faster canbus

Signed-off-by: Steven Palma <imstevenpmwork@ieee.org>

.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 libero and 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\[pi\]|lerobot\[libero\]' pyproject.toml | sed 's/^/::warning:: Removing line: /' || true
+          sed -E -i '/@ git\+https|lerobot\[pi\]|lerobot\[libero\]/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 }}

.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 libero or pi tags, you will have to do: `pip install "lerobot[pi,libero]@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
@@ -19,6 +17,8 @@
     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 +37,17 @@
     title: π₀ (Pi0)
   - local: pi05
     title: π₀.₅ (Pi05)
+  - local: groot
+    title: NVIDIA GR00T N1.5
+  title: "Policies"
+- sections:
+  - 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

docs/source/groot.mdx

@@ -0,0 +1,122 @@
+# 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] # consider also installing libero,dev and test tags
+```
+
+## 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
+
+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 libero or pi, you will have to do: `pip install "lerobot[pi,libero]@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/openarms.mdx

@@ -0,0 +1,328 @@
+# OpenArms Robot
+
+OpenArms is a 7 DOF robotic arm with a gripper, designed by [Enactic, Inc.](https://www.enactic.com/) It uses Damiao motors controlled via CAN bus communication and MIT control mode for smooth, precise motion.
+
+## Hardware Overview
+
+- **7 DOF per arm** (14 DOF total for dual arm setup)
+- **1 gripper per arm** (2 grippers total)
+- **Damiao motors** with 4 different types:
+  - **DM8009** (DM-J8009P-2EC) for shoulders (J1, J2) - high torque
+  - **DM4340** for shoulder rotation and elbow (J3, J4)
+  - **DM4310** (DM-J4310-2EC V1.1) for wrist (J5, J6, J7) and gripper (J8)
+- **24V power supply** required
+- **CAN interface device**:
+  - **Linux**: Any SocketCAN-compatible adapter
+  - **macOS**: CANable, PEAK PCAN-USB, or Kvaser USBcan
+- Proper CAN wiring (CANH, CANL, 120Ω termination)
+
+
+## Motor Configuration
+
+Each arm has the following motor configuration based on the [OpenArm setup guide](https://docs.openarm.dev/software/setup/):
+
+| Joint | Motor | Motor Type | Sender CAN ID | Receiver ID | Description |
+|-------|-------|------------|---------------|-------------|-------------|
+| J1 | joint_1 | DM8009 | 0x01 | 0x11 | Shoulder pan |
+| J2 | joint_2 | DM8009 | 0x02 | 0x12 | Shoulder lift |
+| J3 | joint_3 | DM4340 | 0x03 | 0x13 | Shoulder rotation |
+| J4 | joint_4 | DM4340 | 0x04 | 0x14 | Elbow flex |
+| J5 | joint_5 | DM4310 | 0x05 | 0x15 | Wrist roll |
+| J6 | joint_6 | DM4310 | 0x06 | 0x16 | Wrist pitch |
+| J7 | joint_7 | DM4310 | 0x07 | 0x17 | Wrist rotation |
+| J8 | gripper | DM4310 | 0x08 | 0x18 | Gripper |
+
+For dual arm setups, the left arm uses IDs 0x09-0x10 for joints 1-8 with the same motor types.
+
+## Quick Start 
+
+```bash
+# Install system dependencies
+sudo apt install can-utils iproute2
+
+# Install LeRobot with OpenArms support
+pip install -e ".[openarms]"
+```
+
+## Setup Guide
+
+### Step 1: Motor ID Configuration
+
+**IMPORTANT**: Before using the robot, motors must be configured with the correct CAN IDs.
+
+Refer to the [OpenArm Motor ID Configuration Guide](https://docs.openarm.dev/software/setup/motor-id) for detailed instructions using the Damiao Debugging Tools on Windows.
+
+Key points:
+- Each motor needs a unique **Sender CAN ID** (0x01-0x08)
+- Each motor needs a unique **Receiver/Master ID** (0x11-0x18)
+- Use the Damiao Debugging Tools to set these IDs
+
+### Step 2: Setup CAN Interface
+
+Configure your CAN interface as described in the [OpenArm CAN Setup Guide](https://docs.openarm.dev/software/setup/can-setup):
+
+#### Linux (SocketCAN)
+
+```bash
+# Find your CAN interface
+ip link show
+
+# Configure can0, 1, 2, 3
+sudo ip link set can0 down
+sudo ip link set can0 type can bitrate 1000000
+sudo ip link set can0 up
+
+sudo ip link set can1 down
+sudo ip link set can1 type can bitrate 1000000
+sudo ip link set can1 up
+
+sudo ip link set can2 down
+sudo ip link set can2 type can bitrate 1000000
+sudo ip link set can2 up
+
+sudo ip link set can3 down
+sudo ip link set can3 type can bitrate 1000000
+sudo ip link set can3 up
+
+# Verify configuration
+ip link show can0
+```
+
+or run:
+
+`examples/openarms/setup_can.sh`
+
+### Testing canbus and motor connection
+
+Please run this script to check if all motors can be found and to find your can-fd speed: `python examples/openarms/debug_can_communication.py`
+
+## Usage
+
+### Basic Setup
+
+
+```python
+from lerobot.robots.openarms import OpenArmsFollower
+from lerobot.robots.openarms.config_openarms_follower import OpenArmsFollowerConfig
+
+# Configure for dual arm setup
+config = OpenArmsFollowerConfig(
+    port="can0",
+    can_interface="socketcan",  # Or "auto" for auto-detection
+    id="openarms_dual",
+    is_dual_arm=True,
+)
+
+robot = OpenArmsFollower(config)
+robot.connect()
+```
+
+### Calibration
+
+On first use, you'll need to calibrate the robot:
+
+```python
+robot.calibrate()
+```
+
+The calibration process will:
+1. Disable torque on all motors
+2. Ask you to position arms in **hanging position with grippers closed**
+3. Set this as the zero position
+4. Ask you to move each joint through its full range
+5. Record min/max positions for each joint
+6. Save calibration to file
+
+### Reading Observations
+
+The robot provides comprehensive state information:
+
+```python
+observation = robot.get_observation()
+
+# Observation includes for each motor:
+# - {motor_name}.pos: Position in degrees
+# - {motor_name}.vel: Velocity in degrees/second  
+# - {motor_name}.torque: Motor torque
+# - {camera_name}: Camera images (if configured)
+
+print(f"Right arm joint 1 position: {observation['right_joint_1.pos']:.1f}°")
+print(f"Right arm joint 1 velocity: {observation['right_joint_1.vel']:.1f}°/s")
+print(f"Right arm joint 1 torque: {observation['right_joint_1.torque']:.3f} N·m")
+```
+
+### Sending Actions
+
+```python
+# Send target positions (in degrees)
+action = {
+    "right_joint_1.pos": 45.0,
+    "right_joint_2.pos": -30.0,
+    # ... all joints
+    "right_gripper.pos": 45.0,  # Half-closed
+}
+
+actual_action = robot.send_action(action)
+```
+
+### Gripper Control
+
+```python
+# Open gripper
+robot.open_gripper(arm="right")
+
+# Close gripper  
+robot.close_gripper(arm="right")
+```
+
+## Safety Features
+
+### 1. Maximum Relative Target
+
+Limits how far a joint can move in a single command to prevent sudden movements:
+
+```python
+config = OpenArmsFollowerConfig(
+    port="can0",
+    # Limit all joints to 10 degrees per command
+    max_relative_target=10.0,
+    
+    # Or set per-motor limits
+    max_relative_target={
+        "right_joint_1": 15.0,  # Slower moving joint
+        "right_joint_2": 10.0,
+        "right_gripper": 5.0,   # Very slow gripper
+    }
+)
+```
+
+**How it works**: If current position is 50° and you command 80°, with `max_relative_target=10.0`, the robot will only move to 60° in that step.
+
+### 2. Torque Limits
+
+Control maximum torque output, especially important for grippers and teleoperation:
+
+```python
+config = OpenArmsFollowerConfig(
+    port="can0",
+    # Gripper torque limit (fraction of motor's max torque)
+    gripper_torque_limit=0.5,  # 50% of max torque
+)
+```
+
+Lower torque limits prevent damage when gripping delicate objects.
+
+### 3. MIT Control Gains
+
+Control responsiveness and stability via PID-like gains:
+
+```python
+config = OpenArmsFollowerConfig(
+    port="can0",
+    position_kp=10.0,  # Position gain (higher = more responsive)
+    position_kd=0.5,   # Velocity damping (higher = more damped)
+)
+```
+
+**Guidelines**:
+- **For following (robot)**: Higher gains for responsiveness
+  - `position_kp=10.0`, `position_kd=0.5`
+- **For teleoperation (leader)**: Lower gains or disable torque for manual movement
+  - `manual_control=True` (torque disabled)
+
+### 4. Velocity Limits
+
+Velocity limits are enforced by the Damiao motors based on motor type. For DM4310:
+- Max velocity: 30 rad/s ≈ 1718°/s
+
+The motors will automatically limit velocity to safe values.
+
+## Teleoperation
+
+### Leader Arm Setup
+
+The leader arm is moved manually (torque disabled) to generate commands:
+
+```python
+from lerobot.teleoperators.openarms import OpenArmsLeader
+from lerobot.teleoperators.openarms.config_openarms_leader import OpenArmsLeaderConfig
+
+config = OpenArmsLeaderConfig(
+    port="can1",  # Separate CAN interface for leader
+    id="openarms_leader",
+    manual_control=True,  # Torque disabled for manual movement
+    is_dual_arm=True,
+)
+
+leader = OpenArmsLeader(config)
+leader.connect()
+
+# Read current position as action
+action = leader.get_action()
+# action contains positions for all joints in degrees
+```
+
+### Safety Considerations for Teleoperation
+
+1. **Use separate CAN interfaces** for leader and follower to avoid conflicts
+2. **Enable max_relative_target** on follower to smooth abrupt movements
+3. **Lower torque limits** on follower to prevent damage from tracking errors
+4. **Test with one arm** before enabling dual arm teleoperation
+5. **Have emergency stop** ready (power switch or CAN disable)
+
+```python
+# Recommended follower config for teleoperation
+follower_config = OpenArmsFollowerConfig(
+    port="can0",
+    max_relative_target=5.0,  # Small steps for smooth following
+    gripper_torque_limit=0.3, # Low torque for safety
+    position_kp=5.0,  # Lower gains for gentler following
+    position_kd=0.3,
+)
+```
+
+## Troubleshooting
+
+### Motor Shaking/Unstable
+
+- **Lower control gains**: Reduce `position_kp` and `position_kd`
+- **Check calibration**: Re-run calibration procedure
+- **Verify power**: Insufficient current can cause instability
+- **Check mechanical**: Loose connections, binding, or damaged components
+
+### CAN Bus Errors
+
+```bash
+# Check for errors
+ip -s link show can0
+
+# Reset CAN interface
+sudo ip link set can0 down
+sudo ip link set can0 up
+```
+
+### Control Mode
+
+OpenArms uses **MIT control mode** which allows simultaneous control of:
+- Position (degrees)
+- Velocity (degrees/second)
+- Torque (N·m)
+- Position gain (Kp)
+- Velocity damping (Kd)
+
+### Communication
+
+- **Protocol**: CAN 2.0 at 1 Mbps (or CAN-FD at 5 Mbps)
+- **Frame format**: Standard 11-bit IDs
+- **Update rate**: Typically 50-100 Hz depending on motor count
+- **Latency**: ~10-20ms per motor command
+
+## References
+
+- [OpenArm Official Documentation](https://docs.openarm.dev/)
+- [OpenArm Setup Guide](https://docs.openarm.dev/software/setup/)
+- [Motor ID Configuration](https://docs.openarm.dev/software/setup/motor-id)
+- [CAN Interface Setup](https://docs.openarm.dev/software/setup/can-setup)
+- [Motor Communication Test](https://docs.openarm.dev/software/setup/configure-test)
+- [Damiao Motor Documentation](https://wiki.seeedstudio.com/damiao_series/)
+- [Enactic GitHub](https://github.com/enactic/openarm_can)

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

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/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/openarms/debug_can_communication.py

