Add uniform sampling and transition smoothing

* more changes

* working changes

* more changes

* more fixes

* fix style

* more

* clean

* put axis-1

* more fixes

* more styling fixes:

* iterate on review:

* more changes

* add env processor

* style

* more changes

* add docs

* fix imports

* fix test, add to train

* Update src/lerobot/envs/factory.py

Co-authored-by: Michel Aractingi <michel.aractingi@huggingface.co>
Signed-off-by: Jade Choghari <chogharijade@gmail.com>

* iterate on review

---------

Signed-off-by: Jade Choghari <chogharijade@gmail.com>
Co-authored-by: jade.choghari@huggingface.co <“chogharijade@gmail.com”>
Co-authored-by: Michel Aractingi <michel.aractingi@huggingface.co>

.github/workflows/full_tests.yml

@@ -78,7 +78,7 @@ jobs:
           python-version: ${{ env.PYTHON_VERSION }}
 
       - name: Install lerobot with all extras
-        run: uv sync --all-extras
+        run: uv sync --all-extras --no-extra groot # TODO(Steven): Make flash-attn optional
 
       - name: Run pytest (all extras)
         run: uv run pytest tests -vv --maxfail=10

.github/workflows/nightly.yml

@@ -119,6 +119,7 @@ jobs:
       TRITON_CACHE_DIR: /home/user_lerobot/.cache/triton
     container:
       image: ${{ needs.build-docker-cpu-nightly.outputs.image_tag }} # zizmor: ignore[unpinned-images]
+      options: --shm-size "16gb"
       credentials:
         username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
         password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
@@ -158,3 +159,36 @@ jobs:
         run: pytest tests -vv --maxfail=10
       - name: Run end-to-end tests
         run: make test-end-to-end
+
+  # This job runs multi-GPU training tests with 4 GPUs
+  nightly-multi-gpu-tests:
+    name: Nightly Multi-GPU Tests
+    needs: [build-docker-gpu-nightly]
+    runs-on:
+      group: aws-g4dn-12xlarge  # Instance with 4 GPUs
+    env:
+      HF_HOME: /home/user_lerobot/.cache/huggingface
+      HF_LEROBOT_HOME: /home/user_lerobot/.cache/huggingface/lerobot
+      TORCH_HOME: /home/user_lerobot/.cache/torch
+      TRITON_CACHE_DIR: /home/user_lerobot/.cache/triton
+      CUDA_VISIBLE_DEVICES: "0,1,2,3"
+    container:
+      image: ${{ needs.build-docker-gpu-nightly.outputs.image_tag }} # zizmor: ignore[unpinned-images]
+      options: --gpus all --shm-size "16gb"
+      credentials:
+        username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
+        password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
+    defaults:
+      run:
+        shell: bash
+        working-directory: /lerobot
+    steps:
+      - name: Verify GPU availability
+        run: |
+          nvidia-smi
+          python -c "import torch; print(f'PyTorch CUDA available: {torch.cuda.is_available()}'); print(f'Number of GPUs: {torch.cuda.device_count()}')"
+
+      - name: Run multi-GPU training tests
+      # TODO(Steven): Investigate why motors tests are failing in multi-GPU setup
+        run: pytest tests -vv --maxfail=10 --ignore=tests/motors/
+        timeout-minutes: 10

.github/workflows/release.yml

@@ -82,6 +82,14 @@ jobs:
             exit 1
           fi
 
+      - name: Remove Tags with Git dependencies
+        # TODO(Steven): Temporary patch to remove pi from PyPi 0.4.0 release due to its reliance on git dependencies.
+        run: |
+          echo "::info:: Checking for Git dependencies to remove from pyproject.toml..."
+          grep -E '@ git\+https|lerobot\[pi\]' pyproject.toml | sed 's/^/::warning:: Removing line: /' || true
+          sed -E -i '/@ git\+https|lerobot\[pi\]/d' pyproject.toml
+          echo "::info:: Git dependencies removed. Proceeding with build."
+
       - name: Install build dependencies
         run: python -m pip install build
 
@@ -103,7 +111,7 @@ jobs:
       - name: Publish to TestPyPI for pre-releases
         # True for tags like 'v0.2.0-rc1'
         if: startsWith(github.ref, 'refs/tags/v') && contains(github.ref, '-')
-        uses: pypa/gh-action-pypi-publish@v1.12.4 # zizmor: ignore[unpinned-uses, use-trusted-publishing]
+        uses: pypa/gh-action-pypi-publish@v1.13.0 # zizmor: ignore[unpinned-uses, use-trusted-publishing]
         with:
           repository-url: https://test.pypi.org/legacy/
           verbose: true
@@ -111,7 +119,7 @@ jobs:
 
       - name: Publish to PyPI
         if: startsWith(github.ref, 'refs/tags/v') && !contains(github.ref, '-')
-        uses: pypa/gh-action-pypi-publish@v1.12.4 # zizmor: ignore[unpinned-uses, use-trusted-publishing]
+        uses: pypa/gh-action-pypi-publish@v1.13.0 # zizmor: ignore[unpinned-uses, use-trusted-publishing]
         with:
           verbose: true
           print-hash: true
@@ -138,7 +146,7 @@ jobs:
       - name: Setup uv and Python
         uses: astral-sh/setup-uv@v6 # zizmor: ignore[unpinned-uses]
         with:
-          enable-cache: true
+          enable-cache: true # zizmor: ignore[cache-poisoning]
           version: ${{ env.UV_VERSION }}
           python-version: ${{ env.PYTHON_VERSION }}
       - name: Create uv virtual environment

.github/workflows/stale.yml

@@ -27,15 +27,17 @@ env:
     This issue was closed because it has been stalled for 14 days with no activity.
     Feel free to reopen if is still relevant, or to ping a collaborator if you have any questions.
   CLOSE_PR_MESSAGE: >
-    This PR was closed because it has been stalled for 14 days with no activity.
+    This PR was closed because it has been stalled for 21 days with no activity.
     Feel free to reopen if is still relevant, or to ping a collaborator if you have any questions.
   WARN_ISSUE_MESSAGE: >
     This issue has been automatically marked as stale because it has not had
     recent activity (6 months). It will be closed if no further activity occurs.
+    Any change, comment or update to this issue will reset this count.
     Thank you for your contributions.
   WARN_PR_MESSAGE: >
     This PR has been automatically marked as stale because it has not had
-    recent activity (6 months). It will be closed if no further activity occurs.
+    recent activity (1 year). It will be closed if no further activity occurs.
+    Any change, comment or update to this PR will reset this count.
     Thank you for your contributions.
 
 jobs:
@@ -56,10 +58,10 @@ jobs:
           stale-pr-label: stale
           exempt-issue-labels: never-stale
           exempt-pr-labels: never-stale
-          days-before-issue-stale: 180 # TODO(Steven): Will modify this to 90 after initial cleanup
+          days-before-issue-stale: 180
           days-before-issue-close: 14
-          days-before-pr-stale: 180
-          days-before-pr-close: 14
+          days-before-pr-stale: 365
+          days-before-pr-close: 21
           delete-branch: true
           close-issue-message: ${{ env.CLOSE_ISSUE_MESSAGE }}
           close-pr-message: ${{ env.CLOSE_PR_MESSAGE }}

.github/workflows/unbound_deps_tests.yml

@@ -70,7 +70,7 @@ jobs:
           echo "Dependencies unbound:" && cat pyproject.toml
 
       - name: Install lerobot with all extras
-        run: uv sync --all-extras
+        run: uv sync --all-extras --no-extra groot # TODO(Steven): Make flash-attn optional
 
       - name: Run pytest (all extras)
         run: uv run pytest tests -vv

.pre-commit-config.yaml

@@ -26,7 +26,7 @@ repos:
 
    ##### General Code Quality & Formatting #####
   - repo: https://github.com/pre-commit/pre-commit-hooks
-    rev: v5.0.0
+    rev: v6.0.0
     hooks:
       - id: check-added-large-files
         args: ['--maxkb=1024']
@@ -39,20 +39,20 @@ repos:
       - id: trailing-whitespace
 
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.12.4
+    rev: v0.14.1
     hooks:
       - id: ruff-format
       - id: ruff
         args: [--fix, --exit-non-zero-on-fix]
 
   - repo: https://github.com/adhtruong/mirrors-typos
-    rev: v1.34.0
+    rev: v1.38.1
     hooks:
       - id: typos
         args: [--force-exclude]
 
   - repo: https://github.com/asottile/pyupgrade
-    rev: v3.20.0
+    rev: v3.21.0
     hooks:
     -   id: pyupgrade
         args: [--py310-plus]
@@ -68,12 +68,12 @@ repos:
 
   ##### Security #####
   - repo: https://github.com/gitleaks/gitleaks
-    rev: v8.27.2
+    rev: v8.28.0
     hooks:
       - id: gitleaks
 
   - repo: https://github.com/woodruffw/zizmor-pre-commit
-    rev: v1.11.0
+    rev: v1.15.2
     hooks:
       - id: zizmor
 
@@ -87,7 +87,7 @@ repos:
   # TODO(Steven): Uncomment when ready to use
   ##### Static Analysis & Typing #####
   - repo: https://github.com/pre-commit/mirrors-mypy
-    rev: v1.16.0
+    rev: v1.18.2
     hooks:
       - id: mypy
         args: [--config-file=pyproject.toml]

CONTRIBUTING.md

@@ -72,7 +72,6 @@ post it.
 
 Look at our implementations for [datasets](./src/lerobot/datasets/), [policies](./src/lerobot/policies/),
 environments ([aloha](https://github.com/huggingface/gym-aloha),
-[xarm](https://github.com/huggingface/gym-xarm),
 [pusht](https://github.com/huggingface/gym-pusht))
 and follow the same api design.
 
@@ -138,7 +137,7 @@ Follow these steps to start contributing:
 4. for development, we advise to use a tool like `poetry` or `uv` instead of just `pip` to easily track our dependencies.
    Follow the instructions to [install poetry](https://python-poetry.org/docs/#installation) (use a version >=2.1.0) or to [install uv](https://docs.astral.sh/uv/getting-started/installation/#installation-methods) if you don't have one of them already.
 
-   Set up a development environment with conda or miniconda:
+   Set up a development environment with conda:
 
    ```bash
    conda create -y -n lerobot-dev python=3.10 && conda activate lerobot-dev

Makefile

@@ -119,10 +119,9 @@ test-tdmpc-ete-train:
 		--policy.type=tdmpc \
 		--policy.device=$(DEVICE) \
 		--policy.push_to_hub=false \
-		--env.type=xarm \
-		--env.task=XarmLift-v0 \
+		--env.type=pusht \
 		--env.episode_length=5 \
-		--dataset.repo_id=lerobot/xarm_lift_medium \
+		--dataset.repo_id=lerobot/pusht_image \
 		--dataset.image_transforms.enable=true \
 		--dataset.episodes="[0]" \
 		--batch_size=2 \
@@ -140,9 +139,10 @@ test-tdmpc-ete-eval:
 	lerobot-eval \
 		--policy.path=tests/outputs/tdmpc/checkpoints/000002/pretrained_model \
 		--policy.device=$(DEVICE) \
-		--env.type=xarm \
+		--env.type=pusht \
 		--env.episode_length=5 \
-		--env.task=XarmLift-v0 \
+		--env.observation_height=96 \
+        --env.observation_width=96 \
 		--eval.n_episodes=1 \
 		--eval.batch_size=1
 

README.md

@@ -104,14 +104,14 @@ LeRobot works with Python 3.10+ and PyTorch 2.2+.
 
 ### Environment Setup
 
-Create a virtual environment with Python 3.10 and activate it, e.g. with [`miniconda`](https://docs.anaconda.com/free/miniconda/index.html):
+Create a virtual environment with Python 3.10 and activate it, e.g. with [`miniforge`](https://conda-forge.org/download/):
 
 ```bash
 conda create -y -n lerobot python=3.10
 conda activate lerobot
 ```
 
-When using `miniconda`, install `ffmpeg` in your environment:
+When using `conda`, install `ffmpeg` in your environment:
 
 ```bash
 conda install ffmpeg -c conda-forge
@@ -185,6 +185,11 @@ _Replace `[...]` with your desired features._
 For a full list of optional dependencies, see:
 https://pypi.org/project/lerobot/
 
+> [!NOTE]
+> For lerobot 0.4.0, if you want to install pi tags, you will have to do: `pip install "lerobot[pi]@git+https://github.com/huggingface/lerobot.git"`.
+>
+> This will be solved in the next patch release
+
 ### Weights & Biases
 
 To use [Weights and Biases](https://docs.wandb.ai/quickstart) for experiment tracking, log in with
@@ -207,13 +212,13 @@ lerobot-dataset-viz \
     --episode-index 0
 ```
 
-or from a dataset in a local folder with the `root` option and the `--local-files-only` (in the following case the dataset will be searched for in `./my_local_data_dir/lerobot/pusht`)
+or from a dataset in a local folder with the `root` option and the `--mode local` (in the following case the dataset will be searched for in `./my_local_data_dir/lerobot/pusht`)
 
 ```bash
 lerobot-dataset-viz \
     --repo-id lerobot/pusht \
     --root ./my_local_data_dir \
-    --local-files-only 1 \
+    --mode local \
     --episode-index 0
 ```
 
@@ -310,7 +315,7 @@ To upload these to the hub, run the following:
 huggingface-cli upload ${hf_user}/${repo_name} path/to/pretrained_model
 ```
 
-See [eval.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/eval.py) for an example of how other people may use your policy.
+See [lerobot_eval.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/lerobot_eval.py) for an example of how other people may use your policy.
 
 ### Acknowledgment
 
@@ -337,7 +342,3 @@ If you want, you can cite this work with:
 ## Star History
 
 [![Star History Chart](https://api.star-history.com/svg?repos=huggingface/lerobot&type=Timeline)](https://star-history.com/#huggingface/lerobot&Timeline)
-
-```
-
-```

docs/source/_toctree.yml

@@ -7,8 +7,6 @@
 - sections:
   - local: il_robots
     title: Imitation Learning for Robots
-  - local: il_sim
-    title: Imitation Learning in Sim
   - local: cameras
     title: Cameras
   - local: integrate_hardware
@@ -17,8 +15,8 @@
     title: Train a Robot with RL
   - local: hilserl_sim
     title: Train RL in Simulation
-  - local: async
-    title: Use Async Inference
+  - local: multi_gpu_training
+    title: Multi GPU training
   title: "Tutorials"
 - sections:
   - local: lerobot-dataset-v3
@@ -37,9 +35,25 @@
     title: π₀ (Pi0)
   - local: pi05
     title: π₀.₅ (Pi05)
+  - local: groot
+    title: NVIDIA GR00T N1.5
+  title: "Policies"
+- sections:
+  - local: async
+    title: Use Async Inference
+  - local: rtc
+    title: Real-Time Chunking (RTC)
+  title: "Inference"
+- sections:
+  - local: envhub
+    title: Environments from the Hub
+  - local: il_sim
+    title: Imitation Learning in Sim
   - local: libero
     title: Using Libero
-  title: "Policies"
+  - local: metaworld
+    title: Using MetaWorld
+  title: "Simulation"
 - sections:
   - local: introduction_processors
     title: Introduction to Robot Processors
@@ -49,6 +63,8 @@
     title: Implement your own processor
   - local: processors_robots_teleop
     title: Processors for Robots and Teleoperators
+  - local: env_processor
+    title: Environment Processors
   title: "Robot Processors"
 - sections:
   - local: so101

docs/source/env_processor.mdx

@@ -0,0 +1,418 @@
+# Environment Processors
+
+Environment processors are a critical layer in LeRobot's data processing architecture that handle **environment-specific** transformations, separate from policy-specific processing. This separation of concerns enables cleaner code, better modularity, and easier experimentation with different environments and policies.
+
+## Why Environment Processors?
+
+When working with different robot environments (LIBERO, MetaWorld, Aloha, etc.), each environment often has unique data formats, coordinate systems, and conventions that need standardization **before** policy processing. Without environment processors, these transformations would be:
+
+1. **Hardcoded in environment code** - Making it difficult to experiment with different state representations
+2. **Duplicated across policies** - Each policy would need to handle environment-specific quirks
+3. **Mixed with policy logic** - Violating separation of concerns and making debugging harder
+
+Environment processors solve this by providing a **dedicated processing layer** between raw environment observations and policy inputs.
+
+## The Processing Pipeline
+
+Here's how data flows through the complete processing pipeline during evaluation:
+
+```python
+# In lerobot_eval.py rollout() function:
+
+# 1. Raw environment observation (numpy arrays, various formats)
+raw_observation = env.step(action)
+
+# 2. Convert numpy to torch, normalize images [0,1]
+observation = preprocess_observation(raw_observation)
+
+# 3. Add task metadata (for multi-task environments)
+observation = add_envs_task(env, observation)
+
+# 4. ENVIRONMENT-SPECIFIC preprocessing (NEW!)
+#    - Flatten robot states
+#    - Rotate images to match dataset conventions
+#    - Handle environment-specific coordinate systems
+observation = env_preprocessor(observation)
+
+# 5. POLICY-SPECIFIC preprocessing
+#    - Normalize with dataset statistics
+#    - Add batch dimensions
+#    - Move to GPU
+#    - Tokenize language instructions
+observation = preprocessor(observation)
+
+# 6. Policy inference
+action = policy.select_action(observation)
+
+# 7. POLICY-SPECIFIC postprocessing
+#    - Unnormalize actions
+#    - Remove batch dimensions
+action = postprocessor(action)
+
+# 8. ENVIRONMENT-SPECIFIC postprocessing (NEW!)
+#    - Convert action formats if needed
+#    - Apply environment-specific constraints
+action_transition = {"action": action}
+action_transition = env_postprocessor(action_transition)
+action = action_transition["action"]
+
+# 9. Execute in environment
+env.step(action)
+```
+
+## The Benefits
+
+### 1. **Separation of Concerns**
+
+Environment processors handle transformations specific to the **environment's data format**, while policy processors handle transformations specific to the **model's requirements**.
+
+```python
+# ❌ Before: Mixed concerns
+class LiberoVLAPolicy:
+    def preprocess(self, obs):
+        # Environment-specific: Flatten robot state (shouldn't be in policy!)
+        state = self._flatten_robot_state(obs["robot_state"])
+        # Policy-specific: Normalize with dataset stats
+        state = self.normalizer(state)
+        return state
+
+# ✅ After: Clear separation
+# Environment processor: Handles LIBERO's nested robot state
+env_preprocessor = LiberoProcessorStep()  # Flattens robot_state
+
+# Policy processor: Handles model requirements
+policy_preprocessor = NormalizerProcessorStep(stats=dataset_stats)
+```
+
+### 2. **Flexibility and Reusability**
+
+The same policy can work with different environment processors, and the same environment processor can work with different policies:
+
+```python
+# Use SmolVLA policy with LIBERO environment
+libero_preprocessor, libero_postprocessor = make_env_pre_post_processors(libero_cfg)
+smolvla_preprocessor, smolvla_postprocessor = make_pre_post_processors(smolvla_cfg)
+
+# Or use ACT policy with the same LIBERO environment
+libero_preprocessor, libero_postprocessor = make_env_pre_post_processors(libero_cfg)
+act_preprocessor, act_postprocessor = make_pre_post_processors(act_cfg)
+```
+
+### 3. **Easier Experimentation**
+
+Want to try different state representations for LIBERO? Just create a new processor:
+
+```python
+# Original: 8D state (pos + quat→axisangle + gripper)
+@ProcessorStepRegistry.register("libero_processor")
+class LiberoProcessorStep(ObservationProcessorStep):
+    def _process_observation(self, obs):
+        eef_pos = robot_state["eef"]["pos"]          # 3D
+        eef_axisangle = quat2axisangle(quat)         # 3D
+        gripper = robot_state["gripper"]["qpos"]     # 2D
+        state = torch.cat([eef_pos, eef_axisangle, gripper], dim=-1)  # 8D
+        return state
+
+# Experiment: Add velocity for better control
+@ProcessorStepRegistry.register("libero_velocity_processor")
+class LiberoVelocityProcessorStep(ObservationProcessorStep):
+    def _process_observation(self, obs):
+        # Include velocities for 14D state
+        eef_pos = robot_state["eef"]["pos"]          # 3D
+        eef_axisangle = quat2axisangle(quat)         # 3D
+        eef_vel = robot_state["eef"]["vel"]          # 3D  (NEW)
+        gripper_pos = robot_state["gripper"]["qpos"] # 2D
+        gripper_vel = robot_state["gripper"]["qvel"] # 3D  (NEW)
+        state = torch.cat([eef_pos, eef_axisangle, eef_vel,
+                          gripper_pos, gripper_vel], dim=-1)  # 14D
+        return state
+```
+
+### 4. **Cleaner Environment Code**
+
+Environments expose **all available data** without needing to know what downstream models will use:
+
+```python
+# LIBERO environment exposes full robot state
+observation = {
+    "pixels": {"image": img, "image2": img2},
+    "robot_state": {
+        "eef": {"pos": ..., "quat": ..., "vel": ..., "mat": ..., "axisangle": ...},
+        "gripper": {"qpos": ..., "qvel": ...},
+        "joints": {"pos": ..., "vel": ...}
+    }
+}
+
+# Environment processor decides what to use
+# Policy processor handles model-specific transformations
+```
+
+## Using Environment Processors
+
+### Factory Function
+
+The `make_env_pre_post_processors` function follows the same pattern as `make_pre_post_processors` for policies:
+
+```python
+from lerobot.envs.factory import make_env_pre_post_processors
+from lerobot.envs.configs import LiberoEnv, PushtEnv
+
+# For LIBERO: Returns LiberoProcessorStep in preprocessor
+libero_cfg = LiberoEnv(task="libero_spatial", camera_name=["agentview"])
+env_preprocessor, env_postprocessor = make_env_pre_post_processors(libero_cfg)
+
+# For other environments: Returns identity processors (no-op)
+pusht_cfg = PushtEnv()
+env_preprocessor, env_postprocessor = make_env_pre_post_processors(pusht_cfg)
+```
+
+### Implementation in `envs/factory.py`
+
+```python
+def make_env_pre_post_processors(
+    env_cfg: EnvConfig,
+) -> tuple[
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+]:
+    """
+    Create preprocessor and postprocessor pipelines for environment observations.
+
+    Args:
+        env_cfg: The configuration of the environment.
+
+    Returns:
+        A tuple containing:
+            - preprocessor: Pipeline that processes environment observations
+            - postprocessor: Pipeline that processes environment outputs
+    """
+    # For LIBERO environments, add the LiberoProcessorStep to preprocessor
+    if isinstance(env_cfg, LiberoEnv) or "libero" in env_cfg.type:
+        preprocessor = PolicyProcessorPipeline(steps=[LiberoProcessorStep()])
+    else:
+        # For all other environments, return an identity preprocessor
+        preprocessor = PolicyProcessorPipeline(steps=[])
+
+    # Postprocessor is currently identity for all environments
+    # Future: Could add environment-specific action transformations
+    postprocessor = PolicyProcessorPipeline(steps=[])
+
+    return preprocessor, postprocessor
+```
+
+### Integration in Evaluation
+
+In `lerobot_eval.py`, the environment processors are created once and used throughout:
+
+```python
+def eval_main(cfg: EvalPipelineConfig):
+    # Create environment
+    envs = make_env(cfg.env, n_envs=cfg.eval.batch_size)
+
+    # Create policy
+    policy = make_policy(cfg=cfg.policy, env_cfg=cfg.env)
+
+    # Create policy processors
+    preprocessor, postprocessor = make_pre_post_processors(
+        policy_cfg=cfg.policy,
+        pretrained_path=cfg.policy.pretrained_path,
+    )
+
+    # Create environment processors (NEW!)
+    env_preprocessor, env_postprocessor = make_env_pre_post_processors(env_cfg=cfg.env)
+
+    # Run evaluation with both processor types
+    eval_policy_all(
+        envs=envs,
+        policy=policy,
+        env_preprocessor=env_preprocessor,      # Environment-specific
+        env_postprocessor=env_postprocessor,    # Environment-specific
+        preprocessor=preprocessor,              # Policy-specific
+        postprocessor=postprocessor,            # Policy-specific
+        n_episodes=cfg.eval.n_episodes,
+    )
+```
+
+## Example: LIBERO Environment Processor
+
+The `LiberoProcessorStep` demonstrates a real-world environment processor:
+
+```python
+from lerobot.processor.pipeline import ObservationProcessorStep
+
+@dataclass
+@ProcessorStepRegistry.register(name="libero_processor")
+class LiberoProcessorStep(ObservationProcessorStep):
+    """
+    Processes LIBERO observations into the LeRobot format.
+
+    **State Processing:**
+    - Extracts end-effector position (3D)
+    - Converts quaternion to axis-angle representation (3D)
+    - Extracts gripper joint positions (2D)
+    - Concatenates into 8D state vector
+
+    **Image Processing:**
+    - Rotates images 180° to match HuggingFaceVLA/libero convention
+    """
+
+    def _process_observation(self, observation):
+        processed_obs = observation.copy()
+
+        # Process images: Flip 180° for camera convention
+        for key in list(processed_obs.keys()):
+            if key.startswith("observation.images."):
+                img = processed_obs[key]
+                img = torch.flip(img, dims=[2, 3])  # Flip H and W
+                processed_obs[key] = img
+
+        # Process robot_state: Flatten to 8D vector
+        if "observation.robot_state" in processed_obs:
+            robot_state = processed_obs.pop("observation.robot_state")
+
+            eef_pos = robot_state["eef"]["pos"]           # (B, 3)
+            eef_quat = robot_state["eef"]["quat"]         # (B, 4)
+            gripper_qpos = robot_state["gripper"]["qpos"] # (B, 2)
+
+            # Convert quaternion to axis-angle
+            eef_axisangle = self._quat2axisangle(eef_quat)  # (B, 3)
+
+            # Concatenate into single state vector
+            state = torch.cat((eef_pos, eef_axisangle, gripper_qpos), dim=-1)
+            state = state.float()
+
+            processed_obs["observation.state"] = state
+
+        return processed_obs
+```
+
+### Why These Transformations?
+
+1. **Image Rotation**: The HuggingFaceVLA/libero dataset has images rotated 180° from the raw LIBERO simulator. The processor handles this convention mismatch so policies trained on the dataset work seamlessly.
+
+2. **State Flattening**: The raw LIBERO environment exposes nested dictionaries with all available state information (position, quaternion, velocity, matrix representation, etc.). The processor:
+   - Selects the relevant components (pos, quat, gripper)
+   - Converts quaternion to axis-angle (more suitable for learning)
+   - Flattens to a single 8D vector that policies expect
+
+3. **Flexibility**: The environment still exposes **all** raw data. If you want to try different state representations (e.g., including velocities, using matrix representation instead of axis-angle), you can create a new processor without modifying the environment code.
+
+## Adding Environment Processors for New Environments
+
+To add environment processors for a new environment:
+
+### 1. Create the Processor Step
+
+```python
+# In src/lerobot/processor/env_processor.py
+
+@dataclass
+@ProcessorStepRegistry.register(name="myenv_processor")
+class MyEnvProcessorStep(ObservationProcessorStep):
+    """Process observations from MyEnv."""
+
+    def _process_observation(self, observation):
+        processed = observation.copy()
+
+        # Your environment-specific transformations
+        if "myenv.specific.state" in processed:
+            state = processed.pop("myenv.specific.state")
+            # Transform to standard format
+            processed["observation.state"] = self._transform_state(state)
+
+        return processed
+```
+
+### 2. Update the Factory
+
+```python
+# In src/lerobot/envs/factory.py
+
+def make_env_pre_post_processors(env_cfg: EnvConfig):
+    if isinstance(env_cfg, LiberoEnv) or "libero" in env_cfg.type:
+        preprocessor = PolicyProcessorPipeline(steps=[LiberoProcessorStep()])
+    elif isinstance(env_cfg, MyEnvConfig) or "myenv" in env_cfg.type:
+        preprocessor = PolicyProcessorPipeline(steps=[MyEnvProcessorStep()])
+    else:
+        preprocessor = PolicyProcessorPipeline(steps=[])
+
+    postprocessor = PolicyProcessorPipeline(steps=[])
+    return preprocessor, postprocessor
+```
+
+### 3. Use in Evaluation
+
+No changes needed! The evaluation script automatically uses the appropriate processor:
+
+```bash
+lerobot-eval \
+    --policy.path=lerobot/my_policy \
+    --env.type=myenv \  # Automatically uses MyEnvProcessorStep
+    --eval.n_episodes=10
+```
+
+## Future: Environment Postprocessors
+
+Currently, postprocessors are identity (no-op) for all environments. Future use cases include:
+
+### Action Space Transformations
+
+```python
+@dataclass
+class MyEnvActionPostprocessor(ProcessorStep):
+    """Convert policy actions to environment-specific format."""
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        action = transition["action"]
+
+        # Example: Convert from Cartesian to joint space
+        if self.action_space == "joint":
+            action = self.ik_solver(action)
+
+        # Example: Apply environment-specific safety limits
+        action = torch.clamp(action, self.min_action, self.max_action)
+
+        transition["action"] = action
+        return transition
+```
+
+### Coordinate System Conversions
+
+```python
+@dataclass
+class CoordinateTransformPostprocessor(ProcessorStep):
+    """Transform actions between coordinate systems."""
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        action = transition["action"]
+
+        # Example: Policy outputs in world frame, env expects base frame
+        action = self.world_to_base_transform(action)
+
+        transition["action"] = action
+        return transition
+```
+
+## Best Practices
+
+1. **Keep environment processors simple**: They should only handle environment-specific data format issues, not complex learning-related transformations.
+
+2. **Use policy processors for model requirements**: Normalization, batching, device placement, and tokenization belong in policy processors.
+
+3. **Expose all data from environments**: Let processors decide what to use rather than hardcoding choices in the environment.
+
+4. **Document conventions**: Clearly document any coordinate system conventions, camera orientations, or data formats that your processor handles.
+
+5. **Test independently**: Environment processors should be testable without loading full policies or environments.
+
+## Summary
+
+Environment processors provide a **clean separation** between environment-specific data transformations and policy-specific model requirements. This architecture:
+
+- ✅ Enables easy experimentation with different state representations
+- ✅ Allows policies to work seamlessly across different environments
+- ✅ Keeps environment code focused on simulation/hardware interface
+- ✅ Makes processor pipelines more maintainable and debuggable
+- ✅ Follows the single responsibility principle
+
+The key insight: **Environments define data formats, processors standardize them, policies consume standardized data.** Each layer has a clear, focused responsibility.

docs/source/envhub.mdx

@@ -0,0 +1,424 @@
+# Loading Environments from the Hub
+
+The **EnvHub** feature allows you to load simulation environments directly from the Hugging Face Hub with a single line of code. This unlocks a powerful new model for collaboration: instead of environments being locked away inside monolithic libraries, anyone can publish custom environments and share them with the community.
+
+## Overview
+
+With EnvHub, you can:
+
+- Load environments from the Hub instantly
+- Share your custom simulation tasks with the community
+- Version control your environments using Git
+- Distribute complex physics simulations without packaging hassles
+
+## Quick Start
+
+Loading an environment from the Hub is as simple as:
+
+```python
+from lerobot.envs.factory import make_env
+
+# Load a hub environment (requires explicit consent to run remote code)
+env = make_env("lerobot/cartpole-env", trust_remote_code=True)
+```
+
+<Tip warning={true}>
+  **Security Notice**: Loading environments from the Hub executes Python code
+  from third-party repositories. Only use `trust_remote_code=True` with
+  repositories you trust. We strongly recommend pinning to a specific commit
+  hash for reproducibility and security.
+</Tip>
+
+## What is EnvHub?
+
+EnvHub is a framework that allows researchers and developers to:
+
+1. **Publish environments** to the Hugging Face Hub as Git repositories
+2. **Load environments** dynamically without installing them as packages
+3. **Version and track** environment changes using Git semantics
+4. **Discover** new simulation tasks shared by the community
+
+This design means you can go from discovering an interesting environment on the Hub to running experiments in seconds, without worrying about dependency conflicts or complex installation procedures.
+
+## Repository Structure
+
+To make your environment loadable from the Hub, your repository must contain at minimum:
+
+### Required Files
+
+**`env.py`** (or custom Python file)
+
+- Must expose a `make_env(n_envs: int, use_async_envs: bool)` function
+- This function should return one of:
+  - A `gym.vector.VectorEnv` (most common)
+  - A single `gym.Env` (will be automatically wrapped)
+  - A dict mapping `{suite_name: {task_id: VectorEnv}}` (for multi-task benchmarks)
+
+### Optional Files
+
+**`requirements.txt`**
+
+- List any additional dependencies your environment needs
+- Users will need to install these manually before loading your environment
+
+**`README.md`**
+
+- Document your environment: what task it implements, observation/action spaces, rewards, etc.
+- Include usage examples and any special setup instructions
+
+**`.gitignore`**
+
+- Exclude unnecessary files from your repository
+
+### Example Repository Structure
+
+```
+my-environment-repo/
+├── env.py                 # Main environment definition (required)
+├── requirements.txt       # Dependencies (optional)
+├── README.md             # Documentation (recommended)
+├── assets/               # Images, videos, etc. (optional)
+│   └── demo.gif
+└── configs/              # Config files if needed (optional)
+    └── task_config.yaml
+```
+
+## Creating Your Environment Repository
+
+### Step 1: Define Your Environment
+
+Create an `env.py` file with a `make_env` function:
+
+```python
+# env.py
+import gymnasium as gym
+
+def make_env(n_envs: int = 1, use_async_envs: bool = False):
+    """
+    Create vectorized environments for your custom task.
+
+    Args:
+        n_envs: Number of parallel environments
+        use_async_envs: Whether to use AsyncVectorEnv or SyncVectorEnv
+
+    Returns:
+        gym.vector.VectorEnv or dict mapping suite names to vectorized envs
+    """
+    def _make_single_env():
+        # Create your custom environment
+        return gym.make("CartPole-v1")
+
+    # Choose vector environment type
+    env_cls = gym.vector.AsyncVectorEnv if use_async_envs else gym.vector.SyncVectorEnv
+
+    # Create vectorized environment
+    vec_env = env_cls([_make_single_env for _ in range(n_envs)])
+
+    return vec_env
+```
+
+### Step 2: Test Locally
+
+Before uploading, test your environment locally:
+
+```python
+from lerobot.envs.utils import _load_module_from_path, _call_make_env, _normalize_hub_result
+
+# Load your module
+module = _load_module_from_path("./env.py")
+
+# Test the make_env function
+result = _call_make_env(module, n_envs=2, use_async_envs=False)
+normalized = _normalize_hub_result(result)
+
+# Verify it works
+suite_name = next(iter(normalized))
+env = normalized[suite_name][0]
+obs, info = env.reset()
+print(f"Observation shape: {obs.shape if hasattr(obs, 'shape') else type(obs)}")
+env.close()
+```
+
+### Step 3: Upload to the Hub
+
+Upload your repository to Hugging Face:
+
+```bash
+# Install huggingface_hub if needed
+pip install huggingface_hub
+
+# Login to Hugging Face
+huggingface-cli login
+
+# Create a new repository
+huggingface-cli repo create my-custom-env --type space --org my-org
+
+# Initialize git and push
+git init
+git add .
+git commit -m "Initial environment implementation"
+git remote add origin https://huggingface.co/my-org/my-custom-env
+git push -u origin main
+```
+
+Alternatively, use the `huggingface_hub` Python API:
+
+```python
+from huggingface_hub import HfApi
+
+api = HfApi()
+
+# Create repository
+api.create_repo("my-custom-env", repo_type="space")
+
+# Upload files
+api.upload_folder(
+    folder_path="./my-env-folder",
+    repo_id="username/my-custom-env",
+    repo_type="space",
+)
+```
+
+## Loading Environments from the Hub
+
+### Basic Usage
+
+```python
+from lerobot.envs.factory import make_env
+
+# Load from the hub
+envs_dict = make_env(
+    "username/my-custom-env",
+    n_envs=4,
+    trust_remote_code=True
+)
+
+# Access the environment
+suite_name = next(iter(envs_dict))
+env = envs_dict[suite_name][0]
+
+# Use it like any gym environment
+obs, info = env.reset()
+action = env.action_space.sample()
+obs, reward, terminated, truncated, info = env.step(action)
+```
+
+### Advanced: Pinning to Specific Versions
+
+For reproducibility and security, pin to a specific Git revision:
+
+```python
+# Pin to a specific branch
+env = make_env("username/my-env@main", trust_remote_code=True)
+
+# Pin to a specific commit (recommended for papers/experiments)
+env = make_env("username/my-env@abc123def456", trust_remote_code=True)
+
+# Pin to a tag
+env = make_env("username/my-env@v1.0.0", trust_remote_code=True)
+```
+
+### Custom File Paths
+
+If your environment definition is not in `env.py`:
+
+```python
+# Load from a custom file
+env = make_env("username/my-env:custom_env.py", trust_remote_code=True)
+
+# Combine with version pinning
+env = make_env("username/my-env@v1.0:envs/task_a.py", trust_remote_code=True)
+```
+
+### Async Environments
+
+For better performance with multiple environments:
+
+```python
+envs_dict = make_env(
+    "username/my-env",
+    n_envs=8,
+    use_async_envs=True,  # Use AsyncVectorEnv for parallel execution
+    trust_remote_code=True
+)
+```
+
+## URL Format Reference
+
+The hub URL format supports several patterns:
+
+| Pattern              | Description                    | Example                                |
+| -------------------- | ------------------------------ | -------------------------------------- |
+| `user/repo`          | Load `env.py` from main branch | `make_env("lerobot/pusht-env")`        |
+| `user/repo@revision` | Load from specific revision    | `make_env("lerobot/pusht-env@main")`   |
+| `user/repo:path`     | Load custom file               | `make_env("lerobot/envs:pusht.py")`    |
+| `user/repo@rev:path` | Revision + custom file         | `make_env("lerobot/envs@v1:pusht.py")` |
+
+## Multi-Task Environments
+
+For benchmarks with multiple tasks (like LIBERO), return a nested dictionary:
+
+```python
+def make_env(n_envs: int = 1, use_async_envs: bool = False):
+    env_cls = gym.vector.AsyncVectorEnv if use_async_envs else gym.vector.SyncVectorEnv
+
+    # Return dict: {suite_name: {task_id: VectorEnv}}
+    return {
+        "suite_1": {
+            0: env_cls([lambda: gym.make("Task1-v0") for _ in range(n_envs)]),
+            1: env_cls([lambda: gym.make("Task2-v0") for _ in range(n_envs)]),
+        },
+        "suite_2": {
+            0: env_cls([lambda: gym.make("Task3-v0") for _ in range(n_envs)]),
+        }
+    }
+```
+
+## Security Considerations
+
+<Tip warning={true}>
+  **Important**: The `trust_remote_code=True` flag is required to execute
+  environment code from the Hub. This is by design for security.
+</Tip>
+
+When loading environments from the Hub:
+
+1. **Review the code first**: Visit the repository and inspect `env.py` before loading
+2. **Pin to commits**: Use specific commit hashes for reproducibility
+3. **Check dependencies**: Review `requirements.txt` for suspicious packages
+4. **Use trusted sources**: Prefer official organizations or well-known researchers
+5. **Sandbox if needed**: Run untrusted code in isolated environments (containers, VMs)
+
+Example of safe usage:
+
+```python
+# ❌ BAD: Loading without inspection
+env = make_env("random-user/untrusted-env", trust_remote_code=True)
+
+# ✅ GOOD: Review code, then pin to specific commit
+# 1. Visit https://huggingface.co/trusted-org/verified-env
+# 2. Review the env.py file
+# 3. Copy the commit hash
+env = make_env("trusted-org/verified-env@a1b2c3d4", trust_remote_code=True)
+```
+
+## Example: CartPole from the Hub
+
+Here's a complete example using the reference CartPole environment:
+
+```python
+from lerobot.envs.factory import make_env
+import numpy as np
+
+# Load the environment
+envs_dict = make_env("lerobot/cartpole-env", n_envs=4, trust_remote_code=True)
+
+# Get the vectorized environment
+suite_name = next(iter(envs_dict))
+env = envs_dict[suite_name][0]
+
+# Run a simple episode
+obs, info = env.reset()
+done = np.zeros(env.num_envs, dtype=bool)
+total_reward = np.zeros(env.num_envs)
+
+while not done.all():
+    # Random policy
+    action = env.action_space.sample()
+    obs, reward, terminated, truncated, info = env.step(action)
+    total_reward += reward
+    done = terminated | truncated
+
+print(f"Average reward: {total_reward.mean():.2f}")
+env.close()
+```
+
+## Benefits of EnvHub
+
+### For Environment Authors
+
+- **Easy distribution**: No PyPI packaging required
+- **Version control**: Use Git for environment versioning
+- **Rapid iteration**: Push updates instantly
+- **Documentation**: Hub README renders beautifully
+- **Community**: Reach LeRobot users directly
+
+### For Researchers
+
+- **Quick experiments**: Load any environment in one line
+- **Reproducibility**: Pin to specific commits
+- **Discovery**: Browse environments on the Hub
+- **No conflicts**: No need to install conflicting packages
+
+### For the Community
+
+- **Growing ecosystem**: More diverse simulation tasks
+- **Standardization**: Common `make_env` API
+- **Collaboration**: Fork and improve existing environments
+- **Accessibility**: Lower barrier to sharing research
+
+## Troubleshooting
+
+### "Refusing to execute remote code"
+
+You must explicitly pass `trust_remote_code=True`:
+
+```python
+env = make_env("user/repo", trust_remote_code=True)
+```
+
+### "Module X not found"
+
+The hub environment has dependencies you need to install:
+
+```bash
+# Check the repo's requirements.txt and install dependencies
+pip install gymnasium numpy
+```
+
+### "make_env not found in module"
+
+Your `env.py` must expose a `make_env` function:
+
+```python
+def make_env(n_envs: int, use_async_envs: bool):
+    # Your implementation
+    pass
+```
+
+### Environment returns wrong type
+
+The `make_env` function must return:
+
+- A `gym.vector.VectorEnv`, or
+- A single `gym.Env`, or
+- A dict `{suite_name: {task_id: VectorEnv}}`
+
+## Best Practices
+
+1. **Document your environment**: Include observation/action space descriptions, reward structure, and termination conditions in your README
+2. **Add requirements.txt**: List all dependencies with versions
+3. **Test thoroughly**: Verify your environment works locally before pushing
+4. **Use semantic versioning**: Tag releases with version numbers
+5. **Add examples**: Include usage examples in your README
+6. **Keep it simple**: Minimize dependencies when possible
+7. **License your work**: Add a LICENSE file to clarify usage terms
+
+## Future Directions
+
+The EnvHub ecosystem enables exciting possibilities:
+
+- **GPU-accelerated physics**: Share Isaac Gym or Brax environments
+- **Photorealistic rendering**: Distribute environments with advanced graphics
+- **Multi-agent scenarios**: Complex interaction tasks
+- **Real-world simulators**: Digital twins of physical setups
+- **Procedural generation**: Infinite task variations
+- **Domain randomization**: Pre-configured DR pipelines
+
+As more researchers and developers contribute, the diversity and quality of available environments will grow, benefiting the entire robotics learning community.
+
+## See Also
+
+- [Hugging Face Hub Documentation](https://huggingface.co/docs/hub/en/index)
+- [Gymnasium Documentation](https://gymnasium.farama.org/index.html)
+- [Example Hub Environment](https://huggingface.co/lerobot/cartpole-env)

docs/source/groot.mdx

@@ -0,0 +1,125 @@
+# GR00T N1.5 Policy
+
+GR00T N1.5 is an open foundation model from NVIDIA designed for generalized humanoid robot reasoning and skills. It is a cross-embodiment model that accepts multimodal input, including language and images, to perform manipulation tasks in diverse environments.
+
+This document outlines the specifics of its integration and usage within the LeRobot framework.
+
+## Model Overview
+
+NVIDIA Isaac GR00T N1.5 is an upgraded version of the GR00T N1 foundation model. It is built to improve generalization and language-following abilities for humanoid robots.
+
+Developers and researchers can post-train GR00T N1.5 with their own real or synthetic data to adapt it for specific humanoid robots or tasks.
+
+GR00T N1.5 (specifically the GR00T-N1.5-3B model) is built using pre-trained vision and language encoders. It utilizes a flow matching action transformer to model a chunk of actions, conditioned on vision, language, and proprioception.
+
+Its strong performance comes from being trained on an expansive and diverse humanoid dataset, which includes:
+
+- Real captured data from robots.
+- Synthetic data generated using NVIDIA Isaac GR00T Blueprint.
+- Internet-scale video data.
+
+This approach allows the model to be highly adaptable through post-training for specific embodiments, tasks, and environments.
+
+## Installation Requirements
+
+As of today, GR00T N1.5 requires flash attention for it's internal working.
+
+We are working on making this optional, but in the meantime that means that we require an extra installation step and it can only be used in CUDA enabled devices.
+
+1. Following the Environment Setup of our [Installation Guide](./installation). **Attention** don't install `lerobot` in this step.
+2. Install [Flash Attention](https://github.com/Dao-AILab/flash-attention) by running:
+
+```bash
+# Check https://pytorch.org/get-started/locally/ for your system
+pip install "torch>=2.2.1,<2.8.0" "torchvision>=0.21.0,<0.23.0" # --index-url https://download.pytorch.org/whl/cu1XX
+pip install ninja "packaging>=24.2,<26.0" # flash attention dependencies
+pip install "flash-attn>=2.5.9,<3.0.0" --no-build-isolation
+python -c "import flash_attn; print(f'Flash Attention {flash_attn.__version__} imported successfully')"
+```
+
+3. Install LeRobot by running:
+
+```bash
+pip install lerobot[groot]
+```
+
+## Usage
+
+To use GR00T in your LeRobot configuration, specify the policy type as:
+
+```python
+policy.type=groot
+```
+
+## Training
+
+### Training Command Example
+
+Here's a complete training command for finetuning the base GR00T model on your own dataset:
+
+```bash
+# Using a multi-GPU setup
+accelerate launch \
+  --multi_gpu \
+  --num_processes=$NUM_GPUS \
+  $(which lerobot-train) \
+  --output_dir=$OUTPUT_DIR \
+  --save_checkpoint=true \
+  --batch_size=$BATCH_SIZE \
+  --steps=$NUM_STEPS \
+  --save_freq=$SAVE_FREQ \
+  --log_freq=$LOG_FREQ \
+  --policy.push_to_hub=true \
+  --policy.type=groot \
+  --policy.repo_id=$REPO_ID \
+  --policy.tune_diffusion_model=false \
+  --dataset.repo_id=$DATASET_ID \
+  --wandb.enable=true \
+  --wandb.disable_artifact=true \
+  --job_name=$JOB_NAME
+```
+
+## Performance Results
+
+### Libero Benchmark Results
+
+> [!NOTE]
+> Follow our instructions for Libero usage: [Libero](./libero)
+
+GR00T has demonstrated strong performance on the Libero benchmark suite. To compare and test its LeRobot implementation, we finetuned the GR00T N1.5 model for 30k steps on the Libero dataset and compared the results to the GR00T reference results.
+
+| Benchmark          | LeRobot Implementation | GR00T Reference |
+| ------------------ | ---------------------- | --------------- |
+| **Libero Spatial** | 82.0%                  | 92.0%           |
+| **Libero Object**  | 99.0%                  | 92.0%           |
+| **Libero Long**    | 82.0%                  | 76.0%           |
+| **Average**        | 87.0%                  | 87.0%           |
+
+These results demonstrate GR00T's strong generalization capabilities across diverse robotic manipulation tasks. To reproduce these results, you can follow the instructions in the [Libero](https://huggingface.co/docs/lerobot/libero) section.
+
+### Evaluate in your hardware setup
+
+Once you have trained your model using your parameters you can run inference in your downstream task. Follow the instructions in [Imitation Learning for Robots](./il_robots). For example:
+
+```bash
+lerobot-record \
+  --robot.type=bi_so100_follower \
+  --robot.left_arm_port=/dev/ttyACM1 \
+  --robot.right_arm_port=/dev/ttyACM0 \
+  --robot.id=bimanual_follower \
+  --robot.cameras='{ right: {"type": "opencv", "index_or_path": 0, "width": 640, "height": 480, "fps": 30},
+    left: {"type": "opencv", "index_or_path": 2, "width": 640, "height": 480, "fps": 30},
+    top: {"type": "opencv", "index_or_path": 4, "width": 640, "height": 480, "fps": 30},
+  }' \
+  --display_data=true \
+  --dataset.repo_id=<user>/eval_groot-bimanual  \
+  --dataset.num_episodes=10 \
+  --dataset.single_task="Grab and handover the red cube to the other arm"
+  --policy.path=<user>/groot-bimanual # your trained model
+  --dataset.episode_time_s=30
+  --dataset.reset_time_s=10
+```
+
+## License
+
+This model follows the **Apache 2.0 License**, consistent with the original [GR00T repository](https://github.com/NVIDIA/Isaac-GR00T).

docs/source/il_robots.mdx

@@ -165,7 +165,7 @@ huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 Then store your Hugging Face repository name in a variable:
 
 ```bash
-HF_USER=$(huggingface-cli whoami | head -n 1)
+HF_USER=$(hf auth whoami | head -n 1)
 echo $HF_USER
 ```
 
@@ -513,13 +513,14 @@ from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.datasets.utils import hw_to_dataset_features
 from lerobot.policies.act.modeling_act import ACTPolicy
+from lerobot.policies.factory import make_pre_post_processors
 from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
 from lerobot.robots.so100_follower.so100_follower import SO100Follower
+from lerobot.scripts.lerobot_record import record_loop
 from lerobot.utils.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun
-from lerobot.record import record_loop
-from lerobot.policies.factory import make_processor
+
 
 NUM_EPISODES = 5
 FPS = 30
@@ -562,7 +563,7 @@ init_rerun(session_name="recording")
 # Connect the robot
 robot.connect()
 
-preprocessor, postprocessor = make_processor(
+preprocessor, postprocessor = make_pre_post_processors(
     policy_cfg=policy,
     pretrained_path=HF_MODEL_ID,
     dataset_stats=dataset.meta.stats,

docs/source/installation.mdx

@@ -1,8 +1,15 @@
 # Installation
 
+## Install [`miniforge`](https://conda-forge.org/download/)
+
+```bash
+wget "https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-$(uname)-$(uname -m).sh"
+bash Miniforge3-$(uname)-$(uname -m).sh
+```
+
 ## Environment Setup
 
-Create a virtual environment with Python 3.10, using [`Miniconda`](https://docs.anaconda.com/miniconda/install/#quick-command-line-install)
+Create a virtual environment with Python 3.10, using conda:
 
 ```bash
 conda create -y -n lerobot python=3.10
@@ -14,7 +21,7 @@ Then activate your conda environment, you have to do this each time you open a s
 conda activate lerobot
 ```
 
-When using `miniconda`, install `ffmpeg` in your environment:
+When using `conda`, install `ffmpeg` in your environment:
 
 ```bash
 conda install ffmpeg -c conda-forge
@@ -74,6 +81,9 @@ _Replace `[...]` with your desired features._
 For a full list of optional dependencies, see:
 https://pypi.org/project/lerobot/
 
+> [!NOTE]
+> For lerobot 0.4.0, if you want to install pi, you will have to do: `pip install "lerobot[pi]@git+https://github.com/huggingface/lerobot.git"`
+
 ### Troubleshooting
 
 If you encounter build errors, you may need to install additional dependencies: `cmake`, `build-essential`, and `ffmpeg libs`.
@@ -91,7 +101,7 @@ LeRobot provides optional extras for specific functionalities. Multiple extras c
 
 ### Simulations
 
-Install environment packages: `aloha` ([gym-aloha](https://github.com/huggingface/gym-aloha)), `xarm` ([gym-xarm](https://github.com/huggingface/gym-xarm)), or `pusht` ([gym-pusht](https://github.com/huggingface/gym-pusht))
+Install environment packages: `aloha` ([gym-aloha](https://github.com/huggingface/gym-aloha)), or `pusht` ([gym-pusht](https://github.com/huggingface/gym-pusht))
 Example:
 
 ```bash

docs/source/integrate_hardware.mdx

@@ -208,34 +208,36 @@ LeRobot supports saving and loading calibration data automatically. This is usef
 
 <!-- prettier-ignore-start -->
 ```python
-> @property
-> def is_calibrated(self) -> bool:
->    return True
->
-> def calibrate(self) -> None:
->    pass
-> ```
+@property
+def is_calibrated(self) -> bool:
+    return True
+
+def calibrate(self) -> None:
+    pass
+```
+<!-- prettier-ignore-end -->
 
 ### `is_calibrated`
 
 This should reflect whether your robot has the required calibration loaded.
 
