["latent_res_lst"]),
                update_cache=2,
                cache_name="pos",
                action_mode=False,
            )
            self.transformer(
                self._repeat_input_for_cfg(input_dict["action_res_lst"]),
                update_cache=2,
                cache_name="pos",
                action_mode=True,
            )
        self._frame_st_id += latent_model_input.shape[2]
    # The core dual-stream denoising loop (one chunk)
    @torch.no_grad()
    def _infer(self, init_latent, frame_st_id=0):
        cfg = self.config
        device = self.config.device
        latent_h, latent_w = self._latent_hw
        frame_chunk_size = cfg.frame_chunk_size
        latents = torch.randn(1, 48, frame_chunk_size, latent_h, latent_w, device=device, dtype=self.dtype)
        actions = torch.randn(
            1, cfg.action_dim, frame_chunk_size, cfg.action_per_frame, 1, device=device, dtype=self.dtype
        )
        self._scheduler.set_timesteps(cfg.num_inference_steps)
        self._action_scheduler.set_timesteps(cfg.action_num_inference_steps)
        timesteps = F.pad(self._scheduler.timesteps, (0, 1), mode="constant", value=0)
        if cfg.video_exec_step != -1:
            timesteps = timesteps[: cfg.video_exec_step]
        action_timesteps = F.pad(self._action_scheduler.timesteps, (0, 1), mode="constant", value=0)
        # 1. Video-latent denoising loop
        for i, t in enumerate(timesteps):
            last_step = i == len(timesteps) - 1
            latent_cond = (
                init_latent[:, :, 0:1].to(self.dtype)
                if frame_st_id == 0 and init_latent is not None
                else None
            )
            input_dict = self._prepare_latent_input(
                latents, None, t, t, latent_cond, None, frame_st_id=frame_st_id
            )
            video_noise_pred = self.transformer(
                self._repeat_input_for_cfg(input_dict["latent_res_lst"]),
                update_cache=1 if last_step else 0,
                cache_name="pos",
                action_mode=False,
            )
            if not last_step or cfg.video_exec_step != -1:
                video_noise_pred = data_seq_to_patch(
                    cfg.patch_size,
                    video_noise_pred,
                    frame_chunk_size,
                    latent_h,
                    latent_w,
                    batch_size=2 if self._use_cfg else 1,
                )
                if cfg.guidance_scale > 1:
                    video_noise_pred = video_noise_pred[1:] + cfg.guidance_scale * (
                        video_noise_pred[:1] - video_noise_pred[1:]
                    )
                else:
                    video_noise_pred = video_noise_pred[:1]
                latents = self._scheduler.step(video_noise_pred, t, latents, return_dict=False)
            if frame_st_id == 0 and latent_cond is not None:
                latents[:, :, 0:1] = latent_cond
        # 2. Action denoising loop
        for i, t in enumerate(action_timesteps):
            last_step = i == len(action_timesteps) - 1
            action_cond = (
                torch.zeros([1, cfg.action_dim, 1, cfg.action_per_frame, 1], device=device, dtype=self.dtype)
                if frame_st_id == 0
                else None
            )
            input_dict = self._prepare_latent_input(
                None, actions, t, t, None, action_cond, frame_st_id=frame_st_id
            )
            action_noise_pred = self.transformer(
                self._repeat_input_for_cfg(input_dict["action_res_lst"]),
                update_cache=1 if last_step else 0,
                cache_name="pos",
                action_mode=True,
            )
            if not last_step:
                action_noise_pred = rearrange(action_noise_pred, "b (f n) c -> b c f n 1", f=frame_chunk_size)
                if cfg.action_guidance_scale > 1:
                    action_noise_pred = action_noise_pred[1:] + cfg.action_guidance_scale * (
                        action_noise_pred[:1] - action_noise_pred[1:]
                    )
                else:
                    action_noise_pred = action_noise_pred[:1]
                actions = self._action_scheduler.step(action_noise_pred, t, actions, return_dict=False)
            if frame_st_id == 0 and action_cond is not None:
                actions[:, :, 0:1] = action_cond
        actions[:, ~self._action_mask] *= 0
        return actions, latents
    # Predicted-video decoding (opt-in)
    @torch.no_grad()
    def decode_predicted_latents(self, latents) -> Tensor:
        """Decode a concatenated predicted-latent sequence into ``[T, H, W, 3]`` uint8 frames."""
        return self._decode_predicted_video(latents)
    @torch.no_grad()
    def _decode_predicted_video(self, latents) -> Tensor:
        """VAE-decode predicted latents into a uint8 frame stack ``[T, H, W, 3]`` on CPU."""
        vae = self._vae
        z_dim = vae.config.z_dim
        vae_device = next(vae.parameters()).device
        latents = latents.to(device=vae_device, dtype=vae.dtype)
        latents = denormalize_latents(latents, vae.config.latents_mean, vae.config.latents_std, z_dim)
        video = vae.decode(latents, return_dict=False)[0]  # [B, C, F, H, W] in [-1, 1]
        video = (video.float().clamp(-1, 1) + 1.0) / 2.0
        video = (video[0].permute(1, 2, 3, 0) * 255.0).round().to(torch.uint8)  # [F, H, W, C]
        return video.cpu()
@@ -1,87 +0,0 @@
 # Copyright 2026 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
 #     http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """Pre/post-processor pipelines for the LingBot-VA policy.
 The preprocessor passes inputs through (IDENTITY) and the postprocessor maps the policy's
 ``[-1, 1]`` actions back to physical units with the built-in ``UnnormalizerProcessorStep``
 (QUANTILES) using per-channel q01/q99 restored from the checkpoint.
 """
 from typing import Any
 import torch
 from lerobot.configs.types import FeatureType, NormalizationMode
 from lerobot.processor import (
    AddBatchDimensionProcessorStep,
    DeviceProcessorStep,
    NormalizerProcessorStep,
    PolicyAction,
    PolicyProcessorPipeline,
    ProcessorStep,
    RenameObservationsProcessorStep,
    UnnormalizerProcessorStep,
 )
 from lerobot.processor.converters import policy_action_to_transition, transition_to_policy_action
 from lerobot.utils.constants import (
    POLICY_POSTPROCESSOR_DEFAULT_NAME,
    POLICY_PREPROCESSOR_DEFAULT_NAME,
 )
 from .configuration_lingbot_va import LingBotVAConfig
 def make_lingbot_va_pre_post_processors(
    config: LingBotVAConfig,
    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
 ) -> tuple[
    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
    PolicyProcessorPipeline[PolicyAction, PolicyAction],
 ]:
    """Build the pre/post processor pipelines for LingBot-VA."""
    input_steps: list[ProcessorStep] = [
        RenameObservationsProcessorStep(rename_map={}),
        AddBatchDimensionProcessorStep(),
        NormalizerProcessorStep(
            features={**config.input_features, **config.output_features},
            norm_map=config.normalization_mapping,
            stats=dataset_stats,
        ),
        DeviceProcessorStep(device=config.device),
    ]
    # Unnormalize actions from [-1, 1] to physical units (QUANTILES) using q01/q99 restored from the checkpoint.
    output_steps: list[ProcessorStep] = [
        UnnormalizerProcessorStep(
            features=config.output_features,
            norm_map={FeatureType.ACTION: NormalizationMode.QUANTILES},
            stats=dataset_stats,
        ),
        DeviceProcessorStep(device="cpu"),
    ]
    return (
        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
            steps=input_steps,
            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
        ),
        PolicyProcessorPipeline[PolicyAction, PolicyAction](
            steps=output_steps,
            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
            to_transition=policy_action_to_transition,
            to_output=transition_to_policy_action,
        ),
    )
@@ -226,14 +226,11 @@ class RolloutConfig:
    device: str | None = None
    task: str = ""
    display_data: bool = False
-    # Visualization backend used when display_data is True: "rerun" or "foxglove".
+    # Display data on a remote Rerun server
    display_mode: str = "rerun"
    # For "rerun": IP of a remote server to send to. For "foxglove": interface to bind the WebSocket
    # server to (127.0.0.1 for local only, 0.0.0.0 for all interfaces).
    display_ip: str | None = None
-    # For "rerun": port of the remote server. For "foxglove": port to bind the WebSocket server to.
+    # Port of the remote Rerun server
    display_port: int | None = None
-    # Whether to display compressed (JPEG) images instead of raw frames
+    # Whether to display compressed images in Rerun
    display_compressed_images: bool = False
    # Use vocal synthesis to read events
    play_sounds: bool = True
@@ -26,7 +26,7 @@ from lerobot.utils.action_interpolator import ActionInterpolator
 from lerobot.utils.constants import OBS_STR
 from lerobot.utils.feature_utils import build_dataset_frame
 from lerobot.utils.robot_utils import precise_sleep
-from lerobot.utils.visualization_utils import log_visualization_data
+from lerobot.utils.visualization_utils import log_rerun_data
 from ..inference import InferenceEngine
@@ -162,12 +162,11 @@ class RolloutStrategy(abc.ABC):
        action_dict: dict | None,
        runtime_ctx: RuntimeContext,
    ) -> None:
-        """Log observation/action telemetry to the visualization backend if display_data is enabled."""
+        """Log observation/action telemetry to Rerun if display_data is enabled."""
        cfg = runtime_ctx.cfg
        if not cfg.display_data:
            return
-        log_visualization_data(
+        log_rerun_data(
            cfg.display_mode,
            observation=obs_processed,
            action=action_dict,
            compress_images=cfg.display_compressed_images,
@@ -44,7 +44,7 @@ from lerobot.utils.feature_utils import build_dataset_frame
 from lerobot.utils.keyboard_input import init_keyboard_listener
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import log_say
-from lerobot.utils.visualization_utils import log_visualization_data
+from lerobot.utils.visualization_utils import log_rerun_data
 from ..configs import EpisodicStrategyConfig
 from ..context import RolloutContext
@@ -171,7 +171,6 @@ class EpisodicStrategy(RolloutStrategy):
                            fps=fps,
                            control_time_s=reset_time_s,
                            display_data=cfg.display_data,
                            display_mode=cfg.display_mode,
                            display_compressed=display_compressed,
                        )
@@ -260,7 +259,6 @@ class EpisodicStrategy(RolloutStrategy):
        fps: float,
        control_time_s: float,
        display_data: bool,
        display_mode: str,
        display_compressed: bool,
    ) -> None:
        """Reset-phase loop: teleop drives the robot if available, no recording."""
@@ -290,8 +288,7 @@ class EpisodicStrategy(RolloutStrategy):
                if display_data:
                    obs_processed = processors.robot_observation_processor(obs)
-                    log_visualization_data(
+                    log_rerun_data(
                        display_mode,
                        observation=obs_processed,
                        action=act_teleop,
                        compress_images=display_compressed,
@@ -59,18 +59,6 @@ distant$ lerobot-dataset-viz \
 local$ rerun rerun+http://IP:GRPC_PORT/proxy
 ```
 - Visualize data in Foxglove with a seekable, scrubbable timeline:
 ```
 local$ lerobot-dataset-viz \
    --repo-id lerobot/pusht \
    --episode-index 0 \
    --display-mode foxglove
 # then open the Foxglove app and connect to ws://127.0.0.1:8765
 ```
 This starts a Foxglove WebSocket server that serves the episode on demand from the on-disk dataset,
 so you can play/pause and scrub anywhere in the episode using Foxglove's playback controls.
 """
 import argparse
@@ -84,14 +72,10 @@ import torch
 import torch.utils.data
 import tqdm
 from lerobot.configs import DEPTH_MILLIMETER_UNIT
 from lerobot.datasets import LeRobotDataset
-from lerobot.utils.constants import ACTION, DONE, OBS_STATE, REWARD, SUCCESS
+from lerobot.utils.constants import ACTION, DONE, OBS_STATE, REWARD
 from lerobot.utils.utils import init_logging
 DEFAULT_FOXGLOVE_PORT = 8765
 DEFAULT_RERUN_PORT = 9090
 def get_feature_names(dataset: LeRobotDataset, key: str) -> list[str]:
    """Return per-dimension names for a feature from the dataset metadata.
@@ -124,12 +108,6 @@ def to_hwc_uint8_numpy(chw_float32_torch: torch.Tensor) -> np.ndarray:
    return hwc_uint8_numpy
 def to_hwc_float32_numpy(chw_float32_torch: torch.Tensor) -> np.ndarray:
    check_chw_float32(chw_float32_torch)
    hwc_float32_numpy = chw_float32_torch.permute(1, 2, 0).numpy()
    return hwc_float32_numpy
 def build_blueprint_from_dataset(dataset: LeRobotDataset):
    """Build a Rerun blueprint laying out camera images and time series for the given dataset.
@@ -148,43 +126,32 @@ def build_blueprint_from_dataset(dataset: LeRobotDataset):
            names = get_feature_names(dataset, key)
            styling = rr.SeriesLines(names=names)
            views.append(rrb.TimeSeriesView(origin=origin, name=origin, overrides={origin: styling}))
-    for key in (DONE, REWARD, SUCCESS):
+    for key in (DONE, REWARD, "next.success"):
        if key in dataset.features:
            views.append(rrb.TimeSeriesView(origin=key, name=key))
    return rrb.Blueprint(rrb.Grid(*views))
 def to_hwc_uint16_numpy(chw_float32_torch: torch.Tensor) -> np.ndarray:
    check_chw_float32(chw_float32_torch)
    hwc_uint16_numpy = chw_float32_torch.round().type(torch.uint16).permute(1, 2, 0).numpy()
    return hwc_uint16_numpy
 def visualize_dataset(
    dataset: LeRobotDataset,
    episode_index: int,
    batch_size: int = 32,
    num_workers: int = 0,
    mode: str = "local",
-    web_port: int | None = None,
+    web_port: int = 9090,
    grpc_port: int = 9876,
    save: bool = False,
    output_dir: Path | None = None,
    display_compressed_images: bool = False,
    display_mode: str = "rerun",
    host: str = "127.0.0.1",
    autoplay: bool = True,
    **kwargs,
 ) -> Path | None:
    if display_mode == "foxglove":
        from lerobot.utils.foxglove_visualization import serve_foxglove_dataset_playback
        logging.info("Starting Foxglove server")
        serve_foxglove_dataset_playback(
            dataset,
            episode_index,
            host=host,
            port=web_port if web_port is not None else DEFAULT_FOXGLOVE_PORT,
            compress_images=display_compressed_images,
            autoplay=autoplay,
        )
        return None
    if save:
        assert output_dir is not None, (
            "Set an output directory where to write .rrd files with `--output-dir path/to/directory`."
@@ -221,23 +188,14 @@ def visualize_dataset(
    if mode == "distant":
        server_uri = rr.serve_grpc(grpc_port=grpc_port)
        logging.info(f"Connect to a Rerun Server: rerun rerun+http://IP:{grpc_port}/proxy")
-        rr.serve_web_viewer(
+        rr.serve_web_viewer(open_browser=False, web_port=web_port, connect_to=server_uri)
            open_browser=False,
            web_port=web_port if web_port is not None else DEFAULT_RERUN_PORT,
            connect_to=server_uri,
        )
    logging.info("Logging to Rerun")
    # Depth frames and stats are dequantized to the dataset's depth_output_unit on load.
    depth_meter = 1000.0 if dataset.depth_output_unit == DEPTH_MILLIMETER_UNIT else 1.0
    # Use the dataset's q01/q99 depth statistics for robust depth range bounds
    depth_ranges = {}
    for key in dataset.meta.depth_keys:
-        stats = (dataset.meta.stats or {}).get(key)
+        stats = dataset.meta.stats[key]
        if not stats:
            continue
        lo = stats["q01"] if "q01" in stats else stats["min"]
        hi = stats["q99"] if "q99" in stats else stats["max"]
        depth_ranges[key] = (float(np.asarray(lo).item()), float(np.asarray(hi).item()))
@@ -255,12 +213,11 @@ def visualize_dataset(
            # display each camera image (or depth map)
            for key in dataset.meta.camera_keys:
                if key in dataset.meta.depth_keys:
-                    depth = to_hwc_float32_numpy(batch[key][i])
+                    depth = to_hwc_uint16_numpy(batch[key][i])
                    depth_entity = rr.DepthImage(
                        depth,
                        meter=depth_meter,
                        colormap=rr.components.Colormap.Viridis,
-                        depth_range=depth_ranges.get(key),
+                        depth_range=depth_ranges[key],
                    )
                    rr.log(key, entity=depth_entity)
                else:
@@ -282,8 +239,8 @@ def visualize_dataset(
            if REWARD in batch:
                rr.log(REWARD, rr.Scalars(batch[REWARD][i].item()))
-            if SUCCESS in batch:
+            if "next.success" in batch:
-                rr.log(SUCCESS, rr.Scalars(batch[SUCCESS][i].item()))
+                rr.log("next.success", rr.Scalars(batch["next.success"][i].item()))
    # save .rrd locally
    if mode == "local" and save:
@@ -355,11 +312,13 @@ def main():
    parser.add_argument(
        "--web-port",
        type=int,
-        default=None,
+        default=9090,
-        help=(
+        help="Web port for rerun.io when `--mode distant` is set.",
-            "Web/WebSocket port. For rerun `--mode distant` it is the web viewer port (default 9090); "
+    )
-            "for `--display-mode foxglove` it is the server bind port (default 8765)."
+    parser.add_argument(
-        ),
+        "--ws-port",
        type=int,
        help="deprecated, please use --grpc-port instead.",
    )
    parser.add_argument(
        "--grpc-port",
@@ -392,56 +351,24 @@ def main():
    parser.add_argument(
        "--display-compressed-images",
        action="store_true",
-        help="If set, display compressed (JPEG) images instead of uncompressed ones.",
+        help="If set, display compressed images in Rerun instead of uncompressed ones.",
    )
    parser.add_argument(
        "--display-mode",
        type=str,
        default="rerun",
        choices=["rerun", "foxglove"],
        help=(
            "Visualization backend. 'rerun' uses the Rerun viewer (--mode/--save/--*-port apply). "
            "'foxglove' starts a Foxglove WebSocket server that serves the episode as a seekable, "
            "scrubbable timeline; connect the Foxglove app to ws://HOST:PORT (--host/--web-port)."
        ),
    )
    parser.add_argument(
        "--host",
        type=str,
        default="127.0.0.1",
        help=(
            "Host to bind the Foxglove WebSocket server to when `--display-mode foxglove` is set "
            "(127.0.0.1 for local only, 0.0.0.0 for all interfaces)."
        ),
    )
    parser.add_argument(
        "--no-autoplay",
        dest="autoplay",
        action="store_false",
        help=(
            "For `--display-mode foxglove`: don't start playing automatically when a client "
            "connects; wait for play to be pressed in the Foxglove app instead."
        ),
    )
    args = parser.parse_args()
    if args.display_mode == "foxglove":
        rerun_only = ("mode", "save", "output_dir", "grpc_port", "batch_size", "num_workers")
        ignored = [name for name in rerun_only if getattr(args, name) != parser.get_default(name)]
        if ignored:
            logging.warning(
                "These flags only apply to `--display-mode rerun` and are ignored with "
                "`--display-mode foxglove`: %s.",
                ", ".join(f"--{name.replace('_', '-')}" for name in ignored),
            )
    kwargs = vars(args)
    repo_id = kwargs.pop("repo_id")
    root = kwargs.pop("root")
    tolerance_s = kwargs.pop("tolerance_s")
    if kwargs["ws_port"] is not None:
        logging.warning(
            "--ws-port is deprecated and will be removed in future versions. Please use --grpc-port instead."
        )
        logging.warning("Setting grpc_port to ws_port value.")
        kwargs["grpc_port"] = kwargs.pop("ws_port")
    else:
        kwargs.pop("ws_port")  # Always remove ws_port from kwargs
    init_logging()
    logging.info("Loading dataset")
    dataset = LeRobotDataset(repo_id, episodes=[args.episode_index], root=root, tolerance_s=tolerance_s)
@@ -169,7 +169,6 @@ def rollout(
    env_features: dict | None = None,
    recording_repo_id: str | None = None,
    recording_private: bool = False,
    predicted_latents_callback: Callable[[PreTrainedPolicy], None] | None = None,
 ) -> dict:
    """Run a batched policy rollout once through a batch of environments.
@@ -199,9 +198,6 @@ def rollout(
            are returned optionally because they typically take more memory to cache. Defaults to False.
        render_callback: Optional rendering callback to be used after the environments are reset, and after
            every step.
        predicted_latents_callback: Optional callback invoked after every ``select_action`` with the policy
            itself. World-model policies (e.g. LingBot-VA) stash predicted video latents on
            ``policy.last_predicted_latents``; this lets the caller concatenate chunks and decode once.
    Returns:
        The dictionary described above.
    """
@@ -280,8 +276,6 @@ def rollout(
            observation = preprocessor(observation)
            with torch.inference_mode():
                action = policy.select_action(observation)
            if predicted_latents_callback is not None:
                predicted_latents_callback(policy)
            action = postprocessor(action)
            action_transition = {ACTION: action}
@@ -301,22 +295,12 @@ def rollout(
            # available if none of the envs finished.
            if "final_info" in info:
                final_info = info["final_info"]
-                if isinstance(final_info, dict):
+                if not isinstance(final_info, dict):
-                    is_success = final_info.get("is_success", [False] * env.num_envs)
+                    raise RuntimeError(
-                    successes = (
+                        "Unsupported `final_info` format: expected dict (Gymnasium >= 1.0). "
-                        is_success.tolist()
+                        "You're likely using an older version of gymnasium (< 1.0). Please upgrade."
                        if hasattr(is_success, "tolist")
                        else [bool(is_success)] * env.num_envs
                    )
-                else:
+                successes = final_info["is_success"].tolist()
                    # Gymnasium < 1.0 returns final_info as a per-env sequence/object array,
                    # with entries set to a dict only for envs that just finished.
                    successes = []
                    for item in final_info:
                        if isinstance(item, dict) and "is_success" in item:
                            successes.append(bool(item["is_success"]))
                        else:
                            successes.append(False)
            elif "is_success" in info:
                is_success = info["is_success"]
                successes = (
@@ -416,7 +400,6 @@ def eval_policy(
    env_features: dict | None = None,
    recording_repo_id: str | None = None,
    recording_private: bool = False,
    save_predicted_video: bool = False,
 ) -> dict:
    """
    Args:
@@ -435,11 +418,6 @@ def eval_policy(
    if max_episodes_rendered > 0 and not videos_dir:
        raise ValueError("If max_episodes_rendered > 0, videos_dir must be provided.")
    # World-model policies (e.g. LingBot-VA) opt into predicted-video saving via their config.
    save_predicted_video = save_predicted_video or bool(
        getattr(getattr(policy, "config", None), "save_predicted_video", False)
    )
    if not isinstance(policy, PreTrainedPolicy):
        exc = ValueError(
            f"Policy of type 'PreTrainedPolicy' is expected, but type '{type(policy)}' was provided."
@@ -483,22 +461,6 @@ def eval_policy(
    if max_episodes_rendered > 0:
        video_paths: list[str] = []
    if save_predicted_video:
        if not videos_dir:
            raise ValueError("If save_predicted_video is True, videos_dir must be provided.")
        predicted_video_paths: list[str] = []
        n_predicted_rendered = 0
    # Collect predicted-video latents across a rollout (world-model policies only). The latents are
    # concatenated and decoded once after the rollout, matching upstream LingBot-VA's visualization path.
    def collect_predicted_latents(policy: PreTrainedPolicy):
        latents = getattr(policy, "last_predicted_latents", None)
        if latents is not None:
            pred_latents.append(
                latents.detach().to("cpu") if hasattr(latents, "detach") else torch.as_tensor(latents).cpu()
            )
            policy.last_predicted_latents = None
    if return_episode_data:
        episode_data: dict | None = None
@@ -510,9 +472,6 @@ def eval_policy(
        if max_episodes_rendered > 0:
            ep_frames: list[np.ndarray] = []
        if save_predicted_video:
            pred_latents: list[torch.Tensor] = []
        if start_seed is None:
            seeds = None
        else:
@@ -533,7 +492,6 @@ def eval_policy(
            env_features=env_features,
            recording_repo_id=recording_repo_id,
            recording_private=recording_private,
            predicted_latents_callback=collect_predicted_latents if save_predicted_video else None,
        )
        # Figure out where in each rollout sequence the first done condition was encountered (results after
@@ -599,35 +557,6 @@ def eval_policy(
                threads.append(thread)
                n_episodes_rendered += 1
        # Maybe save the policy's predicted (imagined) video for this batch's rollout.
        if save_predicted_video and len(pred_latents) > 0:
            predicted_latent = torch.cat(pred_latents, dim=2)
            decoder = getattr(policy, "decode_predicted_latents", None) or getattr(
                policy, "_decode_predicted_video", None
            )
            if decoder is None:
                raise AttributeError(
                    "Policy config requested predicted-video saving, but the policy does not expose "
                    "`decode_predicted_latents` or `_decode_predicted_video`."
                )
            predicted_video = decoder(predicted_latent)
            if hasattr(predicted_video, "detach"):
                predicted_video = predicted_video.detach().to("cpu").numpy()
            videos_dir.mkdir(parents=True, exist_ok=True)
            predicted_video_path = videos_dir / f"pred_episode_{n_predicted_rendered}.mp4"
            predicted_video_paths.append(str(predicted_video_path))
            thread = threading.Thread(
                target=write_video,
                args=(
                    str(predicted_video_path),
                    predicted_video,
                    env.unwrapped.metadata["render_fps"],
                ),
            )
            thread.start()
            threads.append(thread)
            n_predicted_rendered += 1
        progbar.set_postfix(
            {"running_success_rate": f"{np.mean(all_successes[:n_episodes]).item() * 100:.1f}%"}
        )
@@ -671,9 +600,6 @@ def eval_policy(
    if max_episodes_rendered > 0:
        info["video_paths"] = video_paths
    if save_predicted_video:
        info["predicted_video_paths"] = predicted_video_paths
    return info
@@ -814,10 +740,9 @@ class TaskMetrics(TypedDict):
    max_rewards: list[float]
    successes: list[bool]
    video_paths: list[str]
    predicted_video_paths: list[str]
-ACC_KEYS = ("sum_rewards", "max_rewards", "successes", "video_paths", "predicted_video_paths")
+ACC_KEYS = ("sum_rewards", "max_rewards", "successes", "video_paths")
 def eval_one(
@@ -866,7 +791,6 @@ def eval_one(
        max_rewards=[ep["max_reward"] for ep in per_episode],
        successes=[ep["success"] for ep in per_episode],
        video_paths=task_result.get("video_paths", []),
        predicted_video_paths=task_result.get("predicted_video_paths", []),
    )
@@ -927,7 +851,6 @@ def run_one(
    if max_episodes_rendered > 0:
        metrics.setdefault("video_paths", [])
    metrics.setdefault("predicted_video_paths", [])
    return task_group, task_id, metrics
@@ -985,11 +908,11 @@ def eval_policy_all(
        _append("sum_rewards", metrics.get("sum_rewards"))
        _append("max_rewards", metrics.get("max_rewards"))
        _append("successes", metrics.get("successes"))
-        for key in ("video_paths", "predicted_video_paths"):
+        # video_paths is list-like
-            paths = metrics.get(key, [])
+        paths = metrics.get("video_paths", [])
-            if paths:
+        if paths:
-                group_acc[group][key].extend(paths)
+            group_acc[group]["video_paths"].extend(paths)
-                overall[key].extend(paths)
+            overall["video_paths"].extend(paths)
    # Choose runner (sequential vs threaded)
    task_runner = partial(
@@ -1061,7 +984,6 @@ def eval_policy_all(
            "pc_success": _agg_from_list(acc["successes"]) * 100 if acc["successes"] else float("nan"),
            "n_episodes": len(acc["sum_rewards"]),
            "video_paths": list(acc["video_paths"]),
            "predicted_video_paths": list(acc["predicted_video_paths"]),
        }
    # overall aggregates
@@ -1073,7 +995,6 @@ def eval_policy_all(
        "eval_s": time.time() - start_t,
        "eval_ep_s": (time.time() - start_t) / max(1, len(overall["sum_rewards"])),
        "video_paths": list(overall["video_paths"]),
        "predicted_video_paths": list(overall["predicted_video_paths"]),
    }
    return {
@@ -38,9 +38,6 @@ lerobot-record \\
    --display_data=true
 ```
 To stream the data to Foxglove instead of Rerun, add ``--display_mode=foxglove`` (then connect the
 Foxglove app to ``ws://127.0.0.1:8765``; override the port with ``--display_port=<port>``).
