fix(robots): restore joint clipping and wrist_yaw fallback in ReBot B601 send_action

- Add configurable arm control mode (mit default, pos_vel fallback) with tunable mit_kp / mit_kd
- Add optional gripper control mode (force_pos default, mit optional) with gripper_mit_kp / gripper_mit_kd
- Update tests for MIT arm routing, gripper mode routing, and revised joint limits

docs/source/_toctree.yml

@@ -9,6 +9,8 @@
 - sections:
   - local: il_robots
     title: Imitation Learning for Robots
+  - local: lelab
+    title: LeLab - Lerobot GUI
   - local: bring_your_own_policies
     title: Adding a Policy
   - local: integrate_hardware
@@ -61,6 +63,8 @@
     title: π₀.₅ (Pi05)
   - local: molmoact2
     title: MolmoAct2
+  - local: vla_jepa
+    title: VLA-JEPA
   - local: eo1
     title: EO-1
   - local: groot
@@ -75,6 +79,8 @@
 - sections:
   - local: sarm
     title: SARM
+  - local: robometer
+    title: ROBOMETER
   - local: topreward
     title: TOPReward
   title: "Reward Models"

docs/source/il_robots.mdx

@@ -647,5 +647,6 @@ The `--strategy.type` flag selects the execution mode:
 - `sentry`: Continuous recording with auto-upload (useful for large-scale evaluation)
 - `highlight`: Ring buffer recording with keystroke save (useful for capturing interesting events)
 - `dagger`: Human-in-the-loop data collection (see [HIL Data Collection](./hil_data_collection))
+- `episodic`: Episode-oriented policy recording with reset phases between episodes
 
 All strategies support `--inference.type=rtc` for smooth execution with slow VLA models (Pi0, Pi0.5, SmolVLA).

docs/source/inference.mdx

@@ -157,6 +157,44 @@ Foot pedal input is also supported via `--strategy.input_device=pedal`. Configur
 | `--strategy.input_device`            | Input device: `keyboard` or `pedal` (default: keyboard) |
 | `--teleop.type`                      | **Required.** Teleoperator type                         |
 
+### Episodic (`--strategy.type=episodic`)
+
+Episode-oriented recording that mirrors the behavior of `lerobot-record`. The policy drives the robot for each episode; an optional teleoperator can drive the robot during the reset phase between episodes.
+
+```bash
+lerobot-rollout \
+    --strategy.type=episodic \
+    --policy.path=${HF_USER}/my_policy \
+    --robot.type=so100_follower \
+    --robot.port=/dev/ttyACM0 \
+    --teleop.type=so100_leader \
+    --teleop.port=/dev/ttyACM1 \
+    --dataset.repo_id=${HF_USER}/my_eval_data \
+    --dataset.num_episodes=20 \
+    --dataset.episode_time_s=30 \
+    --dataset.reset_time_s=10 \
+    --dataset.single_task="Pick up the red cube"
+```
+
+Teleop is optional — if omitted the robot holds its position during the reset phase.
+
+**Keyboard controls:**
+
+| Key         | Action                           |
+| ----------- | -------------------------------- |
+| `→` (right) | End the current episode early    |
+| `←` (left)  | Discard episode and re-record it |
+| `ESC`       | Stop the recording session       |
+
+| Flag                                            | Description                                                                |
+| ----------------------------------------------- | -------------------------------------------------------------------------- |
+| `--dataset.num_episodes`                        | Number of episodes to record                                               |
+| `--dataset.episode_time_s`                      | Duration of each recording episode in seconds                              |
+| `--dataset.reset_time_s`                        | Duration of the reset phase between episodes in seconds                    |
+| `--teleop.type`                                 | Optional. Teleoperator to drive the robot during resets                    |
+| `--strategy.reset_to_initial_position`          | Whether to reset the robot to its initial position between episodes        |
+| `--strategy.smooth_leader_to_follower_handover` | Whether to turn on or off the leader -> follower smooth handover behavior. |
+
 ---
 
 ## Inference Backends

docs/source/lelab.mdx

@@ -0,0 +1,29 @@
+# LeLab - LeRobot Guide
+
+LeLab is a graphical user interface built on top of the LeRobot library, designed to make robotics accessible without needing to memorize CLI commands. From a single app you can configure your robot, teleoperate it, collect datasets, train policies locally or on cloud GPUs via HF Jobs, and deploy trained models back onto your robot. It's the easiest way to go from an unboxed SO-101 to a working policy, and a great companion for anyone learning the LeRobot workflow. Source code and issues live on GitHub: [huggingface/leLab](https://github.com/huggingface/leLab).
+
+> [!TIP]
+> For now LeLab is compatible only with SO-ARM101
+
+<Youtube id="VqyKUuW9V1g" />
+
+### Installation
+
+Requires [`uv`](https://docs.astral.sh/uv/getting-started/installation/). Install and launch in one command:
+
+```
+uv tool install git+https://github.com/huggingface/leLab.git && lelab
+```
+
+After install, run `lelab` from your terminal anytime to start the app.
+
+### Features
+
+- **Add robots** — Select arm type (leader/follower), calibrate each joint from the middle position, and attach cameras.
+- **Teleoperation** — Control the follower arm with the leader and see a live 3D visualization of the arms.
+- **Dataset recording** — Define a task description, number of episodes, and episode/reset durations. Press spacebar to advance between episodes. 30+ episodes recommended.
+- **Local training** — Train a policy directly on your own machine with a selected dataset, policy type, batch size, and step count.
+- **Cloud training with HF Jobs** — Train on powerful GPUs via [HF Jobs](https://huggingface.co/docs/huggingface_hub/en/guides/jobs) with transparent pricing. Run `hf auth login` first. See the [Compute HW Guide](hardware_guide) for hardware/batch size tips.
+- **Training visualization** — Watch progress live in the app, with checkpoints saved automatically.
+- **Run trained policies** — Pick any model from your jobs list and run inference on your robot with one click.
+- **Use community datasets** — Provide any Hugging Face dataset ID to train on datasets you didn't record yourself.

docs/source/lerobot-dataset-v3.mdx

@@ -275,7 +275,7 @@ A converter aggregates per‑episode files into larger shards and writes episode
 pip install "https://github.com/huggingface/lerobot/archive/33cad37054c2b594ceba57463e8f11ee374fa93c.zip"
 
 # Convert an existing v2.1 dataset hosted on the Hub:
-python -m lerobot.datasets.v30.convert_dataset_v21_to_v30 --repo-id=<HF_USER/DATASET_ID>
+python -m lerobot.scripts.convert_dataset_v21_to_v30 --repo-id=<HF_USER/DATASET_ID>
 ```
 
 **What it does**

docs/source/molmoact2.mdx

@@ -238,7 +238,7 @@ your dataset has not been converted with quantile statistics, you can add them
 with:
 
 ```bash
-python src/lerobot/datasets/v30/augment_dataset_quantile_stats.py \
+python src/lerobot/scripts/augment_dataset_quantile_stats.py \
   --repo-id=your_dataset
 ```
 

docs/source/pi05.mdx

@@ -91,7 +91,7 @@ lerobot-train \
 If your dataset is not converted with `quantiles`, you can convert it with the following command:
 
 ```bash
-python src/lerobot/datasets/v30/augment_dataset_quantile_stats.py \
+python src/lerobot/scripts/augment_dataset_quantile_stats.py \
     --repo-id=your_dataset \
 ```
 

docs/source/policy_vla_jepa_README.md

@@ -0,0 +1,39 @@
+# VLA-JEPA
+
+This repository contains the LeRobot port of **VLA-JEPA**, a Vision-Language-Action model that combines a Qwen3-VL language backbone with a self-supervised video world model (V-JEPA2) and a flow-matching DiT action head.
+
+Converted from [ginwind/VLA-JEPA](https://huggingface.co/ginwind/VLA-JEPA).
+
+---
+
+## Architecture Overview
+
+| Component               | Module                            | Role                                                    |
+| ----------------------- | --------------------------------- | ------------------------------------------------------- |
+| **Qwen3-VL backbone**   | `Qwen3VLInterface`                | Fuses images + language instruction into context tokens |
+| **DiT-B action head**   | `VLAJEPAActionHead`               | Flow-matching diffusion over the action chunk           |
+| **V-JEPA2 world model** | `ActionConditionedVideoPredictor` | Self-supervised video prediction loss (training only)   |
+
+At inference time only the Qwen backbone and action head are used; the world model is not needed.
+
+---
+
+## Citation
+
+```bibtex
+@misc{sun2026vlajepaenhancingvisionlanguageactionmodel,
+  title         = {VLA-JEPA: Enhancing Vision-Language-Action Model with Latent World Model},
+  author        = {Jingwen Sun and Wenyao Zhang and Zekun Qi and Shaojie Ren and Zezhi Liu and Hanxin Zhu and Guangzhong Sun and Xin Jin and Zhibo Chen},
+  year          = {2026},
+  eprint        = {2602.10098},
+  archivePrefix = {arXiv},
+  primaryClass  = {cs.RO},
+  url           = {https://arxiv.org/abs/2602.10098},
+}
+```
+
+---
+
+## License
+
+Weights are distributed under the license terms of the original [ginwind/VLA-JEPA](https://huggingface.co/ginwind/VLA-JEPA) repository (**Apache 2.0 License**). The LeRobot integration code follows the **Apache 2.0 License**.

docs/source/porting_datasets_v3.mdx

@@ -300,7 +300,7 @@ This replaces the old episode-per-file structure with efficient, optimally-sized
 If you have existing datasets in v2.1 format, use the migration tool:
 