-```
-<!-- prettier-ignore-end -->python
+<!-- prettier-ignore-start -->
+```python
 @property
 def is_calibrated(self) -> bool:
     return self.bus.is_calibrated
 ```
+<!-- prettier-ignore-end -->
 
 ### `calibrate()`
 
 The goal of the calibration is twofold:
-    - Know the physical range of motion of each motors in order to only send commands within this range.
-    - Normalize raw motors positions to sensible continuous values (e.g. percentages, degrees) instead of arbitrary discrete value dependant on the specific motor used that will not replicate elsewhere.
+
+- Know the physical range of motion of each motors in order to only send commands within this range.
+- Normalize raw motors positions to sensible continuous values (e.g. percentages, degrees) instead of arbitrary discrete value dependant on the specific motor used that will not replicate elsewhere.
 
 It should implement the logic for calibration (if relevant) and update the `self.calibration` dictionary. If you are using Feetech or Dynamixel motors, our bus interfaces already include methods to help with this.
 
-
 <!-- prettier-ignore-start -->
 ```python
 def calibrate(self) -> None:

docs/source/lerobot-dataset-v3.mdx

@@ -279,3 +279,36 @@ python -m lerobot.datasets.v30.convert_dataset_v21_to_v30 --repo-id=<HF_USER/DAT
 - Aggregates parquet files: `episode-0000.parquet`, `episode-0001.parquet`, … → **`file-0000.parquet`**, …
 - Aggregates mp4 files: `episode-0000.mp4`, `episode-0001.mp4`, … → **`file-0000.mp4`**, …
 - Updates `meta/episodes/*` (chunked Parquet) with per‑episode lengths, tasks, and byte/frame offsets.
+
+## Common Issues
+
+### Always call `finalize()` before pushing
+
+When creating or recording datasets, you **must** call `dataset.finalize()` to properly close parquet writers. See the [PR #1903](https://github.com/huggingface/lerobot/pull/1903) for more details.
+
+```python
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+
+# Create dataset and record episodes
+dataset = LeRobotDataset.create(...)
+
+for episode in range(num_episodes):
+    # Record frames
+    for frame in episode_data:
+        dataset.add_frame(frame)
+    dataset.save_episode()
+
+# Call finalize() when done recording and before push_to_hub()
+dataset.finalize()  # Closes parquet writers, writes metadata footers
+dataset.push_to_hub()
+```
+
+**Why is this necessary?**
+
+Dataset v3.0 uses incremental parquet writing with buffered metadata for efficiency. The `finalize()` method:
+
+- Flushes any buffered episode metadata to disk
+- Closes parquet writers to write footer metadata, otherwise the parquet files will be corrupt
+- Ensures the dataset is valid for loading
+
+Without calling `finalize()`, your parquet files will be incomplete and the dataset won't load properly.