 Example recording with bimanual so100:
 ```shell
 lerobot-record \\
@@ -160,11 +157,7 @@ from lerobot.utils.utils import (
    init_logging,
    log_say,
 )
-from lerobot.utils.visualization_utils import (
+from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
    init_visualization,
    log_visualization_data,
    shutdown_visualization,
 )
@dataclass
@@ -175,14 +168,11 @@ class RecordConfig:
    teleop: TeleoperatorConfig | None = None
    # Display all cameras on screen
    display_data: bool = False
-    # Visualization backend used when display_data is True: "rerun" or "foxglove".
+    # Display data on a remote Rerun server
    display_mode: str = "rerun"
    # For "rerun": IP of a remote server to send to. For "foxglove": interface to bind the WebSocket
    # server to (127.0.0.1 for local only, 0.0.0.0 for all interfaces).
    display_ip: str | None = None
-    # For "rerun": port of the remote server. For "foxglove": port to bind the WebSocket server to.
+    # Port of the remote Rerun server
    display_port: int | None = None
-    # Whether to display compressed (JPEG) images instead of raw frames
+    # Whether to  display compressed images in Rerun
    display_compressed_images: bool = False
    # Use vocal synthesis to read events.
    play_sounds: bool = True
@@ -243,7 +233,6 @@ def record_loop(
    control_time_s: int | None = None,
    single_task: str | None = None,
    display_data: bool = False,
    display_mode: str = "rerun",
    display_compressed_images: bool = False,
 ):
    if dataset is not None and dataset.fps != fps:
@@ -338,11 +327,8 @@ def record_loop(
            dataset.add_frame(frame)
        if display_data:
-            log_visualization_data(
+            log_rerun_data(
-                display_mode,
+                observation=obs_processed, action=action_values, compress_images=display_compressed_images
                observation=obs_processed,
                action=action_values,
                compress_images=display_compressed_images,
            )
        dt_s = time.perf_counter() - start_loop_t
@@ -368,9 +354,7 @@ def record(
    init_logging()
    logging.info(pformat(asdict(cfg)))
    if cfg.display_data:
-        init_visualization(
+        init_rerun(session_name="recording", ip=cfg.display_ip, port=cfg.display_port)
            cfg.display_mode, session_name="recording", ip=cfg.display_ip, port=cfg.display_port
        )
    display_compressed_images = (
        True
        if (cfg.display_data and cfg.display_ip is not None and cfg.display_port is not None)
@@ -480,7 +464,6 @@ def record(
                    control_time_s=cfg.dataset.episode_time_s,
                    single_task=cfg.dataset.single_task,
                    display_data=cfg.display_data,
                    display_mode=cfg.display_mode,
                    display_compressed_images=display_compressed_images,
                )
@@ -502,7 +485,6 @@ def record(
                        control_time_s=cfg.dataset.reset_time_s,
                        single_task=cfg.dataset.single_task,
                        display_data=cfg.display_data,
                        display_mode=cfg.display_mode,
                    )
                if events["rerecord_episode"]:
@@ -528,9 +510,6 @@ def record(
        if listener is not None:
            listener.stop()
        if cfg.display_data:
            shutdown_visualization(cfg.display_mode)
        if cfg.dataset.push_to_hub:
            if dataset and dataset.num_episodes > 0:
                dataset.push_to_hub(tags=cfg.dataset.tags, private=cfg.dataset.private)
@@ -145,9 +145,6 @@ Usage examples
        --dataset.rgb_encoder.vcodec=h264 \\
        --dataset.rgb_encoder.preset=fast \\
        --dataset.rgb_encoder.extra_options={"tune": "film", "profile:v": "high", "bf": 2}
    # Stream to Foxglove instead of Rerun:
    # add --display_mode=foxglove, then connect the Foxglove app to ws://127.0.0.1:8765.
 """
 import logging
@@ -193,7 +190,7 @@ from lerobot.teleoperators import (  # noqa: F401
 from lerobot.utils.import_utils import register_third_party_plugins
 from lerobot.utils.process import ProcessSignalHandler
 from lerobot.utils.utils import init_logging
-from lerobot.utils.visualization_utils import init_visualization, shutdown_visualization
+from lerobot.utils.visualization_utils import init_rerun
 logger = logging.getLogger(__name__)
@@ -204,13 +201,8 @@ def rollout(cfg: RolloutConfig):
    init_logging()
    if cfg.display_data:
-        logger.info(
+        logger.info("Initializing Rerun visualization (ip=%s, port=%s)", cfg.display_ip, cfg.display_port)
-            "Initializing %s visualization (ip=%s, port=%s)",
+        init_rerun(session_name="rollout", ip=cfg.display_ip, port=cfg.display_port)
            cfg.display_mode,
            cfg.display_ip,
            cfg.display_port,
        )
        init_visualization(cfg.display_mode, session_name="rollout", ip=cfg.display_ip, port=cfg.display_port)
    signal_handler = ProcessSignalHandler(use_threads=True, display_pid=False)
    shutdown_event = signal_handler.shutdown_event
@@ -235,8 +227,6 @@ def rollout(cfg: RolloutConfig):
        logger.info("Interrupted by user")
    finally:
        strategy.teardown(ctx)
        if cfg.display_data:
            shutdown_visualization(cfg.display_mode)
    logger.info("Rollout finished")
@@ -31,22 +31,6 @@ lerobot-teleoperate \
    --display_data=true
 ```
 To stream the data to Foxglove instead of Rerun, add ``--display_mode=foxglove``
 (then connect the Foxglove app to ``ws://127.0.0.1:8765``; override the port with ``--display_port=<port>``):
 ```shell
 lerobot-teleoperate \
    --robot.type=so101_follower \
    --robot.port=/dev/tty.usbmodem58760431541 \
    --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 1920, height: 1080, fps: 30}}" \
    --robot.id=black \
    --teleop.type=so101_leader \
    --teleop.port=/dev/tty.usbmodem58760431551 \
    --teleop.id=blue \
    --display_data=true \
    --display_mode=foxglove
 ```
 Example teleoperation with bimanual so100:
 ```shell
@@ -124,11 +108,7 @@ from lerobot.teleoperators import (  # noqa: F401
 from lerobot.utils.import_utils import register_third_party_plugins
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import init_logging, move_cursor_up
-from lerobot.utils.visualization_utils import (
+from lerobot.utils.visualization_utils import init_rerun, log_rerun_data, shutdown_rerun
    init_visualization,
    log_visualization_data,
    shutdown_visualization,
 )
@dataclass
@@ -141,14 +121,11 @@ class TeleoperateConfig:
    teleop_time_s: float | None = None
    # Display all cameras on screen
    display_data: bool = False
-    # Visualization backend used when display_data is True: "rerun" or "foxglove".
+    # Display data on a remote Rerun server
    display_mode: str = "rerun"
    # For "rerun": IP of a remote server to send to. For "foxglove": interface to bind the WebSocket
    # server to (127.0.0.1 for local only, 0.0.0.0 for all interfaces).
    display_ip: str | None = None
-    # For "rerun": port of the remote server. For "foxglove": port to bind the WebSocket server to.
+    # Port of the remote Rerun server
    display_port: int | None = None
-    # Whether to display compressed (JPEG) images instead of raw frames
+    # Whether to  display compressed images in Rerun
    display_compressed_images: bool = False
@@ -160,7 +137,6 @@ def teleop_loop(
    robot_action_processor: RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction],
    robot_observation_processor: RobotProcessorPipeline[RobotObservation, RobotObservation],
    display_data: bool = False,
    display_mode: str = "rerun",
    duration: float | None = None,
    display_compressed_images: bool = False,
 ):
@@ -173,10 +149,8 @@ def teleop_loop(
        teleop: The teleoperator device instance providing control actions.
        robot: The robot instance being controlled.
        fps: The target frequency for the control loop in frames per second.
-        display_data: If True, fetches robot observations and displays them in the console and the
+        display_data: If True, fetches robot observations and displays them in the console and Rerun.
-            visualization backend.
+        display_compressed_images: If True, compresses images before sending them to Rerun for display.
        display_mode: Visualization backend to use when display_data is True ("rerun" or "foxglove").
        display_compressed_images: If True, compresses images before sending them to the backend for display.
        duration: The maximum duration of the teleoperation loop in seconds. If None, the loop runs indefinitely.
        teleop_action_processor: An optional pipeline to process raw actions from the teleoperator.
        robot_action_processor: An optional pipeline to process actions before they are sent to the robot.
@@ -213,8 +187,7 @@ def teleop_loop(
            # Process robot observation through pipeline
            obs_transition = robot_observation_processor(obs)
-            log_visualization_data(
+            log_rerun_data(
                display_mode,
                observation=obs_transition,
                action=teleop_action,
                compress_images=display_compressed_images,
@@ -242,9 +215,7 @@ def teleoperate(cfg: TeleoperateConfig):
    init_logging()
    logging.info(pformat(asdict(cfg)))
    if cfg.display_data:
-        init_visualization(
+        init_rerun(session_name="teleoperation", ip=cfg.display_ip, port=cfg.display_port)
            cfg.display_mode, session_name="teleoperation", ip=cfg.display_ip, port=cfg.display_port
        )
    display_compressed_images = (
        True
        if (cfg.display_data and cfg.display_ip is not None and cfg.display_port is not None)
@@ -264,7 +235,6 @@ def teleoperate(cfg: TeleoperateConfig):
            robot=robot,
            fps=cfg.fps,
            display_data=cfg.display_data,
            display_mode=cfg.display_mode,
            duration=cfg.teleop_time_s,
            teleop_action_processor=teleop_action_processor,
            robot_action_processor=robot_action_processor,
@@ -275,7 +245,7 @@ def teleoperate(cfg: TeleoperateConfig):
        pass
    finally:
        if cfg.display_data:
-            shutdown_visualization(cfg.display_mode)
+            shutdown_rerun()
        teleop.disconnect()
        robot.disconnect()
@@ -211,12 +211,8 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
        # Accelerate auto-detects the device based on the available hardware and ignores the policy.device setting.
        # Force the device to be CPU when the active config's device is set to CPU (works for both policy and reward model training).
        force_cpu = cfg.trainable_config.device == "cpu"
        # Drive Accelerate's autocast from policy.dtype (bf16/fp16 activate it; float32/absent -> launcher default).
        policy_dtype = getattr(cfg.trainable_config, "dtype", None)
        mixed_precision = {"bfloat16": "bf16", "float16": "fp16", "float32": "no"}.get(policy_dtype)
        accelerator = Accelerator(
            step_scheduler_with_optimizer=False,
            mixed_precision=mixed_precision,
            kwargs_handlers=[ddp_kwargs],
            cpu=force_cpu,
        )
@@ -37,7 +37,6 @@ ACTION_TOKEN_MASK = ACTION + ".token_mask"
 REWARD = "next.reward"
 TRUNCATED = "next.truncated"
 DONE = "next.done"
 SUCCESS = "next.success"
 INFO = "info"
 ROBOTS = "robots"
@@ -1,651 +0,0 @@
 # Copyright 2024 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
 #     http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """Foxglove visualization backend.
 Live control-loop streaming (:func:`log_foxglove_data`) and seekable dataset playback
 (:func:`serve_foxglove_dataset_playback`) over a Foxglove WebSocket server. Callers usually select a
 backend at runtime through the dispatch in :mod:`lerobot.utils.visualization_utils` rather than
 importing from here directly. Requires the ``viz`` extra (``pip install 'lerobot[viz]'``).
 """
 import logging
 import numbers
 import time
 import cv2
 import numpy as np
 from lerobot.types import RobotAction, RobotObservation
 from .constants import (
    ACTION,
    ACTION_PREFIX,
    DONE,
    OBS_IMAGES,
    OBS_PREFIX,
    OBS_STATE,
    OBS_STR,
    REWARD,
    SUCCESS,
    TRUNCATED,
 )
 from .import_utils import require_package
 # Static schema shared by all scalar topics. Each message carries a flat list of ``{label, value}``
 # pairs rather than one field per feature, so the same schema fits any robot regardless of which
 # observation/action features it reports. The ``label`` field name is what Foxglove looks for to name
 # each series automatically, so a single filtered path plots every feature, e.g.
 # ``/observation/state.scalars[:]``.
 _SCALARS_SCHEMA = {
    "type": "object",
    "title": "lerobot.Scalars",
    "properties": {
        "scalars": {
            "type": "array",
            "items": {
                "type": "object",
                "properties": {
                    "label": {"type": "string"},
                    "value": {"type": "number"},
                },
            },
        }
    },
 }
 def _is_scalar(x):
    return isinstance(x, (float | numbers.Real | np.integer | np.floating)) or (
        isinstance(x, np.ndarray) and x.ndim == 0
    )
 def init_foxglove(host: str = "127.0.0.1", port: int | None = 8765) -> None:
    """
    Starts a Foxglove WebSocket server for visualizing the control loop.
    Connect to it from the Foxglove app at ``ws://<host>:<port>``. Calling this
    more than once is a no-op while a server is already running.
    Args:
        host: Host interface to bind the WebSocket server to.
        port: Port to bind the WebSocket server to (defaults to 8765).
    """
    require_package("foxglove-sdk", extra="viz", import_name="foxglove")
    import foxglove
    # Live-stream state lives as attributes on ``log_foxglove_data``:
    # ``.server`` is the shared WebSocket server and
    # ``.channels`` caches one Foxglove channel per topic
    if getattr(log_foxglove_data, "server", None) is not None:
        return
    log_foxglove_data.server = foxglove.start_server(host=host, port=port or 8765)
    log_foxglove_data.channels = {}
 def shutdown_foxglove() -> None:
    """Stops the Foxglove WebSocket server and clears cached channels."""
    server = getattr(log_foxglove_data, "server", None)
    if server is not None:
        server.stop()
    log_foxglove_data.server = None
    log_foxglove_data.channels = {}
 def _foxglove_safe_name(name: str) -> str:
    """Replace ``.`` with ``_`` so a feature name is a single Foxglove topic-path segment.
    Foxglove treats ``.`` as a path separator, so an unsanitized name like ``observation.images.front``
    would split into nested segments instead of naming one topic.
    """
    return name.replace(".", "_")
 def _foxglove_topic(key: str, *, is_image: bool = False) -> str:
    """Build the Foxglove topic for a feature ``key``.
    Camera features map to a per-source image topic (``/observation/images/<name>``); scalar features
    share one aggregate topic per source: ``/observation/state`` for observations, ``/action/state``
    for actions.
    """
    if is_image:
        name = str(key)
        for prefix in (f"{OBS_IMAGES}.", OBS_PREFIX):
            if name.startswith(prefix):
                name = name[len(prefix) :]
                break
        return f"/{OBS_STR}/images/{_foxglove_safe_name(name)}"
    source = ACTION if (str(key).startswith(ACTION_PREFIX) or str(key) == ACTION) else OBS_STR
    return f"/{source}/state"
 def _log_foxglove_scalars(
    topic: str, values: dict[str, float], *, channels: dict | None = None, log_time: int | None = None
 ) -> None:
    """Log scalars on a typed JSON channel using the static :data:`_SCALARS_SCHEMA`.
    ``values`` is an ordered mapping of feature name to value; it is emitted as a ``scalars`` array of
    ``{label, value}`` objects. Insertion order is preserved so series stay stable across messages.
    ``channels`` is the per-topic channel cache to reuse (defaults to the live-stream cache on
    :func:`log_foxglove_data`; dataset playback passes its own local cache to stay self-contained).
    ``log_time`` is the message time in nanoseconds; when ``None`` the server's receive time is used.
    """
    if not values:
        return
    import foxglove
    if channels is None:
        channels = log_foxglove_data.channels
    channel = channels.get(topic)
    if channel is None:
        channel = channels[topic] = foxglove.Channel(topic, schema=_SCALARS_SCHEMA, message_encoding="json")
    msg = {"scalars": [{"label": label, "value": value} for label, value in values.items()]}
    if log_time is None:
        channel.log(msg)
    else:
        channel.log(msg, log_time=log_time)
 def _labeled_scalars(name: str, values, labels: list[str] | None = None) -> dict[str, float]:
    """Expand a 1D sequence into ``{label: value}`` entries with a consistent fallback."""
    flat = [float(v) for v in values]
    if labels is None or len(labels) != len(flat):
        labels = [f"{name}_{i}" for i in range(len(flat))]
    return dict(zip(labels, flat, strict=True))
 def _log_foxglove_image(
    topic: str,
    frame_id: str,
    arr: np.ndarray,
    *,
    compress_images: bool,
    channels: dict | None = None,
    log_time: int | None = None,
    depth_range: tuple[float, float] | None = None,
    raw_depth_values: bool = False,
 ) -> None:
    """Log an image on a cached per-topic channel.
    The encoding is chosen from the channel count and dtype: a single-channel ``float`` or ``uint16``
    frame is a depth map (``32FC1``/``16UC1``), single-channel ``uint8`` is ``mono8``, 3 => ``rgb8``
    (float input assumed in [0, 1], cast to uint8), 4 => ``rgba8``; other counts are skipped with a
    warning. When ``compress_images`` is set, ``rgb8`` is JPEG-encoded instead.
    Args:
        topic: Foxglove topic to log on.
        frame_id: Frame id stamped on the message.
        arr: Image as HWC or CHW (CHW is transposed to HWC), any dtype.
        compress_images: JPEG-encode ``rgb8`` frames; ignored for other encodings.
        channels: Per-topic channel cache to reuse (see :func:`_log_foxglove_scalars`).
        log_time: Message time in nanoseconds, also written to the header timestamp; when ``None``
            the server's receive time is used.
        depth_range: ``(lo, hi)`` clip bounds in a depth frame's own input units. Depth frames
            (``32FC1``/``16UC1``) are rescaled onto Foxglove's default display max for their encoding
            (``1.0`` / ``10000``) so they show with sensible contrast; ``depth_range`` sets the source
            range, else the frame's own min/max is used. Ignored for ``mono8``/``rgb8``/``rgba8``.
        raw_depth_values: If True, depth values are not rescaled and are logged as is.
    """
    from foxglove.channels import CompressedImageChannel, RawImageChannel
    from foxglove.messages import CompressedImage, RawImage, Timestamp
    if channels is None:
        channels = log_foxglove_data.channels
    time_ns = time.time_ns() if log_time is None else log_time
    timestamp = Timestamp(sec=time_ns // 1_000_000_000, nsec=time_ns % 1_000_000_000)
    log_kwargs = {} if log_time is None else {"log_time": log_time}
    # Convert CHW -> HWC when needed (mirrors log_rerun_data).
    if arr.ndim == 3 and arr.shape[0] in (1, 3, 4) and arr.shape[-1] not in (1, 3, 4):
        arr = np.transpose(arr, (1, 2, 0))
    height, width = arr.shape[0], arr.shape[1]
    n_channels = 1 if arr.ndim == 2 else arr.shape[2]
    if n_channels == 1 and arr.dtype != np.uint8:
        # Depth map: infer the encoding from the dtype.
        encoding, target_dtype, value_max = (
            ("32FC1", np.float32, 1.0)
            if np.issubdtype(arr.dtype, np.floating)
            else ("16UC1", np.uint16, 10000.0)
        )
        if not raw_depth_values:
            # Rescale onto the encoding's display max with respect to the given depth_range.
            lo, hi = depth_range if depth_range is not None else (float(arr.min()), float(arr.max()))
            arr = arr.clip(lo, hi).astype(np.float32)
            arr = (arr - lo) / ((hi - lo) if hi > lo else 1.0) * value_max
        arr = np.ascontiguousarray(arr, dtype=target_dtype)
    else:
        if n_channels == 3 and np.issubdtype(arr.dtype, np.floating):
            arr = (arr * 255.0).clip(0, 255)
        arr = np.ascontiguousarray(arr, dtype=np.uint8)
        if compress_images and n_channels == 3:
            buf_src = cv2.cvtColor(arr, cv2.COLOR_RGB2BGR)
            _, buf = cv2.imencode(".jpg", buf_src)
            channel = channels.get(topic)
            if channel is None:
                channel = channels[topic] = CompressedImageChannel(topic=topic)
            channel.log(
                CompressedImage(timestamp=timestamp, frame_id=frame_id, data=buf.tobytes(), format="jpeg"),
                **log_kwargs,
            )
            return
        encoding = {1: "mono8", 3: "rgb8", 4: "rgba8"}.get(n_channels)
        if encoding is None:
            logging.warning(
                "Foxglove: skipping image on topic '%s' with unsupported shape %s (%d channels); "
                "expected 1 (mono8/16UC1/32FC1), 3 (rgb8), or 4 (rgba8) channels.",
                topic,
                tuple(arr.shape),
                n_channels,
            )
            return
    channel = channels.get(topic)
    if channel is None:
        channel = channels[topic] = RawImageChannel(topic=topic)
    channel.log(
        RawImage(
            timestamp=timestamp,
            frame_id=frame_id,
            width=width,
            height=height,
            encoding=encoding,
            step=width * n_channels * arr.itemsize,
            data=arr.tobytes(),
        ),
        **log_kwargs,
    )
 def log_foxglove_data(
    observation: RobotObservation | None = None,
    action: RobotAction | None = None,
    compress_images: bool = False,
 ) -> None:
    """
    Logs observation and action data to a Foxglove WebSocket server for real-time visualization.
    Mirrors ``log_rerun_data`` but emits Foxglove messages over the server started by
    :func:`init_foxglove`. Data is mapped as follows:
    - Scalars (and elements of 1D arrays) are accumulated per source and logged on the
      ``/observation/state`` and ``/action/state`` topics as typed JSON messages using the static
      ``lerobot.Scalars`` schema: a ``scalars`` array of ``{label, value}`` objects (see
      :data:`_SCALARS_SCHEMA`). The ``label`` field lets Foxglove name each series automatically, so
      ``/observation/state.scalars[:].value`` plots every feature at once.
    - 3D NumPy arrays that resemble images are transposed from CHW to HWC when needed and logged on a
      per-source topic (e.g. ``/observation/images/front``) as a ``RawImage`` (or a JPEG
      ``CompressedImage`` when ``compress_images`` is True).
    Args:
        observation: An optional dictionary containing observation data to log.
        action: An optional dictionary containing action data to log.
        compress_images: Whether to JPEG-compress images before logging to save bandwidth in exchange
            for CPU and quality.
    """
    require_package("foxglove-sdk", extra="viz", import_name="foxglove")
    if getattr(log_foxglove_data, "server", None) is None:
        raise RuntimeError("init_foxglove() must be called before log_foxglove_data().")
    now = time.time_ns()
    if observation:
        obs_scalars: dict[str, float] = {}
        for k, v in observation.items():
            if v is None:
                continue
            key = k[len(OBS_PREFIX) :] if str(k).startswith(OBS_PREFIX) else str(k)
            if _is_scalar(v):
                obs_scalars[key] = float(v)
            elif isinstance(v, np.ndarray):
                if v.ndim == 1:
                    obs_scalars.update(_labeled_scalars(key, v))
                else:
                    _log_foxglove_image(
                        _foxglove_topic(k, is_image=True),
                        key,
                        v,
                        compress_images=compress_images,
                        log_time=now,
                    )
        _log_foxglove_scalars(_foxglove_topic(OBS_STATE), obs_scalars, log_time=now)
    if action:
        action_scalars: dict[str, float] = {}
        for k, v in action.items():
            if v is None:
                continue
            key = k[len(ACTION_PREFIX) :] if str(k).startswith(ACTION_PREFIX) else str(k)
            if _is_scalar(v):
                action_scalars[key] = float(v)
            elif isinstance(v, np.ndarray):
                action_scalars.update(_labeled_scalars(key, v.flatten()))
        _log_foxglove_scalars(_foxglove_topic(ACTION), action_scalars, log_time=now)
 # ── Dataset playback over a Foxglove WebSocket server ─────────────────────
 # A LeRobotDataset is random-access on disk, so rather than fire-and-forget a forward stream we
 # advertise a seekable timeline and serve frames on demand for whatever time the user scrubs/plays
 # to in the Foxglove app. This relies on the SDK's PlaybackControl capability.
 def _feature_dim_names(feature: dict | None) -> list[str] | None:
    """Best-effort per-dimension series labels for a 1D feature, or ``None`` to fall back to indices.
    LeRobot records a feature's ``names`` inconsistently: a flat list (``["x", "y"]``), a category
    mapping (``{"motors": ["motor_0", "motor_1"]}``), or a name->index mapping
    (``{"delta_x": 0, "delta_y": 1}``). Each is handled, but labels are only returned when their count
    matches the feature's 1D shape, so a malformed/mismatched ``names`` can't silently mislabel series.
    """
    if not feature:
        return None
    shape = feature.get("shape")
    dim = shape[0] if shape and len(shape) == 1 else None
    names = feature.get("names")
    labels: list[str] | None = None
    if isinstance(names, dict):
        values = list(names.values())
        if values and all(isinstance(v, (list, tuple)) for v in values):
            labels = [str(n) for group in values for n in group]
        elif values and all(isinstance(v, int) and not isinstance(v, bool) for v in values):
            labels = [name for name, _ in sorted(names.items(), key=lambda kv: kv[1])]
    elif isinstance(names, (list, tuple)):
        labels = [str(n) for n in names]
    if labels is not None and dim is not None and len(labels) == dim:
        return labels
    return None
 def _frame_to_scalars(sample: dict, key: str, labels: list[str] | None = None) -> dict[str, float]:
    """Flatten a frame's vector/scalar feature ``key`` into ``{label: value}`` entries.
    ``labels`` provides one name per dimension (from the dataset's feature metadata); when absent or
    the wrong length, dimensions fall back to ``{name}_{i}`` (the short feature name), matching the
    live stream so series names agree. A scalar feature becomes a single entry. Missing or ``None``
    features yield an empty mapping.
    """
    v = sample.get(key)
    if v is None:
        return {}
    arr = v.numpy() if hasattr(v, "numpy") else np.asarray(v)
    if key.startswith(OBS_PREFIX):
        name = key[len(OBS_PREFIX) :]
    elif key.startswith(ACTION_PREFIX):
        name = key[len(ACTION_PREFIX) :]
    else:
        name = key
    if arr.ndim == 0:
        return {name: float(arr)}
    return _labeled_scalars(name, arr.flatten(), labels)
 def serve_foxglove_dataset_playback(
    dataset,
    episode_index: int,
    *,
    host: str = "127.0.0.1",
    port: int = 8765,
    compress_images: bool = False,
    autoplay: bool = True,
 ) -> None:
    """Serve a single dataset episode to Foxglove as a seekable, scrubbable timeline.
    Starts a Foxglove WebSocket server advertising the ``PlaybackControl`` capability over the
    episode's time range. The Foxglove app drives play/pause/seek/speed; a background thread and a
    ``ServerListener`` read frames from the on-disk ``dataset`` on demand and log them stamped at
    their dataset timestamps, so the user can scrub anywhere in the episode. Blocks until interrupted.
    Args:
        dataset: A ``LeRobotDataset`` loaded for the single episode to visualize.
        episode_index: Index of the episode being visualized (used only for the session name).
        host: Host interface to bind the WebSocket server to.
        port: Port to bind the WebSocket server to.
        compress_images: Whether to JPEG-compress camera frames before logging.
        autoplay: If True, start playing automatically as soon as a client connects, instead of
            waiting for the user to press play in the Foxglove app.