 ```bash
-python src/lerobot/datasets/v30/convert_dataset_v21_to_v30.py \
+python src/lerobot/scripts/convert_dataset_v21_to_v30.py \
     --repo-id your_id/existing_dataset
 ```
 

docs/source/robometer.mdx

@@ -0,0 +1,185 @@
+# ROBOMETER
+
+ROBOMETER is a **general-purpose video-language robotic reward model**. It predicts dense, frame-level task progress and frame-level success from a trajectory video and a task description.
+
+**Paper**: [ROBOMETER: Scaling General-Purpose Robotic Reward Models via Trajectory Comparisons](https://arxiv.org/abs/2603.02115)
+**Project**: [robometer.github.io](https://robometer.github.io/)
+**Original code**: [github.com/robometer/robometer](https://github.com/robometer/robometer)
+**Checkpoint**: [lerobot/Robometer-4B](https://huggingface.co/lerobot/Robometer-4B)
+
+## Overview
+
+ROBOMETER builds on `Qwen/Qwen3-VL-4B-Instruct` and adds three lightweight prediction heads:
+
+- **Progress head**: predicts per-frame task progress in `[0, 1]`.
+- **Success head**: predicts per-frame task success probability.
+- **Preference head**: predicts which of two trajectories better completes the task during training.
+
+The paper trains ROBOMETER with a composite objective:
+
+```text
+L = L_pref + L_prog + L_succ
+```
+
+The LeRobot integration is currently **inference-only**. It preserves the preference head so that the published `Robometer-4B` checkpoint loads without remapping, but `compute_reward()` queries the progress or success head only.
+
+## What the LeRobot Integration Covers
+
+- Standard `reward_model.type=robometer` configuration through LeRobot.
+- Qwen3-VL image and text preprocessing through `RobometerEncoderProcessorStep`.
+- LeRobot reward-model save/load APIs through `PreTrainedRewardModel`.
+- Dense, frame-level progress and success predictions internally.
+- A scalar reward through `compute_reward()` for downstream LeRobot reward-model usage.
+
+This page focuses on using the published ROBOMETER checkpoint as a zero-shot reward model. Training ROBOMETER from scratch is outside the current LeRobot integration.
+
+## Installation Requirements
+
+1. Install LeRobot by following the [Installation Guide](./installation).
+2. Install the ROBOMETER dependencies:
+
+```bash
+pip install -e ".[robometer]"
+```
+
+If you use `uv` directly from a source checkout:
+
+```bash
+uv sync --extra robometer
+```
+
+ROBOMETER uses a Qwen3-VL-4B backbone, so GPU inference is strongly recommended.
+
+## Model Inputs and Outputs
+
+ROBOMETER expects:
+
+- A trajectory video or sequence of frames.
+- A natural-language task description.
+
+In LeRobot datasets, the preprocessor reads:
+
+| Config field              | Default                  | Meaning                                               |
+| ------------------------- | ------------------------ | ----------------------------------------------------- |
+| `reward_model.image_key`  | `observation.images.top` | Camera/video observation used by ROBOMETER            |
+| `reward_model.task_key`   | `task`                   | Key in complementary data that stores the task string |
+| `reward_model.max_frames` | `8`                      | Maximum number of frames passed to ROBOMETER          |
+
+The model predicts per-frame progress and success internally. The LeRobot reward API returns a scalar per sample:
+
+- `reward_output="progress"` (default): return the last-frame progress, clamped to `[0, 1]`.
+- `reward_output="success"`: return `1.0` if the last-frame success probability is above `success_threshold`, otherwise `0.0`.
+
+## Usage
+
+### Load the Reward Model Directly
+
+```python
+from lerobot.rewards.robometer import RobometerConfig, RobometerRewardModel
+
+cfg = RobometerConfig(
+    pretrained_path="lerobot/Robometer-4B",
+    device="cuda",
+    reward_output="progress",
+)
+reward_model = RobometerRewardModel.from_pretrained(cfg.pretrained_path, config=cfg)
+```
+
+### Encode Frames and Compute a Reward
+
+For a direct Python call, provide frames as `uint8` arrays with shape `(T, H, W, C)` and a task string:
+
+```python
+from lerobot.rewards.robometer.modeling_robometer import ROBOMETER_FEATURE_PREFIX
+from lerobot.rewards.robometer.processor_robometer import RobometerEncoderProcessorStep
+
+# frames: np.ndarray, shape (T, H, W, C), dtype uint8
+# task: str
+encoder = RobometerEncoderProcessorStep(
+    base_model_id=cfg.base_model_id,
+    use_multi_image=cfg.use_multi_image,
+    use_per_frame_progress_token=cfg.use_per_frame_progress_token,
+    max_frames=cfg.max_frames,
+)
+
+encoded = encoder.encode_samples([(frames, task)])
+batch = {f"{ROBOMETER_FEATURE_PREFIX}{key}": value for key, value in encoded.items()}
+
+reward = reward_model.compute_reward(batch)
+```
+
+`reward` is a tensor of shape `(batch_size,)`.
+
+### Use the Reward Factory
+
+You can also instantiate ROBOMETER through the reward factory:
+
+```python
+from lerobot.rewards import make_reward_model, make_reward_model_config, make_reward_pre_post_processors
+
+cfg = make_reward_model_config(
+    "robometer",
+    pretrained_path="lerobot/Robometer-4B",
+    device="cuda",
+    image_key="observation.images.top",
+)
+reward_model = make_reward_model(cfg)
+preprocessor, postprocessor = make_reward_pre_post_processors(cfg)
+```
+
+The preprocessor writes Qwen-VL tensors under the `observation.robometer.*` namespace, and `compute_reward()` reads those encoded tensors.
+
+## Configuration Notes
+
+### Backbone and Vocabulary
+
+The published checkpoint uses a Qwen3-VL-4B backbone. ROBOMETER adds five special tokens to the tokenizer in a fixed order:
+
+```text
+<|split_token|>
+<|reward_token|>
+<|pref_token|>
+<|sim_token|>
+<|prog_token|>
+```
+
+`<|prog_token|>` is inserted after each frame and is the hidden-state position used for per-frame progress and success prediction. `<|split_token|>` and `<|pref_token|>` are used by the paper's pairwise trajectory preference objective. `<|reward_token|>` and `<|sim_token|>` are preserved for checkpoint compatibility.
+
+The LeRobot config stores a serialized `vlm_config` with the post-resize vocabulary so the model can reload from `config.json` without downloading the base Qwen weights first. For `Qwen/Qwen3-VL-4B-Instruct`, the tokenizer length is `151669`, and the five ROBOMETER tokens produce the checkpoint vocabulary size `151674`.
+
+### Progress Prediction
+
+In the published checkpoint, progress is discrete. The progress head outputs logits over `progress_discrete_bins=10` uniformly spaced bin centers in `[0, 1]`. LeRobot converts these logits into a continuous value by applying a softmax and taking the expectation over bin centers, matching the upstream ROBOMETER implementation.
+
+### Success Prediction
+
+The success head outputs raw logits per frame. LeRobot converts them to probabilities with `sigmoid`. When `reward_output="success"`, `compute_reward()` thresholds the last-frame success probability using `success_threshold`.
+
+## Limitations
+
+- The current LeRobot integration is inference-only; it does not implement ROBOMETER training or preference-pair training.
+- `compute_reward()` returns a scalar per sample for the LeRobot reward-model API, even though ROBOMETER predicts per-frame progress and success internally.
+- ROBOMETER is video-language based; it does not use privileged robot state such as contact forces or object poses.
+
+## References
+
+- [ROBOMETER project](https://robometer.github.io/)
+- [ROBOMETER paper](https://arxiv.org/abs/2603.02115)
+- [Original ROBOMETER code](https://github.com/robometer/robometer)
+- [Published ROBOMETER-4B checkpoint](https://huggingface.co/lerobot/Robometer-4B)
+- [Qwen3-VL-4B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct)
+
+## Citation
+
+```bibtex
+@inproceedings{liang2026robometer,
+title = {Robometer: Scaling General-Purpose Robotic Reward Models via Trajectory Comparisons},
+author={Anthony Liang and Yigit Korkmaz and Jiahui Zhang and Minyoung Hwang and Abrar Anwar and Sidhant Kaushik and Aditya Shah and Alex S. Huang and Luke Zettlemoyer and Dieter Fox and Yu Xiang and Anqi Li and Andreea Bobu and Abhishek Gupta and Stephen Tu and Erdem Biyik and Jesse Zhang},
+year={2026},
+booktitle={Robotics: Science and Systems 2026},
+}
+```
+
+## License
+
+This LeRobot integration follows the **Apache 2.0 License** used by LeRobot. Check the upstream ROBOMETER code and model pages for the licenses of the original implementation and released checkpoints.

docs/source/vla_jepa.mdx

@@ -0,0 +1,235 @@
+# VLA-JEPA
+
+This is the LeRobot port of **VLA-JEPA**, a Vision-Language-Action model that combines a Qwen3-VL language backbone with a self-supervised video world model (V-JEPA2) and a flow-matching DiT action head.
+
+---
+
+## Architecture Overview
+
+VLA-JEPA has three main components:
+
+| Component               | Module                            | Role                                                    |
+| ----------------------- | --------------------------------- | ------------------------------------------------------- |
+| **Qwen3-VL backbone**   | `Qwen3VLInterface`                | Fuses images + language instruction into context tokens |
+| **DiT-B action head**   | `VLAJEPAActionHead`               | Flow-matching diffusion over the action chunk           |
+| **V-JEPA2 world model** | `ActionConditionedVideoPredictor` | Self-supervised video prediction loss (training only)   |
+
+### Data flow
+
+**Training:**
+
+1. A video clip of `num_video_frames` frames is encoded by V-JEPA2 into per-frame patch tokens.
+2. The Qwen3-VL backbone processes multi-view images + the task instruction and produces a sequence of context tokens that includes special action tokens (for world model conditioning) and embodied tokens.
+3. The action head receives those context tokens as cross-attention keys/values and predicts a denoised action chunk via flow matching.
+4. The world model predictor uses the action tokens extracted from Qwen to predict future V-JEPA2 frame embeddings; a regression loss on those predictions is added to the action loss.
+
+**Inference:**
+Only Qwen + the action head are used. The world model is not needed at inference time.
+
+### Action head details
+
+Available presets via `action_model_type`:
+
+| Preset  | Hidden dim | Heads | Head dim |
+| ------- | ---------- | ----- | -------- |
+| `DiT-B` | 768        | 12    | 64       |
+| `DiT-L` | 1536       | 32    | 48       |
+
+### World model details
+
+The video predictor is a ViT-style transformer (`ActionConditionedVideoPredictor`) that takes:
+
+- **Frame tokens**: V-JEPA2 patch embeddings projected to `predictor_embed_dim`
+- **Action tokens**: Qwen action token embeddings projected to `predictor_embed_dim`
+
+It uses block-causal attention so each temporal step can attend to all previous steps. The predictor's input `embed_dim` equals `num_views × video_encoder_hidden_size` (e.g. 2 views × 1024 = 2048 for the pretrained checkpoints).
+
+---
+
+## Pretrained Checkpoints
+
+Three checkpoints are available directly inside the LeRobot org here: [`lerobot/VLA-JEPA`](https://huggingface.co/collections/lerobot/vla-jepa), converted from [ginwind/VLA-JEPA](https://huggingface.co/ginwind/VLA-JEPA):
+
+| Checkpoint                    | Dataset           | Cameras                 | World model | Action dim |
+| ----------------------------- | ----------------- | ----------------------- | ----------- | ---------- |
+| `lerobot/VLA-JEPA-LIBERO`     | LIBERO-10         | 2 (agentview + wrist)   | Enabled     | 7          |
+| `lerobot/VLA-JEPA-Pretrain`   | DROID 1.0.1       | 2 (exterior left views) | Enabled     | 7          |
+| `lerobot/VLA-JEPA-SimplerEnv` | OXE Bridge / RT-1 | 1 (view duplicated ×2)  | Enabled     | 7          |
+
+All checkpoints use `Qwen/Qwen3-VL-2B-Instruct` as the language backbone.
+
+---
+
+## Configuration
+
+Key parameters in `VLAJEPAConfig`:
+
+| Parameter                 | Default | Description                                                                                                                                                                     |
+| ------------------------- | ------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `chunk_size`              | 7       | Number of actions predicted per inference call                                                                                                                                  |
+| `n_action_steps`          | 7       | Steps executed from the predicted chunk before re-planning                                                                                                                      |
+| `num_video_frames`        | 8       | Video clip length fed to the world model                                                                                                                                        |
+| `enable_world_model`      | `True`  | Whether to load and train the V-JEPA2 predictor                                                                                                                                 |
+| `world_model_loss_weight` | 0.1     | Weight of the JEPA prediction loss relative to the action loss                                                                                                                  |
+| `num_inference_timesteps` | 4       | Euler integration steps for action denoising                                                                                                                                    |
+| `freeze_qwen`             | `False` | Freeze the Qwen3-VL backbone and only train the action head                                                                                                                     |
+| `reinit_modules`          | `None`  | Key prefixes allowed to be randomly re-initialised on load (for cross-embodiment transfer, see [Fine-tuning on a different embodiment](#fine-tuning-on-a-different-embodiment)) |
+| `gripper_dim`             | 6       | Index of the gripper dimension in the action vector (e.g. 6 for a 7-DoF arm with gripper as the last joint)                                                                     |
+| `gripper_threshold`       | 0.5     | Threshold used by `pre_snap_gripper_action` and `binarize_gripper_action` to binarize the gripper dimension                                                                     |
+| `pre_snap_gripper_action` | `True`  | Snap the gripper dim to {0, 1} before unnormalization. Set to `False` for robots without a binary gripper                                                                       |
+| `binarize_gripper_action` | `True`  | Binarize the gripper dim to {-1, 1} after unnormalization. Set to `False` for robots without a binary gripper                                                                   |
+
+---
+
+## Training
+
+Number of training steps may vary based on dataset size and compute budget. The original paper pretrained for 50k on ssv2 + droid jointly, then additional 30k steps for LIBERO, but fewer steps may still yield good performance when fine-tuning from the provided pretrained checkpoints.
+
+### Full training from scratch
+
+```bash
+lerobot-train \
+  policy.type=vla_jepa \
+  policy.repo_id=your_org/your_repo \
+  dataset.repo_id=your_org/your_dataset
+```
+
+### Fine-tuning from a pretrained checkpoint
+
+```bash
+lerobot-train \
+  --policy.path=lerobot/VLA-JEPA-Pretrain \
+  --policy.repo_id=your_org/your_repo \
+  --dataset.repo_id=your_org/your_dataset
+```
+
+If you want to freeze the Qwen backbone and only train the action head, set `policy.freeze_qwen=True`:
+
+```bash
+lerobot-train \
+  --policy.path=lerobot/VLA-JEPA-Pretrain \
+  --policy.repo_id=your_org/your_repo \
+  --policy.freeze_qwen=true \
+  --dataset.repo_id=your_org/your_dataset
+```
+
+### Fine-tuning on a different embodiment
+
+When the target robot has a different action or state dimensionality than the pretrained checkpoint, the input/output projection layers of the action head will have mismatched shapes and cannot be loaded directly. `reinit_modules` lets you list the key prefixes that are allowed to mismatch — those layers are randomly re-initialised while every other weight is reused from the checkpoint. Any shape mismatch outside the listed prefixes raises an error.
+
+The layers that depend on `action_dim` and `state_dim` are:
+
+| Layer                                     | Key prefix                          |
+| ----------------------------------------- | ----------------------------------- |
+| Action encoder (action_dim → inner_dim)   | `model.action_model.action_encoder` |
+| Action decoder (hidden_size → action_dim) | `model.action_model.action_decoder` |
+| State encoder (state_dim → inner_dim)     | `model.action_model.state_encoder`  |
+
+```bash
+lerobot-train \
+  --policy.path=lerobot/VLA-JEPA-Pretrain \
+  --policy.repo_id=your_org/your_repo \
+  --policy.freeze_qwen=true \
+  --policy.reinit_modules='["model.action_model.action_encoder", "model.action_model.action_decoder", "model.action_model.state_encoder"]' \
+  --dataset.repo_id=your_org/your_dataset
+```
+
+If your robot has no proprioceptive state, omit `model.action_model.state_encoder` from the list.
+
+### Reproducing the LIBERO results
+
+**Training on LIBERO:**
+starts the training from the Pretrain checkpoint, trains for 30k steps on the LIBERO dataset.
+Original paper mentions training across 8 GPUs with a batch size of 32, meaning global batch size of 256.
+
+```bash
+lerobot-train \
+  --policy.path=lerobot/VLA-JEPA-Pretrain \
+  --policy.repo_id=your_org/your_repo \
+  --dataset.repo_id=HuggingFaceVLA/libero \
+  --steps=30000
+```
+
+**Evaluating the pretrained LIBERO-10 checkpoint:**
+
+```bash
+lerobot-eval \
+  --policy.path=lerobot/VLA-JEPA-LIBERO \
+  --env.type=libero \
+  --env.task=libero_spatial,libero_object,libero_goal,libero_10 \
+  --eval.n_episodes=10 \
+  --eval.batch_size=5
+```
+
+To evaluate a subset of tasks only:
+
+```bash
+lerobot-eval \
+  --policy.path=lerobot/VLA-JEPA-LIBERO \
+  --env.type=libero \
+  --env.task=libero_10 \
+  --env.task_ids='[0,1,2]' \
+  --eval.n_episodes=10 \
+  --eval.batch_size=5
+```
+
+**Expected results:**
+
+| Suite          | Episodes | Successes | Success Rate |
+| -------------- | -------- | --------- | ------------ |
+| libero_spatial | 100      | 93        | **95.0%**    |
+| libero_object  | 100      | 100       | **100.0%**   |
+| libero_goal    | 100      | 98        | **98.0%**    |
+| libero_10      | 100      | 96        | **93.0%**    |
+| **Overall**    | **400**  | **387**   | **96.5%**    |
+
+---
+
+## Fine-tuning on datasets with a different number of cameras
+
+The pretrained world model predictor was trained with `embed_dim = jepa_tubelet_size × 1024` (default `jepa_tubelet_size=2`).
+
+**Default behaviour — view padding / trimming (no action required)**
+
+When fine-tuning from `VLA-JEPA-Pretrain` the model automatically adjusts the number of views fed to the world model to match `jepa_tubelet_size`:
+
+- **Single-view datasets (e.g. BridgeV2):** the single-view latent is duplicated to produce a two-view world-model input, preserving the JEPA self-supervised signal without any weight mismatch.
+- **>2-view datasets (e.g. DROID with 3 views):** all views are passed to the Qwen backbone (for richer context), but only the first `jepa_tubelet_size` views (one wrist + one third-person, following the configured view order) are used for the world model.
+
+**Option 1 — Disable the world model**
+
+Set `enable_world_model=False` to skip the JEPA loss entirely. Only the Qwen backbone and action head are loaded and trained. This is sufficient for good action performance.
+
+```bash
+lerobot-train \
+  --policy.path=lerobot/VLA-JEPA-Pretrain \
+  --policy.enable_world_model=false \
+  --policy.repo_id=your_org/your_repo \
+  --dataset.repo_id=your_org/single_camera_dataset
+```
+
+**Option 2 — Reinitialize the predictor input projection**
+
+If you want to change `jepa_tubelet_size` to a value other than 2, load the checkpoint with `strict=False` and reinitialize `model.video_predictor.predictor_embed` for the new `embed_dim`. All other predictor block weights (attention, MLP, norm, output projection) are camera-count-agnostic and can be reused from the pretrained checkpoint.
+
+---
+
+## Citation
+
+```bibtex
+@misc{sun2026vlajepaenhancingvisionlanguageactionmodel,
+  title         = {VLA-JEPA: Enhancing Vision-Language-Action Model with Latent World Model},
+  author        = {Jingwen Sun and Wenyao Zhang and Zekun Qi and Shaojie Ren and Zezhi Liu and Hanxin Zhu and Guangzhong Sun and Xin Jin and Zhibo Chen},
+  year          = {2026},
+  eprint        = {2602.10098},
+  archivePrefix = {arXiv},
+  primaryClass  = {cs.RO},
+  url           = {https://arxiv.org/abs/2602.10098},
+}
+```
+
+---
+
+## License
+
+Weights are distributed under the license terms of the original [ginwind/VLA-JEPA](https://huggingface.co/ginwind/VLA-JEPA) repository (**Apache 2.0 License**). The LeRobot integration code follows the **Apache 2.0 License**.

pyproject.toml

@@ -212,10 +212,12 @@ groot = [
     "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]"]
-hilserl = ["lerobot[transformers-dep]", "lerobot[dataset]", "gym-hil>=0.1.13,<0.2.0", "lerobot[grpcio-dep]", "lerobot[placo-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]"]
 
 # Features
 async = ["lerobot[grpcio-dep]", "lerobot[matplotlib-dep]"]
@@ -229,9 +231,9 @@ video_benchmark = ["scikit-image>=0.23.2,<0.26.0", "pandas>=2.2.2,<2.4.0"]
 
 # Simulation
 # NOTE: Explicitly listing scipy helps flatten the dependecy tree.
-aloha = ["lerobot[dataset]", "gym-aloha>=0.1.2,<0.2.0", "lerobot[scipy-dep]"]
+aloha = ["lerobot[dataset]", "gym-aloha>=0.1.4,<0.2.0", "lerobot[scipy-dep]"]
 pusht = ["lerobot[dataset]", "gym-pusht>=0.1.5,<0.2.0", "pymunk>=6.6.0,<7.0.0"] # TODO: Fix pymunk version in gym-pusht instead
-libero = ["lerobot[dataset]", "lerobot[transformers-dep]", "hf-libero>=0.1.3,<0.2.0; sys_platform == 'linux'", "lerobot[scipy-dep]"]
+libero = ["lerobot[dataset]", "lerobot[transformers-dep]", "hf-libero>=0.1.4,<0.2.0; sys_platform == 'linux'", "lerobot[scipy-dep]"]
 metaworld = ["lerobot[dataset]", "metaworld==3.0.0", "lerobot[scipy-dep]"]
 # NOTE: vlabench is NOT exposed as a `lerobot` extra. Its only distribution
 # is the OpenMOSS/VLABench GitHub repo (package name `VLABench`, no PyPI
@@ -281,6 +283,7 @@ all = [
     # "lerobot[groot]", TODO(Steven): Gr00t requires specific installation instructions for flash-attn
     "lerobot[xvla]",
     "lerobot[hilserl]",
+    "lerobot[vla_jepa]",
     "lerobot[async]",
     "lerobot[dev]",
     "lerobot[test]",
@@ -291,6 +294,7 @@ all = [
     "lerobot[libero]; sys_platform == 'linux'",
     "lerobot[metaworld]",
     "lerobot[sarm]",
+    "lerobot[robometer]",
     "lerobot[topreward]",
     "lerobot[peft]",
     # "lerobot[unitree_g1]", TODO: Unitree requires specific installation instructions for unitree_sdk2

src/lerobot/common/control_utils.py

@@ -18,6 +18,7 @@ from __future__ import annotations
 # Utilities
 ########################################################################################
 import logging
+import time
 import traceback
 from contextlib import nullcontext
 from copy import copy
@@ -243,3 +244,72 @@ def sanity_check_dataset_robot_compatibility(
         raise ValueError(
             "Dataset metadata compatibility check failed with mismatches:\n" + "\n".join(mismatches)
         )
+
+
+########################################################################################
+# Teleoperator smooth handover helpers
+# NOTE(Maxime): These functions use minimal type hints to maintain compatibility with utils
+# being a root module.
+########################################################################################
+
+
+def teleop_supports_feedback(teleop) -> bool:
+    """Return True when the teleop can receive position feedback (is actuated).
+
+    Actuated teleops (e.g. SO-101, OpenArmMini) have non-empty ``feedback_features``
+    and expose ``enable_torque`` / ``disable_torque`` motor-control methods.
+
+    TODO(Maxime): See if it is possible to unify this interface across teleops instead of duck-typing.
+    """
+    return (
+        bool(teleop.feedback_features)
+        and hasattr(teleop, "disable_torque")
+        and hasattr(teleop, "enable_torque")
+    )
+
+
+def teleop_smooth_move_to(teleop, target_pos: dict, duration_s: float = 2.0, fps: int = 30) -> None:
+    """Smoothly move an actuated teleop to ``target_pos`` via linear interpolation.
+
+    Requires the teleoperator to support feedback (i.e. have non-empty
+    ``feedback_features`` and implement ``disable_torque`` / ``enable_torque``).
+
+    ``target_pos`` is expected to be in the teleop's action/feedback key space.
+    For homogeneous setups (e.g. SO-101 leader + SO-101 follower) this matches
+    the robot action key space directly.
+
+    TODO(Maxime): This blocks up to ``duration_s`` seconds; during this time the
+    follower robot does not receive new actions, which could be an issue on LeKiwi.
+    """
+    teleop.enable_torque()
+    current = teleop.get_action()
+    steps = max(int(duration_s * fps), 1)
+
+    for step in range(steps + 1):
+        t = step / steps
+        interp = {
+            k: current[k] * (1 - t) + target_pos[k] * t if k in target_pos else current[k] for k in current
+        }
+        teleop.send_feedback(interp)
+        time.sleep(1 / fps)
+
+
+def follower_smooth_move_to(
+    robot, current: dict, target: dict, duration_s: float = 1.0, fps: int = 30
+) -> None:
+    """Smoothly move the follower robot from ``current`` to ``target`` action.
+
+    Used when the teleop is non-actuated: instead of driving the leader arm to
+    the follower, the follower is brought to the teleop's current pose so the
+    robot meets the operator's hand rather than jumping to it on the first frame.
+
+    Both ``current`` and ``target`` must be in the robot action key space
+    (i.e. the output of ``robot_action_processor``).
+    """
+    steps = max(int(duration_s * fps), 1)
+
+    for step in range(steps + 1):
+        t = step / steps
+        interp = {k: current[k] * (1 - t) + target[k] * t if k in target else current[k] for k in current}
+        robot.send_action(interp)
+        time.sleep(1 / fps)

src/lerobot/configs/dataset.py

@@ -41,8 +41,8 @@ class DatasetRecordConfig:
     video: bool = True
     # Upload dataset to Hugging Face hub.
     push_to_hub: bool = True
-    # Upload on private repository on the Hugging Face hub.
-    private: bool = False
+    # If True, upload as private; if None, defer to the org default on the Hub (only affects orgs).
+    private: bool | None = None
     # Add tags to your dataset on the hub.
     tags: list[str] | None = None
     # Number of subprocesses handling the saving of frames as PNG. Set to 0 to use threads only;

src/lerobot/configs/train.py

@@ -177,6 +177,12 @@ class TrainPipelineConfig(HubMixin):
             )
 
         active_cfg = self.trainable_config
+        if self.rename_map and active_cfg.pretrained_path is None:
+            raise ValueError(
+                "`rename_map` requires a pretrained policy checkpoint. "
+                "Fresh initialization derives feature names from the current dataset, so no rename is applied."
+            )
+
         if not self.job_name:
             if self.env is None:
                 self.job_name = f"{active_cfg.type}"

src/lerobot/datasets/lerobot_dataset.py

@@ -524,7 +524,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         license: str | None = "apache-2.0",
         tag_version: bool = True,
         push_videos: bool = True,
-        private: bool = False,
+        private: bool | None = None,
         allow_patterns: list[str] | str | None = None,
         upload_large_folder: bool = False,
         **card_kwargs,
@@ -543,7 +543,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
             tag_version: If ``True``, create a Git tag for the current codebase
                 version.
             push_videos: If ``False``, skip uploading the ``videos/`` directory.
-            private: If ``True``, create a private repository.
+            private: If ``True``, create a private repository. If ``None``
+                (default), defer to the org default on the Hub (only affects orgs).
             allow_patterns: Glob pattern(s) restricting which files to upload.
             upload_large_folder: If ``True``, use ``upload_large_folder`` instead
                 of ``upload_folder`` for very large datasets.

src/lerobot/datasets/sampler.py

@@ -30,6 +30,7 @@ class EpisodeAwareSampler:
         drop_n_first_frames: int = 0,
         drop_n_last_frames: int = 0,
         shuffle: bool = False,
+        generator: torch.Generator | None = None,
     ):
         """Sampler that optionally incorporates episode boundary information.
 
@@ -41,6 +42,10 @@ class EpisodeAwareSampler:
             drop_n_first_frames: Number of frames to drop from the start of each episode.
             drop_n_last_frames: Number of frames to drop from the end of each episode.
             shuffle: Whether to shuffle the indices.
+            generator: Generator used for shuffling. Exposing this attribute (even when None) lets
+                       `accelerate` register it as the synchronized RNG in distributed training, so
+                       every rank draws the same permutation and batch shards stay disjoint. When
+                       None, shuffling falls back to the global torch RNG.
         """
         if drop_n_first_frames < 0:
             raise ValueError(f"drop_n_first_frames must be >= 0, got {drop_n_first_frames}")
@@ -73,10 +78,11 @@ class EpisodeAwareSampler:
 
         self.indices = indices
         self.shuffle = shuffle
+        self.generator = generator
 
     def __iter__(self) -> Iterator[int]:
         if self.shuffle:
-            for i in torch.randperm(len(self.indices)):
+            for i in torch.randperm(len(self.indices), generator=self.generator):
                 yield self.indices[i]
         else:
             for i in self.indices:

src/lerobot/policies/factory.py

@@ -57,6 +57,7 @@ from .pretrained import PreTrainedPolicy
 from .smolvla.configuration_smolvla import SmolVLAConfig
 from .tdmpc.configuration_tdmpc import TDMPCConfig
 from .utils import validate_visual_features_consistency
+from .vla_jepa.configuration_vla_jepa import VLAJEPAConfig
 from .vqbet.configuration_vqbet import VQBeTConfig
 from .wall_x.configuration_wall_x import WallXConfig
 from .xvla.configuration_xvla import XVLAConfig
@@ -157,6 +158,10 @@ def get_policy_class(name: str) -> type[PreTrainedPolicy]:
         from .molmoact2.modeling_molmoact2 import MolmoAct2Policy
 
         return MolmoAct2Policy
+    elif name == "vla_jepa":
+        from .vla_jepa.modeling_vla_jepa import VLAJEPAPolicy
+
+        return VLAJEPAPolicy
     else:
         try:
             return _get_policy_cls_from_policy_name(name=name)
@@ -211,6 +216,8 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
         return EO1Config(**kwargs)
     elif policy_type == "molmoact2":
         return MolmoAct2Config(**kwargs)
+    elif policy_type == "vla_jepa":
+        return VLAJEPAConfig(**kwargs)
     else:
         try:
             config_cls = PreTrainedConfig.get_choice_class(policy_type)
@@ -415,6 +422,7 @@ def make_pre_post_processors(
             config=policy_cfg,
             dataset_stats=kwargs.get("dataset_stats"),
         )
+
     elif isinstance(policy_cfg, EO1Config):
         from .eo1.processor_eo1 import make_eo1_pre_post_processors
 
@@ -432,6 +440,14 @@ def make_pre_post_processors(
             dataset_meta=kwargs.get("dataset_meta"),
         )
 
+    elif isinstance(policy_cfg, VLAJEPAConfig):
+        from .vla_jepa.processor_vla_jepa import make_vla_jepa_pre_post_processors
+
+        processors = make_vla_jepa_pre_post_processors(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+        )
+
     else:
         try:
             processors = _make_processors_from_policy_config(

src/lerobot/policies/vla_jepa/README.md

@@ -0,0 +1 @@
+../../../../docs/source/policy_vla_jepa_README.md

src/lerobot/policies/vla_jepa/__init__.py

@@ -0,0 +1,23 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .configuration_vla_jepa import VLAJEPAConfig
+from .modeling_vla_jepa import VLAJEPAPolicy
+from .processor_vla_jepa import make_vla_jepa_pre_post_processors
+
+__all__ = [
+    "VLAJEPAConfig",
+    "VLAJEPAPolicy",
+    "make_vla_jepa_pre_post_processors",
+]

src/lerobot/policies/vla_jepa/action_head.py

@@ -0,0 +1,337 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+from collections import OrderedDict
+from dataclasses import dataclass
+from typing import TYPE_CHECKING
+
+import torch
+import torch.nn.functional as F  # noqa: N812
+from torch import nn
+from torch.distributions import Beta
+
+from lerobot.utils.import_utils import _diffusers_available, require_package
+
+if TYPE_CHECKING or _diffusers_available:
+    from diffusers import ConfigMixin, ModelMixin
+    from diffusers.configuration_utils import register_to_config
+    from diffusers.models.attention import Attention, FeedForward
+    from diffusers.models.embeddings import TimestepEmbedding, Timesteps
+else:
+
+    class ModelMixin:  # type: ignore[no-redef]
+        pass
+
+    class ConfigMixin:  # type: ignore[no-redef]
+        pass
+
+    register_to_config = lambda f: f  # noqa: E731
+    Attention = FeedForward = TimestepEmbedding = Timesteps = None
+
+from .configuration_vla_jepa import VLAJEPAConfig
+
+
+class SinusoidalPositionalEncoding(nn.Module):
+    def __init__(self, embedding_dim: int):
+        super().__init__()
+        self.embedding_dim = embedding_dim
+
+    def forward(self, timesteps: torch.Tensor) -> torch.Tensor:
+        timesteps = timesteps.float()
+        batch_size, seq_len = timesteps.shape
+        half_dim = self.embedding_dim // 2
+        exponent = -torch.arange(half_dim, dtype=torch.float, device=timesteps.device)
+        exponent = exponent * (torch.log(torch.tensor(10000.0, device=timesteps.device)) / max(half_dim, 1))
+        freqs = timesteps.unsqueeze(-1) * exponent.exp()
+        return torch.cat([torch.sin(freqs), torch.cos(freqs)], dim=-1).view(batch_size, seq_len, -1)
+
+
+class ActionEncoder(nn.Module):
+    def __init__(self, action_dim: int, hidden_size: int):
+        super().__init__()
+        self.layer1 = nn.Linear(action_dim, hidden_size)
+        self.layer2 = nn.Linear(hidden_size * 2, hidden_size)
+        self.layer3 = nn.Linear(hidden_size, hidden_size)
+        self.pos_encoding = SinusoidalPositionalEncoding(hidden_size)
+
+    def forward(self, actions: torch.Tensor, timesteps: torch.Tensor) -> torch.Tensor:
+        batch_size, seq_len, _ = actions.shape
+        if timesteps.ndim != 1 or timesteps.shape[0] != batch_size:
+            raise ValueError("timesteps must have shape [batch_size].")
+        timesteps = timesteps.unsqueeze(1).expand(-1, seq_len)
+        action_emb = self.layer1(actions)
+        time_emb = self.pos_encoding(timesteps).to(dtype=action_emb.dtype)
+        return self.layer3(F.silu(self.layer2(torch.cat([action_emb, time_emb], dim=-1))))
+
+
+class TimestepEncoder(nn.Module):
+    def __init__(self, embedding_dim: int):
+        super().__init__()
+        require_package("diffusers", extra="vla_jepa")
+        self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=1)
+        self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim)
+
+    def forward(self, timesteps: torch.Tensor) -> torch.Tensor:
+        projected = self.time_proj(timesteps).to(dtype=next(self.parameters()).dtype)
+        return self.timestep_embedder(projected)
+
+
+class AdaLayerNorm(nn.Module):
+    def __init__(self, embedding_dim: int):
+        super().__init__()
+        self.linear = nn.Linear(embedding_dim, embedding_dim * 2)
+        self.norm = nn.LayerNorm(embedding_dim, eps=1e-5, elementwise_affine=False)
+        self.silu = nn.SiLU()
+
+    def forward(self, x: torch.Tensor, temb: torch.Tensor) -> torch.Tensor:
+        scale, shift = self.linear(self.silu(temb)).chunk(2, dim=-1)
+        return self.norm(x) * (1 + scale[:, None]) + shift[:, None]
+
+
+class BasicTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout: float,
+        cross_attention_dim: int,
+        is_cross_attention: bool = True,
+    ) -> None:
+        super().__init__()
+        self.is_cross_attention = is_cross_attention
+        self.norm1 = AdaLayerNorm(dim)
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=True,
+            cross_attention_dim=cross_attention_dim,
+            out_bias=True,
+        )
+        self.norm2 = nn.LayerNorm(dim, eps=1e-5, elementwise_affine=False)
+        self.ff = FeedForward(dim, dropout=dropout, activation_fn="gelu-approximate", final_dropout=True)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor | None,
+        temb: torch.Tensor,
+    ) -> torch.Tensor:
+        attn_input = self.norm1(hidden_states, temb)
+        attention_context = encoder_hidden_states if self.is_cross_attention else None
+        hidden_states = hidden_states + self.attn1(attn_input, encoder_hidden_states=attention_context)
+        hidden_states = hidden_states + self.ff(self.norm2(hidden_states))
+        return hidden_states
+
+
+class DiT(ModelMixin, ConfigMixin):
+    _supports_gradient_checkpointing = False
+
+    @register_to_config
+    def __init__(
+        self,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        output_dim: int,
+        num_layers: int,
+        dropout: float,
+        cross_attention_dim: int,
+    ) -> None:
+        super().__init__()
+        self.inner_dim = num_attention_heads * attention_head_dim
+        self.timestep_encoder = TimestepEncoder(self.inner_dim)
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    dim=self.inner_dim,
+                    num_attention_heads=num_attention_heads,
+                    attention_head_dim=attention_head_dim,
+                    dropout=dropout,
+                    cross_attention_dim=cross_attention_dim if layer_idx % 2 == 0 else self.inner_dim,
+                    is_cross_attention=layer_idx % 2 == 0,
+                )
+                for layer_idx in range(num_layers)
+            ]
+        )
+        self.norm_out = nn.LayerNorm(self.inner_dim, eps=1e-6, elementwise_affine=False)
+        self.proj_out_1 = nn.Linear(self.inner_dim, self.inner_dim * 2)
+        self.proj_out_2 = nn.Linear(self.inner_dim, output_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        timestep: torch.Tensor,
+    ) -> torch.Tensor:
+        temb = self.timestep_encoder(timestep)
+        x = hidden_states
+        for block in self.transformer_blocks:
+            x = block(x, encoder_hidden_states=encoder_hidden_states, temb=temb)
+        shift, scale = self.proj_out_1(F.silu(temb)).chunk(2, dim=-1)
+        x = self.norm_out(x) * (1 + scale[:, None]) + shift[:, None]
+        return self.proj_out_2(x)
+
+
+@dataclass
+class ActionModelPreset:
+    hidden_size: int
+    attention_head_dim: int
+    num_attention_heads: int
+
+
+DIT_PRESETS = {
+    "DiT-B": ActionModelPreset(hidden_size=768, attention_head_dim=64, num_attention_heads=12),
+    "DiT-L": ActionModelPreset(hidden_size=1536, attention_head_dim=48, num_attention_heads=32),
+    "DiT-test": ActionModelPreset(hidden_size=16, attention_head_dim=8, num_attention_heads=2),
+}
+
+
+class VLAJEPAActionHead(nn.Module):
+    def __init__(self, config: VLAJEPAConfig, cross_attention_dim: int) -> None:
+        super().__init__()
+        preset = DIT_PRESETS[config.action_model_type]
+        self.config = config
+        num_heads = config.action_num_heads or preset.num_attention_heads
+        head_dim = config.action_attention_head_dim or preset.attention_head_dim
+        inner_dim = num_heads * head_dim  # e.g. DiT-B: 12 × 64 = 768
+
+        self.input_embedding_dim = inner_dim
+        self.action_horizon = config.chunk_size
+        self.num_inference_timesteps = config.num_inference_timesteps
+
+        hidden_size = config.action_hidden_size
+        self.model = DiT(
+            num_attention_heads=num_heads,
+            attention_head_dim=head_dim,
+            output_dim=hidden_size,
+            num_layers=config.action_num_layers,
+            dropout=config.action_dropout,
+            cross_attention_dim=cross_attention_dim,
+        )
+        self.action_encoder = ActionEncoder(config.action_dim, inner_dim)
+        self.action_decoder = nn.Sequential(
+            OrderedDict(
+                [
+                    ("layer1", nn.Linear(hidden_size, hidden_size)),
+                    ("relu", nn.ReLU()),
+                    ("layer2", nn.Linear(hidden_size, config.action_dim)),
+                ]
+            )
+        )
+        self.state_encoder = (
+            nn.Sequential(
+                OrderedDict(
+                    [
+                        ("layer1", nn.Linear(config.state_dim, hidden_size)),
+                        ("relu", nn.ReLU()),
+                        ("layer2", nn.Linear(hidden_size, inner_dim)),
+                    ]
+                )
+            )
+            if config.state_dim > 0
+            else None
+        )
+        self.future_tokens = nn.Embedding(config.num_embodied_action_tokens_per_instruction, inner_dim)
+        self.position_embedding = nn.Embedding(
+            max(1024, config.chunk_size + config.num_action_tokens_per_timestep + 4),
+            inner_dim,
+        )
+        self.beta_dist = Beta(config.action_noise_beta_alpha, config.action_noise_beta_beta)
+
+    def sample_time(self, batch_size: int, device: torch.device, dtype: torch.dtype) -> torch.Tensor:
+        sample = self.beta_dist.sample([batch_size]).to(device=device, dtype=dtype)
+        return (self.config.action_noise_s - sample) / self.config.action_noise_s
+
+    def _build_inputs(
+        self,
+        conditioning_tokens: torch.Tensor,
+        actions: torch.Tensor,
+        state: torch.Tensor | None,
+        timesteps: torch.Tensor,
+    ) -> torch.Tensor:
+        action_features = self.action_encoder(actions, timesteps)
+        pos_ids = torch.arange(action_features.shape[1], device=actions.device)
+        action_features = action_features + self.position_embedding(pos_ids)[None]
+
+        future_tokens = self.future_tokens.weight.unsqueeze(0).expand(actions.shape[0], -1, -1)
+        seq = [future_tokens, action_features]
+        if state is not None and self.state_encoder is not None:
+            if state.ndim == 2:
+                state = state.unsqueeze(1)
+            seq.insert(0, self.state_encoder(state))
+        return torch.cat(seq, dim=1)
+
+    def forward(
+        self,
+        conditioning_tokens: torch.Tensor,
+        actions: torch.Tensor,
+        state: torch.Tensor | None = None,
+        action_is_pad: torch.Tensor | None = None,
+    ) -> torch.Tensor:
+        noise = torch.randn_like(actions)
+        t = self.sample_time(actions.shape[0], actions.device, actions.dtype)
+        noisy_actions = (1 - t[:, None, None]) * noise + t[:, None, None] * actions
+        velocity = actions - noise
+        t_discretized = (t * self.config.action_num_timestep_buckets).long()
+
+        hidden_states = self._build_inputs(conditioning_tokens, noisy_actions, state, t_discretized)
+        pred = self.model(
+            hidden_states=hidden_states,
+            encoder_hidden_states=conditioning_tokens,
+            timestep=t_discretized,
+        )
+        pred_actions = self.action_decoder(pred[:, -actions.shape[1] :])
+
+        if action_is_pad is None:
+            action_is_pad = torch.zeros(actions.shape[:2], dtype=torch.bool, device=actions.device)
+
+        loss = F.mse_loss(pred_actions, velocity, reduction="none")  # [B, T, action_dim]
+        valid_mask = ~action_is_pad.unsqueeze(-1)  # [B, T, 1]
+        num_valid = valid_mask.sum() * loss.shape[-1]
+        return (loss * valid_mask).sum() / num_valid.clamp_min(1)
+
+    @torch.no_grad()
+    def predict_action(
+        self,
+        conditioning_tokens: torch.Tensor,
+        state: torch.Tensor | None = None,
+    ) -> torch.Tensor:
+        batch_size = conditioning_tokens.shape[0]
+        actions = torch.randn(
+            batch_size,
+            self.action_horizon,
+            self.config.action_dim,
+            dtype=conditioning_tokens.dtype,
+            device=conditioning_tokens.device,
+        )
+        dt = 1.0 / max(self.num_inference_timesteps, 1)
+        for step in range(self.num_inference_timesteps):
+            t_cont = step / float(max(self.num_inference_timesteps, 1))
+            t_value = int(t_cont * self.config.action_num_timestep_buckets)
+            timesteps = torch.full(
+                (batch_size,), t_value, device=conditioning_tokens.device, dtype=torch.long
+            )
+            hidden_states = self._build_inputs(conditioning_tokens, actions, state, timesteps)
