Merge branch 'main' into fix/zero-shaped-features

* docs(prettier): making video encoding parameters docs prettier

* chore(format): formatting code

* chore(contrast): removing poor contrast elements

.dockerignore

@@ -22,6 +22,10 @@ outputs
 rl
 media
 
+# Local virtualenvs (the image provides its own)
+.venv
+venv
+
 
 # Logging
 logs

.github/workflows/benchmark_tests.yml

@@ -167,9 +167,9 @@ jobs:
 
       # ── LIBERO TRAIN+EVAL SMOKE ──────────────────────────────────────────────
       # Train SmolVLA for 1 step (batch_size=1, dataset episode 0 only) then
-      # immediately runs eval inside the training loop (eval_freq=1, 1 episode).
+      # immediately runs eval inside the training loop (env_eval_freq=1, 1 episode).
       # Tests the full train→eval-within-training pipeline end-to-end.
-      - name: Run Libero train+eval smoke (1 step, eval_freq=1)
+      - name: Run Libero train+eval smoke (1 step, env_eval_freq=1)
         if: env.HF_USER_TOKEN != ''
         run: |
           docker run --name libero-train-smoke --gpus all \
@@ -196,7 +196,7 @@ jobs:
                 --output_dir=/tmp/train-smoke \
                 --steps=1 \
                 --batch_size=1 \
-                --eval_freq=1 \
+                --env_eval_freq=1 \
                 --eval.n_episodes=1 \
                 --eval.batch_size=1 \
                 --eval.use_async_envs=false \

AGENT_GUIDE.md

@@ -138,7 +138,7 @@ lerobot-replay --robot.type=so101_follower --robot.port=<FOLLOWER_PORT> --robot.
   --dataset.repo_id=${HF_USER}/my_task --dataset.episode=0
 ```
 
-**4.9 Train** (default: ACT — fastest, lowest memory). Apple silicon: `--policy.device=mps`. See §6/§7 for policy and duration.
+**4.9 Train** (default: ACT — fastest, lowest memory). Apple silicon: `--policy.device=mps`. No local GPU? Add `--job.target=<flavor>` (e.g. `a10g-small`, list them with `hf jobs hardware`) to run on Hugging Face Jobs instead. See §6/§7 for policy and duration.
 
 ```bash
 lerobot-train \

Makefile

@@ -58,7 +58,7 @@ test-act-ete-train:
 		--dataset.episodes="[0]" \
 		--batch_size=2 \
 		--steps=4 \
-		--eval_freq=2 \
+		--env_eval_freq=2 \
 		--eval.n_episodes=1 \
 		--eval.batch_size=1 \
 		--save_freq=2 \
@@ -96,7 +96,7 @@ test-diffusion-ete-train:
 		--dataset.episodes="[0]" \
 		--batch_size=2 \
 		--steps=2 \
-		--eval_freq=2 \
+		--env_eval_freq=2 \
 		--eval.n_episodes=1 \
 		--eval.batch_size=1 \
 		--save_checkpoint=true \
@@ -126,7 +126,7 @@ test-tdmpc-ete-train:
 		--dataset.episodes="[0]" \
 		--batch_size=2 \
 		--steps=2 \
-		--eval_freq=2 \
+		--env_eval_freq=2 \
 		--eval.n_episodes=1 \
 		--eval.batch_size=1 \
 		--save_checkpoint=true \
@@ -161,7 +161,7 @@ test-smolvla-ete-train:
 		--dataset.episodes="[0]" \
 		--batch_size=2 \
 		--steps=4 \
-		--eval_freq=2 \
+		--env_eval_freq=2 \
 		--eval.n_episodes=1 \
 		--eval.batch_size=1 \
 		--save_freq=2 \

README.md

@@ -87,7 +87,7 @@ Learn more about it in the [LeRobotDataset Documentation](https://huggingface.co
 
 ## SoTA Models
 
-LeRobot implements state-of-the-art policies in pure PyTorch, covering Imitation Learning, Reinforcement Learning, and Vision-Language-Action (VLA) models, with more coming soon. It also provides you with the tools to instrument and inspect your training process.
+LeRobot implements state-of-the-art policies in pure PyTorch, covering Imitation Learning, Reinforcement Learning, Vision-Language-Action (VLA) models, World Models, and Reward Models, with more coming soon. It also provides you with the tools to instrument and inspect your training process.
 
 <p align="center">
   <img alt="Gr00t Architecture" src="./media/readme/VLA_architecture.jpg" width="640px">
@@ -97,17 +97,17 @@ Training a policy is as simple as running a script configuration:
 
 ```bash
 lerobot-train \
-  --policy=act \
+  --policy.type=act \
   --dataset.repo_id=lerobot/aloha_mobile_cabinet
 ```
 
-| Category                   | Models                                                                                                                                                                                                                                                                                                                                                     |
-| -------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| **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 N1.5](./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)                                                                                                                                                                                                                                         |
+| Category                   | Models                                                                                                                                                                                                                                                                                                                                                                                     |
+| -------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
+| **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), [EVO1](./docs/source/evo1.mdx) |
+| **World Models**           | [VLA-JEPA](./docs/source/vla_jepa.mdx), [LingBot-VA](./docs/source/lingbot_va.mdx), [FastWAM](./docs/source/fastwam.mdx)                                                                                                                                                                                                                                                                   |
+| **Reward Models**          | [SARM](./docs/source/sarm.mdx), [TOPReward](./docs/source/topreward.mdx), [Robometer](./docs/source/robometer.mdx)                                                                                                                                                                                                                                                                         |
 
 Similarly to the hardware, you can easily implement your own policy & leverage LeRobot's data collection, training, and visualization tools, and share your model to the HF Hub
 
@@ -136,6 +136,7 @@ Learn how to implement your own simulation environment or benchmark and distribu
 - **[X](https://x.com/LeRobotHF):** Follow us on X to stay up-to-date with the latest developments.
 - **[Robot Learning Tutorial](https://huggingface.co/spaces/lerobot/robot-learning-tutorial):** A free, hands-on course to learn robot learning using LeRobot.
 - **[T-Shirt Folding Experiment](https://huggingface.co/spaces/lerobot/robot-folding):** An end-to-end demonstration of folding t-shirts with LeRobot.
+- **[LeLab](https://github.com/huggingface/leLab):** A web interface for LeRobot — teleoperate, calibrate, record datasets, replay, and train your SO arm from the browser, no CLI required.
 
 ## Citation
 

docs/source/_toctree.yml

@@ -69,8 +69,14 @@
     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 N1.5
+    title: NVIDIA GR00T
   - local: xvla
     title: X-VLA
   - local: multi_task_dit
@@ -163,6 +169,8 @@
 - sections:
   - local: phone_teleop
     title: Phone
+  - local: isaac_teleop
+    title: Isaac Teleop
   title: "Teleoperators"
 - sections:
   - local: cameras

docs/source/bring_your_own_policies.mdx

@@ -295,11 +295,12 @@ The file names are load-bearing: the factory does lazy imports by name, and the
 
 ### Wiring
 
-Three places need to know about your policy. All by name.
+Four places need to know about your policy. All by name.
 
 1. **`policies/__init__.py`** — re-export `MyPolicyConfig` and add it to `__all__`. **Don't** re-export the modeling class; it loads lazily through the factory (so `import lerobot` stays fast).
 2. **`factory.py:get_policy_class`** — add a branch returning `MyPolicy` from a lazy import.
 3. **`factory.py:make_policy_config`** and **`factory.py:make_pre_post_processors`** — same idea, two more branches.
+4. **`templates/lerobot_modelcard_template.md` and the root `README.md`** — the template is what `push_model_to_hub` renders into the model card of every checkpoint trained with your policy: add a one-line description of your policy in the `model_name` branches, map it in `policy_docs` so cards link to your MDX guide, and optionally add an architecture image to `diagrams`. Then add your policy to the models table in the root `README.md`, under the right category, linking to your doc page.
 
 Mirror an existing policy that's structurally similar to yours; the diff is small.
 
@@ -371,6 +372,8 @@ The general expectations are in [`CONTRIBUTING.md`](https://github.com/huggingfa
 - [ ] Optional deps live behind a `[project.optional-dependencies]` extra and the `TYPE_CHECKING + require_package` guard.
 - [ ] `tests/policies/` updated; backward-compat artifact committed & policy-specific tests.
 - [ ] `src/lerobot/policies/<name>/README.md` symlinked into `docs/source/policy_<name>_README.md`; user-facing `docs/source/<name>.mdx` written and added to `_toctree.yml`.
+- [ ] `templates/lerobot_modelcard_template.md` has a description entry and a `policy_docs` link for your policy.
+- [ ] The models table in the root `README.md` lists your policy in the right category, linking to your doc page.
 - [ ] At least one reproducible benchmark eval in the policy MDX with a published checkpoint (sim benchmark, or real-robot dataset + checkpoint).
 
 The fastest way to get a clean PR is to copy the directory of the existing policy closest to yours, rename, and replace contents method by method. Don't wait until everything is polished — open a draft PR early and iterate with us; reviewers would much rather give feedback on a half-finished branch than a fully-merged one.

docs/source/cameras.mdx

@@ -157,6 +157,14 @@ finally:
 </hfoption>
 </hfoptions>
 
+### Working with depth
+
+The Intel RealSense and Reachy 2 cameras can capture both color and depth in lockstep. Calling `read()` returns the **color** frame as `(H, W, 3)` `uint8`. Calling `read_depth()` returns the **depth map** as `(H, W, 1)` `uint16`, where each pixel value is the distance from the sensor expressed in **millimetres**. A pixel value of `0` typically means "no measurement available" (out-of-range, occluded, or low-confidence).
+
+During recording, the control loop peeks the freshest buffered frames non-blockingly via `read_latest()` (color) and `read_latest_depth()` (depth), adding the depth map as a sibling feature (e.g. `front_depth` next to `front`).
+
+For how depth streams are stored and encoded when recording a dataset, see the [Depth streams](./video_encoding_parameters#depth-streams) section of the video encoding guide.
+
 ## Use your phone's camera
 
 <hfoptions id="use phone">

docs/source/cheat-sheet.mdx

@@ -89,6 +89,36 @@ Control the data recording flow using keyboard shortcuts:
 - Press **Left Arrow (`←`)**: Delete current episode and retry.
 - Press **Escape (`ESC`)**: Stop, encode videos, and upload.
 
+### Recording depth
+
+Intel RealSense cameras (`type: intelrealsense`) record a depth stream when you set `use_depth: true`. Depth is quantized to 12-bit codes and stored as its own video.
+
+```bash
+lerobot-record \
+    ... \
+    --robot.cameras="{ head: {type: intelrealsense, serial_number_or_name: \"0123456789\", width: 640, height: 480, fps: 30, use_depth: true} }" \
+    --dataset.repo_id=${HF_USER}/so101_depth_test \
+    --dataset.single_task="put the red brick in a bowl" \
+    --dataset.depth_encoder.depth_min=0.01 \
+    --dataset.depth_encoder.depth_max=10.0 \
+    --dataset.depth_encoder.shift=0.0 \
+    --dataset.depth_encoder.use_log=true
+```
+
+### Video encoding parameters
+
+RGB and depth streams are encoded independently via the `--dataset.rgb_encoder.*` and `--dataset.depth_encoder.*` keys.
+
+```bash
+lerobot-record \
+    ... \
+    --dataset.rgb_encoder.vcodec=h264 \
+    --dataset.rgb_encoder.pix_fmt=yuv420p \
+    --dataset.rgb_encoder.crf=23 \
+    --dataset.depth_encoder.vcodec=hevc \
+    --dataset.depth_encoder.extra_options='{"x265-params": "lossless=1"}'
+```
+
 ### Training
 
 Depending on your hardware training the policy might take a few hours. That's how you train simple `ACT` policy:
@@ -120,6 +150,14 @@ lerobot-train \
     --steps=20000
 ```
 
+No local GPU? Add `--job.target=<flavor>` (e.g. `a10g-small`) to either command and `lerobot-train` runs it on [Hugging Face Jobs](https://huggingface.co/docs/hub/jobs) instead — it uploads a local-only dataset for you and pushes the trained model. List flavors with `hf jobs hardware`.
+
+To resume, point `--config_path` at a checkpoint and add `--resume=true`. It accepts a local path or a Hub repo id (the latest checkpoint is fetched), and works locally or on a job by adding `--job.target=<flavor>`:
+
+```bash
+lerobot-train --config_path=${HF_USER}/policy_test --resume=true --job.target=a10g-small
+```
+
 ### Inference
 
 Inference means running the trained policy/model on a robot. For that we use `lerobot-rollout`. You will need to provide a path to your policy. It can be a local path or a path to Hugging Face for example "lerobot/folding_latest". Your cameras configuration needs to match what was used when collecting the dataset. Duration is in seconds if unspecified, it will run forever.

docs/source/earthrover_mini_plus.mdx

@@ -194,7 +194,7 @@ lerobot-record \
     --dataset.single_task="Navigate around obstacles" \
     --dataset.streaming_encoding=true \
     --dataset.encoder_threads=2 \
-    # --dataset.camera_encoder.vcodec=auto \
+    # --dataset.rgb_encoder.vcodec=auto \
     --display_data=true
 ```
 

docs/source/envhub_isaaclab_arena.mdx

@@ -193,7 +193,7 @@ To learn more about training policies with LeRobot, please refer to the training
 
 - [SmolVLA](./smolvla)
 - [Pi0.5](./pi05)
-- [GR00T N1.5](./groot)
+- [GR00T N1.7](./groot)
 
 Sample IsaacLab Arena datasets are available on HuggingFace Hub for experimentation:
 

docs/source/evo1.mdx

@@ -0,0 +1,191 @@
+# 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.

docs/source/fastwam.mdx

@@ -0,0 +1,167 @@
+# FastWAM
+
+FastWAM is a World Action Model policy for robot control. The LeRobot integration exposes FastWAM through the standard policy API so it can be configured with `policy.type=fastwam`, trained with `lerobot-train`, and loaded through the LeRobot pretrained policy interface.
+
+## Model Overview
+
+FastWAM keeps video modeling during training, but uses direct action prediction at inference time instead of iteratively generating future observations. This LeRobot policy wraps the FastWAM action model, adapts LeRobot batches to FastWAM training samples, and provides the standard processor pipeline for normalization and action postprocessing.
+
+The implementation initializes the visual world-model components from `Wan-AI/Wan2.2-TI2V-5B` by default and predicts action chunks with shape `[batch, action_horizon, action_dim]`.
+
+### What the LeRobot Integration Covers
+
+- Standard `policy.type=fastwam` configuration through LeRobot
+- Image, state, action, and language-task batch adaptation
+- Action chunk inference through `select_action` and `predict_action_chunk`
+- Checkpoint save/load through the LeRobot policy APIs
+- Configurable LIBERO gripper action postprocessing
+
+## Installation Requirements
+
+Install LeRobot from source, then install FastWAM dependencies:
+
+```bash
+pip install -e ".[fastwam]"
+```
+
+This installs the FastWAM policy extra from `pyproject.toml`: `transformers`,
+`diffusers`, `ftfy`, and `regex`, plus LeRobot's base dependencies.
+
+For LIBERO evaluation, install the benchmark dependencies too:
+
+```bash
+pip install -e ".[fastwam,libero]"
+```
+
+This installs both extras. In addition to the FastWAM dependencies above, the
+`libero` extra installs LeRobot dataset dependencies, `hf-libero` on Linux, and
+`scipy`.
+
+FastWAM uses the Wan2.2 TI2V backbone. The default model id is:
+
+```python
+policy.model_id=Wan-AI/Wan2.2-TI2V-5B
+```
+
+## Data Requirements
+
+FastWAM expects a LeRobot dataset with:
+
+- one or more visual observations whose widths concatenate to `policy.image_size[1]`
+- `observation.state` when `policy.proprio_dim` is not `None`
+- `action`
+- a language task instruction through the dataset task field, or precomputed `context` and `context_mask` tensors
+
+The default visual setup is one image feature named `observation.images.image` with shape `(3, 224, 448)`. If the dataset uses two cameras, configure `policy.input_features` so their heights match `224` and their widths sum to `448`.
+
+## Usage
+
+Create a new FastWAM policy with:
+
+```bash
+lerobot-train \
+  --dataset.repo_id=your-org/your-dataset \
+  --policy.type=fastwam \
+  --policy.action_dim=7 \
+  --policy.proprio_dim=8 \
+  --policy.action_horizon=32 \
+  --policy.n_action_steps=10 \
+  --policy.image_size='[224,448]' \
+  --output_dir=./outputs/fastwam_training \
+  --job_name=fastwam_training \
+  --steps=300000 \
+  --batch_size=8 \
+  --policy.device=cuda
+```
+
+Evaluate an existing LeRobot-format checkpoint on LIBERO-10 with:
+
+```bash
+lerobot-eval \
+  --policy.path=ZibinDong/fastwam_libero_uncond_2cam224 \
+  --policy.device=cuda \
+  --policy.torch_dtype=float32 \
+  --policy.n_action_steps=10 \
+  --env.type=libero \
+  --env.task=libero_10 \
+  --env.observation_height=224 \
+  --env.observation_width=224 \
+  --eval.batch_size=1 \
+  --eval.n_episodes=50 \
+  --seed=0 \
+  --env.episode_length=600
+```
+
+For `libero_goal`, `libero_spatial`, and `libero_object`, use
+`--env.episode_length=300`.
+
+For real-robot rollout, use the same checkpoint path:
+
+```bash
+lerobot-rollout \
+  --robot.type=so101_follower \
+  --robot.port=/dev/ttyACM0 \
+  --policy.path=your-org/fastwam-real-robot
+```
+
+## Configuration Notes
+
+### Image Features
+
+`policy.image_size` is the size of the concatenated FastWAM image tensor as `(height, width)`. Each configured image feature must have shape `(3, height, camera_width)`, and all camera widths must sum to the configured width.
+
+### Action Chunking
+
+`policy.action_horizon` controls the number of future actions supervised during training and predicted during inference. `policy.n_action_steps` controls how many actions are consumed before the policy predicts a fresh chunk. `policy.n_action_steps` must be less than or equal to `policy.action_horizon`.
+
+### Wan Components
+
+FastWAM loads the Wan VAE, video DiT, text encoder, and tokenizer from the configured Wan model directory or Hugging Face Hub model id. LeRobot-format FastWAM checkpoints saved by `save_pretrained` also copy the local Wan component files needed by `from_pretrained`.
+
+### Attention Backend
+
+FastWAM's DiT uses PyTorch's `scaled_dot_product_attention` (SDPA) for all attention. It does **not** use FlashAttention: its Mixture-of-Transformers (MoT) routing needs arbitrary boolean `[query, key]` attention masks, which the FlashAttention varlen API cannot express. Installing the `flash-attn` package therefore has no effect on the FastWAM path. (Note that SDPA itself may still select PyTorch's own flash / memory-efficient / math kernel internally — this is unrelated to the `flash-attn` package.)
+
+### LIBERO Action Toggle
+
+FastWAM LIBERO checkpoints use `policy.toggle_action_dimensions=[-1]` by
+default to match the gripper action convention used by the original FastWAM
+evaluation pipeline:
+
+```bash
+--policy.toggle_action_dimensions='[-1]'
+```
+
+## Results
+
+Evaluated on LIBERO with [`ZibinDong/fastwam_libero_uncond_2cam224`](https://huggingface.co/ZibinDong/fastwam_libero_uncond_2cam224):
+
+| Suite          | Success rate | n_episodes |
+| -------------- | -----------: | ---------: |
+| libero_spatial |        97.6% |        500 |
+| libero_object  |        99.0% |        500 |
+| libero_goal    |        95.0% |        500 |
+| libero_10      |        94.0% |        500 |
+| **average**    |    **96.4%** |       2000 |
+
+Reproduce: `lerobot-eval --policy.path=ZibinDong/fastwam_libero_uncond_2cam224 --policy.device=cuda --policy.torch_dtype=float32 --policy.n_action_steps=10 --env.type=libero --env.task=libero_spatial --env.observation_height=256 --env.observation_width=256 --eval.batch_size=1 --eval.n_episodes=50 --seed=0 --env.episode_length=300` (1x H20 140 GB).
+
+## References
+
+- [Fast-WAM paper](https://arxiv.org/abs/2603.16666)
+- [Fast-WAM project page](https://yuantianyuan01.github.io/FastWAM/)
+- [Fast-WAM code](https://github.com/yuantianyuan01/FastWAM)
+- [Released upstream checkpoints](https://huggingface.co/yuanty/fastwam)
+- [Wan2.2 TI2V 5B](https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B)
+
+## Citation
+
+```bibtex
+@article{yuan2026fastwam,
+  title = {Fast-WAM: Do World Action Models Need Test-time Future Imagination?},
+  author = {Tianyuan Yuan and Zibin Dong and Yicheng Liu and Hang Zhao},
+  journal = {arXiv preprint arXiv:2603.16666},
+  year = {2026},
+  url = {https://arxiv.org/abs/2603.16666}
+}
+```

docs/source/groot.mdx

@@ -1,16 +1,19 @@
-# GR00T N1.5 Policy
+# GR00T Policy
 
-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.
+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.
 
-This document outlines the specifics of its integration and usage within the LeRobot framework.
+LeRobot integrates GR00T N1.7 through the `groot` policy type.
+
+> [!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
 
-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.
+GR00T N1.7 uses a Cosmos-Reason2/Qwen3-VL backbone and provides checkpoints for SimplerEnv, DROID, and LIBERO.
 
-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.
+Developers and researchers can post-train GR00T with their own real or synthetic data to adapt it for specific humanoid robots or tasks.
 
-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.
+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.
 
 <img
   src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/lerobot-groot-paper1%20(1).png"
@@ -28,33 +31,24 @@ This approach allows the model to be highly adaptable through post-training for
 
 ## Installation Requirements
 
-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:
+GR00T is intended for NVIDIA GPU-accelerated systems. Install LeRobot with the GR00T extra:
 
 ```bash
-# 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')"
+pip install "lerobot[groot]"
 ```
 
-3. Install LeRobot by running:
+For a source checkout:
 
 ```bash
-pip install lerobot[groot]
+pip install -e ".[groot]"
 ```
 
 ## Usage
 
-To use GR00T in your LeRobot configuration, specify the policy type as:
+To use GR00T N1.7:
 
-```python
-policy.type=groot
+```bash
+--policy.type=groot
 ```
 
 ## Training
@@ -63,72 +57,171 @@ policy.type=groot
 
 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
-# Using a multi-GPU setup
-accelerate launch \
-  --multi_gpu \
-  --num_processes=$NUM_GPUS \
-  $(which lerobot-train) \
-  --output_dir=$OUTPUT_DIR \
-  --save_checkpoint=true \
-  --batch_size=$BATCH_SIZE \
-  --steps=$NUM_STEPS \
-  --save_freq=$SAVE_FREQ \
-  --log_freq=$LOG_FREQ \
-  --policy.push_to_hub=true \
+# install extra deps for training
+pip install "lerobot[training]"
+
+hf auth login
+wandb login
+
+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 \
-  --policy.tune_diffusion_model=false \
-  --dataset.repo_id=$DATASET_ID \
+  --seed=42 \
+  --batch_size=64 \
+  --steps=20000 \
+  --save_checkpoint=true \
+  --save_freq=5000 \
+  --use_policy_training_preset=true \
+  --env_eval_freq=0 \
+  --eval_steps=0 \
+  --log_freq=10 \
+  --output_dir=$OUTPUT_DIR \
+  --job_name=$DATASET \
   --wandb.enable=true \
-  --wandb.disable_artifact=true \
-  --job_name=$JOB_NAME
+  --wandb.disable_artifact=true
+
 ```
 
 ## Performance Results
 
-### Libero Benchmark Results
+### LIBERO Benchmark Results
 
 > [!NOTE]
-> Follow our instructions for Libero usage: [Libero](./libero)
+> Follow the [LIBERO](./libero) setup instructions before running `lerobot-eval`.
 
-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.
+GR00T N1.7 has demonstrated strong performance on the LIBERO benchmark suite. To reproduce LeRobot results, follow the instructions in the [LIBERO](./libero) section.
 
-| 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%           |
+### Train on LIBERO
 
-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.
+Example training command for a LIBERO suite (here `libero_spatial`):
+
+```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=$JOB_NAME
+```
+
+This will follow the recipe found [here](https://github.com/NVIDIA/Isaac-GR00T/blob/main/examples/LIBERO/README.md).
+
+### GR00T N1.7 LIBERO Results
+
+Preliminary LeRobot integration results (GR00T-LeRobot, `eval.n_episodes >= 50` per suite):
+
+| Suite            | Success rate | Checkpoint                                                                                                    |
+| ---------------- | -----------: | ------------------------------------------------------------------------------------------------------------- |
+| LIBERO Spatial   |          91% | [nvidia/gr00t17-lerobot-libero_spatial-640](https://huggingface.co/nvidia/gr00t17-lerobot-libero_spatial-640) |
+| LIBERO Object    |          81% | [nvidia/gr00t17-lerobot-libero_object-640](https://huggingface.co/nvidia/gr00t17-lerobot-libero_object-640)   |
+| LIBERO Goal      |          97% | [nvidia/gr00t17-lerobot-libero_goal-640](https://huggingface.co/nvidia/gr00t17-lerobot-libero_goal-640)       |
+| LIBERO 10 (Long) |          84% | [nvidia/gr00t17-lerobot-libero_10-640](https://huggingface.co/nvidia/gr00t17-lerobot-libero_10-640)           |
+| **Average**      |   **88.25%** |                                                                                                               |
+
+```bash
+export MODEL_ID=your_trained_model_on_huggingface
+
+lerobot-eval \
+  --policy.type=groot \
+  --policy.base_model_path=$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
-lerobot-rollout\
-  --strategy.type=sentry \
-  --strategy.upload_every_n_episodes=5 \
-  --robot.type=bi_so_follower \
-  --robot.left_arm_port=/dev/ttyACM1 \
-  --robot.right_arm_port=/dev/ttyACM0 \
-  --robot.id=bimanual_follower \
-  --robot.cameras='{ right: {"type": "opencv", "index_or_path": 0, "width": 640, "height": 480, "fps": 30},
-    left: {"type": "opencv", "index_or_path": 2, "width": 640, "height": 480, "fps": 30},
-    top: {"type": "opencv", "index_or_path": 4, "width": 640, "height": 480, "fps": 30},
-  }' \
+# install extra deps for roullout and real hardware
+pip install "lerobot[feetech,viz]"
+
+export MODEL_ID=your_trained_model_on_huggingface
+
+# make sure that camera index matches your setup!
+# find index using `uv run lerobot-find-cameras opencv`
+WRIST_CAM='wrist: {type: opencv, index_or_path: 2, width: 640, height: 480, fps: 30, fourcc: "MJPG"}'
+FRONT_CAM='front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30, fourcc: "MJPG"}'
+export ROBOT_CAMERAS="{ $WRIST_CAM, $FRONT_CAM }"
+export ROBOT_ID=follower_robot
+export ROBOT_PORT=/dev/ttyACM0
+
+uv run lerobot-rollout \
+  --strategy.type=base \
+  --policy.path=$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 \
-  --dataset.repo_id=<user>/eval_groot-bimanual  \
-  --dataset.single_task="Grab and handover the red cube to the other arm" \
-  --dataset.streaming_encoding=true \
-  --dataset.encoder_threads=2 \
-  # --dataset.camera_encoder.vcodec=auto \
-  --policy.path=<user>/groot-bimanual \ # your trained model
-  --duration=600
+  --inference.type=rtc \
+  --inference.rtc.enabled=True \ # set to False if it causes inference instability
+  --inference.rtc.execution_horizon=8 \
+  --inference.queue_threshold=0
 ```
 
+> [!NOTE]
+> Value of `inference.queue_threshold` should not exceed 5 to ensure stable inference.
+
 ## 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**.
+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/).

docs/source/hardware_guide.mdx

@@ -82,17 +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. 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.
+[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:
 
 ```bash
-hf jobs run --flavor a10g-large huggingface/lerobot-gpu:latest \
-  bash -c "nvidia-smi && lerobot-train \
-    --policy.type=act --dataset.repo_id=<USER>/<DATASET> \
-    --policy.repo_id=<USER>/act_<task> --batch_size=8 --steps=50000"
+lerobot-train \
+  --policy.type=act --dataset.repo_id=<USER>/<DATASET> \
+  --policy.repo_id=<USER>/act_<task> \
+  --job.target=a10g-large
 ```
 
 Notes:
 
-- 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.
-- The default Job timeout is 30 minutes; pass `--timeout 4h` (or longer) for real training.
-- `--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).
+- Run `hf auth login` once before submitting, the job runs under your token.
+- `--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 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.
+- 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).

docs/source/hilserl.mdx

@@ -719,7 +719,7 @@ Example configuration for training the [reward classifier](https://huggingface.c
   "num_workers": 4,
   "steps": 5000,
   "log_freq": 10,
-  "eval_freq": 1000,
+  "env_eval_freq": 1000,
   "save_freq": 1000,
   "save_checkpoint": true,
   "seed": 2,

docs/source/hope_jr.mdx

@@ -232,7 +232,7 @@ lerobot-record \
     --dataset.private=true \
     --dataset.streaming_encoding=true \
     --dataset.encoder_threads=2 \
-    # --dataset.camera_encoder.vcodec=auto \
+    # --dataset.rgb_encoder.vcodec=auto \
     --display_data=true
 ```
 
@@ -278,6 +278,6 @@ lerobot-record \
   --dataset.num_episodes=10 \
   --dataset.streaming_encoding=true \
   --dataset.encoder_threads=2 \
-  # --dataset.camera_encoder.vcodec=auto \
+  # --dataset.rgb_encoder.vcodec=auto \
   --policy.path=outputs/train/hopejr_hand/checkpoints/last/pretrained_model
 ```

docs/source/il_robots.mdx

@@ -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_rerun, log_rerun_data, shutdown_rerun
+from lerobot.utils.visualization_utils import init_visualization, log_visualization_data, shutdown_visualization
 
 robot_config = SO101FollowerConfig(
     port="/dev/tty.usbmodem5AB90687491",
@@ -142,7 +142,7 @@ teleop_config = SO101LeaderConfig(
     id="my_leader_arm",
 )
 
-init_rerun(session_name="teleoperation")
+init_visualization("rerun", session_name="teleoperation")  # pass "foxglove" to stream to Foxglove instead
 
 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_rerun_data(observation=observation, action=action)
+    log_visualization_data("rerun", observation=observation, action=action)
 
     elapsed_time = time.perf_counter() - start_time
     sleep_time = TIME_PER_FRAME - elapsed_time
@@ -207,7 +207,7 @@ lerobot-record \
     --dataset.num_episodes=5 \
     --dataset.single_task="Grab the black cube" \
     --dataset.streaming_encoding=true \
-    # --dataset.camera_encoder.vcodec=auto \
+    # --dataset.rgb_encoder.vcodec=auto \
     --dataset.encoder_threads=2
 ```
 </hfoption>
@@ -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_rerun
+from lerobot.utils.visualization_utils import init_visualization
 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_rerun(session_name="recording")
+    init_visualization("rerun", session_name="recording")
 
     # Connect the robot and teleoperator
     robot.connect()
@@ -390,9 +390,17 @@ Set the flow of data recording using command-line arguments:
 
 Control the data recording flow using keyboard shortcuts:
 
-- Press **Right Arrow (`→`)**: Early stop the current episode or reset time and move to the next.
-- Press **Left Arrow (`←`)**: Cancel the current episode and re-record it.
-- Press **Escape (`ESC`)**: Immediately stop the session, encode videos, and upload the dataset.
+- Press **Right Arrow (`→`)** or **`n`**: Early stop the current episode or reset time and move to the next.
+- Press **Left Arrow (`←`)** or **`r`**: Cancel the current episode and re-record it.
+- Press **Escape (`ESC`)** or **`q`**: Immediately stop the session, encode videos, and upload the dataset.
+
+<Tip>
+
+These control-flow shortcuts work on **X11, Wayland, and headless/SSH** sessions. When a global keyboard backend isn't available (Wayland, a headless machine, or macOS without Accessibility permission), `lerobot-record` automatically reads the same keys from the terminal — launch it from an interactive terminal and keep it focused. You can also use the letter equivalents **`n`** (next, same as `→`), **`r`** (re-record, same as `←`) and **`q`** (quit, same as `ESC`). No `$DISPLAY` setup is required.
+
+This applies to the recording control flow only. Keyboard **teleoperation** (driving the robot with the keyboard) still needs a global key backend, so it works only on an X11 session, a Windows desktop, or macOS with Accessibility/Input Monitoring granted — not on Wayland or headless sessions.
+
+</Tip>
 
 #### Tips for gathering data
 
@@ -406,7 +414,7 @@ If you want to dive deeper into this important topic, you can check out the [blo
 
 #### Troubleshooting:
 
-- On Linux, if the left and right arrow keys and escape key don't have any effect during data recording, make sure you've set the `$DISPLAY` environment variable. See [pynput limitations](https://pynput.readthedocs.io/en/latest/limitations.html#linux).
+- On Linux, the recording control-flow keys (arrow keys, Escape) work on X11, Wayland, and headless/SSH sessions as long as `lerobot-record` runs in an interactive terminal — no `$DISPLAY` setup is needed. If the keys have no effect, make sure you are in an interactive (TTY) terminal, not a piped/non-TTY session, and that it is focused; the letter equivalents `n` / `r` / `q` also work. Keyboard _teleoperation_ (as opposed to the recording control flow) still requires a global key backend — an X11 session, a Windows desktop, or macOS with Accessibility/Input Monitoring granted — and is unavailable on Wayland or headless machines. See [pynput limitations](https://pynput.readthedocs.io/en/latest/limitations.html#linux).
 
 ## Visualize a dataset
 
@@ -506,6 +514,12 @@ lerobot-train \
   --resume=true
 ```
 
+`--config_path` also accepts a **Hub repo id**: if a run pushed its checkpoints to the Hub (with `--save_checkpoint_to_hub=true`), you can resume straight from the repo — its latest checkpoint is downloaded and training continues, restoring the optimizer, scheduler, step counter and data order:
+
+```bash
+lerobot-train --config_path=${HF_USER}/my_policy --resume=true
+```
+
 If you do not want to push your model to the hub after training use `--policy.push_to_hub=false`.
 
 Additionally you can provide extra `tags` or specify a `license` for your model or make the model repo `private` by adding this: `--policy.private=true --policy.tags=\[ppo,rl\] --policy.license=mit`
@@ -518,78 +532,48 @@ 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).
 
-To run the training use this command:
+`lerobot-train` runs locally by default. To run on a HuggingFace GPU, pass `--job.target` with a hardware flavor name:
 
-<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
+lerobot-train \
+  --dataset.repo_id=${HF_USER}/so101_test \
+  --policy.type=act \
+  --policy.repo_id=${HF_USER}/my_policy \
+  --job.target=a10g-small
 ```
-</hfoption>
-<hfoption id="API example">
 
-<!-- prettier-ignore-start -->
-```python
-from huggingface_hub import run_job, get_token
+List available flavors and pricing with `hf jobs hardware`. The run streams its logs to your terminal; press Ctrl-C to detach (the job keeps running in the cloud). Re-attach or cancel with:
 
-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}")
+```bash
+hf jobs logs <job-id>
+hf jobs cancel <job-id>
 ```
-<!-- prettier-ignore-end -->
 
-</hfoption>
-</hfoptions>
+If your dataset exists only locally (not yet on the Hub), it is automatically pushed to a **private** Hub repo so the job can download it by `repo_id` (nothing is made public). The trained model is pushed to the model repo at the end of the run. To also push every intermediate checkpoint to the Hub as it is saved (so you can monitor progress mid-run), add `--save_checkpoint_to_hub=true` — this requires a runtime image that includes this feature.
 
-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.
+Every job (and any dataset pushed by the run) is tagged `lerobot` so it's easy to find on the Hub. Add your own with `--job.tags '["my-tag"]'`.
 
-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.
+By default the job is capped at `2d` (48h) of wall-clock. Override it with an HF Jobs duration string, e.g. `--job.timeout=4h` to fail faster or `--job.timeout=7d` for a longer run.
 
-After training the model will be pushed to hub and you can use it as any other model with LeRobot.
+> **Note:** the model repo is created up front (it holds the staged training config the job runs from). If a run fails before the model is pushed, that repo is left on the Hub so you can inspect it — it is not deleted automatically, so repeated failures can leave empty repos behind. Remove one with `hf repo delete <repo-id>`.
+
+**Prerequisites:** run `hf auth login` before submitting. For Weights & Biases integration, run `wandb login` or set `WANDB_API_KEY` on your machine — the key is forwarded to the job automatically.
+
+**Resuming on a job.** Adding `--job.target` to a resume command runs the resume in the cloud — the same command works locally or remotely. The checkpoint repo is the source of truth, and new checkpoints continue the lineage in the same repo:
+
+```bash
+# resume a Hub run on a job (its checkpoints are already on the Hub)
+lerobot-train --config_path=${HF_USER}/my_policy --resume=true --job.target=a10g-small
+
+# resume a LOCAL run on a job — the checkpoint is uploaded to a private Hub repo first,
+# then the job resumes from it (a local-only dataset is uploaded the same way)
+lerobot-train \
+  --config_path=outputs/train/act_so101_test/checkpoints/last/pretrained_model/train_config.json \
+  --resume=true \
+  --job.target=a10g-small
+```
+
+Job settings come from the current command, so override `--job.target`, `--job.timeout`, etc. as needed; for the resumed run to itself be resumable later, keep `--save_checkpoint_to_hub=true`.
 
 #### Upload policy checkpoints
 
@@ -612,6 +596,8 @@ hf upload ${HF_USER}/act_so101_test${CKPT} \
 
 Use `lerobot-rollout` to deploy a trained policy on your robot. You can choose different strategies depending on your needs:
 
+The examples below load the model from `--policy.path`. To pin a specific pushed version — useful once `--save_checkpoint_to_hub=true` has committed several checkpoints — add `--policy.pretrained_revision` with a commit hash, branch, or tag. Each pushed checkpoint is tagged with its step (e.g. `--policy.pretrained_revision=010000`), so you can recover a checkpoint by step without looking up its commit sha.
+
 <hfoptions id="eval">
 <hfoption id="Base mode (no recording)">
 ```bash

docs/source/isaac_teleop.mdx

@@ -0,0 +1,397 @@
+# Isaac Teleop
+
+Control your robot with NVIDIA [Isaac Teleop](https://github.com/NVIDIA/IsaacTeleop), a
+multi-modal teleoperation framework. Isaac Teleop drives a single `TeleopSession` from a range
+of input devices — XR (VR) controllers, hand tracking, full-body tracking, Manus gloves, foot
+pedals, and more.
+
+In LeRobot, Isaac Teleop ships as a self-contained example under
+[`examples/isaac_teleop_to_so101/`](https://github.com/huggingface/lerobot/tree/main/examples/isaac_teleop_to_so101).
+Each Isaac Teleop input device is its own `Teleoperator` subclass in the example's
+`isaac_teleop` package, sharing one session lifecycle (see `IsaacTeleopTeleoperator`). The
+devices available today are the **XR controller** (`XRController`) and a back-drivable
+**SO-101 leader arm** (`SO101LeaderArm`); Manus gloves and hand/full-body tracking are the
+natural next devices. This guide focuses on the XR controller; the SO-101 leader is summarized
+under [Run the example](#step-3-run-the-example).
+
+**In this guide you'll learn:**
+
+- How an Isaac Teleop device drives a robot end‑effector (EE) target
+- How the _clutch_ (squeeze/grip on the XR controller) engages teleoperation without jerking the arm
+- How to run the SO‑101 teleoperation example and tune motion / gripper / IK
+
+## Installation
+
+The example lives in the LeRobot repository (it is not part of the `lerobot` pip package), so
+clone the repo and install from source. The canonical, always-up-to-date install and usage
+reference is the example's
+[`README.md`](https://github.com/huggingface/lerobot/tree/main/examples/isaac_teleop_to_so101/README.md);
+in short:
+
+```bash
+git clone https://github.com/huggingface/lerobot.git
+cd lerobot
+uv pip install -e ".[feetech,kinematics,dataset]" "huggingface_hub>=1.5"
+uv pip install "isaacteleop[cloudxr,retargeters-lite]~=1.3.131" "scipy>=1.14"
+```
+
+`isaacteleop` is published on public PyPI (Linux only). The `cloudxr` extra brings the CloudXR
+runtime bindings; `retargeters-lite` is the scipy-based retargeter path that resolves on both
+x86_64 and ARM (on aarch64 — e.g. a DGX Spark — the full `retargeters` extra does not resolve
+because of its `dex-retargeting`/`nlopt` pins, which is why it is not the default here). On
+x86_64 you can additionally install the full retargeter stack:
+
+```bash
+uv pip install "isaacteleop[retargeters]~=1.3.131"
+```
+
+### Set up CloudXR and connect a headset
+
+Isaac Teleop streams the headset to your machine over **NVIDIA CloudXR**, which provides the
+OpenXR runtime the session connects to. By default LeTeleop **auto-launches the CloudXR runtime
+for you** when you call `teleop_device.connect()` — you no longer have to run `python -m
+isaacteleop.cloudxr` and `source cloudxr.env` in a separate shell. All you need is a supported
+headset connected and the CloudXR firewall ports open. Follow the Isaac Teleop
+[Quick Start](https://nvidia.github.io/IsaacTeleop/main/getting_started/quick_start.html) for the
+headset-pairing and firewall details.
+
+**First run (EULA).** The very first launch must accept the NVIDIA CloudXR EULA. The auto-launch
+prompts for it **on stdin**, so on a headless machine it will hang waiting for input. Bootstrap
+the EULA once, interactively, with:
+
+```bash
+python -m isaacteleop.cloudxr --accept-eula   # one-time: accept the CloudXR EULA
+```
+
+After that, `connect()` launches the runtime non-interactively. The launch **blocks for ~30s**
+while the runtime comes up.
+
+**Configuration.** Two fields on `IsaacTeleopConfig` (shared by every device) control this:
+
+- `auto_launch_cloudxr` (default `True`) — whether `connect()` starts the runtime. Set `False`
+  when CloudXR is already running externally.
+- `cloudxr_env_file` (default `None`) — an optional CloudXR device-profile `.env` selecting the
+  headset transport (e.g. an Apple Vision Pro profile). This is launcher **input**; it is not the
+  `~/.cloudxr/run/cloudxr.env` **output** file the old manual flow told you to `source`. `None`
+  keeps the default auto-WebRTC profile — though the SO-101 example overrides it to the
+  `default.env` shipped next to `teleoperate.py` unless you pass `--teleop.cloudxr_env_file`.
+
+**Opting out.** To skip the auto-launch (CloudXR already running), either set
+`auto_launch_cloudxr=False` or export:
+
+```bash
+export LEROBOT_CLOUDXR_SKIP_AUTOLAUNCH=1
+```
+
+The **env var takes precedence over the config field**: if `LEROBOT_CLOUDXR_SKIP_AUTOLAUNCH=1` is
+set, the auto-launch is skipped even when `auto_launch_cloudxr=True`. This variable is
+**independent** of Isaac Lab's `ISAACLAB_CXR_SKIP_AUTOLAUNCH` — setting one does not affect the
+other.
+
+**One teleoperator per process.** The CloudXR runtime configures the environment process-wide (a
+singleton), so run a single Isaac Teleop teleoperator per process.
+
+**Shutting down.** Always call `teleop_device.disconnect()` on exit — including on Ctrl-C. Wrap
+your teleoperation loop in `try/finally` and call `disconnect()` in the `finally`. This tears down
+the OpenXR session **before** the CloudXR runtime, which is the required order; the launcher's
+`atexit` hook only reaps the runtime and does not run the session's `__exit__`, so without an
+explicit `disconnect()` an interrupted run shuts down in the wrong order.
+
+```python
+teleop_device.connect()
+try:
+    while True:
+        action = teleop_device.get_action()
+        # ... drive the robot ...
+finally:
+    teleop_device.disconnect()
+```
+
+See [System Requirements](https://nvidia.github.io/IsaacTeleop/main/references/requirements.html)
+for supported OS / GPU / CloudXR versions and headsets.
+
+## How it works
+
+The XR controller is one Isaac Teleop **input** device. `XRController` is a deliberately thin
+reader: it exposes the **raw** controller grip pose — already statically rebased into the robot
+base frame — plus the squeeze and trigger analog values. It has **no** retargeters and **no**
+clutch logic of its own. The clutch (engage latch + delta rebasing onto the EE) and the gripper
+mapping live downstream in the example loop, which then feeds LeRobot's existing closed‑loop
+Cartesian IK pipeline — the same one the phone teleoperator uses. The device‑specific pieces are
+`XRController`, the loop's `Clutch`, and `MapXRControllerActionToRobotAction`; everything downstream
+(`EEBoundsAndSafety`, `InverseKinematicsEEToJoints`) is shared, and a future device (e.g. Manus
+gloves) would swap in its own `teleop_<device>.py` + processor while reusing the rest.
+
+`XRController._build_pipeline` wires Isaac Teleop's `ControllersSource` — statically rebased into
+the robot base frame by the native `ControllerTransform` (`base_T_anchor`) — and exposes the
+transformed controller stream verbatim. `get_action()` reads the grip pose, squeeze, and trigger
+straight off it; the session is always stepped `RUNNING` (there is no clutch retargeter to gate).
+
+The `Clutch` class (in `examples/isaac_teleop_to_so101/isaac_teleop/clutch.py`, driven by the
+loop in `common.py`) mirrors Isaac Teleop's `SO101ClutchRetargeter`, but lives in-loop so the
+device can stay a thin reader:
+
+- It latches its engage origin on the squeeze **engage edge** (the frame the squeeze first crosses
+  `clutch_threshold`) and rebases both position and orientation around it, so engaging does not
+  teleport the arm. `Clutch.rebase` returns the absolute base-frame target as a `(pos, quat)`
+  pair, which the loop concatenates into the 7D `ee_pose` fed to the processor.
+- The analog trigger becomes a gripper `closedness` in `[0, 1]` (0 = open, 1 = closed),
+  proportional to the trigger pull, which `MapXRControllerActionToRobotAction` maps to a jaw target.
+
+See the Isaac Teleop
+[Retargeting interface](https://nvidia.github.io/IsaacTeleop/main/references/retargeting/index.html)
+and [architecture overview](https://nvidia.github.io/IsaacTeleop/main/overview/architecture.html)
+for how source nodes and retargeters compose.
+
+```text
+  VR controller (OpenXR)
+        │
+        ▼
+  XRController.get_action()          ── raw base-frame grip_pos / grip_quat + squeeze + trigger
+        │                                (TeleopSession always stepped RUNNING; clutch lives downstream)
+        ▼
+  Clutch.rebase(grip_pos, grip_quat) ── engage-relative delta applied to the EE home (pos + orient)
+        │  ee_pose (7) / closedness       → absolute ee_pose; closedness = trigger
+        ▼
+  MapXRControllerActionToRobotAction ── absolute ee.x/y/z; ee.w* = orientation rotvec target;
+        │  ee.x/y/z / ee.w* / ee.gripper_pos   ee.gripper_pos = (1 - closedness) * 100
+        ▼
+  EEBoundsAndSafety                ── workspace clip + per-frame step clamp (clamp+warn)
+        │
+        ▼
+  InverseKinematicsEEToJoints      ── closed-loop Placo IK; position + soft-orientation
+        │  (orientation_weight=0.01)   (passes ee.gripper_pos → gripper.pos)
+        ▼
+  SO-101 follower joint targets
+```
+
+### The clutch: owned by the example loop
+
+Unlike the phone pipeline (which splits the clutch across `MapPhoneActionToRobotAction` and
+`EEReferenceAndDelta`), the XR clutch lives entirely in the example loop's `Clutch` class. It emits
+an **absolute** EE pose, so there is no `EEReferenceAndDelta` stage and no delta accumulation in the
+processor — `MapXRControllerActionToRobotAction` is a pure, stateless per‑frame mapping.
+
+The clutch latches its engage origin on the squeeze **engage edge** (the moment the squeeze crosses
+`clutch_threshold`) and drives the EE from the motion _relative_ to that origin, so the arm does not
+teleport on engage. On **every** engage — startup and mid‑task re‑clutch alike — the home
+_position_ is latched from forward kinematics on the arm's **measured joints**, so the home equals
+where the arm physically is even if it moved while disengaged, and the engage is jump‑free. The
+home _orientation_ keeps the last commanded rotation: the 5‑DOF arm tracks orientation only
+softly, so latching the measured wrist orientation would inject its tracking offset into the
+command on every re‑clutch.
+
+## Controls
+
+- **Squeeze / grip** — the **clutch** (deadman). Hold it past `clutch_threshold` to engage
+  teleoperation; release to pause. Each engage re‑captures the origin, so you can reposition
+  your hand while paused and re‑engage without the arm jumping (index/clutch style).
+- **Trigger** — the **gripper**, controlled **analog**. The jaw tracks the trigger
+  proportionally — a half‑pressed trigger leaves the jaw half‑closed — via a closedness in
+  `[0, 1]` (0 = open, 1 = closed) that maps to an absolute gripper joint target.
+- **Controller orientation** — the **wrist**. The clutch rebases the controller orientation
+  (engage‑relative, base‑frame) into a soft IK orientation target the wrist tracks alongside
+  position. On the 5‑DOF SO‑101 the wrist follows the hand only partially by design — see
+  `orientation_weight` below.
+
+## Get started
+
+### Step 1: Create the teleoperator
+
+```python
+# Run from the repo root so the `examples` package is importable.
+from examples.isaac_teleop_to_so101.isaac_teleop import XRController, XRControllerConfig
+
+teleop_config = XRControllerConfig(
+    hand_side="right",      # "left" or "right" controller
+    clutch_threshold=0.5,   # squeeze value above which the clutch engages
+)
+teleop_device = XRController(teleop_config)
+```
+
+`XRController.get_action()` returns the **raw** base‑frame controller pose, not a clutch‑rebased
+target: `grip_pos` (3,) `[x, y, z]` [m] and `grip_quat` (4,) `[qx, qy, qz, qw]` in the robot base
+frame, plus scalar `squeeze` and `trigger` analog values in `[0, 1]`. The example loop's `Clutch`
+turns these into the absolute `ee_pose`, and the squeeze is thresholded by the loop against
+`clutch_threshold` to engage.
+
+### Step 2: Connect
+
+Calling `teleop_device.connect()` first auto-launches the CloudXR runtime (unless you opted out —
+see [Set up CloudXR and connect a headset](#set-up-cloudxr-and-connect-a-headset); this blocks for
+~30s and on the first run prompts for the EULA on stdin), then starts the Isaac Teleop
+[`TeleopSession`](https://nvidia.github.io/IsaacTeleop/main/getting_started/teleop_session.html)
+(opens the OpenXR session and discovers the controllers). XR controllers are self‑calibrating, so
+there is no manual calibration step — the clutch handles re‑centering each time you engage. Pair
+`connect()` with a `try/finally` that calls `disconnect()` so the session tears down before the
+runtime on exit/Ctrl-C.
+
+### Step 3: Run the example
+
+The example assumes you configured your robot (SO‑101 follower) and set the correct serial port.
+
+The **robot URDF and its meshes are fetched automatically** on first run: the XR device downloads
+the SO-101 URDF from the
+[`lerobot/robot-urdfs` Hugging Face bucket](https://huggingface.co/buckets/lerobot/robot-urdfs/tree/so101)
+into the LeRobot cache (`HF_LEROBOT_HOME/robot-urdfs/so101/`) and reuses it after, so there is no
+separate download step :
+
+```bash
+python -m examples.isaac_teleop_to_so101.teleoperate --robot.type=so101_follower --robot.port=/dev/ttyACM0 \
+    --robot.id=so101_follower_arm --teleop.type=xr_controller
+```
+
+The CLI is `lerobot-teleoperate`-style (draccus): `--robot.*` configures the SO-101 follower and
+`--teleop.type` selects the Isaac input device (`xr_controller` | `so101_leader`), with
+`--teleop.*` its device knobs. `--teleop.type=xr_controller` runs the XR-controller path described
+above. The startup safety contract: by default it slews all joints to a default reset pose over
+`--reset_duration` seconds (`--reset_to_origin=false` keeps the arm where it is), then seeds the
+clutch home from the arm's measured pose so the first engage is jump-free; the follower is
+commanded only while the clutch is engaged.
+
+**Customizing the reset pose.** The reset pose ships as a built-in default (a comfortable mid-range
+pose) and works out of the box — you do **not** need to record anything. To tailor it to your setup,
+back-drive the arm to the pose you want and run
+`python -m examples.isaac_teleop_to_so101.override_reset_pose --id <robot.id>`; it writes the
+current joints to a per-arm file in the LeRobot cache
+(`HF_LEROBOT_HOME/reset_poses/<robot.name>/<robot.id>.json`, keyed like calibration), which then takes
+priority over the built-in default on the next run. Because it lives in the user-local cache (not
+the repo), your override stays on your machine, and both `teleoperate` and `record` honor it
+when launched with the same `--robot.id`.
+
+The other device, `--teleop.type=so101_leader`, mirrors the follower 1:1 from a back-drivable
+SO-101 _leader arm_ whose joints are streamed by Isaac Teleop's native `so101_leader` plugin (no
+clutch, no IK — the leader and follower share the SO-101 kinematics).
+
+The `so101_leader_plugin` binary is a C++ plugin that is **not** part of the `isaacteleop` pip
+package — you build it from the Isaac Teleop source tree. Follow
+[Build Isaac Teleop from source](https://nvidia.github.io/IsaacTeleop/main/getting_started/build_from_source/index.html)
+(in short, from your Isaac Teleop checkout: `cmake -B build && cmake --build build --parallel &&
+cmake --install build`); the build installs the plugins under `<IsaacTeleop>/install/plugins/`, so
+the binary lands at `install/plugins/so101_leader/so101_leader_plugin` — the `--launch_plugin` path
+below. See the plugin's own `README.md` (next to the binary) for its serial/calibration details.
+
+Point `--teleop.port` at the physical leader's serial port and `--launch_plugin` at that plugin
+binary to have the script spawn it after CloudXR is up:
+
+```bash
+python -m examples.isaac_teleop_to_so101.teleoperate --robot.type=so101_follower --robot.port=/dev/ttyACM0 \
+    --robot.id=so101_follower_arm --teleop.type=so101_leader \
+    --teleop.port=/dev/ttyACM1 --teleop.id=so101_leader_arm \
+    --launch_plugin=/code/Teleop/install/plugins/so101_leader/so101_leader_plugin
+```
+
+(Note `so101_leader` here is the _Isaac_ leader, resolved against the Isaac Teleop device
+registry, distinct from `lerobot-teleoperate`'s serial `so101_leader`.) When a `--teleop.port` is
+set, the plugin's tick→radian calibration is inferred from `--teleop.id` and passed to the plugin
+as its third positional arg — the LeRobot-format JSON at
+`HF_LEROBOT_CALIBRATION/teleoperators/so_leader/<id>.json`, the same file the serial SO-101 leader
+uses (`lerobot-calibrate --teleop.type=so101_leader --teleop.id=<id>`). If it is missing the script
+warns and the plugin uses built-in defaults. Run `python -m examples.isaac_teleop_to_so101.teleoperate --help` for all flags. Its
+startup safety contract: by default the follower is
+slewed to the leader's first reading over `--align_duration` seconds (`--align=false` to skip) so
+the arm does not snap when the mirror begins, and while the leader stream is stale the follower is
+held at its measured pose.
+
+The URDF fetch uses `huggingface_hub` (already a LeRobot dependency) against the public
+`lerobot/robot-urdfs` bucket, so it needs no login. It is cached under
+`HF_LEROBOT_HOME/robot-urdfs/so101/`; delete that folder to force a re‑download.
+
+Then, in your headset: squeeze and hold the grip to engage, move the controller to drive the
+arm, twist/tilt it to orient the wrist, and press the trigger to close the gripper
+(proportionally — release to open).
+
+To record a dataset (not just teleoperate), use `record.py` in the same folder. It dispatches on
+`--teleop.type` (`xr_controller` | `so101_leader`) exactly like `teleoperate.py`, so either device
+can drive the follower, and it saves the commanded joints to a LeRobot dataset (`lerobot-record`-style
+`--dataset.*` flags). See its module docstring for the full CLI and the keyboard recording shortcuts.
+
+## Important pipeline steps and options
+
+The clutch already produces an absolute base‑frame pose, so the processor side is a thin
+**absolute‑pose** path — there is no frame remap, no delta accumulation, and no
+`EEReferenceAndDelta` stage.
+
+- `MapXRControllerActionToRobotAction` is a stateless per‑frame mapping from the device output to
+  the IK input contract. It writes the absolute base‑frame position, encodes the absolute
+  orientation as a rotvec target, and inverts the closedness into a motor gripper target:
+
+  ```python
+  action["ee.x"], action["ee.y"], action["ee.z"] = ee_pose[:3]   # absolute, base frame [m]
+  action["ee.wx"], action["ee.wy"], action["ee.wz"] = orient_rotvec  # orientation target (rotvec)
+  action["ee.gripper_pos"] = (1 - closedness) * 100   # motor units; SO-101 calibrates 100 = open
+  ```
+
+  The gripper polarity (`100 = open, 0 = closed`) is a hardware‑calibration convention in the source — flip it there if the jaw opens when it should close.
+
+- `EEBoundsAndSafety` clamps the EE to a workspace and rate‑limits per‑frame jumps. The clutch's
+  no‑teleport keeps frames small, so `max_ee_step_m` mostly catches transient controller tracking
+  glitches. The z floor is `0.0` (the table plane) so a stray target cannot drive the EE below the
+  table; x/y stay at the loose `[-1, 1]` m box. Set `raise_on_jump=False` so an over‑limit frame is
+  **clamped and warned** instead of raising — a crash mid‑loop would leave the arm uncontrolled:
+
+  ```python
+  EEBoundsAndSafety(
+      end_effector_bounds={"min": [-1.0, -1.0, 0.0], "max": [1.0, 1.0, 1.0]},
+      max_ee_step_m=0.10,
+      raise_on_jump=False,
+  )
+  ```
+
+- `InverseKinematicsEEToJoints(initial_guess_current_joints=False, orientation_weight=0.01)` solves
+  closed‑loop Placo IK. SO‑101 is a 5‑DOF arm, so the IK is position‑dominant; the small
+  `orientation_weight` lets it softly track the orientation target carried in `ee.w*` so the wrist
+  follows the hand, while the under‑determined roll stays partial by design. There is **no**
+  `GripperVelocityToJoint`: the absolute `ee.gripper_pos` is passed straight to `gripper.pos`.
+  `initial_guess_current_joints=False` warm‑starts each solve from the **previous IK solution**
+  rather than re‑seeding from the measured joints, so the joint trajectory stays continuous
+  frame‑to‑frame. Tune `orientation_weight` on hardware — too high fights position tracking, too
+  low ignores the orientation command.
+
+The example also gates safety at the loop level: after the startup reset slew (on by default —
+pass `--reset_to_origin=false` to keep the arm where it is), it commands the robot **only while
+the clutch is engaged**, and re‑sends the measured joints while disengaged, so releasing the
+clutch freezes the arm in place.
+
+See the [Processors for Robots and Teleoperators](./processors_robots_teleop) guide for more on
+adapting the pipeline to other robots.
+
+## Troubleshooting
+
+- **`ModuleNotFoundError: isaacteleop`** — the `isaacteleop` package is not installed in the
+  active environment. Re-run the install command at the top of this guide:
+  `uv pip install "isaacteleop[cloudxr,retargeters-lite]~=1.3.131"`.
+- **No controllers found** — make sure the CloudXR runtime is running, the firewall ports are
+  whitelisted, and the headset is connected (see
+  [Set up CloudXR and connect a headset](#set-up-cloudxr-and-connect-a-headset) and the Isaac
+  Teleop [Quick Start](https://nvidia.github.io/IsaacTeleop/main/getting_started/quick_start.html)).
+- **CloudXR auto-launch failed** — `connect()` raises a `RuntimeError` if the runtime does not
+  come up within its startup timeout. Check the launcher logs under `~/.cloudxr/logs`. Common
+  causes: the EULA was never accepted (run `python -m isaacteleop.cloudxr --accept-eula` once,
+  interactively — the auto-launch prompts on stdin and hangs headless), or the runtime is already
+  running externally (set `LEROBOT_CLOUDXR_SKIP_AUTOLAUNCH=1` or `auto_launch_cloudxr=False` to
+  skip the auto-launch).
+- **Arm does not move** — the clutch is a deadman: you must hold the squeeze/grip past
+  `clutch_threshold`. Lower the threshold if your controller's squeeze is reported softly.
+- **Motion feels misaligned** — confirm the headset/play space orientation. The controller stream
+  is rebased into the robot base frame by the `base_T_anchor` transform on `XRControllerConfig`
+  (default: standard OpenXR → robot axis convention); adjust it if your anchor frame differs.
+
+## Learn more
+
+NVIDIA Isaac Teleop documentation ([docs home](https://nvidia.github.io/IsaacTeleop/),
+[GitHub](https://github.com/NVIDIA/IsaacTeleop)):
+
+- [Quick Start](https://nvidia.github.io/IsaacTeleop/main/getting_started/quick_start.html) —
+  install, run the CloudXR server, connect a headset, run a teleop example.
+- [TeleopSession](https://nvidia.github.io/IsaacTeleop/main/getting_started/teleop_session.html) —
+  the session API `XRController` wraps.
+- [Retargeting interface](https://nvidia.github.io/IsaacTeleop/main/references/retargeting/index.html)
+  and [architecture overview](https://nvidia.github.io/IsaacTeleop/main/overview/architecture.html) —
+  how source nodes and retargeters compose into a pipeline.
+- [Build from source](https://nvidia.github.io/IsaacTeleop/main/getting_started/build_from_source/index.html) —
+  build `isaacteleop` (and its C++ plugins, including the `so101_leader` plugin used above) from a
+  local checkout.
+- [System Requirements](https://nvidia.github.io/IsaacTeleop/main/references/requirements.html) and
+  the [CloudXR SDK docs](https://docs.nvidia.com/cloudxr-sdk) — supported platforms, GPUs,
+  CloudXR/OpenXR runtime versions, and headsets.