@@ -0,0 +1,416 @@
+#!/usr/bin/env python3
+"""
+Comprehensive debug script for OpenArms CAN FD communication.
+Tests all 4 CAN interfaces with CAN FD support.
+"""
+
+import can
+import time
+import sys
+import subprocess
+
+def check_can_interface(port):
+    """Check if CAN interface is UP and configured."""
+    try:
+        result = subprocess.run(['ip', 'link', 'show', port], 
+                              capture_output=True, text=True)
+        if result.returncode != 0:
+            return False, "Interface not found", None
+        
+        output = result.stdout
+        if 'UP' not in output:
+            return False, "Interface is DOWN", None
+        
+        # Check if CAN FD is enabled
+        is_fd = 'fd on' in output.lower() or 'canfd' in output.lower()
+        
+        return True, "Interface is UP", is_fd
+    except FileNotFoundError:
+        return None, "Cannot check (ip command not found)", None
+
+
+def test_motor_on_interface(bus, motor_id, timeout=2.0, use_fd=False):
+    """
+    Test a single motor and return all responses.
+    
+    Returns:
+        list of (arbitration_id, data) tuples for all responses received
+    """
+    # Send enable command
+    enable_msg = can.Message(
+        arbitration_id=motor_id,
+        data=[0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFC],
+        is_extended_id=False,
+        is_fd=use_fd
+    )
+    
+    try:
+        bus.send(enable_msg)
+    except Exception as e:
+        return None, f"Send error: {e}"
+    
+    # Listen for responses
+    responses = []
+    start_time = time.time()
+    
+    while time.time() - start_time < timeout:
+        msg = bus.recv(timeout=0.1)
+        if msg:
+            responses.append((msg.arbitration_id, msg.data, msg.is_fd if hasattr(msg, 'is_fd') else False))
+    
+    # Send disable command
+    disable_msg = can.Message(
+        arbitration_id=motor_id,
+        data=[0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFD],
+        is_extended_id=False,
+        is_fd=use_fd
+    )
+    try:
+        bus.send(disable_msg)
+    except:
+        pass
+    
+    return responses, None
+
+
+def test_interface(port, interface_type="socketcan", use_can_fd=True):
+    """Test all 8 motors on a single CAN interface."""
+    
+    results = {
+        'interface': port,
+        'status': None,
+        'is_fd': use_can_fd,
+        'motors': {}
+    }
+    
+    # Check interface status
+    status_ok, status_msg, interface_has_fd = check_can_interface(port)
+    
+    if interface_has_fd is not None:
+        results['interface_fd_enabled'] = interface_has_fd
+        if use_can_fd and not interface_has_fd:
+            status_msg += " (CAN FD NOT enabled on interface!)"
+        elif interface_has_fd:
+            status_msg += " (CAN FD enabled)"
+    
+    results['status'] = status_msg
+    
+    if status_ok is False:
+        return results
+    
+    # Try to connect
+    try:
+        if use_can_fd:
+            print(f"  Connecting to {port} with CAN FD (1 Mbps / 5 Mbps)...")
+            bus = can.interface.Bus(
+                channel=port,
+                interface=interface_type,
+                bitrate=1000000,
+                data_bitrate=5000000,
+                fd=True
+            )
+        else:
+            print(f"  Connecting to {port} with CAN 2.0 (1 Mbps)...")
+            bus = can.interface.Bus(
+                channel=port,
+                interface=interface_type,
+                bitrate=1000000
+            )
+    except Exception as e:
+        results['status'] = f"Connection failed: {e}"
+        return results
+    
+    try:
+        # Clear any pending messages
+        while bus.recv(timeout=0.01):
+            pass
+        
+        # Test each motor (0x01 to 0x08)
+        for motor_id in range(0x01, 0x09):
+            responses, error = test_motor_on_interface(bus, motor_id, timeout=1.0, use_fd=use_can_fd)
+            
+            if error:
+                results['motors'][motor_id] = {'error': error}
+            elif responses:
+                results['motors'][motor_id] = {
+                    'found': True,
+                    'responses': responses
+                }
+            else:
+                results['motors'][motor_id] = {
+                    'found': False,
+                    'responses': []
+                }
+            
+            time.sleep(0.05)  # Small delay between motors
+        
+    finally:
+        bus.shutdown()
+    
+    return results
+
+
+def print_results(all_results):
+    """Print formatted results for all interfaces."""
+    
+    print("SUMMARY - Motors Found on Each Interface")
+    
+    motor_names = {
+        0x01: "joint_1 (Shoulder pan)",
+        0x02: "joint_2 (Shoulder lift)",
+        0x03: "joint_3 (Shoulder rotation)",
+        0x04: "joint_4 (Elbow flex)",
+        0x05: "joint_5 (Wrist roll)",
+        0x06: "joint_6 (Wrist pitch)",
+        0x07: "joint_7 (Wrist rotation)",
+        0x08: "gripper",
+    }
+    
+    total_found = 0
+    
+    for result in all_results:
+        interface = result['interface']
+        status = result['status']
+        
+        print(f"{interface}: {status}")
+        if result.get('is_fd'):
+            print(f"  Mode: CAN FD")
+        else:
+            print(f"  Mode: CAN 2.0")
+        
+        if 'Connection failed' in status or 'DOWN' in status:
+            print(f"  ⚠ Cannot test {interface}")
+            continue
+        
+        motors_found = 0
+        
+        for motor_id in range(0x01, 0x09):
+            motor_data = result['motors'].get(motor_id, {})
+            motor_name = motor_names.get(motor_id, "Unknown")
+            
+            if motor_data.get('error'):
+                print(f"  Motor 0x{motor_id:02X} ({motor_name}): ✗ {motor_data['error']}")
+            elif motor_data.get('found'):
+                motors_found += 1
+                total_found += 1
+                responses = motor_data['responses']
+                print(f"  Motor 0x{motor_id:02X} ({motor_name}): ✓ FOUND")
+                
+                for resp_id, data, is_fd in responses:
+                    data_hex = data.hex()
+                    fd_flag = " [FD]" if is_fd else " [2.0]"
+                    print(f"    → Response from 0x{resp_id:02X}{fd_flag}: {data_hex}")
+            else:
+                print(f"  Motor 0x{motor_id:02X} ({motor_name}): ✗ No response")
+        
+        print(f"\n  Summary: {motors_found}/8 motors found on {interface}")
+    
+    # Overall summary
+    print("OVERALL SUMMARY")
+    print(f"Total motors found across all interfaces: {total_found}")
+    
+    # Analyze configuration
+    print("DIAGNOSIS")
+    
+    for result in all_results:
+        interface = result['interface']
+        motors_found = sum(1 for m in result['motors'].values() if m.get('found'))
+        
+        if motors_found == 0:
+            print(f"\n⚠ {interface}: NO MOTORS FOUND")
+            print("  Possible issues:")
+            print("    1. CAN FD mode mismatch (interface vs motor configuration)")
+            print("    2. Missing 120Ω termination resistors at BOTH cable ends")
+            print("    3. Motor timeout parameter set incorrectly (should NOT be 0)")
+            print("    4. CANH/CANL wiring issue")
+            print("    5. Cable too long (>40m for CAN FD at 5Mbps)")
+            
+            # Check FD mismatch
+            if result.get('is_fd') and not result.get('interface_fd_enabled'):
+                print("    ⚠️ CRITICAL: Trying CAN FD but interface NOT configured for FD!")
+                print(f"       Fix: sudo ip link set {interface} type can bitrate 1000000 dbitrate 5000000 fd on")
+                
+        elif motors_found < 8:
+            print(f"\n⚠ {interface}: Only {motors_found}/8 motors responding")
+            print("  Check power and connections for missing motors")
+        else:
+            print(f"\n✓ {interface}: All 8 motors responding correctly!")
+    
+    # Check for unexpected response IDs
+    print("RESPONSE ID ANALYSIS")
+    
+    for result in all_results:
+        interface = result['interface']
+        unexpected = []
+        
+        for motor_id, motor_data in result['motors'].items():
+            if motor_data.get('found'):
+                expected_id = motor_id + 0x10
+                actual_ids = [resp[0] for resp in motor_data['responses']]
+                
+                if expected_id not in actual_ids:
+                    unexpected.append((motor_id, actual_ids))
+        
+        if unexpected:
+            print(f"\n⚠ {interface}: Unexpected response IDs detected")
+            for motor_id, actual_ids in unexpected:
+                expected_id = motor_id + 0x10
+                print(f"  Motor 0x{motor_id:02X}: Expected 0x{expected_id:02X}, "
+                      f"got {[f'0x{id:02X}' for id in actual_ids]}")
+            print("  → Motor Master IDs need reconfiguration")
+        else:
+            motors_found = sum(1 for m in result['motors'].values() if m.get('found'))
+            if motors_found > 0:
+                print(f"\n✓ {interface}: All responding motors use correct IDs")
+
+
+def test_communication_speed(interface, motor_id, num_iterations=100):
+    """
+    Test communication speed with a motor.
+    
+    Returns:
+        tuple: (hz, avg_latency_ms) or (None, None) if test failed
+    """
+    try:
+        # Connect to interface
+        bus = can.interface.Bus(
+            channel=interface,
+            interface="socketcan",
+            bitrate=1000000,
+            data_bitrate=5000000,
+            fd=True
+        )
+        
+        # Send refresh commands and measure round-trip time
+        latencies = []
+        successful = 0
+        
+        for _ in range(num_iterations):
+            start = time.perf_counter()
+            
+            # Send enable command (lightweight operation)
+            enable_msg = can.Message(
+                arbitration_id=motor_id,
+                data=[0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFC],
+                is_extended_id=False,
+                is_fd=True
+            )
+            bus.send(enable_msg)
+            
+            # Wait for response
+            msg = bus.recv(timeout=0.1)
+            
+            if msg:
+                latency = (time.perf_counter() - start) * 1000  # Convert to ms
+                latencies.append(latency)
+                successful += 1
+        
+        bus.shutdown()
+        
+        if successful > 0:
+            avg_latency = sum(latencies) / len(latencies)
+            hz = 1000.0 / avg_latency if avg_latency > 0 else 0
+            return hz, avg_latency
+        
+        return None, None
+        
+    except Exception as e:
+        print(f"    Speed test error: {e}")
+        return None, None
+
+
+def main():
+    """Main function to test all CAN interfaces with CAN FD."""
+    
+    print("\nThis will test all 4 CAN interfaces (can0-can3) with CAN FD")
+    print("Testing motors 0x01-0x08 on each interface")
+    print()
+    print("Make sure:")
+    print("  ✓ Motors are powered (24V)")
+    print("  ✓ CAN interfaces configured with FD mode:")
+    print("    ./examples/openarms/setup_can.sh")
+    print("  ✓ Motor 'timeout' parameter NOT set to 0 (use Damiao tools)")
+    print("  ✓ CAN wiring includes 120Ω termination at BOTH ends")
+    print()
+    
+    input("Press ENTER to start testing...")
+    
+    # Test all 4 interfaces with CAN FD
+    all_results = []
+    
+    for i in range(4):
+        interface = f"can{i}"
+        print(f"Testing {interface}...")
+        
+        result = test_interface(interface, use_can_fd=True)
+        all_results.append(result)
+        
+        # Quick status
+        if 'Connection failed' in result['status'] or 'DOWN' in result['status']:
+            print(f"  ⚠ {interface}: {result['status']}")
+        else:
+            motors_found = sum(1 for m in result['motors'].values() if m.get('found'))
+            print(f"  {interface}: {motors_found}/8 motors found")
+        
+        time.sleep(0.2)
+    
+    # Print detailed results
+    print_results(all_results)
+    
+    print("Testing Complete!")
+    
+    all_found = sum(sum(1 for m in r['motors'].values() if m.get('found')) for r in all_results)
+    
+    if all_found == 0:
+        print("\n⚠️ CRITICAL: No motors found on any interface!")
+        print("\nTop issues to check:")
+        print("  1. Motor 'timeout' parameter (use Damiao tools to set > 0)")
+        print("  2. CAN FD not enabled (run ./examples/openarms/setup_can.sh)")
+        print("  3. Missing termination resistors")
+        print("\nTry:")
+        print("  a) Check motor parameters with Damiao Debugging Tools")
+        print("  b) Verify CAN FD is enabled: ip -d link show can0 | grep fd")
+        print("  c) Run setup script: ./examples/openarms/setup_can.sh")
+    else:
+        # Run speed test on interfaces with motors
+        print("COMMUNICATION SPEED TEST")
+        print("\nTesting maximum communication frequency...")
+        
+        for result in all_results:
+            interface = result['interface']
+            
+            # Find first responding motor
+            responding_motor = None
+            for motor_id, motor_data in result['motors'].items():
+                if motor_data.get('found'):
+                    responding_motor = motor_id
+                    break
+            
+            if responding_motor:
+                print(f"\n{interface}: Testing with motor 0x{responding_motor:02X}...")
+                hz, latency = test_communication_speed(interface, responding_motor, num_iterations=100)
+                
+                if hz:
+                    print(f"  ✓ Max frequency: {hz:.1f} Hz")
+                    print(f"  ✓ Avg latency: {latency:.2f} ms")
+                    print(f"  ✓ Commands per second: ~{int(hz)}")
+                else:
+                    print(f"  ✗ Speed test failed")
+            else:
+                print(f"\n{interface}: No motors found, skipping speed test")
+
+        print()
+
+
+if __name__ == "__main__":
+    try:
+        main()
+    except KeyboardInterrupt:
+        print("\n\nTesting interrupted by user.")
+        sys.exit(1)
+    except Exception as e:
+        print(f"\nUnexpected error: {e}")
+        import traceback
+        traceback.print_exc()
+        sys.exit(1)
+

examples/openarms/setup_can.sh

@@ -0,0 +1,73 @@
+#!/bin/bash
+# Setup all OpenArms CAN interfaces with CAN FD
+
+set -e
+
+echo "=========================================="
+echo "OpenArms CAN FD Interface Setup"
+echo "=========================================="
+echo ""
+echo "Mode: CAN FD"
+echo "  - Nominal bitrate: 1 Mbps"
+echo "  - Data bitrate: 5 Mbps"
+echo ""
+echo "Configuring interfaces can0, can1, can2, can3..."
+echo ""
+
+# Configure each CAN interface with CAN FD
+for i in 0 1 2 3; do
+    interface="can$i"
+    
+    # Check if interface exists
+    if ! ip link show "$interface" &> /dev/null; then
+        echo "⚠ $interface: Not found, skipping"
+        continue
+    fi
+    
+    # Bring down interface
+    sudo ip link set "$interface" down 2>/dev/null
+    
+    # Configure CAN FD mode
+    sudo ip link set "$interface" type can \
+        bitrate 1000000 \
+        dbitrate 5000000 \
+        fd on
+    
+    # Bring up interface
+    sudo ip link set "$interface" up
+    
+    # Verify configuration
+    if ip link show "$interface" | grep -q "UP"; then
+        echo "✓ $interface: Configured and UP"
+    else
+        echo "✗ $interface: Failed to bring UP"
+    fi
+done
+
+echo ""
+echo "=========================================="
+echo "Verification"
+echo "=========================================="
+echo ""
+
+# Show detailed status for each interface
+for i in 0 1 2 3; do
+    interface="can$i"
+    if ip link show "$interface" &> /dev/null; then
+        echo "$interface:"
+        # Show key parameters
+        ip -d link show "$interface" | grep -E "can|state|bitrate|dbitrate" | head -3
+        echo ""
+    fi
+done
+
+echo "=========================================="
+echo "Setup Complete!"
+echo "=========================================="
+echo ""
+echo "All interfaces configured for CAN FD mode"
+echo ""
+echo "Next steps:"
+echo "  1. Test motors: python debug_can_communication.py"
+echo "  2. Run teleoperation: python examples/openarms/teleop.py"
+echo ""

examples/openarms/teleop.py

@@ -0,0 +1,148 @@
+"""
+OpenArms Teleoperation Example - Full Dual Arms
+
+This script demonstrates teleoperation of OpenArms follower robot using an OpenArms leader arm.
+It first calibrates both devices, then enters a teleoperation loop for both arms.
+"""
+
+import time
+
+from lerobot.robots.openarms.openarms_follower import OpenArmsFollower
+from lerobot.robots.openarms.config_openarms_follower import OpenArmsFollowerConfig
+from lerobot.teleoperators.openarms.openarms_leader import OpenArmsLeader
+from lerobot.teleoperators.openarms.config_openarms_leader import OpenArmsLeaderConfig
+
+
+follower_config = OpenArmsFollowerConfig(
+    port_left="can0",   # CAN interface for follower left arm
+    port_right="can1",  # CAN interface for follower right arm
+    can_interface="socketcan",  # Linux SocketCAN
+    id="openarms_follower",
+    disable_torque_on_disconnect=True,
+    max_relative_target=5.0,  # Safety limit
+)
+
+
+leader_config = OpenArmsLeaderConfig(
+    port_left="can2",   # CAN interface for leader left arm
+    port_right="can3",  # CAN interface for leader right arm
+    can_interface="socketcan",  # Linux SocketCAN
+    id="openarms_leader",
+    manual_control=True,  # Enable manual control (torque disabled)
+)
+
+print("=" * 60)
+print("OpenArms Teleoperation - Full Dual Arms")
+print("=" * 60)
+
+# Initialize devices
+print("\n[1/4] Initializing devices...")
+follower = OpenArmsFollower(follower_config)
+leader = OpenArmsLeader(leader_config)
+
+# Connect and calibrate follower
+print("\n[2/4] Connecting and calibrating follower robot...")
+print("Note: If you have existing calibration, just press ENTER to use it.")
+follower.connect(calibrate=True)
+
+# Connect and calibrate leader
+print("\n[3/4] Connecting and calibrating leader arm...")
+print("Note: The leader arm will have torque disabled for manual control.")
+leader.connect(calibrate=True)
+
+# Wait for user to be ready
+print("\n[4/4] Ready for teleoperation!")
+print("\nBoth arms will be controlled (16 motors total):")
+print("  RIGHT ARM: joints 1-7 + gripper")
+print("  LEFT ARM: joints 1-7 + gripper")
+
+print("\nPress ENTER to start teleoperation...")
+input()
+
+print("\nTeleoperation started! Move both leader arms.")
+print("Press Ctrl+C to stop.\n")
+
+# All joints for both arms (16 motors total)
+all_joints = [
+    # Right arm
+    "right_joint_1",
+    "right_joint_2",
+    "right_joint_3",
+    "right_joint_4",
+    "right_joint_5",
+    "right_joint_6",
+    "right_joint_7",
+    "right_gripper",
+    # Left arm
+    "left_joint_1",
+    "left_joint_2",
+    "left_joint_3",
+    "left_joint_4",
+    "left_joint_5",
+    "left_joint_6",
+    "left_joint_7",
+    "left_gripper",
+]
+
+# Performance monitoring
+loop_times = []
+start_time = time.perf_counter()
+last_print_time = start_time
+
+try:
+    while True:
+        loop_start = time.perf_counter()
+        
+        # Get action from leader
+        leader_action = leader.get_action()
+        
+        # Filter to only position data for all joints (both arms)
+        joint_action = {}
+        for joint in all_joints:
+            pos_key = f"{joint}.pos"
+            if pos_key in leader_action:
+                joint_action[pos_key] = leader_action[pos_key]
+        
+        # Send action to follower (both arms)
+        if joint_action:
+            follower.send_action(joint_action)
+        
+        # Measure loop time
+        loop_end = time.perf_counter()
+        loop_time = loop_end - loop_start
+        loop_times.append(loop_time)
+        
+        # Print stats every 2 seconds
+        if loop_end - last_print_time >= 2.0:
+            if loop_times:
+                avg_time = sum(loop_times) / len(loop_times)
+                current_hz = 1.0 / avg_time if avg_time > 0 else 0
+                min_time = min(loop_times)
+                max_time = max(loop_times)
+                max_hz = 1.0 / min_time if min_time > 0 else 0
+                min_hz = 1.0 / max_time if max_time > 0 else 0
+                
+                print(f"[Hz Stats] Avg: {current_hz:.1f} Hz | "
+                      f"Range: {min_hz:.1f}-{max_hz:.1f} Hz | "
+                      f"Avg loop time: {avg_time*1000:.1f} ms")
+                
+                # Reset for next measurement window
+                loop_times = []
+                last_print_time = loop_end
+            
+except KeyboardInterrupt:
+    print("\n\nStopping teleoperation...")
+finally:
+    # Disconnect devices
+    print("Disconnecting devices...")
+    try:
+        follower.disconnect()
+    except Exception as e:
+        print(f"Error disconnecting follower: {e}")
+    
+    try:
+        leader.disconnect()
+    except Exception as e:
+        print(f"Error disconnecting leader: {e}")
+    
+    print("Done!")