+            pred = self.model(
+                hidden_states=hidden_states,
+                encoder_hidden_states=conditioning_tokens,
+                timestep=timesteps,
+            )
+            pred_velocity = self.action_decoder(pred[:, -self.action_horizon :])
+            actions = actions + dt * pred_velocity
+        return actions

src/lerobot/policies/vla_jepa/configuration_vla_jepa.py

@@ -0,0 +1,154 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import NormalizationMode
+from lerobot.optim.optimizers import AdamWConfig
+from lerobot.optim.schedulers import CosineDecayWithWarmupSchedulerConfig
+
+
+@PreTrainedConfig.register_subclass("vla_jepa")
+@dataclass
+class VLAJEPAConfig(PreTrainedConfig):
+    n_obs_steps: int = 1
+    chunk_size: int = 7
+    n_action_steps: int = 7
+
+    normalization_mapping: dict[str, NormalizationMode] = field(
+        default_factory=lambda: {
+            "VISUAL": NormalizationMode.IDENTITY,
+            "STATE": NormalizationMode.MEAN_STD,
+            "ACTION": NormalizationMode.MIN_MAX,
+        }
+    )
+
+    qwen_model_name: str = "Qwen/Qwen3-VL-2B-Instruct"
+    jepa_encoder_name: str = "facebook/vjepa2-vitl-fpc64-256"
+    freeze_qwen: bool = False
+    enable_world_model: bool = True
+    # Enables cross-embodiment transfer: when fine-tuning a pretrained model on a robot with a
+    # different action or state dimensionality, the input/output projection layers must be
+    # re-initialised from scratch while the rest of the network keeps its pretrained weights.
+    # List the key prefixes that are allowed to have shape mismatches; anything else raises an error.
+    # e.g. ["model.action_model.action_encoder", "model.action_model.state_encoder"]
+    reinit_modules: list[str] | None = None
+
+    tokenizer_padding_side: str = "left"
+    prompt_template: str = "Your task is {instruction}. Infer the temporal dynamics from frames {actions} and produce the corresponding policy actions {e_actions}."
+    special_action_token: str = "<|action_{}|>"
+    embodied_action_token: str = "<|embodied_action|>"
+
+    action_dim: int = 7
+    state_dim: int = 8
+
+    num_action_tokens_per_timestep: int = 8
+    num_embodied_action_tokens_per_instruction: int = 32
+    num_inference_timesteps: int = 4
+
+    action_hidden_size: int = 1024
+    action_model_type: str = "DiT-B"
+    action_num_layers: int = 16
+    action_num_heads: int | None = None
+    action_attention_head_dim: int | None = None
+    action_dropout: float = 0.2
+    action_num_timestep_buckets: int = 1000
+    action_noise_beta_alpha: float = 1.5
+    action_noise_beta_beta: float = 1.0
+    action_noise_s: float = 0.999
+    num_target_vision_tokens: int = 32
+    action_max_seq_len: int = 1024
+
+    # total video frames loaded per sample
+    num_video_frames: int = 8
+    predictor_depth: int = 12
+    predictor_num_heads: int = 8
+    predictor_mlp_ratio: float = 4.0
+    predictor_dropout: float = 0.0
+    world_model_loss_weight: float = 0.1
+    jepa_tubelet_size: int = 2  # must match the encoder (e.g. 2 for vjepa2-vitl-fpc64-256)
+    repeated_diffusion_steps: int = 8  # independent noise draws per batch item (CogACT-style)
+
+    resize_images_to: tuple[int, int] | None = None
+    binarize_gripper_action: bool = True
+    pre_snap_gripper_action: bool = True
+    clip_normalized_actions: bool = True
+    gripper_dim: int = 6
+    gripper_threshold: float = 0.5
+    torch_dtype: str = "bfloat16"
+
+    optimizer_lr: float = 1e-4
+    optimizer_betas: tuple[float, float] = (0.9, 0.95)
+    optimizer_eps: float = 1e-8
+    optimizer_weight_decay: float = 1e-10
+    optimizer_grad_clip_norm: float = 10.0
+    scheduler_warmup_steps: int = 1_000
+    scheduler_decay_steps: int = 30_000
+    scheduler_decay_lr: float = 2.5e-6
+
+    def __post_init__(self) -> None:
+        super().__post_init__()
+        if self.freeze_qwen and self.enable_world_model:
+            # freezing qwen backbone makes world model training irrelevant since no grad flows
+            self.enable_world_model = False
+        if self.n_action_steps > self.chunk_size:
+            raise ValueError("`n_action_steps` must be <= `chunk_size`.")
+        if self.num_video_frames < 2 * self.jepa_tubelet_size:
+            raise ValueError(
+                f"`video_horizon` ({self.num_video_frames}) must be >= 2 * `jepa_tubelet_size` "
+                f"({self.jepa_tubelet_size}) to have at least one context and one GT temporal position."
+            )
+
+    def validate_features(self) -> None:
+        if not self.image_features:
+            raise ValueError("VLAJEPA requires at least one visual input feature.")
+        if self.action_feature is None:
+            raise ValueError("VLAJEPA requires an action output feature.")
+        self.action_dim = self.action_feature.shape[0]
+        if self.robot_state_feature is not None:
+            self.state_dim = self.robot_state_feature.shape[0]
+
+    def get_optimizer_preset(self) -> AdamWConfig:
+        return AdamWConfig(
+            lr=self.optimizer_lr,
+            betas=self.optimizer_betas,
+            eps=self.optimizer_eps,
+            weight_decay=self.optimizer_weight_decay,
+            grad_clip_norm=self.optimizer_grad_clip_norm,
+        )
+
+    def get_scheduler_preset(self) -> CosineDecayWithWarmupSchedulerConfig:
+        return CosineDecayWithWarmupSchedulerConfig(
+            peak_lr=self.optimizer_lr,
+            decay_lr=self.scheduler_decay_lr,
+            num_warmup_steps=self.scheduler_warmup_steps,
+            num_decay_steps=self.scheduler_decay_steps,
+        )
+
+    @property
+    def observation_delta_indices(self) -> list[int]:
+        # load video_horizon frames starting from current timestep: [t, t+1, ..., t+video_horizon-1]
+        # matches original repo's observation_indices=list(range(video_horizon))
+        return list(range(self.num_video_frames))
+
+    @property
+    def action_delta_indices(self) -> list[int]:
+        return list(range(self.chunk_size))
+
+    @property
+    def reward_delta_indices(self) -> None:
+        return None

src/lerobot/policies/vla_jepa/modeling_vla_jepa.py

@@ -0,0 +1,629 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+import logging
+from collections import deque
+from pathlib import Path
+from typing import TYPE_CHECKING
+
+import numpy as np
+import torch
+import torch.nn.functional as F  # noqa: N812
+from PIL import Image
+from torch import Tensor, nn
+
+from lerobot.policies.pretrained import PreTrainedPolicy, T
+from lerobot.policies.utils import populate_queues
+from lerobot.utils.constants import ACTION, OBS_STATE
+from lerobot.utils.import_utils import _transformers_available, require_package
+
+if TYPE_CHECKING or _transformers_available:
+    from transformers import AutoModel, AutoVideoProcessor
+else:
+    AutoModel = None
+    AutoVideoProcessor = None
+
+from .action_head import VLAJEPAActionHead
+from .configuration_vla_jepa import VLAJEPAConfig
+from .qwen_interface import Qwen3VLInterface
+from .world_model import ActionConditionedVideoPredictor
+
+# ============================================================================
+# Native VLA-JEPA Model - follows original starVLA VLA_JEPA.py implementation
+# ============================================================================
+
+
+class VLAJEPAModel(nn.Module):
+    """
+    Native VLA-JEPA model following the original starVLA VLA_JEPA.py.
+
+    Components:
+      - Qwen3-VL: vision-language backbone for fused embeddings
+      - DiT-B: flow-matching action head for future action prediction
+      - V-JEPA: world model for video frame prediction
+
+    Input: List[dict] native format (same as original starVLA)
+      - "image": List[PIL.Image] (multi-view images)
+      - "video": np.ndarray [V, T, H, W, 3]
+      - "lang": str (task instruction)
+      - "action": np.ndarray [T, action_dim] (optional, training only)
+      - "state": np.ndarray [1, state_dim] (optional)
+    """
+
+    def __init__(self, config: VLAJEPAConfig) -> None:
+        super().__init__()
+        require_package("transformers", extra="vla_jepa")
+        self.config = config
+
+        # Vision-language backbone
+        self.qwen = Qwen3VLInterface(config)
+
+        # Tokenizer expansion for special action tokens
+        self.action_tokens, self.action_token_ids, self.embodied_action_token_id = (
+            self.qwen.expand_tokenizer()
+        )
+
+        # Action head (flow-matching DiT)
+        self.action_model = VLAJEPAActionHead(config, cross_attention_dim=self.qwen.model.config.hidden_size)
+
+        # JEPA world model components
+        if config.enable_world_model:
+            self.video_encoder = AutoModel.from_pretrained(
+                config.jepa_encoder_name,
+                torch_dtype=self.qwen._get_torch_dtype(config.torch_dtype),
+            )
+            self.video_processor = AutoVideoProcessor.from_pretrained(config.jepa_encoder_name)
+            num_views = config.jepa_tubelet_size
+            tubelet_size = self.video_encoder.config.tubelet_size
+            image_size = getattr(self.video_encoder.config, "image_size", None)
+            if image_size is None:
+                first_image_shape = next(iter(config.image_features.values())).shape
+                image_size = first_image_shape[-1]
+            self.video_predictor = ActionConditionedVideoPredictor(
+                num_frames=config.num_video_frames // tubelet_size,
+                img_size=(image_size, image_size),
+                patch_size=16,
+                tubelet_size=1,
+                embed_dim=self.video_encoder.config.hidden_size * num_views,
+                action_embed_dim=self.qwen.model.config.hidden_size,
+                predictor_embed_dim=self.video_encoder.config.hidden_size,
+                depth=config.predictor_depth,
+                num_heads=config.predictor_num_heads,
+                mlp_ratio=config.predictor_mlp_ratio,
+                num_action_tokens_per_step=config.num_action_tokens_per_timestep,
+            )
+        else:
+            self.video_encoder = None
+            self.video_processor = None
+            self.video_predictor = None
+
+        if config.freeze_qwen:
+            self.qwen.requires_grad_(False)
+
+        # Build prompt placeholders.
+        # Use the encoder's actual tubelet_size when available (world model enabled),
+        # otherwise fall back to config.
+        _tubelet_size = (
+            self.video_encoder.config.tubelet_size
+            if config.enable_world_model
+            else self.config.jepa_tubelet_size
+        )
+        num_action_prompt_steps = self.config.num_video_frames // _tubelet_size - 1
+        self.replace_prompt = "".join(
+            token * self.config.num_action_tokens_per_timestep
+            for token in self.action_tokens[:num_action_prompt_steps]
+        )
+        self.embodied_replace_prompt = (
+            self.config.embodied_action_token * self.config.num_embodied_action_tokens_per_instruction
+        )
+
+    def _qwen_last_decoder_hidden(self, qwen_inputs: dict[str, torch.Tensor]) -> torch.Tensor:
+        """Return the last decoder hidden state before the final RMSNorm.
+
+        The model was trained with the output of the last transformer block BEFORE
+        the final RMSNorm. In transformers 5.x, `hidden_states[-1]` from
+        `output_hidden_states=True` is post-norm (tied to `last_hidden_state` via
+        `@capture_outputs`). A forward hook on `language_model.layers[-1]` recovers
+        the correct pre-RMSNorm state, matching the training-time representation.
+        """
+        captured: list[torch.Tensor] = []
+
+        def _hook(module, input, output):
+            h = output[0] if isinstance(output, tuple) else output
+            captured.append(h)
+
+        last_layer = self.qwen.model.model.language_model.layers[-1]
+        handle = last_layer.register_forward_hook(_hook)
+        try:
+            self.qwen.model(
+                **qwen_inputs,
+                output_hidden_states=False,
+                output_attentions=False,
+                return_dict=True,
+            )
+        finally:
+            handle.remove()
+
+        return captured[0]  # [B, seq_len, H]
+
+    # ---- Native VLA-JEPA forward (follows original VLA_JEPA.py) ----
+
+    def forward(self, examples: list[dict]) -> dict[str, Tensor]:
+        """
+        Native forward pass following original starVLA VLA_JEPA.forward.
+
+        Args:
+            examples: List of per-sample dicts with keys:
+                "image"  : List[PIL.Image]  — multi-view images
+                "video"  : np.ndarray [V, T, H, W, 3]
+                "lang"   : str — task instruction
+                "action" : np.ndarray [T, action_dim] (optional)
+                "state"  : np.ndarray [1, state_dim] (optional)
+
+        Returns:
+            dict with "action_loss" and "wm_loss" keys (scalar Tensors).
+        """
+        # Unpack native format (same pattern as original VLA_JEPA.py)
+        batch_images = [ex["image"] for ex in examples]  # List[List[PIL.Image]]
+        batch_videos = [ex["video"] for ex in examples]  # List[np.ndarray]
+        instructions = [ex["lang"] for ex in examples]  # List[str]
+        has_action = "action" in examples[0] and examples[0]["action"] is not None
+        actions = [ex["action"] for ex in examples] if has_action else None
+        has_state = "state" in examples[0] and examples[0]["state"] is not None
+        state = [ex["state"] for ex in examples] if has_state else None
+        action_is_pad = (
+            [ex["action_is_pad"] for ex in examples]
+            if has_action and "action_is_pad" in examples[0] and examples[0]["action_is_pad"] is not None
+            else None
+        )
+
+        # Stack videos: [B, V, T, H, W, 3] -> [B, V, T, 3, H, W]
+        batch_videos = np.stack(batch_videos)
+        batch_videos = batch_videos.transpose(0, 1, 2, 5, 3, 4)  # [B, V, T, 3, H, W]
+
+        # Adjust number of views for the world model:
+        # - fewer views than expected: duplicate the first view to fill up
+        # - more views than expected: keep only the first num_views_world_model views
+        num_views_world_model = self.config.jepa_tubelet_size
+        if batch_videos.shape[1] < num_views_world_model:
+            num_missing_views = num_views_world_model - batch_videos.shape[1]
+            first_view = np.repeat(batch_videos[:, :1], num_missing_views, axis=1)
+            batch_videos = np.concatenate([batch_videos, first_view], axis=1)
+        elif batch_videos.shape[1] > num_views_world_model:
+            batch_videos = batch_videos[:, :num_views_world_model]
+
+        # ---- Step 1: QwenVL encode (same as original) ----
+        qwen_inputs = self.qwen.build_inputs(
+            images=batch_images,
+            instructions=instructions,
+            action_prompt=self.replace_prompt,
+            embodied_prompt=self.embodied_replace_prompt,
+        )
+
+        # Locate embodied-action tokens (always needed for action head)
+        embodied_mask = qwen_inputs["input_ids"] == self.embodied_action_token_id
+        embodied_indices = embodied_mask.nonzero(as_tuple=True)
+
+        # Locate action tokens (only needed for world model predictor)
+        if self.config.enable_world_model:
+            action_mask = torch.isin(
+                qwen_inputs["input_ids"],
+                torch.tensor(self.action_token_ids, device=qwen_inputs["input_ids"].device),
+            )
+            action_indices = action_mask.nonzero(as_tuple=True)
+
+        device_type = next(self.parameters()).device.type
+
+        with torch.autocast(device_type=device_type, dtype=torch.bfloat16):
+            last_hidden = self._qwen_last_decoder_hidden(qwen_inputs)  # [B, seq_len, H]
+            b, _, h = last_hidden.shape
+
+            if self.config.enable_world_model:
+                action_tokens = last_hidden[action_indices[0], action_indices[1], :].view(b, -1, h)
+
+            embodied_action_tokens = last_hidden[embodied_indices[0], embodied_indices[1], :].view(b, -1, h)
+
+        # ---- Step 2+3: JEPA Encoder + Predictor ----
+        device_wm = last_hidden.device
+        if not self.config.enable_world_model:
+            wm_loss = torch.tensor(0.0, device=device_wm)
+        else:
+            b, v, t_frames, c, h_img, w_img = batch_videos.shape
+            batch_videos_flat = batch_videos.reshape(b * v, t_frames, c, h_img, w_img)
+
+            video_pixels = self.video_processor(videos=list(batch_videos_flat), return_tensors="pt")[
+                "pixel_values_videos"
+            ].to(self.video_encoder.device)  # [B*V, T, C, H, W]
+
+            with torch.no_grad():
+                video_embeddings = self.video_encoder.get_vision_features(pixel_values_videos=video_pixels)
+                # Merge views: [B*V, ...] -> [B, ..., V*embed_dim]
+                video_embeddings = torch.cat(torch.chunk(video_embeddings, chunks=v, dim=0), dim=2)
+
+            tubelet_size = self.video_encoder.config.tubelet_size
+            device_wm = video_embeddings.device
+            # num_video_frames raw frames → t_enc_total temporal positions after tubelet compression
+            t_enc_total = self.config.num_video_frames // tubelet_size
+
+            if t_enc_total < 2:
+                wm_loss = torch.tensor(0.0, device=device_wm)
+            else:
+                # Shift-by-one JEPA split (matches original VLA_JEPA.py lines 231-232):
+                # input_states: positions 0..T-2, gt_states: positions 1..T-1
+                t_enc_ctx = t_enc_total - 1
+                tokens_per_frame = video_embeddings.shape[1] // t_enc_total
+
+                input_states = video_embeddings[:, : tokens_per_frame * t_enc_ctx, :]
+                gt_states = video_embeddings[:, tokens_per_frame:, :]
+
+                expected_actions = t_enc_ctx * self.config.num_action_tokens_per_timestep
+                if action_tokens.shape[1] < expected_actions:
+                    pad = action_tokens[:, -1:].repeat(1, expected_actions - action_tokens.shape[1], 1)
+                    action_tokens = torch.cat([action_tokens, pad], dim=1)
+
+                predicted_states = self.video_predictor(
+                    input_states.float(),
+                    action_tokens[:, :expected_actions].float(),
+                )
+
+                wm_loss = F.l1_loss(predicted_states, gt_states.float(), reduction="mean")
+
+        if not has_action:
+            return {"wm_loss": wm_loss}
+
+        # ---- Step 4: Action Head ----
+        with torch.autocast(device_type=device_type, dtype=torch.float32):
+            actions_tensor = torch.tensor(
+                np.array(actions), device=last_hidden.device, dtype=torch.float32
+            )  # [B, T_full, action_dim]
+            action_horizon = self.config.chunk_size
+            actions_target = actions_tensor[:, -action_horizon:, :]
+
+            state_tensor = None
+            if state is not None:
+                state_tensor = torch.tensor(
+                    np.array(state), device=last_hidden.device, dtype=last_hidden.dtype
+                )  # [B, 1, state_dim]
+
+            repeated_diffusion_steps = self.config.repeated_diffusion_steps
+            actions_target = actions_target.repeat(repeated_diffusion_steps, 1, 1)
+            embodied_action_tokens = embodied_action_tokens.repeat(repeated_diffusion_steps, 1, 1)
+            if state_tensor is not None:
+                state_tensor = state_tensor.repeat(repeated_diffusion_steps, 1, 1)
+
+            action_is_pad_rep = None
+            if action_is_pad is not None:
+                pad_tensor = torch.stack(
+                    [
+                        p.to(actions_target.device)
+                        if isinstance(p, Tensor)
+                        else torch.tensor(p, device=actions_target.device)
+                        for p in action_is_pad
+                    ]
+                )  # [B, T_full]
+                pad_tensor = pad_tensor[:, -action_horizon:]  # [B, action_horizon]
+                action_is_pad_rep = pad_tensor.repeat(repeated_diffusion_steps, 1)  # [B*R, action_horizon]
+
+            action_loss = self.action_model(
+                embodied_action_tokens, actions_target, state_tensor, action_is_pad_rep
+            )
+
+        return {"action_loss": action_loss, "wm_loss": wm_loss * self.config.world_model_loss_weight}
+
+    # ---- Native predict_action (follows original VLA_JEPA.predict_action) ----
+
+    @torch.no_grad()
+    def predict_action(
+        self,
+        batch_images: list[list[Image.Image]],
+        instructions: list[str],
+        state: np.ndarray | None = None,
+    ) -> np.ndarray:
+        """
+        Native action prediction following original VLA_JEPA.predict_action.
+
+        Args:
+            batch_images: List of samples; each is List[PIL.Image] (multi-view).
+            instructions: Task instructions, one per sample.
+            state: Optional [B, state_dim] numpy array.
+
+        Returns:
+            np.ndarray [B, action_horizon, action_dim] — predicted actions.
+        """
+        if self.config.resize_images_to is not None:
+            height, width = self.config.resize_images_to
+            resampling = getattr(Image, "Resampling", Image).BOX
+            batch_images = [
+                [image.resize((width, height), resample=resampling) for image in sample_images]
+                for sample_images in batch_images
+            ]
+
+        qwen_inputs = self.qwen.build_inputs(
+            images=batch_images,
+            instructions=instructions,
+            action_prompt=self.replace_prompt,
+            embodied_prompt=self.embodied_replace_prompt,
+        )
+
+        embodied_mask = qwen_inputs["input_ids"] == self.embodied_action_token_id
+        embodied_indices = embodied_mask.nonzero(as_tuple=True)
+
+        device_type = next(self.parameters()).device.type
+
+        with torch.autocast(device_type=device_type, dtype=torch.bfloat16):
+            last_hidden = self._qwen_last_decoder_hidden(qwen_inputs)  # [B, seq_len, H]
+            b, _, h = last_hidden.shape
+            embodied_action_tokens = last_hidden[embodied_indices[0], embodied_indices[1], :].view(b, -1, h)
+
+        state_tensor = None
+        if state is not None:
+            state_tensor = torch.from_numpy(np.array(state)).to(
+                device=last_hidden.device, dtype=last_hidden.dtype
+            )
+
+        pred_actions = self.action_model.predict_action(
+            embodied_action_tokens.float(), state_tensor.float() if state_tensor is not None else None
+        )  # [B, action_horizon, action_dim]
+
+        return pred_actions.detach().cpu().numpy()
+
+
+# ============================================================================
+# LeRobot Adapter Layer - converts between LeRobot batch format and native VLA-JEPA format
+# ============================================================================
+
+
+class VLAJEPAPolicy(PreTrainedPolicy):
+    """
+    LeRobot adapter for VLA-JEPA.
+
+    Converts LeRobot's standard batch format (dict[str, Tensor]) to the native
+    VLA-JEPA format (List[dict]), calls the native model, and converts outputs
+    back to LeRobot format.
+    """
+
+    config_class = VLAJEPAConfig
+    name = "vla_jepa"
+
+    def __init__(self, config: VLAJEPAConfig, **kwargs) -> None:
+        super().__init__(config)
+        config.validate_features()
+        if dataset_meta := kwargs.get("dataset_meta"):
+            # cfg.input_features keeps the pretrained model's feature keys (needed for rename_map
+            # compatibility), so validate_features() may have read stale dims from a pretrained
+            # config. Override state_dim/action_dim from the actual dataset being used.
+            ds_features = dataset_meta.features
+            if OBS_STATE in ds_features:
+                config.state_dim = ds_features[OBS_STATE]["shape"][0]
+            if ACTION in ds_features:
+                config.action_dim = ds_features[ACTION]["shape"][0]
+
+        self.model = VLAJEPAModel(config)
+        self.reset()
+
+    def reset(self) -> None:
+        self._queues = {ACTION: deque(maxlen=self.config.n_action_steps)}
+
+    # ---- Format Conversion: LeRobot → Native ----
+
+    def _prepare_model_inputs(self, batch: dict[str, Tensor]) -> list[dict]:
+        """
+        Convert LeRobot batch format to native VLA-JEPA examples format.
+
+        LeRobot format:
+            batch = {
+                "observation.images.<key>": Tensor [B, C, H, W] or [B, T, C, H, W],
+                "observation.state": Tensor [B, state_dim] or [B, T, state_dim],
+                "action": Tensor [B, chunk_size, action_dim],  (training only)
+                "task": str | List[str],  (optional instruction)
+            }
+
+        Native format (List[dict]):
+            {
+                "image": List[PIL.Image],       # multi-view images per sample
+                "video": np.ndarray [V, T, H, W, 3],
+                "lang": str,                     # task instruction
+                "action": np.ndarray [T, action_dim],  # optional
+                "state": np.ndarray [1, state_dim],    # optional
+            }
+        """
+        # Determine batch size from the first image feature
+        image_keys = list(self.config.image_features.keys())
+        if not image_keys:
+            raise ValueError("VLAJEPA requires at least one image feature.")
+        first_key = image_keys[0]
+        first_tensor = batch[first_key]
+        batch_size = first_tensor.shape[0]
+
+        # ---- Collect images per sample ----
+        # images_per_sample[b][v] = PIL.Image for view v
+        images_per_sample: list[list[Image.Image]] = [[] for _ in range(batch_size)]
+        for key in image_keys:
+            tensor = batch[key]  # [B, C, H, W] or [B, T, C, H, W]
+            if tensor.ndim == 5:
+                # observation_delta_indices = [0, 1, ..., num_video_frames-1]
+                # index 0 is the current observation (delta=0)
+                tensor = tensor[:, 0]
+            for b in range(batch_size):
+                images_per_sample[b].append(self.model.qwen.tensor_to_pil(tensor[b]))
+
+        # ---- Collect videos per sample ----
+        # Build video arrays: for each sample, stack views as [V, T, H, W, 3]
+        # Check whether any image feature has a time dimension
+        video_source = None
+        for k in image_keys:
+            if k in batch:
+                video_source = batch[k]  # Use first available for shape inspection
+                break
+
+        if video_source is None:
+            raise ValueError("No image data found in batch for video construction.")
+
+        videos_per_sample = []
+        for b in range(batch_size):
+            sample_views = []
+            for k in image_keys:
+                t = batch[k][b]  # [C, H, W] or [T, C, H, W]
+                if t.ndim == 3:
+                    t = t.unsqueeze(0)  # [1, C, H, W]
+                # Convert to [T, H, W, 3] numpy
+                t_np = t.permute(0, 2, 3, 1).detach().cpu().float().numpy()
+                # Clamp to [0, 255]
+                if t_np.max() <= 1.0:
+                    t_np = t_np * 255.0
+                t_np = np.rint(t_np.clip(0, 255)).astype(np.uint8)
+                sample_views.append(t_np)
+            # Stack views: [V, T, H, W, 3]
+            videos_per_sample.append(np.stack(sample_views, axis=0))
+
+        # ---- Collect instructions ----
+        tasks = batch.get("task")
+        if tasks is None:
+            instructions = ["Execute the robot action."] * batch_size
+        elif isinstance(tasks, str):
+            instructions = [tasks] * batch_size
+        else:
+            instructions = list(tasks)
+
+        # ---- Collect actions (training only) ----
+        actions_list = None
+        action_is_pad_list = None
+        actions_tensor = batch.get(ACTION)
+        if actions_tensor is not None:
+            if actions_tensor.ndim == 2:
+                actions_tensor = actions_tensor.unsqueeze(1)
+            actions_list = [actions_tensor[b].detach().cpu().float().numpy() for b in range(batch_size)]
+            action_is_pad_tensor = batch.get("action_is_pad")
+            if action_is_pad_tensor is not None:
+                action_is_pad_list = [action_is_pad_tensor[b].detach().cpu() for b in range(batch_size)]
+
+        # ---- Collect state ----
+        state_list = None
+        state_tensor = batch.get(OBS_STATE)
+        if state_tensor is not None:
+            if state_tensor.ndim > 2:
+                state_tensor = state_tensor[:, -1, :]
+            if state_tensor.ndim == 2:
+                state_tensor = state_tensor.unsqueeze(1)  # [B, 1, state_dim]
+            state_list = [state_tensor[b].detach().cpu().float().numpy() for b in range(batch_size)]
+
+        # ---- Assemble native examples ----
+        examples = []
+        for b in range(batch_size):
+            example = {
+                "image": images_per_sample[b],
+                "video": videos_per_sample[b],
+                "lang": instructions[b],
+            }
+            if actions_list is not None:
+                example["action"] = actions_list[b]
+            if action_is_pad_list is not None:
+                example["action_is_pad"] = action_is_pad_list[b]
+            if state_list is not None:
+                example["state"] = state_list[b]
+            examples.append(example)
+
+        return examples
+
+    # ---- LeRobot Policy Interface ----
+
+    def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict]:
+        """LeRobot train forward: convert → native forward → aggregate losses."""
+        examples = self._prepare_model_inputs(batch)
+        native_output = self.model.forward(examples)
+
+        ref = next(iter(native_output.values()))
+        zero = torch.zeros((), device=ref.device, dtype=ref.dtype)
+        total_loss = native_output.get("action_loss", zero) + native_output.get("wm_loss", zero)
+        logs = {k: v.detach().item() for k, v in native_output.items()}
+        logs["loss"] = total_loss.detach().item()
+        return total_loss, logs
+
+    def get_optim_params(self) -> dict:
+        return self.model.parameters()
+
+    @torch.no_grad()
+    def predict_action_chunk(self, batch: dict[str, Tensor], noise: Tensor | None = None) -> Tensor:
+        """LeRobot inference: convert → native predict → return as Tensor."""
+        self.eval()
+        self._queues = populate_queues(self._queues, batch, exclude_keys=[ACTION])
+
+        examples = self._prepare_model_inputs(batch)
+        batch_images = [ex["image"] for ex in examples]
+        instructions = [ex["lang"] for ex in examples]
+
+        state_np = None
+        if "state" in examples[0] and examples[0]["state"] is not None:
+            state_np = np.stack([ex["state"] for ex in examples])
+
+        actions_np = self.model.predict_action(batch_images, instructions, state_np)
+        return torch.from_numpy(actions_np).to(device=self.config.device, dtype=torch.float32)
+
+    @torch.no_grad()
+    def select_action(self, batch: dict[str, Tensor], noise: Tensor | None = None) -> Tensor:
+        """LeRobot select_action with action queue caching."""
+        self.eval()
+        self._queues = populate_queues(self._queues, batch, exclude_keys=[ACTION])
+        if len(self._queues[ACTION]) == 0:
+            actions = self.predict_action_chunk(batch)
+            self._queues[ACTION].extend(actions.transpose(0, 1)[: self.config.n_action_steps])
+        return self._queues[ACTION].popleft()
+
+    @classmethod
+    def from_pretrained(
+        cls: type[T],
+        pretrained_name_or_path: str | Path,
+        **kwargs,
+    ):
+        return super().from_pretrained(pretrained_name_or_path, **kwargs)
+
+    @classmethod
+    def _load_as_safetensor(cls, model: T, model_file: str, map_location: str, strict: bool) -> T:
+        reinit_prefixes = model.config.reinit_modules
+        if not reinit_prefixes:
+            return super()._load_as_safetensor(model, model_file, map_location, strict)
+
+        from safetensors.torch import load_file
+
+        state_dict = load_file(model_file, device=map_location)
+        current = model.state_dict()
+
+        reinitialized: list[str] = []
+        filtered: dict = {}
+        for key, value in state_dict.items():
+            if key in current and value.shape != current[key].shape:
+                if not any(key.startswith(p) for p in reinit_prefixes):
+                    raise ValueError(
+                        f"Shape mismatch for '{key}' (checkpoint {tuple(value.shape)} vs model "
+                        f"{tuple(current[key].shape)}) and its prefix is not in `reinit_modules`."
+                    )
+                reinitialized.append(
+                    f"{key}: checkpoint {tuple(value.shape)} → model {tuple(current[key].shape)}"
+                )
+            else:
+                filtered[key] = value
+
+        if reinitialized:
+            logging.warning(
+                f"reinit_modules: skipping {len(reinitialized)} tensor(s) with mismatched shapes "
+                f"(randomly re-initialised):\n  " + "\n  ".join(reinitialized)
+            )
+
+        from lerobot.policies.utils import log_model_loading_keys
+
+        missing_keys, unexpected_keys = model.load_state_dict(filtered, strict=False)
+        log_model_loading_keys(missing_keys, unexpected_keys)
+        return model

src/lerobot/policies/vla_jepa/processor_vla_jepa.py

@@ -0,0 +1,155 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+from typing import Any
+
+import torch
+
+from lerobot.policies.vla_jepa.configuration_vla_jepa import VLAJEPAConfig
+from lerobot.processor import (
+    AddBatchDimensionProcessorStep,
+    DeviceProcessorStep,
+    EnvTransition,
+    NormalizerProcessorStep,
+    PolicyAction,
+    PolicyProcessorPipeline,
+    ProcessorStep,
+    ProcessorStepRegistry,
+    RenameObservationsProcessorStep,
+    TransitionKey,
+    UnnormalizerProcessorStep,
+)
+from lerobot.processor.converters import policy_action_to_transition, transition_to_policy_action
+from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
+
+
+@ProcessorStepRegistry.register(name="vla_jepa_clip_actions")
+class ClipActionsProcessorStep(ProcessorStep):
+    """Clips action tensor to [-1, 1] before unnormalization."""
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        action = transition.get(TransitionKey.ACTION)
+        if action is not None:
+            transition = dict(transition)
+            transition[TransitionKey.ACTION] = action.clamp(-1.0, 1.0)
+        return transition
+
+    def transform_features(self, features):
+        return features
+
+
+@ProcessorStepRegistry.register(name="vla_jepa_pre_snap_gripper")
+class PreSnapGripperProcessorStep(ProcessorStep):
+    """Snaps a gripper dimension to {0, 1} BEFORE unnormalization.
+
+    Mirrors the original starVLA LIBERO eval:
+      normalized[:, gripper_dim] = np.where(normalized[:, gripper_dim] < threshold, 0, 1)
+    This ensures the unnormalizer receives an exact binary value, which is
+    required when the model was trained with gripper in identity (mask=False)
+    space where 0=open and 1=close.
+    """
+
+    def __init__(self, gripper_dim: int = 6, threshold: float = 0.5):
+        self.gripper_dim = gripper_dim
+        self.threshold = threshold
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        action = transition.get(TransitionKey.ACTION)
+        if action is not None and action.shape[-1] > self.gripper_dim:
+            transition = dict(transition)
+            a = action.clone()
+            a[..., self.gripper_dim] = (a[..., self.gripper_dim] >= self.threshold).float()
+            transition[TransitionKey.ACTION] = a
+        return transition
+
+    def transform_features(self, features):
+        return features
+
+
+@ProcessorStepRegistry.register(name="vla_jepa_binarize_gripper")
+class BinarizeGripperProcessorStep(ProcessorStep):
+    """Binarizes a gripper dimension after unnormalization.
+
+    Maps continuous value to {-1, 1}: > threshold → -1, <= threshold → 1 (matches starVLA convention).
+    Only applied when action has more dimensions than gripper_dim.
+    """
+
+    def __init__(self, gripper_dim: int = 6, threshold: float = 0.5):
+        self.gripper_dim = gripper_dim
+        self.threshold = threshold
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        action = transition.get(TransitionKey.ACTION)
+        if action is not None and action.shape[-1] > self.gripper_dim:
+            transition = dict(transition)
+            a = action.clone()
+            a[..., self.gripper_dim] = 1.0 - 2.0 * (a[..., self.gripper_dim] > self.threshold).float()
+            transition[TransitionKey.ACTION] = a
+        return transition
+
+    def transform_features(self, features):
+        return features
+
+
+def make_vla_jepa_pre_post_processors(
+    config: VLAJEPAConfig,
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+) -> tuple[
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    PolicyProcessorPipeline[PolicyAction, PolicyAction],
+]:
+    features = {**config.input_features, **config.output_features}
+    input_steps = [
+        RenameObservationsProcessorStep(rename_map={}),
+        AddBatchDimensionProcessorStep(),
+        DeviceProcessorStep(device=config.device),
+        NormalizerProcessorStep(
+            features=features,
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+        ),
+    ]
+    output_steps: list[ProcessorStep] = []
+    if config.clip_normalized_actions:
+        output_steps.append(ClipActionsProcessorStep())
+    if config.pre_snap_gripper_action:
+        output_steps.append(
+            PreSnapGripperProcessorStep(gripper_dim=config.gripper_dim, threshold=config.gripper_threshold)
+        )
+    output_steps.append(
+        UnnormalizerProcessorStep(
+            features=features,
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+        )
+    )
+    if config.binarize_gripper_action:
+        output_steps.append(
+            BinarizeGripperProcessorStep(gripper_dim=config.gripper_dim, threshold=config.gripper_threshold)
+        )
+    output_steps.append(DeviceProcessorStep(device="cpu"))
+    return (
+        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
+            steps=input_steps,
+            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
+        ),
+        PolicyProcessorPipeline[PolicyAction, PolicyAction](
+            steps=output_steps,
+            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
+            to_transition=policy_action_to_transition,
+            to_output=transition_to_policy_action,
+        ),
+    )

src/lerobot/policies/vla_jepa/qwen_interface.py

@@ -0,0 +1,117 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+from typing import TYPE_CHECKING
+
+import numpy as np
+import torch
+from PIL import Image
+
+from lerobot.utils.import_utils import _transformers_available
+
+if TYPE_CHECKING or _transformers_available:
+    from transformers import AutoProcessor, Qwen3VLForConditionalGeneration
+else:
+    AutoProcessor = None
+    Qwen3VLForConditionalGeneration = None
+
+from .configuration_vla_jepa import VLAJEPAConfig
+
+
+class Qwen3VLInterface(torch.nn.Module):
+    def __init__(self, config: VLAJEPAConfig) -> None:
+        super().__init__()
+        self.config = config
+        self.model = Qwen3VLForConditionalGeneration.from_pretrained(
+            config.qwen_model_name,
+            torch_dtype=self._get_torch_dtype(config.torch_dtype),
+        )
+        self.processor = AutoProcessor.from_pretrained(config.qwen_model_name)
+        self.processor.tokenizer.padding_side = config.tokenizer_padding_side
+        self.model.config.hidden_size = self.model.config.text_config.hidden_size
+
+    @staticmethod
+    def _get_torch_dtype(dtype_name: str) -> torch.dtype:
+        if dtype_name == "float32":
+            return torch.float32
+        if dtype_name == "float16":
+            return torch.float16
+        return torch.bfloat16
+
+    def expand_tokenizer(self) -> tuple[list[str], list[int], int]:
+        # starVLA/JEVLA checkpoints expand action tokens as action_horizon * 4,
+        # independent of vj2 num_action_tokens_per_timestep. Keeping this count
+        # is required for Qwen embedding/lm_head checkpoint shapes to match.
+        max_action_tokens = self.config.chunk_size * 4
+        tokenizer = self.processor.tokenizer
+        action_tokens = []
+        action_token_ids = []
+        for idx in range(max_action_tokens):
+            token = self.config.special_action_token.format(idx)
+            action_tokens.append(token)
+            if token not in tokenizer.get_vocab():
+                tokenizer.add_tokens([token], special_tokens=True)
+            action_token_ids.append(tokenizer.convert_tokens_to_ids(token))
+
+        embodied_action_token = self.config.embodied_action_token
+        if embodied_action_token not in tokenizer.get_vocab():
+            tokenizer.add_tokens([embodied_action_token], special_tokens=True)
+        embodied_action_token_id = tokenizer.convert_tokens_to_ids(embodied_action_token)
+
+        if self.model.get_input_embeddings().weight.size(0) < len(tokenizer):
+            self.model.resize_token_embeddings(len(tokenizer))
+        return action_tokens, action_token_ids, embodied_action_token_id
+
+    def build_inputs(
+        self,
+        images: Sequence[Sequence[Image.Image]],
+        instructions: Sequence[str],
+        action_prompt: str,
+        embodied_prompt: str,
+    ) -> dict[str, torch.Tensor]:
+        messages = []
+        for sample_images, instruction in zip(images, instructions, strict=True):
+            prompt = self.config.prompt_template.format(
+                instruction=instruction,
+                actions=action_prompt,
+                e_actions=embodied_prompt,
+            )
+            content = [{"type": "image", "image": img} for img in sample_images]
+            content.append({"type": "text", "text": prompt})
+            messages.append([{"role": "user", "content": content}])
+
+        batch_inputs = self.processor.apply_chat_template(
+            messages,
+            tokenize=True,
+            add_generation_prompt=True,
+            return_dict=True,
+            processor_kwargs={"padding": True, "return_tensors": "pt"},
+        )
+        return batch_inputs.to(self.model.device)
+
+    @staticmethod
+    def tensor_to_pil(image_tensor: torch.Tensor) -> Image.Image:
+        image = image_tensor.detach().cpu()
+        if image.ndim == 3 and image.shape[0] in (1, 3):
+            image = image.permute(1, 2, 0)
+        image = image.float()
+        if image.max() <= 1.0:
+            image = image * 255.0
+        image = image.clamp(0, 255).round().to(torch.uint8).numpy()
+        if image.shape[-1] == 1:
+            image = np.repeat(image, 3, axis=-1)
+        return Image.fromarray(image)

src/lerobot/policies/vla_jepa/world_model.py

@@ -0,0 +1,418 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+import torch
+import torch.nn.functional as F  # noqa: N812
+from torch import nn
+
+
+def build_action_block_causal_attention_mask(
+    num_frames: int, grid_height: int, grid_width: int, add_tokens: int = 1
+) -> torch.Tensor:
+    tokens_per_frame = add_tokens + grid_height * grid_width
+    num_tokens = num_frames * tokens_per_frame
+    mask = torch.zeros(num_tokens, num_tokens, dtype=torch.bool)
+    mask_block = torch.ones(tokens_per_frame, tokens_per_frame, dtype=torch.bool)
+    local_window_time = num_frames
+
+    for current_frame in range(num_frames):
+        first_context_frame = max(0, current_frame - local_window_time + 1)
+        for context_frame in range(first_context_frame, current_frame + 1):
+            row = slice(current_frame * tokens_per_frame, (current_frame + 1) * tokens_per_frame)
+            col = slice(context_frame * tokens_per_frame, (context_frame + 1) * tokens_per_frame)
+            mask[row, col] = mask_block
+    return mask
+
+
+def rotate_queries_or_keys(x: torch.Tensor, pos: torch.Tensor) -> torch.Tensor:
+    _, _, _, dim = x.size()
+    if dim % 2 != 0:
+        raise ValueError("Embedding dimension must be even for rotary position encoding.")
+
+    omega = torch.arange(dim // 2, dtype=x.dtype, device=x.device)
+    omega /= dim / 2.0
+    omega = 1.0 / 10000**omega
+    freqs = torch.einsum("..., f -> ... f", pos, omega)
+    emb_sin = freqs.sin().squeeze(-1).repeat(1, 1, 1, 2)
+    emb_cos = freqs.cos().squeeze(-1).repeat(1, 1, 1, 2)
+
+    y = x.unflatten(-1, (-1, 2))
+    y1, y2 = y.unbind(dim=-1)
+    y = torch.stack((-y2, y1), dim=-1).flatten(-2)
+    return x * emb_cos + y * emb_sin