docs/source/lekiwi.mdx

@@ -319,7 +319,7 @@ If you want to dive deeper into this important topic, you can check out the [blo
 
 #### Troubleshooting:
 
-- On Linux, if the left and right arrow keys and escape key don't have any effect during data recording, make sure you've set the `$DISPLAY` environment variable. See [pynput limitations](https://pynput.readthedocs.io/en/latest/limitations.html#linux).
+- On Linux, the recording control-flow keys (arrow keys, Escape) work on X11, Wayland, and headless/SSH sessions as long as you run the recording from an interactive terminal (keep it focused) — no `$DISPLAY` setup is needed; the letter equivalents `n` / `r` / `q` also work. Note that **keyboard teleoperation of the LeKiwi base** is different: it relies on a global key backend and therefore works only on an X11 session, a Windows desktop, or macOS with Accessibility/Input Monitoring granted — not on Wayland or headless machines. See [pynput limitations](https://pynput.readthedocs.io/en/latest/limitations.html#linux).
 
 ## Replay an episode
 

docs/source/lerobot-dataset-v3.mdx

@@ -44,7 +44,7 @@ lerobot-record \
   --dataset.num_episodes=5 \
   --dataset.single_task="Grab the black cube" \
   --dataset.streaming_encoding=true \
-  # --dataset.camera_encoder.vcodec=auto \
+  # --dataset.rgb_encoder.vcodec=auto \
   --dataset.encoder_threads=2
 ```
 

docs/source/libero.mdx

@@ -143,7 +143,7 @@ lerobot-train \
   --batch_size=4 \
   --eval.batch_size=1 \
   --eval.n_episodes=1 \
-  --eval_freq=1000
+  --env_eval_freq=1000
 ```
 
 ## Reproducing published results

docs/source/libero_plus.mdx

@@ -173,7 +173,7 @@ lerobot-train \
     --batch_size=4 \
     --eval.batch_size=1 \
     --eval.n_episodes=1 \
-    --eval_freq=1000
+    --env_eval_freq=1000
 ```
 
 ## Relationship to LIBERO

docs/source/lingbot_va.mdx

@@ -0,0 +1,187 @@
+# 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).

docs/source/metaworld.mdx

@@ -120,11 +120,11 @@ lerobot-train \
   --batch_size=4 \
   --eval.batch_size=1 \
   --eval.n_episodes=1 \
-  --eval_freq=1000
+  --env_eval_freq=1000
 ```
 
 ## Practical tips
 
 - Use the one-hot task conditioning for multi-task training (MT10/MT50 conventions) so policies have explicit task context.
 - Inspect the dataset task descriptions and the `info["is_success"]` keys when writing post-processing or logging so your success metrics line up with the benchmark.
-- Adjust `batch_size`, `steps`, and `eval_freq` to match your compute budget.
+- Adjust `batch_size`, `steps`, and `env_eval_freq` to match your compute budget.

docs/source/molmoact2.mdx

@@ -17,7 +17,7 @@ the paper, see [allenai/molmoact2](https://github.com/allenai/molmoact2).
 Install LeRobot with the MolmoAct2 optional dependencies:
 
 ```bash
-pip install -e ".[molmoact2]"
+uv sync --locked --extra molmoact2
 ```
 
 To run the models in this repository, you need an NVIDIA GPU. The measurements
@@ -46,8 +46,8 @@ The repo has been tested with Ubuntu 22.04.
 
 To use MolmoAct2 in a LeRobot training config, set:
 
-```python
-policy.type=molmoact2
+```bash
+--policy.type=molmoact2
 ```
 
 ## Training
@@ -103,7 +103,7 @@ accelerate launch \
   --batch_size=32 \
   --num_workers=4 \
   --log_freq=20 \
-  --eval_freq=-1 \
+  --env_eval_freq=-1 \
   --save_checkpoint=true \
   --save_freq=2000
 ```
@@ -142,7 +142,7 @@ accelerate launch \
   --batch_size=32 \
   --num_workers=4 \
   --log_freq=20 \
-  --eval_freq=-1 \
+  --env_eval_freq=-1 \
   --save_checkpoint=true \
   --save_freq=2000
 ```
@@ -386,6 +386,68 @@ These results demonstrate MolmoAct2's strong performance across diverse robotic
 manipulation tasks. To reproduce them, follow the instructions in the LIBERO
 evaluation section.
 
+## Hardware Deployment (lerobot-rollout)
+
+LeRobot-format checkpoints are available on the Hub for direct use with
+`lerobot-rollout`. Each checkpoint uses specific camera names that must
+match your robot's camera configuration.
+
+### Camera naming convention
+
+Each checkpoint expects specific `observation.images.*` keys.
+If your robot cameras have different names, use `--rename_map` to map them:
+
+| Checkpoint                    | Camera keys            | Description              |
+| ----------------------------- | ---------------------- | ------------------------ |
+| MolmoAct2-LIBERO-LeRobot      | `image`, `wrist_image` | LIBERO sim cameras       |
+| MolmoAct2-BimanualYAM-LeRobot | `top`, `left`, `right` | YAM 3-camera setup       |
+| MolmoAct2-DROID-LeRobot       | `cam0`, `cam1`         | External + wrist         |
+| MolmoAct2-SO100_101-LeRobot   | `cam0`, `cam1`         | Primary + secondary view |
+
+Example with an SO-100 robot using top and side cameras:
+
+```bash
+lerobot-rollout \
+  --policy.path=lerobot/MolmoAct2-SO100_101-LeRobot \
+  --rename_map='{"observation.images.top": "observation.images.cam0", "observation.images.side": "observation.images.cam1"}' \
+  --robot.type=so100_follower \
+  --robot.port=/dev/ttyACM0 \
+  --robot.cameras='{
+      top: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30},
+      side: {type: opencv, index_or_path: 2, width: 640, height: 480, fps: 30}
+  }' \
+  --task="pick up the red cube" --duration=30
+```
+
+To use a wrist camera instead, just change the rename mapping:
+
+```bash
+--rename_map='{"observation.images.top": "observation.images.cam0", "observation.images.wrist": "observation.images.cam1"}'
+```
+
+### Joint frame transform (SO-100/101 zero-shot)
+
+<Tip warning={true}>
+The MolmoAct2-SO100_101 checkpoint was trained on data that uses a different
+joint calibration convention than LeRobot >= 0.5.0. Without a frame
+correction, the arm may move in the wrong direction.
+
+This affects both **zero-shot deployment** and **fine-tuning** from the
+original checkpoint. The pretrained weights expect the old convention, so
+all joint data (observations and actions) must be transformed to match.
+
+The converted LeRobot checkpoint (`lerobot/MolmoAct2-SO100_101-LeRobot`)
+already includes this correction in its processor pipeline. If you convert
+or fine-tune the checkpoint yourself, set the following in the policy config (`configuration_molmoact2.py`):
+
+- `joint_signs`: `[1, -1, 1, 1, 1, 1]` (flips shoulder_lift direction)
+- `joint_offsets`: `[0, 90, 90, 0, 0, 0]` (shifts shoulder_lift and elbow_flex by 90°)
+
+See the [backward compatibility guide](./backwardcomp) for details on the
+calibration change.
+
+</Tip>
+
 ## Differences From the Original Implementation
 
 This LeRobot port is intended to match MolmoAct2 behavior while using LeRobot's

docs/source/multi_gpu_training.mdx

@@ -95,7 +95,7 @@ If you want to scale your hyperparameters when using multiple GPUs, you should d
 accelerate launch --num_processes=2 $(which lerobot-train) \
   --optimizer.lr=2e-4 \
   --dataset.repo_id=lerobot/pusht \
-  --policy=act
+  --policy.type=act
 ```
 
 **Training Steps Scaling:**
@@ -110,9 +110,64 @@ accelerate launch --num_processes=2 $(which lerobot-train) \
   --batch_size=8 \
   --steps=50000 \
   --dataset.repo_id=lerobot/pusht \
-  --policy=act
+  --policy.type=act
 ```
 
+## Training Large Models with FSDP
+
+DDP replicates the full model on every GPU, so a model that doesn't fit on one GPU won't fit under
+DDP either. For large models, use **FSDP** (Fully Sharded Data Parallel), which shards parameters,
+gradients, and optimizer state across GPUs. See the [accelerate FSDP guide](https://huggingface.co/docs/accelerate/usage_guides/fsdp) for background.
+
+An example on how to launch LeRobot training with FSDP across 4 GPUs (1 machine):
+
+```bash
+accelerate launch --config_file fsdp.yaml --num_processes=4 $(which lerobot-train) \
+  --dataset.repo_id=${HF_USER}/my_dataset \
+  --policy.type=<your_policy> \
+  --output_dir=outputs/train/my_policy_fsdp
+```
+
+A minimal `fsdp.yaml` (FSDP1; shards params/grads/optimizer — ZeRO-3-equivalent):
+
+```yaml
+compute_environment: LOCAL_MACHINE
+distributed_type: FSDP
+mixed_precision: bf16
+num_machines: 1
+num_processes: 4
+fsdp_config:
+  fsdp_version: 1
+  fsdp_sharding_strategy: FULL_SHARD # params + grads + optimizer (ZeRO-3)
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_transformer_layer_cls_to_wrap: <YourTransformerBlock> # repeated block class to shard
+  fsdp_use_orig_params: true # required: optimizer is built pre-prepare
+  fsdp_state_dict_type: FULL_STATE_DICT
+```
+
+Set `fsdp_transformer_layer_cls_to_wrap` to your model's repeated transformer-block class so each
+block is sharded as its own unit. `fsdp_use_orig_params: true` is required because LeRobot builds the
+optimizer before `accelerator.prepare()`.
+
+### FSDP checkpoints
+
+LeRobot gathers the full state dict across all ranks and the main process writes it as a single
+`model.safetensors`, loadable as usual with `Policy.from_pretrained(...)`. Two things to look out for:
+
+- **Checkpoints store fp32 weights.** Under mixed precision (`bf16`/`fp16`) FSDP keeps an fp32 master
+  copy, and the checkpoint saves it (~2× the bf16 size on disk) so training can resume consistently
+  with the fp32 optimizer state; `from_pretrained` casts back to the policy dtype on load. FSDP-specific
+  caveat: an fp32 checkpoint is materialized in full precision on the target device _before_ casting,
+  so loading it for inference on a tight GPU can OOM even when the bf16 model would fit — load on CPU
+  first, or cast `model.safetensors` to the deployment dtype offline.
+- The sharded optimizer state is gathered into a full (world-size-independent) state dict and saved
+  alongside the model in the same `optimizer_state.safetensors` / `optimizer_param_groups.json`
+  format as single-GPU training, so **resume-from-checkpoint is supported** with `--resume=true`.
+  Resume reshards both the model and the optimizer state to the _current_ FSDP topology, so you can
+  resume an FSDP checkpoint on a different number of GPUs. Note that the data sampler is only
+  sample-exact when the world size and batch size match the original run (a warning is logged
+  otherwise); the optimizer/model state itself is unaffected.
+
 ## Notes
 
 - The `--policy.use_amp` flag in `lerobot-train` is only used when **not** running with accelerate. When using accelerate, mixed precision is controlled by accelerate's configuration.

docs/source/multi_task_dit.mdx

@@ -314,7 +314,7 @@ lerobot-train \
   --steps=30000 \
   --save_freq=1000 \
   --log_freq=100 \
-  --eval_freq=1000 \
+  --env_eval_freq=1000 \
   --policy.type=multi_task_dit \
   --policy.device=cuda \
   --policy.horizon=32 \

docs/source/pi0fast.mdx

@@ -96,7 +96,7 @@ lerobot-train \
     --policy.type=pi0_fast \
     --output_dir=./outputs/pi0fast_training \
     --job_name=pi0fast_training \
-    --policy.pretrained_path=lerobot/pi0_fast_base \
+    --policy.pretrained_path=lerobot/pi0fast-base \
     --policy.dtype=bfloat16 \
     --policy.gradient_checkpointing=true \
     --policy.chunk_size=10 \
@@ -187,7 +187,7 @@ lerobot-train \
   --dataset.repo_id=lerobot/libero \
   --output_dir=outputs/libero_pi0fast \
   --job_name=libero_pi0fast \
-  --policy.path=lerobot/pi0fast_base \
+  --policy.path=lerobot/pi0fast-base \
   --policy.dtype=bfloat16 \
   --steps=100000 \
   --save_freq=20000 \

docs/source/policy_evo1_README.md

@@ -0,0 +1,18 @@
+# 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}},
+}
+```

docs/source/policy_fastwam_README.md

@@ -0,0 +1,56 @@
+## Research Paper
+
+Paper: https://arxiv.org/abs/2603.16666
+
+## Repository
+
+Code: https://github.com/yuantianyuan01/FastWAM
+
+Project page: https://yuantianyuan01.github.io/FastWAM/
+
+## Citation
+
+```bibtex
+@article{yuan2026fastwam,
+  title = {Fast-WAM: Do World Action Models Need Test-time Future Imagination?},
+  author = {Tianyuan Yuan and Zibin Dong and Yicheng Liu and Hang Zhao},
+  journal = {arXiv preprint arXiv:2603.16666},
+  year = {2026},
+  url = {https://arxiv.org/abs/2603.16666}
+}
+```
+
+## Additional Resources
+
+Base video model: https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B
+
+Released upstream checkpoints: https://huggingface.co/yuanty/fastwam
+
+## Results
+
+Evaluated on LIBERO with [`ZibinDong/fastwam_libero_uncond_2cam224`](https://huggingface.co/ZibinDong/fastwam_libero_uncond_2cam224):
+
+| Suite          | Success rate | n_episodes |
+| -------------- | -----------: | ---------: |
+| libero_spatial |        97.6% |        500 |
+| libero_object  |        99.0% |        500 |
+| libero_goal    |        95.0% |        500 |
+| libero_10      |        94.0% |        500 |
+| **average**    |    **96.4%** |       2000 |
+
+Reproduce: `lerobot-eval --policy.path=ZibinDong/fastwam_libero_uncond_2cam224 --policy.device=cuda --policy.torch_dtype=float32 --policy.n_action_steps=10 --env.type=libero --env.task=libero_spatial --env.observation_height=256 --env.observation_width=256 --eval.batch_size=1 --eval.n_episodes=50 --seed=0 --env.episode_length=300`.
+
+For LIBERO-10, use `--env.task=libero_10 --env.episode_length=600`:
+
+```bash
+lerobot-eval \
+    --policy.path=ZibinDong/fastwam_libero_uncond_2cam224 \
+    --policy.device=cuda \
+    --policy.torch_dtype=float32 \
+    --policy.n_action_steps=10 \
+    --env.type=libero \
+    --env.task=libero_10 --env.observation_height=256 --env.observation_width=256 \
+    --eval.batch_size=1 \
+    --eval.n_episodes=50 \
+    --seed=0 --env.episode_length=600
+```

docs/source/policy_groot_README.md

@@ -1,6 +1,13 @@
 ## Research Paper
 
-Paper: https://research.nvidia.com/labs/gear/gr00t-n1_5/
+GR00T N1 technical report (covers the GR00T N1.x family, including N1.7): https://arxiv.org/abs/2503.14734
+
+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
 
@@ -24,4 +31,108 @@ Code: https://github.com/NVIDIA/Isaac-GR00T
 
 Blog: https://developer.nvidia.com/isaac/gr00t
 
-Hugging Face Model: https://huggingface.co/nvidia/GR00T-N1.5-3B
+Hugging Face Models:
+
+- 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>

docs/source/reachy2.mdx

@@ -161,7 +161,7 @@ lerobot-record \
     --dataset.private=true \
     --dataset.streaming_encoding=true \
     --dataset.encoder_threads=2 \
-    # --dataset.camera_encoder.vcodec=auto \
+    # --dataset.rgb_encoder.vcodec=auto \
     --display_data=true
 ```
 
@@ -203,7 +203,7 @@ lerobot-record \
     --dataset.private=true \
     --dataset.streaming_encoding=true \
     --dataset.encoder_threads=2 \
-    # --dataset.camera_encoder.vcodec=auto \
+    # --dataset.rgb_encoder.vcodec=auto \
     --display_data=true
 ```
 

docs/source/robocasa.mdx

@@ -166,7 +166,7 @@ lerobot-train \
   --output_dir=./outputs/smolvla_robocasa_CloseFridge \
   --steps=100000 \
   --batch_size=4 \
-  --eval_freq=5000 \
+  --env_eval_freq=5000 \
   --eval.batch_size=1 \
   --eval.n_episodes=5 \
   --save_freq=10000

docs/source/so101.mdx

@@ -122,7 +122,7 @@ The video below shows the sequence of steps for setting the motor ids.
 
 #### Follower
 
-Connect the usb cable from your computer and the power supply to the follower arm's controller board. Then, run the following command or run the API example with the port you got from the previous step. You'll also need to give your leader arm a name with the `id` parameter.
+Connect the usb cable from your computer and the power supply to the follower arm's controller board. Then, run the following command or run the API example with the port you got from the previous step. You'll also need to give your follower arm a name with the `id` parameter.
 
 <hfoptions id="setup_motors">
 <hfoption id="Command">

docs/source/streaming_video_encoding.mdx

@@ -17,7 +17,7 @@ This makes `save_episode()` near-instant (the video is already encoded by the ti
 | Parameter               | CLI Flag                          | Type          | Default       | Description                                                       |
 | ----------------------- | --------------------------------- | ------------- | ------------- | ----------------------------------------------------------------- |
 | `streaming_encoding`    | `--dataset.streaming_encoding`    | `bool`        | `True`        | Enable real-time encoding during capture                          |
-| `vcodec`                | `--dataset.camera_encoder.vcodec` | `str`         | `"libsvtav1"` | Video codec. `"auto"` detects best HW encoder                     |
+| `vcodec`                | `--dataset.rgb_encoder.vcodec`    | `str`         | `"libsvtav1"` | Video codec. `"auto"` detects best HW encoder                     |
 | `encoder_threads`       | `--dataset.encoder_threads`       | `int \| None` | `None` (auto) | Threads per encoder instance. `None` will leave the vcoded decide |
 | `encoder_queue_maxsize` | `--dataset.encoder_queue_maxsize` | `int`         | `30`          | Max buffered frames per camera (~1s at 30fps). Consumes RAM       |
 
@@ -82,15 +82,15 @@ Use HW encoding when:
 
 ### Available HW Encoders
 
-| Encoder             | Platform      | Hardware                                                                                         | CLI Value                                           |
-| ------------------- | ------------- | ------------------------------------------------------------------------------------------------ | --------------------------------------------------- |
-| `h264_videotoolbox` | macOS         | Apple Silicon / Intel                                                                            | `--dataset.camera_encoder.vcodec=h264_videotoolbox` |
-| `hevc_videotoolbox` | macOS         | Apple Silicon / Intel                                                                            | `--dataset.camera_encoder.vcodec=hevc_videotoolbox` |
-| `h264_nvenc`        | Linux/Windows | NVIDIA GPU                                                                                       | `--dataset.camera_encoder.vcodec=h264_nvenc`        |
-| `hevc_nvenc`        | Linux/Windows | NVIDIA GPU                                                                                       | `--dataset.camera_encoder.vcodec=hevc_nvenc`        |
-| `h264_vaapi`        | Linux         | Intel/AMD GPU                                                                                    | `--dataset.camera_encoder.vcodec=h264_vaapi`        |
-| `h264_qsv`          | Linux/Windows | Intel Quick Sync                                                                                 | `--dataset.camera_encoder.vcodec=h264_qsv`          |
-| `auto`              | Any           | Probes the system for available HW encoders. Falls back to `libsvtav1` if no HW encoder is found | `--dataset.camera_encoder.vcodec=auto`              |
+| Encoder             | Platform      | Hardware                                                                                         | CLI Value                                        |
+| ------------------- | ------------- | ------------------------------------------------------------------------------------------------ | ------------------------------------------------ |
+| `h264_videotoolbox` | macOS         | Apple Silicon / Intel                                                                            | `--dataset.rgb_encoder.vcodec=h264_videotoolbox` |
+| `hevc_videotoolbox` | macOS         | Apple Silicon / Intel                                                                            | `--dataset.rgb_encoder.vcodec=hevc_videotoolbox` |
+| `h264_nvenc`        | Linux/Windows | NVIDIA GPU                                                                                       | `--dataset.rgb_encoder.vcodec=h264_nvenc`        |
+| `hevc_nvenc`        | Linux/Windows | NVIDIA GPU                                                                                       | `--dataset.rgb_encoder.vcodec=hevc_nvenc`        |
+| `h264_vaapi`        | Linux         | Intel/AMD GPU                                                                                    | `--dataset.rgb_encoder.vcodec=h264_vaapi`        |
+| `h264_qsv`          | Linux/Windows | Intel Quick Sync                                                                                 | `--dataset.rgb_encoder.vcodec=h264_qsv`          |
+| `auto`              | Any           | Probes the system for available HW encoders. Falls back to `libsvtav1` if no HW encoder is found | `--dataset.rgb_encoder.vcodec=auto`              |
 
 > [!NOTE]
 > In order to use the HW accelerated encoders you might need to upgrade your GPU drivers.
@@ -100,15 +100,15 @@ Use HW encoding when:
 
 ## 5. Troubleshooting
 
-| Symptom                                                            | Likely Cause                                 | Fix                                                                                                                                                                                                                                                                                                 |
-| ------------------------------------------------------------------ | -------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| System freezes or choppy robot movement or Rerun visualization lag | CPU starved (100% load usage)                | Close other apps, reduce encoding throughput, lower `encoder_threads`, use `h264`, use `display_data=False`. If the CPU continues to be at 100% then it might be insufficient for your setup, consider `--dataset.streaming_encoding=false` or HW encoding (`--dataset.camera_encoder.vcodec=auto`) |
-| "Encoder queue full" warnings or dropped frames in dataset         | Encoder can't keep up (Queue overflow)       | If CPU is not at 100%: Increase `encoder_threads`, increase `encoder_queue_maxsize` or use HW encoding (`--dataset.camera_encoder.vcodec=auto`).                                                                                                                                                    |
-| High RAM usage                                                     | Queue filling faster than encoding           | `encoder_threads` too low or CPU insufficient. Reduce `encoder_queue_maxsize` or use HW encoding                                                                                                                                                                                                    |
-| Large video files                                                  | Using HW encoder or H.264                    | Expected trade-off. Switch to `libsvtav1` if CPU allows                                                                                                                                                                                                                                             |
-| `save_episode()` still slow                                        | `streaming_encoding` is `False`              | Set `--dataset.streaming_encoding=true`                                                                                                                                                                                                                                                             |
-| Encoder thread crash                                               | Codec not available or invalid settings      | Check `vcodec` is installed, try `--dataset.camera_encoder.vcodec=auto`                                                                                                                                                                                                                             |
-| Recorded dataset is missing frames                                 | CPU/GPU starvation or occasional load spikes | If ~5% of frames are missing, your system is likely overloaded — follow the recommendations above. If fewer frames are missing (~2%), they are probably due to occasional transient load spikes (often at startup) and can be considered expected.                                                  |
+| Symptom                                                            | Likely Cause                                 | Fix                                                                                                                                                                                                                                                                                              |
+| ------------------------------------------------------------------ | -------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
+| System freezes or choppy robot movement or Rerun visualization lag | CPU starved (100% load usage)                | Close other apps, reduce encoding throughput, lower `encoder_threads`, use `h264`, use `display_data=False`. If the CPU continues to be at 100% then it might be insufficient for your setup, consider `--dataset.streaming_encoding=false` or HW encoding (`--dataset.rgb_encoder.vcodec=auto`) |
+| "Encoder queue full" warnings or dropped frames in dataset         | Encoder can't keep up (Queue overflow)       | If CPU is not at 100%: Increase `encoder_threads`, increase `encoder_queue_maxsize` or use HW encoding (`--dataset.rgb_encoder.vcodec=auto`).                                                                                                                                                    |
+| High RAM usage                                                     | Queue filling faster than encoding           | `encoder_threads` too low or CPU insufficient. Reduce `encoder_queue_maxsize` or use HW encoding                                                                                                                                                                                                 |
+| Large video files                                                  | Using HW encoder or H.264                    | Expected trade-off. Switch to `libsvtav1` if CPU allows                                                                                                                                                                                                                                          |
+| `save_episode()` still slow                                        | `streaming_encoding` is `False`              | Set `--dataset.streaming_encoding=true`                                                                                                                                                                                                                                                          |
+| Encoder thread crash                                               | Codec not available or invalid settings      | Check `vcodec` is installed, try `--dataset.rgb_encoder.vcodec=auto`                                                                                                                                                                                                                             |
+| Recorded dataset is missing frames                                 | CPU/GPU starvation or occasional load spikes | If ~5% of frames are missing, your system is likely overloaded — follow the recommendations above. If fewer frames are missing (~2%), they are probably due to occasional transient load spikes (often at startup) and can be considered expected.                                               |
 
 ## 6. Recommended Configurations
 
@@ -146,7 +146,7 @@ On very constrained systems, streaming encoding may compete too heavily with the
 # 2camsx 640x480x3 @30fps: Requires some tuning.
 
 # Use H.264, disable streaming, consider batching encoding
-lerobot-record --dataset.camera_encoder.vcodec=h264 --dataset.streaming_encoding=false ...
+lerobot-record --dataset.rgb_encoder.vcodec=h264 --dataset.streaming_encoding=false ...
 ```
 
 ## 7. Closing note

docs/source/using_dataset_tools.mdx

@@ -11,8 +11,9 @@ LeRobot provides several utilities for manipulating datasets:
 3. **Merge Datasets** - Combine multiple datasets into one. The datasets must have identical features, and episodes are concatenated in the order specified in `repo_ids`
 4. **Add Features** - Add new features to a dataset
 5. **Remove Features** - Remove features from a dataset
-6. **Convert to Video** - Convert image-based datasets to video format for efficient storage
-7. **Show the Info of Datasets** - Show the summary of datasets information such as number of episode etc.
+6. **Convert to Video** - Convert image-based datasets to video format for efficient storage (RGB and depth cameras are encoded with separate encoders)
+7. **Re-encode Videos** - Re-encode an existing video dataset's RGB and/or depth streams with new encoder settings
+8. **Show the Info of Datasets** - Show the summary of datasets information such as number of episode etc.
 
 The core implementation is in `lerobot.datasets.dataset_tools`.
 An example script detailing how to use the tools API is available in `examples/dataset/use_dataset_tools.py`.
@@ -117,10 +118,19 @@ lerobot-edit-dataset \
     --repo_id lerobot/pusht_image \
     --operation.type convert_image_to_video \
     --operation.output_dir outputs/pusht_video \
-    --operation.camera_encoder.vcodec libsvtav1 \
-    --operation.camera_encoder.pix_fmt yuv420p \
-    --operation.camera_encoder.g 2 \
-    --operation.camera_encoder.crf 30
+    --operation.rgb_encoder.vcodec libsvtav1 \
+    --operation.rgb_encoder.pix_fmt yuv420p \
+    --operation.rgb_encoder.g 2 \
+    --operation.rgb_encoder.crf 30
+
+# Convert a dataset that includes depth maps, customizing the depth encoder
+lerobot-edit-dataset \
+    --repo_id lerobot/pusht_image \
+    --operation.type convert_image_to_video \
+    --operation.output_dir outputs/pusht_video \
+    --operation.depth_encoder.depth_min 0.01 \
+    --operation.depth_encoder.depth_max 10.0 \
+    --operation.depth_encoder.use_log true
 
 # Convert only specific episodes
 lerobot-edit-dataset \
@@ -147,11 +157,42 @@ lerobot-edit-dataset \
 **Parameters:**
 
 - `output_dir`: Custom output directory (optional - by default uses `new_repo_id` or `{repo_id}_video`)
-- `camera_encoder`: Video encoder settings — all sub-fields accessible via `--operation.camera_encoder.<field>. See [Video Encoding Parameters](./video_encoding_parameters) for more details.
+- `rgb_encoder`: Video encoder settings applied to RGB cameras — all sub-fields accessible via `--operation.rgb_encoder.<field>`. See [Video Encoding Parameters](./video_encoding_parameters) for more details.
+- `depth_encoder`: Video encoder settings applied to depth-map cameras (e.g. from an Intel RealSense). In addition to the standard encoder fields it exposes the depth quantization knobs (`depth_min`, `depth_max`, `shift`, `use_log`), accessible via `--operation.depth_encoder.<field>`. These quantization settings are persisted to the dataset metadata so depth can be dequantized back to physical units on load. See the [Depth streams](./video_encoding_parameters#depth-streams) section for details.
 - `episode_indices`: List of specific episodes to convert (default: all episodes)
 - `num_workers`: Number of parallel workers for processing (default: 4)
 
-**Note:** The resulting dataset will be a proper LeRobotDataset with all cameras encoded as videos in the `videos/` directory, with parquet files containing only metadata (no raw image data). All episodes, stats, and tasks are preserved.
+**Note:** The resulting dataset will be a proper LeRobotDataset with all cameras encoded as videos in the `videos/` directory, with parquet files containing only metadata (no raw image data). Depth-map cameras are detected automatically and routed to the `depth_encoder`, while RGB cameras use the `rgb_encoder`. All episodes, stats, and tasks are preserved.
+
+#### Re-encode Videos
+
+Re-encode the videos of an existing video dataset with different encoder settings, without going back to raw frames. RGB videos use the `rgb_encoder` and depth videos use the `depth_encoder`. Provide only the encoder(s) you want to re-encode; the other stream type is left untouched.
+
+```bash
+# Re-encode all RGB videos with new settings (saves to lerobot/pusht_reencoded by default)
+lerobot-edit-dataset \
+    --repo_id lerobot/pusht \
+    --operation.type reencode_videos \
+    --operation.rgb_encoder.vcodec h264 \
+    --operation.rgb_encoder.pix_fmt yuv420p \
+    --operation.rgb_encoder.crf 23
+
+# Re-encode both RGB and depth videos in a dataset with depth maps
+lerobot-edit-dataset \
+    --repo_id lerobot/pusht_depth \
+    --operation.type reencode_videos \
+    --operation.rgb_encoder.vcodec h264 \
+    --operation.depth_encoder.crf 50
+```
+
+**Parameters:**
+
+- `rgb_encoder`: Encoder settings applied to every RGB video. Omit to skip re-encoding RGB videos.
+- `depth_encoder`: Encoder settings applied to every depth video. Omit to skip re-encoding depth videos.
+- `num_workers`: Number of parallel workers for processing.
+
+> [!NOTE]
+> When re-encoding depth videos, the existing depth quantization parameters (`depth_min`, `depth_max`, `shift`, `use_log`) and the `is_depth_map` flag are **preserved** — re-encoding only changes the codec/quality of the stored stream, not how depth is dequantized on load.
 
 ### Show the information of datasets
 
@@ -224,6 +265,8 @@ 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

docs/source/video_encoding_parameters.mdx

@@ -2,16 +2,15 @@
 
 When video storage is enabled, LeRobot stores each camera stream as an **MP4** file instead of saving one image file per timestep. Video encoding compresses across time, which usually cuts dataset size and I/O compared to a pile of PNG, while keeping MP4 — a format every player and loader understands.
 
-Encoding frames into an MP4 is a full FFmpeg pipeline: choice of encoder, pixel format, GOP/keyframes, quality vs. speed, and optional extra encoder flags. Most of these knobs are user-tunable through `camera_encoder`, a nested `VideoEncoderConfig` (`lerobot.configs.video.VideoEncoderConfig`) passed through PyAV.
+Encoding frames into an MP4 is a full FFmpeg pipeline: choice of encoder, pixel format, GOP/keyframes, quality vs. speed, and optional extra encoder flags. Most of these knobs are user-tunable through `rgb_encoder`, a nested `RGBEncoderConfig` (`lerobot.configs.video.RGBEncoderConfig`) passed through PyAV.
 
-You can set these parameters from the CLI with `--dataset.camera_encoder.<field>` (e.g. with `lerobot-record` or `lerobot-rollout`). The same block applies to every camera video stream in that run.
+You can set these parameters from the CLI with `--dataset.rgb_encoder.<field>` (e.g. with `lerobot-record` or `lerobot-rollout`). The same block applies to every camera video stream in that run.
 
-<Tip>
-  Video storage must be on for `camera_encoder` to have any effect —
-  `use_videos=True` in Python APIs, or `--dataset.video=true` on the CLI (the
-  recording default). With video off, inputs stay as images and `camera_encoder`
-  is ignored.
-</Tip>
+> [!TIP]
+> Video storage must be on for `rgb_encoder` to have any effect —
+> `use_videos=True` in Python APIs, or `--dataset.video=true` on the CLI (the
+> recording default). With video off, inputs stay as images and `rgb_encoder` is
+> ignored.
 
 For details on **when** frames are written vs. encoded (streaming vs. post-episode), queues, and other top-level `--dataset.*` switches, see [Streaming Video Encoding](./streaming_video_encoding). For an encoding-parameter comparison and experiments, see the [video-benchmark Space](https://huggingface.co/spaces/lerobot/video-benchmark).
 