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/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 = "0.4.1"
 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",
@@ -95,13 +97,15 @@ 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"]
 dynamixel = ["dynamixel-sdk>=3.7.31,<3.9.0"]
+damiao = ["python-can>=4.2.0,<5.0.0"]
 
 # Robots
+openarms = ["lerobot[damiao]"]
 gamepad = ["lerobot[pygame-dep]", "hidapi>=0.14.0,<0.15.0"]
 hopejr = ["lerobot[feetech]", "lerobot[pygame-dep]"]
 lekiwi = ["lerobot[feetech]", "pyzmq>=26.2.1,<28.0.0"]
@@ -111,17 +115,23 @@ 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"]
@@ -132,15 +142,15 @@ 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"]
-
+metaworld = ["metaworld==3.0.0"]
 
 # All
 all = [
     "lerobot[dynamixel]",
+    "lerobot[openarms]",
     "lerobot[gamepad]",
     "lerobot[hopejr]",
     "lerobot[lekiwi]",
@@ -149,6 +159,7 @@ 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 +167,9 @@ all = [
     "lerobot[video_benchmark]",
     "lerobot[aloha]",
     "lerobot[pusht]",
-    "lerobot[xarm]",
     "lerobot[phone]",
     "lerobot[libero]",
+    "lerobot[metaworld]",
 ]
 
 [project.scripts]
@@ -233,9 +244,6 @@ 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 +258,8 @@ default.extend-ignore-identifiers-re = [
     "pn",
     "ser",
     "ein",
+    "thw",
+    "inpt",
 ]
 
 # TODO: Uncomment when ready to use
@@ -288,7 +298,6 @@ ignore_errors = true
 
 [[tool.mypy.overrides]]
 module = "lerobot.envs.*"
-# Enable type checking only for the envs module
 ignore_errors = false
 
 
@@ -296,17 +305,22 @@ 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 +330,9 @@ 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]

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 +64,18 @@ class TrainPipelineConfig(HubMixin):
     scheduler: LRSchedulerConfig | None = None
     eval: EvalConfig = field(default_factory=EvalConfig)
     wandb: WandBConfig = field(default_factory=WandBConfig)
+    checkpoint_path: Path | None = field(init=False, default=None)
+    # 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):
-        self.checkpoint_path = None
-
-    def validate(self):
+    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 +83,22 @@ 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 +135,8 @@ 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 +148,13 @@ 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 +190,6 @@ 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 +42,4 @@ class NormalizationMode(str, Enum):
 @dataclass
 class PolicyFeature:
     type: FeatureType
-    shape: tuple
+    shape: tuple[int, ...]

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
 
@@ -43,7 +45,6 @@ from lerobot.datasets.utils import (
     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 +269,79 @@ 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 +349,18 @@ 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 +373,46 @@ 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 +427,17 @@ 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 +544,7 @@ 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 +902,25 @@ 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 +936,7 @@ 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 +962,34 @@ def _copy_data_with_feature_changes(
     remove_features: list[str] | None = None,
 ) -> None:
     """Copy data while adding or removing features."""
-    file_paths = set()
+    if dataset.meta.episodes is None:
+        dataset.meta.episodes = load_episodes(dataset.meta.root)
+
+    # Map file paths to episode indices to extract chunk/file indices
+    file_to_episodes: dict[Path, set[int]] = {}
     for ep_idx in range(dataset.meta.total_episodes):
-        file_paths.add(dataset.meta.get_data_file_path(ep_idx))
+        file_path = dataset.meta.get_data_file_path(ep_idx)
+        if file_path not in file_to_episodes:
+            file_to_episodes[file_path] = set()
+        file_to_episodes[file_path].add(ep_idx)
 
     frame_idx = 0
 
-    for src_path in tqdm(sorted(file_paths), desc="Processing data files"):
+    for src_path in tqdm(sorted(file_to_episodes.keys()), desc="Processing data files"):
         df = pd.read_parquet(dataset.root / src_path).reset_index(drop=True)
 
+        # Get chunk_idx and file_idx from the source file's first episode
+        episodes_in_file = file_to_episodes[src_path]
+        first_ep_idx = min(episodes_in_file)
+        src_ep = dataset.meta.episodes[first_ep_idx]
+        chunk_idx = src_ep["data/chunk_index"]
+        file_idx = src_ep["data/file_index"]
+
         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 +1002,18 @@ 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:
                         df[feature_name] = feature_slice
-                    frame_idx = end_idx
+            frame_idx = end_idx
 
-        _save_data_chunk(df, new_meta)
+        # Write using the preserved chunk_idx and file_idx from 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/lerobot_dataset.py

@@ -686,6 +686,7 @@ 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)
 
@@ -708,7 +709,8 @@ 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()
 
@@ -835,14 +837,14 @@ class LeRobotDataset(torch.utils.data.Dataset):
         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
@@ -854,7 +856,18 @@ class LeRobotDataset(torch.utils.data.Dataset):
             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)
@@ -1231,6 +1244,7 @@ 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 +1256,7 @@ 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 +1267,7 @@ 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 +1281,7 @@ 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
@@ -1472,6 +1488,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         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/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/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

@@ -37,6 +37,16 @@ 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 +143,6 @@ 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
@@ -306,3 +277,45 @@ class LiberoEnv(EnvConfig):
             "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}")
+
+    @property
+    def gym_kwargs(self) -> dict:
+        return {
+            "obs_type": self.obs_type,
+            "render_mode": self.render_mode,
+        }

src/lerobot/envs/factory.py

@@ -16,8 +16,9 @@
 import importlib
 
 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
 
 
 def make_env_config(env_type: str, **kwargs) -> EnvConfig:
@@ -25,8 +26,6 @@ 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:
@@ -74,20 +73,38 @@ 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

@@ -260,19 +260,23 @@ 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
 

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/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/__init__.py

@@ -14,4 +14,11 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .motors_bus import Motor, MotorCalibration, MotorNormMode, MotorsBus
+from .motors_bus import (
+    Motor,
+    MotorCalibration,
+    MotorNormMode,
+    MotorsBus,  # Backward compatibility (alias for SerialMotorsBus)
+    MotorsBusBase,
+    SerialMotorsBus,
+)

src/lerobot/motors/damiao/__init__.py

@@ -14,5 +14,5 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .config_viperx import ViperXConfig
-from .viperx import ViperX
+from .damiao import DamiaoMotorsBus
+from .tables import *