    """
    require_package("foxglove-sdk", extra="viz", import_name="foxglove")
    import bisect
    import threading
    import foxglove
    from foxglove.websocket import (
        Capability,
        PlaybackCommand,
        PlaybackControlRequest,
        PlaybackState,
        PlaybackStatus,
        ServerListener,
    )
    # Per-frame timestamps in nanoseconds (read straight from the table, no video decode).
    times_ns = [int(round(float(t) * 1e9)) for t in dataset.hf_dataset["timestamp"]]
    n_frames = len(times_ns)
    if n_frames == 0:
        raise ValueError("Cannot visualize an empty episode.")
    first_ns, last_ns = times_ns[0], times_ns[-1]
    camera_keys = list(dataset.meta.camera_keys)
    # Dataset-wide q01/q99 depth bounds (fallback min/max) used to normalize depth to [0, 1].
    depth_ranges: dict[str, tuple[float, float]] = {}
    for key in dataset.meta.depth_keys:
        stats = (dataset.meta.stats or {}).get(key)
        if not stats:
            continue
        lo = stats["q01"] if "q01" in stats else stats["min"]
        hi = stats["q99"] if "q99" in stats else stats["max"]
        depth_ranges[key] = (float(np.asarray(lo).item()), float(np.asarray(hi).item()))
    # Per-dimension series labels from the dataset metadata (e.g. joint names), computed once.
    scalar_labels = {
        OBS_STATE: _feature_dim_names(dataset.meta.features.get(OBS_STATE)),
        ACTION: _feature_dim_names(dataset.meta.features.get(ACTION)),
    }
    # Local channel cache so the playback server is self-contained and doesn't touch the live-stream cache.
    channels: dict = {}
    def emit_frame(i: int) -> None:
        """Log every channel for frame ``i`` stamped at its dataset timestamp."""
        sample = dataset[i]
        log_time = times_ns[i]
        for key in camera_keys:
            arr = sample.get(key)
            if arr is None:
                continue
            arr = arr.numpy() if hasattr(arr, "numpy") else np.asarray(arr)
            _log_foxglove_image(
                _foxglove_topic(key, is_image=True),
                key,
                arr,
                compress_images=compress_images,
                channels=channels,
                log_time=log_time,
                depth_range=depth_ranges.get(key),
                raw_depth_values=True,
            )
        _log_foxglove_scalars(
            _foxglove_topic(OBS_STATE),
            _frame_to_scalars(sample, OBS_STATE, scalar_labels[OBS_STATE]),
            channels=channels,
            log_time=log_time,
        )
        _log_foxglove_scalars(
            _foxglove_topic(ACTION),
            _frame_to_scalars(sample, ACTION, scalar_labels[ACTION]),
            channels=channels,
            log_time=log_time,
        )
        episode_scalars = {}
        for feat, label in (
            (DONE, "done"),
            (TRUNCATED, "truncated"),
            (REWARD, "reward"),
            (SUCCESS, "success"),
        ):
            v = sample.get(feat)
            if v is not None:
                episode_scalars[label] = float(v)
        _log_foxglove_scalars("/episode/state", episode_scalars, channels=channels, log_time=log_time)
    lock = threading.Lock()
    stop_event = threading.Event()
    # Shared playback state, guarded by ``lock``. ``seek_idx`` is a one-shot request set by the
    # listener and serviced by the playback loop, which is the *only* thread that emits frames (so
    # concurrent random access into the on-disk dataset / video decoder never overlaps).
    state = {
        "status": PlaybackStatus.Paused,
        "cursor": first_ns,
        "speed": 1.0,
        "last_idx": -1,
        "seek_idx": None,
    }
    def index_at(t_ns: int) -> int:
        return max(0, min(n_frames - 1, bisect.bisect_right(times_ns, t_ns) - 1))
    # One-shot latch so autoplay fires only on the first client subscription.
    autoplay_started = threading.Event()
    class _PlaybackListener(ServerListener):
        def on_subscribe(self, client, channel):
            # Start playing automatically once a client actually connects (subscribes). Using the
            # subscribe hook, rather than starting in Playing up front, means the timeline doesn't
            # advance before anyone is watching. Fires once; the user can still pause/seek after.
            if not autoplay:
                return
            with lock:
                if autoplay_started.is_set() or state["status"] != PlaybackStatus.Paused:
                    return
                autoplay_started.set()
                state["status"] = PlaybackStatus.Playing
                cursor, speed = state["cursor"], state["speed"]
            server.broadcast_playback_state(PlaybackState(PlaybackStatus.Playing, cursor, speed, False, ""))
        def on_playback_control_request(self, req: PlaybackControlRequest):
            # Only mutate state here; the playback loop performs all frame emission.
            with lock:
                did_seek = False
                if req.seek_time is not None:
                    cursor = max(first_ns, min(last_ns, req.seek_time))
                    state["cursor"] = cursor
                    state["last_idx"] = state["seek_idx"] = index_at(cursor)
                    did_seek = True
                if req.playback_speed and req.playback_speed > 0:
                    state["speed"] = req.playback_speed
                if req.playback_command == PlaybackCommand.Play:
                    # Restarting from the end replays from the beginning.
                    if state["cursor"] >= last_ns:
                        state["cursor"] = first_ns
                        state["last_idx"] = state["seek_idx"] = 0
                        did_seek = True
                    state["status"] = PlaybackStatus.Playing
                elif req.playback_command == PlaybackCommand.Pause:
                    state["status"] = PlaybackStatus.Paused
                status, cursor, speed = state["status"], state["cursor"], state["speed"]
                request_id = req.request_id or ""
            return PlaybackState(status, cursor, speed, did_seek, request_id)
    server = foxglove.start_server(
        name=f"{dataset.repo_id}/episode_{episode_index}",
        host=host,
        port=port,
        capabilities=[Capability.PlaybackControl, Capability.Time],
        server_listener=_PlaybackListener(),
        playback_time_range=(first_ns, last_ns),
    )
    def playback_loop() -> None:
        # Cap how far the cursor may advance in a single tick. A slow frame decode (or any stall)
        # would otherwise make ``dt`` huge and produce one enormous catch-up batch; clamping it makes
        # playback trail wall-clock under a slow decoder while each tick emits a bounded frame range.
        max_tick_dt_s = 0.25
        prev = time.monotonic()
        while not stop_event.is_set():
            time.sleep(1.0 / 60.0)
            ended = False
            speed = 1.0
            with lock:
                now = time.monotonic()
                dt = min(now - prev, max_tick_dt_s)
                prev = now
                # A queued seek is always serviced, even while paused, so scrubbing updates the view.
                work = []
                seek_idx = state["seek_idx"]
                if seek_idx is not None:
                    state["seek_idx"] = None
                    work.append(seek_idx)
                if state["status"] == PlaybackStatus.Playing:
                    cursor = state["cursor"] + int(dt * 1e9 * state["speed"])
                    start_idx = state["last_idx"] + 1
                    if cursor >= last_ns:
                        cursor, target, ended = last_ns, n_frames - 1, True
                    else:
                        target = index_at(cursor)
                    state["cursor"] = cursor
                    work.extend(range(start_idx, target + 1))
                    # cursor only grows while playing (seeks reset last_idx in the listener), so
                    # target >= last_idx here; a plain assignment is correct and clearer than max().
                    state["last_idx"] = target
                    if ended:
                        state["status"] = PlaybackStatus.Ended
                if not work:
                    continue
                cursor, speed = state["cursor"], state["speed"]
            # Emit outside the lock; this is the only thread that calls emit_frame. Re-check
            # stop_event between frames so shutdown stays responsive even mid-batch.
            for i in work:
                if stop_event.is_set():
                    break
                emit_frame(i)
            server.broadcast_time(cursor)
            if ended:
                server.broadcast_playback_state(PlaybackState(PlaybackStatus.Ended, cursor, speed, False, ""))
    # Emit the first frame so channels are advertised (done before the loop starts, so emission stays
    # single-threaded). Late-connecting clients re-receive frames once they seek/play.
    emit_frame(0)
    with lock:
        state["last_idx"] = 0
    server.broadcast_time(first_ns)
    server.broadcast_playback_state(PlaybackState(PlaybackStatus.Paused, first_ns, 1.0, True, ""))
    thread = threading.Thread(target=playback_loop, name="foxglove-playback", daemon=True)
    thread.start()
    print(f"Foxglove server running. Connect the Foxglove app to ws://{host}:{port}")
    print("Use the playback controls in Foxglove to play/pause and scrub the episode. Ctrl-C to exit.")
    try:
        while not stop_event.is_set():
            time.sleep(0.5)
    except KeyboardInterrupt:
        print("Ctrl-C received. Exiting.")
    finally:
        stop_event.set()
        thread.join(timeout=2.0)
        server.stop()
        channels.clear()
@@ -129,7 +129,6 @@ _placo_available = is_package_available("placo")
 _hidapi_available = is_package_available("hidapi", import_name="hid")
 # Data / serialization
 _datasets_available = is_package_available("datasets")
 _pandas_available = is_package_available("pandas")
 _faker_available = is_package_available("faker")
@@ -1,191 +0,0 @@
 # Copyright 2024 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
 #     http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """Rerun visualization backend.
 Live control-loop streaming to the Rerun viewer (:func:`log_rerun_data`). Callers usually select a
 backend at runtime through the dispatch in :mod:`lerobot.utils.visualization_utils` rather than
 importing from here directly. Requires the ``viz`` extra (``pip install 'lerobot[viz]'``).
 """
 import numbers
 import os
 import numpy as np
 from lerobot.configs import DEPTH_MILLIMETER_UNIT, infer_depth_unit
 from lerobot.types import RobotAction, RobotObservation
 from .constants import ACTION, ACTION_PREFIX, OBS_PREFIX, OBS_STR
 from .import_utils import require_package
 def _is_scalar(x):
    return isinstance(x, (float | numbers.Real | np.integer | np.floating)) or (
        isinstance(x, np.ndarray) and x.ndim == 0
    )
 def init_rerun(
    session_name: str = "lerobot_control_loop", ip: str | None = None, port: int | None = None
 ) -> None:
    """
    Initializes the Rerun SDK for visualizing the control loop.
    Args:
        session_name: Name of the Rerun session.
        ip: Optional IP for connecting to a Rerun server.
        port: Optional port for connecting to a Rerun server.
    """
    require_package("rerun-sdk", extra="viz", import_name="rerun")
    import rerun as rr
    log_rerun_data.blueprint = None  # Reset blueprint cache for new session
    batch_size = os.getenv("RERUN_FLUSH_NUM_BYTES", "8000")
    os.environ["RERUN_FLUSH_NUM_BYTES"] = batch_size
    rr.init(session_name)
    memory_limit = os.getenv("LEROBOT_RERUN_MEMORY_LIMIT", "10%")
    if ip and port:
        rr.connect_grpc(url=f"rerun+http://{ip}:{port}/proxy")
    else:
        rr.spawn(memory_limit=memory_limit)
 def shutdown_rerun() -> None:
    """Shuts down the Rerun SDK gracefully."""
    require_package("rerun-sdk", extra="viz", import_name="rerun")
    import rerun as rr
    rr.rerun_shutdown()
 def _build_blueprint(observation_paths: set[str], action_paths: set[str], image_paths: set[str]):
    """Build a Rerun blueprint laying out camera images, observation and action scalars in separate views.
    Camera images, observation and action scalars are arranged in a grid.
    """
    # Safe + zero-overhead: `log_rerun_data` already ran the `require_package` guard and imported rerun.
    import rerun.blueprint as rrb
    views = [rrb.Spatial2DView(origin=path, name=path) for path in sorted(image_paths)]
    if observation_paths:
        views.append(rrb.TimeSeriesView(name="observation", contents=sorted(observation_paths)))
    if action_paths:
        views.append(rrb.TimeSeriesView(name="action", contents=sorted(action_paths)))
    return rrb.Blueprint(rrb.Grid(*views))
 def _ensure_blueprint(observation_paths: set[str], action_paths: set[str], image_paths: set[str]) -> None:
    """Build and send the blueprint once, from the first observation and action data."""
    if getattr(log_rerun_data, "blueprint", None) is not None:
        return
    if not (observation_paths or action_paths or image_paths):
        return
    # Safe + zero-overhead: `log_rerun_data` already ran the `require_package` guard and imported rerun.
    import rerun as rr
    blueprint = _build_blueprint(observation_paths, action_paths, image_paths)
    log_rerun_data.blueprint = blueprint
    rr.send_blueprint(blueprint)
 def log_rerun_data(
    observation: RobotObservation | None = None,
    action: RobotAction | None = None,
    compress_images: bool = False,
 ) -> None:
    """
    Logs observation and action data to Rerun for real-time visualization.
    This function iterates through the provided observation and action dictionaries and sends their contents
    to the Rerun viewer. It handles different data types appropriately:
    - Scalars values (floats, ints) are logged as `rr.Scalars`.
    - 3D NumPy arrays that resemble images (e.g., with 1, 3, or 4 channels first) are transposed
      from CHW to HWC format, (optionally) compressed to JPEG and logged as `rr.Image` or `rr.EncodedImage`.
    - 1D NumPy arrays are logged as a single `rr.Scalars` batch under one entity path, so that every
      dimension shares the same view instead of being split across one view per element.
    - Multi-dimensional **action** arrays are flattened and logged as a single `rr.Scalars` batch.
    Keys are automatically namespaced with "observation." or "action." if not already present.
    On the first call, a blueprint is built and sent so observation and action scalars get separate
    time-series views and each image gets its own spatial view.
    Args:
        observation: An optional dictionary containing observation data to log.
        action: An optional dictionary containing action data to log.
        compress_images: Whether to compress images before logging to save bandwidth & memory in exchange for cpu and quality.
    """
    require_package("rerun-sdk", extra="viz", import_name="rerun")
    import rerun as rr
    observation_paths: set[str] = set()
    action_paths: set[str] = set()
    image_paths: set[str] = set()
    if observation:
        for k, v in observation.items():
            if v is None:
                continue
            key = k if str(k).startswith(OBS_PREFIX) else f"{OBS_STR}.{k}"
            if _is_scalar(v):
                rr.log(key, rr.Scalars(float(v)))
                observation_paths.add(key)
            elif isinstance(v, np.ndarray):
                arr = v
                # Convert CHW -> HWC when needed
                if arr.ndim == 3 and arr.shape[0] in (1, 3, 4) and arr.shape[-1] not in (1, 3, 4):
                    arr = np.transpose(arr, (1, 2, 0))
                if arr.ndim == 1:
                    rr.log(key, rr.Scalars(arr.astype(float)))
                    observation_paths.add(key)
                else:
                    if arr.shape[-1] == 1:
                        # At record time, the depth unit is inferred from the frame type.
                        depth_unit = infer_depth_unit(arr.dtype)
                        img_entity = rr.DepthImage(
                            arr,
                            meter=1000.0 if depth_unit == DEPTH_MILLIMETER_UNIT else 1.0,
                            colormap=rr.components.Colormap.Viridis,
                        )
                    else:
                        img_entity = rr.Image(arr).compress() if compress_images else rr.Image(arr)
                    rr.log(key, entity=img_entity, static=True)
                    image_paths.add(key)
    if action:
        for k, v in action.items():
            if v is None:
                continue
            key = k if str(k).startswith(ACTION_PREFIX) else f"{ACTION}.{k}"
            if _is_scalar(v):
                rr.log(key, rr.Scalars(float(v)))
                action_paths.add(key)
            elif isinstance(v, np.ndarray):
                # Flatten any (incl. higher-dimensional) array into a single batched Scalars
                rr.log(key, rr.Scalars(v.reshape(-1).astype(float)))
                action_paths.add(key)
    _ensure_blueprint(observation_paths, action_paths, image_paths)
@@ -12,68 +12,166 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
-"""Backend-agnostic visualization dispatch.
+import numbers
 import os
-Selects a visualization backend at runtime via a display-mode string (e.g. a ``--display_mode`` CLI
+import numpy as np
 flag) so callers never branch on the backend. The concrete implementations live in
 :mod:`lerobot.utils.rerun_visualization` and :mod:`lerobot.utils.foxglove_visualization`; importing
 this module does not import ``rerun`` or ``foxglove`` (each backend imports its SDK lazily behind a
 ``require_package`` guard).
 """
 from lerobot.types import RobotAction, RobotObservation
-from .foxglove_visualization import init_foxglove, log_foxglove_data, shutdown_foxglove
+from .constants import ACTION, ACTION_PREFIX, OBS_PREFIX, OBS_STR
-from .rerun_visualization import init_rerun, log_rerun_data, shutdown_rerun
+from .import_utils import require_package
 # Visualization backends selectable at runtime via a display-mode string (e.g. a --display_mode flag).
 VISUALIZATION_MODES = ("rerun", "foxglove")
-def init_visualization(
+def init_rerun(
-    display_mode: str,
+    session_name: str = "lerobot_control_loop", ip: str | None = None, port: int | None = None
    *,
    session_name: str = "lerobot_control_loop",
    ip: str | None = None,
    port: int | None = None,
 ) -> None:
-    """Initializes the visualization backend selected by ``display_mode``.
+    """
    Initializes the Rerun SDK for visualizing the control loop.
-    For ``"rerun"``, ``ip``/``port`` point at an optional remote Rerun server. For ``"foxglove"``,
+    Args:
-    ``ip`` is the interface to bind the WebSocket server to (``127.0.0.1`` for local only, ``0.0.0.0``
+        session_name: Name of the Rerun session.
-    for all interfaces) and ``port`` is its port.
+        ip: Optional IP for connecting to a Rerun server.
        port: Optional port for connecting to a Rerun server.
    """
-    if display_mode == "rerun":
+    require_package("rerun-sdk", extra="viz", import_name="rerun")
-        init_rerun(session_name=session_name, ip=ip, port=port)
+    import rerun as rr
-    elif display_mode == "foxglove":
+
-        init_foxglove(host=ip or "127.0.0.1", port=port)
+    log_rerun_data.blueprint = None  # Reset blueprint cache for new session
    batch_size = os.getenv("RERUN_FLUSH_NUM_BYTES", "8000")
    os.environ["RERUN_FLUSH_NUM_BYTES"] = batch_size
    rr.init(session_name)
    memory_limit = os.getenv("LEROBOT_RERUN_MEMORY_LIMIT", "10%")
    if ip and port:
        rr.connect_grpc(url=f"rerun+http://{ip}:{port}/proxy")
    else:
-        raise ValueError(f"Unknown display_mode '{display_mode}'. Expected one of {VISUALIZATION_MODES}.")
+        rr.spawn(memory_limit=memory_limit)
-def log_visualization_data(
+def shutdown_rerun() -> None:
-    display_mode: str,
+    """Shuts down the Rerun SDK gracefully."""
    require_package("rerun-sdk", extra="viz", import_name="rerun")
    import rerun as rr
    rr.rerun_shutdown()
 def _is_scalar(x):
    return isinstance(x, (float | numbers.Real | np.integer | np.floating)) or (
        isinstance(x, np.ndarray) and x.ndim == 0
    )
 def _build_blueprint(observation_paths: set[str], action_paths: set[str], image_paths: set[str]):
    """Build a Rerun blueprint laying out camera images, observation and action scalars in separate views.
    Camera images, observation and action scalars are arranged in a grid.
    """
    # Safe + zero-overhead: `log_rerun_data` already ran the `require_package` guard and imported rerun.
    import rerun.blueprint as rrb
    views = [rrb.Spatial2DView(origin=path, name=path) for path in sorted(image_paths)]
    if observation_paths:
        views.append(rrb.TimeSeriesView(name="observation", contents=sorted(observation_paths)))
    if action_paths:
        views.append(rrb.TimeSeriesView(name="action", contents=sorted(action_paths)))
    return rrb.Blueprint(rrb.Grid(*views))
 def _ensure_blueprint(observation_paths: set[str], action_paths: set[str], image_paths: set[str]) -> None:
    """Build and send the blueprint once, from the first observation and action data."""
    if getattr(log_rerun_data, "blueprint", None) is not None:
        return
    if not (observation_paths or action_paths or image_paths):
        return
    # Safe + zero-overhead: `log_rerun_data` already ran the `require_package` guard and imported rerun.
    import rerun as rr
    blueprint = _build_blueprint(observation_paths, action_paths, image_paths)
    log_rerun_data.blueprint = blueprint
    rr.send_blueprint(blueprint)
 def log_rerun_data(
    observation: RobotObservation | None = None,
    action: RobotAction | None = None,
    compress_images: bool = False,
 ) -> None:
-    """Logs observation/action data to the backend selected by ``display_mode``."""
+    """
    Logs observation and action data to Rerun for real-time visualization.
-    if display_mode == "rerun":
+    This function iterates through the provided observation and action dictionaries and sends their contents
-        log_rerun_data(observation=observation, action=action, compress_images=compress_images)
+    to the Rerun viewer. It handles different data types appropriately:
-    elif display_mode == "foxglove":
+    - Scalars values (floats, ints) are logged as `rr.Scalars`.
-        log_foxglove_data(observation=observation, action=action, compress_images=compress_images)
+    - 3D NumPy arrays that resemble images (e.g., with 1, 3, or 4 channels first) are transposed
-    else:
+      from CHW to HWC format, (optionally) compressed to JPEG and logged as `rr.Image` or `rr.EncodedImage`.
-        raise ValueError(f"Unknown display_mode '{display_mode}'. Expected one of {VISUALIZATION_MODES}.")
+    - 1D NumPy arrays are logged as a single `rr.Scalars` batch under one entity path, so that every
      dimension shares the same view instead of being split across one view per element.
    - Multi-dimensional **action** arrays are flattened and logged as a single `rr.Scalars` batch.
    Keys are automatically namespaced with "observation." or "action." if not already present.
-def shutdown_visualization(display_mode: str) -> None:
+    On the first call, a blueprint is built and sent so observation and action scalars get separate
-    """Shuts down the backend selected by ``display_mode``."""
+    time-series views and each image gets its own spatial view.
-    if display_mode == "rerun":
+    Args:
-        shutdown_rerun()
+        observation: An optional dictionary containing observation data to log.
-    elif display_mode == "foxglove":
+        action: An optional dictionary containing action data to log.
-        shutdown_foxglove()
+        compress_images: Whether to compress images before logging to save bandwidth & memory in exchange for cpu and quality.
-    else:
+    """
-        raise ValueError(f"Unknown display_mode '{display_mode}'. Expected one of {VISUALIZATION_MODES}.")
+
    require_package("rerun-sdk", extra="viz", import_name="rerun")
    import rerun as rr
    observation_paths: set[str] = set()
    action_paths: set[str] = set()
    image_paths: set[str] = set()
    if observation:
        for k, v in observation.items():
            if v is None:
                continue
            key = k if str(k).startswith(OBS_PREFIX) else f"{OBS_STR}.{k}"
            if _is_scalar(v):
                rr.log(key, rr.Scalars(float(v)))
                observation_paths.add(key)
            elif isinstance(v, np.ndarray):
                arr = v
                # Convert CHW -> HWC when needed
                if arr.ndim == 3 and arr.shape[0] in (1, 3, 4) and arr.shape[-1] not in (1, 3, 4):
                    arr = np.transpose(arr, (1, 2, 0))
                if arr.ndim == 1:
                    rr.log(key, rr.Scalars(arr.astype(float)))
                    observation_paths.add(key)
                else:
                    if arr.shape[-1] == 1:
                        img_entity = rr.DepthImage(arr, colormap=rr.components.Colormap.Viridis)
                    else:
                        img_entity = rr.Image(arr).compress() if compress_images else rr.Image(arr)
                    rr.log(key, entity=img_entity, static=True)
                    image_paths.add(key)
    if action:
        for k, v in action.items():
            if v is None:
                continue
            key = k if str(k).startswith(ACTION_PREFIX) else f"{ACTION}.{k}"
            if _is_scalar(v):
                rr.log(key, rr.Scalars(float(v)))
                action_paths.add(key)
            elif isinstance(v, np.ndarray):
                # Flatten any (incl. higher-dimensional) array into a single batched Scalars
                rr.log(key, rr.Scalars(v.reshape(-1).astype(float)))
                action_paths.add(key)
    _ensure_blueprint(observation_paths, action_paths, image_paths)
@@ -1531,7 +1531,6 @@ def test_valid_video_codecs_constant():
    assert "h264" in VALID_VIDEO_CODECS
    assert "hevc" in VALID_VIDEO_CODECS
    assert "libsvtav1" in VALID_VIDEO_CODECS
    assert "libaom-av1" in VALID_VIDEO_CODECS
    assert "auto" in VALID_VIDEO_CODECS
    assert "h264_videotoolbox" in VALID_VIDEO_CODECS
    assert "h264_nvenc" in VALID_VIDEO_CODECS
@@ -1539,7 +1538,7 @@ def test_valid_video_codecs_constant():
    assert "h264_qsv" in VALID_VIDEO_CODECS
    assert "hevc_videotoolbox" in VALID_VIDEO_CODECS
    assert "hevc_nvenc" in VALID_VIDEO_CODECS
-    assert len(VALID_VIDEO_CODECS) == 11
+    assert len(VALID_VIDEO_CODECS) == 10
 def test_delta_timestamps_with_episodes_filter(tmp_path, empty_lerobot_dataset_factory):
@@ -32,7 +32,6 @@ from lerobot.configs.video import (
 )
 from lerobot.datasets.depth_utils import dequantize_depth, quantize_depth
 from lerobot.datasets.image_writer import image_array_to_pil_image, write_image
 from lerobot.utils.constants import DEFAULT_FEATURES
 from tests.fixtures.constants import (
    DEFAULT_FPS,
    DUMMY_CAMERA_FEATURES,
@@ -248,89 +247,42 @@ class TestFeatureFileRouting:
        dataset.finalize()
-class TestDepthUnitMetadata:
+# ── 5. Depth stats unit canonicalization (millimetres) ────────────────
-    """The depth unit is inferred once from dtype, stored in ``info``, and drives stats + reads."""
+
 class TestDepthStatsUnit:
    """Depth stats are always stored in millimetres, regardless of raw frame dtype."""
    NUM_FRAMES = 4
    def _record(self, root, features_factory, depth_dtype, value, use_videos):
        from lerobot.datasets.lerobot_dataset import LeRobotDataset
        features = features_factory(camera_features=DUMMY_CAMERA_FEATURES_WITH_DEPTH, use_videos=use_videos)
        dataset = LeRobotDataset.create(
            repo_id=DUMMY_REPO_ID,
            fps=DEFAULT_FPS,
            features=features,
            root=root,
            use_videos=use_videos,
            streaming_encoding=use_videos,
        )
        for _ in range(self.NUM_FRAMES):
            frame: dict = {"task": "test"}
            for key, ft in dataset.meta.features.items():
                if key in DEFAULT_FEATURES:
                    continue
                if key in dataset.meta.depth_keys:
                    frame[key] = np.full(ft["shape"], value, dtype=depth_dtype)
                elif key in dataset.meta.camera_keys:
                    frame[key] = np.random.randint(0, 256, ft["shape"], dtype=np.uint8)
                else:
                    frame[key] = np.zeros(ft["shape"], dtype=np.float32)
            dataset.add_frame(frame)
        return dataset
    @pytest.mark.parametrize("use_videos", [False, True])
-    @pytest.mark.parametrize(
+    def test_stats_canonicalized_to_mm(self, tmp_path, features_factory, use_videos):
-        ("depth_dtype", "value", "expected_unit"),
+        """Float (metre) and integer (millimetre) depth over the same physical range
-        [(np.float32, 2.0, DEPTH_METER_UNIT), (np.uint16, 2000, DEPTH_MILLIMETER_UNIT)],
+        yield identical millimetre-scale stats."""