+
+
+class DropPath(nn.Module):
+    def __init__(self, drop_prob: float = 0.0) -> None:
+        super().__init__()
+        self.drop_prob = drop_prob
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        if self.drop_prob == 0.0 or not self.training:
+            return x
+        keep_prob = 1 - self.drop_prob
+        shape = (x.shape[0],) + (1,) * (x.ndim - 1)
+        random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype, device=x.device)
+        random_tensor.floor_()
+        return x.div(keep_prob) * random_tensor
+
+
+class MLP(nn.Module):
+    def __init__(
+        self,
+        in_features: int,
+        hidden_features: int | None = None,
+        out_features: int | None = None,
+        act_layer: type[nn.Module] = nn.GELU,
+        drop: float = 0.0,
+    ) -> None:
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+class ACRoPEAttention(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int = 8,
+        qkv_bias: bool = False,
+        qk_scale: float | None = None,
+        attn_drop: float = 0.0,
+        proj_drop: float = 0.0,
+        use_sdpa: bool = True,
+        is_causal: bool = False,
+        grid_size: int = 16,
+    ) -> None:
+        super().__init__()
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.scale = qk_scale or self.head_dim**-0.5
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop_prob = proj_drop
+        self.proj_drop = nn.Dropout(proj_drop)
+        self.use_sdpa = use_sdpa
+        self.d_dim = int(2 * ((self.head_dim // 3) // 2))
+        self.h_dim = int(2 * ((self.head_dim // 3) // 2))
+        self.w_dim = int(2 * ((self.head_dim // 3) // 2))
+        self.grid_size = grid_size
+        self.is_causal = is_causal
+
+    @staticmethod
+    def _get_frame_pos(ids: torch.Tensor, height: int, width: int) -> torch.Tensor:
+        return ids // int(height * width)
+
+    def _get_height_pos(self, ids: torch.Tensor, height: int, width: int) -> torch.Tensor:
+        frame_ids = self._get_frame_pos(ids, height, width)
+        ids = ids - int(height * width) * frame_ids
+        return ids // width
+
+    def separate_positions(
+        self, ids: torch.Tensor, height: int, width: int
+    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        frame_ids = self._get_frame_pos(ids, height, width)
+        height_ids = self._get_height_pos(ids, height, width)
+        width_ids = ids - int(height * width) * frame_ids - width * height_ids
+        return 1.0 * frame_ids, 1.0 * height_ids, 1.0 * width_ids
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        mask: torch.Tensor | None = None,
+        attn_mask: torch.Tensor | None = None,
+        num_frames: int | None = None,
+        grid_height: int | None = None,
+        grid_width: int | None = None,
+        action_tokens: int = 0,
+    ) -> torch.Tensor:
+        batch_size, num_tokens, channels = x.size()
+        if num_frames is None or grid_height is None or grid_width is None:
+            raise ValueError("num_frames, grid_height and grid_width are required.")
+
+        if mask is not None:
+            mask = mask.unsqueeze(1).repeat(1, self.num_heads, 1)
+            d_mask, h_mask, w_mask = self.separate_positions(mask, grid_height, grid_width)
+        else:
+            mask = torch.arange(int(num_frames * grid_height * grid_width), device=x.device)
+            d_mask, h_mask, w_mask = self.separate_positions(mask, grid_height, grid_width)
+
+        h_mask *= self.grid_size / grid_height
+        w_mask *= self.grid_size / grid_width
+
+        if action_tokens > 0:
+            x = x.view(batch_size, -1, action_tokens + grid_height * grid_width, channels)
+            action_q, action_k, action_v = [], [], []
+            for idx in range(action_tokens):
+                action_token = x[:, :, idx : idx + 1, :].flatten(1, 2)
+                qkv = self.qkv(action_token).unflatten(-1, (3, self.num_heads, -1)).permute(2, 0, 3, 1, 4)
+                q, k, v = qkv[0], qkv[1], qkv[2]
+                qd = rotate_queries_or_keys(
+                    q[..., : self.d_dim], pos=torch.arange(num_frames, device=x.device)
+                )
+                kd = rotate_queries_or_keys(
+                    k[..., : self.d_dim], pos=torch.arange(num_frames, device=x.device)
+                )
+                qr = q[..., self.d_dim :]
+                kr = k[..., self.d_dim :]
+                action_q.append(
+                    torch.cat([qd, qr], dim=-1).view(batch_size, self.num_heads, num_frames, 1, -1)
+                )
+                action_k.append(
+                    torch.cat([kd, kr], dim=-1).view(batch_size, self.num_heads, num_frames, 1, -1)
+                )
+                action_v.append(v.view(batch_size, self.num_heads, num_frames, 1, -1))
+
+            action_q = torch.cat(action_q, dim=3).flatten(2, 3)
+            action_k = torch.cat(action_k, dim=3).flatten(2, 3)
+            action_v = torch.cat(action_v, dim=3).flatten(2, 3)
+            x = x[:, :, action_tokens:, :].flatten(1, 2)
+
+        qkv = self.qkv(x).unflatten(-1, (3, self.num_heads, -1)).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]
+
+        offset = 0
+        qd = rotate_queries_or_keys(q[..., offset : offset + self.d_dim], pos=d_mask)
+        kd = rotate_queries_or_keys(k[..., offset : offset + self.d_dim], pos=d_mask)
+        offset += self.d_dim
+        qh = rotate_queries_or_keys(q[..., offset : offset + self.h_dim], pos=h_mask)
+        kh = rotate_queries_or_keys(k[..., offset : offset + self.h_dim], pos=h_mask)
+        offset += self.h_dim
+        qw = rotate_queries_or_keys(q[..., offset : offset + self.w_dim], pos=w_mask)
+        kw = rotate_queries_or_keys(k[..., offset : offset + self.w_dim], pos=w_mask)
+        offset += self.w_dim
+
+        if offset < self.head_dim:
+            q = torch.cat([qd, qh, qw, q[..., offset:]], dim=-1)
+            k = torch.cat([kd, kh, kw, k[..., offset:]], dim=-1)
+        else:
+            q = torch.cat([qd, qh, qw], dim=-1)
+            k = torch.cat([kd, kh, kw], dim=-1)
+
+        if action_tokens > 0:
+
+            def merge(frame_tokens: torch.Tensor, action_token_values: torch.Tensor) -> torch.Tensor:
+                frame_tokens = frame_tokens.view(
+                    batch_size, self.num_heads, num_frames, grid_height * grid_width, -1
+                )
+                action_token_values = action_token_values.view(
+                    batch_size, self.num_heads, num_frames, action_tokens, -1
+                )
+                return torch.cat([action_token_values, frame_tokens], dim=3).flatten(2, 3)
+
+            q = merge(q, action_q)
+            k = merge(k, action_k)
+            v = merge(v, action_v)
+
+        if attn_mask is not None or self.use_sdpa:
+            x = F.scaled_dot_product_attention(
+                q, k, v, dropout_p=self.proj_drop_prob, is_causal=self.is_causal, attn_mask=attn_mask
+            )
+        else:
+            attn = (q @ k.transpose(-2, -1)) * self.scale
+            attn = attn.softmax(dim=-1)
+            attn = self.attn_drop(attn)
+            x = attn @ v
+
+        x = x.transpose(1, 2).reshape(batch_size, num_tokens, channels)
+        x = self.proj(x)
+        return self.proj_drop(x)
+
+
+class ACBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        qkv_bias: bool = True,
+        qk_scale: float | None = None,
+        drop: float = 0.0,
+        attn_drop: float = 0.0,
+        drop_path: float = 0.0,
+        norm_layer: type[nn.Module] = nn.LayerNorm,
+        use_sdpa: bool = True,
+        is_causal: bool = False,
+        grid_size: int = 16,
+        use_rope: bool = True,
+    ) -> None:
+        super().__init__()
+        self.norm1 = norm_layer(dim)
+        if not use_rope:
+            raise ValueError("JEVLA1 world predictor uses AC RoPE attention.")
+        self.attn = ACRoPEAttention(
+            dim,
+            num_heads=num_heads,
+            qkv_bias=qkv_bias,
+            qk_scale=qk_scale,
+            attn_drop=attn_drop,
+            use_sdpa=use_sdpa,
+            is_causal=is_causal,
+            grid_size=grid_size,
+            proj_drop=drop,
+        )
+        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        self.mlp = MLP(
+            in_features=dim,
+            hidden_features=int(dim * mlp_ratio),
+            act_layer=nn.GELU,
+            drop=drop,
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        attn_mask: torch.Tensor | None = None,
+        num_frames: int | None = None,
+        grid_height: int | None = None,
+        grid_width: int | None = None,
+        action_tokens: int = 0,
+    ) -> torch.Tensor:
+        y = self.norm1(x)
+        y = self.attn(
+            y,
+            mask=None,
+            attn_mask=attn_mask,
+            num_frames=num_frames,
+            grid_height=grid_height,
+            grid_width=grid_width,
+            action_tokens=action_tokens,
+        )
+        x = x + self.drop_path(y)
+        y = self.norm2(x)
+        return x + self.drop_path(self.mlp(y))
+
+
+class ActionConditionedVideoPredictor(nn.Module):
+    """JEVLA1-compatible action-conditioned V-JEPA predictor."""
+
+    def __init__(
+        self,
+        num_frames: int,
+        img_size: tuple[int, int],
+        patch_size: int,
+        tubelet_size: int,
+        embed_dim: int,
+        action_embed_dim: int,
+        predictor_embed_dim: int,
+        depth: int,
+        num_heads: int,
+        mlp_ratio: float,
+        num_action_tokens_per_step: int,
+        use_extrinsics: bool = False,
+    ) -> None:
+        super().__init__()
+        self.is_frame_causal = True
+        self.use_extrinsics = use_extrinsics
+        self.predictor_embed = nn.Linear(embed_dim, predictor_embed_dim, bias=True)
+        self.action_encoder = nn.Linear(action_embed_dim, predictor_embed_dim, bias=True)
+        self.state_encoder = nn.Linear(action_embed_dim, predictor_embed_dim, bias=True)
+        self.extrinsics_encoder = nn.Linear(action_embed_dim - 1, predictor_embed_dim, bias=True)
+
+        self.img_height, self.img_width = img_size
+        self.patch_size = patch_size
+        self.num_frames = num_frames
+        self.tubelet_size = tubelet_size
+        self.grid_height = self.img_height // self.patch_size
+        self.grid_width = self.img_width // self.patch_size
+
+        self.predictor_blocks = nn.ModuleList(
+            [
+                ACBlock(
+                    dim=predictor_embed_dim,
+                    num_heads=num_heads,
+                    mlp_ratio=mlp_ratio,
+                    qkv_bias=True,
+                    drop=0.0,
+                    attn_drop=0.0,
+                    drop_path=0.0,
+                    norm_layer=lambda dim: nn.LayerNorm(dim, eps=1e-6),
+                    grid_size=self.grid_height,
+                    use_rope=True,
+                )
+                for _ in range(depth)
+            ]
+        )
+        self.predictor_norm = nn.LayerNorm(predictor_embed_dim, eps=1e-6)
+        self.predictor_proj = nn.Linear(predictor_embed_dim, embed_dim, bias=True)
+        self.num_action_tokens_per_step = num_action_tokens_per_step
+
+    @property
+    def norm(self) -> nn.LayerNorm:
+        return self.predictor_norm
+
+    @property
+    def proj(self) -> nn.Linear:
+        return self.predictor_proj
+
+    def forward(
+        self,
+        frame_tokens: torch.Tensor,
+        action_tokens: torch.Tensor,
+        extrinsics: torch.Tensor | None = None,
+    ) -> torch.Tensor:
+        # starVLA input convention: frame_tokens [B, T*H*W, D], actions [B, T*A, D].
+        x = self.predictor_embed(frame_tokens)
+        batch_size, num_context_tokens, hidden_dim = x.size()
+        num_frames = num_context_tokens // (self.grid_height * self.grid_width)
+
+        actions = self.action_encoder(action_tokens)
+        actions = actions.view(batch_size, num_frames, -1, hidden_dim)
+        cond_tokens = actions.shape[2]
+
+        x = x.view(batch_size, num_frames, self.grid_height * self.grid_width, hidden_dim)
+        if self.use_extrinsics:
+            if extrinsics is None:
+                raise ValueError("extrinsics are required when use_extrinsics=True.")
+            cond_tokens += 1
+            extrinsic_tokens = self.extrinsics_encoder(extrinsics).unsqueeze(2)
+            x = torch.cat([actions, extrinsic_tokens, x], dim=2).flatten(1, 2)
+        else:
+            x = torch.cat([actions, x], dim=2).flatten(1, 2)
+
+        attn_mask = build_action_block_causal_attention_mask(
+            num_frames, self.grid_height, self.grid_width, add_tokens=cond_tokens
+        )
+        attn_mask = attn_mask[: x.size(1), : x.size(1)].to(x.device, non_blocking=True)
+
+        for block in self.predictor_blocks:
+            x = block(
+                x,
+                attn_mask=attn_mask,
+                num_frames=num_frames,
+                grid_height=self.grid_height,
+                grid_width=self.grid_width,
+                action_tokens=cond_tokens,
+            )
+
+        x = x.view(batch_size, num_frames, cond_tokens + self.grid_height * self.grid_width, hidden_dim)
+        x = x[:, :, cond_tokens:, :].flatten(1, 2)
+        x = self.predictor_norm(x)
+        return self.predictor_proj(x)

src/lerobot/processor/pipeline.py

@@ -32,7 +32,6 @@ from __future__ import annotations
 
 import importlib
 import json
-import os
 import re
 from abc import ABC, abstractmethod
 from collections.abc import Callable, Iterable, Sequence
@@ -281,6 +280,11 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
 
     before_step_hooks: list[Callable[[int, EnvTransition], None]] = field(default_factory=list, repr=False)
     after_step_hooks: list[Callable[[int, EnvTransition], None]] = field(default_factory=list, repr=False)
+    _serialized_state_filenames: tuple[str | None, ...] | None = field(
+        default=None,
+        init=False,
+        repr=False,
+    )
 
     def __call__(self, data: TInput) -> TOutput:
         """Processes input data through the full pipeline.
@@ -338,30 +342,108 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
             transition = processor_step(transition)
             yield transition
 
-    def _save_pretrained(self, save_directory: Path, **kwargs):
-        """Internal method to comply with `HubMixin`'s saving mechanism.
+    def _get_sanitized_name(self) -> str:
+        """Return a filename-safe version of the pipeline name.
 
-        This method does the actual saving work and is called by HubMixin.save_pretrained.
+        Returns:
+            The lower-cased pipeline name with non-alphanumeric characters replaced by underscores.
         """
-        config_filename = kwargs.pop("config_filename", None)
+        return re.sub(r"[^a-zA-Z0-9_]", "_", self.name.lower())
 
-        # Sanitize the pipeline name to create a valid filename prefix.
-        sanitized_name = re.sub(r"[^a-zA-Z0-9_]", "_", self.name.lower())
+    @staticmethod
+    def _get_state_filename(
+        *,
+        step_index: int,
+        registry_name: str | None,
+        sanitized_name: str,
+    ) -> str:
+        """Return the safetensors filename for one stateful processor step.
 
-        if config_filename is None:
-            config_filename = f"{sanitized_name}.json"
+        Args:
+            step_index: The index of the processor step in this pipeline.
+            registry_name: The registered processor step name, if available.
+            sanitized_name: The filename-safe pipeline name.
 
-        config: dict[str, Any] = {
+        Returns:
+            The state filename used by the existing disk serialization format.
+        """
+        if registry_name:
+            return f"{sanitized_name}_step_{step_index}_{registry_name}.safetensors"
+
+        return f"{sanitized_name}_step_{step_index}.safetensors"
+
+    @staticmethod
+    def _get_state_key(state_filename: str) -> str:
+        """Return the in-memory state key for a serialized state filename.
+
+        Args:
+            state_filename: The `.safetensors` filename from the serialized config.
+
+        Returns:
+            The state key used by the in-memory pipeline state dictionary.
+        """
+        return state_filename.removesuffix(".safetensors")
+
+    @staticmethod
+    def _get_state_filenames_from_config(loaded_config: dict[str, Any]) -> tuple[str | None, ...]:
+        """Return serialized state filenames in step order.
+
+        Args:
+            loaded_config: A validated processor pipeline config.
+
+        Returns:
+            A tuple containing each step's serialized state filename, or None for stateless steps.
+        """
+        return tuple(step_entry.get("state_file") for step_entry in loaded_config["steps"])
+
+    def _get_state_filenames_for_loading(self) -> tuple[str | None, ...]:
+        """Return expected state filenames in step order for `load_state_dict()`.
+
+        Returns:
+            The preserved serialized state filenames when available, otherwise filenames derived from
+            current non-empty step state.
+        """
+        if self._serialized_state_filenames is not None and len(self._serialized_state_filenames) == len(
+            self.steps
+        ):
+            return self._serialized_state_filenames
+
+        sanitized_name = self._get_sanitized_name()
+        state_filenames: list[str | None] = []
+
+        for step_index, processor_step in enumerate(self.steps):
+            step_state_dict = processor_step.state_dict()
+            if not step_state_dict:
+                state_filenames.append(None)
+                continue
+
+            registry_name = getattr(processor_step.__class__, "_registry_name", None)
+            state_filenames.append(
+                self._get_state_filename(
+                    step_index=step_index,
+                    registry_name=registry_name,
+                    sanitized_name=sanitized_name,
+                )
+            )
+
+        return tuple(state_filenames)
+
+    def get_config(self) -> dict[str, Any]:
+        """Return the JSON-serializable pipeline configuration.
+
+        Returns:
+            A dictionary with the same content that `save_pretrained()` writes as JSON.
+        """
+        sanitized_name = self._get_sanitized_name()
+        pipeline_config: dict[str, Any] = {
             "name": self.name,
             "steps": [],
         }
 
-        # Iterate through each step to build its configuration entry.
         for step_index, processor_step in enumerate(self.steps):
             registry_name = getattr(processor_step.__class__, "_registry_name", None)
-
             step_entry: dict[str, Any] = {}
-            # Prefer registry name for portability, otherwise fall back to full class path.
+
             if registry_name:
                 step_entry["registry_name"] = registry_name
             else:
@@ -369,31 +451,110 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
                     f"{processor_step.__class__.__module__}.{processor_step.__class__.__name__}"
                 )
 
-            # Save step configuration if `get_config` is implemented.
-            if hasattr(processor_step, "get_config"):
-                step_entry["config"] = processor_step.get_config()
+            step_entry["config"] = processor_step.get_config()
 
-            # Save step state if `state_dict` is implemented and returns a non-empty dict.
-            if hasattr(processor_step, "state_dict"):
-                state = processor_step.state_dict()
-                if state:
-                    # Clone tensors to avoid modifying the original state.
-                    cloned_state = {key: tensor.clone() for key, tensor in state.items()}
+            step_state_dict = processor_step.state_dict()
+            if step_state_dict:
+                step_entry["state_file"] = self._get_state_filename(
+                    step_index=step_index,
+                    registry_name=registry_name,
+                    sanitized_name=sanitized_name,
+                )
 
-                    # Create a unique filename for the state file.
-                    if registry_name:
-                        state_filename = f"{sanitized_name}_step_{step_index}_{registry_name}.safetensors"
-                    else:
-                        state_filename = f"{sanitized_name}_step_{step_index}.safetensors"
+            pipeline_config["steps"].append(step_entry)
 
-                    save_file(cloned_state, os.path.join(str(save_directory), state_filename))
-                    step_entry["state_file"] = state_filename
+        return pipeline_config
 
-            config["steps"].append(step_entry)
+    def state_dict(self) -> dict[str, dict[str, torch.Tensor]]:
+        """Return pipeline state tensors grouped by state key.
 
-        # Write the main configuration JSON file.
-        with open(os.path.join(str(save_directory), config_filename), "w") as file_pointer:
-            json.dump(config, file_pointer, indent=2)
+        Returns:
+            A dictionary mapping suffixless state keys to cloned step state dictionaries.
+        """
+        sanitized_name = self._get_sanitized_name()
+        pipeline_state_dict: dict[str, dict[str, torch.Tensor]] = {}
+
+        for step_index, processor_step in enumerate(self.steps):
+            step_state_dict = processor_step.state_dict()
+            if not step_state_dict:
+                continue
+
+            registry_name = getattr(processor_step.__class__, "_registry_name", None)
+            state_filename = self._get_state_filename(
+                step_index=step_index,
+                registry_name=registry_name,
+                sanitized_name=sanitized_name,
+            )
+            state_key = self._get_state_key(state_filename)
+            pipeline_state_dict[state_key] = {
+                tensor_name: tensor.clone() for tensor_name, tensor in step_state_dict.items()
+            }
+
+        return pipeline_state_dict
+
+    def load_state_dict(
+        self,
+        state_dict: dict[str, dict[str, torch.Tensor]],
+    ) -> None:
+        """Load pipeline state tensors into the existing steps.
+
+        Args:
+            state_dict: A dictionary mapping suffixless state keys to step state dictionaries.
+
+        Raises:
+            KeyError: If loading finds missing expected state or unexpected extra state.
+        """
+        expected_state_filenames = self._get_state_filenames_for_loading()
+        used_state_keys: set[str] = set()
+
+        for step_index, (processor_step, state_filename) in enumerate(
+            zip(self.steps, expected_state_filenames, strict=True)
+        ):
+            if state_filename is None:
+                continue
+
+            state_key = self._get_state_key(state_filename)
+            if state_key not in state_dict:
+                raise KeyError(
+                    f"Missing state key '{state_key}' for processor step {step_index}. "
+                    f"Available state keys: {sorted(state_dict.keys())}"
+                )
+
+            processor_step.load_state_dict(state_dict[state_key])
+            used_state_keys.add(state_key)
+
+        unexpected_state_keys = set(state_dict) - used_state_keys
+        if unexpected_state_keys:
+            expected_state_key_set = {
+                self._get_state_key(state_filename)
+                for state_filename in expected_state_filenames
+                if state_filename is not None
+            }
+            raise KeyError(
+                f"Unexpected processor state keys: {sorted(unexpected_state_keys)}. "
+                f"Expected state keys: {sorted(expected_state_key_set)}"
+            )
+
+    def _save_pretrained(self, save_directory: Path, **kwargs) -> None:
+        """Internal method to comply with `HubMixin`'s saving mechanism.
+
+        This method does the actual saving work and is called by HubMixin.save_pretrained.
+        """
+        config_filename = kwargs.pop("config_filename", None)
+        sanitized_name = self._get_sanitized_name()
+
+        if config_filename is None:
+            config_filename = f"{sanitized_name}.json"
+
+        pipeline_config = self.get_config()
+        pipeline_state_dict = self.state_dict()
+
+        for state_key, step_state_dict in pipeline_state_dict.items():
+            state_filename = f"{state_key}.safetensors"
+            save_file(step_state_dict, save_directory / state_filename)
+
+        with open(save_directory / config_filename, "w") as file_pointer:
+            json.dump(pipeline_config, file_pointer, indent=2)
 
     def save_pretrained(
         self,
@@ -577,12 +738,54 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
         cls._validate_overrides_used(validated_overrides, loaded_config)
 
         # 5. Construct and return the final pipeline instance
-        return cls(
+        pipeline = cls(
             steps=steps,
             name=loaded_config.get("name", "DataProcessorPipeline"),
             to_transition=to_transition or cast(Callable[[TInput], EnvTransition], batch_to_transition),
             to_output=to_output or cast(Callable[[EnvTransition], TOutput], transition_to_batch),
         )
+        pipeline._serialized_state_filenames = cls._get_state_filenames_from_config(loaded_config)
+        return pipeline
+
+    @classmethod
+    def from_config(
+        cls,
+        config: dict[str, Any],
+        *,
+        state_dict: dict[str, dict[str, torch.Tensor]] | None = None,
+        overrides: dict[str, Any] | None = None,
+        to_transition: Callable[[TInput], EnvTransition] | None = None,
+        to_output: Callable[[EnvTransition], TOutput] | None = None,
+    ) -> DataProcessorPipeline[TInput, TOutput]:
+        """Build a pipeline from an in-memory config and optional state tensors.
+
+        Args:
+            config: A config dictionary with the same structure as the saved processor JSON.
+            state_dict: Optional in-memory pipeline state grouped by suffixless state key.
+            overrides: Optional constructor overrides keyed by registry name or class name.
+            to_transition: Optional converter from input data to `EnvTransition`.
+            to_output: Optional converter from `EnvTransition` to output data.
+
+        Returns:
+            A processor pipeline built from the config and optional state.
+        """
+        cls._validate_loaded_config("<in-memory config>", config, "<in-memory config>")
+
+        steps, remaining_override_keys = cls._build_steps_from_config(config, overrides or {})
+        cls._validate_overrides_used(remaining_override_keys, config)
+
+        pipeline = cls(
+            steps=steps,
+            name=config.get("name", "DataProcessorPipeline"),
+            to_transition=to_transition or cast(Callable[[TInput], EnvTransition], batch_to_transition),
+            to_output=to_output or cast(Callable[[EnvTransition], TOutput], transition_to_batch),
+        )
+        pipeline._serialized_state_filenames = cls._get_state_filenames_from_config(config)
+
+        if state_dict is not None:
+            pipeline.load_state_dict(state_dict)
+
+        return pipeline
 
     @classmethod
     def _load_config(
@@ -666,9 +869,7 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
                 ) from e
 
     @classmethod
-    def _validate_loaded_config(
-        cls, model_id: str, loaded_config: dict[str, Any], config_filename: str
-    ) -> None:
+    def _validate_loaded_config(cls, model_id: str, loaded_config: Any, config_filename: str) -> None:
         """Validate that a config was loaded and is a valid processor config.
 
         This method validates processor config format with intelligent migration detection:
@@ -688,7 +889,7 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
 
         Args:
             model_id: The model identifier (used for migration detection)
-            loaded_config: The loaded config dictionary (guaranteed non-None)
+            loaded_config: The loaded config value to validate (may be non-dict)
             config_filename: The config filename that was loaded (for error messages)
 
         Raises:
@@ -702,9 +903,14 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
                     model_id,
                     f"Config file '{config_filename}' is not a valid processor configuration",
                 )
+            loaded_config_description = (
+                list(loaded_config.keys())
+                if isinstance(loaded_config, dict)
+                else type(loaded_config).__name__
+            )
             raise ValueError(
                 f"Config file '{config_filename}' is not a valid processor configuration. "
-                f"Expected a config with 'steps' field, but got: {list(loaded_config.keys())}"
+                f"Expected a config with 'steps' field, but got: {loaded_config_description}"
             )
 
     @classmethod
@@ -766,26 +972,41 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
             ImportError: If a step class cannot be imported or found in registry
             ValueError: If a step cannot be instantiated with its configuration
         """
-        steps: list[ProcessorStep] = []
-        override_keys = set(overrides.keys())
+        steps, remaining_override_keys = cls._build_steps_from_config(loaded_config, overrides)
 
-        for step_entry in loaded_config["steps"]:
-            # 1. Get step class and key
-            step_class, step_key = cls._resolve_step_class(step_entry)
-
-            # 2. Instantiate step with overrides
-            step_instance = cls._instantiate_step(step_entry, step_class, step_key, overrides)
-
-            # 3. Load step state if available
+        for step_instance, step_entry in zip(steps, loaded_config["steps"], strict=True):
             cls._load_step_state(step_instance, step_entry, model_id, base_path, hub_download_kwargs)
 
-            # 4. Track used overrides
-            if step_key in override_keys:
-                override_keys.discard(step_key)
+        return steps, remaining_override_keys
 
-            steps.append(step_instance)
+    @classmethod
+    def _build_steps_from_config(
+        cls,
+        loaded_config: dict[str, Any],
+        overrides: dict[str, Any],
+    ) -> tuple[list[ProcessorStep], set[str]]:
+        """Build processor steps from config without loading tensor state.
 
-        return steps, override_keys
+        Args:
+            loaded_config: The loaded processor configuration.
+            overrides: User-provided constructor overrides keyed by step key.
+
+        Returns:
+            A tuple containing instantiated steps and override keys that did not match a step.
+        """
+        processor_steps: list[ProcessorStep] = []
+        remaining_override_keys = set(overrides.keys())
+
+        for step_entry in loaded_config["steps"]:
+            step_class, step_key = cls._resolve_step_class(step_entry)
+            processor_step = cls._instantiate_step(step_entry, step_class, step_key, overrides)
+
+            if step_key in remaining_override_keys:
+                remaining_override_keys.discard(step_key)
+
+            processor_steps.append(processor_step)
+
+        return processor_steps, remaining_override_keys
 
     @classmethod
     def _resolve_step_class(cls, step_entry: dict[str, Any]) -> tuple[type[ProcessorStep], str]:
@@ -1096,7 +1317,7 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
         return True
 
     @classmethod
-    def _is_processor_config(cls, config: dict) -> bool:
+    def _is_processor_config(cls, config: Any) -> bool:
         """Check if config follows DataProcessorPipeline format.
 
         This method validates the processor configuration structure:
@@ -1147,6 +1368,9 @@ class DataProcessorPipeline[TInput, TOutput](HubMixin):
         Returns:
             True if config follows valid DataProcessorPipeline format, False otherwise
         """
+        if not isinstance(config, dict):
+            return False
+
         # Must have a "steps" field with a list of step configurations
         if not isinstance(config.get("steps"), list):
             return False

src/lerobot/processor/relative_action_processor.py

@@ -81,7 +81,7 @@ def to_absolute_actions(actions: Tensor, state: Tensor, mask: Sequence[bool]) ->
     return actions
 
 
-@ProcessorStepRegistry.register("delta_actions_processor")
+@ProcessorStepRegistry.register("relative_actions_processor")
 @dataclass
 class RelativeActionsProcessorStep(ProcessorStep):
     """Converts absolute actions to relative actions (action -= state) for masked dimensions.

src/lerobot/rewards/__init__.py

@@ -20,12 +20,14 @@ from .factory import (
     make_reward_pre_post_processors as make_reward_pre_post_processors,
 )
 from .pretrained import PreTrainedRewardModel as PreTrainedRewardModel
+from .robometer.configuration_robometer import RobometerConfig as RobometerConfig
 from .sarm.configuration_sarm import SARMConfig as SARMConfig
 from .topreward.configuration_topreward import TOPRewardConfig as TOPRewardConfig
 
 __all__ = [
     # Configuration classes
     "RewardClassifierConfig",
+    "RobometerConfig",
     "SARMConfig",
     "TOPRewardConfig",
     # Base class

src/lerobot/rewards/factory.py

@@ -25,6 +25,7 @@ from lerobot.processor import PolicyAction, PolicyProcessorPipeline
 
 from .classifier.configuration_classifier import RewardClassifierConfig
 from .pretrained import PreTrainedRewardModel
+from .robometer.configuration_robometer import RobometerConfig
 from .sarm.configuration_sarm import SARMConfig
 from .topreward.configuration_topreward import TOPRewardConfig
 
@@ -38,7 +39,7 @@ def get_reward_model_class(name: str) -> type[PreTrainedRewardModel]:
 
     Args:
         name: The name of the reward model. Supported names are "reward_classifier",
-              "sarm", "topreward".
+              "sarm", "robometer", "topreward".
 
     Returns:
         The reward model class corresponding to the given name.
@@ -54,6 +55,10 @@ def get_reward_model_class(name: str) -> type[PreTrainedRewardModel]:
         from lerobot.rewards.sarm.modeling_sarm import SARMRewardModel
 
         return SARMRewardModel
+    elif name == "robometer":
+        from lerobot.rewards.robometer.modeling_robometer import RobometerRewardModel
+
+        return RobometerRewardModel
     elif name == "topreward":
         from lerobot.rewards.topreward.modeling_topreward import TOPRewardModel
 
@@ -74,7 +79,7 @@ def make_reward_model_config(reward_type: str, **kwargs) -> RewardModelConfig:
 
     Args:
         reward_type: The type of the reward model. Supported types include
-                     "reward_classifier", "sarm", "topreward".
+                     "reward_classifier", "sarm", "robometer", "topreward".
         **kwargs: Keyword arguments to be passed to the configuration class constructor.
 
     Returns:
@@ -87,6 +92,8 @@ def make_reward_model_config(reward_type: str, **kwargs) -> RewardModelConfig:
         return RewardClassifierConfig(**kwargs)
     elif reward_type == "sarm":
         return SARMConfig(**kwargs)
+    elif reward_type == "robometer":
+        return RobometerConfig(**kwargs)
     elif reward_type == "topreward":
         return TOPRewardConfig(**kwargs)
     else:
@@ -168,6 +175,13 @@ def make_reward_pre_post_processors(
             dataset_stats=kwargs.get("dataset_stats"),
             dataset_meta=kwargs.get("dataset_meta"),
         )
+    elif isinstance(reward_cfg, RobometerConfig):
+        from lerobot.rewards.robometer.processor_robometer import make_robometer_pre_post_processors
+
+        return make_robometer_pre_post_processors(
+            config=reward_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+        )
 
     elif isinstance(reward_cfg, TOPRewardConfig):
         from lerobot.rewards.topreward.processor_topreward import make_topreward_pre_post_processors

src/lerobot/rewards/robometer/__init__.py

@@ -0,0 +1,19 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .configuration_robometer import RobometerConfig
+from .modeling_robometer import RobometerRewardModel
+from .processor_robometer import make_robometer_pre_post_processors
+
+__all__ = ["RobometerConfig", "RobometerRewardModel", "make_robometer_pre_post_processors"]

src/lerobot/rewards/robometer/compute_rabc_weights.py

@@ -0,0 +1,320 @@
+#!/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.
+
+"""Compute per-frame Robometer progress and success curves for a LeRobot dataset.
+
+For each episode, builds per-frame sub-samples using the frame-steps
+strategy from the Robometer eval server: for each original frame ``t``,
+linspace-subsample ``[0, t]`` into ``K`` frames (default 4, matching
+``NUM_SUBSAMPLED_FRAMES`` in the eval server), run one forward through
+the Robometer processor + model, and keep the last-frame progress value.
+All sub-samples are the same size ``K`` so they batch cleanly.
+
+The parquet uses the same schema as SARM's
+:mod:`lerobot.rewards.sarm.compute_rabc_weights` so existing consumers —
+:class:`lerobot.rewards.sarm.rabc.RABCWeights` (which reads
+``progress_sparse``) and the progress-overlay script in
+``examples/dataset/create_progress_videos.py`` — work without modification.
+
+Usage:
+    # Dense per-frame progress for one episode
+    python -m lerobot.rewards.robometer.compute_rabc_weights \\
+        --dataset-repo-id lerobot/libero_10_image \\
+        --reward-model-path lerobot/Robometer-4B \\
+        --episodes 0
+
+    # All episodes with batching
+    python -m lerobot.rewards.robometer.compute_rabc_weights \\
+        --dataset-repo-id lerobot/libero_10_image \\
+        --reward-model-path lerobot/Robometer-4B \\
+        --batch-size 16
+"""
+
+from __future__ import annotations
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+import pyarrow as pa
+import pyarrow.parquet as pq
+import torch
+from tqdm import tqdm
+
+from lerobot.datasets import LeRobotDataset
+from lerobot.rewards.robometer.configuration_robometer import RobometerConfig
+from lerobot.rewards.robometer.modeling_robometer import RobometerRewardModel
+from lerobot.rewards.robometer.processor_robometer import RobometerEncoderProcessorStep
+from lerobot.types import TransitionKey
+
+DEFAULT_OUTPUT_FILENAME = "robometer_progress.parquet"
+
+# Upstream Robometer eval server uses K=4 for frame-steps sub-samples.
+DEFAULT_NUM_SUBSAMPLED_FRAMES = 4
+
+
+def get_reward_model_path_from_parquet(parquet_path: Path) -> str | None:
+    """Read ``reward_model_path`` from parquet metadata if available."""
+    if not parquet_path.exists():
+        return None
+    try:
+        metadata = pq.read_metadata(parquet_path).schema.to_arrow_schema().metadata
+        if metadata and b"reward_model_path" in metadata:
+            return metadata[b"reward_model_path"].decode()
+    except Exception:  # nosec B110
+        return None
+    return None
+
+
+def _resolve_task(sample: dict[str, Any], default: str) -> str:
+    """Best-effort task extraction from a dataset sample."""
+    task = sample.get("task")
+    if isinstance(task, str) and task:
+        return task
+    return default
+
+
+def _build_subsample_indices(num_frames: int, num_subsampled_frames: int) -> list[np.ndarray]:
+    """Frame-steps linspace expansion.
+
+    For each ``t in [0, num_frames - 1]`` returns ``num_subsampled_frames``
+    indices from ``np.linspace(0, t, num_subsampled_frames)`` — the first
+    and last frames are always included. Each entry is a fixed-size array
+    so the model can batch them.
+    """
+    return [np.linspace(0, t, num_subsampled_frames).round().astype(np.int64) for t in range(num_frames)]
+
+
+def compute_robometer_progress(
+    dataset_repo_id: str,
+    reward_model_path: str,
+    output_path: str | None = None,
+    device: str = "cuda",
+    batch_size: int = 32,
+    num_subsampled_frames: int = DEFAULT_NUM_SUBSAMPLED_FRAMES,
+    episodes: list[int] | None = None,
+    image_key: str | None = None,
+) -> Path:
+    """Run Robometer over a dataset and write per-frame progress + success."""
+    logging.info(f"Loading Robometer: {reward_model_path}")
+    config = RobometerConfig(pretrained_path=reward_model_path, device=device)
+    if image_key is not None:
+        config.image_key = image_key
+    model = RobometerRewardModel.from_pretrained(reward_model_path, config=config)
+    model.to(device).eval()
+
+    encoder = RobometerEncoderProcessorStep(
+        base_model_id=config.base_model_id,
+        image_key=config.image_key,
+        task_key=config.task_key,
+        default_task=config.default_task,
+        max_frames=num_subsampled_frames,
+        use_multi_image=config.use_multi_image,
+        use_per_frame_progress_token=config.use_per_frame_progress_token,
+    )
+
+    image_key = config.image_key
+
+    logging.info(f"Loading dataset: {dataset_repo_id}")
+    dataset = LeRobotDataset(dataset_repo_id, download_videos=True)
+    logging.info(f"Dataset: {dataset.num_episodes} episodes, {dataset.num_frames} frames")
+
+    episode_indices = list(range(dataset.num_episodes)) if episodes is None else episodes
+    logging.info(f"Processing {len(episode_indices)} episode(s)")
+
+    all_index: list[int] = []
+    all_episode: list[int] = []
+    all_frame: list[int] = []
+    all_progress: list[float] = []
+
+    for episode_idx in tqdm(episode_indices, desc="Episodes"):
+        ep = dataset.meta.episodes[episode_idx]
+        ep_start = int(ep["dataset_from_index"])
+        ep_end = int(ep["dataset_to_index"])
+        num_frames = ep_end - ep_start
+        if num_frames <= 0:
+            continue
+
+        first_sample = dataset[ep_start]
+        task = _resolve_task(first_sample, default=config.default_task or "perform the task")
+
+        ep_frames = torch.stack([dataset[ep_start + i][image_key] for i in range(num_frames)])
+
+        sub_indices = _build_subsample_indices(num_frames, num_subsampled_frames)
+
+        progress_per_frame = np.zeros(num_frames, dtype=np.float32)
+
+        for start in tqdm(range(0, num_frames, batch_size), desc=f"  Ep {episode_idx}", leave=False):
+            end = min(start + batch_size, num_frames)
+            frames_batch = torch.stack([ep_frames[sub_indices[i]] for i in range(start, end)])
+
+            transition = {
+                TransitionKey.OBSERVATION: {image_key: frames_batch},
+                TransitionKey.COMPLEMENTARY_DATA: {"task": task},
+            }
+            encoded = encoder(transition)
+            obs = encoded[TransitionKey.OBSERVATION]
+            batch = {
+                key: value.to(device) if isinstance(value, torch.Tensor) else value
+                for key, value in obs.items()
+            }
+
+            with torch.no_grad():
+                rewards = model.compute_reward(batch)
+            progress_per_frame[start:end] = rewards.cpu().numpy()
+
+        for local in range(num_frames):
+            all_index.append(ep_start + local)
+            all_episode.append(episode_idx)
+            all_frame.append(local)
+            all_progress.append(float(progress_per_frame[local]))
+
+        if device.startswith("cuda"):
+            torch.cuda.empty_cache()
+
+    table = pa.table(
+        {
+            "index": np.asarray(all_index, dtype=np.int64),
+            "episode_index": np.asarray(all_episode, dtype=np.int64),
+            "frame_index": np.asarray(all_frame, dtype=np.int64),
+            "progress_sparse": np.asarray(all_progress, dtype=np.float32),
+        }
+    ).replace_schema_metadata({b"reward_model_path": reward_model_path.encode()})
+
+    out = Path(dataset.root) / DEFAULT_OUTPUT_FILENAME if output_path is None else Path(output_path)
+    out.parent.mkdir(parents=True, exist_ok=True)
+    pq.write_table(table, out)
+    logging.info(f"Saved {len(table)} frame values to {out}")
+
+    progress_arr = np.asarray(all_progress, dtype=np.float32)
+    if progress_arr.size:
+        logging.info(
+            f"Progress: mean={float(progress_arr.mean()):.4f}, "
+            f"std={float(progress_arr.std()):.4f}, "
+            f"min={float(progress_arr.min()):.4f}, "
+            f"max={float(progress_arr.max()):.4f}"
+        )
+    return out
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Compute per-frame Robometer progress curves for RA-BC weighting.",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+        epilog="""
+Examples:
+    # Dense per-frame progress for one episode
+    python -m lerobot.rewards.robometer.compute_rabc_weights \\
+        --dataset-repo-id lerobot/libero_10_image \\
+        --reward-model-path lerobot/Robometer-4B \\
+        --episodes 0
+
+    # All episodes, smaller batches for memory-constrained GPUs
+    python -m lerobot.rewards.robometer.compute_rabc_weights \\
+        --dataset-repo-id lerobot/libero_10_image \\
+        --reward-model-path lerobot/Robometer-4B \\
+        --batch-size 16
+        """,