src/lerobot/motors/damiao/damiao.py

@@ -0,0 +1,787 @@
+# 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.
+
+# TODO(pepijn): add license of: https://github.com/cmjang/DM_Control_Python?tab=MIT-1-ov-file#readme
+
+import logging
+import time
+from contextlib import contextmanager
+from copy import deepcopy
+from functools import cached_property
+from typing import Dict, List, Optional, Tuple, Union
+
+import can
+import numpy as np
+
+from lerobot.motors import Motor, MotorCalibration, MotorNormMode, MotorsBusBase
+from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+from lerobot.utils.utils import enter_pressed, move_cursor_up
+
+from .tables import (
+    AVAILABLE_BAUDRATES,
+    CAN_CMD_DISABLE,
+    CAN_CMD_ENABLE,
+    CAN_CMD_REFRESH,
+    CAN_CMD_SET_ZERO,
+    CAN_PARAM_ID,
+    DEFAULT_BAUDRATE,
+    DEFAULT_TIMEOUT_MS,
+    MODEL_RESOLUTION,
+    MOTOR_LIMIT_PARAMS,
+    NORMALIZED_DATA,
+    MotorType,
+)
+
+logger = logging.getLogger(__name__)
+
+NameOrID = Union[str, int]
+Value = Union[int, float]
+
+
+class DamiaoMotorsBus(MotorsBusBase):
+    """
+    The Damiao implementation for a MotorsBus using CAN bus communication.
+    
+    This class uses python-can for CAN bus communication with Damiao motors.
+    For more info, see:
+    - python-can documentation: https://python-can.readthedocs.io/en/stable/
+    - Seedstudio documentation: https://wiki.seeedstudio.com/damiao_series/
+    - DM_Control_Python repo: https://github.com/cmjang/DM_Control_Python
+    """
+
+    # CAN-specific settings
+    available_baudrates = deepcopy(AVAILABLE_BAUDRATES)
+    default_baudrate = DEFAULT_BAUDRATE
+    default_timeout = DEFAULT_TIMEOUT_MS
+    
+    # Motor configuration
+    model_resolution_table = deepcopy(MODEL_RESOLUTION)
+    normalized_data = deepcopy(NORMALIZED_DATA)
+
+    def __init__(
+        self,
+        port: str,
+        motors: dict[str, Motor],
+        calibration: dict[str, MotorCalibration] | None = None,
+        can_interface: str = "auto",
+        use_can_fd: bool = True,
+        bitrate: int = 1000000,
+        data_bitrate: int | None = 5000000,
+    ):
+        """
+        Initialize the Damiao motors bus.
+
+        Args:
+            port: CAN interface name (e.g., "can0" for Linux, "/dev/cu.usbmodem*" for macOS)
+            motors: Dictionary mapping motor names to Motor objects
+            calibration: Optional calibration data
+            can_interface: CAN interface type - "auto" (default), "socketcan" (Linux), or "slcan" (macOS/serial)
+            use_can_fd: Whether to use CAN FD mode (default: True for OpenArms)
+            bitrate: Nominal bitrate in bps (default: 1000000 = 1 Mbps)
+            data_bitrate: Data bitrate for CAN FD in bps (default: 5000000 = 5 Mbps), ignored if use_can_fd is False
+        """
+        super().__init__(port, motors, calibration)
+        self.port = port
+        self.can_interface = can_interface
+        self.use_can_fd = use_can_fd
+        self.bitrate = bitrate
+        self.data_bitrate = data_bitrate
+        self.canbus = None
+        self._is_connected = False
+        
+        # Map motor names to CAN IDs
+        self._motor_can_ids = {}
+        self._recv_id_to_motor = {}
+        
+        # Store motor types and recv IDs
+        self._motor_types = {}
+        for name, motor in self.motors.items():
+            if hasattr(motor, "motor_type"):
+                self._motor_types[name] = motor.motor_type
+            else:
+                # Default to DM4310 if not specified
+                self._motor_types[name] = MotorType.DM4310
+            
+            # Map recv_id to motor name for filtering responses
+            if hasattr(motor, "recv_id"):
+                self._recv_id_to_motor[motor.recv_id] = name
+
+    @property
+    def is_connected(self) -> bool:
+        """Check if the CAN bus is connected."""
+        return self._is_connected and self.canbus is not None
+
+    def connect(self, handshake: bool = True) -> None:
+        """
+        Open the CAN bus and initialize communication.
+
+        Args:
+            handshake: If True, ping all motors to verify they're present
+        """
+        if self.is_connected:
+            raise DeviceAlreadyConnectedError(
+                f"{self.__class__.__name__}('{self.port}') is already connected."
+            )
+
+        try:
+            # Auto-detect interface type based on port name
+            if self.can_interface == "auto":
+                if self.port.startswith("/dev/"):
+                    # Serial device (macOS/Windows)
+                    self.can_interface = "slcan"
+                    logger.info(f"Auto-detected slcan interface for port {self.port}")
+                else:
+                    # Network interface (Linux)
+                    self.can_interface = "socketcan"
+                    logger.info(f"Auto-detected socketcan interface for port {self.port}")
+            
+            # Connect to CAN bus
+            if self.can_interface == "socketcan":
+                # Linux SocketCAN with CAN FD support
+                if self.use_can_fd and self.data_bitrate is not None:
+                    self.canbus = can.interface.Bus(
+                        channel=self.port,
+                        interface="socketcan",
+                        bitrate=self.bitrate,
+                        data_bitrate=self.data_bitrate,
+                        fd=True
+                    )
+                    logger.info(f"Connected to {self.port} with CAN FD (bitrate={self.bitrate}, data_bitrate={self.data_bitrate})")
+                else:
+                    self.canbus = can.interface.Bus(
+                        channel=self.port,
+                        interface="socketcan",
+                        bitrate=self.bitrate
+                    )
+                    logger.info(f"Connected to {self.port} with CAN 2.0 (bitrate={self.bitrate})")
+            elif self.can_interface == "slcan":
+                # Serial Line CAN (macOS, Windows, or USB adapters)
+                # Note: SLCAN typically doesn't support CAN FD
+                self.canbus = can.interface.Bus(
+                    channel=self.port,
+                    interface="slcan",
+                    bitrate=self.bitrate
+                )
+                logger.info(f"Connected to {self.port} with SLCAN (bitrate={self.bitrate})")
+            else:
+                # Generic interface (vector, pcan, etc.)
+                if self.use_can_fd and self.data_bitrate is not None:
+                    self.canbus = can.interface.Bus(
+                        channel=self.port,
+                        interface=self.can_interface,
+                        bitrate=self.bitrate,
+                        data_bitrate=self.data_bitrate,
+                        fd=True
+                    )
+                else:
+                    self.canbus = can.interface.Bus(
+                        channel=self.port,
+                        interface=self.can_interface,
+                        bitrate=self.bitrate
+                    )
+            
+            self._is_connected = True
+            
+            if handshake:
+                self._handshake()
+                
+            logger.debug(f"{self.__class__.__name__} connected via {self.can_interface}.")
+        except Exception as e:
+            self._is_connected = False
+            raise ConnectionError(f"Failed to connect to CAN bus: {e}")
+
+    def _handshake(self) -> None:
+        """Verify all motors are present by refreshing their status."""
+        for motor_name in self.motors:
+            self._refresh_motor(motor_name)
+            time.sleep(0.01)  # Small delay between motors
+
+    def disconnect(self, disable_torque: bool = True) -> None:
+        """
+        Close the CAN bus connection.
+
+        Args:
+            disable_torque: If True, disable torque on all motors before disconnecting
+        """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(
+                f"{self.__class__.__name__}('{self.port}') is not connected."
+            )
+
+        if disable_torque:
+            try:
+                self.disable_torque()
+            except Exception as e:
+                logger.warning(f"Failed to disable torque during disconnect: {e}")
+
+        if self.canbus:
+            self.canbus.shutdown()
+            self.canbus = None
+        self._is_connected = False
+        logger.debug(f"{self.__class__.__name__} disconnected.")
+
+    def configure_motors(self) -> None:
+        """Configure all motors with default settings."""
+        # Damiao motors don't require much configuration in MIT mode
+        # Just ensure they're enabled
+        for motor in self.motors:
+            self._enable_motor(motor)
+            time.sleep(0.01)
+
+    def _enable_motor(self, motor: NameOrID) -> None:
+        """Enable a single motor."""
+        motor_id = self._get_motor_id(motor)
+        recv_id = self._get_motor_recv_id(motor)
+        data = [0xFF] * 7 + [CAN_CMD_ENABLE]
+        msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False)
+        self.canbus.send(msg)
+        self._recv_motor_response(expected_recv_id=recv_id)
+
+    def _disable_motor(self, motor: NameOrID) -> None:
+        """Disable a single motor."""
+        motor_id = self._get_motor_id(motor)
+        recv_id = self._get_motor_recv_id(motor)
+        data = [0xFF] * 7 + [CAN_CMD_DISABLE]
+        msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False)
+        self.canbus.send(msg)
+        self._recv_motor_response(expected_recv_id=recv_id)
+
+    def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+        """Enable torque on selected motors."""
+        motors = self._get_motors_list(motors)
+        for motor in motors:
+            for _ in range(num_retry + 1):
+                try:
+                    self._enable_motor(motor)
+                    break
+                except Exception as e:
+                    if _ == num_retry:
+                        raise e
+                    time.sleep(0.01)
+
+    def disable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+        """Disable torque on selected motors."""
+        motors = self._get_motors_list(motors)
+        for motor in motors:
+            for _ in range(num_retry + 1):
+                try:
+                    self._disable_motor(motor)
+                    break
+                except Exception as e:
+                    if _ == num_retry:
+                        raise e
+                    time.sleep(0.01)
+
+    @contextmanager
+    def torque_disabled(self, motors: str | list[str] | None = None):
+        """
+        Context manager that guarantees torque is re-enabled.
+        
+        This helper is useful to temporarily disable torque when configuring motors.
+        
+        Examples:
+            >>> with bus.torque_disabled():
+            ...     # Safe operations here with torque disabled
+            ...     pass
+        """
+        self.disable_torque(motors)
+        try:
+            yield
+        finally:
+            self.enable_torque(motors)
+
+    def set_zero_position(self, motors: str | list[str] | None = None) -> None:
+        """Set current position as zero for selected motors."""
+        motors = self._get_motors_list(motors)
+        for motor in motors:
+            motor_id = self._get_motor_id(motor)
+            recv_id = self._get_motor_recv_id(motor)
+            data = [0xFF] * 7 + [CAN_CMD_SET_ZERO]
+            msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False)
+            self.canbus.send(msg)
+            self._recv_motor_response(expected_recv_id=recv_id)
+            time.sleep(0.01)
+
+    def _refresh_motor(self, motor: NameOrID) -> Optional[can.Message]:
+        """Refresh motor status and return the response."""
+        motor_id = self._get_motor_id(motor)
+        recv_id = self._get_motor_recv_id(motor)
+        data = [motor_id & 0xFF, (motor_id >> 8) & 0xFF, CAN_CMD_REFRESH, 0, 0, 0, 0, 0]
+        msg = can.Message(arbitration_id=CAN_PARAM_ID, data=data, is_extended_id=False)
+        self.canbus.send(msg)
+        return self._recv_motor_response(expected_recv_id=recv_id)
+
+    def _recv_motor_response(self, expected_recv_id: Optional[int] = None, timeout: float = 0.001) -> Optional[can.Message]:
+        """
+        Receive a response from a motor.
+        
+        Args:
+            expected_recv_id: If provided, only return messages from this CAN ID
+            timeout: Timeout in seconds (default: 1ms for high-speed operation)
+        
+        Returns:
+            CAN message if received, None otherwise
+        """
+        try:
+            start_time = time.time()
+            messages_seen = []
+            while time.time() - start_time < timeout:
+                msg = self.canbus.recv(timeout=0.0001)  # 100us timeout for fast polling
+                if msg:
+                    messages_seen.append(f"0x{msg.arbitration_id:02X}")
+                    # If no filter specified, return any message
+                    if expected_recv_id is None:
+                        return msg
+                    # Otherwise, only return if it matches the expected recv_id
+                    if msg.arbitration_id == expected_recv_id:
+                        return msg
+                    else:
+                        logger.debug(f"Ignoring message from CAN ID 0x{msg.arbitration_id:02X}, expected 0x{expected_recv_id:02X}")
+            
+            # Only log warnings if we're in debug mode to reduce overhead
+            if logger.isEnabledFor(logging.DEBUG):
+                if messages_seen:
+                    logger.debug(f"Received {len(messages_seen)} message(s) from IDs {set(messages_seen)}, but expected 0x{expected_recv_id:02X}")
+                else:
+                    logger.debug(f"No CAN messages received (expected from 0x{expected_recv_id:02X})")
+        except Exception as e:
+            logger.debug(f"Failed to receive CAN message: {e}")
+        return None
+    
+    def _recv_all_responses(self, expected_recv_ids: list[int], timeout: float = 0.002) -> dict[int, can.Message]:
+        """
+        Efficiently receive responses from multiple motors at once.
+        Uses the OpenArms pattern: collect all available messages within timeout.
+        
+        Args:
+            expected_recv_ids: List of CAN IDs we expect responses from
+            timeout: Total timeout in seconds (default: 2ms)
+        
+        Returns:
+            Dictionary mapping recv_id to CAN message
+        """
+        responses = {}
+        expected_set = set(expected_recv_ids)
+        start_time = time.time()
+        
+        try:
+            while len(responses) < len(expected_recv_ids) and (time.time() - start_time) < timeout:
+                msg = self.canbus.recv(timeout=0.0001)  # 100us poll timeout
+                if msg and msg.arbitration_id in expected_set:
+                    responses[msg.arbitration_id] = msg
+                    if len(responses) == len(expected_recv_ids):
+                        break  # Got all responses, exit early
+        except Exception as e:
+            logger.debug(f"Error receiving responses: {e}")
+        
+        return responses
+
+    def _mit_control(
+        self,
+        motor: NameOrID,
+        kp: float,
+        kd: float,
+        position_degrees: float,
+        velocity_deg_per_sec: float,
+        torque: float,
+    ) -> None:
+        """
+        Send MIT control command to a motor.
+
+        Args:
+            motor: Motor name or ID
+            kp: Position gain
+            kd: Velocity gain
+            position_degrees: Target position (degrees)
+            velocity_deg_per_sec: Target velocity (degrees/s)
+            torque: Target torque (N·m)
+        """
+        motor_id = self._get_motor_id(motor)
+        motor_name = self._get_motor_name(motor)
+        motor_type = self._motor_types.get(motor_name, MotorType.DM4310)
+        
+        # Convert degrees to radians for motor control
+        position_rad = np.radians(position_degrees)
+        velocity_rad_per_sec = np.radians(velocity_deg_per_sec)
+        
+        # Get motor limits
+        pmax, vmax, tmax = MOTOR_LIMIT_PARAMS[motor_type]
+        
+        # Encode parameters
+        kp_uint = self._float_to_uint(kp, 0, 500, 12)
+        kd_uint = self._float_to_uint(kd, 0, 5, 12)
+        q_uint = self._float_to_uint(position_rad, -pmax, pmax, 16)
+        dq_uint = self._float_to_uint(velocity_rad_per_sec, -vmax, vmax, 12)
+        tau_uint = self._float_to_uint(torque, -tmax, tmax, 12)
+        
+        # Pack data
+        data = [0] * 8
+        data[0] = (q_uint >> 8) & 0xFF
+        data[1] = q_uint & 0xFF
+        data[2] = dq_uint >> 4
+        data[3] = ((dq_uint & 0xF) << 4) | ((kp_uint >> 8) & 0xF)
+        data[4] = kp_uint & 0xFF
+        data[5] = kd_uint >> 4
+        data[6] = ((kd_uint & 0xF) << 4) | ((tau_uint >> 8) & 0xF)
+        data[7] = tau_uint & 0xFF
+        
+        msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False)
+        self.canbus.send(msg)
+        recv_id = self._get_motor_recv_id(motor)
+        self._recv_motor_response(expected_recv_id=recv_id)
+
+    def _float_to_uint(self, x: float, x_min: float, x_max: float, bits: int) -> int:
+        """Convert float to unsigned integer for CAN transmission."""
+        x = max(x_min, min(x_max, x))  # Clamp to range
+        span = x_max - x_min
+        data_norm = (x - x_min) / span
+        return int(data_norm * ((1 << bits) - 1))
+
+    def _uint_to_float(self, x: int, x_min: float, x_max: float, bits: int) -> float:
+        """Convert unsigned integer from CAN to float."""
+        span = x_max - x_min
+        data_norm = float(x) / ((1 << bits) - 1)
+        return data_norm * span + x_min
+
+    def _decode_motor_state(self, data: bytes, motor_type: MotorType) -> Tuple[float, float, float, int, int]:
+        """
+        Decode motor state from CAN data.
+
+        Returns:
+            Tuple of (position_degrees, velocity_deg_per_sec, torque, temp_mos, temp_rotor)
+        """
+        if len(data) < 8:
+            raise ValueError("Invalid motor state data")
+        
+        # Extract encoded values
+        q_uint = (data[1] << 8) | data[2]
+        dq_uint = (data[3] << 4) | (data[4] >> 4)
+        tau_uint = ((data[4] & 0x0F) << 8) | data[5]
+        t_mos = data[6]
+        t_rotor = data[7]
+        
+        # Get motor limits
+        pmax, vmax, tmax = MOTOR_LIMIT_PARAMS[motor_type]
+        
+        # Decode to physical values (radians)
+        position_rad = self._uint_to_float(q_uint, -pmax, pmax, 16)
+        velocity_rad_per_sec = self._uint_to_float(dq_uint, -vmax, vmax, 12)
+        torque = self._uint_to_float(tau_uint, -tmax, tmax, 12)
+        
+        # Convert to degrees
+        position_degrees = np.degrees(position_rad)
+        velocity_deg_per_sec = np.degrees(velocity_rad_per_sec)
+        
+        return position_degrees, velocity_deg_per_sec, torque, t_mos, t_rotor
+
+    def read(
+        self,
+        data_name: str,
+        motor: str,
+        *,
+        normalize: bool = True,
+        num_retry: int = 0,
+    ) -> Value:
+        """Read a value from a single motor. Positions are always in degrees."""
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+        
+        # Refresh motor to get latest state
+        msg = self._refresh_motor(motor)
+        if msg is None:
+            motor_id = self._get_motor_id(motor)
+            recv_id = self._get_motor_recv_id(motor)
+            raise ConnectionError(
+                f"No response from motor '{motor}' (send ID: 0x{motor_id:02X}, recv ID: 0x{recv_id:02X}). "
+                f"Check that: 1) Motor is powered (24V), 2) CAN wiring is correct, "
+                f"3) Motor IDs are configured correctly using Damiao Debugging Tools"
+            )
+        
+        motor_type = self._motor_types.get(motor, MotorType.DM4310)
+        position_degrees, velocity_deg_per_sec, torque, t_mos, t_rotor = self._decode_motor_state(msg.data, motor_type)
+        
+        # Return requested data (already in degrees for position/velocity)
+        if data_name == "Present_Position":
+            value = position_degrees
+        elif data_name == "Present_Velocity":
+            value = velocity_deg_per_sec
+        elif data_name == "Present_Torque":
+            value = torque
+        elif data_name == "Temperature_MOS":
+            value = t_mos
+        elif data_name == "Temperature_Rotor":
+            value = t_rotor
+        else:
+            raise ValueError(f"Unknown data_name: {data_name}")
+        
+        # For Damiao, positions are always in degrees, no normalization needed
+        # We keep the normalize parameter for compatibility but don't use it
+        return value
+
+    def write(
+        self,
+        data_name: str,
+        motor: str,
+        value: Value,
+        *,
+        normalize: bool = True,
+        num_retry: int = 0,
+    ) -> None:
+        """Write a value to a single motor. Positions are always in degrees."""
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+        
+        # Value is expected to be in degrees for positions
+        if data_name == "Goal_Position":
+            # Use MIT control with position in degrees
+            self._mit_control(motor, 10.0, 0.5, value, 0, 0)
+        else:
+            raise ValueError(f"Writing {data_name} not supported in MIT mode")
+
+    def sync_read(
+        self,
+        data_name: str,
+        motors: str | list[str] | None = None,
+        *,
+        normalize: bool = True,
+        num_retry: int = 0,
+    ) -> Dict[str, Value]:
+        """
+        Read the same value from multiple motors simultaneously.
+        Uses batched operations: sends all refresh commands, then collects all responses.
+        This is MUCH faster than sequential reads (OpenArms pattern).
+        """
+        motors = self._get_motors_list(motors)
+        result = {}
+
+        # Step 1: Send refresh commands to ALL motors first (no waiting)
+        for motor in motors:
+            motor_id = self._get_motor_id(motor)
+            data = [motor_id & 0xFF, (motor_id >> 8) & 0xFF, CAN_CMD_REFRESH, 0, 0, 0, 0, 0]
+            msg = can.Message(arbitration_id=CAN_PARAM_ID, data=data, is_extended_id=False)
+            self.canbus.send(msg)
+        
+        # Step 2: Collect all responses at once (batch receive)
+        expected_recv_ids = [self._get_motor_recv_id(motor) for motor in motors]
+        responses = self._recv_all_responses(expected_recv_ids, timeout=0.003)  # 3ms total timeout
+        
+        # Step 3: Parse responses
+        for motor in motors:
+            try:
+                recv_id = self._get_motor_recv_id(motor)
+                msg = responses.get(recv_id)
+                
+                if msg is None:
+                    logger.warning(f"No response from motor '{motor}' (recv ID: 0x{recv_id:02X})")
+                    result[motor] = 0.0
+                    continue
+                
+                motor_type = self._motor_types.get(motor, MotorType.DM4310)
+                position_degrees, velocity_deg_per_sec, torque, t_mos, t_rotor = self._decode_motor_state(msg.data, motor_type)
+                
+                # Return requested data
+                if data_name == "Present_Position":
+                    value = position_degrees
+                elif data_name == "Present_Velocity":
+                    value = velocity_deg_per_sec
+                elif data_name == "Present_Torque":
+                    value = torque
+                elif data_name == "Temperature_MOS":
+                    value = t_mos
+                elif data_name == "Temperature_Rotor":
+                    value = t_rotor
+                else:
+                    raise ValueError(f"Unknown data_name: {data_name}")
+                
+                result[motor] = value
+                
+            except Exception as e:
+                logger.warning(f"Failed to read {data_name} from {motor}: {e}")
+                result[motor] = 0.0
+        
+        return result
+
+    def sync_write(
+        self,
+        data_name: str,
+        values: Dict[str, Value],
+        *,
+        normalize: bool = True,
+        num_retry: int = 0,
+    ) -> None:
+        """
+        Write different values to multiple motors simultaneously. Positions are always in degrees.
+        Uses batched operations: sends all commands first, then collects responses (OpenArms pattern).
+        """
+        if data_name == "Goal_Position":
+            # Step 1: Send all MIT control commands first (no waiting)
+            for motor, value_degrees in values.items():
+                motor_id = self._get_motor_id(motor)
+                motor_name = self._get_motor_name(motor)
+                motor_type = self._motor_types.get(motor_name, MotorType.DM4310)
+                
+                # Convert degrees to radians
+                position_rad = np.radians(value_degrees)
+                
+                # Default gains for position control
+                kp, kd = 10.0, 0.5
+                
+                # Get motor limits and encode parameters
+                pmax, vmax, tmax = MOTOR_LIMIT_PARAMS[motor_type]
+                kp_uint = self._float_to_uint(kp, 0, 500, 12)
+                kd_uint = self._float_to_uint(kd, 0, 5, 12)
+                q_uint = self._float_to_uint(position_rad, -pmax, pmax, 16)
+                dq_uint = self._float_to_uint(0, -vmax, vmax, 12)
+                tau_uint = self._float_to_uint(0, -tmax, tmax, 12)
+                
+                # Pack data
+                data = [0] * 8
+                data[0] = (q_uint >> 8) & 0xFF
+                data[1] = q_uint & 0xFF
+                data[2] = dq_uint >> 4
+                data[3] = ((dq_uint & 0xF) << 4) | ((kp_uint >> 8) & 0xF)
+                data[4] = kp_uint & 0xFF
+                data[5] = kd_uint >> 4
+                data[6] = ((kd_uint & 0xF) << 4) | ((tau_uint >> 8) & 0xF)
+                data[7] = tau_uint & 0xFF
+                
+                msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False)
+                self.canbus.send(msg)
+            
+            # Step 2: Collect all responses at once
+            expected_recv_ids = [self._get_motor_recv_id(motor) for motor in values.keys()]
+            self._recv_all_responses(expected_recv_ids, timeout=0.002)  # 2ms timeout
+        else:
+            # Fall back to individual writes for other data types
+            for motor, value in values.items():
+                self.write(data_name, motor, value, normalize=normalize, num_retry=num_retry)
+
+    def read_calibration(self) -> dict[str, MotorCalibration]:
+        """Read calibration data from motors."""
+        # Damiao motors don't store calibration internally
+        # Return existing calibration or empty dict
+        return self.calibration if self.calibration else {}
+
+    def write_calibration(self, calibration_dict: dict[str, MotorCalibration], cache: bool = True) -> None:
+        """Write calibration data to motors."""
+        # Damiao motors don't store calibration internally
+        # Just cache it in memory
+        if cache:
+            self.calibration = calibration_dict
+
+    def record_ranges_of_motion(
+        self, motors: NameOrID | list[NameOrID] | None = None, display_values: bool = True
+    ) -> tuple[dict[NameOrID, Value], dict[NameOrID, Value]]:
+        """
+        Interactively record the min/max values of each motor in degrees.
+
+        Move the joints by hand (with torque disabled) while the method streams live positions.
+        Press Enter to finish.
+        """
+        if motors is None:
+            motors = list(self.motors.keys())
+        elif isinstance(motors, (str, int)):
+            motors = [motors]
+
+        # Disable torque for manual movement
+        self.disable_torque(motors)
+        time.sleep(0.1)
+
+        # Get initial positions (already in degrees)
+        start_positions = self.sync_read("Present_Position", motors, normalize=False)
+        mins = start_positions.copy()
+        maxes = start_positions.copy()
+
+        print("\nMove joints through their full range of motion. Press ENTER when done.")
+        user_pressed_enter = False
+        
+        while not user_pressed_enter:
+            positions = self.sync_read("Present_Position", motors, normalize=False)
+            
+            for motor in motors:
+                if motor in positions:
+                    mins[motor] = min(positions[motor], mins.get(motor, positions[motor]))
+                    maxes[motor] = max(positions[motor], maxes.get(motor, positions[motor]))
+
+            if display_values:
+                print("\n" + "=" * 50)
+                print(f"{'MOTOR':<20} | {'MIN (deg)':>12} | {'POS (deg)':>12} | {'MAX (deg)':>12}")
+                print("-" * 50)
+                for motor in motors:
+                    if motor in positions:
+                        print(f"{motor:<20} | {mins[motor]:>12.1f} | {positions[motor]:>12.1f} | {maxes[motor]:>12.1f}")
+
+            if enter_pressed():
+                user_pressed_enter = True
+
+            if display_values and not user_pressed_enter:
+                # Move cursor up to overwrite the previous output
+                move_cursor_up(len(motors) + 4)
+
+            time.sleep(0.05)
+
+        # Re-enable torque
+        self.enable_torque(motors)
+        
+        # Validate ranges
+        for motor in motors:
+            if motor in mins and motor in maxes:
+                if abs(maxes[motor] - mins[motor]) < 5.0:  # At least 5 degrees of range
+                    raise ValueError(f"Motor {motor} has insufficient range of motion (< 5 degrees)")
+
+        return mins, maxes
+
+    def _get_motors_list(self, motors: str | list[str] | None) -> list[str]:
+        """Convert motor specification to list of motor names."""
+        if motors is None:
+            return list(self.motors.keys())
+        elif isinstance(motors, str):
+            return [motors]
+        elif isinstance(motors, list):
+            return motors
+        else:
+            raise TypeError(f"Invalid motors type: {type(motors)}")
+
+    def _get_motor_id(self, motor: NameOrID) -> int:
+        """Get CAN ID for a motor."""
+        if isinstance(motor, str):
+            if motor in self.motors:
+                return self.motors[motor].id
+            else:
+                raise ValueError(f"Unknown motor: {motor}")
+        else:
+            return motor
+
+    def _get_motor_name(self, motor: NameOrID) -> str:
+        """Get motor name from name or ID."""
+        if isinstance(motor, str):
+            return motor
+        else:
+            for name, m in self.motors.items():
+                if m.id == motor:
+                    return name
+            raise ValueError(f"Unknown motor ID: {motor}")
+    
+    def _get_motor_recv_id(self, motor: NameOrID) -> Optional[int]:
+        """Get motor recv_id from name or ID."""
+        motor_name = self._get_motor_name(motor)
+        motor_obj = self.motors.get(motor_name)
+        if motor_obj and hasattr(motor_obj, "recv_id"):
+            return motor_obj.recv_id
+        return None
+
+    @cached_property
+    def is_calibrated(self) -> bool:
+        """Check if motors are calibrated."""
+        return bool(self.calibration)