docs/source/libero.mdx

@@ -137,7 +137,7 @@ The finetuned model can be found here:
 We then evaluate the finetuned model using the LeRobot LIBERO implementation, by running the following command:
 
 ```bash
-python src/lerobot/scripts/eval.py \
+lerobot-eval \
   --output_dir=/logs/ \
   --env.type=libero \
   --env.task=libero_spatial,libero_object,libero_goal,libero_10 \

docs/source/metaworld.mdx

@@ -0,0 +1,80 @@
+# Meta-World
+
+Meta-World is a well-designed, open-source simulation benchmark for multi-task and meta reinforcement learning in continuous-control robotic manipulation. It gives researchers a shared, realistic playground to test whether algorithms can _learn many different tasks_ and _generalize quickly to new ones_ — two central challenges for real-world robotics.
+
+- 📄 [MetaWorld paper](https://arxiv.org/pdf/1910.10897)
+- 💻 [Original MetaWorld repo](https://github.com/Farama-Foundation/Metaworld)
+
+![MetaWorld MT10 demo](https://meta-world.github.io/figures/ml45.gif)
+
+## Why Meta-World matters
+
+- **Diverse, realistic tasks.** Meta-World bundles a large suite of simulated manipulation tasks (50 in the MT50 suite) using everyday objects and a common tabletop Sawyer arm. This diversity exposes algorithms to a wide variety of dynamics, contacts and goal specifications while keeping a consistent control and observation structure.
+- **Focus on generalization and multi-task learning.** By evaluating across task distributions that share structure but differ in goals and objects, Meta-World reveals whether an agent truly learns transferable skills rather than overfitting to a narrow task.
+- **Standardized evaluation protocol.** It provides clear evaluation modes and difficulty splits, so different methods can be compared fairly across easy, medium, hard and very-hard regimes.
+- **Empirical insight.** Past evaluations on Meta-World show impressive progress on some fronts, but also highlight that current multi-task and meta-RL methods still struggle with large, diverse task sets. That gap points to important research directions.
+
+## What it enables in LeRobot
+
+In LeRobot, you can evaluate any policy or vision-language-action (VLA) model on Meta-World tasks and get a clear success-rate measure. The integration is designed to be straightforward:
+
+- We provide a LeRobot-ready dataset for Meta-World (MT50) on the HF Hub: `https://huggingface.co/datasets/lerobot/metaworld_mt50`.
+  - This dataset is formatted for the MT50 evaluation that uses all 50 tasks (the most challenging multi-task setting).
+  - MT50 gives the policy a one-hot task vector and uses fixed object/goal positions for consistency.
+
+- Task descriptions and the exact keys required for evaluation are available in the repo/dataset — use these to ensure your policy outputs the right success signals.
+
+## Quick start, train a SmolVLA policy on Meta-World
+
+Example command to train a SmolVLA policy on a subset of tasks:
+
+```bash
+lerobot-train \
+  --policy.type=smolvla \
+  --policy.repo_id=${HF_USER}/metaworld-test \
+  --policy.load_vlm_weights=true \
+  --dataset.repo_id=lerobot/metaworld_mt50 \
+  --env.type=metaworld \
+  --env.task=assembly-v3,dial-turn-v3,handle-press-side-v3 \
+  --output_dir=./outputs/ \
+  --steps=100000 \
+  --batch_size=4 \
+  --eval.batch_size=1 \
+  --eval.n_episodes=1 \
+  --eval_freq=1000
+```
+
+Notes:
+
+- `--env.task` accepts explicit task lists (comma separated) or difficulty groups (e.g., `env.task="hard"`).
+- Adjust `batch_size`, `steps`, and `eval_freq` to match your compute budget.
+- **Gymnasium Assertion Error**: if you encounter an error like
+  `AssertionError: ['human', 'rgb_array', 'depth_array']` when running MetaWorld environments, this comes from a mismatch between MetaWorld and your Gymnasium version.
+  We recommend using:
+
+```bash
+  pip install "gymnasium==1.1.0"
+```
+
+to ensure proper compatibility.
+
+## Quick start — evaluate a trained policy
+
+To evaluate a trained policy on the Meta-World medium difficulty split:
+
+```bash
+lerobot-eval \
+  --policy.path="your-policy-id" \
+  --env.type=metaworld \
+  --env.task=medium \
+  --eval.batch_size=1 \
+  --eval.n_episodes=2
+```
+
+This will run episodes and return per-task success rates using the standard Meta-World evaluation keys.
+
+## Practical tips
+
+- If you care about generalization, run on the full MT50 suite — it’s intentionally challenging and reveals strengths/weaknesses better than a few narrow tasks.
+- Use the one-hot task conditioning for multi-task training (MT10 / MT50 conventions) so policies have explicit task context.
+- Inspect the dataset task descriptions and the `info["is_success"]` keys when writing post-processing or logging so your success metrics line up with the benchmark.

docs/source/multi_gpu_training.mdx

@@ -0,0 +1,125 @@
+# Multi-GPU Training
+
+This guide shows you how to train policies on multiple GPUs using [Hugging Face Accelerate](https://huggingface.co/docs/accelerate).
+
+## Installation
+
+First, ensure you have accelerate installed:
+
+```bash
+pip install accelerate
+```
+
+## Training with Multiple GPUs
+
+You can launch training in two ways:
+
+### Option 1: Without config (specify parameters directly)
+
+You can specify all parameters directly in the command without running `accelerate config`:
+
+```bash
+accelerate launch \
+  --multi_gpu \
+  --num_processes=2 \
+  $(which lerobot-train) \
+  --dataset.repo_id=${HF_USER}/my_dataset \
+  --policy.type=act \
+  --policy.repo_id=${HF_USER}/my_trained_policy \
+  --output_dir=outputs/train/act_multi_gpu \
+  --job_name=act_multi_gpu \
+  --wandb.enable=true
+```
+
+**Key accelerate parameters:**
+
+- `--multi_gpu`: Enable multi-GPU training
+- `--num_processes=2`: Number of GPUs to use
+- `--mixed_precision=fp16`: Use fp16 mixed precision (or `bf16` if supported)
+
+### Option 2: Using accelerate config
+
+If you prefer to save your configuration, you can optionally configure accelerate for your hardware setup by running:
+
+```bash
+accelerate config
+```
+
+This interactive setup will ask you questions about your training environment (number of GPUs, mixed precision settings, etc.) and saves the configuration for future use. For a simple multi-GPU setup on a single machine, you can use these recommended settings:
+
+- Compute environment: This machine
+- Number of machines: 1
+- Number of processes: (number of GPUs you want to use)
+- GPU ids to use: (leave empty to use all)
+- Mixed precision: fp16 or bf16 (recommended for faster training)
+
+Then launch training with:
+
+```bash
+accelerate launch $(which lerobot-train) \
+  --dataset.repo_id=${HF_USER}/my_dataset \
+  --policy.type=act \
+  --policy.repo_id=${HF_USER}/my_trained_policy \
+  --output_dir=outputs/train/act_multi_gpu \
+  --job_name=act_multi_gpu \
+  --wandb.enable=true
+```
+
+## How It Works
+
+When you launch training with accelerate:
+
+1. **Automatic detection**: LeRobot automatically detects if it's running under accelerate
+2. **Data distribution**: Your batch is automatically split across GPUs
+3. **Gradient synchronization**: Gradients are synchronized across GPUs during backpropagation
+4. **Single process logging**: Only the main process logs to wandb and saves checkpoints
+
+## Learning Rate and Training Steps Scaling
+
+**Important:** LeRobot does **NOT** automatically scale learning rates or training steps based on the number of GPUs. This gives you full control over your training hyperparameters.
+
+### Why No Automatic Scaling?
+
+Many distributed training frameworks automatically scale the learning rate by the number of GPUs (e.g., `lr = base_lr × num_gpus`).
+However, LeRobot keeps the learning rate exactly as you specify it.
+
+### When and How to Scale
+
+If you want to scale your hyperparameters when using multiple GPUs, you should do it manually:
+
+**Learning Rate Scaling:**
+
+```bash
+# Example: 2 GPUs with linear LR scaling
+# Base LR: 1e-4, with 2 GPUs -> 2e-4
+accelerate launch --num_processes=2 $(which lerobot-train) \
+  --optimizer.lr=2e-4 \
+  --dataset.repo_id=lerobot/pusht \
+  --policy=act
+```
+
+**Training Steps Scaling:**
+
+Since the effective batch size `bs` increases with multiple GPUs (batch_size × num_gpus), you may want to reduce the number of training steps proportionally:
+
+```bash
+# Example: 2 GPUs with effective batch size 2x larger
+# Original: batch_size=8, steps=100000
+# With 2 GPUs: batch_size=8 (16 in total), steps=50000
+accelerate launch --num_processes=2 $(which lerobot-train) \
+  --batch_size=8 \
+  --steps=50000 \
+  --dataset.repo_id=lerobot/pusht \
+  --policy=act
+```
+
+## Notes
+
+- The `--policy.use_amp` flag in `lerobot-train` is only used when **not** running with accelerate. When using accelerate, mixed precision is controlled by accelerate's configuration.
+- Training logs, checkpoints, and hub uploads are only done by the main process to avoid conflicts. Non-main processes have console logging disabled to prevent duplicate output.
+- The effective batch size is `batch_size × num_gpus`. If you use 4 GPUs with `--batch_size=8`, your effective batch size is 32.
+- Learning rate scheduling is handled correctly across multiple processes—LeRobot sets `step_scheduler_with_optimizer=False` to prevent accelerate from adjusting scheduler steps based on the number of processes.
+- When saving or pushing models, LeRobot automatically unwraps the model from accelerate's distributed wrapper to ensure compatibility.
+- WandB integration automatically initializes only on the main process, preventing multiple runs from being created.
+
+For more advanced configurations and troubleshooting, see the [Accelerate documentation](https://huggingface.co/docs/accelerate). If you want to learn more about how to train on a large number of GPUs, checkout this awesome guide: [Ultrascale Playbook](https://huggingface.co/spaces/nanotron/ultrascale-playbook).

docs/source/pi0.mdx

@@ -28,6 +28,11 @@ As described by Physical Intelligence, while AI has achieved remarkable success
    pip install -e ".[pi]"
    ```
 
+   > [!NOTE]
+   > For lerobot 0.4.0, if you want to install pi tag, you will have to do: `pip install "lerobot[pi]@git+https://github.com/huggingface/lerobot.git"`.
+   >
+   > This will be solved in the next patch release
+
 ## Training Data and Capabilities
 
 π₀ is trained on the largest robot interaction dataset to date, combining three key data sources:

docs/source/pi05.mdx

@@ -36,6 +36,11 @@ This diverse training mixture creates a "curriculum" that enables generalization
    pip install -e ".[pi]"
    ```
 
+   > [!NOTE]
+   > For lerobot 0.4.0, if you want to install pi tag, you will have to do: `pip install "lerobot[pi]@git+https://github.com/huggingface/lerobot.git"`.
+   >
+   > This will be solved in the next patch release
+
 ## Usage
 
 To use π₀.₅ in your LeRobot configuration, specify the policy type as:

docs/source/policy_groot_README.md

@@ -0,0 +1,27 @@
+## Research Paper
+
+Paper: https://research.nvidia.com/labs/gear/gr00t-n1_5/
+
+## Repository
+
+Code: https://github.com/NVIDIA/Isaac-GR00T
+
+## Citation
+
+```bibtex
+@inproceedings{gr00tn1_2025,
+  archivePrefix = {arxiv},
+  eprint     = {2503.14734},
+  title      = {{GR00T} {N1}: An Open Foundation Model for Generalist Humanoid Robots},
+  author     = {NVIDIA and Johan Bjorck andFernando Castañeda, Nikita Cherniadev and Xingye Da and Runyu Ding and Linxi "Jim" Fan and Yu Fang and Dieter Fox and Fengyuan Hu and Spencer Huang and Joel Jang and Zhenyu Jiang and Jan Kautz and Kaushil Kundalia and Lawrence Lao and Zhiqi Li and Zongyu Lin and Kevin Lin and Guilin Liu and Edith Llontop and Loic Magne and Ajay Mandlekar and Avnish Narayan and Soroush Nasiriany and Scott Reed and You Liang Tan and Guanzhi Wang and Zu Wang and Jing Wang and Qi Wang and Jiannan Xiang and Yuqi Xie and Yinzhen Xu and Zhenjia Xu and Seonghyeon Ye and Zhiding Yu and Ao Zhang and Hao Zhang and Yizhou Zhao and Ruijie Zheng and Yuke Zhu},
+  month      = {March},
+  year       = {2025},
+  booktitle  = {ArXiv Preprint},
+}
+```
+
+## Additional Resources
+
+Blog: https://developer.nvidia.com/isaac/gr00t
+
+Hugging Face Model: https://huggingface.co/nvidia/GR00T-N1.5-3B

docs/source/rtc.mdx

@@ -0,0 +1,188 @@
+# Real-Time Chunking (RTC)
+
+Real-Time Chunking (RTC) is an inference-time method that allows large, flow-matching based robotic policies, such as [Pi0](./pi0), [Pi0.5](./pi05), and [SmolVLA](./smolvla), to produce smooth, continuous, and reactive motion despite having high inference latency.
+
+These policies generate chunks of future actions (e.g., 50 steps at a time) instead of single actions.
+Because the models are large, producing each chunk takes longer than the time it takes the robot to execute it.
+Naively executing chunks leads to problems such as pauses, jerky transitions, or sudden changes in strategy whenever the next chunk arrives late or disagrees with the previously executed actions.
+
+RTC solves this by asynchronously generating the next chunk while the robot continues executing the current one, and by guiding the new chunk so it aligns smoothly with the portion of the previous chunk that has already been executed.
+
+## How RTC Works (simplified)
+
+RTC lets the robot think ahead while it’s still moving. When the robot is carrying out one chunk of actions, RTC starts creating the next chunk early.
+But since the robot has already moved a bit by the time the new chunk is ready, RTC has to make sure the new chunk still lines up smoothly with what the robot is currently doing.
+
+To do this, RTC treats the beginning of the new chunk like an inpainting or “fill-in-the-gaps” problem:
+it gently adjusts the first part of the new chunk so it blends naturally with the robot’s ongoing motion. The result is no pauses, no sudden jumps.
+
+In technical terms, RTC adds a guidance term to the flow-matching denoising process that forces the overlapping timesteps of the new chunk to stay close to the executed portion of the previous chunk, typically using a soft transition mask.
+
+## Quick Start
+
+### Installation
+
+RTC is built into LeRobot. Just install the policy dependencies you need:
+
+```bash
+# For Pi0 or Pi0.5
+pip install -e ".[pi]"
+
+# For SmolVLA
+pip install -e ".[smolvla]"
+```
+
+### Using RTC with Pi0
+
+You can find a complete reference implementation in [eval_with_real_robot.py](examples/rtc/eval_with_real_robot.py).
+The snippet below provides a simplified pseudo-example of how RTC operates with Pi0 in your pipeline:
+
+```python
+from lerobot.policies.pi0 import PI0Policy, PI0Config
+from lerobot.configs.types import RTCAttentionSchedule
+from lerobot.policies.rtc.configuration_rtc import RTCConfig
+from lerobot.policies.rtc.action_queue import ActionQueue
+
+# Load Pi0 with RTC enabled
+policy_cfg = PI0Config()
+
+# Enable RTC
+policy_cfg.rtc_config = RTCConfig(
+    enabled=True,
+    execution_horizon=10,  # How many steps to blend with previous chunk
+    max_guidance_weight=10.0,  # How strongly to enforce consistency
+    prefix_attention_schedule=RTCAttentionSchedule.EXP,  # Exponential blend
+)
+
+# Load the policy
+policy = PI0Policy.from_pretrained("lerobot/pi0_base", policy_cfg=policy_cfg, device="cuda")
+
+# Now use predict_action_chunk with RTC parameters
+inference_delay = 4  # How many steps of inference latency, this values should be calculated based on the inference latency of the policy
+
+# Initialize the action queue
+action_queue = ActionQueue(policy_cfg.rtc_config)
+
+# Start in a separate thread with the following function
+def get_actions():
+  while True:
+    if should_get_actions:
+
+      prev_actions = action_queue.get_left_over()
+      obs = get_robot_observations(robot)
+
+      # Generate actions WITH RTC
+      actions = policy.predict_action_chunk(
+          obs,
+          inference_delay=inference_delay,
+          prev_chunk_left_over=prev_actions,
+      )
+
+      action_queue.merge(
+          actions, actions, inference_delay
+      )
+
+for step in range(num_steps):
+    action = action_queue.get()
+
+    # Execute the first N actions
+    execute_actions(action)
+```
+
+## Key Parameters
+
+`RTCConfig` has the following parameters to tune:
+
+**`execution_horizon`**: How many timesteps from the previous chunk to maintain consistency with. Higher values mean smoother transitions but potentially less reactivity.
+
+Typical values: 8-12 steps
+
+```python
+RTCConfig(execution_horizon=10)
+```
+
+**`max_guidance_weight`**: How strongly to enforce consistency with the previous chunk. This is a hyperparameter that can be tuned to balance the smoothness of the transitions and the reactivity of the policy. For 10 steps flow matching (SmolVLA, Pi0, Pi0.5), a value of 10.0 is a optimal value.
+
+**`prefix_attention_schedule`**: How to weight consistency across the overlap region.
+
+- `LINEAR`: Linear decay from inference_delay to execution_horizon
+- `EXP`: Exponential decay (recommended for getting started)
+- `ONES`: Full weight across entire execution_horizon
+- `ZEROS`: Binary (full weight up to inference_delay, then zero)
+
+**`inference_delay`**: How many timesteps of inference latency your system has. This is passed to `predict_action_chunk()` rather than the config, since it may vary at runtime.
+
+## Testing RTC Offline
+
+Before running on a real robot, test RTC with dataset samples to visualize how it works:
+
+```bash
+python examples/rtc/eval_dataset.py \
+    --policy.path=lerobot/pi0_libero_finetuned \
+    --dataset.repo_id=HuggingFaceVLA/libero \
+    --rtc.execution_horizon=10 \
+    --rtc.max_guidance_weight=10.0 \
+    --device=cuda
+```
+
+The script generates a visualization of the denoising process, comparing standard generation (left) with RTC (right). In the RTC plots, you can see how the first few steps (blue/purple lines) are guided to match the red ground truth trajectory (previous chunk's tail), ensuring a smooth transition between chunks.
+
+<p align="center">
+  <img
+    src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/flow_matching.png"
+    alt="Denoising steps with and without RTC"
+    width="100%"
+  />
+</p>
+
+## Testing RTC with a Real Robot
+
+```bash
+python examples/rtc/eval_with_real_robot.py \
+    --policy.path=${HF_USERNAME}/policy_repo_id \
+    --robot.type=so100_follower \
+    --robot.port=/dev/tty.usbmodem58FA0834591 \
+    --robot.cameras="{ gripper: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30}, front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
+    --task="Move green small object into the purple platform" \
+    --duration=120 \
+    --device=cuda
+```
+
+## How It Differs from the Async Inference in LeRobot
+
+Both RTC and [async inference](./async) improve real-time robot control, but they solve different problems.
+
+| Aspect        | Async Inference                                                            | RTC                                                 |
+| ------------- | -------------------------------------------------------------------------- | --------------------------------------------------- |
+| **Problem**   | Idle frames while waiting for inference                                    | Discontinuities between action chunks               |
+| **Solution**  | Decouple prediction from execution                                         | Guide new chunks to continue smoothly from previous |
+| **Benefit**   | No waiting, continuous action                                              | Smooth transitions, natural motion                  |
+| **Best Used** | Async inference is best used with large models with high inference latency | Flow-matching based policies                        |
+
+**Use both together** for maximum smoothness and reactivity!
+
+## Advanced: Debug Tracking
+
+RTC includes built-in debug tracking to help you understand what's happening during inference:
+
+```python
+# Enable debug tracking
+policy_cfg.rtc_config.debug = True
+policy_cfg.rtc_config.debug_maxlen = 100
+
+# After inference, access debug data
+debug_data = policy.rtc_processor.get_debug_data()
+
+# Visualize denoising steps, corrections, etc.
+from lerobot.policies.rtc.debug_visualizer import RTCDebugVisualizer
+visualizer = RTCDebugVisualizer()
+# ... create plots
+```
+
+See `examples/rtc/eval_dataset.py` for a complete example of visualization.
+
+## References
+
+- [Smooth-As-Butter Robot Policies](https://alexander-soare.github.io/robotics/2025/08/05/smooth-as-butter-robot-policies.html) - Excellent technical explanation with real robot results
+- [Physical Intelligence - Real-Time Chunking](https://www.physicalintelligence.company/research/real_time_chunking) - Original paper and research
+- [Kinetix RTC Implementation](https://github.com/Physical-Intelligence/real-time-chunking-kinetix) - Reference implementation from Physical Intelligence

examples/dataset/load_lerobot_dataset.py

@@ -132,17 +132,15 @@ print(f"\n{dataset[0][camera_key].shape=}")  # (4, c, h, w)
 print(f"{dataset[0]['observation.state'].shape=}")  # (6, c)
 print(f"{dataset[0]['action'].shape=}\n")  # (64, c)
 
-# Finally, our datasets are fully compatible with PyTorch dataloaders and samplers because they are just
-# PyTorch datasets.
-dataloader = torch.utils.data.DataLoader(
-    dataset,
-    num_workers=4,
-    batch_size=32,
-    shuffle=True,
-)
-
-for batch in dataloader:
-    print(f"{batch[camera_key].shape=}")  # (32, 4, c, h, w)
-    print(f"{batch['observation.state'].shape=}")  # (32, 6, c)
-    print(f"{batch['action'].shape=}")  # (32, 64, c)
-    break
+if __name__ == "__main__":
+    dataloader = torch.utils.data.DataLoader(
+        dataset,
+        num_workers=4,
+        batch_size=32,
+        shuffle=True,
+    )
+    for batch in dataloader:
+        print(f"{batch[camera_key].shape=}")  # (32, 4, c, h, w)
+        print(f"{batch['observation.state'].shape=}")  # (32, 6, c)
+        print(f"{batch['action'].shape=}")  # (32, 64, c)
+        break

examples/dataset/use_dataset_tools.py

@@ -30,9 +30,10 @@ Usage:
 import numpy as np
 
 from lerobot.datasets.dataset_tools import (
-    add_feature,
+    add_features,
     delete_episodes,
     merge_datasets,
+    modify_features,
     remove_feature,
     split_dataset,
 )
@@ -57,50 +58,56 @@ def main():
     print(f"Train split: {splits['train'].meta.total_episodes} episodes")
     print(f"Val split: {splits['val'].meta.total_episodes} episodes")
 
-    print("\n3. Adding a reward feature...")
+    print("\n3. Adding features...")
 
     reward_values = np.random.randn(dataset.meta.total_frames).astype(np.float32)
-    dataset_with_reward = add_feature(
-        dataset,
-        feature_name="reward",
-        feature_values=reward_values,
-        feature_info={
-            "dtype": "float32",
-            "shape": (1,),
-            "names": None,
-        },
-        repo_id="lerobot/pusht_with_reward",
-    )
 
     def compute_success(row_dict, episode_index, frame_index):
         episode_length = 10
         return float(frame_index >= episode_length - 10)
 
-    dataset_with_success = add_feature(
-        dataset_with_reward,
-        feature_name="success",
-        feature_values=compute_success,
-        feature_info={
-            "dtype": "float32",
-            "shape": (1,),
-            "names": None,
+    dataset_with_features = add_features(
+        dataset,
+        features={
+            "reward": (
+                reward_values,
+                {"dtype": "float32", "shape": (1,), "names": None},
+            ),
+            "success": (
+                compute_success,
+                {"dtype": "float32", "shape": (1,), "names": None},
+            ),
         },
-        repo_id="lerobot/pusht_with_reward_and_success",
+        repo_id="lerobot/pusht_with_features",
     )
 
-    print(f"New features: {list(dataset_with_success.meta.features.keys())}")
+    print(f"New features: {list(dataset_with_features.meta.features.keys())}")
 
     print("\n4. Removing the success feature...")
     dataset_cleaned = remove_feature(
-        dataset_with_success, feature_names="success", repo_id="lerobot/pusht_cleaned"
+        dataset_with_features, feature_names="success", repo_id="lerobot/pusht_cleaned"
     )
     print(f"Features after removal: {list(dataset_cleaned.meta.features.keys())}")
 
-    print("\n5. Merging train and val splits back together...")
+    print("\n5. Using modify_features to add and remove features simultaneously...")
+    dataset_modified = modify_features(
+        dataset_with_features,
+        add_features={
+            "discount": (
+                np.ones(dataset.meta.total_frames, dtype=np.float32) * 0.99,
+                {"dtype": "float32", "shape": (1,), "names": None},
+            ),
+        },
+        remove_features="reward",
+        repo_id="lerobot/pusht_modified",
+    )
+    print(f"Modified features: {list(dataset_modified.meta.features.keys())}")
+
+    print("\n6. Merging train and val splits back together...")
     merged = merge_datasets([splits["train"], splits["val"]], output_repo_id="lerobot/pusht_merged")
     print(f"Merged dataset: {merged.meta.total_episodes} episodes")
 
-    print("\n6. Complex workflow example...")
+    print("\n7. Complex workflow example...")
 
     if len(dataset.meta.camera_keys) > 1:
         camera_to_remove = dataset.meta.camera_keys[0]

examples/dataset_annotation/visualize_subtask_annotations.py

@@ -0,0 +1,525 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Visualize SARM Subtask Annotations
+
+This script creates visualizations of the subtask annotations generated by subtask_annotation.py.
+For each episode, it shows:
+- A timeline with dashed vertical lines at subtask boundaries
+- Sample frames from the episode at key points (start, middle, end of each subtask)
+- Color-coded subtask segments
+
+Usage:
+    python visualize_subtask_annotations.py --repo-id pepijn223/mydataset --video-key observation.images.top --num-episodes 5
+"""
+
+import argparse
+import random
+from pathlib import Path
+
+import cv2
+import matplotlib.pyplot as plt
+import matplotlib.patches as mpatches
+import numpy as np
+import pandas as pd
+from matplotlib.lines import Line2D
+from rich.console import Console
+
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.utils import load_episodes
+from lerobot.policies.sarm.sarm_utils import SubtaskAnnotation, Subtask, Timestamp
+
+
+def timestamp_to_seconds(timestamp: str) -> float:
+    """Convert MM:SS or SS timestamp to seconds"""
+    parts = timestamp.split(":")
+    if len(parts) == 2:
+        return int(parts[0]) * 60 + int(parts[1])
+    else:
+        return int(parts[0])
+
+
+def load_annotations_from_dataset(dataset_path: Path) -> dict[int, SubtaskAnnotation]:
+    """
+    Load annotations from LeRobot dataset parquet files.
+    
+    Reads subtask annotations from the episodes metadata parquet files.
+    """
+    episodes_dataset = load_episodes(dataset_path)
+    
+    if episodes_dataset is None or len(episodes_dataset) == 0:
+        return {}
+    
+    # Check if subtask columns exist
+    if "subtask_names" not in episodes_dataset.column_names:
+        return {}
+    
+    # Convert to pandas DataFrame for easier access
+    episodes_df = episodes_dataset.to_pandas()
+    
+    annotations = {}
+    
+    for ep_idx in episodes_df.index:
+        subtask_names = episodes_df.loc[ep_idx, "subtask_names"]
+        
+        # Skip episodes without annotations
+        if subtask_names is None or (isinstance(subtask_names, float) and pd.isna(subtask_names)):
+            continue
+        
+        start_times = episodes_df.loc[ep_idx, "subtask_start_times"]
+        end_times = episodes_df.loc[ep_idx, "subtask_end_times"]
+        
+        # Reconstruct SubtaskAnnotation from stored data
+        subtasks = []
+        for i, name in enumerate(subtask_names):
+            # Convert seconds back to MM:SS format
+            start_sec = int(start_times[i])
+            end_sec = int(end_times[i])
+            start_str = f"{start_sec // 60:02d}:{start_sec % 60:02d}"
+            end_str = f"{end_sec // 60:02d}:{end_sec % 60:02d}"
+            
+            subtasks.append(
+                Subtask(
+                    name=name,
+                    timestamps=Timestamp(start=start_str, end=end_str)
+                )
+            )
+        
+        annotations[int(ep_idx)] = SubtaskAnnotation(subtasks=subtasks)
+    
+    return annotations
+
+
+# Color palette for subtasks (colorblind-friendly)
+SUBTASK_COLORS = [
+    "#E69F00",  # Orange
+    "#56B4E9",  # Sky blue
+    "#009E73",  # Bluish green
+    "#F0E442",  # Yellow
+    "#0072B2",  # Blue
+    "#D55E00",  # Vermillion
+    "#CC79A7",  # Reddish purple
+    "#999999",  # Gray
+]
+
+
+def extract_frame_from_video(video_path: Path, timestamp: float) -> np.ndarray | None:
+    """Extract a single frame from video at given timestamp."""
+    cap = cv2.VideoCapture(str(video_path))
+    if not cap.isOpened():
+        return None
+    
+    # Set position to timestamp
+    cap.set(cv2.CAP_PROP_POS_MSEC, timestamp * 1000)
+    ret, frame = cap.read()
+    cap.release()
+    
+    if ret:
+        # Convert BGR to RGB
+        return cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+    return None
+
+
+def visualize_episode(
+    episode_idx: int,
+    annotation,
+    video_path: Path,
+    video_start_timestamp: float,
+    video_end_timestamp: float,
+    fps: int,
+    output_path: Path,
+    video_key: str,
+):
+    """
+    Create visualization for a single episode.
+    
+    Shows:
+    - Top row: Sample frames from the episode (one per subtask)
+    - Bottom: Timeline with subtask segments and boundary lines
+    """
+    subtasks = annotation.subtasks
+    num_subtasks = len(subtasks)
+    
+    if num_subtasks == 0:
+        print(f"No subtasks found for episode {episode_idx}")
+        return
+    
+    # Calculate episode duration
+    episode_duration = video_end_timestamp - video_start_timestamp
+    
+    # Extract sample frames - get frame from middle of each subtask
+    sample_frames = []
+    frame_timestamps = []
+    
+    for subtask in subtasks:
+        start_sec = timestamp_to_seconds(subtask.timestamps.start)
+        end_sec = timestamp_to_seconds(subtask.timestamps.end)
+        mid_sec = (start_sec + end_sec) / 2
+        
+        # Convert to video timestamp (add video_start_timestamp offset)
+        video_timestamp = video_start_timestamp + mid_sec
+        frame_timestamps.append(mid_sec)
+        
+        frame = extract_frame_from_video(video_path, video_timestamp)
+        sample_frames.append(frame)
+    
+    # Create figure
+    fig = plt.figure(figsize=(16, 10))
+    
+    # Use a dark background for better contrast
+    fig.patch.set_facecolor('#1a1a2e')
+    
+    # Calculate grid layout
+    # Top section: frames (variable number of columns based on subtasks)
+    # Bottom section: timeline
+    
+    # Create gridspec
+    gs = fig.add_gridspec(
+        2, max(num_subtasks, 1), 
+        height_ratios=[2, 1],
+        hspace=0.3,
+        wspace=0.1,
+        left=0.05, right=0.95,
+        top=0.88, bottom=0.1
+    )
+    
+    # Add title
+    fig.suptitle(
+        f"Episode {episode_idx} - Subtask Annotations",
+        fontsize=18,
+        fontweight='bold',
+        color='white',
+        y=0.96
+    )
+    
+    # Add subtitle with video info
+    fig.text(
+        0.5, 0.91,
+        f"Camera: {video_key} | Duration: {episode_duration:.1f}s | {num_subtasks} subtasks",
+        ha='center',
+        fontsize=11,
+        color='#888888'
+    )
+    
+    # Plot sample frames
+    for i, (frame, subtask) in enumerate(zip(sample_frames, subtasks)):
+        ax = fig.add_subplot(gs[0, i])
+        ax.set_facecolor('#16213e')
+        
+        if frame is not None:
+            ax.imshow(frame)
+        else:
+            ax.text(0.5, 0.5, "Frame\nN/A", ha='center', va='center', 
+                   fontsize=12, color='white', transform=ax.transAxes)
+        
+        ax.set_title(
+            f"{subtask.name}",
+            fontsize=10,
+            fontweight='bold',
+            color=SUBTASK_COLORS[i % len(SUBTASK_COLORS)],
+            pad=8
+        )
+        ax.axis('off')
+        
+        # Add frame timestamp below
+        ax.text(
+            0.5, -0.08,
+            f"t={frame_timestamps[i]:.1f}s",
+            ha='center',
+            fontsize=9,
+            color='#888888',
+            transform=ax.transAxes
+        )
+    
+    # Create timeline subplot spanning all columns
+    ax_timeline = fig.add_subplot(gs[1, :])
+    ax_timeline.set_facecolor('#16213e')
+    
+    # Get total duration from last subtask end time
+    total_duration = timestamp_to_seconds(subtasks[-1].timestamps.end)
+    
+    # Draw subtask segments as colored bars
+    bar_height = 0.6
+    bar_y = 0.5
+    
+    for i, subtask in enumerate(subtasks):
+        start_sec = timestamp_to_seconds(subtask.timestamps.start)
+        end_sec = timestamp_to_seconds(subtask.timestamps.end)
+        color = SUBTASK_COLORS[i % len(SUBTASK_COLORS)]
+        
+        # Draw segment bar
+        rect = mpatches.FancyBboxPatch(
+            (start_sec, bar_y - bar_height/2),
+            end_sec - start_sec,
+            bar_height,
+            boxstyle="round,pad=0.02,rounding_size=0.1",
+            facecolor=color,
+            edgecolor='white',
+            linewidth=1.5,
+            alpha=0.85
+        )
+        ax_timeline.add_patch(rect)
+        
+        # Add subtask label inside bar
+        mid_x = (start_sec + end_sec) / 2
+        duration = end_sec - start_sec
+        
+        # Only add text if segment is wide enough
+        if duration > total_duration * 0.08:
+            ax_timeline.text(
+                mid_x, bar_y,
+                subtask.name,
+                ha='center', va='center',
+                fontsize=9,
+                fontweight='bold',
+                color='black' if i in [3] else 'white',  # Yellow needs dark text
+                rotation=0 if duration > total_duration * 0.15 else 45
+            )
+    
+    # Draw boundary lines (dashed vertical lines between subtasks)
+    boundary_times = []
+    for i, subtask in enumerate(subtasks):
+        start_sec = timestamp_to_seconds(subtask.timestamps.start)
+        end_sec = timestamp_to_seconds(subtask.timestamps.end)
+        
+        # Add start boundary (except for first subtask at t=0)
+        if i == 0 and start_sec > 0:
+            boundary_times.append(start_sec)
+        elif i > 0:
+            boundary_times.append(start_sec)
+        
+        # Add end boundary for last subtask
+        if i == len(subtasks) - 1:
+            boundary_times.append(end_sec)
+    
+    # Draw dashed lines at boundaries
+    for t in boundary_times:
+        ax_timeline.axvline(
+            x=t, 
+            ymin=0.1, ymax=0.9,
+            color='white', 
+            linestyle='--', 
+            linewidth=2,
+            alpha=0.9
+        )
+        
+        # Add time label below line
+        ax_timeline.text(
+            t, 0.0,
+            f"{int(t//60):02d}:{int(t%60):02d}",
+            ha='center', va='top',
+            fontsize=8,
+            color='#cccccc'
+        )
+    
+    # Add start line at t=0
+    ax_timeline.axvline(x=0, ymin=0.1, ymax=0.9, color='#00ff00', linestyle='-', linewidth=2.5, alpha=0.9)
+    ax_timeline.text(0, 0.0, "00:00", ha='center', va='top', fontsize=8, color='#00ff00', fontweight='bold')
+    
+    # Configure timeline axes
+    ax_timeline.set_xlim(-total_duration * 0.02, total_duration * 1.02)
+    ax_timeline.set_ylim(-0.3, 1.2)
+    ax_timeline.set_xlabel("Time (seconds)", fontsize=11, color='white', labelpad=10)
+    ax_timeline.set_ylabel("")
+    
+    # Style the axes
+    ax_timeline.spines['top'].set_visible(False)
+    ax_timeline.spines['right'].set_visible(False)
+    ax_timeline.spines['left'].set_visible(False)
+    ax_timeline.spines['bottom'].set_color('#444444')
+    ax_timeline.tick_params(axis='x', colors='#888888', labelsize=9)
+    ax_timeline.tick_params(axis='y', left=False, labelleft=False)
+    
+    # Add x-axis ticks at regular intervals
+    tick_interval = max(1, int(total_duration / 10))
+    ax_timeline.set_xticks(np.arange(0, total_duration + tick_interval, tick_interval))
+    
+    # Add legend explaining line styles
+    legend_elements = [
+        Line2D([0], [0], color='#00ff00', linewidth=2.5, linestyle='-', label='Start'),
+        Line2D([0], [0], color='white', linewidth=2, linestyle='--', label='Subtask boundary'),
+    ]
+    ax_timeline.legend(
+        handles=legend_elements, 
+        loc='upper right',
+        framealpha=0.3,
+        facecolor='#16213e',
+        edgecolor='#444444',
+        fontsize=9,
+        labelcolor='white'
+    )
+    
+    # Save figure
+    plt.savefig(output_path, dpi=150, facecolor=fig.get_facecolor(), edgecolor='none', bbox_inches='tight')
+    plt.close()
+    
+    return output_path
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Visualize SARM subtask annotations",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+    )
+    parser.add_argument(
+        "--repo-id",
+        type=str,
+        required=True,
+        help="HuggingFace dataset repository ID",
+    )
+    parser.add_argument(
+        "--num-episodes",
+        type=int,
+        default=5,
+        help="Number of random episodes to visualize (default: 5)",
+    )
+    parser.add_argument(
+        "--episodes",
+        type=int,
+        nargs="+",
+        default=None,
+        help="Specific episode indices to visualize (overrides --num-episodes)",
+    )
+    parser.add_argument(
+        "--video-key",
+        type=str,
+        default=None,
+        help="Camera/video key to use. If not specified, uses first available.",
+    )
+    parser.add_argument(
+        "--output-dir",
+        type=str,
+        default="./subtask_viz",
+        help="Output directory for visualizations (default: ./subtask_viz)",
+    )
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=None,
+        help="Random seed for reproducibility",
+    )
+    
+    args = parser.parse_args()
+    
+    console = Console()
+    
+    # Set random seed if specified
+    if args.seed is not None:
+        random.seed(args.seed)
+    
+    console.print(f"\n[cyan]Loading dataset: {args.repo_id}[/cyan]")
+    dataset = LeRobotDataset(args.repo_id, download_videos=True)
+    fps = dataset.fps
+    
+    # Get video key
+    if args.video_key:
+        if args.video_key not in dataset.meta.video_keys:
+            console.print(f"[red]Error: Video key '{args.video_key}' not found[/red]")
+            console.print(f"[yellow]Available: {', '.join(dataset.meta.video_keys)}[/yellow]")
+            return
+        video_key = args.video_key
+    else:
+        video_key = dataset.meta.video_keys[0]
+    
+    console.print(f"[cyan]Using camera: {video_key}[/cyan]")
+    console.print(f"[cyan]FPS: {fps}[/cyan]")
+    
+    # Load annotations
+    console.print(f"\n[cyan]Loading annotations...[/cyan]")
+    annotations = load_annotations_from_dataset(dataset.root)
+    
+    if not annotations:
+        console.print("[red]Error: No annotations found in dataset[/red]")
+        console.print("[yellow]Run subtask_annotation.py first to generate annotations[/yellow]")
+        return
+    
+    console.print(f"[green]Found {len(annotations)} annotated episodes[/green]")
+    
+    # Determine which episodes to visualize
+    if args.episodes:
+        episode_indices = args.episodes
+        # Validate episodes exist
+        for ep in episode_indices:
+            if ep not in annotations:
+                console.print(f"[yellow]Warning: Episode {ep} has no annotation, skipping[/yellow]")
+        episode_indices = [ep for ep in episode_indices if ep in annotations]
+    else:
+        # Random selection
+        available_episodes = list(annotations.keys())
+        num_to_select = min(args.num_episodes, len(available_episodes))
+        episode_indices = random.sample(available_episodes, num_to_select)
+        episode_indices.sort()
+    
+    if not episode_indices:
+        console.print("[red]Error: No valid episodes to visualize[/red]")
+        return
+    
+    console.print(f"[cyan]Visualizing episodes: {episode_indices}[/cyan]")
+    
+    # Create output directory
+    output_dir = Path(args.output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+    
+    # Generate visualizations
+    for ep_idx in episode_indices:
+        console.print(f"\n[cyan]Processing episode {ep_idx}...[/cyan]")
+        
+        annotation = annotations[ep_idx]
+        
+        # Get video path and timestamps
+        video_path = dataset.root / dataset.meta.get_video_file_path(ep_idx, video_key)
+        
+        if not video_path.exists():
+            console.print(f"[red]Video not found: {video_path}[/red]")
+            continue
+        
+        # Get episode-specific timestamps within the video file
+        video_path_key = f"videos/{video_key}/from_timestamp"
+        video_path_key_to = f"videos/{video_key}/to_timestamp"
+        
+        video_start_timestamp = float(dataset.meta.episodes[video_path_key][ep_idx])
+        video_end_timestamp = float(dataset.meta.episodes[video_path_key_to][ep_idx])
+        
+        # Create visualization
+        output_path = output_dir / f"episode_{ep_idx:04d}_subtasks.png"
+        
+        try:
+            visualize_episode(
+                episode_idx=ep_idx,
+                annotation=annotation,
+                video_path=video_path,
+                video_start_timestamp=video_start_timestamp,
+                video_end_timestamp=video_end_timestamp,
+                fps=fps,
+                output_path=output_path,
+                video_key=video_key,
+            )
+            console.print(f"[green]✓ Saved: {output_path}[/green]")
+        except Exception as e:
+            console.print(f"[red]✗ Failed to visualize episode {ep_idx}: {e}[/red]")
+    
+    # Print summary
+    console.print(f"\n[bold green]{'=' * 50}[/bold green]")
+    console.print(f"[bold green]Visualization Complete![/bold green]")
+    console.print(f"[bold green]{'=' * 50}[/bold green]")
+    console.print(f"Output directory: {output_dir.absolute()}")
+    console.print(f"Episodes visualized: {len(episode_indices)}")
+
+
+if __name__ == "__main__":
+    main()
+

examples/lekiwi/evaluate.py

@@ -133,4 +133,6 @@ while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
 log_say("Stop recording")
 robot.disconnect()
 listener.stop()
+
+dataset.finalize()
 dataset.push_to_hub()

examples/lekiwi/record.py

@@ -130,4 +130,6 @@ robot.disconnect()
 leader_arm.disconnect()
 keyboard.disconnect()
 listener.stop()
+
+dataset.finalize()
 dataset.push_to_hub()

examples/phone_to_so100/evaluate.py

@@ -194,4 +194,6 @@ for episode_idx in range(NUM_EPISODES):
 log_say("Stop recording")
 robot.disconnect()
 listener.stop()
+
+dataset.finalize()
 dataset.push_to_hub()

examples/phone_to_so100/record.py

@@ -200,4 +200,6 @@ log_say("Stop recording")
 robot.disconnect()
 phone.disconnect()
 listener.stop()
+
+dataset.finalize()
 dataset.push_to_hub()

examples/port_datasets/port_droid.py

@@ -362,6 +362,8 @@ def port_droid(
         lerobot_dataset.save_episode()
         logging.info("Save_episode")
 
+    lerobot_dataset.finalize()
+
     if push_to_hub:
         lerobot_dataset.push_to_hub(
             # Add openx tag, since it belongs to the openx collection of datasets

examples/port_datasets/slurm_aggregate_shards.py

@@ -15,16 +15,12 @@
 # limitations under the License.
 
 import argparse
-import logging
 from pathlib import Path
 
 from datatrove.executor import LocalPipelineExecutor
 from datatrove.executor.slurm import SlurmPipelineExecutor
 from datatrove.pipeline.base import PipelineStep
-from port_datasets.droid_rlds.port_droid import DROID_SHARDS
-
-from lerobot.datasets.aggregate import aggregate_datasets
-from lerobot.utils.utils import init_logging
+from port_droid import DROID_SHARDS
 
 
 class AggregateDatasets(PipelineStep):
@@ -38,6 +34,11 @@ class AggregateDatasets(PipelineStep):
         self.aggr_repo_id = aggregated_repo_id
 
     def run(self, data=None, rank: int = 0, world_size: int = 1):
+        import logging
+
+        from lerobot.datasets.aggregate import aggregate_datasets
+        from lerobot.utils.utils import init_logging
+
         init_logging()
 
         # Since aggregate_datasets already handles parallel processing internally,

examples/port_datasets/slurm_port_shards.py

@@ -20,7 +20,7 @@ from pathlib import Path
 from datatrove.executor import LocalPipelineExecutor
 from datatrove.executor.slurm import SlurmPipelineExecutor
 from datatrove.pipeline.base import PipelineStep
-from port_datasets.droid_rlds.port_droid import DROID_SHARDS
+from port_droid import DROID_SHARDS
 
 
 class PortDroidShards(PipelineStep):
@@ -35,7 +35,7 @@ class PortDroidShards(PipelineStep):
 
     def run(self, data=None, rank: int = 0, world_size: int = 1):
         from datasets.utils.tqdm import disable_progress_bars
-        from port_datasets.droid_rlds.port_droid import port_droid, validate_dataset
+        from port_droid import port_droid, validate_dataset
 
         from lerobot.utils.utils import init_logging
 

examples/port_datasets/slurm_upload.py

@@ -24,7 +24,7 @@ from datatrove.executor.slurm import SlurmPipelineExecutor
 from datatrove.pipeline.base import PipelineStep
 from huggingface_hub import HfApi
 from huggingface_hub.constants import REPOCARD_NAME
-from port_datasets.droid_rlds.port_droid import DROID_SHARDS
+from port_droid import DROID_SHARDS
 
 from lerobot.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDatasetMetadata
 from lerobot.datasets.utils import create_lerobot_dataset_card
@@ -185,11 +185,11 @@ class UploadDataset(PipelineStep):
 
 
 def make_upload_executor(
-    repo_id, job_name, logs_dir, workers, partition, cpus_per_task, mem_per_cpu, slurm=True
+    repo_id, job_name, logs_dir, workers, partition, cpus_per_task, mem_per_cpu, private=False, slurm=True
 ):
     kwargs = {
         "pipeline": [
-            UploadDataset(repo_id),
+            UploadDataset(repo_id, private=private),
         ],
         "logging_dir": str(logs_dir / job_name),
     }
@@ -267,6 +267,12 @@ def main():
         default="1950M",
         help="Memory per cpu that each worker will use.",
     )
+    parser.add_argument(
+        "--private",
+        action="store_true",
+        default=False,
+        help="Whether to create a private repository.",
+    )
 
     init_logging()
 

examples/rtc/eval_dataset.py

@@ -0,0 +1,951 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Evaluate Real-Time Chunking (RTC) performance on dataset samples.
+
+This script takes two random samples from a dataset:
+- Uses actions from the first sample as previous chunk
+- Generates new actions for the second sample with and without RTC
+
+It compares action predictions with and without RTC on dataset samples,
+measuring consistency and ground truth alignment.
+
+Usage:
+    # Basic usage with smolvla policy
+    uv run python examples/rtc/eval_dataset.py \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --dataset.repo_id=helper2424/check_rtc \
+        --rtc.execution_horizon=8 \
+        --device=mps \
+        --rtc.max_guidance_weight=10.0 \
+        --rtc.prefix_attention_schedule=EXP \
+        --seed=10
+
+    # Basic usage with pi0.5 policy
+    uv run python examples/rtc/eval_dataset.py \
+        --policy.path=lerobot/pi05_libero_finetuned \
+        --dataset.repo_id=HuggingFaceVLA/libero \
+        --rtc.execution_horizon=10 \
+        --device=mps
+        --seed=10
+
+    # Basic usage with pi0.5 policy with cuda device
+    uv run python examples/rtc/eval_dataset.py \
+        --policy.path=lerobot/pi05_libero_finetuned \
+        --dataset.repo_id=HuggingFaceVLA/libero \
+        --rtc.execution_horizon=8 \
+        --device=cuda
+
+    # Basic usage with pi0 policy with cuda device
+    uv run python examples/rtc/eval_dataset.py \
+        --policy.path=lerobot/pi0_libero_finetuned \
+        --dataset.repo_id=HuggingFaceVLA/libero \
+        --rtc.execution_horizon=8 \
+        --device=cuda
+
+    uv run python examples/rtc/eval_dataset.py \
+        --policy.path=lipsop/reuben_pi0 \
+        --dataset.repo_id=ReubenLim/so101_cube_in_cup \
+        --rtc.execution_horizon=8 \
+        --device=cuda
+
+    # With torch.compile for faster inference (PyTorch 2.0+)
+    # Note: CUDA graphs disabled by default due to in-place ops in denoising loop
+    uv run python examples/rtc/eval_dataset.py \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --dataset.repo_id=helper2424/check_rtc \
+        --rtc.execution_horizon=8 \
+        --device=mps \
+        --use_torch_compile=true \
+        --torch_compile_mode=max-autotune
+
+    # With torch.compile on CUDA (CUDA graphs disabled by default)
+    uv run python examples/rtc/eval_dataset.py \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --dataset.repo_id=helper2424/check_rtc \
+        --rtc.execution_horizon=8 \
+        --device=cuda \
+        --use_torch_compile=true \
+        --torch_compile_mode=reduce-overhead
+
+    # Enable CUDA graphs (advanced - may cause tensor aliasing errors)
+    uv run python examples/rtc/eval_dataset.py \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --dataset.repo_id=helper2424/check_rtc \
+        --use_torch_compile=true \
+        --torch_compile_backend=inductor \
+        --torch_compile_mode=max-autotune \
+        --torch_compile_disable_cudagraphs=false
+"""
+
+import gc
+import logging
+import os
+import random
+from dataclasses import dataclass, field
+
+import numpy as np
+import torch
+
+try:
+    import matplotlib.pyplot as plt
+
+    MATPLOTLIB_AVAILABLE = True
+except ImportError:
+    MATPLOTLIB_AVAILABLE = False
+    plt = None
+
+from lerobot.configs import parser
+from lerobot.configs.default import DatasetConfig
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import RTCAttentionSchedule
+from lerobot.datasets.factory import resolve_delta_timestamps
+from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.policies.factory import get_policy_class, make_pre_post_processors
+from lerobot.policies.rtc.configuration_rtc import RTCConfig
+from lerobot.policies.rtc.debug_visualizer import RTCDebugVisualizer
+from lerobot.utils.hub import HubMixin
+from lerobot.utils.utils import init_logging
+
+
+def set_seed(seed: int):
+    """Set random seed for reproducibility."""
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed(seed)
+        torch.cuda.manual_seed_all(seed)
+    if torch.backends.mps.is_available():
+        torch.mps.manual_seed(seed)
+    torch.backends.cudnn.deterministic = True
+    torch.backends.cudnn.benchmark = False
+
+
+def _check_matplotlib_available():
+    """Check if matplotlib is available, raise helpful error if not."""
+    if not MATPLOTLIB_AVAILABLE:
+        raise ImportError(
+            "matplotlib is required for RTC debug visualizations. "
+            "Please install it by running:\n"
+            "  uv pip install matplotlib"
+        )
+
+
+@dataclass
+class RTCEvalConfig(HubMixin):
+    """Configuration for RTC evaluation."""
+
+    # Policy configuration
+    policy: PreTrainedConfig | None = None
+
+    # Dataset configuration
+    dataset: DatasetConfig = field(default_factory=DatasetConfig)
+
+    # RTC configuration
+    rtc: RTCConfig = field(
+        default_factory=lambda: RTCConfig(
+            enabled=True,
+            execution_horizon=20,
+            max_guidance_weight=10.0,
+            prefix_attention_schedule=RTCAttentionSchedule.EXP,
+            debug=True,
+            debug_maxlen=1000,
+        )
+    )
+
+    # Device configuration
+    device: str | None = field(
+        default=None,
+        metadata={"help": "Device to run on (cuda, cpu, mps, auto)"},
+    )
+
+    # Output configuration
+    output_dir: str = field(
+        default="rtc_debug_output",
+        metadata={"help": "Directory to save debug visualizations"},
+    )
+
+    # Seed configuration
+    seed: int = field(
+        default=42,
+        metadata={"help": "Random seed for reproducibility"},
+    )
+
+    inference_delay: int = field(
+        default=4,
+        metadata={"help": "Inference delay for RTC"},
+    )
+
+    # Torch compile configuration
+    use_torch_compile: bool = field(
+        default=False,
+        metadata={"help": "Use torch.compile for faster inference (PyTorch 2.0+)"},
+    )
+
+    torch_compile_backend: str = field(
+        default="inductor",
+        metadata={"help": "Backend for torch.compile (inductor, aot_eager, cudagraphs)"},
+    )
+
+    torch_compile_mode: str = field(
+        default="default",
+        metadata={"help": "Compilation mode (default, reduce-overhead, max-autotune)"},
+    )
+
+    torch_compile_disable_cudagraphs: bool = field(
+        default=True,
+        metadata={
+            "help": "Disable CUDA graphs in torch.compile. Required due to in-place tensor "
+            "operations in denoising loop (x_t += dt * v_t) which cause tensor aliasing issues."
+        },
+    )
+
+    def __post_init__(self):
+        # Parse policy path
+        policy_path = parser.get_path_arg("policy")
+        if policy_path:
+            cli_overrides = parser.get_cli_overrides("policy")
+            self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
+            self.policy.pretrained_path = policy_path
+        else:
+            raise ValueError("Policy path is required (--policy.path)")
+
+        # Auto-detect device if not specified
+        if self.device is None or self.device == "auto":
+            if torch.cuda.is_available():
+                self.device = "cuda"
+            elif torch.backends.mps.is_available():
+                self.device = "mps"
+            else:
+                self.device = "cpu"
+            logging.info(f"Auto-detected device: {self.device}")
+
+    @classmethod
+    def __get_path_fields__(cls) -> list[str]:
+        """This enables the parser to load config from the policy using `--policy.path=local/dir`"""
+        return ["policy"]
+
+
+class RTCEvaluator:
+    """Evaluator for RTC on dataset samples."""
+
+    def __init__(self, cfg: RTCEvalConfig):
+        self.cfg = cfg
+        self.device = cfg.device
+
+        # Load dataset with proper delta_timestamps based on policy configuration
+        # Calculate delta_timestamps using the same logic as make_dataset factory
+        logging.info(f"Loading dataset: {cfg.dataset.repo_id}")
+
+        # Get dataset metadata to extract FPS
+        ds_meta = LeRobotDatasetMetadata(cfg.dataset.repo_id)
+
+        # Calculate delta_timestamps from policy's delta_indices
+        delta_timestamps = resolve_delta_timestamps(cfg.policy, ds_meta)
+
+        # Create dataset with calculated delta_timestamps
+        self.dataset = LeRobotDataset(
+            cfg.dataset.repo_id,
+            delta_timestamps=delta_timestamps,
+        )
+        logging.info(f"Dataset loaded: {len(self.dataset)} samples, {self.dataset.num_episodes} episodes")
+
+        # Create preprocessor/postprocessor
+        self.preprocessor, self.postprocessor = make_pre_post_processors(
+            policy_cfg=cfg.policy,
+            pretrained_path=cfg.policy.pretrained_path,
+            preprocessor_overrides={
+                "device_processor": {"device": self.device},
+            },
+        )
+
+        logging.info("=" * 80)
+        logging.info("Ready to run evaluation with sequential policy loading:")
+        logging.info("  1. policy_prev_chunk - Generate reference chunk, then destroy")
+        logging.info("  2. policy_no_rtc - Generate without RTC, then destroy")
+        logging.info("  3. policy_rtc - Generate with RTC, then destroy")
+        logging.info("  Note: Only one policy in memory at a time for efficient memory usage")
+        logging.info("=" * 80)
+
+    def _init_policy(self, name: str, rtc_enabled: bool, rtc_debug: bool):
+        """Initialize a single policy instance with specified RTC configuration.
+
+        Args:
+            name: Name identifier for logging purposes
+            rtc_enabled: Whether to enable RTC for this policy
+            rtc_debug: Whether to enable debug tracking for this policy
+
+        Returns:
+            Configured policy instance with optional torch.compile applied
+        """
+        logging.info(f"Initializing {name}...")
+
+        # Load policy from pretrained
+        policy_class = get_policy_class(self.cfg.policy.type)
+
+        config = PreTrainedConfig.from_pretrained(self.cfg.policy.pretrained_path)
+
+        if self.cfg.policy.type == "pi05" or self.cfg.policy.type == "pi0":
+            config.compile_model = self.cfg.use_torch_compile
+
+        policy = policy_class.from_pretrained(self.cfg.policy.pretrained_path, config=config)
+        policy = policy.to(self.device)
+        policy.eval()
+
+        # Configure RTC
+        rtc_config = RTCConfig(
+            enabled=rtc_enabled,
+            execution_horizon=self.cfg.rtc.execution_horizon,
+            max_guidance_weight=self.cfg.rtc.max_guidance_weight,
+            prefix_attention_schedule=self.cfg.rtc.prefix_attention_schedule,
+            debug=rtc_debug,
+            debug_maxlen=self.cfg.rtc.debug_maxlen,
+        )
+        policy.config.rtc_config = rtc_config
+        policy.init_rtc_processor()
+
+        logging.info(f"  RTC enabled: {rtc_enabled}")
+        logging.info(f"  RTC debug: {rtc_debug}")
+        logging.info(f"  Policy config: {config}")
+
+        # Apply torch.compile to predict_action_chunk method if enabled
+        if self.cfg.use_torch_compile:
+            policy = self._apply_torch_compile(policy, name)
+
+        logging.info(f"✓ {name} initialized successfully")
+        return policy
+
+    def _apply_torch_compile(self, policy, policy_name: str):
+        """Apply torch.compile to the policy's predict_action_chunk method.
+
+        Args:
+            policy: Policy instance to compile
+            policy_name: Name for logging purposes
+
+        Returns:
+            Policy with compiled predict_action_chunk method
+        """
+
+        # PI models handle their own compilation
+        if policy.type == "pi05" or policy.type == "pi0":
+            return policy
+
+        try:
+            # Check if torch.compile is available (PyTorch 2.0+)
+            if not hasattr(torch, "compile"):
+                logging.warning(
+                    f"  [{policy_name}] torch.compile is not available. Requires PyTorch 2.0+. "
+                    f"Current version: {torch.__version__}. Skipping compilation."
+                )
+                return policy
+
+            logging.info(f"  [{policy_name}] Applying torch.compile to predict_action_chunk...")
+            logging.info(f"    Backend: {self.cfg.torch_compile_backend}")
+            logging.info(f"    Mode: {self.cfg.torch_compile_mode}")
+            logging.info(f"    Disable CUDA graphs: {self.cfg.torch_compile_disable_cudagraphs}")
+            logging.info("    Note: Debug tracker excluded from compilation via @torch._dynamo.disable")
+
+            # Compile the predict_action_chunk method
+            # - Debug tracker is excluded from compilation via @torch._dynamo.disable
+            # - CUDA graphs disabled to prevent tensor aliasing from in-place ops (x_t += dt * v_t)
+            compile_kwargs = {
+                "backend": self.cfg.torch_compile_backend,
+                "mode": self.cfg.torch_compile_mode,
+            }
+
+            # Disable CUDA graphs if requested (prevents tensor aliasing issues)
+            if self.cfg.torch_compile_disable_cudagraphs:
+                compile_kwargs["options"] = {"triton.cudagraphs": False}
+
+            original_method = policy.predict_action_chunk
+            compiled_method = torch.compile(original_method, **compile_kwargs)
+            policy.predict_action_chunk = compiled_method
+            logging.info(f"  ✓ [{policy_name}] Successfully compiled predict_action_chunk")
+
+        except Exception as e:
+            logging.error(f"  [{policy_name}] Failed to apply torch.compile: {e}")
+            logging.warning(f"  [{policy_name}] Continuing without torch.compile")
+
+        return policy
+
+    def _destroy_policy(self, policy, policy_name: str):
+        """Explicitly destroy a policy and free all associated memory.
+
+        This method performs aggressive cleanup to ensure maximum memory is freed,
+        which is critical for large models (e.g., VLAs with billions of parameters).
+
+        Args:
+            policy: Policy instance to destroy
+            policy_name: Name for logging purposes
+        """
+        logging.info(f"  Destroying {policy_name} and freeing memory...")
+
+        try:
+            # Step 1: Move policy to CPU to free GPU/MPS memory
+            policy.cpu()
+
+            # Step 2: Delete the policy object
+            del policy
+
+            # Step 3: Force garbage collection to reclaim memory immediately
+            gc.collect()
+
+            # Step 4: Clear device-specific caches
+            if torch.cuda.is_available():
+                torch.cuda.empty_cache()
+                torch.cuda.synchronize()  # Ensure all operations complete
+
+            if torch.backends.mps.is_available():
+                torch.mps.empty_cache()
+
+            logging.info(f"  ✓ {policy_name} destroyed and memory freed")
+
+        except Exception as e:
+            logging.warning(f"  Warning: Error during {policy_name} cleanup: {e}")
+
+    def run_evaluation(self):
+        """Run evaluation on two random dataset samples using three separate policies.
+
+        Note: Policies are deinitalized after each step to free memory. Large models
+        (e.g., VLA models with billions of parameters) cannot fit three instances in
+        memory simultaneously. By deleting and garbage collecting after each step,
+        we ensure only one policy is loaded at a time.
+        """
+        # Create output directory
+        os.makedirs(self.cfg.output_dir, exist_ok=True)
+        logging.info(f"Output directory: {self.cfg.output_dir}")
+
+        logging.info("=" * 80)
+        logging.info("Starting RTC evaluation")
+        logging.info(f"Inference delay: {self.cfg.inference_delay}")
+        logging.info("=" * 80)
+
+        # Load two random samples from dataset
+        data_loader = torch.utils.data.DataLoader(self.dataset, batch_size=1, shuffle=True)
+        loader_iter = iter(data_loader)
+        first_sample = next(loader_iter)
+        second_sample = next(loader_iter)
+
+        preprocessed_first_sample = self.preprocessor(first_sample)
+        preprocessed_second_sample = self.preprocessor(second_sample)
+
+        # ============================================================================
+        # Step 1: Generate previous chunk using policy_prev_chunk
+        # ============================================================================
+        # This policy is only used to generate the reference chunk and then freed
+        logging.info("=" * 80)
+        logging.info("Step 1: Generating previous chunk with policy_prev_chunk")
+        logging.info("=" * 80)
+
+        # Initialize policy 1
+        policy_prev_chunk_policy = self._init_policy(
+            name="policy_prev_chunk",
+            rtc_enabled=False,
+            rtc_debug=False,
+        )
+        with torch.no_grad():
+            prev_chunk_left_over = policy_prev_chunk_policy.predict_action_chunk(