+    )
+    parser.add_argument(
+        "--dataset-repo-id", type=str, required=True, help="HuggingFace dataset repo id or local path."
+    )
+    parser.add_argument(
+        "--reward-model-path", type=str, default=None, help="Robometer checkpoint repo id or local path."
+    )
+    parser.add_argument("--output-path", type=str, default=None, help="Output parquet path.")
+    parser.add_argument("--device", type=str, default="cuda", help="Device to use (default: cuda).")
+    parser.add_argument(
+        "--batch-size", type=int, default=32, help="Sub-samples per Qwen forward (default: 32)."
+    )
+    parser.add_argument(
+        "--num-subsampled-frames",
+        type=int,
+        default=DEFAULT_NUM_SUBSAMPLED_FRAMES,
+        help=f"Frames per sub-sample (default: {DEFAULT_NUM_SUBSAMPLED_FRAMES}, matches eval server).",
+    )
+    parser.add_argument(
+        "--episodes", type=int, nargs="+", default=None, help="Process only these episode indices."
+    )
+    parser.add_argument(
+        "--image-key", type=str, default=None, help="Image observation key (default: from config)."
+    )
+    parser.add_argument(
+        "--push-to-hub", action="store_true", help="Upload to the dataset repo on HuggingFace Hub."
+    )
+
+    args = parser.parse_args()
+
+    logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(message)s")
+
+    reward_model_path = args.reward_model_path
+    if reward_model_path is None:
+        temp_dataset = LeRobotDataset(args.dataset_repo_id, download_videos=False)
+        parquet_path = Path(temp_dataset.root) / DEFAULT_OUTPUT_FILENAME
+        reward_model_path = get_reward_model_path_from_parquet(parquet_path)
+        if reward_model_path:
+            logging.info(f"Using reward model from parquet metadata: {reward_model_path}")
+        else:
+            raise ValueError(
+                "--reward-model-path is required (no existing parquet with model metadata found)."
+            )
+
+    output_path = compute_robometer_progress(
+        dataset_repo_id=args.dataset_repo_id,
+        reward_model_path=reward_model_path,
+        output_path=args.output_path,
+        device=args.device,
+        batch_size=args.batch_size,
+        num_subsampled_frames=args.num_subsampled_frames,
+        episodes=args.episodes,
+        image_key=args.image_key,
+    )
+
+    print(f"\nRobometer progress saved to: {output_path}")
+
+    if args.push_to_hub:
+        from huggingface_hub import HfApi
+
+        api = HfApi()
+        hub_path = DEFAULT_OUTPUT_FILENAME
+
+        print(f"\nUploading to Hub: {args.dataset_repo_id}/{hub_path}")
+        api.upload_file(
+            path_or_fileobj=str(output_path),
+            path_in_repo=hub_path,
+            repo_id=args.dataset_repo_id,
+            repo_type="dataset",
+        )
+        print(
+            "Successfully uploaded to: "
+            f"https://huggingface.co/datasets/{args.dataset_repo_id}/blob/main/{hub_path}"
+        )
+
+        print("\nTo use in training, add to your config:")
+        print("  use_rabc: true")
+        print(f"  rabc_progress_path: hf://datasets/{args.dataset_repo_id}/{hub_path}")
+        print("  rabc_head_mode: sparse")
+    else:
+        print("\nTo use in training, add to your config:")
+        print("  use_rabc: true")
+        print(f"  rabc_progress_path: {output_path}")
+        print("  rabc_head_mode: sparse")
+
+
+if __name__ == "__main__":
+    main()

src/lerobot/rewards/robometer/configuration_robometer.py

@@ -0,0 +1,158 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+from copy import deepcopy
+from dataclasses import dataclass, field
+from typing import TYPE_CHECKING, Any
+
+from lerobot.configs import FeatureType, NormalizationMode, PolicyFeature
+from lerobot.configs.rewards import RewardModelConfig
+from lerobot.utils.constants import OBS_IMAGES
+from lerobot.utils.import_utils import _transformers_available, require_package
+
+if TYPE_CHECKING or _transformers_available:
+    from transformers import AutoConfig, AutoTokenizer
+else:
+    AutoConfig = None  # type: ignore[assignment]
+    AutoTokenizer = None  # type: ignore[assignment]
+
+
+# Special tokens Robometer adds to the Qwen-VL tokenizer at construction time.
+# The order is part of the data contract: upstream resized ``embed_tokens``
+# after adding these tokens in this exact order, so changing the set or order
+# would silently misalign the saved embedding rows with their token ids.
+# ``<|reward_token|>`` and ``<|sim_token|>`` are leftover from earlier upstream
+# heads (never read at inference) but still occupy rows the checkpoint expects.
+ROBOMETER_SPECIAL_TOKENS = (
+    "<|split_token|>",
+    "<|reward_token|>",
+    "<|pref_token|>",
+    "<|sim_token|>",
+    "<|prog_token|>",
+)
+
+
+@RewardModelConfig.register_subclass("robometer")
+@dataclass
+class RobometerConfig(RewardModelConfig):
+    """Configuration for the Robometer reward model."""
+
+    pretrained_path: str | None = "lerobot/Robometer-4B"
+    image_key: str = OBS_IMAGES + ".top"
+    task_key: str = "task"
+    default_task: str | None = None
+
+    max_frames: int | None = 8
+    reward_output: str = "progress"  # "progress" or "success"
+    success_threshold: float = 0.5
+
+    license: str | None = "apache-2.0"
+    tags: list[str] | None = field(
+        default_factory=lambda: ["reward-model", "vision-language", "qwen3-vl", "zero-shot"]
+    )
+
+    base_model_id: str = "Qwen/Qwen3-VL-4B-Instruct"
+    torch_dtype: str = "bfloat16"
+    use_multi_image: bool = True
+    use_per_frame_progress_token: bool = True
+    average_temporal_patches: bool = True
+    frame_pooling: str = "mean"  # "mean" | "boundary" | "attention"
+    frame_pooling_attn_temperature: float = 1.0
+    progress_loss_type: str = "discrete"  # "l1" | "l2" | "discrete"
+    progress_discrete_bins: int = 10
+
+    # Serialised Qwen backbone config (post-resize). Always populated by
+    # ``__post_init__`` from ``base_model_id`` + ``len(tokenizer) + 5``, so it
+    # is non-empty after construction. Saved into ``config.json`` automatically
+    # by the base ``_save_pretrained``.
+    vlm_config: dict[str, Any] = field(default_factory=dict)
+
+    input_features: dict[str, PolicyFeature] = field(default_factory=dict)
+    output_features: dict[str, PolicyFeature] = field(default_factory=dict)
+    normalization_mapping: dict[str, NormalizationMode] = field(
+        default_factory=lambda: {
+            "VISUAL": NormalizationMode.IDENTITY,
+            "REWARD": NormalizationMode.IDENTITY,
+        }
+    )
+
+    def __post_init__(self) -> None:
+        super().__post_init__()
+        if self.reward_output not in {"progress", "success"}:
+            raise ValueError(f"reward_output must be 'progress' or 'success', got {self.reward_output!r}")
+        if self.max_frames is not None and self.max_frames < 1:
+            raise ValueError(f"max_frames must be >= 1, got {self.max_frames}")
+        if self.frame_pooling not in {"mean", "boundary", "attention"}:
+            raise ValueError(f"frame_pooling must be mean/boundary/attention; got {self.frame_pooling!r}")
+        if self.frame_pooling_attn_temperature <= 0:
+            raise ValueError("frame_pooling_attn_temperature must be > 0")
+        if self.progress_loss_type not in {"l1", "l2", "discrete"}:
+            raise ValueError(f"progress_loss_type must be l1/l2/discrete; got {self.progress_loss_type!r}")
+        if self.use_per_frame_progress_token and not self.use_multi_image:
+            raise ValueError("use_per_frame_progress_token=True requires use_multi_image=True")
+
+        if self.image_key not in self.input_features:
+            self.input_features[self.image_key] = PolicyFeature(shape=(3, 224, 224), type=FeatureType.VISUAL)
+        self.output_features.setdefault("progress", PolicyFeature(shape=(1,), type=FeatureType.REWARD))
+        self.output_features.setdefault("success", PolicyFeature(shape=(1,), type=FeatureType.REWARD))
+
+        # Deterministically populate ``vlm_config`` so it is non-empty after
+        # construction. For ``Qwen/Qwen3-VL-4B-Instruct`` this gives
+        # ``len(tokenizer) + 5 = 151,669 + 5 = 151,674`` — the exact post-resize
+        # vocab the published ``Robometer-4B`` checkpoint was saved with.
+        if not self.vlm_config:
+            require_package("transformers", extra="robometer")
+            vlm = AutoConfig.from_pretrained(self.base_model_id).to_dict()
+            tokenizer = AutoTokenizer.from_pretrained(self.base_model_id)
+            text_config = vlm.get("text_config")
+            if not isinstance(text_config, dict):
+                raise ValueError(
+                    f"Backbone config for {self.base_model_id!r} has no nested `text_config`; "
+                    "Robometer expects a Qwen-VL-style config."
+                )
+            text_config["vocab_size"] = len(tokenizer) + len(ROBOMETER_SPECIAL_TOKENS)
+            self.vlm_config = vlm
+
+    @property
+    def use_discrete_progress(self) -> bool:
+        """Whether the progress head outputs distribution logits over bins."""
+        return self.progress_loss_type.lower() == "discrete"
+
+    @property
+    def vlm_backbone_config(self):
+        """Reconstruct the Qwen backbone config from :attr:`vlm_config`."""
+        require_package("transformers", extra="robometer")
+        config_dict = deepcopy(self.vlm_config)
+        model_type = config_dict.pop("model_type", None)
+        if model_type is None:
+            raise ValueError("vlm_config must include `model_type` to reconstruct the backbone config")
+        return AutoConfig.for_model(model_type, **config_dict)
+
+    @property
+    def observation_delta_indices(self) -> list[int] | None:
+        return None
+
+    @property
+    def action_delta_indices(self) -> None:
+        return None
+
+    @property
+    def reward_delta_indices(self) -> None:
+        return None
+
+    def validate_features(self) -> None:
+        if self.image_key not in self.input_features:
+            raise ValueError(f"Robometer requires image input feature {self.image_key!r}")

src/lerobot/rewards/robometer/modeling_robometer.py

@@ -0,0 +1,481 @@
+# Copyright 2026 Anthony Liang, Yigit Korkmaz, Stephen Tu, Erdem Bıyık, Jesse Zhang
+# 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.
+
+"""ROBOMETER: Scaling General-Purpose Robotic Reward Models via Trajectory Comparisons.
+
+Paper:         https://arxiv.org/abs/2603.02115
+Project:       https://robometer.github.io
+Original code: https://github.com/aliang8/robometer
+Model:         https://huggingface.co/robometer/Robometer-4B
+
+Robometer is a general-purpose, video-language-input reward model built on
+``Qwen/Qwen3-VL-4B-Instruct``. It is trained with a dual reward-prediction
+objective:
+
+- A frame-level progress loss anchoring reward magnitude on expert data.
+- A trajectory-comparison preference loss imposing global ordering constraints
+  across trajectories sharing the same instruction.
+
+To support downstream RL it also predicts a frame-level binary success. The
+training prompt inserts three learnable tokens:
+
+- ``<|prog_token|>`` after each frame to read per-frame progress and success.
+- ``<|pref_token|>`` at the end to read pairwise preference (training-only).
+- ``<|split_token|>`` between two trajectories in preference samples
+  (training-only).
+
+Progress is modeled as a categorical distribution over ``progress_discrete_bins``
+uniformly-spaced centers in ``[0, 1]`` (C51-style), and the continuous estimate
+is recovered as the softmax-weighted mean of those centers — see
+:func:`convert_bins_to_continuous`.
+
+This LeRobot port is **inference-only**: the preference head is preserved in
+the state dict for byte-equivalence with the published ``Robometer-4B``
+checkpoint but is not queried by :meth:`RobometerRewardModel.compute_reward`,
+which returns the last-frame progress (clamped to ``[0, 1]``) or sigmoid'd
+success probability depending on :attr:`RobometerConfig.reward_output`.
+"""
+
+from __future__ import annotations
+
+import logging
+from typing import TYPE_CHECKING, Any
+
+import torch
+from torch import Tensor, nn
+
+from lerobot.rewards.pretrained import PreTrainedRewardModel
+from lerobot.rewards.robometer.configuration_robometer import RobometerConfig
+from lerobot.utils.constants import OBS_PREFIX
+from lerobot.utils.import_utils import _transformers_available, require_package
+
+if TYPE_CHECKING or _transformers_available:
+    from transformers import AutoModelForImageTextToText
+else:
+    AutoModelForImageTextToText = None  # type: ignore[assignment]
+
+logger = logging.getLogger(__name__)
+
+# Namespace for Robometer's pre-encoded Qwen-VL observation tensors.
+ROBOMETER_FEATURE_PREFIX = f"{OBS_PREFIX}robometer."
+ROBOMETER_QWEN_INPUT_KEYS = (
+    "input_ids",
+    "attention_mask",
+    "pixel_values",
+    "pixel_values_videos",
+    "image_grid_thw",
+    "video_grid_thw",
+    "second_per_grid_ts",
+    "mm_token_type_ids",
+)
+ROBOMETER_METADATA_KEYS = (
+    "prog_token_id",
+    "vision_start_token_id",
+    "vision_end_token_id",
+    "video_merge_size",
+)
+ROBOMETER_INPUT_KEYS = ROBOMETER_QWEN_INPUT_KEYS + ROBOMETER_METADATA_KEYS
+
+
+def convert_bins_to_continuous(bin_logits: Tensor) -> Tensor:
+    """Collapse per-bin logits into a single value in ``[0, 1]``.
+
+    The discrete progress head outputs ``num_bins`` logits per frame. Bins are
+    evenly spaced centers in ``[0, 1]``; the continuous prediction is the
+    softmax-weighted mean of those centers.
+    """
+    bin_probs = torch.softmax(bin_logits, dim=-1)
+    num_bins = bin_logits.shape[-1]
+    bin_centers = torch.linspace(0.0, 1.0, num_bins, device=bin_logits.device, dtype=bin_logits.dtype)
+    return (bin_probs * bin_centers).sum(dim=-1)
+
+
+def _squeeze_last_safe(x: Tensor) -> Tensor:
+    """Drop a trailing singleton dim only when present."""
+    return x.squeeze(-1) if x.ndim > 1 and x.shape[-1] == 1 else x
+
+
+def _torch_dtype(name: str) -> torch.dtype:
+    dtype = getattr(torch, name, None)
+    if isinstance(dtype, torch.dtype):
+        return dtype
+    raise ValueError(f"Unknown torch dtype: {name!r}")
+
+
+class RobometerPredictionHead(nn.Sequential):
+    """Small MLP head used for Robometer's progress / success / preference outputs."""
+
+    def __init__(self, hidden_dim: int, output_size: int, *, dropout: float, with_sigmoid: bool) -> None:
+        layers: list[nn.Module] = [
+            nn.Linear(hidden_dim, hidden_dim // 2),
+            nn.LayerNorm(hidden_dim // 2),
+            nn.GELU(),
+            nn.Dropout(dropout),
+            nn.Linear(hidden_dim // 2, output_size),
+        ]
+        if with_sigmoid:
+            layers.append(nn.Sigmoid())
+        super().__init__(*layers)
+
+
+def decode_progress_outputs(
+    progress_logits: Tensor | None,
+    success_logits: Tensor | None,
+    *,
+    is_discrete_mode: bool,
+) -> dict[str, list[list[float]]]:
+    """Decode RBM head outputs into per-frame floats.
+
+    Args:
+        progress_logits: ``(B, T)`` (continuous) or ``(B, T, num_bins)`` (discrete).
+        success_logits: ``(B, T)`` raw logits, ``sigmoid``-ed to probabilities.
+        is_discrete_mode: if True the progress logits get a softmax over bins
+            and are projected onto bin centers via :func:`convert_bins_to_continuous`.
+
+    Returns:
+        Dict with ``progress_pred`` and ``success_probs``, each a list of
+        length ``B`` of per-frame float lists.
+    """
+    progress_pred: list[list[float]] = []
+    success_probs: list[list[float]] = []
+
+    if progress_logits is not None:
+        for sample_logits in progress_logits:
+            if is_discrete_mode:
+                continuous = convert_bins_to_continuous(sample_logits.detach().float().cpu())
+                progress_pred.append(continuous.flatten().tolist())
+            else:
+                progress_pred.append(sample_logits.detach().float().cpu().flatten().tolist())
+
+    if success_logits is not None:
+        for sample_logits in success_logits:
+            success_probs.append(torch.sigmoid(sample_logits.detach().float().cpu()).flatten().tolist())
+
+    return {"progress_pred": progress_pred, "success_probs": success_probs}
+
+
+class RobometerRewardModel(PreTrainedRewardModel):
+    """Robometer (RBM) reward model — inference-only LeRobot port.
+
+    Wraps a Qwen-VL backbone (default: ``Qwen/Qwen3-VL-4B-Instruct``) with three
+    prediction heads from the paper (progress, success, preference). At
+    inference time only the progress and success heads are queried; the
+    preference head is kept on the module so the published ``Robometer-4B``
+    safetensors load unchanged.
+    """
+
+    name = "robometer"
+    config_class = RobometerConfig
+
+    def __init__(self, config: RobometerConfig, *, dropout: float = 0.1) -> None:
+        require_package("transformers", extra="robometer")
+        super().__init__(config)
+        self.config = config
+
+        # Two backbone-build paths (EO-1 style, branched on ``pretrained_path``):
+        #
+        #   - Fresh training (``pretrained_path is None``): download the base
+        #     Qwen weights and resize the embed table to match
+        #     ``vlm_config.text_config.vocab_size`` — populated deterministically
+        #     in ``RobometerConfig.__post_init__`` as
+        #     ``len(tokenizer) + len(ROBOMETER_SPECIAL_TOKENS)``
+        #
+        #   - Loading a saved checkpoint (``pretrained_path`` is set): rebuild
+        #     the empty architecture from ``vlm_config`` via
+        #     ``AutoModelForImageTextToText.from_config`` so the subsequent
+        #     ``model.safetensors`` load is a direct fill of the right shape —
+        #     no redundant Qwen weight download.
+        torch_dtype = _torch_dtype(config.torch_dtype)
+        if config.pretrained_path is None:
+            self.model = AutoModelForImageTextToText.from_pretrained(
+                config.base_model_id,
+                dtype=torch_dtype,
+                trust_remote_code=True,
+            )
+            target_vocab = config.vlm_config["text_config"]["vocab_size"]
+            self.model.resize_token_embeddings(target_vocab)
+        else:
+            self.model = AutoModelForImageTextToText.from_config(
+                config.vlm_backbone_config,
+                dtype=torch_dtype,
+                trust_remote_code=True,
+            )
+
+        # All Qwen-VL backbones Robometer supports expose `text_config.hidden_size`.
+        # Falls back to the top-level `hidden_size` so future non-multimodal
+        # variants would still resolve.
+        backbone_config = self.model.config
+        text_config = getattr(backbone_config, "text_config", None)
+        hidden_size = getattr(text_config, "hidden_size", None) if text_config is not None else None
+        if hidden_size is None:
+            hidden_size = getattr(backbone_config, "hidden_size", None)
+        if hidden_size is None:
+            raise AttributeError(
+                f"Could not infer hidden_size from backbone config of {config.base_model_id}"
+            )
+        hidden_dim = int(hidden_size)
+
+        # Robometer's three prediction heads + frame-pool attention.
+        progress_output = config.progress_discrete_bins if config.use_discrete_progress else 1
+        self.progress_head = RobometerPredictionHead(
+            hidden_dim,
+            progress_output,
+            dropout=dropout,
+            with_sigmoid=not config.use_discrete_progress,
+        )
+        self.preference_head = RobometerPredictionHead(hidden_dim, 1, dropout=dropout, with_sigmoid=False)
+        self.success_head = RobometerPredictionHead(hidden_dim, 1, dropout=dropout, with_sigmoid=False)
+        self.frame_pool_attn = nn.Linear(hidden_dim, 1, bias=False)
+
+        # Match the dtype of the loaded base model so weight loading is a no-op cast.
+        model_dtype = next(self.model.parameters()).dtype
+        self.progress_head.to(dtype=model_dtype)
+        self.preference_head.to(dtype=model_dtype)
+        self.success_head.to(dtype=model_dtype)
+        self.frame_pool_attn.to(dtype=model_dtype)
+
+    def compute_reward(self, batch: dict[str, Tensor]) -> Tensor:
+        inputs = {
+            key: batch[f"{ROBOMETER_FEATURE_PREFIX}{key}"]
+            for key in ROBOMETER_INPUT_KEYS
+            if f"{ROBOMETER_FEATURE_PREFIX}{key}" in batch
+        }
+        if "input_ids" not in inputs:
+            raise KeyError(
+                f"Robometer batch missing pre-encoded inputs (expected "
+                f"`{ROBOMETER_FEATURE_PREFIX}input_ids`). Make sure the "
+                "RobometerEncoderProcessorStep ran before `compute_reward`."
+            )
+
+        device = next(self.model.parameters()).device
+        inputs = {key: value.to(device) if hasattr(value, "to") else value for key, value in inputs.items()}
+
+        self.eval()
+        with torch.no_grad():
+            progress_logits, success_logits = self._compute_rbm_logits(inputs)
+
+        decoded = decode_progress_outputs(
+            progress_logits,
+            success_logits,
+            is_discrete_mode=self.config.use_discrete_progress,
+        )
+        values = (
+            decoded["success_probs"] if self.config.reward_output == "success" else decoded["progress_pred"]
+        )
+
+        rewards = torch.stack([torch.as_tensor(seq, dtype=torch.float32)[-1] for seq in values])
+        if self.config.reward_output == "success":
+            rewards = (rewards > self.config.success_threshold).float()
+        else:
+            # Match upstream Robometer's ``extract_rewards_from_output``: per-frame
+            # progress predictions are clamped to ``[0, 1]`` before being returned.
+            rewards = rewards.clamp(0.0, 1.0)
+        return rewards.to(self.config.device or "cpu")
+
+    def _compute_rbm_logits(
+        self,
+        inputs: dict[str, Any],
+    ) -> tuple[Tensor, Tensor]:
+        """Run the Qwen3-VL backbone and apply Robometer's heads.
+
+        ``inputs`` is the encoded batch produced by
+        :class:`RobometerEncoderProcessorStep`. It carries Qwen tensors as well
+        as Robometer-specific metadata (``prog_token_id``,
+        ``vision_start_token_id``, ``vision_end_token_id``, ``video_merge_size``)
+        — the metadata is popped here so the rest can be forwarded straight to
+        the Qwen model.
+
+        Returns ``(progress_logits, success_logits)``. Shapes:
+
+        - ``progress_logits``: ``(B, T)`` (continuous) or ``(B, T, num_bins)`` (discrete).
+        - ``success_logits``: ``(B, T)`` raw logits (sigmoid happens at decode time).
+        """
+        prog_token_id = inputs.pop("prog_token_id", None)
+        vision_start_token_id = inputs.pop("vision_start_token_id", None)
+        vision_end_token_id = inputs.pop("vision_end_token_id", None)
+        video_merge_size = inputs.pop("video_merge_size", 14)
+
+        # Qwen3-VL doesn't reliably populate `last_hidden_state`; ask for the
+        # full hidden-state tuple and take the last layer. This matches the
+        # `is_qwen3` path in upstream Robometer's `RBM.forward_qwen` (main).
+        outputs = self.model(**inputs, output_hidden_states=True, return_dict=True)
+        hidden_state = (
+            outputs.hidden_states[-1]
+            if getattr(outputs, "hidden_states", None)
+            else outputs.last_hidden_state
+        )
+
+        input_ids = inputs["input_ids"]
+        if self.config.use_per_frame_progress_token:
+            if prog_token_id is None:
+                raise KeyError("`prog_token_id` missing in batch (run RobometerEncoderProcessorStep first)")
+            return self._process_token_extraction(hidden_state, input_ids, prog_token_id=prog_token_id)
+        if self.config.use_multi_image:
+            if vision_start_token_id is None or vision_end_token_id is None:
+                raise KeyError(
+                    "`vision_start_token_id` / `vision_end_token_id` missing in batch "
+                    "(run RobometerEncoderProcessorStep first)"
+                )
+            return self._process_multi_image_frames(
+                hidden_state,
+                input_ids,
+                start_id=vision_start_token_id,
+                end_id=vision_end_token_id,
+            )
+        video_grid_thw = inputs.get("video_grid_thw")
+        if video_grid_thw is None:
+            raise ValueError("video_grid_thw is required for video-mode Robometer inference")
+        if vision_start_token_id is None:
+            raise KeyError("`vision_start_token_id` missing in batch")
+        return self._process_video_frames(
+            hidden_state,
+            input_ids,
+            video_grid_thw,
+            start_id=vision_start_token_id,
+            merge_size=video_merge_size,
+        )
+
+    def _apply_heads_to_hidden_states(self, frame_embeddings: Tensor) -> tuple[Tensor, Tensor]:
+        """Apply progress + success heads to a tensor of frame embeddings."""
+        progress_out = self.progress_head(frame_embeddings)
+        progress = progress_out if self.config.use_discrete_progress else _squeeze_last_safe(progress_out)
+        success = _squeeze_last_safe(self.success_head(frame_embeddings))
+        return progress, success
+
+    def _process_token_extraction(
+        self,
+        hidden_state: Tensor,
+        input_ids: Tensor,
+        *,
+        prog_token_id: int,
+    ) -> tuple[Tensor, Tensor]:
+        """Per-frame progress/success from ``<|prog_token|>`` positions."""
+        token_mask = input_ids == prog_token_id
+        batch_indices, positions = token_mask.nonzero(as_tuple=True)
+        if positions.numel() == 0:
+            raise ValueError("`<|prog_token|>` not found in any sequence")
+
+        per_sample_hidden = [
+            hidden_state[i, positions[batch_indices == i]] for i in range(input_ids.shape[0])
+        ]
+        progress_list, success_list = [], []
+        for embeddings in per_sample_hidden:
+            if embeddings.shape[0] == 0:
+                raise ValueError("`<|prog_token|>` missing in a sequence")
+            progress, success = self._apply_heads_to_hidden_states(embeddings)
+            progress_list.append(progress)
+            success_list.append(success)
+
+        return torch.stack(progress_list), torch.stack(success_list)
+
+    def _process_multi_image_frames(
+        self,
+        hidden_state: Tensor,
+        input_ids: Tensor,
+        *,
+        start_id: int,
+        end_id: int,
+    ) -> tuple[Tensor, Tensor]:
+        """Per-frame progress/success in multi-image mode (Qwen-VL)."""
+        progress_list, success_list = [], []
+        for batch_idx in range(input_ids.shape[0]):
+            seq_ids = input_ids[batch_idx]
+            seq_hidden = hidden_state[batch_idx]
+            frame_embeddings = self._extract_hidden_states_from_token_pairs(
+                seq_hidden, seq_ids, start_id, end_id
+            )
+            progress, success = self._apply_heads_to_hidden_states(frame_embeddings)
+            progress_list.append(progress)
+            success_list.append(success)
+
+        return torch.stack(progress_list), torch.stack(success_list)
+
+    def _extract_hidden_states_from_token_pairs(
+        self,
+        hidden_state: Tensor,
+        input_ids: Tensor,
+        start_id: int,
+        end_id: int,
+    ) -> Tensor:
+        start_positions = (input_ids == start_id).nonzero(as_tuple=True)[0]
+        end_positions = (input_ids == end_id).nonzero(as_tuple=True)[0]
+        if start_positions.numel() == 0:
+            raise ValueError("`<|vision_start|>` not found in sequence")
+        if start_positions.numel() != end_positions.numel():
+            raise ValueError(
+                f"Mismatched vision token counts: {start_positions.numel()} start vs "
+                f"{end_positions.numel()} end"
+            )
+
+        frames: list[Tensor] = []
+        for start, end in zip(start_positions.tolist(), end_positions.tolist(), strict=True):
+            if start >= end:
+                raise ValueError(f"Invalid vision token pair: start={start} end={end}")
+            patch_tokens = hidden_state[start + 1 : end]
+            if patch_tokens.shape[0] == 0:
+                frames.append((hidden_state[start] + hidden_state[end]) / 2.0)
+                continue
+
+            pooling = self.config.frame_pooling
+            if pooling == "mean":
+                frames.append(patch_tokens.mean(dim=0))
+            elif pooling == "boundary":
+                frames.append(patch_tokens[-1])
+            else:  # attention
+                scores = (
+                    self.frame_pool_attn(patch_tokens).squeeze(-1)
+                    / self.config.frame_pooling_attn_temperature
+                )
+                weights = torch.softmax(scores, dim=0).unsqueeze(-1)
+                frames.append((weights * patch_tokens).sum(dim=0))
+
+        return torch.stack(frames)
+
+    def _process_video_frames(
+        self,
+        hidden_state: Tensor,
+        input_ids: Tensor,
+        video_grid_thw: Tensor,
+        *,
+        start_id: int,
+        merge_size: int,
+    ) -> tuple[Tensor, Tensor]:
+        """Per-frame progress/success in video mode (Qwen-VL)."""
+        progress_list, success_list = [], []
+        for batch_idx in range(input_ids.shape[0]):
+            seq_ids = input_ids[batch_idx]
+            seq_hidden = hidden_state[batch_idx]
+            start_positions = (seq_ids == start_id).nonzero(as_tuple=True)[0]
+            if start_positions.numel() == 0:
+                raise ValueError("`<|vision_start|>` not found in sequence")
+            t_dim, h_dim, w_dim = (int(x) for x in video_grid_thw[batch_idx].tolist())
+            tokens_per_frame = (h_dim * w_dim) // (merge_size**2)
+
+            cursor = start_positions[0].item()
+            frame_embeddings: list[Tensor] = []
+            for _ in range(t_dim):
+                if self.config.average_temporal_patches:
+                    patch = seq_hidden[cursor : cursor + tokens_per_frame]
+                    frame_embeddings.append(patch.mean(dim=0))
+                else:
+                    frame_embeddings.append(seq_hidden[cursor + tokens_per_frame])
+                cursor += tokens_per_frame
+
+            stacked = torch.stack(frame_embeddings)
+            progress, success = self._apply_heads_to_hidden_states(stacked)
+            progress_list.append(progress)
+            success_list.append(success)
+
+        return torch.stack(progress_list), torch.stack(success_list)

src/lerobot/rewards/robometer/processor_robometer.py

@@ -0,0 +1,338 @@
+# 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.
+
+"""Robometer pre/post processing pipelines."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import TYPE_CHECKING, Any
+
+import numpy as np
+import torch
+from PIL import Image
+from torch import Tensor
+
+from lerobot.configs import PipelineFeatureType, PolicyFeature
+from lerobot.processor import (
+    AddBatchDimensionProcessorStep,
+    DeviceProcessorStep,
+    PolicyAction,
+    PolicyProcessorPipeline,
+    ProcessorStep,
+    ProcessorStepRegistry,
+    policy_action_to_transition,
+)
+from lerobot.rewards.robometer.configuration_robometer import (
+    ROBOMETER_SPECIAL_TOKENS,
+    RobometerConfig,
+)
+from lerobot.rewards.robometer.modeling_robometer import ROBOMETER_FEATURE_PREFIX
+from lerobot.types import EnvTransition, TransitionKey
+from lerobot.utils.constants import (
+    OBS_IMAGES,
+    POLICY_POSTPROCESSOR_DEFAULT_NAME,
+    POLICY_PREPROCESSOR_DEFAULT_NAME,
+)
+from lerobot.utils.import_utils import _transformers_available, require_package
+
+if TYPE_CHECKING or _transformers_available:
+    from transformers import AutoProcessor
+else:
+    AutoProcessor = None
+
+PROGRESS_PROMPT = (
+    "The task for the robot is '{task}'. Given the trajectory video, predict "
+    "the task progress at each frame, how far along the robot is towards "
+    "completing the task, a float between 0 and 1, where 0 is the starting "
+    "state and 1 is when the task is completed. If the robot is not "
+    "performing the same task, predict 0 progress."
+)
+
+
+def _frames_to_pil(frames: np.ndarray) -> list[Image.Image]:
+    """Convert ``(T, H, W, C)`` uint8 frames to a list of PIL images."""
+    if frames.ndim != 4:
+        raise ValueError(f"Expected (T,H,W,C) frames; got shape {frames.shape}")
+    if frames.dtype != np.uint8:
+        frames = np.clip(frames, 0, 255).astype(np.uint8)
+    return [Image.fromarray(frames[i]) for i in range(frames.shape[0])]
+
+
+def _video_to_numpy(video: Tensor, *, max_frames: int | None) -> np.ndarray:
+    """Convert one trajectory tensor to a ``(T, H, W, C) uint8`` numpy array."""
+    if max_frames is not None:
+        video = video[-max_frames:]
+    if video.shape[1] in (1, 3):
+        video = video.permute(0, 2, 3, 1)
+    elif video.shape[-1] not in (1, 3):
+        raise ValueError(f"Expected channel dim of size 1 or 3, got shape {tuple(video.shape)}")
+
+    array = video.detach().cpu().numpy()
+    if np.issubdtype(array.dtype, np.floating) and array.size > 0 and array.max() <= 1.0:
+        array = array * 255.0
+    return np.clip(array, 0, 255).astype(np.uint8)
+
+
+def _expand_tasks(task: Any, *, batch_size: int, default: str | None) -> list[str]:
+    if task is None:
+        task = default
+    if task is None:
+        raise KeyError("Robometer expected a task description in complementary data")
+    if isinstance(task, str):
+        return [task] * batch_size
+    if isinstance(task, tuple):
+        task = list(task)
+    if not (isinstance(task, list) and all(isinstance(item, str) for item in task)):
+        raise TypeError(f"Robometer task must be a string or list of strings, got {type(task)}")
+    if len(task) == 1 and batch_size > 1:
+        return task * batch_size
+    if len(task) != batch_size:
+        raise ValueError(f"Expected {batch_size} tasks, got {len(task)}")
+    return task
+
+
+@dataclass
+@ProcessorStepRegistry.register(name="robometer_encoder")
+class RobometerEncoderProcessorStep(ProcessorStep):
+    """Encode raw frames + task into Qwen-VL tensors for the Robometer model.
+
+    Loads a :class:`~transformers.AutoProcessor` matching ``base_model_id`` and
+    registers Robometer's special tokens on the tokenizer. The matching
+    embedding resize happens model-side in
+    :meth:`RobometerRewardModel.__init__`.
+
+    At call time the step reads:
+
+    - ``observation[image_key]``: ``(B, T, C, H, W)`` or ``(B, C, H, W)`` frames.
+    - ``complementary_data[task_key]``: a string or list of strings.
+
+    and writes ``observation[f"{ROBOMETER_FEATURE_PREFIX}<name>"]`` for:
+
+    - the Qwen-VL processor outputs: ``input_ids``, ``attention_mask``,
+      ``pixel_values``, ``image_grid_thw``, ``video_grid_thw``, ...
+    - Robometer-specific token ids consumed by the model heads:
+      ``prog_token_id``, ``vision_start_token_id``, ``vision_end_token_id``,
+      ``video_merge_size``.
+    """
+
+    base_model_id: str = "Qwen/Qwen3-VL-4B-Instruct"
+    image_key: str = OBS_IMAGES + ".top"
+    task_key: str = "task"
+    default_task: str | None = None
+    max_frames: int | None = 8
+    use_multi_image: bool = True
+    use_per_frame_progress_token: bool = True
+    max_length: int = 1024
+
+    _processor: Any = field(default=None, init=False, repr=False)
+
+    def __post_init__(self) -> None:
+        require_package("transformers", extra="robometer")
+        require_package("qwen-vl-utils", extra="robometer", import_name="qwen_vl_utils")
+
+        self._processor = AutoProcessor.from_pretrained(
+            self.base_model_id,
+            trust_remote_code=True,
+            do_sample_frames=False,
+            padding_side="right",
+        )
+
+        # Register Robometer's special tokens on the tokenizer. The matching
+        # embedding resize happens model-side in `RobometerRewardModel.__init__`.
+        tokenizer = self._processor.tokenizer
+        # Qwen tokenizers may not define a pad token, but batched prompts/videos
+        # require padding, so reuse EOS as the padding token.
+        if tokenizer.pad_token is None:
+            tokenizer.pad_token = tokenizer.eos_token
+        for token in ROBOMETER_SPECIAL_TOKENS:
+            if token not in tokenizer.get_vocab():
+                tokenizer.add_special_tokens({"additional_special_tokens": [token]})
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        observation = transition.get(TransitionKey.OBSERVATION)
+        complementary = transition.get(TransitionKey.COMPLEMENTARY_DATA) or {}
+        if not isinstance(observation, dict):
+            raise ValueError("RobometerEncoderProcessorStep requires an observation dict")
+
+        if self.image_key not in observation:
+            raise KeyError(f"Robometer expected image key {self.image_key!r} in observation")
+
+        frames = observation[self.image_key]
+        tensor = frames.detach().cpu() if isinstance(frames, Tensor) else torch.as_tensor(frames)
+        if tensor.ndim == 4:
+            tensor = tensor.unsqueeze(1)
+        elif tensor.ndim != 5:
+            raise ValueError(
+                f"Expected Robometer frames with shape (B,C,H,W) or (B,T,C,H,W); got {tuple(tensor.shape)}"
+            )
+
+        batch_size = tensor.shape[0]
+        tasks = _expand_tasks(
+            complementary.get(self.task_key, self.default_task),
+            batch_size=batch_size,
+            default=self.default_task,
+        )
+
+        samples = [
+            (_video_to_numpy(tensor[i], max_frames=self.max_frames), tasks[i]) for i in range(batch_size)
+        ]
+        encoded = self.encode_samples(samples)
+
+        new_observation = dict(observation)
+        for key, value in encoded.items():
+            new_observation[f"{ROBOMETER_FEATURE_PREFIX}{key}"] = value
+
+        new_transition = transition.copy()
+        new_transition[TransitionKey.OBSERVATION] = new_observation
+        return new_transition
+
+    def encode_samples(self, samples: list[tuple[np.ndarray, str]]) -> dict[str, Tensor]:
+        """Run the Qwen-VL processor on a list of ``(frames, task)`` samples."""
+        from qwen_vl_utils import process_vision_info
+
+        conversations = [self._build_conversation(frames, task) for frames, task in samples]
+
+        texts = [
+            self._processor.apply_chat_template(
+                msg,
+                tokenize=False,
+                add_generation_prompt=False,
+                add_vision_id=True,
+                enable_thinking=False,
+                fps=1,
+            )
+            for msg in conversations
+        ]
+
+        process_kwargs: dict[str, Any] = {
+            "return_video_kwargs": True,
+            "return_video_metadata": True,
+        }
+        image_processor = getattr(self._processor, "image_processor", None)
+        if image_processor is not None and hasattr(image_processor, "patch_size"):
+            process_kwargs["image_patch_size"] = image_processor.patch_size
+
+        image_inputs, video_inputs, video_kwargs = process_vision_info(conversations, **process_kwargs)
+
+        videos: list[Any] | None = None
+        video_metadatas: list[Any] | None = None
+        if video_inputs:
+            if isinstance(video_inputs[0], tuple) and len(video_inputs[0]) == 2:
+                videos_seq, metadatas_seq = zip(*video_inputs, strict=False)
+                videos = list(videos_seq)
+                video_metadatas = list(metadatas_seq)
+            else:
+                videos = list(video_inputs)
+
+        processor_kwargs: dict[str, Any] = {
+            "text": texts,
+            "images": image_inputs,
+            "padding": True,
+            "truncation": False,
+            "max_length": self.max_length,
+            "return_tensors": "pt",
+            "do_resize": False,
+        }
+        if videos is not None:
+            processor_kwargs["videos"] = videos
+        if video_metadatas is not None:
+            processor_kwargs["video_metadata"] = video_metadatas
+        if video_kwargs:
+            processor_kwargs.update(video_kwargs)
+
+        encoded = self._processor(**processor_kwargs)
+
+        # Write Robometer-specific token ids and the video patch merge size into
+        # the encoded batch so `RobometerRewardModel` doesn't need its own
+        # tokenizer at inference (EO1-style separation: the processor owns the
+        # tokenizer, the model owns the backbone and heads).
+        tokenizer = self._processor.tokenizer
+        encoded["prog_token_id"] = tokenizer.convert_tokens_to_ids("<|prog_token|>")
+        encoded["vision_start_token_id"] = tokenizer.convert_tokens_to_ids("<|vision_start|>")
+        encoded["vision_end_token_id"] = tokenizer.convert_tokens_to_ids("<|vision_end|>")
+        video_processor = getattr(self._processor, "video_processor", None)
+        encoded["video_merge_size"] = int(getattr(video_processor, "merge_size", 14))
+        return encoded
+
+    def _build_conversation(self, frames: np.ndarray, task: str) -> list[dict[str, Any]]:
+        pil_frames = _frames_to_pil(frames)
+        prompt = PROGRESS_PROMPT.format(task=task)
+        content: list[dict[str, Any]] = [{"type": "text", "text": prompt}]
+
+        if self.use_multi_image:
+            for image in pil_frames:
+                content.append({"type": "image", "image": image})
+                if self.use_per_frame_progress_token:
+                    content.append({"type": "text", "text": "<|prog_token|>"})
+        else:
+            content.append({"type": "video", "video": pil_frames, "sample_fps": 1.0})
+
+        return [{"role": "user", "content": content}]
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        return features
+
+    def get_config(self) -> dict[str, Any]:
+        return {
+            "base_model_id": self.base_model_id,
+            "image_key": self.image_key,
+            "task_key": self.task_key,
+            "default_task": self.default_task,
+            "max_frames": self.max_frames,
+            "use_multi_image": self.use_multi_image,
+            "use_per_frame_progress_token": self.use_per_frame_progress_token,
+            "max_length": self.max_length,
+        }
+
+
+def make_robometer_pre_post_processors(
+    config: RobometerConfig,
+    dataset_stats: dict[str, dict[str, Any]] | None = None,
+) -> tuple[
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    PolicyProcessorPipeline[PolicyAction, PolicyAction],
+]:
+    """Pipeline that pre-encodes frames + task into Qwen-VL tensors.
+
+    The preprocessor adds a batch dimension if needed, runs Robometer's
+    encoder, and moves everything to the configured device. The
+    postprocessor is the identity since Robometer outputs a single reward
+    tensor.
+    """
+    del dataset_stats  # Robometer has its own normalisation inside the Qwen-VL processor.
+
+    preprocessor = PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
+        steps=[
+            AddBatchDimensionProcessorStep(),
+            RobometerEncoderProcessorStep(
+                base_model_id=config.base_model_id,
+                image_key=config.image_key,
+                task_key=config.task_key,
+                default_task=config.default_task,
+                max_frames=config.max_frames,
+                use_multi_image=config.use_multi_image,
+                use_per_frame_progress_token=config.use_per_frame_progress_token,
+            ),
+            DeviceProcessorStep(device=config.device or "cpu"),
+        ],
+        name=POLICY_PREPROCESSOR_DEFAULT_NAME,
+    )
+    postprocessor = PolicyProcessorPipeline(
+        name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
+        to_transition=policy_action_to_transition,
+    )
+    return preprocessor, postprocessor

src/lerobot/robots/bi_rebot_b601_follower/bi_rebot_b601_follower.py

@@ -52,7 +52,13 @@ class BiRebotB601Follower(Robot):
             cameras=config.left_arm_config.cameras,
             motor_can_ids=config.left_arm_config.motor_can_ids,
             pos_vel_velocity=config.left_arm_config.pos_vel_velocity,
+            control_mode=config.left_arm_config.control_mode,
+            mit_kp=config.left_arm_config.mit_kp,
+            mit_kd=config.left_arm_config.mit_kd,
+            gripper_control_mode=config.left_arm_config.gripper_control_mode,
             gripper_torque_ratio=config.left_arm_config.gripper_torque_ratio,
+            gripper_mit_kp=config.left_arm_config.gripper_mit_kp,
+            gripper_mit_kd=config.left_arm_config.gripper_mit_kd,
             joint_limits=config.left_arm_config.joint_limits,
         )
 
@@ -67,7 +73,13 @@ class BiRebotB601Follower(Robot):
             cameras=config.right_arm_config.cameras,
             motor_can_ids=config.right_arm_config.motor_can_ids,