src/lerobot/motors/damiao/seed_damiao.py

@@ -0,0 +1,833 @@
+## This is a derivative of the following software.
+## https://github.com/cmjang/DM_Control_Python/blob/main/DM_CAN.py
+
+import can
+from time import sleep, time
+import numpy as np
+from enum import IntEnum
+from struct import unpack
+from struct import pack
+
+class Motor:
+    def __init__(self, MotorType, SlaveID, MasterID):
+        """
+        define Motor object 定义电机对象
+        :param MotorType: Motor type 电机类型
+        :param SlaveID: CANID 电机ID
+        :param MasterID: MasterID 主机ID 建议不要设为0
+        """
+        self.Pd = float(0)
+        self.Vd = float(0)
+        self.goal_position = float(0)
+        self.goal_tau = float(0)
+        self.state_q = float(0)
+        self.state_dq = float(0)
+        self.state_tau = float(0)
+        self.state_tmos = int(0)
+        self.state_trotor = int(0)
+        self.SlaveID = SlaveID
+        self.MasterID = MasterID
+        self.MotorType = MotorType
+        self.isEnable = False
+        self.NowControlMode = Control_Type.MIT
+        self.temp_param_dict = {}
+
+    def recv_data(self, q: float, dq: float, tau: float, tmos: int, trotor: int):
+        self.state_q = q
+        self.state_dq = dq
+        self.state_tau = tau
+        self.state_tmos = tmos
+        self.state_trotor = trotor
+
+    def getPosition(self):
+        """
+        get the position of the motor 获取电机位置
+        :return: the position of the motor 电机位置
+        """
+        return self.state_q
+
+    def getVelocity(self):
+        """
+        get the velocity of the motor 获取电机速度
+        :return: the velocity of the motor 电机速度
+        """
+        return self.state_dq
+
+    def getTorque(self):
+        """
+        get the torque of the motor 获取电机力矩
+        :return: the torque of the motor 电机力矩
+        """
+        return self.state_tau
+
+    def getParam(self, RID):
+        """
+        get the parameter of the motor 获取电机内部的参数，需要提前读取
+        :param RID: DM_variable 电机参数
+        :return: the parameter of the motor 电机参数
+        """
+        if RID in self.temp_param_dict:
+            return self.temp_param_dict[RID]
+        else:
+            return None
+
+
+class MotorControl:
+    #send_data_frame = np.array(
+    #    [0x55, 0xAA, 0x1e, 0x03, 0x01, 0x00, 0x00, 0x00, 0x0a, 0x00, 0x00, 0x00, 0x00, 0, 0, 0, 0, 0x00, 0x08, 0x00,
+    #     0x00, 0, 0, 0, 0, 0, 0, 0, 0, 0x00], np.uint8)
+    #                4310           4310_48        4340           4340_48
+    Limit_Param = [[12.5, 30, 10], [12.5, 50, 10], [12.5, 8, 28], [12.5, 10, 28],
+                   # 6006           8006           8009            10010L         10010
+                   [12.5, 45, 20], [12.5, 45, 40], [12.5, 45, 54], [12.5, 25, 200], [12.5, 20, 200],
+                   # H3510            DMG62150      DMH6220
+                   [12.5 , 280 , 1],[12.5 , 45 , 10],[12.5 , 45 , 10]]
+
+    def __init__(self, channel: str, bitrate: int = 1000000):
+        """
+        define MotorControl object 定义电机控制对象
+        :param serial_device: serial object 串口对象
+        """
+        #self.serial_ = serial_device
+        self.motors_map = dict()
+        self.data_save = bytes()  # save data
+        #if self.serial_.is_open:  # open the serial port
+        #    print("Serial port is open")
+        #    serial_device.close()
+        #self.serial_.open()
+        self.canbus = can.interface.Bus(channel=channel, interface='socketcan', bitrate=bitrate)
+
+        #print("can is open")
+        
+
+    def controlMIT(self, DM_Motor, kp: float, kd: float, q: float, dq: float, tau: float):
+        """
+        MIT Control Mode Function 达妙电机MIT控制模式函数
+        :param DM_Motor: Motor object 电机对象
+        :param kp: kp
+        :param kd:  kd
+        :param q:  position  期望位置
+        :param dq:  velocity  期望速度
+        :param tau: torque  期望力矩
+        :return: None
+        """
+        if DM_Motor.SlaveID not in self.motors_map:
+            print("controlMIT ERROR : Motor ID not found")
+            return
+        kp_uint = float_to_uint(kp, 0, 500, 12)
+        kd_uint = float_to_uint(kd, 0, 5, 12)
+        MotorType = DM_Motor.MotorType
+        Q_MAX = self.Limit_Param[MotorType][0]
+        DQ_MAX = self.Limit_Param[MotorType][1]
+        TAU_MAX = self.Limit_Param[MotorType][2]
+        q_uint = float_to_uint(q, -Q_MAX, Q_MAX, 16)
+        dq_uint = float_to_uint(dq, -DQ_MAX, DQ_MAX, 12)
+        tau_uint = float_to_uint(tau, -TAU_MAX, TAU_MAX, 12)
+        data_buf = np.array([0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00], np.uint8)
+        data_buf[0] = (q_uint >> 8) & 0xff
+        data_buf[1] = q_uint & 0xff
+        data_buf[2] = dq_uint >> 4
+        data_buf[3] = ((dq_uint & 0xf) << 4) | ((kp_uint >> 8) & 0xf)
+        data_buf[4] = kp_uint & 0xff
+        data_buf[5] = kd_uint >> 4
+        data_buf[6] = ((kd_uint & 0xf) << 4) | ((tau_uint >> 8) & 0xf)
+        data_buf[7] = tau_uint & 0xff
+        self.__send_data(DM_Motor.SlaveID, data_buf)
+        self.recv()  # receive the data from serial port
+
+    def control_delay(self, DM_Motor, kp: float, kd: float, q: float, dq: float, tau: float, delay: float):
+        """
+        MIT Control Mode Function with delay 达妙电机MIT控制模式函数带延迟
+        :param DM_Motor: Motor object 电机对象
+        :param kp: kp
+        :param kd: kd
+        :param q:  position  期望位置
+        :param dq:  velocity  期望速度
+        :param tau: torque  期望力矩
+        :param delay: delay time 延迟时间 单位秒
+        """
+        self.controlMIT(DM_Motor, kp, kd, q, dq, tau)
+        sleep(delay)
+
+    def control_Pos_Vel(self, Motor, P_desired: float, V_desired: float):
+        """
+        control the motor in position and velocity control mode 电机位置速度控制模式
+        :param Motor: Motor object 电机对象
+        :param P_desired: desired position 期望位置
+        :param V_desired: desired velocity 期望速度
+        :return: None
+        """
+        if Motor.SlaveID not in self.motors_map:
+            print("Control Pos_Vel Error : Motor ID not found")
+            return
+        motorid = 0x100 + Motor.SlaveID
+        data_buf = np.array([0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00], np.uint8)
+        P_desired_uint8s = float_to_uint8s(P_desired)
+        V_desired_uint8s = float_to_uint8s(V_desired)
+        data_buf[0:4] = P_desired_uint8s
+        data_buf[4:8] = V_desired_uint8s
+        self.__send_data(motorid, data_buf)
+        self.recv()  # receive the data from serial port
+
+    def control_Vel(self, Motor, Vel_desired):
+        """
+        control the motor in velocity control mode 电机速度控制模式
+        :param Motor: Motor object 电机对象
+        :param Vel_desired: desired velocity 期望速度
+        """
+        if Motor.SlaveID not in self.motors_map:
+            print("control_VEL ERROR : Motor ID not found")
+            return
+        motorid = 0x200 + Motor.SlaveID
+        data_buf = np.array([0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00], np.uint8)
+        Vel_desired_uint8s = float_to_uint8s(Vel_desired)
+        data_buf[0:4] = Vel_desired_uint8s
+        self.__send_data(motorid, data_buf)
+        self.recv()  # receive the data from serial port
+
+    def control_pos_force(self, Motor, Pos_des: float, Vel_des, i_des):
+        """
+        control the motor in EMIT control mode 电机力位混合模式
+        :param Pos_des: desired position rad  期望位置 单位为rad
+        :param Vel_des: desired velocity rad/s  期望速度 为放大100倍
+        :param i_des: desired current rang 0-10000 期望电流标幺值放大10000倍
+        电流标幺值：实际电流值除以最大电流值，最大电流见上电打印
+        """
+        if Motor.SlaveID not in self.motors_map:
+            print("control_pos_vel ERROR : Motor ID not found")
+            return
+        motorid = 0x300 + Motor.SlaveID
+        data_buf = np.array([0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00], np.uint8)
+        Pos_desired_uint8s = float_to_uint8s(Pos_des)
+        data_buf[0:4] = Pos_desired_uint8s
+        Vel_uint = np.uint16(Vel_des)
+        ides_uint = np.uint16(i_des)
+        data_buf[4] = Vel_uint & 0xff
+        data_buf[5] = Vel_uint >> 8
+        data_buf[6] = ides_uint & 0xff
+        data_buf[7] = ides_uint >> 8
+        self.__send_data(motorid, data_buf)
+        self.recv()  # receive the data from serial port
+
+    def enable(self, Motor):
+        """
+        enable motor 使能电机
+        最好在上电后几秒后再使能电机
+        :param Motor: Motor object 电机对象
+        """
+        self.__control_cmd(Motor, np.uint8(0xFC))
+        sleep(0.1)
+        self.recv()  # receive the data from serial port
+
+    def enable_old(self, Motor ,ControlMode):
+        """
+        enable motor old firmware 使能电机旧版本固件，这个是为了旧版本电机固件的兼容性
+        可恶的旧版本固件使能需要加上偏移量
+        最好在上电后几秒后再使能电机
+        :param Motor: Motor object 电机对象
+        """
+        data_buf = np.array([0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xfc], np.uint8)
+        enable_id = ((int(ControlMode)-1) << 2) + Motor.SlaveID
+        self.__send_data(enable_id, data_buf)
+        sleep(0.1)
+        self.recv()  # receive the data from serial port
+
+    def disable(self, Motor):
+        """
+        disable motor 失能电机
+        :param Motor: Motor object 电机对象
+        """
+        self.__control_cmd(Motor, np.uint8(0xFD))
+        sleep(0.1)
+        self.recv()  # receive the data from serial port
+
+    def set_zero_position(self, Motor):
+        """
+        set the zero position of the motor 设置电机0位
+        :param Motor: Motor object 电机对象
+        """
+        self.__control_cmd(Motor, np.uint8(0xFE))
+        sleep(0.1)
+        self.recv()  # receive the data from serial port
+
+    def recv(self):
+        # 把上次没有解析完的剩下的也放进来
+        # data_recv = b''.join([self.data_save, self.serial_.read_all()])
+        #data_recv = b''.join([self.data_save, self.canbus.recv()])
+
+
+        # packets = self.__extract_packets(data_recv)
+        # for packet in packets:
+        #     data = packet[7:15]
+        #     CANID = (packet[6] << 24) | (packet[5] << 16) | (packet[4] << 8) | packet[3]
+        #     CMD = packet[1]
+        #     self.__process_packet(data, CANID, CMD)
+        
+        data_recv = self.canbus.recv(0.1)
+
+        if data_recv is not None:
+            # data = data_recv.data
+            # err = data[0] >> 12
+            # id = data[0] & 0x7f
+            # pos = (data[1] << 8) + data[2]
+            # vel = (data[3] << 4) + (data[4] >> 4)
+            # tau = ((data[4] & 0x0f) << 8) + data[5]
+            # t_mos = data[6]
+            # t_rotor = data[7]
+            # print(hex(id), err, id, pos, vel, tau, goal_tau, t_mos, t_rotor)
+            # CANID = data_recv.arbitration_id
+            CANID = data_recv.data[0]
+            # CMD = data_recv.data[3]
+            CMD = 0x11                  # 飯田：修正の必要あり
+            self.__process_packet(data_recv.data, CANID, CMD)
+
+            # 飯田：Debug print
+            # print(hex(CANID),hex(CMD))
+            # print(hex(data_recv.data[0]),hex(data_recv.data[1]),hex(data_recv.data[2]),hex(data_recv.data[3]),hex(data_recv.data[4]),hex(data_recv.data[5]),hex(data_recv.data[6]),hex(data_recv.data[7]))
+            #return data
+
+    def recv_set_param_data(self):
+        #data_recv = self.serial_.read_all()
+
+        # packets = self.__extract_packets(data_recv)
+        # for packet in packets:
+        #     data = packet[7:15]
+        #     CANID = (packet[6] << 24) | (packet[5] << 16) | (packet[4] << 8) | packet[3]
+        #     CMD = packet[1]
+        #     self.__process_set_param_packet(data, CANID, CMD)
+        
+        data_recv = self.canbus.recv(0.1)
+
+
+        if data_recv is not None:
+            data = data_recv.data
+            CANID = data_recv.arbitration_id
+            # CANID = data_recv.data[0]
+            # CMD = data_recv.data[3]  
+            CMD = 0x11                  # 飯田：修正の必要あり
+            self.__process_packet(data, CANID, CMD)
+
+
+            # 飯田：Debug print
+            print(hex(CANID),hex(CMD))
+            print(hex(data_recv.data[0]),hex(data_recv.data[1]),hex(data_recv.data[2]),hex(data_recv.data[3]),hex(data_recv.data[4]),hex(data_recv.data[5]),hex(data_recv.data[6]),hex(data_recv.data[7]))
+
+    def __process_packet(self, data, CANID, CMD):
+        if CMD == 0x11:
+            if CANID != 0x00:
+                if CANID in self.motors_map:
+                    q_uint = np.uint16((np.uint16(data[1]) << 8) | data[2])
+                    dq_uint = np.uint16((np.uint16(data[3]) << 4) | (data[4] >> 4))
+                    tau_uint = np.uint16(((data[4] & 0xf) << 8) | data[5])
+                    t_mos = data[6]
+                    t_rotor = data[7]
+                    MotorType_recv = self.motors_map[CANID].MotorType
+                    Q_MAX = self.Limit_Param[MotorType_recv][0]
+                    DQ_MAX = self.Limit_Param[MotorType_recv][1]
+                    TAU_MAX = self.Limit_Param[MotorType_recv][2]
+                    recv_q = uint_to_float(q_uint, -Q_MAX, Q_MAX, 16)
+                    recv_dq = uint_to_float(dq_uint, -DQ_MAX, DQ_MAX, 12)
+                    recv_tau = uint_to_float(tau_uint, -TAU_MAX, TAU_MAX, 12)
+                    self.motors_map[CANID].recv_data(recv_q, recv_dq, recv_tau, t_mos, t_rotor)
+            else:
+                MasterID=data[0] & 0x0f
+                if MasterID in self.motors_map:
+                    q_uint = np.uint16((np.uint16(data[1]) << 8) | data[2])
+                    dq_uint = np.uint16((np.uint16(data[3]) << 4) | (data[4] >> 4))
+                    tau_uint = np.uint16(((data[4] & 0xf) << 8) | data[5])
+                    t_mos = data[6]
+                    t_rotor = data[7]
+                    MotorType_recv = self.motors_map[MasterID].MotorType
+                    Q_MAX = self.Limit_Param[MotorType_recv][0]
+                    DQ_MAX = self.Limit_Param[MotorType_recv][1]
+                    TAU_MAX = self.Limit_Param[MotorType_recv][2]
+                    recv_q = uint_to_float(q_uint, -Q_MAX, Q_MAX, 16)
+                    recv_dq = uint_to_float(dq_uint, -DQ_MAX, DQ_MAX, 12)
+                    recv_tau = uint_to_float(tau_uint, -TAU_MAX, TAU_MAX, 12)
+                    self.motors_map[MasterID].recv_data(recv_q, recv_dq, recv_tau, t_mos, t_rotor)
+
+
+    def __process_set_param_packet(self, data, CANID, CMD):
+        if CMD == 0x11 and (data[2] == 0x33 or data[2] == 0x55):
+            masterid=CANID
+            slaveId = ((data[1] << 8) | data[0])
+            if CANID==0x00:  #防止有人把MasterID设为0稳一手
+                masterid=slaveId
+
+            if masterid not in self.motors_map:
+                if slaveId not in self.motors_map:
+                    return
+                else:
+                    masterid=slaveId
+
+            RID = data[3]
+            # 读取参数得到的数据
+            if is_in_ranges(RID):
+                #uint32类型
+                num = uint8s_to_uint32(data[4], data[5], data[6], data[7])
+                self.motors_map[masterid].temp_param_dict[RID] = num
+
+            else:
+                #float类型
+                num = uint8s_to_float(data[4], data[5], data[6], data[7])
+                self.motors_map[masterid].temp_param_dict[RID] = num
+
+
+    def addMotor(self, Motor):
+        """
+        add motor to the motor control object 添加电机到电机控制对象
+        :param Motor: Motor object 电机对象
+        """
+        self.motors_map[Motor.SlaveID] = Motor
+        if Motor.MasterID != 0:
+            self.motors_map[Motor.MasterID] = Motor
+        return True
+
+    def __control_cmd(self, Motor, cmd: np.uint8):     # 飯田：コマンドは通ります
+        data_buf = np.array([0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, cmd], np.uint8)
+        self.__send_data(Motor.SlaveID, data_buf)
+
+    def __send_data(self, motor_id, data):
+        """
+        send data to the motor 发送数据到电机
+        :param motor_id:
+        :param data:
+        :return:
+        """
+        #self.send_data_frame[13] = motor_id & 0xff
+        #self.send_data_frame[14] = (motor_id >> 8)& 0xff  #id high 8 bits
+        #self.send_data_frame[21:29] = data
+        #self.serial_.write(bytes(self.send_data_frame.T))
+
+        msg =can.Message(is_extended_id=False,arbitration_id=motor_id,data=data,is_remote_frame = False)
+        self.canbus.send(msg)
+
+    def __read_RID_param(self, Motor, RID):             # 飯田：修正の必要あり?
+        can_id_l = Motor.SlaveID & 0xff #id low 8 bits
+        can_id_h = (Motor.SlaveID >> 8)& 0xff  #id high 8 bits
+        data_buf = np.array([np.uint8(can_id_l), np.uint8(can_id_h), 0x33, np.uint8(RID), 0x00, 0x00, 0x00, 0x00], np.uint8)
+        self.__send_data(0x7FF, data_buf)
+        
+
+
+    def __write_motor_param(self, Motor, RID, data):             # 飯田：修正の必要あり?
+        can_id_l = Motor.SlaveID & 0xff #id low 8 bits
+        can_id_h = (Motor.SlaveID >> 8)& 0xff  #id high 8 bits
+        data_buf = np.array([np.uint8(can_id_l), np.uint8(can_id_h), 0x55, np.uint8(RID), 0x00, 0x00, 0x00, 0x00], np.uint8)
+        if not is_in_ranges(RID):
+            # data is float
+            data_buf[4:8] = float_to_uint8s(data)
+        else:
+            # data is int
+            data_buf[4:8] = data_to_uint8s(int(data))
+        self.__send_data(0x7FF, data_buf)
+
+    def switchControlMode(self, Motor, ControlMode):
+        """
+        switch the control mode of the motor 切换电机控制模式
+        :param Motor: Motor object 电机对象
+        :param ControlMode: Control_Type 电机控制模式 example:MIT:Control_Type.MIT MIT模式
+        """
+        max_retries = 20
+        retry_interval = 0.1  #retry times
+        RID = 10
+        self.__write_motor_param(Motor, RID, np.uint8(ControlMode))
+        for _ in range(max_retries):
+            sleep(retry_interval)
+            self.recv_set_param_data()
+            if Motor.SlaveID in self.motors_map:
+                if RID in self.motors_map[Motor.SlaveID].temp_param_dict:
+                    if abs(self.motors_map[Motor.SlaveID].temp_param_dict[RID] - ControlMode) < 0.1:
+                        return True
+                    else:
+                        return False
+        return False
+
+    def save_motor_param(self, Motor):
+        """
+        save the all parameter  to flash 保存所有电机参数
+        :param Motor: Motor object 电机对象
+        :return:
+        """
+        can_id_l = Motor.SlaveID & 0xff #id low 8 bits
+        can_id_h = (Motor.SlaveID >> 8)& 0xff  #id high 8 bits
+        data_buf = np.array([np.uint8(can_id_l), np.uint8(can_id_h), 0xAA, 0x00, 0x00, 0x00, 0x00, 0x00], np.uint8)
+        self.disable(Motor)  # before save disable the motor
+        self.__send_data(0x7FF, data_buf)
+        sleep(0.001)
+
+    def change_limit_param(self, Motor_Type, PMAX, VMAX, TMAX):
+        """
+        change the PMAX VMAX TMAX of the motor 改变电机的PMAX VMAX TMAX
+        :param Motor_Type:
+        :param PMAX: 电机的PMAX
+        :param VMAX: 电机的VMAX
+        :param TMAX: 电机的TMAX
+        :return:
+        """
+        self.Limit_Param[Motor_Type][0] = PMAX
+        self.Limit_Param[Motor_Type][1] = VMAX
+        self.Limit_Param[Motor_Type][2] = TMAX
+
+    def refresh_motor_status(self,Motor):
+        """
+        get the motor status 获得电机状态
+        """
+        can_id_l = Motor.SlaveID & 0xff #id low 8 bits
+        can_id_h = (Motor.SlaveID >> 8) & 0xff  #id high 8 bits
+        data_buf = np.array([np.uint8(can_id_l), np.uint8(can_id_h), 0xCC, 0x00, 0x00, 0x00, 0x00, 0x00], np.uint8)
+        self.__send_data(0x7FF, data_buf)
+        self.recv()  # receive the data from serial port
+
+    def change_motor_param(self, Motor, RID, data):
+        """
+        change the RID of the motor 改变电机的参数
+        :param Motor: Motor object 电机对象
+        :param RID: DM_variable 电机参数
+        :param data: 电机参数的值
+        :return: True or False ,True means success, False means fail
+        """
+        max_retries = 20
+        retry_interval = 0.05  #retry times
+
+        self.__write_motor_param(Motor, RID, data)
+        for _ in range(max_retries):
+            self.recv_set_param_data()
+            if Motor.SlaveID in self.motors_map and RID in self.motors_map[Motor.SlaveID].temp_param_dict:
+                if abs(self.motors_map[Motor.SlaveID].temp_param_dict[RID] - data) < 0.1:
+                    return True
+                else:
+                    return False
+            sleep(retry_interval)
+        return False
+
+    def read_motor_param(self, Motor, RID):
+        """
+        read only the RID of the motor 读取电机的内部信息例如 版本号等
+        :param Motor: Motor object 电机对象
+        :param RID: DM_variable 电机参数
+        :return: 电机参数的值
+        """
+        max_retries = 5
+        retry_interval = 0.05  #retry times
+        self.__read_RID_param(Motor, RID)
+        for _ in range(max_retries):
+            sleep(retry_interval)
+            self.recv_set_param_data()
+            if Motor.SlaveID in self.motors_map:
+                if RID in self.motors_map[Motor.SlaveID].temp_param_dict:
+                    return self.motors_map[Motor.SlaveID].temp_param_dict[RID]
+        return None
+
+    # -------------------------------------------------
+    # Extract packets from the serial data
+    def __extract_packets(self, data):          
+        frames = []
+        header = 0xAA
+        tail = 0x55
+        frame_length = 16
+        i = 0
+        remainder_pos = 0
+
+        while i <= len(data) - frame_length:
+            if data[i] == header and data[i + frame_length - 1] == tail:
+                frame = data[i:i + frame_length]
+                frames.append(frame)
+                i += frame_length
+                remainder_pos = i
+            else:
+                i += 1
+        self.data_save = data[remainder_pos:]
+        return frames
+
+
+def LIMIT_MIN_MAX(x, min, max):
+    if x <= min:
+        x = min
+    elif x > max:
+        x = max
+
+
+def float_to_uint(x: float, x_min: float, x_max: float, bits):
+    LIMIT_MIN_MAX(x, x_min, x_max)
+    span = x_max - x_min
+    data_norm = (x - x_min) / span
+    return np.uint16(data_norm * ((1 << bits) - 1))
+
+
+def uint_to_float(x: np.uint16, min: float, max: float, bits):
+    span = max - min
+    data_norm = float(x) / ((1 << bits) - 1)
+    temp = data_norm * span + min
+    return np.float32(temp)
+
+
+def float_to_uint8s(value):
+    # Pack the float into 4 bytes
+    packed = pack('f', value)
+    # Unpack the bytes into four uint8 values
+    return unpack('4B', packed)
+
+
+def data_to_uint8s(value):
+    # Check if the value is within the range of uint32
+    if isinstance(value, int) and (0 <= value <= 0xFFFFFFFF):
+        # Pack the uint32 into 4 bytes
+        packed = pack('I', value)
+    else:
+        raise ValueError("Value must be an integer within the range of uint32")
+
+    # Unpack the bytes into four uint8 values
+    return unpack('4B', packed)
+
+
+def is_in_ranges(number):
+    """
+    check if the number is in the range of uint32
+    :param number:
+    :return:
+    """
+    if (7 <= number <= 10) or (13 <= number <= 16) or (35 <= number <= 36):
+        return True
+    return False
+
+
+def uint8s_to_uint32(byte1, byte2, byte3, byte4):
+    # Pack the four uint8 values into a single uint32 value in little-endian order
+    packed = pack('<4B', byte1, byte2, byte3, byte4)
+    # Unpack the packed bytes into a uint32 value
+    return unpack('<I', packed)[0]
+
+
+def uint8s_to_float(byte1, byte2, byte3, byte4):
+    # Pack the four uint8 values into a single float value in little-endian order
+    packed = pack('<4B', byte1, byte2, byte3, byte4)
+    # Unpack the packed bytes into a float value
+    return unpack('<f', packed)[0]
+
+
+def print_hex(data):
+    hex_values = [f'{byte:02X}' for byte in data]
+    print(' '.join(hex_values))
+
+
+def get_enum_by_index(index, enum_class):
+    try:
+        return enum_class(index)
+    except ValueError:
+        return None
+
+
+class DM_Motor_Type(IntEnum):
+    DM4310 = 0
+    DM4310_48V = 1
+    DM4340 = 2
+    DM4340_48V = 3
+    DM6006 = 4
+    DM8006 = 5
+    DM8009 = 6
+    DM10010L = 7
+    DM10010 = 8
+    DMH3510 = 9
+    DMH6215 = 10
+    DMG6220 = 11
+
+
+class DM_variable(IntEnum):
+    UV_Value = 0
+    KT_Value = 1
+    OT_Value = 2
+    OC_Value = 3
+    ACC = 4
+    DEC = 5
+    MAX_SPD = 6
+    MST_ID = 7
+    ESC_ID = 8
+    TIMEOUT = 9
+    CTRL_MODE = 10
+    Damp = 11
+    Inertia = 12
+    hw_ver = 13
+    sw_ver = 14
+    SN = 15
+    NPP = 16
+    Rs = 17
+    LS = 18
+    Flux = 19
+    Gr = 20
+    PMAX = 21
+    VMAX = 22
+    TMAX = 23
+    I_BW = 24
+    KP_ASR = 25
+    KI_ASR = 26
+    KP_APR = 27
+    KI_APR = 28
+    OV_Value = 29
+    GREF = 30
+    Deta = 31
+    V_BW = 32
+    IQ_c1 = 33
+    VL_c1 = 34
+    can_br = 35
+    sub_ver = 36
+    u_off = 50
+    v_off = 51
+    k1 = 52
+    k2 = 53
+    m_off = 54
+    dir = 55
+    p_m = 80
+    xout = 81
+
+
+class Control_Type(IntEnum):
+    MIT = 1
+    POS_VEL = 2
+    VEL = 3
+    Torque_Pos = 4
+
+class DamiaoPort:
+    def __init__(self, device, types, can_ids, master_ids, motor_with_torque, control_mode=Control_Type.MIT):
+        self.device = device
+        self.types = types
+        self.can_ids = can_ids
+        self.master_ids = master_ids
+        self.control = MotorControl(self.device, bitrate=4000000)
+        self.motors = [Motor(type, can_id, master_id) for type, can_id, master_id in zip(types, can_ids, master_ids)]
+        self.stat_data = []
+        self.stat_time = []
+        for motor in self.motors:
+            self.control.addMotor(motor)
+            self.control.enable(motor)
+
+    def get_present_status(self):
+        self.stat_time.append(time())
+        stat = [[
+            motor.goal_position,
+            motor.goal_tau,
+            motor.getPosition(),
+            motor.getVelocity(),
+            motor.getTorque(),
+            motor.state_tmos,
+            motor.state_trotor,
+        ] for motor in self.motors]
+        self.stat_data.append(stat)
+
+        return stat
+
+    def save_status(self, filename):
+        np.savez(filename, np.array(self.stat_time), np.array(self.stat_data))
+
+    def disable(self):
+        for motor in self.motors:
+            self.control.disable(motor)
+
+    def shutdown(self):
+        for motor in self.motors:
+            self.control.controlMIT(motor, 0, 0, 0, 0, 0)
+        self.control.canbus.shutdown()
+
+    def set_zero_position(self):
+        for motor in self.motors:
+            self.control.disable(motor)
+        sleep(1)
+        for motor in self.motors:
+            self.control.set_zero_position(motor)
+        sleep(1)
+        for motor in self.motors:
+            self.control.enable(motor)
+        return 0
+
+    async def move_towards(self, goal_positions, kps, kds):
+        for motor, goal_position, kp, kd in zip(self.motors, goal_positions, kps, kds):
+            delta = goal_position - motor.getPosition()
+            v = motor.getVelocity()
+            tau = kp * delta - kd * v
+            motor.goal_position = goal_position
+            motor.goal_tau = tau
+            self.control.controlMIT(motor, 0, 0, 0, 0, tau)
+            await asyncio.sleep(0.00003)
+
+    def move_regressor_sync(self, regs, search_range, search_step, goal_positions, kps, kds):
+        TORQUE_SCALER=30
+        if len(self.stat_data) == 0:
+            return self.move_towards_sync(goal_positions, kps, kds)
+        for motor, reg, goal_position, kp, kd, stat in zip(
+                self.motors, regs, goal_positions, kps, kds, self.stat_data[-1]):
+            pos = motor.getPosition()
+            vel = motor.getVelocity()
+            delta = goal_position - pos
+            goal_tau = kp * delta - kd * vel
+            _goal_pos, _goal_tau, _pos, _vel, _tau = stat
+            x = np.array([[_pos, _vel, _tau, _goal_pos, _goal_tau],
+                          [pos, vel, motor.getTorque(), goal_position, goal_tau]])
+            x /= np.array([[np.pi, 10, TORQUE_SCALER, np.pi, TORQUE_SCALER]])
+            xs = []
+            for tau in np.linspace(goal_tau/TORQUE_SCALER - search_range,
+                                   goal_tau/TORQUE_SCALER + search_range,
+                                   num=search_step):
+                x_ = x.copy()
+                x_[0,4] = tau
+                xs.append(x_.flatten())
+            h = reg.predict(xs)
+            diff = h - goal_position
+            tau = TORQUE_SCALER * xs[np.argmin(diff ** 2)][4]
+            goal_tau = tau
+            motor.goal_position = goal_position
+            motor.goal_tau = goal_tau
+            self.control.controlMIT(motor, 0, 0, 0, 0, goal_tau)
+            sleep(0.00003)
+
+    def move_towards_sync(self, goal_positions, kps, kds):
+        for motor, goal_position, kp, kd in zip(self.motors, goal_positions, kps, kds):
+            delta = goal_position - motor.getPosition()
+            v = motor.getVelocity()
+            tau = kp * delta - kd * v
+            motor.goal_position = goal_position
+            motor.goal_tau = tau
+            self.control.controlMIT(motor, 0, 0, 0, 0, tau)
+            # sleep(0.00003)
+
+    def set_goal_torque_sync(self, goal_taus):
+        for motor, goal_tau in zip(self.motors, goal_taus):
+            motor.goal_position = 0
+            motor.goal_tau = goal_tau
+            self.control.controlMIT(motor, 0, 0, 0, 0, motor.goal_tau)
+            sleep(0.00003)
+
+    def move_torque_sync(self, taus):
+        for motor,tau in zip(self.motors, taus):
+            motor.goal_position = 0
+            motor.goal_tau = tau
+            self.control.controlMIT(motor, 0, 0, 0, 0, motor.goal_tau)
+            sleep(0.00003)
+
+    def keep_torque_sync(self):
+        for motor in self.motors:
+            self.control.controlMIT(motor, 0, 0, 0, 0, motor.goal_tau)
+            sleep(0.00003)
+
+    async def set_goal_positions(self, goal_positions, kps):
+        for motor, goal_position, kp in zip(self.motors, goal_positions, kps):
+            motor.goal_position = goal_position
+            motor.goal_tau = 0
+            self.control.controlMIT(motor, kp, 1.2, goal_position, 0, 0)
+            await asyncio.sleep(0.00003)
+
+    def set_goal_positions_sync(self, goal_positions, kps, kds):
+        for motor, goal_position, kp, kd in zip(self.motors, goal_positions, kps, kds):
+            motor.goal_position = goal_position
+            motor.goal_tau = 0
+            self.control.controlMIT(motor, kp, kd, goal_position, 0, 0)
+            sleep(0.00003)
+
+    def set_goal_posvel(self, goal_positions):
+        for motor, goal_position in zip(self.motors, goal_positions):
+            motor.goal_position = goal_position
+            motor.goal_tau = 0
+            self.control.control_pos_force(motor, goal_position, 1, 1)
+
+    def controlMIT(self, motor, kp, kd, q, dq, tau):
+        self.control.controlMIT(self.motors[motor], kp, kd, q, dq, tau)
+