    )
    def test_recorded_unit_inferred_persisted_and_kept_in_stats(
        self, tmp_path, features_factory, use_videos, depth_dtype, value, expected_unit
    ):
        """Unit is inferred from the first frame's dtype, drives stats (raw, never canonicalized), and survives a reload."""
        from lerobot.datasets.lerobot_dataset import LeRobotDataset
-        dataset = self._record(tmp_path / "ds", features_factory, depth_dtype, value, use_videos)
+        def _record(depth_dtype, root):
-        assert dataset.meta.features[DEPTH_KEY]["info"]["depth_unit"] == expected_unit
+            features = features_factory(
-        dataset.save_episode()
+                camera_features=DUMMY_CAMERA_FEATURES_WITH_DEPTH, use_videos=use_videos
        mean = float(np.asarray(dataset.meta.stats[DEPTH_KEY]["mean"]).reshape(-1)[0])
        np.testing.assert_allclose(mean, value, rtol=0.05)
        dataset.finalize()
        reloaded = LeRobotDataset(repo_id=DUMMY_REPO_ID, root=tmp_path / "ds")
        assert reloaded.meta.features[DEPTH_KEY]["info"]["depth_unit"] == expected_unit
    @pytest.mark.parametrize("use_videos", [False, True])
    @pytest.mark.parametrize(
        ("output_unit", "expected"),
        [(DEPTH_MILLIMETER_UNIT, 2000.0), (DEPTH_METER_UNIT, 2.0)],
    )
    def test_read_honors_output_unit_for_frames_and_stats(
        self, tmp_path, features_factory, use_videos, output_unit, expected
    ):
        """Reloading with a ``depth_output_unit`` converts metre frames (image mode) and rescales stats while preserving count."""
        from lerobot.datasets.lerobot_dataset import LeRobotDataset
        dataset = self._record(tmp_path / "ds", features_factory, np.float32, 2.0, use_videos=use_videos)
        dataset.save_episode()
        count = float(np.asarray(dataset.meta.stats[DEPTH_KEY]["count"]).reshape(-1)[0])
        dataset.finalize()
        read_dataset = LeRobotDataset(
            repo_id=DUMMY_REPO_ID, root=tmp_path / "ds", depth_output_unit=output_unit
        )
        stats = read_dataset.meta.stats[DEPTH_KEY]
        np.testing.assert_allclose(float(np.asarray(stats["mean"]).reshape(-1)[0]), expected, rtol=0.05)
        np.testing.assert_allclose(float(np.asarray(stats["count"]).reshape(-1)[0]), count)
        if not use_videos:
            depth = read_dataset[0][DEPTH_KEY]
            assert torch.allclose(depth, torch.full_like(depth, expected))
            from lerobot.datasets.streaming_dataset import StreamingLeRobotDataset
            stream_dataset = StreamingLeRobotDataset(
                repo_id=DUMMY_REPO_ID, root=tmp_path / "ds", depth_output_unit=output_unit
            )
-            stream_depth = next(iter(stream_dataset))[DEPTH_KEY]
+            dataset = LeRobotDataset.create(
-            assert torch.allclose(stream_depth, torch.full_like(stream_depth, expected))
+                repo_id=DUMMY_REPO_ID,
                fps=DEFAULT_FPS,
                features=features,
                root=root,
                use_videos=use_videos,
                streaming_encoding=use_videos,
            )
            add_frames(dataset, num_frames=self.NUM_FRAMES, depth_dtype=depth_dtype)
            dataset.save_episode()
            dataset.finalize()
            return np.asarray(dataset.meta.stats[DEPTH_KEY]["mean"]).reshape(-1)
        # add_frames ramps float depth over 0.1–10 m and integer depth over 100–10000 mm
        # (the same physical range), so canonicalized stats must match.
        mean_m = _record(np.float32, tmp_path / "ds_m")
        mean_mm = _record(np.uint16, tmp_path / "ds_mm")
        # Float (metre) input is scaled to millimetres, not left in the single-digit metre range.
        assert mean_m.item() > 50.0
        np.testing.assert_allclose(mean_m, mean_mm, rtol=0.05)
@@ -345,9 +345,7 @@ class TestExtraOptions:
        opts = cfg.get_codec_options()
        assert opts["qp"] == 20
        assert isinstance(opts["qp"], int)
-        str_opts = cfg.get_codec_options(as_strings=True)
+        assert cfg.get_codec_options(as_strings=True)["qp"] == "20"
        assert str_opts["qp"] == "20"
        assert all(isinstance(v, str) for v in str_opts.values())
    @require_libsvtav1
    def test_structured_fields_win_on_collision(self):
@@ -7,14 +7,11 @@ from dataclasses import dataclass, field
 import gymnasium as gym
 import pytest
 import torch
 from gymnasium.envs.registration import register, registry as gym_registry
 from lerobot.configs.types import PolicyFeature
-from lerobot.envs.configs import EnvConfig, LiberoEnv
+from lerobot.envs.configs import EnvConfig
 from lerobot.envs.factory import make_env, make_env_config, make_env_pre_post_processors
 from lerobot.processor import LiberoProcessorStep
 from lerobot.utils.constants import OBS_PREFIX, OBS_STATE
 logger = logging.getLogger(__name__)
@@ -64,31 +61,6 @@ def test_processors_delegation():
    assert len(pre.steps) == 0
 def test_libero_processors_are_policy_agnostic():
    cfg = LiberoEnv()
    pre, post = make_env_pre_post_processors(cfg, policy_cfg=object())
    assert isinstance(pre.steps[0], LiberoProcessorStep)
    assert len(post.steps) == 0
 def test_libero_processor_flattens_state_to_raw_8_dim():
    step = LiberoProcessorStep()
    observation = {
        OBS_PREFIX + "robot_state": {
            "eef": {
                "pos": torch.tensor([[1.0, 2.0, 3.0]]),
                "quat": torch.tensor([[0.0, 0.0, 0.0, 1.0]]),
            },
            "gripper": {"qpos": torch.tensor([[4.0, 5.0]])},
        }
    }
    state = step.observation(observation)[OBS_STATE]
    assert state.shape == (1, 8)
    assert torch.allclose(state, torch.tensor([[1.0, 2.0, 3.0, 0.0, 0.0, 0.0, 4.0, 5.0]]))
 def test_base_create_envs():
    """Base class create_envs() should build a single-task VectorEnv via gym.make()."""
    gym_id = "_dispatch_test/CartPole-v99"
@@ -26,7 +26,6 @@ import pytest
 import torch
 from datasets import Dataset
 from lerobot.configs.video import infer_depth_unit
 from lerobot.datasets.dataset_metadata import CODEBASE_VERSION, LeRobotDatasetMetadata
 from lerobot.datasets.feature_utils import get_hf_features_from_features
 from lerobot.datasets.io_utils import flatten_dict, hf_transform_to_torch
@@ -50,16 +49,18 @@ from tests.fixtures.constants import (
 )
-def add_frames(dataset: LeRobotDataset, num_frames: int) -> None:
+def add_frames(dataset: LeRobotDataset, num_frames: int, depth_dtype: np.dtype = np.uint16) -> None:
    """Append ``num_frames`` synthetic frames to ``dataset``.
-    Generates per-feature payloads from ``dataset.meta``: uint16 depth ramps for
+    Generates per-feature payloads from ``dataset.meta``: depth ramps (``depth_dtype``,
-    keys in ``dataset.meta.depth_keys``, uint8 random noise for video/image keys,
+    default ``uint16`` millimetres; pass ``np.float32`` for metres) for keys in
-    and float32 zeros for everything else. ``DEFAULT_FEATURES`` (timestamp,
+    ``dataset.meta.depth_keys``, uint8 random noise for video/image keys, and float32
-    frame_index, ...) are auto-populated by ``add_frame`` and skipped here.
+    zeros for everything else. ``DEFAULT_FEATURES`` (timestamp, frame_index, ...) are
    auto-populated by ``add_frame`` and skipped here.
    """
    video_keys = dataset.meta.video_keys
    depth_keys = dataset.meta.depth_keys
    depth_is_float = np.issubdtype(depth_dtype, np.floating)
    # Smooth gradient base reused per (H, W) to keep depth frames cheap to
    # encode (HEVC Main 12 hates white noise).
    _depth_base_cache: dict[tuple[int, int], np.ndarray] = {}
@@ -71,11 +72,14 @@ def add_frames(dataset: LeRobotDataset, num_frames: int) -> None:
            shape = ft["shape"]
            if key in depth_keys:
                h, w, _ = shape
                # Float depth is expressed in metres, integer depth in millimetres.
                lo, hi = (0.1, 10.0) if depth_is_float else (100.0, 10_000.0)
                base = _depth_base_cache.setdefault(
                    (h, w),
-                    np.linspace(100.0, 10_000.0, h * w, dtype=np.float32).reshape(h, w, 1),
+                    np.linspace(lo, hi, h * w, dtype=np.float32).reshape(h, w, 1),
                )
-                frame[key] = (base + 50.0 * i).clip(0, 65535).astype(np.uint16)
+                step = (0.05 if depth_is_float else 50.0) * i
                frame[key] = (base + step).clip(0, 65535).astype(depth_dtype)
            elif key in video_keys:
                frame[key] = np.random.randint(0, 256, shape, dtype=np.uint8)
            else:
@@ -536,13 +540,6 @@ def lerobot_dataset_factory(
                chunks_size=chunks_size,
                **info_kwargs,
            )
            # This synthetic path skips add_frame, so record the depth unit the writer would
            # have stored (dummy depth is uint16) to keep ``depth_unit`` present in info.json.
            # Reassign a fresh info dict to avoid mutating the shared feature constants.
            for ft in info.features.values():
                ft_info = ft.get("info")
                if ft_info is not None and ft_info.get("is_depth_map") and "depth_unit" not in ft_info:
                    ft["info"] = {**ft_info, "depth_unit": infer_depth_unit(np.uint16)}
        if stats is None:
            stats = stats_factory(features=info.features)
        if tasks is None:
@@ -1,840 +0,0 @@
 #!/usr/bin/env python
 # Copyright 2026 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
 #     http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from __future__ import annotations
 import pytest
 import torch
 from torch import nn
 import lerobot.policies.evo1.evo1_model as evo1_model
 import lerobot.policies.evo1.modeling_evo1 as modeling_evo1
 from lerobot.configs.types import FeatureType, PolicyFeature
 from lerobot.policies.evo1.configuration_evo1 import Evo1Config
 from lerobot.policies.evo1.flow_matching import FlowmatchingActionHead
 from lerobot.policies.evo1.internvl3_embedder import (
    IMAGENET_MEAN,
    IMAGENET_STD,
    _batched_pixel_values,
 )
 from lerobot.policies.evo1.processor_evo1 import (
    Evo1ActionProcessorStep,
    Evo1PadActionProcessorStep,
    Evo1PadStateProcessorStep,
    evo1_batch_to_transition,
    make_evo1_pre_post_processors,
    reconcile_evo1_processors,
 )
 from lerobot.policies.factory import get_policy_class, make_policy_config
 from lerobot.policies.rtc.configuration_rtc import RTCConfig
 from lerobot.policies.rtc.modeling_rtc import RTCProcessor
 from lerobot.processor import (
    DeviceProcessorStep,
    NormalizerProcessorStep,
    PolicyProcessorPipeline,
    UnnormalizerProcessorStep,
 )
 from lerobot.processor.converters import (
    batch_to_transition,
    policy_action_to_transition,
    transition_to_batch,
    transition_to_policy_action,
 )
 from lerobot.utils.constants import (
    ACTION,
    OBS_IMAGES,
    OBS_STATE,
    POLICY_POSTPROCESSOR_DEFAULT_NAME,
    POLICY_PREPROCESSOR_DEFAULT_NAME,
 )
 STATE_DIM = 4
 ACTION_DIM = 3
 MAX_STATE_DIM = 6
 MAX_ACTION_DIM = 5
 CHUNK_SIZE = 2
 EMBED_DIM = 8
 class DummyEvo1Model(nn.Module):
    def __init__(self, config, vlm_hub_kwargs=None):
        super().__init__()
        self.config = config
        self.embedder = nn.Dropout(p=0.0)
        self.action_head = nn.Linear(1, 1)
        self.get_vl_embeddings_calls = 0
        self.grad_enabled_calls = []
        self.embedder_training_calls = []
    def set_finetune_flags(self):
        return None
    def get_vl_embeddings(self, images, image_mask, prompt=None, return_cls_only=False):
        self.get_vl_embeddings_calls += 1
        self.grad_enabled_calls.append(torch.is_grad_enabled())
        self.embedder_training_calls.append(self.embedder.training)
        # images is a list of per-camera (B, C, H, W) tensors, so the batch dim is images[0].shape[0].
        batch_size = images[0].shape[0]
        tokens = torch.ones(batch_size, 4, EMBED_DIM, requires_grad=torch.is_grad_enabled())
        valid_mask = torch.ones(batch_size, 4, dtype=torch.bool)
        return tokens, valid_mask
    def forward(
        self,
        fused_tokens,
        state=None,
        actions_gt=None,
        action_mask=None,
        embodiment_ids=None,
        context_mask=None,
        **kwargs,
    ):
        batch_size = fused_tokens.shape[0]
        if actions_gt is None:
            return torch.ones(batch_size, CHUNK_SIZE * MAX_ACTION_DIM)
        pred_velocity = torch.zeros(batch_size, CHUNK_SIZE * MAX_ACTION_DIM)
        noise = torch.zeros_like(actions_gt)
        return pred_velocity, noise
 class ChunkCountingDummyModel(DummyEvo1Model):
    """Emits per-step distinguishable actions so queue ordering and re-prediction are observable."""
    def __init__(self, config, vlm_hub_kwargs=None):
        super().__init__(config, vlm_hub_kwargs)
        self.chunks_predicted = 0
    def forward(
        self,
        fused_tokens,
        state=None,
        actions_gt=None,
        action_mask=None,
        embodiment_ids=None,
        context_mask=None,
        **kwargs,
    ):
        if actions_gt is not None:
            return super().forward(fused_tokens, state, actions_gt, action_mask, embodiment_ids, context_mask)
        self.chunks_predicted += 1
        batch_size = fused_tokens.shape[0]
        step_values = torch.arange(CHUNK_SIZE, dtype=torch.float32) + 10.0 * self.chunks_predicted
        chunk = step_values.repeat_interleave(MAX_ACTION_DIM).unsqueeze(0).repeat(batch_size, 1)
        return chunk
 def make_config(training_stage="stage1", **kwargs):
    config_kwargs = {
        "device": "cpu",
        "vlm_model_name": "dummy-internvl3",
        "training_stage": training_stage,
        "chunk_size": CHUNK_SIZE,
        "n_action_steps": 1,
        "max_state_dim": MAX_STATE_DIM,
        "max_action_dim": MAX_ACTION_DIM,
        "max_views": 2,
        "embed_dim": EMBED_DIM,
        "hidden_dim": 16,
        "state_hidden_dim": 16,
        "num_heads": 2,
        "num_layers": 1,
        "num_inference_timesteps": 2,
        "input_features": {
            OBS_STATE: PolicyFeature(type=FeatureType.STATE, shape=(STATE_DIM,)),
            f"{OBS_IMAGES}.front": PolicyFeature(type=FeatureType.VISUAL, shape=(3, 16, 16)),
        },
        "output_features": {
            ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(ACTION_DIM,)),
        },
    }
    config_kwargs.update(kwargs)
    return Evo1Config(**config_kwargs)
 def make_batch(include_action=True):
    batch = {
        "task": ["pick the block", "place the block"],
        OBS_STATE: torch.randn(2, STATE_DIM),
        f"{OBS_IMAGES}.front": torch.rand(2, 3, 16, 16),
    }
    if include_action:
        batch[ACTION] = torch.randn(2, CHUNK_SIZE, ACTION_DIM)
    return batch
 def make_stats(state_dim=STATE_DIM, action_dim=ACTION_DIM):
    return {
        OBS_STATE: {
            "min": torch.full((state_dim,), -2.0),
            "max": torch.full((state_dim,), 2.0),
        },
        ACTION: {
            "min": torch.full((action_dim,), -1.0),
            "max": torch.full((action_dim,), 1.0),
        },
    }
 def make_flowmatching_head(**overrides):
    kwargs = {
        "embed_dim": EMBED_DIM,
        "hidden_dim": 16,
        "action_dim": CHUNK_SIZE * ACTION_DIM,
        "horizon": CHUNK_SIZE,
        "per_action_dim": ACTION_DIM,
        "num_heads": 2,
        "num_layers": 1,
        "num_inference_timesteps": 2,
        "state_dim": STATE_DIM,
        "state_hidden_dim": 16,
        "num_categories": 1,
    }
    kwargs.update(overrides)
    return FlowmatchingActionHead(**kwargs)
 def test_evo1_factory_registration():
    cfg = make_policy_config(
        "evo1",
        device="cpu",
        vlm_model_name="dummy-internvl3",
        input_features={
            OBS_STATE: PolicyFeature(type=FeatureType.STATE, shape=(STATE_DIM,)),
            f"{OBS_IMAGES}.front": PolicyFeature(type=FeatureType.VISUAL, shape=(3, 16, 16)),
        },
        output_features={ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(ACTION_DIM,))},
    )
    assert isinstance(cfg, Evo1Config)
    assert get_policy_class("evo1") is modeling_evo1.Evo1Policy
 def test_evo1_stage_defaults_and_consistency():
    stage1 = make_config(training_stage="stage1")
    assert (stage1.finetune_vlm, stage1.finetune_language_model, stage1.finetune_vision_model) == (
        False,
        False,
        False,
    )
    assert stage1.finetune_action_head is True
    stage2 = make_config(training_stage="stage2")
    assert (stage2.finetune_vlm, stage2.finetune_language_model, stage2.finetune_vision_model) == (
        True,
        True,
        True,
    )
    assert stage2.finetune_action_head is True
    stage2_from_stage1_checkpoint_flags = make_config(
        training_stage="stage2",
        finetune_vlm=False,
        finetune_language_model=False,
        finetune_vision_model=False,
        finetune_action_head=False,
    )
    assert (
        stage2_from_stage1_checkpoint_flags.finetune_vlm,
        stage2_from_stage1_checkpoint_flags.finetune_language_model,
        stage2_from_stage1_checkpoint_flags.finetune_vision_model,
    ) == (
        True,
        True,
        True,
    )
    assert stage2_from_stage1_checkpoint_flags.finetune_action_head is True
    explicit_off = make_config(
        training_stage="stage2",
        apply_training_stage_defaults=False,
        finetune_vlm=False,
        finetune_language_model=False,
        finetune_vision_model=False,
        finetune_action_head=False,
    )
    assert (
        explicit_off.finetune_vlm,
        explicit_off.finetune_language_model,
        explicit_off.finetune_vision_model,
    ) == (
        False,
        False,
        False,
    )
    assert explicit_off.finetune_action_head is False
    # An explicit finetune_vlm=False without branch-level flags freezes both branches instead of
    # raising an inconsistency error.
    frozen_vlm = make_config(
        training_stage="stage2",
        apply_training_stage_defaults=False,
        finetune_vlm=False,
    )
    assert (
        frozen_vlm.finetune_vlm,
        frozen_vlm.finetune_language_model,
        frozen_vlm.finetune_vision_model,
    ) == (False, False, False)
    try:
        make_config(
            training_stage="stage2",
            apply_training_stage_defaults=False,
            finetune_vlm=True,
            finetune_language_model=False,
        )
    except ValueError as exc:
        assert "Inconsistent EVO1 finetune config" in str(exc)
    else:
        raise AssertionError("Expected inconsistent finetune config to raise ValueError")
 def test_evo1_rejects_non_square_image_resolution():
    with pytest.raises(ValueError, match="square image_resolution"):
        make_config(image_resolution=(448, 320))
 def test_evo1_rejects_out_of_range_default_embodiment_id():
    with pytest.raises(ValueError, match="default_embodiment_id"):
        make_config(default_embodiment_id=3, num_categories=2)
 def test_evo1_model_uses_image_resolution_and_trainable_checkpointing(monkeypatch):
    captured: dict = {}
    class SpyEmbedder(nn.Module):
        def __init__(self, **kwargs):
            super().__init__()
            captured.clear()
            captured.update(kwargs)
    monkeypatch.setattr(evo1_model, "InternVL3Embedder", SpyEmbedder)
    stage1 = make_config(training_stage="stage1", image_resolution=(224, 224))
    evo1_model.Evo1Model(stage1)
    assert captured["image_size"] == 224
    # VLM is frozen in stage1, so gradient checkpointing is gated off.
    assert captured["enable_gradient_checkpointing"] is False
    stage2 = make_config(training_stage="stage2", image_resolution=(224, 224))
    evo1_model.Evo1Model(stage2)
    assert captured["enable_gradient_checkpointing"] is True
 class FakeInternVLModel(nn.Module):
    """Minimal stand-in with the native HF InternVL submodule layout."""
    def __init__(self):
        super().__init__()
        self.language_model = nn.Linear(2, 2)
        self.vision_tower = nn.Linear(2, 2)
        self.multi_modal_projector = nn.Linear(2, 2)
 class FakeEmbedder(nn.Module):
    def __init__(self, **kwargs):
        super().__init__()
        self.model = FakeInternVLModel()
 def test_set_finetune_flags_targets_native_hf_internvl_submodules(monkeypatch):
    monkeypatch.setattr(evo1_model, "InternVL3Embedder", FakeEmbedder)
    stage2_model = evo1_model.Evo1Model(make_config(training_stage="stage2"))
    stage2_model.set_finetune_flags()
    vlm = stage2_model.embedder.model
    assert all(p.requires_grad for p in vlm.language_model.parameters())
    assert all(p.requires_grad for p in vlm.vision_tower.parameters())
    assert all(p.requires_grad for p in vlm.multi_modal_projector.parameters())
    assert all(p.requires_grad for p in stage2_model.action_head.parameters())
    stage1_model = evo1_model.Evo1Model(make_config(training_stage="stage1"))
    stage1_model.set_finetune_flags()
    vlm = stage1_model.embedder.model
    assert not any(p.requires_grad for p in vlm.parameters())
    assert all(p.requires_grad for p in stage1_model.action_head.parameters())
 def test_set_finetune_flags_fails_loudly_on_unknown_vlm_layout(monkeypatch):
    class LegacyLayoutModel(nn.Module):
        def __init__(self):
            super().__init__()
            self.language_model = nn.Linear(2, 2)
            self.vision_model = nn.Linear(2, 2)  # trust_remote_code-era attribute name
            self.mlp1 = nn.Linear(2, 2)
    class FakeEmbedder(nn.Module):
        def __init__(self, **kwargs):
            super().__init__()
            self.model = LegacyLayoutModel()
    monkeypatch.setattr(evo1_model, "InternVL3Embedder", FakeEmbedder)
    model = evo1_model.Evo1Model(make_config(training_stage="stage2"))
    with pytest.raises(AttributeError, match="vision_tower"):
        model.set_finetune_flags()
 def test_evo1_policy_processors_pad_state_crop_action_and_binarize_gripper():
    libero_action_dim = 7
    config = make_config(
        max_state_dim=MAX_STATE_DIM,
        max_action_dim=8,
        postprocess_action_dim=libero_action_dim,
        binarize_gripper=True,
        output_features={ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(libero_action_dim,))},
    )
    stats = make_stats(action_dim=libero_action_dim)
    preprocessor, postprocessor = make_evo1_pre_post_processors(config, dataset_stats=stats)
    assert isinstance(preprocessor.steps[2], Evo1PadStateProcessorStep)
    assert isinstance(preprocessor.steps[3], Evo1PadActionProcessorStep)
    assert isinstance(preprocessor.steps[4], NormalizerProcessorStep)
    assert isinstance(postprocessor.steps[0], UnnormalizerProcessorStep)
    assert isinstance(postprocessor.steps[1], Evo1ActionProcessorStep)
    normalizer = preprocessor.steps[4]
    assert normalizer.features[OBS_STATE].shape == (MAX_STATE_DIM,)
    assert normalizer.features[ACTION].shape == (8,)
    assert normalizer._tensor_stats[OBS_STATE]["min"].shape == (MAX_STATE_DIM,)
    assert normalizer._tensor_stats[ACTION]["min"].shape == (8,)
    processed_batch = preprocessor(
        {
            "task": "pick the block",
            OBS_STATE: torch.zeros(STATE_DIM),
            ACTION: torch.zeros(libero_action_dim),
            f"{OBS_IMAGES}.front": torch.rand(3, 16, 16),
        }
    )
    processed_state = processed_batch[OBS_STATE]
    assert processed_state.shape == (1, MAX_STATE_DIM)
    assert torch.allclose(processed_state, torch.zeros_like(processed_state))
    assert processed_batch[ACTION].shape == (1, 8)
    assert torch.allclose(processed_batch[ACTION], torch.zeros_like(processed_batch[ACTION]))
    assert processed_batch["action_mask"].shape == (1, 8)
    assert processed_batch["action_mask"][:, :libero_action_dim].all()
    assert not processed_batch["action_mask"][:, libero_action_dim:].any()
    action = torch.tensor(
        [
            [0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.5, 0.7],
            [0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7],
        ],
        dtype=torch.float32,
    )
    processed = postprocessor(action)
    assert processed.shape == (2, 7)
    assert processed.dtype == torch.float32
    assert torch.allclose(processed[:, :6], action[:, :6])
    assert torch.equal(processed[:, 6], torch.tensor([1.0, -1.0]))
 def test_evo1_postprocessor_returns_float32_for_bf16_actions():
    config = make_config()
    _preprocessor, postprocessor = make_evo1_pre_post_processors(config, dataset_stats=make_stats())
    processed = postprocessor(torch.zeros(2, MAX_ACTION_DIM, dtype=torch.bfloat16))
    assert processed.dtype == torch.float32
 def test_evo1_processor_save_load_round_trip_applies_config_overrides(tmp_path):
    train_config = make_config()
    preprocessor, postprocessor = make_evo1_pre_post_processors(train_config, dataset_stats=make_stats())
    preprocessor.save_pretrained(tmp_path)
    postprocessor.save_pretrained(tmp_path)
    loaded_pre = PolicyProcessorPipeline.from_pretrained(
        tmp_path,
        config_filename=f"{POLICY_PREPROCESSOR_DEFAULT_NAME}.json",
        to_transition=batch_to_transition,
        to_output=transition_to_batch,
    )
    loaded_post = PolicyProcessorPipeline.from_pretrained(
        tmp_path,
        config_filename=f"{POLICY_POSTPROCESSOR_DEFAULT_NAME}.json",
        to_transition=policy_action_to_transition,
        to_output=transition_to_policy_action,
    )
    # Simulate eval-time CLI overrides applied on top of the loaded pipelines.
    eval_config = make_config(binarize_gripper=True, postprocess_action_dim=ACTION_DIM)
    loaded_pre, loaded_post = reconcile_evo1_processors(eval_config, loaded_pre, loaded_post)
    assert loaded_pre.to_transition is evo1_batch_to_transition
    assert sum(isinstance(step, Evo1ActionProcessorStep) for step in loaded_post.steps) == 1
    action_step = next(step for step in loaded_post.steps if isinstance(step, Evo1ActionProcessorStep))
    assert action_step.binarize_gripper is True
    assert action_step.action_dim == ACTION_DIM
    # The float32 output dtype is part of the serialized pipeline itself.
    device_step = next(step for step in loaded_post.steps if isinstance(step, DeviceProcessorStep))
    assert device_step.float_dtype == "float32"
    # Non-observation extras (embodiment_id, ...) must survive the reloaded preprocessor.
    processed = loaded_pre(
        {
            "task": "pick the block",
            OBS_STATE: torch.zeros(STATE_DIM),
            f"{OBS_IMAGES}.front": torch.rand(3, 16, 16),
            "embodiment_id": torch.tensor([0]),
        }
    )
    assert "embodiment_id" in processed
 def test_evo1_policy_forward_and_inference_use_batched_embedding(monkeypatch):
    monkeypatch.setattr(modeling_evo1, "Evo1Model", DummyEvo1Model)
    policy = modeling_evo1.Evo1Policy(make_config())
    preprocessor, _postprocessor = make_evo1_pre_post_processors(policy.config, dataset_stats=make_stats())
    training_batch = preprocessor(make_batch(include_action=True))
    assert training_batch[ACTION].shape == (2, CHUNK_SIZE, MAX_ACTION_DIM)
    assert training_batch["action_mask"].shape == (2, CHUNK_SIZE, MAX_ACTION_DIM)
    assert training_batch["action_mask"][:, :, :ACTION_DIM].all()
    assert not training_batch["action_mask"][:, :, ACTION_DIM:].any()
    loss, metrics = policy.forward(training_batch)
    assert loss.ndim == 0
    assert torch.isfinite(loss)
    assert metrics["active_action_dims"] == ACTION_DIM * CHUNK_SIZE
    assert policy.model.get_vl_embeddings_calls == 1
    action_chunk = policy.predict_action_chunk(make_batch(include_action=False))
    assert action_chunk.shape == (2, CHUNK_SIZE, MAX_ACTION_DIM)
    assert action_chunk.dtype == torch.float32
    policy.reset()
    selected = policy.select_action(make_batch(include_action=False))
    assert selected.shape == (2, MAX_ACTION_DIM)
 def test_evo1_forward_masks_padded_action_timesteps(monkeypatch):
    monkeypatch.setattr(modeling_evo1, "Evo1Model", DummyEvo1Model)
    policy = modeling_evo1.Evo1Policy(make_config())
    batch = make_batch(include_action=True)
    batch[ACTION] = torch.ones(2, CHUNK_SIZE, ACTION_DIM)