@@ -33,9 +32,9 @@ lerobot-record \
     --dataset.single_task="Grab the cube" \
     --dataset.streaming_encoding=true \
     --dataset.encoder_threads=2 \
-    --dataset.camera_encoder.vcodec=h264 \
-    --dataset.camera_encoder.preset=fast \
-    --dataset.camera_encoder.extra_options={"tune": "film", "profile:v": "high", "bf": 2} \
+    --dataset.rgb_encoder.vcodec=h264 \
+    --dataset.rgb_encoder.preset=fast \
+    --dataset.rgb_encoder.extra_options={"tune": "film", "profile:v": "high", "bf": 2} \
     --display_data=true
 ```
 
@@ -43,14 +42,12 @@ lerobot-record \
 
 ## Tuning parameters
 
-<Tip warning={true}>
-The defaults are tuned to balance **compression ratio**, **visual quality**, and **decoding/seek speed** for typical robotics datasets. Changing them can affect both recording (CPU load, frame drops) and training (decoding throughput, image quality).
+> [!WARNING]
+> The defaults are tuned to balance **compression ratio**, **visual quality**, and **decoding/seek speed** for typical robotics datasets. Changing them can affect both recording (CPU load, frame drops) and training (decoding throughput, image quality).
+>
+> Only override these parameters if you have a specific reason to, and measure the impact on your pipeline before relying on the new settings.
 
-Only override these parameters if you have a specific reason to, and measure the impact on your pipeline before relying on the new settings.
-
-</Tip>
-
-All flags below are prefixed with `--dataset.camera_encoder.` on the CLI.
+All flags below are prefixed with `--dataset.rgb_encoder.` on the CLI.
 
 | Parameter       | Type             | Default       | Description                                                                                                                                                                            |
 | --------------- | ---------------- | ------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
@@ -65,6 +62,144 @@ All flags below are prefixed with `--dataset.camera_encoder.` on the CLI.
 
 ---
 
+## Depth streams
+
+Depth maps (Intel RealSense, Reachy 2) are stored as their **own video streams** alongside the RGB streams. Raw depth (`uint16` millimetres or `float32` metres) can't survive an 8-bit codec, so LeRobot **quantizes** each map to a 12-bit code (`[0, 4095]`) — logarithmically by default, to match the `1/depth` error profile of depth sensors — then packs it into a high-bit-depth pixel format (`gray12le`) and encodes it with a 12-bit codec.
+
+<div style="margin:28px 0;padding:14px 0;">
+  <div style="margin:0 auto;display:flex;flex-wrap:wrap;justify-content:center;align-items:stretch;gap:6px;font-family:'Source Sans 3',ui-sans-serif,system-ui,sans-serif;font-size:14px;font-weight:600;color:#1B1B1D;">
+    <span style="display:flex;flex-direction:column;justify-content:center;align-items:center;text-align:center;gap:2px;background:#DBEAFE;color:#1D4ED8;border-radius:9px;padding:8px 12px;">
+      <span>Raw depth</span>
+      <span style="font-size:11px;font-weight:400;color:#3B6FD4;white-space:nowrap;">
+        uint16 mm
+        <br />
+        float32 m
+      </span>
+    </span>
+    <span style="display:flex;align-items:center;font-size:16px;color:#C3CBD9;">
+      →
+    </span>
+    <div style="border:2px dashed #C4B5FD;border-radius:13px;padding:18px 12px 12px;position:relative;display:flex;align-items:stretch;gap:6px;">
+      <span style="position:absolute;top:-10px;left:12px;background:#fff;padding:0 6px;font-size:11px;font-weight:700;color:#7E22CE;text-transform:uppercase;letter-spacing:0.5px;white-space:nowrap;">
+        Record time
+      </span>
+      <span style="display:flex;flex-direction:column;justify-content:center;align-items:center;text-align:center;gap:2px;background:#F3E8FF;color:#7E22CE;border-radius:9px;padding:8px 12px;">
+        <span>Clip</span>
+        <span style="font-size:11px;font-weight:400;color:#9061C2;white-space:nowrap;">
+          to [depth_min,
+          <br />
+          depth_max]
+        </span>
+      </span>
+      <span style="display:flex;align-items:center;font-size:16px;color:#C3CBD9;">
+        →
+      </span>
+      <span style="display:flex;flex-direction:column;justify-content:center;align-items:center;text-align:center;gap:2px;background:#F3E8FF;color:#7E22CE;border-radius:9px;padding:8px 12px;">
+        <span>Quantize</span>
+        <span style="font-size:11px;font-weight:400;color:#9061C2;white-space:nowrap;">
+          12-bit codes 0–4095
+          <br />
+          log (default) or linear
+        </span>
+      </span>
+      <span style="display:flex;align-items:center;font-size:16px;color:#C3CBD9;">
+        →
+      </span>
+      <span style="display:flex;flex-direction:column;justify-content:center;align-items:center;text-align:center;gap:2px;background:#F3E8FF;color:#7E22CE;border-radius:9px;padding:8px 12px;">
+        <span>Pack</span>
+        <span style="font-size:11px;font-weight:400;color:#9061C2;white-space:nowrap;">
+          into gray12le
+          <br />
+          plane
+        </span>
+      </span>
+      <span style="display:flex;align-items:center;font-size:16px;color:#C3CBD9;">
+        →
+      </span>
+      <span style="display:flex;flex-direction:column;justify-content:center;align-items:center;text-align:center;gap:2px;background:#F3E8FF;color:#7E22CE;border-radius:9px;padding:8px 12px;">
+        <span>Encode</span>
+        <span style="font-size:11px;font-weight:400;color:#9061C2;white-space:nowrap;">
+          HEVC
+          <br />
+          Main 12
+        </span>
+      </span>
+    </div>
+    <span style="display:flex;align-items:center;font-size:16px;color:#C3CBD9;">
+      →
+    </span>
+    <span style="display:flex;flex-direction:column;justify-content:center;align-items:center;text-align:center;gap:2px;background:#FEF3C7;color:#B45309;border-radius:9px;padding:8px 12px;">
+      <span>MP4</span>
+      <span style="font-size:11px;font-weight:400;color:#C77D18;white-space:nowrap;">
+        stored
+        <br />
+        stream
+      </span>
+    </span>
+    <span style="display:flex;align-items:center;font-size:16px;color:#34A06B;">
+      →
+    </span>
+    <div style="border:2px dashed #6EE7B7;border-radius:13px;padding:18px 12px 12px;position:relative;display:flex;align-items:center;gap:6px;">
+      <span style="position:absolute;top:-10px;left:12px;background:#fff;padding:0 6px;font-size:11px;font-weight:700;color:#047857;text-transform:uppercase;letter-spacing:0.5px;white-space:nowrap;">
+        Load time
+      </span>
+      <span style="display:flex;flex-direction:column;justify-content:center;align-items:center;text-align:center;gap:2px;background:#D1FAE5;color:#047857;border-radius:9px;padding:8px 12px;">
+        <span>Dequantize</span>
+        <span style="font-size:11px;font-weight:400;color:#059669;white-space:nowrap;">
+          to mm / m
+        </span>
+      </span>
+    </div>
+  </div>
+</div>
+
+Configure the depth pipeline through a parallel **`depth_encoder`** block (`DepthEncoderConfig`). It shares every `RGBEncoderConfig` field (`vcodec`, `pix_fmt`, `crf`, …) and adds four quantizer knobs, set via `--dataset.depth_encoder.<field>`:
+
+```bash
+lerobot-record \
+    ... \
+    --dataset.depth_encoder.vcodec=hevc \
+    --dataset.depth_encoder.depth_min=0.05 \
+    --dataset.depth_encoder.depth_max=5.0 \
+    --dataset.depth_encoder.use_log=true
+```
+
+| Parameter       | Type    | Default                         | Description                                                                                                                            |
+| --------------- | ------- | ------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------- |
+| `vcodec`        | `str`   | `"hevc"`                        | HEVC Main 12 (a 12-bit-capable codec, MP4-compatible).                                                                                 |
+| `extra_options` | `dict`  | `{"x265-params": "lossless=1"}` | **Depth defaults to lossless** (exact round-trip); `crf` is ignored. Pass `extra_options={}` and set `crf` for a smaller lossy stream. |
+| `pix_fmt`       | `str`   | `"gray12le"`                    | Single-channel 12-bit pixel format used to carry the quantized codes.                                                                  |
+| `depth_min`     | `float` | `0.01`                          | Depth in metres mapped to quantum `0`. Values below are clipped on decode.                                                             |
+| `depth_max`     | `float` | `10.0`                          | Depth in metres mapped to quantum `4095`. Values above are clipped on decode.                                                          |
+| `shift`         | `float` | `3.5`                           | Pre-log offset (metres) used in logarithmic quantization for numerical stability near zero. Must satisfy `depth_min + shift > 0`.      |
+| `use_log`       | `bool`  | `True`                          | If `true`, quantize in log-space (recommended for typical depth sensors). Set to `false` for uniform/linear quantization.              |
+
+> [!TIP]
+> `depth_min`, `depth_max`, and `shift` are always interpreted in **metres**, regardless of the input depth's unit. Inputs are auto-detected: integer arrays (e.g. `uint16` millimetres straight from a RealSense) are treated as millimetres, floating arrays as metres.
+> Pick `depth_min` / `depth_max` to bracket the actual working range of your sensor — quanta outside that range saturate, which can crush detail at the boundaries.
+
+Depth features are flagged with `"is_depth_map": true` in `meta/info.json`, and their quantizer settings (`video.depth_min`, `video.depth_max`, `video.shift`, `video.use_log`) are persisted — which is what lets depth be **dequantized back to physical units** on load.
+
+### Output unit at load time
+
+`depth_encoder` is a **record-time** concern. The unit that depth maps are dequantized to on _load_ (e.g. during training) is set separately by the read-time flag `--dataset.depth_output_unit`:
+
+```bash
+lerobot-train \
+    --dataset.repo_id=<my_username>/<my_dataset_name> \
+    --dataset.depth_output_unit=m \
+    --policy.type=act
+```
+
+| Parameter           | Type  | Default | Description                                                                                  |
+| ------------------- | ----- | ------- | -------------------------------------------------------------------------------------------- |
+| `depth_output_unit` | `str` | `"mm"`  | Physical unit depth maps are dequantized to on load: `"mm"` (millimetres) or `"m"` (metres). |
+
+> [!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.
+
+---
+
 ## Persistence in dataset metadata
 
 After the first episode of a video stream is encoded, the encoder configuration is **persisted into the dataset metadata** (`meta/info.json`) under each video feature, alongside the values probed from the file itself. For a video feature `observation.images.<camera>`, the layout in `info.json` is:
@@ -82,7 +217,7 @@ After the first episode of a video stream is encoded, the encoder configuration
         "video.pix_fmt": "yuv420p",
         "video.fps": 30,
         "video.channels": 3,
-        "video.is_depth_map": false,
+        "is_depth_map": false,
         "video.g": 2,
         "video.crf": 30,
         "video.preset": "fast",
@@ -97,15 +232,16 @@ After the first episode of a video stream is encoded, the encoder configuration
 
 Two sources contribute to the `info` block:
 
-- **Stream-derived** (read back from the encoded MP4 with PyAV): `video.height`, `video.width`, `video.codec`, `video.pix_fmt`, `video.fps`, `video.channels`, `video.is_depth_map`, plus `audio.*` if an audio stream is present.
-- **Encoder-derived** (taken from `VideoEncoderConfig`): `video.g`, `video.crf`, `video.preset`, `video.fast_decode`, `video.video_backend`, `video.extra_options`.
+| Source              | Where it comes from                                   | Fields                                                                                                                  |
+| ------------------- | ----------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------- |
+| **Stream-derived**  | Read back from the encoded MP4 with PyAV.             | `video.height`, `video.width`, `video.codec`, `video.pix_fmt`, `video.fps`, `video.channels`, `is_depth_map`, `audio.*` |
+| **Encoder-derived** | Taken from `RGBEncoderConfig` / `DepthEncoderConfig`. | `video.g`, `video.crf`, `video.preset`, `video.fast_decode`, `video.video_backend`, `video.extra_options`               |
 
-<Tip>
-  This block is populated **once**, from the **first** episode. It assumes every
-  episode in the dataset was encoded with the same `camera_encoder`. Changing
-  encoder settings partway through a recording is not supported — the
-  `info.json` will only reflect the parameters used for the first episode.
-</Tip>
+> [!IMPORTANT]
+> This block is populated **once**, from the **first** episode. It assumes every
+> episode in the dataset was encoded with the same `rgb_encoder`. Changing
+> encoder settings partway through a recording is not supported — the
+> `info.json` will only reflect the parameters used for the first episode.
 
 ---
 
@@ -113,5 +249,7 @@ Two sources contribute to the `info` block:
 
 When aggregating datasets with `merge_datasets`, video files are concatenated as-is (no re-encoding), and encoder fields in `info.json` are merged per-key:
 
-- **Stream-derived fields must match** across sources: `video.codec`, `video.pix_fmt`, `video.height`, `video.width`, `video.fps`. Otherwise FFmpeg's concat demuxer fails.
-- **Encoder-tuning fields are merged loosely**: `video.g`, `video.crf`, `video.preset`, `video.fast_decode`, `video.extra_options`. If every source agrees, the value is kept; if not, it's set to `null` (or `{}` for `video.extra_options`) and a warning is logged.
+| Merge rule         | Fields                                                                             | Behaviour                                                                                                                                                 |
+| ------------------ | ---------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| **Must match**     | `video.codec`, `video.pix_fmt`, `video.height`, `video.width`, `video.fps`         | Stream-derived fields must match across sources, otherwise FFmpeg's concat demuxer fails.                                                                 |
+| **Merged loosely** | `video.g`, `video.crf`, `video.preset`, `video.fast_decode`, `video.extra_options` | Encoder-tuning fields. If every source agrees, the value is kept; if not, it's set to `null` (or `{}` for `video.extra_options`) and a warning is logged. |

docs/source/vlabench.mdx

@@ -165,7 +165,7 @@ lerobot-train \
   --output_dir=./outputs/smolvla_vlabench_primitive \
   --steps=100000 \
   --batch_size=4 \
-  --eval_freq=5000 \
+  --env_eval_freq=5000 \
   --eval.batch_size=1 \
   --eval.n_episodes=1 \
   --save_freq=10000

examples/isaac_teleop_to_so101/README.md

@@ -0,0 +1,131 @@
+# Isaac Teleop → SO-101
+
+Teleoperate an SO-101/SO-100 follower arm — and record LeRobot datasets — with NVIDIA
+[Isaac Teleop](https://github.com/NVIDIA/IsaacTeleop). Two input devices ship today:
+
+- **XR (VR) controller** (`--teleop.type=xr_controller`) — the controller's grip pose drives the
+  end-effector through a squeeze-to-engage clutch and LeRobot's Cartesian IK pipeline; the analog
+  trigger drives the gripper.
+- **SO-101 leader arm** (`--teleop.type=so101_leader`) — a back-drivable leader arm mirrored 1:1
+  onto the follower via Isaac Teleop's native `so101_leader` plugin (no clutch, no IK).
+
+The full narrative guide (how the clutch works, CloudXR setup, headset pairing, tuning, and
+troubleshooting) is in the [LeRobot docs](https://huggingface.co/docs/lerobot/isaac_teleop)
+(source: `docs/source/isaac_teleop.mdx`). This README is the canonical install and usage
+reference.
+
+## Requirements
+
+- Linux workstation (see NVIDIA's
+  [system requirements](https://nvidia.github.io/IsaacTeleop/main/references/requirements.html)
+  for supported OS/GPU/headset combinations; `isaacteleop` publishes Linux wheels only).
+- An SO-101 (or SO-100) follower arm, calibrated with `lerobot-calibrate`.
+- For the XR device: a CloudXR-capable headset (e.g. Quest 3, Pico 4, Apple Vision Pro) on the
+  same network.
+- For the leader device: a second, back-drivable SO-101 leader arm and the `so101_leader` plugin
+  binary built from the Isaac Teleop source tree (see
+  [Build from source](https://nvidia.github.io/IsaacTeleop/main/getting_started/build_from_source/index.html)).
+
+## Installation
+
+This example lives in the LeRobot repository and is not part of the `lerobot` pip package, so
+work from a source checkout. From the repo root:
+
+```bash
+# LeRobot with the extras this example uses:
+#   feetech    - SO-101 serial motor bus
+#   kinematics - Placo IK solver (XR controller path)
+#   dataset    - dataset recording (record.py)
+# huggingface_hub >= 1.5 is needed by the automatic URDF fetch (Buckets API).
+uv pip install -e ".[feetech,kinematics,dataset]" "huggingface_hub>=1.5"
+
+# Isaac Teleop from public PyPI. `cloudxr` brings the CloudXR runtime bindings;
+# `retargeters-lite` is the scipy-based retargeter path that resolves on both
+# x86_64 and ARM (the full `retargeters` extra does not resolve on aarch64).
+uv pip install "isaacteleop[cloudxr,retargeters-lite]~=1.3.131" "scipy>=1.14"
+
+# Optional, x86_64 only: the full retargeter stack.
+uv pip install "isaacteleop[retargeters]~=1.3.131"
+```
+
+One-time CloudXR EULA (the auto-launch prompts on stdin and would hang on a headless machine):
+
+```bash
+python -m isaacteleop.cloudxr --accept-eula
+```
+
+## Usage
+
+Run everything from the repo root with `python -m` so the `examples` package resolves.
+
+### Teleoperate — XR controller
+
+```bash
+python -m examples.isaac_teleop_to_so101.teleoperate \
+    --robot.type=so101_follower \
+    --robot.port=/dev/ttyACM0 \
+    --robot.id=so101_follower_arm \
+    --teleop.type=xr_controller
+```
+
+On startup the script launches the CloudXR runtime (~30 s), prints the workstation IP to enter in
+the headset's CloudXR web client, waits for the controllers to stream, slews the arm to a reset
+pose (`--reset_to_origin=false` to skip), and then: **hold the squeeze/grip** to engage, move the
+controller to drive the arm, pull the trigger to close the gripper. Releasing the squeeze freezes
+the arm. The SO-101 URDF is fetched automatically from the `lerobot/robot-urdfs` Hugging Face
+bucket into the LeRobot cache on first run.
+
+To customize the reset pose: back-drive the arm to the pose you want, then
+
+```bash
+python -m examples.isaac_teleop_to_so101.override_reset_pose --port /dev/ttyACM0 --id so101_follower_arm
+```
+
+which writes it to `HF_LEROBOT_HOME/reset_poses/<robot.name>/<robot.id>.json`; runs with the same
+`--robot.id` use it automatically.
+
+### Teleoperate — SO-101 leader arm
+
+```bash
+python -m examples.isaac_teleop_to_so101.teleoperate \
+    --robot.type=so101_follower --robot.port=/dev/ttyACM0 --robot.id=so101_follower_arm \
+    --teleop.type=so101_leader --teleop.port=/dev/ttyACM1 --teleop.id=so101_leader_arm \
+    --launch_plugin=/path/to/IsaacTeleop/install/plugins/so101_leader/so101_leader_plugin
+```
+
+The follower is first slewed to the leader's pose over `--align_duration` seconds
+(`--align=false` to skip), then mirrors it 1:1. The plugin reuses the serial leader's calibration
+(`HF_LEROBOT_CALIBRATION/teleoperators/so_leader/<teleop.id>.json`).
+
+### Record a dataset
+
+`record.py` takes the same `--robot.*`/`--teleop.*`/loop flags plus `lerobot-record`-style
+`--dataset.*` flags:
+
+```bash
+python -m examples.isaac_teleop_to_so101.record \
+    --robot.type=so101_follower --robot.port=/dev/ttyACM0 --robot.id=so101_follower_arm \
+    --teleop.type=xr_controller \
+    --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
+    --dataset.repo_id=<hf_user>/<dataset_name> \
+    --dataset.single_task="Pick up the cube" \
+    --dataset.num_episodes=3 --dataset.episode_time_s=20 --dataset.reset_time_s=5
+```
+
+Keyboard shortcuts (terminal-first, so they work over SSH): **Right/n** end episode early,
+**Left/r** re-record, **Esc/q** stop after the current episode.
+
+Run either script with `--help` for all flags.
+
+## Layout
+
+```
+isaac_teleop/            device library: session lifecycle (base.py), XRController,
+                         SO101LeaderArm, Clutch, configs, and the XR→IK processor step
+common.py                shared loop infra: device bundles, clutch/IK pipeline wiring,
+                         reset/align slews, URDF fetch, keyboard listener
+teleoperate.py           teleoperation CLI (device selected via --teleop.type)
+record.py                dataset-recording CLI (same device selection + --dataset.*)
+override_reset_pose.py   save the current joints as the per-arm reset pose
+default.env              CloudXR device-profile overrides passed to the launcher
+```

examples/isaac_teleop_to_so101/__init__.py

@@ -0,0 +1,17 @@
+#!/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.
+
+"""Isaac Teleop -> SO-101 example package."""

examples/isaac_teleop_to_so101/common.py

@@ -0,0 +1,650 @@
+#!/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.
+
+"""Shared device + control-loop infrastructure for the Isaac Teleop -> SO-101 examples.
+
+Consumed by ``teleoperate.py`` and ``record.py``, which both build a per-device
+:class:`Device` bundle and run the same loop: read -> (maybe command) -> hold-when-idle ->
+sleep. A :class:`Device` bundles three closures: ``compute(obs) -> RobotAction | None``
+(``None`` = hold at the measured pose while idle), ``startup``, and ``cleanup``. The devices:
+
+* ``xr_controller`` — a thin :class:`XRController` whose raw grip pose an in-loop
+  :class:`Clutch` turns into an EE target for LeRobot's Cartesian IK pipeline.
+* ``so101_leader`` — a back-drivable leader arm mirrored 1:1 into the follower.
+
+Requires the ``isaacteleop`` package and an OpenXR runtime (install instructions in this
+folder's ``README.md``). User-facing guide: ``docs/source/isaac_teleop.mdx``.
+"""
+
+import json
+import logging
+import socket
+import subprocess
+import sys
+import time
+from collections.abc import Callable
+from contextlib import suppress
+from dataclasses import dataclass
+from importlib.resources import files
+from pathlib import Path
+from typing import Protocol
+
+import numpy as np
+
+from lerobot.model.kinematics import RobotKinematics
+from lerobot.processor import (
+    RobotProcessorPipeline,
+    robot_action_observation_to_transition,
+    transition_to_robot_action,
+)
+from lerobot.robots import RobotConfig, make_robot_from_config
+from lerobot.robots.so_follower import SOFollowerConfig  # noqa: F401  (registers so101_follower)
+from lerobot.robots.so_follower.robot_kinematic_processor import (
+    EEBoundsAndSafety,
+    InverseKinematicsEEToJoints,
+)
+from lerobot.types import RobotAction, RobotObservation
+from lerobot.utils.constants import HF_LEROBOT_CALIBRATION, HF_LEROBOT_HOME, TELEOPERATORS
+from lerobot.utils.robot_utils import precise_sleep
+
+from .isaac_teleop import (
+    Clutch,
+    IsaacTeleopConfig,
+    MapXRControllerActionToRobotAction,
+    SO101LeaderArm,
+    SO101LeaderArmConfig,
+    XRController,
+)
+
+# Fixed rate [Hz] for the teleoperate loop and the pre-loop slews / connect-wait poll sleeps.
+FPS = 30
+
+# CloudXR device-profile env file passed to the launcher (see default.env in this package).
+CLOUDXR_ENV_FILE = str(files(__package__) / "default.env")
+
+
+class LoopConfig(Protocol):
+    """Structural type for the loop/launch knobs ``build_device`` and the ``setup_*`` read.
+
+    Both ``TeleoperateConfig`` and ``RecordConfig`` satisfy it, keeping ``common`` decoupled
+    from either entry point's concrete config.
+    """
+
+    teleop: IsaacTeleopConfig
+    robot: RobotConfig
+    launch_plugin: str | None
+    reset_to_origin: bool
+    reset_duration: float
+    align: bool
+    align_duration: float
+
+
+# Per-device bundle consumed by the shared loop. ``compute`` returns None to mean
+# "idle -> hold at the measured pose"; ``startup`` warms up; ``cleanup`` reaps/disconnects.
+@dataclass(frozen=True)
+class Device:
+    compute: Callable[[RobotObservation | None], RobotAction | None]
+    startup: Callable[[], None]
+    cleanup: Callable[[], None]
+
+
+def hold_action(obs: RobotObservation, motor_names: list[str]) -> dict[str, float]:
+    """Re-send the measured joints — the explicit hold when a device is idle."""
+    return {f"{name}.pos": float(obs[f"{name}.pos"]) for name in motor_names}
+
+
+class HoldLatch:
+    """Resolve the per-frame action, holding one LATCHED pose while the device is idle.
+
+    Re-sending the freshly measured joints on every idle frame would ratchet the arm
+    downward: under gravity the P-only servo settles below its goal by a steady-state
+    error, so each re-command of the measurement lowers the goal by that error again.
+    Latching the target once on the active->idle transition holds a fixed pose instead.
+    """
+
+    def __init__(self, motor_names: list[str]):
+        self._motor_names = motor_names
+        self._held: dict[str, float] | None = None
+
+    def resolve(self, action: RobotAction | None, obs: RobotObservation) -> RobotAction:
+        """Pass through an active action (clearing the latch); latch + hold when idle."""
+        if action is not None:
+            self._held = None
+            return action
+        if self._held is None:
+            self._held = hold_action(obs, self._motor_names)
+        return self._held
+
+
+def slew(
+    robot,
+    motor_names: list[str],
+    target_fn: Callable[[], dict[str, float]],
+    duration_s: float,
+) -> None:
+    """Linearly slew all joints from their current measured pose toward a target.
+
+    ``target_fn`` is called EACH step, so the leader can pass a live re-read (landing on its
+    current pose at ``alpha == 1`` for a continuous handoff) while XR passes a constant.
+    """
+    obs = robot.get_observation()
+    start = {name: float(obs[f"{name}.pos"]) for name in motor_names}
+    n_steps = max(1, int(duration_s * FPS))
+    for step in range(1, n_steps + 1):
+        alpha = step / n_steps
+        target = target_fn()
+        action = {f"{name}.pos": start[name] + alpha * (target[name] - start[name]) for name in motor_names}
+        robot.send_action(action)
+        precise_sleep(1.0 / FPS)
+
+
+# ============================================================================
+# XR controller device
+# ============================================================================
+
+# Per-frame EE rate limit [m]. With raise_on_jump=False, EEBoundsAndSafety clamps an
+# over-limit step instead of raising, absorbing a tracking glitch as one slow frame. At
+# FPS=30, 0.1 m/frame caps EE speed at ~3 m/s. (end_effector_bounds clips the absolute target.)
+MAX_EE_STEP_M = 0.1
+
+# Soft-orientation IK weight: small but nonzero so the wrist follows the hand while position
+# dominates (the 5-DOF SO-101 cannot realize an arbitrary orientation). 0.0 = position-only.
+IK_ORIENTATION_WEIGHT = 0.01
+
+
+def _ensure_so101_urdf() -> str:
+    """Return the cached SO-101 URDF path, fetching the ``so101`` folder (URDF + meshes) from
+    the public ``lerobot/robot-urdfs`` HF bucket into the LeRobot cache on first use."""
+    dest_dir = HF_LEROBOT_HOME / "robot-urdfs" / "so101"
+    urdf_path = dest_dir / "so101_new_calib.urdf"
+    # Completeness marker written only after a FULL sync: the URDF file alone is not a
+    # completeness signal (an interrupted first sync can leave the meshes it references
+    # missing, which the URDF's mere existence would then hide forever). Re-syncing is
+    # idempotent and repairs a partial cache; delete the folder to force a re-download.
+    marker = dest_dir / ".sync_complete"
+    if not marker.exists():
+        from huggingface_hub import sync_bucket
+
+        sync_bucket("hf://buckets/lerobot/robot-urdfs/so101", str(dest_dir), quiet=True)
+        marker.touch()
+    return str(urdf_path)
+
+
+# Default duration [s] for the startup reset-to-origin slew.
+RESET_DURATION_S = 5.0
+
+# Optional cached file written by override_reset_pose.py. When present it takes priority over RESET_ORIGIN_DEG.
+RESET_POSE_FILE = str(HF_LEROBOT_HOME / "reset_poses" / "{robot_name}" / "{robot_id}.json")
+
+# Reset target in each motor's native units (arm joints in degrees, gripper RANGE_0_100,
+# 100 = open). An empirically comfortable pose (elbow/wrist bent) avoiding the singularity of
+# a fully-extended arm; assumes standard calibration. Override per-arm via override_reset_pose.py.
+RESET_ORIGIN_DEG: dict[str, float] = {
+    "shoulder_pan": -4.0,
+    "shoulder_lift": -103.0,
+    "elbow_flex": 97.0,
+    "wrist_flex": 78.0,
+    "wrist_roll": -65.0,
+    "gripper": 0.0,
+}
+
+
+def _load_reset_target(reset_pose_file: Path, motor_names: list[str]) -> dict[str, float]:
+    """Return reset targets: the saved reset pose if present, else RESET_ORIGIN_DEG."""
+    if reset_pose_file.exists():
+        saved = json.loads(reset_pose_file.read_text())
+        # Fill any missing motors from the fallback dict.
+        return {name: float(saved.get(name, RESET_ORIGIN_DEG.get(name, 0.0))) for name in motor_names}
+    return {name: RESET_ORIGIN_DEG.get(name, 0.0) for name in motor_names}
+
+
+# CloudXR web client URL opened in the headset (Isaac Teleop quick start, step 5).
+_CLOUDXR_WEB_CLIENT_URL = "https://nvidia.github.io/IsaacTeleop/client"
+# WSS-proxy / self-signed-cert port the operator accepts in-browser before connecting.
+_CLOUDXR_WSS_PORT = 48322
+# How often to re-print the connection hint while waiting for the headset [s].
+_XR_CONNECT_REMINDER_S = 15.0
+# Virtual / bridge / USB-gadget interfaces a headset can't reach over the network — skip
+# by name prefix (``docker0``, compose ``br-*``, ``veth*``, libvirt ``virbr*``, and the
+# Tegra USB device-mode bridge ``l4tbr0``).
+_SKIP_IFACE_PREFIXES = ("docker", "br-", "veth", "virbr", "l4tbr")
+
+
+def _primary_ipv4() -> str | None:
+    """The workstation's primary outbound IPv4, via the UDP-socket trick (``connect()`` on a
+    datagram socket selects the egress interface without sending packets)."""
+    with socket.socket(socket.AF_INET, socket.SOCK_DGRAM) as s:
+        try:
+            s.connect(("8.8.8.8", 80))
+            return s.getsockname()[0]
+        except OSError:
+            return None
+
+
+def _candidate_ipv4s() -> list[tuple[str, str]]:
+    """Return ``[(interface, ipv4), ...]`` the headset might reach this workstation at.
+
+    Lists each interface's IPv4 via ``psutil`` (dropping loopback, link-local, and the
+    virtual/bridge interfaces in ``_SKIP_IFACE_PREFIXES``), primary outbound first. Falls
+    back to just the primary IP when ``psutil`` is unavailable.
+    """
+    primary = _primary_ipv4()
+    found: list[tuple[str, str]] = []
+    try:
+        import psutil
+
+        for iface, addrs in psutil.net_if_addrs().items():
+            if iface.startswith(_SKIP_IFACE_PREFIXES):
+                continue
+            for addr in addrs:
+                if addr.family != socket.AF_INET:
+                    continue
+                ip = addr.address
+                if ip.startswith("127.") or ip.startswith("169.254."):
+                    continue
+                found.append((iface, ip))
+    except Exception:
+        if primary:
+            found.append(("default", primary))
+    found.sort(key=lambda t: t[1] != primary)  # primary outbound interface first
+    return found
+
+
+def _print_xr_connect_help() -> None:
+    """Print how to connect the headset to this workstation over CloudXR."""
+    ips = _candidate_ipv4s()
+    print("\n" + "=" * 76)
+    print("Connect your XR headset to this workstation over NVIDIA CloudXR:")
+    print(f"  1. In the headset, open the CloudXR web client:  {_CLOUDXR_WEB_CLIENT_URL}")
+    print("  2. Enter this workstation's IP address:")
+    if ips:
+        for iface, ip in ips:
+            print(f"        {ip:<15}  ({iface})")
+        if len(ips) > 1:
+            print("     (use the address on the same network as your headset)")
+    else:
+        print("        <could not determine — check `hostname -I` / `ip addr`>")
+    print(f"  3. Accept the self-signed cert at https://<that-ip>:{_CLOUDXR_WSS_PORT}/ , then Connect.")
+    print("=" * 76 + "\n")
+
+
+def _wait_for_xr_controller(teleop_device: XRController) -> None:
+    """Block until the XR controller is tracked, polling ``get_action()`` and re-printing a
+    reminder every ``_XR_CONNECT_REMINDER_S``. User-paced; ``Ctrl-C`` aborts (no hard timeout).
+    """
+    _print_xr_connect_help()
+    print("Waiting for the headset controllers to start streaming…  (Ctrl-C to abort)")
+    last_reminder = time.time()
+    while True:
+        teleop_device.get_action()  # steps the session; updates is_tracking
+        if teleop_device.is_tracking:
+            print("Headset connected — controllers are streaming.")
+            return
+        if time.time() - last_reminder >= _XR_CONNECT_REMINDER_S:
+            print("…still waiting for the headset to connect (Ctrl-C to abort).")
+            last_reminder = time.time()
+        time.sleep(1.0 / FPS)
+
+
+def setup_xr(cfg: LoopConfig, robot, motor_names: list[str]) -> Device:
+    """Build the XR controller device bundle (clutch + soft-orientation IK pipeline)."""
+    kinematics_solver = RobotKinematics(
+        urdf_path=_ensure_so101_urdf(),
+        target_frame_name="gripper_frame_link",
+        joint_names=motor_names,
+    )
+
+    teleop_config = cfg.teleop  # XRControllerConfig (selected via --teleop.type=xr_controller)
+    teleop_device = XRController(teleop_config)
+
+    # The clutch (below) turns the raw grip pose into an absolute base-frame ee_pose; this
+    # pipeline maps it to joint targets: rename -> bounds/rate-limit -> IK.
+    xr_to_robot_joints_processor = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+        steps=[
+            MapXRControllerActionToRobotAction(),
+            # raise_on_jump=False: an over-limit step (e.g. a tracking glitch) is clamped +
+            # warned instead of raised, since a crash mid-loop would leave the arm uncontrolled.
+            # z floor 0.0 keeps a stray target above the table; x/y stay at a loose [-1,1]m box.
+            EEBoundsAndSafety(
+                end_effector_bounds={"min": [-1.0, -1.0, 0.0], "max": [1.0, 1.0, 1.0]},
+                max_ee_step_m=MAX_EE_STEP_M,
+                raise_on_jump=False,
+            ),
+            # initial_guess_current_joints=False: warm-start from the previous IK solution so
+            # the joint trajectory stays continuous frame-to-frame.
+            InverseKinematicsEEToJoints(
+                kinematics=kinematics_solver,
+                motor_names=motor_names,
+                initial_guess_current_joints=False,
+                orientation_weight=IK_ORIENTATION_WEIGHT,
+            ),
+        ],
+        to_transition=robot_action_observation_to_transition,
+        to_output=transition_to_robot_action,
+    )
+
+    # The clutch is built in startup() (after the optional reset slew, seeded from the
+    # post-slew MEASURED pose) and shared with compute() via nonlocal.
+    clutch: Clutch | None = None
+    prev_enabled = False
+
+    def startup() -> None:
+        nonlocal clutch
+        # Connect and wait for the operator to don the headset BEFORE moving the arm, so the
+        # reset slew happens while they are watching in VR.
+        teleop_device.connect()
+        if not teleop_device.is_connected:
+            raise ValueError("Teleop is not connected!")
+        _wait_for_xr_controller(teleop_device)
+
+        if cfg.reset_to_origin:
+            reset_pose_file = Path(RESET_POSE_FILE.format(robot_name=robot.name, robot_id=robot.id))
+            target = _load_reset_target(reset_pose_file, motor_names)
+            source = str(reset_pose_file) if reset_pose_file.exists() else "hardcoded defaults"
+            print(f"Reset target source: {source}")
+            print(f"Resetting to origin over {cfg.reset_duration:.1f} s…")
+            slew(robot, motor_names, lambda: target, cfg.reset_duration)
+            print("Reset complete.")
+
+        # Seed the clutch home from the arm's measured pose (FK of the current joints) so the
+        # first engage is jump-free, whether or not a reset slew ran.
+        obs0 = robot.get_observation()
+        q_measured_deg = np.array([float(obs0[f"{name}.pos"]) for name in motor_names], dtype=float)
+        home_base_T_ee = kinematics_solver.forward_kinematics(q_measured_deg)  # noqa: N806
+        clutch = Clutch(home_base_T_ee)
+
+        print("Starting teleop loop. Squeeze and move the controller to teleoperate the robot...")
+
+    def compute(robot_obs: RobotObservation | None) -> RobotAction | None:
+        nonlocal prev_enabled
+        if clutch is None:  # set in startup(), which runs before compute()
+            raise RuntimeError("compute() called before startup(); the clutch is not initialized")
+        xr_action = teleop_device.get_action()
+        grip_pos = np.asarray(xr_action["grip_pos"], dtype=float)
+        grip_quat = np.asarray(xr_action["grip_quat"], dtype=float)
+        squeeze = float(xr_action["squeeze"])
+        trigger = float(xr_action["trigger"])
+        enabled = squeeze > teleop_config.clutch_threshold
+
+        # On the engage edge, latch the clutch home at the arm's MEASURED EE pose (FK of
+        # the live joints) and the controller origin so the per-frame delta starts at zero.
+        # Latching the last commanded pose instead would snap the arm back to it at full
+        # servo speed if the arm moved while disengaged (gravity sag, external contact).
+        is_engage_frame = enabled and not prev_enabled
+        if is_engage_frame:
+            q_measured = np.array([float(robot_obs[f"{name}.pos"]) for name in motor_names], dtype=float)
+            measured_base_T_ee = kinematics_solver.forward_kinematics(q_measured)  # noqa: N806
+            clutch.engage(grip_pos, grip_quat, measured_base_T_ee=measured_base_T_ee)
+            # Re-anchor the pipeline state at the measured pose as well: EEBoundsAndSafety's
+            # rate limiter and the IK warm start otherwise still reference the stale
+            # pre-disengage command and would fight the fresh home for several frames.
+            xr_to_robot_joints_processor.reset()
+        prev_enabled = enabled
+
+        # SAFETY GATE: command the robot ONLY while the clutch is engaged; otherwise return
+        # None so the loop holds the measured joints (releasing the clutch freezes the arm).
+        if not enabled:
+            return None
+
+        # Rebase the raw grip pose onto the EE, then run the pipeline. closedness = trigger.
+        ee_pos, ee_quat = clutch.rebase(grip_pos, grip_quat)
+        ee_action = {
+            "ee_pose": np.concatenate([ee_pos, ee_quat]).astype(np.float32),
+            "closedness": trigger,
+        }
+        return xr_to_robot_joints_processor((ee_action, robot_obs))
+
+    return Device(compute=compute, startup=startup, cleanup=teleop_device.disconnect)
+
+
+# ============================================================================
+# SO-101 leader arm device
+# ============================================================================
+
+# Default duration [s] for the startup alignment slew (follower current -> leader first pose).
+ALIGN_DURATION_S = 3.0
+
+# How long to wait for the leader plugin to start streaming before aligning / looping.
+LEADER_WARMUP_TIMEOUT_S = 20.0
+
+# The plugin converts the leader's servo ticks to radians, so it reuses the serial SO-101
+# leader's calibration, stored by lerobot-calibrate under SO101Leader.name == "so_leader".
+SO_LEADER_CALIBRATION_NAME = "so_leader"
+
+
+def _leader_calibration_path(cfg: LoopConfig) -> Path | None:
+    """Infer the calibration JSON the launched plugin should read, or None.
+
+    Path convention: ``HF_LEROBOT_CALIBRATION / teleoperators / so_leader / {--teleop.id}.json``
+    (or ``--teleop.calibration_dir`` if set). Returns None (plugin falls back to defaults) when
+    it does not exist, warning if an id was given, or when no ``--teleop.id`` is set.
+    """
+    if not cfg.teleop.id:
+        return None
+    calib_dir = cfg.teleop.calibration_dir or (
+        HF_LEROBOT_CALIBRATION / TELEOPERATORS / SO_LEADER_CALIBRATION_NAME
+    )
+    calib_path = Path(calib_dir) / f"{cfg.teleop.id}.json"
+    if calib_path.is_file():
+        return calib_path
+    print(
+        f"WARNING: no leader calibration at {calib_path}; the plugin will use built-in defaults. "
+        f"Calibrate with the serial leader (`lerobot-calibrate --teleop.type=so101_leader "
+        f"--teleop.id={cfg.teleop.id}`) or the plugin's `calibrate` subcommand."
+    )
+    return None
+
+
+def _wait_for_leader(teleop: SO101LeaderArm, timeout_s: float) -> dict[str, float]:
+    """Poll the leader until it streams a live frame; return that frame's ``{joint}.pos``.
+
+    Raises ``SystemExit`` if no live frame arrives within ``timeout_s`` (plugin not pushing,
+    wrong ``--teleop.collection_id``, or CloudXR not up).
+    """
+    print(f"Waiting up to {timeout_s:.0f}s for the so101_leader plugin to stream…")
+    deadline = time.time() + timeout_s
+    while time.time() < deadline:
+        action = teleop.get_action()
+        if teleop.is_tracking:
+            print("Leader is streaming.")
+            return action
+        time.sleep(1.0 / FPS)
+    raise SystemExit(
+        f"FAILED: leader did not stream within {timeout_s:.0f}s. Is the so101_leader plugin "
+        "running and pushing (check --teleop.collection_id)? Is CloudXR up?"
+    )
+
+
+def _maybe_launch_plugin(cfg: LoopConfig) -> subprocess.Popen | None:
+    """Spawn the so101_leader plugin if ``--launch_plugin <path>`` was given (after connect())."""
+    if cfg.launch_plugin is None:
+        return None
+    if not Path(cfg.launch_plugin).exists():
+        raise SystemExit(
+            f"plugin binary not found: {cfg.launch_plugin} (build it in the IsaacTeleop repo first)"
+        )
+    leader_port = cfg.teleop.port  # SO101LeaderArmConfig.port, forwarded to the plugin
+    backend = f"leader on {leader_port}" if leader_port else "synthetic trajectory"
+    print(f"launching plugin: {cfg.launch_plugin} ({backend})")
+    # Positional args: [device_path] [collection_id] [calibration_file]. Empty device_path ->
+    # synthetic backend. Calibration (only real hardware needs it) is appended when a port is set.
+    argv = [cfg.launch_plugin, leader_port, cfg.teleop.collection_id]
+    if leader_port:
+        calib_path = _leader_calibration_path(cfg)
+        if calib_path is not None:
+            argv.append(str(calib_path))
+            print(f"  leader calibration: {calib_path}")
+    # Spawned after connect() so it inherits the CloudXR runtime env (XR_RUNTIME_JSON, ...).
+    proc = subprocess.Popen(argv)
+    time.sleep(1.5)  # let it create its OpenXR session and start pushing
+    return proc
+
+
+def setup_leader(cfg: LoopConfig, robot, motor_names: list[str]) -> Device:
+    """Build the SO-101 leader arm device bundle (1:1 joint mirror)."""
+    teleop_config = cfg.teleop  # SO101LeaderArmConfig (selected via --teleop.type=so101_leader)
+    teleop = SO101LeaderArm(teleop_config)
+
+    plugin_proc: subprocess.Popen | None = None
+
+    def startup() -> None:
+        nonlocal plugin_proc
+        # connect() auto-launches CloudXR (unless opted out); spawn the plugin AFTER so it
+        # inherits the runtime env. The plugin is reaped in cleanup().
+        teleop.connect()
+        plugin_proc = _maybe_launch_plugin(cfg)
+
+        if not teleop.is_connected:
+            raise ValueError("Teleop is not connected!")
+
+        # Block until the leader streams a live frame (clear error if it never does).
+        _wait_for_leader(teleop, LEADER_WARMUP_TIMEOUT_S)
+
+        if cfg.align:
+            print(f"Aligning follower to leader over {cfg.align_duration:.1f}s…")
+
+            # Re-read the live leader pose once per step so alpha=1 lands on its current pose
+            # from a single coherent frame.
+            def _leader_target() -> dict[str, float]:
+                leader_now = teleop.get_action()
+                return {name: float(leader_now[f"{name}.pos"]) for name in motor_names}
+
+            slew(robot, motor_names, _leader_target, cfg.align_duration)
+            print("Alignment complete.")
+
+        print(
+            "Starting joint-mirror loop. Back-drive the leader to teleoperate the follower… (Ctrl-C to stop)"
+        )
+
+    def compute(robot_obs: RobotObservation | None) -> RobotAction | None:
+        leader_action = teleop.get_action()
+        # Hold the follower at its measured pose when the leader drops out (stale stream)
+        # rather than commanding a possibly-old target.
+        if not teleop.is_tracking:
+            return None
+        return leader_action
+
+    def cleanup() -> None:
+        # A plugin-reaping failure must not skip the session disconnect (and vice versa
+        # the disconnect runs after the plugin stops pushing on it).
+        try:
+            if plugin_proc is not None:
+                plugin_proc.terminate()
+                try:
+                    plugin_proc.wait(timeout=5)
+                except subprocess.TimeoutExpired:
+                    plugin_proc.kill()
+        finally:
+            teleop.disconnect()
+
+    return Device(compute=compute, startup=startup, cleanup=cleanup)
+
+
+# ============================================================================
+# Shared setup
+# ============================================================================
+
+
+def build_device(cfg: LoopConfig) -> tuple:
+    """Connect the follower, build the selected Isaac device, and run its pre-loop startup.
+
+    Connects the follower FIRST (so the startup slew / clutch-home seed can read live joints),
+    dispatches on ``--teleop.type``, then runs ``device.startup()`` before returning. On any
+    failure after ``connect()`` the follower is disconnected so the connection never leaks.
+
+    Returns ``(robot, device, motor_names)``.
+    """
+    # Default the CloudXR input profile to this example's default.env unless the user overrode
+    # it via --teleop.cloudxr_env_file.
+    if cfg.teleop.cloudxr_env_file is None:
+        cfg.teleop.cloudxr_env_file = CLOUDXR_ENV_FILE
+
+    # SO-101/SO-100 only (both share the SO-101 URDF), reject other followers.
+    supported_robots = {"so101_follower", "so100_follower"}
+    if cfg.robot.type not in supported_robots:
+        raise ValueError(
+            f"This example only supports SO-101/SO-100 followers ({sorted(supported_robots)}), "
+            f"but got --robot.type={cfg.robot.type}."
+        )
+
+    # The degree-based pipeline relies on --robot.use_degrees (default True).
+    robot = make_robot_from_config(cfg.robot)
+    # Connect FIRST so the startup slew and clutch-home seed can read live joints.
+    robot.connect()
+    # Everything after connect() can fail; this runs outside the callers' try/finally, so
+    # disconnect the follower on any failure to avoid leaking the connection.
+    device: Device | None = None
+    try:
+        # Joint names in action order, read from {name}.pos action features (robot-agnostic).
+        motor_names = [key.removesuffix(".pos") for key in robot.action_features if key.endswith(".pos")]
+
+        if isinstance(cfg.teleop, SO101LeaderArmConfig):
+            device = setup_leader(cfg, robot, motor_names)
+        else:
+            device = setup_xr(cfg, robot, motor_names)
+
+        device.startup()
+    except BaseException:
+        # Reap a partially-started device, then always disconnect the follower.
+        if device is not None:
+            with suppress(Exception):
+                device.cleanup()
+        robot.disconnect()
+        raise
+
+    return robot, device, motor_names
+
+
+# ============================================================================
+# Keyboard control
+# ============================================================================
+
+
+def init_keyboard_listener():
+    """Recording shortcuts, terminal-first so they work over SSH.
+
+    Whenever stdin is a TTY we use the stdlib :class:`TerminalKeyListener` directly rather
+    than upstream's pynput-first :func:`init_keyboard_listener`, whose global listener would
+    capture the workstation console instead of this (often SSH) terminal. With no TTY we defer
+    to upstream (pynput on a GUI, else headless no-op).
+    """
+    if not (sys.stdin is not None and sys.stdin.isatty()):
+        from lerobot.utils.keyboard_input import init_keyboard_listener as _upstream
+
+        return _upstream()
+
+    from lerobot.utils.keyboard_input import TerminalKeyListener, apply_recording_control
+
+    events = {"exit_early": False, "rerecord_episode": False, "stop_recording": False}
+
+    # n/r/q are the arrow/Esc equivalents that survive escape-sequence splitting over laggy
+    # SSH/VNC links. Case-insensitive so Shift+letter still works.
+    def on_key(name: str) -> None:
+        key = name.lower()
+        if key in ("right", "n"):
+            apply_recording_control("right", events)
+        elif key in ("left", "r"):
+            apply_recording_control("left", events)
+        elif key in ("esc", "q"):
+            apply_recording_control("esc", events)
+
+    listener = TerminalKeyListener(on_key)
+    listener.start()
+    logging.info(
+        "Keyboard control via terminal — keep this terminal focused: "
+        "Right/n = end episode early, Left/r = re-record, Esc/q = stop."
+    )
+    return listener, events

examples/isaac_teleop_to_so101/default.env

@@ -0,0 +1,21 @@
+# CloudXR device-profile overrides for the Isaac Teleop XR -> SO-101 example.
+#
+# Passed to isaacteleop's CloudXRLauncher as `env_config` (via
+# XRControllerConfig.cloudxr_env_file). Format: KEY=value, one per line; `#`
+# comments and blank lines ignored; $VARS / ~ expanded. See
+# isaacteleop/cloudxr/env_config.py::_load_env_file.
+#
+# Runtime-resolved keys (XR_RUNTIME_JSON, XRT_NO_STDIN, NV_CXR_RUNTIME_DIR,
+# NV_CXR_OUTPUT_DIR) are reserved and ignored if set here.
+
+# Transport profile the runtime advertises (CloudXR default: auto-webrtc).
+# "Quest3" also covers the Pico 4. Other values: auto-native, AppleVisionPro.
+NV_DEVICE_PROFILE=Quest3
+
+# Input device discovery channels (both default to true; pinned for clarity).
+NV_CXR_ENABLE_PUSH_DEVICES=true
+NV_CXR_ENABLE_TENSOR_DATA=true
+
+# Runtime logs to ~/.cloudxr/logs — helps debug connection issues
+# (e.g. "Failed to get OpenXR system: -35").
+NV_CXR_FILE_LOGGING=true

examples/isaac_teleop_to_so101/isaac_teleop/__init__.py

@@ -0,0 +1,40 @@
+#!/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.
+
+"""NVIDIA Isaac Teleop teleoperators for LeRobot.
+