+                preprocessed_first_sample,
+            )[:, :25, :].squeeze(0)
+        logging.info(f"  Generated prev_chunk shape: {prev_chunk_left_over.shape}")
+
+        # Destroy policy_prev_chunk to free memory for large models
+        self._destroy_policy(policy_prev_chunk_policy, "policy_prev_chunk")
+
+        # ============================================================================
+        # Step 2: Generate actions WITHOUT RTC using policy_no_rtc
+        # ============================================================================
+        logging.info("=" * 80)
+        logging.info("Step 2: Generating actions WITHOUT RTC with policy_no_rtc")
+        logging.info("=" * 80)
+
+        set_seed(self.cfg.seed)
+
+        # Initialize policy 2
+        policy_no_rtc_policy = self._init_policy(
+            name="policy_no_rtc",
+            rtc_enabled=False,
+            rtc_debug=True,
+        )
+
+        # Sample noise (use same noise for both RTC and non-RTC for fair comparison)
+        noise_size = (1, policy_no_rtc_policy.config.chunk_size, policy_no_rtc_policy.config.max_action_dim)
+        noise = policy_no_rtc_policy.model.sample_noise(noise_size, self.device)
+        noise_clone = noise.clone()
+        policy_no_rtc_policy.rtc_processor.reset_tracker()
+        with torch.no_grad():
+            no_rtc_actions = policy_no_rtc_policy.predict_action_chunk(
+                preprocessed_second_sample,
+                noise=noise,
+            )
+        no_rtc_tracked_steps = policy_no_rtc_policy.rtc_processor.tracker.get_all_steps()
+        logging.info(f"  Tracked {len(no_rtc_tracked_steps)} steps without RTC")
+        logging.info(f"  Generated no_rtc_actions shape: {no_rtc_actions.shape}")
+
+        # Destroy policy_no_rtc to free memory before loading policy_rtc
+        self._destroy_policy(policy_no_rtc_policy, "policy_no_rtc")
+
+        # ============================================================================
+        # Step 3: Generate actions WITH RTC using policy_rtc
+        # ============================================================================
+        logging.info("=" * 80)
+        logging.info("Step 3: Generating actions WITH RTC with policy_rtc")
+        logging.info("=" * 80)
+
+        set_seed(self.cfg.seed)
+
+        # Initialize policy 3
+        policy_rtc_policy = self._init_policy(
+            name="policy_rtc",
+            rtc_enabled=True,
+            rtc_debug=True,
+        )
+        policy_rtc_policy.rtc_processor.reset_tracker()
+        with torch.no_grad():
+            rtc_actions = policy_rtc_policy.predict_action_chunk(
+                preprocessed_second_sample,
+                noise=noise_clone,
+                inference_delay=self.cfg.inference_delay,
+                prev_chunk_left_over=prev_chunk_left_over,
+                execution_horizon=self.cfg.rtc.execution_horizon,
+            )
+        rtc_tracked_steps = policy_rtc_policy.rtc_processor.get_all_debug_steps()
+        logging.info(f"  Tracked {len(rtc_tracked_steps)} steps with RTC")
+        logging.info(f"  Generated rtc_actions shape: {rtc_actions.shape}")
+
+        # Save num_steps before destroying policy (needed for plotting)
+        try:
+            num_steps = policy_rtc_policy.config.num_steps
+        except Exception as e:
+            logging.error(f"  Error getting num_steps: {e}")
+            num_steps = policy_rtc_policy.config.num_inference_steps
+            logging.warning(f"  Using num_inference_steps: {num_steps} instead of num_steps")
+
+        # Destroy policy_rtc after final use
+        self._destroy_policy(policy_rtc_policy, "policy_rtc")
+
+        # Plot and save results
+        logging.info("=" * 80)
+        logging.info("Plotting results...")
+        self.plot_tracked_data(rtc_tracked_steps, no_rtc_tracked_steps, prev_chunk_left_over, num_steps)
+
+        # Plot final actions comparison
+        logging.info("=" * 80)
+        logging.info("Plotting final actions comparison...")
+        self.plot_final_actions_comparison(rtc_actions, no_rtc_actions, prev_chunk_left_over)
+
+        logging.info("=" * 80)
+        logging.info("Evaluation completed successfully")
+
+    def plot_final_actions_comparison(self, rtc_actions, no_rtc_actions, prev_chunk_left_over):
+        """Plot final action predictions comparison on a single chart.
+
+        Args:
+            rtc_actions: Final actions from RTC policy
+            no_rtc_actions: Final actions from non-RTC policy
+            prev_chunk_left_over: Previous chunk used as ground truth
+        """
+        _check_matplotlib_available()
+
+        # Remove batch dimension if present
+        rtc_actions_plot = rtc_actions.squeeze(0).cpu() if len(rtc_actions.shape) == 3 else rtc_actions.cpu()
+        no_rtc_actions_plot = (
+            no_rtc_actions.squeeze(0).cpu() if len(no_rtc_actions.shape) == 3 else no_rtc_actions.cpu()
+        )
+        prev_chunk_plot = prev_chunk_left_over.cpu()
+
+        # Create figure with 6 subplots (one per action dimension)
+        fig, axes = plt.subplots(6, 1, figsize=(16, 12))
+        fig.suptitle("Final Action Predictions Comparison (Raw)", fontsize=16)
+
+        # Plot each action dimension
+        for dim_idx, ax in enumerate(axes):
+            # Plot previous chunk (ground truth) in red
+            RTCDebugVisualizer.plot_waypoints(
+                [ax],
+                prev_chunk_plot[:, dim_idx : dim_idx + 1],
+                start_from=0,
+                color="red",
+                label="Previous Chunk (Ground Truth)",
+                linewidth=2.5,
+                alpha=0.8,
+            )
+
+            # Plot no-RTC actions in blue
+            RTCDebugVisualizer.plot_waypoints(
+                [ax],
+                no_rtc_actions_plot[:, dim_idx : dim_idx + 1],
+                start_from=0,
+                color="blue",
+                label="No RTC",
+                linewidth=2,
+                alpha=0.7,
+            )
+
+            # Plot RTC actions in green
+            RTCDebugVisualizer.plot_waypoints(
+                [ax],
+                rtc_actions_plot[:, dim_idx : dim_idx + 1],
+                start_from=0,
+                color="green",
+                label="RTC",
+                linewidth=2,
+                alpha=0.7,
+            )
+
+            # Add vertical lines for inference delay and execution horizon
+            inference_delay = self.cfg.inference_delay
+            execution_horizon = self.cfg.rtc.execution_horizon
+
+            if inference_delay > 0:
+                ax.axvline(
+                    x=inference_delay - 1,
+                    color="orange",
+                    linestyle="--",
+                    alpha=0.5,
+                    label=f"Inference Delay ({inference_delay})",
+                )
+
+            if execution_horizon > 0:
+                ax.axvline(
+                    x=execution_horizon,
+                    color="purple",
+                    linestyle="--",
+                    alpha=0.5,
+                    label=f"Execution Horizon ({execution_horizon})",
+                )
+
+            ax.set_ylabel(f"Dim {dim_idx}", fontsize=10)
+            ax.grid(True, alpha=0.3)
+
+            # Set x-axis ticks to show all integer values
+            max_len = max(rtc_actions_plot.shape[0], no_rtc_actions_plot.shape[0], prev_chunk_plot.shape[0])
+            ax.set_xticks(range(0, max_len, max(1, max_len // 20)))  # Show ~20 ticks
+            ax.set_xlim(-0.5, max_len - 0.5)
+
+        axes[-1].set_xlabel("Step", fontsize=10)
+
+        # Collect legend handles and labels from first subplot
+        handles, labels = axes[0].get_legend_handles_labels()
+        # Remove duplicates while preserving order
+        seen = set()
+        unique_handles = []
+        unique_labels = []
+        for handle, label in zip(handles, labels, strict=True):
+            if label not in seen:
+                seen.add(label)
+                unique_handles.append(handle)
+                unique_labels.append(label)
+
+        # Add legend outside the plot area (to the right)
+        fig.legend(
+            unique_handles,
+            unique_labels,
+            loc="center right",
+            fontsize=9,
+            bbox_to_anchor=(1.0, 0.5),
+            framealpha=0.9,
+        )
+
+        # Save figure
+        output_path = os.path.join(self.cfg.output_dir, "final_actions_comparison.png")
+        fig.tight_layout(rect=[0, 0, 0.85, 1])  # Leave space for legend on right
+        fig.savefig(output_path, dpi=150, bbox_inches="tight")
+        logging.info(f"Saved final actions comparison to {output_path}")
+        plt.close(fig)
+
+    def plot_tracked_data(self, rtc_tracked_steps, no_rtc_tracked_steps, prev_chunk_left_over, num_steps):
+        _check_matplotlib_available()
+
+        # Create side-by-side figures for denoising visualization
+        fig_xt, axs_xt = self._create_figure("x_t Denoising: No RTC (left) vs RTC (right)")
+        fig_vt, axs_vt = self._create_figure("v_t Denoising: No RTC (left) vs RTC (right)")
+        fig_corr, axs_corr = self._create_figure("Correction: No RTC (left) vs RTC (right)")
+        fig_x1t, axs_x1t = self._create_figure(
+            "x1_t Predicted State & Error: No RTC (left - empty) vs RTC (right)"
+        )
+        self._plot_denoising_steps_from_tracker(
+            rtc_tracked_steps,
+            axs_xt[:, 1],  # Right column for x_t
+            axs_vt[:, 1],  # Right column for v_t
+            axs_corr[:, 1],  # Right column for correction
+            axs_x1t[:, 1],  # Right column for x1_t
+            num_steps,
+            add_labels=True,  # Add labels for RTC (right column)
+        )
+
+        self._plot_denoising_steps_from_tracker(
+            no_rtc_tracked_steps,
+            axs_xt[:, 0],  # Left column for x_t
+            axs_vt[:, 0],  # Left column for v_t
+            axs_corr[:, 0],  # Left column for correction
+            axs_x1t[:, 0],  # Left column for x1_t
+            num_steps,
+            add_labels=False,  # No labels for No RTC (left column)
+        )
+
+        # Plot no-RTC x_t data on right chart as orange dashed line for comparison
+        self._plot_no_rtc_xt_reference(no_rtc_tracked_steps, axs_xt[:, 1], num_steps)
+
+        # Plot ground truth on x_t axes
+        RTCDebugVisualizer.plot_waypoints(
+            axs_xt[:, 1], prev_chunk_left_over, start_from=0, color="red", label="Ground truth"
+        )
+
+        # Plot ground truth on x1_t axes
+        RTCDebugVisualizer.plot_waypoints(
+            axs_x1t[:, 1], prev_chunk_left_over, start_from=0, color="red", label="Ground truth"
+        )
+
+        # Plot ground truth on x_t axes (no labels for left column)
+        RTCDebugVisualizer.plot_waypoints(
+            axs_xt[:, 0], prev_chunk_left_over, start_from=0, color="red", label=None
+        )
+
+        RTCDebugVisualizer.plot_waypoints(
+            axs_x1t[:, 0], prev_chunk_left_over, start_from=0, color="red", label=None
+        )
+
+        # Add legends outside the plot area for each figure
+        self._add_figure_legend(fig_xt, axs_xt)
+        self._add_figure_legend(fig_vt, axs_vt)
+        self._add_figure_legend(fig_corr, axs_corr)
+        self._add_figure_legend(fig_x1t, axs_x1t)
+
+        # Save denoising plots
+        self._save_figure(fig_xt, os.path.join(self.cfg.output_dir, "denoising_xt_comparison.png"))
+        self._save_figure(fig_vt, os.path.join(self.cfg.output_dir, "denoising_vt_comparison.png"))
+        self._save_figure(fig_corr, os.path.join(self.cfg.output_dir, "denoising_correction_comparison.png"))
+        self._save_figure(fig_x1t, os.path.join(self.cfg.output_dir, "denoising_x1t_comparison.png"))
+
+    def _create_figure(self, title):
+        fig, axs = plt.subplots(6, 2, figsize=(24, 12))
+        fig.suptitle(title, fontsize=16)
+
+        for ax in axs[:, 0]:
+            ax.set_title("No RTC (N/A)" if ax == axs[0, 0] else "", fontsize=12)
+        for ax in axs[:, 1]:
+            ax.set_title("RTC" if ax == axs[0, 1] else "", fontsize=12)
+
+        return fig, axs
+
+    def _add_figure_legend(self, fig, axs):
+        """Add a legend outside the plot area on the right side.
+
+        Args:
+            fig: Matplotlib figure to add legend to
+            axs: Array of axes to collect legend handles from
+        """
+        # Collect all handles and labels from the first row of axes (right column)
+        handles, labels = axs[0, 1].get_legend_handles_labels()
+
+        # Remove duplicates while preserving order
+        seen = set()
+        unique_handles = []
+        unique_labels = []
+        for handle, label in zip(handles, labels, strict=True):
+            if label not in seen:
+                seen.add(label)
+                unique_handles.append(handle)
+                unique_labels.append(label)
+
+        # Add legend outside the plot area (to the right, close to charts)
+        if unique_handles:
+            fig.legend(
+                unique_handles,
+                unique_labels,
+                loc="center left",
+                fontsize=8,
+                bbox_to_anchor=(0.87, 0.5),
+                framealpha=0.9,
+                ncol=1,
+            )
+
+    def _save_figure(self, fig, path):
+        fig.tight_layout(rect=[0, 0, 0.85, 1])  # Leave space for legend/colorbar on right
+        fig.savefig(path, dpi=150, bbox_inches="tight")
+        logging.info(f"Saved figure to {path}")
+        plt.close(fig)
+
+    def _plot_denoising_steps_from_tracker(
+        self, tracked_steps, xt_axs, vt_axs, corr_axs, x1t_axs, num_steps, add_labels=True
+    ):
+        """Plot denoising steps from tracker data.
+
+        Args:
+            tracked_steps: List of DebugStep objects containing debug steps
+            xt_axs: Matplotlib axes for x_t plots (array of 6 axes)
+            vt_axs: Matplotlib axes for v_t plots (array of 6 axes)
+            corr_axs: Matplotlib axes for correction plots (array of 6 axes)
+            x1t_axs: Matplotlib axes for x1_t plots (array of 6 axes)
+            num_steps: Total number of denoising steps for colormap
+            add_labels: Whether to add legend labels for the plots
+        """
+
+        logging.info("=" * 80)
+        logging.info(f"Plotting {len(tracked_steps)} steps")
+
+        debug_steps = tracked_steps
+        if not debug_steps:
+            return
+
+        # Define colors for different denoise steps (using a colormap)
+        colors = plt.cm.viridis(np.linspace(0, 1, num_steps))
+
+        for step_idx, debug_step in enumerate(debug_steps):
+            color = colors[step_idx % len(colors)]
+            label = f"Step {step_idx}" if add_labels else None
+
+            # Plot x_t
+            if debug_step.x_t is not None:
+                RTCDebugVisualizer.plot_waypoints(
+                    xt_axs, debug_step.x_t, start_from=0, color=color, label=label
+                )
+
+            # Plot v_t
+            if debug_step.v_t is not None:
+                RTCDebugVisualizer.plot_waypoints(
+                    vt_axs, debug_step.v_t, start_from=0, color=color, label=label
+                )
+
+            # Plot correction on separate axes
+            if debug_step.correction is not None:
+                RTCDebugVisualizer.plot_waypoints(
+                    corr_axs,
+                    debug_step.correction,
+                    start_from=0,
+                    color=color,
+                    label=label,
+                )
+
+            # Plot x1_t (predicted state)
+            if x1t_axs is not None and debug_step.x1_t is not None:
+                x1t_label = f"x1_t Step {step_idx}" if add_labels else None
+                RTCDebugVisualizer.plot_waypoints(
+                    x1t_axs,
+                    debug_step.x1_t,
+                    start_from=0,
+                    color=color,
+                    label=x1t_label,
+                )
+
+            # Plot error in orange dashed
+            if x1t_axs is not None and debug_step.err is not None:
+                error_chunk = (
+                    debug_step.err[0].cpu().numpy()
+                    if len(debug_step.err.shape) == 3
+                    else debug_step.err.cpu().numpy()
+                )
+
+                num_dims = min(error_chunk.shape[-1], 6)
+                error_label = f"error Step {step_idx}" if add_labels else None
+                for j in range(num_dims):
+                    x1t_axs[j].plot(
+                        np.arange(0, error_chunk.shape[0]),
+                        error_chunk[:, j],
+                        color="orange",
+                        linestyle="--",
+                        alpha=0.7,
+                        label=error_label,
+                    )
+
+        # Recalculate axis limits after plotting to ensure proper scaling
+        self._rescale_axes(xt_axs)
+        self._rescale_axes(vt_axs)
+        self._rescale_axes(corr_axs)
+        self._rescale_axes(x1t_axs)
+
+    def _plot_no_rtc_xt_reference(self, no_rtc_tracked_steps, xt_axs, num_steps):
+        """Plot final no-RTC x_t data as orange dashed line on the RTC chart for comparison.
+
+        Args:
+            no_rtc_tracked_steps: List of DebugStep objects containing no-RTC debug steps
+            xt_axs: Matplotlib axes for x_t plots (array of 6 axes, right column)
+            num_steps: Total number of denoising steps for colormap
+        """
+        debug_steps = no_rtc_tracked_steps
+        if not debug_steps:
+            return
+
+        # Plot only the final x_t step as orange dashed line
+        final_step = debug_steps[-1]
+        logging.info("Plotting final no-RTC x_t step as orange dashed reference")
+
+        if final_step.x_t is not None:
+            x_t_chunk = (
+                final_step.x_t[0].cpu().numpy()
+                if len(final_step.x_t.shape) == 3
+                else final_step.x_t.cpu().numpy()
+            )
+
+            num_dims = min(x_t_chunk.shape[-1], 6)
+            for j in range(num_dims):
+                xt_axs[j].plot(
+                    np.arange(0, x_t_chunk.shape[0]),
+                    x_t_chunk[:, j],
+                    color="orange",
+                    linestyle="--",
+                    alpha=0.7,
+                    linewidth=2,
+                    label="No RTC (final)" if j == 0 else "",
+                )
+
+    def _rescale_axes(self, axes):
+        """Rescale axes to show all data with proper margins.
+
+        Args:
+            axes: Array of matplotlib axes to rescale
+        """
+        for ax in axes:
+            ax.relim()
+            ax.autoscale_view()
+
+            # Add 10% margin to y-axis for better visualization
+            ylim = ax.get_ylim()
+            y_range = ylim[1] - ylim[0]
+            if y_range > 0:  # Avoid division by zero
+                margin = y_range * 0.1
+                ax.set_ylim(ylim[0] - margin, ylim[1] + margin)
+
+            # Set x-axis ticks to show all integer values
+            xlim = ax.get_xlim()
+            max_len = int(xlim[1]) + 1
+            if max_len > 0:
+                ax.set_xticks(range(0, max_len, max(1, max_len // 20)))  # Show ~20 ticks
+                ax.set_xlim(-0.5, max_len - 0.5)
+
+
+@parser.wrap()
+def main(cfg: RTCEvalConfig):
+    """Main entry point for RTC evaluation."""
+    # Set random seed for reproducibility
+    set_seed(cfg.seed)
+
+    init_logging()
+
+    logging.info("=" * 80)
+    logging.info("RTC Dataset Evaluation")
+    logging.info(f"Config: {cfg}")
+    logging.info("=" * 80)
+
+    evaluator = RTCEvaluator(cfg)
+    evaluator.run_evaluation()
+
+
+if __name__ == "__main__":
+    main()

examples/rtc/eval_with_real_robot.py

@@ -0,0 +1,549 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Demo script showing how to use Real-Time Chunking (RTC) with action chunking policies on real robots.
+
+This script demonstrates:
+1. Creating a robot and policy (SmolVLA, Pi0, etc.) with RTC
+2. Consuming actions from the policy while the robot executes
+3. Periodically requesting new action chunks in the background using threads
+4. Managing action buffers and timing for real-time operation
+
+For simulation environments, see eval_with_simulation.py
+
+Usage:
+    # Run RTC with Real robot with RTC
+    uv run examples/rtc/eval_with_real_robot.py \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --policy.device=mps \
+        --rtc.enabled=true \
+        --rtc.execution_horizon=20 \
+        --robot.type=so100_follower \
+        --robot.port=/dev/tty.usbmodem58FA0834591 \
+        --robot.id=so100_follower \
+        --robot.cameras="{ gripper: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30}, front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
+        --task="Move green small object into the purple platform" \
+        --duration=120
+
+    # Run RTC with Real robot without RTC
+    uv run examples/rtc/eval_with_real_robot.py \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --policy.device=mps \
+        --rtc.enabled=false \
+        --robot.type=so100_follower \
+        --robot.port=/dev/tty.usbmodem58FA0834591 \
+        --robot.id=so100_follower \
+        --robot.cameras="{ gripper: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30}, front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
+        --task="Move green small object into the purple platform" \
+        --duration=120
+
+    # Run RTC with Real robot with pi0.5 policy
+    uv run examples/rtc/eval_with_real_robot.py \
+        --policy.path=helper2424/pi05_check_rtc \
+        --policy.device=mps \
+        --rtc.enabled=true \
+        --rtc.execution_horizon=20 \
+        --robot.type=so100_follower \
+        --robot.port=/dev/tty.usbmodem58FA0834591 \
+        --robot.id=so100_follower \
+        --robot.cameras="{ gripper: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}, front: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30}}" \
+        --task="Move green small object into the purple platform" \
+        --duration=120
+"""
+
+import logging
+import math
+import sys
+import time
+import traceback
+from dataclasses import dataclass, field
+from threading import Event, Lock, Thread
+
+import torch
+from torch import Tensor
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig  # noqa: F401
+from lerobot.cameras.realsense.configuration_realsense import RealSenseCameraConfig  # noqa: F401
+from lerobot.configs import parser
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import RTCAttentionSchedule
+from lerobot.datasets.utils import build_dataset_frame, hw_to_dataset_features
+from lerobot.policies.factory import get_policy_class, make_pre_post_processors
+from lerobot.policies.rtc.action_queue import ActionQueue
+from lerobot.policies.rtc.configuration_rtc import RTCConfig
+from lerobot.policies.rtc.latency_tracker import LatencyTracker
+from lerobot.processor.factory import (
+    make_default_robot_action_processor,
+    make_default_robot_observation_processor,
+)
+from lerobot.rl.process import ProcessSignalHandler
+from lerobot.robots import (  # noqa: F401
+    Robot,
+    RobotConfig,
+    koch_follower,
+    so100_follower,
+    so101_follower,
+)
+from lerobot.robots.utils import make_robot_from_config
+from lerobot.utils.constants import OBS_IMAGES
+from lerobot.utils.hub import HubMixin
+from lerobot.utils.utils import init_logging
+
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
+
+
+class RobotWrapper:
+    def __init__(self, robot: Robot):
+        self.robot = robot
+        self.lock = Lock()
+
+    def get_observation(self) -> dict[str, Tensor]:
+        with self.lock:
+            return self.robot.get_observation()
+
+    def send_action(self, action: Tensor):
+        with self.lock:
+            self.robot.send_action(action)
+
+    def observation_features(self) -> list[str]:
+        with self.lock:
+            return self.robot.observation_features
+
+    def action_features(self) -> list[str]:
+        with self.lock:
+            return self.robot.action_features
+
+
+@dataclass
+class RTCDemoConfig(HubMixin):
+    """Configuration for RTC demo with action chunking policies and real robots."""
+
+    # Policy configuration
+    policy: PreTrainedConfig | None = None
+
+    # Robot configuration
+    robot: RobotConfig | None = None
+
+    # RTC configuration
+    rtc: RTCConfig = field(
+        default_factory=lambda: RTCConfig(
+            execution_horizon=10,
+            max_guidance_weight=1.0,
+            prefix_attention_schedule=RTCAttentionSchedule.EXP,
+        )
+    )
+
+    # Demo parameters
+    duration: float = 30.0  # Duration to run the demo (seconds)
+    fps: float = 10.0  # Action execution frequency (Hz)
+
+    # Compute device
+    device: str | None = None  # Device to run on (cuda, cpu, auto)
+
+    # Get new actions horizon. The amount of executed steps after which will be requested new actions.
+    # It should be higher than inference delay + execution horizon.
+    action_queue_size_to_get_new_actions: int = 30
+
+    # Task to execute
+    task: str = field(default="", metadata={"help": "Task to execute"})
+
+    # Torch compile configuration
+    use_torch_compile: bool = field(
+        default=False,
+        metadata={"help": "Use torch.compile for faster inference (PyTorch 2.0+)"},
+    )
+
+    torch_compile_backend: str = field(
+        default="inductor",
+        metadata={"help": "Backend for torch.compile (inductor, aot_eager, cudagraphs)"},
+    )
+
+    torch_compile_mode: str = field(
+        default="default",
+        metadata={"help": "Compilation mode (default, reduce-overhead, max-autotune)"},
+    )
+
+    torch_compile_disable_cudagraphs: bool = field(
+        default=True,
+        metadata={
+            "help": "Disable CUDA graphs in torch.compile. Required due to in-place tensor "
+            "operations in denoising loop (x_t += dt * v_t) which cause tensor aliasing issues."
+        },
+    )
+
+    def __post_init__(self):
+        # HACK: We parse again the cli args here to get the pretrained path if there was one.
+        policy_path = parser.get_path_arg("policy")
+        if policy_path:
+            cli_overrides = parser.get_cli_overrides("policy")
+            self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
+            self.policy.pretrained_path = policy_path
+        else:
+            raise ValueError("Policy path is required")
+
+        # Validate that robot configuration is provided
+        if self.robot is None:
+            raise ValueError("Robot configuration must be provided")
+
+    @classmethod
+    def __get_path_fields__(cls) -> list[str]:
+        """This enables the parser to load config from the policy using `--policy.path=local/dir`"""
+        return ["policy"]
+
+
+def is_image_key(k: str) -> bool:
+    return k.startswith(OBS_IMAGES)
+
+
+def get_actions(
+    policy,
+    robot: RobotWrapper,
+    robot_observation_processor,
+    action_queue: ActionQueue,
+    shutdown_event: Event,
+    cfg: RTCDemoConfig,
+):
+    """Thread function to request action chunks from the policy.
+
+    Args:
+        policy: The policy instance (SmolVLA, Pi0, etc.)
+        robot: The robot instance for getting observations
+        robot_observation_processor: Processor for raw robot observations
+        action_queue: Queue to put new action chunks
+        shutdown_event: Event to signal shutdown
+        cfg: Demo configuration
+    """
+    try:
+        logger.info("[GET_ACTIONS] Starting get actions thread")
+
+        latency_tracker = LatencyTracker()  # Track latency of action chunks
+        fps = cfg.fps
+        time_per_chunk = 1.0 / fps
+
+        dataset_features = hw_to_dataset_features(robot.observation_features(), "observation")
+        policy_device = policy.config.device
+
+        # Load preprocessor and postprocessor from pretrained files
+        # The stats are embedded in the processor .safetensors files
+        logger.info(f"[GET_ACTIONS] Loading preprocessor/postprocessor from {cfg.policy.pretrained_path}")
+
+        preprocessor, postprocessor = make_pre_post_processors(
+            policy_cfg=cfg.policy,
+            pretrained_path=cfg.policy.pretrained_path,
+            dataset_stats=None,  # Will load from pretrained processor files
+            preprocessor_overrides={
+                "device_processor": {"device": cfg.policy.device},
+            },
+        )
+
+        logger.info("[GET_ACTIONS] Preprocessor/postprocessor loaded successfully with embedded stats")
+
+        get_actions_threshold = cfg.action_queue_size_to_get_new_actions
+
+        if not cfg.rtc.enabled:
+            get_actions_threshold = 0
+
+        while not shutdown_event.is_set():
+            if action_queue.qsize() <= get_actions_threshold:
+                current_time = time.perf_counter()
+                action_index_before_inference = action_queue.get_action_index()
+                prev_actions = action_queue.get_left_over()
+
+                inference_latency = latency_tracker.max()
+                inference_delay = math.ceil(inference_latency / time_per_chunk)
+
+                obs = robot.get_observation()
+
+                # Apply robot observation processor
+                obs_processed = robot_observation_processor(obs)
+
+                obs_with_policy_features = build_dataset_frame(
+                    dataset_features, obs_processed, prefix="observation"
+                )
+
+                for name in obs_with_policy_features:
+                    obs_with_policy_features[name] = torch.from_numpy(obs_with_policy_features[name])
+                    if "image" in name:
+                        obs_with_policy_features[name] = (
+                            obs_with_policy_features[name].type(torch.float32) / 255
+                        )
+                        obs_with_policy_features[name] = (
+                            obs_with_policy_features[name].permute(2, 0, 1).contiguous()
+                        )
+                    obs_with_policy_features[name] = obs_with_policy_features[name].unsqueeze(0)
+                    obs_with_policy_features[name] = obs_with_policy_features[name].to(policy_device)
+
+                obs_with_policy_features["task"] = [cfg.task]  # Task should be a list, not a string!
+                obs_with_policy_features["robot_type"] = (
+                    robot.robot.name if hasattr(robot.robot, "name") else ""
+                )
+
+                preproceseded_obs = preprocessor(obs_with_policy_features)
+
+                # Generate actions WITH RTC
+                actions = policy.predict_action_chunk(
+                    preproceseded_obs,
+                    inference_delay=inference_delay,
+                    prev_chunk_left_over=prev_actions,
+                )
+
+                # Store original actions (before postprocessing) for RTC
+                original_actions = actions.squeeze(0).clone()
+
+                postprocessed_actions = postprocessor(actions)
+
+                postprocessed_actions = postprocessed_actions.squeeze(0)
+
+                new_latency = time.perf_counter() - current_time
+                new_delay = math.ceil(new_latency / time_per_chunk)
+                latency_tracker.add(new_latency)
+
+                if cfg.action_queue_size_to_get_new_actions < cfg.rtc.execution_horizon + new_delay:
+                    logger.warning(
+                        "[GET_ACTIONS] cfg.action_queue_size_to_get_new_actions Too small, It should be higher than inference delay + execution horizon."
+                    )
+
+                action_queue.merge(
+                    original_actions, postprocessed_actions, new_delay, action_index_before_inference
+                )
+            else:
+                # Small sleep to prevent busy waiting
+                time.sleep(0.1)
+
+        logger.info("[GET_ACTIONS] get actions thread shutting down")
+    except Exception as e:
+        logger.error(f"[GET_ACTIONS] Fatal exception in get_actions thread: {e}")
+        logger.error(traceback.format_exc())
+        sys.exit(1)
+
+
+def actor_control(
+    robot: RobotWrapper,
+    robot_action_processor,
+    action_queue: ActionQueue,
+    shutdown_event: Event,
+    cfg: RTCDemoConfig,
+):
+    """Thread function to execute actions on the robot.
+
+    Args:
+        robot: The robot instance
+        action_queue: Queue to get actions from
+        shutdown_event: Event to signal shutdown
+        cfg: Demo configuration
+    """
+    try:
+        logger.info("[ACTOR] Starting actor thread")
+
+        action_count = 0
+        action_interval = 1.0 / cfg.fps
+
+        while not shutdown_event.is_set():
+            start_time = time.perf_counter()
+
+            # Try to get an action from the queue with timeout
+            action = action_queue.get()
+
+            if action is not None:
+                action = action.cpu()
+                action_dict = {key: action[i].item() for i, key in enumerate(robot.action_features())}
+                action_processed = robot_action_processor((action_dict, None))
+                robot.send_action(action_processed)
+
+                action_count += 1
+
+            dt_s = time.perf_counter() - start_time
+            time.sleep(max(0, (action_interval - dt_s) - 0.001))
+
+        logger.info(f"[ACTOR] Actor thread shutting down. Total actions executed: {action_count}")
+    except Exception as e:
+        logger.error(f"[ACTOR] Fatal exception in actor_control thread: {e}")
+        logger.error(traceback.format_exc())
+        sys.exit(1)
+
+
+def _apply_torch_compile(policy, cfg: RTCDemoConfig):
+    """Apply torch.compile to the policy's predict_action_chunk method.
+
+    Args:
+        policy: Policy instance to compile
+        cfg: Configuration containing torch compile settings
+
+    Returns:
+        Policy with compiled predict_action_chunk method
+    """
+
+    # PI models handle their own compilation
+    if policy.type == "pi05" or policy.type == "pi0":
+        return policy
+
+    try:
+        # Check if torch.compile is available (PyTorch 2.0+)
+        if not hasattr(torch, "compile"):
+            logger.warning(
+                f"torch.compile is not available. Requires PyTorch 2.0+. "
+                f"Current version: {torch.__version__}. Skipping compilation."
+            )
+            return policy
+
+        logger.info("Applying torch.compile to predict_action_chunk...")
+        logger.info(f"  Backend: {cfg.torch_compile_backend}")
+        logger.info(f"  Mode: {cfg.torch_compile_mode}")
+        logger.info(f"  Disable CUDA graphs: {cfg.torch_compile_disable_cudagraphs}")
+
+        # Compile the predict_action_chunk method
+        # - CUDA graphs disabled to prevent tensor aliasing from in-place ops (x_t += dt * v_t)
+        compile_kwargs = {
+            "backend": cfg.torch_compile_backend,
+            "mode": cfg.torch_compile_mode,
+        }
+
+        # Disable CUDA graphs if requested (prevents tensor aliasing issues)
+        if cfg.torch_compile_disable_cudagraphs:
+            compile_kwargs["options"] = {"triton.cudagraphs": False}
+
+        original_method = policy.predict_action_chunk
+        compiled_method = torch.compile(original_method, **compile_kwargs)
+        policy.predict_action_chunk = compiled_method
+        logger.info("✓ Successfully compiled predict_action_chunk")
+
+    except Exception as e:
+        logger.error(f"Failed to apply torch.compile: {e}")
+        logger.warning("Continuing without torch.compile")
+
+    return policy
+
+
+@parser.wrap()
+def demo_cli(cfg: RTCDemoConfig):
+    """Main entry point for RTC demo with draccus configuration."""
+
+    # Initialize logging
+    init_logging()
+
+    logger.info(f"Using device: {cfg.device}")
+
+    # Setup signal handler for graceful shutdown
+    signal_handler = ProcessSignalHandler(use_threads=True, display_pid=False)
+    shutdown_event = signal_handler.shutdown_event
+
+    policy = None
+    robot = None
+    get_actions_thread = None
+    actor_thread = None
+
+    policy_class = get_policy_class(cfg.policy.type)
+
+    # Load config and set compile_model for pi0/pi05 models
+    config = PreTrainedConfig.from_pretrained(cfg.policy.pretrained_path)
+
+    if cfg.policy.type == "pi05" or cfg.policy.type == "pi0":
+        config.compile_model = cfg.use_torch_compile
+
+    policy = policy_class.from_pretrained(cfg.policy.pretrained_path, config=config)
+
+    # Turn on RTC
+    policy.config.rtc_config = cfg.rtc
+
+    # Init RTC processort, as by default if RTC disabled in the config
+    # The processor won't be created
+    policy.init_rtc_processor()
+
+    assert policy.name in ["smolvla", "pi05", "pi0"], "Only smolvla, pi05, and pi0 are supported for RTC"
+
+    policy = policy.to(cfg.device)
+    policy.eval()
+
+    # Apply torch.compile to predict_action_chunk method if enabled
+    if cfg.use_torch_compile:
+        policy = _apply_torch_compile(policy, cfg)
+
+    # Create robot
+    logger.info(f"Initializing robot: {cfg.robot.type}")
+    robot = make_robot_from_config(cfg.robot)
+    robot.connect()
+    robot_wrapper = RobotWrapper(robot)
+
+    # Create robot observation processor
+    robot_observation_processor = make_default_robot_observation_processor()
+    robot_action_processor = make_default_robot_action_processor()
+
+    # Create action queue for communication between threads
+    action_queue = ActionQueue(cfg.rtc)
+
+    # Start chunk requester thread
+    get_actions_thread = Thread(
+        target=get_actions,
+        args=(policy, robot_wrapper, robot_observation_processor, action_queue, shutdown_event, cfg),
+        daemon=True,
+        name="GetActions",
+    )
+    get_actions_thread.start()
+    logger.info("Started get actions thread")
+
+    # Start action executor thread
+    actor_thread = Thread(
+        target=actor_control,
+        args=(robot_wrapper, robot_action_processor, action_queue, shutdown_event, cfg),
+        daemon=True,
+        name="Actor",
+    )
+    actor_thread.start()
+    logger.info("Started actor thread")
+
+    logger.info("Started stop by duration thread")
+
+    # Main thread monitors for duration or shutdown
+    logger.info(f"Running demo for {cfg.duration} seconds...")
+    start_time = time.time()
+
+    while not shutdown_event.is_set() and (time.time() - start_time) < cfg.duration:
+        time.sleep(10)
+
+        # Log queue status periodically
+        if int(time.time() - start_time) % 5 == 0:
+            logger.info(f"[MAIN] Action queue size: {action_queue.qsize()}")
+
+        if time.time() - start_time > cfg.duration:
+            break
+
+    logger.info("Demo duration reached or shutdown requested")
+
+    # Signal shutdown
+    shutdown_event.set()
+
+    # Wait for threads to finish
+    if get_actions_thread and get_actions_thread.is_alive():
+        logger.info("Waiting for chunk requester thread to finish...")
+        get_actions_thread.join()
+
+    if actor_thread and actor_thread.is_alive():
+        logger.info("Waiting for action executor thread to finish...")
+        actor_thread.join()
+
+    # Cleanup robot
+    if robot:
+        robot.disconnect()
+        logger.info("Robot disconnected")
+
+    logger.info("Cleanup completed")
+
+
+if __name__ == "__main__":
+    demo_cli()
+    logging.info("RTC demo finished")

examples/so100_to_so100_EE/evaluate.py

@@ -195,4 +195,6 @@ for episode_idx in range(NUM_EPISODES):
 log_say("Stop recording")
 robot.disconnect()
 listener.stop()
+
+dataset.finalize()
 dataset.push_to_hub()

examples/so100_to_so100_EE/record.py

@@ -199,4 +199,6 @@ log_say("Stop recording")
 leader.disconnect()
 follower.disconnect()
 listener.stop()
+
+dataset.finalize()
 dataset.push_to_hub()

examples/tutorial/act/act_training_example.py

@@ -0,0 +1,98 @@
+"""This script demonstrates how to train ACT Policy on a real-world dataset."""
+
+from pathlib import Path
+
+import torch
+
+from lerobot.configs.types import FeatureType
+from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.datasets.utils import dataset_to_policy_features
+from lerobot.policies.act.configuration_act import ACTConfig
+from lerobot.policies.act.modeling_act import ACTPolicy
+from lerobot.policies.factory import make_pre_post_processors
+
+
+def make_delta_timestamps(delta_indices: list[int] | None, fps: int) -> list[float]:
+    if delta_indices is None:
+        return [0]
+
+    return [i / fps for i in delta_indices]
+
+
+output_directory = Path("outputs/robot_learning_tutorial/act")
+output_directory.mkdir(parents=True, exist_ok=True)
+
+# Select your device
+device = torch.device("mps")  # or "cuda" or "cpu"
+
+dataset_id = "lerobot/svla_so101_pickplace"
+
+# This specifies the inputs the model will be expecting and the outputs it will produce
+dataset_metadata = LeRobotDatasetMetadata(dataset_id)
+features = dataset_to_policy_features(dataset_metadata.features)
+
+output_features = {key: ft for key, ft in features.items() if ft.type is FeatureType.ACTION}
+input_features = {key: ft for key, ft in features.items() if key not in output_features}
+
+cfg = ACTConfig(input_features=input_features, output_features=output_features)
+policy = ACTPolicy(cfg)
+preprocessor, postprocessor = make_pre_post_processors(cfg, dataset_stats=dataset_metadata.stats)
+
+policy.train()
+policy.to(device)
+
+# To perform action chunking, ACT expects a given number of actions as targets
+delta_timestamps = {
+    "action": make_delta_timestamps(cfg.action_delta_indices, dataset_metadata.fps),
+}
+
+# add image features if they are present
+delta_timestamps |= {
+    k: make_delta_timestamps(cfg.observation_delta_indices, dataset_metadata.fps) for k in cfg.image_features
+}
+
+# Instantiate the dataset
+dataset = LeRobotDataset(dataset_id, delta_timestamps=delta_timestamps)
+
+# Create the optimizer and dataloader for offline training
+optimizer = cfg.get_optimizer_preset().build(policy.parameters())
+batch_size = 32
+dataloader = torch.utils.data.DataLoader(
+    dataset,
+    batch_size=batch_size,
+    shuffle=True,
+    pin_memory=device.type != "cpu",
+    drop_last=True,
+)
+
+# Number of training steps and logging frequency
+training_steps = 1
+log_freq = 1
+
+# Run training loop
+step = 0
+done = False
+while not done:
+    for batch in dataloader:
+        batch = preprocessor(batch)
+        loss, _ = policy.forward(batch)
+        loss.backward()
+        optimizer.step()
+        optimizer.zero_grad()
+
+        if step % log_freq == 0:
+            print(f"step: {step} loss: {loss.item():.3f}")
+        step += 1
+        if step >= training_steps:
+            done = True
+            break
+
+# Save the policy checkpoint, alongside the pre/post processors
+policy.save_pretrained(output_directory)
+preprocessor.save_pretrained(output_directory)
+postprocessor.save_pretrained(output_directory)
+
+# Save all assets to the Hub
+policy.push_to_hub("fracapuano/robot_learning_tutorial_act")
+preprocessor.push_to_hub("fracapuano/robot_learning_tutorial_act")
+postprocessor.push_to_hub("fracapuano/robot_learning_tutorial_act")

examples/tutorial/act/act_using_example.py

@@ -0,0 +1,57 @@
+import torch
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
+from lerobot.policies.act.modeling_act import ACTPolicy
+from lerobot.policies.factory import make_pre_post_processors
+from lerobot.policies.utils import build_inference_frame, make_robot_action
+from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
+from lerobot.robots.so100_follower.so100_follower import SO100Follower
+
+device = torch.device("mps")  # or "cuda" or "cpu"
+model_id = "fracapuano/robot_learning_tutorial_act"
+model = ACTPolicy.from_pretrained(model_id)
+
+dataset_id = "lerobot/svla_so101_pickplace"
+# This only downloads the metadata for the dataset, ~10s of MB even for large-scale datasets
+dataset_metadata = LeRobotDatasetMetadata(dataset_id)
+preprocess, postprocess = make_pre_post_processors(model.config, dataset_stats=dataset_metadata.stats)
+
+# # find ports using lerobot-find-port
+follower_port = ...  # something like "/dev/tty.usbmodem58760431631"
+
+# # the robot ids are used the load the right calibration files
+follower_id = ...  # something like "follower_so100"
+
+MAX_EPISODES = 5
+MAX_STEPS_PER_EPISODE = 20
+
+# Robot and environment configuration
+# Camera keys must match the name and resolutions of the ones used for training!
+# You can check the camera keys expected by a model in the info.json card on the model card on the Hub
+camera_config = {
+    "side": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
+    "up": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30),
+}
+
+robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_config)
+robot = SO100Follower(robot_cfg)
+robot.connect()
+
+for _ in range(MAX_EPISODES):
+    for _ in range(MAX_STEPS_PER_EPISODE):
+        obs = robot.get_observation()
+        obs_frame = build_inference_frame(
+            observation=obs, ds_features=dataset_metadata.features, device=device
+        )
+
+        obs = preprocess(obs_frame)
+
+        action = model.select_action(obs)
+        action = postprocess(action)
+
+        action = make_robot_action(action, dataset_metadata.features)
+
+        robot.send_action(action)
+
+    print("Episode finished! Starting new episode...")

examples/tutorial/async-inf/policy_server.py

@@ -0,0 +1,11 @@
+from lerobot.async_inference.configs import PolicyServerConfig
+from lerobot.async_inference.policy_server import serve
+
+host = ...  # something like "127.0.0.1" if you're exposing to localhost
+port = ...  # something like 8080
+
+config = PolicyServerConfig(
+    host=host,
+    port=port,
+)
+serve(config)

examples/tutorial/async-inf/robot_client.py

@@ -0,0 +1,55 @@
+import threading
+
+from lerobot.async_inference.configs import RobotClientConfig
+from lerobot.async_inference.helpers import visualize_action_queue_size
+from lerobot.async_inference.robot_client import RobotClient
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.robots.so100_follower import SO100FollowerConfig
+
+# these cameras must match the ones expected by the policy - find your cameras with lerobot-find-cameras
+# check the config.json on the Hub for the policy you are using to see the expected camera specs
+camera_cfg = {
+    "up": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
+    "side": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30),
+}
+
+# # find ports using lerobot-find-port
+follower_port = ...  # something like "/dev/tty.usbmodem58760431631"
+
+# # the robot ids are used the load the right calibration files
+follower_id = ...  # something like "follower_so100"
+
+robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_cfg)
+
+server_address = ...  # something like "127.0.0.1:8080" if using localhost
+
+# 3. Create client configuration
+client_cfg = RobotClientConfig(
+    robot=robot_cfg,
+    server_address=server_address,
+    policy_device="mps",
+    policy_type="act",
+    pretrained_name_or_path="fracapuano/robot_learning_tutorial_act",
+    chunk_size_threshold=0.5,  # g
+    actions_per_chunk=50,  # make sure this is less than the max actions of the policy
+)
+
+# 4. Create and start client
+client = RobotClient(client_cfg)
+
+# 5. Provide a textual description of the task
+task = ...
+
+if client.start():
+    # Start action receiver thread
+    action_receiver_thread = threading.Thread(target=client.receive_actions, daemon=True)
+    action_receiver_thread.start()
+
+    try:
+        # Run the control loop
+        client.control_loop(task)
+    except KeyboardInterrupt:
+        client.stop()
+        action_receiver_thread.join()
+        # (Optionally) plot the action queue size
+        visualize_action_queue_size(client.action_queue_size)

examples/tutorial/diffusion/diffusion_training_example.py

@@ -0,0 +1,99 @@
+"""This script demonstrates how to train Diffusion Policy on a real-world dataset."""
+
+from pathlib import Path
+
+import torch
+
+from lerobot.configs.types import FeatureType
+from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.datasets.utils import dataset_to_policy_features
+from lerobot.policies.diffusion.configuration_diffusion import DiffusionConfig
+from lerobot.policies.diffusion.modeling_diffusion import DiffusionPolicy
+from lerobot.policies.factory import make_pre_post_processors
+
+
+def make_delta_timestamps(delta_indices: list[int] | None, fps: int) -> list[float]:
+    if delta_indices is None:
+        return [0]
+
+    return [i / fps for i in delta_indices]
+
+
+output_directory = Path("outputs/robot_learning_tutorial/diffusion")
+output_directory.mkdir(parents=True, exist_ok=True)
+
+# Select your device
+device = torch.device("mps")  # or "cuda" or "cpu"
+
+dataset_id = "lerobot/svla_so101_pickplace"
+
+# This specifies the inputs the model will be expecting and the outputs it will produce
+dataset_metadata = LeRobotDatasetMetadata(dataset_id)
+features = dataset_to_policy_features(dataset_metadata.features)
+
+output_features = {key: ft for key, ft in features.items() if ft.type is FeatureType.ACTION}
+input_features = {key: ft for key, ft in features.items() if key not in output_features}
+
+cfg = DiffusionConfig(input_features=input_features, output_features=output_features)
+policy = DiffusionPolicy(cfg)
+preprocessor, postprocessor = make_pre_post_processors(cfg, dataset_stats=dataset_metadata.stats)
+
+policy.train()
+policy.to(device)
+
+# To perform action chunking, ACT expects a given number of actions as targets
+delta_timestamps = {
+    "observation.state": make_delta_timestamps(cfg.observation_delta_indices, dataset_metadata.fps),
+    "action": make_delta_timestamps(cfg.action_delta_indices, dataset_metadata.fps),
+}
+
+# add image features if they are present
+delta_timestamps |= {
+    k: make_delta_timestamps(cfg.observation_delta_indices, dataset_metadata.fps) for k in cfg.image_features
+}
+
+# Instantiate the dataset
+dataset = LeRobotDataset(dataset_id, delta_timestamps=delta_timestamps)
+
+# Create the optimizer and dataloader for offline training
+optimizer = cfg.get_optimizer_preset().build(policy.parameters())
+batch_size = 32
+dataloader = torch.utils.data.DataLoader(
+    dataset,
+    batch_size=batch_size,
+    shuffle=True,
+    pin_memory=device.type != "cpu",
+    drop_last=True,
+)
+
+# Number of training steps and logging frequency
+training_steps = 1
+log_freq = 1
+
+# Run training loop
+step = 0
+done = False
+while not done:
+    for batch in dataloader:
+        batch = preprocessor(batch)
+        loss, _ = policy.forward(batch)
+        loss.backward()
+        optimizer.step()
+        optimizer.zero_grad()
+
+        if step % log_freq == 0:
+            print(f"step: {step} loss: {loss.item():.3f}")
+        step += 1
+        if step >= training_steps:
+            done = True
+            break
+
+# Save the policy checkpoint, alongside the pre/post processors
+policy.save_pretrained(output_directory)
+preprocessor.save_pretrained(output_directory)
+postprocessor.save_pretrained(output_directory)
+
+# Save all assets to the Hub
+policy.push_to_hub("fracapuano/robot_learning_tutorial_diffusion")
+preprocessor.push_to_hub("fracapuano/robot_learning_tutorial_diffusion")
+postprocessor.push_to_hub("fracapuano/robot_learning_tutorial_diffusion")

examples/tutorial/diffusion/diffusion_using_example.py

@@ -0,0 +1,60 @@
+import torch
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
+from lerobot.policies.diffusion.modeling_diffusion import DiffusionPolicy
+from lerobot.policies.factory import make_pre_post_processors
+from lerobot.policies.utils import build_inference_frame, make_robot_action
+from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
+from lerobot.robots.so100_follower.so100_follower import SO100Follower
+
+device = torch.device("mps")  # or "cuda" or "cpu"
+model_id = "fracapuano/robot_learning_tutorial_diffusion"
+
+model = DiffusionPolicy.from_pretrained(model_id)
+
+dataset_id = "lerobot/svla_so101_pickplace"
+# This only downloads the metadata for the dataset, ~10s of MB even for large-scale datasets
+dataset_metadata = LeRobotDatasetMetadata(dataset_id)
+preprocess, postprocess = make_pre_post_processors(
+    model.config, model_id, dataset_stats=dataset_metadata.stats
+)
+
+MAX_EPISODES = 5
+MAX_STEPS_PER_EPISODE = 20
+
+
+# # find ports using lerobot-find-port
+follower_port = ...  # something like "/dev/tty.usbmodem58760431631"
+
+# # the robot ids are used the load the right calibration files
+follower_id = ...  # something like "follower_so100"
+
+# Robot and environment configuration
+# Camera keys must match the name and resolutions of the ones used for training!
+# You can check the camera keys expected by a model in the info.json card on the model card on the Hub
+camera_config = {
+    "side": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
+    "up": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30),
+}
+
+robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_config)
+robot = SO100Follower(robot_cfg)
+robot.connect()
+
+
+for _ in range(MAX_EPISODES):
+    for _ in range(MAX_STEPS_PER_EPISODE):
+        obs = robot.get_observation()
+        obs_frame = build_inference_frame(
+            observation=obs, ds_features=dataset_metadata.features, device=device
+        )
+
+        obs = preprocess(obs_frame)
+
+        action = model.select_action(obs)
+        action = postprocess(action)
+        action = make_robot_action(action, dataset_metadata.features)
+        robot.send_action(action)
+
+    print("Episode finished! Starting new episode...")

examples/tutorial/pi0/using_pi0_example.py

@@ -0,0 +1,67 @@
+import torch
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.datasets.utils import hw_to_dataset_features
+from lerobot.policies.factory import make_pre_post_processors
+from lerobot.policies.pi0.modeling_pi0 import PI0Policy
+from lerobot.policies.utils import build_inference_frame, make_robot_action
+from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
+from lerobot.robots.so100_follower.so100_follower import SO100Follower
+
+MAX_EPISODES = 5
+MAX_STEPS_PER_EPISODE = 20
+
+device = torch.device("mps")  # or "cuda" or "cpu"
+model_id = "lerobot/pi0_base"
+
+model = PI0Policy.from_pretrained(model_id)
+
+preprocess, postprocess = make_pre_post_processors(
+    model.config,
+    model_id,
+    # This overrides allows to run on MPS, otherwise defaults to CUDA (if available)
+    preprocessor_overrides={"device_processor": {"device": str(device)}},
+)
+
+# find ports using lerobot-find-port
+follower_port = ...  # something like "/dev/tty.usbmodem58760431631"
+
+# the robot ids are used the load the right calibration files
+follower_id = ...  # something like "follower_so100"
+
+# Robot and environment configuration
+# Camera keys must match the name and resolutions of the ones used for training!
+# You can check the camera keys expected by a model in the info.json card on the model card on the Hub
+camera_config = {
+    "base_0_rgb": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
+    "left_wrist_0_rgb": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30),
+    "right_wrist_0_rgb": OpenCVCameraConfig(index_or_path=2, width=640, height=480, fps=30),
+}
+
+robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_config)
+robot = SO100Follower(robot_cfg)
+robot.connect()
+
+task = ""  # something like "pick the red block"
+robot_type = ""  # something like "so100_follower" for multi-embodiment datasets
+
+# This is used to match the raw observation keys to the keys expected by the policy
+action_features = hw_to_dataset_features(robot.action_features, "action")
+obs_features = hw_to_dataset_features(robot.observation_features, "observation")
+dataset_features = {**action_features, **obs_features}
+
+for _ in range(MAX_EPISODES):
+    for _ in range(MAX_STEPS_PER_EPISODE):
+        obs = robot.get_observation()
+        obs_frame = build_inference_frame(
+            observation=obs, ds_features=dataset_features, device=device, task=task, robot_type=robot_type
+        )
+
+        obs = preprocess(obs_frame)
+
+        action = model.select_action(obs)
+        action = postprocess(action)
+        action = make_robot_action(action, dataset_features)
+        robot.send_action(action)
+
+    print("Episode finished! Starting new episode...")

examples/tutorial/rl/hilserl_example.py

@@ -0,0 +1,345 @@
+import multiprocessing as mp
+import signal
+from pathlib import Path
+from queue import Empty, Full
+
+import torch
+import torch.optim as optim
+
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.utils import hw_to_dataset_features
+from lerobot.envs.configs import HILSerlProcessorConfig, HILSerlRobotEnvConfig
+from lerobot.policies.sac.configuration_sac import SACConfig
+from lerobot.policies.sac.modeling_sac import SACPolicy
+from lerobot.policies.sac.reward_model.modeling_classifier import Classifier
+from lerobot.rl.buffer import ReplayBuffer
+from lerobot.rl.gym_manipulator import make_robot_env
+from lerobot.robots.so100_follower import SO100FollowerConfig
+from lerobot.teleoperators.so100_leader import SO100LeaderConfig
+from lerobot.teleoperators.utils import TeleopEvents
+
+LOG_EVERY = 10
+SEND_EVERY = 10
+
+
+def run_learner(
+    transitions_queue: mp.Queue,
+    parameters_queue: mp.Queue,
+    shutdown_event: mp.Event,
+    policy_learner: SACPolicy,
+    online_buffer: ReplayBuffer,
+    offline_buffer: ReplayBuffer,
+    lr: float = 3e-4,
+    batch_size: int = 32,
+    device: torch.device = "mps",
+):
+    """The learner process - trains SAC policy on transitions streamed from the actor, updating parameters
+    for the actor to adopt."""
+    policy_learner.train()
+    policy_learner.to(device)
+
+    # Create Adam optimizer from scratch - simple and clean
+    optimizer = optim.Adam(policy_learner.parameters(), lr=lr)
+
+    print(f"[LEARNER] Online buffer capacity: {online_buffer.capacity}")
+    print(f"[LEARNER] Offline buffer capacity: {offline_buffer.capacity}")
+
+    training_step = 0
+
+    while not shutdown_event.is_set():
+        # retrieve incoming transitions from the actor process
+        try:
+            transitions = transitions_queue.get(timeout=0.1)
+            for transition in transitions:
+                # HIL-SERL: Add ALL transitions to online buffer
+                online_buffer.add(**transition)
+
+                # HIL-SERL: Add ONLY human intervention transitions to offline buffer
+                is_intervention = transition.get("complementary_info", {}).get("is_intervention", False)
+                if is_intervention:
+                    offline_buffer.add(**transition)
+                    print(
+                        f"[LEARNER] Human intervention detected! Added to offline buffer (now {len(offline_buffer)} transitions)"
+                    )
+
+        except Empty:
+            pass  # No transitions available, continue
+
+        # Train if we have enough data
+        if len(online_buffer) >= policy_learner.config.online_step_before_learning:
+            # Sample from online buffer (autonomous + human data)
+            online_batch = online_buffer.sample(batch_size // 2)
+
+            # Sample from offline buffer (human demonstrations only, either precollected or at runtime)
+            offline_batch = offline_buffer.sample(batch_size // 2)
+
+            # Combine batches - this is the key HIL-SERL mechanism!
+            batch = {}
+            for key in online_batch:
+                if key in offline_batch:
+                    batch[key] = torch.cat([online_batch[key], offline_batch[key]], dim=0)
+                else:
+                    batch[key] = online_batch[key]
+
+            loss, _ = policy_learner.forward(batch)
+
+            optimizer.zero_grad()
+            loss.backward()
+            optimizer.step()
+            training_step += 1
+
+            if training_step % LOG_EVERY == 0:
+                print(
+                    f"[LEARNER] Training step {training_step}, Loss: {loss.item():.4f}, "
+                    f"Buffers: Online={len(online_buffer)}, Offline={len(offline_buffer)}"
+                )
+
+            # Send updated parameters to actor every 10 training steps
+            if training_step % SEND_EVERY == 0:
+                try:
+                    state_dict = {k: v.cpu() for k, v in policy_learner.state_dict().items()}
+                    parameters_queue.put_nowait(state_dict)
+                    print("[LEARNER] Sent updated parameters to actor")
+                except Full:
+                    # Missing write due to queue not being consumed (should happen rarely)
+                    pass
+
+    print("[LEARNER] Learner process finished")
+
+
+def run_actor(
+    transitions_queue: mp.Queue,
+    parameters_queue: mp.Queue,
+    shutdown_event: mp.Event,
+    policy_actor: SACPolicy,
+    reward_classifier: Classifier,
+    env_cfg: HILSerlRobotEnvConfig,
+    device: torch.device = "mps",
+    output_directory: Path | None = None,
+):
+    """The actor process - interacts with environment and collects data.
+    The policy is frozen and only the parameters are updated, popping the most recent ones from a queue."""
+    policy_actor.eval()
+    policy_actor.to(device)
+
+    reward_classifier.eval()
+    reward_classifier.to(device)
+
+    # Create robot environment inside the actor process
+    env, teleop_device = make_robot_env(env_cfg)
+
+    try:
+        for episode in range(MAX_EPISODES):
+            if shutdown_event.is_set():
+                break
+
+            obs, _info = env.reset()
+            episode_reward = 0.0
+            step = 0
+            episode_transitions = []
+
+            print(f"[ACTOR] Starting episode {episode + 1}")
+
+            while step < MAX_STEPS_PER_EPISODE and not shutdown_event.is_set():
+                try:
+                    new_params = parameters_queue.get_nowait()
+                    policy_actor.load_state_dict(new_params)
+                    print("[ACTOR] Updated policy parameters from learner")
+                except Empty:  # No new updated parameters available from learner, waiting
+                    pass
+
+                # Get action from policy
+                policy_obs = make_policy_obs(obs, device=device)
+                action_tensor = policy_actor.select_action(policy_obs)  # predicts a single action
+                action = action_tensor.squeeze(0).cpu().numpy()
+
+                # Step environment
+                next_obs, _env_reward, terminated, truncated, _info = env.step(action)
+                done = terminated or truncated
+
+                # Predict reward
+                policy_next_obs = make_policy_obs(next_obs, device=device)
+                reward = reward_classifier.predict_reward(policy_next_obs)
+
+                if reward >= 1.0 and not done:  # success detected! halt episode
+                    terminated = True
+                    done = True
+
+                # In HIL-SERL, human interventions come from the teleop device
+                is_intervention = False
+                if hasattr(teleop_device, "get_teleop_events"):
+                    # Real intervention detection from teleop device
+                    teleop_events = teleop_device.get_teleop_events()
+                    is_intervention = teleop_events.get(TeleopEvents.IS_INTERVENTION, False)
+
+                # Store transition with intervention metadata
+                transition = {
+                    "state": policy_obs,