    # Give the padded (past-episode-end) timestep a huge value: if it leaked into the loss, the
    # loss would blow up far beyond 1.0.
    batch[ACTION][:, -1, :] = 100.0
    batch["action_is_pad"] = torch.zeros(2, CHUNK_SIZE, dtype=torch.bool)
    batch["action_is_pad"][:, -1] = True
    loss, metrics = policy.forward(batch)
    # DummyEvo1Model predicts zero velocity and zero noise, so each active element contributes
    # (0 - action)^2 = 1.0 for the in-episode ones-valued actions.
    assert metrics["active_action_dims"] == ACTION_DIM * (CHUNK_SIZE - 1)
    assert torch.isclose(loss, torch.tensor(1.0))
 def test_evo1_select_action_queue_orders_steps_and_repredicts(monkeypatch):
    monkeypatch.setattr(modeling_evo1, "Evo1Model", ChunkCountingDummyModel)
    policy = modeling_evo1.Evo1Policy(make_config(n_action_steps=CHUNK_SIZE))
    batch = make_batch(include_action=False)
    first = policy.select_action(batch)
    second = policy.select_action(batch)
    third = policy.select_action(batch)
    # First chunk provides steps 10, 11 in order; the third call triggers a fresh prediction (20).
    assert torch.all(first == 10.0)
    assert torch.all(second == 11.0)
    assert torch.all(third == 20.0)
    assert policy.model.chunks_predicted == 2
 def test_evo1_predict_action_chunk_rejects_rtc_kwargs_without_rtc_config(monkeypatch):
    monkeypatch.setattr(modeling_evo1, "Evo1Model", DummyEvo1Model)
    policy = modeling_evo1.Evo1Policy(make_config())
    with pytest.raises(RuntimeError, match="RTC"):
        policy.predict_action_chunk(make_batch(include_action=False), inference_delay=2)
 def test_evo1_rtc_processor_wiring(monkeypatch):
    monkeypatch.setattr(evo1_model, "InternVL3Embedder", FakeEmbedder)
    policy = modeling_evo1.Evo1Policy(make_config())
    assert policy.rtc_processor is None
    assert policy.model.rtc_processor is None
    # The RTC rollout backend assigns rtc_config after loading and re-inits the processor.
    policy.config.rtc_config = RTCConfig(execution_horizon=CHUNK_SIZE)
    policy.init_rtc_processor()
    assert isinstance(policy.rtc_processor, RTCProcessor)
    assert policy.model.rtc_processor is policy.rtc_processor
    # RTC drives predict_action_chunk directly; the select_action queue path is unsupported.
    with pytest.raises(AssertionError, match="select_action"):
        policy.select_action(make_batch(include_action=False))
 def test_flowmatching_rtc_guidance_pulls_prefix_toward_previous_chunk():
    head = make_flowmatching_head(num_inference_timesteps=16)
    processor = RTCProcessor(RTCConfig(execution_horizon=CHUNK_SIZE))
    fused = torch.randn(2, 4, EMBED_DIM)
    state = torch.randn(2, STATE_DIM)
    action_mask = torch.ones(2, ACTION_DIM, dtype=torch.bool)
    prev_chunk = torch.tensor([0.7, -0.4, 0.2]).expand(2, CHUNK_SIZE, ACTION_DIM).contiguous()
    torch.manual_seed(0)
    unguided = head.get_action(fused, state=state, action_mask=action_mask)
    unguided = unguided.view(2, CHUNK_SIZE, ACTION_DIM)
    torch.manual_seed(0)
    guided = head.get_action(
        fused,
        state=state,
        action_mask=action_mask,
        inference_delay=1,
        prev_chunk_left_over=prev_chunk,
        rtc_processor=processor,
    )
    guided = guided.view(2, CHUNK_SIZE, ACTION_DIM)
    # The frozen prefix (first inference_delay steps) must land far closer to the previous chunk
    # than the unguided sample from the same noise does.
    guided_dist = (guided[:, 0] - prev_chunk[:, 0]).abs().mean()
    unguided_dist = (unguided[:, 0] - prev_chunk[:, 0]).abs().mean()
    assert guided_dist < 0.5 * unguided_dist
    assert torch.isfinite(guided).all()
 def test_flowmatching_rtc_first_chunk_without_leftover_matches_unguided():
    head = make_flowmatching_head(num_inference_timesteps=4)
    processor = RTCProcessor(RTCConfig(execution_horizon=CHUNK_SIZE))
    fused = torch.randn(2, 4, EMBED_DIM)
    state = torch.randn(2, STATE_DIM)
    action_mask = torch.ones(2, ACTION_DIM, dtype=torch.bool)
    torch.manual_seed(0)
    unguided = head.get_action(fused, state=state, action_mask=action_mask)
    torch.manual_seed(0)
    first_chunk = head.get_action(
        fused,
        state=state,
        action_mask=action_mask,
        inference_delay=2,
        prev_chunk_left_over=None,
        rtc_processor=processor,
    )
    assert torch.allclose(unguided, first_chunk)
 def test_evo1_missing_configured_camera_needs_empty_cameras_budget(monkeypatch):
    monkeypatch.setattr(modeling_evo1, "Evo1Model", DummyEvo1Model)
    batch = make_batch(include_action=False)  # only provides the front camera
    two_camera_features = {
        OBS_STATE: PolicyFeature(type=FeatureType.STATE, shape=(STATE_DIM,)),
        f"{OBS_IMAGES}.front": PolicyFeature(type=FeatureType.VISUAL, shape=(3, 16, 16)),
        f"{OBS_IMAGES}.wrist": PolicyFeature(type=FeatureType.VISUAL, shape=(3, 16, 16)),
    }
    strict_policy = modeling_evo1.Evo1Policy(make_config(input_features=dict(two_camera_features)))
    with pytest.raises(ValueError, match="empty_cameras"):
        strict_policy._collect_image_batches(batch)
    # empty_cameras adds placeholder camera features that are never present in the batch; they
    # become masked-out views instead of crashing with a KeyError.
    padded_policy = modeling_evo1.Evo1Policy(make_config(empty_cameras=1))
    assert len(padded_policy.config.image_features) == 2
    camera_images, image_masks = padded_policy._collect_image_batches(batch)
    assert len(camera_images) == 1
    assert image_masks.tolist() == [[True, False], [True, False]]
 def test_stage1_frozen_vlm_embeddings_do_not_track_gradients(monkeypatch):
    monkeypatch.setattr(modeling_evo1, "Evo1Model", DummyEvo1Model)
    policy = modeling_evo1.Evo1Policy(make_config(training_stage="stage1"))
    policy.train()
    image_batches, image_masks = policy._collect_image_batches(make_batch(include_action=False))
    fused_tokens, context_mask = policy._compute_fused_tokens(["pick", "place"], image_batches, image_masks)
    assert policy.model.grad_enabled_calls == [False]
    assert policy.model.embedder_training_calls == [False]
    assert not fused_tokens.requires_grad
    assert context_mask is not None
    assert policy.model.embedder.training is False
 def test_stage2_vlm_embeddings_track_gradients(monkeypatch):
    monkeypatch.setattr(modeling_evo1, "Evo1Model", DummyEvo1Model)
    policy = modeling_evo1.Evo1Policy(make_config(training_stage="stage2"))
    policy.train()
    image_batches, image_masks = policy._collect_image_batches(make_batch(include_action=False))
    fused_tokens, _context_mask = policy._compute_fused_tokens(["pick", "place"], image_batches, image_masks)
    assert policy.model.grad_enabled_calls == [True]
    assert policy.model.embedder_training_calls == [True]
    assert fused_tokens.requires_grad
 def test_collect_image_batches_handles_unbatched_chw(monkeypatch):
    # Regression for an issue where batch_size was read from shape[0] before normalizing
    # per-camera tensor dims, so an unbatched (C, H, W) input was treated as batch_size=C.
    monkeypatch.setattr(modeling_evo1, "Evo1Model", DummyEvo1Model)
    policy = modeling_evo1.Evo1Policy(make_config())
    batch = {
        OBS_STATE: torch.randn(1, STATE_DIM),
        f"{OBS_IMAGES}.front": torch.rand(3, 16, 16),
    }
    camera_images, image_masks = policy._collect_image_batches(batch)
    # One present camera, returned as a batched (B, C, H, W) tensor with the unbatched CHW frame
    # promoted to batch_size=1 (not read as batch_size=C).
    assert len(camera_images) == 1
    assert camera_images[0].shape == (1, 3, 16, 16)
    assert image_masks.tolist() == [[True, False]]
 def test_evo1_state_mask_zeroes_masked_dims(monkeypatch):
    monkeypatch.setattr(modeling_evo1, "Evo1Model", DummyEvo1Model)
    policy = modeling_evo1.Evo1Policy(make_config())
    batch = {
        OBS_STATE: torch.ones(2, STATE_DIM),
        "state_mask": torch.tensor([[True, True, False, False]] * 2),
    }
    states, mask = policy._prepare_state(batch)
    assert torch.all(states[:, :2] == 1.0)
    assert torch.all(states[:, 2:] == 0.0)
    assert mask[:, :2].all()
    assert not mask[:, 2:].any()
 def test_evo1_action_mask_accepts_chunk_size_one(monkeypatch):
    monkeypatch.setattr(modeling_evo1, "Evo1Model", DummyEvo1Model)
    config = make_config(chunk_size=1, n_action_steps=1)
    policy = modeling_evo1.Evo1Policy(config)
    batch = make_batch(include_action=True)
    batch[ACTION] = torch.randn(2, ACTION_DIM)
    batch["action_mask"] = torch.ones(2, ACTION_DIM, dtype=torch.bool)
    actions, action_mask = policy._prepare_actions(batch)
    assert actions.shape == (2, 1, MAX_ACTION_DIM)
    assert action_mask.shape == (2, 1, MAX_ACTION_DIM)
    assert action_mask[:, :, :ACTION_DIM].all()
    assert not action_mask[:, :, ACTION_DIM:].any()
 def test_flowmatching_state_encoder_for_horizon_one():
    head = make_flowmatching_head(action_dim=ACTION_DIM, horizon=1)
    assert head.state_encoder is not None
    pred_velocity, noise = head(
        torch.randn(2, 4, EMBED_DIM),
        state=torch.randn(2, STATE_DIM),
        actions_gt=torch.randn(2, 1, ACTION_DIM),
        action_mask=torch.ones(2, 1, ACTION_DIM, dtype=torch.bool),
    )
    assert pred_velocity.shape == (2, ACTION_DIM)
    assert noise.shape == (2, 1, ACTION_DIM)
 def test_flowmatching_get_action_real_path_respects_action_mask():
    torch.manual_seed(0)
    head = make_flowmatching_head()
    action_mask = torch.zeros(2, ACTION_DIM, dtype=torch.bool)
    action_mask[:, :2] = True
    actions = head.get_action(
        torch.randn(2, 4, EMBED_DIM),
        state=torch.randn(2, STATE_DIM),
        action_mask=action_mask,
    )
    assert actions.shape == (2, CHUNK_SIZE * ACTION_DIM)
    assert torch.isfinite(actions).all()
    action_seq = actions.view(2, CHUNK_SIZE, ACTION_DIM)
    assert torch.all(action_seq[..., 2] == 0.0)
 def test_flowmatching_context_mask_blocks_masked_context_tokens():
    head = make_flowmatching_head()
    state = torch.randn(2, STATE_DIM)
    action_mask = torch.ones(2, ACTION_DIM, dtype=torch.bool)
    fused = torch.randn(2, 4, EMBED_DIM)
    context_mask = torch.ones(2, 4, dtype=torch.bool)
    context_mask[:, -1] = False
    corrupted = fused.clone()
    corrupted[:, -1] = 1e4
    torch.manual_seed(0)
    reference = head.get_action(fused, state=state, action_mask=action_mask, context_mask=context_mask)
    torch.manual_seed(0)
    with_garbage = head.get_action(corrupted, state=state, action_mask=action_mask, context_mask=context_mask)
    assert torch.allclose(reference, with_garbage)
 def test_flowmatching_head_accepts_pooled_2d_context():
    head = make_flowmatching_head()
    pred_velocity, noise = head(
        torch.randn(2, EMBED_DIM),  # pooled (B, E) context from return_cls_only
        state=torch.randn(2, STATE_DIM),
        actions_gt=torch.randn(2, CHUNK_SIZE, ACTION_DIM),
        action_mask=torch.ones(2, CHUNK_SIZE, ACTION_DIM, dtype=torch.bool),
    )
    assert pred_velocity.shape == (2, CHUNK_SIZE * ACTION_DIM)
    actions = head.get_action(
        torch.randn(2, EMBED_DIM),
        state=torch.randn(2, STATE_DIM),
        action_mask=torch.ones(2, ACTION_DIM, dtype=torch.bool),
    )
    assert actions.shape == (2, CHUNK_SIZE * ACTION_DIM)
 def test_flowmatching_rejects_out_of_range_embodiment_ids():
    head = make_flowmatching_head(num_categories=2)
    with pytest.raises(ValueError, match="num_categories"):
        head.get_action(
            torch.randn(2, 4, EMBED_DIM),
            state=torch.randn(2, STATE_DIM),
            action_mask=torch.ones(2, ACTION_DIM, dtype=torch.bool),
            embodiment_id=torch.tensor([0, 5]),
        )
 def test_evo1_batched_pixel_values_shape_and_zero_padding():
    torch.manual_seed(0)
    batch_size, image_size, max_views = 2, 448, 3
    camera_images = [torch.rand(batch_size, 3, 40, 50)]  # a single present camera
    mean = torch.tensor(IMAGENET_MEAN)
    std = torch.tensor(IMAGENET_STD)
    pixel_values = _batched_pixel_values(
        camera_images, max_views, image_size, mean, std, torch.float32, torch.device("cpu")
    )
    assert pixel_values.shape == (batch_size * max_views, 3, image_size, image_size)
    grouped = pixel_values.reshape(batch_size, max_views, 3, image_size, image_size)
    # Absent views (indices 1, 2) are zero images, normalized to the constant -mean/std.
    expected_pad = (-mean / std).view(1, 3, 1, 1)
    for view in (1, 2):
        assert torch.allclose(
            grouped[:, view], expected_pad.expand(batch_size, 3, image_size, image_size), atol=1e-5
        )
    # The present view is genuinely different from the constant pad value.
    assert not torch.allclose(
        grouped[:, 0], expected_pad.expand(batch_size, 3, image_size, image_size), atol=1e-3
    )
@@ -14,7 +14,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
-"""Test script for LeRobot's GR00T N1.7 policy forward and inference passes."""
+"""Test script for LeRobot's Groot policy forward and inference passes."""
 import gc
 import os
@@ -25,8 +25,6 @@ import numpy as np
 import pytest
 import torch
 pytest.importorskip("transformers", reason="groot requires the `groot` extra (transformers)")
 from lerobot.policies.groot.configuration_groot import GrootConfig
 from lerobot.policies.groot.modeling_groot import GrootPolicy
 from lerobot.policies.groot.processor_groot import make_groot_pre_post_processors
@@ -35,26 +33,21 @@ from lerobot.types import PolicyAction
 from lerobot.utils.device_utils import auto_select_torch_device
 from tests.utils import require_cuda
 pytest.importorskip("transformers")
 pytestmark = pytest.mark.skipif(
    os.environ.get("CI") == "true" or os.environ.get("GITHUB_ACTIONS") == "true",
    reason="This test requires local Groot installation and is not meant for CI",
 )
-# Define constants for dummy data (GR00T N1.7 native conventions).
+# Define constants for dummy data
 # N1.7 internally uses a 40-step action chunk, 132-dim state/action, and 256px images
 # (see GrootConfig.__post_init__). Use a chunk-sized action horizon so the dummy batch
 # matches the model's native action space.
 DUMMY_STATE_DIM = 44
 DUMMY_ACTION_DIM = 44
-DUMMY_ACTION_HORIZON = 40
+DUMMY_ACTION_HORIZON = 16
 IMAGE_SIZE = 256
 DEVICE = auto_select_torch_device()
-# GR00T N1.7 checkpoint (N1.5 is no longer supported). The N1.7-3B base model loads
+MODEL_PATH = "aractingi/bimanual-handover-groot-10k"
 # via GrootPolicy.from_pretrained with root-level sharded safetensors.
 MODEL_PATH = "nvidia/GR00T-N1.7-3B"
 # Valid N1.7 embodiment tag carried by the checkpoint metadata.
 EMBODIMENT_TAG = "gr1_unified"
 def cleanup_memory():
@@ -95,13 +88,13 @@ def instantiate_lerobot_groot(
    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
    PolicyProcessorPipeline[PolicyAction, PolicyAction],
 ]:
-    """Instantiate LeRobot GR00T N1.7 policy with preprocessor and postprocessor."""
+    """Instantiate LeRobot Groot policy with preprocessor and postprocessor."""
    if from_pretrained:
        policy = GrootPolicy.from_pretrained(
            pretrained_name_or_path=model_path,
            strict=False,
        )
-        policy.config.embodiment_tag = EMBODIMENT_TAG
+        policy.config.embodiment_tag = "gr1"
    else:
        config = GrootConfig(
            base_model_path=model_path,
@@ -109,7 +102,7 @@ def instantiate_lerobot_groot(
            chunk_size=DUMMY_ACTION_HORIZON,
            image_size=[IMAGE_SIZE, IMAGE_SIZE],
            device=DEVICE,
-            embodiment_tag=EMBODIMENT_TAG,
+            embodiment_tag="gr1",
        )
        policy = GrootPolicy(config)
@@ -155,8 +148,8 @@ def create_dummy_data(device=DEVICE):
@require_cuda
 def test_lerobot_groot_inference():
-    """Test the inference pass (select_action) of LeRobot's GR00T N1.7 policy."""
+    """Test the inference pass (select_action) of LeRobot's Groot policy."""
-    print("Test: LeRobot GR00T N1.7 Inference Pass")
+    print("Test: LeRobot Groot Inference Pass")
    set_seed_all(42)
@@ -188,9 +181,9 @@ def test_lerobot_groot_inference():
@require_cuda
 def test_lerobot_groot_forward_pass():
-    """Test the forward pass of LeRobot's GR00T N1.7 policy."""
+    """Test the forward pass of LeRobot's Groot policy."""
    print("\n" + "=" * 50)
-    print("Test: LeRobot GR00T N1.7 Forward Pass (Training Mode)")
+    print("Test: LeRobot Groot Forward Pass (Training Mode)")
    set_seed_all(42)
@@ -1,259 +0,0 @@
 #!/usr/bin/env python
 # Copyright 2026 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
 #     http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import hashlib
 import os
 from pathlib import Path
 import numpy as np
 import pytest
 import torch
 from lerobot.policies.groot.action_head.cross_attention_dit import AlternateVLDiT
 from lerobot.policies.groot.groot_n1_7 import GR00TN17
 from lerobot.policies.groot.processor_groot import (
    GrootN17ActionDecodeStep,
    GrootN17PackInputsStep,
    GrootN17VLMEncodeStep,
    _transform_n1_7_image_for_vlm_albumentations,
 )
 from lerobot.types import TransitionKey
 from lerobot.utils.constants import OBS_STATE
 OSS_REFERENCE_COMMIT = "ab88b50c718f6528e1df9dcbaf75865d1b604760"
 def _fixture_path(filename: str) -> Path:
    fixture_dir = os.environ.get("GROOT_N17_OSS_PARITY_FIXTURE_DIR")
    if fixture_dir is None:
        pytest.skip("Set GROOT_N17_OSS_PARITY_FIXTURE_DIR to run external OSS parity fixtures.")
    path = Path(fixture_dir) / filename
    if not path.is_file():
        pytest.skip(f"External OSS parity fixture not found: {path}")
    return path
 def test_groot_n1_7_eval_image_transform_matches_oss_reference():
    """Match the native N1.7 eval transform for a non-square SO-101 frame."""
    y, x = np.indices((480, 640), dtype=np.uint16)
    image = np.stack(
        ((x + 3 * y) % 256, (2 * x + y) % 256, (x + 5 * y) % 256),
        axis=-1,
    ).astype(np.uint8)
    actual = _transform_n1_7_image_for_vlm_albumentations(
        image,
        image_crop_size=[230, 230],
        image_target_size=[256, 256],
        shortest_image_edge=256,
        crop_fraction=0.95,
    )
    assert actual.shape == (256, 340, 3)
    assert hashlib.sha256(actual.tobytes()).hexdigest() == (
        "c17e47af68a812aa79db3bb7b64b549ddf10148ac1b204a9686095018561ae9e"
    )
 def test_groot_n1_7_vlm_chat_content_order_matches_oss_reference():
    """Native OSS places all image items before the language item."""
    class RecordingProcessor:
        def __init__(self):
            self.content_types = None
        def apply_chat_template(self, conversation, tokenize, add_generation_prompt):
            assert tokenize is False
            assert add_generation_prompt is False
            self.content_types = [item["type"] for item in conversation[0]["content"]]
            return "rendered"
        def __call__(self, **kwargs):
            return {}
    processor = RecordingProcessor()
    step = GrootN17VLMEncodeStep(
        image_crop_size=[230, 230],
        image_target_size=[256, 256],
        shortest_image_edge=256,
        crop_fraction=0.95,
        use_albumentations=True,
        device="cpu",
    )
    step._proc = processor
    transition = {
        TransitionKey.OBSERVATION: {
            "video": np.zeros((1, 1, 2, 480, 640, 3), dtype=np.uint8),
        },
        TransitionKey.COMPLEMENTARY_DATA: {"language": ["pick up the vial"]},
    }
    step(transition)
    assert processor.content_types == ["image", "image", "text"]
 def test_groot_n1_7_alternate_vl_dit_matches_oss_reference():
    """Run the LeRobot DiT with native OSS weights and identical inputs."""
    pytest.importorskip("diffusers")
    fixture = torch.load(_fixture_path("alternate_vl_dit_small.pt"), map_location="cpu", weights_only=True)
    model = AlternateVLDiT(
        output_dim=8,
        num_attention_heads=2,
        attention_head_dim=4,
        num_layers=4,
        dropout=0.0,
        final_dropout=False,
        max_num_positional_embeddings=16,
        compute_dtype=torch.float32,
        interleave_self_attention=True,
        cross_attention_dim=6,
    ).eval()
    model.load_state_dict(fixture["state_dict"], strict=True)
    actual = model(
        hidden_states=fixture["hidden_states"],
        encoder_hidden_states=fixture["encoder_hidden_states"],
        timestep=fixture["timestep"],
        image_mask=fixture["image_mask"],
        backbone_attention_mask=fixture["backbone_attention_mask"],
    )
    torch.testing.assert_close(actual, fixture["output"], atol=1e-6, rtol=1e-6)
 def _state_decode_reference():
    fixture = np.load(_fixture_path("state_and_action_decode.npz"))
    raw_stats = {
        "state": {
            "single_arm": {"q01": fixture["state_single_arm_q01"], "q99": fixture["state_single_arm_q99"]},
            "gripper": {"q01": fixture["state_gripper_q01"], "q99": fixture["state_gripper_q99"]},
        },
        "action": {
            "single_arm": {"q01": fixture["action_single_arm_q01"], "q99": fixture["action_single_arm_q99"]},
            "gripper": {"q01": fixture["action_gripper_q01"], "q99": fixture["action_gripper_q99"]},
        },
        "relative_action": {
            "single_arm": {
                "min": fixture["relative_single_arm_min"],
                "max": fixture["relative_single_arm_max"],
            },
        },
    }
    for modality_stats in raw_stats.values():
        for entry in modality_stats.values():
            for key, value in entry.items():
                if isinstance(value, np.ndarray):
                    entry[key] = value.tolist()
    modality_config = {
        "state": {"modality_keys": ["single_arm", "gripper"]},
        "action": {
            "delta_indices": list(range(16)),
            "modality_keys": ["single_arm", "gripper"],
            "action_configs": [
                {"rep": "RELATIVE", "type": "NON_EEF", "format": "DEFAULT", "state_key": None},
                {"rep": "ABSOLUTE", "type": "NON_EEF", "format": "DEFAULT", "state_key": None},
            ],
        },
    }
    state_min = np.concatenate((fixture["state_single_arm_q01"], fixture["state_gripper_q01"]))
    state_max = np.concatenate((fixture["state_single_arm_q99"], fixture["state_gripper_q99"]))
    pack_step = GrootN17PackInputsStep(
        normalize_min_max=True,
        stats={OBS_STATE: {"min": state_min, "max": state_max}},
        raw_stats=raw_stats,
        modality_config=modality_config,
        use_percentiles=True,
    )
    raw_state = np.concatenate((fixture["state_single_arm"], fixture["state_gripper"]), axis=-1)
    transition = {
        TransitionKey.OBSERVATION: {OBS_STATE: torch.from_numpy(raw_state)},
        TransitionKey.COMPLEMENTARY_DATA: {},
    }
    packed = pack_step(transition)
    return fixture, raw_stats, modality_config, pack_step, packed
 def test_groot_n1_7_state_normalization_matches_oss_checkpoint_reference():
    fixture, _raw_stats, _modality_config, _pack_step, packed = _state_decode_reference()
    expected = np.concatenate(
        (fixture["normalized_state_single_arm"], fixture["normalized_state_gripper"]), axis=-1
    )
    actual = packed[TransitionKey.OBSERVATION]["state"][:, 0, :6]
    torch.testing.assert_close(actual, torch.from_numpy(expected), atol=1e-6, rtol=1e-6)
 def test_groot_n1_7_relative_action_decode_matches_oss_checkpoint_reference():
    fixture, raw_stats, modality_config, pack_step, _packed = _state_decode_reference()
    decode_step = GrootN17ActionDecodeStep(
        env_action_dim=6,
        raw_stats=raw_stats,
        modality_config=modality_config,
        use_percentiles=True,
        use_relative_action=True,
        pack_step=pack_step,
    )
    decoded = decode_step({TransitionKey.ACTION: torch.from_numpy(fixture["normalized_action"])})[
        TransitionKey.ACTION
    ]
    expected = np.concatenate((fixture["decoded_single_arm"], fixture["decoded_gripper"]), axis=-1).astype(
        np.float32
    )
    torch.testing.assert_close(decoded, torch.from_numpy(expected), atol=1e-5, rtol=1e-5)
 def test_groot_n1_7_qwen_backbone_matches_oss_checkpoint_reference():
    """Compare the actual 3B checkpoint backbone when explicitly enabled."""
    checkpoint = os.environ.get("GROOT_N17_PARITY_CHECKPOINT")
    if checkpoint is None:
        pytest.skip("Set GROOT_N17_PARITY_CHECKPOINT to run the 3B OSS Qwen parity test.")
    if not torch.cuda.is_available():
        pytest.skip("The 3B OSS Qwen parity test requires CUDA.")
    pytest.importorskip("transformers")
    from transformers.feature_extraction_utils import BatchFeature
    fixture = torch.load(_fixture_path("qwen_backbone_so101.pt"), map_location="cpu", weights_only=True)
    model = GR00TN17.from_pretrained(checkpoint).to(device="cuda", dtype=torch.bfloat16).eval()
    backbone_input = BatchFeature(
        data={
            key.removeprefix("input."): value.to("cuda")
            for key, value in fixture.items()
            if key.startswith("input.")