+Each input device is an :class:`IsaacTeleopTeleoperator` subclass: :class:`XRController`
+(XR/VR controller) and :class:`SO101LeaderArm` (back-drivable SO-101 leader arm) ship today.
+"""
+
+from .base import IsaacTeleopTeleoperator
+from .clutch import Clutch
+from .config_isaac_teleop import IsaacTeleopConfig, SO101LeaderArmConfig, XRControllerConfig
+from .teleop_so101_leader_arm import SO101LeaderArm, leader_joints_to_robot_action
+from .teleop_xr_controller import XRController
+from .xr_controller_processor import MapXRControllerActionToRobotAction
+
+__all__ = [
+    "Clutch",
+    "IsaacTeleopConfig",
+    "IsaacTeleopTeleoperator",
+    "MapXRControllerActionToRobotAction",
+    "SO101LeaderArm",
+    "SO101LeaderArmConfig",
+    "XRController",
+    "XRControllerConfig",
+    "leader_joints_to_robot_action",
+]

examples/isaac_teleop_to_so101/isaac_teleop/base.py

@@ -0,0 +1,282 @@
+#!/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.
+
+"""Shared base for NVIDIA Isaac Teleop-backed LeRobot teleoperators.
+
+Isaac Teleop is a multi-modal framework: a single ``TeleopSession`` can be driven by
+XR controllers, hand tracking, Manus gloves, etc. Each modality is a
+:class:`Teleoperator` subclass in its own ``teleop_<device>.py``.
+
+:class:`IsaacTeleopTeleoperator` owns what those devices share — the session
+lifecycle, the per-step staleness/worker-health guard, and the no-op calibration
+tracking devices need. A concrete device implements :meth:`_build_pipeline` (its
+retargeting graph) and :meth:`get_action` (usually via :meth:`_step`).
+
+``isaacteleop`` is an optional NVIDIA dependency (install instructions in the example's
+``README.md``); its imports are guarded behind an availability check at module top, so this
+module imports without it and constructing a device fails fast with install instructions.
+"""
+
+from __future__ import annotations
+
+import abc
+import logging
+import os
+from collections.abc import Mapping
+from pathlib import Path
+from typing import TYPE_CHECKING, Any
+
+from lerobot.teleoperators.teleoperator import Teleoperator
+from lerobot.utils.import_utils import is_package_available
+
+from .config_isaac_teleop import IsaacTeleopConfig
+
+_isaacteleop_available = is_package_available("isaacteleop")
+
+if TYPE_CHECKING or _isaacteleop_available:
+    from isaacteleop.cloudxr import CloudXRLauncher
+    from isaacteleop.retargeting_engine.interface import (
+        ExecutionEvents,
+        ExecutionState,
+        GraphExecutable,
+        RetargeterIO,
+    )
+    from isaacteleop.teleop_session_manager import TeleopSession, TeleopSessionConfig
+else:
+    CloudXRLauncher = None
+    ExecutionEvents = None
+    ExecutionState = None
+    GraphExecutable = None
+    RetargeterIO = None
+    TeleopSession = None
+    TeleopSessionConfig = None
+
+logger = logging.getLogger(__name__)
+
+# Gripper closedness [0, 1] -> SO-101 follower motor units [0, 100] (RANGE_0_100, 100 = OPEN).
+# Shared by the XR processor and leader device, which invert via ``pos = (1 - c) * SCALE``.
+_GRIPPER_MOTOR_SCALE = 100.0
+
+
+def _require_isaacteleop() -> None:
+    """Fail fast with install pointers when the optional ``isaacteleop`` package is missing."""
+    if not _isaacteleop_available:
+        raise ImportError(
+            "The 'isaacteleop' package is required for Isaac Teleop devices but is not "
+            "installed. See examples/isaac_teleop_to_so101/README.md for install instructions."
+        )
+
+
+class IsaacTeleopTeleoperator(Teleoperator):
+    """Abstract base for teleoperators backed by an Isaac Teleop ``TeleopSession``.
+
+    Owns the session lifecycle and the per-step health guard; subclasses supply
+    :meth:`_build_pipeline` and :meth:`get_action`.
+    """
+
+    config_class = IsaacTeleopConfig
+
+    def __init__(self, config: IsaacTeleopConfig):
+        _require_isaacteleop()
+        super().__init__(config)
+        self.config: IsaacTeleopConfig = config
+        self._session: TeleopSession | None = None
+        self._cloudxr_launcher: CloudXRLauncher | None = None
+
+    # ------------------------------------------------------------------
+    # Pipeline construction (device override point)
+    # ------------------------------------------------------------------
+
+    @abc.abstractmethod
+    def _build_pipeline(self) -> GraphExecutable:
+        """Build this device's retargeting pipeline (the ``GraphExecutable`` for
+        ``TeleopSessionConfig.pipeline``). Called once in :meth:`connect`; its output
+        keys must match what :meth:`get_action` unpacks.
+        """
+        raise NotImplementedError
+
+    # ------------------------------------------------------------------
+    # Lifecycle (shared)
+    # ------------------------------------------------------------------
+
+    @property
+    def is_connected(self) -> bool:
+        return self._session is not None
+
+    @property
+    def is_calibrated(self) -> bool:
+        return True  # Tracking devices are self-calibrating.
+
+    def calibrate(self) -> None:
+        pass
+
+    def configure(self) -> None:
+        pass
+
+    def connect(self, calibrate: bool = True) -> None:
+        """Auto-launch the CloudXR runtime (unless opted out) and open the session.
+
+        The CloudXR launch blocks ~30s and, on the first run, prompts on stdin for the
+        EULA (accept once via ``python -m isaacteleop.cloudxr --accept-eula``). Opt out
+        when CloudXR runs externally via ``config.auto_launch_cloudxr=False`` or
+        ``LEROBOT_CLOUDXR_SKIP_AUTOLAUNCH=1`` (env var wins).
+        """
+        if self._session is not None:
+            raise RuntimeError("Already connected. Call disconnect() first.")
+
+        self._ensure_cloudxr_runtime()
+
+        try:
+            pipeline = self._build_pipeline()
+            session_config = TeleopSessionConfig(app_name=self.config.app_name, pipeline=pipeline)
+            self._session = TeleopSession(session_config)
+            self._session.__enter__()
+        except Exception:
+            self._session = None
+            try:
+                self._stop_cloudxr_runtime()
+            except Exception:
+                logger.exception("Failed to stop CloudXR runtime during connect() rollback")
+            raise
+        logger.info("Isaac Teleop session started: %s", self.config.app_name)
+
+    def disconnect(self) -> None:
+        try:
+            if self._session is not None:
+                # Null the handle BEFORE __exit__: even a failed session teardown must not
+                # wedge the device as is_connected (blocking every later connect/disconnect).
+                session = self._session
+                self._session = None
+                session.__exit__(None, None, None)
+                logger.info("Isaac Teleop session ended")
+        finally:
+            # Reap the CloudXR runtime even if session teardown raised, and even if no
+            # session was ever established (e.g. the launcher came up but session creation
+            # failed before this point); a no-op when we never launched CloudXR (opt-out /
+            # externally-owned runtime), so we never stop a runtime we don't own.
+            self._stop_cloudxr_runtime()
+
+    # ------------------------------------------------------------------
+    # CloudXR runtime (shared)
+    # ------------------------------------------------------------------
+
+    def _ensure_cloudxr_runtime(self) -> None:
+        """Auto-launch the CloudXR runtime once, unless opted out.
+
+        Idempotent (no-op once the launcher is up). ``LEROBOT_CLOUDXR_SKIP_AUTOLAUNCH``
+        is checked first and wins over ``config.auto_launch_cloudxr``. Constructing
+        :class:`CloudXRLauncher` mutates the process env (``XR_RUNTIME_JSON`` etc.) and
+        blocks until the runtime is ready or raises :class:`RuntimeError`.
+        """
+        if self._cloudxr_launcher is not None:
+            return
+
+        if os.environ.get("LEROBOT_CLOUDXR_SKIP_AUTOLAUNCH", "").strip() == "1":
+            logger.info(
+                "LEROBOT_CLOUDXR_SKIP_AUTOLAUNCH=1 set; skipping CloudXR auto-launch "
+                "(assuming CloudXR is already running externally)"
+            )
+            return
+
+        if not self.config.auto_launch_cloudxr:
+            logger.info(
+                "config.auto_launch_cloudxr is False; skipping CloudXR auto-launch "
+                "(assuming CloudXR is already running externally)"
+            )
+            return
+
+        logger.info("Launching CloudXR runtime (first run may prompt for EULA and take ~30s)...")
+
+        self._cloudxr_launcher = CloudXRLauncher(
+            install_dir=str(Path.home() / ".cloudxr"),
+            env_config=self.config.cloudxr_env_file,
+            accept_eula=False,
+        )
+
+    def _stop_cloudxr_runtime(self) -> None:
+        """Stop the auto-launched CloudXR runtime, if any.
+
+        Clean stop nulls the handle. On :class:`RuntimeError` the handle is RETAINED so
+        the launcher's ``atexit`` hook owns the retry — a later :meth:`connect` then
+        treats the retained runtime as still up and will not relaunch.
+        """
+        if self._cloudxr_launcher is None:
+            return
+        try:
+            self._cloudxr_launcher.stop()
+        except RuntimeError:
+            logger.warning("CloudXR runtime could not be terminated; handle retained for atexit cleanup")
+        else:
+            self._cloudxr_launcher = None
+            logger.info("CloudXR runtime stopped")
+
+    def send_feedback(self, feedback: dict[str, Any]) -> None:
+        pass  # Haptic feedback not yet implemented.
+
+    # ------------------------------------------------------------------
+    # Stepping (shared)
+    # ------------------------------------------------------------------
+
+    def _running_events(self) -> ExecutionEvents:
+        """Constant ``RUNNING`` ``ExecutionEvents`` for a device with no clutch lifecycle.
+
+        Keeps the stream flowing; ``reset`` stays ``False``. A clutched device that needs
+        a real lifecycle should build its own ``ExecutionEvents`` instead.
+        """
+        return ExecutionEvents(execution_state=ExecutionState.RUNNING, reset=False)
+
+    def _step(
+        self,
+        *,
+        execution_events: ExecutionEvents | None = None,
+        external_inputs: Mapping[str, Any] | None = None,
+    ) -> RetargeterIO:
+        """Step the session once and return the raw pipeline outputs.
+
+        Applies the shared guard: re-raises a retargeting-worker exception and warns on a
+        stale frame. Subclasses call this from :meth:`get_action`.
+
+        Args:
+            execution_events: The ``ExecutionEvents`` driving the session this frame.
+                Devices with a lifecycle (clutch) MUST pass this every frame — when
+                ``None``, ``TeleopSession.step`` auto-fires ``RUNNING`` (the clutch would
+                latch immediately and never stop).
+            external_inputs: Per-step inputs (e.g. a static ``base_T_anchor``) in the
+                ``{leaf_node_name: {output_port_name: TensorGroup}}`` shape ``step`` expects.
+
+        Raises:
+            RuntimeError: If not connected, or if the retargeting worker raised.
+        """
+        if self._session is None:
+            raise RuntimeError("Not connected. Call connect() first.")
+
+        result = self._session.step(
+            execution_events=execution_events,
+            external_inputs=external_inputs,
+        )
+
+        info = self._session.last_step_info
+        if info is not None:
+            if info.worker_exception is not None:
+                raise RuntimeError(
+                    "Isaac Teleop retargeting worker raised an exception"
+                ) from info.worker_exception
+            if info.frame_deadline_miss:
+                logger.warning(
+                    "Isaac Teleop frame deadline miss (returned_age_frames=%s)",
+                    info.returned_age_frames,
+                )
+        return result

examples/isaac_teleop_to_so101/isaac_teleop/clutch.py

@@ -0,0 +1,102 @@
+#!/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.
+
+"""Engage-relative clutch for the XR -> SO-101 teleop loop.
+
+Turns the raw controller grip pose into an absolute base-frame EE target, so the XR
+device can stay a thin raw-pose reader. Pure numpy + the local ``Rotation`` helper (no
+``isaacteleop``), so it is unit-testable without the XR runtime.
+"""
+
+from __future__ import annotations
+
+import numpy as np
+
+from lerobot.utils.rotation import Rotation
+
+
+class Clutch:
+    """Engage-relative clutch for both position AND orientation.
+
+    Latch an origin on engage, then track the base-frame delta from it, applied
+    independently to position and orientation. State:
+
+    - ``_last_commanded_pos`` / ``_last_commanded_rot``: last commanded EE pose; held
+      while disengaged so the arm freezes where it was left.
+    - ``_home_pos`` / ``_home_rot``: latched on engage — the EE pose the delta applies to.
+      The position comes from the arm's MEASURED pose when the caller provides it (so an
+      arm that moved while disengaged is not snapped back to a stale command); the
+      orientation always comes from the last commanded rotation (see NOTE below).
+    - ``_origin_pos`` / ``_origin_rot``: latched on engage — the controller pose the delta
+      is measured against.
+
+    Each engaged frame :meth:`rebase` returns::
+
+        pos = home_pos + (grip_pos - origin_pos)  # 1:1 controller -> EE translation
+        rot = (R_ctrl @ R_origin ^ -1) @ R_home  # base-frame delta, left-composed
+
+    On the engage edge the output is exactly the home pose (no teleport). The orientation
+    delta is left-composed (base frame), so hand rotation about base Z maps to EE rotation
+    about base Z. A re-clutch latches a fresh home/origin.
+
+    NOTE: ``_home_rot`` is the last *commanded* orientation even when the measured pose is
+    supplied: the 5-DOF SO-101 tracks orientation only softly, so its measured wrist
+    orientation persistently differs from the command, and latching the measurement would
+    inject that offset into the commanded signal on every re-clutch. Position has no such
+    tracking gap, and there latching the measurement is what prevents the snap-back.
+    """
+
+    def __init__(self, home_base_T_ee: np.ndarray):  # noqa: N803
+        # Seed the held pose from the arm's measured startup EE pose so the first
+        # engage latches home there (no jump on the first squeeze).
+        home = np.asarray(home_base_T_ee, dtype=float)
+        self._last_commanded_pos = home[:3, 3].copy()
+        self._last_commanded_rot = Rotation.from_matrix(home[:3, :3])
+        self._home_pos = self._last_commanded_pos.copy()
+        self._home_rot = self._last_commanded_rot
+        self._origin_pos = np.zeros(3, dtype=float)
+        self._origin_rot = Rotation.from_quat(np.array([0.0, 0.0, 0.0, 1.0]))
+
+    def engage(
+        self,
+        grip_pos: np.ndarray,
+        grip_quat: np.ndarray,
+        measured_base_T_ee: np.ndarray | None = None,  # noqa: N803
+    ) -> None:
+        """Latch the engage home (where the arm is now) and controller origin.
+
+        Pass ``measured_base_T_ee`` (FK of the measured joints) so the home POSITION is
+        where the arm physically is — if the arm moved while disengaged (gravity sag,
+        external contact), latching the stale last-commanded position would make the
+        first engaged frame command a full-speed jump back to it. The home ORIENTATION
+        always stays the last commanded one (see the class NOTE).
+        """
+        if measured_base_T_ee is not None:
+            self._home_pos = np.asarray(measured_base_T_ee, dtype=float)[:3, 3].copy()
+        else:
+            self._home_pos = self._last_commanded_pos.copy()
+        self._home_rot = self._last_commanded_rot
+        self._origin_pos = np.asarray(grip_pos, dtype=float).copy()
+        self._origin_rot = Rotation.from_quat(np.asarray(grip_quat, dtype=float))
+
+    def rebase(self, grip_pos: np.ndarray, grip_quat: np.ndarray) -> tuple[np.ndarray, np.ndarray]:
+        """Return the absolute base-frame EE target ``(pos [m], quat [xyzw])`` for this frame."""
+        pos = self._home_pos + (np.asarray(grip_pos, dtype=float) - self._origin_pos)
+        rot_ctrl = Rotation.from_quat(np.asarray(grip_quat, dtype=float))
+        rot = (rot_ctrl * self._origin_rot.inv()) * self._home_rot
+        self._last_commanded_pos = pos.copy()
+        self._last_commanded_rot = rot
+        return pos, rot.as_quat()

examples/isaac_teleop_to_so101/isaac_teleop/config_isaac_teleop.py

@@ -0,0 +1,135 @@
+#!/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.
+
+"""Configuration dataclasses for NVIDIA Isaac Teleop-backed teleoperators.
+
+:class:`IsaacTeleopConfig` holds the shared fields; each device adds its own subclass
+(e.g. :class:`XRControllerConfig`, :class:`SO101LeaderArmConfig`).
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import ClassVar
+
+from lerobot.teleoperators.config import TeleoperatorConfig
+
+
+@dataclass(kw_only=True)
+class IsaacTeleopConfig(TeleoperatorConfig):
+    """Shared config for all Isaac Teleop-backed teleoperators.
+
+    Uses its own draccus ``_choice_registry`` (decoupled from the global
+    :class:`TeleoperatorConfig` one) so ``--teleop.type`` on a field typed
+    ``IsaacTeleopConfig`` resolves against ONLY the Isaac devices — letting them claim
+    short names (``xr_controller``, ``so101_leader``) without colliding with the global
+    registry. These devices are selected by the example scripts, not routed through
+    ``make_teleoperator_from_config``.
+    """
+
+    _choice_registry: ClassVar[dict] = {}
+
+    app_name: str = "LeTeleop"
+    """Application name for the OpenXR / Isaac Teleop session."""
+
+    auto_launch_cloudxr: bool = True
+    """Auto-launch the CloudXR runtime on :meth:`connect`. Set ``False`` (or export
+    ``LEROBOT_CLOUDXR_SKIP_AUTOLAUNCH=1``, which wins) when CloudXR runs externally.
+    """
+
+    cloudxr_env_file: str | None = None
+    """Optional CloudXR device-profile ``.env`` (an INPUT profile selecting the headset
+    transport) passed to ``CloudXRLauncher``. ``None`` keeps the default auto-WebRTC profile.
+    """
+
+
+# Static rebase from the OpenXR controller anchor frame (X=Right, Y=Up, Z=Backward) into the
+# robot base frame (X=Forward, Y=Left, Z=Up). A proper rotation (det=+1): controller motion
+# forward -> robot +X, right -> robot -Y (i.e. rightward), up -> robot +Z.
+_DEFAULT_BASE_T_ANCHOR: list[list[float]] = [
+    [0.0, 0.0, -1.0, 0.0],
+    [-1.0, 0.0, 0.0, 0.0],
+    [0.0, 1.0, 0.0, 0.0],
+    [0.0, 0.0, 0.0, 1.0],
+]
+
+
+@IsaacTeleopConfig.register_subclass("xr_controller")
+@dataclass(kw_only=True)
+class XRControllerConfig(IsaacTeleopConfig):
+    """Config for Isaac Teleop XR (VR) controller teleoperation.
+
+    Exposes the raw base-frame grip pose, squeeze, and trigger via ``ControllersSource``.
+    No retargeters: the clutch and gripper mapping live in the owning loop.
+    """
+
+    hand_side: str = "right"
+    """Which controller hand to use: ``"left"`` or ``"right"``. A plain ``str`` (validated in
+    ``__post_init__``) because draccus cannot decode ``Literal``-typed fields from the CLI."""
+
+    clutch_threshold: float = 0.5
+    """Squeeze value above which the owning loop's clutch engages (held-to-enable). The
+    device reports only the raw squeeze; the threshold is applied by the loop."""
+
+    base_T_anchor: list[list[float]] = field(  # noqa: N815  (frameA_T_frameB transform-matrix convention)
+        # Fresh copy per instance: returning the module-level list itself would alias one
+        # mutable matrix across every config.
+        default_factory=lambda: [row.copy() for row in _DEFAULT_BASE_T_ANCHOR]
+    )
+    """Static 4x4 [row-major] transform rebasing the OpenXR controller anchor frame into
+    the robot base frame. Defaults to OpenXR (X=Right, Y=Up, Z=Backward) -> robot
+    (X=Forward, Y=Left, Z=Up). Plain nested lists so the config stays serializable.
+    """
+
+    def __post_init__(self):
+        if self.hand_side not in ("left", "right"):
+            raise ValueError(f"hand_side must be 'left' or 'right', got {self.hand_side!r}")
+
+
+# Provisional gripper open/close endpoints [rad], normalizing the streamed gripper angle
+# into the follower's RANGE_0_100 jaw target. Derived from the so101_leader plugin README's
+# example calibration (home_ticks=2048, range 2000..3000; angle = (ticks-home)*2*pi/4096).
+_DEFAULT_GRIPPER_OPEN_RAD = -0.074
+_DEFAULT_GRIPPER_CLOSE_RAD = 1.460
+
+
+@IsaacTeleopConfig.register_subclass("so101_leader")
+@dataclass(kw_only=True)
+class SO101LeaderArmConfig(IsaacTeleopConfig):
+    """Config for an Isaac Teleop SO-101 *leader arm* (generic joint-space device).
+
+    Mirrors the leader's joint angles 1:1 onto a follower SO-101. The leader state is
+    streamed in radians by the native ``so101_leader`` plugin and read via a
+    ``JointStateSource``; the device converts arm joints to degrees and the gripper to the
+    follower's RANGE_0_100 jaw target (no IK/clutch/retargeter on the LeRobot side).
+    """
+
+    port: str = ""
+    """Serial port of the physical LEADER arm (e.g. ``/dev/ttyACM1``), forwarded to the
+    plugin (which reads the servos) when the example launches it. Empty -> the plugin runs
+    its synthetic trajectory."""
+
+    collection_id: str = "so101_leader"
+    """Tensor collection id the leader plugin pushes on; must match the running
+    ``so101_leader`` plugin (its second positional arg, default ``"so101_leader"``)."""
+
+    gripper_open_rad: float = _DEFAULT_GRIPPER_OPEN_RAD
+    """Leader gripper angle [rad] at fully OPEN -> follower jaw 100. Provisional default;
+    set from the plugin's ``calibrate`` subcommand. See ``_DEFAULT_GRIPPER_OPEN_RAD``."""
+
+    gripper_close_rad: float = _DEFAULT_GRIPPER_CLOSE_RAD
+    """Leader gripper angle [rad] at fully CLOSED -> follower jaw 0. Provisional default;
+    set from the plugin's ``calibrate`` subcommand. See ``_DEFAULT_GRIPPER_CLOSE_RAD``."""

examples/isaac_teleop_to_so101/isaac_teleop/teleop_so101_leader_arm.py

@@ -0,0 +1,186 @@
+#!/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.
+
+"""SO-101 leader-arm device for NVIDIA Isaac Teleop, exposed to LeRobot.
+
+The leader is a back-drivable SO-101 whose six joint angles are streamed (in radians) by
+the native ``so101_leader`` plugin; this device reads them via a ``JointStateSource`` and
+converts them into follower-ready ``{joint}.pos``. Same kinematics as the follower, so it
+needs no retargeting — a 1:1 joint mirror, direct joint drive.
+
+Units (converted in the device so the output is always follower-valid):
+
+* arm joints: ``rad2deg`` — correct only if the leader's calibrated zero and the follower's
+  homing map to the same physical zero (the standard same-hardware assumption).
+* gripper: normalized from ``[gripper_open_rad, gripper_close_rad]`` to RANGE_0_100.
+
+``isaacteleop`` imports are guarded behind the availability flag so this module — and the
+pure :func:`leader_joints_to_robot_action` converter — import without it (construction
+fails fast via the base class).
+"""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+import numpy as np
+
+from lerobot.types import RobotAction
+
+from .base import _GRIPPER_MOTOR_SCALE, IsaacTeleopTeleoperator, _isaacteleop_available
+from .config_isaac_teleop import SO101LeaderArmConfig
+
+if TYPE_CHECKING or _isaacteleop_available:
+    from isaacteleop.retargeting_engine.deviceio_source_nodes import JointStateSource
+    from isaacteleop.retargeting_engine.interface import OutputCombiner
+else:
+    JointStateSource = None
+    OutputCombiner = None
+
+# Canonical SO-101 DOF names and order — matches the plugin stream and the follower's motor
+# order. Passed to the ``JointStateSource`` as its output layout; the source maps by name and
+# :func:`_joints_group_to_rad` reads back by name, so a layout mismatch can't mislabel a DOF.
+SO101_LEADER_JOINTS = [
+    "shoulder_pan",
+    "shoulder_lift",
+    "elbow_flex",
+    "wrist_flex",
+    "wrist_roll",
+    "gripper",
+]
+
+
+def leader_joints_to_robot_action(
+    joints_rad: dict[str, float],
+    *,
+    gripper_joint: str,
+    gripper_open_rad: float,
+    gripper_close_rad: float,
+) -> RobotAction:
+    """Convert streamed leader joint angles [rad] to follower-ready ``{joint}.pos``.
+
+    Pure (no ``isaacteleop``, no I/O). Iteration follows ``joints_rad`` insertion order, so
+    pass it in :data:`SO101_LEADER_JOINTS` order for a stable layout. Arm joints are
+    converted ``rad2deg``; ``gripper_joint`` is normalized from
+    ``[gripper_open_rad, gripper_close_rad]`` to RANGE_0_100 (clipped).
+    """
+    action: RobotAction = {}
+    span = gripper_close_rad - gripper_open_rad
+    for name, rad in joints_rad.items():
+        if name == gripper_joint:
+            # Closedness c=0 at open, c=1 at closed; invert to the follower's 100=open jaw.
+            closedness = 0.0 if span == 0.0 else (rad - gripper_open_rad) / span
+            closedness = min(1.0, max(0.0, closedness))
+            action[f"{name}.pos"] = (1.0 - closedness) * _GRIPPER_MOTOR_SCALE
+        else:
+            action[f"{name}.pos"] = float(np.rad2deg(rad))
+    return action
+
+
+def _joints_group_to_rad(joints) -> dict[str, float]:
+    """Read a ``JointStateSource`` output group into ``{joint_name: angle [rad]}``.
+
+    Pure (duck-typed on the group). The group is positional but each slot carries its joint
+    name in ``group.group_type.types``; we key off those names (not a positional index) so a
+    layout mismatch surfaces as a wrong/missing key here rather than a mislabeled DOF.
+    """
+    names = [t.name for t in joints.group_type.types]
+    return {name: float(joints[i]) for i, name in enumerate(names)}
+
+
+class SO101LeaderArm(IsaacTeleopTeleoperator):
+    """SO-101 leader-arm teleoperator (joint-space), direct joint mirror to the follower.
+
+    Reads the six joint angles off a single ``JointStateSource`` each frame; no retargeter,
+    no clutch. When the leader is not streaming, :meth:`get_action` returns the held-last
+    joints and :attr:`is_tracking` is ``False`` so the owning loop can hold the follower.
+    """
+
+    config_class = SO101LeaderArmConfig
+    name = "isaac_teleop_so101_leader"
+
+    def __init__(self, config: SO101LeaderArmConfig):
+        super().__init__(config)
+        self.config: SO101LeaderArmConfig = config
+        # Held-last joint angles [rad], seeded at zero (URDF/home pose) so the first frames
+        # before the plugin starts pushing read as the home pose, not garbage.
+        self._last_joints_rad: dict[str, float] = dict.fromkeys(SO101_LEADER_JOINTS, 0.0)
+        # Whether the most recent get_action() read live leader data (vs held-last).
+        self._is_tracking = False
+
+    # ------------------------------------------------------------------
+    # Pipeline construction
+    # ------------------------------------------------------------------
+
+    def _build_pipeline(self) -> OutputCombiner:
+        """Build the joint-mirror pipeline: a single ``JointStateSource`` leaf that converts
+        the raw stream into a name-keyed joint group. No retargeter (shared kinematics)."""
+        source = JointStateSource(
+            name="so101_leader",
+            collection_id=self.config.collection_id,
+            joint_names=SO101_LEADER_JOINTS,
+        )
+        return OutputCombiner({"joints": source.output(JointStateSource.JOINTS)})
+
+    # ------------------------------------------------------------------
+    # Action features
+    # ------------------------------------------------------------------
+
+    @property
+    def action_features(self) -> dict[str, type]:
+        # Matches the serial SOLeader's action features so this is a drop-in joint-space
+        # leader: one float `{joint}.pos` per DOF, sendable straight to an SO-101 follower.
+        return {f"{name}.pos": float for name in SO101_LEADER_JOINTS}
+
+    @property
+    def feedback_features(self) -> dict[str, type]:
+        return {}
+
+    @property
+    def is_tracking(self) -> bool:
+        """Whether the last :meth:`get_action` read live leader data (vs held-last)."""
+        return self._is_tracking
+
+    # ------------------------------------------------------------------
+    # Action extraction
+    # ------------------------------------------------------------------
+
+    def get_action(self) -> RobotAction:
+        """Step the session and return the leader joints as follower-ready ``{joint}.pos``.
+
+        When the leader is streaming, the live angles are cached and converted; otherwise the
+        held-last angles are reused and :attr:`is_tracking` is set ``False``.
+        """
+        result = self._step(execution_events=self._running_events())
+
+        joints = result["joints"]
+        # The JointStateSource output is Optional: absent (is_none) when the device is
+        # inactive. Treat that as "not tracking" and reuse the held-last angles.
+        self._is_tracking = not getattr(joints, "is_none", False)
+        if self._is_tracking:
+            try:
+                self._last_joints_rad = _joints_group_to_rad(joints)
+            except (AttributeError, IndexError, KeyError, TypeError, ValueError):
+                # A partially-populated / malformed group on an odd frame: keep held-last, but
+                # report it as not-tracking so the loop holds the follower rather than trusting it.
+                self._is_tracking = False
+
+        return leader_joints_to_robot_action(
+            self._last_joints_rad,
+            gripper_joint="gripper",
+            gripper_open_rad=self.config.gripper_open_rad,
+            gripper_close_rad=self.config.gripper_close_rad,
+        )

examples/isaac_teleop_to_so101/isaac_teleop/teleop_xr_controller.py

@@ -0,0 +1,204 @@
+#!/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.
+
+"""XR (VR) controller device for NVIDIA Isaac Teleop, exposed to LeRobot.
+
+A deliberately thin reader: exposes the raw controller grip pose off
+``ControllersSource`` (statically rebased into the robot base frame by
+``ControllerTransform``), plus squeeze and trigger. No retargeters and no clutch —
+the clutch rebasing and gripper mapping live downstream in the owning loop, so this
+device is stateless across frames.
+
+``isaacteleop`` imports are guarded behind the availability flag so this module imports
+without it (construction fails fast via the base class).
+"""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Any
+
+import numpy as np
+
+from lerobot.types import RobotAction
+
+from .base import IsaacTeleopTeleoperator, _isaacteleop_available
+from .config_isaac_teleop import XRControllerConfig
+
+if TYPE_CHECKING or _isaacteleop_available:
+    from isaacteleop.retargeting_engine.deviceio_source_nodes import ControllersSource
+    from isaacteleop.retargeting_engine.interface import OutputCombiner, TensorGroup, ValueInput
+    from isaacteleop.retargeting_engine.tensor_types import TransformMatrix
+    from isaacteleop.retargeting_engine.tensor_types.indices import ControllerInputIndex
+else:
+    ControllersSource = None
+    OutputCombiner = None
+    TensorGroup = None
+    ValueInput = None
+    TransformMatrix = None
+    ControllerInputIndex = None
+
+# Source-node name for the static base_T_anchor rebase input fed via
+# ``TeleopSession.step(external_inputs=...)`` each frame.
+_BASE_T_ANCHOR_INPUT = "base_T_anchor"
+
+
+class XRController(IsaacTeleopTeleoperator):
+    """Raw XR controller grip-pose teleoperator (base-frame), no retargeters.
+
+    Reads the raw grip pose + squeeze + trigger off a ``ControllersSource`` rebased into
+    the robot base frame. :meth:`get_action` returns the absolute base-frame grip pose
+    untouched; the owning loop owns the clutch and gripper mapping.
+    """
+
+    config_class = XRControllerConfig
+    name = "isaac_teleop_controller"
+
+    def __init__(self, config: XRControllerConfig):
+        super().__init__(config)
+        self.config: XRControllerConfig = config
+
+        # Constant base_T_anchor input, built once in connect() (a TensorGroup is heavy and
+        # isaacteleop-backed) and reused every step.
+        self._external_inputs: dict[str, Any] | None = None
+        # Whether the last get_action() read a tracked controller; the owning loop polls this
+        # to wait for the operator to connect before driving the arm.
+        self._is_tracking = False
+
+    # ------------------------------------------------------------------
+    # Pipeline construction
+    # ------------------------------------------------------------------
+
+    def _build_pipeline(self) -> OutputCombiner:
+        """Build the raw-grip-pose pipeline: a ``ControllersSource`` rebased into the base
+        frame by ``ControllerTransform``, exposed verbatim as ``"controller"``. No retargeters.
+        """
+        side = self.config.hand_side
+        controller_key = f"controller_{side}"
+
+        controllers = ControllersSource(name="controllers")
+        # Static base_T_anchor rebase fed via external_inputs each step.
+        xform = ValueInput(_BASE_T_ANCHOR_INPUT, TransformMatrix())
+        transformed = controllers.transformed(xform.output("value"))
+        ctrl = transformed.output(controller_key)
+
+        return OutputCombiner({"controller": ctrl})
+
+    def _build_external_inputs(self) -> dict[str, Any]:
+        """Materialize the constant ``base_T_anchor`` external input (once, in connect)."""
+        tg = TensorGroup(TransformMatrix())
+        tg[0] = np.asarray(self.config.base_T_anchor, dtype=np.float32)
+        return {_BASE_T_ANCHOR_INPUT: {"value": tg}}
+
+    def connect(self, calibrate: bool = True) -> None:
+        super().connect(calibrate=calibrate)
+        try:
+            self._external_inputs = self._build_external_inputs()
+        except Exception:
+            # Roll the session/runtime back so a failed connect() leaves no half-state
+            # (a live session behind a raised connect would leak the CloudXR runtime).
+            self.disconnect()
+            raise
+
+    # ------------------------------------------------------------------
+    # Action features
+    # ------------------------------------------------------------------
+
+    @property
+    def action_features(self) -> dict:
+        return {
+            "grip_pos": {
+                "dtype": "float32",
+                "shape": (3,),
+                "names": {"x": 0, "y": 1, "z": 2},
+            },
+            "grip_quat": {
+                "dtype": "float32",
+                "shape": (4,),
+                "names": {"qx": 0, "qy": 1, "qz": 2, "qw": 3},
+            },
+            # ``get_action`` returns scalars for these two, so the advertised
+            # shape is () (0-d) to stay consistent with the returned values.
+            "squeeze": {
+                "dtype": "float32",
+                "shape": (),
+                "names": None,
+            },
+            "trigger": {
+                "dtype": "float32",
+                "shape": (),
+                "names": None,
+            },
+        }
+
+    @property
+    def feedback_features(self) -> dict:
+        return {}
+
+    @property
+    def is_tracking(self) -> bool:
+        """Whether the last :meth:`get_action` read a tracked controller. ``False`` until the
+        headset is connected over CloudXR and its controllers are live; the owning loop polls
+        it to wait for the operator before commanding the arm."""
+        return self._is_tracking
+
+    # ------------------------------------------------------------------
+    # Action extraction
+    # ------------------------------------------------------------------
+
+    def get_action(self) -> RobotAction:
+        """Step the session and return the raw base-frame grip pose.
+
+        Reads the grip pose + squeeze + trigger off the transformed controller stream (with
+        the constant ``base_T_anchor`` rebase). When the controller is not tracked, returns
+        identity pose and squeeze/trigger = 0.0 so the owning loop freezes the arm.
+
+        Returns:
+            ``{"grip_pos": (3,) [m], "grip_quat": (4,) [qx,qy,qz,qw], "squeeze": float,
+            "trigger": float}`` — pose in the robot base frame; squeeze/trigger in ``[0, 1]``.
+        """
+        result = self._step(execution_events=self._running_events(), external_inputs=self._external_inputs)
+
+        # Optional controller group is None until the headset is connected and its controllers
+        # are live; expose that as is_tracking so the loop can wait before driving the arm.
+        controller = result["controller"]
+        grip_pos = np.zeros(3, dtype=np.float32)
+        grip_quat = np.array([0.0, 0.0, 0.0, 1.0], dtype=np.float32)
+        squeeze = 0.0
+        trigger = 0.0
+        self._is_tracking = not getattr(controller, "is_none", False)
+        if self._is_tracking:
+            # Read ALL four fields into locals before committing any of them: a failure on a
+            # partially-populated frame must not mix live values with the safe defaults (a
+            # live squeeze paired with a defaulted trigger=0.0 would keep the clutch engaged
+            # while commanding the gripper fully open, dropping whatever is grasped). On
+            # failure the defaults stand untouched and the frame reports not-tracked.
+            try:
+                pos = np.asarray(controller[ControllerInputIndex.GRIP_POSITION], dtype=np.float32)
+                quat = np.asarray(controller[ControllerInputIndex.GRIP_ORIENTATION], dtype=np.float32)
+                squeeze_val = float(controller[ControllerInputIndex.SQUEEZE_VALUE])
+                trigger_val = float(controller[ControllerInputIndex.TRIGGER_VALUE])
+            except (IndexError, KeyError, TypeError, ValueError):
+                self._is_tracking = False
+            else:
+                grip_pos, grip_quat = pos, quat
+                squeeze, trigger = squeeze_val, trigger_val
+
+        return {
+            "grip_pos": grip_pos,
+            "grip_quat": grip_quat,
+            "squeeze": squeeze,
+            "trigger": trigger,
+        }

examples/isaac_teleop_to_so101/isaac_teleop/xr_controller_processor.py

@@ -0,0 +1,87 @@
+#!/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.
+
+"""Processor step that maps XR controller actions to robot EE targets.
+
+Analogous to ``MapPhoneActionToRobotAction``, this bridges the clutch-rebased EE pose to
+the IK pipeline's input contract (``EEBoundsAndSafety`` -> ``InverseKinematicsEEToJoints``).
+Pure (no ``isaacteleop``), so it is unit-testable without the XR runtime.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+
+from lerobot.configs.types import FeatureType, PipelineFeatureType, PolicyFeature
+from lerobot.processor import ProcessorStepRegistry, RobotActionProcessorStep
+from lerobot.types import RobotAction
+from lerobot.utils.rotation import Rotation
+
+from .base import _GRIPPER_MOTOR_SCALE
+
+
+@ProcessorStepRegistry.register("map_xr_controller_action_to_robot_action")
+@dataclass
+class MapXRControllerActionToRobotAction(RobotActionProcessorStep):
+    """Maps an absolute base-frame EE pose + gripper closedness to the IK input contract.
+
+    Pure, stateless rename (the owning loop's clutch already produced the absolute base-frame
+    target). Each frame it writes:
+
+    - ``ee.x/y/z`` = ``ee_pose[:3]`` (position [m]);
+    - ``ee.wx/wy/wz`` = rotvec of ``ee_pose[3:7]`` (orientation; the IK tracks it softly at a
+      small ``orientation_weight`` on the 5-DOF SO-101);
+    - ``ee.gripper_pos`` = ``(1 - closedness) * _GRIPPER_MOTOR_SCALE`` (jaw target [0, 100],
+      RANGE_0_100 where 100 = open, so closedness is inverted).
+
+    Input keys: ``ee_pose`` ``(7,)`` ``[x,y,z,qx,qy,qz,qw]``, ``closedness`` float in [0, 1].
+    """
+
+    def action(self, action: RobotAction) -> RobotAction:
+        ee_pose = action.pop("ee_pose")
+        closedness = float(action.pop("closedness"))
+
+        action["ee.x"] = float(ee_pose[0])
+        action["ee.y"] = float(ee_pose[1])
+        action["ee.z"] = float(ee_pose[2])
+        # Orientation target as a rotvec (quat [qx,qy,qz,qw] -> axis-angle); the IK
+        # consumes ee.w* as a rotvec and tracks it with orientation_weight.
+        rotvec = Rotation.from_quat(ee_pose[3:7]).as_rotvec()
+        action["ee.wx"] = float(rotvec[0])
+        action["ee.wy"] = float(rotvec[1])
+        action["ee.wz"] = float(rotvec[2])
+        # Inverted: closedness c=1 (closed) -> 0, c=0 (open) -> 100 (SO-101 calibration).
+        action["ee.gripper_pos"] = (1.0 - closedness) * _GRIPPER_MOTOR_SCALE
+        return action
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        for feat in ["ee_pose", "closedness"]:
+            features[PipelineFeatureType.ACTION].pop(feat, None)
+
+        for feat in [
+            "ee.x",
+            "ee.y",
+            "ee.z",
+            "ee.wx",
+            "ee.wy",
+            "ee.wz",
+            "ee.gripper_pos",
+        ]:
+            features[PipelineFeatureType.ACTION][feat] = PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+
+        return features

examples/isaac_teleop_to_so101/override_reset_pose.py

@@ -0,0 +1,73 @@
+#!/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.
+
+"""Save the current SO-101 joint positions as the reset-origin pose (override).
+
+Move the arm to the desired reset pose by hand (torque off), then run this script to write
+those joints to a per-arm file in the LeRobot cache. ``teleoperate.py`` / ``record.py`` load
+it on startup (matched by ``--robot.id``) as the reset target instead of the defaults.
+
+Usage::
+
+    # 1. Move arm to desired reset pose by hand
+    python -m examples.isaac_teleop_to_so101.override_reset_pose [--port /dev/ttyACM0] [--id so101_follower_arm]
+
+    # 2. Launch teleop with the SAME --robot.id — it will now reset to this pose on startup
+    python -m examples.isaac_teleop_to_so101.teleoperate --robot.type=so101_follower --robot.port=/dev/ttyACM0 --robot.id=so101_follower_arm --teleop.type=xr_controller
+"""
+
+import argparse
+import json
+from pathlib import Path
+
+from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
+
+from .common import RESET_POSE_FILE
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(
+        description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter
+    )
+    parser.add_argument("--port", type=str, default="/dev/ttyACM0")
+    parser.add_argument("--id", type=str, default="so101_follower_arm")
+    return parser.parse_args()
+
+
+def main():
+    args = parse_args()
+    robot = SO100Follower(SO100FollowerConfig(port=args.port, id=args.id, use_degrees=True))
+    robot.connect()
+    # Always disconnect the follower so a failure never leaks the serial connection.
+    try:
+        obs = robot.get_observation()
+        motor_names = list(robot.bus.motors.keys())
+        pose = {name: float(obs[f"{name}.pos"]) for name in motor_names}
+    finally:
+        robot.disconnect()
+
+    print("Current joint positions:")
+    for name, val in pose.items():
+        print(f"  {name:20s}: {val:.2f}")
+
+    reset_pose_file = Path(RESET_POSE_FILE.format(robot_name=robot.name, robot_id=robot.id))
+    reset_pose_file.parent.mkdir(parents=True, exist_ok=True)
+    reset_pose_file.write_text(json.dumps(pose, indent=2))
+    print(f"\nSaved to {reset_pose_file}")
+
+
+if __name__ == "__main__":
+    main()

examples/isaac_teleop_to_so101/record.py

@@ -0,0 +1,321 @@
+#!/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.
+
+"""Record a LeRobot dataset via NVIDIA Isaac Teleop -> SO-101.
+
+Runs ``teleoperate.py``'s control loop while also saving each frame to a LeRobot dataset.
+``--teleop.type`` selects the device (``xr_controller`` | ``so101_leader``) as in
+``teleoperate.py``.
+
+Usage::
+
+    # XR (VR) controller: clutch + soft-orientation IK
+    python -m examples.isaac_teleop_to_so101.record \\
+        --robot.type=so101_follower \\
+        --robot.port=/dev/ttyACM0 \\
+        --robot.id=so101_follower_arm \\
+        --teleop.type=xr_controller \\
+        --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \\
+        --dataset.repo_id=<hf_user>/<dataset_name> \\
+        --dataset.single_task="Pick up vial from rack on the left side" \\
+        --dataset.num_episodes=3 \\
+        --dataset.episode_time_s=20 \\
+        --dataset.reset_time_s=5
+
+    # SO-101 leader arm: 1:1 joint mirror (real leader on /dev/ttyACM1)
+    python -m examples.isaac_teleop_to_so101.record \\
+        --robot.type=so101_follower --robot.port=/dev/ttyACM0 --robot.id=so101_follower_arm \\
+        --teleop.type=so101_leader --teleop.port=/dev/ttyACM1 --teleop.id=so101_leader_arm \\
+        --launch_plugin=/path/to/IsaacTeleop/install/plugins/so101_leader/so101_leader_plugin \\
+        --dataset.repo_id=<hf_user>/<dataset_name> --dataset.single_task="Pick up the cube" \\
+        --dataset.num_episodes=3 --dataset.episode_time_s=20 --dataset.reset_time_s=5
+
+The loop/launch knobs mirror ``teleoperate.py`` (tagged ``[xr]`` / ``[leader]`` below).
+
+Keyboard shortcuts: Right/n = end episode early and save, Left/r = discard + re-record,
+Esc/q = stop after the current episode. All frames are recorded (including hold frames).
+"""
+
+import logging
+import time
+from dataclasses import asdict, dataclass
+from pprint import pformat
+
+from lerobot.cameras import CameraConfig  # noqa: F401
+from lerobot.cameras.opencv import OpenCVCameraConfig  # noqa: F401
+from lerobot.common.control_utils import sanity_check_dataset_robot_compatibility
+from lerobot.configs import parser
+from lerobot.configs.dataset import DatasetRecordConfig
+from lerobot.datasets import (
+    LeRobotDataset,
+    VideoEncodingManager,
+    aggregate_pipeline_dataset_features,
+    create_initial_features,
+    safe_stop_image_writer,
+)
+from lerobot.processor import make_default_processors
+from lerobot.robots import RobotConfig
+from lerobot.robots.so_follower import SOFollowerConfig  # noqa: F401  (registers so101_follower)
+from lerobot.utils.constants import ACTION, OBS_STR
+from lerobot.utils.feature_utils import build_dataset_frame, combine_feature_dicts
+from lerobot.utils.robot_utils import precise_sleep
+from lerobot.utils.utils import init_logging
+
+from .common import (
+    ALIGN_DURATION_S,
+    RESET_DURATION_S,
+    Device,
+    HoldLatch,
+    build_device,
+    init_keyboard_listener,
+)
+from .isaac_teleop import IsaacTeleopConfig
+
+
+@dataclass
+class RecordConfig:
+    """CLI config for Isaac Teleop -> SO-101 dataset recording.
+
+    ``--robot.*`` / ``--teleop.*`` / ``--dataset.*`` configure the follower, device, and
+    recording; the loop/launch knobs below carry the same ``[xr]`` / ``[leader]`` tags as
+    ``teleoperate.py``. Use ``--flag=false`` for booleans (draccus style).
+    """
+
+    robot: RobotConfig
+    # --teleop.type=xr_controller|so101_leader, resolved against IsaacTeleopConfig's registry.
+    teleop: IsaacTeleopConfig
+    dataset: DatasetRecordConfig
+
+    # [leader] Path to the so101_leader plugin binary to spawn after CloudXR is up (it then
+    # inherits the runtime env). None (default) -> assume the plugin already runs externally.
+    launch_plugin: str | None = None
+
+    # [xr] Slew all joints to the reset pose before the first episode (--reset_to_origin=false to
+    # keep the arm where it is). After the slew the clutch seeds its home from the measured pose.
+    reset_to_origin: bool = True
+    # [xr] Duration [s] of the reset-to-origin slew (passed through to setup_xr).
+    reset_duration: float = RESET_DURATION_S
+
+    # [leader] Slew the follower to the leader's first pose before mirroring (--align=false to
+    # begin the 1:1 mirror immediately; the follower may snap).
+    align: bool = True
+    # [leader] Duration [s] of the startup alignment slew.
+    align_duration: float = ALIGN_DURATION_S
+
+    # Resume recording on an existing (previously interrupted) dataset.
+    resume: bool = False
+
+
+@safe_stop_image_writer
+def _record_loop(
+    robot,
+    device: Device,
+    motor_names: list[str],
+    events: dict,
+    fps: int,
+    dataset: LeRobotDataset | None = None,
+    control_time_s: float = 0.0,
+    single_task: str | None = None,
+) -> None:
+    """Run one episode (or reset phase) of the control loop.
+
+    When ``dataset`` is None the loop still controls the robot (so the operator
+    can reposition the arm during the reset window) but does not record frames.
+    """
+    control_interval = 1.0 / fps
+    timestamp = 0.0
+    start_t = time.perf_counter()
+    record_frames = dataset is not None
+    hold = HoldLatch(motor_names)
+
+    while timestamp < control_time_s:
+        loop_start = time.perf_counter()
+
+        if events["exit_early"]:
+            events["exit_early"] = False
+            break
+
+        obs = robot.get_observation()
+
+        if record_frames:
+            observation_frame = build_dataset_frame(dataset.features, obs, prefix=OBS_STR)
+
+        # Device idle (XR clutch disengaged, or leader stream stale) -> hold the pose
+        # latched on the active->idle edge.
+        action = hold.resolve(device.compute(obs), obs)
+
+        robot.send_action(action)
+
+        if record_frames:
+            action_frame = build_dataset_frame(dataset.features, action, prefix=ACTION)
+            dataset.add_frame({**observation_frame, **action_frame, "task": single_task})
+
+        dt_s = time.perf_counter() - loop_start
+        precise_sleep(max(control_interval - dt_s, 0.0))
+        timestamp = time.perf_counter() - start_t
+
+
+@parser.wrap()
+def record(cfg: RecordConfig) -> LeRobotDataset:
+    init_logging()
+    logging.info(pformat(asdict(cfg)))
+
+    # Connect the follower, build the selected Isaac device, and run its pre-loop startup
+    # (reset slew / leader align) — shared with teleoperate.py.
+    robot, device, motor_names = build_device(cfg)
+
+    # Build dataset feature spec.  The IK pipeline lives inside device.compute(), so the
+    # action features are exactly robot.action_features (joint positions in degrees).
+    teleop_proc, _, obs_proc = make_default_processors()
+    dataset_features = combine_feature_dicts(
+        aggregate_pipeline_dataset_features(
+            pipeline=teleop_proc,
+            initial_features=create_initial_features(action=robot.action_features),
+            use_videos=cfg.dataset.video,
+        ),
+        aggregate_pipeline_dataset_features(
+            pipeline=obs_proc,
+            initial_features=create_initial_features(observation=robot.observation_features),
+            use_videos=cfg.dataset.video,
+        ),
+    )
+
+    num_cameras = len(robot.cameras) if hasattr(robot, "cameras") else 0
+    image_writer_threads = cfg.dataset.num_image_writer_threads_per_camera * num_cameras
+
+    dataset: LeRobotDataset | None = None
+    listener = None
+    try:
+        if cfg.resume:
+            dataset = LeRobotDataset.resume(
+                cfg.dataset.repo_id,