+                    "action": action,
+                    "reward": float(reward) if hasattr(reward, "item") else reward,
+                    "next_state": policy_next_obs,
+                    "done": done,
+                    "truncated": truncated,
+                    "complementary_info": {
+                        "is_intervention": is_intervention,
+                    },
+                }
+
+                episode_transitions.append(transition)
+
+                episode_reward += reward
+                step += 1
+
+                obs = next_obs
+
+                if done:
+                    break
+
+            # Send episode transitions to learner
+            transitions_queue.put_nowait(episode_transitions)
+
+    except KeyboardInterrupt:
+        print("[ACTOR] Interrupted by user")
+    finally:
+        # Clean up
+        if hasattr(env, "robot") and env.robot.is_connected:
+            env.robot.disconnect()
+        if teleop_device and hasattr(teleop_device, "disconnect"):
+            teleop_device.disconnect()
+        if output_directory is not None:
+            policy_actor.save_pretrained(output_directory)
+            print(f"[ACTOR] Latest actor policy saved at: {output_directory}")
+
+        print("[ACTOR] Actor process finished")
+
+
+def make_policy_obs(obs, device: torch.device = "cpu"):
+    return {
+        "observation.state": torch.from_numpy(obs["agent_pos"]).float().unsqueeze(0).to(device),
+        **{
+            f"observation.image.{k}": torch.from_numpy(obs["pixels"][k]).float().unsqueeze(0).to(device)
+            for k in obs["pixels"]
+        },
+    }
+
+
+"""Main function - coordinates actor and learner processes."""
+
+device = "mps"  # or "cuda" or "cpu"
+output_directory = Path("outputs/robot_learning_tutorial/hil_serl")
+output_directory.mkdir(parents=True, exist_ok=True)
+
+# find ports using lerobot-find-port
+follower_port = ...
+leader_port = ...
+
+# the robot ids are used the load the right calibration files
+follower_id = ...
+leader_id = ...
+
+# A pretrained model (to be used in-distribution!)
+reward_classifier_id = "fracapuano/reward_classifier_hil_serl_example"
+reward_classifier = Classifier.from_pretrained(reward_classifier_id)
+
+reward_classifier.to(device)
+reward_classifier.eval()
+
+MAX_EPISODES = 5
+MAX_STEPS_PER_EPISODE = 20
+
+# Robot and environment configuration
+robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id)
+teleop_cfg = SO100LeaderConfig(port=leader_port, id=leader_id)
+processor_cfg = HILSerlProcessorConfig(control_mode="leader")
+
+env_cfg = HILSerlRobotEnvConfig(robot=robot_cfg, teleop=teleop_cfg, processor=processor_cfg)
+
+# Create robot environment
+env, teleop_device = make_robot_env(env_cfg)
+
+obs_features = hw_to_dataset_features(env.robot.observation_features, "observation")
+action_features = hw_to_dataset_features(env.robot.action_features, "action")
+
+# Create SAC policy for action selection
+policy_cfg = SACConfig(
+    device=device,
+    input_features=obs_features,
+    output_features=action_features,
+)
+
+policy_actor = SACPolicy(policy_cfg)
+policy_learner = SACPolicy(policy_cfg)
+
+demonstrations_repo_id = "lerobot/example_hil_serl_dataset"
+offline_dataset = LeRobotDataset(repo_id=demonstrations_repo_id)
+
+# Online buffer: initialized from scratch
+online_replay_buffer = ReplayBuffer(device=device, state_keys=list(obs_features.keys()))
+# Offline buffer: Created from dataset (pre-populated it with demonstrations)
+offline_replay_buffer = ReplayBuffer.from_lerobot_dataset(
+    lerobot_dataset=offline_dataset, device=device, state_keys=list(obs_features.keys())
+)
+
+# Create communication channels between learner and actor processes
+transitions_queue = mp.Queue(maxsize=10)
+parameters_queue = mp.Queue(maxsize=2)
+shutdown_event = mp.Event()
+
+
+# Signal handler for graceful shutdown
+def signal_handler(sig):
+    print(f"\nSignal {sig} received, shutting down...")
+    shutdown_event.set()
+
+
+signal.signal(signal.SIGINT, signal_handler)
+signal.signal(signal.SIGTERM, signal_handler)
+
+# Create processes
+learner_process = mp.Process(
+    target=run_learner,
+    args=(
+        transitions_queue,
+        parameters_queue,
+        shutdown_event,
+        policy_learner,
+        online_replay_buffer,
+        offline_replay_buffer,
+    ),
+    kwargs={"device": device},  # can run on accelerated hardware for training
+)
+
+actor_process = mp.Process(
+    target=run_actor,
+    args=(
+        transitions_queue,
+        parameters_queue,
+        shutdown_event,
+        policy_actor,
+        reward_classifier,
+        env_cfg,
+        output_directory,
+    ),
+    kwargs={"device": "cpu"},  # actor is frozen, can run on CPU or accelerate for inference
+)
+
+learner_process.start()
+actor_process.start()
+
+try:
+    # Wait for actor to finish (it controls the episode loop)
+    actor_process.join()
+    shutdown_event.set()
+    learner_process.join(timeout=10)
+
+except KeyboardInterrupt:
+    print("Main process interrupted")
+    shutdown_event.set()
+    actor_process.join(timeout=5)
+    learner_process.join(timeout=10)
+
+finally:
+    if learner_process.is_alive():
+        learner_process.terminate()
+    if actor_process.is_alive():
+        actor_process.terminate()

examples/tutorial/rl/reward_classifier_example.py

@@ -0,0 +1,62 @@
+import torch
+
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.policies.factory import make_policy, make_pre_post_processors
+from lerobot.policies.sac.reward_model.configuration_classifier import RewardClassifierConfig
+
+# Device to use for training
+device = "mps"  # or "cuda", or "cpu"
+
+# Load the dataset used for training
+repo_id = "lerobot/example_hil_serl_dataset"
+dataset = LeRobotDataset(repo_id)
+
+# Configure the policy to extract features from the image frames
+camera_keys = dataset.meta.camera_keys
+
+config = RewardClassifierConfig(
+    num_cameras=len(camera_keys),
+    device=device,
+    # backbone model to extract features from the image frames
+    model_name="microsoft/resnet-18",
+)
+
+# Make policy, preprocessor, and optimizer
+policy = make_policy(config, ds_meta=dataset.meta)
+optimizer = config.get_optimizer_preset().build(policy.parameters())
+preprocessor, _ = make_pre_post_processors(policy_cfg=config, dataset_stats=dataset.meta.stats)
+
+
+classifier_id = "fracapuano/reward_classifier_hil_serl_example"
+
+# Instantiate a dataloader
+dataloader = torch.utils.data.DataLoader(dataset, batch_size=16, shuffle=True)
+
+# Training loop
+num_epochs = 5
+for epoch in range(num_epochs):
+    total_loss = 0
+    total_accuracy = 0
+    for batch in dataloader:
+        # Preprocess the batch and move it to the correct device.
+        batch = preprocessor(batch)
+
+        # Forward pass
+        loss, output_dict = policy.forward(batch)
+
+        # Backward pass and optimization
+        optimizer.zero_grad()
+        loss.backward()
+        optimizer.step()
+
+        total_loss += loss.item()
+        total_accuracy += output_dict["accuracy"]
+
+    avg_loss = total_loss / len(dataloader)
+    avg_accuracy = total_accuracy / len(dataloader)
+    print(f"Epoch {epoch + 1}/{num_epochs}, Loss: {avg_loss:.4f}, Accuracy: {avg_accuracy:.2f}%")
+
+print("Training finished!")
+
+# You can now save the trained policy.
+policy.push_to_hub(classifier_id)

examples/tutorial/smolvla/using_smolvla_example.py

@@ -0,0 +1,66 @@
+import torch
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.datasets.utils import hw_to_dataset_features
+from lerobot.policies.factory import make_pre_post_processors
+from lerobot.policies.smolvla.modeling_smolvla import SmolVLAPolicy
+from lerobot.policies.utils import build_inference_frame, make_robot_action
+from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
+from lerobot.robots.so100_follower.so100_follower import SO100Follower
+
+MAX_EPISODES = 5
+MAX_STEPS_PER_EPISODE = 20
+
+device = torch.device("mps")  # or "cuda" or "cpu"
+model_id = "lerobot/smolvla_base"
+
+model = SmolVLAPolicy.from_pretrained(model_id)
+
+preprocess, postprocess = make_pre_post_processors(
+    model.config,
+    model_id,
+    # This overrides allows to run on MPS, otherwise defaults to CUDA (if available)
+    preprocessor_overrides={"device_processor": {"device": str(device)}},
+)
+
+# find ports using lerobot-find-port
+follower_port = ...  # something like "/dev/tty.usbmodem58760431631"
+
+# the robot ids are used the load the right calibration files
+follower_id = ...  # something like "follower_so100"
+
+# Robot and environment configuration
+# Camera keys must match the name and resolutions of the ones used for training!
+# You can check the camera keys expected by a model in the info.json card on the model card on the Hub
+camera_config = {
+    "camera1": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
+    "camera2": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30),
+}
+
+robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_config)
+robot = SO100Follower(robot_cfg)
+robot.connect()
+
+task = ""  # something like "pick the red block"
+robot_type = ""  # something like "so100_follower" for multi-embodiment datasets
+
+# This is used to match the raw observation keys to the keys expected by the policy
+action_features = hw_to_dataset_features(robot.action_features, "action")
+obs_features = hw_to_dataset_features(robot.observation_features, "observation")
+dataset_features = {**action_features, **obs_features}
+
+for _ in range(MAX_EPISODES):
+    for _ in range(MAX_STEPS_PER_EPISODE):
+        obs = robot.get_observation()
+        obs_frame = build_inference_frame(
+            observation=obs, ds_features=dataset_features, device=device, task=task, robot_type=robot_type
+        )
+
+        obs = preprocess(obs_frame)
+
+        action = model.select_action(obs)
+        action = postprocess(action)
+        action = make_robot_action(action, dataset_features)
+        robot.send_action(action)
+
+    print("Episode finished! Starting new episode...")

pyproject.toml

@@ -25,7 +25,7 @@ discord = "https://discord.gg/s3KuuzsPFb"
 
 [project]
 name = "lerobot"
-version = "0.3.4"
+version = "0.4.2"
 description = "🤗 LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch"
 readme = "README.md"
 license = { text = "Apache-2.0" }
@@ -62,8 +62,10 @@ dependencies = [
     "datasets>=4.0.0,<4.2.0",
     "diffusers>=0.27.2,<0.36.0",
     "huggingface-hub[hf-transfer,cli]>=0.34.2,<0.36.0",
+    "accelerate>=1.10.0,<2.0.0",
 
     # Core dependencies
+    "setuptools>=71.0.0,<81.0.0",
     "cmake>=3.29.0.1,<4.2.0",
     "einops>=0.8.0,<0.9.0",
     "opencv-python-headless>=4.9.0,<4.13.0",
@@ -72,15 +74,15 @@ dependencies = [
     "packaging>=24.2,<26.0",
     "pynput>=1.7.7,<1.9.0",
     "pyserial>=3.5,<4.0",
-    "wandb>=0.20.0,<0.23.0",
+    "wandb>=0.20.0,<0.22.0", # TODO: Bumb dependency (compatible with protobuf)
 
     "torch>=2.2.1,<2.8.0", # TODO: Bumb dependency
     "torchcodec>=0.2.1,<0.6.0; sys_platform != 'win32' and (sys_platform != 'linux' or (platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')) and (sys_platform != 'darwin' or platform_machine != 'x86_64')", # TODO: Bumb dependency
     "torchvision>=0.21.0,<0.23.0", # TODO: Bumb dependency
 
     "draccus==0.10.0", # TODO: Remove ==
-    "gymnasium>=0.29.1,<1.0.0", # TODO: Bumb dependency
-    "rerun-sdk>=0.21.0,<0.23.0", # TODO: Bumb dependency
+    "gymnasium>=1.1.1,<2.0.0",
+    "rerun-sdk>=0.24.0,<0.27.0",
 
     # Support dependencies
     "deepdiff>=7.0.1,<9.0.0",
@@ -95,7 +97,7 @@ dependencies = [
 pygame-dep = ["pygame>=2.5.1,<2.7.0"]
 placo-dep = ["placo>=0.9.6,<0.10.0"]
 transformers-dep = ["transformers>=4.53.0,<5.0.0"]
-grpcio-dep = ["grpcio==1.73.1", "protobuf==6.31.0"]
+grpcio-dep = ["grpcio==1.73.1", "protobuf==6.31.0"] # TODO: Bumb dependency (compatible with wandb)
 
 # Motors
 feetech = ["feetech-servo-sdk>=1.0.0,<2.0.0"]
@@ -111,17 +113,23 @@ intelrealsense = [
     "pyrealsense2>=2.55.1.6486,<2.57.0 ; sys_platform != 'darwin'",
     "pyrealsense2-macosx>=2.54,<2.55.0 ; sys_platform == 'darwin'",
 ]
-phone = ["hebi-py>=2.8.0,<2.12.0", "teleop>=0.1.0,<0.2.0"]
-# stretch = [
-#     "hello-robot-stretch-body>=0.7.27 ; sys_platform == 'linux'",
-#     "pyrender @ git+https://github.com/mmatl/pyrender.git ; sys_platform == 'linux'",
-#     "pyrealsense2>=2.55.1.6486 ; sys_platform != 'darwin'"
-# ] # TODO: Currently not supported
+phone = ["hebi-py>=2.8.0,<2.12.0", "teleop>=0.1.0,<0.2.0", "fastapi<1.0"]
 
 # Policies
 pi = ["transformers @ git+https://github.com/huggingface/transformers.git@fix/lerobot_openpi"]
 smolvla = ["lerobot[transformers-dep]", "num2words>=0.5.14,<0.6.0", "accelerate>=1.7.0,<2.0.0", "safetensors>=0.4.3,<1.0.0"]
-hilserl = ["lerobot[transformers-dep]", "gym-hil>=0.1.11,<0.2.0", "lerobot[grpcio-dep]", "lerobot[placo-dep]"]
+groot = [
+    "lerobot[transformers-dep]",
+    "peft>=0.13.0,<1.0.0",
+    "dm-tree>=0.1.8,<1.0.0",
+    "timm>=1.0.0,<1.1.0",
+    "safetensors>=0.4.3,<1.0.0",
+    "Pillow>=10.0.0,<13.0.0",
+    "decord>=0.6.0,<1.0.0; (platform_machine == 'AMD64' or platform_machine == 'x86_64')",
+    "ninja>=1.11.1,<2.0.0",
+    "flash-attn>=2.5.9,<3.0.0 ; sys_platform != 'darwin'"
+]
+hilserl = ["lerobot[transformers-dep]", "gym-hil>=0.1.13,<0.2.0", "lerobot[grpcio-dep]", "lerobot[placo-dep]"]
 
 # Features
 async = ["lerobot[grpcio-dep]", "matplotlib>=3.10.3,<4.0.0"]
@@ -132,11 +140,10 @@ test = ["pytest>=8.1.0,<9.0.0", "pytest-timeout>=2.4.0,<3.0.0", "pytest-cov>=5.0
 video_benchmark = ["scikit-image>=0.23.2,<0.26.0", "pandas>=2.2.2,<2.4.0"]
 
 # Simulation
-aloha = ["gym-aloha>=0.1.1,<0.2.0"]
+aloha = ["gym-aloha>=0.1.2,<0.2.0"]
 pusht = ["gym-pusht>=0.1.5,<0.2.0", "pymunk>=6.6.0,<7.0.0"] # TODO: Fix pymunk version in gym-pusht instead
-xarm = ["gym-xarm>=0.1.1,<0.2.0"]
-libero = ["lerobot[transformers-dep]", "libero @ git+https://github.com/huggingface/lerobot-libero.git@main#egg=libero"]
-
+libero = ["lerobot[transformers-dep]", "hf-libero>=0.1.3,<0.2.0"]
+metaworld = ["metaworld==3.0.0"]
 
 # All
 all = [
@@ -149,6 +156,7 @@ all = [
     "lerobot[intelrealsense]",
     "lerobot[pi]",
     "lerobot[smolvla]",
+    # "lerobot[groot]", TODO(Steven): Gr00t requires specific installation instructions for flash-attn
     "lerobot[hilserl]",
     "lerobot[async]",
     "lerobot[dev]",
@@ -156,9 +164,9 @@ all = [
     "lerobot[video_benchmark]",
     "lerobot[aloha]",
     "lerobot[pusht]",
-    "lerobot[xarm]",
     "lerobot[phone]",
     "lerobot[libero]",
+    "lerobot[metaworld]",
 ]
 
 [project.scripts]
@@ -233,9 +241,6 @@ exclude_dirs = [
     "tests",
     "benchmarks",
     "src/lerobot/datasets/push_dataset_to_hub",
-    "src/lerobot/datasets/v2/convert_dataset_v1_to_v2",
-    "src/lerobot/policies/pi0/conversion_scripts",
-    "src/lerobot/scripts/push_dataset_to_hub.py",
 ]
 skips = ["B101", "B311", "B404", "B603", "B615"]
 
@@ -250,6 +255,8 @@ default.extend-ignore-identifiers-re = [
     "pn",
     "ser",
     "ein",
+    "thw",
+    "inpt",
 ]
 
 # TODO: Uncomment when ready to use
@@ -288,7 +295,6 @@ ignore_errors = true
 
 [[tool.mypy.overrides]]
 module = "lerobot.envs.*"
-# Enable type checking only for the envs module
 ignore_errors = false
 
 
@@ -296,17 +302,22 @@ ignore_errors = false
 # module = "lerobot.utils.*"
 # ignore_errors = false
 
-# [[tool.mypy.overrides]]
-# module = "lerobot.configs.*"
-# ignore_errors = false
+[[tool.mypy.overrides]]
+module = "lerobot.configs.*"
+ignore_errors = false
+
+# extra strictness for configs
+disallow_untyped_defs = true
+disallow_incomplete_defs = true
+check_untyped_defs = true
 
 # [[tool.mypy.overrides]]
 # module = "lerobot.optim.*"
 # ignore_errors = false
 
-# [[tool.mypy.overrides]]
-# module = "lerobot.model.*"
-# ignore_errors = false
+[[tool.mypy.overrides]]
+module = "lerobot.model.*"
+ignore_errors = false
 
 # [[tool.mypy.overrides]]
 # module = "lerobot.processor.*"
@@ -316,9 +327,9 @@ ignore_errors = false
 # module = "lerobot.datasets.*"
 # ignore_errors = false
 
-# [[tool.mypy.overrides]]
-# module = "lerobot.cameras.*"
-# ignore_errors = false
+[[tool.mypy.overrides]]
+module = "lerobot.cameras.*"
+ignore_errors = false
 
 # [[tool.mypy.overrides]]
 # module = "lerobot.motors.*"

requirements-macos.txt

@@ -1,3 +1,4 @@
+#
 # This file is autogenerated by pip-compile with Python 3.10
 # by the following command:
 #
@@ -12,47 +13,62 @@ absl-py==2.3.1
     #   dm-tree
     #   labmaze
     #   mujoco
-accelerate==1.9.0
-    # via lerobot
+    #   tensorboard
+accelerate==1.11.0
+    # via
+    #   lerobot
+    #   peft
 aiohappyeyeballs==2.6.1
     # via aiohttp
-aiohttp==3.12.15
+aiohttp==3.13.1
     # via fsspec
 aiosignal==1.4.0
     # via aiohttp
 annotated-types==0.7.0
     # via pydantic
+antlr4-python3-runtime==4.9.3
+    # via
+    #   hydra-core
+    #   omegaconf
+anyio==4.11.0
+    # via
+    #   starlette
+    #   watchfiles
 asttokens==3.0.0
     # via stack-data
 async-timeout==5.0.1
     # via aiohttp
-attrs==25.3.0
+attrs==25.4.0
     # via
     #   aiohttp
     #   dm-tree
     #   jsonlines
+    #   jsonschema
+    #   referencing
     #   rerun-sdk
-av==15.0.0
+av==15.1.0
     # via lerobot
-blinker==1.9.0
-    # via flask
-certifi==2025.7.14
+bddl==1.0.1
+    # via libero
+certifi==2025.10.5
     # via
     #   requests
     #   sentry-sdk
-cffi==1.17.1
+cffi==2.0.0
     # via pymunk
 cfgv==3.4.0
     # via pre-commit
-charset-normalizer==3.4.2
+charset-normalizer==3.4.4
     # via requests
-click==8.2.1
+click==8.3.0
     # via
-    #   flask
+    #   uvicorn
     #   wandb
 cloudpickle==3.1.1
-    # via gymnasium
-cmake==4.0.3
+    # via
+    #   gymnasium
+    #   libero
+cmake==4.1.0
     # via lerobot
 cmeel==0.57.3
     # via
@@ -94,27 +110,27 @@ coal-library==3.0.1
     # via pin
 contourpy==1.3.2
     # via matplotlib
-coverage[toml]==7.10.1
+coverage[toml]==7.11.0
     # via pytest-cov
 cycler==0.12.1
     # via matplotlib
-datasets==3.6.0
+datasets==4.1.1
     # via lerobot
-debugpy==1.8.15
+debugpy==1.8.17
     # via lerobot
 decorator==5.2.1
     # via ipython
-deepdiff==8.5.0
+deepdiff==8.6.1
     # via lerobot
-diffusers==0.34.0
+diffusers==0.35.2
     # via lerobot
-dill==0.3.8
+dill==0.4.0
     # via
     #   datasets
     #   multiprocess
 distlib==0.4.0
     # via virtualenv
-dm-control==1.0.14
+dm-control==1.0.34
     # via gym-aloha
 dm-env==1.6
     # via dm-control
@@ -122,29 +138,45 @@ dm-tree==0.1.9
     # via
     #   dm-control
     #   dm-env
+    #   lerobot
 docopt==0.6.2
     # via num2words
 draccus==0.10.0
     # via lerobot
-dynamixel-sdk==3.7.31
+dynamixel-sdk==3.8.4
     # via lerobot
+easydict==1.13
+    # via libero
+egl-probe @ git+https://github.com/huggingface/egl_probe.git
+    # via
+    #   libero
+    #   robomimic
 eigenpy==3.10.3
     # via coal-library
 einops==0.8.1
-    # via lerobot
+    # via
+    #   lerobot
+    #   libero
 eiquadprog==1.2.9
     # via placo
+etils[epath,epy]==1.13.0
+    # via mujoco
 exceptiongroup==1.3.0
     # via
+    #   anyio
     #   ipython
     #   pytest
-executing==2.2.0
+executing==2.2.1
     # via stack-data
 farama-notifications==0.0.4
     # via gymnasium
+fastapi==0.119.1
+    # via teleop
+fastjsonschema==2.21.2
+    # via nbformat
 feetech-servo-sdk==1.0.0
     # via lerobot
-filelock==3.18.0
+filelock==3.20.0
     # via
     #   datasets
     #   diffusers
@@ -152,24 +184,25 @@ filelock==3.18.0
     #   torch
     #   transformers
     #   virtualenv
-flask==3.1.1
-    # via lerobot
-fonttools==4.59.0
+fonttools==4.60.1
     # via matplotlib
-frozenlist==1.7.0
+frozenlist==1.8.0
     # via
     #   aiohttp
     #   aiosignal
-fsspec[http]==2025.3.0
+fsspec[http]==2025.9.0
     # via
     #   datasets
+    #   etils
     #   huggingface-hub
     #   torch
+future==1.0.0
+    # via libero
 gitdb==4.0.12
     # via gitpython
 gitpython==3.1.45
     # via wandb
-glfw==2.9.0
+glfw==2.10.0
     # via
     #   dm-control
     #   mujoco
@@ -177,61 +210,79 @@ grpcio==1.73.1
     # via
     #   grpcio-tools
     #   lerobot
+    #   reachy2-sdk
+    #   reachy2-sdk-api
+    #   tensorboard
 grpcio-tools==1.73.1
+    # via
+    #   lerobot
+    #   reachy2-sdk-api
+gym-aloha==0.1.3
     # via lerobot
-gym-aloha==0.1.1
+gym-hil==0.1.13
     # via lerobot
-gym-hil==0.1.10
+gym-pusht==0.1.6
     # via lerobot
-gym-pusht==0.1.5
-    # via lerobot
-gym-xarm==0.1.1
-    # via lerobot
-gymnasium==0.29.1
+gymnasium==1.2.1
     # via
     #   gym-aloha
     #   gym-hil
     #   gym-pusht
-    #   gym-xarm
-    #   gymnasium-robotics
     #   lerobot
-    #   pettingzoo
-gymnasium-robotics==1.2.4
-    # via gym-xarm
+    #   libero
+    #   metaworld
+h11==0.16.0
+    # via uvicorn
+h5py==3.15.1
+    # via robomimic
+hebi-py==2.11.0
+    # via lerobot
 hf-transfer==0.1.9
     # via huggingface-hub
-hf-xet==1.1.5
+hf-xet==1.1.10
     # via huggingface-hub
 hidapi==0.14.0.post4
     # via
     #   gym-hil
     #   lerobot
-huggingface-hub[cli,hf-transfer]==0.34.3
+httptools==0.7.1
+    # via uvicorn
+huggingface-hub[cli,hf-transfer]==0.35.3
     # via
     #   accelerate
     #   datasets
     #   diffusers
     #   lerobot
+    #   peft
+    #   timm
     #   tokenizers
     #   transformers
-identify==2.6.12
+hydra-core==1.3.2
+    # via libero
+identify==2.6.15
     # via pre-commit
-idna==3.10
+idna==3.11
     # via
+    #   anyio
     #   requests
     #   yarl
 imageio[ffmpeg]==2.37.0
     # via
     #   gym-aloha
     #   gym-hil
-    #   gymnasium-robotics
     #   lerobot
+    #   metaworld
+    #   robomimic
     #   scikit-image
 imageio-ffmpeg==0.6.0
-    # via imageio
+    # via
+    #   imageio
+    #   robomimic
 importlib-metadata==8.7.0
     # via diffusers
-iniconfig==2.1.0
+importlib-resources==6.5.2
+    # via etils
+iniconfig==2.3.0
     # via pytest
 inquirerpy==0.3.4
     # via huggingface-hub
@@ -239,50 +290,71 @@ ipython==8.37.0
     # via meshcat
 ischedule==1.2.7
     # via placo
-itsdangerous==2.2.0
-    # via flask
 jedi==0.19.2
     # via ipython
 jinja2==3.1.6
-    # via
-    #   flask
-    #   gymnasium-robotics
-    #   torch
+    # via torch
 jsonlines==4.0.0
     # via lerobot
-kiwisolver==1.4.8
+jsonschema==4.25.1
+    # via nbformat
+jsonschema-specifications==2025.9.1
+    # via jsonschema
+jupyter-core==5.9.1
+    # via nbformat
+jupytext==1.18.1
+    # via bddl
+kiwisolver==1.4.9
     # via matplotlib
 labmaze==1.0.6
     # via dm-control
 lazy-loader==0.4
     # via scikit-image
-lxml==6.0.0
+libero @ git+https://github.com/huggingface/lerobot-libero.git@main
+    # via lerobot
+llvmlite==0.45.1
+    # via numba
+lxml==6.0.2
     # via dm-control
-markupsafe==3.0.2
+markdown==3.9
+    # via tensorboard
+markdown-it-py==4.0.0
+    # via
+    #   jupytext
+    #   mdit-py-plugins
+markupsafe==3.0.3
     # via
-    #   flask
     #   jinja2
     #   werkzeug
-matplotlib==3.10.5
-    # via lerobot
-matplotlib-inline==0.1.7
+matplotlib==3.10.7
+    # via
+    #   lerobot
+    #   libero
+matplotlib-inline==0.2.1
     # via ipython
+mdit-py-plugins==0.5.0
+    # via jupytext
+mdurl==0.1.2
+    # via markdown-it-py
 mergedeep==1.3.4
     # via draccus
 meshcat==0.3.2
     # via placo
+metaworld==3.0.0
+    # via lerobot
 mock-serial==0.0.1
     # via lerobot
 mpmath==1.3.0
     # via sympy
-mujoco==2.3.7
+mujoco==3.3.7
     # via
     #   dm-control
     #   gym-aloha
     #   gym-hil
-    #   gym-xarm
-    #   gymnasium-robotics
-multidict==6.6.3
+    #   libero
+    #   metaworld
+    #   robosuite
+multidict==6.7.0
     # via
     #   aiohttp
     #   yarl
@@ -290,17 +362,25 @@ multiprocess==0.70.16
     # via datasets
 mypy-extensions==1.1.0
     # via typing-inspect
+nbformat==5.10.4
+    # via jupytext
 networkx==3.4.2
     # via
+    #   bddl
     #   scikit-image
     #   torch
+ninja==1.13.0
+    # via lerobot
 nodeenv==1.9.1
     # via pre-commit
 num2words==0.5.14
     # via lerobot
+numba==0.62.1
+    # via robosuite
 numpy==2.2.6
     # via
     #   accelerate
+    #   bddl
     #   cmeel-boost
     #   contourpy
     #   datasets
@@ -309,25 +389,43 @@ numpy==2.2.6
     #   dm-env
     #   dm-tree
     #   gymnasium
-    #   gymnasium-robotics
+    #   h5py
+    #   hebi-py
     #   imageio
     #   labmaze
+    #   libero
     #   matplotlib
     #   meshcat
+    #   metaworld
     #   mujoco
+    #   numba
     #   opencv-python
     #   opencv-python-headless
     #   pandas
-    #   pettingzoo
+    #   peft
+    #   pyquaternion
+    #   reachy2-sdk
     #   rerun-sdk
+    #   robomimic
+    #   robosuite
     #   scikit-image
     #   scipy
     #   shapely
+    #   teleop
+    #   tensorboard
+    #   tensorboardx
     #   tifffile
     #   torchvision
     #   transformers
+    #   transforms3d
+omegaconf==2.3.0
+    # via hydra-core
 opencv-python==4.12.0.88
-    # via gym-pusht
+    # via
+    #   gym-pusht
+    #   libero
+    #   reachy2-sdk
+    #   robosuite
 opencv-python-headless==4.12.0.88
     # via lerobot
 orderly-set==5.5.0
@@ -337,53 +435,63 @@ packaging==25.0
     #   accelerate
     #   datasets
     #   huggingface-hub
+    #   hydra-core
+    #   jupytext
     #   lazy-loader
     #   lerobot
     #   matplotlib
+    #   peft
     #   pytest
+    #   reachy2-sdk
     #   scikit-image
+    #   tensorboard
+    #   tensorboardx
     #   transformers
     #   wandb
-pandas==2.3.1
+pandas==2.3.3
     # via
     #   datasets
     #   lerobot
-parso==0.8.4
+parso==0.8.5
     # via jedi
-pettingzoo==1.24.3
-    # via gymnasium-robotics
+peft==0.17.1
+    # via lerobot
 pexpect==4.9.0
     # via ipython
 pfzy==0.3.4
     # via inquirerpy
-pillow==11.3.0
+pillow==12.0.0
     # via
     #   diffusers
     #   imageio
+    #   lerobot
     #   matplotlib
     #   meshcat
     #   rerun-sdk
+    #   robosuite
     #   scikit-image
+    #   tensorboard
     #   torchvision
 pin==3.4.0
     # via placo
 placo==0.9.14
     # via lerobot
-platformdirs==4.3.8
+platformdirs==4.5.0
     # via
+    #   jupyter-core
     #   virtualenv
     #   wandb
 pluggy==1.6.0
     # via
     #   pytest
     #   pytest-cov
-pre-commit==4.2.0
+pre-commit==4.3.0
     # via lerobot
-prompt-toolkit==3.0.51
+prompt-toolkit==3.0.52
     # via
     #   inquirerpy
     #   ipython
-propcache==0.3.2
+propcache==0.4.1
     # via
     #   aiohttp
     #   yarl
@@ -392,11 +500,17 @@ protobuf==6.31.0
     #   dm-control
     #   grpcio-tools
     #   lerobot
+    #   reachy2-sdk
+    #   reachy2-sdk-api
+    #   tensorboard
+    #   tensorboardx
     #   wandb
-psutil==7.0.0
+psutil==7.1.1
     # via
     #   accelerate
     #   imageio
+    #   peft
+    #   robomimic
 ptyprocess==0.7.0
     # via pexpect
 pure-eval==0.2.3
@@ -405,11 +519,13 @@ pyarrow==21.0.0
     # via
     #   datasets
     #   rerun-sdk
-pycparser==2.22
+pycparser==2.23
     # via cffi
-pydantic==2.11.7
-    # via wandb
-pydantic-core==2.33.2
+pydantic==2.12.3
+    # via
+    #   fastapi
+    #   wandb
+pydantic-core==2.41.4
     # via pydantic
 pygame==2.6.1
     # via
@@ -424,40 +540,42 @@ pymunk==6.11.1
     # via
     #   gym-pusht
     #   lerobot
-pyngrok==7.2.12
+pyngrok==7.4.1
     # via meshcat
 pynput==1.8.1
     # via
     #   gym-hil
     #   lerobot
-pyobjc-core==11.1
+pyobjc-core==12.0
     # via
     #   pyobjc-framework-applicationservices
     #   pyobjc-framework-cocoa
     #   pyobjc-framework-coretext
     #   pyobjc-framework-quartz
-pyobjc-framework-applicationservices==11.1
+pyobjc-framework-applicationservices==12.0
     # via pynput
-pyobjc-framework-cocoa==11.1
+pyobjc-framework-cocoa==12.0
     # via
     #   pyobjc-framework-applicationservices
     #   pyobjc-framework-coretext
     #   pyobjc-framework-quartz
-pyobjc-framework-coretext==11.1
+pyobjc-framework-coretext==12.0
     # via pyobjc-framework-applicationservices
-pyobjc-framework-quartz==11.1
+pyobjc-framework-quartz==12.0
     # via
     #   pynput
     #   pyobjc-framework-applicationservices
     #   pyobjc-framework-coretext
-pyopengl==3.1.9
+pyopengl==3.1.10
     # via
     #   dm-control
     #   mujoco
-pyparsing==3.2.3
+pyparsing==3.2.5
     # via
     #   dm-control
     #   matplotlib
+pyquaternion==0.9.9
+    # via reachy2-sdk
 pyrealsense2-macosx==2.54.2
     # via lerobot
 pyserial==3.5
@@ -465,12 +583,14 @@ pyserial==3.5
     #   dynamixel-sdk
     #   feetech-servo-sdk
     #   lerobot
-pytest==8.4.1
+pytest==8.4.2
     # via
+    #   bddl
     #   lerobot
     #   pytest-cov
     #   pytest-timeout
-pytest-cov==6.2.1
+    #   teleop
+pytest-cov==7.0.0
     # via lerobot
 pytest-timeout==2.4.0
     # via lerobot
@@ -478,46 +598,73 @@ python-dateutil==2.9.0.post0
     # via
     #   matplotlib
     #   pandas
+python-dotenv==1.1.1
+    # via uvicorn
 pytz==2025.2
     # via pandas
-pyyaml==6.0.2
+pyyaml==6.0.3
     # via
     #   accelerate
     #   datasets
     #   draccus
+    #   hebi-py
     #   huggingface-hub
+    #   jupytext
+    #   omegaconf
+    #   peft
     #   pre-commit
     #   pyngrok
     #   pyyaml-include
+    #   timm
     #   transformers
+    #   uvicorn
     #   wandb
 pyyaml-include==1.4.1
     # via draccus
-pyzmq==27.0.0
+pyzmq==27.1.0
     # via
     #   lerobot
     #   meshcat
-regex==2025.7.34
+reachy2-sdk==1.0.14
+    # via lerobot
+reachy2-sdk-api==1.0.21
+    # via reachy2-sdk
+referencing==0.37.0
+    # via
+    #   jsonschema
+    #   jsonschema-specifications
+regex==2025.10.23
     # via
     #   diffusers
     #   transformers
-requests==2.32.4
+requests==2.32.5
     # via
     #   datasets
     #   diffusers
     #   dm-control
     #   huggingface-hub
+    #   teleop
     #   transformers
     #   wandb
-rerun-sdk==0.22.1
+rerun-sdk==0.26.1
     # via lerobot
 rhoban-cmeel-jsoncpp==1.9.4.9
     # via placo
-safetensors==0.5.3
+robomimic==0.2.0
+    # via libero
+robosuite==1.4.0
+    # via libero
+rpds-py==0.28.0
+    # via
+    #   jsonschema
+    #   referencing
+safetensors==0.6.2
     # via
     #   accelerate
     #   diffusers
     #   lerobot
+    #   peft
+    #   timm
     #   transformers
 scikit-image==0.25.2
     # via
@@ -526,10 +673,12 @@ scikit-image==0.25.2
 scipy==1.15.3
     # via
     #   dm-control
+    #   metaworld
+    #   robosuite
     #   scikit-image
-sentry-sdk==2.34.1
+sentry-sdk==2.42.1
     # via wandb
-shapely==2.1.1
+shapely==2.1.2
     # via gym-pusht
 six==1.17.0
     # via
@@ -537,64 +686,106 @@ six==1.17.0
     #   python-dateutil
 smmap==5.0.2
     # via gitdb
+sniffio==1.3.1
+    # via anyio
 stack-data==0.6.3
     # via ipython
+starlette==0.48.0
+    # via fastapi
 sympy==1.14.0
     # via torch
-termcolor==3.1.0
+teleop==0.1.2
     # via lerobot
+tensorboard==2.20.0
+    # via robomimic
+tensorboard-data-server==0.7.2
+    # via tensorboard
+tensorboardx==2.6.4
+    # via robomimic
+termcolor==3.1.0
+    # via
+    #   lerobot
+    #   robomimic
+thop==0.1.1.post2209072238
+    # via libero
 tifffile==2025.5.10
     # via scikit-image
-tokenizers==0.21.4
+timm==1.0.20
+    # via lerobot
+tokenizers==0.22.1
     # via transformers
 toml==0.10.2
     # via draccus
-tomli==2.2.1
+tomli==2.3.0
     # via
     #   cmeel
     #   coverage
+    #   jupytext
     #   pytest
 torch==2.7.1
     # via
     #   accelerate
     #   lerobot
+    #   peft
+    #   robomimic
+    #   thop
+    #   timm
     #   torchvision
 torchcodec==0.5
     # via lerobot
 torchvision==0.22.1
-    # via lerobot
-tornado==6.5.1
+    # via
+    #   lerobot
+    #   robomimic
+    #   timm
+tornado==6.5.2
     # via meshcat
 tqdm==4.67.1
     # via
     #   datasets
     #   dm-control
     #   huggingface-hub
+    #   peft
+    #   robomimic
     #   transformers
 traitlets==5.14.3
     # via
     #   ipython
+    #   jupyter-core
     #   matplotlib-inline
-transformers==4.51.3
-    # via lerobot
-typing-extensions==4.14.1
+    #   nbformat
+transformers==4.57.1
+    # via
+    #   lerobot
+    #   libero
+    #   peft
+transforms3d==0.4.2
+    # via teleop
+typing-extensions==4.15.0
     # via
     #   aiosignal
+    #   anyio
+    #   etils
     #   exceptiongroup
+    #   fastapi
     #   gymnasium
     #   huggingface-hub
     #   ipython
     #   multidict
     #   pydantic
     #   pydantic-core
+    #   referencing
     #   rerun-sdk
+    #   starlette
     #   torch
     #   typing-inspect
     #   typing-inspection
+    #   uvicorn
+    #   virtualenv
     #   wandb
 typing-inspect==0.9.0
     # via draccus
-typing-inspection==0.4.1
+typing-inspection==0.4.2
     # via pydantic
 tzdata==2025.2
     # via pandas
@@ -604,22 +795,36 @@ urllib3==2.5.0
     # via
     #   requests
     #   sentry-sdk
-virtualenv==20.32.0
+uvicorn[standard]==0.38.0
+    # via teleop
+uvloop==0.22.1
+    # via uvicorn
+virtualenv==20.35.3
     # via pre-commit
-wandb==0.21.0
-    # via lerobot
-wcwidth==0.2.13
+wandb==0.21.4
+    # via
+    #   lerobot
+    #   libero
+watchfiles==1.1.1
+    # via uvicorn
+wcwidth==0.2.14
     # via prompt-toolkit
+websocket-client==1.9.0
+    # via teleop
+websockets==15.0.1
+    # via uvicorn
 werkzeug==3.1.3
-    # via flask
-wrapt==1.17.2
+    # via tensorboard
+wrapt==2.0.0
     # via dm-tree
-xxhash==3.5.0
+xxhash==3.6.0
     # via datasets
-yarl==1.20.1
+yarl==1.22.0
     # via aiohttp
 zipp==3.23.0
-    # via importlib-metadata
+    # via
+    #   etils
+    #   importlib-metadata
 
 # The following packages are considered to be unsafe in a requirements file:
 # setuptools

requirements-ubuntu.txt

@@ -13,47 +13,62 @@ absl-py==2.3.1
     #   dm-tree
     #   labmaze
     #   mujoco
-accelerate==1.9.0
-    # via lerobot
+    #   tensorboard
+accelerate==1.11.0
+    # via
+    #   lerobot
+    #   peft
 aiohappyeyeballs==2.6.1
     # via aiohttp
-aiohttp==3.12.15
+aiohttp==3.13.1
     # via fsspec
 aiosignal==1.4.0
     # via aiohttp
 annotated-types==0.7.0
     # via pydantic
+antlr4-python3-runtime==4.9.3
+    # via
+    #   hydra-core
+    #   omegaconf
+anyio==4.11.0
+    # via
+    #   starlette
+    #   watchfiles
 asttokens==3.0.0
     # via stack-data
 async-timeout==5.0.1
     # via aiohttp
-attrs==25.3.0
+attrs==25.4.0
     # via
     #   aiohttp
     #   dm-tree
     #   jsonlines
+    #   jsonschema
+    #   referencing
     #   rerun-sdk
-av==15.0.0
+av==15.1.0
     # via lerobot
-blinker==1.9.0
-    # via flask
-certifi==2025.7.14
+bddl==1.0.1
+    # via libero
+certifi==2025.10.5
     # via
     #   requests
     #   sentry-sdk
-cffi==1.17.1
+cffi==2.0.0
     # via pymunk
 cfgv==3.4.0
     # via pre-commit
-charset-normalizer==3.4.2
+charset-normalizer==3.4.4
     # via requests
-click==8.2.1
+click==8.3.0
     # via
-    #   flask
+    #   uvicorn
     #   wandb
 cloudpickle==3.1.1
-    # via gymnasium
-cmake==4.0.3
+    # via
+    #   gymnasium
+    #   libero
+cmake==4.1.0
     # via lerobot
 cmeel==0.57.3
     # via
@@ -95,27 +110,29 @@ coal-library==3.0.1
     # via pin
 contourpy==1.3.2
     # via matplotlib
-coverage[toml]==7.10.1
+coverage[toml]==7.11.0
     # via pytest-cov
 cycler==0.12.1
     # via matplotlib
-datasets==3.6.0
+datasets==4.1.1
     # via lerobot
-debugpy==1.8.15
+debugpy==1.8.17
     # via lerobot
 decorator==5.2.1
     # via ipython
-deepdiff==8.5.0
+decord==0.6.0
     # via lerobot
-diffusers==0.34.0
+deepdiff==8.6.1
     # via lerobot
-dill==0.3.8
+diffusers==0.35.2
+    # via lerobot
+dill==0.4.0
     # via
     #   datasets
     #   multiprocess
 distlib==0.4.0
     # via virtualenv
-dm-control==1.0.14
+dm-control==1.0.34
     # via gym-aloha
 dm-env==1.6
     # via dm-control
@@ -123,31 +140,48 @@ dm-tree==0.1.9
     # via
     #   dm-control
     #   dm-env
+    #   lerobot
 docopt==0.6.2
     # via num2words
 draccus==0.10.0
     # via lerobot
-dynamixel-sdk==3.7.31
+dynamixel-sdk==3.8.4
     # via lerobot
+easydict==1.13
+    # via libero
+egl-probe @ git+https://github.com/huggingface/egl_probe.git
+    # via
+    #   libero
+    #   robomimic
 eigenpy==3.10.3
     # via coal-library
 einops==0.8.1
-    # via lerobot
+    # via
+    #   flash-attn
+    #   lerobot
+    #   libero
 eiquadprog==1.2.9
     # via placo
+etils[epath,epy]==1.13.0
+    # via mujoco
 evdev==1.9.2
     # via pynput
 exceptiongroup==1.3.0
     # via
+    #   anyio
     #   ipython
     #   pytest
-executing==2.2.0
+executing==2.2.1
     # via stack-data
 farama-notifications==0.0.4
     # via gymnasium
+fastapi==0.119.1
+    # via teleop
+fastjsonschema==2.21.2
+    # via nbformat
 feetech-servo-sdk==1.0.0
     # via lerobot
-filelock==3.18.0
+filelock==3.20.0
     # via
     #   datasets
     #   diffusers
@@ -155,24 +189,27 @@ filelock==3.18.0
     #   torch
     #   transformers
     #   virtualenv
-flask==3.1.1
+flash-attn==2.8.3
     # via lerobot
-fonttools==4.59.0
+fonttools==4.60.1
     # via matplotlib
-frozenlist==1.7.0
+frozenlist==1.8.0
     # via
     #   aiohttp
     #   aiosignal
-fsspec[http]==2025.3.0
+fsspec[http]==2025.9.0
     # via
     #   datasets
+    #   etils
     #   huggingface-hub
     #   torch
+future==1.0.0
+    # via libero
 gitdb==4.0.12
     # via gitpython
 gitpython==3.1.45
     # via wandb
-glfw==2.9.0
+glfw==2.10.0
     # via
     #   dm-control
     #   mujoco
@@ -180,61 +217,79 @@ grpcio==1.73.1
     # via
     #   grpcio-tools
     #   lerobot
+    #   reachy2-sdk
+    #   reachy2-sdk-api
+    #   tensorboard
 grpcio-tools==1.73.1
+    # via
+    #   lerobot
+    #   reachy2-sdk-api
+gym-aloha==0.1.3
     # via lerobot
-gym-aloha==0.1.1
+gym-hil==0.1.13
     # via lerobot
-gym-hil==0.1.10
+gym-pusht==0.1.6
     # via lerobot
-gym-pusht==0.1.5
-    # via lerobot
-gym-xarm==0.1.1
-    # via lerobot
-gymnasium==0.29.1
+gymnasium==1.2.1
     # via
     #   gym-aloha
     #   gym-hil
     #   gym-pusht
-    #   gym-xarm
-    #   gymnasium-robotics
     #   lerobot
-    #   pettingzoo
-gymnasium-robotics==1.2.4
-    # via gym-xarm
+    #   libero
+    #   metaworld
+h11==0.16.0
+    # via uvicorn
+h5py==3.15.1
+    # via robomimic
+hebi-py==2.11.0
+    # via lerobot
 hf-transfer==0.1.9
     # via huggingface-hub
-hf-xet==1.1.5
+hf-xet==1.1.10
     # via huggingface-hub
 hidapi==0.14.0.post4
     # via
     #   gym-hil
     #   lerobot
-huggingface-hub[cli,hf-transfer]==0.34.3
+httptools==0.7.1
+    # via uvicorn
+huggingface-hub[cli,hf-transfer]==0.35.3
     # via
     #   accelerate
     #   datasets
     #   diffusers
     #   lerobot
+    #   peft
+    #   timm
     #   tokenizers
     #   transformers
-identify==2.6.12
+hydra-core==1.3.2
+    # via libero
+identify==2.6.15
     # via pre-commit
-idna==3.10
+idna==3.11
     # via
+    #   anyio
     #   requests
     #   yarl
 imageio[ffmpeg]==2.37.0
     # via
     #   gym-aloha
     #   gym-hil
-    #   gymnasium-robotics
     #   lerobot
+    #   metaworld
+    #   robomimic
     #   scikit-image
 imageio-ffmpeg==0.6.0
-    # via imageio
+    # via
+    #   imageio
+    #   robomimic
 importlib-metadata==8.7.0
     # via diffusers
-iniconfig==2.1.0
+importlib-resources==6.5.2
+    # via etils
+iniconfig==2.3.0
     # via pytest
 inquirerpy==0.3.4
     # via huggingface-hub
@@ -242,50 +297,71 @@ ipython==8.37.0
     # via meshcat
 ischedule==1.2.7
     # via placo
-itsdangerous==2.2.0
-    # via flask
 jedi==0.19.2
     # via ipython
 jinja2==3.1.6
-    # via
-    #   flask
-    #   gymnasium-robotics
-    #   torch
+    # via torch
 jsonlines==4.0.0
     # via lerobot
-kiwisolver==1.4.8
+jsonschema==4.25.1
+    # via nbformat
+jsonschema-specifications==2025.9.1
+    # via jsonschema
+jupyter-core==5.9.1
+    # via nbformat
+jupytext==1.18.1
+    # via bddl
+kiwisolver==1.4.9
     # via matplotlib
 labmaze==1.0.6
     # via dm-control
 lazy-loader==0.4
     # via scikit-image
-lxml==6.0.0
+libero @ git+https://github.com/huggingface/lerobot-libero.git@main
+    # via lerobot
+llvmlite==0.45.1
+    # via numba
+lxml==6.0.2
     # via dm-control
-markupsafe==3.0.2
+markdown==3.9
+    # via tensorboard
+markdown-it-py==4.0.0
+    # via
+    #   jupytext
+    #   mdit-py-plugins
+markupsafe==3.0.3
     # via
-    #   flask
     #   jinja2
     #   werkzeug
-matplotlib==3.10.5
-    # via lerobot
-matplotlib-inline==0.1.7
+matplotlib==3.10.7
+    # via
+    #   lerobot
+    #   libero
+matplotlib-inline==0.2.1
     # via ipython
+mdit-py-plugins==0.5.0
+    # via jupytext
+mdurl==0.1.2
+    # via markdown-it-py
 mergedeep==1.3.4
     # via draccus
 meshcat==0.3.2
     # via placo
+metaworld==3.0.0
+    # via lerobot
 mock-serial==0.0.1
     # via lerobot
 mpmath==1.3.0
     # via sympy
-mujoco==2.3.7
+mujoco==3.3.7
     # via
     #   dm-control
     #   gym-aloha
     #   gym-hil
-    #   gym-xarm
-    #   gymnasium-robotics
-multidict==6.6.3
+    #   libero
+    #   metaworld
+    #   robosuite
+multidict==6.7.0
     # via
     #   aiohttp
     #   yarl
@@ -293,42 +369,63 @@ multiprocess==0.70.16
     # via datasets
 mypy-extensions==1.1.0
     # via typing-inspect
+nbformat==5.10.4
+    # via jupytext
 networkx==3.4.2
     # via
+    #   bddl
     #   scikit-image
     #   torch
+ninja==1.13.0
+    # via lerobot
 nodeenv==1.9.1
     # via pre-commit
 num2words==0.5.14
     # via lerobot
+numba==0.62.1
+    # via robosuite
 numpy==2.2.6
     # via
     #   accelerate
+    #   bddl
     #   cmeel-boost
     #   contourpy
     #   datasets
+    #   decord
     #   diffusers
     #   dm-control
     #   dm-env
     #   dm-tree
     #   gymnasium
-    #   gymnasium-robotics
+    #   h5py
+    #   hebi-py
     #   imageio
     #   labmaze
+    #   libero
     #   matplotlib
     #   meshcat
+    #   metaworld
     #   mujoco
+    #   numba
     #   opencv-python
     #   opencv-python-headless
     #   pandas
-    #   pettingzoo
+    #   peft
+    #   pyquaternion
+    #   reachy2-sdk
     #   rerun-sdk
+    #   robomimic
+    #   robosuite
     #   scikit-image
     #   scipy
     #   shapely
+    #   teleop
+    #   tensorboard
+    #   tensorboardx
     #   tifffile
     #   torchvision
     #   transformers
+    #   transforms3d
 nvidia-cublas-cu12==12.6.4.1
     # via
     #   nvidia-cudnn-cu12
@@ -366,8 +463,14 @@ nvidia-nvjitlink-cu12==12.6.85
     #   torch
 nvidia-nvtx-cu12==12.6.77
     # via torch
+omegaconf==2.3.0
+    # via hydra-core
 opencv-python==4.12.0.88
-    # via gym-pusht
+    # via
+    #   gym-pusht
+    #   libero
+    #   reachy2-sdk
+    #   robosuite
 opencv-python-headless==4.12.0.88
     # via lerobot
 orderly-set==5.5.0
@@ -377,53 +480,63 @@ packaging==25.0
     #   accelerate
     #   datasets
     #   huggingface-hub
+    #   hydra-core
+    #   jupytext
     #   lazy-loader
     #   lerobot
     #   matplotlib
+    #   peft
     #   pytest
+    #   reachy2-sdk
     #   scikit-image
+    #   tensorboard
+    #   tensorboardx
     #   transformers
     #   wandb
-pandas==2.3.1
+pandas==2.3.3
     # via
     #   datasets
     #   lerobot
-parso==0.8.4
+parso==0.8.5
     # via jedi
-pettingzoo==1.24.3
-    # via gymnasium-robotics
+peft==0.17.1
+    # via lerobot
 pexpect==4.9.0
     # via ipython
 pfzy==0.3.4
     # via inquirerpy
-pillow==11.3.0
+pillow==12.0.0
     # via
     #   diffusers
     #   imageio
+    #   lerobot
     #   matplotlib
     #   meshcat
     #   rerun-sdk
+    #   robosuite
     #   scikit-image
+    #   tensorboard
     #   torchvision
 pin==3.4.0
     # via placo
 placo==0.9.14
     # via lerobot
-platformdirs==4.3.8
+platformdirs==4.5.0
     # via
+    #   jupyter-core
     #   virtualenv
     #   wandb
 pluggy==1.6.0
     # via
     #   pytest
     #   pytest-cov
-pre-commit==4.2.0
+pre-commit==4.3.0
     # via lerobot
-prompt-toolkit==3.0.51
+prompt-toolkit==3.0.52
     # via
     #   inquirerpy
     #   ipython
-propcache==0.3.2
+propcache==0.4.1
     # via
     #   aiohttp
     #   yarl
@@ -432,11 +545,17 @@ protobuf==6.31.0
     #   dm-control
     #   grpcio-tools
     #   lerobot
+    #   reachy2-sdk
+    #   reachy2-sdk-api
+    #   tensorboard
+    #   tensorboardx
     #   wandb
-psutil==7.0.0
+psutil==7.1.1
     # via
     #   accelerate
     #   imageio
+    #   peft
+    #   robomimic
 ptyprocess==0.7.0
     # via pexpect
 pure-eval==0.2.3
@@ -445,11 +564,13 @@ pyarrow==21.0.0
     # via
     #   datasets
     #   rerun-sdk
-pycparser==2.22
+pycparser==2.23
     # via cffi
-pydantic==2.11.7
-    # via wandb
-pydantic-core==2.33.2
+pydantic==2.12.3
+    # via
+    #   fastapi
+    #   wandb
+pydantic-core==2.41.4
     # via pydantic
 pygame==2.6.1
     # via
@@ -464,20 +585,22 @@ pymunk==6.11.1
     # via
     #   gym-pusht
     #   lerobot
-pyngrok==7.2.12
+pyngrok==7.4.1
     # via meshcat
 pynput==1.8.1
     # via
     #   gym-hil
     #   lerobot
-pyopengl==3.1.9
+pyopengl==3.1.10
     # via
     #   dm-control
     #   mujoco
-pyparsing==3.2.3
+pyparsing==3.2.5
     # via
     #   dm-control
     #   matplotlib
+pyquaternion==0.9.9
+    # via reachy2-sdk
 pyrealsense2==2.56.5.9235
     # via lerobot
 pyserial==3.5
@@ -485,12 +608,14 @@ pyserial==3.5
     #   dynamixel-sdk
     #   feetech-servo-sdk
     #   lerobot
-pytest==8.4.1
+pytest==8.4.2
     # via
+    #   bddl
     #   lerobot
     #   pytest-cov
     #   pytest-timeout
-pytest-cov==6.2.1
+    #   teleop
+pytest-cov==7.0.0
     # via lerobot
 pytest-timeout==2.4.0
     # via lerobot
@@ -498,48 +623,75 @@ python-dateutil==2.9.0.post0
     # via
     #   matplotlib
     #   pandas
+python-dotenv==1.1.1
+    # via uvicorn
 python-xlib==0.33
     # via pynput
 pytz==2025.2
     # via pandas
-pyyaml==6.0.2
+pyyaml==6.0.3
     # via
     #   accelerate
     #   datasets
     #   draccus
+    #   hebi-py
     #   huggingface-hub
+    #   jupytext
+    #   omegaconf
+    #   peft
     #   pre-commit
     #   pyngrok
     #   pyyaml-include
+    #   timm
     #   transformers
+    #   uvicorn
     #   wandb
 pyyaml-include==1.4.1
     # via draccus
-pyzmq==27.0.0
+pyzmq==27.1.0
     # via
     #   lerobot
     #   meshcat
-regex==2025.7.34
+reachy2-sdk==1.0.14
+    # via lerobot
+reachy2-sdk-api==1.0.21
+    # via reachy2-sdk
+referencing==0.37.0
+    # via
+    #   jsonschema
+    #   jsonschema-specifications
+regex==2025.10.23
     # via
     #   diffusers
     #   transformers
-requests==2.32.4
+requests==2.32.5
     # via
     #   datasets
     #   diffusers
     #   dm-control
     #   huggingface-hub
+    #   teleop
     #   transformers
     #   wandb
-rerun-sdk==0.22.1
+rerun-sdk==0.26.1
     # via lerobot
 rhoban-cmeel-jsoncpp==1.9.4.9
     # via placo
-safetensors==0.5.3
+robomimic==0.2.0
+    # via libero
+robosuite==1.4.0
+    # via libero
+rpds-py==0.28.0
+    # via
+    #   jsonschema
+    #   referencing
+safetensors==0.6.2
     # via
     #   accelerate
     #   diffusers
     #   lerobot
+    #   peft
+    #   timm
     #   transformers
 scikit-image==0.25.2
     # via
@@ -548,10 +700,12 @@ scikit-image==0.25.2
 scipy==1.15.3
     # via
     #   dm-control
+    #   metaworld
+    #   robosuite
     #   scikit-image
-sentry-sdk==2.34.1
+sentry-sdk==2.42.1
     # via wandb
-shapely==2.1.1
+shapely==2.1.2
     # via gym-pusht
 six==1.17.0
     # via
@@ -560,66 +714,109 @@ six==1.17.0
     #   python-xlib
 smmap==5.0.2
     # via gitdb
+sniffio==1.3.1
+    # via anyio
 stack-data==0.6.3
     # via ipython
+starlette==0.48.0
+    # via fastapi
 sympy==1.14.0
     # via torch
-termcolor==3.1.0
+teleop==0.1.2
     # via lerobot
+tensorboard==2.20.0
+    # via robomimic
+tensorboard-data-server==0.7.2
+    # via tensorboard
+tensorboardx==2.6.4
+    # via robomimic
+termcolor==3.1.0
+    # via
+    #   lerobot
+    #   robomimic
+thop==0.1.1.post2209072238
+    # via libero
 tifffile==2025.5.10
     # via scikit-image
-tokenizers==0.21.4
+timm==1.0.20
+    # via lerobot
+tokenizers==0.22.1
     # via transformers
 toml==0.10.2
     # via draccus
-tomli==2.2.1
+tomli==2.3.0
     # via
     #   cmeel
     #   coverage
+    #   jupytext
     #   pytest
 torch==2.7.1
     # via
     #   accelerate
+    #   flash-attn
     #   lerobot
+    #   peft
+    #   robomimic
+    #   thop
+    #   timm
     #   torchvision
 torchcodec==0.5
     # via lerobot
 torchvision==0.22.1
-    # via lerobot
-tornado==6.5.1
+    # via
+    #   lerobot
+    #   robomimic
+    #   timm
+tornado==6.5.2
     # via meshcat
 tqdm==4.67.1
     # via
     #   datasets
     #   dm-control
     #   huggingface-hub
+    #   peft
+    #   robomimic
     #   transformers
 traitlets==5.14.3
     # via
     #   ipython
+    #   jupyter-core
     #   matplotlib-inline
-transformers==4.51.3
-    # via lerobot
+    #   nbformat
+transformers==4.57.1
+    # via
+    #   lerobot
+    #   libero
+    #   peft
+transforms3d==0.4.2
+    # via teleop
 triton==3.3.1
     # via torch
-typing-extensions==4.14.1
+typing-extensions==4.15.0
     # via
     #   aiosignal
+    #   anyio
+    #   etils
     #   exceptiongroup
+    #   fastapi
     #   gymnasium
     #   huggingface-hub
     #   ipython
     #   multidict
     #   pydantic
     #   pydantic-core
+    #   referencing
     #   rerun-sdk
+    #   starlette
     #   torch
     #   typing-inspect
     #   typing-inspection
+    #   uvicorn
+    #   virtualenv
     #   wandb
 typing-inspect==0.9.0
     # via draccus
-typing-inspection==0.4.1
+typing-inspection==0.4.2
     # via pydantic
 tzdata==2025.2
     # via pandas
@@ -629,22 +826,36 @@ urllib3==2.5.0
     # via
     #   requests
     #   sentry-sdk
-virtualenv==20.32.0
+uvicorn[standard]==0.38.0
+    # via teleop
+uvloop==0.22.1
+    # via uvicorn
+virtualenv==20.35.3
     # via pre-commit
-wandb==0.21.0
-    # via lerobot
-wcwidth==0.2.13
+wandb==0.21.4
+    # via
+    #   lerobot
+    #   libero
+watchfiles==1.1.1
+    # via uvicorn
+wcwidth==0.2.14
     # via prompt-toolkit
+websocket-client==1.9.0
+    # via teleop
+websockets==15.0.1
+    # via uvicorn
 werkzeug==3.1.3
-    # via flask
-wrapt==1.17.2
+    # via tensorboard
+wrapt==2.0.0
     # via dm-tree
-xxhash==3.5.0
+xxhash==3.6.0
     # via datasets
-yarl==1.20.1
+yarl==1.22.0
     # via aiohttp
 zipp==3.23.0
-    # via importlib-metadata
+    # via
+    #   etils
+    #   importlib-metadata
 
 # The following packages are considered to be unsafe in a requirements file:
 # setuptools

requirements.in

@@ -1,9 +1,9 @@
 # requirements.in
 
-# requirements-macos.txt was generated on macOS and is platform-specific (macOS 15.5 24F74 arm64).
-# Darwin MacBook-Pro.local 24.5.0 Darwin Kernel Version 24.5.0: Tue Apr 22 19:54:43 PDT 2025; root:xnu-11417.121.6~2/RELEASE_ARM64_T8132 arm64
+# requirements-macos.txt was generated on macOS and is platform-specific (macOS 26.0.1 25A362 arm64).
+# Darwin MacBook-Pro.local 25.0.0 Darwin Kernel Version 25.0.0: Wed Sep 17 21:42:08 PDT 2025; root:xnu-12377.1.9~141/RELEASE_ARM64_T8132 arm64
 
-# requirements-ubuntu.txt was generated on Linux and is platform-specific (Ubuntu 24.04.2 LTS x86_64).
-# Linux mlerobot-linux 6.14.0-27-generic #27~24.04.1-Ubuntu SMP PREEMPT_DYNAMIC Tue Jul 22 17:38:49 UTC 2 x86_64 x86_64 x86_64 GNU/Linux
+# requirements-ubuntu.txt was generated on Linux and is platform-specific (Ubuntu 24.04.3 LTS x86_64).
+# Linux mlerobot-linux 6.14.0-33-generic #33~24.04.1-Ubuntu SMP PREEMPT_DYNAMIC Fri Sep 19 17:02:30 UTC 2 x86_64 x86_64 x86_64 GNU/Linux
 
 -e .[all]

scripts/visualize_sarm_predictions.py

@@ -0,0 +1,761 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Inference script for SARM (Stage-Aware Reward Model).
+
+This script loads a trained SARM model and runs inference on a dataset episode,
+generating visualizations of the predicted task stages and progress over time.
+
+Example usage:
+    python scripts/visualize_sarm_predictions.py \
+        --model-id username/sarm-model \
+        --dataset-repo lerobot/aloha_sim_insertion_human \
+        --episode-index 0 \
+        --output-dir outputs/sarm_viz \
+        --task-description "insert the peg into the socket"
+"""
+
+import argparse
+import json
+import logging
+from pathlib import Path
+from typing import Optional
+
+import matplotlib.pyplot as plt
+import matplotlib.gridspec as gridspec
+import matplotlib.patches as mpatches
+import numpy as np
+import pandas as pd
+import torch
+from tqdm import tqdm
+
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.policies.sarm.modeling_sarm import SARMRewardModel
+from lerobot.policies.sarm.sarm_utils import (
+    pad_state_to_max_dim,
+    compute_tau,
+    compute_cumulative_progress_batch,
+)
+from lerobot.datasets.utils import load_stats
+
+
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description="Run SARM inference and visualize predictions")
+    
+    # Model arguments
+    parser.add_argument(
+        "--model-id",
+        type=str,
+        required=True,
+        help="HuggingFace model ID or local path to trained SARM model"
+    )
+    
+    # Dataset arguments
+    parser.add_argument(
+        "--dataset-repo",
+        type=str,
+        required=True,
+        help="HuggingFace dataset repository ID (e.g., lerobot/aloha_sim_insertion_human)"
+    )
+    parser.add_argument(
+        "--episode-index",
+        type=int,
+        default=0,
+        help="Index of the episode to visualize (default: 0)"
+    )
+    parser.add_argument(
+        "--task-description",
+        type=str,
+        default="perform the task",
+        help="Task description for the reward model (default: 'perform the task')"
+    )
+    
+    # Output arguments
+    parser.add_argument(
+        "--output-dir",
+        type=str,
+        default="outputs/sarm_inference",
+        help="Directory to save visualization outputs (default: outputs/sarm_inference)"
+    )
+    parser.add_argument(
+        "--image-key",
+        type=str,
+        default=None,
+        help="Key for images in dataset (e.g., observation.images.image). If not specified, uses model config's image_key"
+    )
+    parser.add_argument(
+        "--state-key",
+        type=str,
+        default=None,
+        help="Key for joint states in dataset. If None, auto-detects from dataset"
+    )
+    
+    # Visualization options
+    parser.add_argument(
+        "--show-frames",
+        action="store_true",
+        help="Include sample frames in the visualization"
+    )
+    parser.add_argument(
+        "--num-sample-frames",
+        type=int,
+        default=8,
+        help="Number of sample frames to show (default: 8)"
+    )
+    parser.add_argument(
+        "--figsize",
+        type=int,
+        nargs=2,
+        default=[14, 8],
+        help="Figure size as width height (default: 14 8)"
+    )
+    
+    # Device
+    parser.add_argument(
+        "--device",
+        type=str,
+        default=None,
+        help="Device to run inference on (cuda/cpu, default: auto-detect)"
+    )
+    
+    return parser.parse_args()
+
+
+def load_episode_data(
+    dataset: LeRobotDataset,
+    episode_index: int,
+    image_key: str,
+    state_key: str | None = None
+) -> tuple[np.ndarray, np.ndarray, int, int, str]:
+    """
+    Load all frames and states from a specific episode.
+    
+    Args:
+        dataset: LeRobotDataset instance
+        episode_index: Index of the episode to load
+        image_key: Key for accessing images in the dataset
+        state_key: Key for accessing joint states (auto-detected if None)
+        
+    Returns:
+        Tuple of (frames, states, start_index, end_index, task_description)
+    """
+    # Get episode boundaries
+    episode_data = dataset.meta.episodes
+    start_idx = episode_data["dataset_from_index"][episode_index]
+    end_idx = episode_data["dataset_to_index"][episode_index]
+    
+    logger.info(f"Loading episode {episode_index}: frames {start_idx} to {end_idx} ({end_idx - start_idx} frames)")
+    
+    # Auto-detect state key if not provided
+    if state_key is None:
+        first_item = dataset[start_idx]
+        state_keys = [k for k in first_item.keys() if 'state' in k.lower() or 'qpos' in k.lower()]
+        if state_keys:
+            state_key = state_keys[0]
+            logger.info(f"Auto-detected state key: {state_key}")
+    
+    # Get task description from the dataset if available
+    task_description = None
+    first_item = dataset[start_idx]
+    if "task" in first_item:
+        task_description = first_item["task"]
+        logger.info(f"✓ Extracted task from episode {episode_index}: '{task_description}'")
+    
+    # Load all frames and states from the episode
+    frames = []
+    states = []
+    for idx in tqdm(range(start_idx, end_idx), desc="Loading frames"):
+        item = dataset[idx]
+        
+        # Get image
+        img = item[image_key]
+        
+        # Convert to numpy if needed
+        if isinstance(img, torch.Tensor):
+            img = img.cpu().numpy()
+        
+        # Handle different image formats (C, H, W) or (H, W, C)
+        if img.shape[0] in [1, 3]:  # Channel first
+            img = np.transpose(img, (1, 2, 0))
+        
+        # Convert to uint8 if needed
+        if img.dtype != np.uint8:
+            if img.max() <= 1.0:
+                img = (img * 255).astype(np.uint8)
+            else:
+                img = img.astype(np.uint8)
+        
+        frames.append(img)
+        
+        # Get state if available
+        if state_key and state_key in item:
+            state = item[state_key]
+            if isinstance(state, torch.Tensor):
+                state = state.cpu().numpy()
+            states.append(state)
+    
+    frames = np.array(frames)
+    states = np.array(states) if states else None
+    logger.info(f"Loaded {len(frames)} frames with shape {frames[0].shape}")
+    if states is not None:
+        logger.info(f"Loaded states with shape {states.shape}")
+    
+    return frames, states, start_idx, end_idx, task_description
+
+
+@torch.no_grad()
+def run_inference(
+    model: SARMRewardModel,
+    frames: np.ndarray,
+    states: Optional[np.ndarray],
+    task_description: str,
+    dataset_stats: dict | None = None,
+    state_key: str = "observation.state",
+    batch_size: int = 32
+) -> tuple[np.ndarray, np.ndarray]:
+    """
+    Run SARM inference on video frames and joint states.
+    
+    (per SARM paper Section A.4):
+    - Frame 0: Initial frame of the episode (frame 0)
+    - Frames 1-8: 8 consecutive frames with frame_gap spacing ending at current frame t
+    Pattern: [frame_0, t-(7*gap), t-(6*gap), ..., t-gap, t]
+    
+    Args:
+        model: SARM model
+        frames: Video frames (num_frames, H, W, C) - all frames from ONE episode
+        states: Joint states (num_frames, state_dim)
+        task_description: Task description text
+        dataset_stats: Dataset statistics for state normalization (same as training)
+        state_key: Key for state in dataset_stats
+        batch_size: Batch size for processing slices
+        
+    Returns:
+        Tuple of (progress_predictions, stage_predictions)
+            - progress_predictions: (num_frames,)
+            - stage_predictions: (num_frames, num_stages)
+    """
+    logger.info("Encoding video frames with CLIP...")
+    video_embeddings = model.encode_images(frames)
+    
+    logger.info("Encoding task description with CLIP...")
+    text_embedding = model.encode_text(task_description)
+    
+    # Get config values
+    num_frames_model = model.config.num_frames  # 9
+    frame_gap = model.config.frame_gap  # 30
+    
+    logger.info("Creating video slices (SARM paper: initial frame + 8 consecutive)...")
+    
+    # Convert to tensors
+    video_embeddings = torch.tensor(video_embeddings, dtype=torch.float32)
+    text_embedding = torch.tensor(text_embedding, dtype=torch.float32)
+    if states is not None:
+        state_embeddings = torch.tensor(states, dtype=torch.float32)
+        
+        # Normalize states using dataset stats (same as training processor)
+        if dataset_stats is not None and state_key in dataset_stats:
+            mean = torch.tensor(dataset_stats[state_key]["mean"], dtype=torch.float32)
+            std = torch.tensor(dataset_stats[state_key]["std"], dtype=torch.float32)
+            state_embeddings = (state_embeddings - mean) / (std + 1e-8)
+            logger.info(f"✓ Applied MEAN_STD normalization to states using {state_key}")
+        else:
+            logger.warning("⚠ No dataset_stats provided - states not normalized (may differ from training)")
+    else:
+        state_embeddings = None
+    
+    video_slices = []
+    state_slices = []
+    
+    for current_frame in tqdm(range(len(video_embeddings)), desc="Creating slices"):
+        # Compute frame indices using symmetric bidirectional pattern:
+        # [initial (0), t-4*gap, t-3*gap, t-2*gap, t-gap, t, t+gap, t+2*gap, t+3*gap]
+        # Boundary handling: clamp to [0, last_valid]
+        deltas = model.config.observation_delta_indices
+        last_valid = len(video_embeddings) - 1
+        
+        frame_indices = []
+        for delta in deltas:
+            idx = current_frame + delta
+            idx = max(0, min(idx, last_valid))  # Clamp to valid range
+            frame_indices.append(idx)
+
+        video_slice = video_embeddings[frame_indices]
+        video_slices.append(video_slice)
+        
+        if state_embeddings is not None:
+            state_slice = state_embeddings[frame_indices]
+            state_slices.append(state_slice)
+    
+    video_slices = torch.stack(video_slices)  # (num_frames, num_frames_model, 512)
+    if state_embeddings is not None:
+        state_slices = torch.stack(state_slices)  # (num_frames, num_frames_model, state_dim)
+        # Pad states to max_state_dim (same as training processor)
+        state_slices = pad_state_to_max_dim(state_slices, model.config.max_state_dim)
+    else:
+        state_slices = None
+    
+    logger.info("Running SARM inference on all slices...")
+    # Process in batches
+    all_progress = []
+    all_stages = []
+    
+    for i in tqdm(range(0, len(video_slices), batch_size), desc="Inference"):
+        batch_video = video_slices[i:i + batch_size].to(model.device)
+        batch_states = state_slices[i:i + batch_size].to(model.device) if state_slices is not None else None
+        batch_size_actual = batch_video.shape[0]
+        
+        # Replicate text embedding for batch
+        batch_text = text_embedding.unsqueeze(0).repeat(batch_size_actual, 1).to(model.device)
+        
+        # Get predictions
+        stage_logits, stage_probs, progress_preds = model.sarm_transformer(
+            batch_video, batch_text, batch_states
+        )
+        
+        # Extract predictions at the "current frame" position
+        # With symmetric pattern [initial, t-4g, t-3g, t-2g, t-g, t, t+g, t+2g, t+3g],
+        # the current frame is at position 5 (0-indexed)
+        current_frame_idx = 5
+        batch_progress = progress_preds[:, current_frame_idx, 0].cpu().numpy()
+        batch_stages = stage_probs[:, current_frame_idx, :].cpu().numpy()
+        
+        all_progress.extend(batch_progress)
+        all_stages.extend(batch_stages)
+    
+    return np.array(all_progress), np.array(all_stages)
+
+
+def compute_ground_truth_progress(
+    dataset: LeRobotDataset,
+    episode_index: int,
+    temporal_proportions: dict[str, float],
+    subtask_names_ordered: list[str],
+) -> tuple[np.ndarray, np.ndarray] | tuple[None, None]:
+    """
+    Compute ground truth progress and stage labels for an episode using annotations.
+    
+    Uses SARM Paper Formula (2):
+        y_t = P_{k-1} + ᾱ_k × τ_t
+    
+    where:
+        - τ_t = (t - s_k) / (e_k - s_k) is within-subtask progress
+        - P_{k-1} is cumulative prior (sum of previous subtask proportions)
+        - ᾱ_k is the temporal proportion for subtask k
+    
+    Args:
+        dataset: LeRobotDataset instance
+        episode_index: Index of the episode
+        temporal_proportions: Dict mapping subtask name to proportion
+        subtask_names_ordered: Ordered list of subtask names (for consistent stage indexing)
+        
+    Returns:
+        Tuple of (ground_truth_progress, ground_truth_stages) arrays, or (None, None) if no annotations