             pos_vel_velocity=config.right_arm_config.pos_vel_velocity,
+            control_mode=config.right_arm_config.control_mode,
+            mit_kp=config.right_arm_config.mit_kp,
+            mit_kd=config.right_arm_config.mit_kd,
+            gripper_control_mode=config.right_arm_config.gripper_control_mode,
             gripper_torque_ratio=config.right_arm_config.gripper_torque_ratio,
+            gripper_mit_kp=config.right_arm_config.gripper_mit_kp,
+            gripper_mit_kd=config.right_arm_config.gripper_mit_kd,
             joint_limits=config.right_arm_config.joint_limits,
         )
 

src/lerobot/robots/rebot_b601_follower/config_rebot_b601_follower.py

@@ -65,18 +65,33 @@ class RebotB601FollowerConfig:
         }
     )
 
-    # Target velocity for joints running in POS_VEL mode, in degrees/s. A scalar is
-    # applied to every joint; a list provides one value per joint (in motor order).
-    pos_vel_velocity: float | list[float] = field(default_factory=lambda: [150.0] * 7)
+    # Max speed (deg/s) per joint for POS_VEL arms and FORCE_POS gripper (motor order).
+    pos_vel_velocity: float | list[float] = field(
+        default_factory=lambda: [150.0, 150.0, 150.0, 150.0, 150.0, 150.0, 500.0]
+    )
 
-    # Torque/current ratio for the gripper's FORCE_POS mode, in range [0, 1].
-    gripper_torque_ratio: float = 0.1
+    # Arm control: "mit" or "pos_vel".
+    control_mode: str = "mit"
+
+    # MIT kp/kd per arm joint (motor order). Unused when control_mode="pos_vel".
+    mit_kp: float | list[float] = field(default_factory=lambda: [45.0, 45.0, 45.0, 8.0, 9.0, 8.0, 8.0])
+    mit_kd: float | list[float] = field(default_factory=lambda: [12.0, 12.0, 12.0, 1.0, 1.0, 1.0, 1.0])
+
+    # Gripper control: "force_pos" or "mit".
+    gripper_control_mode: str = "force_pos"
+
+    # FORCE_POS only: max grip force, in [0, 1].
+    gripper_torque_ratio: float = 0.05
+
+    # MIT only.
+    gripper_mit_kp: float = 8.0
+    gripper_mit_kd: float = 0.3
 
     # Soft joint limits (degrees). These are clipped against on every action.
     joint_limits: dict[str, tuple[float, float]] = field(
         default_factory=lambda: {
-            "shoulder_pan": (-145.0, 145.0),
-            "shoulder_lift": (-170.0, 1.0),
+            "shoulder_pan": (-150.0, 150.0),
+            "shoulder_lift": (-200.0, 1.0),
             "elbow_flex": (-200.0, 1.0),
             "wrist_flex": (-80.0, 90.0),
             "wrist_yaw": (-90.0, 90.0),

src/lerobot/robots/rebot_b601_follower/rebot_b601_follower.py

@@ -169,11 +169,25 @@ class RebotB601Follower(Robot):
         print(f"Calibration saved to {self.calibration_fpath}")
 
     def configure(self) -> None:
+        if self.config.control_mode not in ("pos_vel", "mit"):
+            raise ValueError(
+                f"Unsupported control_mode '{self.config.control_mode}'. Use 'pos_vel' or 'mit'."
+            )
+        if self.config.gripper_control_mode not in ("force_pos", "mit"):
+            raise ValueError(
+                f"Unsupported gripper_control_mode '{self.config.gripper_control_mode}'. "
+                "Use 'force_pos' or 'mit'."
+            )
+        use_mit = self.config.control_mode == "mit"
+        gripper_use_mit = self.config.gripper_control_mode == "mit"
         self.bus.enable_all()
         for motor_name, motor in self.motors.items():
-            target_mode = (
-                MotorBridgeMode.FORCE_POS if motor_name == GRIPPER_MOTOR else MotorBridgeMode.POS_VEL
-            )
+            if motor_name == GRIPPER_MOTOR:
+                target_mode = MotorBridgeMode.MIT if gripper_use_mit else MotorBridgeMode.FORCE_POS
+            elif use_mit:
+                target_mode = MotorBridgeMode.MIT
+            else:
+                target_mode = MotorBridgeMode.POS_VEL
             for attempt in range(_ENSURE_MODE_RETRIES + 1):
                 try:
                     motor.ensure_mode(target_mode)
@@ -252,22 +266,34 @@ class RebotB601Follower(Robot):
             goal_present_pos = {key: (g, present_pos.get(key, g)) for key, g in goal_pos.items()}
             goal_pos = ensure_safe_goal_position(goal_present_pos, self.config.max_relative_target)
 
+        use_mit = self.config.control_mode == "mit"
         for motor_name, position_deg in goal_pos.items():
             motor = self.motors.get(motor_name)
             if motor is None:
                 continue
             idx = self.motor_names.index(motor_name)
-            vel_deg_s = (
-                self.config.pos_vel_velocity[idx]
-                if isinstance(self.config.pos_vel_velocity, list)
-                else self.config.pos_vel_velocity
-            )
             pos_rad = math.radians(position_deg)
-            vel_rad = math.radians(vel_deg_s)
             if motor_name == GRIPPER_MOTOR:
-                motor.send_force_pos(pos_rad, vel_rad, self.config.gripper_torque_ratio)
+                if self.config.gripper_control_mode == "mit":
+                    motor.send_mit(pos_rad, 0.0, self.config.gripper_mit_kp, self.config.gripper_mit_kd, 0.0)
+                else:
+                    vel_deg_s = (
+                        self.config.pos_vel_velocity[idx]
+                        if isinstance(self.config.pos_vel_velocity, list)
+                        else self.config.pos_vel_velocity
+                    )
+                    motor.send_force_pos(pos_rad, math.radians(vel_deg_s), self.config.gripper_torque_ratio)
+            elif use_mit:
+                kp = self.config.mit_kp[idx] if isinstance(self.config.mit_kp, list) else self.config.mit_kp
+                kd = self.config.mit_kd[idx] if isinstance(self.config.mit_kd, list) else self.config.mit_kd
+                motor.send_mit(pos_rad, 0.0, kp, kd, 0.0)
             else:
-                motor.send_pos_vel(pos_rad, vel_rad)
+                vel_deg_s = (
+                    self.config.pos_vel_velocity[idx]
+                    if isinstance(self.config.pos_vel_velocity, list)
+                    else self.config.pos_vel_velocity
+                )
+                motor.send_pos_vel(pos_rad, math.radians(vel_deg_s))
 
         return {f"{motor}.pos": val for motor, val in goal_pos.items()}
 

src/lerobot/rollout/__init__.py

@@ -23,6 +23,7 @@ from .configs import (
     DAggerKeyboardConfig,
     DAggerPedalConfig,
     DAggerStrategyConfig,
+    EpisodicStrategyConfig,
     HighlightStrategyConfig,
     RolloutConfig,
     RolloutStrategyConfig,
@@ -49,6 +50,7 @@ from .inference import (
 from .strategies import (
     BaseStrategy,
     DAggerStrategy,
+    EpisodicStrategy,
     HighlightStrategy,
     RolloutStrategy,
     SentryStrategy,
@@ -66,6 +68,8 @@ __all__ = [
     "HardwareContext",
     "HighlightStrategy",
     "HighlightStrategyConfig",
+    "EpisodicStrategy",
+    "EpisodicStrategyConfig",
     "InferenceEngine",
     "InferenceEngineConfig",
     "PolicyContext",

src/lerobot/rollout/configs.py

@@ -121,6 +121,35 @@ class DAggerPedalConfig:
     upload: str = "KEY_C"
 
 
+@RolloutStrategyConfig.register_subclass("episodic")
+@dataclass
+class EpisodicStrategyConfig(RolloutStrategyConfig):
+    """Episode-oriented recording that mirrors the behavior of ``lerobot-record``.
+
+    Records ``dataset.num_episodes`` episodes of maximum ``dataset.episode_time_s`` each.
+    After each episode, runs ``dataset.reset_time_s`` seconds of reset time.
+
+    Keyboard controls:
+        Right arrow  — end current episode or reset phase early
+        Left arrow   — discard current episode and re-record
+        Escape       — stop recording session
+
+    In between episodes:
+    - if there is no teleop leader, the robot is held at its initial joint positions captured at startup.
+    - else, the robot is moved smoothly to the position of the teleop leader.
+    """
+
+    # This only applies if there are no teleop leaders specified.
+    # When True (default), moves the robot back to the joint positions captured at startup.
+    # Otherwise, leave the robot in its current position.
+    reset_to_initial_position: bool = True
+
+    # Whether to turn on or off the leader -> follower smooth handover behavior.
+    # When False, fallback to follower -> leader handover.
+    # Note that leader -> follower handover is only supported when the leader has `send_feedback` capability.
+    smooth_leader_to_follower_handover: bool = True
+
+
 @RolloutStrategyConfig.register_subclass("dagger")
 @dataclass
 class DAggerStrategyConfig(RolloutStrategyConfig):
@@ -229,7 +258,13 @@ class RolloutConfig:
 
         # TODO(Steven): DAgger shouldn't require a dataset (user may want to just rollout+intervene without recording), but for now we require it to simplify the implementation.
         needs_dataset = isinstance(
-            self.strategy, (SentryStrategyConfig, HighlightStrategyConfig, DAggerStrategyConfig)
+            self.strategy,
+            (
+                SentryStrategyConfig,
+                HighlightStrategyConfig,
+                DAggerStrategyConfig,
+                EpisodicStrategyConfig,
+            ),
         )
         if needs_dataset and (self.dataset is None or not self.dataset.repo_id):
             raise ValueError(f"{self.strategy.type} strategy requires --dataset.repo_id to be set")

src/lerobot/rollout/strategies/__init__.py

@@ -17,6 +17,7 @@
 from .base import BaseStrategy
 from .core import RolloutStrategy, estimate_max_episode_seconds, safe_push_to_hub, send_next_action
 from .dagger import DAggerEvents, DAggerPhase, DAggerStrategy
+from .episodic import EpisodicStrategy
 from .factory import create_strategy
 from .highlight import HighlightStrategy
 from .sentry import SentryStrategy
@@ -27,6 +28,7 @@ __all__ = [
     "DAggerPhase",
     "DAggerStrategy",
     "HighlightStrategy",
+    "EpisodicStrategy",
     "RolloutStrategy",
     "SentryStrategy",
     "create_strategy",

src/lerobot/rollout/strategies/dagger.py

@@ -56,10 +56,14 @@ from typing import Any
 
 import numpy as np
 
-from lerobot.common.control_utils import is_headless
+from lerobot.common.control_utils import (
+    follower_smooth_move_to,
+    is_headless,
+    teleop_smooth_move_to,
+    teleop_supports_feedback,
+)
 from lerobot.datasets import VideoEncodingManager
 from lerobot.datasets.utils import DEFAULT_VIDEO_FILE_SIZE_IN_MB
-from lerobot.teleoperators import Teleoperator
 from lerobot.utils.constants import ACTION, OBS_STR
 from lerobot.utils.feature_utils import build_dataset_frame
 from lerobot.utils.import_utils import _pynput_available
@@ -69,7 +73,6 @@ from lerobot.utils.utils import log_say
 
 from ..configs import DAggerKeyboardConfig, DAggerPedalConfig, DAggerStrategyConfig
 from ..context import RolloutContext
-from ..robot_wrapper import ThreadSafeRobot
 from .core import RolloutStrategy, estimate_max_episode_seconds, safe_push_to_hub, send_next_action
 
 PYNPUT_AVAILABLE = _pynput_available
@@ -171,64 +174,6 @@ class DAggerEvents:
         self.upload_requested.clear()
 
 
-# ---------------------------------------------------------------------------
-# Teleoperator helpers
-# ---------------------------------------------------------------------------
-
-
-def _teleop_supports_feedback(teleop: Teleoperator) -> bool:
-    """Return True when the teleop can receive position feedback (is actuated).
-    TODO(Maxime): See if it is possible to unify this interface across teleops instead of duck-typing.
-    """
-    return (
-        bool(teleop.feedback_features)
-        and hasattr(teleop, "disable_torque")
-        and hasattr(teleop, "enable_torque")
-    )
-
-
-def _teleop_smooth_move_to(
-    teleop: Teleoperator, target_pos: dict, duration_s: float = 2.0, fps: int = 30
-) -> None:
-    """Smoothly move an actuated teleop to ``target_pos`` via linear interpolation.
-
-    Requires the teleoperator to support feedback
-    (i.e. have non-empty ``feedback_features`` and implement ``disable_torque`` / ``enable_torque``).
-
-    TODO(Maxime): This blocks up to ``duration_s`` seconds, during this time
-    the follower robot doesn't receive new actions, this could be an issue on LeKiwi.
-    """
-    teleop.enable_torque()
-    current = teleop.get_action()
-    steps = max(int(duration_s * fps), 1)
-
-    for step in range(steps + 1):
-        t = step / steps
-        interp = {
-            k: current[k] * (1 - t) + target_pos[k] * t if k in target_pos else current[k] for k in current
-        }
-        teleop.send_feedback(interp)
-        time.sleep(1 / fps)
-
-
-def _follower_smooth_move_to(
-    robot: ThreadSafeRobot, current: dict, target: dict, duration_s: float = 1.0, fps: int = 30
-) -> None:
-    """Smoothly move the follower robot from ``current`` to ``target`` action.
-
-    Used when the teleop is non-actuated: instead of driving the leader arm
-    to the follower, we bring the follower to the teleop's current pose.
-    Both ``current`` and ``target`` must be in robot-action key space.
-    """
-    steps = max(int(duration_s * fps), 1)
-
-    for step in range(steps + 1):
-        t = step / steps
-        interp = {k: current[k] * (1 - t) + target[k] * t if k in target else current[k] for k in current}
-        robot.send_action(interp)
-        time.sleep(1 / fps)
-
-
 # ---------------------------------------------------------------------------
 # Input device handlers
 # ---------------------------------------------------------------------------
@@ -756,31 +701,31 @@ class DAggerStrategy(RolloutStrategy):
             logger.info("Pausing engine - robot holds position")
             engine.pause()
 
-            if _teleop_supports_feedback(teleop) and prev_action is not None:
+            if teleop_supports_feedback(teleop) and prev_action is not None:
                 # TODO(Maxime): prev_action is in robot action key space (output of robot_action_processor).
                 # send_feedback expects teleop feedback key space. For homogeneous setups (e.g. SO-101
                 # leader + SO-101 follower) the keys are identical so this works. If the processor pipeline
                 # does non-trivial key renaming (e.g. a rename_map on action keys), the interpolation in
-                # _teleop_smooth_move_to silently no-ops and the arm doesn't move.
+                # teleop_smooth_move_to silently no-ops and the arm doesn't move.
                 logger.info("Smooth handover: moving leader arm to follower position")
-                _teleop_smooth_move_to(teleop, prev_action)
+                teleop_smooth_move_to(teleop, prev_action)
 
         elif old_phase == DAggerPhase.PAUSED and new_phase == DAggerPhase.CORRECTING:
             logger.info("Entering correction mode - human teleop control")
-            if not _teleop_supports_feedback(teleop) and prev_action is not None:
+            if not teleop_supports_feedback(teleop) and prev_action is not None:
                 logger.info("Smooth handover: sliding follower to teleop position")
                 obs = robot.get_observation()
                 teleop_action = teleop.get_action()
                 processed = ctx.processors.teleop_action_processor((teleop_action, obs))
                 target = ctx.processors.robot_action_processor((processed, obs))
-                _follower_smooth_move_to(robot, prev_action, target)
+                follower_smooth_move_to(robot, prev_action, target)
 
             # unlock the teleop for human control
-            if _teleop_supports_feedback(teleop):
+            if teleop_supports_feedback(teleop):
                 teleop.disable_torque()
 
         elif old_phase == DAggerPhase.CORRECTING and new_phase == DAggerPhase.PAUSED:
-            if _teleop_supports_feedback(teleop):
+            if teleop_supports_feedback(teleop):
                 teleop.enable_torque()
 
         elif new_phase == DAggerPhase.AUTONOMOUS:
@@ -790,7 +735,7 @@ class DAggerStrategy(RolloutStrategy):
             engine.resume()
 
             # release teleop before resuming the policy
-            if _teleop_supports_feedback(teleop):
+            if teleop_supports_feedback(teleop):
                 teleop.disable_torque()
 
     # ------------------------------------------------------------------

src/lerobot/rollout/strategies/episodic.py

@@ -0,0 +1,335 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Episodic rollout strategy: mirrors the behavior of ``lerobot-record``.
+
+- Policy drives the robot during each recording episode.
+- An optional teleoperator can drive the robot during reset phases so the
+  operator can bring the environment back to its starting configuration.
+  If no teleop is connected the robot stays in its current position.
+- Keyboard controls:
+
+      Right arrow  — end the current episode or reset phase early
+      Left arrow   — discard the current episode and re-record it
+      Escape       — stop the recording session
+
+Dataset naming follows the rollout convention: repo names must start with ``rollout_``.
+"""
+
+from __future__ import annotations
+
+import contextlib
+import logging
+import time
+
+from lerobot.common.control_utils import (
+    follower_smooth_move_to,
+    init_keyboard_listener,
+    is_headless,
+    teleop_smooth_move_to,
+    teleop_supports_feedback,
+)
+from lerobot.datasets import VideoEncodingManager
+from lerobot.utils.constants import ACTION, OBS_STR
+from lerobot.utils.feature_utils import build_dataset_frame
+from lerobot.utils.robot_utils import precise_sleep
+from lerobot.utils.utils import log_say
+from lerobot.utils.visualization_utils import log_rerun_data
+
+from ..configs import EpisodicStrategyConfig
+from ..context import RolloutContext
+from .core import RolloutStrategy, safe_push_to_hub, send_next_action
+
+logger = logging.getLogger(__name__)
+
+
+class EpisodicStrategy(RolloutStrategy):
+    """Policy-driven multi-episode recording, mirrors the behavior of ``lerobot-record``.
+
+    Each recording episode runs the policy for maximum ``dataset.episode_time_s``
+    seconds, recording every frame.  A reset phase of ``dataset.reset_time_s``
+    follows every episode (except the last) so the operator can manually
+    reset the environment.  During the reset phase, an optional teleoperator
+    drives the robot; if none is present the robot returns to its initial joint positions captured at startup.
+
+    The policy state (hidden state, RTC queue, interpolator) is reset at
+    the start of each recording episode.
+
+    Keyboard events:
+        right arrow  → end current episode or reset phase early
+        left arrow   → discard & re-record current episode
+        ESC          → stop the session
+    """
+
+    config: EpisodicStrategyConfig
+
+    def __init__(self, config: EpisodicStrategyConfig) -> None:
+        super().__init__(config)
+        self._listener = None
+        self._events: dict | None = None
+
+    def setup(self, ctx: RolloutContext) -> None:
+        """Start the inference engine and attach the keyboard listener."""
+        self._init_engine(ctx)
+        self._listener, self._events = init_keyboard_listener()
+        logger.info("Episodic strategy ready")
+
+    def run(self, ctx: RolloutContext) -> None:
+        """Main multi-episode recording loop."""
+        cfg = ctx.runtime.cfg
+        dataset_cfg = cfg.dataset
+        robot = ctx.hardware.robot_wrapper
+        teleop = ctx.hardware.teleop
+        dataset = ctx.data.dataset
+        events = self._events
+        features = ctx.data.dataset_features
+
+        fps = cfg.fps
+        episode_time_s = dataset_cfg.episode_time_s
+        reset_time_s = dataset_cfg.reset_time_s
+        num_episodes = dataset_cfg.num_episodes
+        single_task = dataset_cfg.single_task or cfg.task
+        play_sounds = cfg.play_sounds
+
+        display_compressed = (
+            True
+            if (cfg.display_data and cfg.display_ip is not None and cfg.display_port is not None)
+            else cfg.display_compressed_images
+        )
+
+        with VideoEncodingManager(dataset):
+            try:
+                recorded_episodes = 0
+                while recorded_episodes < num_episodes and not events["stop_recording"]:
+                    if ctx.runtime.shutdown_event.is_set():
+                        break
+
+                    # Reset policy state at episode start (discard leftover hidden state / queue)
+                    self._engine.reset()
+                    self._interpolator.reset()
+                    self._engine.resume()
+
+                    log_say(f"Recording episode {dataset.num_episodes}", play_sounds)
+                    self._policy_loop(
+                        ctx=ctx,
+                        robot=robot,
+                        events=events,
+                        features=features,
+                        fps=fps,
+                        control_time_s=episode_time_s,
+                        dataset=dataset,
+                        single_task=single_task,
+                    )
+
+                    # Reset phase, skip after the last episode (but run when re-recording)
+                    if not events["stop_recording"] and (
+                        recorded_episodes < num_episodes - 1 or events["rerecord_episode"]
+                    ):
+                        log_say("Reset the environment", play_sounds)
+
+                        if teleop:
+                            # Smooth handover so the transition to teleop control is jerk-free.
+                            # For actuated teleops: drive the leader arm to the follower's current
+                            # position so the operator takes over without fighting the arm.
+                            # For non-actuated teleops: slide the follower to the teleop's current
+                            # pose instead, since the leader cannot be driven.
+                            obs = robot.get_observation()
+                            current_pos = {k: v for k, v in obs.items() if k.endswith(".pos")}
+                            if (
+                                teleop_supports_feedback(teleop)
+                                and self.config.smooth_leader_to_follower_handover
+                            ):
+                                logger.info("Smooth handover: moving leader arm to follower position")
+                                teleop_smooth_move_to(teleop, current_pos, duration_s=2)
+                                teleop.disable_torque()
+                            else:
+                                logger.info("Smooth handover: sliding follower to teleop position")
+                                teleop_action = teleop.get_action()
+                                processed = ctx.processors.teleop_action_processor((teleop_action, obs))
+                                target = ctx.processors.robot_action_processor((processed, obs))
+                                follower_smooth_move_to(robot, current_pos, target, duration_s=1)
+
+                        elif self.config.reset_to_initial_position:
+                            # No teleop: return the robot to its startup position.
+                            self._return_to_initial_position(hw=ctx.hardware, duration_s=1)
+
+                        self._reset_loop(
+                            ctx=ctx,
+                            robot=robot,
+                            teleop=teleop,
+                            events=events,
+                            fps=fps,
+                            control_time_s=reset_time_s,
+                            display_data=cfg.display_data,
+                            display_compressed=display_compressed,
+                        )
+
+                    if events["rerecord_episode"]:
+                        log_say("Re-record episode", play_sounds)
+                        events["rerecord_episode"] = False
+                        events["exit_early"] = False
+                        dataset.clear_episode_buffer()
+
+                        # returns to its initial joint positions captured at startup
+                        if not teleop and self.config.reset_to_initial_position:
+                            self._return_to_initial_position(hw=ctx.hardware, duration_s=1)
+
+                        continue
+
+                    dataset.save_episode()
+                    recorded_episodes += 1
+            finally:
+                # Save any frames buffered in the current episode so an unexpected
+                # exception or KeyboardInterrupt does not silently drop recorded data.
+                # suppress: save_episode raises if the buffer is empty (nothing to lose).
+                logger.info("Episodic control loop ended — saving any in-progress episode")
+                with contextlib.suppress(Exception):
+                    dataset.save_episode()
+
+    def _policy_loop(
+        self,
+        ctx: RolloutContext,
+        robot,
+        events: dict,
+        features: dict,
+        fps: float,
+        control_time_s: float,
+        dataset,
+        single_task: str,
+    ) -> None:
+        """Policy-driven recording loop for a single episode."""
+        interpolator = self._interpolator
+        control_interval = interpolator.get_control_interval(fps)
+
+        timestamp = 0.0
+        start_t = time.perf_counter()
+
+        while timestamp < control_time_s:
+            loop_start = time.perf_counter()
+
+            if events["exit_early"]:
+                events["exit_early"] = False
+                break
+
+            if ctx.runtime.shutdown_event.is_set():
+                break
+
+            obs = robot.get_observation()
+            obs_processed = self._process_observation_and_notify(ctx.processors, obs)
+
+            if self._handle_warmup(ctx.runtime.cfg.use_torch_compile, loop_start, control_interval):
+                continue
+
+            action_dict = send_next_action(obs_processed, obs, ctx, interpolator)
+
+            if action_dict is not None:
+                obs_frame = build_dataset_frame(features, obs_processed, prefix=OBS_STR)
+                action_frame = build_dataset_frame(features, action_dict, prefix=ACTION)
+                dataset.add_frame({**obs_frame, **action_frame, "task": single_task})
+                self._log_telemetry(obs_processed, action_dict, ctx.runtime)
+
+            dt = time.perf_counter() - loop_start
+            sleep_t = control_interval - dt
+            if sleep_t < 0:
+                logger.warning(
+                    f"Record loop is running slower ({1 / dt:.1f} Hz) than the target FPS ({fps} Hz). "
+                    "Dataset frames might be dropped and robot control might be unstable. "
+                    "Common causes are: 1) Camera FPS not keeping up 2) Policy inference taking too long "
+                    "3) CPU starvation"
+                )
+            precise_sleep(max(sleep_t, 0.0))
+            timestamp = time.perf_counter() - start_t
+
+    def _reset_loop(
+        self,
+        ctx: RolloutContext,
+        robot,
+        teleop,
+        events: dict,
+        fps: float,
+        control_time_s: float,
+        display_data: bool,
+        display_compressed: bool,
+    ) -> None:
+        """Reset-phase loop: teleop drives the robot if available, no recording."""
+        processors = ctx.processors
+        control_interval = 1.0 / fps
+
+        timestamp = 0.0
+        start_t = time.perf_counter()
+
+        while timestamp < control_time_s:
+            loop_start = time.perf_counter()
+
+            if events["exit_early"]:
+                events["exit_early"] = False
+                break
+
+            if ctx.runtime.shutdown_event.is_set():
+                break
+
+            obs = robot.get_observation()
+
+            if teleop is not None:
+                act = teleop.get_action()
+                act_teleop = processors.teleop_action_processor((act, obs))
+                robot_action = processors.robot_action_processor((act_teleop, obs))
+                robot.send_action(robot_action)
+
+                if display_data:
+                    obs_processed = processors.robot_observation_processor(obs)
+                    log_rerun_data(
+                        observation=obs_processed,
+                        action=act_teleop,
+                        compress_images=display_compressed,
+                    )
+
+            dt = time.perf_counter() - loop_start
+            sleep_t = control_interval - dt
+            precise_sleep(max(sleep_t, 0.0))
+            timestamp = time.perf_counter() - start_t
+
+    def teardown(self, ctx: RolloutContext) -> None:
+        """Finalise dataset, stop listener, push to hub, and disconnect hardware."""
+        cfg = ctx.runtime.cfg
+        play_sounds = cfg.play_sounds
+
+        log_say("Stop recording", play_sounds, blocking=True)
+
+        if not is_headless() and self._listener is not None:
+            self._listener.stop()
+
+        if ctx.data.dataset is not None:
+            logger.info("Finalizing dataset...")
+            ctx.data.dataset.finalize()
+
+        if (
+            cfg.dataset is not None
+            and cfg.dataset.push_to_hub
+            and ctx.data.dataset is not None
+            and safe_push_to_hub(
+                ctx.data.dataset,
+                tags=cfg.dataset.tags,
+                private=cfg.dataset.private,
+            )
+        ):
+            logger.info("Dataset uploaded to hub")
+            log_say("Dataset uploaded to hub", play_sounds)
+
+        self._teardown_hardware(
+            ctx.hardware,
+            return_to_initial_position=cfg.return_to_initial_position,
+        )
+        log_say("Exiting", play_sounds)
+        logger.info("Episodic strategy teardown complete")

src/lerobot/rollout/strategies/factory.py

@@ -21,6 +21,7 @@ from typing import TYPE_CHECKING
 from .base import BaseStrategy
 from .core import RolloutStrategy
 from .dagger import DAggerStrategy
+from .episodic import EpisodicStrategy
 from .highlight import HighlightStrategy
 from .sentry import SentryStrategy
 
@@ -42,4 +43,8 @@ def create_strategy(config: RolloutStrategyConfig) -> RolloutStrategy:
         return HighlightStrategy(config)
     if config.type == "dagger":
         return DAggerStrategy(config)
-    raise ValueError(f"Unknown strategy type '{config.type}'. Available: base, sentry, highlight, dagger")
+    if config.type == "episodic":
+        return EpisodicStrategy(config)
+    raise ValueError(
+        f"Unknown strategy type '{config.type}'. Available: base, sentry, highlight, dagger, episodic"
+    )

src/lerobot/scripts/lerobot_rollout.py

@@ -25,6 +25,7 @@ Strategies
     --strategy.type=sentry     Continuous recording with auto-upload
     --strategy.type=highlight  Ring buffer + keystroke save
     --strategy.type=dagger     Human-in-the-loop (DAgger / RaC)
+    --strategy.type=episodic   Episode-oriented recording with reset phases
 
 Inference backends
 ------------------
@@ -111,6 +112,18 @@ Usage examples
         --display_data=true \\
         --use_torch_compile=true
 
+    # Episodic mode — episode-oriented recording with reset phases
+    lerobot-rollout \\
+        --strategy.type=episodic \\
+        --policy.path=user/my_policy \\
+        --robot.type=so100_follower \\
+        --robot.port=/dev/ttyACM0 \\
+        --teleop.type=so100_leader \\
+        --teleop.port=/dev/ttyACM1 \\
+        --dataset.repo_id=user/rollout_episodic_data \\
+        --dataset.num_episodes=20 \\
+        --dataset.single_task="Grab the cube"
+
     # Resume a previous sentry recording session
     lerobot-rollout \\
         --strategy.type=sentry \\

src/lerobot/scripts/lerobot_train.py

@@ -232,15 +232,18 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
         torch.backends.cudnn.benchmark = True
     torch.backends.cuda.matmul.allow_tf32 = True
 
-    # Dataset loading synchronization: main process downloads first to avoid race conditions
-    if is_main_process:
-        logging.info("Creating dataset")
+    # Dataset loading synchronization: each node's local main process downloads first to avoid
+    # race conditions (the global main process only exists on node 0, so gating on it would let
+    # all ranks of the other nodes download and build the Arrow cache concurrently).
+    if accelerator.is_local_main_process:
+        if is_main_process:
+            logging.info("Creating dataset")
         dataset = make_dataset(cfg)
 
     accelerator.wait_for_everyone()
 
-    # Now all other processes can safely load the dataset
-    if not is_main_process:
+    # Now all other processes can safely load the dataset from the local cache
+    if not accelerator.is_local_main_process:
         dataset = make_dataset(cfg)
 
     # Create environment used for evaluating checkpoints during training on simulation data.
@@ -292,19 +295,8 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
 
     active_cfg = cfg.trainable_config
     processor_pretrained_path = active_cfg.pretrained_path
-    if (
-        getattr(active_cfg, "use_relative_actions", False)
-        and processor_pretrained_path is not None
-        and not cfg.resume
-    ):
-        logging.warning(
-            "use_relative_actions=true with pretrained processors can skip relative transforms if "
-            "the checkpoint processors do not define them. Building processors from current policy config."
-        )
-        processor_pretrained_path = None
 
     processor_kwargs = {}
-    postprocessor_kwargs = {}
     if (processor_pretrained_path and not cfg.resume) or not processor_pretrained_path:
         processor_kwargs["dataset_stats"] = dataset.meta.stats
 
@@ -312,24 +304,31 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
         processor_kwargs["dataset_meta"] = dataset.meta
 
     if not cfg.is_reward_model_training and processor_pretrained_path is not None:
-        processor_kwargs["preprocessor_overrides"] = {
+        preprocessor_overrides = {
             "device_processor": {"device": device.type},
             "normalizer_processor": {
                 "stats": dataset.meta.stats,
                 "features": {**policy.config.input_features, **policy.config.output_features},
                 "norm_map": policy.config.normalization_mapping,
             },
+            "rename_observations_processor": {"rename_map": cfg.rename_map},
         }
-        processor_kwargs["preprocessor_overrides"]["rename_observations_processor"] = {
-            "rename_map": cfg.rename_map
-        }
-        postprocessor_kwargs["postprocessor_overrides"] = {
+        postprocessor_overrides = {
             "unnormalizer_processor": {
                 "stats": dataset.meta.stats,
                 "features": policy.config.output_features,
                 "norm_map": policy.config.normalization_mapping,
             },
         }
+        if getattr(active_cfg, "use_relative_actions", False):
+            preprocessor_overrides["relative_actions_processor"] = {
+                "enabled": True,
+                "exclude_joints": getattr(active_cfg, "relative_exclude_joints", []),
+                "action_names": getattr(active_cfg, "action_feature_names", None),
+            }
+            postprocessor_overrides["absolute_actions_processor"] = {"enabled": True}
+        processor_kwargs["preprocessor_overrides"] = preprocessor_overrides
+        processor_kwargs["postprocessor_overrides"] = postprocessor_overrides
 
     if cfg.is_reward_model_training:
         preprocessor, postprocessor = make_reward_pre_post_processors(
@@ -341,7 +340,6 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
             policy_cfg=cfg.policy,
             pretrained_path=processor_pretrained_path,
             **processor_kwargs,
-            **postprocessor_kwargs,
         )
 
     if is_main_process:
@@ -391,12 +389,19 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
     # create dataloader for offline training
     if hasattr(active_cfg, "drop_n_last_frames"):
         shuffle = False
+        # A dedicated generator (rather than the global torch RNG) lets accelerator.prepare
+        # synchronize the shuffle permutation across ranks, keeping batch shards disjoint even
+        # when ranks consume the global RNG asymmetrically (e.g. eval on the main process only).
+        sampler_generator = torch.Generator()
+        if cfg.seed is not None:
+            sampler_generator.manual_seed(cfg.seed)
         sampler = EpisodeAwareSampler(
             dataset.meta.episodes["dataset_from_index"],
             dataset.meta.episodes["dataset_to_index"],
             episode_indices_to_use=dataset.episodes,
             drop_n_last_frames=active_cfg.drop_n_last_frames,
             shuffle=True,
+            generator=sampler_generator,
         )
     else:
         shuffle = True

src/lerobot/teleoperators/rebot_102_leader/config_rebot_102_leader.py

@@ -65,7 +65,7 @@ class RebotArm102LeaderConfig:
     joint_ranges: dict[str, list[int]] = field(
         default_factory=lambda: {
             "shoulder_pan": [-150, 150],
-            "shoulder_lift": [-170, 1],
+            "shoulder_lift": [-200, 1],
             "elbow_flex": [-200, 1],
             "wrist_flex": [-80, 90],
             "wrist_yaw": [-90, 90],

src/lerobot/templates/lerobot_rewardmodel_modelcard_template.md

@@ -13,6 +13,8 @@
 A reward classifier is a lightweight neural network that scores observations or trajectories for task success, providing a learned reward signal or offline evaluation when explicit rewards are unavailable.
 {% elif model_name == "sarm" %}
 A Success-Aware Reward Model (SARM) predicts a dense reward signal from observations, typically used downstream for reinforcement learning or human-in-the-loop fine-tuning when task success is not directly observable.
+{% elif model_name == "robometer" %}
+ROBOMETER is a general-purpose video-language robotic reward model built on a fine-tuned Qwen3-VL-4B backbone with progress, preference, and success heads. Given a trajectory video and a task description, it predicts dense, frame-level task progress in [0, 1] and frame-level success probabilities for downstream robot learning, including offline RL, online RL, data filtering and retrieval, and automated failure detection.
 {% elif model_name == "topreward" %}
 TOPReward is a **zero-shot** reward model that extracts token log-probabilities from an off-the-shelf vision-language model (default Qwen3-VL) as a reward signal. Given a video trajectory and a task instruction, it returns the VLM's log-likelihood of the instruction being true, with no fine-tuning required.
 {% else %}

tests/datasets/test_sampler.py

@@ -114,6 +114,30 @@ def test_shuffle():
     assert set(sampler) == {0, 1, 2, 3, 4, 5}
 
 
+def test_shuffle_with_generator_is_deterministic():
+    # Two samplers shuffling with same-seed generators must yield identical permutations.
+    # This is what keeps batch shards disjoint across ranks in distributed training, where
+    # accelerate synchronizes the sampler's generator state instead of the global torch RNG.
+    sampler_a = EpisodeAwareSampler([0], [6], shuffle=True, generator=torch.Generator().manual_seed(42))
+    sampler_b = EpisodeAwareSampler([0], [6], shuffle=True, generator=torch.Generator().manual_seed(42))
+    assert list(sampler_a) == list(sampler_b)
+
+    # Desyncing the global RNG must not affect the permutation.
+    sampler_c = EpisodeAwareSampler([0], [6], shuffle=True, generator=torch.Generator().manual_seed(42))
+    order_before = list(sampler_c)
+    sampler_c.generator.manual_seed(42)
+    torch.randperm(1000)  # consume global RNG, as rank-asymmetric code (e.g. eval) would
+    assert list(sampler_c) == order_before
+
+
+def test_generator_attribute_defaults_to_none():
+    # accelerate detects synchronizable samplers via `hasattr(sampler, "generator")`,
+    # so the attribute must exist even when no generator is passed.
+    sampler = EpisodeAwareSampler([0], [6], shuffle=True)
+    assert sampler.generator is None
+    assert set(sampler) == {0, 1, 2, 3, 4, 5}
+
+
 def test_negative_drop_first_frames_raises():
     with pytest.raises(ValueError, match="drop_n_first_frames must be >= 0"):
         EpisodeAwareSampler([0], [10], drop_n_first_frames=-1)

tests/policies/vla_jepa/conftest.py

@@ -0,0 +1,273 @@
+#!/usr/bin/env python
+"""Shared fixtures and helpers for VLA-JEPA tests."""
+
+from __future__ import annotations
+
+from types import SimpleNamespace
+
+import numpy as np
+import pytest
+import torch
+from PIL import Image
+from torch import Tensor, nn
+
+from lerobot.configs.types import FeatureType, PolicyFeature
+from lerobot.policies.vla_jepa.configuration_vla_jepa import VLAJEPAConfig
+from lerobot.utils.constants import ACTION, OBS_IMAGES, OBS_STATE
+
+# ---------------------------------------------------------------------------
+# Shared constants
+# ---------------------------------------------------------------------------
+
+BATCH_SIZE = 2
+ACTION_DIM = 3
+STATE_DIM = 4
+IMAGE_SIZE = 8
+ACTION_HORIZON = 4
+N_ACTION_STEPS = 2
+NUM_VIDEO_FRAMES = 3
+QWEN_HIDDEN_SIZE = 16  # hidden size produced by _FakeQwenBackbone
+
+EXPECTED_ACTION_CHUNK_SHAPE = (BATCH_SIZE, ACTION_HORIZON, ACTION_DIM)
+EXPECTED_SELECT_ACTION_SHAPE = (BATCH_SIZE, ACTION_DIM)
+
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+
+def set_seed_all(seed: int) -> None:
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed(seed)
+        torch.cuda.manual_seed_all(seed)
+
+
+def make_config(
+    action_dim: int = ACTION_DIM,
+    state_dim: int = STATE_DIM,
+    action_horizon: int = ACTION_HORIZON,
+    num_video_frames: int = NUM_VIDEO_FRAMES,
+) -> VLAJEPAConfig:
+    config = VLAJEPAConfig(
+        input_features={
+            f"{OBS_IMAGES}.laptop": PolicyFeature(type=FeatureType.VISUAL, shape=(3, IMAGE_SIZE, IMAGE_SIZE)),
+            OBS_STATE: PolicyFeature(type=FeatureType.STATE, shape=(state_dim,)),
+        },
+        output_features={
+            ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(action_dim,)),
+        },
+        device="cpu",
+        chunk_size=action_horizon,
+        n_action_steps=min(N_ACTION_STEPS, action_horizon),
+        action_dim=action_dim,
+        state_dim=state_dim,
+        num_video_frames=num_video_frames,
+        num_action_tokens_per_timestep=2,
+        num_embodied_action_tokens_per_instruction=3,
+        num_inference_timesteps=2,
+        action_hidden_size=QWEN_HIDDEN_SIZE,
+        action_model_type="DiT-test",
+        action_num_layers=1,
+        predictor_depth=1,
+        predictor_num_heads=2,
+        predictor_mlp_ratio=2.0,
+        jepa_tubelet_size=1,
+    )
+    config.validate_features()
+    return config
+
+
+def make_train_batch(
+    batch_size: int = BATCH_SIZE,
+    action_dim: int = ACTION_DIM,
+    state_dim: int = STATE_DIM,
+    action_horizon: int = ACTION_HORIZON,
+    num_video_frames: int = NUM_VIDEO_FRAMES,
+) -> dict[str, Tensor | list[str]]:
+    return {
+        f"{OBS_IMAGES}.laptop": torch.rand(batch_size, num_video_frames, 3, IMAGE_SIZE, IMAGE_SIZE),
+        OBS_STATE: torch.randn(batch_size, 1, state_dim),
+        ACTION: torch.randn(batch_size, action_horizon, action_dim),
+        "task": ["pick up the cube"] * batch_size,
+    }
+
+
+def make_inference_batch(
+    batch_size: int = BATCH_SIZE,
+    state_dim: int = STATE_DIM,
+) -> dict[str, Tensor | list[str]]:
+    return {
+        f"{OBS_IMAGES}.laptop": torch.rand(batch_size, 3, IMAGE_SIZE, IMAGE_SIZE),
+        OBS_STATE: torch.randn(batch_size, state_dim),
+        "task": ["pick up the cube"] * batch_size,
+    }
+
+
+# ---------------------------------------------------------------------------
+# Fake external models (replace Qwen3-VL and V-JEPA at test time)
+# ---------------------------------------------------------------------------
+
+
+class _FakeLanguageLayer(nn.Module):
+    """Leaf module whose forward hook is captured by _qwen_last_decoder_hidden."""
+
+    def __init__(self, hidden_size: int) -> None:
+        super().__init__()
+        self._hidden_size = hidden_size
+
+    def forward(self, hidden: Tensor, **_: object) -> tuple[Tensor, ...]:
+        return (hidden,)
+
+
+class _FakeLanguageModel(nn.Module):
+    def __init__(self, hidden_size: int) -> None:
+        super().__init__()
+        self._hidden_size = hidden_size
+        self.layers = nn.ModuleList([_FakeLanguageLayer(hidden_size)])
+
+    def forward(self, input_ids: Tensor, **_: object) -> SimpleNamespace:
+        batch_size, seq_len = input_ids.shape
+        hidden = torch.zeros(batch_size, seq_len, self._hidden_size, device=input_ids.device)
+        self.layers[-1](hidden)
+        return SimpleNamespace()
+
+
+class _FakeQwenInnerModel(nn.Module):
+    """Mimics the `.model.model` level that _qwen_last_decoder_hidden walks into."""
+
+    def __init__(self, hidden_size: int) -> None:
+        super().__init__()
+        self.language_model = _FakeLanguageModel(hidden_size)
+
+    def forward(self, input_ids: Tensor, **kwargs: object) -> SimpleNamespace:
+        return self.language_model(input_ids)
+
+
+class _FakeQwenBackbone(nn.Module):
+    def __init__(self, hidden_size: int) -> None:
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(1))
+        self.config = SimpleNamespace(
+            hidden_size=hidden_size,
+            text_config=SimpleNamespace(hidden_size=hidden_size),
+        )
+        self.model = _FakeQwenInnerModel(hidden_size)
+
+    @property
+    def device(self) -> torch.device:
+        return self.weight.device
+
+    def forward(self, input_ids: Tensor, **_: object) -> SimpleNamespace:
+        batch_size, seq_len = input_ids.shape
+        hidden_size = self.config.hidden_size
+        values = torch.arange(
+            batch_size * seq_len * hidden_size,
+            device=input_ids.device,
+            dtype=torch.float32,
+        ).view(batch_size, seq_len, hidden_size)
+        hidden = values / values.numel() + self.weight
+        self.model(input_ids)  # call through so the forward hook on layers[-1] fires
+        return SimpleNamespace(hidden_states=[hidden])
+
+
+class _FakeQwenInterface(nn.Module):
+    def __init__(self, config: VLAJEPAConfig) -> None:
+        super().__init__()
+        self.config = config
+        self.model = _FakeQwenBackbone(hidden_size=QWEN_HIDDEN_SIZE)
+
+    @staticmethod
+    def _get_torch_dtype(dtype_name: str) -> torch.dtype:
+        return torch.float32 if dtype_name == "float32" else torch.bfloat16
+
+    def expand_tokenizer(self) -> tuple[list[str], list[int], int]:
+        max_action_tokens = self.config.chunk_size * self.config.num_action_tokens_per_timestep
+        action_tokens = [self.config.special_action_token.format(idx) for idx in range(max_action_tokens)]
+        action_token_ids = list(range(1000, 1000 + max_action_tokens))
+        return action_tokens, action_token_ids, 2000
+
+    def build_inputs(
+        self,
+        images: list[list[Image.Image]],
+        instructions: list[str],
+        action_prompt: str,
+        embodied_prompt: str,
+    ) -> dict[str, Tensor]:
+        batch_size = len(images)
+        del images, instructions, action_prompt, embodied_prompt
+        action_count = (self.config.num_video_frames - 1) * self.config.num_action_tokens_per_timestep
+        token_ids = (
+            [10]
+            + list(range(1000, 1000 + action_count))
+            + [2000] * self.config.num_embodied_action_tokens_per_instruction
+            + [11]
+        )
+        return {
+            "input_ids": torch.tensor(
+                [token_ids] * batch_size,
+                device=self.model.device,
+                dtype=torch.long,
+            )
+        }
+
+    @staticmethod
+    def tensor_to_pil(image_tensor: Tensor) -> Image.Image:
+        image = image_tensor.detach().cpu()
+        if image.ndim == 3 and image.shape[0] in (1, 3):
+            image = image.permute(1, 2, 0)
+        image = (image.float().clamp(0, 1) * 255).to(torch.uint8).numpy()
+        return Image.fromarray(image)
+
+
+class _FakeVideoEncoder(nn.Module):
+    def __init__(self, hidden_size: int = 8, tubelet_size: int = 1) -> None:
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(1))
+        # image_size must be >= patch_size (16) so the predictor grid is non-zero.
+        # Setting image_size=16 gives a 1x1 grid (1 patch per frame).
+        self.config = SimpleNamespace(hidden_size=hidden_size, tubelet_size=tubelet_size, image_size=16)
+
+    @property
+    def device(self) -> torch.device:
+        return self.weight.device
+
+    def get_vision_features(self, pixel_values_videos: Tensor) -> Tensor:
+        batch_size, num_frames = pixel_values_videos.shape[:2]
+        hidden_size = self.config.hidden_size
+        frame_values = pixel_values_videos.float().mean(dim=(2, 3, 4), keepdim=False)
+        return frame_values[:, :, None].expand(batch_size, num_frames, hidden_size)
+
+
+class _FakeVideoProcessor:
+    def __call__(self, videos, return_tensors: str) -> dict[str, Tensor]:
+        assert return_tensors == "pt"
+        if isinstance(videos, list):
+            pixel_values = torch.stack([torch.as_tensor(v) for v in videos])
+        else:
+            pixel_values = torch.as_tensor(videos).unsqueeze(0)
+        return {"pixel_values_videos": pixel_values}
+
+
+# ---------------------------------------------------------------------------
+# Fixtures
+# ---------------------------------------------------------------------------
+
+
+@pytest.fixture
+def patch_vla_jepa_external_models(monkeypatch: pytest.MonkeyPatch) -> None:
+    from lerobot.policies.vla_jepa import modeling_vla_jepa
+
+    monkeypatch.setattr(modeling_vla_jepa, "Qwen3VLInterface", _FakeQwenInterface)
+    monkeypatch.setattr(
+        modeling_vla_jepa.AutoModel,
+        "from_pretrained",
+        lambda *args, **kwargs: _FakeVideoEncoder(),
+    )
+    monkeypatch.setattr(
+        modeling_vla_jepa.AutoVideoProcessor,
+        "from_pretrained",
+        lambda *args, **kwargs: _FakeVideoProcessor(),
+    )

tests/policies/vla_jepa/test_action_head.py

@@ -0,0 +1,157 @@
+#!/usr/bin/env python
+
+from __future__ import annotations
+
+import pytest
+import torch
+
+pytest.importorskip("diffusers")
+
+from conftest import (
+    ACTION_DIM,
+    ACTION_HORIZON,
+    BATCH_SIZE,
+    QWEN_HIDDEN_SIZE,
+    STATE_DIM,
+    make_config,
+    set_seed_all,
+)  # noqa: E402
+
+from lerobot.policies.vla_jepa.action_head import (  # noqa: E402
+    VLAJEPAActionHead,
+)
+
+# ---------------------------------------------------------------------------
+# VLAJEPAActionHead
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.parametrize(
+    "action_dim,state_dim,action_horizon",
+    [
+        (3, 4, 4),  # default test dims
+        (7, 0, 16),  # no proprioceptive state, production-like action space
+        (6, 8, 8),  # medium dims
+    ],
+)
+def test_action_head_sample_time_range(action_dim: int, state_dim: int, action_horizon: int) -> None:
+    config = make_config(action_dim=action_dim, state_dim=state_dim, action_horizon=action_horizon)
+    head = VLAJEPAActionHead(config, cross_attention_dim=QWEN_HIDDEN_SIZE)
+    t = head.sample_time(batch_size=200, device=torch.device("cpu"), dtype=torch.float32)
+    assert t.shape == (200,)
+    assert torch.isfinite(t).all()
+
+
+@pytest.mark.parametrize(
+    "action_dim,state_dim,action_horizon",
+    [
+        (3, 4, 4),
+        (7, 0, 16),
+        (6, 8, 8),
+    ],
+)
+def test_action_head_build_inputs_shape(action_dim: int, state_dim: int, action_horizon: int) -> None:
+    config = make_config(action_dim=action_dim, state_dim=state_dim, action_horizon=action_horizon)
+    head = VLAJEPAActionHead(config, cross_attention_dim=QWEN_HIDDEN_SIZE)
+    conditioning = torch.randn(2, 4, QWEN_HIDDEN_SIZE)
+    actions = torch.randn(2, action_horizon, action_dim)
+    timesteps = torch.randint(0, 100, (2,))
+
+    state = torch.randn(2, state_dim) if state_dim > 0 else None
+    out_with = head._build_inputs(conditioning, actions, state, timesteps)
+    out_none = head._build_inputs(conditioning, actions, None, timesteps)
+
+    assert out_with.ndim == 3 and out_none.ndim == 3
+    if state_dim > 0:
+        assert out_with.shape[1] > out_none.shape[1]
+    assert torch.isfinite(out_with).all() and torch.isfinite(out_none).all()
+
+
+@pytest.mark.parametrize(
+    "action_dim,state_dim,action_horizon",
+    [
+        (3, 4, 4),
+        (7, 0, 16),
+        (6, 8, 8),
+    ],
+)
+def test_action_head_forward_loss_valid(action_dim: int, state_dim: int, action_horizon: int) -> None:
+    set_seed_all(42)
+    config = make_config(action_dim=action_dim, state_dim=state_dim, action_horizon=action_horizon)
+    head = VLAJEPAActionHead(config, cross_attention_dim=QWEN_HIDDEN_SIZE)
+    conditioning = torch.randn(2, 4, QWEN_HIDDEN_SIZE)
+    actions = torch.randn(2, action_horizon, action_dim)
+    state = torch.randn(2, state_dim) if state_dim > 0 else None
+    loss = head.forward(conditioning, actions, state)
+    assert loss.shape == ()
+    assert torch.isfinite(loss) and loss > 0
+
+
+def test_action_head_forward_gradient_flows() -> None:
+    set_seed_all(42)
+    config = make_config()
+    head = VLAJEPAActionHead(config, cross_attention_dim=QWEN_HIDDEN_SIZE)
+    conditioning = torch.randn(BATCH_SIZE, 4, QWEN_HIDDEN_SIZE)
+    actions = torch.randn(BATCH_SIZE, ACTION_HORIZON, ACTION_DIM)
+    state = torch.randn(BATCH_SIZE, STATE_DIM)
+    loss = head.forward(conditioning, actions, state)
+    loss.backward()
+    assert any(p.grad is not None for p in head.parameters() if p.requires_grad)
+
+
+@torch.no_grad()
+@pytest.mark.parametrize(
+    "action_dim,state_dim,action_horizon",
+    [
+        (3, 4, 4),
+        (7, 0, 16),
+        (6, 8, 8),
+    ],
+)
+def test_action_head_predict_action_shape(action_dim: int, state_dim: int, action_horizon: int) -> None:
+    set_seed_all(42)
+    config = make_config(action_dim=action_dim, state_dim=state_dim, action_horizon=action_horizon)
+    head = VLAJEPAActionHead(config, cross_attention_dim=QWEN_HIDDEN_SIZE)
+    conditioning = torch.randn(2, 4, QWEN_HIDDEN_SIZE)
+    state = torch.randn(2, state_dim) if state_dim > 0 else None
+    pred = head.predict_action(conditioning, state)
+    assert tuple(pred.shape) == (2, action_horizon, action_dim)
+    assert torch.isfinite(pred).all()
+
+
+# ---------------------------------------------------------------------------
+# action_is_pad masking
+# ---------------------------------------------------------------------------
+
+
+def test_action_head_loss_fully_padded_is_zero() -> None:
+    """Loss is 0 when every timestep is padded (exercises the clamp_min guard)."""
+    set_seed_all(42)
+    config = make_config()
+    head = VLAJEPAActionHead(config, cross_attention_dim=QWEN_HIDDEN_SIZE)
+    conditioning = torch.randn(BATCH_SIZE, 4, QWEN_HIDDEN_SIZE)
+    actions = torch.randn(BATCH_SIZE, ACTION_HORIZON, ACTION_DIM)
+    state = torch.randn(BATCH_SIZE, STATE_DIM)
+
+    action_is_pad = torch.ones(BATCH_SIZE, ACTION_HORIZON, dtype=torch.bool)
+    loss = head.forward(conditioning, actions, state, action_is_pad)
+    assert loss.item() == 0.0
+
+
+def test_action_head_loss_none_matches_no_padding() -> None:
+    """action_is_pad=None is equivalent to an all-False (no padding) mask."""
+    set_seed_all(42)
+    config = make_config()
+    head = VLAJEPAActionHead(config, cross_attention_dim=QWEN_HIDDEN_SIZE)
+    conditioning = torch.randn(BATCH_SIZE, 4, QWEN_HIDDEN_SIZE)
+    actions = torch.randn(BATCH_SIZE, ACTION_HORIZON, ACTION_DIM)
+    state = torch.randn(BATCH_SIZE, STATE_DIM)
+
+    set_seed_all(0)
+    loss_none = head.forward(conditioning, actions, state, action_is_pad=None)
+
+    set_seed_all(0)
+    no_pad = torch.zeros(BATCH_SIZE, ACTION_HORIZON, dtype=torch.bool)
+    loss_zeros = head.forward(conditioning, actions, state, action_is_pad=no_pad)
+
+    assert torch.isclose(loss_none, loss_zeros)

tests/policies/vla_jepa/test_configuration.py

@@ -0,0 +1,57 @@
+#!/usr/bin/env python
+
+from __future__ import annotations
+
+import pytest
+from conftest import ACTION_DIM, ACTION_HORIZON, IMAGE_SIZE, NUM_VIDEO_FRAMES, STATE_DIM, make_config
+
+from lerobot.configs.types import FeatureType, PolicyFeature
+from lerobot.policies.vla_jepa.configuration_vla_jepa import VLAJEPAConfig
+from lerobot.utils.constants import ACTION, OBS_IMAGES, OBS_STATE
+
+
+def test_delta_indices() -> None:
+    config = make_config()
+    assert config.observation_delta_indices == list(range(NUM_VIDEO_FRAMES))
+    assert config.action_delta_indices == list(range(ACTION_HORIZON))
+
+
+def test_n_action_steps_exceeds_chunk_size_raises() -> None:
+    with pytest.raises(ValueError, match="n_action_steps"):
+        VLAJEPAConfig(chunk_size=4, n_action_steps=8)
+
+
+def test_too_few_video_frames_raises() -> None:
+    with pytest.raises(ValueError, match="video_horizon"):
+        VLAJEPAConfig(
+            chunk_size=16,
+            n_action_steps=16,
+            num_video_frames=2,
+            jepa_tubelet_size=2,  # needs >= 4 frames (2 for current, 2 for future) to have a window of size > 0
+        )
+
+
+def test_validate_features_no_image_raises() -> None:
+    config = VLAJEPAConfig(
+        input_features={OBS_STATE: PolicyFeature(type=FeatureType.STATE, shape=(STATE_DIM,))},
+        output_features={ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(ACTION_DIM,))},
+    )
+    with pytest.raises(ValueError, match="at least one visual input feature"):
+        config.validate_features()
+
+
+def test_validate_features_no_action_raises() -> None:
+    config = VLAJEPAConfig(
+        input_features={
+            f"{OBS_IMAGES}.cam": PolicyFeature(type=FeatureType.VISUAL, shape=(3, IMAGE_SIZE, IMAGE_SIZE)),
+        },
+        output_features={},
+    )
+    with pytest.raises(ValueError, match="action output feature"):
+        config.validate_features()
+
+
+def test_validate_features_sets_action_dim_from_feature() -> None:
+    config = make_config(action_dim=6, state_dim=10)
+    assert config.action_dim == 6
+    assert config.state_dim == 10

tests/policies/vla_jepa/test_vla_jepa.py

@@ -0,0 +1,598 @@
+#!/usr/bin/env python
+
+from __future__ import annotations
+
+import os
+from copy import deepcopy
+
+import numpy as np
+import pytest
+import torch
+from torch import Tensor
+
+pytest.importorskip("transformers")
+pytest.importorskip("diffusers")
+
+pytestmark = pytest.mark.filterwarnings(
+    "ignore:In CPU autocast, but the target dtype is not supported:UserWarning"
+)
+
+from conftest import (  # noqa: E402
+    ACTION_DIM,
+    ACTION_HORIZON,
+    BATCH_SIZE,
+    EXPECTED_ACTION_CHUNK_SHAPE,
+    EXPECTED_SELECT_ACTION_SHAPE,
+    IMAGE_SIZE,
+    N_ACTION_STEPS,
+    QWEN_HIDDEN_SIZE,
+    STATE_DIM,
+    make_config,
+    make_inference_batch,
+    make_train_batch,
+    set_seed_all,
+)
+
+from lerobot.policies.vla_jepa.configuration_vla_jepa import VLAJEPAConfig  # noqa: E402
+from lerobot.policies.vla_jepa.modeling_vla_jepa import VLAJEPAPolicy  # noqa: E402
+from lerobot.utils.constants import ACTION  # noqa: E402
+
+PRETRAINED_REPO_ID = "ginwind/VLA-JEPA"
+PRETRAINED_SUBFOLDER = "LIBERO"
+
+# extended hub tests load the full converted safetensors checkpoints (~5 GB) and are
+# skipped by default.  Set VLA_JEPA_EXTENDED=1 to opt in.
+_VLA_JEPA_EXTENDED = os.environ.get("VLA_JEPA_EXTENDED", "0") != "0"
+extended_test = pytest.mark.skipif(not _VLA_JEPA_EXTENDED, reason="Set VLA_JEPA_EXTENDED=1 to run hub tests")
+
+
+# ---------------------------------------------------------------------------
+# Core training / inference tests
+# ---------------------------------------------------------------------------
+
+
+def test_training_forward_pass(patch_vla_jepa_external_models: None) -> None:
+    set_seed_all(42)
+    policy = VLAJEPAPolicy(make_config())
+    policy.train()
+
+    batch = make_train_batch()
+    batch_before = deepcopy(batch)
+
+    loss, logs = policy.forward(batch)
+
+    assert loss.shape == ()
+    assert torch.isfinite(loss)
+    assert set(logs) == {"action_loss", "wm_loss", "loss"}
+    assert logs["action_loss"] > 0
+    assert logs["wm_loss"] >= 0
+
+    loss.backward()
+    assert any(p.grad is not None for p in policy.model.action_model.parameters() if p.requires_grad)
+    # Batch must not be mutated.
+    assert set(batch) == set(batch_before)
+    for key, value in batch.items():
+        if isinstance(value, Tensor):
+            assert torch.equal(value, batch_before[key])
+        else:
+            assert value == batch_before[key]
+
+
+@pytest.mark.parametrize("batch_size", [1, 2, 4])
+def test_training_forward_various_batch_sizes(patch_vla_jepa_external_models: None, batch_size: int) -> None:
+    set_seed_all(42)
+    policy = VLAJEPAPolicy(make_config())
+    policy.train()
+    loss, logs = policy.forward(make_train_batch(batch_size=batch_size))
+    assert torch.isfinite(loss) and loss > 0
+    assert set(logs) == {"action_loss", "wm_loss", "loss"}
+
+
+@pytest.mark.parametrize(
+    "action_dim,state_dim,action_horizon",
+    [
+        (3, 4, 4),
+        (7, 0, 16),
+        (6, 8, 8),
+    ],
+)
+def test_training_forward_various_dims(
+    patch_vla_jepa_external_models: None,
+    action_dim: int,
+    state_dim: int,
+    action_horizon: int,
+) -> None:
+    set_seed_all(42)
+    config = make_config(action_dim=action_dim, state_dim=state_dim, action_horizon=action_horizon)
+    policy = VLAJEPAPolicy(config)
+    policy.train()
+    batch = make_train_batch(action_dim=action_dim, state_dim=state_dim, action_horizon=action_horizon)
+    loss, _ = policy.forward(batch)
+    assert torch.isfinite(loss) and loss > 0
+
+
+@torch.no_grad()
+def test_action_generation_shape(patch_vla_jepa_external_models: None) -> None:
+    set_seed_all(42)
+    policy = VLAJEPAPolicy(make_config())
+    policy.eval()
+    batch = make_inference_batch()
+
+    chunk = policy.predict_action_chunk(batch)
+    assert tuple(chunk.shape) == EXPECTED_ACTION_CHUNK_SHAPE
+    assert chunk.device.type == "cpu"
+    assert torch.isfinite(chunk).all()
+
+    a1 = policy.select_action(batch)
+    a2 = policy.select_action(batch)
+    assert tuple(a1.shape) == EXPECTED_SELECT_ACTION_SHAPE
+    assert tuple(a2.shape) == EXPECTED_SELECT_ACTION_SHAPE
+    assert torch.isfinite(a1).all() and torch.isfinite(a2).all()
+
+
+@torch.no_grad()
+@pytest.mark.parametrize("action_dim,state_dim", [(3, 4), (7, 0), (6, 8)])
+def test_action_generation_various_dims(
+    patch_vla_jepa_external_models: None, action_dim: int, state_dim: int
+) -> None:
+    set_seed_all(42)
+    config = make_config(action_dim=action_dim, state_dim=state_dim)
+    policy = VLAJEPAPolicy(config)
+    policy.eval()
+    batch = make_inference_batch(state_dim=state_dim)
+    chunk = policy.predict_action_chunk(batch)
+    assert chunk.shape[-1] == action_dim
+    assert torch.isfinite(chunk).all()
+
+
+@torch.no_grad()
+def test_inference_reproducibility(patch_vla_jepa_external_models: None) -> None:
+    set_seed_all(42)
+    policy = VLAJEPAPolicy(make_config())
+    policy.eval()
+    batch = make_inference_batch()
+
+    set_seed_all(123)
+    actions_1 = policy.predict_action_chunk(batch)
+    set_seed_all(123)
+    actions_2 = policy.predict_action_chunk(batch)
+
+    assert tuple(actions_1.shape) == EXPECTED_ACTION_CHUNK_SHAPE
+    assert torch.allclose(actions_1, actions_2, atol=1e-6)
+
+
+@torch.no_grad()
+def test_predict_action_chunk_always_finite(patch_vla_jepa_external_models: None) -> None:
+    policy = VLAJEPAPolicy(make_config())
+    policy.eval()
+    for seed in [0, 42, 123]:
+        set_seed_all(seed)
+        chunk = policy.predict_action_chunk(make_inference_batch())
+        assert torch.isfinite(chunk).all(), f"non-finite actions with seed={seed}"
+
+
+# ---------------------------------------------------------------------------
+# Action queue behaviour
+# ---------------------------------------------------------------------------
+
+
+@torch.no_grad()
+def test_select_action_queue_drains_before_refill(patch_vla_jepa_external_models: None) -> None:
+    set_seed_all(42)
+    policy = VLAJEPAPolicy(make_config())
+    policy.eval()
+    batch = make_inference_batch()
+
+    # First call fills the queue (n_action_steps items) and pops one.
+    a1 = policy.select_action(batch)
+    assert len(policy._queues[ACTION]) == N_ACTION_STEPS - 1
+
+    # Second call pops from the existing queue without calling predict_action_chunk.
+    a2 = policy.select_action(batch)
+    assert tuple(a1.shape) == EXPECTED_SELECT_ACTION_SHAPE
+    assert tuple(a2.shape) == EXPECTED_SELECT_ACTION_SHAPE
+
+
+@torch.no_grad()
+def test_reset_clears_action_queue(patch_vla_jepa_external_models: None) -> None:
+    set_seed_all(42)
+    policy = VLAJEPAPolicy(make_config())
+    policy.eval()
+    policy.select_action(make_inference_batch())
+    assert len(policy._queues[ACTION]) > 0
+
+    policy.reset()
+    assert len(policy._queues[ACTION]) == 0
+
+
+# ---------------------------------------------------------------------------
+# Format conversion
+# ---------------------------------------------------------------------------
+
+
+def test_prepare_model_inputs_training_format(patch_vla_jepa_external_models: None) -> None:
+    from PIL import Image
+
+    policy = VLAJEPAPolicy(make_config())
+    examples = policy._prepare_model_inputs(make_train_batch())
+
+    assert len(examples) == BATCH_SIZE
+    for ex in examples:
+        assert set(ex) >= {"image", "video", "lang", "action", "state"}
+        assert len(ex["image"]) == 1 and isinstance(ex["image"][0], Image.Image)
+        assert ex["video"].ndim == 5 and ex["video"].dtype == np.uint8  # [V,T,H,W,C]
+        assert ex["action"].shape == (ACTION_HORIZON, ACTION_DIM)
+        assert ex["state"].shape == (1, STATE_DIM)
+
+
+def test_prepare_model_inputs_inference_omits_action(patch_vla_jepa_external_models: None) -> None:
+    policy = VLAJEPAPolicy(make_config())
+    for ex in policy._prepare_model_inputs(make_inference_batch()):
+        assert "action" not in ex
+        assert "image" in ex and "video" in ex and "lang" in ex
+
+
+def test_prepare_model_inputs_missing_task_uses_default(patch_vla_jepa_external_models: None) -> None:
+    policy = VLAJEPAPolicy(make_config())
+    batch = make_inference_batch()
+    del batch["task"]
+    examples = policy._prepare_model_inputs(batch)
+    assert all(isinstance(ex["lang"], str) and len(ex["lang"]) > 0 for ex in examples)
+
+
+def test_prepare_model_inputs_string_task_broadcast(patch_vla_jepa_external_models: None) -> None:
+    policy = VLAJEPAPolicy(make_config())
+    batch = make_inference_batch()
+    batch["task"] = "open the drawer"
+    assert all(ex["lang"] == "open the drawer" for ex in policy._prepare_model_inputs(batch))
+
+
+def test_prepare_model_inputs_no_state_omitted(patch_vla_jepa_external_models: None) -> None:
+    from lerobot.utils.constants import OBS_STATE
+
+    policy = VLAJEPAPolicy(make_config())
+    batch = make_inference_batch()
+    del batch[OBS_STATE]
+    assert all("state" not in ex for ex in policy._prepare_model_inputs(batch))
+
+
+# ---------------------------------------------------------------------------
+# Pretrained checkpoint
+# Hub tests (opt-in: VLA_JEPA_EXTENDED=1)
+# ---------------------------------------------------------------------------
+
+
+def _make_hub_train_batch(policy: VLAJEPAPolicy, batch_size: int = 1) -> dict:
+    """Build a training batch whose keys/shapes match a hub-loaded policy config."""
+    cfg = policy.config
+    batch: dict = {"task": ["pick up the cube"] * batch_size}
+    for key, feat in cfg.image_features.items():
+        h, w = feat.shape[-2], feat.shape[-1]
+        batch[key] = torch.rand(batch_size, cfg.num_video_frames, 3, h, w)
+    if cfg.robot_state_feature is not None:
+        batch["observation.state"] = torch.randn(batch_size, 1, cfg.robot_state_feature.shape[0])
+    batch[ACTION] = torch.randn(batch_size, cfg.chunk_size, cfg.action_dim)
+    return batch
+
+
+def _make_hub_inference_batch(policy: VLAJEPAPolicy, batch_size: int = 1) -> dict:
+    """Build an inference batch whose keys/shapes match a hub-loaded policy config."""
+    cfg = policy.config
+    batch: dict = {"task": ["pick up the cube"] * batch_size}
+    for key, feat in cfg.image_features.items():
+        h, w = feat.shape[-2], feat.shape[-1]
+        batch[key] = torch.rand(batch_size, 3, h, w)
+    if cfg.robot_state_feature is not None:
+        batch["observation.state"] = torch.randn(batch_size, cfg.robot_state_feature.shape[0])
+    return batch
+
+
+_CP_ROOT = "lerobot"
+
+# Each tuple: (repo_id, enable_world_model)
+_HUB_VARIANTS = [
+    (f"{_CP_ROOT}/VLA-JEPA-LIBERO", True),
+    (f"{_CP_ROOT}/VLA-JEPA-Pretrain", True),
+    (f"{_CP_ROOT}/VLA-JEPA-SimplerEnv", False),
+]
+
+
+@extended_test
+@pytest.mark.parametrize("repo_id,enable_world_model", _HUB_VARIANTS)
+def test_hub_checkpoint_loads(repo_id: str, enable_world_model: bool) -> None:
+    """Policy loads from the converted safetensors checkpoint on the Hub."""
+    policy = VLAJEPAPolicy.from_pretrained(repo_id)
+    assert policy.config.enable_world_model == enable_world_model
+    assert sum(p.numel() for p in policy.parameters()) > 0
+
+
+@extended_test
+@pytest.mark.parametrize("repo_id,enable_world_model", _HUB_VARIANTS)
+def test_hub_checkpoint_forward_pass(repo_id: str, enable_world_model: bool) -> None:
+    """Policy loaded from hub produces finite losses with a correctly-shaped batch."""
+    policy = VLAJEPAPolicy.from_pretrained(repo_id)
+    policy.train()
+
+    batch = _make_hub_train_batch(policy)
+    loss, logs = policy.forward(batch)
+    assert torch.isfinite(loss)
+    assert "action_loss" in logs
+    if enable_world_model:
+        assert "wm_loss" in logs
+
+
+@extended_test
+def test_hub_freeze_qwen_disables_world_model() -> None:
+    """freeze_qwen=True (via cli_overrides) freezes qwen and disables the world model."""
+    policy = VLAJEPAPolicy.from_pretrained(f"{_CP_ROOT}/VLA-JEPA-LIBERO", cli_overrides=["freeze_qwen=true"])
+    assert not policy.config.enable_world_model
+    assert policy.model.video_predictor is None
+    qwen_params = list(policy.model.qwen.parameters())
+    assert all(not p.requires_grad for p in qwen_params)
+    assert any(p.requires_grad for p in policy.model.action_model.parameters())
+
+
+@extended_test
+def test_hub_disable_world_model_loads_simpler_env() -> None:
+    """SimplerEnv checkpoint (world model disabled) loads cleanly and runs inference."""
+    policy = VLAJEPAPolicy.from_pretrained(f"{_CP_ROOT}/VLA-JEPA-SimplerEnv")
+    assert not policy.config.enable_world_model
+    assert policy.model.video_predictor is None
+    assert policy.model.video_encoder is None
+
+
+@extended_test
+def test_hub_libero_inference_shape() -> None:
+    """select_action returns the expected shape using the LIBERO hub checkpoint."""
+    policy = VLAJEPAPolicy.from_pretrained(f"{_CP_ROOT}/VLA-JEPA-LIBERO")
+    policy.eval()
+    batch = _make_hub_inference_batch(policy)
+    action = policy.select_action(batch)
+    assert action.shape[-1] == policy.config.action_dim
+
+
+# ---------------------------------------------------------------------------
+# Postprocessor unnormalization tests
+#
+# These tests verify that the postprocessor pipeline (clip → unnorm → binarize)
+# correctly applies MIN_MAX unnormalization after predict_action_chunk.
+# ---------------------------------------------------------------------------
+
+
+def _make_dataset_stats(action_dim: int = ACTION_DIM) -> dict:
+    """Returns sample dataset_stats with a simple [i, i+10] range per action dim."""
+    from lerobot.utils.constants import ACTION
+
+    return {
+        ACTION: {
+            "min": torch.tensor([float(i) for i in range(action_dim)], dtype=torch.float32),
+            "max": torch.tensor([float(i) + 10.0 for i in range(action_dim)], dtype=torch.float32),
+        }
+    }
+
+
+@torch.no_grad()
+def test_postprocessor_unnormalizes_actions(patch_vla_jepa_external_models: None) -> None:
+    """UnnormalizerProcessorStep with MIN_MAX produces the correct inverse of MIN_MAX normalization."""
+    from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
+    from lerobot.processor import UnnormalizerProcessorStep
+    from lerobot.processor.converters import policy_action_to_transition, transition_to_policy_action
+    from lerobot.utils.constants import ACTION
+
+    dataset_stats = _make_dataset_stats()
+
+    rng = np.random.default_rng(7)
+    actions_np = rng.uniform(-1.0, 1.0, (2, ACTION_HORIZON, ACTION_DIM)).astype(np.float32)
+
+    a_min = dataset_stats[ACTION]["min"].numpy()
+    a_max = dataset_stats[ACTION]["max"].numpy()
+    expected = (actions_np + 1.0) / 2.0 * (a_max - a_min) + a_min
+
+    features = {ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(ACTION_DIM,))}
+    unnorm_step = UnnormalizerProcessorStep(
+        features=features,
+        norm_map={FeatureType.ACTION: NormalizationMode.MIN_MAX},
+        stats=dataset_stats,
+    )
+
+    actions_tensor = torch.from_numpy(actions_np)
+    transition = policy_action_to_transition(actions_tensor)
+    result = transition_to_policy_action(unnorm_step(transition)).numpy()
+
+    np.testing.assert_allclose(result, expected, rtol=1e-5, atol=1e-6)
+
+
+@torch.no_grad()
+def test_postprocessor_clip_clamps_before_unnorm(patch_vla_jepa_external_models: None) -> None:
+    """ClipActionsProcessorStep clamps to [-1, 1] before unnormalization."""
+    from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
+    from lerobot.policies.vla_jepa.processor_vla_jepa import ClipActionsProcessorStep
+    from lerobot.processor import UnnormalizerProcessorStep
+    from lerobot.processor.converters import policy_action_to_transition, transition_to_policy_action
+    from lerobot.utils.constants import ACTION
+
+    dataset_stats = _make_dataset_stats()
+    a_min = dataset_stats[ACTION]["min"].numpy()
+    a_max = dataset_stats[ACTION]["max"].numpy()
+
+    # Deliberately out-of-range inputs
+    actions_np = np.array([[[2.0] * ACTION_DIM, [-3.0] * ACTION_DIM]], dtype=np.float32)
+    clipped = np.clip(actions_np, -1.0, 1.0)
+    expected = (clipped + 1.0) / 2.0 * (a_max - a_min) + a_min
+
+    features = {ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(ACTION_DIM,))}
+    clip_step = ClipActionsProcessorStep()
+    unnorm_step = UnnormalizerProcessorStep(
+        features=features,
+        norm_map={FeatureType.ACTION: NormalizationMode.MIN_MAX},
+        stats=dataset_stats,
+    )
+
+    transition = policy_action_to_transition(torch.from_numpy(actions_np))
+    transition = clip_step(transition)
+    result = transition_to_policy_action(unnorm_step(transition)).numpy()
+
+    np.testing.assert_allclose(result, expected, rtol=1e-5, atol=1e-6)
+
+
+@torch.no_grad()
+def test_postprocessor_applied_after_predict_action_chunk(
+    patch_vla_jepa_external_models: None, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    """predict_action_chunk returns raw actions; the postprocessor applies unnormalization.
+
+    Verifies the split: predict_action_chunk returns normalized actions, and calling the
+    postprocessor on them produces the correctly unnormalized result.
+    """
+    from lerobot.policies.vla_jepa.processor_vla_jepa import make_vla_jepa_pre_post_processors
+
+    raw_actions = np.zeros((BATCH_SIZE, ACTION_HORIZON, ACTION_DIM), dtype=np.float32)
+
+    cfg = make_config()
+    cfg.clip_normalized_actions = False
+    cfg.binarize_gripper_action = False
+    policy = VLAJEPAPolicy(cfg)
+    policy.eval()
+    monkeypatch.setattr(policy.model, "predict_action", lambda *a, **kw: raw_actions.copy())
+
+    dataset_stats = _make_dataset_stats()
+    _, postprocessor = make_vla_jepa_pre_post_processors(cfg, dataset_stats)
+
+    batch = make_inference_batch()
+    chunk = policy.predict_action_chunk(batch)
+
+    # predict_action_chunk returns raw (normalized) actions
+    assert torch.allclose(chunk, torch.zeros_like(chunk), atol=1e-6), (
+        "predict_action_chunk should return raw actions without unnormalization applied."
+    )
+
+    # Postprocessor applies unnormalization: 0 → (0+1)/2 * (max-min) + min = 5 + i
+    unnormed = postprocessor(chunk)
+    from lerobot.utils.constants import ACTION
+
+    a_min = dataset_stats[ACTION]["min"].numpy()
+    a_max = dataset_stats[ACTION]["max"].numpy()
+    expected_first = 0.5 * (0.0 + 1.0) * (a_max[0] - a_min[0]) + a_min[0]
+    assert unnormed[0, 0, 0].item() == pytest.approx(expected_first, abs=1e-5)
+
+
+# ---------------------------------------------------------------------------
+# World-model view adjustment (padding / trimming) tests
+# ---------------------------------------------------------------------------
+
+
+_MULTIVIEW_NUM_FRAMES = 4  # must be >= 2 * jepa_tubelet_size (=2) for world-model tests
+
+
+def _make_multiview_config(num_views: int, jepa_tubelet_size: int = 2) -> VLAJEPAConfig:
+    from lerobot.configs.types import FeatureType, PolicyFeature
+    from lerobot.utils.constants import OBS_IMAGES, OBS_STATE
+
+    config = VLAJEPAConfig(
+        input_features={
+            **{
+                f"{OBS_IMAGES}.cam{i}": PolicyFeature(
+                    type=FeatureType.VISUAL, shape=(3, IMAGE_SIZE, IMAGE_SIZE)
+                )
+                for i in range(num_views)
+            },
+            OBS_STATE: PolicyFeature(type=FeatureType.STATE, shape=(STATE_DIM,)),
+        },
+        output_features={ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(ACTION_DIM,))},
+        device="cpu",
+        chunk_size=ACTION_HORIZON,
+        n_action_steps=N_ACTION_STEPS,
+        action_dim=ACTION_DIM,
+        state_dim=STATE_DIM,
+        num_video_frames=_MULTIVIEW_NUM_FRAMES,
+        num_action_tokens_per_timestep=2,
+        num_embodied_action_tokens_per_instruction=3,
+        num_inference_timesteps=2,
+        action_hidden_size=QWEN_HIDDEN_SIZE,
+        action_model_type="DiT-test",
+        action_num_layers=1,
+        predictor_depth=1,
+        predictor_num_heads=2,
+        predictor_mlp_ratio=2.0,
+        jepa_tubelet_size=jepa_tubelet_size,
+    )
+    config.validate_features()
+    return config
+
+
+def _make_multiview_train_batch(num_views: int, batch_size: int = BATCH_SIZE) -> dict:
+    from lerobot.utils.constants import OBS_IMAGES, OBS_STATE
+
+    batch = {
+        f"{OBS_IMAGES}.cam{i}": torch.rand(batch_size, _MULTIVIEW_NUM_FRAMES, 3, IMAGE_SIZE, IMAGE_SIZE)
+        for i in range(num_views)
+    }
+    batch[OBS_STATE] = torch.randn(batch_size, 1, STATE_DIM)
+    batch[ACTION] = torch.randn(batch_size, ACTION_HORIZON, ACTION_DIM)
+    batch["task"] = ["pick up the cube"] * batch_size
+    return batch
+
+
+@pytest.mark.parametrize(
+    "num_views",
+    [
+        1,  # fewer views than jepa_tubelet_size → first view duplicated
+        2,  # exact match → unchanged
+        3,  # more views than jepa_tubelet_size → trimmed to first two
+    ],
+)
+def test_training_forward_world_model_view_adjustment(
+    patch_vla_jepa_external_models: None,
+    num_views: int,
+) -> None:
+    """World-model view padding/trimming must not break the training forward pass."""
+    set_seed_all(42)
+    policy = VLAJEPAPolicy(_make_multiview_config(num_views=num_views, jepa_tubelet_size=2))
+    policy.train()
+    loss, logs = policy.forward(_make_multiview_train_batch(num_views=num_views))
+    assert torch.isfinite(loss)
+    assert logs["wm_loss"] >= 0
+
+
+def test_single_view_is_duplicated_for_world_model(patch_vla_jepa_external_models: None) -> None:
+    """With one dataset view and jepa_tubelet_size=2, the view must be duplicated before encoding."""
+    set_seed_all(42)
+    policy = VLAJEPAPolicy(_make_multiview_config(num_views=1, jepa_tubelet_size=2))
+    policy.train()
+
+    captured_videos: list = []
+    original_processor = policy.model.video_processor
+
+    class _CapturingProcessor:
+        def __call__(self, videos: list, return_tensors: str) -> dict:
+            captured_videos.extend(videos)
+            return original_processor(videos=videos, return_tensors=return_tensors)
+
+    policy.model.video_processor = _CapturingProcessor()
+    policy.forward(_make_multiview_train_batch(num_views=1))
+
+    # reshape is batch-major: (b0v0, b0v1, b1v0, b1v1, …)
+    assert len(captured_videos) == BATCH_SIZE * 2
+    for i in range(BATCH_SIZE):
+        np.testing.assert_array_equal(captured_videos[2 * i], captured_videos[2 * i + 1])
+
+
+def test_excess_views_trimmed_for_world_model(patch_vla_jepa_external_models: None) -> None:
+    """With three dataset views and jepa_tubelet_size=2, only the first two views reach the encoder."""
+    set_seed_all(42)
+    policy = VLAJEPAPolicy(_make_multiview_config(num_views=3, jepa_tubelet_size=2))
+    policy.train()
+
+    captured_videos: list = []
+    original_processor = policy.model.video_processor
+
+    class _CapturingProcessor:
+        def __call__(self, videos: list, return_tensors: str) -> dict:
+            captured_videos.extend(videos)
+            return original_processor(videos=videos, return_tensors=return_tensors)
+
+    policy.model.video_processor = _CapturingProcessor()
+    policy.forward(_make_multiview_train_batch(num_views=3))
+
+    # Only B*2 items must reach the encoder, not B*3.
+    assert len(captured_videos) == BATCH_SIZE * 2

tests/policies/vla_jepa/test_world_model.py

@@ -0,0 +1,60 @@
+#!/usr/bin/env python
+
+from __future__ import annotations
+
+import pytest
+import torch
+
+from lerobot.policies.vla_jepa.world_model import (
+    ActionConditionedVideoPredictor,
+)
+
+_ACTION_EMBED_DIM = 8
+
+
+def _make_predictor(
+    embed_dim: int = 8,
+    action_embed_dim: int = _ACTION_EMBED_DIM,
+    predictor_embed_dim: int = 24,
+    num_action_tokens: int = 2,
+    tokens_per_frame: int = 1,
+) -> ActionConditionedVideoPredictor:
+    return ActionConditionedVideoPredictor(
+        num_frames=1,
+        img_size=(1, tokens_per_frame),
+        patch_size=1,
+        tubelet_size=1,
+        embed_dim=embed_dim,
+        action_embed_dim=action_embed_dim,
+        predictor_embed_dim=predictor_embed_dim,
+        depth=1,
+        num_heads=2,
+        mlp_ratio=2.0,
+        num_action_tokens_per_step=num_action_tokens,
+    )
+
+
+@pytest.mark.parametrize(
+    "batch,num_steps,tokens_per_frame,embed_dim",
+    [
+        (1, 2, 1, 8),
+        (2, 3, 4, 8),
+        (4, 5, 2, 16),
+    ],
+)
+def test_predictor_output_shape(batch: int, num_steps: int, tokens_per_frame: int, embed_dim: int) -> None:
+    predictor = _make_predictor(
+        embed_dim=embed_dim, action_embed_dim=_ACTION_EMBED_DIM, tokens_per_frame=tokens_per_frame
+    )
+    frame_tokens = torch.randn(batch, num_steps * tokens_per_frame, embed_dim)
+    action_tokens = torch.randn(batch, num_steps * 2, _ACTION_EMBED_DIM)
+    out = predictor(frame_tokens, action_tokens)
+    assert tuple(out.shape) == (batch, num_steps * tokens_per_frame, embed_dim)
+    assert torch.isfinite(out).all()
+
+
+def test_predictor_step_mismatch_raises() -> None:
+    predictor = _make_predictor(tokens_per_frame=4)
+    frame_tokens = torch.randn(2, 3 * 4, 8)  # 3 steps, 4 tokens each
+    with pytest.raises(RuntimeError):
+        predictor(frame_tokens, torch.randn(2, 2 * 2, 8))  # 2 steps → mismatch

tests/processor/test_pipeline.py

@@ -24,6 +24,7 @@ from typing import Any
 import pytest
 import torch
 import torch.nn as nn
+from safetensors.torch import load_file
 
 pytest.importorskip("datasets", reason="datasets is required (install lerobot[dataset])")
 
@@ -174,6 +175,53 @@ class MockStepWithTensorState(ProcessorStep):
         return features
 
 
+class MockLazyTensorStateStep(ProcessorStep):
+    """Mock step whose tensor state is not present in constructor config."""
+
+    def __init__(
+        self, name: str = "lazy_tensor_step", scale: float = 1.0, initial_value: float | None = None
+    ):
+        self.name = name
+        self.scale = scale
+        self.tensor_state: torch.Tensor | None = None
+
+        if initial_value is not None:
+            self.tensor_state = torch.tensor([initial_value], dtype=torch.float32)
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        """Return the transition unchanged."""
+        return transition
+
+    def get_config(self) -> dict[str, Any]:
+        """Return constructor config while intentionally omitting tensor state."""
+        return {
+            "name": self.name,
+            "scale": self.scale,
+        }
+
+    def state_dict(self) -> dict[str, torch.Tensor]:
+        """Return tensor state only after it has been initialized or loaded."""
+        if self.tensor_state is None:
+            return {}
+
+        return {"tensor_state": self.tensor_state}
+
+    def load_state_dict(self, state: dict[str, torch.Tensor]) -> None:
+        """Load tensor state."""
+        self.tensor_state = state["tensor_state"].clone()
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        """Return features unchanged."""
+        return features
+
+
+@ProcessorStepRegistry.register("registered_lazy_tensor_state_step")
+class RegisteredLazyTensorStateStep(MockLazyTensorStateStep):
+    """Registered lazy tensor state step for registry-based serialization tests."""
+
+
 def test_empty_pipeline():
     """Test pipeline with no steps."""
     pipeline = DataProcessorPipeline([], to_transition=identity_transition, to_output=identity_transition)
@@ -620,6 +668,178 @@ def test_mixed_json_and_tensor_state():
         assert torch.allclose(loaded_step.running_mean, step.running_mean)
 
 
+def test_get_config_matches_saved_json():
+    """Test that in-memory config matches the config written by save_pretrained."""
+    stateless_step = MockStep(name="stateless")
+    stateful_step = MockLazyTensorStateStep(name="stateful", initial_value=4.0)
+    pipeline = DataProcessorPipeline([stateless_step, stateful_step], name="Memory Pipeline")
+
+    in_memory_config = pipeline.get_config()
+
+    assert pipeline.get_config() == in_memory_config
+
+    with tempfile.TemporaryDirectory() as tmp_dir:
+        pipeline.save_pretrained(tmp_dir)
+
+        config_path = Path(tmp_dir) / "memory_pipeline.json"
+        with open(config_path) as file_pointer:
+            saved_config = json.load(file_pointer)
+
+    assert in_memory_config == saved_config
+    assert "state_file" not in in_memory_config["steps"][0]
+    assert in_memory_config["steps"][1]["state_file"] == "memory_pipeline_step_1.safetensors"
+
+
+def test_state_dict_matches_saved_safetensors():
+    """Test that in-memory state matches the safetensors written by save_pretrained."""
+    stateful_step = MockLazyTensorStateStep(initial_value=7.0)
+    pipeline = DataProcessorPipeline([stateful_step], name="Stateful Pipeline")
+
+    in_memory_state_dict = pipeline.state_dict()
+    state_filename = "stateful_pipeline_step_0.safetensors"
+    state_key = "stateful_pipeline_step_0"
+
+    assert set(in_memory_state_dict) == {state_key}
+    assert set(in_memory_state_dict[state_key]) == {"tensor_state"}
+
+    in_memory_state_dict[state_key]["tensor_state"].add_(1)
+    assert stateful_step.tensor_state is not None
+    assert torch.equal(stateful_step.tensor_state, torch.tensor([7.0]))
+
+    with tempfile.TemporaryDirectory() as tmp_dir:
+        pipeline.save_pretrained(tmp_dir)
+        saved_state_dict = load_file(Path(tmp_dir) / state_filename)
+
+    torch.testing.assert_close(saved_state_dict["tensor_state"], torch.tensor([7.0]))
+
+
+def test_save_pretrained_still_writes_expected_serialization_files():
+    """Test that save_pretrained keeps the existing config and state filenames."""
+    stateful_step = MockLazyTensorStateStep(initial_value=3.0)
+    pipeline = DataProcessorPipeline([stateful_step], name="Policy Preprocessor")
+
+    with tempfile.TemporaryDirectory() as tmp_dir:
+        pipeline.save_pretrained(tmp_dir)
+
+        save_path = Path(tmp_dir)
+        assert (save_path / "policy_preprocessor.json").exists()
+        assert (save_path / "policy_preprocessor_step_0.safetensors").exists()
+
+
+def test_from_config_round_trips_stateful_pipeline():
+    """Test that from_config rebuilds a stateful pipeline from in-memory artifacts."""
+    stateful_step = MockLazyTensorStateStep(name="roundtrip", initial_value=11.0)
+    pipeline = DataProcessorPipeline([stateful_step], name="Roundtrip Pipeline")
+    config = pipeline.get_config()
+    pipeline_state_dict = pipeline.state_dict()
+
+    loaded_pipeline = DataProcessorPipeline.from_config(config, state_dict=pipeline_state_dict)
+    loaded_step = loaded_pipeline.steps[0]
+
+    assert len(loaded_pipeline) == 1
+    assert isinstance(loaded_step, MockLazyTensorStateStep)
+    torch.testing.assert_close(loaded_step.tensor_state, torch.tensor([11.0]))
+
+
+def test_from_config_round_trips_registered_stateful_pipeline():
+    """Test that from_config resolves registry steps and loads their named tensor state."""
+    stateful_step = RegisteredLazyTensorStateStep(name="registered", initial_value=29.0)
+    pipeline = DataProcessorPipeline([stateful_step], name="Registry Pipeline")
+    config = pipeline.get_config()
+    pipeline_state_dict = pipeline.state_dict()
+    state_filename = "registry_pipeline_step_0_registered_lazy_tensor_state_step.safetensors"
+    state_key = "registry_pipeline_step_0_registered_lazy_tensor_state_step"
+
+    assert config["steps"][0]["registry_name"] == "registered_lazy_tensor_state_step"
+    assert config["steps"][0]["state_file"] == state_filename
+    assert set(pipeline_state_dict) == {state_key}
+
+    loaded_pipeline = DataProcessorPipeline.from_config(config, state_dict=pipeline_state_dict)
+    loaded_step = loaded_pipeline.steps[0]
+
+    assert isinstance(loaded_step, RegisteredLazyTensorStateStep)
+    assert loaded_step.tensor_state is not None
+    torch.testing.assert_close(loaded_step.tensor_state, torch.tensor([29.0]))
+
+