        }
    )
    with torch.inference_mode():
        actual = model.backbone(backbone_input)
    feature_error = (
        actual.backbone_features.cpu().float() - fixture["output.backbone_features"].float()
    ).abs()
    # Native OSS and LeRobot use different Torch/Transformers/Flash-Attention releases.
    # Require the measured BF16 accumulation envelope while rejecting structural drift.
    assert feature_error.mean().item() <= 0.04
    assert feature_error.max().item() <= 2.0
    torch.testing.assert_close(
        actual.backbone_attention_mask.cpu(), fixture["output.backbone_attention_mask"]
    )
    torch.testing.assert_close(actual.image_mask.cpu(), fixture["output.image_mask"])
@@ -1,100 +0,0 @@
 #!/usr/bin/env python
 # Copyright 2026 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
 #     http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """Isaac-GR00T N1.7 raw-state dropout training contract.
 Isaac-GR00T zeroes the entire proprioceptive state of a sample with probability
 ``state_dropout_prob`` (configured in the checkpoint's processor sidecar) during
 training only. Baseline LeRobot kept the processor deterministic, so this
 regularization never activated. These tests pin the train/eval split.
 """
 import torch
 from lerobot.policies.groot.processor_groot import GrootN17PackInputsStep
 from lerobot.types import TransitionKey
 from lerobot.utils.constants import OBS_STATE
 def _make_transition():
    return {
        TransitionKey.OBSERVATION: {OBS_STATE: torch.tensor([[1.0, 2.0], [3.0, 4.0]])},
        TransitionKey.COMPLEMENTARY_DATA: {"task": ["Move", "Move"]},
    }
 def test_groot_n1_7_training_applies_raw_state_dropout_before_encoder():
    step = GrootN17PackInputsStep(
        max_state_dim=4,
        max_action_dim=4,
        normalize_min_max=False,
        training=True,
        state_dropout_prob=1.0,
    )
    output = step(_make_transition())
    expected = torch.zeros(2, 1, 4)
    torch.testing.assert_close(output[TransitionKey.OBSERVATION]["state"], expected)
 def test_groot_n1_7_training_state_dropout_is_disabled_under_no_grad():
    step = GrootN17PackInputsStep(
        max_state_dim=4,
        max_action_dim=4,
        normalize_min_max=False,
        training=True,
        state_dropout_prob=1.0,
    )
    with torch.no_grad():
        output = step(_make_transition())
    expected = torch.tensor([[[1.0, 2.0, 0.0, 0.0]], [[3.0, 4.0, 0.0, 0.0]]])
    torch.testing.assert_close(output[TransitionKey.OBSERVATION]["state"], expected)
 def test_groot_n1_7_eval_mode_state_dropout_is_inactive():
    step = GrootN17PackInputsStep(
        max_state_dim=4,
        max_action_dim=4,
        normalize_min_max=False,
        training=False,
        state_dropout_prob=1.0,
    )
    output = step(_make_transition())
    expected = torch.tensor([[[1.0, 2.0, 0.0, 0.0]], [[3.0, 4.0, 0.0, 0.0]]])
    torch.testing.assert_close(output[TransitionKey.OBSERVATION]["state"], expected)
 def test_groot_n1_7_pack_step_serializes_dropout_prob_but_not_training_mode():
    step = GrootN17PackInputsStep(
        max_state_dim=4,
        max_action_dim=4,
        normalize_min_max=False,
        training=True,
        state_dropout_prob=0.2,
    )
    serialized = step.get_config()
    restored = GrootN17PackInputsStep(**serialized)
    assert "training" not in serialized
    assert serialized["state_dropout_prob"] == 0.2
    assert restored.training is False
    assert restored.state_dropout_prob == 0.2
@@ -1,169 +0,0 @@
 #!/usr/bin/env python
 # Copyright 2026 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
 #     http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """Isaac-GR00T N1.7 train-time random crop contract (crop geometry only).
 Isaac-GR00T crops a random ``crop_fraction`` window during training and the
 deterministic center window at eval, replaying the sampled window across all
 camera views of a sample (gr00t/data/transform/video.py, n1.5-release onward:
 "If mode is 'train', return a random crop transform. If mode is 'eval', return
 a center crop transform."). This mirrors LeRobot's own Diffusion/VQBeT
 ``crop_is_random`` pattern. Color jitter is intentionally out of scope here.
 """
 import random
 import numpy as np
 import torch
 from lerobot.policies.groot.processor_groot import (
    GrootN17VLMEncodeStep,
    _transform_n1_7_image_for_vlm_albumentations,
 )
 def _structured_image(h=480, w=640):
    yy, xx = np.mgrid[0:h, 0:w]
    return np.stack([(xx * 255 / w), (yy * 255 / h), ((xx + yy) * 255 / (h + w))], axis=-1).astype(np.uint8)
 def test_crop_position_none_is_bitexact_center_crop():
    """crop_position=None must remain byte-identical to the pre-change eval path."""
    img = _structured_image()
    ref = _transform_n1_7_image_for_vlm_albumentations(
        img,
        image_crop_size=None,
        image_target_size=[256, 256],
        shortest_image_edge=256,
        crop_fraction=0.95,
    )
    out = _transform_n1_7_image_for_vlm_albumentations(
        img,
        image_crop_size=None,
        image_target_size=[256, 256],
        shortest_image_edge=256,
        crop_fraction=0.95,
        crop_position=None,
    )
    np.testing.assert_array_equal(ref, out)
 def test_crop_position_center_matches_center_crop():
    img = _structured_image()
    center = _transform_n1_7_image_for_vlm_albumentations(
        img,
        image_crop_size=None,
        image_target_size=[256, 256],
        shortest_image_edge=256,
        crop_fraction=0.95,
        crop_position=None,
    )
    explicit = _transform_n1_7_image_for_vlm_albumentations(
        img,
        image_crop_size=None,
        image_target_size=[256, 256],
        shortest_image_edge=256,
        crop_fraction=0.95,
        crop_position=(0.5, 0.5),
    )
    # int-floor center vs rounded positional center may differ by <=1 px of grid
    assert center.shape == explicit.shape
    diff = np.abs(center.astype(np.int16) - explicit.astype(np.int16))
    assert diff.mean() < 3.0
 def test_crop_position_corners_differ_from_center():
    img = _structured_image()
    def crop_at(position):
        return _transform_n1_7_image_for_vlm_albumentations(
            img,
            image_crop_size=None,
            image_target_size=[256, 256],
            shortest_image_edge=256,
            crop_fraction=0.95,
            crop_position=position,
        )
    center = crop_at(None)
    tl = crop_at((0.0, 0.0))
    br = crop_at((1.0, 1.0))
    assert not np.array_equal(center, tl)
    assert not np.array_equal(tl, br)
 def _video(img, views=2):
    return np.stack([img] * views, axis=0).reshape(1, 1, views, *img.shape)
 def _step(training):
    return GrootN17VLMEncodeStep(
        image_target_size=[256, 256],
        shortest_image_edge=256,
        crop_fraction=0.95,
        use_albumentations=True,
        training=training,
    )
 def test_training_crop_replays_one_window_across_views():
    video = _video(_structured_image())
    frames = _step(training=True)._build_sample_images(video, batch_size=1, target_device=None)[0]
    np.testing.assert_array_equal(np.asarray(frames[0]), np.asarray(frames[1]))
 def test_training_crop_differs_from_eval_center_crop():
    video = _video(_structured_image())
    random.seed(3)  # a draw that is not the exact center
    train_frame = np.asarray(
        _step(training=True)._build_sample_images(video, batch_size=1, target_device=None)[0][0]
    )
    eval_frame = np.asarray(
        _step(training=False)._build_sample_images(video, batch_size=1, target_device=None)[0][0]
    )
    assert not np.array_equal(train_frame, eval_frame)
 def test_training_crop_is_disabled_under_no_grad():
    video = _video(_structured_image())
    with torch.no_grad():
        no_grad_frame = np.asarray(
            _step(training=True)._build_sample_images(video, batch_size=1, target_device=None)[0][0]
        )
    eval_frame = np.asarray(
        _step(training=False)._build_sample_images(video, batch_size=1, target_device=None)[0][0]
    )
    np.testing.assert_array_equal(no_grad_frame, eval_frame)
 def test_training_mode_is_not_serialized():
    step = _step(training=True)
    serialized = step.get_config()
    assert "training" not in serialized
    restored = GrootN17VLMEncodeStep(**serialized)
    assert restored.training is False
 def test_training_crop_respects_global_seed():
    video = _video(_structured_image())
    def draw():
        random.seed(11)
        return np.asarray(
            _step(training=True)._build_sample_images(video, batch_size=1, target_device=None)[0][0]
        )
    np.testing.assert_array_equal(draw(), draw())
@@ -1,125 +0,0 @@
 #!/usr/bin/env python
 # Copyright 2026 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
 #     http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """Isaac-GR00T N1.7 optimizer/scheduler/precision training contract.
 Pins the LeRobot GR00T fine-tuning recipe to the native Isaac-GR00T contract:
 AdamW(lr=1e-4, betas=(0.9, 0.999), eps=1e-8, weight_decay=1e-5, grad clip 1.0),
 HF cosine schedule with ~5% warmup over the actual update count, FP32 master
 parameters under BF16 autocast, transformers-style weight-decay grouping, the
 frozen LM-head weight tie, and episode-tail exclusion for incomplete chunks.
 """
 import pytest
 import torch
 from lerobot.optim.schedulers import DiffuserSchedulerConfig
 from lerobot.policies.groot.configuration_groot import GrootConfig
 from lerobot.policies.groot.groot_n1_7 import _tie_unused_qwen_lm_head
 from lerobot.policies.groot.modeling_groot import GrootPolicy
 def test_groot_n1_7_optimizer_matches_isaac_training_contract():
    optimizer = GrootConfig().get_optimizer_preset()
    assert optimizer.lr == pytest.approx(1e-4)
    assert optimizer.betas == pytest.approx((0.9, 0.999))
    assert optimizer.eps == pytest.approx(1e-8)
    assert optimizer.weight_decay == pytest.approx(1e-5)
    assert optimizer.grad_clip_norm == pytest.approx(1.0)
 def test_groot_n1_7_sampler_excludes_incomplete_action_tails():
    config = GrootConfig(chunk_size=16, n_action_steps=16)
    assert len(config.action_delta_indices) == 16
    assert config.drop_n_last_frames == 15
 def test_groot_n1_7_scheduler_matches_isaac_hf_cosine_contract():
    pytest.importorskip("diffusers", reason="the scheduler preset requires the `groot` extra (diffusers)")
    config = GrootConfig(max_steps=20_000)
    scheduler_config = config.get_scheduler_preset()
    assert isinstance(scheduler_config, DiffuserSchedulerConfig)
    assert scheduler_config.name == "cosine"
    assert scheduler_config.num_warmup_steps == 1_000
    parameter = torch.nn.Parameter(torch.ones(()))
    optimizer = torch.optim.AdamW([parameter], lr=config.optimizer_lr)
    scheduler = scheduler_config.build(optimizer, num_training_steps=20_000)
    lr_factor = scheduler.lr_lambdas[0]
    assert lr_factor(0) == pytest.approx(0.0)
    assert lr_factor(1_000) == pytest.approx(1.0)
    assert lr_factor(10_500) == pytest.approx(0.5)
    assert lr_factor(20_000) == pytest.approx(0.0, abs=1e-12)
 def test_groot_n1_7_scheduler_rounds_fractional_warmup_up_like_transformers():
    scheduler_config = GrootConfig(max_steps=777).get_scheduler_preset()
    assert scheduler_config.num_warmup_steps == 39
 def test_groot_n1_7_model_parameters_use_fp32_checkpoint_and_optimizer_precision():
    module = torch.nn.Module()
    module.trainable = torch.nn.Parameter(torch.ones(3, dtype=torch.bfloat16))
    module.frozen = torch.nn.Parameter(torch.ones(3, dtype=torch.bfloat16), requires_grad=False)
    GrootPolicy._cast_model_parameters_to_fp32(module)
    assert module.trainable.dtype == torch.float32
    assert module.frozen.dtype == torch.float32
 def test_groot_n1_7_ties_unused_qwen_lm_head_to_frozen_input_embeddings():
    class DummyQwen(torch.nn.Module):
        def __init__(self):
            super().__init__()
            self.embed_tokens = torch.nn.Embedding(7, 3)
            self.lm_head = torch.nn.Linear(3, 7, bias=False)
        def get_input_embeddings(self):
            return self.embed_tokens
    model = DummyQwen()
    _tie_unused_qwen_lm_head(model)
    assert model.lm_head.weight is model.embed_tokens.weight
    assert len(list(model.parameters())) == 1
 def test_groot_n1_7_optimizer_groups_match_transformers_weight_decay_rules():
    pytest.importorskip(
        "transformers", reason="weight-decay grouping requires the `groot` extra (transformers)"
    )
    module = torch.nn.Module()
    module.linear = torch.nn.Linear(3, 2)
    module.norm = torch.nn.LayerNorm(2)
    module.frozen = torch.nn.Parameter(torch.ones(1), requires_grad=False)
    groups = GrootPolicy._build_weight_decay_parameter_groups(module)
    assert len(groups) == 2
    assert "weight_decay" not in groups[0]
    assert groups[1]["weight_decay"] == 0.0
    assert groups[0]["params"] == [module.linear.weight]
    assert {id(parameter) for parameter in groups[1]["params"]} == {
        id(module.linear.bias),
        id(module.norm.weight),
        id(module.norm.bias),
    }
@@ -1,6 +1,6 @@
 #!/usr/bin/env python
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -14,194 +14,431 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
-"""Parity test: original NVIDIA GR00T N1.7 vs the GR00T N1.7 integration in LeRobot.
+"""Test script to verify Groot policy integration with LeRobot vs the original implementation, only meant to be run locally!"""
 Verifies that the self-contained LeRobot reimplementation of the GR00T N1.7 action
 head + Qwen3-VL backbone produces the SAME raw model output (``action_pred``, the
 normalized flow-matching prediction before any action decoding) as NVIDIA's original
 ``gr00t`` package, given byte-identical pre-processed inputs and the same
 flow-matching seed. The comparison is parametrized over every embodiment tag present
 in the checkpoint.
 To keep the comparison fair, the original outputs + the exact collated inputs are
 produced once per embodiment in the original ``gr00t`` env via the companion script
 ``utils/dump_original_n1_7.py`` (in the ``utils`` package next to this file) and saved
 to per-tag ``.npz`` files.
 This test discovers those artifacts, replays the identical inputs through the LeRobot
 model, and compares.
 This test is LOCAL-only and skips on CI, when ``gr00t``-side prerequisites are not
 present, or when no artifact has been generated. By default it looks for artifacts in
 ``<this dir>/artifacts/``; override with ``GROOT_N1_7_PARITY_DIR``. See the
 "Original-vs-LeRobot parity test" section of ``src/lerobot/policies/groot/README.md``
 for the full run procedure.
 """
 import gc
 import os
-from pathlib import Path
+from copy import deepcopy
 from typing import Any
 import numpy as np
 import pytest
 import torch
 from lerobot.policies.groot.configuration_groot import GrootConfig
 from lerobot.policies.groot.modeling_groot import GrootPolicy
 from lerobot.policies.groot.processor_groot import make_groot_pre_post_processors
 from lerobot.processor import PolicyProcessorPipeline
 from lerobot.types import PolicyAction
 pytest.importorskip("gr00t")
 pytest.importorskip("transformers")
 pytestmark = pytest.mark.skipif(
    os.environ.get("CI") == "true" or os.environ.get("GITHUB_ACTIONS") == "true",
-    reason="Requires a local GR00T N1.7 checkpoint + pre-generated artifacts; not for CI.",
+    reason="This test requires local Groot installation and is not meant for CI",
 )
 from lerobot.policies.groot.configuration_groot import GROOT_N1_7  # noqa: E402,F401
-SEED = 42
+from gr00t.data.dataset import ModalityConfig  # noqa: E402
-DEVICE = os.environ.get("GROOT_PARITY_DEVICE", "cuda" if torch.cuda.is_available() else "cpu")
+from gr00t.data.embodiment_tags import EmbodimentTag  # noqa: E402
-ATOL = float(os.environ.get("GROOT_PARITY_ATOL", "1e-3"))
+from gr00t.data.transform.base import ComposedModalityTransform  # noqa: E402
-RTOL = float(os.environ.get("GROOT_PARITY_RTOL", "1e-3"))
+from gr00t.model.policy import Gr00tPolicy  # noqa: E402
-# Artifact filenames are original_n1_7_<embodiment_tag>.npz
+# GR1 humanoid dimensions (from pretrained model metadata)
-_ARTIFACT_PREFIX = "original_n1_7_"
+# The actual GR1 robot has 44 dimensions for both state and action
-_ARTIFACT_SUFFIX = ".npz"
+# GR00TTransform will pad state to 64 and truncate action to 32
 DUMMY_STATE_DIM = 44
 DUMMY_ACTION_DIM = 44
 DUMMY_ACTION_HORIZON = 16
 IMAGE_SIZE = 256
 DEVICE = "cpu"
 MODEL_PATH = "nvidia/GR00T-N1.5-3B"
 GR1_BODY_PARTS = {
    "left_arm": 7,
    "left_hand": 6,
    "left_leg": 6,
    "neck": 3,
    "right_arm": 7,
    "right_hand": 6,
    "right_leg": 6,
    "waist": 3,
 }
-def _artifact_dir() -> Path:
+def cleanup_memory():
-    """Directory holding the per-embodiment .npz artifacts.
+    """Clean up GPU/MPS memory to prevent OOM errors between tests."""
-
+    print("\nCleaning up memory...")
-    Self-contained by default: a sibling ``artifacts/`` directory next to this test.
+    gc.collect()
    Override with ``GROOT_N1_7_PARITY_DIR`` (e.g. to point at a scratch location).
    The directory is read-only here -- it is populated by ``utils/dump_original_n1_7.py``
    run in the original gr00t environment; the test never creates it.
    """
    env = os.environ.get("GROOT_N1_7_PARITY_DIR")
    if env:
        return Path(env)
    return Path(__file__).resolve().parent / "artifacts"
 def _discover_artifacts() -> list[tuple[str, Path]]:
    """Return [(embodiment_tag, npz_path), ...] for every dumped artifact."""
    d = _artifact_dir()
    if not d.is_dir():
        return []
    out = []
    for p in sorted(d.glob(f"{_ARTIFACT_PREFIX}*{_ARTIFACT_SUFFIX}")):
        tag = p.name[len(_ARTIFACT_PREFIX) : -len(_ARTIFACT_SUFFIX)]
        out.append((tag, p))
    return out
 def _resolve_checkpoint() -> str:
    env = os.environ.get("GROOT_N1_7_LIBERO_CKPT")
    if env:
        if not Path(env).exists():
            pytest.skip(f"GROOT_N1_7_LIBERO_CKPT={env} does not exist")
        return env
    try:
        from huggingface_hub import snapshot_download
        root = snapshot_download(
            "nvidia/GR00T-N1.7-LIBERO",
            local_files_only=True,
            allow_patterns=["libero_10/*"],
        )
    except Exception as exc:  # noqa: BLE001
        pytest.skip(f"GR00T N1.7 LIBERO checkpoint not available locally: {exc}")
    ckpt = Path(root) / "libero_10"
    if not (ckpt / "config.json").exists():
        pytest.skip(f"GR00T N1.7 LIBERO checkpoint incomplete at {ckpt}")
    return str(ckpt)
 def _load_artifact(path: Path):
    data = np.load(path, allow_pickle=True)
    original_action = torch.from_numpy(data["action_pred"]).float()
    dtypes = dict(zip(data["meta_keys"].tolist(), data["meta_dtypes"].tolist(), strict=False))
    inputs = {}
    for key in data.files:
        if not key.startswith("in::"):
            continue
        name = key[4:]
        arr = data[key]
        t = torch.from_numpy(np.asarray(arr))
        declared = dtypes.get(key, "")
        if "int" in declared or "long" in declared:
            t = t.long()
        inputs[name] = t
    return original_action, inputs
 def _unflatten(inputs: dict[str, torch.Tensor]) -> dict:
    """Rebuild the nested model-input dict from dot-prefixed flat keys."""
    nested: dict = {}
    for dotted, value in inputs.items():
        parts = dotted.split(".")
        cur = nested
        for p in parts[:-1]:
            cur = cur.setdefault(p, {})
        cur[parts[-1]] = value
    return nested.get("inputs", nested)
@pytest.fixture(scope="module")
 def lerobot_model():
    """Load the LeRobot GR00T N1.7 model once (fp32 + SDPA) and reuse across tags."""
    ckpt = _resolve_checkpoint()
    from lerobot.policies.groot.groot_n1_7 import GR00TN17
    model = GR00TN17.from_pretrained(
        ckpt,
        tune_llm=False,
        tune_visual=False,
        tune_projector=False,
        tune_diffusion_model=False,
        tune_vlln=False,
        transformers_loading_kwargs={"trust_remote_code": True},
    )
    # fp32 + SDPA on both sides: bf16 + differing attention kernels otherwise introduce
    # ~1e-2 numerical noise unrelated to the implementations.
    model.compute_dtype = "float32"
    model.config.compute_dtype = model.compute_dtype
    model.to(device=DEVICE, dtype=torch.float32)
    model.eval()
    return model
 _ARTIFACTS = _discover_artifacts()
@pytest.mark.skipif(
    not _ARTIFACTS,
    reason=(
        "No GR00T N1.7 parity artifacts found. Generate them first in the original gr00t "
        "env:\n  .venv-original/bin/python tests/policies/groot/utils/dump_original_n1_7.py "
        "--ckpt <ckpt> --out-dir tests/policies/groot/artifacts --device cuda"
    ),
 )
@pytest.mark.parametrize("embodiment_tag,artifact", _ARTIFACTS, ids=[t for t, _ in _ARTIFACTS])
 def test_groot_get_action_parity(embodiment_tag, artifact, lerobot_model):
    """Raw model.get_action(action_pred) parity per embodiment: original vs LeRobot."""
    original_action, flat_inputs = _load_artifact(artifact)
    model_inputs = _unflatten(flat_inputs)
    # Align the flow-matching RNG exactly as the producer did (seed right before sampling).
    torch.manual_seed(SEED)
    if torch.cuda.is_available():
-        torch.cuda.manual_seed_all(SEED)
+        torch.cuda.empty_cache()
-    with torch.inference_mode():
+        torch.cuda.synchronize()
-        out = lerobot_model.get_action(model_inputs)
+    if torch.backends.mps.is_available():
-    lerobot_action = out["action_pred"].float().cpu()
+        torch.mps.empty_cache()
    print("Memory cleanup complete.")
    t = min(original_action.shape[1], lerobot_action.shape[1])
    d = min(original_action.shape[2], lerobot_action.shape[2])
    original_action = original_action[:, :t, :d]
    lerobot_action = lerobot_action[:, :t, :d]
-    diff = torch.abs(lerobot_action - original_action)
+def set_seed_all(seed: int):
-    max_diff = diff.max().item()
+    """Set random seed for all RNG sources to ensure reproducibility."""
-    print(
+    import random
-        f"\n[{embodiment_tag}] shapes lerobot={tuple(lerobot_action.shape)} "
+
-        f"original={tuple(original_action.shape)}  "
+    random.seed(seed)
-        f"max|diff|={max_diff:.6e}  mean|diff|={diff.mean().item():.6e}"
+    np.random.seed(seed)
    torch.manual_seed(seed)
    if torch.cuda.is_available():
        torch.cuda.manual_seed(seed)
        torch.cuda.manual_seed_all(seed)
    # Set deterministic behavior
    torch.backends.cudnn.deterministic = True
    torch.backends.cudnn.benchmark = False
    torch.use_deterministic_algorithms(True, warn_only=True)
 def instantiate_lerobot_groot(
    from_pretrained: bool = False,
    model_path: str = MODEL_PATH,
 ) -> tuple[
    GrootPolicy,
    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
    PolicyProcessorPipeline[PolicyAction, PolicyAction],
 ]:
    """Instantiate LeRobot Groot policy with preprocessor and postprocessor."""
    if from_pretrained:
        policy = GrootPolicy.from_pretrained(
            pretrained_name_or_path=model_path,
            strict=False,
        )
        policy.config.embodiment_tag = "gr1"
    else:
        config = GrootConfig(
            base_model_path=model_path,
            n_action_steps=DUMMY_ACTION_HORIZON,
            chunk_size=DUMMY_ACTION_HORIZON,
            image_size=[IMAGE_SIZE, IMAGE_SIZE],
            device=DEVICE,
            embodiment_tag="gr1",
        )
        policy = GrootPolicy(config)
    policy.to(DEVICE)
    policy.config.device = DEVICE
    preprocessor, postprocessor = make_groot_pre_post_processors(
        config=policy.config,
        dataset_stats=None,  # Pass None for dataset_stats to disable normalization (original GR00T doesn't normalize)
    )
-    assert torch.allclose(lerobot_action, original_action, atol=ATOL, rtol=RTOL), (
+    return (policy, preprocessor, postprocessor)
-        f"GR00T N1.7 raw action_pred differs for embodiment '{embodiment_tag}' beyond "
+
-        f"atol={ATOL}, rtol={RTOL}: max|diff|={max_diff:.6e}"
+
 def instantiate_original_groot(
    from_pretrained: bool = False,
    model_path: str = MODEL_PATH,
 ):
    """Instantiate original Groot policy from NVIDIA's implementation."""
    from gr00t.data.transform.concat import ConcatTransform
    from gr00t.data.transform.state_action import StateActionToTensor
    from gr00t.data.transform.video import VideoToNumpy, VideoToTensor
    from gr00t.model.transforms import GR00TTransform
    video_keys = ["video.ego_view"]
    state_keys = [
        "state"
    ]  # Important: Use single concatenated "state" key (not split body parts) to match preprocessing
    action_keys = [
        "action.left_arm",
        "action.right_arm",
        "action.left_hand",
        "action.right_hand",
        "action.left_leg",
        "action.right_leg",
        "action.neck",
        "action.waist",
    ]
    language_keys = ["annotation.human.action.task_description"]
    modality_config = {
        "video": ModalityConfig(
            delta_indices=[0],  # Current frame only
            modality_keys=video_keys,
        ),
        "state": ModalityConfig(
            delta_indices=[0],
            modality_keys=state_keys,
        ),
        "action": ModalityConfig(
            delta_indices=list(range(DUMMY_ACTION_HORIZON)),
            modality_keys=action_keys,
        ),
        "language": ModalityConfig(
            delta_indices=[0],
            modality_keys=language_keys,
        ),
    }
    modality_transform = ComposedModalityTransform(
        transforms=[
            VideoToTensor(apply_to=video_keys),
            VideoToNumpy(apply_to=video_keys),  # Convert to numpy (GR00TTransform expects numpy arrays)
            # State is already a single concatenated key, so no StateActionToTensor needed
            # Convert action from numpy to tensor
            StateActionToTensor(apply_to=action_keys),
            # Concatenate only video and actions (state is already single key)
            ConcatTransform(
                video_concat_order=video_keys,
                state_concat_order=[],  # Empty:state is already single key
                action_concat_order=action_keys,
            ),
            GR00TTransform(
                max_state_dim=64,
                max_action_dim=32,
                state_horizon=1,
                action_horizon=DUMMY_ACTION_HORIZON,
                training=False,
            ),
        ]
    )
    policy = Gr00tPolicy(
        model_path=model_path,
        embodiment_tag=EmbodimentTag.GR1,
        modality_config=modality_config,
        modality_transform=modality_transform,
        device=DEVICE,
    )
    return policy, modality_config, modality_transform
 def create_dummy_data(device=DEVICE):
    """Create dummy data for testing both implementations."""
    batch_size = 2
    prompt = "Pick up the red cube and place it in the bin"
    state = torch.randn(batch_size, DUMMY_STATE_DIM, dtype=torch.float32, device=device)
    batch = {
        "observation.state": state,
        "action": torch.randn(
            batch_size,
            DUMMY_ACTION_HORIZON,
            DUMMY_ACTION_DIM,
            dtype=torch.float32,
            device=device,  # Action ground truth (for training)
        ),
        "observation.images.ego_view": torch.rand(
            batch_size,
            3,
            IMAGE_SIZE,
            IMAGE_SIZE,
            dtype=torch.float32,
            device=device,  # Images in [0, 1] range as expected by LeRobot
        ),
        "task": [prompt for _ in range(batch_size)],
    }
    return batch
 def convert_lerobot_to_original_format(batch, modality_config):
    """Convert LeRobot batch format to original Groot format.