+                root=cfg.dataset.root,
+                batch_encoding_size=cfg.dataset.video_encoding_batch_size,
+                rgb_encoder=cfg.dataset.rgb_encoder,
+                depth_encoder=cfg.dataset.depth_encoder,
+                encoder_threads=cfg.dataset.encoder_threads,
+                streaming_encoding=cfg.dataset.streaming_encoding,
+                encoder_queue_maxsize=cfg.dataset.encoder_queue_maxsize,
+                image_writer_processes=cfg.dataset.num_image_writer_processes if num_cameras > 0 else 0,
+                image_writer_threads=image_writer_threads if num_cameras > 0 else 0,
+            )
+            sanity_check_dataset_robot_compatibility(dataset, robot, cfg.dataset.fps, dataset_features)
+        else:
+            cfg.dataset.stamp_repo_id()
+            dataset = LeRobotDataset.create(
+                cfg.dataset.repo_id,
+                cfg.dataset.fps,
+                root=cfg.dataset.root,
+                robot_type=robot.name,
+                features=dataset_features,
+                use_videos=cfg.dataset.video,
+                image_writer_processes=cfg.dataset.num_image_writer_processes,
+                image_writer_threads=image_writer_threads,
+                batch_encoding_size=cfg.dataset.video_encoding_batch_size,
+                rgb_encoder=cfg.dataset.rgb_encoder,
+                depth_encoder=cfg.dataset.depth_encoder,
+                encoder_threads=cfg.dataset.encoder_threads,
+                streaming_encoding=cfg.dataset.streaming_encoding,
+                encoder_queue_maxsize=cfg.dataset.encoder_queue_maxsize,
+            )
+
+        listener, events = init_keyboard_listener()
+
+        loop_kwargs = {
+            "robot": robot,
+            "device": device,
+            "motor_names": motor_names,
+            "events": events,
+            "fps": cfg.dataset.fps,
+            "single_task": cfg.dataset.single_task,
+        }
+
+        with VideoEncodingManager(dataset):
+            recorded_episodes = 0
+            while recorded_episodes < cfg.dataset.num_episodes and not events["stop_recording"]:
+                logging.info(f"Recording episode {dataset.num_episodes}")
+                _record_loop(
+                    **loop_kwargs,
+                    dataset=dataset,
+                    control_time_s=cfg.dataset.episode_time_s,
+                )
+
+                # Reset window: give the operator time to reposition the scene.
+                # Skipped for the last episode (or if stop_recording was set).
+                if not events["stop_recording"] and (
+                    recorded_episodes < cfg.dataset.num_episodes - 1 or events["rerecord_episode"]
+                ):
+                    logging.info("Reset the environment")
+                    _record_loop(
+                        **loop_kwargs,
+                        dataset=None,
+                        control_time_s=cfg.dataset.reset_time_s,
+                    )
+
+                if events["rerecord_episode"]:
+                    logging.info("Re-record episode")
+                    events["rerecord_episode"] = False
+                    events["exit_early"] = False
+                    dataset.clear_episode_buffer()
+                    continue
+
+                dataset.save_episode()
+                recorded_episodes += 1
+
+    finally:
+        logging.info("Stop recording")
+
+        # Hardware teardown FIRST, each step guarded: the arm must be freed promptly (not
+        # after a potentially long finalize/encode), a cleanup failure must not skip the
+        # follower disconnect (which is what disables torque), and neither must prevent
+        # the dataset from being finalized below.
+        try:
+            device.cleanup()
+        except Exception:
+            logging.exception("Device cleanup failed")
+        try:
+            if robot.is_connected:
+                robot.disconnect()
+        except Exception:
+            logging.exception("Robot disconnect failed")
+
+        # Restore the terminal before the (potentially long) finalize/encode.
+        if listener is not None:
+            try:
+                listener.stop()
+            except Exception:
+                logging.exception("Keyboard listener stop failed")
+
+        if dataset is not None:
+            dataset.finalize()
+
+        if cfg.dataset.push_to_hub:
+            if dataset is not None and dataset.num_episodes > 0:
+                dataset.push_to_hub(tags=cfg.dataset.tags, private=cfg.dataset.private)
+            else:
+                logging.warning("No episodes saved — skipping push to hub")
+
+        logging.info("Exiting")
+
+    return dataset
+
+
+def main():
+    record()
+
+
+if __name__ == "__main__":
+    main()

examples/isaac_teleop_to_so101/teleoperate.py

@@ -0,0 +1,117 @@
+#!/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.
+
+"""Teleoperate an SO-101 follower arm via NVIDIA Isaac Teleop.
+
+``lerobot-teleoperate``-style CLI (draccus): ``--teleop.type`` selects the Isaac device
+(``xr_controller`` | ``so101_leader``), ``--robot.*`` the follower::
+
+    # XR (VR) controller: clutch + soft-orientation IK
+    python -m examples.isaac_teleop_to_so101.teleoperate --robot.type=so101_follower \
+        --robot.port=/dev/ttyACM0 --robot.id=so101_follower_arm --teleop.type=xr_controller
+
+    # SO-101 leader arm: 1:1 joint mirror (real leader on /dev/ttyACM1)
+    python -m examples.isaac_teleop_to_so101.teleoperate --robot.type=so101_follower \
+        --robot.port=/dev/ttyACM0 --robot.id=so101_follower_arm --teleop.type=so101_leader \
+        --teleop.port=/dev/ttyACM1 --teleop.id=so101_leader_arm \
+        --launch_plugin=/code/Teleop/install/plugins/so101_leader/so101_leader_plugin
+
+``--teleop.type`` resolves against the Isaac device registry (see :class:`IsaacTeleopConfig`),
+distinct from the serial ``so101_leader``. The pipelines, clutch/IK/align internals, and
+reset-pose behavior live in ``common.py``. Requires the ``isaacteleop`` package and an OpenXR
+runtime (install instructions in this folder's ``README.md``).
+"""
+
+import time
+from dataclasses import dataclass
+
+from lerobot.configs import parser
+from lerobot.robots import RobotConfig
+from lerobot.robots.so_follower import SOFollowerConfig  # noqa: F401  (registers so101_follower)
+from lerobot.utils.robot_utils import precise_sleep
+
+from .common import (
+    ALIGN_DURATION_S,
+    FPS,
+    RESET_DURATION_S,
+    HoldLatch,
+    build_device,
+)
+from .isaac_teleop import IsaacTeleopConfig
+
+
+@dataclass
+class TeleoperateConfig:
+    """``lerobot-teleoperate``-style CLI for the Isaac Teleop -> SO-101 example.
+
+    The fields below are the loop/launch knobs (not part of either device's config); the
+    ``[xr]`` / ``[leader]`` tags mark which device a knob applies to. Use ``--flag=false``
+    for booleans (draccus style).
+    """
+
+    # Isaac Teleop input device + its knobs (--teleop.type=xr_controller|so101_leader,
+    # then --teleop.<field>=...). Resolved against IsaacTeleopConfig's own choice registry.
+    teleop: IsaacTeleopConfig
+    # SO-101 FOLLOWER arm (--robot.type=so101_follower --robot.port=/dev/ttyACM0 --robot.id=...).
+    robot: RobotConfig
+
+    # [leader] Path to the so101_leader plugin binary to spawn AFTER CloudXR is up (it then
+    # inherits the runtime env). None (default) -> assume the plugin already runs externally.
+    # The leader's serial port is --teleop.port (forwarded to the plugin; empty -> synthetic).
+    launch_plugin: str | None = None
+
+    # [xr] Slew all joints to a default reset pose before the loop (--reset_to_origin=false to
+    # keep the arm where it is). After the slew the clutch seeds its home from the measured pose.
+    reset_to_origin: bool = True
+    # [xr] Duration [s] of the reset-to-origin slew.
+    reset_duration: float = RESET_DURATION_S
+
+    # [leader] Slew the follower to the leader's first pose before mirroring (--align=false to
+    # begin the 1:1 mirror immediately; the follower may snap).
+    align: bool = True
+    # [leader] Duration [s] of the startup alignment slew.
+    align_duration: float = ALIGN_DURATION_S
+
+
+@parser.wrap()
+def teleoperate(cfg: TeleoperateConfig):
+    robot, device, motor_names = build_device(cfg)
+    hold = HoldLatch(motor_names)
+    try:
+        while True:
+            t0 = time.perf_counter()
+            obs = robot.get_observation()
+            # Idle (compute() -> None) holds the pose latched on the active->idle edge.
+            action = hold.resolve(device.compute(obs), obs)
+            robot.send_action(action)
+            precise_sleep(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))
+    except KeyboardInterrupt:
+        pass
+    finally:
+        # A failing device cleanup must not skip the follower disconnect (which is what
+        # disables torque on the arm).
+        try:
+            device.cleanup()
+        finally:
+            robot.disconnect()
+
+
+def main():
+    teleoperate()
+
+
+if __name__ == "__main__":
+    main()

examples/lekiwi/evaluate.py

@@ -17,7 +17,7 @@
 import logging
 import time
 
-from lerobot.common.control_utils import init_keyboard_listener, predict_action
+from lerobot.common.control_utils import predict_action
 from lerobot.datasets import LeRobotDataset
 from lerobot.policies import make_pre_post_processors
 from lerobot.policies.act import ACTPolicy
@@ -26,6 +26,7 @@ from lerobot.processor import make_default_processors
 from lerobot.robots.lekiwi import LeKiwiClient, LeKiwiClientConfig
 from lerobot.utils.constants import ACTION, OBS_STR
 from lerobot.utils.feature_utils import build_dataset_frame, hw_to_dataset_features
+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 init_rerun, log_rerun_data

examples/lekiwi/record.py

@@ -14,7 +14,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from lerobot.common.control_utils import init_keyboard_listener
 from lerobot.datasets import LeRobotDataset
 from lerobot.processor import make_default_processors
 from lerobot.robots.lekiwi import LeKiwiClient, LeKiwiClientConfig
@@ -23,6 +22,7 @@ from lerobot.teleoperators.keyboard import KeyboardTeleop, KeyboardTeleopConfig
 from lerobot.teleoperators.so_leader import SO100Leader, SO100LeaderConfig
 from lerobot.utils.constants import ACTION, OBS_STR
 from lerobot.utils.feature_utils import hw_to_dataset_features
+from lerobot.utils.keyboard_input import init_keyboard_listener
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun
 

examples/phone_to_so100/evaluate.py

@@ -18,7 +18,7 @@ import logging
 import time
 
 from lerobot.cameras.opencv import OpenCVCameraConfig
-from lerobot.common.control_utils import init_keyboard_listener, predict_action
+from lerobot.common.control_utils import predict_action
 from lerobot.configs import FeatureType, PolicyFeature
 from lerobot.datasets import LeRobotDataset, aggregate_pipeline_dataset_features, create_initial_features
 from lerobot.model.kinematics import RobotKinematics
@@ -41,6 +41,7 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
 from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.constants import ACTION, OBS_STR
 from lerobot.utils.feature_utils import build_dataset_frame, combine_feature_dicts
+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 init_rerun, log_rerun_data

examples/phone_to_so100/record.py

@@ -15,7 +15,6 @@
 # limitations under the License.
 
 from lerobot.cameras.opencv import OpenCVCameraConfig
-from lerobot.common.control_utils import init_keyboard_listener
 from lerobot.datasets import LeRobotDataset, aggregate_pipeline_dataset_features, create_initial_features
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.processor import (
@@ -39,6 +38,7 @@ from lerobot.teleoperators.phone.config_phone import PhoneOS
 from lerobot.teleoperators.phone.phone_processor import MapPhoneActionToRobotAction
 from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.feature_utils import combine_feature_dicts
+from lerobot.utils.keyboard_input import init_keyboard_listener
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun
 

examples/so100_to_so100_EE/evaluate.py

@@ -18,7 +18,7 @@ import logging
 import time
 
 from lerobot.cameras.opencv import OpenCVCameraConfig
-from lerobot.common.control_utils import init_keyboard_listener, predict_action
+from lerobot.common.control_utils import predict_action
 from lerobot.configs import FeatureType, PolicyFeature
 from lerobot.datasets import LeRobotDataset, aggregate_pipeline_dataset_features, create_initial_features
 from lerobot.model.kinematics import RobotKinematics
@@ -41,6 +41,7 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
 from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.constants import ACTION, OBS_STR
 from lerobot.utils.feature_utils import build_dataset_frame, combine_feature_dicts
+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 init_rerun, log_rerun_data

examples/so100_to_so100_EE/record.py

@@ -16,7 +16,6 @@
 
 
 from lerobot.cameras.opencv import OpenCVCameraConfig
-from lerobot.common.control_utils import init_keyboard_listener
 from lerobot.datasets import LeRobotDataset, aggregate_pipeline_dataset_features, create_initial_features
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.processor import (
@@ -36,6 +35,7 @@ from lerobot.scripts.lerobot_record import record_loop
 from lerobot.teleoperators.so_leader import SO100Leader, SO100LeaderConfig
 from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.feature_utils import combine_feature_dicts
+from lerobot.utils.keyboard_input import init_keyboard_listener
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun
 

pyproject.toml

@@ -25,7 +25,7 @@ discord = "https://discord.gg/s3KuuzsPFb"
 
 [project]
 name = "lerobot"
-version = "0.5.2"
+version = "0.6.1"
 description = "🤗 LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch"
 dynamic = ["readme"]
 license = { text = "Apache-2.0" }
@@ -124,7 +124,8 @@ hardware = [
     "lerobot[deepdiff-dep]",
 ]
 viz = [
-    "rerun-sdk>=0.24.0,<0.27.0",
+    "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.
@@ -140,7 +141,14 @@ av-dep = ["av>=15.0.0,<16.0.0"]
 pygame-dep = ["pygame>=2.5.1,<2.7.0"]
 # NOTE: 0.9.16 links against liburdfdom_sensor.so.4, which is unavailable on Ubuntu 24.04
 # (noble ships urdfdom 3.x). Cap below 0.9.16 until system urdfdom 4.x is broadly available.
-placo-dep = ["placo>=0.9.6,<0.9.16"]
+#
+# NOTE: placo pulls in pin (Pinocchio), whose binary wheels dlopen specific cmeel sonames
+# (liburdfdom_sensor.so.4.0, libtinyxml2.so.10) but declare only `>=` floors on their cmeel
+# packages. The 2026-05-21 major bumps (cmeel-urdfdom 6.0.0 -> .so.6, cmeel-tinyxml2 11.0.0
+# -> .so.11) ship newer sonames, so left unpinned the resolver grabs them and `import placo`
+# fails at load with "liburdfdom_sensor.so.4.0: cannot open shared object file" (see #3755).
+# There is no cmeel-urdfdom 5.x; <5 selects the 4.x ABI the placo/pin wheels are built against.
+placo-dep = ["placo>=0.9.6,<0.9.16", "cmeel-urdfdom>=4,<5", "cmeel-tinyxml2<11"]
 transformers-dep = ["transformers>=5.4.0,<5.6.0"]
 grpcio-dep = ["grpcio>=1.73.1,<2.0.0", "protobuf>=6.31.1,<8.0.0"]
 accelerate-dep = ["accelerate>=1.14.0,<2.0.0"]
@@ -156,6 +164,7 @@ 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]"]
@@ -211,19 +220,24 @@ 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",
-    "timm>=1.0.0,<1.1.0",
+    "lerobot[timm-dep]",
     "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]"]
 topreward = ["lerobot[transformers-dep]"]
 xvla = ["lerobot[transformers-dep]"]
 eo1 = ["lerobot[transformers-dep]", "lerobot[qwen-vl-utils-dep]"]
+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]"]
@@ -301,10 +315,13 @@ all = [
     "lerobot[pi]",
     "lerobot[molmoact2]",
     "lerobot[smolvla]",
-    # "lerobot[groot]", TODO(Steven): Gr00t requires specific installation instructions for flash-attn
+    "lerobot[fastwam]",
+    "lerobot[groot]",
     "lerobot[xvla]",
+    "lerobot[evo1]",
     "lerobot[hilserl]",
     "lerobot[vla_jepa]",
+    "lerobot[lingbot_va]",
     "lerobot[async]",
     "lerobot[dev]",
     "lerobot[test]",
@@ -355,8 +372,6 @@ explicit = true
 [tool.uv.sources]
 torch = [{ index = "pytorch-cu128", marker = "sys_platform == 'linux'" }]
 torchvision = [{ index = "pytorch-cu128", marker = "sys_platform == 'linux'" }]
-huggingface-hub = { git = "https://github.com/huggingface/huggingface_hub.git", branch = "feat/hffs-cache-cdn-range-reads" }
-datasets = { git = "https://github.com/huggingface/datasets.git", branch = "main" }
 
 [tool.setuptools.package-data]
 lerobot = ["envs/*.json", "annotations/steerable_pipeline/prompts/*.txt"]
@@ -423,7 +438,6 @@ exclude_dirs = [
 skips = ["B101", "B311", "B404", "B603", "B615"]
 
 [tool.typos]
-default.extend-words = { trak = "trak" }
 default.extend-ignore-re = [
     "(?Rm)^.*(#|//)\\s*spellchecker:disable-line$",                      # spellchecker:disable-line
     "(?s)(#|//)\\s*spellchecker:off.*?\\n\\s*(#|//)\\s*spellchecker:on", # spellchecker:<on|off>
@@ -440,7 +454,8 @@ default.extend-ignore-identifiers-re = [
     "is_compileable",
     "ROBOTIS",
     "OT_VALUE",
-    "VanderBilt"
+    "VanderBilt",
+    "seperated_timestep",
 ]
 
 # TODO: Uncomment when ready to use

requirements-macos.txt

@@ -1,729 +0,0 @@
-#
-# 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

requirements-ubuntu.txt

@@ -1,882 +0,0 @@
-#
-# 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

requirements.in

@@ -1,9 +0,0 @@
-# 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]

scripts/build_mp4_sidecar.py

@@ -1,93 +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
-
-from __future__ import annotations
-
-import argparse
-import time
-from pathlib import Path
-
-import fsspec
-
-from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
-from lerobot.datasets.episode_video_streaming import EpisodeVideoManifest, assert_hf_hub_range_cache_branch
-
-DEFAULT_REPO = "allenai/MolmoAct2-BimanualYAM-Dataset"
-DEFAULT_REVISION = "e9f21ae15074330839f2ac25ed4b49d76dfa1f9c"
-DEFAULT_DATA_ROOT = "hf://buckets/pepijn223/MolmoAct2-BimanualYAM-Dataset-bucket"
-
-
-def parse_args() -> argparse.Namespace:
-    parser = argparse.ArgumentParser(description="Build a reusable MP4 byte-index sidecar for streaming.")
-    parser.add_argument("--repo-id", default=DEFAULT_REPO)
-    parser.add_argument("--revision", default=DEFAULT_REVISION)
-    parser.add_argument("--data-root", default=DEFAULT_DATA_ROOT)
-    parser.add_argument("--output", required=True)
-    parser.add_argument("--episodes", type=int, default=None)
-    parser.add_argument("--workers", type=int, default=8)
-    parser.add_argument("--range-backend", choices=("fsspec", "native-http"), default="native-http")
-    parser.add_argument("--max-probe-mb", type=int, default=64)
-    parser.add_argument(
-        "--no-push", action="store_true", help="Do not upload the sidecar to data_root/meta/mp4-sidecars."
-    )
-    parser.add_argument("--no-hub-branch-assert", action="store_true")
-    return parser.parse_args()
-
-
-def push_sidecar(local_path: str, data_root: str) -> list[str]:
-    if not data_root.startswith("hf://"):
-        return []
-
-    local = Path(local_path)
-    fs = fsspec.filesystem("hf")
-    remote_dir = f"{data_root.rstrip('/')}/meta/mp4-sidecars"
-    remote_paths = [f"{remote_dir}/{local.name}"]
-
-    for remote in remote_paths:
-        fs.put(str(local), remote)
-    return remote_paths
-
-
-def main() -> None:
-    args = parse_args()
-    if args.data_root.startswith("hf://") and not args.no_hub_branch_assert:
-        assert_hf_hub_range_cache_branch()
-
-    meta = LeRobotDatasetMetadata(args.repo_id, revision=args.revision)
-    meta.ensure_readable()
-    total = (
-        int(meta.total_episodes) if args.episodes is None else min(args.episodes, int(meta.total_episodes))
-    )
-    rel_paths = sorted(
-        {str(meta.get_video_file_path(ep_idx, key)) for ep_idx in range(total) for key in meta.video_keys}
-    )
-
-    start = time.perf_counter()
-    EpisodeVideoManifest.write_file_sidecar(
-        args.output,
-        rel_paths,
-        args.data_root,
-        range_backend=args.range_backend,
-        workers=args.workers,
-        max_probe_bytes=args.max_probe_mb * 1024 * 1024,
-    )
-    elapsed = time.perf_counter() - start
-    print(f"wrote {args.output}")
-    print(f"episodes={total} files={len(rel_paths)} elapsed_s={elapsed:.2f}")
-    if args.no_push:
-        print("push_skipped: --no-push")
-    else:
-        pushed = push_sidecar(args.output, args.data_root)
-        for remote in pushed:
-            print(f"pushed {remote}")
-
-
-if __name__ == "__main__":
-    main()

scripts/summarize_episode_pool_bench.py

@@ -1,65 +0,0 @@
-#!/usr/bin/env python
-
-from __future__ import annotations
-
-import argparse
-import json
-from pathlib import Path
-
-
-def parse_args() -> argparse.Namespace:
-    parser = argparse.ArgumentParser(description="Summarize distributed episode pool benchmark JSON files.")
-    parser.add_argument("summaries", nargs="+", help="Rank summary JSON files.")
-    return parser.parse_args()
-
-
-def _load(path: str) -> dict:
-    return json.loads(Path(path).read_text())
-
-
-def _fmt(value: float) -> str:
-    return f"{value:.1f}"
-
-
-def main() -> None:
-    args = parse_args()
-    rows = [_load(path) for path in args.summaries]
-    rows.sort(key=lambda row: int(row.get("distributed_shard_index", 0)))
-    total_bytes = sum(float(row.get("fetch_bytes", 0.0)) for row in rows)
-    max_fetch_s = max(float(row.get("fetch_s", 0.0)) for row in rows)
-    aggregate_mib_s = total_bytes / max_fetch_s / 1024**2 if max_fetch_s > 0 else float("inf")
-    summed_rank_mib_s = sum(float(row.get("fetch_mib_s", 0.0)) for row in rows)
-    total_decode_samples_s = sum(float(row.get("pool_decode_training_samples_s", 0.0)) for row in rows)
-    total_stream_samples_s = sum(float(row.get("pool_stream_actual_samples_s", 0.0)) for row in rows)
-    kept_up = all(bool(row.get("pool_stream_kept_up", 0.0)) for row in rows)
-
-    print("| Aggregate | value |")
-    print("|---|---:|")
-    print(f"| ranks | {len(rows)} |")
-    print(f"| total fetched GiB | {total_bytes / 1024**3:.2f} |")
-    print(f"| aggregate fetch MiB/s | {_fmt(aggregate_mib_s)} |")
-    print(f"| summed rank fetch MiB/s | {_fmt(summed_rank_mib_s)} |")
-    if total_decode_samples_s:
-        print(f"| aggregate resident decode samples/s | {_fmt(total_decode_samples_s)} |")
-    if total_stream_samples_s:
-        print(f"| aggregate stream samples/s | {_fmt(total_stream_samples_s)} |")
-        print(f"| all ranks kept up | {'yes' if kept_up else 'no'} |")
-
-    print()
-    print("| Rank | host | fetch MiB/s | fetch s | GiB | decode samples/s | stream samples/s | kept up |")
-    print("|---:|---|---:|---:|---:|---:|---:|---|")
-    for row in rows:
-        rank = int(row.get("distributed_shard_index", 0))
-        print(
-            f"| {rank} | {row.get('hostname', '')} | "
-            f"{_fmt(float(row.get('fetch_mib_s', 0.0)))} | "
-            f"{_fmt(float(row.get('fetch_s', 0.0)))} | "
-            f"{float(row.get('fetch_gib', 0.0)):.2f} | "
-            f"{_fmt(float(row.get('pool_decode_training_samples_s', 0.0)))} | "
-            f"{_fmt(float(row.get('pool_stream_actual_samples_s', 0.0)))} | "
-            f"{'yes' if row.get('pool_stream_kept_up', 0.0) else 'no'} |"
-        )
-
-
-if __name__ == "__main__":
-    main()

src/lerobot/annotations/steerable_pipeline/frames.py

@@ -36,7 +36,7 @@ from typing import Any, Protocol
 import PIL.Image
 import torch
 
-from lerobot.configs.video import VideoEncoderConfig
+from lerobot.configs import RGBEncoderConfig
 from lerobot.datasets.video_utils import decode_video_frames, reencode_video
 
 from .reader import EpisodeRecord, snap_to_frame
@@ -164,7 +164,9 @@ class VideoFrameProvider:
         # only for video-stored cameras. Image-stored cameras (also in
         # ``camera_keys``) would KeyError, so restrict the list — and the
         # default — to video keys.
-        keys = list(self._meta.video_keys)
+        # Depth cameras are excluded from the annotation pipeline for now.
+        depth_keys = set(self._meta.depth_keys)
+        keys = [key for key in self._meta.video_keys if key not in depth_keys]
         # Last-resort fallback: if metadata didn't surface any video keys but
         # the caller explicitly named a camera (``--vlm.camera_key=...``),
         # trust them — the key is by definition known to exist on the dataset.
@@ -276,12 +278,12 @@ class VideoFrameProvider:
         from_timestamp = float(ep[f"videos/{self.camera_key}/from_timestamp"])
         to_timestamp = float(ep[f"videos/{self.camera_key}/to_timestamp"])
         src = self.root / self._meta.get_video_file_path(record.episode_index, self.camera_key)
-        encoder = VideoEncoderConfig(vcodec="h264", pix_fmt="yuv420p", g=None, crf=23, preset="ultrafast")
+        encoder = RGBEncoderConfig(vcodec="h264", pix_fmt="yuv420p", g=None, crf=23, preset="ultrafast")
         try:
             reencode_video(
                 src,
                 out_path,
-                camera_encoder=encoder,
+                video_encoder=encoder,
                 overwrite=True,
                 start_time_s=from_timestamp,
                 end_time_s=to_timestamp,

src/lerobot/annotations/steerable_pipeline/writer.py

@@ -54,6 +54,7 @@ from typing import Any
 import pyarrow as pa
 import pyarrow.parquet as pq
 
+from lerobot.datasets.io_utils import write_table_one_row_group_per_episode
 from lerobot.datasets.language import (
     EVENT_ONLY_STYLES,
     LANGUAGE_EVENTS,
@@ -274,12 +275,11 @@ class LanguageColumnsWriter:
         new_table = self._materialize_table(
             table, per_row_persistent, per_row_events, drop_old=self.drop_existing_subtask_index
         )
-        # Atomic replace: write to a sibling tmp path and rename so a crash
-        # mid-write can't leave a half-written shard that ``pq.read_table``
-        # would then fail to open. ``Path.replace`` is atomic on POSIX +
-        # Windows when source and target sit on the same filesystem.
+        # Re-emit one row group per episode (a bulk pq.write_table would collapse
+        # them into one). Write to a sibling tmp path and atomically rename so a
+        # crash mid-write can't leave a half-written shard.
         tmp_path = path.with_suffix(path.suffix + ".tmp")
-        pq.write_table(new_table, tmp_path)
+        write_table_one_row_group_per_episode(new_table, tmp_path)
         tmp_path.replace(path)
 
     def _materialize_table(

src/lerobot/async_inference/helpers.py

@@ -105,8 +105,9 @@ def raw_observation_to_observation(
 
 
 def prepare_image(image: torch.Tensor) -> torch.Tensor:
-    """Minimal preprocessing to turn int8 images to float32 in [0, 1], and create a memory-contiguous tensor"""
-    image = image.type(torch.float32) / 255
+    """Minimal preprocessing to turn RGB uint8 images to float32 in [0, 1], and create a memory-contiguous tensor"""
+    if image.dtype == torch.uint8:
+        image = image.type(torch.float32) / 255
     image = image.contiguous()
 
     return image

src/lerobot/cameras/opencv/camera_opencv.py

@@ -436,17 +436,18 @@ class OpenCVCamera(Camera):
         Internal loop run by the background thread for asynchronous reading.
 
         On each iteration:
-        1. Reads a color frame
+        1. Reads a color frame (blocking call)
         2. Stores result in latest_frame and updates timestamp (thread-safe)
         3. Sets new_frame_event to notify listeners
 
         Stops on DeviceNotConnectedError, logs other errors and continues.
         """
-        if self.stop_event is None:
+        stop_event = self.stop_event
+        if stop_event is None:
             raise RuntimeError(f"{self}: stop_event is not initialized before starting read loop.")
 
         failure_count = 0
-        while not self.stop_event.is_set():
+        while not stop_event.is_set():
             try:
                 raw_frame = self._read_from_hardware()
                 processed_frame = self._postprocess_image(raw_frame)
@@ -484,6 +485,8 @@ class OpenCVCamera(Camera):
 
         if self.thread is not None and self.thread.is_alive():
             self.thread.join(timeout=2.0)
+            if self.thread.is_alive():
+                logger.warning(f"{self} read thread did not terminate within timeout.")
 
         self.thread = None
         self.stop_event = None

src/lerobot/cameras/realsense/camera_realsense.py

@@ -128,6 +128,7 @@ class RealSenseCamera(Camera):
 
         self.fps = config.fps
         self.color_mode = config.color_mode
+        self.use_rgb = config.use_rgb
         self.use_depth = config.use_depth
         self.warmup_s = config.warmup_s
 
@@ -195,12 +196,15 @@ class RealSenseCamera(Camera):
         # NOTE(Steven/Caroline): Enforcing at least one second of warmup as RS cameras need a bit of time before the first read. If we don't wait, the first read from the warmup will raise.
         self.warmup_s = max(self.warmup_s, 1)
 
+        warmup_read = self.async_read if self.use_rgb else self.async_read_depth
         start_time = time.time()
         while time.time() - start_time < self.warmup_s:
-            self.async_read(timeout_ms=self.warmup_s * 1000)
+            warmup_read(timeout_ms=self.warmup_s * 1000)
             time.sleep(0.1)
         with self.frame_lock:
-            if self.latest_color_frame is None or self.use_depth and self.latest_depth_frame is None:
+            if (self.use_rgb and self.latest_color_frame is None) or (
+                self.use_depth and self.latest_depth_frame is None
+            ):
                 raise ConnectionError(f"{self} failed to capture frames during warmup.")
 
         logger.info(f"{self} connected.")
@@ -268,13 +272,13 @@ class RealSenseCamera(Camera):
             )
 
         if len(found_devices) > 1:
-            serial_numbers = [dev["serial_number"] for dev in found_devices]
+            serial_numbers = [dev["id"] for dev in found_devices]
             raise ValueError(
                 f"Multiple RealSense cameras found with name '{name}'. "
                 f"Please use a unique serial number instead. Found SNs: {serial_numbers}"
             )
 
-        serial_number = str(found_devices[0]["serial_number"])
+        serial_number = str(found_devices[0]["id"])
         return serial_number
 
     def _configure_rs_pipeline_config(self, rs_config: Any) -> None:
@@ -282,15 +286,17 @@ class RealSenseCamera(Camera):
         rs.config.enable_device(rs_config, self.serial_number)
 
         if self.width and self.height and self.fps:
-            rs_config.enable_stream(
-                rs.stream.color, self.capture_width, self.capture_height, rs.format.rgb8, self.fps
-            )
+            if self.use_rgb:
+                rs_config.enable_stream(
+                    rs.stream.color, self.capture_width, self.capture_height, rs.format.rgb8, self.fps
+                )
             if self.use_depth:
                 rs_config.enable_stream(
                     rs.stream.depth, self.capture_width, self.capture_height, rs.format.z16, self.fps
                 )
         else:
-            rs_config.enable_stream(rs.stream.color)
+            if self.use_rgb:
+                rs_config.enable_stream(rs.stream.color)
             if self.use_depth:
                 rs_config.enable_stream(rs.stream.depth)
 
@@ -298,8 +304,9 @@ class RealSenseCamera(Camera):
     def _configure_capture_settings(self) -> None:
         """Sets fps, width, and height from device stream if not already configured.
 
-        Uses the color stream profile to update unset attributes. Handles rotation by
-        swapping width/height when needed. Original capture dimensions are always stored.
+        Uses the color stream profile (or the depth stream profile when the color
+        stream is disabled) to update unset attributes. Handles rotation by swapping
+        width/height when needed. Original capture dimensions are always stored.
 
         Raises:
             DeviceNotConnectedError: If device is not connected.
@@ -308,7 +315,8 @@ class RealSenseCamera(Camera):
         if self.rs_profile is None:
             raise RuntimeError(f"{self}: rs_profile must be initialized before use.")
 
-        stream = self.rs_profile.get_stream(rs.stream.color).as_video_stream_profile()
+        rs_stream = rs.stream.color if self.use_rgb else rs.stream.depth
+        stream = self.rs_profile.get_stream(rs_stream).as_video_stream_profile()
 
         if self.fps is None:
             self.fps = stream.fps()
@@ -323,6 +331,14 @@ class RealSenseCamera(Camera):
                 self.width, self.height = actual_width, actual_height
                 self.capture_width, self.capture_height = actual_width, actual_height
 
+    def _read(self, read_depth: bool = False) -> NDArray[Any]:
+        """Shared helper for :meth:`read`/:meth:`read_depth`: wait for a fresh color or depth frame."""
+        if self.thread is None or not self.thread.is_alive():
+            raise RuntimeError(f"{self} read thread is not running.")
+
+        self.new_frame_event.clear()
+        return self._async_read(timeout_ms=10000, read_depth=read_depth)
+
     @check_if_not_connected
     def read_depth(self, timeout_ms: int = 200) -> NDArray[Any]:
         """
@@ -332,8 +348,8 @@ class RealSenseCamera(Camera):
         from the camera hardware via the RealSense pipeline.
 
         Returns:
-            np.ndarray: The depth map as a NumPy array (height, width)
-                  of type `np.uint16` (raw depth values in millimeters) and rotation.
+            np.ndarray: The depth map as a NumPy array (height, width, 1)
+                  of type `np.uint16` (raw depth values in millimeters).
 
         Raises:
             DeviceNotConnectedError: If the camera is not connected.
@@ -349,20 +365,7 @@ class RealSenseCamera(Camera):
                 f"Failed to capture depth frame '.read_depth()'. Depth stream is not enabled for {self}."
             )
 
-        if self.thread is None or not self.thread.is_alive():
-            raise RuntimeError(f"{self} read thread is not running.")
-
-        self.new_frame_event.clear()
-
-        _ = self.async_read(timeout_ms=10000)
-
-        with self.frame_lock:
-            depth_map = self.latest_depth_frame
-
-        if depth_map is None:
-            raise RuntimeError("No depth frame available. Ensure camera is streaming.")
-
-        return depth_map
+        return self._read(read_depth=True)
 
     def _read_from_hardware(self):
         if self.rs_pipeline is None:
@@ -405,12 +408,10 @@ class RealSenseCamera(Camera):
                 f"{self} read() timeout_ms parameter is deprecated and will be removed in future versions."
             )
 
-        if self.thread is None or not self.thread.is_alive():
-            raise RuntimeError(f"{self} read thread is not running.")
+        if not self.use_rgb:
+            raise RuntimeError(f"{self}: cannot read color — camera was configured with use_rgb=False.")
 
-        self.new_frame_event.clear()
-
-        frame = self.async_read(timeout_ms=10000)
+        frame = self._read()
 
         read_duration_ms = (time.perf_counter() - start_time) * 1e3
         logger.debug(f"{self} read took: {read_duration_ms:.1f}ms")
@@ -465,32 +466,38 @@ class RealSenseCamera(Camera):
         Internal loop run by the background thread for asynchronous reading.
 
         On each iteration:
-        1. Reads a color frame with 500ms timeout
-        2. Stores result in latest_frame and updates timestamp (thread-safe)
+        1. Reads a color/depth frame (blocking call with 10s timeout)
+        2. Stores result in latest_color_frame/latest_depth_frame and updates timestamp (thread-safe)
         3. Sets new_frame_event to notify listeners
 
         Stops on DeviceNotConnectedError, logs other errors and continues.
         """
-        if self.stop_event is None:
+        stop_event = self.stop_event
+        if stop_event is None:
             raise RuntimeError(f"{self}: stop_event is not initialized before starting read loop.")
 
         failure_count = 0
-        while not self.stop_event.is_set():
+        while not stop_event.is_set():
             try:
                 frame = self._read_from_hardware()
-                color_frame_raw = frame.get_color_frame()
-                color_frame = np.asanyarray(color_frame_raw.get_data())
-                processed_color_frame = self._postprocess_image(color_frame)
+
+                if self.use_rgb:
+                    color_frame_raw = frame.get_color_frame()
+                    color_frame = np.asanyarray(color_frame_raw.get_data())
+                    processed_color_frame = self._postprocess_image(color_frame)
 
                 if self.use_depth:
                     depth_frame_raw = frame.get_depth_frame()
                     depth_frame = np.asanyarray(depth_frame_raw.get_data())
                     processed_depth_frame = self._postprocess_image(depth_frame, depth_frame=True)
+                    if processed_depth_frame.ndim == 2:  # (H, W) -> (H, W, 1)
+                        processed_depth_frame = processed_depth_frame[..., np.newaxis]
 
                 capture_time = time.perf_counter()
 
                 with self.frame_lock:
-                    self.latest_color_frame = processed_color_frame
+                    if self.use_rgb:
+                        self.latest_color_frame = processed_color_frame
                     if self.use_depth:
                         self.latest_depth_frame = processed_depth_frame
                     self.latest_timestamp = capture_time
@@ -522,6 +529,8 @@ class RealSenseCamera(Camera):
 
         if self.thread is not None and self.thread.is_alive():
             self.thread.join(timeout=2.0)
+            if self.thread.is_alive():  # pragma: no cover
+                logger.warning(f"{self} read thread did not terminate within timeout.")
 
         self.thread = None
         self.stop_event = None
@@ -532,7 +541,26 @@ class RealSenseCamera(Camera):
             self.latest_timestamp = None
             self.new_frame_event.clear()
 
-    # NOTE(Steven): Missing implementation for depth for now
+    def _async_read(self, timeout_ms: float, read_depth: bool = False) -> NDArray[Any]:
+        """Shared helper for :meth:`async_read`/:meth:`async_read_depth`: return the latest buffered frame."""
+        if self.thread is None or not self.thread.is_alive():
+            raise RuntimeError(f"{self} read thread is not running.")
+
+        if not self.new_frame_event.wait(timeout=timeout_ms / 1000.0):
+            raise TimeoutError(
+                f"Timed out waiting for frame from camera {self} after {timeout_ms} ms. "
+                f"Read thread alive: {self.thread.is_alive()}."
+            )
+
+        with self.frame_lock:
+            frame = self.latest_depth_frame if read_depth else self.latest_color_frame
+            self.new_frame_event.clear()
+
+        if frame is None:
+            raise RuntimeError(f"Internal error: Event set but no frame available for {self}.")
+
+        return frame
+
     @check_if_not_connected
     def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
         """
@@ -557,25 +585,31 @@ class RealSenseCamera(Camera):
             RuntimeError: If the background thread died unexpectedly or another error occurs.
         """
 
+        if not self.use_rgb:
+            raise RuntimeError(f"{self}: cannot read color — camera was configured with use_rgb=False.")
+
+        return self._async_read(timeout_ms=timeout_ms)
+
+    def _read_latest(self, max_age_ms: int, read_depth: bool = False) -> NDArray[Any]:
+        """Shared helper for :meth:`read_latest`/:meth:`read_latest_depth`: peek the latest buffered frame."""
         if self.thread is None or not self.thread.is_alive():
             raise RuntimeError(f"{self} read thread is not running.")
 
-        if not self.new_frame_event.wait(timeout=timeout_ms / 1000.0):
-            raise TimeoutError(
-                f"Timed out waiting for frame from camera {self} after {timeout_ms} ms. "
-                f"Read thread alive: {self.thread.is_alive()}."
-            )
-
         with self.frame_lock:
-            frame = self.latest_color_frame
-            self.new_frame_event.clear()
+            frame = self.latest_depth_frame if read_depth else self.latest_color_frame
+            timestamp = self.latest_timestamp
 
-        if frame is None:
-            raise RuntimeError(f"Internal error: Event set but no frame available for {self}.")
+        if frame is None or timestamp is None:
+            raise RuntimeError(f"{self} has not captured any frames yet.")
+
+        age_ms = (time.perf_counter() - timestamp) * 1e3
+        if age_ms > max_age_ms:
+            raise TimeoutError(
+                f"{self} latest frame is too old: {age_ms:.1f} ms (max allowed: {max_age_ms} ms)."
+            )
 
         return frame
 
-    # NOTE(Steven): Missing implementation for depth for now
     @check_if_not_connected
     def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
         """Return the most recent (color) frame captured immediately (Peeking).
@@ -592,24 +626,48 @@ class RealSenseCamera(Camera):
             DeviceNotConnectedError: If the camera is not connected.
             RuntimeError: If the camera is connected but has not captured any frames yet.
         """
+        if not self.use_rgb:
+            raise RuntimeError(f"{self}: cannot read color — camera was configured with use_rgb=False.")
 
-        if self.thread is None or not self.thread.is_alive():
-            raise RuntimeError(f"{self} read thread is not running.")
+        return self._read_latest(max_age_ms=max_age_ms)
 
-        with self.frame_lock:
-            frame = self.latest_color_frame
-            timestamp = self.latest_timestamp
+    @check_if_not_connected
+    def async_read_depth(self, timeout_ms: float = 200) -> NDArray[np.uint16]:
+        """Read the latest depth frame asynchronously, in millimeters.
 
-        if frame is None or timestamp is None:
-            raise RuntimeError(f"{self} has not captured any frames yet.")
+        Mirrors :meth:`async_read` but returns the depth stream rather than the
+        color stream. Output is ``np.uint16`` of shape ``(H, W, 1)``, where each
+        pixel is the distance from the sensor in millimeters.
 
-        age_ms = (time.perf_counter() - timestamp) * 1e3
-        if age_ms > max_age_ms:
-            raise TimeoutError(
-                f"{self} latest frame is too old: {age_ms:.1f} ms (max allowed: {max_age_ms} ms)."
-            )
+        Raises:
+            DeviceNotConnectedError: If the camera is not connected.
+            RuntimeError: If ``use_depth`` is ``False`` for this camera, or if
+                the background read thread is not running.
+            TimeoutError: If no frame becomes available within ``timeout_ms``.
+        """
+        if not self.use_depth:
+            raise RuntimeError(f"{self}: cannot read depth — camera was configured with use_depth=False.")
 
-        return frame
+        return self._async_read(timeout_ms=timeout_ms, read_depth=True)
+
+    @check_if_not_connected
+    def read_latest_depth(self, max_age_ms: int = 500) -> NDArray[Any]:
+        """Return the most recent depth frame in millimeters (peeking).
+
+        Non-blocking counterpart of :meth:`read_latest` for the depth stream.
+        Output is ``np.uint16`` of shape ``(H, W, 1)``, where each pixel is the
+        distance from the sensor in millimeters.
+
+        Raises:
+            DeviceNotConnectedError: If the camera is not connected.
+            RuntimeError: If ``use_depth`` is ``False`` for this camera, or if
+                no depth frame has been captured yet.
+            TimeoutError: If the latest depth frame is older than ``max_age_ms``.
+        """
+        if not self.use_depth:
+            raise RuntimeError(f"{self}: cannot read depth — camera was configured with use_depth=False.")
+
+        return self._read_latest(max_age_ms=max_age_ms, read_depth=True)
 
     def disconnect(self) -> None:
         """

src/lerobot/cameras/realsense/configuration_realsense.py

@@ -42,12 +42,14 @@ class RealSenseCameraConfig(CameraConfig):
         height: Requested frame height in pixels for the color stream.
         serial_number_or_name: Unique serial number or human-readable name to identify the camera.
         color_mode: Color mode for image output (RGB or BGR). Defaults to RGB.
+        use_rgb: Whether to enable the color stream. Defaults to True.
         use_depth: Whether to enable depth stream. Defaults to False.
         rotation: Image rotation setting (0°, 90°, 180°, or 270°). Defaults to no rotation.
         warmup_s: Time reading frames before returning from connect (in seconds)
 
     Note:
         - Either name or serial_number must be specified.
+        - At least one of `use_rgb` or `use_depth` must be enabled.
         - Depth stream configuration (if enabled) will use the same FPS as the color stream.
         - The actual resolution and FPS may be adjusted by the camera to the nearest supported mode.
         - For `fps`, `width` and `height`, either all of them need to be set, or none of them.
@@ -55,6 +57,7 @@ class RealSenseCameraConfig(CameraConfig):
 
     serial_number_or_name: str
     color_mode: ColorMode = ColorMode.RGB
+    use_rgb: bool = True
     use_depth: bool = False
     rotation: Cv2Rotation = Cv2Rotation.NO_ROTATION
     warmup_s: int = 1
@@ -63,6 +66,9 @@ class RealSenseCameraConfig(CameraConfig):
         self.color_mode = ColorMode(self.color_mode)
         self.rotation = Cv2Rotation(self.rotation)
 
+        if not self.use_rgb and not self.use_depth:
+            raise ValueError("At least one of `use_rgb` or `use_depth` must be enabled.")
+
         values = (self.fps, self.width, self.height)
         if any(v is not None for v in values) and any(v is None for v in values):
             raise ValueError(

src/lerobot/cameras/zmq/camera_zmq.py

@@ -246,11 +246,12 @@ class ZMQCamera(Camera):
         """
         Internal loop run by the background thread for asynchronous reading.
         """
-        if self.stop_event is None:
+        stop_event = self.stop_event
+        if stop_event is None:
             raise RuntimeError(f"{self}: stop_event is not initialized.")
 
         failure_count = 0
-        while not self.stop_event.is_set():
+        while not stop_event.is_set():
             try:
                 frame = self._read_from_hardware()
                 capture_time = time.perf_counter()
@@ -292,6 +293,8 @@ class ZMQCamera(Camera):
 
         if self.thread is not None and self.thread.is_alive():
             self.thread.join(timeout=2.0)
+            if self.thread.is_alive():
+                logger.warning(f"{self} read thread did not terminate within timeout.")
 
         self.thread = None
         self.stop_event = None

src/lerobot/common/control_utils.py

@@ -17,12 +17,9 @@ from __future__ import annotations
 ########################################################################################
 # Utilities
 ########################################################################################
-import logging
 import time
-import traceback
 from contextlib import nullcontext
 from copy import copy
-from functools import cache
 from typing import TYPE_CHECKING, Any
 
 import numpy as np
@@ -43,34 +40,6 @@ from lerobot.robots import Robot
 from lerobot.types import PolicyAction
 
 
-@cache
-def is_headless():
-    """
-    Detects if the Python script is running in a headless environment (e.g., without a display).
-
-    This function attempts to import `pynput`, a library that requires a graphical environment.
-    If the import fails, it assumes the environment is headless. The result is cached to avoid
-    re-running the check.
-
-    Returns:
-        True if the environment is determined to be headless, False otherwise.
-    """
-    try:
-        import pynput  # noqa
-
-        return False
-    except Exception:
-        print(
-            "Error trying to import pynput. Switching to headless mode. "
-            "As a result, the video stream from the cameras won't be shown, "
-            "and you won't be able to change the control flow with keyboards. "
-            "For more info, see traceback below.\n"
-        )
-        traceback.print_exc()
-        print()
-        return True
-
-
 def predict_action(
     observation: dict[str, np.ndarray],
     policy: PreTrainedPolicy,
@@ -122,59 +91,6 @@ def predict_action(
     return action
 
 
-def init_keyboard_listener():
-    """
-    Initializes a non-blocking keyboard listener for real-time user interaction.
-
-    This function sets up a listener for specific keys (right arrow, left arrow, escape) to control
-    the program flow during execution, such as stopping recording or exiting loops. It gracefully
-    handles headless environments where keyboard listening is not possible.
-
-    Returns:
-        A tuple containing:
-        - The `pynput.keyboard.Listener` instance, or `None` if in a headless environment.
-        - A dictionary of event flags (e.g., `exit_early`) that are set by key presses.
-    """
-    # Allow to exit early while recording an episode or resetting the environment,
-    # by tapping the right arrow key '->'. This might require a sudo permission
-    # to allow your terminal to monitor keyboard events.
-    events = {}
-    events["exit_early"] = False
-    events["rerecord_episode"] = False
-    events["stop_recording"] = False
-
-    if is_headless():
-        logging.warning(
-            "Headless environment detected. On-screen cameras display and keyboard inputs will not be available."
-        )
-        listener = None
-        return listener, events
-
-    # Only import pynput if not in a headless environment
-    from pynput import keyboard
-
-    def on_press(key):
-        try:
-            if key == keyboard.Key.right:
-                print("Right arrow key pressed. Exiting loop...")
-                events["exit_early"] = True
-            elif key == keyboard.Key.left:
-                print("Left arrow key pressed. Exiting loop and rerecord the last episode...")
-                events["rerecord_episode"] = True
-                events["exit_early"] = True
-            elif key == keyboard.Key.esc:
-                print("Escape key pressed. Stopping data recording...")
-                events["stop_recording"] = True
-                events["exit_early"] = True
-        except Exception as e:
-            print(f"Error handling key press: {e}")
-
-    listener = keyboard.Listener(on_press=on_press)
-    listener.start()
-
-    return listener, events
-
-
 def sanity_check_dataset_name(repo_id, policy_cfg):
     """
     Validates the dataset repository name against the presence of a policy configuration.

src/lerobot/common/train_utils.py

@@ -15,12 +15,14 @@
 # limitations under the License.
 from pathlib import Path
 
+from huggingface_hub import HfApi, snapshot_download
 from torch.optim import Optimizer
 from torch.optim.lr_scheduler import LRScheduler
 