+    """
+    # Load episode metadata
+    episodes_df = dataset.meta.episodes.to_pandas()
+    
+    # Check if annotations exist
+    if "subtask_names" not in episodes_df.columns:
+        logger.warning("No subtask_names column found in episodes metadata")
+        return None, None
+    
+    ep_subtask_names = episodes_df.loc[episode_index, "subtask_names"]
+    if ep_subtask_names is None or (isinstance(ep_subtask_names, float) and pd.isna(ep_subtask_names)):
+        logger.warning(f"No annotations found for episode {episode_index}")
+        return None, None
+    
+    subtask_start_frames = episodes_df.loc[episode_index, "subtask_start_frames"]
+    subtask_end_frames = episodes_df.loc[episode_index, "subtask_end_frames"]
+    
+    # Get episode boundaries
+    ep_start = dataset.meta.episodes["dataset_from_index"][episode_index]
+    ep_end = dataset.meta.episodes["dataset_to_index"][episode_index]
+    num_frames = ep_end - ep_start
+    
+    # Get temporal proportions as ordered list
+    temporal_proportions_list = [
+        temporal_proportions.get(name, 0.0) for name in subtask_names_ordered
+    ]
+    
+    logger.info(f"Computing ground truth for {num_frames} frames using {len(ep_subtask_names)} annotated subtasks")
+    logger.info(f"Subtask names in episode: {ep_subtask_names}")
+    logger.info(f"Subtask start frames: {subtask_start_frames}")
+    logger.info(f"Subtask end frames: {subtask_end_frames}")
+    logger.info(f"Temporal proportions (ordered): {dict(zip(subtask_names_ordered, temporal_proportions_list))}")
+    
+    # Compute ground truth for each frame
+    gt_progress = np.zeros(num_frames)
+    gt_stages = np.zeros(num_frames, dtype=np.int32)
+    
+    for frame_rel in range(num_frames):
+        # Find which subtask this frame belongs to
+        found = False
+        for j, (name, start_frame, end_frame) in enumerate(zip(ep_subtask_names, subtask_start_frames, subtask_end_frames)):
+            if frame_rel >= start_frame and frame_rel <= end_frame:
+                # Found the subtask - get its global index
+                stage_idx = subtask_names_ordered.index(name) if name in subtask_names_ordered else 0
+                
+                # Compute τ_t using utility function
+                tau = compute_tau(frame_rel, start_frame, end_frame)
+                
+                # Compute cumulative progress using utility function
+                progress = compute_cumulative_progress_batch(tau, stage_idx, temporal_proportions_list)
+                
+                gt_progress[frame_rel] = progress
+                gt_stages[frame_rel] = stage_idx
+                found = True
+                break
+        
+        if not found:
+            # Handle frames outside annotated subtasks
+            if frame_rel < subtask_start_frames[0]:
+                gt_progress[frame_rel] = 0.0
+                gt_stages[frame_rel] = 0
+            elif frame_rel > subtask_end_frames[-1]:
+                gt_progress[frame_rel] = 1.0
+                gt_stages[frame_rel] = len(subtask_names_ordered) - 1
+            else:
+                # Between subtasks - find previous subtask
+                for j in range(len(ep_subtask_names) - 1):
+                    if frame_rel > subtask_end_frames[j] and frame_rel < subtask_start_frames[j + 1]:
+                        name = ep_subtask_names[j]
+                        stage_idx = subtask_names_ordered.index(name) if name in subtask_names_ordered else j
+                        progress = compute_cumulative_progress_batch(1.0, stage_idx, temporal_proportions_list)
+                        gt_progress[frame_rel] = progress
+                        gt_stages[frame_rel] = stage_idx
+                        break
+    
+    logger.info(f"✓ Ground truth computed: final={gt_progress[-1]:.3f}, max={gt_progress.max():.3f}")
+    return gt_progress, gt_stages
+
+
+def visualize_predictions(
+    frames: np.ndarray,
+    progress_predictions: np.ndarray,
+    stage_predictions: np.ndarray,
+    task_description: str,
+    output_path: Path,
+    num_sample_frames: int = 8,
+    figsize: tuple = (14, 8),
+    subtask_names: list[str] | None = None,
+    temporal_proportions: dict[str, float] | None = None,
+    ground_truth_progress: np.ndarray | None = None,
+    ground_truth_stages: np.ndarray | None = None,
+):
+    """
+    Create visualization of SARM predictions with optional ground truth comparison.
+    
+    Args:
+        frames: Video frames (num_frames, H, W, C)
+        progress_predictions: Progress predictions (num_frames,)
+        stage_predictions: Stage probabilities (num_frames, num_stages)
+        task_description: Task description
+        output_path: Path to save the figure
+        num_sample_frames: Number of frames to show
+        figsize: Figure size (width, height)
+        subtask_names: Optional list of subtask names for labeling
+        temporal_proportions: Optional dict of temporal proportions for each subtask
+        ground_truth_progress: Optional ground truth progress array (num_frames,)
+        ground_truth_stages: Optional ground truth stage indices array (num_frames,)
+    """
+    num_stages = stage_predictions.shape[1]
+    stage_colors = plt.cm.tab10(np.linspace(0, 1, num_stages))
+    
+    # Use subtask names if available, otherwise use generic labels
+    if subtask_names is not None and len(subtask_names) == num_stages:
+        stage_labels = subtask_names
+    else:
+        stage_labels = [f'Stage {i+1}' for i in range(num_stages)]
+    
+    # Create figure with progress plot, stage plot, and sample frames
+    fig = plt.figure(figsize=(figsize[0], figsize[1] + 4))
+    gs = gridspec.GridSpec(3, 1, height_ratios=[2, 1, 1], hspace=0.3)
+    
+    ax_progress = fig.add_subplot(gs[0])
+    ax_stages = fig.add_subplot(gs[1], sharex=ax_progress)
+    ax_frames = fig.add_subplot(gs[2])
+    
+    frame_indices = np.arange(len(progress_predictions))
+    
+    # Plot 1: Progress over time
+    ax_progress.plot(frame_indices, progress_predictions, linewidth=2, color='#2E86AB', label='Predicted Progress')
+    ax_progress.fill_between(frame_indices, 0, progress_predictions, alpha=0.3, color='#2E86AB')
+    
+    # Plot ground truth if available
+    if ground_truth_progress is not None:
+        ax_progress.plot(frame_indices, ground_truth_progress, linewidth=2, color='#28A745', 
+                        linestyle='--', label='Ground Truth Progress')
+        ax_progress.fill_between(frame_indices, 0, ground_truth_progress, alpha=0.15, color='#28A745')
+    
+    ax_progress.axhline(y=1.0, color='gray', linestyle='--', alpha=0.5, linewidth=1)
+    ax_progress.set_ylabel('Task Progress', fontsize=12)
+    ax_progress.set_title(f'Task: "{task_description}"', fontsize=14, fontweight='bold')
+    ax_progress.grid(True, alpha=0.3)
+    ax_progress.set_ylim(-0.05, 1.1)
+    ax_progress.legend(loc='upper left')
+    
+    # Add statistics box
+    stats_text = (
+        f'Frames: {len(progress_predictions)}\n'
+        f'Final Progress: {progress_predictions[-1]:.3f}\n'
+        f'Max Progress: {progress_predictions.max():.3f}\n'
+        f'Mean Progress: {progress_predictions.mean():.3f}'
+    )
+    if ground_truth_progress is not None:
+        mse = np.mean((progress_predictions - ground_truth_progress) ** 2)
+        stats_text += f'\nMSE vs GT: {mse:.4f}'
+        stats_text += f'\nGT Final: {ground_truth_progress[-1]:.3f}'
+    
+    ax_progress.text(0.98, 0.02, stats_text, transform=ax_progress.transAxes,
+                    fontsize=10, verticalalignment='bottom', horizontalalignment='right',
+                    bbox=dict(boxstyle='round', facecolor='wheat', alpha=0.5))
+    
+    # Plot 2: Stage predictions (stacked area plot)
+    ax_stages.stackplot(frame_indices, *[stage_predictions[:, i] for i in range(num_stages)],
+                        colors=stage_colors, alpha=0.8, labels=stage_labels)
+    
+    # Plot ground truth stage as vertical bands or markers
+    if ground_truth_stages is not None:
+        # Find stage transition points in ground truth
+        stage_changes = np.where(np.diff(ground_truth_stages) != 0)[0] + 1
+        for change_idx in stage_changes:
+            ax_stages.axvline(x=change_idx, color='black', linestyle='-', alpha=0.7, linewidth=1.5)
+            ax_progress.axvline(x=change_idx, color='black', linestyle='-', alpha=0.3, linewidth=1)
+        
+        # Add small markers at bottom showing GT stage
+        gt_stage_normalized = ground_truth_stages / max(num_stages - 1, 1)
+        ax_stages.scatter(frame_indices[::30], np.zeros(len(frame_indices[::30])) + 0.02, 
+                         c=[stage_colors[s] for s in ground_truth_stages[::30]], 
+                         s=20, marker='|', alpha=0.8, label='GT Stage Markers')
+    
+    ax_stages.set_xlabel('Frame Index', fontsize=12)
+    ax_stages.set_ylabel('Stage Probability', fontsize=12)
+    ax_stages.set_ylim(0, 1)
+    ax_stages.grid(True, alpha=0.3)
+    
+    # Adjust legend based on number of stages and label lengths
+    if num_stages <= 5:
+        ax_stages.legend(loc='upper left', ncol=num_stages, fontsize=8)
+    else:
+        ax_stages.legend(loc='upper left', ncol=3, fontsize=7)
+    
+    # Add vertical lines and labels for expected stage transitions (if temporal proportions available)
+    if temporal_proportions is not None and subtask_names is not None:
+        cumulative_progress = 0.0
+        for i, name in enumerate(stage_labels):
+            if name in temporal_proportions:
+                # Find approximate frame where this stage should end
+                stage_end_progress = cumulative_progress + temporal_proportions[name]
+                
+                # Find frame index closest to this progress
+                progress_diffs = np.abs(progress_predictions - stage_end_progress)
+                stage_end_frame = np.argmin(progress_diffs)
+                
+                # Draw vertical line
+                ax_progress.axvline(x=stage_end_frame, color='gray', linestyle=':', alpha=0.5, linewidth=1)
+                ax_stages.axvline(x=stage_end_frame, color='gray', linestyle=':', alpha=0.5, linewidth=1)
+                
+                cumulative_progress = stage_end_progress
+    
+    # Plot 3: Sample frames (if requested)
+    frame_indices_to_show = np.linspace(0, len(frames) - 1, num_sample_frames, dtype=int)
+    
+    ax_frames.axis('off')
+    
+    # Create grid for frames
+    frame_height = frames[0].shape[0]
+    frame_width = frames[0].shape[1]
+    
+    combined_width = frame_width * num_sample_frames
+    combined_image = np.zeros((frame_height, combined_width, 3), dtype=np.uint8)
+    
+    for i, frame_idx in enumerate(frame_indices_to_show):
+        frame = frames[frame_idx]
+        if frame.shape[-1] == 1:
+            frame = np.repeat(frame, 3, axis=-1)
+        
+        # Add frame to combined image
+        x_start = i * frame_width
+        x_end = (i + 1) * frame_width
+        combined_image[:, x_start:x_end] = frame
+        
+        # Add frame number, progress, and stage
+        progress_val = progress_predictions[frame_idx]
+        stage_idx = np.argmax(stage_predictions[frame_idx])
+        stage_name = stage_labels[stage_idx] if stage_idx < len(stage_labels) else f'{stage_idx+1}'
+        
+        # Truncate long stage names for display
+        if len(stage_name) > 15:
+            stage_name = stage_name[:12] + '...'
+        
+        label = f'Frame {frame_idx}\nProg: {progress_val:.2f}\n{stage_name}'
+        
+        # Draw label on image
+        ax_frames.text(x_start + frame_width / 2, -10, label, 
+                      ha='center', va='top', fontsize=7, 
+                      bbox=dict(boxstyle='round', facecolor='white', alpha=0.7))
+    
+    ax_frames.imshow(combined_image)
+    ax_frames.set_title('Sample Frames', fontsize=12, pad=20)
+    
+    plt.tight_layout()
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    plt.savefig(output_path, dpi=150, bbox_inches='tight')
+    logger.info(f"Saved visualization to {output_path}")
+    
+    plt.close()
+
+
+def main():
+    args = parse_args()
+    
+    # Setup device
+    if args.device is None:
+        device = "cuda" if torch.cuda.is_available() else "cpu"
+    else:
+        device = args.device
+    logger.info(f"Using device: {device}")
+    
+    # Load model
+    logger.info(f"Loading SARM model from {args.model_id}...")
+    model = SARMRewardModel.from_pretrained(args.model_id)
+    model.to(device)
+    model.eval()
+    logger.info("Model loaded successfully")
+    
+    # Load dataset
+    logger.info(f"Loading dataset {args.dataset_repo}...")
+    dataset = LeRobotDataset(args.dataset_repo)
+    logger.info(f"Dataset loaded: {len(dataset.meta.episodes)} episodes, {len(dataset)} frames")
+    
+    # Validate episode index
+    if args.episode_index >= len(dataset.meta.episodes):
+        raise ValueError(
+            f"Episode index {args.episode_index} out of range. "
+            f"Dataset has {len(dataset.meta.episodes)} episodes."
+        )
+    
+    image_key = args.image_key if args.image_key is not None else model.config.image_key
+    state_key = args.state_key if args.state_key is not None else model.config.state_key
+    logger.info(f"Using image key: {image_key}")
+    logger.info(f"Using state key: {state_key}")
+    
+    # Load dataset stats for state normalization (same as training)
+    dataset_stats = load_stats(dataset.root)
+    if dataset_stats:
+        logger.info(f"✓ Loaded dataset stats from {dataset.root}")
+    else:
+        logger.warning("⚠ Could not load dataset stats - states will not be normalized")
+    
+    # Load episode data
+    frames, states, start_idx, end_idx, dataset_task = load_episode_data(
+        dataset, args.episode_index, image_key, state_key
+    )
+    
+    # Use task description from dataset if available, otherwise use command-line argument
+    task_description = dataset_task if dataset_task is not None else args.task_description
+    logger.info(f"Using task description: '{task_description}'")
+    
+    # Run inference
+    progress_predictions, stage_predictions = run_inference(
+        model, frames, states, task_description, 
+        dataset_stats=dataset_stats, state_key=state_key
+    )
+    
+    # Extract subtask names and temporal proportions from model config if available
+    subtask_names = None
+    temporal_proportions = None
+    
+    if hasattr(model.config, 'subtask_names') and model.config.subtask_names is not None:
+        subtask_names = model.config.subtask_names
+        logger.info(f"✓ Found {len(subtask_names)} subtask names in model config: {subtask_names}")
+    
+    # Try to load temporal proportions from model config
+    if hasattr(model.config, 'temporal_proportions') and model.config.temporal_proportions is not None:
+        temporal_proportions = {
+            name: prop for name, prop in zip(model.config.subtask_names, model.config.temporal_proportions)
+        }
+        logger.info(f"✓ Loaded temporal proportions from model config: {temporal_proportions}")
+    
+    # Fallback: try to load from dataset meta
+    if temporal_proportions is None:
+        proportions_path = dataset.root / "meta" / "temporal_proportions.json"
+        if proportions_path.exists():
+            with open(proportions_path, 'r') as f:
+                temporal_proportions = json.load(f)
+                logger.info(f"✓ Loaded temporal proportions from dataset: {temporal_proportions}")
+                
+                # Also extract subtask names from proportions if not already set
+                if subtask_names is None:
+                    subtask_names = sorted(temporal_proportions.keys())
+                    logger.info(f"✓ Extracted subtask names from proportions: {subtask_names}")
+    
+    # Compute ground truth progress if annotations are available
+    ground_truth_progress = None
+    ground_truth_stages = None
+    
+    if temporal_proportions is not None and subtask_names is not None:
+        logger.info("Attempting to compute ground truth progress from annotations...")
+        ground_truth_progress, ground_truth_stages = compute_ground_truth_progress(
+            dataset,
+            args.episode_index,
+            temporal_proportions,
+            subtask_names
+        )
+        if ground_truth_progress is None:
+            logger.warning("⚠ Ground truth not available - annotations may be missing for this episode")
+    else:
+        logger.warning("⚠ Cannot compute ground truth - temporal_proportions or subtask_names not available")
+    
+    output_dir = Path(args.output_dir)
+    output_path = output_dir / f"sarm_prediction_ep{args.episode_index}.png"
+    
+    visualize_predictions(
+        frames,
+        progress_predictions,
+        stage_predictions,
+        task_description,
+        output_path,
+        num_sample_frames=args.num_sample_frames,
+        figsize=tuple(args.figsize),
+        subtask_names=subtask_names,
+        temporal_proportions=temporal_proportions,
+        ground_truth_progress=ground_truth_progress,
+        ground_truth_stages=ground_truth_stages,
+    )
+    
+    predictions_path = output_dir / f"predictions_ep{args.episode_index}.npz"
+    save_dict = {
+        'progress': progress_predictions, 
+        'stages': stage_predictions
+    }
+    if ground_truth_progress is not None:
+        save_dict['gt_progress'] = ground_truth_progress
+        save_dict['gt_stages'] = ground_truth_stages
+    np.savez(predictions_path, **save_dict)
+    logger.info(f"Saved predictions to {predictions_path}")
+    logger.info(f"\nVisualization: {output_path}")
+
+
+if __name__ == "__main__":
+    main()
+

src/lerobot/__init__.py

@@ -57,7 +57,6 @@ available_tasks_per_env = {
         "AlohaTransferCube-v0",
     ],
     "pusht": ["PushT-v0"],
-    "xarm": ["XarmLift-v0"],
 }
 available_envs = list(available_tasks_per_env.keys())
 
@@ -75,16 +74,6 @@ available_datasets_per_env = {
     # TODO(alexander-soare): Add "lerobot/pusht_keypoints". Right now we can't because this is too tightly
     # coupled with tests.
     "pusht": ["lerobot/pusht", "lerobot/pusht_image"],
-    "xarm": [
-        "lerobot/xarm_lift_medium",
-        "lerobot/xarm_lift_medium_replay",
-        "lerobot/xarm_push_medium",
-        "lerobot/xarm_push_medium_replay",
-        "lerobot/xarm_lift_medium_image",
-        "lerobot/xarm_lift_medium_replay_image",
-        "lerobot/xarm_push_medium_image",
-        "lerobot/xarm_push_medium_replay_image",
-    ],
 }
 
 available_real_world_datasets = [
@@ -195,7 +184,6 @@ available_motors = [
 available_policies_per_env = {
     "aloha": ["act"],
     "pusht": ["diffusion", "vqbet"],
-    "xarm": ["tdmpc"],
     "koch_real": ["act_koch_real"],
     "aloha_real": ["act_aloha_real"],
 }

src/lerobot/async_inference/helpers.py

@@ -16,7 +16,7 @@ import logging
 import logging.handlers
 import os
 import time
-from dataclasses import dataclass
+from dataclasses import dataclass, field
 from pathlib import Path
 
 import torch
@@ -268,6 +268,7 @@ class RemotePolicyConfig:
     lerobot_features: dict[str, PolicyFeature]
     actions_per_chunk: int
     device: str = "cpu"
+    rename_map: dict[str, str] = field(default_factory=dict)
 
 
 def _compare_observation_states(obs1_state: torch.Tensor, obs2_state: torch.Tensor, atol: float) -> bool:

src/lerobot/async_inference/policy_server.py

@@ -159,7 +159,10 @@ class PolicyServer(services_pb2_grpc.AsyncInferenceServicer):
         self.preprocessor, self.postprocessor = make_pre_post_processors(
             self.policy.config,
             pretrained_path=policy_specs.pretrained_name_or_path,
-            preprocessor_overrides={"device_processor": device_override},
+            preprocessor_overrides={
+                "device_processor": device_override,
+                "rename_observations_processor": {"rename_map": policy_specs.rename_map},
+            },
             postprocessor_overrides={"device_processor": device_override},
         )
 

src/lerobot/cameras/camera.py

@@ -17,7 +17,7 @@
 import abc
 from typing import Any
 
-import numpy as np
+from numpy.typing import NDArray  # type: ignore  # TODO: add type stubs for numpy.typing
 
 from .configs import CameraConfig, ColorMode
 
@@ -89,7 +89,7 @@ class Camera(abc.ABC):
         pass
 
     @abc.abstractmethod
-    def read(self, color_mode: ColorMode | None = None) -> np.ndarray:
+    def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
         """Capture and return a single frame from the camera.
 
         Args:
@@ -102,7 +102,7 @@ class Camera(abc.ABC):
         pass
 
     @abc.abstractmethod
-    def async_read(self, timeout_ms: float = ...) -> np.ndarray:
+    def async_read(self, timeout_ms: float = ...) -> NDArray[Any]:
         """Asynchronously capture and return a single frame from the camera.
 
         Args:

src/lerobot/cameras/configs.py

@@ -18,7 +18,7 @@ import abc
 from dataclasses import dataclass
 from enum import Enum
 
-import draccus
+import draccus  # type: ignore  # TODO: add type stubs for draccus
 
 
 class ColorMode(str, Enum):
@@ -34,11 +34,11 @@ class Cv2Rotation(int, Enum):
 
 
 @dataclass(kw_only=True)
-class CameraConfig(draccus.ChoiceRegistry, abc.ABC):
+class CameraConfig(draccus.ChoiceRegistry, abc.ABC):  # type: ignore  # TODO: add type stubs for draccus
     fps: int | None = None
     width: int | None = None
     height: int | None = None
 
     @property
     def type(self) -> str:
-        return self.get_choice_name(self.__class__)
+        return str(self.get_choice_name(self.__class__))

src/lerobot/cameras/opencv/__init__.py

@@ -14,3 +14,5 @@
 
 from .camera_opencv import OpenCVCamera
 from .configuration_opencv import OpenCVCameraConfig
+
+__all__ = ["OpenCVCamera", "OpenCVCameraConfig"]

src/lerobot/cameras/opencv/camera_opencv.py

@@ -25,11 +25,12 @@ from pathlib import Path
 from threading import Event, Lock, Thread
 from typing import Any
 
+from numpy.typing import NDArray  # type: ignore  # TODO: add type stubs for numpy.typing
+
 # Fix MSMF hardware transform compatibility for Windows before importing cv2
 if platform.system() == "Windows" and "OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS" not in os.environ:
     os.environ["OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS"] = "0"
-import cv2
-import numpy as np
+import cv2  # type: ignore  # TODO: add type stubs for OpenCV
 
 from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
 
@@ -121,7 +122,7 @@ class OpenCVCamera(Camera):
         self.thread: Thread | None = None
         self.stop_event: Event | None = None
         self.frame_lock: Lock = Lock()
-        self.latest_frame: np.ndarray | None = None
+        self.latest_frame: NDArray[Any] | None = None
         self.new_frame_event: Event = Event()
 
         self.rotation: int | None = get_cv2_rotation(config.rotation)
@@ -140,7 +141,7 @@ class OpenCVCamera(Camera):
         """Checks if the camera is currently connected and opened."""
         return isinstance(self.videocapture, cv2.VideoCapture) and self.videocapture.isOpened()
 
-    def connect(self, warmup: bool = True):
+    def connect(self, warmup: bool = True) -> None:
         """
         Connects to the OpenCV camera specified in the configuration.
 
@@ -180,12 +181,14 @@ class OpenCVCamera(Camera):
 
     def _configure_capture_settings(self) -> None:
         """
-        Applies the specified FPS, width, and height settings to the connected camera.
+        Applies the specified FOURCC, FPS, width, and height settings to the connected camera.
 
         This method attempts to set the camera properties via OpenCV. It checks if
         the camera successfully applied the settings and raises an error if not.
+        FOURCC is set first (if specified) as it can affect the available FPS and resolution options.
 
         Args:
+            fourcc: The desired FOURCC code (e.g., "MJPG", "YUYV"). If None, auto-detect.
             fps: The desired frames per second. If None, the setting is skipped.
             width: The desired capture width. If None, the setting is skipped.
             height: The desired capture height. If None, the setting is skipped.
@@ -199,10 +202,11 @@ class OpenCVCamera(Camera):
         if not self.is_connected:
             raise DeviceNotConnectedError(f"Cannot configure settings for {self} as it is not connected.")
 
-        if self.fps is None:
-            self.fps = self.videocapture.get(cv2.CAP_PROP_FPS)
-        else:
-            self._validate_fps()
+        # Set FOURCC first (if specified) as it can affect available FPS/resolution options
+        if self.config.fourcc is not None:
+            self._validate_fourcc()
+        if self.videocapture is None:
+            raise DeviceNotConnectedError(f"{self} videocapture is not initialized")
 
         default_width = int(round(self.videocapture.get(cv2.CAP_PROP_FRAME_WIDTH)))
         default_height = int(round(self.videocapture.get(cv2.CAP_PROP_FRAME_HEIGHT)))
@@ -216,18 +220,56 @@ class OpenCVCamera(Camera):
         else:
             self._validate_width_and_height()
 
+        if self.fps is None:
+            self.fps = self.videocapture.get(cv2.CAP_PROP_FPS)
+        else:
+            self._validate_fps()
+
     def _validate_fps(self) -> None:
         """Validates and sets the camera's frames per second (FPS)."""
 
+        if self.videocapture is None:
+            raise DeviceNotConnectedError(f"{self} videocapture is not initialized")
+
+        if self.fps is None:
+            raise ValueError(f"{self} FPS is not set")
+
         success = self.videocapture.set(cv2.CAP_PROP_FPS, float(self.fps))
         actual_fps = self.videocapture.get(cv2.CAP_PROP_FPS)
         # Use math.isclose for robust float comparison
         if not success or not math.isclose(self.fps, actual_fps, rel_tol=1e-3):
             raise RuntimeError(f"{self} failed to set fps={self.fps} ({actual_fps=}).")
 
+    def _validate_fourcc(self) -> None:
+        """Validates and sets the camera's FOURCC code."""
+
+        fourcc_code = cv2.VideoWriter_fourcc(*self.config.fourcc)
+
+        if self.videocapture is None:
+            raise DeviceNotConnectedError(f"{self} videocapture is not initialized")
+
+        success = self.videocapture.set(cv2.CAP_PROP_FOURCC, fourcc_code)
+        actual_fourcc_code = self.videocapture.get(cv2.CAP_PROP_FOURCC)
+
+        # Convert actual FOURCC code back to string for comparison
+        actual_fourcc_code_int = int(actual_fourcc_code)
+        actual_fourcc = "".join([chr((actual_fourcc_code_int >> 8 * i) & 0xFF) for i in range(4)])
+
+        if not success or actual_fourcc != self.config.fourcc:
+            logger.warning(
+                f"{self} failed to set fourcc={self.config.fourcc} (actual={actual_fourcc}, success={success}). "
+                f"Continuing with default format."
+            )
+
     def _validate_width_and_height(self) -> None:
         """Validates and sets the camera's frame capture width and height."""
 
+        if self.videocapture is None:
+            raise DeviceNotConnectedError(f"{self} videocapture is not initialized")
+
+        if self.capture_width is None or self.capture_height is None:
+            raise ValueError(f"{self} capture_width or capture_height is not set")
+
         width_success = self.videocapture.set(cv2.CAP_PROP_FRAME_WIDTH, float(self.capture_width))
         height_success = self.videocapture.set(cv2.CAP_PROP_FRAME_HEIGHT, float(self.capture_height))
 
@@ -258,11 +300,12 @@ class OpenCVCamera(Camera):
         """
         found_cameras_info = []
 
+        targets_to_scan: list[str | int]
         if platform.system() == "Linux":
             possible_paths = sorted(Path("/dev").glob("video*"), key=lambda p: p.name)
             targets_to_scan = [str(p) for p in possible_paths]
         else:
-            targets_to_scan = list(range(MAX_OPENCV_INDEX))
+            targets_to_scan = [int(i) for i in range(MAX_OPENCV_INDEX)]
 
         for target in targets_to_scan:
             camera = cv2.VideoCapture(target)
@@ -271,6 +314,12 @@ class OpenCVCamera(Camera):
                 default_height = int(camera.get(cv2.CAP_PROP_FRAME_HEIGHT))
                 default_fps = camera.get(cv2.CAP_PROP_FPS)
                 default_format = camera.get(cv2.CAP_PROP_FORMAT)
+
+                # Get FOURCC code and convert to string
+                default_fourcc_code = camera.get(cv2.CAP_PROP_FOURCC)
+                default_fourcc_code_int = int(default_fourcc_code)
+                default_fourcc = "".join([chr((default_fourcc_code_int >> 8 * i) & 0xFF) for i in range(4)])
+
                 camera_info = {
                     "name": f"OpenCV Camera @ {target}",
                     "type": "OpenCV",
@@ -278,6 +327,7 @@ class OpenCVCamera(Camera):
                     "backend_api": camera.getBackendName(),
                     "default_stream_profile": {
                         "format": default_format,
+                        "fourcc": default_fourcc,
                         "width": default_width,
                         "height": default_height,
                         "fps": default_fps,
@@ -289,7 +339,7 @@ class OpenCVCamera(Camera):
 
         return found_cameras_info
 
-    def read(self, color_mode: ColorMode | None = None) -> np.ndarray:
+    def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
         """
         Reads a single frame synchronously from the camera.
 
@@ -317,6 +367,9 @@ class OpenCVCamera(Camera):
 
         start_time = time.perf_counter()
 
+        if self.videocapture is None:
+            raise DeviceNotConnectedError(f"{self} videocapture is not initialized")
+
         ret, frame = self.videocapture.read()
 
         if not ret or frame is None:
@@ -329,7 +382,7 @@ class OpenCVCamera(Camera):
 
         return processed_frame
 
-    def _postprocess_image(self, image: np.ndarray, color_mode: ColorMode | None = None) -> np.ndarray:
+    def _postprocess_image(self, image: NDArray[Any], color_mode: ColorMode | None = None) -> NDArray[Any]:
         """
         Applies color conversion, dimension validation, and rotation to a raw frame.
 
@@ -372,7 +425,7 @@ class OpenCVCamera(Camera):
 
         return processed_image
 
-    def _read_loop(self):
+    def _read_loop(self) -> None:
         """
         Internal loop run by the background thread for asynchronous reading.
 
@@ -383,6 +436,9 @@ class OpenCVCamera(Camera):
 
         Stops on DeviceNotConnectedError, logs other errors and continues.
         """
+        if self.stop_event is None:
+            raise RuntimeError(f"{self}: stop_event is not initialized before starting read loop.")
+
         while not self.stop_event.is_set():
             try:
                 color_image = self.read()
@@ -419,7 +475,7 @@ class OpenCVCamera(Camera):
         self.thread = None
         self.stop_event = None
 
-    def async_read(self, timeout_ms: float = 200) -> np.ndarray:
+    def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
         """
         Reads the latest available frame asynchronously.
 
@@ -462,7 +518,7 @@ class OpenCVCamera(Camera):
 
         return frame
 
-    def disconnect(self):
+    def disconnect(self) -> None:
         """
         Disconnects from the camera and cleans up resources.
 

src/lerobot/cameras/opencv/configuration_opencv.py

@@ -17,6 +17,8 @@ from pathlib import Path
 
 from ..configs import CameraConfig, ColorMode, Cv2Rotation
 
+__all__ = ["OpenCVCameraConfig", "ColorMode", "Cv2Rotation"]
+
 
 @CameraConfig.register_subclass("opencv")
 @dataclass
@@ -33,8 +35,9 @@ class OpenCVCameraConfig(CameraConfig):
     OpenCVCameraConfig(0, 30, 1280, 720)   # 1280x720 @ 30FPS
     OpenCVCameraConfig(/dev/video4, 60, 640, 480)   # 640x480 @ 60FPS
 
-    # Advanced configurations
-    OpenCVCameraConfig(128422271347, 30, 640, 480, rotation=Cv2Rotation.ROTATE_90)     # With 90° rotation
+    # Advanced configurations with FOURCC format
+    OpenCVCameraConfig(128422271347, 30, 640, 480, rotation=Cv2Rotation.ROTATE_90, fourcc="MJPG")     # With 90° rotation and MJPG format
+    OpenCVCameraConfig(0, 30, 1280, 720, fourcc="YUYV")     # With YUYV format
     ```
 
     Attributes:
@@ -46,17 +49,21 @@ class OpenCVCameraConfig(CameraConfig):
         color_mode: Color mode for image output (RGB or BGR). Defaults to RGB.
         rotation: Image rotation setting (0°, 90°, 180°, or 270°). Defaults to no rotation.
         warmup_s: Time reading frames before returning from connect (in seconds)
+        fourcc: FOURCC code for video format (e.g., "MJPG", "YUYV", "I420"). Defaults to None (auto-detect).
 
     Note:
         - Only 3-channel color output (RGB/BGR) is currently supported.
+        - FOURCC codes must be 4-character strings (e.g., "MJPG", "YUYV"). Some common FOUCC codes: https://learn.microsoft.com/en-us/windows/win32/medfound/video-fourccs#fourcc-constants
+        - Setting FOURCC can help achieve higher frame rates on some cameras.
     """
 
     index_or_path: int | Path
     color_mode: ColorMode = ColorMode.RGB
     rotation: Cv2Rotation = Cv2Rotation.NO_ROTATION
     warmup_s: int = 1
+    fourcc: str | None = None
 
-    def __post_init__(self):
+    def __post_init__(self) -> None:
         if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
             raise ValueError(
                 f"`color_mode` is expected to be {ColorMode.RGB.value} or {ColorMode.BGR.value}, but {self.color_mode} is provided."
@@ -71,3 +78,8 @@ class OpenCVCameraConfig(CameraConfig):
             raise ValueError(
                 f"`rotation` is expected to be in {(Cv2Rotation.NO_ROTATION, Cv2Rotation.ROTATE_90, Cv2Rotation.ROTATE_180, Cv2Rotation.ROTATE_270)}, but {self.rotation} is provided."
             )
+
+        if self.fourcc is not None and (not isinstance(self.fourcc, str) or len(self.fourcc) != 4):
+            raise ValueError(
+                f"`fourcc` must be a 4-character string (e.g., 'MJPG', 'YUYV'), but '{self.fourcc}' is provided."
+            )

src/lerobot/cameras/reachy2_camera/configuration_reachy2_camera.py

@@ -16,6 +16,8 @@ from dataclasses import dataclass
 
 from ..configs import CameraConfig, ColorMode
 
+__all__ = ["CameraConfig", "ColorMode", "Reachy2CameraConfig"]
+
 
 @CameraConfig.register_subclass("reachy2_camera")
 @dataclass
@@ -62,7 +64,7 @@ class Reachy2CameraConfig(CameraConfig):
     port: int = 50065
     # use_depth: bool = False
 
-    def __post_init__(self):
+    def __post_init__(self) -> None:
         if self.name not in ["teleop", "depth"]:
             raise ValueError(f"`name` is expected to be 'teleop' or 'depth', but {self.name} is provided.")
         if (self.name == "teleop" and self.image_type not in ["left", "right"]) or (

src/lerobot/cameras/reachy2_camera/reachy2_camera.py

@@ -23,13 +23,17 @@ import time
 from threading import Event, Lock, Thread
 from typing import Any
 
+from numpy.typing import NDArray  # type: ignore  # TODO: add type stubs for numpy.typing
+
 # Fix MSMF hardware transform compatibility for Windows before importing cv2
 if platform.system() == "Windows" and "OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS" not in os.environ:
     os.environ["OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS"] = "0"
-import cv2
-import numpy as np
-from reachy2_sdk.media.camera import CameraView
-from reachy2_sdk.media.camera_manager import CameraManager
+import cv2  # type: ignore  # TODO: add type stubs for OpenCV
+import numpy as np  # type: ignore  # TODO: add type stubs for numpy
+from reachy2_sdk.media.camera import CameraView  # type: ignore  # TODO: add type stubs for reachy2_sdk
+from reachy2_sdk.media.camera_manager import (  # type: ignore  # TODO: add type stubs for reachy2_sdk
+    CameraManager,
+)
 
 from lerobot.utils.errors import DeviceNotConnectedError
 
@@ -73,7 +77,7 @@ class Reachy2Camera(Camera):
         self.thread: Thread | None = None
         self.stop_event: Event | None = None
         self.frame_lock: Lock = Lock()
-        self.latest_frame: np.ndarray | None = None
+        self.latest_frame: NDArray[Any] | None = None
         self.new_frame_event: Event = Event()
 
     def __str__(self) -> str:
@@ -83,13 +87,17 @@ class Reachy2Camera(Camera):
     def is_connected(self) -> bool:
         """Checks if the camera is currently connected and opened."""
         if self.config.name == "teleop":
-            return self.cam_manager._grpc_connected and self.cam_manager.teleop if self.cam_manager else False
+            return bool(
+                self.cam_manager._grpc_connected and self.cam_manager.teleop if self.cam_manager else False
+            )
         elif self.config.name == "depth":
-            return self.cam_manager._grpc_connected and self.cam_manager.depth if self.cam_manager else False
+            return bool(
+                self.cam_manager._grpc_connected and self.cam_manager.depth if self.cam_manager else False
+            )
         else:
             raise ValueError(f"Invalid camera name '{self.config.name}'. Expected 'teleop' or 'depth'.")
 
-    def connect(self, warmup: bool = True):
+    def connect(self, warmup: bool = True) -> None:
         """
         Connects to the Reachy2 CameraManager as specified in the configuration.
         """
@@ -131,7 +139,7 @@ class Reachy2Camera(Camera):
         camera_manager.disconnect()
         return initialized_cameras
 
-    def read(self, color_mode: ColorMode | None = None) -> np.ndarray:
+    def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
         """
         Reads a single frame synchronously from the camera.
 
@@ -152,7 +160,7 @@ class Reachy2Camera(Camera):
 
         start_time = time.perf_counter()
 
-        frame = None
+        frame: NDArray[Any] = np.empty((0, 0, 3), dtype=np.uint8)
 
         if self.cam_manager is None:
             raise DeviceNotConnectedError(f"{self} is not connected.")
@@ -179,7 +187,7 @@ class Reachy2Camera(Camera):
 
         return frame
 
-    def _read_loop(self):
+    def _read_loop(self) -> None:
         """
         Internal loop run by the background thread for asynchronous reading.
 
@@ -190,6 +198,9 @@ class Reachy2Camera(Camera):
 
         Stops on DeviceNotConnectedError, logs other errors and continues.
         """
+        if self.stop_event is None:
+            raise RuntimeError(f"{self}: stop_event is not initialized before starting read loop.")
+
         while not self.stop_event.is_set():
             try:
                 color_image = self.read()
@@ -226,7 +237,7 @@ class Reachy2Camera(Camera):
         self.thread = None
         self.stop_event = None
 
-    def async_read(self, timeout_ms: float = 200) -> np.ndarray:
+    def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
         """
         Reads the latest available frame asynchronously.
 
@@ -269,7 +280,7 @@ class Reachy2Camera(Camera):
 
         return frame
 
-    def disconnect(self):
+    def disconnect(self) -> None:
         """
         Stops the background read thread (if running).
 

src/lerobot/cameras/realsense/camera_realsense.py

@@ -21,11 +21,12 @@ import time
 from threading import Event, Lock, Thread
 from typing import Any
 
-import cv2
-import numpy as np
+import cv2  # type: ignore  # TODO: add type stubs for OpenCV
+import numpy as np  # type: ignore  # TODO: add type stubs for numpy
+from numpy.typing import NDArray  # type: ignore  # TODO: add type stubs for numpy.typing
 
 try:
-    import pyrealsense2 as rs
+    import pyrealsense2 as rs  # type: ignore  # TODO: add type stubs for pyrealsense2
 except Exception as e:
     logging.info(f"Could not import realsense: {e}")
 
@@ -132,7 +133,7 @@ class RealSenseCamera(Camera):
         self.thread: Thread | None = None
         self.stop_event: Event | None = None
         self.frame_lock: Lock = Lock()
-        self.latest_frame: np.ndarray | None = None
+        self.latest_frame: NDArray[Any] | None = None
         self.new_frame_event: Event = Event()
 
         self.rotation: int | None = get_cv2_rotation(config.rotation)
@@ -150,7 +151,7 @@ class RealSenseCamera(Camera):
         """Checks if the camera pipeline is started and streams are active."""
         return self.rs_pipeline is not None and self.rs_profile is not None
 
-    def connect(self, warmup: bool = True):
+    def connect(self, warmup: bool = True) -> None:
         """
         Connects to the RealSense camera specified in the configuration.
 
@@ -264,7 +265,7 @@ class RealSenseCamera(Camera):
         serial_number = str(found_devices[0]["serial_number"])
         return serial_number
 
-    def _configure_rs_pipeline_config(self, rs_config):
+    def _configure_rs_pipeline_config(self, rs_config: Any) -> None:
         """Creates and configures the RealSense pipeline configuration object."""
         rs.config.enable_device(rs_config, self.serial_number)
 
@@ -293,6 +294,9 @@ class RealSenseCamera(Camera):
         if not self.is_connected:
             raise DeviceNotConnectedError(f"Cannot validate settings for {self} as it is not connected.")
 
+        if self.rs_profile is None:
+            raise RuntimeError(f"{self}: rs_profile must be initialized before use.")
+
         stream = self.rs_profile.get_stream(rs.stream.color).as_video_stream_profile()
 
         if self.fps is None:
@@ -308,7 +312,7 @@ class RealSenseCamera(Camera):
                 self.width, self.height = actual_width, actual_height
                 self.capture_width, self.capture_height = actual_width, actual_height
 
-    def read_depth(self, timeout_ms: int = 200) -> np.ndarray:
+    def read_depth(self, timeout_ms: int = 200) -> NDArray[Any]:
         """
         Reads a single frame (depth) synchronously from the camera.
 
@@ -336,6 +340,9 @@ class RealSenseCamera(Camera):
 
         start_time = time.perf_counter()
 
+        if self.rs_pipeline is None:
+            raise RuntimeError(f"{self}: rs_pipeline must be initialized before use.")
+
         ret, frame = self.rs_pipeline.try_wait_for_frames(timeout_ms=timeout_ms)
 
         if not ret or frame is None:
@@ -351,7 +358,7 @@ class RealSenseCamera(Camera):
 
         return depth_map_processed
 
-    def read(self, color_mode: ColorMode | None = None, timeout_ms: int = 200) -> np.ndarray:
+    def read(self, color_mode: ColorMode | None = None, timeout_ms: int = 200) -> NDArray[Any]:
         """
         Reads a single frame (color) synchronously from the camera.
 
@@ -376,6 +383,9 @@ class RealSenseCamera(Camera):
 
         start_time = time.perf_counter()
 
+        if self.rs_pipeline is None:
+            raise RuntimeError(f"{self}: rs_pipeline must be initialized before use.")
+
         ret, frame = self.rs_pipeline.try_wait_for_frames(timeout_ms=timeout_ms)
 
         if not ret or frame is None:
@@ -392,8 +402,8 @@ class RealSenseCamera(Camera):
         return color_image_processed
 
     def _postprocess_image(
-        self, image: np.ndarray, color_mode: ColorMode | None = None, depth_frame: bool = False
-    ) -> np.ndarray:
+        self, image: NDArray[Any], color_mode: ColorMode | None = None, depth_frame: bool = False
+    ) -> NDArray[Any]:
         """
         Applies color conversion, dimension validation, and rotation to a raw color frame.
 
@@ -438,7 +448,7 @@ class RealSenseCamera(Camera):
 
         return processed_image
 
-    def _read_loop(self):
+    def _read_loop(self) -> None:
         """
         Internal loop run by the background thread for asynchronous reading.
 
@@ -449,6 +459,9 @@ class RealSenseCamera(Camera):
 
         Stops on DeviceNotConnectedError, logs other errors and continues.
         """
+        if self.stop_event is None:
+            raise RuntimeError(f"{self}: stop_event is not initialized before starting read loop.")
+
         while not self.stop_event.is_set():
             try:
                 color_image = self.read(timeout_ms=500)
@@ -474,7 +487,7 @@ class RealSenseCamera(Camera):
         self.thread.daemon = True
         self.thread.start()
 
-    def _stop_read_thread(self):
+    def _stop_read_thread(self) -> None:
         """Signals the background read thread to stop and waits for it to join."""
         if self.stop_event is not None:
             self.stop_event.set()
@@ -486,7 +499,7 @@ class RealSenseCamera(Camera):
         self.stop_event = None
 
     # NOTE(Steven): Missing implementation for depth for now
-    def async_read(self, timeout_ms: float = 200) -> np.ndarray:
+    def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
         """
         Reads the latest available frame data (color) asynchronously.
 
@@ -529,7 +542,7 @@ class RealSenseCamera(Camera):
 
         return frame
 
-    def disconnect(self):
+    def disconnect(self) -> None:
         """
         Disconnects from the camera, stops the pipeline, and cleans up resources.
 

src/lerobot/cameras/realsense/configuration_realsense.py

@@ -59,7 +59,7 @@ class RealSenseCameraConfig(CameraConfig):
     rotation: Cv2Rotation = Cv2Rotation.NO_ROTATION
     warmup_s: int = 1
 
-    def __post_init__(self):
+    def __post_init__(self) -> None:
         if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
             raise ValueError(
                 f"`color_mode` is expected to be {ColorMode.RGB.value} or {ColorMode.BGR.value}, but {self.color_mode} is provided."

src/lerobot/cameras/utils.py

@@ -53,14 +53,14 @@ def make_cameras_from_configs(camera_configs: dict[str, CameraConfig]) -> dict[s
 
 
 def get_cv2_rotation(rotation: Cv2Rotation) -> int | None:
-    import cv2
+    import cv2  # type: ignore  # TODO: add type stubs for OpenCV
 
     if rotation == Cv2Rotation.ROTATE_90:
-        return cv2.ROTATE_90_CLOCKWISE
+        return int(cv2.ROTATE_90_CLOCKWISE)
     elif rotation == Cv2Rotation.ROTATE_180:
-        return cv2.ROTATE_180
+        return int(cv2.ROTATE_180)
     elif rotation == Cv2Rotation.ROTATE_270:
-        return cv2.ROTATE_90_COUNTERCLOCKWISE
+        return int(cv2.ROTATE_90_COUNTERCLOCKWISE)
     else:
         return None
 
@@ -69,8 +69,8 @@ def get_cv2_backend() -> int:
     import cv2
 
     if platform.system() == "Windows":
-        return cv2.CAP_MSMF  # Use MSMF for Windows instead of AVFOUNDATION
+        return int(cv2.CAP_MSMF)  # Use MSMF for Windows instead of AVFOUNDATION
     # elif platform.system() == "Darwin":  # macOS
     #     return cv2.CAP_AVFOUNDATION
     else:  # Linux and others
-        return cv2.CAP_ANY
+        return int(cv2.CAP_ANY)

src/lerobot/configs/default.py

@@ -57,7 +57,7 @@ class EvalConfig:
     # `use_async_envs` specifies whether to use asynchronous environments (multiprocessing).
     use_async_envs: bool = False
 
-    def __post_init__(self):
+    def __post_init__(self) -> None:
         if self.batch_size > self.n_episodes:
             raise ValueError(
                 "The eval batch size is greater than the number of eval episodes "

src/lerobot/configs/eval.py

@@ -13,8 +13,8 @@
 # limitations under the License.
 
 import datetime as dt
-import logging
 from dataclasses import dataclass, field
+from logging import getLogger
 from pathlib import Path
 
 from lerobot import envs, policies  # noqa: F401
@@ -22,6 +22,8 @@ from lerobot.configs import parser
 from lerobot.configs.default import EvalConfig
 from lerobot.configs.policies import PreTrainedConfig
 
+logger = getLogger(__name__)
+
 
 @dataclass
 class EvalPipelineConfig:
@@ -34,25 +36,31 @@ class EvalPipelineConfig:
     output_dir: Path | None = None
     job_name: str | None = None
     seed: int | None = 1000
+    # Rename map for the observation to override the image and state keys
+    rename_map: dict[str, str] = field(default_factory=dict)
 
-    def __post_init__(self):
+    def __post_init__(self) -> None:
         # HACK: We parse again the cli args here to get the pretrained path if there was one.
         policy_path = parser.get_path_arg("policy")
         if policy_path:
             cli_overrides = parser.get_cli_overrides("policy")
             self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
-            self.policy.pretrained_path = policy_path
+            self.policy.pretrained_path = Path(policy_path)
 
         else:
-            logging.warning(
+            logger.warning(
                 "No pretrained path was provided, evaluated policy will be built from scratch (random weights)."
             )
 
         if not self.job_name:
             if self.env is None:
-                self.job_name = f"{self.policy.type}"
+                self.job_name = f"{self.policy.type if self.policy is not None else 'scratch'}"
             else:
-                self.job_name = f"{self.env.type}_{self.policy.type}"
+                self.job_name = (
+                    f"{self.env.type}_{self.policy.type if self.policy is not None else 'scratch'}"
+                )
+
+            logger.warning(f"No job name provided, using '{self.job_name}' as job name.")
 
         if not self.output_dir:
             now = dt.datetime.now()

src/lerobot/configs/parser.py

@@ -16,14 +16,19 @@ import inspect
 import pkgutil
 import sys
 from argparse import ArgumentError
-from collections.abc import Sequence
+from collections.abc import Callable, Iterable, Sequence
 from functools import wraps
 from pathlib import Path
+from pkgutil import ModuleInfo
+from types import ModuleType
+from typing import Any, TypeVar, cast
 
 import draccus
 
 from lerobot.utils.utils import has_method
 
+F = TypeVar("F", bound=Callable[..., object])
+
 PATH_KEY = "path"
 PLUGIN_DISCOVERY_SUFFIX = "discover_packages_path"
 
@@ -60,7 +65,7 @@ def parse_arg(arg_name: str, args: Sequence[str] | None = None) -> str | None:
     return None
 
 
-def parse_plugin_args(plugin_arg_suffix: str, args: Sequence[str]) -> dict:
+def parse_plugin_args(plugin_arg_suffix: str, args: Sequence[str]) -> dict[str, str]:
     """Parse plugin-related arguments from command-line arguments.
 
     This function extracts arguments from command-line arguments that match a specified suffix pattern.
@@ -127,7 +132,7 @@ def load_plugin(plugin_path: str) -> None:
             f"Failed to load plugin '{plugin_path}'. Verify the path and installation: {str(e)}"
         ) from e
 
-    def iter_namespace(ns_pkg):
+    def iter_namespace(ns_pkg: ModuleType) -> Iterable[ModuleInfo]:
         return pkgutil.iter_modules(ns_pkg.__path__, ns_pkg.__name__ + ".")
 
     try:
@@ -148,6 +153,8 @@ def get_type_arg(field_name: str, args: Sequence[str] | None = None) -> str | No
 
 
 def filter_arg(field_to_filter: str, args: Sequence[str] | None = None) -> list[str]:
+    if args is None:
+        return []
     return [arg for arg in args if not arg.startswith(f"--{field_to_filter}=")]
 
 
@@ -171,7 +178,8 @@ def filter_path_args(fields_to_filter: str | list[str], args: Sequence[str] | No
     if isinstance(fields_to_filter, str):
         fields_to_filter = [fields_to_filter]
 
-    filtered_args = args
+    filtered_args = [] if args is None else list(args)
+
     for field in fields_to_filter:
         if get_path_arg(field, args):
             if get_type_arg(field, args):
@@ -184,7 +192,7 @@ def filter_path_args(fields_to_filter: str | list[str], args: Sequence[str] | No
     return filtered_args
 
 
-def wrap(config_path: Path | None = None):
+def wrap(config_path: Path | None = None) -> Callable[[F], F]:
     """
     HACK: Similar to draccus.wrap but does three additional things:
         - Will remove '.path' arguments from CLI in order to process them later on.
@@ -195,9 +203,9 @@ def wrap(config_path: Path | None = None):
             from the CLI '.type' arguments
     """
 
-    def wrapper_outer(fn):
+    def wrapper_outer(fn: F) -> F:
         @wraps(fn)
-        def wrapper_inner(*args, **kwargs):
+        def wrapper_inner(*args: Any, **kwargs: Any) -> Any:
             argspec = inspect.getfullargspec(fn)
             argtype = argspec.annotations[argspec.args[0]]
             if len(args) > 0 and type(args[0]) is argtype:
@@ -225,6 +233,6 @@ def wrap(config_path: Path | None = None):
             response = fn(cfg, *args, **kwargs)
             return response
 
-        return wrapper_inner
+        return cast(F, wrapper_inner)
 
-    return wrapper_outer
+    return cast(Callable[[F], F], wrapper_outer)

src/lerobot/configs/policies.py

@@ -14,12 +14,12 @@
 import abc
 import builtins
 import json
-import logging
 import os
 import tempfile
 from dataclasses import dataclass, field
+from logging import getLogger
 from pathlib import Path
-from typing import TypeVar
+from typing import Any, TypeVar
 
 import draccus
 from huggingface_hub import hf_hub_download
@@ -34,10 +34,11 @@ from lerobot.utils.hub import HubMixin
 from lerobot.utils.utils import auto_select_torch_device, is_amp_available, is_torch_device_available
 
 T = TypeVar("T", bound="PreTrainedConfig")
+logger = getLogger(__name__)
 
 
 @dataclass
-class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):
+class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):  # type: ignore[misc,name-defined] #TODO: draccus issue
     """
     Base configuration class for policy models.
 
@@ -57,12 +58,12 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):
     input_features: dict[str, PolicyFeature] = field(default_factory=dict)
     output_features: dict[str, PolicyFeature] = field(default_factory=dict)
 
-    device: str | None = None  # cuda | cpu | mp
+    device: str | None = None  # e.g. "cuda", "cuda:0", "cpu", or "mps"
     # `use_amp` determines whether to use Automatic Mixed Precision (AMP) for training and evaluation. With AMP,
     # automatic gradient scaling is used.
     use_amp: bool = False
 
-    push_to_hub: bool = True
+    push_to_hub: bool = True  # type: ignore[assignment] # TODO: use a different name to avoid override
     repo_id: str | None = None
 
     # Upload on private repository on the Hugging Face hub.
@@ -73,38 +74,41 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):
     license: str | None = None
     # Either the repo ID of a model hosted on the Hub or a path to a directory containing weights
     # saved using `Policy.save_pretrained`. If not provided, the policy is initialized from scratch.
-    pretrained_path: str | None = None
+    pretrained_path: Path | None = None
 
-    def __post_init__(self):
+    def __post_init__(self) -> None:
         if not self.device or not is_torch_device_available(self.device):
             auto_device = auto_select_torch_device()
-            logging.warning(f"Device '{self.device}' is not available. Switching to '{auto_device}'.")
+            logger.warning(f"Device '{self.device}' is not available. Switching to '{auto_device}'.")
             self.device = auto_device.type
 
         # Automatically deactivate AMP if necessary
         if self.use_amp and not is_amp_available(self.device):
-            logging.warning(
+            logger.warning(
                 f"Automatic Mixed Precision (amp) is not available on device '{self.device}'. Deactivating AMP."
             )
             self.use_amp = False
 
     @property
     def type(self) -> str:
-        return self.get_choice_name(self.__class__)
+        choice_name = self.get_choice_name(self.__class__)
+        if not isinstance(choice_name, str):
+            raise TypeError(f"Expected string from get_choice_name, got {type(choice_name)}")
+        return choice_name
 
     @property
     @abc.abstractmethod
-    def observation_delta_indices(self) -> list | None:
+    def observation_delta_indices(self) -> list | None:  # type: ignore[type-arg] #TODO: No implementation
         raise NotImplementedError
 
     @property
     @abc.abstractmethod
-    def action_delta_indices(self) -> list | None:
+    def action_delta_indices(self) -> list | None:  # type: ignore[type-arg]    #TODO: No implementation
         raise NotImplementedError
 
     @property
     @abc.abstractmethod
-    def reward_delta_indices(self) -> list | None:
+    def reward_delta_indices(self) -> list | None:  # type: ignore[type-arg]    #TODO: No implementation
         raise NotImplementedError
 
     @abc.abstractmethod
@@ -154,13 +158,13 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):
         pretrained_name_or_path: str | Path,
         *,
         force_download: bool = False,
-        resume_download: bool = None,
-        proxies: dict | None = None,
+        resume_download: bool | None = None,
+        proxies: dict[Any, Any] | None = None,
         token: str | bool | None = None,
         cache_dir: str | Path | None = None,
         local_files_only: bool = False,
         revision: str | None = None,
-        **policy_kwargs,
+        **policy_kwargs: Any,
     ) -> T:
         model_id = str(pretrained_name_or_path)
         config_file: str | None = None
@@ -168,7 +172,7 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):
             if CONFIG_NAME in os.listdir(model_id):
                 config_file = os.path.join(model_id, CONFIG_NAME)
             else:
-                print(f"{CONFIG_NAME} not found in {Path(model_id).resolve()}")
+                logger.error(f"{CONFIG_NAME} not found in {Path(model_id).resolve()}")
         else:
             try:
                 config_file = hf_hub_download(
@@ -194,6 +198,9 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):
         with draccus.config_type("json"):
             orig_config = draccus.parse(cls, config_file, args=[])
 
+        if config_file is None:
+            raise FileNotFoundError(f"{CONFIG_NAME} not found in {model_id}")
+
         with open(config_file) as f:
             config = json.load(f)
 

src/lerobot/configs/train.py

@@ -16,6 +16,7 @@ import datetime as dt
 import os
 from dataclasses import dataclass, field
 from pathlib import Path
+from typing import Any
 
 import draccus
 from huggingface_hub import hf_hub_download
@@ -63,18 +64,35 @@ class TrainPipelineConfig(HubMixin):
     scheduler: LRSchedulerConfig | None = None
     eval: EvalConfig = field(default_factory=EvalConfig)
     wandb: WandBConfig = field(default_factory=WandBConfig)
+    
+    # RA-BC (Reward-Aligned Behavior Cloning) parameters
+    use_rabc: bool = False  # Enable reward-weighted training
+    reward_model_path: str | None = None  # Path to pre-trained reward model (e.g., SARM)
+    rabc_kappa: float = 0.01  # Hard threshold for high-quality samples
+    rabc_epsilon: float = 1e-6  # Small constant for numerical stability
+    rabc_update_freq: int = 1  # Compute rewards every N batches (1 = every batch)
 
-    def __post_init__(self):
-        self.checkpoint_path = None
+    # Rename map for the observation to override the image and state keys
+    rename_map: dict[str, str] = field(default_factory=dict)       
+    checkpoint_path: Path | None = field(init=False, default=None)
+        
 
     def validate(self):
+        # Validate RA-BC configuration
+        if self.use_rabc and not self.reward_model_path:
+            raise ValueError(
+                "RA-BC is enabled (use_rabc=True) but no reward_model_path provided. "
+                "Please specify a pre-trained reward model (e.g., SARM) path."
+            )
+
+    def validate(self) -> None:
         # HACK: We parse again the cli args here to get the pretrained paths if there was some.
         policy_path = parser.get_path_arg("policy")
         if policy_path:
             # Only load the policy config
             cli_overrides = parser.get_cli_overrides("policy")
             self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
-            self.policy.pretrained_path = policy_path
+            self.policy.pretrained_path = Path(policy_path)
         elif self.resume:
             # The entire train config is already loaded, we just need to get the checkpoint dir
             config_path = parser.parse_arg("config_path")
@@ -82,14 +100,22 @@ class TrainPipelineConfig(HubMixin):
                 raise ValueError(
                     f"A config_path is expected when resuming a run. Please specify path to {TRAIN_CONFIG_NAME}"
                 )
+
             if not Path(config_path).resolve().exists():
                 raise NotADirectoryError(
                     f"{config_path=} is expected to be a local path. "
                     "Resuming from the hub is not supported for now."
                 )
-            policy_path = Path(config_path).parent
-            self.policy.pretrained_path = policy_path
-            self.checkpoint_path = policy_path.parent
+
+            policy_dir = Path(config_path).parent
+            if self.policy is not None:
+                self.policy.pretrained_path = policy_dir
+            self.checkpoint_path = policy_dir.parent
+
+        if self.policy is None:
+            raise ValueError(
+                "Policy is not configured. Please specify a pretrained policy with `--policy.path`."
+            )
 
         if not self.job_name:
             if self.env is None:
@@ -126,8 +152,8 @@ class TrainPipelineConfig(HubMixin):
         """This enables the parser to load config from the policy using `--policy.path=local/dir`"""
         return ["policy"]
 
-    def to_dict(self) -> dict:
-        return draccus.encode(self)
+    def to_dict(self) -> dict[str, Any]:
+        return draccus.encode(self)  # type: ignore[no-any-return]  # because of the third-party library draccus uses Any as the return type
 
     def _save_pretrained(self, save_directory: Path) -> None:
         with open(save_directory / TRAIN_CONFIG_NAME, "w") as f, draccus.config_type("json"):
@@ -139,13 +165,13 @@ class TrainPipelineConfig(HubMixin):
         pretrained_name_or_path: str | Path,
         *,
         force_download: bool = False,
-        resume_download: bool = None,
-        proxies: dict | None = None,
+        resume_download: bool | None = None,
+        proxies: dict[Any, Any] | None = None,
         token: str | bool | None = None,
         cache_dir: str | Path | None = None,
         local_files_only: bool = False,
         revision: str | None = None,
-        **kwargs,
+        **kwargs: Any,
     ) -> "TrainPipelineConfig":
         model_id = str(pretrained_name_or_path)
         config_file: str | None = None
@@ -181,4 +207,6 @@ class TrainPipelineConfig(HubMixin):
 
 @dataclass(kw_only=True)
 class TrainRLServerPipelineConfig(TrainPipelineConfig):
-    dataset: DatasetConfig | None = None  # NOTE: In RL, we don't need an offline dataset
+    # NOTE: In RL, we don't need an offline dataset
+    # TODO: Make `TrainPipelineConfig.dataset` optional
+    dataset: DatasetConfig | None = None  # type: ignore[assignment] # because the parent class has made it's type non-optional

src/lerobot/configs/types.py

@@ -42,4 +42,11 @@ class NormalizationMode(str, Enum):
 @dataclass
 class PolicyFeature:
     type: FeatureType
-    shape: tuple
+    shape: tuple[int, ...]
+
+
+class RTCAttentionSchedule(str, Enum):
+    ZEROS = "ZEROS"
+    ONES = "ONES"
+    LINEAR = "LINEAR"
+    EXP = "EXP"

src/lerobot/datasets/dataset_tools.py

@@ -28,8 +28,10 @@ import shutil
 from collections.abc import Callable
 from pathlib import Path
 
+import datasets
 import numpy as np
 import pandas as pd
+import pyarrow.parquet as pq
 import torch
 from tqdm import tqdm
 
@@ -37,13 +39,13 @@ from lerobot.datasets.aggregate import aggregate_datasets
 from lerobot.datasets.compute_stats import aggregate_stats
 from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
 from lerobot.datasets.utils import (
+    DATA_DIR,
     DEFAULT_CHUNK_SIZE,
     DEFAULT_DATA_FILE_SIZE_IN_MB,
     DEFAULT_DATA_PATH,
     DEFAULT_EPISODES_PATH,