src/lerobot/motors/damiao/tables.py

@@ -0,0 +1,209 @@
+# 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.
+
+"""Configuration tables for Damiao motors."""
+
+from enum import IntEnum
+from typing import Dict, List, Tuple
+
+# Motor type definitions
+class MotorType(IntEnum):
+    DM3507 = 0
+    DM4310 = 1
+    DM4310_48V = 2
+    DM4340 = 3
+    DM4340_48V = 4
+    DM6006 = 5
+    DM8006 = 6
+    DM8009 = 7
+    DM10010L = 8
+    DM10010 = 9
+    DMH3510 = 10
+    DMH6215 = 11
+    DMG6220 = 12
+
+# Control modes
+class ControlMode(IntEnum):
+    MIT = 1
+    POS_VEL = 2
+    VEL = 3
+    TORQUE_POS = 4
+
+# Motor variable IDs (RID)
+class MotorVariable(IntEnum):
+    UV_VALUE = 0
+    KT_VALUE = 1
+    OT_VALUE = 2
+    OC_VALUE = 3
+    ACC = 4
+    DEC = 5
+    MAX_SPD = 6
+    MST_ID = 7
+    ESC_ID = 8
+    TIMEOUT = 9
+    CTRL_MODE = 10
+    DAMP = 11
+    INERTIA = 12
+    HW_VER = 13
+    SW_VER = 14
+    SN = 15
+    NPP = 16
+    RS = 17
+    LS = 18
+    FLUX = 19
+    GR = 20
+    PMAX = 21
+    VMAX = 22
+    TMAX = 23
+    I_BW = 24
+    KP_ASR = 25
+    KI_ASR = 26
+    KP_APR = 27
+    KI_APR = 28
+    OV_VALUE = 29
+    GREF = 30
+    DETA = 31
+    V_BW = 32
+    IQ_C1 = 33
+    VL_C1 = 34
+    CAN_BR = 35
+    SUB_VER = 36
+    U_OFF = 50
+    V_OFF = 51
+    K1 = 52
+    K2 = 53
+    M_OFF = 54
+    DIR = 55
+    P_M = 80
+    XOUT = 81
+
+# Motor limit parameters [PMAX, VMAX, TMAX] 
+# PMAX: Maximum position (rad)
+# VMAX: Maximum velocity (rad/s)
+# TMAX: Maximum torque (N·m)
+MOTOR_LIMIT_PARAMS = {
+    MotorType.DM3507: (12.5, 30, 10),
+    MotorType.DM4310: (12.5, 30, 10),
+    MotorType.DM4310_48V: (12.5, 50, 10),
+    MotorType.DM4340: (12.5, 8, 28),
+    MotorType.DM4340_48V: (12.5, 10, 28),
+    MotorType.DM6006: (12.5, 45, 20),
+    MotorType.DM8006: (12.5, 45, 40),
+    MotorType.DM8009: (12.5, 45, 54),
+    MotorType.DM10010L: (12.5, 25, 200),
+    MotorType.DM10010: (12.5, 20, 200),
+    MotorType.DMH3510: (12.5, 280, 1),
+    MotorType.DMH6215: (12.5, 45, 10),
+    MotorType.DMG6220: (12.5, 45, 10),
+}
+
+# Motor model names
+MODEL_NAMES = {
+    MotorType.DM3507: "dm3507",
+    MotorType.DM4310: "dm4310",
+    MotorType.DM4310_48V: "dm4310_48v",
+    MotorType.DM4340: "dm4340",
+    MotorType.DM4340_48V: "dm4340_48v",
+    MotorType.DM6006: "dm6006",
+    MotorType.DM8006: "dm8006",
+    MotorType.DM8009: "dm8009",
+    MotorType.DM10010L: "dm10010l",
+    MotorType.DM10010: "dm10010",
+    MotorType.DMH3510: "dmh3510",
+    MotorType.DMH6215: "dmh6215",
+    MotorType.DMG6220: "dmg6220",
+}
+
+# Motor resolution table (encoder counts per revolution)
+MODEL_RESOLUTION = {
+    "dm3507": 65536,
+    "dm4310": 65536,
+    "dm4310_48v": 65536,
+    "dm4340": 65536,
+    "dm4340_48v": 65536,
+    "dm6006": 65536,
+    "dm8006": 65536,
+    "dm8009": 65536,
+    "dm10010l": 65536,
+    "dm10010": 65536,
+    "dmh3510": 65536,
+    "dmh6215": 65536,
+    "dmg6220": 65536,
+}
+
+# CAN baudrates supported by Damiao motors
+AVAILABLE_BAUDRATES = [
+    125000,   # 0: 125 kbps
+    200000,   # 1: 200 kbps
+    250000,   # 2: 250 kbps
+    500000,   # 3: 500 kbps
+    1000000,  # 4: 1 mbps (default for OpenArms)
+    2000000,  # 5: 2 mbps
+    2500000,  # 6: 2.5 mbps
+    3200000,  # 7: 3.2 mbps
+    4000000,  # 8: 4 mbps
+    5000000,  # 9: 5 mbps
+]
+DEFAULT_BAUDRATE = 1000000  # 1 Mbps is standard for OpenArms
+
+# Default timeout in milliseconds
+DEFAULT_TIMEOUT_MS = 1000
+
+# Data that should be normalized
+NORMALIZED_DATA = ["Present_Position", "Goal_Position"]
+
+# OpenArms specific configurations
+# Based on: https://docs.openarm.dev/software/setup/configure-test
+# OpenArms has 7 DOF per arm (14 total for dual arm)
+OPENARMS_ARM_MOTOR_IDS = {
+    "joint_1": {"send": 0x01, "recv": 0x11},  # J1 - Shoulder pan
+    "joint_2": {"send": 0x02, "recv": 0x12},  # J2 - Shoulder lift
+    "joint_3": {"send": 0x03, "recv": 0x13},  # J3 - Elbow flex
+    "joint_4": {"send": 0x04, "recv": 0x14},  # J4 - Wrist flex
+    "joint_5": {"send": 0x05, "recv": 0x15},  # J5 - Wrist roll
+    "joint_6": {"send": 0x06, "recv": 0x16},  # J6 - Wrist pitch
+    "joint_7": {"send": 0x07, "recv": 0x17},  # J7 - Wrist rotation
+}
+
+OPENARMS_GRIPPER_MOTOR_IDS = {
+    "gripper": {"send": 0x08, "recv": 0x18},  # J8 - Gripper
+}
+
+# Default motor types for OpenArms
+OPENARMS_DEFAULT_MOTOR_TYPES = {
+    "joint_1": MotorType.DM8009,   # Shoulder pan - high torque
+    "joint_2": MotorType.DM8009,   # Shoulder lift - high torque
+    "joint_3": MotorType.DM4340,   # Shoulder rotation
+    "joint_4": MotorType.DM4340,   # Elbow flex
+    "joint_5": MotorType.DM4310,   # Wrist roll
+    "joint_6": MotorType.DM4310,   # Wrist pitch
+    "joint_7": MotorType.DM4310,   # Wrist rotation
+    "gripper": MotorType.DM4310,   # Gripper
+}
+
+# MIT control parameter ranges
+MIT_KP_RANGE = (0.0, 500.0)
+MIT_KD_RANGE = (0.0, 5.0)
+
+# CAN frame command IDs
+CAN_CMD_ENABLE = 0xFC
+CAN_CMD_DISABLE = 0xFD
+CAN_CMD_SET_ZERO = 0xFE
+CAN_CMD_REFRESH = 0xCC
+CAN_CMD_QUERY_PARAM = 0x33
+CAN_CMD_WRITE_PARAM = 0x55
+CAN_CMD_SAVE_PARAM = 0xAA
+
+# CAN ID for parameter operations
+CAN_PARAM_ID = 0x7FF