 from lerobot.configs.train import TrainPipelineConfig
 from lerobot.optim import (
     load_optimizer_state,
+    load_optimizer_state_dict,
     load_scheduler_state,
     save_optimizer_state,
     save_scheduler_state,
@@ -34,6 +36,7 @@ from lerobot.utils.constants import (
     TRAINING_STATE_DIR,
     TRAINING_STEP,
 )
+from lerobot.utils.hub import find_latest_hub_checkpoint
 from lerobot.utils.io_utils import load_json, write_json
 from lerobot.utils.random_utils import load_rng_state, save_rng_state
 
@@ -98,6 +101,8 @@ def save_checkpoint(
     postprocessor: PolicyProcessorPipeline | None = None,
     num_processes: int | None = None,
     batch_size: int | None = None,
+    model_state_dict: dict | None = None,
+    optim_state_dict: dict | None = None,
 ) -> None:
     """This function creates the following directory structure:
 
@@ -127,9 +132,18 @@ def save_checkpoint(
             resume. Defaults to None (not recorded).
         batch_size (int | None, optional): Per-process batch size to record for sample-exact
             resume. Defaults to None (not recorded).
+        model_state_dict: Pre-gathered full (unsharded) model state dict. Required under FSDP,
+            where `policy.state_dict()` would return sharded tensors; the caller gathers it via a
+            cross-rank collective and passes it here so rank 0 can write it directly. It holds
+            FSDP's fp32 master weights and is saved as-is (the loader casts to the policy dtype on
+            read). When None (DDP / single-GPU), the model is saved the normal way. Defaults to None.
+        optim_state_dict: Pre-gathered full (unsharded) optimizer state dict. Required under FSDP
+            (gathered alongside `model_state_dict` via `gather_fsdp_state_dicts`); saved in the same
+            safetensors format as the single-GPU path. When None, `optimizer.state_dict()` is used.
+            Defaults to None.
     """
     pretrained_dir = checkpoint_dir / PRETRAINED_MODEL_DIR
-    policy.save_pretrained(pretrained_dir)
+    policy.save_pretrained(pretrained_dir, state_dict=model_state_dict)
     cfg.save_pretrained(pretrained_dir)
     if cfg.peft is not None:
         # When using PEFT, policy.save_pretrained will only write the adapter weights + config, not the
@@ -140,7 +154,13 @@ def save_checkpoint(
     if postprocessor is not None:
         postprocessor.save_pretrained(pretrained_dir)
     save_training_state(
-        checkpoint_dir, step, optimizer, scheduler, num_processes=num_processes, batch_size=batch_size
+        checkpoint_dir,
+        step,
+        optimizer,
+        scheduler,
+        num_processes=num_processes,
+        batch_size=batch_size,
+        optim_state_dict=optim_state_dict,
     )
 
 
@@ -151,6 +171,7 @@ def save_training_state(
     scheduler: LRScheduler | None = None,
     num_processes: int | None = None,
     batch_size: int | None = None,
+    optim_state_dict: dict | None = None,
 ) -> None:
     """
     Saves the training step, optimizer state, scheduler state, and rng state.
@@ -164,19 +185,21 @@ def save_training_state(
             Defaults to None.
         num_processes (int | None, optional): Distributed world size to record. Defaults to None.
         batch_size (int | None, optional): Per-process batch size to record. Defaults to None.
+        optim_state_dict: Pre-gathered full optimizer state dict (for FSDP). Saved instead of
+            `optimizer.state_dict()` when provided. Defaults to None.
     """
     save_dir = checkpoint_dir / TRAINING_STATE_DIR
     save_dir.mkdir(parents=True, exist_ok=True)
     save_training_step(train_step, save_dir, num_processes=num_processes, batch_size=batch_size)
     save_rng_state(save_dir)
     if optimizer is not None:
-        save_optimizer_state(optimizer, save_dir)
+        save_optimizer_state(optimizer, save_dir, optim_state_dict=optim_state_dict)
     if scheduler is not None:
         save_scheduler_state(scheduler, save_dir)
 
 
 def load_training_state(
-    checkpoint_dir: Path, optimizer: Optimizer, scheduler: LRScheduler | None
+    checkpoint_dir: Path, optimizer: Optimizer, scheduler: LRScheduler | None, load_optimizer: bool = True
 ) -> tuple[int, Optimizer, LRScheduler | None]:
     """
     Loads the training step, optimizer state, scheduler state, and rng state.
@@ -186,6 +209,10 @@ def load_training_state(
         checkpoint_dir (Path): The checkpoint directory. Should contain a 'training_state' dir.
         optimizer (Optimizer): The optimizer to load the state_dict to.
         scheduler (LRScheduler | None): The scheduler to load the state_dict to (can be None).
+        load_optimizer (bool, optional): Whether to load the optimizer state from disk. Defaults to
+            True. Set to False under FSDP, where the sharded optimizer state must be loaded after
+            `accelerator.prepare()` via `load_fsdp_optimizer_state` (the optimizer is returned
+            untouched here).
 
     Raises:
         NotADirectoryError: If 'checkpoint_dir' doesn't contain a 'training_state' dir
@@ -200,8 +227,119 @@ def load_training_state(
 
     load_rng_state(training_state_dir)
     step = load_training_step(training_state_dir)
-    optimizer = load_optimizer_state(optimizer, training_state_dir)
+    if load_optimizer:
+        optimizer = load_optimizer_state(optimizer, training_state_dir)
     if scheduler is not None:
         scheduler = load_scheduler_state(scheduler, training_state_dir)
 
     return step, optimizer, scheduler
+
+
+def gather_fsdp_state_dicts(model, optimizer) -> tuple[dict, dict]:
+    """Gather the full (unsharded) model and optimizer state dicts under FSDP.
+
+    `model.state_dict()` and `FSDP.optim_state_dict(...)` are cross-rank collectives, so this must be
+    called on *every* rank with the prepared (FSDP-wrapped) `model` and `optimizer`. With
+    `rank0_only=True` and `offload_to_cpu=True`, every rank runs the all-gather but only rank 0
+    materializes the full dicts (the others get empty dicts) and they are kept on CPU to bound GPU
+    memory. The returned optimizer state dict is keyed by parameter FQNs and is world-size
+    independent; `load_fsdp_optimizer_state` reshards it on resume.
+
+    Returns:
+        (model_state_dict, optim_state_dict): full dicts on rank 0, empty dicts on other ranks.
+    """
+    from torch.distributed.fsdp import (
+        FullOptimStateDictConfig,
+        FullStateDictConfig,
+        FullyShardedDataParallel as FSDP,  # noqa F401
+        StateDictType,
+    )
+
+    state_cfg = FullStateDictConfig(offload_to_cpu=True, rank0_only=True)
+    optim_cfg = FullOptimStateDictConfig(offload_to_cpu=True, rank0_only=True)
+    with FSDP.state_dict_type(model, StateDictType.FULL_STATE_DICT, state_cfg, optim_cfg):
+        model_state_dict = model.state_dict()
+        optim_state_dict = FSDP.optim_state_dict(model, optimizer)
+    return model_state_dict, optim_state_dict
+
+
+def load_fsdp_optimizer_state(model, optimizer, checkpoint_dir: Path) -> None:
+    """Load the FSDP optimizer state (saved as safetensors) and reshard it into the optimizer.
+
+    This is a cross-rank collective and must be called on every rank *after* `accelerator.prepare()`
+    with the prepared (FSDP-wrapped) `model` and `optimizer`. The saved state is the full,
+    world-size-independent optimizer state (keyed by parameter FQNs); `FSDP.optim_state_dict_to_load`
+    reshards it to the current FSDP topology, so resume on a different number of GPUs works.
+    """
+    from torch.distributed.fsdp import (
+        FullOptimStateDictConfig,
+        FullStateDictConfig,
+        FullyShardedDataParallel as FSDP,  # noqa F401
+        StateDictType,
+    )
+
+    # Every rank reads the same full state from the (shared) checkpoint dir, so rank0_only=False.
+    full_osd = load_optimizer_state_dict(checkpoint_dir / TRAINING_STATE_DIR)
+    state_cfg = FullStateDictConfig(rank0_only=False)
+    optim_cfg = FullOptimStateDictConfig(rank0_only=False)
+    with FSDP.state_dict_type(model, StateDictType.FULL_STATE_DICT, state_cfg, optim_cfg):
+        sharded_osd = FSDP.optim_state_dict_to_load(model=model, optim=optimizer, optim_state_dict=full_osd)
+    optimizer.load_state_dict(sharded_osd)
+
+
+def push_checkpoint_to_hub(
+    checkpoint_dir: Path,
+    repo_id: str,
+    *,
+    private: bool | None = None,
+) -> None:
+    """Upload a saved checkpoint directory to the Hub under checkpoints/<name>/.
+
+    Called once per save step when save_checkpoint_to_hub is enabled, so a
+    timed-out or crashed run still leaves recoverable checkpoints on the Hub.
+    The model repo is created idempotently, and the commit is tagged with the
+    checkpoint step so a checkpoint can be recovered with
+    --policy.pretrained_revision=<step> instead of a commit sha.
+    """
+    api = HfApi()
+    api.create_repo(repo_id=repo_id, repo_type="model", private=private, exist_ok=True)
+    commit = api.upload_folder(
+        folder_path=str(checkpoint_dir),
+        repo_id=repo_id,
+        repo_type="model",
+        path_in_repo=f"checkpoints/{checkpoint_dir.name}",
+        commit_message=f"checkpoint {checkpoint_dir.name}",
+    )
+    api.create_tag(
+        repo_id=repo_id,
+        tag=checkpoint_dir.name,
+        revision=commit.oid,
+        repo_type="model",
+        exist_ok=True,
+    )
+
+
+def resolve_resume_checkpoint(repo_id: str, output_dir: Path) -> Path:
+    """Download the latest checkpoint of a Hub training repo into a local run dir.
+
+    The symmetric counterpart to `push_checkpoint_to_hub`: given a model repo holding
+    `checkpoints/<step>/{pretrained_model,training_state}` subtrees, download the highest-numbered step
+    into `output_dir/checkpoints/<step>/`, recreate the local `last` symlink, and return that local
+    checkpoint dir. Used to resume training from the Hub on a machine (or HF Jobs pod) that does not
+    have the original local run dir.
+    """
+    latest = find_latest_hub_checkpoint(repo_id)
+    if latest is None:
+        raise FileNotFoundError(
+            f"No checkpoint found in '{repo_id}' under '{CHECKPOINTS_DIR}/'. "
+            "Was the run trained with --save_checkpoint_to_hub?"
+        )
+    snapshot_download(
+        repo_id=repo_id,
+        repo_type="model",
+        allow_patterns=f"{latest}/*",
+        local_dir=str(output_dir),
+    )
+    checkpoint_dir = output_dir / latest
+    update_last_checkpoint(checkpoint_dir)
+    return checkpoint_dir

src/lerobot/common/wandb_utils.py

@@ -180,24 +180,26 @@ class WandBLogger:
                 self._wandb_custom_step_key.add(new_custom_key)
                 self._wandb.define_metric(new_custom_key, hidden=True)
 
+        batch_data = {}
         for k, v in d.items():
+            # Skip the custom step key here, it's added to the batch below.
+            if custom_step_key is not None and k == custom_step_key:
+                continue
+
             if not isinstance(v, (int | float | str)):
                 logging.warning(
                     f'WandB logging of key "{k}" was ignored as its type "{type(v)}" is not handled by this wrapper.'
                 )
                 continue
 
-            # Do not log the custom step key itself.
-            if self._wandb_custom_step_key is not None and k in self._wandb_custom_step_key:
-                continue
+            batch_data[f"{mode}/{k}"] = v
 
+        if batch_data:
             if custom_step_key is not None:
-                value_custom_step = d[custom_step_key]
-                data = {f"{mode}/{k}": v, f"{mode}/{custom_step_key}": value_custom_step}
-                self._wandb.log(data)
-                continue
-
-            self._wandb.log(data={f"{mode}/{k}": v}, step=step)
+                batch_data[f"{mode}/{custom_step_key}"] = d[custom_step_key]
+                self._wandb.log(batch_data)
+            else:
+                self._wandb.log(data=batch_data, step=step)
 
     def log_video(self, video_path: str, step: int, mode: str = "train"):
         if mode not in {"train", "eval"}:

src/lerobot/configs/__init__.py

@@ -22,7 +22,7 @@ Import them directly: ``from lerobot.configs.train import TrainPipelineConfig``
 """
 
 from .dataset import DatasetRecordConfig
-from .default import DatasetConfig, EvalConfig, PeftConfig, WandBConfig
+from .default import DatasetConfig, EvalConfig, JobConfig, PeftConfig, WandBConfig
 from .policies import PreTrainedConfig
 from .recipe import MessageTurn, TrainingRecipe, load_recipe
 from .types import (
@@ -33,10 +33,18 @@ from .types import (
     RTCAttentionSchedule,
 )
 from .video import (
+    DEFAULT_DEPTH_UNIT,
+    DEPTH_METER_UNIT,
+    DEPTH_MILLIMETER_UNIT,
     VALID_VIDEO_CODECS,
     VIDEO_ENCODER_INFO_KEYS,
+    DepthEncoderConfig,
+    RGBEncoderConfig,
     VideoEncoderConfig,
-    camera_encoder_defaults,
+    depth_encoder_defaults,
+    encoder_config_from_video_info,
+    infer_depth_unit,
+    rgb_encoder_defaults,
 )
 
 __all__ = [
@@ -50,6 +58,7 @@ __all__ = [
     "DatasetRecordConfig",
     "DatasetConfig",
     "EvalConfig",
+    "JobConfig",
     "MessageTurn",
     "PeftConfig",
     "PreTrainedConfig",
@@ -57,9 +66,18 @@ __all__ = [
     "WandBConfig",
     "load_recipe",
     "VideoEncoderConfig",
+    "RGBEncoderConfig",
+    "DepthEncoderConfig",
     # Defaults
-    "camera_encoder_defaults",
+    "rgb_encoder_defaults",
+    "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",
 ]

src/lerobot/configs/dataset.py

@@ -18,7 +18,7 @@ from dataclasses import dataclass, field
 from datetime import datetime
 from pathlib import Path
 
-from .video import VideoEncoderConfig, camera_encoder_defaults
+from .video import DepthEncoderConfig, RGBEncoderConfig, depth_encoder_defaults, rgb_encoder_defaults
 
 
 @dataclass
@@ -58,8 +58,10 @@ class DatasetRecordConfig:
     # Set to 1 for immediate encoding (default behavior), or higher for batched encoding
     video_encoding_batch_size: int = 1
     # Video encoder settings for camera MP4s (codec, quality, GOP, etc.). Tuned via CLI nested keys,
-    # e.g. ``--dataset.camera_encoder.vcodec=h264`` (see ``VideoEncoderConfig``).
-    camera_encoder: VideoEncoderConfig = field(default_factory=camera_encoder_defaults)
+    # e.g. ``--dataset.rgb_encoder.vcodec=h264`` (see ``RGBEncoderConfig``).
+    rgb_encoder: RGBEncoderConfig = field(default_factory=rgb_encoder_defaults)
+    # Video encoder settings for depth-map MP4s (codec, quality, GOP, etc.). Tuned via CLI nested keys.
+    depth_encoder: DepthEncoderConfig = field(default_factory=depth_encoder_defaults)
     # Enable streaming video encoding: encode frames in real-time during capture instead
     # of writing PNG images first. Makes save_episode() near-instant. More info in the documentation: https://huggingface.co/docs/lerobot/streaming_video_encoding
     streaming_encoding: bool = False

src/lerobot/configs/default.py

@@ -19,6 +19,8 @@ from dataclasses import dataclass, field
 from lerobot.transforms import ImageTransformsConfig
 from lerobot.utils.import_utils import get_safe_default_video_backend
 
+from .video import DEFAULT_DEPTH_UNIT, DEPTH_METER_UNIT, DEPTH_MILLIMETER_UNIT
+
 
 @dataclass
 class DatasetConfig:
@@ -35,12 +37,23 @@ class DatasetConfig:
     revision: str | None = None
     use_imagenet_stats: bool = True
     video_backend: str = field(default_factory=get_safe_default_video_backend)
-    # When True, video frames are returned as uint8 tensors (0-255) instead of float32 (0.0-1.0).
+    # When True, RGB video frames are returned as uint8 tensors (0-255) instead of float32 (0.0-1.0).
     # This reduces memory and speeds up DataLoader IPC. The training pipeline handles the conversion.
     return_uint8: bool = False
+    # Physical unit depth maps are dequantized to at load time: "mm" (millimeters) or "m" (metres).
+    # Has no effect on datasets without depth cameras.
+    depth_output_unit: str = DEFAULT_DEPTH_UNIT
     streaming: bool = False
+    # Fraction of episodes held out per task for offline evaluation (0.0 = disabled).
+    eval_split: float = 0.0
 
     def __post_init__(self) -> None:
+        if self.depth_output_unit not in (DEPTH_METER_UNIT, DEPTH_MILLIMETER_UNIT):
+            raise ValueError(
+                f"depth_output_unit must be '{DEPTH_METER_UNIT}' or '{DEPTH_MILLIMETER_UNIT}', got {self.depth_output_unit!r}"
+            )
+        if not (0.0 <= self.eval_split < 1.0):
+            raise ValueError(f"eval_split must be in [0.0, 1.0), got {self.eval_split}")
         if self.episodes is not None:
             if any(ep < 0 for ep in self.episodes):
                 raise ValueError(
@@ -73,8 +86,17 @@ class EvalConfig:
     # `use_async_envs` specifies whether to use asynchronous environments (multiprocessing).
     # Defaults to True; automatically downgraded to SyncVectorEnv when batch_size=1.
     use_async_envs: bool = True
+    # Whether to record eval rollouts as a LeRobot dataset on disk.
+    recording: bool = False
+    # If set, push recorded eval datasets to the Hub under this repo id (one repo per task,
+    # suffixed by task and env index). Requires recording=true.
+    recording_repo_id: str | None = None
+    # Whether the pushed recording repositories should be private.
+    recording_private: bool = False
 
     def __post_init__(self) -> None:
+        if self.recording_repo_id is not None and not self.recording:
+            raise ValueError("eval.recording_repo_id requires eval.recording=true.")
         if self.batch_size == 0:
             self.batch_size = self._auto_batch_size()
         if self.batch_size > self.n_episodes:
@@ -123,3 +145,35 @@ class PeftConfig:
     # If None, the PEFT library defaults to alpha=8, which may dampen high-rank adapters.
     # Common values are r (alpha == rank) or 2*r.
     lora_alpha: int | None = None
+
+
+@dataclass
+class JobConfig:
+    # Where training runs. None (omitted) or "local" runs on this machine.
+    # Any other value is an HF Jobs flavor and submits the run to HF Jobs.
+    # List available flavors + pricing with `hf jobs hardware` command.
+    target: str | None = None
+    # Runtime image for the remote job (ignored for local runs).
+    image: str = "huggingface/lerobot-gpu:latest"
+    # Max wall-clock for the remote job as an HF Jobs duration string (e.g. "2h").
+    # Defaults to "2d": We pass an explicit, generous cap instead. Set a smaller
+    # value to fail fast, or a larger one for long runs.
+    timeout: str | None = "2d"
+    # Submit and exit instead of streaming the job logs in the foreground.
+    detach: bool = False
+    # Extra tags attached to the HF job and to any dataset this run pushes to the
+    # Hub. A "lerobot" tag is always added; e.g. --job.tags '["lelab"]' adds more.
+    tags: list[str] = field(default_factory=list)
+
+    # Two entry points to the same predicate: the staticmethod tests a raw target string
+    # straight from argv (before any JobConfig exists, to decide dispatch early), while the
+    # property is the ergonomic accessor for code that already holds a config instance.
+    @staticmethod
+    def is_remote_target(target: str | None) -> bool:
+        """True when `target` names an HF Jobs flavor rather than a local run."""
+        return target not in (None, "local")
+
+    @property
+    def is_remote(self) -> bool:
+        """True when training should run on HF Jobs rather than this machine."""
+        return self.is_remote_target(self.target)

src/lerobot/configs/policies.py

@@ -79,6 +79,8 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):  # type: igno
     # Either the repo ID of a model hosted on the Hub or a path to a directory containing weights
     # saved using `Policy.save_pretrained`. If not provided, the policy is initialized from scratch.
     pretrained_path: Path | None = None
+    # Optional Hub revision (commit hash, branch, or tag) to pin the pretrained model version.
+    pretrained_revision: str | None = None
 
     def __post_init__(self) -> None:
         if not self.device or not is_torch_device_available(self.device):

src/lerobot/configs/rewards.py

@@ -56,6 +56,8 @@ class RewardModelConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):
     device: str | None = None
 
     pretrained_path: str | None = None
+    # Optional Hub revision (commit hash, branch, or tag) to pin the pretrained reward model version.
+    pretrained_revision: str | None = None
 
     push_to_hub: bool = False
     repo_id: str | None = None

src/lerobot/configs/train.py

@@ -26,11 +26,12 @@ from huggingface_hub.errors import HfHubHTTPError
 
 from lerobot import envs
 from lerobot.optim import LRSchedulerConfig, OptimizerConfig
-from lerobot.utils.hub import HubMixin
+from lerobot.utils.constants import PRETRAINED_MODEL_DIR
+from lerobot.utils.hub import HubMixin, find_latest_hub_checkpoint
 from lerobot.utils.sample_weighting import SampleWeightingConfig
 
 from . import parser
-from .default import DatasetConfig, EvalConfig, PeftConfig, WandBConfig
+from .default import DatasetConfig, EvalConfig, JobConfig, PeftConfig, WandBConfig
 from .policies import PreTrainedConfig
 from .rewards import RewardModelConfig
 
@@ -83,10 +84,11 @@ class TrainPipelineConfig(HubMixin):
     # with the same value for `dir` its contents will be overwritten unless you set `resume` to true.
     output_dir: Path | None = None
     job_name: str | None = None
-    # Set `resume` to true to resume a previous run. In order for this to work, you will need to make sure
-    # `dir` is the directory of an existing run with at least one checkpoint in it.
-    # Note that when resuming a run, the default behavior is to use the configuration from the checkpoint,
-    # regardless of what's provided with the training command at the time of resumption.
+    # Set `resume` to true to resume a previous run. Pass `--config_path` pointing at either a local
+    # checkpoint's train_config.json or a Hub repo id holding `checkpoints/<step>/` subtrees (the
+    # latest checkpoint is downloaded and resumed from). Note that when resuming, the default behavior
+    # is to use the configuration from the checkpoint, regardless of what's provided with the training
+    # command at the time of resumption (CLI `--*` flags still override).
     resume: bool = False
     # `seed` is used for training (eg: model initialization, dataset shuffling)
     # AND for the evaluation environments.
@@ -100,8 +102,13 @@ class TrainPipelineConfig(HubMixin):
     prefetch_factor: int = 4
     persistent_workers: bool = True
     steps: int = 100_000
-    eval_freq: int = 20_000
+    # Run policy in the simulation environment every N steps to measure reward/success (0 = disabled).
+    env_eval_freq: int = 20_000
     log_freq: int = 200
+    # Compute eval loss on held-out episodes every N steps (0 = disabled). Requires eval_split > 0.
+    eval_steps: int = 0
+    # Cap on total eval samples, split uniformly across tasks (0 = use all held-out data).
+    max_eval_samples: int = 0
     tolerance_s: float = 1e-4
     save_checkpoint: bool = True
     # Checkpoint is saved every `save_freq` training iterations and after the last training step.
@@ -113,6 +120,13 @@ class TrainPipelineConfig(HubMixin):
     wandb: WandBConfig = field(default_factory=WandBConfig)
     peft: PeftConfig | None = None
 
+    # Where to run training (local default, or an HF Jobs flavor). See JobConfig.
+    job: JobConfig = field(default_factory=JobConfig)
+    # Push each saved checkpoint to the Hub (policy.repo_id) as it is written, not
+    # just the final model (useful to monitor progress mid-run). Optional; the
+    # final model is pushed regardless. Works the same locally and remotely.
+    save_checkpoint_to_hub: bool = False
+
     # Sample weighting configuration (e.g., for RA-BC training)
     sample_weighting: SampleWeightingConfig | None = None
 
@@ -132,10 +146,17 @@ class TrainPipelineConfig(HubMixin):
             return self.reward_model  # type: ignore[return-value]
         return self.policy  # type: ignore[return-value]
 
-    def validate(self) -> None:
-        # HACK: We parse again the cli args here to get the pretrained paths if there was some.
-        policy_path = parser.get_path_arg("policy")
+    def _resolve_pretrained_from_cli(self) -> None:
+        """Resolve the pretrained source passed on the CLI into a loaded config.
+
+        The pretrained paths (`--policy.path`, `--reward_model.path`) and
+        `--config_path` are only recoverable by re-reading the CLI args: draccus
+        has already consumed them by the time `validate()` runs, so they are not
+        reflected on `self`. Exactly one source applies, in priority order:
+        reward-model path, policy path, then resume.
+        """
         reward_model_path = parser.get_path_arg("reward_model")
+        policy_path = parser.get_path_arg("policy")
 
         if reward_model_path:
             cli_overrides = parser.get_cli_overrides("reward_model")
@@ -144,31 +165,54 @@ class TrainPipelineConfig(HubMixin):
             )
             self.reward_model.pretrained_path = str(Path(reward_model_path))
         elif policy_path:
-            yaml_overrides = parser.get_yaml_overrides("policy")
-            cli_overrides = parser.get_cli_overrides("policy") or []
-            self.policy = PreTrainedConfig.from_pretrained(
-                policy_path, cli_overrides=yaml_overrides + cli_overrides
-            )
+            overrides = parser.get_yaml_overrides("policy") + (parser.get_cli_overrides("policy") or [])
+            self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=overrides)
             self.policy.pretrained_path = Path(policy_path)
         elif self.resume:
-            config_path = parser.parse_arg("config_path")
-            if not config_path:
-                raise ValueError(
-                    f"A config_path is expected when resuming a run. Please specify path to {TRAIN_CONFIG_NAME}"
-                )
+            self._resolve_resume_checkpoint()
 
-            if not Path(config_path).resolve().exists():
-                raise NotADirectoryError(
-                    f"{config_path=} is expected to be a local path. "
-                    "Resuming from the hub is not supported for now."
-                )
+    def _resolve_resume_checkpoint(self) -> None:
+        """Point the trainable config at the checkpoint named by `--config_path`.
 
+        `config_path` is either a local path (to a checkpoint's train_config.json or its
+        pretrained_model/ dir) or a Hub repo id. For a Hub repo, the latest checkpoint is downloaded
+        into a fresh local run dir and resumed from there. The download is skipped when dispatching to
+        an HF Job (`job.is_remote`): the pod performs it when it runs the resume locally, and
+        `submit_to_hf` resolves the source repo for the remote command.
+        """
+        config_path = parser.parse_arg("config_path")
+        if not config_path:
+            raise ValueError(
+                f"A config_path is expected when resuming a run. Please specify path to {TRAIN_CONFIG_NAME}"
+            )
+
+        if Path(config_path).resolve().exists():
             policy_dir = Path(config_path).parent
-            if self.policy is not None:
-                self.policy.pretrained_path = policy_dir
-            if self.reward_model is not None:
-                self.reward_model.pretrained_path = str(policy_dir)
             self.checkpoint_path = policy_dir.parent
+        elif self.job.is_remote:
+            return
+        else:
+            from lerobot.common.train_utils import resolve_resume_checkpoint
+
+            # `self.output_dir` was loaded from the checkpoint's config and points at the original
+            # run's (now-absent) local dir. Resume into a fresh local dir instead, unless the user
+            # passed --output_dir explicitly.
+            cli_output_dir = parser.parse_arg("output_dir")
+            if cli_output_dir:
+                self.output_dir = Path(cli_output_dir)
+            else:
+                now = dt.datetime.now()
+                self.output_dir = Path("outputs/train") / f"{now:%Y-%m-%d}/{now:%H-%M-%S}_resume"
+            self.checkpoint_path = resolve_resume_checkpoint(config_path, self.output_dir)
+            policy_dir = self.checkpoint_path / PRETRAINED_MODEL_DIR
+
+        if self.policy is not None:
+            self.policy.pretrained_path = policy_dir
+        if self.reward_model is not None:
+            self.reward_model.pretrained_path = str(policy_dir)
+
+    def validate(self) -> None:
+        self._resolve_pretrained_from_cli()
 
         if self.policy is None and self.reward_model is None:
             raise ValueError(
@@ -208,9 +252,22 @@ class TrainPipelineConfig(HubMixin):
             self.optimizer = active_cfg.get_optimizer_preset()
             self.scheduler = active_cfg.get_scheduler_preset()
 
-        if hasattr(active_cfg, "push_to_hub") and active_cfg.push_to_hub and not active_cfg.repo_id:
+        if self.eval_steps > 0 and self.dataset.eval_split == 0.0:
+            raise ValueError("eval_steps > 0 requires dataset.eval_split > 0.0 to hold out eval data.")
+
+        # Remote runs auto-generate the repo_id in submit_to_hf (the policy may only be
+        # resolved here, from --policy.path), so don't demand it up front for them.
+        if (
+            hasattr(active_cfg, "push_to_hub")
+            and active_cfg.push_to_hub
+            and not active_cfg.repo_id
+            and not self.job.is_remote
+        ):
             raise ValueError("'repo_id' argument missing. Please specify it to push the model to the hub.")
 
+        if self.save_checkpoint_to_hub and not (self.policy is not None and self.policy.repo_id):
+            raise ValueError("save_checkpoint_to_hub requires --policy.repo_id.")
+
     @classmethod
     def __get_path_fields__(cls) -> list[str]:
         """Keys for draccus pretrained-path loading."""
@@ -247,22 +304,30 @@ class TrainPipelineConfig(HubMixin):
         elif Path(model_id).is_file():
             config_file = model_id
         else:
+            dl_kwargs = {
+                "repo_id": model_id,
+                "revision": revision,
+                "cache_dir": cache_dir,
+                "force_download": force_download,
+                "proxies": proxies,
+                "resume_download": resume_download,
+                "token": token,
+                "local_files_only": local_files_only,
+            }
             try:
-                config_file = hf_hub_download(
-                    repo_id=model_id,
-                    filename=TRAIN_CONFIG_NAME,
-                    revision=revision,
-                    cache_dir=cache_dir,
-                    force_download=force_download,
-                    proxies=proxies,
-                    resume_download=resume_download,
-                    token=token,
-                    local_files_only=local_files_only,
-                )
+                config_file = hf_hub_download(filename=TRAIN_CONFIG_NAME, **dl_kwargs)
             except HfHubHTTPError as e:
-                raise FileNotFoundError(
-                    f"{TRAIN_CONFIG_NAME} not found on the HuggingFace Hub in {model_id}"
-                ) from e
+                # No root train_config.json: this is a repo of periodic checkpoints from an
+                # interrupted run. Fall back to the latest checkpoint's config so the run can be
+                # resumed straight from the repo with `--config_path=<repo>`.
+                latest = find_latest_hub_checkpoint(model_id, token=token, revision=revision)
+                if latest is None:
+                    raise FileNotFoundError(
+                        f"{TRAIN_CONFIG_NAME} not found on the HuggingFace Hub in {model_id}"
+                    ) from e
+                config_file = hf_hub_download(
+                    filename=f"{latest}/{PRETRAINED_MODEL_DIR}/{TRAIN_CONFIG_NAME}", **dl_kwargs
+                )
 
         cli_args = kwargs.pop("cli_args", [])
         # Legacy RA-BC migration only applies to framework-saved checkpoints (always JSON).

src/lerobot/configs/video.py

@@ -20,7 +20,9 @@ from __future__ import annotations
 
 import logging
 from dataclasses import dataclass, field
-from typing import Any
+from typing import Any, ClassVar, Self
+
+import numpy as np
 
 from lerobot.utils.import_utils import require_package
 
@@ -36,11 +38,12 @@ HW_VIDEO_CODECS = [
     "h264_vaapi",  # Linux Intel/AMD
     "h264_qsv",  # Intel Quick Sync
 ]
-VALID_VIDEO_CODECS: frozenset[str] = frozenset({"h264", "hevc", "libsvtav1", "auto", *HW_VIDEO_CODECS})
+VALID_VIDEO_CODECS: frozenset[str] = frozenset(
+    {"h264", "hevc", "libsvtav1", "libaom-av1", "auto", *HW_VIDEO_CODECS}
+)
 # Aliases for legacy video codec names.
 VIDEO_CODECS_ALIASES: dict[str, str] = {"av1": "libsvtav1"}
 
-
 LIBSVTAV1_DEFAULT_PRESET: int = 12
 
 # Keys persisted under ``features[*]["info"]`` as ``video.<name>`` (from :class:`VideoEncoderConfig`).
@@ -52,40 +55,54 @@ VIDEO_ENCODER_INFO_KEYS: frozenset[str] = frozenset(
     f"video.{name}" for name in VIDEO_ENCODER_INFO_FIELD_NAMES
 )
 
+# Default depth quantization and encoding parameters.
+DEPTH_QUANT_BITS: int = 12
+DEPTH_QMAX: int = (1 << DEPTH_QUANT_BITS) - 1  # 4095
+
+DEFAULT_DEPTH_MIN: float = 0.01
+DEFAULT_DEPTH_MAX: float = 10.0
+DEFAULT_DEPTH_SHIFT: float = 3.5
+DEFAULT_DEPTH_USE_LOG: bool = True
+DEFAULT_DEPTH_PIX_FMT: str = "gray12le"
+
+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"})
+
 
 @dataclass
 class VideoEncoderConfig:
-    """Video encoder configuration.
+    """Video encoder configuration."""
 
-    Attributes:
-        vcodec: Video encoder name. ``"auto"`` is resolved during
-            construction (HW encoder if available, else ``libsvtav1``).
-        pix_fmt: Pixel format (e.g. ``"yuv420p"``).
-        g: GOP size (keyframe interval).
-        crf: Quality level — mapped to the native quality parameter of the
-            codec (``crf`` for software, ``qp`` for NVENC/VAAPI,
-            ``q:v`` for VideoToolbox, ``global_quality`` for QSV).
-        preset: Speed/quality preset. Accepted type is per-codec.
-        fast_decode: Fast-decode tuning. For ``libsvtav1`` this is a level (0-2)
-            embedded in ``svtav1-params``. For ``h264`` and ``hevc`` non-zero values
-            set ``tune=fastdecode``. Ignored for other codecs.
-        video_backend: Python to be used for encoding. Only ``"pyav"``
-            is currently supported.
-        extra_options: Free-form dictionary of additional video encoder options
-            (e.g. ``{"tune": "film", "profile:v": "high", "bf": 2}``).
-    """
-
-    vcodec: str = "libsvtav1"  # TODO(CarolinePascal): rename to codec ?
-    pix_fmt: str = "yuv420p"
-    g: int | None = 2
-    crf: int | float | None = 30
-    preset: int | str | None = None
-    fast_decode: int = 0
+    vcodec: str = "libsvtav1"  # Video codec name. "auto" picks a hardware codec if available, else libsvtav1.
+    pix_fmt: str = "yuv420p"  # Pixel format (e.g. yuv420p).
+    g: int | None = 2  # GOP size (keyframe interval).
+    crf: int | float | None = 30  # Quality level. Lower means better quality and larger files.
+    preset: int | str | None = None  # Speed/quality preset. Accepted values are codec-specific.
+    fast_decode: int = 0  # Fast-decode tuning. Accepted values are codec-specific, 0 disables it.
     # TODO(CarolinePascal): add torchcodec support + find a way to unify the
     # two backends (encoding and decoding).
-    video_backend: str = "pyav"
+    video_backend: str = "pyav"  # Encoding backend. Only "pyav" is currently supported.
+    # Extra codec options merged last, e.g. {"tune": "film"}.
     extra_options: dict[str, Any] = field(default_factory=dict)
 
+    # Source-data channel count this encoder is expected to handle. ``None``
+    # disables the pix_fmt channel-count check; concrete subclasses set it
+    # (3 for RGB, 1 for depth, etc.).
+    _DEFAULT_CHANNELS: ClassVar[int | None] = None
+
     def __post_init__(self) -> None:
         self.resolve_vcodec()
         # Empty-constructor ergonomics: ``VideoEncoderConfig()`` must "just work".
@@ -94,9 +111,9 @@ class VideoEncoderConfig:
         self.validate()
 
     @classmethod
-    def from_video_info(cls, video_info: dict | None) -> VideoEncoderConfig:
-        """Reconstruct a :class:`VideoEncoderConfig` from a video feature's ``info`` block.
-        Missing or ``None`` values fall back to the class defaults.
+    def _kwargs_from_video_info(cls, video_info: dict | None) -> dict[str, Any]:
+        """Parse the ``video.*`` keys of a feature ``info`` block into
+        constructor kwargs.
         """
         video_info = video_info or {}
         kwargs: dict[str, Any] = {}
@@ -115,7 +132,15 @@ class VideoEncoderConfig:
                 continue
             kwargs[field_name] = value
 
-        return cls(**kwargs)
+        return kwargs
+
+    @classmethod
+    def from_video_info(cls, video_info: dict | None) -> Self:
+        """Reconstruct an encoder config from a video feature's ``info`` block.
+
+        Missing or ``None`` values fall back to the class defaults.
+        """
+        return cls(**cls._kwargs_from_video_info(video_info))
 
     def detect_available_encoders(self, encoders: list[str] | str) -> list[str]:
         """Return the subset of available encoders based on the specified video backend.
@@ -138,7 +163,9 @@ class VideoEncoderConfig:
             require_package("av", extra="dataset")
             from lerobot.datasets import check_video_encoder_parameters_pyav
 
-            check_video_encoder_parameters_pyav(self.vcodec, self.pix_fmt, self.get_codec_options())
+            check_video_encoder_parameters_pyav(
+                self.vcodec, self.pix_fmt, self.get_codec_options(), channels=self._DEFAULT_CHANNELS
+            )
 
     def resolve_vcodec(self) -> None:
         """Check ``vcodec`` and, when it is ``"auto"``, pick a concrete encoder.
@@ -199,18 +226,24 @@ class VideoEncoderConfig:
             if encoder_threads is not None:
                 svtav1_parts.append(f"lp={encoder_threads}")
             if svtav1_parts:
-                opts["svtav1-params"] = ":".join(svtav1_parts)
+                set_if("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:
-                opts["tune"] = "fastdecode"
+                set_if("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:
-                opts["q:v"] = max(1, min(100, 100 - self.crf * 2))
+                set_if("q:v", max(1, min(100, 100 - self.crf * 2)))
         elif self.vcodec in ("h264_nvenc", "hevc_nvenc"):
-            opts["rc"] = 0
+            set_if("rc", 0)
             set_if("qp", self.crf)
             set_if("preset", self.preset)
         elif self.vcodec == "h264_vaapi":
@@ -230,6 +263,79 @@ class VideoEncoderConfig:
         return opts
 
 
-def camera_encoder_defaults() -> VideoEncoderConfig:
-    """Return a :class:`VideoEncoderConfig` with RGB-camera defaults."""
-    return VideoEncoderConfig()
+@dataclass
+class RGBEncoderConfig(VideoEncoderConfig):
+    """Encoder configuration for RGB camera streams.
+
+    Identical to :class:`VideoEncoderConfig` but declares the 3-channel
+    source-data layout so ``pix_fmt`` is validated against RGB inputs.
+    """
+
+    _DEFAULT_CHANNELS: ClassVar[int] = 3
+
+
+def rgb_encoder_defaults() -> RGBEncoderConfig:
+    """Return a :class:`RGBEncoderConfig` with RGB-camera defaults."""
+    return RGBEncoderConfig()
+
+
+@dataclass
+class DepthEncoderConfig(VideoEncoderConfig):
+    """Encoder configuration for depth-map streams.
+
+    Inherits the full :class:`VideoEncoderConfig` surface (codec, GOP, CRF,
+    preset, ``extra_options``…) and adds the parameters of the depth quantizer.
+    Defaults flip ``vcodec`` to ``"hevc"`` (Main 12 profile) and ``pix_fmt`` to
+    ``"gray12le"``.
+    """
+
+    vcodec: str = "hevc"  # Video codec name. Defaults to HEVC Main 12 (a 12-bit-capable codec).
+    pix_fmt: str = "gray12le"  # Pixel format. Defaults to 12-bit grayscale.
+    extra_options: dict[str, Any] = field(default_factory=lambda: {"x265-params": "lossless=1"})
+
+    depth_min: float = DEFAULT_DEPTH_MIN  # Minimum depth in meters, mapped to the lowest quantum.
+    depth_max: float = DEFAULT_DEPTH_MAX  # Maximum depth in meters, mapped to the highest quantum.
+    shift: float = DEFAULT_DEPTH_SHIFT  # Pre-log offset in meters for numerical stability near zero.
+    use_log: bool = DEFAULT_DEPTH_USE_LOG  # Use logarithmic quantization (True) or linear (False).
+
+    _DEFAULT_CHANNELS: ClassVar[int] = 1
+
+    @classmethod
+    def _kwargs_from_video_info(cls, video_info: dict | None) -> dict[str, Any]:
+        """Layer the depth-specific tuning (``depth_min`` / ``depth_max`` /
+        ``shift`` / ``use_log``) on top of the base parser. Missing keys
+        fall back to the class defaults.
+        """
+        kwargs = super()._kwargs_from_video_info(video_info)
+        video_info = video_info or {}
+        for name in DEPTH_ENCODER_INFO_FIELD_NAMES:
+            value = video_info.get(f"video.{name}")
+            if value is not None:
+                kwargs[name] = value
+        return kwargs
+
+
+def depth_encoder_defaults() -> DepthEncoderConfig:
+    """Return a :class:`DepthEncoderConfig` with depth-camera defaults."""
+    return DepthEncoderConfig()
+
+
+def encoder_config_from_video_info(video_info: dict | None) -> VideoEncoderConfig:
+    """Build the appropriate encoder config from a feature's ``info`` block.
+
+    Dispatches to :class:`DepthEncoderConfig` when the dict marks the feature
+    as a depth map and to :class:`RGBEncoderConfig`
+    otherwise.
+
+    Args:
+        video_info: A feature's ``info`` dict as persisted in ``info.json``,
+            or ``None`` (treated as an empty dict).
+
+    Returns:
+        A :class:`DepthEncoderConfig` for depth features, otherwise a
+        :class:`RGBEncoderConfig`.
+    """
+    video_info = video_info or {}
+    is_depth = bool(video_info.get("is_depth_map") or video_info.get("video.is_depth_map"))
+    cls: type[VideoEncoderConfig] = DepthEncoderConfig if is_depth else RGBEncoderConfig
+    return cls.from_video_info(video_info)

src/lerobot/datasets/__init__.py

@@ -35,7 +35,7 @@ from .dataset_tools import (
     remove_feature,
     split_dataset,
 )
-from .factory import make_dataset, resolve_delta_timestamps
+from .factory import make_dataset, make_train_eval_datasets, resolve_delta_timestamps
 from .image_writer import safe_stop_image_writer
 from .io_utils import load_episodes, write_stats
 from .language import (
@@ -89,6 +89,7 @@ __all__ = [
     "get_feature_stats",
     "load_episodes",
     "make_dataset",
+    "make_train_eval_datasets",
     "merge_datasets",
     "modify_features",
     "modify_tasks",

src/lerobot/datasets/aggregate.py

@@ -32,6 +32,7 @@ from .feature_utils import features_equal_for_merge, get_hf_features_from_featur
 from .io_utils import (
     get_file_size_in_mb,
     get_parquet_file_size_in_mb,
+    to_parquet_one_row_group_per_episode,
     to_parquet_with_hf_images,
     write_info,
     write_stats,
@@ -551,6 +552,7 @@ def aggregate_data(src_meta, dst_meta, data_idx, data_files_size_in_mb, chunk_si
             aggr_root=dst_meta.root,
             hf_features=hf_features,
             concatenate=concatenate_data,
+            one_row_group_per_episode=True,
         )
 
         # Record the mapping from source to actual destination
@@ -628,6 +630,7 @@ def append_or_create_parquet_file(
     aggr_root: Path = None,
     hf_features: datasets.Features | None = None,
     concatenate: bool = True,
+    one_row_group_per_episode: bool = False,
 ) -> tuple[dict[str, int], tuple[int, int]]:
     """Appends data to an existing parquet file or creates a new one based on size constraints.
 
@@ -645,6 +648,8 @@ def append_or_create_parquet_file(
         aggr_root: Root path for the aggregated dataset.
         hf_features: Optional HuggingFace Features schema for proper image typing.
         concatenate: When False, always rotate to a new file instead of appending to the current one.
+        one_row_group_per_episode: True for DATA parquet (emit one row group per episode); False for
+            the episodes-metadata parquet (already one row per episode).
 
     Returns:
         tuple: (updated_idx, (dst_chunk, dst_file)) where updated_idx is the index dict
@@ -657,6 +662,8 @@ def append_or_create_parquet_file(
         dst_path.parent.mkdir(parents=True, exist_ok=True)
         if contains_images:
             to_parquet_with_hf_images(df, dst_path, features=hf_features)
+        elif one_row_group_per_episode:
+            to_parquet_one_row_group_per_episode(df, dst_path)
         else:
             df.to_parquet(dst_path)
         return idx, (dst_chunk, dst_file)
@@ -683,6 +690,8 @@ def append_or_create_parquet_file(
 
     if contains_images:
         to_parquet_with_hf_images(final_df, target_path, features=hf_features)
+    elif one_row_group_per_episode:
+        to_parquet_one_row_group_per_episode(final_df, target_path)
     else:
         final_df.to_parquet(target_path)
 

src/lerobot/datasets/compute_stats.py

@@ -242,12 +242,12 @@ def sample_images(image_paths: list[str]) -> np.ndarray:
     images = None
     for i, idx in enumerate(sampled_indices):
         path = image_paths[idx]
-        # we load as uint8 to reduce memory usage
+        # we load RGB images as uint8 to reduce memory usage; depth keeps its native dtype
         img = load_image_as_numpy(path, dtype=np.uint8, channel_first=True)
         img = auto_downsample_height_width(img)
 
         if images is None:
-            images = np.empty((len(sampled_indices), *img.shape), dtype=np.uint8)
+            images = np.empty((len(sampled_indices), *img.shape), dtype=img.dtype)
 
         images[i] = img
 
@@ -506,8 +506,10 @@ def compute_episode_stats(
         Each statistics dictionary contains min, max, mean, std, count, and quantiles.
 
     Note:
-        Image statistics are normalized to [0,1] range and have shape (3,1,1) for
-        per-channel values when dtype is 'image' or 'video'.
+        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
+        this rescaling and remain in their stored units (stored in ``depth_unit``).
     """
     if quantile_list is None:
         quantile_list = DEFAULT_QUANTILES
@@ -517,6 +519,13 @@ def compute_episode_stats(
         if features[key]["dtype"] in {"string", "language"}:
             continue
 
+        # Features with zero-width shapes are skipped (no data to compute stats on)
+        if any(d == 0 for d in features[key].get("shape", ())):
+            logging.debug(
+                f"Skipping statistics computation for feature '{key}' with a zero-width shape {features[key]['shape']}."
+            )
+            continue
+
         if features[key]["dtype"] in ["image", "video"]:
             ep_ft_array = sample_images(data)
             axes_to_reduce = (0, 2, 3)
@@ -531,8 +540,12 @@ def compute_episode_stats(
         )
 
         if features[key]["dtype"] in ["image", "video"]:
+            normalization_factor = (
+                255.0 if not (features[key].get("info") or {}).get("is_depth_map", False) else 1.0
+            )
             ep_stats[key] = {
-                k: v if k == "count" else np.squeeze(v / 255.0, axis=0) for k, v in ep_stats[key].items()
+                k: v if k == "count" else np.squeeze(v / normalization_factor, axis=0)
+                for k, v in ep_stats[key].items()
             }
 
     return ep_stats
@@ -552,8 +565,10 @@ def _validate_stat_value(value: np.ndarray, key: str, feature_key: str) -> None:
     if key == "count" and value.shape != (1,):
         raise ValueError(f"Shape of 'count' must be (1), but is {value.shape} instead.")
 
-    if "image" in feature_key and key != "count" and value.shape != (3, 1, 1):
-        raise ValueError(f"Shape of quantile '{key}' must be (3,1,1), but is {value.shape} instead.")
+    if "image" in feature_key and key != "count" and value.shape not in ((3, 1, 1), (1, 1, 1)):
+        raise ValueError(
+            f"Shape of quantile '{key}' must be (3,1,1) or (1,1,1) but is {value.shape} instead."
+        )
 
 
 def _assert_type_and_shape(stats_list: list[dict[str, dict]]):

src/lerobot/datasets/dataset_metadata.py

@@ -14,7 +14,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import contextlib
-from collections.abc import Callable
+import logging
+from collections.abc import Callable, Iterable
+from copy import deepcopy
 from pathlib import Path
 
 import numpy as np
@@ -24,12 +26,13 @@ import pyarrow as pa
 import pyarrow.parquet as pq
 from huggingface_hub import snapshot_download
 
-from lerobot.configs import VideoEncoderConfig
+from lerobot.configs import DEPTH_METER_UNIT, 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,
@@ -337,6 +340,54 @@ class LeRobotDatasetMetadata:
         """Keys to access visual modalities stored as videos."""
         return [key for key, ft in self.features.items() if ft["dtype"] == "video"]
 
+    @property
+    def depth_keys(self) -> list[str]:
+        """Keys to access depth-map modalities stored as videos or images.
+
+        A depth key is a feature whose ``info`` dict carries ``"is_depth_map": True``
+        (or the legacy ``"video.is_depth_map"`` inside ``info`` or ``video_info``).
+        """
+
+        def _is_depth(ft: dict) -> bool:
+            info = ft.get("info") or {}
+            video_info = ft.get("video_info") or {}
+            return (
+                info.get("is_depth_map", False)
+                or info.get("video.is_depth_map", False)
+                or video_info.get("video.is_depth_map", False)
+            )
+
+        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)."""
@@ -580,29 +631,48 @@ class LeRobotDatasetMetadata:
     def update_video_info(
         self,
         video_key: str | None = None,
-        camera_encoder: VideoEncoderConfig | None = None,
+        video_encoder: VideoEncoderConfig | None = None,
+        preserve_keys: Iterable[str] | None = None,
     ) -> None:
-        """Populate per-feature video info in ``info.json``.
+        """Populate or refresh per-feature video info in ``info.json``.
 
         Warning: this function writes info from first episode videos, implicitly assuming that all videos have
         been encoded the same way. Also, this means it assumes the first episode exists.
 
+        Always re-probes the videos and overwrites existing info for every recomputed
+        key. ``preserve_keys`` lists keys whose existing values must be kept (e.g.
+        data-intrinsic entries like ``is_depth_map`` and depth quantization params)
+        instead of being recomputed.
+
         Args:
             video_key: If provided, only update this video key. Otherwise update
                 all video keys in the dataset.
-            camera_encoder: Encoder configuration used to produce the
+            video_encoder: Encoder configuration used to produce the
                 videos. When provided, its fields are recorded as
                 ``video.<field>`` entries alongside the stream-derived