     get_parquet_file_size_in_mb,
     load_episodes,
-    to_parquet_with_hf_images,
     update_chunk_file_indices,
     write_info,
     write_stats,
@@ -268,39 +270,79 @@ def merge_datasets(
     return merged_dataset
 
 
-def add_feature(
+def modify_features(
     dataset: LeRobotDataset,
-    feature_name: str,
-    feature_values: np.ndarray | torch.Tensor | Callable,
-    feature_info: dict,
+    add_features: dict[str, tuple[np.ndarray | torch.Tensor | Callable, dict]] | None = None,
+    remove_features: str | list[str] | None = None,
     output_dir: str | Path | None = None,
     repo_id: str | None = None,
 ) -> LeRobotDataset:
-    """Add a new feature to a LeRobotDataset.
+    """Modify a LeRobotDataset by adding and/or removing features in a single pass.
+
+    This is the most efficient way to modify features, as it only copies the dataset once
+    regardless of how many features are being added or removed.
 
     Args:
         dataset: The source LeRobotDataset.
-        feature_name: Name of the new feature.
-        feature_values: Either:
-            - Array/tensor of shape (num_frames, ...) with values for each frame
-            - Callable that takes (frame_dict, episode_index, frame_index) and returns feature value
-        feature_info: Dictionary with feature metadata (dtype, shape, names).
+        add_features: Optional dict mapping feature names to (feature_values, feature_info) tuples.
+        remove_features: Optional feature name(s) to remove. Can be a single string or list.
         output_dir: Directory to save the new dataset. If None, uses default location.
         repo_id: Repository ID for the new dataset. If None, appends "_modified" to original.
+
+    Returns:
+        New dataset with features modified.
+
+    Example:
+        new_dataset = modify_features(
+            dataset,
+            add_features={
+                "reward": (reward_array, {"dtype": "float32", "shape": [1], "names": None}),
+            },
+            remove_features=["old_feature"],
+            output_dir="./output",
+        )
     """
-    if feature_name in dataset.meta.features:
-        raise ValueError(f"Feature '{feature_name}' already exists in dataset")
+    if add_features is None and remove_features is None:
+        raise ValueError("Must specify at least one of add_features or remove_features")
+
+    remove_features_list: list[str] = []
+    if remove_features is not None:
+        remove_features_list = [remove_features] if isinstance(remove_features, str) else remove_features
+
+    if add_features:
+        required_keys = {"dtype", "shape"}
+        for feature_name, (_, feature_info) in add_features.items():
+            if feature_name in dataset.meta.features:
+                raise ValueError(f"Feature '{feature_name}' already exists in dataset")
+
+            if not required_keys.issubset(feature_info.keys()):
+                raise ValueError(f"feature_info for '{feature_name}' must contain keys: {required_keys}")
+
+    if remove_features_list:
+        for name in remove_features_list:
+            if name not in dataset.meta.features:
+                raise ValueError(f"Feature '{name}' not found in dataset")
+
+        required_features = {"timestamp", "frame_index", "episode_index", "index", "task_index"}
+        if any(name in required_features for name in remove_features_list):
+            raise ValueError(f"Cannot remove required features: {required_features}")
 
     if repo_id is None:
         repo_id = f"{dataset.repo_id}_modified"
     output_dir = Path(output_dir) if output_dir is not None else HF_LEROBOT_HOME / repo_id
 
-    required_keys = {"dtype", "shape"}
-    if not required_keys.issubset(feature_info.keys()):
-        raise ValueError(f"feature_info must contain keys: {required_keys}")
-
     new_features = dataset.meta.features.copy()
-    new_features[feature_name] = feature_info
+
+    if remove_features_list:
+        for name in remove_features_list:
+            new_features.pop(name, None)
+
+    if add_features:
+        for feature_name, (_, feature_info) in add_features.items():
+            new_features[feature_name] = feature_info
+
+    video_keys_to_remove = [name for name in remove_features_list if name in dataset.meta.video_keys]
+    remaining_video_keys = [k for k in dataset.meta.video_keys if k not in video_keys_to_remove]
 
     new_meta = LeRobotDatasetMetadata.create(
         repo_id=repo_id,
@@ -308,17 +350,18 @@ def add_feature(
         features=new_features,
         robot_type=dataset.meta.robot_type,
         root=output_dir,
-        use_videos=len(dataset.meta.video_keys) > 0,
+        use_videos=len(remaining_video_keys) > 0,
     )
 
     _copy_data_with_feature_changes(
         dataset=dataset,
         new_meta=new_meta,
-        add_features={feature_name: (feature_values, feature_info)},
+        add_features=add_features,
+        remove_features=remove_features_list if remove_features_list else None,
     )
 
-    if dataset.meta.video_keys:
-        _copy_videos(dataset, new_meta)
+    if new_meta.video_keys:
+        _copy_videos(dataset, new_meta, exclude_keys=video_keys_to_remove if video_keys_to_remove else None)
 
     new_dataset = LeRobotDataset(
         repo_id=repo_id,
@@ -331,6 +374,46 @@ def add_feature(
     return new_dataset
 
 
+def add_features(
+    dataset: LeRobotDataset,
+    features: dict[str, tuple[np.ndarray | torch.Tensor | Callable, dict]],
+    output_dir: str | Path | None = None,
+    repo_id: str | None = None,
+) -> LeRobotDataset:
+    """Add multiple features to a LeRobotDataset in a single pass.
+
+    This is more efficient than calling add_feature() multiple times, as it only
+    copies the dataset once regardless of how many features are being added.
+
+    Args:
+        dataset: The source LeRobotDataset.
+        features: Dictionary mapping feature names to (feature_values, feature_info) tuples.
+        output_dir: Directory to save the new dataset. If None, uses default location.
+        repo_id: Repository ID for the new dataset. If None, appends "_modified" to original.
+
+    Returns:
+        New dataset with all features added.
+
+    Example:
+        features = {
+            "task_embedding": (task_emb_array, {"dtype": "float32", "shape": [384], "names": None}),
+            "cam1_embedding": (cam1_emb_array, {"dtype": "float32", "shape": [768], "names": None}),
+            "cam2_embedding": (cam2_emb_array, {"dtype": "float32", "shape": [768], "names": None}),
+        }
+        new_dataset = add_features(dataset, features, output_dir="./output", repo_id="my_dataset")
+    """
+    if not features:
+        raise ValueError("No features provided")
+
+    return modify_features(
+        dataset=dataset,
+        add_features=features,
+        remove_features=None,
+        output_dir=output_dir,
+        repo_id=repo_id,
+    )
+
+
 def remove_feature(
     dataset: LeRobotDataset,
     feature_names: str | list[str],
@@ -345,56 +428,17 @@ def remove_feature(
         output_dir: Directory to save the new dataset. If None, uses default location.
         repo_id: Repository ID for the new dataset. If None, appends "_modified" to original.
 
+    Returns:
+        New dataset with features removed.
     """
-    if isinstance(feature_names, str):
-        feature_names = [feature_names]
-
-    for name in feature_names:
-        if name not in dataset.meta.features:
-            raise ValueError(f"Feature '{name}' not found in dataset")
-
-    required_features = {"timestamp", "frame_index", "episode_index", "index", "task_index"}
-    if any(name in required_features for name in feature_names):
-        raise ValueError(f"Cannot remove required features: {required_features}")
-
-    if repo_id is None:
-        repo_id = f"{dataset.repo_id}_modified"
-    output_dir = Path(output_dir) if output_dir is not None else HF_LEROBOT_HOME / repo_id
-
-    new_features = {k: v for k, v in dataset.meta.features.items() if k not in feature_names}
-
-    video_keys_to_remove = [name for name in feature_names if name in dataset.meta.video_keys]
-
-    remaining_video_keys = [k for k in dataset.meta.video_keys if k not in video_keys_to_remove]
-
-    new_meta = LeRobotDatasetMetadata.create(
-        repo_id=repo_id,
-        fps=dataset.meta.fps,
-        features=new_features,
-        robot_type=dataset.meta.robot_type,
-        root=output_dir,
-        use_videos=len(remaining_video_keys) > 0,
-    )
-
-    _copy_data_with_feature_changes(
+    return modify_features(
         dataset=dataset,
-        new_meta=new_meta,
+        add_features=None,
         remove_features=feature_names,
-    )
-
-    if new_meta.video_keys:
-        _copy_videos(dataset, new_meta, exclude_keys=video_keys_to_remove)
-
-    new_dataset = LeRobotDataset(
+        output_dir=output_dir,
         repo_id=repo_id,
-        root=output_dir,
-        image_transforms=dataset.image_transforms,
-        delta_timestamps=dataset.delta_timestamps,
-        tolerance_s=dataset.tolerance_s,
     )
 
-    return new_dataset
-
 
 def _fractions_to_episode_indices(
     total_episodes: int,
@@ -501,10 +545,7 @@ def _copy_and_reindex_data(
         dst_path = dst_meta.root / DEFAULT_DATA_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
         dst_path.parent.mkdir(parents=True, exist_ok=True)
 
-        if len(dst_meta.image_keys) > 0:
-            to_parquet_with_hf_images(df, dst_path)
-        else:
-            df.to_parquet(dst_path, index=False)
+        _write_parquet(df, dst_path, dst_meta)
 
         for ep_old_idx in episodes_to_keep:
             ep_new_idx = episode_mapping[ep_old_idx]
@@ -862,6 +903,25 @@ def _copy_and_reindex_episodes_metadata(
     write_stats(filtered_stats, dst_meta.root)
 
 
+def _write_parquet(df: pd.DataFrame, path: Path, meta: LeRobotDatasetMetadata) -> None:
+    """Write DataFrame to parquet
+
+    This ensures images are properly embedded and the file can be loaded correctly by HF datasets.
+    """
+    from lerobot.datasets.utils import embed_images, get_hf_features_from_features
+
+    hf_features = get_hf_features_from_features(meta.features)
+    ep_dataset = datasets.Dataset.from_dict(df.to_dict(orient="list"), features=hf_features, split="train")
+
+    if len(meta.image_keys) > 0:
+        ep_dataset = embed_images(ep_dataset)
+
+    table = ep_dataset.with_format("arrow")[:]
+    writer = pq.ParquetWriter(path, schema=table.schema, compression="snappy", use_dictionary=True)
+    writer.write_table(table)
+    writer.close()
+
+
 def _save_data_chunk(
     df: pd.DataFrame,
     meta: LeRobotDatasetMetadata,
@@ -877,10 +937,7 @@ def _save_data_chunk(
     path = meta.root / DEFAULT_DATA_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
     path.parent.mkdir(parents=True, exist_ok=True)
 
-    if len(meta.image_keys) > 0:
-        to_parquet_with_hf_images(df, path)
-    else:
-        df.to_parquet(path, index=False)
+    _write_parquet(df, path, meta)
 
     episode_metadata = {}
     for ep_idx in df["episode_index"].unique():
@@ -906,19 +963,29 @@ def _copy_data_with_feature_changes(
     remove_features: list[str] | None = None,
 ) -> None:
     """Copy data while adding or removing features."""
-    file_paths = set()
-    for ep_idx in range(dataset.meta.total_episodes):
-        file_paths.add(dataset.meta.get_data_file_path(ep_idx))
+    data_dir = dataset.root / DATA_DIR
+    parquet_files = sorted(data_dir.glob("*/*.parquet"))
+
+    if not parquet_files:
+        raise ValueError(f"No parquet files found in {data_dir}")
 
     frame_idx = 0
 
-    for src_path in tqdm(sorted(file_paths), desc="Processing data files"):
-        df = pd.read_parquet(dataset.root / src_path).reset_index(drop=True)
+    for src_path in tqdm(parquet_files, desc="Processing data files"):
+        df = pd.read_parquet(src_path).reset_index(drop=True)
+
+        relative_path = src_path.relative_to(dataset.root)
+        chunk_dir = relative_path.parts[1]
+        file_name = relative_path.parts[2]
+
+        chunk_idx = int(chunk_dir.split("-")[1])
+        file_idx = int(file_name.split("-")[1].split(".")[0])
 
         if remove_features:
             df = df.drop(columns=remove_features, errors="ignore")
 
         if add_features:
+            end_idx = frame_idx + len(df)
             for feature_name, (values, _) in add_features.items():
                 if callable(values):
                     feature_values = []
@@ -931,15 +998,26 @@ def _copy_data_with_feature_changes(
                         feature_values.append(value)
                     df[feature_name] = feature_values
                 else:
-                    end_idx = frame_idx + len(df)
                     feature_slice = values[frame_idx:end_idx]
-                    if len(feature_slice.shape) > 1 and feature_slice.shape[1] == 1:
-                        df[feature_name] = feature_slice.flatten()
-                    else:
+                    if len(feature_slice.shape) == 1:
+                        # 1D array - can assign directly
                         df[feature_name] = feature_slice
-                    frame_idx = end_idx
+                    elif len(feature_slice.shape) == 2 and feature_slice.shape[1] == 1:
+                        # 2D array with single column - flatten it
+                        df[feature_name] = feature_slice.flatten()
+                    elif len(feature_slice.shape) == 2:
+                        # 2D array with multiple columns (e.g., embeddings) - convert to list of lists
+                        df[feature_name] = feature_slice.tolist()
+                    else:
+                        # Higher dimensional - convert to list
+                        df[feature_name] = [row.tolist() for row in feature_slice]
+            frame_idx = end_idx
 
-        _save_data_chunk(df, new_meta)
+        # Write using the same chunk/file structure as source
+        dst_path = new_meta.root / DEFAULT_DATA_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
+        dst_path.parent.mkdir(parents=True, exist_ok=True)
+
+        _write_parquet(df, dst_path, new_meta)
 
     _copy_episodes_metadata_and_stats(dataset, new_meta)
 

src/lerobot/datasets/generating_embeddings/README.md

@@ -0,0 +1,146 @@
+# LeRobot Embedding Generation Script
+
+Generate embeddings for LeRobot datasets to make them more lightweight and efficient for training.
+
+## Overview
+
+This script processes v3.0 LeRobot datasets and adds pre-computed embeddings for:
+
+- **Task embeddings**: Language command embeddings using MiniLM
+- **Image embeddings**: Frame embeddings using DinoV2
+
+The resulting dataset can be used more efficiently during training by loading pre-computed embeddings instead of running encoders on-the-fly.
+
+## Supported Encoders
+
+### Image Encoders (DinoV2)
+
+DinoV2 is a self-supervised vision transformer that produces high-quality image embeddings:
+
+- **`dinov2_vits14`**: ViT-S/14 (384-dim) - Fastest, smaller model
+- **`dinov2_vitb14`**: ViT-B/14 (768-dim) - **Recommended** - Good balance
+- **`dinov2_vitl14`**: ViT-L/14 (1024-dim) - Best quality, slower
+
+### Language Encoders (MiniLM)
+
+MiniLM is a lightweight sentence transformer model:
+
+- **`minilm-l6`**: MiniLM-L6-v2 (384-dim) - Faster
+- **`minilm-l12`**: MiniLM-L12-v2 (384-dim) - **Recommended** - Better quality
+
+## Usage
+
+### Basic Command
+
+```bash
+python src/lerobot/datasets/generating_embeddings/generate_embeddings.py \
+    --repo-id lerobot/utokyo_xarm_bimanual \
+    --output-repo-id your-username/utokyo_xarm_bimanual_embeddings \
+    --image-encoder dinov2_vitb14 \
+    --language-encoder minilm-l12 \
+    --push-to-hub
+```
+
+### Lightweight Version (No Videos)
+
+Removes video files to significantly reduce storage:
+
+```bash
+python src/lerobot/datasets/generating_embeddings/generate_embeddings.py \
+    --repo-id lerobot/utokyo_xarm_bimanual \
+    --output-repo-id your-username/utokyo_xarm_bimanual_lightweight \
+    --image-encoder dinov2_vitb14 \
+    --language-encoder minilm-l12 \
+    --remove-videos \
+    --push-to-hub
+```
+
+## Output
+
+The script adds new features to your dataset:
+
+### New Features
+
+1. **`task_embedding`**: Language embedding for each frame
+   - Shape: `[384]` (MiniLM)
+   - One embedding per frame based on its task
+
+2. **`{camera_key}_embedding`**: Image embedding for each camera view
+   - Shape: `[384]`, `[768]`, or `[1024]` depending on DinoV2 model
+   - Examples: `observation.images.top_embedding`, `observation.images.wrist_embedding`
+
+### Using Embeddings in Training
+
+```python
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+
+# Load dataset with embeddings
+dataset = LeRobotDataset("your-username/utokyo_xarm_bimanual_embeddings")
+
+# Access embeddings
+item = dataset[0]
+task_emb = item["task_embedding"]  # Shape: [384]
+img_emb = item["observation.images.top_embedding"]  # Shape: [768]
+
+# Use in your policy
+# Instead of running encoders during training, use pre-computed embeddings
+```
+
+## Extending with New Encoders
+
+The script is designed to be easily extensible. To add a new encoder:
+
+### 1. Create Encoder Class
+
+```python
+class MyCustomImageEncoder(ImageEncoder):
+    """Your custom image encoder."""
+
+    def __init__(self, device: str = "cuda"):
+        super().__init__(device)
+        # Load your model
+        self.model = load_my_model()
+        self.model = self.model.to(self.device)
+        self.model.eval()
+
+    def encode(self, images: list[np.ndarray]) -> np.ndarray:
+        """Encode a batch of images."""
+        # Your encoding logic here
+        embeddings = []
+        for img in images:
+            emb = self.model(img)
+            embeddings.append(emb)
+        return np.array(embeddings)
+
+    @property
+    def embedding_dim(self) -> int:
+        """Return embedding dimension."""
+        return 512  # Your embedding dimension
+```
+
+### 2. Add to Factory Function
+
+```python
+def get_image_encoder(encoder_name: str, device: str = "cuda") -> ImageEncoder:
+    encoders = {
+        "dinov2_vits14": lambda: DinoV2Encoder(model_name="dinov2_vits14", device=device),
+        "dinov2_vitb14": lambda: DinoV2Encoder(model_name="dinov2_vitb14", device=device),
+        "dinov2_vitl14": lambda: DinoV2Encoder(model_name="dinov2_vitl14", device=device),
+        # Add your encoder
+        "my_custom": lambda: MyCustomImageEncoder(device=device),
+    }
+    # ... rest of function
+```
+
+## Validating Embeddings
+
+After generating embeddings, you can validate them using `validate_embeddings.py`:
+
+```bash
+python src/lerobot/datasets/generating_embeddings/validate_embeddings.py \
+    --original-repo-id lerobot/utokyo_xarm_bimanual \
+    --embeddings-repo-id pepijn223/utokyo_xarm_bimanual_embeddings \
+    --image-encoder dinov2_vitb14 \
+    --language-encoder minilm-l12 \
+    --num-samples 20
+```

src/lerobot/datasets/generating_embeddings/encoders.py

@@ -0,0 +1,147 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+
+import numpy as np
+import torch
+from PIL import Image
+
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
+
+
+class ImageEncoder:
+    """Base class for image encoders."""
+
+    def __init__(self, device: str = "cuda"):
+        self.device = torch.device(device if torch.cuda.is_available() else "cpu")
+
+    def encode(self, images: list[np.ndarray]) -> np.ndarray:
+        """Encode a batch of images."""
+        raise NotImplementedError
+
+
+class DinoV2Encoder(ImageEncoder):
+    """DinoV2 image encoder.
+
+    DinoV2 is a self-supervised vision transformer that produces high-quality image embeddings.
+    Supports multiple model sizes (ViT-S/14, ViT-B/14, ViT-L/14).
+    """
+
+    def __init__(self, model_name: str = "dinov2_vitb14", device: str = "cuda", batch_size: int = 32):
+        super().__init__(device)
+        self.batch_size = batch_size
+        self.model_name = model_name
+        logger.info(f"Loading DinoV2 model: {model_name}")
+        self.model = torch.hub.load("facebookresearch/dinov2", model_name)  # nosec B614
+        self.model = self.model.to(self.device)
+        self.model.eval()
+
+        # DinoV2 preprocessing
+        from torchvision import transforms
+
+        self.transform = transforms.Compose(
+            [
+                transforms.Resize(256, interpolation=transforms.InterpolationMode.BICUBIC),
+                transforms.CenterCrop(224),
+                transforms.ToTensor(),
+                transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
+            ]
+        )
+
+    def encode(self, images: list[np.ndarray]) -> np.ndarray:
+        """Encode a batch of images."""
+        embeddings = []
+
+        with torch.inference_mode():
+            for i in range(0, len(images), self.batch_size):
+                batch_images = images[i : i + self.batch_size]
+                # Convert numpy arrays to PIL Images and apply transforms
+                pil_images = [Image.fromarray(img.astype(np.uint8)) for img in batch_images]
+                tensors = torch.stack([self.transform(img) for img in pil_images]).to(self.device)
+
+                # Get embeddings
+                batch_embeddings = self.model(tensors).cpu().numpy()
+                embeddings.append(batch_embeddings)
+
+        return np.concatenate(embeddings, axis=0)
+
+    @property
+    def embedding_dim(self) -> int:
+        """Return the embedding dimension based on model size."""
+        if "vits14" in self.model_name:
+            return 384  # DinoV2 ViT-S/14
+        elif "vitb14" in self.model_name:
+            return 768  # DinoV2 ViT-B/14
+        elif "vitl14" in self.model_name:
+            return 1024  # DinoV2 ViT-L/14
+        else:
+            return 768  # Default to ViT-B/14
+
+
+class LanguageEncoder:
+    """Base class for language encoders."""
+
+    def __init__(self, device: str = "cuda"):
+        self.device = torch.device(device if torch.cuda.is_available() else "cpu")
+
+    def encode(self, texts: list[str]) -> np.ndarray:
+        """Encode a batch of texts."""
+        raise NotImplementedError
+
+
+class MiniLMEncoder(LanguageEncoder):
+    """MiniLM language encoder.
+
+    MiniLM is a lightweight sentence transformer model that produces high-quality text embeddings.
+    Supports L6 and L12 model sizes.
+    """
+
+    def __init__(self, model_name: str = "sentence-transformers/all-MiniLM-L12-v2", device: str = "cuda"):
+        super().__init__(device)
+        self.model_name = model_name
+        logger.info(f"Loading MiniLM model: {model_name}")
+
+        from transformers import AutoModel, AutoTokenizer
+
+        self.tokenizer = AutoTokenizer.from_pretrained(model_name)
+        self.model = AutoModel.from_pretrained(model_name).to(self.device)
+        self.model.eval()
+
+    def _mean_pooling(self, model_output, attention_mask):
+        """Mean pooling to get sentence embeddings."""
+        token_embeddings = model_output[0]
+        input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
+        return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(
+            input_mask_expanded.sum(1), min=1e-9
+        )
+
+    def encode(self, texts: list[str]) -> np.ndarray:
+        """Encode a batch of texts."""
+        with torch.inference_mode():
+            encoded_input = self.tokenizer(texts, padding=True, truncation=True, return_tensors="pt")
+            encoded_input = {k: v.to(self.device) for k, v in encoded_input.items()}
+
+            model_output = self.model(**encoded_input)
+            embeddings = self._mean_pooling(model_output, encoded_input["attention_mask"])
+
+            return embeddings.cpu().numpy()
+
+    @property
+    def embedding_dim(self) -> int:
+        """Return the embedding dimension."""
+        return 384  # Both MiniLM-L6 and L12 output 384-dim embeddings

src/lerobot/datasets/generating_embeddings/generate_embeddings.py

@@ -0,0 +1,329 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Generate embeddings for LeRobot datasets to make them more lightweight and efficient.
+
+This script:
+1. Loads a v3.0 LeRobot dataset from the hub
+2. Computes embeddings for tasks (language commands) and frames (images)
+3. Stores embeddings as new features in the dataset
+4. Optionally removes video files to reduce size
+5. Pushes the converted dataset to the hub
+
+Current supported encoders:
+- Image: DinoV2 (dinov2_vits14, dinov2_vitb14, dinov2_vitl14)
+- Language: MiniLM (minilm-l6, minilm-l12)
+
+The architecture is extensible - you can add more encoders by:
+1. Creating a new encoder class inheriting from ImageEncoder or LanguageEncoder
+2. Implementing the encode() method and embedding_dim property
+3. Adding it to the get_image_encoder() or get_language_encoder() factory function
+
+Usage example:
+    python src/lerobot/datasets/generating_embeddings/generate_embeddings.py \
+        --repo-id lerobot/utokyo_xarm_bimanual \
+        --output-repo-id lerobot/utokyo_xarm_bimanual_embeddings \
+        --image-encoder dinov2_vitb14 \
+        --language-encoder minilm-l12 \
+        --remove-videos \
+        --push-to-hub
+"""
+
+import argparse
+import shutil
+from pathlib import Path
+
+import numpy as np
+import torch
+from tqdm import tqdm
+
+from lerobot.datasets.generating_embeddings.encoders import (
+    DinoV2Encoder,
+    ImageEncoder,
+    LanguageEncoder,
+    MiniLMEncoder,
+)
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+
+
+def get_image_encoder(encoder_name: str, device: str = "cuda") -> ImageEncoder:
+    """Factory function to get image encoder.
+
+    To add a new encoder:
+    1. Create a new class inheriting from ImageEncoder
+    2. Implement encode() and embedding_dim property
+    3. Add it to the encoders dictionary below
+    """
+    encoders = {
+        "dinov2_vits14": lambda: DinoV2Encoder(model_name="dinov2_vits14", device=device),
+        "dinov2_vitb14": lambda: DinoV2Encoder(model_name="dinov2_vitb14", device=device),
+        "dinov2_vitl14": lambda: DinoV2Encoder(model_name="dinov2_vitl14", device=device),
+    }
+
+    if encoder_name not in encoders:
+        raise ValueError(f"Unknown image encoder: {encoder_name}. Available options: {list(encoders.keys())}")
+
+    return encoders[encoder_name]()
+
+
+def get_language_encoder(encoder_name: str, device: str = "cuda") -> LanguageEncoder:
+    """Factory function to get language encoder.
+
+    To add a new encoder:
+    1. Create a new class inheriting from LanguageEncoder
+    2. Implement encode() and embedding_dim property
+    3. Add it to the encoders dictionary below
+    """
+    encoders = {
+        "minilm-l6": lambda: MiniLMEncoder(
+            model_name="sentence-transformers/all-MiniLM-L6-v2", device=device
+        ),
+        "minilm-l12": lambda: MiniLMEncoder(
+            model_name="sentence-transformers/all-MiniLM-L12-v2", device=device
+        ),
+    }
+
+    if encoder_name not in encoders:
+        raise ValueError(
+            f"Unknown language encoder: {encoder_name}. Available options: {list(encoders.keys())}"
+        )
+
+    return encoders[encoder_name]()
+
+
+def generate_embeddings_for_dataset(
+    repo_id: str,
+    output_repo_id: str,
+    image_encoder: ImageEncoder,
+    language_encoder: LanguageEncoder,
+    remove_videos: bool = False,
+    local_dir: Path | None = None,
+    output_local_dir: Path | None = None,
+    push_to_hub: bool = False,
+):
+    """Generate embeddings for a LeRobot dataset.
+
+    Args:
+        repo_id: Source dataset repository ID
+        output_repo_id: Output dataset repository ID
+        image_encoder: Image encoder instance
+        language_encoder: Language encoder instance
+        remove_videos: Whether to remove video files
+        local_dir: Local directory for source dataset
+        output_local_dir: Local directory for output dataset
+        push_to_hub: Whether to push to hub after conversion
+    """
+    from lerobot.datasets.dataset_tools import modify_features
+
+    print(f"Loading dataset: {repo_id}")
+
+    dataset = LeRobotDataset(repo_id, root=local_dir, download_videos=True)
+    print(f"Dataset: {dataset.num_episodes} episodes, {dataset.num_frames} frames")
+
+    print("Computing task embeddings...")
+    unique_tasks = dataset.meta.tasks.index.tolist()
+    task_embeddings = {}
+
+    for task in tqdm(unique_tasks, desc="Encoding tasks"):
+        # Clean up task text
+        task_clean = task.strip().capitalize().strip(" .,!?-_")
+        embedding = language_encoder.encode([task_clean])[0]
+        task_embeddings[task] = embedding
+
+    print(f"Computed {len(task_embeddings)} task embeddings")
+
+    print("Processing frames and computing embeddings...")
+    all_task_embeddings = []
+    all_image_embeddings_dict = {cam_key: [] for cam_key in dataset.meta.camera_keys}
+
+    for frame_idx in tqdm(range(dataset.num_frames), desc="Processing frames"):
+        item = dataset.hf_dataset[frame_idx]
+        ep_idx = item["episode_index"].item()
+
+        task = dataset.meta.tasks.iloc[item["task_index"].item()].name
+        task_emb = task_embeddings[task]
+        all_task_embeddings.append(task_emb)
+
+        for cam_key in dataset.meta.camera_keys:
+            if cam_key in dataset.meta.video_keys:
+                current_ts = item["timestamp"].item()
+                video_frames = dataset._query_videos({cam_key: [current_ts]}, ep_idx)
+                img = video_frames[cam_key]
+
+                if isinstance(img, torch.Tensor):
+                    if img.ndim == 4:
+                        img = img[0]  # (T, C, H, W) -> (C, H, W)
+                    elif img.ndim != 3:
+                        raise ValueError(f"Unexpected video frame shape {img.shape} for camera {cam_key}")
+                    img_np = (img.permute(1, 2, 0).numpy() * 255).astype(np.uint8)
+                else:
+                    img_np = np.array(img)
+            else:
+                img = item[cam_key]
+                if isinstance(img, torch.Tensor):
+                    if img.ndim == 3:
+                        img_np = (img.permute(1, 2, 0).numpy() * 255).astype(np.uint8)
+                    else:
+                        raise ValueError(f"Unexpected image shape {img.shape} for camera {cam_key}")
+                else:
+                    img_np = np.array(img)
+
+            all_image_embeddings_dict[cam_key].append(img_np)
+
+    print("Computing image embeddings...")
+    image_embeddings_dict = {}
+    for cam_key, images in all_image_embeddings_dict.items():
+        print(f"  {cam_key}: {len(images)} images")
+        embeddings = image_encoder.encode(images)
+        image_embeddings_dict[cam_key] = embeddings
+
+    all_task_embeddings = np.array(all_task_embeddings)
+    for cam_key in dataset.meta.camera_keys:
+        image_embeddings_dict[cam_key] = np.array(image_embeddings_dict[cam_key])
+
+    img_emb_dim = image_encoder.embedding_dim
+    lang_emb_dim = language_encoder.embedding_dim
+
+    add_features_dict = {
+        "task_embedding": (
+            all_task_embeddings,
+            {"dtype": "float32", "shape": [lang_emb_dim], "names": None},
+        ),
+    }
+
+    for cam_key in dataset.meta.camera_keys:
+        add_features_dict[f"{cam_key}_embedding"] = (
+            image_embeddings_dict[cam_key],
+            {"dtype": "float32", "shape": [img_emb_dim], "names": None},
+        )
+
+    print("Adding embeddings to dataset...")
+    remove_features_list = None
+    if remove_videos:
+        remove_features_list = dataset.meta.video_keys
+
+    output_dataset = modify_features(
+        dataset=dataset,
+        add_features=add_features_dict,
+        remove_features=remove_features_list,
+        output_dir=output_local_dir,
+        repo_id=output_repo_id,
+    )
+
+    if remove_videos:
+        print("Removing video files...")
+        videos_dir = output_dataset.root / "videos"
+        if videos_dir.exists():
+            shutil.rmtree(videos_dir)
+
+    print(f"Saved to: {output_dataset.root}")
+
+    if push_to_hub:
+        print(f"Pushing to hub: {output_repo_id}")
+        output_dataset.push_to_hub(push_videos=not remove_videos)
+        print("Done!")
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Generate embeddings for LeRobot datasets",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+        epilog="""
+Examples:
+  # Basic usage with default encoders (DinoV2 ViT-B/14 + MiniLM-L12)
+  python src/lerobot/datasets/generating_embeddings/generate_embeddings.py \\
+      --repo-id lerobot/utokyo_xarm_bimanual \\
+      --output-repo-id your-username/utokyo_xarm_bimanual_embeddings \\
+      --image-encoder dinov2_vitb14 \\
+      --language-encoder minilm-l12 \\
+      --push-to-hub
+
+  # Generate embeddings and remove videos
+  python src/lerobot/datasets/generating_embeddings/generate_embeddings.py \\
+      --repo-id lerobot/utokyo_xarm_bimanual \\
+      --output-repo-id your-username/utokyo_xarm_bimanual_lightweight \\
+      --image-encoder dinov2_vitb14 \\
+      --language-encoder minilm-l12 \\
+      --remove-videos \\
+      --push-to-hub
+
+Available image encoders:
+  - dinov2_vits14: DinoV2 ViT-S/14 (384-dim, faster)
+  - dinov2_vitb14: DinoV2 ViT-B/14 (768-dim, recommended)
+  - dinov2_vitl14: DinoV2 ViT-L/14 (1024-dim, best quality)
+
+Available language encoders:
+  - minilm-l6: MiniLM-L6-v2 (384-dim, faster)
+  - minilm-l12: MiniLM-L12-v2 (384-dim, recommended)
+        """,
+    )
+    parser.add_argument("--repo-id", type=str, required=True, help="Source dataset repository ID")
+    parser.add_argument("--output-repo-id", type=str, required=True, help="Output dataset repository ID")
+    parser.add_argument(
+        "--image-encoder",
+        type=str,
+        default="dinov2_vitb14",
+        help="Image encoder to use (default: dinov2_vitb14)",
+    )
+    parser.add_argument(
+        "--language-encoder",
+        type=str,
+        default="minilm-l12",
+        help="Language encoder to use (default: minilm-l12)",
+    )
+    parser.add_argument(
+        "--remove-videos",
+        action="store_true",
+        help="Remove video files after generating embeddings",
+    )
+    parser.add_argument("--local-dir", type=str, default=None, help="Local directory for source dataset")
+    parser.add_argument(
+        "--output-local-dir", type=str, default=None, help="Local directory for output dataset"
+    )
+    parser.add_argument(
+        "--push-to-hub",
+        action="store_true",
+        help="Push the converted dataset to the hub",
+    )
+    parser.add_argument(
+        "--device",
+        type=str,
+        default="cuda",
+        help="Device to use for encoding (default: cuda)",
+    )
+
+    args = parser.parse_args()
+
+    # Load encoders
+    image_encoder = get_image_encoder(args.image_encoder, device=args.device)
+    language_encoder = get_language_encoder(args.language_encoder, device=args.device)
+
+    # Generate embeddings
+    generate_embeddings_for_dataset(
+        repo_id=args.repo_id,
+        output_repo_id=args.output_repo_id,
+        image_encoder=image_encoder,
+        language_encoder=language_encoder,
+        remove_videos=args.remove_videos,
+        local_dir=Path(args.local_dir) if args.local_dir else None,
+        output_local_dir=Path(args.output_local_dir) if args.output_local_dir else None,
+        push_to_hub=args.push_to_hub,
+    )
+
+
+if __name__ == "__main__":
+    main()

src/lerobot/datasets/generating_embeddings/validate_embeddings.py

@@ -0,0 +1,222 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Validate pre-computed embeddings against on-the-fly computed embeddings.
+
+Usage:
+    python src/lerobot/datasets/generating_embeddings/validate_embeddings.py \
+        --original-repo-id lerobot/utokyo_xarm_bimanual \
+        --embeddings-repo-id <your_username>/utokyo_xarm_bimanual_embeddings \
+        --image-encoder dinov2_vitb14 \
+        --language-encoder minilm-l12 \
+        --num-samples 10
+"""
+
+import argparse
+
+import numpy as np
+import torch
+from tqdm import tqdm
+
+from lerobot.datasets.generating_embeddings.encoders import ImageEncoder, LanguageEncoder
+from lerobot.datasets.generating_embeddings.generate_embeddings import (
+    get_image_encoder,
+    get_language_encoder,
+)
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+
+
+def cosine_similarity(a: np.ndarray, b: np.ndarray) -> float:
+    """Compute cosine similarity between two vectors."""
+    return np.dot(a, b) / (np.linalg.norm(a) * np.linalg.norm(b))
+
+
+def validate_embeddings(
+    original_repo_id: str,
+    embeddings_repo_id: str,
+    image_encoder: ImageEncoder,
+    language_encoder: LanguageEncoder,
+    num_samples: int = 10,
+    device: str = "cuda",
+):
+    """Validate pre-computed embeddings against on-the-fly embeddings.
+
+    Args:
+        original_repo_id: Original dataset repository ID
+        embeddings_repo_id: Dataset with pre-computed embeddings repository ID
+        image_encoder: Image encoder instance
+        language_encoder: Language encoder instance
+        num_samples: Number of samples to validate
+        device: Device to use for encoding
+    """
+    # Load both datasets
+    print("Loading datasets...")
+    original_dataset = LeRobotDataset(original_repo_id, download_videos=True)
+    embeddings_dataset = LeRobotDataset(embeddings_repo_id, download_videos=False)
+
+    # Verify both datasets have the same number of frames
+    assert original_dataset.num_frames == embeddings_dataset.num_frames, (
+        f"Frame count mismatch: original={original_dataset.num_frames}, "
+        f"embeddings={embeddings_dataset.num_frames}"
+    )
+
+    camera_keys = original_dataset.meta.camera_keys
+
+    # Check embedding features exist
+    expected_features = ["task_embedding"] + [f"{cam}_embedding" for cam in camera_keys]
+    for feat in expected_features:
+        if feat not in embeddings_dataset.features:
+            raise ValueError(f"Embedding feature not found: {feat}")
+
+    # Select random sample indices
+    sample_indices = np.random.choice(
+        original_dataset.num_frames, size=min(num_samples, original_dataset.num_frames), replace=False
+    )
+    print(f"Validating {len(sample_indices)} samples...")
+
+    # Track statistics
+    task_similarities = []
+    image_similarities = {cam: [] for cam in camera_keys}
+
+    for idx in tqdm(sample_indices, desc="Validating"):
+        idx = int(idx)
+
+        embeddings_item = embeddings_dataset[idx]
+        precomputed_task_emb = embeddings_item["task_embedding"].numpy()
+        precomputed_image_embs = {cam: embeddings_item[f"{cam}_embedding"].numpy() for cam in camera_keys}
+
+        original_item = original_dataset[idx]
+
+        # Get task and compute embedding
+        task = original_item["task"]
+        # Clean up task text (same as in generate_embeddings.py)
+        task_clean = task.strip().capitalize().strip(" .,!?-_")
+        onthefly_task_emb = language_encoder.encode([task_clean])[0]
+
+        # Get images and compute embeddings
+        onthefly_image_embs = {}
+        for cam in camera_keys:
+            img = original_item[cam]
+            # Convert to numpy if needed
+            if isinstance(img, torch.Tensor):
+                if img.ndim == 3:  # (C, H, W)
+                    img_np = (img.permute(1, 2, 0).numpy() * 255).astype(np.uint8)
+                else:
+                    raise ValueError(f"Unexpected image shape: {img.shape}")
+            else:
+                img_np = np.array(img)
+
+            onthefly_image_embs[cam] = image_encoder.encode([img_np])[0]
+
+        # Task embedding comparison
+        task_sim = cosine_similarity(precomputed_task_emb, onthefly_task_emb)
+        task_similarities.append(task_sim)
+
+        # Image embedding comparison
+        for cam in camera_keys:
+            img_sim = cosine_similarity(precomputed_image_embs[cam], onthefly_image_embs[cam])
+            image_similarities[cam].append(img_sim)
+
+    # Results
+    print("\nResults:")
+    task_sim_threshold = 0.99
+    img_sim_threshold = 0.99
+
+    task_mean_sim = np.mean(task_similarities)
+    task_pass = task_mean_sim >= task_sim_threshold
+
+    print(f"  Task: {task_mean_sim:.4f} {'✓' if task_pass else '✗'}")
+
+    for cam in camera_keys:
+        cam_mean_sim = np.mean(image_similarities[cam])
+        cam_pass = cam_mean_sim >= img_sim_threshold
+        print(f"  {cam}: {cam_mean_sim:.4f} {'✓' if cam_pass else '✗'}")
+
+    image_pass = all(np.mean(image_similarities[cam]) >= img_sim_threshold for cam in camera_keys)
+
+    print()
+    if task_pass and image_pass:
+        print("✓ PASSED")
+    else:
+        print("✗ FAILED")
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Validate and compare pre-computed embeddings with on-the-fly embeddings",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+        epilog="""
+Example:
+  python src/lerobot/datasets/generating_embeddings/validate_embeddings.py \\
+      --original-repo-id lerobot/utokyo_xarm_bimanual \\
+      --embeddings-repo-id lerobot/utokyo_xarm_bimanual_embeddings \\
+      --image-encoder dinov2_vitb14 \\
+      --language-encoder minilm-l12 \\
+      --num-samples 20
+        """,
+    )
+    parser.add_argument("--original-repo-id", type=str, required=True, help="Original dataset repository ID")
+    parser.add_argument(
+        "--embeddings-repo-id",
+        type=str,
+        required=True,
+        help="Dataset with pre-computed embeddings repository ID",
+    )
+    parser.add_argument(
+        "--image-encoder",
+        type=str,
+        default="dinov2_vitb14",
+        help="Image encoder to use (default: dinov2_vitb14)",
+    )
+    parser.add_argument(
+        "--language-encoder",
+        type=str,
+        default="minilm-l12",
+        help="Language encoder to use (default: minilm-l12)",
+    )
+    parser.add_argument(
+        "--num-samples",
+        type=int,
+        default=10,
+        help="Number of samples to validate (default: 10)",
+    )
+    parser.add_argument(
+        "--device",
+        type=str,
+        default="cuda",
+        help="Device to use for encoding (default: cuda)",
+    )
+
+    args = parser.parse_args()
+
+    # Load encoders
+    image_encoder = get_image_encoder(args.image_encoder, device=args.device)
+    language_encoder = get_language_encoder(args.language_encoder, device=args.device)
+
+    # Validate embeddings
+    validate_embeddings(
+        original_repo_id=args.original_repo_id,
+        embeddings_repo_id=args.embeddings_repo_id,
+        image_encoder=image_encoder,
+        language_encoder=language_encoder,
+        num_samples=args.num_samples,
+        device=args.device,
+    )
+
+
+if __name__ == "__main__":
+    main()

src/lerobot/datasets/lerobot_dataset.py

@@ -430,9 +430,7 @@ class LeRobotDatasetMetadata:
         video_keys = [video_key] if video_key is not None else self.video_keys
         for key in video_keys:
             if not self.features[key].get("info", None):
-                video_path = self.root / self.video_path.format(
-                    video_key=video_key, chunk_index=0, file_index=0
-                )
+                video_path = self.root / self.video_path.format(video_key=key, chunk_index=0, file_index=0)
                 self.info["features"][key]["info"] = get_video_info(video_path)
 
     def update_chunk_settings(
@@ -686,6 +684,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self.episode_buffer = None
         self.writer = None
         self.latest_episode = None
+        self._current_file_start_frame = None  # Track the starting frame index of the current parquet file
 
         self.root.mkdir(exist_ok=True, parents=True)
 
@@ -708,10 +707,20 @@ class LeRobotDataset(torch.utils.data.Dataset):
             if not self._check_cached_episodes_sufficient():
                 raise FileNotFoundError("Cached dataset doesn't contain all requested episodes")
         except (AssertionError, FileNotFoundError, NotADirectoryError):
-            self.revision = get_safe_version(self.repo_id, self.revision)
+            if is_valid_version(self.revision):
+                self.revision = get_safe_version(self.repo_id, self.revision)
             self.download(download_videos)
             self.hf_dataset = self.load_hf_dataset()
 
+        # Create mapping from absolute indices to relative indices when only a subset of the episodes are loaded
+        # Build a mapping: absolute_index -> relative_index_in_filtered_dataset
+        self._absolute_to_relative_idx = None
+        if self.episodes is not None:
+            self._absolute_to_relative_idx = {
+                abs_idx.item() if isinstance(abs_idx, torch.Tensor) else abs_idx: rel_idx
+                for rel_idx, abs_idx in enumerate(self.hf_dataset["index"])
+            }
+
         # Setup delta_indices
         if self.delta_timestamps is not None:
             check_delta_timestamps(self.delta_timestamps, self.fps, self.tolerance_s)
@@ -830,31 +839,40 @@ class LeRobotDataset(torch.utils.data.Dataset):
     def load_hf_dataset(self) -> datasets.Dataset:
         """hf_dataset contains all the observations, states, actions, rewards, etc."""
         features = get_hf_features_from_features(self.features)
-        hf_dataset = load_nested_dataset(self.root / "data", features=features)
+        hf_dataset = load_nested_dataset(self.root / "data", features=features, episodes=self.episodes)
         hf_dataset.set_transform(hf_transform_to_torch)
         return hf_dataset
 
     def _check_cached_episodes_sufficient(self) -> bool:
-        """Check if the cached dataset contains all requested episodes."""
+        """Check if the cached dataset contains all requested episodes and their video files."""
         if self.hf_dataset is None or len(self.hf_dataset) == 0:
             return False
 
         # Get available episode indices from cached dataset
         available_episodes = {
             ep_idx.item() if isinstance(ep_idx, torch.Tensor) else ep_idx
-            for ep_idx in self.hf_dataset["episode_index"]
+            for ep_idx in self.hf_dataset.unique("episode_index")
         }
 
         # Determine requested episodes
         if self.episodes is None:
-            # Requesting all episodes - check if we have all episodes from metadata
             requested_episodes = set(range(self.meta.total_episodes))
         else:
-            # Requesting specific episodes
             requested_episodes = set(self.episodes)
 
         # Check if all requested episodes are available in cached data
-        return requested_episodes.issubset(available_episodes)
+        if not requested_episodes.issubset(available_episodes):
+            return False
+
+        # Check if all required video files exist
+        if len(self.meta.video_keys) > 0:
+            for ep_idx in requested_episodes:
+                for vid_key in self.meta.video_keys:
+                    video_path = self.root / self.meta.get_video_file_path(ep_idx, vid_key)
+                    if not video_path.exists():
+                        return False
+
+        return True
 
     def create_hf_dataset(self) -> datasets.Dataset:
         features = get_hf_features_from_features(self.features)
@@ -921,7 +939,11 @@ class LeRobotDataset(torch.utils.data.Dataset):
         query_timestamps = {}
         for key in self.meta.video_keys:
             if query_indices is not None and key in query_indices:
-                timestamps = self.hf_dataset[query_indices[key]]["timestamp"]
+                if self._absolute_to_relative_idx is not None:
+                    relative_indices = [self._absolute_to_relative_idx[idx] for idx in query_indices[key]]
+                    timestamps = self.hf_dataset[relative_indices]["timestamp"]
+                else:
+                    timestamps = self.hf_dataset[query_indices[key]]["timestamp"]
                 query_timestamps[key] = torch.stack(timestamps).tolist()
             else:
                 query_timestamps[key] = [current_ts]
@@ -929,11 +951,32 @@ class LeRobotDataset(torch.utils.data.Dataset):
         return query_timestamps
 
     def _query_hf_dataset(self, query_indices: dict[str, list[int]]) -> dict:
-        return {
-            key: torch.stack(self.hf_dataset[q_idx][key])
-            for key, q_idx in query_indices.items()
-            if key not in self.meta.video_keys
-        }
+        """
+        Query dataset for indices across keys, skipping video keys.
+
+        Tries column-first [key][indices] for speed, falls back to row-first.
+
+        Args:
+            query_indices: Dict mapping keys to index lists to retrieve
+
+        Returns:
+            Dict with stacked tensors of queried data (video keys excluded)
+        """
+        result: dict = {}
+        for key, q_idx in query_indices.items():
+            if key in self.meta.video_keys:
+                continue
+            # Map absolute indices to relative indices if needed
+            relative_indices = (
+                q_idx
+                if self._absolute_to_relative_idx is None
+                else [self._absolute_to_relative_idx[idx] for idx in q_idx]
+            )
+            try:
+                result[key] = torch.stack(self.hf_dataset[key][relative_indices])
+            except (KeyError, TypeError, IndexError):
+                result[key] = torch.stack(self.hf_dataset[relative_indices][key])
+        return result
 
     def _query_videos(self, query_timestamps: dict[str, list[float]], ep_idx: int) -> dict[str, torch.Tensor]:
         """Note: When using data workers (e.g. DataLoader with num_workers>0), do not call this function
@@ -1231,6 +1274,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             # Initialize indices and frame count for a new dataset made of the first episode data
             chunk_idx, file_idx = 0, 0
             global_frame_index = 0
+            self._current_file_start_frame = 0
             # However, if the episodes already exists
             # It means we are resuming recording, so we need to load the latest episode
             # Update the indices to avoid overwriting the latest episode
@@ -1242,6 +1286,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
 
                 # When resuming, move to the next file
                 chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, self.meta.chunks_size)
+                self._current_file_start_frame = global_frame_index
         else:
             # Retrieve information from the latest parquet file
             latest_ep = self.latest_episode
@@ -1252,7 +1297,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             latest_path = self.root / self.meta.data_path.format(chunk_index=chunk_idx, file_index=file_idx)
             latest_size_in_mb = get_file_size_in_mb(latest_path)
 
-            frames_in_current_file = global_frame_index - latest_ep["dataset_from_index"]
+            frames_in_current_file = global_frame_index - self._current_file_start_frame
             av_size_per_frame = (
                 latest_size_in_mb / frames_in_current_file if frames_in_current_file > 0 else 0
             )
@@ -1266,6 +1311,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, self.meta.chunks_size)
                 self._close_writer()
                 self._writer_closed_for_reading = False
+                self._current_file_start_frame = global_frame_index
 
         ep_dict["data/chunk_index"] = chunk_idx
         ep_dict["data/file_index"] = file_idx
@@ -1469,9 +1515,11 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj.image_transforms = None
         obj.delta_timestamps = None
         obj.delta_indices = None
+        obj._absolute_to_relative_idx = None
         obj.video_backend = video_backend if video_backend is not None else get_safe_default_codec()
         obj.writer = None
         obj.latest_episode = None
+        obj._current_file_start_frame = None
         # Initialize tracking for incremental recording
         obj._lazy_loading = False
         obj._recorded_frames = 0

src/lerobot/datasets/temporal_sampler.py

@@ -0,0 +1,151 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+SARM Temporal Sampler for reward model training.
+
+Samples frames uniformly from episodes for SARM's 9-frame symmetric pattern:
+- 1 initial frame + 4 frames before + current + 3 frames after
+
+Boundary handling: clamp to first/last frame when indices go out of bounds.
+This enables truly uniform sampling across entire episodes.
+"""
+
+import logging
+from typing import Iterator, Optional
+import numpy as np
+import torch
+from torch.utils.data import Sampler
+import random
+
+
+class SARMTemporalSampler(Sampler):
+    """
+    Temporal sampler for SARM reward model training with symmetric/bidirectional sampling.
+    
+    SARM uses 9 frames per sample:
+    - Frame 0: Initial frame of the episode (always frame 0)
+    - Frames 1-8: Symmetric context around current frame
+      Pattern: [t-4*gap, t-3*gap, t-2*gap, t-gap, t, t+gap, t+2*gap, t+3*gap]
+    
+    Boundary handling:
+    - Early frames: backward indices clamp to 0 (e.g., [0,0,0,5,35,65,95,125])
+    - Late frames: forward indices clamp to last frame (e.g., [850,880,910,940,970,1000,1000,1000])
+    
+    This enables truly uniform sampling across entire episodes.
+    
+    Args:
+        dataset_from_index: Start indices of episodes (global dataset indices)
+        dataset_to_index: End indices of episodes (global dataset indices)
+        frame_gap: Gap between consecutive frames (default: 30 = 1 second at 30fps)
+        shuffle: Whether to shuffle sampling order
+        seed: Random seed for reproducibility
+        samples_per_epoch: Number of samples per epoch (default: 6400)
+        min_episode_length: Minimum episode length to include (default: 1)
+    """
+    
+    def __init__(
+        self,
+        dataset_from_index: np.ndarray,
+        dataset_to_index: np.ndarray,
+        frame_gap: int = 30,
+        shuffle: bool = True,
+        seed: Optional[int] = None,
+        samples_per_epoch: int = 6400,
+        min_episode_length: int = 1,
+    ):
+        self.dataset_from_index = np.array(dataset_from_index)
+        self.dataset_to_index = np.array(dataset_to_index)
+        self.frame_gap = frame_gap
+        self.shuffle = shuffle
+        self.samples_per_epoch = samples_per_epoch
+        self.min_episode_length = min_episode_length
+        
+        if seed is not None:
+            self.seed = seed
+            random.seed(seed)
+            np.random.seed(seed)
+            self.generator = torch.Generator().manual_seed(seed)
+        else:
+            self.generator = torch.Generator()
+        
+        # Compute valid episodes and sampling positions (ALL frames for uniform sampling)
+        self._compute_valid_positions()
+        
+        logging.info(
+            f"SARMTemporalSampler: {len(self.valid_episodes)} valid episodes, "
+            f"{len(self.all_valid_positions)} positions (uniform sampling), "
+            f"{self.samples_per_epoch} samples per epoch, "
+            f"frame_gap={frame_gap}, symmetric bidirectional pattern"
+        )
+    
+    def _compute_valid_positions(self):
+        """Compute valid episodes and ALL sampling positions for uniform sampling.
+        
+        With symmetric bidirectional sampling, we can sample from ANY frame:
+        - Early frames: backward indices clamp to first frame
+        - Late frames: forward indices clamp to last frame
+        """
+        self.valid_episodes = []
+        self.all_valid_positions = []
+        
+        for ep_idx in range(len(self.dataset_from_index)):
+            ep_start = self.dataset_from_index[ep_idx]
+            ep_end = self.dataset_to_index[ep_idx]
+            episode_length = ep_end - ep_start
+            
+            # Include all episodes with at least min_episode_length frames
+            if episode_length >= self.min_episode_length:
+                self.valid_episodes.append((ep_idx, ep_start, ep_end))
+                
+                # Include ALL positions in the episode (truly uniform sampling)
+                for pos in range(ep_start, ep_end):
+                    self.all_valid_positions.append(pos)
+        
+        self.valid_episodes = np.array(self.valid_episodes)
+        self.all_valid_positions = np.array(self.all_valid_positions)
+        
+        if len(self.all_valid_positions) == 0:
+            raise ValueError(
+                f"No valid sampling positions found! "
+                f"Check that episodes have at least {self.min_episode_length} frames."
+            )
+    
+    def __len__(self) -> int:
+        return self.samples_per_epoch
+    
+    def __iter__(self) -> Iterator[int]:
+        """
+        Yields global dataset indices for uniform sampling across episodes.
+        
+        Each yielded index represents the "current frame" position.
+        The dataset's observation_delta_indices then handles loading:
+        - Frame 0: Episode initial frame (via large negative delta clamping)
+        - Frames 1-8: Symmetric context around current frame (with boundary clamping)
+        
+        For early frames: backward indices clamp to first frame (progress ~0%)
+        For late frames: forward indices clamp to last frame (progress ~100%)
+        """
+        if self.shuffle:
+            # Randomly sample from all valid positions
+            for _ in range(self.samples_per_epoch):
+                idx = np.random.randint(0, len(self.all_valid_positions))
+                yield int(self.all_valid_positions[idx])
+        else:
+            # Sequential sampling with wrap-around
+            for i in range(self.samples_per_epoch):
+                idx = i % len(self.all_valid_positions)
+                yield int(self.all_valid_positions[idx])

src/lerobot/datasets/transforms.py

@@ -206,6 +206,11 @@ class ImageTransformsConfig:
                 type="SharpnessJitter",
                 kwargs={"sharpness": (0.5, 1.5)},
             ),
+            "affine": ImageTransformConfig(
+                weight=1.0,
+                type="RandomAffine",
+                kwargs={"degrees": (-5.0, 5.0), "translate": (0.05, 0.05)},
+            ),
         }
     )
 
@@ -217,6 +222,8 @@ def make_transform_from_config(cfg: ImageTransformConfig):
         return v2.ColorJitter(**cfg.kwargs)
     elif cfg.type == "SharpnessJitter":
         return SharpnessJitter(**cfg.kwargs)
+    elif cfg.type == "RandomAffine":
+        return v2.RandomAffine(**cfg.kwargs)
     else:
         raise ValueError(f"Transform '{cfg.type}' is not valid.")
 

src/lerobot/datasets/utils.py

@@ -28,6 +28,7 @@ import numpy as np
 import packaging.version
 import pandas
 import pandas as pd
+import pyarrow.dataset as pa_ds
 import pyarrow.parquet as pq
 import torch
 from datasets import Dataset
@@ -103,7 +104,9 @@ def update_chunk_file_indices(chunk_idx: int, file_idx: int, chunks_size: int) -
     return chunk_idx, file_idx
 
 
-def load_nested_dataset(pq_dir: Path, features: datasets.Features | None = None) -> Dataset:
+def load_nested_dataset(
+    pq_dir: Path, features: datasets.Features | None = None, episodes: list[int] | None = None
+) -> Dataset:
     """Find parquet files in provided directory {pq_dir}/chunk-xxx/file-xxx.parquet
     Convert parquet files to pyarrow memory mapped in a cache folder for efficient RAM usage
     Concatenate all pyarrow references to return HF Dataset format
@@ -111,15 +114,26 @@ def load_nested_dataset(pq_dir: Path, features: datasets.Features | None = None)
     Args:
         pq_dir: Directory containing parquet files
         features: Optional features schema to ensure consistent loading of complex types like images
+        episodes: Optional list of episode indices to filter. Uses PyArrow predicate pushdown for efficiency.
     """
     paths = sorted(pq_dir.glob("*/*.parquet"))
     if len(paths) == 0:
         raise FileNotFoundError(f"Provided directory does not contain any parquet file: {pq_dir}")
 
-    # TODO(rcadene): set num_proc to accelerate conversion to pyarrow
     with SuppressProgressBars():
-        datasets = Dataset.from_parquet([str(path) for path in paths], features=features)
-    return datasets
+        # When no filtering needed, Dataset uses memory-mapped loading for efficiency
+        # PyArrow loads the entire dataset into memory
+        if episodes is None:
+            return Dataset.from_parquet([str(path) for path in paths], features=features)
+
+        arrow_dataset = pa_ds.dataset(paths, format="parquet")
+        filter_expr = pa_ds.field("episode_index").isin(episodes)
+        table = arrow_dataset.to_table(filter=filter_expr)
+
+        if features is not None:
+            table = table.cast(features.arrow_schema)
+
+        return Dataset(table)
 
 
 def get_parquet_num_frames(parquet_path: str | Path) -> int:

src/lerobot/datasets/v30/convert_dataset_v21_to_v30.py

@@ -69,9 +69,9 @@ from lerobot.datasets.utils import (
     LEGACY_TASKS_PATH,
     cast_stats_to_numpy,
     flatten_dict,
+    get_file_size_in_mb,
     get_parquet_file_size_in_mb,
     get_parquet_num_frames,
-    get_video_size_in_mb,
     load_info,
     update_chunk_file_indices,
     write_episodes,
@@ -98,7 +98,7 @@ OLD
 videos/chunk-000/CAMERA/episode_000000.mp4
 
 NEW
-videos/chunk-000/file_000.mp4
+videos/CAMERA/chunk-000/file_000.mp4
 -------------------------
 OLD
 episodes.jsonl
@@ -310,7 +310,7 @@ def convert_videos_of_camera(root: Path, new_root: Path, video_key: str, video_f
     episodes_metadata = []
 
     for ep_path in tqdm.tqdm(ep_paths, desc=f"convert videos of {video_key}"):
-        ep_size_in_mb = get_video_size_in_mb(ep_path)
+        ep_size_in_mb = get_file_size_in_mb(ep_path)
         ep_duration_in_s = get_video_duration_in_s(ep_path)
 
         # Check if adding this episode would exceed the limit

src/lerobot/datasets/video_utils.py

@@ -342,8 +342,8 @@ def encode_video_frames(
     # Define video output frame size (assuming all input frames are the same size)
     if len(input_list) == 0:
         raise FileNotFoundError(f"No images found in {imgs_dir}.")
-    dummy_image = Image.open(input_list[0])
-    width, height = dummy_image.size
+    with Image.open(input_list[0]) as dummy_image:
+        width, height = dummy_image.size
 
     # Define video codec options
     video_options = {}
@@ -373,11 +373,12 @@ def encode_video_frames(
 
         # Loop through input frames and encode them
         for input_data in input_list:
-            input_image = Image.open(input_data).convert("RGB")
-            input_frame = av.VideoFrame.from_image(input_image)
-            packet = output_stream.encode(input_frame)
-            if packet:
-                output.mux(packet)
+            with Image.open(input_data) as input_image:
+                input_image = input_image.convert("RGB")
+                input_frame = av.VideoFrame.from_image(input_image)
+                packet = output_stream.encode(input_frame)
+                if packet:
+                    output.mux(packet)
 
         # Flush the encoder
         packet = output_stream.encode()

src/lerobot/envs/__init__.py

@@ -12,4 +12,4 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .configs import AlohaEnv, EnvConfig, PushtEnv, XarmEnv  # noqa: F401
+from .configs import AlohaEnv, EnvConfig, PushtEnv  # noqa: F401

src/lerobot/envs/configs.py

@@ -21,7 +21,22 @@ import draccus
 from lerobot.configs.types import FeatureType, PolicyFeature
 from lerobot.robots import RobotConfig
 from lerobot.teleoperators.config import TeleoperatorConfig