+def test_from_config_preserves_state_metadata_for_empty_initial_state():
+    """Test in-memory loading when rebuilt steps start without tensor state."""
+    stateful_step = MockLazyTensorStateStep(name="lazy", initial_value=13.0)
+    pipeline = DataProcessorPipeline([stateful_step], name="Lazy Pipeline")
+    config = pipeline.get_config()
+    pipeline_state_dict = pipeline.state_dict()
+
+    loaded_pipeline = DataProcessorPipeline.from_config(config)
+    loaded_step = loaded_pipeline.steps[0]
+
+    assert isinstance(loaded_step, MockLazyTensorStateStep)
+    assert loaded_step.state_dict() == {}
+    assert "state_file" not in loaded_pipeline.get_config()["steps"][0]
+
+    loaded_pipeline.load_state_dict(pipeline_state_dict)
+
+    torch.testing.assert_close(loaded_step.tensor_state, torch.tensor([13.0]))
+
+
+def test_from_config_applies_overrides_before_state_loading():
+    """Test that constructor overrides and tensor state loading are separate operations."""
+    stateful_step = MockLazyTensorStateStep(name="override", scale=1.0, initial_value=17.0)
+    pipeline = DataProcessorPipeline([stateful_step], name="Override Pipeline")
+    config = pipeline.get_config()
+    pipeline_state_dict = pipeline.state_dict()
+
+    loaded_pipeline = DataProcessorPipeline.from_config(
+        config,
+        state_dict=pipeline_state_dict,
+        overrides={"MockLazyTensorStateStep": {"scale": 5.0}},
+    )
+    loaded_step = loaded_pipeline.steps[0]
+
+    assert isinstance(loaded_step, MockLazyTensorStateStep)
+    assert loaded_step.scale == 5.0
+    torch.testing.assert_close(loaded_step.tensor_state, torch.tensor([17.0]))
+
+
+def test_load_state_dict_raises_on_missing_expected_state():
+    """Test loading raises when serialized config expects missing state."""
+    stateful_step = MockLazyTensorStateStep(initial_value=19.0)
+    pipeline = DataProcessorPipeline([stateful_step], name="Missing Pipeline")
+    loaded_pipeline = DataProcessorPipeline.from_config(pipeline.get_config())
+
+    with pytest.raises(KeyError, match="missing_pipeline_step_0"):
+        loaded_pipeline.load_state_dict({})
+
+
+def test_load_state_dict_raises_on_unexpected_extra_state():
+    """Test loading raises on unexpected top-level state keys."""
+    pipeline = DataProcessorPipeline([MockStep(name="stateless")], name="Unexpected Pipeline")
+
+    with pytest.raises(KeyError, match="extra"):
+        pipeline.load_state_dict({"extra": {"tensor_state": torch.tensor([1.0])}})
+
+
+def test_stateless_pipeline_in_memory_serialization_returns_empty_state():
+    """Test stateless in-memory serialization and loading."""
+    pipeline = DataProcessorPipeline([MockStep(name="stateless")], name="Stateless Pipeline")
+    config = pipeline.get_config()
+    config_without_name = {"steps": config["steps"]}
+
+    assert pipeline.state_dict() == {}
+    assert all("state_file" not in step_entry for step_entry in config["steps"])
+
+    loaded_pipeline = DataProcessorPipeline.from_config(config_without_name, state_dict={})
+
+    assert loaded_pipeline.name == "DataProcessorPipeline"
+    assert loaded_pipeline.state_dict() == {}
+
+
+@pytest.mark.parametrize("invalid_config", [None, [], "not config"])
+def test_from_config_rejects_non_dict_config(invalid_config):
+    """Test from_config reports invalid top-level config values cleanly."""
+    with pytest.raises(ValueError, match="not a valid processor configuration"):
+        DataProcessorPipeline.from_config(invalid_config)  # type: ignore[arg-type]
+
+
 class MockModuleStep(ProcessorStep, nn.Module):
     """Mock step that inherits from nn.Module to test state_dict handling of module parameters."""
 

tests/rewards/test_modeling_robometer.py

@@ -0,0 +1,340 @@
+# 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.
+
+"""Tests for Robometer reward model."""
+
+from __future__ import annotations
+
+from types import SimpleNamespace
+
+import pytest
+import torch
+
+from lerobot.configs.rewards import RewardModelConfig
+from lerobot.rewards.factory import get_reward_model_class, make_reward_model_config
+from lerobot.rewards.robometer import RobometerConfig
+from lerobot.rewards.robometer.configuration_robometer import ROBOMETER_SPECIAL_TOKENS
+from lerobot.rewards.robometer.modeling_robometer import (
+    ROBOMETER_FEATURE_PREFIX,
+    convert_bins_to_continuous,
+    decode_progress_outputs,
+)
+from tests.utils import skip_if_package_missing
+
+# Length of the fake tokenizer used in `_patch_build`. The deterministic
+# resize target derived in ``RobometerConfig.__post_init__`` is therefore
+# ``_FAKE_TOKENIZER_LEN + len(ROBOMETER_SPECIAL_TOKENS)``.
+_FAKE_TOKENIZER_LEN = 100
+_EXPECTED_RESIZED_VOCAB = _FAKE_TOKENIZER_LEN + len(ROBOMETER_SPECIAL_TOKENS)
+
+
+class _FakeQwenConfig:
+    """Stand-in for a Qwen3-VL config (the `model.config` attribute).
+
+    ``to_dict`` matches HF's ``PretrainedConfig.to_dict`` closely enough for
+    ``RobometerConfig.__post_init__`` to snapshot a meaningful ``vlm_config``
+    into the saved ``config.json`` and for the reload path to round-trip
+    through ``AutoConfig.for_model``.
+    """
+
+    def __init__(self, hidden_dim: int = 8, vocab_size: int = _FAKE_TOKENIZER_LEN) -> None:
+        # `vocab_size` here is the *pre-resize* value the fake backbone advertises.
+        # `__post_init__` is expected to overwrite it with `len(tokenizer) + 5`.
+        self.text_config = SimpleNamespace(hidden_size=hidden_dim, vocab_size=vocab_size)
+        self._hidden_dim = hidden_dim
+        self._vocab_size = vocab_size
+
+    def to_dict(self) -> dict:
+        return {
+            "model_type": "fake_qwen",
+            "text_config": {
+                "hidden_size": self._hidden_dim,
+                "vocab_size": self._vocab_size,
+            },
+        }
+
+
+class _FakeEmbeddings(torch.nn.Module):
+    def __init__(self, num_embeddings: int = _FAKE_TOKENIZER_LEN) -> None:
+        super().__init__()
+        self.num_embeddings = num_embeddings
+
+
+class _FakeBaseModel(torch.nn.Module):
+    """Stand-in for the Qwen3-VL backbone during tests.
+
+    Provides the minimum surface `RobometerRewardModel.__init__` and
+    `_compute_rbm_logits` rely on: a `parameters()` iterator (for dtype +
+    device), a `config.text_config.hidden_size`, a `config.to_dict()` so
+    `_save_pretrained` can snapshot `vlm_config`,
+    `get_input_embeddings()` / `resize_token_embeddings()` so the fresh-init
+    embed resize is a no-op, and a forward that returns a `SimpleNamespace`
+    with a `hidden_states` tuple.
+    """
+
+    def __init__(self, hidden_dim: int = 8) -> None:
+        super().__init__()
+        self._param = torch.nn.Parameter(torch.zeros(1))
+        self.hidden_dim = hidden_dim
+        self.config = _FakeQwenConfig(hidden_dim)
+        self._embeddings = _FakeEmbeddings()
+
+    def get_input_embeddings(self) -> _FakeEmbeddings:
+        return self._embeddings
+
+    def resize_token_embeddings(self, new_size: int) -> None:
+        self._embeddings.num_embeddings = new_size
+
+    def forward(self, **kwargs):  # noqa: ARG002 - intentional kwargs sink
+        input_ids = kwargs["input_ids"]
+        return SimpleNamespace(
+            hidden_states=(torch.zeros(input_ids.shape[0], input_ids.shape[1], self.hidden_dim),),
+            last_hidden_state=torch.zeros(input_ids.shape[0], input_ids.shape[1], self.hidden_dim),
+        )
+
+
+class _FakeTokenizer:
+    """Minimal stand-in for an HF tokenizer.
+
+    ``RobometerConfig.__post_init__`` uses ``len(tokenizer)`` to compute the
+    deterministic resize target ``len(tokenizer) + len(ROBOMETER_SPECIAL_TOKENS)``,
+    so a working ``__len__`` is all we need.
+    """
+
+    def __init__(self, length: int = _FAKE_TOKENIZER_LEN) -> None:
+        self._length = length
+
+    def __len__(self) -> int:
+        return self._length
+
+
+def _patch_build(monkeypatch) -> None:
+    """Stub out the HF AutoX calls so Robometer construction stays cheap in tests.
+
+    Covers (EO-1 style — no model-side override hooks):
+    * ``AutoConfig.from_pretrained`` (config side) — used by
+      ``RobometerConfig.__post_init__`` to snapshot the backbone config.
+    * ``AutoTokenizer.from_pretrained`` (config side) — used by
+      ``__post_init__`` to compute ``len(tokenizer) + 5``.
+    * ``AutoConfig.for_model``                       — used by
+      ``RobometerConfig.vlm_backbone_config`` when rebuilding for ``from_config``.
+    * ``AutoModelForImageTextToText.from_pretrained`` — fresh-training path
+      (``pretrained_path is None``).
+    * ``AutoModelForImageTextToText.from_config``    — checkpoint-reload path
+      (``pretrained_path`` is set).
+    """
+    from lerobot.rewards.robometer import configuration_robometer, modeling_robometer
+
+    monkeypatch.setattr(
+        modeling_robometer.AutoModelForImageTextToText,
+        "from_pretrained",
+        lambda *args, **kwargs: _FakeBaseModel(hidden_dim=8),
+    )
+    monkeypatch.setattr(
+        modeling_robometer.AutoModelForImageTextToText,
+        "from_config",
+        lambda *args, **kwargs: _FakeBaseModel(hidden_dim=8),
+    )
+    monkeypatch.setattr(
+        configuration_robometer.AutoConfig,
+        "for_model",
+        lambda *args, **kwargs: _FakeQwenConfig(hidden_dim=8),
+    )
+    monkeypatch.setattr(
+        configuration_robometer.AutoConfig,
+        "from_pretrained",
+        lambda *args, **kwargs: _FakeQwenConfig(hidden_dim=8),
+    )
+    monkeypatch.setattr(
+        configuration_robometer.AutoTokenizer,
+        "from_pretrained",
+        lambda *args, **kwargs: _FakeTokenizer(length=_FAKE_TOKENIZER_LEN),
+    )
+
+
+def _make_batch(features: dict[str, torch.Tensor]) -> dict[str, torch.Tensor]:
+    """Build a `compute_reward`-ready batch using Robometer's namespaced keys."""
+    return {f"{ROBOMETER_FEATURE_PREFIX}{key}": value for key, value in features.items()}
+
+
+@skip_if_package_missing("transformers")
+def test_robometer_config_registered(monkeypatch):
+    _patch_build(monkeypatch)
+    assert "robometer" in RewardModelConfig.get_known_choices()
+    assert RewardModelConfig.get_choice_class("robometer") is RobometerConfig
+    assert isinstance(make_reward_model_config("robometer", device="cpu"), RobometerConfig)
+
+
+def test_robometer_factory_returns_in_tree_class():
+    from lerobot.rewards.robometer.modeling_robometer import RobometerRewardModel
+
+    assert get_reward_model_class("robometer") is RobometerRewardModel
+
+
+def test_convert_bins_to_continuous_returns_expected_values():
+    # Two frames: first peaks at bin 0 (center 0.0), second peaks at bin 9 (center 1.0).
+    bin_logits = torch.full((2, 10), -10.0)
+    bin_logits[0, 0] = 10.0
+    bin_logits[1, -1] = 10.0
+    values = convert_bins_to_continuous(bin_logits)
+    assert values.shape == (2,)
+    assert torch.allclose(values, torch.tensor([0.0, 1.0]), atol=1e-3)
+
+
+def test_decode_progress_outputs_returns_last_frame_values():
+    progress = torch.tensor([[0.1, 0.9], [0.4, 0.6]])
+    success_logits = torch.tensor([[0.0, 5.0], [0.0, -5.0]])
+
+    outputs = decode_progress_outputs(progress, success_logits, is_discrete_mode=False)
+
+    assert outputs["progress_pred"] == [pytest.approx([0.1, 0.9]), pytest.approx([0.4, 0.6])]
+    assert outputs["success_probs"][0][-1] == pytest.approx(torch.sigmoid(torch.tensor(5.0)).item(), abs=1e-3)
+    assert outputs["success_probs"][1][-1] == pytest.approx(
+        torch.sigmoid(torch.tensor(-5.0)).item(), abs=1e-3
+    )
+
+
+def test_decode_progress_outputs_discrete_mode_softmaxes_over_bins():
+    # 2 frames, peaks at bin 0 and bin 9 → continuous predictions 0.0 and 1.0
+    bin_logits = torch.full((1, 2, 10), -10.0)
+    bin_logits[0, 0, 0] = 10.0
+    bin_logits[0, 1, -1] = 10.0
+
+    outputs = decode_progress_outputs(bin_logits, success_logits=None, is_discrete_mode=True)
+
+    assert outputs["success_probs"] == []
+    assert outputs["progress_pred"][0] == pytest.approx([0.0, 1.0], abs=1e-3)
+
+
+@skip_if_package_missing("transformers")
+def test_robometer_post_init_overwrites_vocab_size_with_tokenizer_length(monkeypatch):
+    """``RobometerConfig.__post_init__`` must overwrite the backbone's stale
+    ``text_config.vocab_size`` (which on the real Qwen3-VL config is the
+    padded embedding size, ``151,936``) with ``len(tokenizer) + 5``. This is
+    the contract that makes the published ``Robometer-4B`` checkpoint load
+    byte-equivalently."""
+    _patch_build(monkeypatch)
+
+    cfg = RobometerConfig(device="cpu", progress_loss_type="l2")
+
+    assert cfg.vlm_config["text_config"]["vocab_size"] == _EXPECTED_RESIZED_VOCAB
+
+
+@skip_if_package_missing("transformers")
+def test_robometer_compute_reward_reads_pre_encoded_inputs(monkeypatch):
+    from lerobot.rewards.robometer.modeling_robometer import RobometerRewardModel
+
+    progress = torch.tensor([[0.1, 0.9], [0.4, 0.6]])
+    success_logits = torch.tensor([[0.0, 5.0], [0.0, -5.0]])
+    _patch_build(monkeypatch)
+
+    cfg = RobometerConfig(device="cpu", reward_output="progress", progress_loss_type="l2")
+    model = RobometerRewardModel(cfg)
+    # Bypass the Qwen3-VL forward + head extraction with deterministic logits.
+    monkeypatch.setattr(model, "_compute_rbm_logits", lambda _inputs: (progress, success_logits))
+
+    batch = _make_batch({"input_ids": torch.zeros(2, 2, dtype=torch.long)})
+    rewards = model.compute_reward(batch)
+
+    assert torch.allclose(rewards, torch.tensor([0.9, 0.6]))
+
+
+@skip_if_package_missing("transformers")
+def test_robometer_compute_reward_can_return_binary_success(monkeypatch):
+    from lerobot.rewards.robometer.modeling_robometer import RobometerRewardModel
+
+    progress = torch.tensor([[0.1, 0.9], [0.4, 0.6]])
+    success_logits = torch.tensor([[0.0, 5.0], [0.0, -5.0]])  # sigmoid(5) > 0.5; sigmoid(-5) < 0.5
+    _patch_build(monkeypatch)
+
+    cfg = RobometerConfig(
+        device="cpu",
+        reward_output="success",
+        success_threshold=0.5,
+        progress_loss_type="l2",
+    )
+    model = RobometerRewardModel(cfg)
+    monkeypatch.setattr(model, "_compute_rbm_logits", lambda _inputs: (progress, success_logits))
+
+    batch = _make_batch({"input_ids": torch.zeros(2, 2, dtype=torch.long)})
+    rewards = model.compute_reward(batch)
+
+    assert torch.equal(rewards, torch.tensor([1.0, 0.0]))
+
+
+@skip_if_package_missing("transformers")
+def test_robometer_compute_reward_errors_when_inputs_missing(monkeypatch):
+    from lerobot.rewards.robometer.modeling_robometer import RobometerRewardModel
+
+    _patch_build(monkeypatch)
+
+    cfg = RobometerConfig(device="cpu", progress_loss_type="l2")
+    model = RobometerRewardModel(cfg)
+
+    with pytest.raises(KeyError, match=r"observation\.robometer\.input_ids"):
+        model.compute_reward({})
+
+
+@skip_if_package_missing("transformers")
+def test_robometer_save_pretrained_roundtrips(monkeypatch, tmp_path):
+    """Saving and reloading a Robometer model in LeRobot HF format must produce
+    a single ``model.safetensors`` + ``config.json`` (no Hydra ``config.yaml``),
+    must round-trip user-tunable config fields, and must persist all three
+    prediction heads (``progress_head``, ``success_head``, ``preference_head``)
+    so the published ``Robometer-4B`` checkpoint loads byte-equivalently.
+    """
+    from huggingface_hub.constants import CONFIG_NAME, SAFETENSORS_SINGLE_FILE
+    from safetensors.torch import load_file
+
+    from lerobot.rewards.robometer.modeling_robometer import RobometerRewardModel
+
+    _patch_build(monkeypatch)
+    cfg = RobometerConfig(
+        device="cpu",
+        pretrained_path="robometer/Robometer-4B",
+        # Knobs the user might tweak — must survive the round-trip.
+        image_key="observation.images.cam_top",
+        task_key="task",
+        reward_output="success",
+        success_threshold=0.7,
+        progress_loss_type="l2",
+    )
+    model = RobometerRewardModel(cfg)
+    model.save_pretrained(str(tmp_path))
+
+    # Exactly the files LeRobot's HubMixin promises.
+    assert (tmp_path / CONFIG_NAME).exists()
+    assert (tmp_path / SAFETENSORS_SINGLE_FILE).exists()
+    assert not (tmp_path / "config.yaml").exists()  # we want HF-style, not Hydra
+
+    # All three heads must be present in the saved safetensors. The preference
+    # head is unused at inference but the published checkpoint expects its
+    # rows — losing it would silently break weight loading.
+    state = load_file(str(tmp_path / SAFETENSORS_SINGLE_FILE))
+    assert any(k.startswith("progress_head.") for k in state), "progress_head weights missing"
+    assert any(k.startswith("success_head.") for k in state), "success_head weights missing"
+    assert any(k.startswith("preference_head.") for k in state), "preference_head weights missing"
+
+    # Reload from the local directory: no Hub fetch, no YAML overlay. The
+    # base class drives subclass dispatch via the `type` field in config.json.
+    reloaded_cfg = RewardModelConfig.from_pretrained(str(tmp_path))
+    assert isinstance(reloaded_cfg, RobometerConfig)
+    reloaded_cfg.pretrained_path = str(tmp_path)  # mimic lerobot-train's `validate()`
+    reloaded = RobometerRewardModel.from_pretrained(str(tmp_path), config=reloaded_cfg)
+
+    assert reloaded.config.image_key == "observation.images.cam_top"
+    assert reloaded.config.task_key == "task"
+    assert reloaded.config.reward_output == "success"
+    assert reloaded.config.success_threshold == 0.7
+    assert reloaded.config.progress_loss_type == "l2"  # came back from config.json

tests/rewards/test_robometer_processor.py

@@ -0,0 +1,354 @@
+# 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.
+
+"""Tests for Robometer's pre-processing helpers and encoder step.
+
+Covers the pure helpers (``_video_to_numpy`` and ``_expand_tasks``) directly,
+and exercises :class:`RobometerEncoderProcessorStep` with a stubbed
+``AutoProcessor`` so we don't need to download Qwen-VL just to test the
+dataclass plumbing (``transform_features`` / ``get_config``).
+
+The full ``__call__`` path that runs ``process_vision_info`` + the Qwen
+processor is intentionally *not* covered here — it is essentially HF glue
+that's exercised by the integration / parity scripts.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+import numpy as np
+import pytest
+import torch
+
+from lerobot.configs import FeatureType, PipelineFeatureType, PolicyFeature
+from lerobot.rewards.robometer.processor_robometer import (
+    PROGRESS_PROMPT,
+    _expand_tasks,
+    _frames_to_pil,
+    _video_to_numpy,
+)
+from tests.utils import skip_if_package_missing
+
+
+def _skip_if_robometer_extras_missing(func):
+    """Apply both optional-dependency guards in one shot.
+
+    ``RobometerEncoderProcessorStep.__post_init__`` calls
+    ``require_package("transformers", ...)`` *and*
+    ``require_package("qwen-vl-utils", ...)``, so both need to be present
+    before we can instantiate the step.
+    """
+    func = skip_if_package_missing("qwen-vl-utils", import_name="qwen_vl_utils")(func)
+    func = skip_if_package_missing("transformers")(func)
+    return func
+
+
+# ---------------------------------------------------------------------------
+# _video_to_numpy — pure tensor → uint8 (T, H, W, C) conversion
+# ---------------------------------------------------------------------------
+
+
+def test_video_to_numpy_chw_float_is_converted_to_thwc_uint8():
+    video = torch.rand(4, 3, 8, 8)  # (T, C, H, W) floats in [0, 1]
+    array = _video_to_numpy(video, max_frames=None)
+
+    assert array.shape == (4, 8, 8, 3)
+    assert array.dtype == np.uint8
+    assert array.min() >= 0 and array.max() <= 255
+
+
+def test_video_to_numpy_already_thwc_uint8_passes_through():
+    video = torch.randint(0, 256, (3, 8, 8, 3), dtype=torch.uint8)  # (T, H, W, C)
+    array = _video_to_numpy(video, max_frames=None)
+
+    assert array.shape == (3, 8, 8, 3)
+    assert array.dtype == np.uint8
+
+
+def test_video_to_numpy_max_frames_tail_crops_recent_frames():
+    """``max_frames`` should keep the **last** K frames (most recent)."""
+    video = torch.zeros(10, 3, 4, 4)
+    for t in range(10):
+        video[t] = t / 9.0  # marker: 0 at t=0, ≈1 at t=9
+
+    array = _video_to_numpy(video, max_frames=3)
+
+    assert array.shape == (3, 4, 4, 3)
+    # The first kept frame is t=7 → marker ≈ 7/9 → uint8 ≈ 198
+    assert int(array[0, 0, 0, 0]) == int(round(7 / 9 * 255))
+    # The last kept frame is t=9 → marker = 1.0 → uint8 = 255
+    assert int(array[-1, 0, 0, 0]) == 255
+
+
+def test_video_to_numpy_rejects_3d_input():
+    with pytest.raises(ValueError, match="Expected channel dim"):
+        _video_to_numpy(torch.zeros(4, 8, 8), max_frames=None)
+
+
+def test_video_to_numpy_floats_above_one_pass_through_without_rescaling():
+    """If ``array.max() > 1`` the helper assumes the tensor is already in the
+    [0, 255] range (uint8-as-float), so values pass through unchanged."""
+    video = torch.full((1, 3, 2, 2), 5.0)
+    array = _video_to_numpy(video, max_frames=None)
+
+    assert array.shape == (1, 2, 2, 3)
+    assert int(array.max()) == 5
+
+
+def test_video_to_numpy_clips_very_large_floats_to_uint8_max():
+    """Out-of-uint8-range floats are clipped at 255 before the cast."""
+    video = torch.full((1, 3, 2, 2), 300.0)
+    array = _video_to_numpy(video, max_frames=None)
+
+    assert int(array.max()) == 255
+
+
+# ---------------------------------------------------------------------------
+# _expand_tasks — string / list / tuple broadcasting to batch size
+# ---------------------------------------------------------------------------
+
+
+def test_expand_tasks_string_is_broadcast_to_batch_size():
+    assert _expand_tasks("pick up", batch_size=3, default=None) == ["pick up", "pick up", "pick up"]
+
+
+def test_expand_tasks_list_of_matching_size_passes_through():
+    assert _expand_tasks(["a", "b", "c"], batch_size=3, default=None) == ["a", "b", "c"]
+
+
+def test_expand_tasks_tuple_is_normalised_to_list():
+    assert _expand_tasks(("a", "b"), batch_size=2, default=None) == ["a", "b"]
+
+
+def test_expand_tasks_single_element_list_is_broadcast():
+    assert _expand_tasks(["only one"], batch_size=3, default=None) == ["only one"] * 3
+
+
+def test_expand_tasks_size_mismatch_raises():
+    with pytest.raises(ValueError, match="Expected 3 tasks"):
+        _expand_tasks(["a", "b"], batch_size=3, default=None)
+
+
+def test_expand_tasks_missing_uses_default():
+    assert _expand_tasks(None, batch_size=2, default="fallback") == ["fallback", "fallback"]
+
+
+def test_expand_tasks_missing_without_default_raises():
+    with pytest.raises(KeyError, match="task description"):
+        _expand_tasks(None, batch_size=1, default=None)
+
+
+def test_expand_tasks_wrong_type_raises():
+    with pytest.raises(TypeError, match="must be a string or list"):
+        _expand_tasks(42, batch_size=1, default=None)
+
+
+# ---------------------------------------------------------------------------
+# _frames_to_pil — uint8 (T, H, W, C) → list[PIL.Image]
+# ---------------------------------------------------------------------------
+
+
+def test_frames_to_pil_returns_one_image_per_frame():
+    frames = np.zeros((4, 8, 8, 3), dtype=np.uint8)
+    images = _frames_to_pil(frames)
+
+    assert len(images) == 4
+    assert all(img.size == (8, 8) for img in images)
+
+
+def test_frames_to_pil_casts_floats_to_uint8():
+    frames = np.full((2, 4, 4, 3), 200.0, dtype=np.float32)
+    images = _frames_to_pil(frames)
+
+    assert len(images) == 2
+    # PIL converted from clipped uint8 - sanity check pixel values come through.
+    assert np.asarray(images[0]).dtype == np.uint8
+
+
+def test_frames_to_pil_rejects_non_4d_input():
+    with pytest.raises(ValueError, match=r"\(T,H,W,C\)"):
+        _frames_to_pil(np.zeros((4, 8, 8), dtype=np.uint8))
+
+
+# ---------------------------------------------------------------------------
+# Encoder step plumbing — exercise dataclass surface with a stubbed AutoProcessor
+# ---------------------------------------------------------------------------
+
+
+class _FakeTokenizer:
+    """Tokenizer surface the encoder step touches in ``__post_init__``."""
+
+    def __init__(self) -> None:
+        self.pad_token: str | None = None
+        self.eos_token = "<|endoftext|>"
+        self._vocab: dict[str, int] = {"<|endoftext|>": 0}
+        self.added: list[str] = []
+
+    def get_vocab(self) -> dict[str, int]:
+        return self._vocab
+
+    def add_special_tokens(self, payload: dict[str, Any]) -> int:
+        for token in payload.get("additional_special_tokens", []):
+            if token not in self._vocab:
+                self._vocab[token] = len(self._vocab)
+                self.added.append(token)
+        return len(self.added)
+
+
+class _FakeAutoProcessor:
+    """Stand-in returned by ``AutoProcessor.from_pretrained`` during tests."""
+
+    def __init__(self) -> None:
+        self.tokenizer = _FakeTokenizer()
+        self.image_processor = None
+        self.video_processor = None
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):  # noqa: ARG003
+        return cls()
+
+
+def _build_step(monkeypatch, **overrides):
+    from lerobot.rewards.robometer import processor_robometer
+
+    monkeypatch.setattr(processor_robometer, "AutoProcessor", _FakeAutoProcessor)
+
+    return processor_robometer.RobometerEncoderProcessorStep(**overrides)
+
+
+@_skip_if_robometer_extras_missing
+def test_encoder_step_registers_special_tokens_on_tokenizer(monkeypatch):
+    """``__post_init__`` must register Robometer's five special tokens on the
+    tokenizer that ships with the chosen Qwen-VL checkpoint."""
+    from lerobot.rewards.robometer.configuration_robometer import ROBOMETER_SPECIAL_TOKENS
+
+    step = _build_step(monkeypatch)
+
+    vocab = step._processor.tokenizer.get_vocab()
+    for token in ROBOMETER_SPECIAL_TOKENS:
+        assert token in vocab, f"{token} not registered on the tokenizer"
+
+
+@_skip_if_robometer_extras_missing
+def test_encoder_step_sets_pad_token_to_eos_when_missing(monkeypatch):
+    """Qwen tokenizers ship without a pad token; the step must reuse EOS so
+    batched processing doesn't crash on padding."""
+    step = _build_step(monkeypatch)
+
+    assert step._processor.tokenizer.pad_token == "<|endoftext|>"
+
+
+@_skip_if_robometer_extras_missing
+def test_encoder_step_get_config_roundtrips_user_fields(monkeypatch):
+    """``get_config`` must serialise every user-tunable field — these are what
+    the processor pipeline saves under ``preprocessor_config.json``."""
+    step = _build_step(
+        monkeypatch,
+        base_model_id="Qwen/Qwen3-VL-4B-Instruct",
+        image_key="observation.images.cam_top",
+        task_key="task",
+        default_task="do the thing",
+        max_frames=12,
+        use_multi_image=True,
+        use_per_frame_progress_token=True,
+        max_length=2048,
+    )
+
+    cfg = step.get_config()
+    assert cfg == {
+        "base_model_id": "Qwen/Qwen3-VL-4B-Instruct",
+        "image_key": "observation.images.cam_top",
+        "task_key": "task",
+        "default_task": "do the thing",
+        "max_frames": 12,
+        "use_multi_image": True,
+        "use_per_frame_progress_token": True,
+        "max_length": 2048,
+    }
+
+
+@_skip_if_robometer_extras_missing
+def test_encoder_step_transform_features_is_identity(monkeypatch):
+    """The encoder step writes Qwen tensors into ``observation`` at call time,
+    but it does **not** advertise new typed features at pipeline-build time —
+    the downstream model consumes them via the ``ROBOMETER_FEATURE_PREFIX``
+    namespace, not via the typed feature map.
+    """
+    step = _build_step(monkeypatch)
+
+    features = {
+        PipelineFeatureType.OBSERVATION: {
+            "observation.images.top": PolicyFeature(shape=(3, 224, 224), type=FeatureType.VISUAL),
+        }
+    }
+    assert step.transform_features(features) == features
+
+
+@_skip_if_robometer_extras_missing
+def test_encoder_step_build_conversation_inserts_prog_token_per_frame(monkeypatch):
+    """In multi-image mode with per-frame progress tokens, the conversation
+    must alternate ``image`` and ``<|prog_token|>`` text entries, one pair
+    per frame, after the task prompt."""
+    step = _build_step(
+        monkeypatch,
+        use_multi_image=True,
+        use_per_frame_progress_token=True,
+    )
+
+    frames = np.zeros((3, 8, 8, 3), dtype=np.uint8)
+    conversation = step._build_conversation(frames, task="pick up the cube")
+
+    assert len(conversation) == 1 and conversation[0]["role"] == "user"
+    content = conversation[0]["content"]
+
+    # First entry is the task prompt.
+    assert content[0] == {"type": "text", "text": PROGRESS_PROMPT.format(task="pick up the cube")}
+
+    # Then 3 (image, <|prog_token|>) pairs.
+    expected_tail = [
+        item
+        for _ in range(3)
+        for item in (
+            {"type": "image"},  # value asserted below
+            {"type": "text", "text": "<|prog_token|>"},
+        )
+    ]
+    assert len(content) == 1 + len(expected_tail)
+    for got, exp in zip(content[1:], expected_tail, strict=True):
+        assert got["type"] == exp["type"]
+        if exp["type"] == "text":
+            assert got["text"] == exp["text"]
+
+
+@_skip_if_robometer_extras_missing
+def test_encoder_step_build_conversation_video_mode_uses_single_video_entry(monkeypatch):
+    """When ``use_multi_image=False``, frames are bundled into a single
+    ``video`` content entry instead of individual ``image`` entries."""
+    step = _build_step(
+        monkeypatch,
+        use_multi_image=False,
+        use_per_frame_progress_token=False,
+    )
+
+    frames = np.zeros((4, 8, 8, 3), dtype=np.uint8)
+    conversation = step._build_conversation(frames, task="pour the water")
+
+    content = conversation[0]["content"]
+    # Exactly two entries: the prompt and one video entry.
+    assert len(content) == 2
+    assert content[0]["type"] == "text"
+    assert content[1]["type"] == "video"
+    # The video entry carries all four frames.
+    assert len(content[1]["video"]) == 4

tests/robots/test_rebot_b601_follower.py

@@ -91,10 +91,11 @@ def test_get_observation_converts_to_degrees(follower):
 
 
 def test_send_action_clips_to_joint_limits(follower):
-    # shoulder_pan limit is (-145, 145); request beyond the upper bound.
+    # shoulder_pan limit is (-150, 150); request beyond the upper bound.
     returned = follower.send_action({"shoulder_pan.pos": 999.0})
-    assert returned["shoulder_pan.pos"] == 145.0
-    follower.motors["shoulder_pan"].send_pos_vel.assert_called_once()
+    assert returned["shoulder_pan.pos"] == 150.0
+    # Default control_mode is "mit", so arm joints are driven via send_mit.
+    follower.motors["shoulder_pan"].send_mit.assert_called_once()
 
 
 def test_send_action_routes_gripper_to_force_pos(follower):
@@ -103,6 +104,22 @@ def test_send_action_routes_gripper_to_force_pos(follower):
     follower.motors["gripper"].send_pos_vel.assert_not_called()
 
 
+def test_gripper_mit_mode_routes_to_send_mit():
+    bus_mock = _make_bus_mock()
+    with (
+        patch(f"{_MODULE}.require_package", lambda *a, **kw: None),
+        patch(f"{_MODULE}.MotorBridgeController") as controller_cls,
+        patch(f"{_MODULE}.MotorBridgeMode", MagicMock()),
+    ):
+        controller_cls.from_dm_serial.return_value = bus_mock
+        cfg = RebotB601FollowerRobotConfig(port="/dev/null", gripper_control_mode="mit")
+        robot = RebotB601Follower(cfg)
+        robot.connect(calibrate=False)
+        robot.send_action({"gripper.pos": -10.0})
+        robot.motors["gripper"].send_mit.assert_called_once()
+        robot.motors["gripper"].send_force_pos.assert_not_called()
+
+
 def test_bimanual_prefixes_features():
     with patch(f"{_MODULE}.require_package", lambda *a, **kw: None):
         cfg = BiRebotB601FollowerConfig(

tests/test_rollout.py

@@ -59,6 +59,7 @@ def test_strategy_config_types():
     from lerobot.rollout import (
         BaseStrategyConfig,
         DAggerStrategyConfig,
+        EpisodicStrategyConfig,
         HighlightStrategyConfig,
         SentryStrategyConfig,
     )
@@ -67,6 +68,7 @@ def test_strategy_config_types():
     assert SentryStrategyConfig().type == "sentry"
     assert HighlightStrategyConfig().type == "highlight"
     assert DAggerStrategyConfig().type == "dagger"
+    assert EpisodicStrategyConfig().type == "episodic"
 
 
 def test_dagger_config_invalid_input_device():
@@ -203,6 +205,8 @@ def test_create_strategy_dispatches():
         BaseStrategyConfig,
         DAggerStrategy,
         DAggerStrategyConfig,
+        EpisodicStrategy,
+        EpisodicStrategyConfig,
         SentryStrategy,
         SentryStrategyConfig,
         create_strategy,
@@ -211,6 +215,7 @@ def test_create_strategy_dispatches():
     assert isinstance(create_strategy(BaseStrategyConfig()), BaseStrategy)
     assert isinstance(create_strategy(SentryStrategyConfig()), SentryStrategy)
     assert isinstance(create_strategy(DAggerStrategyConfig()), DAggerStrategy)
+    assert isinstance(create_strategy(EpisodicStrategyConfig()), EpisodicStrategy)
 
 
 def test_create_strategy_unknown_raises():

uv.lock

@@ -1764,7 +1764,7 @@ wheels = [
 
 [[package]]
 name = "gym-aloha"
-version = "0.1.3"
+version = "0.1.4"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "dm-control" },
@@ -1772,14 +1772,14 @@ dependencies = [
     { name = "imageio", extra = ["ffmpeg"] },
     { name = "mujoco" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/b5/5e/4bb7204730501c2f645e0532a2df4339206948b2882f77cbf0eaf75bc5fe/gym_aloha-0.1.3.tar.gz", hash = "sha256:b794b246a2e6da6ce5f75e152f553fbd4412704bc217fe6311d0ede3bb72a75e", size = 443468, upload-time = "2025-10-09T14:02:35.024Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/4a/c5/a5b8bdbddfcadec0b52b50e6d1a70325e09e6b594e5f55929d67d9122e2c/gym_aloha-0.1.4.tar.gz", hash = "sha256:0dc4e645045aeb3e74e3c320872d28df6dc93a8751d6ab2f266a2ca11323131f", size = 443466, upload-time = "2026-06-10T09:13:25.525Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/57/6c/10da397177c48ce360efa66ec21b10b10ef5fa2766256fcd8d7d9b5fa6fc/gym_aloha-0.1.3-py3-none-any.whl", hash = "sha256:a94e5747e71307897ded7ae17ed97fab05e814dcb714a16d320f110444f9d0c3", size = 447908, upload-time = "2025-10-09T14:02:33.253Z" },
+    { url = "https://files.pythonhosted.org/packages/35/e3/3afd0e517a503aabe255bf65f5136490acb79c43189e8d56a3aa63081a10/gym_aloha-0.1.4-py3-none-any.whl", hash = "sha256:d9044290fbccddf0be4246b5287cf0eb6b9ddee545a3d222ce8d78c93ce7125e", size = 447908, upload-time = "2026-06-10T09:13:23.868Z" },
 ]
 
 [[package]]
 name = "gym-hil"
-version = "0.1.13"
+version = "0.1.14"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "gymnasium" },
@@ -1789,9 +1789,9 @@ dependencies = [
     { name = "pygame" },
     { name = "pynput" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/f3/41/e89c87b3c66fb2f8ab5818bff4aa552977911eabaee7c12a8a336dcc406f/gym_hil-0.1.13.tar.gz", hash = "sha256:b9eab7a0acc811f181254e3ad72865830fdbb292c236895f374135d3d62f1b27", size = 5668001, upload-time = "2025-10-21T09:57:24.01Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/0c/64/b5cfe59d6a69d20497218f01ad2bdaa2a5a72b850bdb1a445d804ecc9948/gym_hil-0.1.14.tar.gz", hash = "sha256:aeee688dcb3ec72e7bcbe604df4a3f990cce49c8a2da469dd67c3a4eeb4c6bbb", size = 5667991, upload-time = "2026-06-10T09:16:38.98Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/c2/8d/9e3ab53f9aac7bd542f339efd0a9283fa76e034474987e0705379274dfcf/gym_hil-0.1.13-py3-none-any.whl", hash = "sha256:b6444fc43ce1a68ce403df14f99100d9c903ae05d822959e9cd0b76a50b93320", size = 5750805, upload-time = "2025-10-21T09:57:22.068Z" },
+    { url = "https://files.pythonhosted.org/packages/72/97/a7a9c3886306a89046ba5c989bc8b79008e7ec973228bad1fa20d7a94bba/gym_hil-0.1.14-py3-none-any.whl", hash = "sha256:9a2799d47a4561e0b0bb8d37fb3d84934657240be328d13991ea06758726533d", size = 5750805, upload-time = "2026-06-10T09:16:36.827Z" },
 ]
 
 [[package]]
@@ -1881,7 +1881,7 @@ sdist = { url = "https://files.pythonhosted.org/packages/e6/3e/ffad88145b342d5a9
 
 [[package]]
 name = "hf-libero"
-version = "0.1.3"
+version = "0.1.4"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "bddl", marker = "sys_platform == 'linux'" },
@@ -1902,7 +1902,10 @@ dependencies = [
     { name = "transformers", marker = "sys_platform == 'linux'" },
     { name = "wandb", marker = "sys_platform == 'linux'" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/7e/ca/7f1c90aedcd067d608681cf03469ae548990ba0806f68a67927dcc801f04/hf_libero-0.1.3.tar.gz", hash = "sha256:0d6b9a215a658db86f66c03d063d6d877d2e9f96d2d326cfa9f43ba4da4a6d5a", size = 2960521, upload-time = "2025-11-03T17:58:00.003Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/af/aa/4e9eb8715e0bff9cb6553db563a35d253393097d446f82bd53575e8b253d/hf_libero-0.1.4.tar.gz", hash = "sha256:c058d67ad5a2b589529c14d614282ef4cca3a7763dafa134f58a6c9039657e34", size = 2961319, upload-time = "2026-06-10T09:56:13.994Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/2a/79/c286b894c051988d062241682834df915c945bcf51009ffdffbe5ecf69bf/hf_libero-0.1.4-py3-none-any.whl", hash = "sha256:207f76e2f28bff30f78132223d8592fe8f64b1f8fd90ce7024948ada0d7e2c27", size = 3169084, upload-time = "2026-06-10T09:56:12.441Z" },
+]
 
 [[package]]
 name = "hf-xet"
@@ -2989,6 +2992,11 @@ rebot = [
     { name = "motorbridge" },
     { name = "motorbridge-smart-servo" },
 ]
+robometer = [
+    { name = "peft" },
+    { name = "qwen-vl-utils" },
+    { name = "transformers" },
+]
 robstride = [
     { name = "python-can" },
 ]
@@ -3047,6 +3055,11 @@ video-benchmark = [
 viz = [
     { name = "rerun-sdk" },
 ]
+vla-jepa = [
+    { name = "diffusers" },
+    { name = "qwen-vl-utils" },
+    { name = "transformers" },
+]
 wallx = [
     { name = "peft" },
     { name = "qwen-vl-utils" },
@@ -3080,12 +3093,12 @@ requires-dist = [
     { name = "flash-attn", marker = "sys_platform != 'darwin' and extra == 'groot'", specifier = ">=2.5.9,<3.0.0" },
     { name = "grpcio", marker = "extra == 'grpcio-dep'", specifier = "==1.73.1" },
     { name = "grpcio-tools", marker = "extra == 'dev'", specifier = "==1.73.1" },
-    { name = "gym-aloha", marker = "extra == 'aloha'", specifier = ">=0.1.2,<0.2.0" },
-    { name = "gym-hil", marker = "extra == 'hilserl'", specifier = ">=0.1.13,<0.2.0" },
+    { name = "gym-aloha", marker = "extra == 'aloha'", specifier = ">=0.1.4,<0.2.0" },
+    { name = "gym-hil", marker = "extra == 'hilserl'", specifier = ">=0.1.14,<0.2.0" },
     { name = "gym-pusht", marker = "extra == 'pusht'", specifier = ">=0.1.5,<0.2.0" },
     { name = "gymnasium", specifier = ">=1.1.1,<2.0.0" },
     { name = "hebi-py", marker = "extra == 'phone'", specifier = ">=2.8.0,<2.12.0" },
-    { name = "hf-libero", marker = "sys_platform == 'linux' and extra == 'libero'", specifier = ">=0.1.3,<0.2.0" },
+    { name = "hf-libero", marker = "sys_platform == 'linux' and extra == 'libero'", specifier = ">=0.1.4,<0.2.0" },
     { name = "hidapi", marker = "extra == 'gamepad'", specifier = ">=0.14.0,<0.15.0" },
     { name = "huggingface-hub", specifier = ">=1.0.0,<2.0.0" },
     { name = "ipykernel", marker = "extra == 'notebook'", specifier = ">=6.0.0,<7.0.0" },
@@ -3115,6 +3128,7 @@ requires-dist = [
     { name = "lerobot", extras = ["diffusers-dep"], marker = "extra == 'diffusion'" },
     { name = "lerobot", extras = ["diffusers-dep"], marker = "extra == 'groot'" },
     { name = "lerobot", extras = ["diffusers-dep"], marker = "extra == 'multi-task-dit'" },
+    { name = "lerobot", extras = ["diffusers-dep"], marker = "extra == 'vla-jepa'" },
     { name = "lerobot", extras = ["diffusion"], marker = "extra == 'all'" },
     { name = "lerobot", extras = ["dynamixel"], marker = "extra == 'all'" },
     { name = "lerobot", extras = ["feetech"], marker = "extra == 'all'" },
@@ -3146,6 +3160,7 @@ requires-dist = [
     { name = "lerobot", extras = ["peft-dep"], marker = "extra == 'groot'" },
     { name = "lerobot", extras = ["peft-dep"], marker = "extra == 'molmoact2'" },
     { name = "lerobot", extras = ["peft-dep"], marker = "extra == 'peft'" },
+    { name = "lerobot", extras = ["peft-dep"], marker = "extra == 'robometer'" },
     { name = "lerobot", extras = ["peft-dep"], marker = "extra == 'wallx'" },
     { name = "lerobot", extras = ["phone"], marker = "extra == 'all'" },
     { name = "lerobot", extras = ["pi"], marker = "extra == 'all'" },
@@ -3163,10 +3178,13 @@ requires-dist = [
     { name = "lerobot", extras = ["pyzmq-dep"], marker = "extra == 'lekiwi'" },
     { name = "lerobot", extras = ["pyzmq-dep"], marker = "extra == 'unitree-g1'" },
     { name = "lerobot", extras = ["qwen-vl-utils-dep"], marker = "extra == 'eo1'" },
+    { name = "lerobot", extras = ["qwen-vl-utils-dep"], marker = "extra == 'robometer'" },
     { name = "lerobot", extras = ["qwen-vl-utils-dep"], marker = "extra == 'sarm'" },
+    { name = "lerobot", extras = ["qwen-vl-utils-dep"], marker = "extra == 'vla-jepa'" },
     { name = "lerobot", extras = ["qwen-vl-utils-dep"], marker = "extra == 'wallx'" },
     { name = "lerobot", extras = ["reachy2"], marker = "extra == 'all'" },
     { name = "lerobot", extras = ["rebot"], marker = "extra == 'all'" },
+    { name = "lerobot", extras = ["robometer"], marker = "extra == 'all'" },
     { name = "lerobot", extras = ["robstride"], marker = "extra == 'all'" },
     { name = "lerobot", extras = ["sarm"], marker = "extra == 'all'" },
     { name = "lerobot", extras = ["scipy-dep"], marker = "extra == 'aloha'" },
@@ -3188,15 +3206,18 @@ requires-dist = [
     { name = "lerobot", extras = ["transformers-dep"], marker = "extra == 'multi-task-dit'" },
     { name = "lerobot", extras = ["transformers-dep"], marker = "extra == 'peft'" },
     { name = "lerobot", extras = ["transformers-dep"], marker = "extra == 'pi'" },
+    { name = "lerobot", extras = ["transformers-dep"], marker = "extra == 'robometer'" },
     { name = "lerobot", extras = ["transformers-dep"], marker = "extra == 'sarm'" },
     { name = "lerobot", extras = ["transformers-dep"], marker = "extra == 'smolvla'" },
     { name = "lerobot", extras = ["transformers-dep"], marker = "extra == 'topreward'" },
+    { name = "lerobot", extras = ["transformers-dep"], marker = "extra == 'vla-jepa'" },
     { name = "lerobot", extras = ["transformers-dep"], marker = "extra == 'wallx'" },
     { name = "lerobot", extras = ["transformers-dep"], marker = "extra == 'xvla'" },
     { name = "lerobot", extras = ["video-benchmark"], marker = "extra == 'all'" },
     { name = "lerobot", extras = ["viz"], marker = "extra == 'all'" },
     { name = "lerobot", extras = ["viz"], marker = "extra == 'core-scripts'" },
     { name = "lerobot", extras = ["viz"], marker = "extra == 'dataset-viz'" },
+    { name = "lerobot", extras = ["vla-jepa"], marker = "extra == 'all'" },
     { name = "lerobot", extras = ["wallx"], marker = "extra == 'all'" },
     { name = "lerobot", extras = ["xvla"], marker = "extra == 'all'" },
     { name = "matplotlib", marker = "extra == 'matplotlib-dep'", specifier = ">=3.10.3,<4.0.0" },
@@ -3258,7 +3279,7 @@ requires-dist = [
     { name = "transformers", marker = "extra == 'transformers-dep'", specifier = ">=5.4.0,<5.6.0" },
     { name = "wandb", marker = "extra == 'training'", specifier = ">=0.24.0,<0.25.0" },
 ]
-provides-extras = ["dataset", "training", "hardware", "viz", "core-scripts", "evaluation", "dataset-viz", "av-dep", "pygame-dep", "placo-dep", "transformers-dep", "grpcio-dep", "can-dep", "peft-dep", "scipy-dep", "diffusers-dep", "qwen-vl-utils-dep", "matplotlib-dep", "pyserial-dep", "deepdiff-dep", "pynput-dep", "pyzmq-dep", "motorbridge-dep", "motorbridge-smart-servo-dep", "feetech", "dynamixel", "damiao", "robstride", "openarms", "gamepad", "hopejr", "lekiwi", "unitree-g1", "reachy2", "rebot", "kinematics", "intelrealsense", "phone", "diffusion", "wallx", "pi", "molmoact2", "smolvla", "multi-task-dit", "groot", "sarm", "topreward", "xvla", "eo1", "hilserl", "async", "peft", "dev", "notebook", "test", "video-benchmark", "aloha", "pusht", "libero", "metaworld", "all"]
+provides-extras = ["dataset", "training", "hardware", "viz", "core-scripts", "evaluation", "dataset-viz", "av-dep", "pygame-dep", "placo-dep", "transformers-dep", "grpcio-dep", "can-dep", "peft-dep", "scipy-dep", "diffusers-dep", "qwen-vl-utils-dep", "matplotlib-dep", "pyserial-dep", "deepdiff-dep", "pynput-dep", "pyzmq-dep", "motorbridge-dep", "motorbridge-smart-servo-dep", "feetech", "dynamixel", "damiao", "robstride", "openarms", "gamepad", "hopejr", "lekiwi", "unitree-g1", "reachy2", "rebot", "kinematics", "intelrealsense", "phone", "diffusion", "wallx", "pi", "molmoact2", "smolvla", "multi-task-dit", "groot", "sarm", "robometer", "topreward", "xvla", "eo1", "hilserl", "vla-jepa", "async", "peft", "dev", "notebook", "test", "video-benchmark", "aloha", "pusht", "libero", "metaworld", "all"]
 
 [[package]]
 name = "librt"