    The original Groot expects observations in this format:
    {
        "video.<camera_name>": np.ndarray (T, H, W, C) or (B, T, H, W, C)
        "state.<state_component>": np.ndarray (T, D) or (B, T, D)
        "action.<action_component>": np.ndarray (T, D) or (B, T, D)
        "annotation.<annotation_type>": str or list[str]
    }
    """
    # Original Groot expects (T, H, W, C) format for images
    # LeRobot has (B, C, H, W) format, so we need to convert
    observation = {}
    for img_key in ["ego_view"]:
        lerobot_key = f"observation.images.{img_key}"
        if lerobot_key in batch:
            img = batch[lerobot_key]
            # Convert from (B, C, H, W) to (B, T=1, H, W, C)
            img_np = img.permute(0, 2, 3, 1).unsqueeze(1).cpu().numpy()
            # Convert [0, 1] to [0, 255] uint8 as expected by original
            img_np = (img_np * 255).astype(np.uint8)
            observation[f"video.{img_key}"] = img_np
    # Important: The Original's GR00TTransform expects "state" as (B, T, D), not split body parts
    if "observation.state" in batch:
        state = batch["observation.state"]
        state_np = state.unsqueeze(1).cpu().numpy()  # (B, 1, D)
        observation["state"] = state_np
    if "action" in batch:
        action = batch["action"]
        action_np = action.cpu().numpy()
        start_idx = 0
        for part_name, part_dim in GR1_BODY_PARTS.items():
            end_idx = start_idx + part_dim
            observation[f"action.{part_name}"] = action_np[:, :, start_idx:end_idx]
            start_idx = end_idx
    if "task" in batch:
        task_list = batch["task"]
        # GR00TTransform expects language with (B, T) shape for batched data
        # Create a (B, T=1) array where each element is the string directly
        bsz = len(task_list)
        task_array = np.empty((bsz, 1), dtype=object)
        for i in range(bsz):
            task_array[i, 0] = task_list[i]  # Assign string directly to each (i, 0) position
        observation["annotation.human.action.task_description"] = task_array
    return observation
 def test_groot_original_vs_lerobot_pretrained():
    """Test Groot original implementation vs LeRobot implementation with pretrained weights."""
    print("Test: Groot Original vs LeRobot with Pretrained Weights (Inference)")
    set_seed_all(42)
    lerobot_policy, lerobot_preprocessor, lerobot_postprocessor = instantiate_lerobot_groot(
        from_pretrained=True
    )
    original_policy, modality_config, modality_transform = instantiate_original_groot(from_pretrained=True)
    batch = create_dummy_data()
    batch_lerobot = deepcopy(batch)
    print("\n[LeRobot] Running inference...")
    lerobot_policy.eval()
    batch_lerobot_processed = lerobot_preprocessor(batch_lerobot)
    # Important: Reset seed immediately before inference to ensure identical RNG state
    torch.manual_seed(42)
    with torch.no_grad():
        lerobot_actions = lerobot_policy.select_action(batch_lerobot_processed)
    print("\n[Original] Running inference...")
    original_policy.model.eval()
    observation = convert_lerobot_to_original_format(batch, modality_config)
    original_obs_transformed = modality_transform(deepcopy(observation))
    # Important: Reset seed immediately before inference to ensure identical RNG state
    torch.manual_seed(42)
    with torch.no_grad():
        original_model_output = original_policy.model.get_action(original_obs_transformed)
        original_actions_raw = original_model_output["action_pred"]  # [2, 16, 32]
    # Take first timestep
    original_actions = original_actions_raw[:, 0, :].to(lerobot_actions.device).to(lerobot_actions.dtype)
    print("Action Comparison:")
    diff = lerobot_actions - original_actions
    abs_diff = torch.abs(diff)
    for batch_idx in range(lerobot_actions.shape[0]):
        print(f"\n{'=' * 60}")
        print(f"Batch {batch_idx}")
        print(f"{'=' * 60}")
        print(f"{'Idx':<5} {'LeRobot':<14} {'Original':<14} {'Difference':<14}")
        print("-" * 60)
        for action_idx in range(lerobot_actions.shape[1]):
            lr_val = lerobot_actions[batch_idx, action_idx].item()
            orig_val = original_actions[batch_idx, action_idx].item()
            diff_val = abs(lr_val - orig_val)
            sign = "+" if (lr_val - orig_val) > 0 else "-"
            print(f"{action_idx:<5} {lr_val:>13.6f} {orig_val:>13.6f} {sign}{diff_val:>12.6f}")
    max_diff = abs_diff.max().item()
    tolerance = 0.001
    assert torch.allclose(lerobot_actions, original_actions, atol=tolerance), (
        f"Actions differ by more than tolerance ({tolerance}): max diff = {max_diff:.6f}"
    )
    print(f"\nSuccess: Actions match within tolerance ({tolerance})!")
    del lerobot_policy, lerobot_preprocessor, lerobot_postprocessor
    del original_policy, modality_config, modality_transform
    del batch, batch_lerobot, observation
    cleanup_memory()
 def test_groot_forward_pass_comparison():
    """Test forward pass comparison between LeRobot and Original Groot implementations."""
    print("Test: Forward Pass Comparison (Training Mode)")
    set_seed_all(42)
    lerobot_policy, lerobot_preprocessor, lerobot_postprocessor = instantiate_lerobot_groot(
        from_pretrained=True
    )
    original_policy, modality_config, modality_transform = instantiate_original_groot(from_pretrained=True)
    batch = create_dummy_data()
    lerobot_policy.eval()
    original_policy.model.eval()
    print("\n[LeRobot] Running forward pass...")
    batch_lerobot = deepcopy(batch)
    batch_lerobot_processed = lerobot_preprocessor(batch_lerobot)
    set_seed_all(42)
    with torch.no_grad():
        lerobot_loss, lerobot_metrics = lerobot_policy.forward(batch_lerobot_processed)
    print(f"  Loss: {lerobot_loss.item():.6f}")
    print("\n[Original] Running forward pass...")
    observation = convert_lerobot_to_original_format(batch, modality_config)
    transformed_obs = modality_transform(observation)
    if "action" not in transformed_obs:
        action_for_forward = batch_lerobot_processed["action"]
        action_mask_for_forward = batch_lerobot_processed["action_mask"]
        # Match action horizon if needed
        if action_for_forward.shape[1] != original_policy.model.action_horizon:
            if action_for_forward.shape[1] < original_policy.model.action_horizon:
                pad_size = original_policy.model.action_horizon - action_for_forward.shape[1]
                last_action = action_for_forward[:, -1:, :]
                padding = last_action.repeat(1, pad_size, 1)
                action_for_forward = torch.cat([action_for_forward, padding], dim=1)
                mask_padding = torch.zeros(
                    action_mask_for_forward.shape[0],
                    pad_size,
                    action_mask_for_forward.shape[2],
                    dtype=action_mask_for_forward.dtype,
                    device=action_mask_for_forward.device,
                )
                action_mask_for_forward = torch.cat([action_mask_for_forward, mask_padding], dim=1)
            else:
                action_for_forward = action_for_forward[:, : original_policy.model.action_horizon, :]
                action_mask_for_forward = action_mask_for_forward[
                    :, : original_policy.model.action_horizon, :
                ]
        transformed_obs["action"] = action_for_forward
        transformed_obs["action_mask"] = action_mask_for_forward
    set_seed_all(42)
    with torch.no_grad():
        original_outputs = original_policy.model.forward(transformed_obs)
    original_loss = original_outputs["loss"]
    print(f"  Loss: {original_loss.item():.6f}")
    loss_diff = abs(lerobot_loss.item() - original_loss.item())
    loss_rel_diff = loss_diff / (abs(original_loss.item()) + 1e-8) * 100
    print("\nLoss Values:")
    print(f"  LeRobot: {lerobot_loss.item():.6f}")
    print(f"  Original: {original_loss.item():.6f}")
    print(f"  Absolute difference: {loss_diff:.6f}")
    print(f"  Relative difference: {loss_rel_diff:.2f}%")
    del lerobot_policy, lerobot_preprocessor, lerobot_postprocessor
    del original_policy, modality_config, modality_transform
    del batch, batch_lerobot, observation, transformed_obs
    cleanup_memory()
@@ -1,212 +0,0 @@
 #!/usr/bin/env python
 # Copyright 2026 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
 #     http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """Producer (run in the ORIGINAL gr00t env): dump original GR00T N1.7 outputs + inputs.
 The original NVIDIA ``gr00t`` package pins ``transformers==4.57.3`` (py3.10) and its
 model-config dataclasses are incompatible with the ``transformers==5.x`` that the
 LeRobot GR00T N1.7 integration requires. The two implementations therefore cannot be
 imported in the same Python process. To keep the parity comparison FAIR, we run the
 original model in its native env here and serialize, PER EMBODIMENT TAG:
  * the exact pre-processed/collated model inputs (so the LeRobot side consumes the
    byte-identical tensors -- same image preprocessing, tokenization, normalization),
  * the random seed used right before the flow-matching sampler,
  * the raw ``action_pred`` tensor returned by ``model.get_action`` (normalized space,
    before any per-implementation action decoding).
 Inputs are built GENERICALLY from the checkpoint metadata (no per-tag hardcoding):
 state keys + dims come from ``statistics.json``; video + language keys come from the
 processor's per-embodiment modality configs. This lets us test many embodiment tags
 from the SAME checkpoint and confirm the LeRobot integration is not overfit to
 ``libero_sim``.
 The companion pytest (run in the LeRobot env) loads each .npz, replays the identical
 inputs + seed through the LeRobot GR00T N1.7 model, and asserts the outputs match.
 Usage:
    .venv-original/bin/python tests/policies/groot/utils/dump_original_n1_7.py \
        --ckpt <path-to-GR00T-N1.7-LIBERO/libero_10> \
        --out-dir tests/policies/groot/artifacts \
        [--tags libero_sim,oxe_droid_relative_eef_relative_joint,...] \
        [--device cuda] [--seed 42]
 If --tags is omitted, every embodiment present in the checkpoint statistics is dumped.
 """
 import argparse
 import json
 import os
 from pathlib import Path
 import numpy as np
 import torch
 IMAGE_SIZE = 256
 BATCH_SIZE = 2
 PROMPT = "pick up the black bowl and place it on the plate"
 def load_statistics(ckpt: str) -> dict:
    with open(os.path.join(ckpt, "statistics.json")) as f:
        return json.load(f)
 def make_observation(seed: int, video_keys, lang_key, state_spec):
    """Build a dummy observation dict generically from the embodiment metadata."""
    rng = np.random.default_rng(seed)
    video = {
        k: rng.integers(0, 256, (BATCH_SIZE, 1, IMAGE_SIZE, IMAGE_SIZE, 3), dtype=np.uint8)
        for k in video_keys
    }
    # One ndarray per state key, shape (B, T=1, key_dim); dim taken from statistics.
    # Keys with dim 0 (e.g. disabled eef on some embodiments) are still emitted as
    # present-but-empty so the processor's state transform finds every expected key.
    state = {k: rng.standard_normal((BATCH_SIZE, 1, dim)).astype(np.float32) for k, dim in state_spec}
    language = {lang_key: [[PROMPT] for _ in range(BATCH_SIZE)]}
    return {"video": video, "state": state, "language": language}
 def dump_one_tag(policy, fair_model, tag, modality_cfg, state_spec, args, out_path):
    from gr00t.data.types import MessageType
    video_keys = modality_cfg["video"].modality_keys
    lang_key = modality_cfg["language"].modality_keys[0]
    observation = make_observation(args.seed, video_keys, lang_key, state_spec)
    # Point the policy preprocessing at this embodiment (mirrors Gr00tPolicy.__init__).
    policy.embodiment_tag = type(policy.embodiment_tag)(tag)
    policy.modality_configs = {
        k: v for k, v in policy.processor.get_modality_configs()[tag].items() if k != "rl_info"
    }
    policy.language_key = policy.modality_configs["language"].modality_keys[0]
    torch.manual_seed(args.seed)
    np.random.seed(args.seed)
    unbatched = policy._unbatch_observation(observation)
    processed = []
    for obs in unbatched:
        vla = policy._to_vla_step_data(obs)
        processed.append(policy.processor([{"type": MessageType.EPISODE_STEP.value, "content": vla}]))
    collated = policy.collate_fn(processed)
    def to_dev(x):
        if isinstance(x, torch.Tensor) and torch.is_floating_point(x):
            return x.to(args.device, torch.float32)
        if isinstance(x, torch.Tensor):
            return x.to(args.device)
        if isinstance(x, dict):
            return {k: to_dev(v) for k, v in x.items()}
        return x
    collated = {k: to_dev(v) for k, v in collated.items()}
    torch.manual_seed(args.seed)
    with torch.inference_mode():
        out = fair_model.get_action(**collated)
    action_pred = out["action_pred"].float().cpu().numpy()
    flat, meta = {}, {}
    def flatten(prefix, obj):
        if isinstance(obj, torch.Tensor):
            arr = obj.float().cpu().numpy() if torch.is_floating_point(obj) else obj.cpu().numpy()
            flat[f"in::{prefix}"] = arr
            meta[f"in::{prefix}"] = str(obj.dtype)
        elif isinstance(obj, dict):
            for k, v in obj.items():
                flatten(f"{prefix}.{k}" if prefix else k, v)
        elif isinstance(obj, (list, tuple)):
            flat[f"in::{prefix}"] = np.array(obj, dtype=object)
        else:
            flat[f"in::{prefix}"] = np.array(obj)
    flatten("", collated)
    out_path.parent.mkdir(parents=True, exist_ok=True)
    np.savez(
        out_path,
        action_pred=action_pred,
        seed=np.array(args.seed),
        device=np.array(args.device),
        embodiment_tag=np.array(tag),
        meta_keys=np.array(list(meta.keys()), dtype=object),
        meta_dtypes=np.array(list(meta.values()), dtype=object),
        **flat,
    )
    print(f"[{tag}] action_pred {action_pred.shape} -> {out_path.name} ({os.path.getsize(out_path)} B)")
 def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--ckpt", required=True)
    ap.add_argument("--out-dir", required=True, help="directory for per-tag .npz files")
    ap.add_argument("--tags", default="", help="comma-separated embodiment tags (default: all in stats)")
    ap.add_argument("--device", default="cuda")
    ap.add_argument("--seed", type=int, default=42)
    args = ap.parse_args()
    from gr00t.policy.gr00t_policy import Gr00tPolicy
    from transformers import AutoConfig, AutoModel
    stats = load_statistics(args.ckpt)
    requested = [t.strip() for t in args.tags.split(",") if t.strip()] or list(stats.keys())
    # Load the policy once (for its processor/preprocessing) on any valid tag.
    bootstrap_tag = "libero_sim" if "libero_sim" in stats else requested[0]
    policy = Gr00tPolicy(embodiment_tag=bootstrap_tag, model_path=args.ckpt, device=args.device)
    all_modality = policy.processor.get_modality_configs()
    # Load a FAIR model (SDPA + fp32) once and reuse across tags. Otherwise the
    # original checkpoint default (flash_attention_2 + bf16) introduces kernel/rounding
    # noise vs the LeRobot env (which has no flash_attn and runs SDPA).
    cfg = AutoConfig.from_pretrained(args.ckpt, trust_remote_code=True)
    cfg.use_flash_attention = False
    cfg.load_bf16 = False
    fair_model = AutoModel.from_pretrained(args.ckpt, config=cfg, trust_remote_code=True)
    fair_model.to(device=args.device, dtype=torch.float32)
    fair_model.eval()
    out_dir = Path(args.out_dir)
    done, skipped = [], []
    for tag in requested:
        if tag not in stats or tag not in all_modality:
            print(f"[skip] {tag}: not present in checkpoint statistics/modality configs")
            skipped.append(tag)
            continue
        state_spec = [(k, len(v["min"])) for k, v in stats[tag]["state"].items()]
        try:
            dump_one_tag(
                policy,
                fair_model,
                tag,
                all_modality[tag],
                state_spec,
                args,
                out_dir / f"original_n1_7_{tag}.npz",
            )
            done.append(tag)
        except Exception as exc:  # noqa: BLE001
            print(f"[fail] {tag}: {type(exc).__name__}: {exc}")
            skipped.append(tag)
    print(f"\nDumped {len(done)} tags: {done}")
    if skipped:
        print(f"Skipped/failed {len(skipped)} tags: {skipped}")
 if __name__ == "__main__":
    main()
@@ -1,78 +0,0 @@
 #!/usr/bin/env python
 # Copyright 2026 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
 #     http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from __future__ import annotations
 import pytest
 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.configs.types import FeatureType, PolicyFeature
 from lerobot.policies.lingbot_va.configuration_lingbot_va import LingBotVAConfig
 from lerobot.utils.constants import ACTION, OBS_IMAGES
 def make_config(**overrides) -> LingBotVAConfig:
    kwargs = {"device": "cpu"}
    kwargs.update(overrides)
    return LingBotVAConfig(**kwargs)
 def test_registered_in_choice_registry() -> None:
    assert "lingbot_va" in PreTrainedConfig.get_known_choices()
    assert PreTrainedConfig.get_choice_class("lingbot_va") is LingBotVAConfig
 def test_type_property() -> None:
    assert make_config().type == "lingbot_va"
 def test_chunk_size_and_action_steps() -> None:
    cfg = make_config(frame_chunk_size=4, action_per_frame=4)
    assert cfg.chunk_size == 16
    assert cfg.n_action_steps == 16
    assert cfg.action_delta_indices == list(range(16))
    assert cfg.observation_delta_indices == list(range(16))
    assert cfg.reward_delta_indices is None
 def test_optimizer_and_scheduler_presets() -> None:
    cfg = make_config()
    opt = cfg.get_optimizer_preset()
    assert opt.lr == cfg.optimizer_lr
    sched = cfg.get_scheduler_preset()
    assert sched.num_warmup_steps == cfg.scheduler_warmup_steps
 def test_validate_features_sets_action_feature() -> None:
    cfg = make_config()
    cfg.input_features = {f"{OBS_IMAGES}.image": PolicyFeature(type=FeatureType.VISUAL, shape=(3, 128, 128))}
    cfg.output_features = {}
    cfg.validate_features()
    assert ACTION in cfg.output_features
    assert cfg.output_features[ACTION].shape == (len(cfg.used_action_channel_ids),)
 def test_validate_features_no_visual_raises() -> None:
    cfg = make_config()
    cfg.input_features = {}
    cfg.output_features = {}
    with pytest.raises(ValueError, match="at least one visual input feature"):
        cfg.validate_features()
 def test_invalid_attn_mode_raises() -> None:
    with pytest.raises(ValueError, match="attn_mode"):
        make_config(attn_mode="banana")
@@ -1,38 +0,0 @@
 #!/usr/bin/env python
 # Copyright 2026 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
 #     http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from __future__ import annotations
 import pytest
 from lerobot.policies.factory import make_policy_config
 from lerobot.policies.lingbot_va.configuration_lingbot_va import LingBotVAConfig
 def test_make_policy_config_returns_lingbot_va() -> None:
    cfg = make_policy_config("lingbot_va", device="cpu")
    assert isinstance(cfg, LingBotVAConfig)
 def test_get_policy_class_resolves_lazily() -> None:
    # Importing the policy class pulls in diffusers (Wan2.2 stack); skip if unavailable.
    pytest.importorskip("diffusers")
    pytest.importorskip("transformers")
    from lerobot.policies.factory import get_policy_class
    cls = get_policy_class("lingbot_va")
    assert cls.name == "lingbot_va"
    assert cls.config_class is LingBotVAConfig
@@ -1,128 +0,0 @@
 #!/usr/bin/env python
 # Copyright 2026 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
 #     http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """Unit tests for the vendored LingBot-VA helper code (scheduler + grid utilities)."""
 from __future__ import annotations
 import pytest
 import torch
 pytest.importorskip("diffusers")  # the model code lives in modeling_lingbot_va, which imports diffusers
 from lerobot.policies.lingbot_va.modeling_lingbot_va import FlowMatchScheduler
 from lerobot.policies.lingbot_va.utils import data_seq_to_patch, get_mesh_id
 def test_flow_match_scheduler_timesteps_monotone_decreasing() -> None:
    sch = FlowMatchScheduler(shift=5.0, sigma_min=0.0, extra_one_step=True)
    sch.set_timesteps(20)
    assert sch.timesteps.shape == (20,)
    diffs = sch.timesteps[1:] - sch.timesteps[:-1]
    assert torch.all(diffs <= 0)  # decreasing
 def test_flow_match_scheduler_step_preserves_shape() -> None:
    sch = FlowMatchScheduler(shift=5.0, sigma_min=0.0, extra_one_step=True)
    sch.set_timesteps(20)
    sample = torch.zeros(1, 48, 4, 8, 16)
    out = sch.step(torch.ones_like(sample), sch.timesteps[0], sample)
    assert out.shape == sample.shape
 def test_flow_match_scheduler_add_noise() -> None:
    sch = FlowMatchScheduler(shift=5.0, sigma_min=0.0, extra_one_step=True)
    sch.set_timesteps(20)
    sample = torch.randn(1, 48, 4, 8, 16)
    noise = torch.randn_like(sample)
    noisy = sch.add_noise(sample, noise, sch.timesteps[:4], t_dim=2)
    assert noisy.shape == sample.shape
 def test_get_mesh_id_latent_shape() -> None:
    grid = get_mesh_id(4, 8, 16, 0, 1, 0)
    assert grid.shape == (4, 4 * 8 * 16)  # (f, h, w, stream) x tokens
 def test_get_mesh_id_action_shape() -> None:
    grid = get_mesh_id(4, 4, 1, 1, 1, 0, action=True)
    assert grid.shape == (4, 4 * 4 * 1)
    # Action rows for h/w are sentinel -1.
    assert torch.all(grid[1] < 0)
    assert torch.all(grid[2] < 0)
 def test_data_seq_to_patch_roundtrip_shape() -> None:
    b, f, h, w, c = 1, 4, 8, 16, 48
    seq = torch.arange(b * f * h * w * c, dtype=torch.float32).reshape(b, f * h * w, c)
    out = data_seq_to_patch((1, 2, 2), seq, f, h, w, batch_size=b)
    assert out.shape == (b, c, f, h, w)
 def test_training_step_reduces_loss_tiny_flex() -> None:
    """End-to-end single training step (flow-matching loss -> backward -> AdamW) on a tiny config.
    Exercises the flex-attention training path; requires a CUDA GPU with flex-attention support.
    """
    if not torch.cuda.is_available():
        import pytest
        pytest.skip("training step test requires a CUDA GPU (flex-attention)")
    from lerobot.configs.types import FeatureType, PolicyFeature
    from lerobot.policies.lingbot_va.configuration_lingbot_va import LingBotVAConfig
    from lerobot.policies.lingbot_va.modeling_lingbot_va import LingBotVAPolicy
    from lerobot.utils.constants import ACTION, OBS_IMAGES
    cfg = LingBotVAConfig(
        attn_mode="flex",
        dtype="bfloat16",
        in_channels=16,
        out_channels=16,
        action_dim=8,
        text_dim=32,
        freq_dim=64,
        ffn_dim=64,
        num_attention_heads=2,
        attention_head_dim=24,
        num_layers=2,
        frame_chunk_size=2,
        action_per_frame=4,
        used_action_channel_ids=[0, 1, 2, 3],
        obs_cam_keys=[f"{OBS_IMAGES}.image"],
        device="cuda",
    )
    cfg.input_features = {f"{OBS_IMAGES}.image": PolicyFeature(type=FeatureType.VISUAL, shape=(3, 64, 64))}
    cfg.output_features = {ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(4,))}
    cfg.validate_features()
    policy = LingBotVAPolicy(cfg).to("cuda")
    policy.train()
    opt = torch.optim.AdamW(policy.get_optim_params(), lr=1e-4)
    b, fc, apf = 1, cfg.frame_chunk_size, cfg.action_per_frame
    latents = torch.randn(b, cfg.in_channels, fc, 4, 4, device="cuda", dtype=torch.bfloat16)
    actions = torch.randn(b, cfg.action_dim, fc, apf, 1, device="cuda", dtype=torch.bfloat16)
    amask = torch.zeros(cfg.action_dim, device="cuda")
    amask[cfg.used_action_channel_ids] = 1.0
    actions_mask = amask.view(1, -1, 1, 1, 1).expand_as(actions)
    text_emb = torch.randn(b, cfg.max_sequence_length, cfg.text_dim, device="cuda", dtype=torch.bfloat16)
    loss, metrics = policy.training_loss_from_streams(latents, actions, actions_mask, text_emb)
    assert torch.isfinite(loss) and {"latent_loss", "action_loss"} <= set(metrics)
    loss.backward()
    assert any(p.grad is not None and torch.isfinite(p.grad).all() for p in policy.get_optim_params())
    opt.step()
@@ -1,88 +0,0 @@
 #!/usr/bin/env python
 # Copyright 2026 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
 #     http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from __future__ import annotations
 import torch
 from lerobot.configs.types import FeatureType, PolicyFeature
 from lerobot.policies.lingbot_va.configuration_lingbot_va import LingBotVAConfig
 from lerobot.policies.lingbot_va.processor_lingbot_va import make_lingbot_va_pre_post_processors
 from lerobot.processor import PolicyProcessorPipeline, UnnormalizerProcessorStep
 from lerobot.processor.converters import policy_action_to_transition, transition_to_policy_action
 from lerobot.utils.constants import (
    ACTION,
    OBS_IMAGES,
    POLICY_POSTPROCESSOR_DEFAULT_NAME,
    POLICY_PREPROCESSOR_DEFAULT_NAME,
 )
 def _make_config() -> LingBotVAConfig:
    cfg = LingBotVAConfig(device="cpu")
    cfg.input_features = {f"{OBS_IMAGES}.image": PolicyFeature(type=FeatureType.VISUAL, shape=(3, 128, 128))}
    cfg.output_features = {}
    cfg.validate_features()
    return cfg
 def test_make_pre_post_processors_names_and_steps() -> None:
    cfg = _make_config()
    pre, post = make_lingbot_va_pre_post_processors(cfg, dataset_stats=None)
    assert pre.name == POLICY_PREPROCESSOR_DEFAULT_NAME
    assert post.name == POLICY_POSTPROCESSOR_DEFAULT_NAME
    # Actions are unnormalized by the standard built-in quantile unnormalizer.
    assert any(isinstance(s, UnnormalizerProcessorStep) for s in post.steps)
 def test_freshly_built_postprocessor_is_identity() -> None:
    # Without action stats the quantile unnormalizer is a no-op (identity passthrough): the real
    # per-benchmark q01/q99 are restored from the saved checkpoint on load, not hardcoded here.
    cfg = _make_config()
    _, post = make_lingbot_va_pre_post_processors(cfg, dataset_stats=None)
    normed = torch.tensor([[0.3, -0.5, 1.0, -1.0, 0.0, 0.7, -0.2]])
    assert torch.allclose(post(normed), normed, atol=1e-6)
 def test_postprocessor_quantile_unnormalization() -> None:
    # QUANTILES unnormalize maps [-1, 1] -> [q01, q99]: -1 -> q01, +1 -> q99.
    cfg = _make_config()
    q01 = [-1.0, -0.5, 0.0, -1.0, -1.0, -1.0, -1.0]
    q99 = [1.0, 0.5, 2.0, 1.0, 1.0, 1.0, 1.0]
    stats = {ACTION: {"q01": q01, "q99": q99}}
    _, post = make_lingbot_va_pre_post_processors(cfg, dataset_stats=stats)
    out_lo = post(torch.full((1, 7), -1.0))
    out_hi = post(torch.full((1, 7), 1.0))
    assert torch.allclose(out_lo, torch.tensor(q01).unsqueeze(0), atol=1e-4)
    assert torch.allclose(out_hi, torch.tensor(q99).unsqueeze(0), atol=1e-4)
 def test_postprocessor_stats_survive_save_load(tmp_path) -> None:
    # Regression guard for the Hub mechanism: the q01/q99 stats live in the saved post-processor
    # state and must round-trip through save_pretrained / from_pretrained.
    cfg = _make_config()
    q01 = [-0.6, -0.8, -0.9, -0.1, -0.15, -0.25, -1.0]
    q99 = [0.9, 0.85, 0.9, 0.17, 0.18, 0.34, 1.0]
    _, post = make_lingbot_va_pre_post_processors(cfg, dataset_stats={ACTION: {"q01": q01, "q99": q99}})
    post.save_pretrained(tmp_path)
    loaded = PolicyProcessorPipeline.from_pretrained(
        tmp_path,
        config_filename=f"{POLICY_POSTPROCESSOR_DEFAULT_NAME}.json",
        to_transition=policy_action_to_transition,
        to_output=transition_to_policy_action,
    )
    out = loaded(torch.full((1, 7), -1.0))
    assert torch.allclose(out, torch.tensor(q01).unsqueeze(0), atol=1e-4)
@@ -1,101 +0,0 @@
 #!/usr/bin/env python
 # Copyright 2025 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
 #     http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """Tests for the Foxglove backend's pure helpers.
 These cover topic naming, series labelling and feature-name parsing. They import
 ``foxglove_visualization`` directly and need NO ``foxglove`` extra: the SDK is imported lazily inside
 the functions that talk to the server, so the helpers below run in the base test tier.