-from lerobot.utils.constants import ACTION, OBS_ENV_STATE, OBS_IMAGE, OBS_IMAGES, OBS_STATE
+from lerobot.utils.constants import (
+    ACTION,
+    LIBERO_KEY_EEF_MAT,
+    LIBERO_KEY_EEF_POS,
+    LIBERO_KEY_EEF_QUAT,
+    LIBERO_KEY_GRIPPER_QPOS,
+    LIBERO_KEY_GRIPPER_QVEL,
+    LIBERO_KEY_JOINTS_POS,
+    LIBERO_KEY_JOINTS_VEL,
+    LIBERO_KEY_PIXELS_AGENTVIEW,
+    LIBERO_KEY_PIXELS_EYE_IN_HAND,
+    OBS_ENV_STATE,
+    OBS_IMAGE,
+    OBS_IMAGES,
+    OBS_STATE,
+)
 
 
 @dataclass
@@ -37,6 +52,16 @@ class EnvConfig(draccus.ChoiceRegistry, abc.ABC):
     def type(self) -> str:
         return self.get_choice_name(self.__class__)
 
+    @property
+    def package_name(self) -> str:
+        """Package name to import if environment not found in gym registry"""
+        return f"gym_{self.type}"
+
+    @property
+    def gym_id(self) -> str:
+        """ID string used in gym.make() to instantiate the environment"""
+        return f"{self.package_name}/{self.task}"
+
     @property
     @abc.abstractmethod
     def gym_kwargs(self) -> dict:
@@ -133,45 +158,6 @@ class PushtEnv(EnvConfig):
         }
 
 
-@EnvConfig.register_subclass("xarm")
-@dataclass
-class XarmEnv(EnvConfig):
-    task: str | None = "XarmLift-v0"
-    fps: int = 15
-    episode_length: int = 200
-    obs_type: str = "pixels_agent_pos"
-    render_mode: str = "rgb_array"
-    visualization_width: int = 384
-    visualization_height: int = 384
-    features: dict[str, PolicyFeature] = field(
-        default_factory=lambda: {
-            ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(4,)),
-            "pixels": PolicyFeature(type=FeatureType.VISUAL, shape=(84, 84, 3)),
-        }
-    )
-    features_map: dict[str, str] = field(
-        default_factory=lambda: {
-            ACTION: ACTION,
-            "agent_pos": OBS_STATE,
-            "pixels": OBS_IMAGE,
-        }
-    )
-
-    def __post_init__(self):
-        if self.obs_type == "pixels_agent_pos":
-            self.features["agent_pos"] = PolicyFeature(type=FeatureType.STATE, shape=(4,))
-
-    @property
-    def gym_kwargs(self) -> dict:
-        return {
-            "obs_type": self.obs_type,
-            "render_mode": self.render_mode,
-            "visualization_width": self.visualization_width,
-            "visualization_height": self.visualization_height,
-            "max_episode_steps": self.episode_length,
-        }
-
-
 @dataclass
 class ImagePreprocessingConfig:
     crop_params_dict: dict[str, tuple[int, int, int, int]] | None = None
@@ -275,28 +261,103 @@ class LiberoEnv(EnvConfig):
     features_map: dict[str, str] = field(
         default_factory=lambda: {
             ACTION: ACTION,
-            "agent_pos": OBS_STATE,
-            "pixels/agentview_image": f"{OBS_IMAGES}.image",
-            "pixels/robot0_eye_in_hand_image": f"{OBS_IMAGES}.image2",
+            LIBERO_KEY_EEF_POS: f"{OBS_STATE}.eef_pos",
+            LIBERO_KEY_EEF_QUAT: f"{OBS_STATE}.eef_quat",
+            LIBERO_KEY_EEF_MAT: f"{OBS_STATE}.eef_mat",
+            LIBERO_KEY_GRIPPER_QPOS: f"{OBS_STATE}.gripper_qpos",
+            LIBERO_KEY_GRIPPER_QVEL: f"{OBS_STATE}.gripper_qvel",
+            LIBERO_KEY_JOINTS_POS: f"{OBS_STATE}.joint_pos",
+            LIBERO_KEY_JOINTS_VEL: f"{OBS_STATE}.joint_vel",
+            LIBERO_KEY_PIXELS_AGENTVIEW: f"{OBS_IMAGES}.image",
+            LIBERO_KEY_PIXELS_EYE_IN_HAND: f"{OBS_IMAGES}.image2",
         }
     )
 
     def __post_init__(self):
         if self.obs_type == "pixels":
-            self.features["pixels/agentview_image"] = PolicyFeature(
+            self.features[LIBERO_KEY_PIXELS_AGENTVIEW] = PolicyFeature(
                 type=FeatureType.VISUAL, shape=(self.observation_height, self.observation_width, 3)
             )
-            self.features["pixels/robot0_eye_in_hand_image"] = PolicyFeature(
+            self.features[LIBERO_KEY_PIXELS_EYE_IN_HAND] = PolicyFeature(
                 type=FeatureType.VISUAL, shape=(self.observation_height, self.observation_width, 3)
             )
         elif self.obs_type == "pixels_agent_pos":
-            self.features["agent_pos"] = PolicyFeature(type=FeatureType.STATE, shape=(8,))
-            self.features["pixels/agentview_image"] = PolicyFeature(
+            self.features[LIBERO_KEY_PIXELS_AGENTVIEW] = PolicyFeature(
                 type=FeatureType.VISUAL, shape=(self.observation_height, self.observation_width, 3)
             )
-            self.features["pixels/robot0_eye_in_hand_image"] = PolicyFeature(
+            self.features[LIBERO_KEY_PIXELS_EYE_IN_HAND] = PolicyFeature(
                 type=FeatureType.VISUAL, shape=(self.observation_height, self.observation_width, 3)
             )
+            self.features[LIBERO_KEY_EEF_POS] = PolicyFeature(
+                type=FeatureType.STATE,
+                shape=(3,),
+            )
+            self.features[LIBERO_KEY_EEF_QUAT] = PolicyFeature(
+                type=FeatureType.STATE,
+                shape=(4,),
+            )
+            self.features[LIBERO_KEY_EEF_MAT] = PolicyFeature(
+                type=FeatureType.STATE,
+                shape=(3, 3),
+            )
+            self.features[LIBERO_KEY_GRIPPER_QPOS] = PolicyFeature(
+                type=FeatureType.STATE,
+                shape=(2,),
+            )
+            self.features[LIBERO_KEY_GRIPPER_QVEL] = PolicyFeature(
+                type=FeatureType.STATE,
+                shape=(2,),
+            )
+            self.features[LIBERO_KEY_JOINTS_POS] = PolicyFeature(
+                type=FeatureType.STATE,
+                shape=(7,),
+            )
+            self.features[LIBERO_KEY_JOINTS_VEL] = PolicyFeature(
+                type=FeatureType.STATE,
+                shape=(7,),
+            )
+        else:
+            raise ValueError(f"Unsupported obs_type: {self.obs_type}")
+
+    @property
+    def gym_kwargs(self) -> dict:
+        return {
+            "obs_type": self.obs_type,
+            "render_mode": self.render_mode,
+        }
+
+
+@EnvConfig.register_subclass("metaworld")
+@dataclass
+class MetaworldEnv(EnvConfig):
+    task: str = "metaworld-push-v2"  # add all tasks
+    fps: int = 80
+    episode_length: int = 400
+    obs_type: str = "pixels_agent_pos"
+    render_mode: str = "rgb_array"
+    multitask_eval: bool = True
+    features: dict[str, PolicyFeature] = field(
+        default_factory=lambda: {
+            "action": PolicyFeature(type=FeatureType.ACTION, shape=(4,)),
+        }
+    )
+    features_map: dict[str, str] = field(
+        default_factory=lambda: {
+            "action": ACTION,
+            "agent_pos": OBS_STATE,
+            "top": f"{OBS_IMAGE}",
+            "pixels/top": f"{OBS_IMAGE}",
+        }
+    )
+
+    def __post_init__(self):
+        if self.obs_type == "pixels":
+            self.features["top"] = PolicyFeature(type=FeatureType.VISUAL, shape=(480, 480, 3))
+
+        elif self.obs_type == "pixels_agent_pos":
+            self.features["agent_pos"] = PolicyFeature(type=FeatureType.STATE, shape=(4,))
+            self.features["pixels/top"] = PolicyFeature(type=FeatureType.VISUAL, shape=(480, 480, 3))
+
         else:
             raise ValueError(f"Unsupported obs_type: {self.obs_type}")
 

src/lerobot/envs/factory.py

@@ -14,10 +14,16 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import importlib
+from typing import Any
 
 import gymnasium as gym
+from gymnasium.envs.registration import registry as gym_registry
 
-from lerobot.envs.configs import AlohaEnv, EnvConfig, LiberoEnv, PushtEnv, XarmEnv
+from lerobot.envs.configs import AlohaEnv, EnvConfig, LiberoEnv, PushtEnv
+from lerobot.envs.utils import _call_make_env, _download_hub_file, _import_hub_module, _normalize_hub_result
+from lerobot.processor import ProcessorStep
+from lerobot.processor.env_processor import LiberoProcessorStep
+from lerobot.processor.pipeline import PolicyProcessorPipeline
 
 
 def make_env_config(env_type: str, **kwargs) -> EnvConfig:
@@ -25,24 +31,66 @@ def make_env_config(env_type: str, **kwargs) -> EnvConfig:
         return AlohaEnv(**kwargs)
     elif env_type == "pusht":
         return PushtEnv(**kwargs)
-    elif env_type == "xarm":
-        return XarmEnv(**kwargs)
     elif env_type == "libero":
         return LiberoEnv(**kwargs)
     else:
         raise ValueError(f"Policy type '{env_type}' is not available.")
 
 
-def make_env(
-    cfg: EnvConfig, n_envs: int = 1, use_async_envs: bool = False
-) -> dict[str, dict[int, gym.vector.VectorEnv]]:
-    """Makes a gym vector environment according to the config.
+def make_env_pre_post_processors(
+    env_cfg: EnvConfig,
+) -> tuple[
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+]:
+    """
+    Create preprocessor and postprocessor pipelines for environment observations.
+
+    This function creates processor pipelines that transform raw environment
+    observations and actions. By default, it returns identity processors that do nothing.
+    For specific environments like LIBERO, it adds environment-specific processing steps.
 
     Args:
-        cfg (EnvConfig): the config of the environment to instantiate.
+        env_cfg: The configuration of the environment.
+
+    Returns:
+        A tuple containing:
+            - preprocessor: Pipeline that processes environment observations
+            - postprocessor: Pipeline that processes environment outputs (currently identity)
+    """
+    # Preprocessor and Postprocessor steps are Identity for most environments
+    preprocessor_steps: list[ProcessorStep] = []
+    postprocessor_steps: list[ProcessorStep] = []
+
+    # For LIBERO environments, add the LiberoProcessorStep to preprocessor
+    if isinstance(env_cfg, LiberoEnv) or "libero" in env_cfg.type:
+        preprocessor_steps.append(LiberoProcessorStep())
+
+    preprocessor = PolicyProcessorPipeline(steps=preprocessor_steps)
+    postprocessor = PolicyProcessorPipeline(steps=postprocessor_steps)
+
+    return preprocessor, postprocessor
+
+
+def make_env(
+    cfg: EnvConfig | str,
+    n_envs: int = 1,
+    use_async_envs: bool = False,
+    hub_cache_dir: str | None = None,
+    trust_remote_code: bool = False,
+) -> dict[str, dict[int, gym.vector.VectorEnv]]:
+    """Makes a gym vector environment according to the config or Hub reference.
+
+    Args:
+        cfg (EnvConfig | str): Either an `EnvConfig` object describing the environment to build locally,
+            or a Hugging Face Hub repository identifier (e.g. `"username/repo"`). In the latter case,
+            the repo must include a Python file (usually `env.py`).
         n_envs (int, optional): The number of parallelized env to return. Defaults to 1.
         use_async_envs (bool, optional): Whether to return an AsyncVectorEnv or a SyncVectorEnv. Defaults to
             False.
+        hub_cache_dir (str | None): Optional cache path for downloaded hub files.
+        trust_remote_code (bool): **Explicit consent** to execute remote code from the Hub.
+            Default False — must be set to True to import/exec hub `env.py`.
 
     Raises:
         ValueError: if n_envs < 1
@@ -55,6 +103,21 @@ def make_env(
             - For single-task environments: a single suite entry (cfg.type) with task_id=0.
 
     """
+    # if user passed a hub id string (e.g., "username/repo", "username/repo@main:env.py")
+    # simplified: only support hub-provided `make_env`
+    if isinstance(cfg, str):
+        # _download_hub_file will raise the same RuntimeError if trust_remote_code is False
+        repo_id, file_path, local_file, revision = _download_hub_file(cfg, trust_remote_code, hub_cache_dir)
+
+        # import and surface clear import errors
+        module = _import_hub_module(local_file, repo_id)
+
+        # call the hub-provided make_env
+        raw_result = _call_make_env(module, n_envs=n_envs, use_async_envs=use_async_envs)
+
+        # normalize the return into {suite: {task_id: vec_env}}
+        return _normalize_hub_result(raw_result)
+
     if n_envs < 1:
         raise ValueError("`n_envs` must be at least 1")
 
@@ -74,20 +137,38 @@ def make_env(
             gym_kwargs=cfg.gym_kwargs,
             env_cls=env_cls,
         )
+    elif "metaworld" in cfg.type:
+        from lerobot.envs.metaworld import create_metaworld_envs
 
-    package_name = f"gym_{cfg.type}"
-    try:
-        importlib.import_module(package_name)
-    except ModuleNotFoundError as e:
-        print(f"{package_name} is not installed. Please install it with `pip install 'lerobot[{cfg.type}]'`")
-        raise e
+        if cfg.task is None:
+            raise ValueError("MetaWorld requires a task to be specified")
 
-    gym_handle = f"{package_name}/{cfg.task}"
+        return create_metaworld_envs(
+            task=cfg.task,
+            n_envs=n_envs,
+            gym_kwargs=cfg.gym_kwargs,
+            env_cls=env_cls,
+        )
+
+    if cfg.gym_id not in gym_registry:
+        print(f"gym id '{cfg.gym_id}' not found, attempting to import '{cfg.package_name}'...")
+        try:
+            importlib.import_module(cfg.package_name)
+        except ModuleNotFoundError as e:
+            raise ModuleNotFoundError(
+                f"Package '{cfg.package_name}' required for env '{cfg.type}' not found. "
+                f"Please install it or check PYTHONPATH."
+            ) from e
+
+        if cfg.gym_id not in gym_registry:
+            raise gym.error.NameNotFound(
+                f"Environment '{cfg.gym_id}' not registered even after importing '{cfg.package_name}'."
+            )
 
     def _make_one():
-        return gym.make(gym_handle, disable_env_checker=cfg.disable_env_checker, **(cfg.gym_kwargs or {}))
+        return gym.make(cfg.gym_id, disable_env_checker=cfg.disable_env_checker, **(cfg.gym_kwargs or {}))
 
-    vec = env_cls([_make_one for _ in range(n_envs)])
+    vec = env_cls([_make_one for _ in range(n_envs)], autoreset_mode=gym.vector.AutoresetMode.SAME_STEP)
 
     # normalize to {suite: {task_id: vec_env}} for consistency
     suite_name = cfg.type  # e.g., "pusht", "aloha"

src/lerobot/envs/libero.py

@@ -28,7 +28,6 @@ import torch
 from gymnasium import spaces
 from libero.libero import benchmark, get_libero_path
 from libero.libero.envs import OffScreenRenderEnv
-from robosuite.utils.transform_utils import quat2axisangle
 
 
 def _parse_camera_names(camera_name: str | Sequence[str]) -> list[str]:
@@ -175,11 +174,36 @@ class LiberoEnv(gym.Env):
             self.observation_space = spaces.Dict(
                 {
                     "pixels": spaces.Dict(images),
-                    "agent_pos": spaces.Box(
-                        low=AGENT_POS_LOW,
-                        high=AGENT_POS_HIGH,
-                        shape=(OBS_STATE_DIM,),
-                        dtype=np.float64,
+                    "robot_state": spaces.Dict(
+                        {
+                            "eef": spaces.Dict(
+                                {
+                                    "pos": spaces.Box(low=-np.inf, high=np.inf, shape=(3,), dtype=np.float64),
+                                    "quat": spaces.Box(
+                                        low=-np.inf, high=np.inf, shape=(4,), dtype=np.float64
+                                    ),
+                                    "mat": spaces.Box(
+                                        low=-np.inf, high=np.inf, shape=(3, 3), dtype=np.float64
+                                    ),
+                                }
+                            ),
+                            "gripper": spaces.Dict(
+                                {
+                                    "qpos": spaces.Box(
+                                        low=-np.inf, high=np.inf, shape=(2,), dtype=np.float64
+                                    ),
+                                    "qvel": spaces.Box(
+                                        low=-np.inf, high=np.inf, shape=(2,), dtype=np.float64
+                                    ),
+                                }
+                            ),
+                            "joints": spaces.Dict(
+                                {
+                                    "pos": spaces.Box(low=-np.inf, high=np.inf, shape=(7,), dtype=np.float64),
+                                    "vel": spaces.Box(low=-np.inf, high=np.inf, shape=(7,), dtype=np.float64),
+                                }
+                            ),
+                        }
                     ),
                 }
             )
@@ -191,6 +215,7 @@ class LiberoEnv(gym.Env):
     def render(self):
         raw_obs = self._env.env._get_observations()
         image = self._format_raw_obs(raw_obs)["pixels"]["image"]
+        image = image[::-1, ::-1]  # flip both H and W for visualization
         return image
 
     def _make_envs_task(self, task_suite: Any, task_id: int = 0):
@@ -212,23 +237,48 @@ class LiberoEnv(gym.Env):
         images = {}
         for camera_name in self.camera_name:
             image = raw_obs[camera_name]
-            image = image[::-1, ::-1]  # rotate 180 degrees
             images[self.camera_name_mapping[camera_name]] = image
-        state = np.concatenate(
-            (
-                raw_obs["robot0_eef_pos"],
-                quat2axisangle(raw_obs["robot0_eef_quat"]),
-                raw_obs["robot0_gripper_qpos"],
-            )
-        )
-        agent_pos = state
+
+        eef_pos = raw_obs.get("robot0_eef_pos")
+        eef_quat = raw_obs.get("robot0_eef_quat")
+
+        # rotation matrix from controller
+        eef_mat = self._env.robots[0].controller.ee_ori_mat if eef_pos is not None else None
+        gripper_qpos = raw_obs.get("robot0_gripper_qpos")
+        gripper_qvel = raw_obs.get("robot0_gripper_qvel")
+        joint_pos = raw_obs.get("robot0_joint_pos")
+        joint_vel = raw_obs.get("robot0_joint_vel")
+        obs = {
+            "pixels": images,
+            "robot_state": {
+                "eef": {
+                    "pos": eef_pos,  # (3,)
+                    "quat": eef_quat,  # (4,)
+                    "mat": eef_mat,  # (3, 3)
+                },
+                "gripper": {
+                    "qpos": gripper_qpos,  # (2,)
+                    "qvel": gripper_qvel,  # (2,)
+                },
+                "joints": {
+                    "pos": joint_pos,  # (7,)
+                    "vel": joint_vel,  # (7,)
+                },
+            },
+        }
         if self.obs_type == "pixels":
             return {"pixels": images.copy()}
+
         if self.obs_type == "pixels_agent_pos":
-            return {
-                "pixels": images.copy(),
-                "agent_pos": agent_pos,
-            }
+            # Validate required fields are present
+            if eef_pos is None or eef_quat is None or gripper_qpos is None:
+                raise ValueError(
+                    f"Missing required robot state fields in raw observation. "
+                    f"Got eef_pos={eef_pos is not None}, eef_quat={eef_quat is not None}, "
+                    f"gripper_qpos={gripper_qpos is not None}"
+                )
+            return obs
+
         raise NotImplementedError(
             f"The observation type '{self.obs_type}' is not supported in LiberoEnv. "
             "Please switch to an image-based obs_type (e.g. 'pixels', 'pixels_agent_pos')."
@@ -260,19 +310,23 @@ class LiberoEnv(gym.Env):
 
         is_success = self._env.check_success()
         terminated = done or is_success
-        info["is_success"] = is_success
-
+        info.update(
+            {
+                "task": self.task,
+                "task_id": self.task_id,
+                "done": done,
+                "is_success": is_success,
+            }
+        )
         observation = self._format_raw_obs(raw_obs)
-        if done:
+        if terminated:
+            info["final_info"] = {
+                "task": self.task,
+                "task_id": self.task_id,
+                "done": bool(done),
+                "is_success": bool(is_success),
+            }
             self.reset()
-            info.update(
-                {
-                    "task": self.task,
-                    "task_id": self.task_id,
-                    "done": done,
-                    "is_success": is_success,
-                }
-            )
         truncated = False
         return observation, reward, terminated, truncated, info
 
@@ -351,12 +405,10 @@ def create_libero_envs(
         print(f"Restricting to task_ids={task_ids_filter}")
 
     out: dict[str, dict[int, Any]] = defaultdict(dict)
-
     for suite_name in suite_names:
         suite = _get_suite(suite_name)
         total = len(suite.tasks)
         selected = _select_task_ids(total, task_ids_filter)
-
         if not selected:
             raise ValueError(f"No tasks selected for suite '{suite_name}' (available: {total}).")
 

src/lerobot/envs/metaworld.py

@@ -0,0 +1,313 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import json
+from collections import defaultdict
+from collections.abc import Callable, Sequence
+from pathlib import Path
+from typing import Any
+
+import gymnasium as gym
+import metaworld
+import metaworld.policies as policies
+import numpy as np
+from gymnasium import spaces
+
+# ---- Load configuration data from the external JSON file ----
+CONFIG_PATH = Path(__file__).parent / "metaworld_config.json"
+try:
+    with open(CONFIG_PATH) as f:
+        data = json.load(f)
+except FileNotFoundError as err:
+    raise FileNotFoundError(
+        "Could not find 'metaworld_config.json'. "
+        "Please ensure the configuration file is in the same directory as the script."
+    ) from err
+except json.JSONDecodeError as err:
+    raise ValueError(
+        "Failed to decode 'metaworld_config.json'. Please ensure it is a valid JSON file."
+    ) from err
+
+# ---- Process the loaded data ----
+
+# extract and type-check top-level dicts
+task_descriptions_obj = data.get("TASK_DESCRIPTIONS")
+if not isinstance(task_descriptions_obj, dict):
+    raise TypeError("Expected TASK_DESCRIPTIONS to be a dict[str, str]")
+TASK_DESCRIPTIONS: dict[str, str] = task_descriptions_obj
+
+task_name_to_id_obj = data.get("TASK_NAME_TO_ID")
+if not isinstance(task_name_to_id_obj, dict):
+    raise TypeError("Expected TASK_NAME_TO_ID to be a dict[str, int]")
+TASK_NAME_TO_ID: dict[str, int] = task_name_to_id_obj
+
+# difficulty -> tasks mapping
+difficulty_to_tasks = data.get("DIFFICULTY_TO_TASKS")
+if not isinstance(difficulty_to_tasks, dict):
+    raise TypeError("Expected 'DIFFICULTY_TO_TASKS' to be a dict[str, list[str]]")
+DIFFICULTY_TO_TASKS: dict[str, list[str]] = difficulty_to_tasks
+
+# convert policy strings -> actual policy classes
+task_policy_mapping = data.get("TASK_POLICY_MAPPING")
+if not isinstance(task_policy_mapping, dict):
+    raise TypeError("Expected 'TASK_POLICY_MAPPING' to be a dict[str, str]")
+TASK_POLICY_MAPPING: dict[str, Any] = {
+    task_name: getattr(policies, policy_class_name)
+    for task_name, policy_class_name in task_policy_mapping.items()
+}
+ACTION_DIM = 4
+OBS_DIM = 4
+
+
+class MetaworldEnv(gym.Env):
+    metadata = {"render_modes": ["rgb_array"], "render_fps": 80}
+
+    def __init__(
+        self,
+        task,
+        camera_name="corner2",
+        obs_type="pixels",
+        render_mode="rgb_array",
+        observation_width=480,
+        observation_height=480,
+        visualization_width=640,
+        visualization_height=480,
+    ):
+        super().__init__()
+        self.task = task.replace("metaworld-", "")
+        self.obs_type = obs_type
+        self.render_mode = render_mode
+        self.observation_width = observation_width
+        self.observation_height = observation_height
+        self.visualization_width = visualization_width
+        self.visualization_height = visualization_height
+        self.camera_name = camera_name
+
+        self._env = self._make_envs_task(self.task)
+        self._max_episode_steps = self._env.max_path_length
+        self.task_description = TASK_DESCRIPTIONS[self.task]
+
+        self.expert_policy = TASK_POLICY_MAPPING[self.task]()
+
+        if self.obs_type == "state":
+            raise NotImplementedError()
+        elif self.obs_type == "pixels":
+            self.observation_space = spaces.Dict(
+                {
+                    "pixels": spaces.Box(
+                        low=0,
+                        high=255,
+                        shape=(self.observation_height, self.observation_width, 3),
+                        dtype=np.uint8,
+                    )
+                }
+            )
+        elif self.obs_type == "pixels_agent_pos":
+            self.observation_space = spaces.Dict(
+                {
+                    "pixels": spaces.Box(
+                        low=0,
+                        high=255,
+                        shape=(self.observation_height, self.observation_width, 3),
+                        dtype=np.uint8,
+                    ),
+                    "agent_pos": spaces.Box(
+                        low=-1000.0,
+                        high=1000.0,
+                        shape=(OBS_DIM,),
+                        dtype=np.float64,
+                    ),
+                }
+            )
+
+        self.action_space = spaces.Box(low=-1, high=1, shape=(ACTION_DIM,), dtype=np.float32)
+
+    def render(self) -> np.ndarray:
+        """
+        Render the current environment frame.
+
+        Returns:
+            np.ndarray: The rendered RGB image from the environment.
+        """
+        image = self._env.render()
+        if self.camera_name == "corner2":
+            # Images from this camera are flipped — correct them
+            image = np.flip(image, (0, 1))
+        return image
+
+    def _make_envs_task(self, env_name: str):
+        mt1 = metaworld.MT1(env_name, seed=42)
+        env = mt1.train_classes[env_name](render_mode="rgb_array", camera_name=self.camera_name)
+        env.set_task(mt1.train_tasks[0])
+        if self.camera_name == "corner2":
+            env.model.cam_pos[2] = [
+                0.75,
+                0.075,
+                0.7,
+            ]  # corner2 position, similar to https://arxiv.org/pdf/2206.14244
+        env.reset()
+        env._freeze_rand_vec = False  # otherwise no randomization
+        return env
+
+    def _format_raw_obs(self, raw_obs: np.ndarray) -> dict[str, Any]:
+        image = None
+        if self._env is not None:
+            image = self._env.render()
+            if self.camera_name == "corner2":
+                # NOTE: The "corner2" camera in MetaWorld environments outputs images with both axes inverted.
+                image = np.flip(image, (0, 1))
+        agent_pos = raw_obs[:4]
+        if self.obs_type == "state":
+            raise NotImplementedError(
+                "'state' obs_type not implemented for MetaWorld. Use pixel modes instead."
+            )
+
+        elif self.obs_type in ("pixels", "pixels_agent_pos"):
+            assert image is not None, (
+                "Expected `image` to be rendered before constructing pixel-based observations. "
+                "This likely means `env.render()` returned None or the environment was not provided."
+            )
+
+            if self.obs_type == "pixels":
+                obs = {"pixels": image.copy()}
+
+            else:  # pixels_agent_pos
+                obs = {
+                    "pixels": image.copy(),
+                    "agent_pos": agent_pos,
+                }
+        else:
+            raise ValueError(f"Unknown obs_type: {self.obs_type}")
+        return obs
+
+    def reset(
+        self,
+        seed: int | None = None,
+        **kwargs,
+    ) -> tuple[dict[str, Any], dict[str, Any]]:
+        """
+        Reset the environment to its initial state.
+
+        Args:
+            seed (Optional[int]): Random seed for environment initialization.
+
+        Returns:
+            observation (Dict[str, Any]): The initial formatted observation.
+            info (Dict[str, Any]): Additional info about the reset state.
+        """
+        super().reset(seed=seed)
+
+        raw_obs, info = self._env.reset(seed=seed)
+
+        observation = self._format_raw_obs(raw_obs)
+
+        info = {"is_success": False}
+        return observation, info
+
+    def step(self, action: np.ndarray) -> tuple[dict[str, Any], float, bool, bool, dict[str, Any]]:
+        """
+        Perform one environment step.
+
+        Args:
+            action (np.ndarray): The action to execute, must be 1-D with shape (action_dim,).
+
+        Returns:
+            observation (Dict[str, Any]): The formatted observation after the step.
+            reward (float): The scalar reward for this step.
+            terminated (bool): Whether the episode terminated successfully.
+            truncated (bool): Whether the episode was truncated due to a time limit.
+            info (Dict[str, Any]): Additional environment info.
+        """
+        if action.ndim != 1:
+            raise ValueError(
+                f"Expected action to be 1-D (shape (action_dim,)), "
+                f"but got shape {action.shape} with ndim={action.ndim}"
+            )
+        raw_obs, reward, done, truncated, info = self._env.step(action)
+
+        # Determine whether the task was successful
+        is_success = bool(info.get("success", 0))
+        terminated = done or is_success
+        info.update(
+            {
+                "task": self.task,
+                "done": done,
+                "is_success": is_success,
+            }
+        )
+
+        # Format the raw observation into the expected structure
+        observation = self._format_raw_obs(raw_obs)
+        if terminated:
+            info["final_info"] = {
+                "task": self.task,
+                "done": bool(done),
+                "is_success": bool(is_success),
+            }
+            self.reset()
+
+        return observation, reward, terminated, truncated, info
+
+    def close(self):
+        self._env.close()
+
+
+# ---- Main API ----------------------------------------------------------------
+
+
+def create_metaworld_envs(
+    task: str,
+    n_envs: int,
+    gym_kwargs: dict[str, Any] | None = None,
+    env_cls: Callable[[Sequence[Callable[[], Any]]], Any] | None = None,
+) -> dict[str, dict[int, Any]]:
+    """
+    Create vectorized Meta-World environments with a consistent return shape.
+
+    Returns:
+        dict[task_group][task_id] -> vec_env (env_cls([...]) with exactly n_envs factories)
+    Notes:
+        - n_envs is the number of rollouts *per task* (episode_index = 0..n_envs-1).
+        - `task` can be a single difficulty group (e.g., "easy", "medium", "hard") or a comma-separated list.
+        - If a task name is not in DIFFICULTY_TO_TASKS, we treat it as a single custom task.
+    """
+    if env_cls is None or not callable(env_cls):
+        raise ValueError("env_cls must be a callable that wraps a list of environment factory callables.")
+    if not isinstance(n_envs, int) or n_envs <= 0:
+        raise ValueError(f"n_envs must be a positive int; got {n_envs}.")
+
+    gym_kwargs = dict(gym_kwargs or {})
+    task_groups = [t.strip() for t in task.split(",") if t.strip()]
+    if not task_groups:
+        raise ValueError("`task` must contain at least one Meta-World task or difficulty group.")
+
+    print(f"Creating Meta-World envs | task_groups={task_groups} | n_envs(per task)={n_envs}")
+
+    out: dict[str, dict[int, Any]] = defaultdict(dict)
+
+    for group in task_groups:
+        # if not in difficulty presets, treat it as a single custom task
+        tasks = DIFFICULTY_TO_TASKS.get(group, [group])
+
+        for tid, task_name in enumerate(tasks):
+            print(f"Building vec env | group={group} | task_id={tid} | task={task_name}")
+
+            # build n_envs factories
+            fns = [(lambda tn=task_name: MetaworldEnv(task=tn, **gym_kwargs)) for _ in range(n_envs)]
+
+            out[group][tid] = env_cls(fns)
+
+    # return a plain dict for consistency
+    return {group: dict(task_map) for group, task_map in out.items()}

src/lerobot/envs/metaworld_config.json

@@ -0,0 +1,121 @@
+{
+  "TASK_DESCRIPTIONS": {
+    "assembly-v3": "Pick up a nut and place it onto a peg",
+    "basketball-v3": "Dunk the basketball into the basket",
+    "bin-picking-v3": "Grasp the puck from one bin and place it into another bin",
+    "box-close-v3": "Grasp the cover and close the box with it",
+    "button-press-topdown-v3": "Press a button from the top",
+    "button-press-topdown-wall-v3": "Bypass a wall and press a button from the top",
+    "button-press-v3": "Press a button",
+    "button-press-wall-v3": "Bypass a wall and press a button",
+    "coffee-button-v3": "Push a button on the coffee machine",
+    "coffee-pull-v3": "Pull a mug from a coffee machine",
+    "coffee-push-v3": "Push a mug under a coffee machine",
+    "dial-turn-v3": "Rotate a dial 180 degrees",
+    "disassemble-v3": "Pick a nut out of a peg",
+    "door-close-v3": "Close a door with a revolving joint",
+    "door-lock-v3": "Lock the door by rotating the lock clockwise",
+    "door-open-v3": "Open a door with a revolving joint",
+    "door-unlock-v3": "Unlock the door by rotating the lock counter-clockwise",
+    "hand-insert-v3": "Insert the gripper into a hole",
+    "drawer-close-v3": "Push and close a drawer",
+    "drawer-open-v3": "Open a drawer",
+    "faucet-open-v3": "Rotate the faucet counter-clockwise",
+    "faucet-close-v3": "Rotate the faucet clockwise",
+    "hammer-v3": "Hammer a screw on the wall",
+    "handle-press-side-v3": "Press a handle down sideways",
+    "handle-press-v3": "Press a handle down",
+    "handle-pull-side-v3": "Pull a handle up sideways",
+    "handle-pull-v3": "Pull a handle up",
+    "lever-pull-v3": "Pull a lever down 90 degrees",
+    "peg-insert-side-v3": "Insert a peg sideways",
+    "pick-place-wall-v3": "Pick a puck, bypass a wall and place the puck",
+    "pick-out-of-hole-v3": "Pick up a puck from a hole",
+    "reach-v3": "Reach a goal position",
+    "push-back-v3": "Push the puck to a goal",
+    "push-v3": "Push the puck to a goal",
+    "pick-place-v3": "Pick and place a puck to a goal",
+    "plate-slide-v3": "Slide a plate into a cabinet",
+    "plate-slide-side-v3": "Slide a plate into a cabinet sideways",
+    "plate-slide-back-v3": "Get a plate from the cabinet",
+    "plate-slide-back-side-v3": "Get a plate from the cabinet sideways",
+    "peg-unplug-side-v3": "Unplug a peg sideways",
+    "soccer-v3": "Kick a soccer into the goal",
+    "stick-push-v3": "Grasp a stick and push a box using the stick",
+    "stick-pull-v3": "Grasp a stick and pull a box with the stick",
+    "push-wall-v3": "Bypass a wall and push a puck to a goal",
+    "reach-wall-v3": "Bypass a wall and reach a goal",
+    "shelf-place-v3": "Pick and place a puck onto a shelf",
+    "sweep-into-v3": "Sweep a puck into a hole",
+    "sweep-v3": "Sweep a puck off the table",
+    "window-open-v3": "Push and open a window",
+    "window-close-v3": "Push and close a window"
+  },
+  "TASK_NAME_TO_ID": {
+    "assembly-v3": 0, "basketball-v3": 1, "bin-picking-v3": 2, "box-close-v3": 3,
+    "button-press-topdown-v3": 4, "button-press-topdown-wall-v3": 5, "button-press-v3": 6,
+    "button-press-wall-v3": 7, "coffee-button-v3": 8, "coffee-pull-v3": 9, "coffee-push-v3": 10,
+    "dial-turn-v3": 11, "disassemble-v3": 12, "door-close-v3": 13, "door-lock-v3": 14,
+    "door-open-v3": 15, "door-unlock-v3": 16, "drawer-close-v3": 17, "drawer-open-v3": 18,
+    "faucet-close-v3": 19, "faucet-open-v3": 20, "hammer-v3": 21, "hand-insert-v3": 22,
+    "handle-press-side-v3": 23, "handle-press-v3": 24, "handle-pull-side-v3": 25,
+    "handle-pull-v3": 26, "lever-pull-v3": 27, "peg-insert-side-v3": 28, "peg-unplug-side-v3": 29,
+    "pick-out-of-hole-v3": 30, "pick-place-v3": 31, "pick-place-wall-v3": 32,
+    "plate-slide-back-side-v3": 33, "plate-slide-back-v3": 34, "plate-slide-side-v3": 35,
+    "plate-slide-v3": 36, "push-back-v3": 37, "push-v3": 38, "push-wall-v3": 39, "reach-v3": 40,
+    "reach-wall-v3": 41, "shelf-place-v3": 42, "soccer-v3": 43, "stick-pull-v3": 44,
+    "stick-push-v3": 45, "sweep-into-v3": 46, "sweep-v3": 47, "window-open-v3": 48,
+    "window-close-v3": 49
+  },
+  "DIFFICULTY_TO_TASKS": {
+    "easy": [
+      "button-press-v3", "button-press-topdown-v3", "button-press-topdown-wall-v3",
+      "button-press-wall-v3", "coffee-button-v3", "dial-turn-v3", "door-close-v3",
+      "door-lock-v3", "door-open-v3", "door-unlock-v3", "drawer-close-v3", "drawer-open-v3",
+      "faucet-close-v3", "faucet-open-v3", "handle-press-v3", "handle-press-side-v3",
+      "handle-pull-v3", "handle-pull-side-v3", "lever-pull-v3", "plate-slide-v3",
+      "plate-slide-back-v3", "plate-slide-back-side-v3", "plate-slide-side-v3", "reach-v3",
+      "reach-wall-v3", "window-close-v3", "window-open-v3", "peg-unplug-side-v3"
+    ],
+    "medium": [
+      "basketball-v3", "bin-picking-v3", "box-close-v3", "coffee-pull-v3", "coffee-push-v3",
+      "hammer-v3", "peg-insert-side-v3", "push-wall-v3", "soccer-v3", "sweep-v3", "sweep-into-v3"
+    ],
+    "hard": [
+      "assembly-v3", "hand-insert-v3", "pick-out-of-hole-v3", "pick-place-v3", "push-v3", "push-back-v3"
+    ],
+    "very_hard": [
+      "shelf-place-v3", "disassemble-v3", "stick-pull-v3", "stick-push-v3", "pick-place-wall-v3"
+    ]
+  },
+  "TASK_POLICY_MAPPING": {
+    "assembly-v3": "SawyerAssemblyV3Policy", "basketball-v3": "SawyerBasketballV3Policy",
+    "bin-picking-v3": "SawyerBinPickingV3Policy", "box-close-v3": "SawyerBoxCloseV3Policy",
+    "button-press-topdown-v3": "SawyerButtonPressTopdownV3Policy",
+    "button-press-topdown-wall-v3": "SawyerButtonPressTopdownWallV3Policy",
+    "button-press-v3": "SawyerButtonPressV3Policy", "button-press-wall-v3": "SawyerButtonPressWallV3Policy",
+    "coffee-button-v3": "SawyerCoffeeButtonV3Policy", "coffee-pull-v3": "SawyerCoffeePullV3Policy",
+    "coffee-push-v3": "SawyerCoffeePushV3Policy", "dial-turn-v3": "SawyerDialTurnV3Policy",
+    "disassemble-v3": "SawyerDisassembleV3Policy", "door-close-v3": "SawyerDoorCloseV3Policy",
+    "door-lock-v3": "SawyerDoorLockV3Policy", "door-open-v3": "SawyerDoorOpenV3Policy",
+    "door-unlock-v3": "SawyerDoorUnlockV3Policy", "drawer-close-v3": "SawyerDrawerCloseV3Policy",
+    "drawer-open-v3": "SawyerDrawerOpenV3Policy", "faucet-close-v3": "SawyerFaucetCloseV3Policy",
+    "faucet-open-v3": "SawyerFaucetOpenV3Policy", "hammer-v3": "SawyerHammerV3Policy",
+    "hand-insert-v3": "SawyerHandInsertV3Policy", "handle-press-side-v3": "SawyerHandlePressSideV3Policy",
+    "handle-press-v3": "SawyerHandlePressV3Policy", "handle-pull-side-v3": "SawyerHandlePullSideV3Policy",
+    "handle-pull-v3": "SawyerHandlePullV3Policy", "lever-pull-v3": "SawyerLeverPullV3Policy",
+    "peg-insert-side-v3": "SawyerPegInsertionSideV3Policy", "peg-unplug-side-v3": "SawyerPegUnplugSideV3Policy",
+    "pick-out-of-hole-v3": "SawyerPickOutOfHoleV3Policy", "pick-place-v3": "SawyerPickPlaceV3Policy",
+    "pick-place-wall-v3": "SawyerPickPlaceWallV3Policy",
+    "plate-slide-back-side-v3": "SawyerPlateSlideBackSideV3Policy",
+    "plate-slide-back-v3": "SawyerPlateSlideBackV3Policy",
+    "plate-slide-side-v3": "SawyerPlateSlideSideV3Policy", "plate-slide-v3": "SawyerPlateSlideV3Policy",
+    "push-back-v3": "SawyerPushBackV3Policy", "push-v3": "SawyerPushV3Policy",
+    "push-wall-v3": "SawyerPushWallV3Policy", "reach-v3": "SawyerReachV3Policy",
+    "reach-wall-v3": "SawyerReachWallV3Policy", "shelf-place-v3": "SawyerShelfPlaceV3Policy",
+    "soccer-v3": "SawyerSoccerV3Policy", "stick-pull-v3": "SawyerStickPullV3Policy",
+    "stick-push-v3": "SawyerStickPushV3Policy", "sweep-into-v3": "SawyerSweepIntoV3Policy",
+    "sweep-v3": "SawyerSweepV3Policy", "window-open-v3": "SawyerWindowOpenV3Policy",
+    "window-close-v3": "SawyerWindowCloseV3Policy"
+  }
+}

src/lerobot/envs/utils.py

@@ -13,6 +13,8 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+import importlib.util
+import os
 import warnings
 from collections.abc import Mapping, Sequence
 from functools import singledispatch
@@ -22,14 +24,27 @@ import einops
 import gymnasium as gym
 import numpy as np
 import torch
+from huggingface_hub import hf_hub_download, snapshot_download
 from torch import Tensor
 
 from lerobot.configs.types import FeatureType, PolicyFeature
 from lerobot.envs.configs import EnvConfig
-from lerobot.utils.constants import OBS_ENV_STATE, OBS_IMAGE, OBS_IMAGES, OBS_STATE
+from lerobot.utils.constants import OBS_ENV_STATE, OBS_IMAGE, OBS_IMAGES, OBS_STATE, OBS_STR
 from lerobot.utils.utils import get_channel_first_image_shape
 
 
+def _convert_nested_dict(d):
+    result = {}
+    for k, v in d.items():
+        if isinstance(v, dict):
+            result[k] = _convert_nested_dict(v)
+        elif isinstance(v, np.ndarray):
+            result[k] = torch.from_numpy(v)
+        else:
+            result[k] = v
+    return result
+
+
 def preprocess_observation(observations: dict[str, np.ndarray]) -> dict[str, Tensor]:
     # TODO(aliberts, rcadene): refactor this to use features from the environment (no hardcoding)
     """Convert environment observation to LeRobot format observation.
@@ -75,12 +90,14 @@ def preprocess_observation(observations: dict[str, np.ndarray]) -> dict[str, Ten
 
         return_observations[OBS_ENV_STATE] = env_state
 
-    # TODO(rcadene): enable pixels only baseline with `obs_type="pixels"` in environment by removing
-    agent_pos = torch.from_numpy(observations["agent_pos"]).float()
-    if agent_pos.dim() == 1:
-        agent_pos = agent_pos.unsqueeze(0)
-    return_observations[OBS_STATE] = agent_pos
+    if "agent_pos" in observations:
+        agent_pos = torch.from_numpy(observations["agent_pos"]).float()
+        if agent_pos.dim() == 1:
+            agent_pos = agent_pos.unsqueeze(0)
+        return_observations[OBS_STATE] = agent_pos
 
+    if "robot_state" in observations:
+        return_observations[f"{OBS_STR}.robot_state"] = _convert_nested_dict(observations["robot_state"])
     return return_observations
 
 
@@ -195,3 +212,132 @@ def _(envs: Sequence) -> None:
 @close_envs.register
 def _(env: gym.Env) -> None:
     _close_single_env(env)
+
+
+# helper to safely load a python file as a module
+def _load_module_from_path(path: str, module_name: str | None = None):
+    module_name = module_name or f"hub_env_{os.path.basename(path).replace('.', '_')}"
+    spec = importlib.util.spec_from_file_location(module_name, path)
+    if spec is None:
+        raise ImportError(f"Could not load module spec for {module_name} from {path}")
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)  # type: ignore
+    return module
+
+
+# helper to parse hub string (supports "user/repo", "user/repo@rev", optional path)
+# examples:
+#   "user/repo" -> will look for env.py at repo root
+#   "user/repo@main:envs/my_env.py" -> explicit revision and path
+def _parse_hub_url(hub_uri: str):
+    # very small parser: [repo_id][@revision][:path]
+    # repo_id is required (user/repo or org/repo)
+    revision = None
+    file_path = "env.py"
+    if "@" in hub_uri:
+        repo_and_rev, *rest = hub_uri.split(":", 1)
+        repo_id, rev = repo_and_rev.split("@", 1)
+        revision = rev
+        if rest:
+            file_path = rest[0]
+    else:
+        repo_id, *rest = hub_uri.split(":", 1)
+        if rest:
+            file_path = rest[0]
+    return repo_id, revision, file_path
+
+
+def _download_hub_file(
+    cfg_str: str,
+    trust_remote_code: bool,
+    hub_cache_dir: str | None,
+) -> tuple[str, str, str, str]:
+    """
+    Parse `cfg_str` (hub URL), enforce `trust_remote_code`, and return
+    (repo_id, file_path, local_file, revision).
+    """
+    if not trust_remote_code:
+        raise RuntimeError(
+            f"Refusing to execute remote code from the Hub for '{cfg_str}'. "
+            "Executing hub env modules runs arbitrary Python code from third-party repositories. "
+            "If you trust this repo and understand the risks, call `make_env(..., trust_remote_code=True)` "
+            "and prefer pinning to a specific revision: 'user/repo@<commit-hash>:env.py'."
+        )
+
+    repo_id, revision, file_path = _parse_hub_url(cfg_str)
+
+    try:
+        local_file = hf_hub_download(
+            repo_id=repo_id, filename=file_path, revision=revision, cache_dir=hub_cache_dir
+        )
+    except Exception as e:
+        # fallback to snapshot download
+        snapshot_dir = snapshot_download(repo_id=repo_id, revision=revision, cache_dir=hub_cache_dir)
+        local_file = os.path.join(snapshot_dir, file_path)
+        if not os.path.exists(local_file):
+            raise FileNotFoundError(
+                f"Could not find {file_path} in repository {repo_id}@{revision or 'main'}"
+            ) from e
+
+    return repo_id, file_path, local_file, revision
+
+
+def _import_hub_module(local_file: str, repo_id: str) -> Any:
+    """
+    Import the downloaded file as a module and surface helpful import error messages.
+    """
+    module_name = f"hub_env_{repo_id.replace('/', '_')}"
+    try:
+        module = _load_module_from_path(local_file, module_name=module_name)
+    except ModuleNotFoundError as e:
+        missing = getattr(e, "name", None) or str(e)
+        raise ModuleNotFoundError(
+            f"Hub env '{repo_id}:{os.path.basename(local_file)}' failed to import because the dependency "
+            f"'{missing}' is not installed locally.\n\n"
+        ) from e
+    except ImportError as e:
+        raise ImportError(
+            f"Failed to load hub env module '{repo_id}:{os.path.basename(local_file)}'. Import error: {e}\n\n"
+        ) from e
+    return module
+
+
+def _call_make_env(module: Any, n_envs: int, use_async_envs: bool) -> Any:
+    """
+    Ensure module exposes make_env and call it.
+    """
+    if not hasattr(module, "make_env"):
+        raise AttributeError(
+            f"The hub module {getattr(module, '__name__', 'hub_module')} must expose `make_env(n_envs=int, use_async_envs=bool)`."
+        )
+    entry_fn = module.make_env
+    return entry_fn(n_envs=n_envs, use_async_envs=use_async_envs)
+
+
+def _normalize_hub_result(result: Any) -> dict[str, dict[int, gym.vector.VectorEnv]]:
+    """
+    Normalize possible return types from hub `make_env` into the mapping:
+      { suite_name: { task_id: vector_env } }
+    Accepts:
+      - dict (assumed already correct)
+      - gym.vector.VectorEnv
+      - gym.Env (will be wrapped into SyncVectorEnv)
+    """
+    if isinstance(result, dict):
+        return result
+
+    # VectorEnv: use its spec.id if available
+    if isinstance(result, gym.vector.VectorEnv):
+        suite_name = getattr(result, "spec", None) and getattr(result.spec, "id", None) or "hub_env"
+        return {suite_name: {0: result}}
+
+    # Single Env: wrap into SyncVectorEnv
+    if isinstance(result, gym.Env):
+        vec = gym.vector.SyncVectorEnv([lambda: result])
+        suite_name = getattr(result, "spec", None) and getattr(result.spec, "id", None) or "hub_env"
+        return {suite_name: {0: vec}}
+
+    raise ValueError(
+        "Hub `make_env` must return either a mapping {suite: {task_id: vec_env}}, "
+        "a gym.vector.VectorEnv, or a single gym.Env."
+    )

src/lerobot/model/kinematics.py

@@ -22,18 +22,18 @@ class RobotKinematics:
         self,
         urdf_path: str,
         target_frame_name: str = "gripper_frame_link",
-        joint_names: list[str] = None,
+        joint_names: list[str] | None = None,
     ):
         """
         Initialize placo-based kinematics solver.
 
         Args:
-            urdf_path: Path to the robot URDF file
-            target_frame_name: Name of the end-effector frame in the URDF
-            joint_names: List of joint names to use for the kinematics solver
+            urdf_path (str): Path to the robot URDF file
+            target_frame_name (str): Name of the end-effector frame in the URDF
+            joint_names (list[str] | None): List of joint names to use for the kinematics solver
         """
         try:
-            import placo
+            import placo  # type: ignore[import-not-found] # C++ library with Python bindings, no type stubs available. TODO: Create stub file or request upstream typing support.
         except ImportError as e:
             raise ImportError(
                 "placo is required for RobotKinematics. "
@@ -52,7 +52,7 @@ class RobotKinematics:
         # Initialize frame task for IK
         self.tip_frame = self.solver.add_frame_task(self.target_frame_name, np.eye(4))
 
-    def forward_kinematics(self, joint_pos_deg):
+    def forward_kinematics(self, joint_pos_deg: np.ndarray) -> np.ndarray:
         """
         Compute forward kinematics for given joint configuration given the target frame name in the constructor.
 
@@ -77,8 +77,12 @@ class RobotKinematics:
         return self.robot.get_T_world_frame(self.target_frame_name)
 
     def inverse_kinematics(
-        self, current_joint_pos, desired_ee_pose, position_weight=1.0, orientation_weight=0.01
-    ):
+        self,
+        current_joint_pos: np.ndarray,
+        desired_ee_pose: np.ndarray,
+        position_weight: float = 1.0,
+        orientation_weight: float = 0.01,
+    ) -> np.ndarray:
         """
         Compute inverse kinematics using placo solver.
 

src/lerobot/motors/dynamixel/dynamixel.py

@@ -60,7 +60,7 @@ class OperatingMode(Enum):
 
     # This mode controls position. This mode is identical to the Multi-turn Position Control from existing
     # DYNAMIXEL. 512 turns are supported(-256[rev] ~ 256[rev]). This mode is ideal for multi-turn wrists or
-    # conveyer systems or a system that requires an additional reduction gear. Note that Max Position
+    # conveyor systems or a system that requires an additional reduction gear. Note that Max Position
     # Limit(48), Min Position Limit(52) are not used on Extended Position Control Mode.
     EXTENDED_POSITION = 4
 

src/lerobot/motors/feetech/tables.py

@@ -206,8 +206,12 @@ MODEL_BAUDRATE_TABLE = {
 # Sign-Magnitude encoding bits
 STS_SMS_SERIES_ENCODINGS_TABLE = {
     "Homing_Offset": 11,
+    "Goal_Position": 15,
     "Goal_Velocity": 15,
+    "Goal_Speed": 15,
+    "Present_Position": 15,
     "Present_Velocity": 15,
+    "Present_Speed": 15,
 }
 
 MODEL_ENCODING_TABLE = {

src/lerobot/optim/schedulers.py

@@ -14,6 +14,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import abc
+import logging
 import math
 from dataclasses import asdict, dataclass
 from pathlib import Path
@@ -79,7 +80,11 @@ class VQBeTSchedulerConfig(LRSchedulerConfig):
 @LRSchedulerConfig.register_subclass("cosine_decay_with_warmup")
 @dataclass
 class CosineDecayWithWarmupSchedulerConfig(LRSchedulerConfig):
-    """Used by Physical Intelligence to train Pi0"""
+    """Used by Physical Intelligence to train Pi0.
+
+    Automatically scales warmup and decay steps if num_training_steps < num_decay_steps.
+    This ensures the learning rate schedule completes properly even with shorter training runs.
+    """
 
     num_warmup_steps: int
     num_decay_steps: int
@@ -87,23 +92,39 @@ class CosineDecayWithWarmupSchedulerConfig(LRSchedulerConfig):
     decay_lr: float
 
     def build(self, optimizer: Optimizer, num_training_steps: int) -> LambdaLR:
-        del num_training_steps
+        # Auto-scale scheduler parameters if training steps are shorter than configured decay steps
+        actual_warmup_steps = self.num_warmup_steps
+        actual_decay_steps = self.num_decay_steps
+
+        if num_training_steps < self.num_decay_steps:
+            # Calculate scaling factor to fit the schedule into the available training steps
+            scale_factor = num_training_steps / self.num_decay_steps
+            actual_warmup_steps = int(self.num_warmup_steps * scale_factor)
+            actual_decay_steps = num_training_steps
+
+            logging.info(
+                f"Auto-scaling LR scheduler: "
+                f"num_training_steps ({num_training_steps}) < num_decay_steps ({self.num_decay_steps}). "
+                f"Scaling warmup: {self.num_warmup_steps} → {actual_warmup_steps}, "
+                f"decay: {self.num_decay_steps} → {actual_decay_steps} "
+                f"(scale factor: {scale_factor:.3f})"
+            )
 
         def lr_lambda(current_step):
             def linear_warmup_schedule(current_step):
                 if current_step <= 0:
-                    return 1 / (self.num_warmup_steps + 1)
-                frac = 1 - current_step / self.num_warmup_steps
-                return (1 / (self.num_warmup_steps + 1) - 1) * frac + 1
+                    return 1 / (actual_warmup_steps + 1)
+                frac = 1 - current_step / actual_warmup_steps
+                return (1 / (actual_warmup_steps + 1) - 1) * frac + 1
 
             def cosine_decay_schedule(current_step):
-                step = min(current_step, self.num_decay_steps)
-                cosine_decay = 0.5 * (1 + math.cos(math.pi * step / self.num_decay_steps))
+                step = min(current_step, actual_decay_steps)
+                cosine_decay = 0.5 * (1 + math.cos(math.pi * step / actual_decay_steps))
                 alpha = self.decay_lr / self.peak_lr
                 decayed = (1 - alpha) * cosine_decay + alpha
                 return decayed
 
-            if current_step < self.num_warmup_steps:
+            if current_step < actual_warmup_steps:
                 return linear_warmup_schedule(current_step)
 
             return cosine_decay_schedule(current_step)

src/lerobot/policies/__init__.py

@@ -14,6 +14,7 @@
 
 from .act.configuration_act import ACTConfig as ACTConfig
 from .diffusion.configuration_diffusion import DiffusionConfig as DiffusionConfig
+from .groot.configuration_groot import GrootConfig as GrootConfig
 from .pi0.configuration_pi0 import PI0Config as PI0Config
 from .pi05.configuration_pi05 import PI05Config as PI05Config
 from .smolvla.configuration_smolvla import SmolVLAConfig as SmolVLAConfig
@@ -29,4 +30,5 @@ __all__ = [
     "SmolVLAConfig",
     "TDMPCConfig",
     "VQBeTConfig",
+    "GrootConfig",
 ]

src/lerobot/policies/act/modeling_act.py

@@ -626,8 +626,8 @@ class ACTDecoderLayer(nn.Module):
             x: (Decoder Sequence, Batch, Channel) tensor of input tokens.
             encoder_out: (Encoder Sequence, B, C) output features from the last layer of the encoder we are
                 cross-attending with.
-            decoder_pos_embed: (ES, 1, C) positional embedding for keys (from the encoder).
-            encoder_pos_embed: (DS, 1, C) Positional_embedding for the queries (from the decoder).
+            encoder_pos_embed: (ES, 1, C) positional embedding for keys (from the encoder).
+            decoder_pos_embed: (DS, 1, C) positional embedding for the queries (from the decoder).
         Returns:
             (DS, B, C) tensor of decoder output features.
         """

src/lerobot/policies/factory.py

@@ -30,14 +30,16 @@ from lerobot.envs.configs import EnvConfig
 from lerobot.envs.utils import env_to_policy_features
 from lerobot.policies.act.configuration_act import ACTConfig
 from lerobot.policies.diffusion.configuration_diffusion import DiffusionConfig
+from lerobot.policies.groot.configuration_groot import GrootConfig
 from lerobot.policies.pi0.configuration_pi0 import PI0Config
-from lerobot.policies.pi0fast.configuration_pi0fast import PI0FASTConfig
 from lerobot.policies.pi05.configuration_pi05 import PI05Config
 from lerobot.policies.pretrained import PreTrainedPolicy
 from lerobot.policies.sac.configuration_sac import SACConfig
+from lerobot.policies.sarm.configuration_sarm import SARMConfig
 from lerobot.policies.sac.reward_model.configuration_classifier import RewardClassifierConfig
 from lerobot.policies.smolvla.configuration_smolvla import SmolVLAConfig
 from lerobot.policies.tdmpc.configuration_tdmpc import TDMPCConfig
+from lerobot.policies.utils import validate_visual_features_consistency
 from lerobot.policies.vqbet.configuration_vqbet import VQBeTConfig
 from lerobot.processor import PolicyAction, PolicyProcessorPipeline
 from lerobot.processor.converters import (
@@ -58,7 +60,7 @@ def get_policy_class(name: str) -> type[PreTrainedPolicy]:
 
     Args:
         name: The name of the policy. Supported names are "tdmpc", "diffusion", "act",
-              "vqbet", "pi0", "pi0fast", "sac", "reward_classifier", "smolvla".
+              "vqbet", "pi0", "pi05", "sac", "reward_classifier", "smolvla".
 
     Returns:
         The policy class corresponding to the given name.
@@ -82,10 +84,6 @@ def get_policy_class(name: str) -> type[PreTrainedPolicy]:
         from lerobot.policies.vqbet.modeling_vqbet import VQBeTPolicy
 
         return VQBeTPolicy
-    elif name == "pi0fast":
-        from lerobot.policies.pi0fast.modeling_pi0fast import PI0FASTPolicy
-
-        return PI0FASTPolicy
     elif name == "pi0":
         from lerobot.policies.pi0.modeling_pi0 import PI0Policy
 
@@ -106,6 +104,14 @@ def get_policy_class(name: str) -> type[PreTrainedPolicy]:
         from lerobot.policies.smolvla.modeling_smolvla import SmolVLAPolicy
 
         return SmolVLAPolicy
+    elif name == "sarm":
+        from lerobot.policies.sarm.modeling_sarm import SARMRewardModel
+
+        return SARMRewardModel
+    elif name == "groot":
+        from lerobot.policies.groot.modeling_groot import GrootPolicy
+
+        return GrootPolicy
     else:
         raise NotImplementedError(f"Policy with name {name} is not implemented.")
 
@@ -119,7 +125,7 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
 
     Args:
         policy_type: The type of the policy. Supported types include "tdmpc",
-                     "diffusion", "act", "vqbet", "pi0", "pi0fast", "sac", "smolvla",
+                     "diffusion", "act", "vqbet", "pi0", "pi05", "sac", "smolvla",
                      "reward_classifier".
         **kwargs: Keyword arguments to be passed to the configuration class constructor.
 
@@ -137,8 +143,6 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
         return ACTConfig(**kwargs)
     elif policy_type == "vqbet":
         return VQBeTConfig(**kwargs)
-    elif policy_type == "pi0fast":
-        return PI0FASTConfig(**kwargs)
     elif policy_type == "pi0":
         return PI0Config(**kwargs)
     elif policy_type == "pi05":
@@ -149,6 +153,8 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
         return SmolVLAConfig(**kwargs)
     elif policy_type == "reward_classifier":
         return RewardClassifierConfig(**kwargs)
+    elif policy_type == "groot":
+        return GrootConfig(**kwargs)
     else:
         raise ValueError(f"Policy type '{policy_type}' is not available.")
 
@@ -206,6 +212,27 @@ def make_pre_post_processors(
             policy configuration type.
     """
     if pretrained_path:
+        # TODO(Steven): Temporary patch, implement correctly the processors for Gr00t
+        if isinstance(policy_cfg, GrootConfig):
+            # GROOT handles normalization in groot_pack_inputs_v3 step
+            # Need to override both stats AND normalize_min_max since saved config might be empty
+            preprocessor_overrides = {}
+            postprocessor_overrides = {}
+            preprocessor_overrides["groot_pack_inputs_v3"] = {
+                "stats": kwargs.get("dataset_stats"),
+                "normalize_min_max": True,
+            }
+
+            # Also ensure postprocessing slices to env action dim and unnormalizes with dataset stats
+            env_action_dim = policy_cfg.output_features["action"].shape[0]
+            postprocessor_overrides["groot_action_unpack_unnormalize_v1"] = {
+                "stats": kwargs.get("dataset_stats"),
+                "normalize_min_max": True,
+                "env_action_dim": env_action_dim,
+            }
+            kwargs["preprocessor_overrides"] = preprocessor_overrides
+            kwargs["postprocessor_overrides"] = postprocessor_overrides
+
         return (
             PolicyProcessorPipeline.from_pretrained(
                 pretrained_model_name_or_path=pretrained_path,
@@ -260,14 +287,6 @@ def make_pre_post_processors(
             dataset_stats=kwargs.get("dataset_stats"),
         )
 
-    elif isinstance(policy_cfg, PI0FASTConfig):
-        from lerobot.policies.pi0fast.processor_pi0fast import make_pi0fast_pre_post_processors
-
-        processors = make_pi0fast_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-
     elif isinstance(policy_cfg, PI0Config):
         from lerobot.policies.pi0.processor_pi0 import make_pi0_pre_post_processors
 
@@ -308,6 +327,22 @@ def make_pre_post_processors(
             dataset_stats=kwargs.get("dataset_stats"),
         )
 
+    elif isinstance(policy_cfg, SARMConfig):
+        from lerobot.policies.sarm.processor_sarm import make_sarm_pre_post_processors
+
+        processors = make_sarm_pre_post_processors(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+            dataset_meta=kwargs.get("dataset_meta"),
+        )
+    elif isinstance(policy_cfg, GrootConfig):
+        from lerobot.policies.groot.processor_groot import make_groot_pre_post_processors
+
+        processors = make_groot_pre_post_processors(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+        )
+
     else:
         raise NotImplementedError(f"Processor for policy type '{policy_cfg.type}' is not implemented.")
 
@@ -318,6 +353,7 @@ def make_policy(
     cfg: PreTrainedConfig,
     ds_meta: LeRobotDatasetMetadata | None = None,
     env_cfg: EnvConfig | None = None,
+    rename_map: dict[str, str] | None = None,
 ) -> PreTrainedPolicy:
     """
     Instantiate a policy model.
@@ -334,6 +370,8 @@ def make_policy(
                  statistics for normalization layers.
         env_cfg: Environment configuration used to infer feature shapes and types.
                  One of `ds_meta` or `env_cfg` must be provided.
+        rename_map: Optional mapping of dataset or environment feature keys to match
+                 expected policy feature names (e.g., `"left"` → `"camera1"`).
 
     Returns:
         An instantiated and device-placed policy model.
@@ -375,9 +413,18 @@ def make_policy(
             raise ValueError("env_cfg cannot be None when ds_meta is not provided")
         features = env_to_policy_features(env_cfg)
 
-    cfg.output_features = {key: ft for key, ft in features.items() if ft.type is FeatureType.ACTION}
-    cfg.input_features = {key: ft for key, ft in features.items() if key not in cfg.output_features}
+    if not cfg.output_features:
+        cfg.output_features = {key: ft for key, ft in features.items() if ft.type is FeatureType.ACTION}
+    if not cfg.input_features:
+        cfg.input_features = {key: ft for key, ft in features.items() if key not in cfg.output_features}
     kwargs["config"] = cfg
+    
+    # Pass dataset_stats to the policy if available (needed for some policies like SARM)
+    if ds_meta is not None and hasattr(ds_meta, 'stats'):
+        kwargs["dataset_stats"] = ds_meta.stats
+    
+    if ds_meta is not None:
+        kwargs["dataset_meta"] = ds_meta
 
     if cfg.pretrained_path:
         # Load a pretrained policy and override the config if needed (for example, if there are inference-time
@@ -393,4 +440,8 @@ def make_policy(
 
     # policy = torch.compile(policy, mode="reduce-overhead")
 
+    if not rename_map:
+        validate_visual_features_consistency(cfg, features)
+        # TODO: (jadechoghari) - add a check_state(cfg, features) and check_action(cfg, features)
+
     return policy

src/lerobot/policies/groot/README.md

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