                 ``video.*`` entries (see :func:`get_video_info`).
+            preserve_keys: Keys whose existing values are kept instead of being
+                recomputed. ``None`` (default) recomputes every key.
         """
         if video_key is not None and video_key not in self.video_keys:
             raise ValueError(f"Video key {video_key} not found in dataset")
 
         video_keys = [video_key] if video_key is not None else self.video_keys
+        preserve_set = set(preserve_keys or ())
         for key in video_keys:
-            if not self.features[key].get("info", None):
-                video_path = self.root / self.video_path.format(video_key=key, chunk_index=0, file_index=0)
-                self.info.features[key]["info"] = get_video_info(video_path, camera_encoder=camera_encoder)
+            existing = self.features[key].get("info") or {}
+            video_path = self.root / self.video_path.format(video_key=key, chunk_index=0, file_index=0)
+            new_info = get_video_info(video_path, video_encoder=video_encoder)
+            # Drop preserved keys so the existing values win on merge.
+            new_info = {k: v for k, v in new_info.items() if k not in preserve_set}
+            merged = {**existing, **new_info}
+            # Migrate the legacy depth marker to the canonical key.
+            if "video.is_depth_map" in merged:
+                logging.warning(
+                    f"Migrating legacy 'video.is_depth_map' to 'is_depth_map' for feature {key!r}."
+                )
+                merged.setdefault("is_depth_map", merged.pop("video.is_depth_map"))
+            self.info.features[key]["info"] = merged
 
     def update_chunk_settings(
         self,
@@ -709,7 +779,7 @@ class LeRobotDatasetMetadata:
 
         obj.root.mkdir(parents=True, exist_ok=False)
 
-        features = {**features, **DEFAULT_FEATURES}
+        features = {**deepcopy(features), **DEFAULT_FEATURES}
         _validate_feature_names(features)
 
         obj.tasks = None

src/lerobot/datasets/dataset_reader.py

@@ -22,7 +22,14 @@ from pathlib import Path
 import datasets
 import torch
 
+from lerobot.configs import (
+    DEFAULT_DEPTH_UNIT,
+    DEPTH_METER_UNIT,
+    DepthEncoderConfig,
+)
+
 from .dataset_metadata import LeRobotDatasetMetadata
+from .depth_utils import MM_PER_METRE, dequantize_depth
 from .feature_utils import (
     check_delta_timestamps,
     get_delta_indices,
@@ -51,6 +58,7 @@ class DatasetReader:
         delta_timestamps: dict[str, list[float]] | None,
         image_transforms: Callable | None,
         return_uint8: bool = False,
+        depth_output_unit: str = DEFAULT_DEPTH_UNIT,
     ):
         """Initialize the reader with metadata, filtering, and transform config.
 
@@ -68,14 +76,21 @@ class DatasetReader:
                 relative timestamp offsets for temporal context windows.
             image_transforms: Optional torchvision v2 transform applied to
                 visual features.
+            return_uint8: If True, return RGB video frames as raw uint8 tensors
+                instead of normalized float32.
+            depth_output_unit: Physical unit depth maps are dequantized to
+                (``"m"`` or ``"mm"``). Defaults to ``"mm"``.
         """
         self._meta = meta
         self.root = root
         self.episodes = episodes
         self._tolerance_s = tolerance_s
         self._video_backend = video_backend
+        if image_transforms is not None and not callable(image_transforms):
+            raise TypeError("image_transforms must be callable or None.")
         self._image_transforms = image_transforms
         self._return_uint8 = return_uint8
+        self._depth_output_unit = depth_output_unit
 
         self.hf_dataset: datasets.Dataset | None = None
         self._absolute_to_relative_idx: dict[int, int] | None = None
@@ -86,6 +101,28 @@ class DatasetReader:
             check_delta_timestamps(delta_timestamps, meta.fps, tolerance_s)
             self.delta_indices = get_delta_indices(delta_timestamps, meta.fps)
 
+        self._depth_encoder_configs: dict[str, DepthEncoderConfig] = {
+            vid_key: DepthEncoderConfig.from_video_info(self._meta.features[vid_key].get("info"))
+            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):
+            raise TypeError("image_transforms must be callable or None.")
+        self._image_transforms = image_transforms
+
+    def clear_image_transforms(self) -> None:
+        """Remove the transform applied to visual observations."""
+        self._image_transforms = None
+
     def try_load(self) -> bool:
         """Attempt to load from local cache. Returns True if data is sufficient."""
         try:
@@ -247,7 +284,18 @@ class DatasetReader:
                 self._tolerance_s,
                 self._video_backend,
                 return_uint8=self._return_uint8,
+                is_depth=vid_key in self._meta.depth_keys,
             )
+            if vid_key in self._meta.depth_keys:
+                depth_encoder = self._depth_encoder_configs[vid_key]
+                frames = dequantize_depth(
+                    frames,
+                    depth_min=depth_encoder.depth_min,
+                    depth_max=depth_encoder.depth_max,
+                    shift=depth_encoder.shift,
+                    use_log=depth_encoder.use_log,
+                    output_unit=self._depth_output_unit,
+                )
             return vid_key, frames.squeeze(0)
 
         items = list(query_timestamps.items())
@@ -287,10 +335,18 @@ class DatasetReader:
             item = {**video_frames, **item}
 
         if self._image_transforms is not None:
-            image_keys = self._meta.camera_keys
-            for cam in image_keys:
+            for cam in self._meta.camera_keys:
+                if cam in self._meta.depth_keys:
+                    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

src/lerobot/datasets/dataset_tools.py

@@ -27,6 +27,7 @@ import logging
 import shutil
 from collections.abc import Callable
 from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor, as_completed
+from copy import deepcopy
 from pathlib import Path
 
 import datasets
@@ -36,7 +37,15 @@ import pyarrow.parquet as pq
 import torch
 from tqdm import tqdm
 
-from lerobot.configs import VideoEncoderConfig, camera_encoder_defaults
+from lerobot.configs import (
+    DepthEncoderConfig,
+    RGBEncoderConfig,
+    VideoEncoderConfig,
+    depth_encoder_defaults,
+    encoder_config_from_video_info,
+    rgb_encoder_defaults,
+)
+from lerobot.configs.video import DEPTH_ENCODER_INFO_FIELD_NAMES
 from lerobot.utils.constants import ACTION, HF_LEROBOT_HOME, OBS_IMAGE, OBS_STATE
 from lerobot.utils.utils import flatten_dict
 
@@ -47,6 +56,7 @@ from .compute_stats import (
     compute_relative_action_stats,
 )
 from .dataset_metadata import LeRobotDatasetMetadata
+from .image_writer import write_image
 from .io_utils import (
     get_parquet_file_size_in_mb,
     load_episodes,
@@ -61,12 +71,13 @@ from .utils import (
     DEFAULT_DATA_FILE_SIZE_IN_MB,
     DEFAULT_DATA_PATH,
     DEFAULT_EPISODES_PATH,
+    DEPTH_FILE_PATTERN,
+    IMAGE_FILE_PATTERN,
     VIDEO_DIR,
     update_chunk_file_indices,
 )
 from .video_utils import (
     encode_video_frames,
-    get_video_info,
     reencode_video,
 )
 
@@ -600,7 +611,7 @@ def _keep_episodes_from_video_with_av(
     output_path: Path,
     episodes_to_keep: list[tuple[int, int]],
     fps: float,
-    camera_encoder: VideoEncoderConfig,
+    video_encoder: VideoEncoderConfig,
 ) -> None:
     """Keep only specified episodes from a video file using PyAV.
 
@@ -614,7 +625,7 @@ def _keep_episodes_from_video_with_av(
             Ranges are half-open intervals: [start_frame, end_frame), where start_frame
             is inclusive and end_frame is exclusive.
         fps: Frame rate of the video.
-        camera_encoder: Video encoder settings used to re-encode the kept frames.
+        video_encoder: Video encoder settings used to re-encode the kept frames.
     """
     from fractions import Fraction
 
@@ -639,13 +650,13 @@ def _keep_episodes_from_video_with_av(
 
     # Convert fps to Fraction for PyAV compatibility.
     fps_fraction = Fraction(fps).limit_denominator(1000)
-    codec_options = camera_encoder.get_codec_options(as_strings=True)
-    v_out = out.add_stream(camera_encoder.vcodec, rate=fps_fraction, options=codec_options)
+    codec_options = video_encoder.get_codec_options(as_strings=True)
+    v_out = out.add_stream(video_encoder.vcodec, rate=fps_fraction, options=codec_options)
 
     # PyAV type stubs don't distinguish video streams from audio/subtitle streams.
     v_out.width = v_in.codec_context.width
     v_out.height = v_in.codec_context.height
-    v_out.pix_fmt = camera_encoder.pix_fmt
+    v_out.pix_fmt = video_encoder.pix_fmt
 
     # Set time_base to match the frame rate for proper timestamp handling.
     v_out.time_base = Fraction(1, int(fps))
@@ -732,7 +743,7 @@ def _copy_and_reindex_videos(
 
     for video_key in src_dataset.meta.video_keys:
         logging.info(f"Processing videos for {video_key}")
-        camera_encoder = VideoEncoderConfig.from_video_info(
+        video_encoder = encoder_config_from_video_info(
             src_dataset.meta.info.features.get(video_key, {}).get("info")
         )
 
@@ -816,7 +827,7 @@ def _copy_and_reindex_videos(
                     dst_video_path,
                     episodes_to_keep_ranges,
                     src_dataset.meta.fps,
-                    camera_encoder,
+                    video_encoder,
                 )
 
                 cumulative_ts = 0.0
@@ -873,11 +884,11 @@ def _copy_and_reindex_episodes_metadata(
             episode_meta.update(video_metadata[new_idx])
 
         # Extract episode statistics from parquet metadata.
-        # Note (maractingi): When pandas/pyarrow serializes numpy arrays with shape (3, 1, 1) to parquet,
+        # When pandas/pyarrow serializes numpy arrays with shape (C, 1, 1) to parquet,
         # they are being deserialized as nested object arrays like:
         #   array([array([array([0.])]), array([array([0.])]), array([array([0.])])])
         # This happens particularly with image/video statistics. We need to detect and flatten
-        # these nested structures back to proper (3, 1, 1) arrays so aggregate_stats can process them.
+        # these nested structures back to proper (C, 1, 1) arrays so aggregate_stats can process them.
         episode_stats = {}
         for key in src_episode_full:
             if key.startswith("stats/"):
@@ -893,15 +904,16 @@ def _copy_and_reindex_episodes_metadata(
                     if feature_name in src_dataset.meta.features:
                         feature_dtype = src_dataset.meta.features[feature_name]["dtype"]
                         if feature_dtype in ["image", "video"] and stat_name != "count":
+                            # Stats are channel-first (C, 1, 1)
                             if isinstance(value, np.ndarray) and value.dtype == object:
                                 flat_values = []
                                 for item in value:
                                     while isinstance(item, np.ndarray):
                                         item = item.flatten()[0]
                                     flat_values.append(item)
-                                value = np.array(flat_values, dtype=np.float64).reshape(3, 1, 1)
-                            elif isinstance(value, np.ndarray) and value.shape == (3,):
-                                value = value.reshape(3, 1, 1)
+                                value = np.array(flat_values, dtype=np.float64).reshape(-1, 1, 1)
+                            elif isinstance(value, np.ndarray) and value.ndim == 1:
+                                value = value.reshape(-1, 1, 1)
 
                     episode_stats[feature_name][stat_name] = value
 
@@ -1101,7 +1113,9 @@ def _copy_episodes_metadata_and_stats(
     if dst_meta.video_keys and src_dataset.meta.video_keys:
         for key in dst_meta.video_keys:
             if key in src_dataset.meta.features:
-                dst_meta.info.features[key]["info"] = src_dataset.meta.info.features[key].get("info", {})
+                dst_meta.info.features[key]["info"] = deepcopy(
+                    src_dataset.meta.info.features[key].get("info", {})
+                )
 
     write_info(dst_meta.info, dst_meta.root)
 
@@ -1150,15 +1164,15 @@ def _save_episode_images_for_video(
     # Get all items for this episode
     episode_dataset = imgs_dataset.select(range(from_idx, to_idx))
 
+    is_depth = img_key in dataset.meta.depth_keys
+    frame_pattern = DEPTH_FILE_PATTERN if is_depth else IMAGE_FILE_PATTERN
+
     # Define function to save a single image
     def save_single_image(i_item_tuple):
         i, item = i_item_tuple
-        img = item[img_key]
-        # Use frame-XXXXXX.png format to match encode_video_frames expectations
-        img.save(str(imgs_dir / f"frame-{i:06d}.png"), quality=100)
+        write_image(item[img_key], imgs_dir / frame_pattern.format(frame_index=i))
         return i
 
-    # Save images with proper naming convention for encode_video_frames (frame-XXXXXX.png)
     items = list(enumerate(episode_dataset))
 
     with ThreadPoolExecutor(max_workers=num_workers) as executor:
@@ -1190,13 +1204,14 @@ def _save_batch_episodes_images(
     hf_dataset = dataset.hf_dataset.with_format(None)
     imgs_dataset = hf_dataset.select_columns(img_key)
 
+    is_depth = img_key in dataset.meta.depth_keys
+    frame_pattern = DEPTH_FILE_PATTERN if is_depth else IMAGE_FILE_PATTERN
+
     # Define function to save a single image with global frame index
     # Defined once outside the loop to avoid repeated closure creation
     def save_single_image(i_item_tuple, base_frame_idx, img_key_param):
         i, item = i_item_tuple
-        img = item[img_key_param]
-        # Use global frame index for naming
-        img.save(str(imgs_dir / f"frame-{base_frame_idx + i:06d}.png"), quality=100)
+        write_image(item[img_key_param], imgs_dir / frame_pattern.format(frame_index=base_frame_idx + i))
         return i
 
     episode_durations = []
@@ -1287,7 +1302,7 @@ def _estimate_frame_size_via_calibration(
     episode_indices: list[int],
     temp_dir: Path,
     fps: int,
-    camera_encoder: VideoEncoderConfig,
+    video_encoder: VideoEncoderConfig,
     num_calibration_frames: int = 30,
 ) -> float:
     """Estimate MB per frame by encoding a small calibration sample.
@@ -1301,7 +1316,7 @@ def _estimate_frame_size_via_calibration(
         episode_indices: List of episode indices being processed.
         temp_dir: Temporary directory for calibration files.
         fps: Frames per second for video encoding.
-        camera_encoder: Video encoder settings used for calibration encoding.
+        video_encoder: Video encoder settings used for calibration encoding.
         num_calibration_frames: Number of frames to use for calibration (default: 30).
 
     Returns:
@@ -1326,10 +1341,11 @@ def _estimate_frame_size_via_calibration(
         hf_dataset = dataset.hf_dataset.with_format(None)
         sample_indices = range(from_idx, from_idx + num_frames)
 
-        # Save calibration frames
+        # Save calibration frames using the suffix/format the encoder expects.
+        is_depth = img_key in dataset.meta.depth_keys
+        frame_pattern = DEPTH_FILE_PATTERN if is_depth else IMAGE_FILE_PATTERN
         for i, idx in enumerate(sample_indices):
-            img = hf_dataset[idx][img_key]
-            img.save(str(calibration_dir / f"frame-{i:06d}.png"), quality=100)
+            write_image(hf_dataset[idx][img_key], calibration_dir / frame_pattern.format(frame_index=i))
 
         # Encode calibration video
         calibration_video_path = calibration_dir / "calibration.mp4"
@@ -1337,7 +1353,7 @@ def _estimate_frame_size_via_calibration(
             imgs_dir=calibration_dir,
             video_path=calibration_video_path,
             fps=fps,
-            camera_encoder=camera_encoder,
+            video_encoder=video_encoder,
             overwrite=True,
         )
 
@@ -1610,6 +1626,7 @@ def recompute_stats(
         raise ValueError(f"No parquet files found in {data_dir}")
 
     all_episode_stats = []
+    # TODO: enable image and video stats re-computation
     numeric_keys = [k for k, v in features_to_compute.items() if v["dtype"] not in ["image", "video"]]
 
     for parquet_path in tqdm(parquet_files, desc="Computing stats from data files"):
@@ -1655,7 +1672,8 @@ def convert_image_to_video_dataset(
     dataset: LeRobotDataset,
     output_dir: Path | None = None,
     repo_id: str | None = None,
-    camera_encoder: VideoEncoderConfig | None = None,
+    rgb_encoder: RGBEncoderConfig | None = None,
+    depth_encoder: DepthEncoderConfig | None = None,
     episode_indices: list[int] | None = None,
     num_workers: int = 4,
     max_episodes_per_batch: int | None = None,
@@ -1667,21 +1685,32 @@ def convert_image_to_video_dataset(
     LeRobot dataset structure with videos stored in chunked MP4 files.
 
     Args:
-        dataset: The source LeRobot dataset with images
-        output_dir: Root directory where the edited dataset will be stored. If not specified, defaults to $HF_LEROBOT_HOME/repo_id. Equivalent to new_root in EditDatasetConfig.
-        repo_id: Edited dataset identifier. Equivalent to new_repo_id in EditDatasetConfig.
-        camera_encoder: Video encoder settings
-            (``None`` uses :func:`~lerobot.configs.camera_encoder_defaults`).
-        episode_indices: List of episode indices to convert (None = all episodes)
-        num_workers: Number of threads for parallel processing (default: 4)
-        max_episodes_per_batch: Maximum episodes per video batch to avoid memory issues (None = no limit)
-        max_frames_per_batch: Maximum frames per video batch to avoid memory issues (None = no limit)
+        dataset: The source LeRobot dataset with images.
+        output_dir: Root directory where the converted dataset will be stored. When
+            ``None``, defaults to ``$HF_LEROBOT_HOME/repo_id``. Equivalent to
+            ``new_root`` in ``EditDatasetConfig``.
+        repo_id: Converted dataset identifier. Equivalent to ``new_repo_id`` in
+            ``EditDatasetConfig``.
+        rgb_encoder: Video encoder settings applied to RGB cameras. When ``None``,
+            :func:`~lerobot.configs.video.rgb_encoder_defaults` is used.
+        depth_encoder: Video encoder settings applied to depth-map cameras, including
+            the quantization parameters persisted to the dataset metadata. When
+            ``None``, :func:`~lerobot.configs.video.depth_encoder_defaults` is used.
+        episode_indices: Episode indices to convert. When ``None``, all episodes are
+            converted.
+        num_workers: Number of threads for parallel processing.
+        max_episodes_per_batch: Maximum episodes per video batch, to bound memory use.
+            ``None`` means no limit.
+        max_frames_per_batch: Maximum frames per video batch, to bound memory use.
+            ``None`` means no limit.
 
     Returns:
-        New LeRobotDataset with images encoded as videos
+        A new :class:`LeRobotDataset` with images encoded as videos.
     """
-    if camera_encoder is None:
-        camera_encoder = camera_encoder_defaults()
+    if rgb_encoder is None:
+        rgb_encoder = rgb_encoder_defaults()
+    if depth_encoder is None:
+        depth_encoder = depth_encoder_defaults()
 
     # Check that it's an image dataset
     if len(dataset.meta.video_keys) > 0:
@@ -1706,10 +1735,7 @@ def convert_image_to_video_dataset(
     logging.info(
         f"Converting {len(episode_indices)} episodes with {len(img_keys)} cameras from {dataset.repo_id}"
     )
-    logging.info(
-        f"Video codec: {camera_encoder.vcodec}, pixel format: {camera_encoder.pix_fmt}, "
-        f"GOP: {camera_encoder.g}, CRF: {camera_encoder.crf}"
-    )
+    logging.info(f"RGB video encoder: {rgb_encoder}, depth video encoder: {depth_encoder}")
 
     # Create new features dict, converting image features to video features
     new_features = {}
@@ -1771,6 +1797,8 @@ def convert_image_to_video_dataset(
         episode_lengths = {ep_idx: dataset.meta.episodes["length"][ep_idx] for ep_idx in episode_indices}
 
         for img_key in tqdm(img_keys, desc="Processing cameras"):
+            target_encoder = depth_encoder if img_key in dataset.meta.depth_keys else rgb_encoder
+
             # Estimate size per frame by encoding a small calibration sample
             # This provides accurate compression ratio for the specific codec parameters
             size_per_frame_mb = _estimate_frame_size_via_calibration(
@@ -1779,7 +1807,7 @@ def convert_image_to_video_dataset(
                 episode_indices=episode_indices,
                 temp_dir=temp_dir,
                 fps=fps,
-                camera_encoder=camera_encoder,
+                video_encoder=target_encoder,
             )
 
             logging.info(f"Processing camera: {img_key}")
@@ -1821,7 +1849,7 @@ def convert_image_to_video_dataset(
                     imgs_dir=imgs_dir,
                     video_path=video_path,
                     fps=fps,
-                    camera_encoder=camera_encoder,
+                    video_encoder=target_encoder,
                     overwrite=True,
                 )
 
@@ -1860,16 +1888,11 @@ def convert_image_to_video_dataset(
         new_meta.info.total_tasks = dataset.meta.total_tasks
         new_meta.info.splits = {"train": f"0:{len(episode_indices)}"}
 
-        # Update video info for all image keys (now videos)
-        # We need to manually set video info since update_video_info() checks video_keys first
+        # Update video info for all image keys (now videos). They are registered as
+        # video features above, so update_video_info populates their (still-empty) info.
         for img_key in img_keys:
-            if not new_meta.features[img_key].get("info", None):
-                video_path = new_meta.root / new_meta.video_path.format(
-                    video_key=img_key, chunk_index=0, file_index=0
-                )
-                new_meta.info.features[img_key]["info"] = get_video_info(
-                    video_path, camera_encoder=camera_encoder
-                )
+            target_encoder = depth_encoder if img_key in dataset.meta.depth_keys else rgb_encoder
+            new_meta.update_video_info(video_key=img_key, video_encoder=target_encoder)
 
         write_info(new_meta.info, new_meta.root)
 
@@ -1896,11 +1919,11 @@ def convert_image_to_video_dataset(
 
 def _reencode_video_worker(args: tuple) -> Path:
     """Picklable worker for :func:`reencode_dataset`'s process pool."""
-    video_path, camera_encoder, encoder_threads = args
+    video_path, video_encoder, encoder_threads = args
     reencode_video(
         input_video_path=video_path,
         output_video_path=video_path,
-        camera_encoder=camera_encoder,
+        video_encoder=video_encoder,
         encoder_threads=encoder_threads,
         overwrite=True,
     )
@@ -1909,7 +1932,8 @@ def _reencode_video_worker(args: tuple) -> Path:
 
 def reencode_dataset(
     dataset: LeRobotDataset,
-    camera_encoder: VideoEncoderConfig,
+    rgb_encoder: RGBEncoderConfig | None = None,
+    depth_encoder: DepthEncoderConfig | None = None,
     encoder_threads: int | None = None,
     num_workers: int | None = None,
 ) -> LeRobotDataset:
@@ -1920,8 +1944,11 @@ def reencode_dataset(
     Args:
         dataset: An existing :class:`LeRobotDataset` whose videos will be
             re-encoded.
-        camera_encoder: Target encoder configuration applied to every video
-            file.
+        rgb_encoder: Target encoder configuration applied to every RGB video
+            file. If ``None``, re-encoding is skipped for RGB videos.
+        depth_encoder: Target encoder configuration applied to every depth video
+            file. If ``None``, re-encoding is skipped for depth videos.
+            Quantization parameters will not override the ones in the current dataset.
         encoder_threads: Per-encoder thread count forwarded to
             :func:`reencode_video`. ``None`` lets the codec decide.
         num_workers: Number of parallel processes. ``None`` or ``0`` means
@@ -1933,23 +1960,35 @@ def reencode_dataset(
         on disk.
     """
     meta = dataset.meta
-    video_paths_list = []
+    video_keys_encoders_dict = {}
+    video_keys_paths_dict = {}
+
+    if rgb_encoder is None and depth_encoder is None:
+        raise ValueError("Either rgb_encoder or depth_encoder must be provided")
 
     # Only re-encode if the videos are not already encoded with the given video encoding parameters
     for video_key in meta.video_keys:
         current_info = meta.info.features[video_key].get("info", {})
-        current_encoder = VideoEncoderConfig.from_video_info(current_info)
-        if current_encoder != camera_encoder:
-            video_paths_list.extend((meta.root / VIDEO_DIR / video_key).rglob("*.mp4"))
+        current_encoder = encoder_config_from_video_info(current_info)
+        target_encoder = depth_encoder if video_key in meta.depth_keys else rgb_encoder
+        if target_encoder is None:
+            logging.info(f"No encoder provided for {video_key} video. Skipping re-encoding.")
+        elif current_encoder != target_encoder:
+            video_keys_paths_dict[video_key] = list((meta.root / VIDEO_DIR / video_key).rglob("*.mp4"))
+            video_keys_encoders_dict[video_key] = target_encoder
         else:
-            logging.info(f"{video_key} videos are already encoded with {camera_encoder}. Nothing to do.")
+            logging.info(f"{video_key} videos are already encoded with {target_encoder}. Nothing to do.")
 
-    if len(video_paths_list) == 0:
+    if len(video_keys_paths_dict) == 0:
         logging.warning("Dataset has no videos to re-encode.")
         return dataset
-    logging.info(f"Re-encoding {len(video_paths_list)} video file(s) with {camera_encoder}")
+    logging.info(f"Re-encoding {sum(len(paths) for paths in video_keys_paths_dict.values())} video file(s).")
 
-    worker_args = [(vp, camera_encoder, encoder_threads) for vp in video_paths_list]
+    worker_args = [
+        (path, encoder, encoder_threads)
+        for video_key, encoder in video_keys_encoders_dict.items()
+        for path in video_keys_paths_dict[video_key]
+    ]
     if num_workers and num_workers > 1:
         with ProcessPoolExecutor(max_workers=num_workers) as pool:
             futures = [pool.submit(_reencode_video_worker, args) for args in worker_args]
@@ -1963,10 +2002,15 @@ def reencode_dataset(
         for args in tqdm(worker_args, desc="Re-encoding videos"):
             _reencode_video_worker(args)
 
-    # Refresh video info in metadata for every video key.
-    for vid_key in meta.video_keys:
-        video_path = meta.root / meta.get_video_file_path(0, vid_key)
-        meta.info.features[vid_key]["info"] = get_video_info(video_path, camera_encoder=camera_encoder)
+    # Refresh video info in metadata for every re-encoded key. Re-encoding only
+    # changes codec/container params, so for depth videos we preserve ``is_depth_map``
+    # and the depth quantization params (``video.depth_min`` / ``video.depth_max`` /
+    # ...), which describe the data rather than the codec and must survive a transcode.
+    # RGB videos pass an empty set: still a refresh, but nothing to preserve.
+    depth_preserve_keys = {"is_depth_map", *(f"video.{n}" for n in DEPTH_ENCODER_INFO_FIELD_NAMES)}
+    for video_key, encoder in video_keys_encoders_dict.items():
+        preserve_keys = depth_preserve_keys if video_key in meta.depth_keys else set()
+        meta.update_video_info(video_key=video_key, video_encoder=encoder, preserve_keys=preserve_keys)
 
     write_info(meta.info, meta.root)
     logging.info("Dataset metadata updated.")

src/lerobot/datasets/dataset_writer.py

@@ -31,7 +31,14 @@ import PIL.Image
 import pyarrow.parquet as pq
 import torch
 
-from lerobot.configs import VideoEncoderConfig, camera_encoder_defaults
+from lerobot.configs import (
+    DepthEncoderConfig,
+    RGBEncoderConfig,
+    VideoEncoderConfig,
+    depth_encoder_defaults,
+    infer_depth_unit,
+    rgb_encoder_defaults,
+)
 
 from .compute_stats import compute_episode_stats
 from .dataset_metadata import LeRobotDatasetMetadata
@@ -48,6 +55,7 @@ from .io_utils import (
     write_info,
 )
 from .utils import (
+    DEFAULT_DEPTH_PATH,
     DEFAULT_EPISODES_PATH,
     DEFAULT_IMAGE_PATH,
     update_chunk_file_indices,
@@ -67,17 +75,22 @@ def _encode_video_worker(
     episode_index: int,
     root: Path,
     fps: int,
-    camera_encoder: VideoEncoderConfig | None = None,
+    video_encoder: VideoEncoderConfig | None = None,
     encoder_threads: int | None = None,
 ) -> Path:
     temp_path = Path(tempfile.mkdtemp(dir=root)) / f"{video_key}_{episode_index:03d}.mp4"
-    fpath = DEFAULT_IMAGE_PATH.format(image_key=video_key, episode_index=episode_index, frame_index=0)
+    path_template = (
+        DEFAULT_DEPTH_PATH
+        if video_encoder is not None and isinstance(video_encoder, DepthEncoderConfig)
+        else DEFAULT_IMAGE_PATH
+    )
+    fpath = path_template.format(image_key=video_key, episode_index=episode_index, frame_index=0)
     img_dir = (root / fpath).parent
     encode_video_frames(
         img_dir,
         temp_path,
         fps,
-        camera_encoder=camera_encoder,
+        video_encoder=video_encoder,
         encoder_threads=encoder_threads,
         overwrite=True,
     )
@@ -96,7 +109,8 @@ class DatasetWriter:
         self,
         meta: LeRobotDatasetMetadata,
         root: Path,
-        camera_encoder: VideoEncoderConfig | None,
+        rgb_encoder: RGBEncoderConfig | None,
+        depth_encoder: DepthEncoderConfig | None,
         encoder_threads: int | None,
         batch_encoding_size: int,
         streaming_encoder: StreamingVideoEncoder | None = None,
@@ -108,8 +122,11 @@ class DatasetWriter:
             meta: Dataset metadata instance (used for feature schema, chunk
                 settings, and episode persistence).
             root: Local dataset root directory.
-            camera_encoder: Video encoder settings applied to all cameras.
-                ``None`` uses :func:`~lerobot.configs.camera_encoder_defaults`.
+            rgb_encoder: Video encoder settings applied to RGB cameras. When
+                ``None``, :func:`~lerobot.configs.video.rgb_encoder_defaults` is used.
+            depth_encoder: Video encoder settings applied to depth cameras, including
+                the quantization parameters. When ``None``,
+                :func:`~lerobot.configs.video.depth_encoder_defaults` is used.
             encoder_threads: Number of encoder threads (global). ``None``
                 lets the codec decide.
             batch_encoding_size: Number of episodes to accumulate before
@@ -120,7 +137,8 @@ class DatasetWriter:
         """
         self._meta = meta
         self._root = root
-        self._camera_encoder = camera_encoder or camera_encoder_defaults()
+        self._rgb_encoder = rgb_encoder or rgb_encoder_defaults()
+        self._depth_encoder = depth_encoder or depth_encoder_defaults()
         self._encoder_threads = encoder_threads
         self._batch_encoding_size = batch_encoding_size
         self._streaming_encoder = streaming_encoder
@@ -145,7 +163,8 @@ class DatasetWriter:
         return ep_buffer
 
     def _get_image_file_path(self, episode_index: int, image_key: str, frame_index: int) -> Path:
-        fpath = DEFAULT_IMAGE_PATH.format(
+        path_template = DEFAULT_DEPTH_PATH if image_key in self._meta.depth_keys else DEFAULT_IMAGE_PATH
+        fpath = path_template.format(
             image_key=image_key, episode_index=episode_index, frame_index=frame_index
         )
         return self._root / fpath
@@ -191,10 +210,20 @@ 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(
                 video_keys=list(self._meta.video_keys),
+                depth_video_keys=list(self._meta.depth_keys),
                 temp_dir=self._root,
             )
 
@@ -282,10 +311,13 @@ class DatasetWriter:
         if use_streaming:
             streaming_results = self._streaming_encoder.finish_episode()
             for video_key in self._meta.video_keys:
+                normalization_factor = 255.0 if video_key not in self._meta.depth_keys else 1.0
                 temp_path, video_stats = streaming_results[video_key]
                 if video_stats is not None:
                     ep_stats[video_key] = {
-                        k: v if k == "count" else np.squeeze(v.reshape(1, -1, 1, 1) / 255.0, axis=0)
+                        k: v
+                        if k == "count"
+                        else np.squeeze(v.reshape(1, -1, 1, 1) / normalization_factor, axis=0)
                         for k, v in video_stats.items()
                     }
                 ep_metadata.update(self._save_episode_video(video_key, episode_index, temp_path=temp_path))
@@ -300,7 +332,7 @@ class DatasetWriter:
                             episode_index,
                             self._root,
                             self._meta.fps,
-                            self._camera_encoder,
+                            self._depth_encoder if video_key in self._meta.depth_keys else self._rgb_encoder,
                             self._encoder_threads,
                         ): video_key
                         for video_key in self._meta.video_keys
@@ -511,7 +543,12 @@ class DatasetWriter:
 
         # Update video info (only needed when first episode is encoded)
         if episode_index == 0:
-            self._meta.update_video_info(video_key, camera_encoder=self._camera_encoder)
+            self._meta.update_video_info(
+                video_key,
+                video_encoder=self._depth_encoder
+                if video_key in self._meta.depth_keys
+                else self._rgb_encoder,
+            )
             write_info(self._meta.info, self._meta.root)
 
         metadata = {
@@ -578,13 +615,14 @@ class DatasetWriter:
             self.image_writer.wait_until_done()
 
     def _encode_temporary_episode_video(self, video_key: str, episode_index: int) -> Path:
-        """Use ffmpeg to convert frames stored as png into mp4 videos."""
+        """Use ffmpeg to convert frames stored as png/tiff into mp4 videos."""
+        is_depth = video_key in self._meta.depth_keys
         return _encode_video_worker(
             video_key,
             episode_index,
             self._root,
             self._meta.fps,
-            self._camera_encoder,
+            self._depth_encoder if is_depth else self._rgb_encoder,
             self._encoder_threads,
         )
 

src/lerobot/datasets/depth_utils.py

@@ -0,0 +1,265 @@
+#!/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.
+"""
+Depth encoding/decoding helpers for :class:`DepthEncoderConfig`.
+"""
+
+import math
+from typing import Literal
+
+import av
+import numpy as np
+import torch
+from numpy.typing import NDArray
+
+from lerobot.configs.video import (
+    DEFAULT_DEPTH_MAX,
+    DEFAULT_DEPTH_MIN,
+    DEFAULT_DEPTH_PIX_FMT,
+    DEFAULT_DEPTH_SHIFT,
+    DEFAULT_DEPTH_USE_LOG,
+    DEPTH_METER_UNIT,
+    DEPTH_MILLIMETER_UNIT,
+    DEPTH_QMAX,
+    infer_depth_unit,
+)
+
+from .image_writer import squeeze_single_channel
+from .pyav_utils import write_u16_plane
+
+MM_PER_METRE = 1000.0
+_UINT16_MAX = 65535
+
+
+def _validate_log_quant_params(depth_min: float, shift: float) -> None:
+    """Ensure ``log(depth_min + shift)`` is finite."""
+    if depth_min + shift <= 0:
+        raise ValueError(
+            f"depth_min + shift must be positive for logarithmic quantization, "
+            f"got depth_min={depth_min} + shift={shift} = {depth_min + shift}"
+        )
+
+
+def _depth_input_to_float32_and_unit(
+    depth: NDArray[np.integer] | NDArray[np.floating],
+    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
+    return depth.astype(np.float32, order="K"), resolved_unit
+
+
+def quantize_depth(
+    depth: NDArray[np.uint16] | NDArray[np.float32] | torch.Tensor,
+    depth_min: float = DEFAULT_DEPTH_MIN,
+    depth_max: float = DEFAULT_DEPTH_MAX,
+    shift: float = DEFAULT_DEPTH_SHIFT,
+    use_log: bool = DEFAULT_DEPTH_USE_LOG,
+    pix_fmt: str = DEFAULT_DEPTH_PIX_FMT,
+    video_backend: str | None = "pyav",
+    input_unit: Literal["auto", DEPTH_METER_UNIT, DEPTH_MILLIMETER_UNIT] = "auto",
+) -> NDArray[np.uint16] | av.VideoFrame:
+    """Quantize depth to 12-bit codes (``uint16``, values ``0…DEPTH_QMAX``).
+
+    Depth maps are packed into 12-bit integer frames so they fit in standard
+    high-bit-depth pixel formats (e.g. ``yuv420p12le`` / ``gray12le``)
+    and can be encoded by widely supported video codecs (e.g. HEVC Main 12).
+    Logarithmic quantization is the default because it allocates more quanta
+    to near-range depth, which matches the (1/depth) error profile of typical
+    depth sensors. Math is ported from BEHAVIOR-1K's ``obs_utils.py``.
+
+    **Input units**:
+
+    - ``input_unit="auto"`` (default): infer from dtype (floating = m, non-floating = mm).
+    - ``input_unit="mm"``: interpret input values as millimetres.
+    - ``input_unit="m"``: interpret input values as metres.
+
+    Quantization math runs in the **resolved input unit**.
+
+    ``depth_min``, ``depth_max``, and ``shift`` are always in **metres**.
+
+    Args:
+        depth: Depth map; ``torch.Tensor`` is moved to CPU for conversion.
+        depth_min: Depth (metres) at quantum ``0``.
+        depth_max: Depth (metres) at quantum :data:`DEPTH_QMAX`.
+        shift: Depth shift (metres); used in log mode. Must satisfy ``depth_min + shift > 0``.
+        use_log: If ``True`` (default), quantize in log space.
+        video_backend: Video backend to use for encoding. Defaults to "pyav".
+        input_unit: Input unit policy (``"auto"``, ``"mm"``, ``"m"``).
+
+    Returns:
+        ``numpy.ndarray``, ``dtype=uint16``, same shape as ``depth``, values in
+        ``[0, DEPTH_QMAX]``.
+
+    Raises:
+        ValueError: If ``input_unit`` is not ``"auto"``, ``"mm"``, or ``"m"``.
+        ValueError: If ``use_log=True`` and ``depth_min + shift <= 0``.
+    """
+    if input_unit not in ("auto", DEPTH_METER_UNIT, DEPTH_MILLIMETER_UNIT):
+        raise ValueError(
+            f"input_unit must be 'auto', '{DEPTH_METER_UNIT}', or '{DEPTH_MILLIMETER_UNIT}', got {input_unit!r}"
+        )
+
+    if isinstance(depth, torch.Tensor):
+        depth = depth.detach().cpu().numpy()
+
+    # Squeeze single-channel dim: (H, W, 1) or (1, H, W) → (H, W)
+    depth = squeeze_single_channel(depth)
+
+    depth_f, resolved_unit = _depth_input_to_float32_and_unit(depth, input_unit=input_unit)
+
+    # Convert depth_min, depth_max, and shift to the resolved input unit.
+    depth_min_u = (
+        np.float32(depth_min) if resolved_unit == DEPTH_METER_UNIT else np.float32(depth_min * MM_PER_METRE)
+    )
+    depth_max_u = (
+        np.float32(depth_max) if resolved_unit == DEPTH_METER_UNIT else np.float32(depth_max * MM_PER_METRE)
+    )
+    shift_u = np.float32(shift) if resolved_unit == DEPTH_METER_UNIT else np.float32(shift * MM_PER_METRE)
+
+    # Normalization and quantization is performed in the resolved input unit.
+    if use_log:
+        _validate_log_quant_params(depth_min, shift)
+        log_min = math.log(float(depth_min_u + shift_u))
+        log_max = math.log(float(depth_max_u + shift_u))
+        norm = (np.log(depth_f + shift_u) - log_min) / (log_max - log_min)
+    else:
+        norm = (depth_f - depth_min_u) / (depth_max_u - depth_min_u)
+
+    quantized = np.rint(norm * DEPTH_QMAX).clip(0, DEPTH_QMAX).astype(np.uint16, copy=False)
+
+    if video_backend == "pyav":
+        frame = av.VideoFrame.from_ndarray(quantized, format=pix_fmt)
+        write_u16_plane(frame.planes[0], quantized)
+        return frame
+    else:
+        return quantized
+
+
+def dequantize_depth(
+    quantized: NDArray[np.uint16] | av.VideoFrame | torch.Tensor,
+    depth_min: float = DEFAULT_DEPTH_MIN,
+    depth_max: float = DEFAULT_DEPTH_MAX,
+    shift: float = DEFAULT_DEPTH_SHIFT,
+    use_log: bool = DEFAULT_DEPTH_USE_LOG,
+    pix_fmt: str = DEFAULT_DEPTH_PIX_FMT,
+    output_unit: Literal[DEPTH_METER_UNIT, DEPTH_MILLIMETER_UNIT] = DEPTH_MILLIMETER_UNIT,
+    output_tensor: bool = True,
+    output_channel_last: bool = False,
+) -> NDArray[np.uint16] | NDArray[np.float32] | torch.Tensor:
+    """Inverse of :func:`quantize_depth`.
+
+    Decoding inverts the same normalized code mapping as :func:`quantize_depth`
+    using ``depth_min`` / ``depth_max`` / ``shift`` (in metres), then returns
+    the requested output unit. Tuning arguments **must match** :func:`quantize_depth`.
+
+    Accepted input layouts :
+
+    - ``(H, W, 1)`` or ``(H, W)`` — single frame with channel-last.
+    - ``(..., 1, H, W)`` — batched frames with channel-first.
+    - ``(..., H, W, 1)`` — batched frames with channel-last.
+    Output layout is determined by ``output_channel_last``.
+
+    Args:
+        quantized: 12-bit codes in ``[0, DEPTH_QMAX]``. ``np.ndarray``,
+            ``av.VideoFrame``, or ``torch.Tensor`` (any integer or float dtype).
+        depth_min, depth_max, shift, use_log: Same as :func:`quantize_depth` (metres).
+        pix_fmt: Pixel format used to extract the plane from an ``av.VideoFrame``.
+        output_unit: ``"mm"`` returns ``uint16`` millimetres (rint, clip
+            ``[0, 65535]``) when returning a numpy array, or ``float32`` mm when
+            ``output_tensor=True``. ``"m"`` returns ``float32`` metres in
+            ``[depth_min, depth_max]``.
+        output_tensor: If True, return a ``torch.Tensor`` instead of a numpy array.
+
+    Returns:
+        Depth map in the requested unit and dtype.
+
+    Raises:
+        ValueError: If ``output_unit`` is not ``"m"`` or ``"mm"``.
+        ValueError: If ``use_log=True`` and ``depth_min + shift <= 0``.
+    """
+    if output_unit not in (DEPTH_METER_UNIT, DEPTH_MILLIMETER_UNIT):
+        raise ValueError(
+            f"output_unit must be '{DEPTH_METER_UNIT}' or '{DEPTH_MILLIMETER_UNIT}', got {output_unit!r}"
+        )
+    if use_log:
+        _validate_log_quant_params(depth_min, shift)
+
+    if isinstance(quantized, av.VideoFrame):
+        quantized = quantized.to_ndarray(format=pix_fmt)
+
+    # Compute the scale and offset first.
+    depth_min_m = float(depth_min)
+    depth_max_m = float(depth_max)
+    shift_m = float(shift)
+    if use_log:
+        log_min = math.log(depth_min_m + shift_m)
+        log_max = math.log(depth_max_m + shift_m)
+        scale = (log_max - log_min) / DEPTH_QMAX
+        offset = log_min
+    else:
+        scale = (depth_max_m - depth_min_m) / DEPTH_QMAX
+        offset = depth_min_m
+
+    # ── Torch path: stay on the input device, single fp32 allocation. ────────
+    if isinstance(quantized, torch.Tensor):
+        if quantized.ndim >= 3:
+            # Drop the single-channel dimension so the math runs on (..., H, W).
+            quantized = quantized.squeeze(-3) if quantized.shape[-3] == 1 else quantized.squeeze(-1)
+
+        # Single allocation we own; everything else is in-place.
+        buf = quantized.to(dtype=torch.float32, copy=True)
+        buf.mul_(scale).add_(offset)
+        if use_log:
+            buf.exp_().sub_(shift_m)
+        buf.clamp_(depth_min_m, depth_max_m)
+        buf.unsqueeze_(-1) if output_channel_last else buf.unsqueeze_(-3)
+
+        if output_unit == DEPTH_METER_UNIT:
+            return buf if output_tensor else buf.cpu().numpy()
+
+        # mm path: round + clamp in float32, skipping the uint16 round-trip
+        # when returning a tensor (torch.uint16 is poorly supported).
+        buf.mul_(MM_PER_METRE).round_().clamp_(0.0, _UINT16_MAX)
+        if output_tensor:
+            return buf
+        return buf.cpu().numpy().astype(np.uint16, copy=False)
+
+    # ── NumPy path: single fp32 allocation, ``out=`` for in-place math. ─────
+    arr = np.asarray(quantized)
+    if arr.ndim >= 3:
+        # Drop the single-channel dimension so the math runs on (..., H, W).
+        arr = np.squeeze(arr, axis=-3) if arr.shape[-3] == 1 else np.squeeze(arr, axis=-1)
+
+    buf = np.empty(arr.shape, dtype=np.float32)
+    np.multiply(arr, scale, out=buf)
+    np.add(buf, offset, out=buf)
+    if use_log:
+        np.exp(buf, out=buf)
+        np.subtract(buf, shift_m, out=buf)
+    np.clip(buf, depth_min_m, depth_max_m, out=buf)
+    buf = np.expand_dims(buf, axis=-1) if output_channel_last else np.expand_dims(buf, axis=-3)
+
+    if output_unit == DEPTH_METER_UNIT:
+        return torch.from_numpy(buf) if output_tensor else buf
+
+    np.multiply(buf, MM_PER_METRE, out=buf)
+    np.rint(buf, out=buf)
+    np.clip(buf, 0.0, _UINT16_MAX, out=buf)
+    if output_tensor:
+        # torch.uint16 support is very limited; return float32 millimetres.
+        return torch.from_numpy(buf)
+    return buf.astype(np.uint16, copy=False)

src/lerobot/datasets/factory.py

@@ -14,6 +14,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import logging
+import math
 from pprint import pformat
 
 import torch
@@ -96,6 +97,7 @@ def make_dataset(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDatas
                 revision=cfg.dataset.revision,
                 video_backend=cfg.dataset.video_backend,
                 return_uint8=True,
+                depth_output_unit=cfg.dataset.depth_output_unit,
                 tolerance_s=cfg.tolerance_s,
             )
         else:
@@ -126,7 +128,87 @@ def make_dataset(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDatas
 
     if cfg.dataset.use_imagenet_stats:
         for key in dataset.meta.camera_keys:
+            if key in dataset.meta.depth_keys:
+                continue  # Exclude depth keys from ImageNet stats
             for stats_type, stats in IMAGENET_STATS.items():
                 dataset.meta.stats[key][stats_type] = torch.tensor(stats, dtype=torch.float32)
 
     return dataset
+
+
+def make_train_eval_datasets(
+    cfg: TrainPipelineConfig,
+) -> tuple[LeRobotDataset | MultiLeRobotDataset, LeRobotDataset | None]:
+    """Create train and optional eval datasets by splitting episodes based on eval_split.
+
+    The last ceil(n_episodes * eval_split) episodes per task are held out for evaluation.
+    If eval_split == 0.0, returns (full_dataset, None).
+    """
+    full_dataset = make_dataset(cfg)
+
+    if cfg.dataset.eval_split == 0.0:
+        return full_dataset, None
+
+    base_episodes = (
+        full_dataset.episodes if full_dataset.episodes is not None else list(range(full_dataset.num_episodes))
+    )
+
+    episode_tasks = full_dataset.meta.episodes["tasks"]
+    task_to_episodes: dict[str, list[int]] = {}
+    for ep_idx in base_episodes:
+        task_key = episode_tasks[ep_idx][0] if episode_tasks[ep_idx] else ""
+        task_to_episodes.setdefault(task_key, []).append(ep_idx)
+
+    train_episodes, eval_episodes = [], []
+    for eps in task_to_episodes.values():
+        n_eval = math.ceil(len(eps) * cfg.dataset.eval_split)
+        train_episodes.extend(eps[: len(eps) - n_eval])
+        eval_episodes.extend(eps[len(eps) - n_eval :])
+
+    if not train_episodes:
+        raise ValueError(
+            f"eval_split={cfg.dataset.eval_split} leaves 0 training episodes from {len(base_episodes)} total."
+        )
+
+    logging.info(
+        f"Train/eval split: {len(train_episodes)} train, {len(eval_episodes)} eval "
+        f"(eval_split={cfg.dataset.eval_split}, {len(task_to_episodes)} tasks)"
+    )
+
+    delta_timestamps = resolve_delta_timestamps(cfg.trainable_config, full_dataset.meta)
+
+    train_image_transforms = (
+        ImageTransforms(cfg.dataset.image_transforms) if cfg.dataset.image_transforms.enable else None
+    )
+
+    train_dataset = LeRobotDataset(
+        cfg.dataset.repo_id,
+        root=cfg.dataset.root,
+        episodes=train_episodes,
+        delta_timestamps=delta_timestamps,
+        image_transforms=train_image_transforms,
+        revision=cfg.dataset.revision,
+        video_backend=cfg.dataset.video_backend,
+        return_uint8=True,
+        tolerance_s=cfg.tolerance_s,
+    )
+
+    eval_dataset = LeRobotDataset(
+        cfg.dataset.repo_id,
+        root=cfg.dataset.root,
+        episodes=eval_episodes,
+        delta_timestamps=delta_timestamps,
+        image_transforms=None,
+        revision=cfg.dataset.revision,
+        video_backend=cfg.dataset.video_backend,
+        return_uint8=True,
+        tolerance_s=cfg.tolerance_s,
+    )
+
+    if cfg.dataset.use_imagenet_stats:
+        for ds in (train_dataset, eval_dataset):
+            for key in ds.meta.camera_keys:
+                for stats_type, stats in IMAGENET_STATS.items():
+                    ds.meta.stats[key][stats_type] = torch.tensor(stats, dtype=torch.float32)
+
+    return train_dataset, eval_dataset

src/lerobot/datasets/feature_utils.py

@@ -67,9 +67,9 @@ def get_hf_features_from_features(features: dict) -> datasets.Features:
         elif ft["shape"] == (1,):
             hf_features[key] = datasets.Value(dtype=ft["dtype"])
         elif len(ft["shape"]) == 1:
-            hf_features[key] = datasets.Sequence(
-                length=ft["shape"][0], feature=datasets.Value(dtype=ft["dtype"])
-            )
+            # pyarrow rejects fixed-size lists of length 0, so use a variable length list instead
+            length = ft["shape"][0] if ft["shape"][0] > 0 else -1
+            hf_features[key] = datasets.Sequence(length=length, feature=datasets.Value(dtype=ft["dtype"]))
         elif len(ft["shape"]) == 2:
             hf_features[key] = datasets.Array2D(shape=ft["shape"], dtype=ft["dtype"])
         elif len(ft["shape"]) == 3:
@@ -336,7 +336,7 @@ def validate_feature_image_or_video(
 
     Args:
         name (str): The name of the feature.
-        expected_shape (list[str]): The expected shape (C, H, W).
+        expected_shape (list[str]): The expected shape, e.g. (C, H, W) or (H, W, C).
         value: The image data to validate.
 
     Returns:

src/lerobot/datasets/image_writer.py

@@ -41,11 +41,51 @@ def safe_stop_image_writer(func):
     return wrapper
 
 
-def image_array_to_pil_image(image_array: np.ndarray, range_check: bool = True) -> PIL.Image.Image:
-    # TODO(aliberts): handle 1 channel and 4 for depth images
-    if image_array.ndim != 3:
-        raise ValueError(f"The array has {image_array.ndim} dimensions, but 3 is expected for an image.")
+def squeeze_single_channel(array: np.ndarray) -> np.ndarray:
+    """Drop a leading or trailing singleton channel dim: ``(1, H, W)`` / ``(H, W, 1)`` -> ``(H, W)``.
 
+    Unlike ``array.squeeze()``, this only removes the channel axis, never an ``H`` or ``W`` of size 1.
+    """
+    if array.ndim == 3:
+        if array.shape[0] == 1:
+            return array[0]
+        if array.shape[-1] == 1:
+            return array[..., 0]
+    return array
+
+
+def image_array_to_pil_image(image_array: np.ndarray, range_check: bool = True) -> PIL.Image.Image:
+    """Convert a NumPy array to a PIL Image, preserving precision for grayscale.
+
+    Behaviour by shape:
+
+    - ``(H, W)`` or ``(1, H, W)`` / ``(H, W, 1)``: single-channel grayscale.
+      The native dtype is preserved using the matching PIL mode
+      (``I;16`` / ``F``). This is the path used for raw depth maps (no rescaling, clamping, or downcasting)
+    - ``(3, H, W)`` / ``(H, W, 3)``: RGB. Channels-first inputs are transposed
+      to channels-last. Float inputs in ``[0, 1]`` are scaled to ``uint8``
+      (existing behaviour, gated by ``range_check``).
+
+    Other shapes / channel counts raise ``NotImplementedError`` or
+    ``ValueError``.
+    """
+    # TODO(CarolinePascal): 4 dimensions RGB-D images
+    if image_array.ndim not in (2, 3):
+        raise ValueError(f"The array has {image_array.ndim} dimensions, but 2 or 3 is expected for an image.")
+
+    # Squeeze 3D single-channel inputs to 2D so depth maps work whether the
+    # caller emits (H, W), (1, H, W), or (H, W, 1).
+    image_array = squeeze_single_channel(image_array)
+
+    if image_array.ndim == 2:
+        if image_array.dtype not in [np.uint16, np.float32]:
+            raise ValueError(
+                f"Unsupported single-channel image dtype: {image_array.dtype}. "
+                f"Supported dtypes: {sorted(str(d) for d in [np.uint16, np.float32])}."
+            )
+        return PIL.Image.fromarray(np.ascontiguousarray(image_array))
+
+    # 3D path: must be RGB (3 channels), channels-first or channels-last.
     if image_array.shape[0] == 3:
         # Transpose from pytorch convention (C, H, W) to (H, W, C)
         image_array = image_array.transpose(1, 2, 0)
@@ -71,13 +111,29 @@ def image_array_to_pil_image(image_array: np.ndarray, range_check: bool = True)
     return PIL.Image.fromarray(image_array)
 
 
+def save_kwargs_for_path(fpath: Path, compress_level: int) -> dict:
+    """Pick the right format-specific kwargs for :meth:`PIL.Image.Image.save`.
+
+    PNG uses ``compress_level`` (0-9, zlib). TIFF uses ``compression`` (raw) for lossless raw depth maps.
+    """
+    suffix = Path(fpath).suffix.lower()
+    if suffix == ".png":
+        return {"compress_level": compress_level}
+    if suffix in (".tif", ".tiff"):
+        return {"compression": "raw"}
+    else:
+        raise ValueError(f"Unsupported image file extension: {suffix}")
+
+
 def write_image(image: np.ndarray | PIL.Image.Image, fpath: Path, compress_level: int = 1):
     """
     Saves a NumPy array or PIL Image to a file.
 
     This function handles both NumPy arrays and PIL Image objects, converting
     the former to a PIL Image before saving. It includes error handling for
-    the save operation.
+    the save operation. The output format is inferred from the *fpath*
+    extension: ``.png`` → PNG with ``compress_level``, ``.tiff`` / ``.tif``
+    → lossless raw depth maps (TIFF).
 
     Args:
         image (np.ndarray | PIL.Image.Image): The image data to save.
@@ -101,7 +157,7 @@ def write_image(image: np.ndarray | PIL.Image.Image, fpath: Path, compress_level
             img = image
         else:
             raise TypeError(f"Unsupported image type: {type(image)}")
-        img.save(fpath, compress_level=compress_level)
+        img.save(fpath, **save_kwargs_for_path(fpath, compress_level))
     except Exception as e:
         logger.error("Error writing image %s: %s", fpath, e)
 

src/lerobot/datasets/io_utils.py

@@ -20,6 +20,7 @@ import datasets
 import numpy as np
 import pandas
 import pandas as pd
+import pyarrow as pa
 import pyarrow.dataset as pa_ds
 import pyarrow.parquet as pq
 import torch
@@ -153,7 +154,7 @@ def cast_stats_to_numpy(stats: dict) -> dict[str, dict[str, np.ndarray]]:
     Returns:
         dict: The statistics dictionary with values cast to numpy arrays.
     """
-    stats = {key: np.array(value) for key, value in flatten_dict(stats).items()}
+    stats = {key: np.atleast_1d(np.array(value)) for key, value in flatten_dict(stats).items()}
     return unflatten_dict(stats)
 
 
@@ -225,28 +226,50 @@ def load_image_as_numpy(
     Args:
         fpath (str | Path): Path to the image file.
         dtype (np.dtype): The desired data type of the output array. If floating,
-            pixels are scaled to [0, 1].
+            pixels are scaled to [0, 1]. Only used for RGB images.
         channel_first (bool): If True, converts the image to (C, H, W) format.
             Otherwise, it remains in (H, W, C) format.
 
     Returns:
         np.ndarray: The image as a numpy array.
     """
-    img = PILImage.open(fpath).convert("RGB")
-    img_array = np.array(img, dtype=dtype)
+    is_depth = fpath.endswith(".tiff") or fpath.endswith(".tif")
+    if is_depth:
+        # Preserve the native depth dtype (uint16 -> "I;16", float32 -> "F").
+        img = PILImage.open(fpath)
+        img_array = np.array(img)
+    else:
+        img = PILImage.open(fpath).convert("RGB")
+        img_array = np.array(img, dtype=dtype)
+        if np.issubdtype(dtype, np.floating):
+            img_array /= 255.0
     if channel_first:  # (H, W, C) -> (C, H, W)
-        img_array = np.transpose(img_array, (2, 0, 1))
-    if np.issubdtype(dtype, np.floating):
-        img_array /= 255.0
+        img_array = img_array[np.newaxis, ...] if img_array.ndim == 2 else np.transpose(img_array, (2, 0, 1))
     return img_array
 
 
+# PIL modes for 16-bit unsigned depth maps.
+UINT16_PIL_MODES = {"I;16", "I;16B", "I;16L"}
+
+
+def pil_to_chw_tensor(img: PILImage.Image) -> torch.Tensor:
+    """Convert a PIL image to a channel-first tensor.
+
+    ``uint16`` depth maps become ``float32 (1, H, W)`` in native units (``ToTensor``
+    would overflow them to ``int16``); all other modes use the standard ``ToTensor`` path.
+    """
+    if img.mode in UINT16_PIL_MODES:
+        return torch.from_numpy(np.array(img, dtype=np.float32))[None, ...]
+    return transforms.ToTensor()(img)
+
+
 def hf_transform_to_torch(items_dict: dict[str, list[Any]]) -> dict[str, list[torch.Tensor | str]]:
     """Convert a batch from a Hugging Face dataset to torch tensors.
 
     This transform function converts items from Hugging Face dataset format (pyarrow)
-    to torch tensors. Importantly, images are converted from PIL objects (H, W, C, uint8)
-    to a torch image representation (C, H, W, float32) in the range [0, 1]. Other
+    to torch tensors. RGB images are converted from PIL objects (H, W, C, uint8)
+    to a torch image representation (C, H, W, float32) in the range [0, 1]. Depth
+    maps are returned as float32 (1, H, W) in their native units. Other
     types are converted to torch.tensor.
 
     Args:
@@ -261,8 +284,7 @@ def hf_transform_to_torch(items_dict: dict[str, list[Any]]) -> dict[str, list[to
             continue
         first_item = items_dict[key][0]
         if isinstance(first_item, PILImage.Image):
-            to_tensor = transforms.ToTensor()
-            items_dict[key] = [to_tensor(img) for img in items_dict[key]]
+            items_dict[key] = [pil_to_chw_tensor(img) for img in items_dict[key]]
         elif first_item is None or isinstance(first_item, dict):
             pass
         else:
@@ -270,21 +292,49 @@ def hf_transform_to_torch(items_dict: dict[str, list[Any]]) -> dict[str, list[to
     return items_dict
 
 
+def write_table_one_row_group_per_episode(table: pa.Table, path: Path) -> None:
+    """Write ``table`` with one parquet row group per episode (in episode order).
+
+    Keeps shards random-access friendly (``read_row_group(i)`` fetches episode i),
+    mirroring the recording writer. ``table`` must carry a contiguous
+    ``episode_index`` column.
+    """
+    episode_index = table.column("episode_index").to_numpy(zero_copy_only=False)
+    starts = np.concatenate(([0], np.nonzero(np.diff(episode_index))[0] + 1))
+    writer = pq.ParquetWriter(str(path), table.schema, compression="snappy", use_dictionary=True)
+    try:
+        for start, stop in zip(starts, np.append(starts[1:], len(episode_index)), strict=True):
+            writer.write_table(table.slice(start, stop - start))  # one episode -> one row group
+    finally:
+        writer.close()
+
+
 def to_parquet_with_hf_images(
     df: pandas.DataFrame, path: Path, features: datasets.Features | None = None
 ) -> None:
-    """This function correctly writes to parquet a panda DataFrame that contains images encoded by HF dataset.
-    This way, it can be loaded by HF dataset and correctly formatted images are returned.
+    """Write a DataFrame with HF-encoded images to parquet, one row group per episode.
 
-    Args:
-        df: DataFrame to write to parquet.
-        path: Path to write the parquet file.
-        features: Optional HuggingFace Features schema. If provided, ensures image columns
-                  are properly typed as Image() in the parquet schema.
+    Images are embedded into the arrow table first (``ParquetWriter.write_table``
+    does not embed external image files like ``Dataset.to_parquet`` does).
+    ``features`` types image columns as ``Image()`` in the parquet schema.
     """
-    # TODO(qlhoest): replace this weird synthax by `df.to_parquet(path)` only
     ds = datasets.Dataset.from_dict(df.to_dict(orient="list"), features=features)
-    ds.to_parquet(path)
+    ds = embed_images(ds)
+    table = ds.with_format("arrow")[:]
+    if "episode_index" in table.column_names:
+        write_table_one_row_group_per_episode(table, path)
+    else:
+        # No episode boundaries to align row groups to — keep a single write.
+        pq.write_table(table, str(path))
+
+
+def to_parquet_one_row_group_per_episode(df: pandas.DataFrame, path: Path) -> None:
+    """Write a (non-image) DataFrame to parquet with one row group per episode."""
+    table = pa.Table.from_pandas(df, preserve_index=False)
+    if "episode_index" in table.column_names:
+        write_table_one_row_group_per_episode(table, path)
+    else:
+        pq.write_table(table, str(path))
 
 
 def item_to_torch(item: dict) -> dict:
@@ -300,7 +350,11 @@ def item_to_torch(item: dict) -> dict:
     """
     skip_keys = {"task", *LANGUAGE_COLUMNS}
     for key, val in item.items():
-        if isinstance(val, (np.ndarray | list)) and key not in skip_keys:
+        if key in skip_keys:
+            continue
+        if isinstance(val, PILImage.Image):
+            item[key] = pil_to_chw_tensor(val)
+        elif isinstance(val, (np.ndarray | list)):
             # Convert numpy arrays and lists to torch tensors
             item[key] = torch.tensor(val)
     return item

src/lerobot/datasets/lerobot_dataset.py

@@ -24,7 +24,7 @@ import torch.utils
 from huggingface_hub import HfApi, snapshot_download
 from huggingface_hub.errors import RevisionNotFoundError
 
-from lerobot.configs import VideoEncoderConfig
+from lerobot.configs import DEFAULT_DEPTH_UNIT, DepthEncoderConfig, RGBEncoderConfig
 from lerobot.utils.constants import HF_LEROBOT_HUB_CACHE
 
 from .dataset_metadata import CODEBASE_VERSION, LeRobotDatasetMetadata
@@ -58,8 +58,10 @@ class LeRobotDataset(torch.utils.data.Dataset):
         download_videos: bool = True,
         video_backend: str | None = None,
         return_uint8: bool = False,
+        depth_output_unit: str = DEFAULT_DEPTH_UNIT,
         batch_encoding_size: int = 1,
-        camera_encoder: VideoEncoderConfig | None = None,
+        rgb_encoder: RGBEncoderConfig | None = None,
+        depth_encoder: DepthEncoderConfig | None = None,
         encoder_threads: int | None = None,
         streaming_encoding: bool = False,
         encoder_queue_maxsize: int = 30,
@@ -183,8 +185,11 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 You can also use the 'pyav' decoder used by Torchvision, which used to be the default option, or 'video_reader' which is another decoder of Torchvision.
             batch_encoding_size (int, optional): Number of episodes to accumulate before batch encoding videos.
                 Set to 1 for immediate encoding (default), or higher for batched encoding. Defaults to 1.
-            camera_encoder (VideoEncoderConfig | None, optional): Video encoder settings for cameras
-                (codec, quality, etc.). When ``None``, :func:`~lerobot.configs.video.camera_encoder_defaults`
+            rgb_encoder (RGBEncoderConfig | None, optional): Video encoder settings for cameras
+                (codec, quality, etc.). When ``None``, :func:`~lerobot.configs.video.rgb_encoder_defaults`
+                is used by the writer.
+            depth_encoder (DepthEncoderConfig | None, optional): Video encoder settings for depth cameras
+                (codec, quality, etc.). When ``None``, :func:`~lerobot.configs.video.depth_encoder_defaults`
                 is used by the writer.
             encoder_threads (int | None, optional): Number of encoder threads (global). ``None`` lets the
                 codec decide.
@@ -201,13 +206,12 @@ class LeRobotDataset(torch.utils.data.Dataset):
         super().__init__()
         self.repo_id = repo_id
         self._requested_root = Path(root) if root else None
-        self.reader = None
-        self.set_image_transforms(image_transforms)
         self.delta_timestamps = delta_timestamps
         self.tolerance_s = tolerance_s
         self.revision = revision if revision else CODEBASE_VERSION
         self._video_backend = video_backend if video_backend else get_safe_default_video_backend()
         self._return_uint8 = return_uint8
+        self._depth_output_unit = depth_output_unit
         self._batch_encoding_size = batch_encoding_size
         self._encoder_threads = encoder_threads
 
@@ -220,6 +224,7 @@ 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
@@ -248,7 +253,9 @@ class LeRobotDataset(torch.utils.data.Dataset):
             delta_timestamps=delta_timestamps,
             image_transforms=image_transforms,
             return_uint8=self._return_uint8,
+            depth_output_unit=self._depth_output_unit,
         )
+        self.image_transforms = image_transforms
 
         # Load actual data
         if force_cache_sync or not self.reader.try_load():
@@ -272,14 +279,16 @@ class LeRobotDataset(torch.utils.data.Dataset):
             if streaming_encoding and len(self.meta.video_keys) > 0:
                 streaming_enc = self._build_streaming_encoder(
                     self.meta.fps,
-                    camera_encoder,
+                    rgb_encoder,
+                    depth_encoder,
                     encoder_queue_maxsize,
                     encoder_threads,
                 )
             self.writer = DatasetWriter(
                 meta=self.meta,
                 root=self.root,
-                camera_encoder=camera_encoder,
+                rgb_encoder=rgb_encoder,
+                depth_encoder=depth_encoder,
                 encoder_threads=encoder_threads,
                 batch_encoding_size=batch_encoding_size,
                 streaming_encoder=streaming_enc,
@@ -315,19 +324,22 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 delta_timestamps=self.delta_timestamps,
                 image_transforms=self.image_transforms,
                 return_uint8=self._return_uint8,
+                depth_output_unit=self._depth_output_unit,
             )
         return self.reader
 
     @staticmethod
     def _build_streaming_encoder(
         fps: int,
-        camera_encoder: VideoEncoderConfig | None,
+        rgb_encoder: RGBEncoderConfig | None,
+        depth_encoder: DepthEncoderConfig | None,
         encoder_queue_maxsize: int,
         encoder_threads: int | None,
     ) -> StreamingVideoEncoder:
         return StreamingVideoEncoder(
             fps=fps,
-            camera_encoder=camera_encoder,
+            rgb_encoder=rgb_encoder,
+            depth_encoder=depth_encoder,
             queue_maxsize=encoder_queue_maxsize,
             encoder_threads=encoder_threads,
         )
@@ -339,6 +351,11 @@ 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."""
@@ -370,6 +387,18 @@ class LeRobotDataset(torch.utils.data.Dataset):
             self.reader.load_and_activate()
         return self.reader.hf_dataset
 
+    @property
+    def absolute_to_relative_idx(self) -> dict[int, int] | None:
+        """Mapping from absolute frame indices to HF dataset row positions.
+
+        Non-None only for episode-filtered datasets where absolute indices
+        (from metadata) differ from row positions in the loaded HF dataset.
+        """
+        reader = self._ensure_reader()
+        if reader.hf_dataset is None:
+            reader.load_and_activate()
+        return reader._absolute_to_relative_idx
+
     # ── Writer-delegated methods ──────────────────────────────────────
 
     def add_frame(self, frame: dict) -> None:
@@ -505,15 +534,14 @@ class LeRobotDataset(torch.utils.data.Dataset):
 
     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):
-            raise TypeError("image_transforms must be callable or None.")
+        self._ensure_reader().set_image_transforms(image_transforms)
         self.image_transforms = image_transforms
-        if self.reader is not None:
-            self.reader._image_transforms = image_transforms
 
     def clear_image_transforms(self) -> None:
         """Remove the transform applied to visual observations."""
-        self.set_image_transforms(None)
+        if self.reader is not None:
+            self.reader.set_image_transforms(None)
+        self.image_transforms = None
 
     # ── Hub methods (stay on facade) ──────────────────────────────────
 
@@ -645,7 +673,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
         image_writer_threads: int = 0,
         video_backend: str | None = None,
         batch_encoding_size: int = 1,
-        camera_encoder: VideoEncoderConfig | None = None,
+        rgb_encoder: RGBEncoderConfig | None = None,
+        depth_encoder: DepthEncoderConfig | None = None,
         metadata_buffer_size: int = 10,
         streaming_encoding: bool = False,
         encoder_queue_maxsize: int = 30,
@@ -676,8 +705,10 @@ class LeRobotDataset(torch.utils.data.Dataset):
             video_backend: Video decoding backend (used when reading back).
             batch_encoding_size: Number of episodes to accumulate before
                 batch-encoding videos. ``1`` means encode immediately.
-            camera_encoder: Video encoder settings for cameras (codec, quality, etc.).
-                When ``None``, :func:`~lerobot.configs.video.camera_encoder_defaults` is used.
+            rgb_encoder: Video encoder settings for cameras (codec, quality, etc.).
+                When ``None``, :func:`~lerobot.configs.video.rgb_encoder_defaults` is used.
+            depth_encoder: Video encoder settings for depth cameras (codec, quality, etc.).
+                When ``None``, :func:`~lerobot.configs.video.depth_encoder_defaults` is used.
             encoder_threads: Number of encoder threads (global). ``None``
                 lets the codec decide.
             metadata_buffer_size: Number of episode metadata records to buffer
@@ -712,6 +743,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj.episodes = None
         obj._video_backend = video_backend if video_backend is not None else get_safe_default_video_backend()
         obj._return_uint8 = False
+        obj._depth_output_unit = DEFAULT_DEPTH_UNIT
         obj._batch_encoding_size = batch_encoding_size
         obj._encoder_threads = encoder_threads
 
@@ -721,12 +753,13 @@ class LeRobotDataset(torch.utils.data.Dataset):
         streaming_enc = None
         if streaming_encoding and len(obj.meta.video_keys) > 0:
             streaming_enc = cls._build_streaming_encoder(
-                fps, camera_encoder, encoder_queue_maxsize, encoder_threads
+                fps, rgb_encoder, depth_encoder, encoder_queue_maxsize, encoder_threads
             )
         obj.writer = DatasetWriter(
             meta=obj.meta,
             root=obj.root,
-            camera_encoder=camera_encoder,
+            rgb_encoder=rgb_encoder,
+            depth_encoder=depth_encoder,
             encoder_threads=encoder_threads,
             batch_encoding_size=batch_encoding_size,
             streaming_encoder=streaming_enc,
@@ -749,7 +782,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
         force_cache_sync: bool = False,
         video_backend: str | None = None,
         batch_encoding_size: int = 1,
-        camera_encoder: VideoEncoderConfig | None = None,
+        rgb_encoder: RGBEncoderConfig | None = None,
+        depth_encoder: DepthEncoderConfig | None = None,
         encoder_threads: int | None = None,
         image_writer_processes: int = 0,
         image_writer_threads: int = 0,
@@ -777,8 +811,10 @@ class LeRobotDataset(torch.utils.data.Dataset):
             video_backend: Video decoding backend for reading back data.
             batch_encoding_size: Number of episodes to accumulate before
                 batch-encoding videos.
-            camera_encoder: Video encoder settings for cameras (codec, quality, etc.).
-                When ``None``, :func:`~lerobot.configs.video.camera_encoder_defaults` is used.
+            rgb_encoder: Video encoder settings for cameras (codec, quality, etc.).
+                When ``None``, :func:`~lerobot.configs.video.rgb_encoder_defaults` is used.
+            depth_encoder: Video encoder settings for depth cameras (codec, quality, etc.).
+                When ``None``, :func:`~lerobot.configs.video.depth_encoder_defaults` is used.
             encoder_threads: Number of encoder threads (global). ``None``
                 lets the codec decide.
             image_writer_processes: Subprocesses for async image writing.
@@ -806,6 +842,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj.episodes = None
         obj._video_backend = video_backend if video_backend else get_safe_default_video_backend()
         obj._return_uint8 = False
+        obj._depth_output_unit = DEFAULT_DEPTH_UNIT
         obj._batch_encoding_size = batch_encoding_size
 
         if obj._requested_root is not None:
@@ -825,12 +862,13 @@ class LeRobotDataset(torch.utils.data.Dataset):
         streaming_enc = None
         if streaming_encoding and len(obj.meta.video_keys) > 0:
             streaming_enc = cls._build_streaming_encoder(
-                obj.meta.fps, camera_encoder, encoder_queue_maxsize, encoder_threads
+                obj.meta.fps, rgb_encoder, depth_encoder, encoder_queue_maxsize, encoder_threads
             )
         obj.writer = DatasetWriter(
             meta=obj.meta,
             root=obj.root,
-            camera_encoder=camera_encoder,
+            rgb_encoder=rgb_encoder,
+            depth_encoder=depth_encoder,
             encoder_threads=encoder_threads,
             batch_encoding_size=batch_encoding_size,
             streaming_encoder=streaming_enc,

src/lerobot/datasets/mp4.py

@@ -1,666 +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
-
-from __future__ import annotations
-
-import struct
-from collections.abc import Callable, Iterable
-from dataclasses import dataclass
-
-import numpy as np
-
-
-@dataclass(frozen=True)
-class Box:
-    type: bytes
-    start: int
-    header_size: int
-    end: int
-
-    @property
-    def payload_start(self) -> int:
-        return self.start + self.header_size
-
-    @property
-    def size(self) -> int:
-        return self.end - self.start
-
-
-@dataclass(frozen=True)
-class Mp4SampleSlice:
-    sample_lo: int
-    sample_hi: int
-    byte_offset: int
-    byte_length: int
-    source_start_pts: float
-
-
-@dataclass(frozen=True)
-class Mp4Index:
-    file_path: str
-    file_size: int
-    ftyp: bytes
-    moov_offset: int
-    mdat_offset: int
-    mdat_payload_offset: int
-    mdat_payload_size: int
-    faststart: bool
-    codec: str
-    timescale: int
-    duration: int
-    track_id: int
-    width: int
-    height: int
-    stsd_body: bytes
-    sample_pts: np.ndarray
-    sample_durations: np.ndarray
-    sample_sizes: np.ndarray
-    sample_offsets: np.ndarray
-    sync_samples: np.ndarray
-
-    def sample_slice(
-        self,
-        from_ts: float,
-        to_ts: float,
-        *,
-        keyframe_pad_s: float = 0.1,
-        keyframe_pad_fraction: float = 0.05,
-        file_size: int | None = None,
-    ) -> Mp4SampleSlice:
-        if to_ts < from_ts:
-            raise ValueError(f"Invalid timestamp span: {from_ts=} {to_ts=}")
-        if len(self.sample_pts) == 0:
-            raise ValueError(f"{self.file_path} contains no indexed samples")
-
-        pad = max(keyframe_pad_s, (to_ts - from_ts) * keyframe_pad_fraction)
-        lo_ts = max(0.0, from_ts - pad)
-        hi_ts = to_ts + pad
-        lo = int(np.searchsorted(self.sample_pts, lo_ts, side="left"))
-        hi = int(np.searchsorted(self.sample_pts, hi_ts, side="right")) - 1
-        lo = min(max(lo, 0), len(self.sample_pts) - 1)
-        hi = min(max(hi, lo), len(self.sample_pts) - 1)
-
-        if len(self.sync_samples):
-            prev_sync = self.sync_samples[self.sync_samples <= lo]
-            if len(prev_sync):
-                lo = int(prev_sync[-1])
-            else:
-                lo = int(self.sync_samples[0])
-                if lo > hi:
-                    hi = lo
-
-        offsets = self.sample_offsets[lo : hi + 1]
-        sizes = self.sample_sizes[lo : hi + 1]
-        slice_lo = int(offsets.min())
-        slice_hi = int((offsets + sizes).max())
-        if file_size is not None:
-            slice_hi = min(slice_hi, int(file_size))
-        return Mp4SampleSlice(
-            sample_lo=lo,
-            sample_hi=hi,
-            byte_offset=slice_lo,
-            byte_length=slice_hi - slice_lo,
-            source_start_pts=float(self.sample_pts[lo]),
-        )
-
-    def to_dict(self) -> dict:
-        return {
-            "file_path": self.file_path,
-            "file_size": self.file_size,
-            "ftyp": self.ftyp.hex(),
-            "moov_offset": self.moov_offset,
-            "mdat_offset": self.mdat_offset,
-            "mdat_payload_offset": self.mdat_payload_offset,
-            "mdat_payload_size": self.mdat_payload_size,
-            "faststart": self.faststart,
-            "codec": self.codec,
-            "timescale": self.timescale,
-            "duration": self.duration,
-            "track_id": self.track_id,
-            "width": self.width,
-            "height": self.height,
-            "stsd_body": self.stsd_body.hex(),
-        }
-
-    @classmethod
-    def from_dict(cls, data: dict, arrays: dict[str, np.ndarray]) -> Mp4Index:
-        return cls(
-            file_path=data["file_path"],
-            file_size=int(data["file_size"]),
-            ftyp=bytes.fromhex(data["ftyp"]),
-            moov_offset=int(data["moov_offset"]),
-            mdat_offset=int(data["mdat_offset"]),
-            mdat_payload_offset=int(data["mdat_payload_offset"]),
-            mdat_payload_size=int(data["mdat_payload_size"]),
-            faststart=bool(data["faststart"]),
-            codec=data["codec"],
-            timescale=int(data["timescale"]),
-            duration=int(data["duration"]),
-            track_id=int(data["track_id"]),
-            width=int(data["width"]),
-            height=int(data["height"]),
-            stsd_body=bytes.fromhex(data["stsd_body"]),
-            sample_pts=arrays["sample_pts"],
-            sample_durations=arrays["sample_durations"],
-            sample_sizes=arrays["sample_sizes"],
-            sample_offsets=arrays["sample_offsets"],
-            sync_samples=arrays["sync_samples"],
-        )
-
-
-def fetch_mp4_index(
-    path: str,
-    read_range: Callable[[str, int, int], bytes],
-    *,
-    file_size: int,
-    header_probe_bytes: int = 4 * 1024 * 1024,
-    max_probe_bytes: int = 64 * 1024 * 1024,
-) -> Mp4Index:
-    probe_size = min(header_probe_bytes, file_size)
-    while True:
-        data = read_range(path, 0, probe_size)
-        top = list(iter_boxes(data, 0, len(data), absolute_base=0, allow_truncated=True))
-        has_mdat = any(box.type == b"mdat" for box in top)
-        has_moov = any(box.type == b"moov" and box.end <= len(data) for box in top)
-        if has_mdat and has_moov:
-            return parse_mp4_index(path, data, file_size=file_size)
-        if probe_size >= min(max_probe_bytes, file_size):
-            if has_mdat and not has_moov:
-                tail_index = _fetch_tail_moov_index(path, read_range, data, top, file_size, max_probe_bytes)
-                if tail_index is not None:
-                    return tail_index
-            missing = []
-            if not has_mdat:
-                missing.append("mdat")
-            if not has_moov:
-                missing.append("moov")
-            raise ValueError(
-                f"Could not find complete {'/'.join(missing)} in first {probe_size} bytes of {path}"
-            )
-        probe_size = min(probe_size * 2, max_probe_bytes, file_size)
-
-
-def _fetch_tail_moov_index(
-    path: str,
-    read_range: Callable[[str, int, int], bytes],
-    prefix: bytes,
-    top_boxes: list[Box],
-    file_size: int,
-    max_probe_bytes: int,
-) -> Mp4Index | None:
-    mdat_box = _one(top_boxes, b"mdat")
-    if mdat_box is None or mdat_box.end >= file_size:
-        return None
-    tail_offset = mdat_box.end
-    tail_length = min(max_probe_bytes, file_size - tail_offset)
-    tail = read_range(path, tail_offset, tail_length)
-    tail_boxes = list(iter_boxes(tail, 0, len(tail), absolute_base=tail_offset, allow_truncated=True))
-    moov_box = next(
-        (box for box in tail_boxes if box.type == b"moov" and box.end <= tail_offset + len(tail)), None
-    )
-    if moov_box is None:
-        return None
-    ftyp_box = _one(top_boxes, b"ftyp", required=False)
-    ftyp = (
-        prefix[ftyp_box.start : ftyp_box.end]
-        if ftyp_box is not None
-        else _box(b"ftyp", b"isom\0\0\2\0isomiso2mp41")
-    )
-    moov_start = moov_box.payload_start - tail_offset
-    moov_end = moov_box.end - tail_offset
-    return _parse_mp4_index_from_layout(
-        path,
-        file_size=file_size,
-        ftyp=ftyp,
-        moov_offset=moov_box.start,
-        moov=tail[moov_start:moov_end],
-        mdat_box=mdat_box,
-    )
-
-
-def parse_mp4_index(path: str, data: bytes, *, file_size: int | None = None) -> Mp4Index:
-    if file_size is None:
-        file_size = len(data)
-    top = list(iter_boxes(data, 0, len(data), absolute_base=0, allow_truncated=True))
-    ftyp_box = _one(top, b"ftyp", required=False)
-    moov_box = _one(top, b"moov")
-    mdat_box = _one(top, b"mdat")
-    if moov_box.end > len(data):
-        raise ValueError(f"{path}: moov box is truncated")
-
-    moov = data[moov_box.payload_start : moov_box.end]
-    ftyp = (
-        data[ftyp_box.start : ftyp_box.end]
-        if ftyp_box is not None
-        else _box(b"ftyp", b"isom\0\0\2\0isomiso2mp41")
-    )
-    return _parse_mp4_index_from_layout(
-        path,
-        file_size=file_size,
-        ftyp=ftyp,
-        moov_offset=moov_box.start,
-        moov=moov,
-        mdat_box=mdat_box,
-    )
-
-
-def _parse_mp4_index_from_layout(
-    path: str,
-    *,
-    file_size: int,
-    ftyp: bytes,
-    moov_offset: int,
-    moov: bytes,
-    mdat_box: Box,
-) -> Mp4Index:
-    mvhd_timescale, mvhd_duration = _parse_mvhd(_find_descendant(moov, [b"mvhd"]))
-    trak_box, trak_payload = _find_video_trak(moov)
-    _ = trak_box
-    tkhd = _parse_tkhd(_find_descendant(trak_payload, [b"tkhd"]))
-    mdhd_timescale, mdhd_duration = _parse_mdhd(_find_descendant(trak_payload, [b"mdia", b"mdhd"]))
-    stbl = _find_descendant(trak_payload, [b"mdia", b"minf", b"stbl"])
-
-    stsd = _find_child(stbl, b"stsd")
-    stsd_body = stbl[stsd.payload_start : stsd.end]
-    codec = _parse_stsd_codec(stsd_body)
-    stts = _parse_stts(_payload(stbl, b"stts"))
-    sample_sizes = _parse_stsz(_payload(stbl, b"stsz"))
-    stsc = _parse_stsc(_payload(stbl, b"stsc"))
-    chunk_offsets = _parse_chunk_offsets(stbl)
-    sync_samples = _parse_stss(stbl, len(sample_sizes))
-
-    sample_durations = _expand_stts(stts, len(sample_sizes))
-    sample_pts_units = np.empty(len(sample_durations), dtype=np.int64)
-    if len(sample_durations):
-        sample_pts_units[0] = 0
-        if len(sample_durations) > 1:
-            sample_pts_units[1:] = np.cumsum(sample_durations[:-1], dtype=np.int64)
-    sample_pts = sample_pts_units.astype(np.float64) / float(mdhd_timescale)
-    sample_offsets = _sample_offsets(stsc, chunk_offsets, sample_sizes)
-
-    return Mp4Index(
-        file_path=path,
-        file_size=file_size,
-        ftyp=ftyp,
-        moov_offset=moov_offset,
-        mdat_offset=mdat_box.start,
-        mdat_payload_offset=mdat_box.payload_start,
-        mdat_payload_size=mdat_box.end - mdat_box.payload_start
-        if mdat_box.end <= file_size
-        else file_size - mdat_box.payload_start,
-        faststart=moov_offset < mdat_box.start,
-        codec=codec,
-        timescale=mdhd_timescale,
-        duration=mdhd_duration or mvhd_duration,
-        track_id=tkhd["track_id"],
-        width=tkhd["width"],
-        height=tkhd["height"],
-        stsd_body=stsd_body,
-        sample_pts=sample_pts,
-        sample_durations=sample_durations,
-        sample_sizes=sample_sizes,
-        sample_offsets=sample_offsets,
-        sync_samples=sync_samples,
-    )
-
-
-def synthesize_mp4(index: Mp4Index, sample_slice: Mp4SampleSlice, mdat_payload: bytes) -> bytes:
-    lo = sample_slice.sample_lo
-    hi = sample_slice.sample_hi + 1
-    if lo < 0 or hi > len(index.sample_sizes) or lo >= hi:
-        raise ValueError(f"Invalid sample range [{lo}, {hi}) for {index.file_path}")
-
-    offsets = index.sample_offsets[lo:hi]
-    sizes = index.sample_sizes[lo:hi]
-    rel_offsets = offsets - sample_slice.byte_offset
-    if int(rel_offsets.min()) != 0:
-        raise ValueError("Sample slice must start at the minimum referenced sample offset")
-    if int((rel_offsets + sizes).max()) > len(mdat_payload):
-        raise ValueError("Sample slice does not cover all referenced samples")
-
-    durations = index.sample_durations[lo:hi]
-    sync = index.sync_samples[(index.sync_samples >= lo) & (index.sync_samples < hi)] - lo + 1
-    moov = _make_moov(index, durations, sizes, rel_offsets, sync, mdat_data_offset=0)
-    header_size = len(index.ftyp) + len(moov)
-    moov = _make_moov(index, durations, sizes, rel_offsets, sync, mdat_data_offset=header_size + 8)
-    return index.ftyp + moov + _box(b"mdat", mdat_payload)
-
-
-def iter_boxes(
-    data: bytes,
-    start: int,
-    end: int,
-    *,
-    absolute_base: int = 0,
-    allow_truncated: bool = False,
-) -> Iterable[Box]:
-    pos = start
-    while pos + 8 <= end:
-        size = struct.unpack_from(">I", data, pos)[0]
-        typ = data[pos + 4 : pos + 8]
-        header_size = 8
-        if size == 1:
-            if pos + 16 > end:
-                break
-            size = struct.unpack_from(">Q", data, pos + 8)[0]
-            header_size = 16
-        elif size == 0:
-            size = end - pos
-        if size < header_size:
-            break
-        box_end = pos + size
-        if box_end > end and not allow_truncated:
-            break
-        yield Box(typ, absolute_base + pos, header_size, absolute_base + box_end)
-        pos = box_end
-
-
-def _find_video_trak(moov: bytes) -> tuple[Box, bytes]:
-    for trak in _children(moov, 0, len(moov)):
-        if trak.type != b"trak":
-            continue
-        payload = moov[trak.payload_start : trak.end]
-        hdlr = _find_descendant(payload, [b"mdia", b"hdlr"])
-        if hdlr[8:12] == b"vide":
-            return trak, payload
-    raise ValueError("No video track found")
-
-
-def _find_descendant(data: bytes, path: list[bytes]) -> bytes:
-    current = data
-    for typ in path:
-        box = _find_child(current, typ)
-        current = current[box.payload_start : box.end]
-    return current
-
-
-def _find_child(data: bytes, typ: bytes) -> Box:
-    for box in _children(data, 0, len(data)):
-        if box.type == typ:
-            return box
-    raise ValueError(f"Missing MP4 box {typ.decode('latin1')}")
-
-
-def _children(data: bytes, start: int, end: int) -> Iterable[Box]:
-    return iter_boxes(data, start, end, absolute_base=0)
-
-
-def _one(boxes: list[Box], typ: bytes, *, required: bool = True) -> Box | None:
-    matches = [box for box in boxes if box.type == typ]
-    if not matches and required:
-        raise ValueError(f"Missing MP4 box {typ.decode('latin1')}")
-    return matches[0] if matches else None
-
-
-def _payload(parent: bytes, typ: bytes) -> bytes:
-    box = _find_child(parent, typ)
-    return parent[box.payload_start : box.end]
-
-
-def _parse_mvhd(payload: bytes) -> tuple[int, int]:
-    version = payload[0]
-    if version == 1:
-        return struct.unpack_from(">IQ", payload, 20)
-    return struct.unpack_from(">II", payload, 12)
-
-
-def _parse_mdhd(payload: bytes) -> tuple[int, int]:
-    version = payload[0]
-    if version == 1:
-        return struct.unpack_from(">IQ", payload, 20)
-    return struct.unpack_from(">II", payload, 12)
-
-
-def _parse_tkhd(payload: bytes) -> dict[str, int]:
-    version = payload[0]
-    if version == 1:
-        track_id = struct.unpack_from(">I", payload, 20)[0]
-        duration = struct.unpack_from(">Q", payload, 28)[0]
-        width, height = struct.unpack_from(">II", payload, 88)
-    else:
-        track_id = struct.unpack_from(">I", payload, 12)[0]
-        duration = struct.unpack_from(">I", payload, 20)[0]
-        width, height = struct.unpack_from(">II", payload, 76)
-    return {"track_id": track_id, "duration": duration, "width": width >> 16, "height": height >> 16}
-
-
-def _parse_stsd_codec(stsd_body: bytes) -> str:
-    if len(stsd_body) < 16:
-        return "unknown"
-    return stsd_body[12:16].decode("latin1")
-
-
-def _parse_stts(payload: bytes) -> list[tuple[int, int]]:
-    count = struct.unpack_from(">I", payload, 4)[0]
-    out = []
-    offset = 8
-    for _ in range(count):
-        out.append(struct.unpack_from(">II", payload, offset))
-        offset += 8
-    return out
-
-
-def _expand_stts(entries: list[tuple[int, int]], sample_count: int) -> np.ndarray:
-    values = np.empty(sample_count, dtype=np.int64)
-    pos = 0
-    for count, delta in entries:
-        values[pos : pos + count] = delta
-        pos += count
-    if pos != sample_count:
-        raise ValueError(f"stts describes {pos} samples, stsz describes {sample_count}")
-    return values
-
-
-def _parse_stsz(payload: bytes) -> np.ndarray:
-    sample_size, sample_count = struct.unpack_from(">II", payload, 4)
-    if sample_size:
-        return np.full(sample_count, sample_size, dtype=np.int64)
-    offset = 12
-    values = np.empty(sample_count, dtype=np.int64)
-    for idx in range(sample_count):
-        values[idx] = struct.unpack_from(">I", payload, offset)[0]
-        offset += 4
-    return values
-
-
-def _parse_stsc(payload: bytes) -> list[tuple[int, int, int]]:
-    count = struct.unpack_from(">I", payload, 4)[0]
-    out = []
-    offset = 8
-    for _ in range(count):
-        out.append(struct.unpack_from(">III", payload, offset))
-        offset += 12
-    return out
-
-
-def _parse_chunk_offsets(stbl: bytes) -> np.ndarray:
-    with_stco = None
-    with_co64 = None
-    for box in _children(stbl, 0, len(stbl)):
-        if box.type == b"stco":
-            with_stco = stbl[box.payload_start : box.end]
-        elif box.type == b"co64":
-            with_co64 = stbl[box.payload_start : box.end]
-    if with_co64 is not None:
-        count = struct.unpack_from(">I", with_co64, 4)[0]
-        return np.array(
-            [struct.unpack_from(">Q", with_co64, 8 + idx * 8)[0] for idx in range(count)], dtype=np.int64
-        )
-    if with_stco is None:
-        raise ValueError("Missing stco/co64 chunk offsets")
-    count = struct.unpack_from(">I", with_stco, 4)[0]
-    return np.array(
-        [struct.unpack_from(">I", with_stco, 8 + idx * 4)[0] for idx in range(count)], dtype=np.int64
-    )
-
-
-def _parse_stss(stbl: bytes, sample_count: int) -> np.ndarray:
-    for box in _children(stbl, 0, len(stbl)):
-        if box.type == b"stss":
-            payload = stbl[box.payload_start : box.end]
-            count = struct.unpack_from(">I", payload, 4)[0]
-            return np.array(
-                [struct.unpack_from(">I", payload, 8 + idx * 4)[0] - 1 for idx in range(count)],
-                dtype=np.int64,
-            )
-    return np.arange(sample_count, dtype=np.int64)
-
-
-def _sample_offsets(
-    stsc: list[tuple[int, int, int]], chunk_offsets: np.ndarray, sample_sizes: np.ndarray
-) -> np.ndarray:
-    if not stsc:
-        raise ValueError("stsc is empty")
-    offsets = np.empty(len(sample_sizes), dtype=np.int64)
-    sample_idx = 0
-    for entry_idx, (first_chunk, samples_per_chunk, _desc_idx) in enumerate(stsc):
-        next_first = stsc[entry_idx + 1][0] if entry_idx + 1 < len(stsc) else len(chunk_offsets) + 1
-        for chunk_number in range(first_chunk, next_first):
-            if chunk_number < 1 or chunk_number > len(chunk_offsets):
-                raise ValueError("stsc references a chunk outside stco/co64")
-            chunk_pos = int(chunk_offsets[chunk_number - 1])
-            for _ in range(samples_per_chunk):
-                if sample_idx >= len(sample_sizes):
-                    return offsets
-                offsets[sample_idx] = chunk_pos
-                chunk_pos += int(sample_sizes[sample_idx])
-                sample_idx += 1
-    if sample_idx != len(sample_sizes):
-        raise ValueError(f"stsc describes {sample_idx} samples, stsz describes {len(sample_sizes)}")
-    return offsets
-
-
-def _make_moov(
-    index: Mp4Index,
-    durations: np.ndarray,
-    sizes: np.ndarray,
-    rel_offsets: np.ndarray,
-    sync_samples: np.ndarray,
-    *,
-    mdat_data_offset: int,
-) -> bytes:
-    duration = int(durations.sum())
-    stco_values = [int(mdat_data_offset + value) for value in rel_offsets]
-    if any(value > 0xFFFFFFFF for value in stco_values):
-        offset_box = _co64(stco_values)
-    else:
-        offset_box = _stco(stco_values)
-    stbl = _box(
-        b"stbl",
-        _box(b"stsd", index.stsd_body)
-        + _stts(durations)
-        + _stsc_one_sample_per_chunk(len(sizes))
-        + _stsz(sizes)
-        + offset_box
-        + (_stss(sync_samples) if len(sync_samples) else b""),
-    )
-    minf = _box(b"minf", _vmhd() + _dinf() + stbl)
-    mdia = _box(b"mdia", _mdhd(index.timescale, duration) + _hdlr() + minf)
-    trak = _box(b"trak", _tkhd(index.track_id, duration, index.width, index.height) + mdia)
-    return _box(b"moov", _mvhd(index.timescale, duration, index.track_id + 1) + trak)
-
-
-def _full_box(typ: bytes, version: int, flags: int, payload: bytes = b"") -> bytes:
-    return _box(typ, bytes([version]) + flags.to_bytes(3, "big") + payload)
-
-
-def _box(typ: bytes, payload: bytes) -> bytes:
-    size = len(payload) + 8
-    if size <= 0xFFFFFFFF:
-        return struct.pack(">I4s", size, typ) + payload
-    return struct.pack(">I4sQ", 1, typ, size + 8) + payload
-
-
-def _mvhd(timescale: int, duration: int, next_track_id: int) -> bytes:
-    matrix = struct.pack(">9I", 0x00010000, 0, 0, 0, 0x00010000, 0, 0, 0, 0x40000000)
-    payload = (
-        struct.pack(">IIII", 0, 0, timescale, duration)
-        + struct.pack(">IHH", 0x00010000, 0x0100, 0)
-        + b"\0" * 8
-        + matrix
-        + b"\0" * 24
-        + struct.pack(">I", next_track_id)
-    )
-    return _full_box(b"mvhd", 0, 0, payload)
-
-
-def _tkhd(track_id: int, duration: int, width: int, height: int) -> bytes:
-    matrix = struct.pack(">9I", 0x00010000, 0, 0, 0, 0x00010000, 0, 0, 0, 0x40000000)
-    payload = (
-        struct.pack(">IIIII", 0, 0, track_id, 0, duration)
-        + b"\0" * 8
-        + struct.pack(">hhhh", 0, 0, 0, 0)
-        + matrix
-        + struct.pack(">II", width << 16, height << 16)
-    )
-    return _full_box(b"tkhd", 0, 7, payload)
-
-
-def _mdhd(timescale: int, duration: int) -> bytes:
-    return _full_box(b"mdhd", 0, 0, struct.pack(">IIIIH", 0, 0, timescale, duration, 0x55C4) + b"\0\0")
-
-
-def _hdlr() -> bytes:
-    return _full_box(b"hdlr", 0, 0, b"\0" * 4 + b"vide" + b"\0" * 12 + b"VideoHandler\0")
-
-
-def _vmhd() -> bytes:
-    return _full_box(b"vmhd", 0, 1, struct.pack(">HHHH", 0, 0, 0, 0))
-
-
-def _dinf() -> bytes:
-    url = _full_box(b"url ", 0, 1)
-    dref = _full_box(b"dref", 0, 0, struct.pack(">I", 1) + url)
-    return _box(b"dinf", dref)
-
-
-def _stts(durations: np.ndarray) -> bytes:
-    runs = []
-    for duration in durations.tolist():
-        if runs and runs[-1][1] == int(duration):
-            runs[-1][0] += 1
-        else:
-            runs.append([1, int(duration)])
-    payload = struct.pack(">I", len(runs)) + b"".join(
-        struct.pack(">II", count, delta) for count, delta in runs
-    )
-    return _full_box(b"stts", 0, 0, payload)
-
-
-def _stsc_one_sample_per_chunk(sample_count: int) -> bytes:
-    return _full_box(b"stsc", 0, 0, struct.pack(">IIII", 1, 1, 1, 1))
-
-
-def _stsz(sizes: np.ndarray) -> bytes:
-    return _full_box(
-        b"stsz",
-        0,
-        0,
-        struct.pack(">II", 0, len(sizes)) + b"".join(struct.pack(">I", int(size)) for size in sizes.tolist()),
-    )
-
-
-def _stco(values: list[int]) -> bytes:
-    return _full_box(
-        b"stco", 0, 0, struct.pack(">I", len(values)) + b"".join(struct.pack(">I", v) for v in values)
-    )
-
-
-def _co64(values: list[int]) -> bytes:
-    return _full_box(
-        b"co64", 0, 0, struct.pack(">I", len(values)) + b"".join(struct.pack(">Q", v) for v in values)
-    )
-
-
-def _stss(values: np.ndarray) -> bytes:
-    return _full_box(
-        b"stss",
-        0,
-        0,
-        struct.pack(">I", len(values)) + b"".join(struct.pack(">I", int(value)) for value in values.tolist()),
-    )

src/lerobot/datasets/pyav_utils.py

@@ -24,6 +24,7 @@ import logging
 from typing import Any
 
 import av
+import numpy as np
 
 logger = logging.getLogger(__name__)
 
@@ -31,6 +32,34 @@ FFMPEG_NUMERIC_OPTION_TYPES = ("INT", "INT64", "UINT64", "FLOAT", "DOUBLE")
 FFMPEG_INTEGER_OPTION_TYPES = ("INT", "INT64", "UINT64")
 
 
+def write_u16_plane(plane: av.video.plane.VideoPlane, src: np.ndarray, fill_value: int | None = None) -> None:
+    """Copy a 2D ``uint16`` image into the plane's memory buffer, row by row.
+
+    For speed, each row is padded to a wider size than ``width``, so the true row width in
+    memory is ``plane.line_size`` (bytes), not ``width``. Copying as one straight stream
+    would skew the image, so we write only the first ``width`` columns of each row and
+    leave the padding untouched.
+
+    Args:
+        plane: Destination 16-bit plane.
+        src: Source image, shape ``(height, width)``, dtype ``uint16``.
+        fill_value: If given, every pixel (padding included) is set to this first, so the
+            padding holds clean data instead of garbage.
+    """
+    height, width = src.shape
+    stride_u16 = plane.line_size // np.dtype(np.uint16).itemsize
+    dst = np.frombuffer(plane, dtype=np.uint16).reshape(height, stride_u16)
+    if fill_value is not None:
+        dst.fill(fill_value)
+    dst[:, :width] = src
+
+
+@functools.cache
+def get_pix_fmt_channels(pix_fmt: str) -> int:
+    """Return the number of components (channels) for *pix_fmt*."""
+    return len(av.VideoFormat(pix_fmt).components)
+
+
 @functools.cache
 def get_codec(vcodec: str) -> av.codec.Codec | None:
     """PyAV write-mode ``Codec`` for *vcodec*, or ``None`` if unavailable."""
@@ -92,7 +121,7 @@ def _check_option_value(vcodec: str, label: str, value: Any, opt: av.option.Opti
                     f"{label}={value!r} is not numeric; codec {vcodec!r} expects a number for this option."
                 ) from e
         elif isinstance(value, (float, int)):
-            num_val = value
+            num_val = float(value)
         else:
             raise ValueError(
                 f"{label}={value!r} is not numeric; codec {vcodec!r} expects a number for this option."
@@ -142,6 +171,16 @@ def _check_pixel_format(vcodec: str, pix_fmt: str) -> None:
         )
 
 
+def _check_pix_fmt_channels(pix_fmt: str, channels: int) -> None:
+    """Ensure *pix_fmt* can carry at least *channels* components."""
+    pix_fmt_channels = get_pix_fmt_channels(pix_fmt)
+    if pix_fmt_channels < channels:
+        raise ValueError(
+            f"pix_fmt={pix_fmt!r} carries only {pix_fmt_channels} component(s) "
+            f"but the source data has {channels} channel(s)."
+        )
+
+
 def _check_codec_options(vcodec: str, codec_options: dict[str, Any]) -> None:
     """Validate merged encoder options (typed) against the codec's published AVOptions."""
     supported_options = _get_codec_options_by_name(vcodec)
@@ -156,12 +195,18 @@ def _check_codec_options(vcodec: str, codec_options: dict[str, Any]) -> None:
         _check_option_value(vcodec, key, value, supported_options[key])
 
 
-def check_video_encoder_parameters_pyav(vcodec: str, pix_fmt: str, codec_options: dict[str, Any]) -> None:
+def check_video_encoder_parameters_pyav(
+    vcodec: str,
+    pix_fmt: str,
+    codec_options: dict[str, Any],
+    channels: int | None = None,
+) -> None:
     """Verify *config* is compatible with the bundled FFmpeg build.
 
     Checks pixel format, abstract tuning-field compatibility, and each merged
     encoder option from :meth:`~lerobot.configs.video.VideoEncoderConfig.get_codec_options`
     against PyAV (including numeric ``extra_options`` present in that dict).
+    When given, additionally verify that *pix_fmt* carries as many components as the source data channels.
     No-op when ``config.vcodec`` isn't in the local FFmpeg build.
 
     Raises:
@@ -171,4 +216,6 @@ def check_video_encoder_parameters_pyav(vcodec: str, pix_fmt: str, codec_options
     if not options:
         raise ValueError(f"Codec {vcodec!r} is not available in the bundled FFmpeg build")
     _check_pixel_format(vcodec, pix_fmt)
+    if channels is not None:
+        _check_pix_fmt_channels(pix_fmt, channels)
     _check_codec_options(vcodec, codec_options)

src/lerobot/datasets/sampler.py

@@ -53,6 +53,7 @@ class EpisodeAwareSampler:
         drop_n_last_frames: int = 0,
         shuffle: bool = False,
         seed: int = 0,
+        absolute_to_relative_idx: dict[int, int] | None = None,
     ):
         """
         Args:
@@ -107,6 +108,7 @@ class EpisodeAwareSampler:
         self.seed = seed
         self._epoch = 0
         self._start_index = 0
+        self._absolute_to_relative = absolute_to_relative_idx
 
     @property
     def indices(self) -> list[int]:
@@ -132,7 +134,10 @@ class EpisodeAwareSampler:
     def _frame_index(self, position: int) -> int:
         episode = int(np.searchsorted(self._cum_lengths, position, side="right"))
         position_in_episode = position - (int(self._cum_lengths[episode - 1]) if episode > 0 else 0)
-        return int(self._starts[episode]) + position_in_episode
+        absolute_idx = int(self._starts[episode]) + position_in_episode
+        if self._absolute_to_relative is not None:
+            return self._absolute_to_relative[absolute_idx]
+        return absolute_idx
 
     def __iter__(self) -> Iterator[int]:
         # Advance epoch state eagerly, not on first consumption of the generator.