 """
 import numpy as np
 from lerobot.utils import foxglove_visualization as fv
 from lerobot.utils.constants import ACTION, OBS_STATE
 def test_foxglove_safe_name_collapses_dots():
    assert fv._foxglove_safe_name("observation.images.front") == "observation_images_front"
    assert fv._foxglove_safe_name("plain") == "plain"
 def test_foxglove_topic_image_strips_prefix_without_doubling_images():
    # Fully-qualified camera key -> single clean segment (no doubled "images").
    assert fv._foxglove_topic("observation.images.front", is_image=True) == "/observation/images/front"
    # A nested camera name keeps its structure via safe-name collapsing.
    assert (
        fv._foxglove_topic("observation.images.wrist.left", is_image=True) == "/observation/images/wrist_left"
    )
    # Bare camera name (as real robots emit).
    assert fv._foxglove_topic("front", is_image=True) == "/observation/images/front"
 def test_foxglove_topic_scalar_sources():
    assert fv._foxglove_topic(OBS_STATE) == "/observation/state"
    assert fv._foxglove_topic("observation.environment_state") == "/observation/state"
    assert fv._foxglove_topic(ACTION) == "/action/state"
    assert fv._foxglove_topic("action.delta") == "/action/state"
 def test_labeled_scalars_uses_labels_then_index_fallback():
    assert fv._labeled_scalars("state", np.array([1.0, 2.0, 3.0])) == {
        "state_0": 1.0,
        "state_1": 2.0,
        "state_2": 3.0,
    }
    assert fv._labeled_scalars("state", [1.0, 2.0], ["pan", "lift"]) == {"pan": 1.0, "lift": 2.0}
    # Wrong-length labels fall back to index naming (never silently mislabels).
    assert fv._labeled_scalars("q", [1.0, 2.0], ["only_one"]) == {"q_0": 1.0, "q_1": 2.0}
 def test_frame_to_scalars_matches_live_labeling_and_handles_scalar():
    frame = {OBS_STATE: np.array([1.0, 2.0])}
    # No metadata -> {short_name}_{i}, identical to the live-stream fallback.
    assert fv._frame_to_scalars(frame, OBS_STATE) == fv._labeled_scalars("state", np.array([1.0, 2.0]))
    assert fv._frame_to_scalars(frame, OBS_STATE) == {"state_0": 1.0, "state_1": 2.0}
    # Metadata labels are honored.
    assert fv._frame_to_scalars(frame, OBS_STATE, ["pan", "lift"]) == {"pan": 1.0, "lift": 2.0}
    # A 0-d scalar becomes a single entry named by the short feature name.
    assert fv._frame_to_scalars({ACTION: np.array(5.0)}, ACTION) == {"action": 5.0}
    # A missing feature yields an empty mapping.
    assert fv._frame_to_scalars({}, OBS_STATE) == {}
 def test_feature_dim_names_formats():
    # Flat list of names.
    assert fv._feature_dim_names({"shape": [2], "names": ["x", "y"]}) == ["x", "y"]
    # Category mapping (dict of lists).
    assert fv._feature_dim_names({"shape": [2], "names": {"motors": ["m0", "m1"]}}) == ["m0", "m1"]
    # name -> index mapping (returned sorted by index).
    assert fv._feature_dim_names({"shape": [2], "names": {"delta_x": 0, "delta_y": 1}}) == [
        "delta_x",
        "delta_y",
    ]
    # Bool values must NOT be treated as an index map (bool is a subclass of int).
    assert fv._feature_dim_names({"shape": [2], "names": {"a": True, "b": False}}) is None
    # Mismatched length -> None (won't silently mislabel).
    assert fv._feature_dim_names({"shape": [3], "names": ["x", "y"]}) is None
    # Missing / absent names -> None.
    assert fv._feature_dim_names(None) is None
    assert fv._feature_dim_names({"shape": [2]}) is None
 def test_is_scalar():
    assert fv._is_scalar(1.0)
    assert fv._is_scalar(np.float32(2.0))
    assert fv._is_scalar(np.array(3.0))  # 0-d array
    assert not fv._is_scalar(np.array([1.0, 2.0]))
    assert not fv._is_scalar("x")
@@ -1,311 +0,0 @@
 #!/usr/bin/env python
 # Copyright 2025 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
 #     http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import importlib
 import sys
 from types import SimpleNamespace
 import numpy as np
 import pytest
 pytest.importorskip("rerun", reason="rerun-sdk is required (install lerobot[viz])")
 from lerobot.types import TransitionKey
 from lerobot.utils.constants import OBS_STATE
@pytest.fixture
 def mock_rerun(monkeypatch):
    """
    Provide a mock `rerun` module (and `rerun.blueprint` submodule) so tests don't
    depend on the real library. Also reload the module-under-test so it binds to
    this mock `rr`.
    """
    calls = []
    blueprints = []
    class DummyScalar:
        def __init__(self, value):
            # Scalars may be built from a single float or from a 1D array batch.
            self.value = value
    class DummyImage:
        def __init__(self, arr):
            self.arr = arr
        def compress(self, *a, **k):
            return self
    class DummyDepthImage:
        def __init__(self, arr, meter=None, colormap=None):
            self.arr = arr
            self.meter = meter
            self.colormap = colormap
    def dummy_log(key, obj=None, **kwargs):
        # Accept either positional `obj` or keyword `entity` and record remaining kwargs.
        if obj is None and "entity" in kwargs:
            obj = kwargs.pop("entity")
        calls.append((key, obj, kwargs))
    def dummy_send_blueprint(blueprint, *a, **k):
        blueprints.append(blueprint)
    # Mock the `rerun.blueprint` submodule used to build the layout.
    dummy_rrb = SimpleNamespace(
        Spatial2DView=lambda origin=None, name=None: SimpleNamespace(
            kind="Spatial2DView", origin=origin, name=name
        ),
        TimeSeriesView=lambda name=None, contents=None: SimpleNamespace(
            kind="TimeSeriesView", name=name, contents=contents
        ),
        Grid=lambda *views: SimpleNamespace(kind="Grid", views=list(views)),
        Blueprint=lambda root: SimpleNamespace(kind="Blueprint", root=root),
    )
    dummy_rr = SimpleNamespace(
        __name__="rerun",
        __package__="rerun",
        __spec__=SimpleNamespace(name="rerun", submodule_search_locations=None),
        Scalars=DummyScalar,
        Image=DummyImage,
        DepthImage=DummyDepthImage,
        components=SimpleNamespace(Colormap=SimpleNamespace(Viridis="viridis")),
        log=dummy_log,
        send_blueprint=dummy_send_blueprint,
        init=lambda *a, **k: None,
        spawn=lambda *a, **k: None,
        blueprint=dummy_rrb,
    )
    # Inject fake modules into sys.modules (both `rerun` and `rerun.blueprint`).
    monkeypatch.setitem(sys.modules, "rerun", dummy_rr)
    monkeypatch.setitem(sys.modules, "rerun.blueprint", dummy_rrb)
    # Now import and reload the module under test, to bind to our rerun mock
    import lerobot.utils.rerun_visualization as rv
    importlib.reload(rv)
    # Expose the reloaded module, the call recorder and the captured blueprints
    yield rv, calls, blueprints
 def _keys(calls):
    """Helper to extract just the keys logged to rr.log"""
    return [k for (k, _obj, _kw) in calls]
 def _obj_for(calls, key):
    """Find the first object logged under a given key."""
    for k, obj, _kw in calls:
        if k == key:
            return obj
    raise KeyError(f"Key {key} not found in calls: {calls}")
 def _kwargs_for(calls, key):
    for k, _obj, kw in calls:
        if k == key:
            return kw
    raise KeyError(f"Key {key} not found in calls: {calls}")
 def _views_by_kind(blueprint, kind):
    """Return the views of a given kind from the (single) blueprint's grid."""
    return [v for v in blueprint.root.views if v.kind == kind]
 def test_log_rerun_data_envtransition_scalars_and_image(mock_rerun):
    rv, calls, blueprints = mock_rerun
    # Build EnvTransition dict
    obs = {
        f"{OBS_STATE}.temperature": np.float32(25.0),
        # CHW image should be converted to HWC for rr.Image
        "observation.camera": np.zeros((3, 10, 20), dtype=np.uint8),
    }
    act = {
        "action.throttle": 0.7,
        # 1D array should be logged as a single Scalars batch under one entity path
        "action.vector": np.array([1.0, 2.0], dtype=np.float32),
    }
    transition = {
        TransitionKey.OBSERVATION: obs,
        TransitionKey.ACTION: act,
    }
    # Extract observation and action data from transition like in the real call sites
    obs_data = transition.get(TransitionKey.OBSERVATION, {})
    action_data = transition.get(TransitionKey.ACTION, {})
    rv.log_rerun_data(observation=obs_data, action=action_data)
    # We expect:
    # - observation.state.temperature -> Scalars
    # - observation.camera -> Image (HWC) with static=True
    # - action.throttle -> Scalars
    # - action.vector -> single Scalars batch (no per-element suffix)
    expected_keys = {
        f"{OBS_STATE}.temperature",
        "observation.camera",
        "action.throttle",
        "action.vector",
    }
    assert set(_keys(calls)) == expected_keys
    # Check scalar types and values
    temp_obj = _obj_for(calls, f"{OBS_STATE}.temperature")
    assert type(temp_obj).__name__ == "DummyScalar"
    assert float(temp_obj.value) == pytest.approx(25.0)
    throttle_obj = _obj_for(calls, "action.throttle")
    assert type(throttle_obj).__name__ == "DummyScalar"
    assert float(throttle_obj.value) == pytest.approx(0.7)
    # 1D vector logged as a single batched Scalars under one entity path
    vec = _obj_for(calls, "action.vector")
    assert type(vec).__name__ == "DummyScalar"
    np.testing.assert_allclose(np.asarray(vec.value), [1.0, 2.0])
    # Check image handling: CHW -> HWC
    img_obj = _obj_for(calls, "observation.camera")
    assert type(img_obj).__name__ == "DummyImage"
    assert img_obj.arr.shape == (10, 20, 3)  # transposed
    assert _kwargs_for(calls, "observation.camera").get("static", False) is True  # static=True for images
    # A blueprint should have been built and sent exactly once, and cached on the function.
    assert len(blueprints) == 1
    assert rv.log_rerun_data.blueprint is blueprints[0]
    bp = blueprints[0]
    # One spatial view per image path
    spatial_views = _views_by_kind(bp, "Spatial2DView")
    assert {v.origin for v in spatial_views} == {"observation.camera"}
    # One time-series view each for observation and action scalars
    ts_views = {v.name: v for v in _views_by_kind(bp, "TimeSeriesView")}
    assert set(ts_views) == {"observation", "action"}
    assert ts_views["observation"].contents == [f"{OBS_STATE}.temperature"]
    assert ts_views["action"].contents == ["action.throttle", "action.vector"]
 def test_log_rerun_data_plain_list_ordering_and_prefixes(mock_rerun):
    rv, calls, blueprints = mock_rerun
    # First dict without prefixes treated as observation
    # Second dict without prefixes treated as action
    obs_plain = {
        "temp": 1.5,
        # Already HWC image => should stay as-is
        "img": np.zeros((5, 6, 3), dtype=np.uint8),
        "none": None,  # should be skipped
    }
    act_plain = {
        "throttle": 0.3,
        "vec": np.array([9, 8, 7], dtype=np.float32),
    }
    # Extract observation and action data from list like the old function logic did
    # First dict was treated as observation, second as action
    rv.log_rerun_data(observation=obs_plain, action=act_plain)
    # Expected keys with auto-prefixes. The 1D vector is a single batched Scalars.
    expected = {
        "observation.temp",
        "observation.img",
        "action.throttle",
        "action.vec",
    }
    logged = set(_keys(calls))
    assert logged == expected
    # Scalars
    t = _obj_for(calls, "observation.temp")
    assert type(t).__name__ == "DummyScalar"
    assert float(t.value) == pytest.approx(1.5)
    throttle = _obj_for(calls, "action.throttle")
    assert type(throttle).__name__ == "DummyScalar"
    assert float(throttle.value) == pytest.approx(0.3)
    # Image stays HWC
    img = _obj_for(calls, "observation.img")
    assert type(img).__name__ == "DummyImage"
    assert img.arr.shape == (5, 6, 3)
    assert _kwargs_for(calls, "observation.img").get("static", False) is True
    # Vector logged as a single batched Scalars under one entity path
    vec = _obj_for(calls, "action.vec")
    assert type(vec).__name__ == "DummyScalar"
    np.testing.assert_allclose(np.asarray(vec.value), [9, 8, 7])
    # Blueprint sent once with the expected view layout
    assert len(blueprints) == 1
    bp = blueprints[0]
    spatial_views = _views_by_kind(bp, "Spatial2DView")
    assert {v.origin for v in spatial_views} == {"observation.img"}
    ts_views = {v.name: v for v in _views_by_kind(bp, "TimeSeriesView")}
    assert ts_views["observation"].contents == ["observation.temp"]
    assert ts_views["action"].contents == ["action.throttle", "action.vec"]
 def test_log_rerun_data_kwargs_only(mock_rerun):
    rv, calls, blueprints = mock_rerun
    rv.log_rerun_data(
        observation={"observation.temp": 10.0, "observation.gray": np.zeros((8, 8, 1), dtype=np.uint8)},
        action={"action.a": 1.0},
    )
    keys = set(_keys(calls))
    assert "observation.temp" in keys
    assert "observation.gray" in keys
    assert "action.a" in keys
    temp = _obj_for(calls, "observation.temp")
    assert type(temp).__name__ == "DummyScalar"
    assert float(temp.value) == pytest.approx(10.0)
    img = _obj_for(calls, "observation.gray")
    assert type(img).__name__ == "DummyDepthImage"  # single-channel -> DepthImage
    assert img.arr.shape == (8, 8, 1)  # remains HWC
    assert _kwargs_for(calls, "observation.gray").get("static", False) is True
    a = _obj_for(calls, "action.a")
    assert type(a).__name__ == "DummyScalar"
    assert float(a.value) == pytest.approx(1.0)
    # Blueprint sent once, with a spatial view for the image and time-series views for scalars
    assert len(blueprints) == 1
    bp = blueprints[0]
    assert {v.origin for v in _views_by_kind(bp, "Spatial2DView")} == {"observation.gray"}
    ts_views = {v.name: v for v in _views_by_kind(bp, "TimeSeriesView")}
    assert ts_views["observation"].contents == ["observation.temp"]
    assert ts_views["action"].contents == ["action.a"]
 def test_log_rerun_data_blueprint_sent_only_once(mock_rerun):
    """The blueprint is built from the first call and not resent on subsequent calls."""
    rv, calls, blueprints = mock_rerun
    rv.log_rerun_data(observation={"temp": 1.0}, action={"a": 2.0})
    assert len(blueprints) == 1
    first_blueprint = rv.log_rerun_data.blueprint
    rv.log_rerun_data(observation={"temp": 3.0}, action={"a": 4.0})
    # Still only one blueprint, and the cached one is unchanged.
    assert len(blueprints) == 1
    assert rv.log_rerun_data.blueprint is first_blueprint
@@ -14,23 +14,297 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
-"""Tests for the backend-agnostic visualization dispatch.
+import importlib
-
+import sys
-These exercise the display-mode routing/validation only; they need neither ``rerun`` nor
+from types import SimpleNamespace
 ``foxglove`` installed since the unknown-mode branch raises before touching any backend. Backend
 behavior is covered in ``test_rerun_visualization.py`` and ``test_foxglove_visualization.py``.
 """
 import numpy as np
 import pytest
-from lerobot.utils import visualization_utils as vu
+pytest.importorskip("rerun", reason="rerun-sdk is required (install lerobot[viz])")
 from lerobot.types import TransitionKey
 from lerobot.utils.constants import OBS_STATE
-def test_visualization_modes():
+@pytest.fixture
-    assert vu.VISUALIZATION_MODES == ("rerun", "foxglove")
+def mock_rerun(monkeypatch):
    """
    Provide a mock `rerun` module (and `rerun.blueprint` submodule) so tests don't
    depend on the real library. Also reload the module-under-test so it binds to
    this mock `rr`.
    """
    calls = []
    blueprints = []
    class DummyScalar:
        def __init__(self, value):
            # Scalars may be built from a single float or from a 1D array batch.
            self.value = value
    class DummyImage:
        def __init__(self, arr):
            self.arr = arr
        def compress(self, *a, **k):
            return self
    class DummyDepthImage:
        def __init__(self, arr, colormap=None):
            self.arr = arr
            self.colormap = colormap
    def dummy_log(key, obj=None, **kwargs):
        # Accept either positional `obj` or keyword `entity` and record remaining kwargs.
        if obj is None and "entity" in kwargs:
            obj = kwargs.pop("entity")
        calls.append((key, obj, kwargs))
    def dummy_send_blueprint(blueprint, *a, **k):
        blueprints.append(blueprint)
    # Mock the `rerun.blueprint` submodule used to build the layout.
    dummy_rrb = SimpleNamespace(
        Spatial2DView=lambda origin=None, name=None: SimpleNamespace(
            kind="Spatial2DView", origin=origin, name=name
        ),
        TimeSeriesView=lambda name=None, contents=None: SimpleNamespace(
            kind="TimeSeriesView", name=name, contents=contents
        ),
        Grid=lambda *views: SimpleNamespace(kind="Grid", views=list(views)),
        Blueprint=lambda root: SimpleNamespace(kind="Blueprint", root=root),
    )
    dummy_rr = SimpleNamespace(
        __name__="rerun",
        __package__="rerun",
        __spec__=SimpleNamespace(name="rerun", submodule_search_locations=None),
        Scalars=DummyScalar,
        Image=DummyImage,
        DepthImage=DummyDepthImage,
        components=SimpleNamespace(Colormap=SimpleNamespace(Viridis="viridis")),
        log=dummy_log,
        send_blueprint=dummy_send_blueprint,
        init=lambda *a, **k: None,
        spawn=lambda *a, **k: None,
        blueprint=dummy_rrb,
    )
    # Inject fake modules into sys.modules (both `rerun` and `rerun.blueprint`).
    monkeypatch.setitem(sys.modules, "rerun", dummy_rr)
    monkeypatch.setitem(sys.modules, "rerun.blueprint", dummy_rrb)
    # Now import and reload the module under test, to bind to our rerun mock
    import lerobot.utils.visualization_utils as vu
    importlib.reload(vu)
    # Expose the reloaded module, the call recorder and the captured blueprints
    yield vu, calls, blueprints
-@pytest.mark.parametrize("func", ["init_visualization", "log_visualization_data", "shutdown_visualization"])
+def _keys(calls):
-def test_dispatch_rejects_unknown_mode(func):
+    """Helper to extract just the keys logged to rr.log"""
-    with pytest.raises(ValueError, match="Unknown display_mode"):
+    return [k for (k, _obj, _kw) in calls]
-        getattr(vu, func)("bogus")
+
 def _obj_for(calls, key):
    """Find the first object logged under a given key."""
    for k, obj, _kw in calls:
        if k == key:
            return obj
    raise KeyError(f"Key {key} not found in calls: {calls}")
 def _kwargs_for(calls, key):
    for k, _obj, kw in calls:
        if k == key:
            return kw
    raise KeyError(f"Key {key} not found in calls: {calls}")
 def _views_by_kind(blueprint, kind):
    """Return the views of a given kind from the (single) blueprint's grid."""
    return [v for v in blueprint.root.views if v.kind == kind]
 def test_log_rerun_data_envtransition_scalars_and_image(mock_rerun):
    vu, calls, blueprints = mock_rerun
    # Build EnvTransition dict
    obs = {
        f"{OBS_STATE}.temperature": np.float32(25.0),
        # CHW image should be converted to HWC for rr.Image
        "observation.camera": np.zeros((3, 10, 20), dtype=np.uint8),
    }
    act = {
        "action.throttle": 0.7,
        # 1D array should be logged as a single Scalars batch under one entity path
        "action.vector": np.array([1.0, 2.0], dtype=np.float32),
    }
    transition = {
        TransitionKey.OBSERVATION: obs,
        TransitionKey.ACTION: act,
    }
    # Extract observation and action data from transition like in the real call sites
    obs_data = transition.get(TransitionKey.OBSERVATION, {})
    action_data = transition.get(TransitionKey.ACTION, {})
    vu.log_rerun_data(observation=obs_data, action=action_data)
    # We expect:
    # - observation.state.temperature -> Scalars
    # - observation.camera -> Image (HWC) with static=True
    # - action.throttle -> Scalars
    # - action.vector -> single Scalars batch (no per-element suffix)
    expected_keys = {
        f"{OBS_STATE}.temperature",
        "observation.camera",
        "action.throttle",
        "action.vector",
    }
    assert set(_keys(calls)) == expected_keys
    # Check scalar types and values
    temp_obj = _obj_for(calls, f"{OBS_STATE}.temperature")
    assert type(temp_obj).__name__ == "DummyScalar"
    assert float(temp_obj.value) == pytest.approx(25.0)
    throttle_obj = _obj_for(calls, "action.throttle")
    assert type(throttle_obj).__name__ == "DummyScalar"
    assert float(throttle_obj.value) == pytest.approx(0.7)
    # 1D vector logged as a single batched Scalars under one entity path
    vec = _obj_for(calls, "action.vector")
    assert type(vec).__name__ == "DummyScalar"
    np.testing.assert_allclose(np.asarray(vec.value), [1.0, 2.0])
    # Check image handling: CHW -> HWC
    img_obj = _obj_for(calls, "observation.camera")
    assert type(img_obj).__name__ == "DummyImage"
    assert img_obj.arr.shape == (10, 20, 3)  # transposed
    assert _kwargs_for(calls, "observation.camera").get("static", False) is True  # static=True for images
    # A blueprint should have been built and sent exactly once, and cached on the function.
    assert len(blueprints) == 1
    assert vu.log_rerun_data.blueprint is blueprints[0]
    bp = blueprints[0]
    # One spatial view per image path
    spatial_views = _views_by_kind(bp, "Spatial2DView")
    assert {v.origin for v in spatial_views} == {"observation.camera"}
    # One time-series view each for observation and action scalars
    ts_views = {v.name: v for v in _views_by_kind(bp, "TimeSeriesView")}
    assert set(ts_views) == {"observation", "action"}
    assert ts_views["observation"].contents == [f"{OBS_STATE}.temperature"]
    assert ts_views["action"].contents == ["action.throttle", "action.vector"]
 def test_log_rerun_data_plain_list_ordering_and_prefixes(mock_rerun):
    vu, calls, blueprints = mock_rerun
    # First dict without prefixes treated as observation
    # Second dict without prefixes treated as action
    obs_plain = {
        "temp": 1.5,
        # Already HWC image => should stay as-is
        "img": np.zeros((5, 6, 3), dtype=np.uint8),
        "none": None,  # should be skipped
    }
    act_plain = {
        "throttle": 0.3,
        "vec": np.array([9, 8, 7], dtype=np.float32),
    }
    # Extract observation and action data from list like the old function logic did
    # First dict was treated as observation, second as action
    vu.log_rerun_data(observation=obs_plain, action=act_plain)
    # Expected keys with auto-prefixes. The 1D vector is a single batched Scalars.
    expected = {
        "observation.temp",
        "observation.img",
        "action.throttle",
        "action.vec",
    }
    logged = set(_keys(calls))
    assert logged == expected
    # Scalars
    t = _obj_for(calls, "observation.temp")
    assert type(t).__name__ == "DummyScalar"
    assert float(t.value) == pytest.approx(1.5)
    throttle = _obj_for(calls, "action.throttle")
    assert type(throttle).__name__ == "DummyScalar"
    assert float(throttle.value) == pytest.approx(0.3)
    # Image stays HWC
    img = _obj_for(calls, "observation.img")
    assert type(img).__name__ == "DummyImage"
    assert img.arr.shape == (5, 6, 3)
    assert _kwargs_for(calls, "observation.img").get("static", False) is True
    # Vector logged as a single batched Scalars under one entity path
    vec = _obj_for(calls, "action.vec")
    assert type(vec).__name__ == "DummyScalar"
    np.testing.assert_allclose(np.asarray(vec.value), [9, 8, 7])
    # Blueprint sent once with the expected view layout
    assert len(blueprints) == 1
    bp = blueprints[0]
    spatial_views = _views_by_kind(bp, "Spatial2DView")
    assert {v.origin for v in spatial_views} == {"observation.img"}
    ts_views = {v.name: v for v in _views_by_kind(bp, "TimeSeriesView")}
    assert ts_views["observation"].contents == ["observation.temp"]
    assert ts_views["action"].contents == ["action.throttle", "action.vec"]
 def test_log_rerun_data_kwargs_only(mock_rerun):
    vu, calls, blueprints = mock_rerun
    vu.log_rerun_data(
        observation={"observation.temp": 10.0, "observation.gray": np.zeros((8, 8, 1), dtype=np.uint8)},
        action={"action.a": 1.0},
    )
    keys = set(_keys(calls))
    assert "observation.temp" in keys
    assert "observation.gray" in keys
    assert "action.a" in keys
    temp = _obj_for(calls, "observation.temp")
    assert type(temp).__name__ == "DummyScalar"
    assert float(temp.value) == pytest.approx(10.0)
    img = _obj_for(calls, "observation.gray")
    assert type(img).__name__ == "DummyDepthImage"  # single-channel -> DepthImage
    assert img.arr.shape == (8, 8, 1)  # remains HWC
    assert _kwargs_for(calls, "observation.gray").get("static", False) is True
    a = _obj_for(calls, "action.a")
    assert type(a).__name__ == "DummyScalar"
    assert float(a.value) == pytest.approx(1.0)
    # Blueprint sent once, with a spatial view for the image and time-series views for scalars
    assert len(blueprints) == 1
    bp = blueprints[0]
    assert {v.origin for v in _views_by_kind(bp, "Spatial2DView")} == {"observation.gray"}
    ts_views = {v.name: v for v in _views_by_kind(bp, "TimeSeriesView")}
    assert ts_views["observation"].contents == ["observation.temp"]
    assert ts_views["action"].contents == ["action.a"]
 def test_log_rerun_data_blueprint_sent_only_once(mock_rerun):
    """The blueprint is built from the first call and not resent on subsequent calls."""
    vu, calls, blueprints = mock_rerun
    vu.log_rerun_data(observation={"temp": 1.0}, action={"a": 2.0})
    assert len(blueprints) == 1
    first_blueprint = vu.log_rerun_data.blueprint
    vu.log_rerun_data(observation={"temp": 3.0}, action={"a": 4.0})
    # Still only one blueprint, and the cached one is unchanged.
    assert len(blueprints) == 1
    assert vu.log_rerun_data.blueprint is first_blueprint
Author	SHA1	Message	Date
CarolinePascal	f8728bde84	docs(depth stats): updating docs	2026-07-01 18:14:01 +02:00
CarolinePascal	d3fd459f81	test(depth stats): updating tests	2026-07-01 18:14:01 +02:00
CarolinePascal	ed29db6d22	feat(depth stats): enforcing all depth stats to be in millimeters (default unit) for consistency	2026-07-01 18:14:01 +02:00
		`@@ -1 +0,0 @@`
			`../../../../docs/source/policy_evo1_README.md`