src/lerobot/motors/dynamixel/dynamixel.py

@@ -24,7 +24,7 @@ from enum import Enum
 
 from lerobot.motors.encoding_utils import decode_twos_complement, encode_twos_complement
 
-from ..motors_bus import Motor, MotorCalibration, MotorsBus, NameOrID, Value, get_address
+from ..motors_bus import Motor, MotorCalibration, NameOrID, SerialMotorsBus, Value, get_address
 from .tables import (
     AVAILABLE_BAUDRATES,
     MODEL_BAUDRATE_TABLE,
@@ -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
 
@@ -100,7 +100,7 @@ def _split_into_byte_chunks(value: int, length: int) -> list[int]:
     return data
 
 
-class DynamixelMotorsBus(MotorsBus):
+class DynamixelMotorsBus(SerialMotorsBus):
     """
     The Dynamixel implementation for a MotorsBus. It relies on the python dynamixel sdk to communicate with
     the motors. For more info, see the Dynamixel SDK Documentation:

src/lerobot/motors/feetech/feetech.py

@@ -19,7 +19,7 @@ from pprint import pformat
 
 from lerobot.motors.encoding_utils import decode_sign_magnitude, encode_sign_magnitude
 
-from ..motors_bus import Motor, MotorCalibration, MotorsBus, NameOrID, Value, get_address
+from ..motors_bus import Motor, MotorCalibration, NameOrID, SerialMotorsBus, Value, get_address
 from .tables import (
     FIRMWARE_MAJOR_VERSION,
     FIRMWARE_MINOR_VERSION,
@@ -96,7 +96,7 @@ def patch_setPacketTimeout(self, packet_length):  # noqa: N802
     self.packet_timeout = (self.tx_time_per_byte * packet_length) + (self.tx_time_per_byte * 3.0) + 50
 
 
-class FeetechMotorsBus(MotorsBus):
+class FeetechMotorsBus(SerialMotorsBus):
     """
     The FeetechMotorsBus class allows to efficiently read and write to the attached motors. It relies on the
     python feetech sdk to communicate with the motors, which is itself based on the dynamixel sdk.

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/motors/motors_bus.py

@@ -19,6 +19,8 @@
 # TODO(aliberts): Add block noqa when feature below is available
 # https://github.com/astral-sh/ruff/issues/3711
 
+from __future__ import annotations
+
 import abc
 import logging
 from contextlib import contextmanager
@@ -41,6 +43,92 @@ Value: TypeAlias = int | float
 logger = logging.getLogger(__name__)
 
 
+class MotorsBusBase(abc.ABC):
+    """
+    Base class for all motor bus implementations.
+    
+    This is a minimal interface that all motor buses must implement, regardless of their
+    communication protocol (serial, CAN, etc.).
+    """
+
+    def __init__(
+        self,
+        port: str,
+        motors: dict[str, Motor],
+        calibration: dict[str, MotorCalibration] | None = None,
+    ):
+        self.port = port
+        self.motors = motors
+        self.calibration = calibration if calibration else {}
+
+    @abc.abstractmethod
+    def connect(self, handshake: bool = True) -> None:
+        """Establish connection to the motors."""
+        pass
+
+    @abc.abstractmethod
+    def disconnect(self, disable_torque: bool = True) -> None:
+        """Disconnect from the motors."""
+        pass
+
+    @property
+    @abc.abstractmethod
+    def is_connected(self) -> bool:
+        """Check if connected to the motors."""
+        pass
+
+    @abc.abstractmethod
+    def read(self, data_name: str, motor: str, *, normalize: bool = True, num_retry: int = 0) -> Value:
+        """Read a value from a single motor."""
+        pass
+
+    @abc.abstractmethod
+    def write(
+        self, data_name: str, motor: str, value: Value, *, normalize: bool = True, num_retry: int = 0
+    ) -> None:
+        """Write a value to a single motor."""
+        pass
+
+    @abc.abstractmethod
+    def sync_read(
+        self, data_name: str, motors: str | list[str] | None = None, *, normalize: bool = True
+    ) -> dict[str, Value]:
+        """Read a value from multiple motors."""
+        pass
+
+    @abc.abstractmethod
+    def sync_write(
+        self,
+        data_name: str,
+        values: Value | dict[str, Value],
+        motors: str | list[str] | None = None,
+        *,
+        normalize: bool = True,
+    ) -> None:
+        """Write values to multiple motors."""
+        pass
+
+    @abc.abstractmethod
+    def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+        """Enable torque on selected motors."""
+        pass
+
+    @abc.abstractmethod
+    def disable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
+        """Disable torque on selected motors."""
+        pass
+
+    @abc.abstractmethod
+    def read_calibration(self) -> dict[str, MotorCalibration]:
+        """Read calibration parameters from the motors."""
+        pass
+
+    @abc.abstractmethod
+    def write_calibration(self, calibration_dict: dict[str, MotorCalibration], cache: bool = True) -> None:
+        """Write calibration parameters to the motors."""
+        pass
+
+
 def get_ctrl_table(model_ctrl_table: dict[str, dict], model: str) -> dict[str, tuple[int, int]]:
     ctrl_table = model_ctrl_table.get(model)
     if ctrl_table is None:
@@ -203,15 +291,15 @@ class GroupSyncWrite(Protocol):
     def txPacket(self): ...
 
 
-class MotorsBus(abc.ABC):
+class SerialMotorsBus(MotorsBusBase):
     """
-    A MotorsBus allows to efficiently read and write to the attached motors.
+    A SerialMotorsBus allows to efficiently read and write to motors connected via serial communication.
     It represents several motors daisy-chained together and connected through a serial port.
-    There are currently two implementations of this abstract class:
+    There are currently two implementations of this class:
         - DynamixelMotorsBus
         - FeetechMotorsBus
 
-    Note: This class may evolve in the future should we add support for other types of bus.
+    This class is specifically for serial-based motor protocols (Dynamixel, Feetech, etc.).
 
     A MotorsBus subclass instance requires a port (e.g. `FeetechMotorsBus(port="/dev/tty.usbmodem575E0031751"`)).
     To find the port, you can run our utility script:
@@ -1212,3 +1300,7 @@ class MotorsBus(abc.ABC):
         for id_, value in ids_values.items():
             data = self._serialize_data(value, length)
             self.sync_writer.addParam(id_, data)
+
+
+# Backward compatibility alias
+MotorsBus = SerialMotorsBus

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/diffusion/configuration_diffusion.py

@@ -45,7 +45,7 @@ class DiffusionConfig(PreTrainedConfig):
     Args:
         n_obs_steps: Number of environment steps worth of observations to pass to the policy (takes the
             current step and additional steps going back).
-        chunk_size: Diffusion model action prediction size as detailed in `DiffusionPolicy.select_action`.
+        horizon: Diffusion model action prediction size as detailed in `DiffusionPolicy.select_action`.
         n_action_steps: The number of action steps to run in the environment for one invocation of the policy.
             See `DiffusionPolicy.select_action` for more details.
         input_shapes: A dictionary defining the shapes of the input data for the policy. The key represents
@@ -105,7 +105,7 @@ class DiffusionConfig(PreTrainedConfig):
 
     # Inputs / output structure.
     n_obs_steps: int = 2
-    chunk_size: int = 16
+    horizon: int = 16
     n_action_steps: int = 8
 
     normalization_mapping: dict[str, NormalizationMode] = field(
@@ -118,7 +118,7 @@ class DiffusionConfig(PreTrainedConfig):
 
     # The original implementation doesn't sample frames for the last 7 steps,
     # which avoids excessive padding and leads to improved training results.
-    drop_n_last_frames: int = 7  # chunk_size - n_action_steps - n_obs_steps + 1
+    drop_n_last_frames: int = 7  # horizon - n_action_steps - n_obs_steps + 1
 
     # Architecture / modeling.
     # Vision backbone.
@@ -180,13 +180,13 @@ class DiffusionConfig(PreTrainedConfig):
                 f"Got {self.noise_scheduler_type}."
             )
 
-        # Check that the chunk size and U-Net downsampling is compatible.
+        # Check that the horizon size and U-Net downsampling is compatible.
         # U-Net downsamples by 2 with each stage.
         downsampling_factor = 2 ** len(self.down_dims)
-        if self.chunk_size % downsampling_factor != 0:
+        if self.horizon % downsampling_factor != 0:
             raise ValueError(
-                "The chunk_size should be an integer multiple of the downsampling factor (which is determined "
-                f"by `len(down_dims)`). Got {self.chunk_size=} and {self.down_dims=}"
+                "The horizon should be an integer multiple of the downsampling factor (which is determined "
+                f"by `len(down_dims)`). Got {self.horizon=} and {self.down_dims=}"
             )
 
     def get_optimizer_preset(self) -> AdamConfig:
@@ -231,7 +231,7 @@ class DiffusionConfig(PreTrainedConfig):
 
     @property
     def action_delta_indices(self) -> list:
-        return list(range(1 - self.n_obs_steps, 1 - self.n_obs_steps + self.chunk_size))
+        return list(range(1 - self.n_obs_steps, 1 - self.n_obs_steps + self.horizon))
 
     @property
     def reward_delta_indices(self) -> None:

src/lerobot/policies/diffusion/modeling_diffusion.py

@@ -99,25 +99,25 @@ class DiffusionPolicy(PreTrainedPolicy):
         return actions
 
     @torch.no_grad()
-    def select_action(self, batch: dict[str, Tensor], noise: Tensor | None = None, **kwargs) -> Tensor:
+    def select_action(self, batch: dict[str, Tensor], noise: Tensor | None = None) -> Tensor:
         """Select a single action given environment observations.
 
         This method handles caching a history of observations and an action trajectory generated by the
         underlying diffusion model. Here's how it works:
           - `n_obs_steps` steps worth of observations are cached (for the first steps, the observation is
             copied `n_obs_steps` times to fill the cache).
-          - The diffusion model generates `chunk_size` steps worth of actions.
+          - The diffusion model generates `horizon` steps worth of actions.
           - `n_action_steps` worth of actions are actually kept for execution, starting from the current step.
         Schematically this looks like:
             ----------------------------------------------------------------------------------------------
-            (legend: o = n_obs_steps, c = chunk_size, a = n_action_steps)
+            (legend: o = n_obs_steps, h = horizon, a = n_action_steps)
             |timestep            | n-o+1 | n-o+2 | ..... | n     | ..... | n+a-1 | n+a   | ..... | n-o+h |
             |observation is used | YES   | YES   | YES   | YES   | NO    | NO    | NO    | NO    | NO    |
             |action is generated | YES   | YES   | YES   | YES   | YES   | YES   | YES   | YES   | YES   |
             |action is used      | NO    | NO    | NO    | YES   | YES   | YES   | NO    | NO    | NO    |
             ----------------------------------------------------------------------------------------------
-        Note that this means we require: `n_action_steps <= chunk_size - n_obs_steps + 1`. Also, note that
-        this period is
+        Note that this means we require: `n_action_steps <= horizon - n_obs_steps + 1`. Also, note that
+        "horizon" may not the best name to describe what the variable actually means, because this period is
         actually measured from the first observation which (if `n_obs_steps` > 1) happened in the past.
         """
         # NOTE: for offline evaluation, we have action in the batch, so we need to pop it out
@@ -213,7 +213,7 @@ class DiffusionModel(nn.Module):
             noise
             if noise is not None
             else torch.randn(
-                size=(batch_size, self.config.chunk_size, self.config.action_feature.shape[0]),
+                size=(batch_size, self.config.horizon, self.config.action_feature.shape[0]),
                 dtype=dtype,
                 device=device,
                 generator=generator,
@@ -309,16 +309,16 @@ class DiffusionModel(nn.Module):
                 AND/OR
             "observation.environment_state": (B, n_obs_steps, environment_dim)
 
-            "action": (B, chunk_size, action_dim)
-            "action_is_pad": (B, chunk_size)
+            "action": (B, horizon, action_dim)
+            "action_is_pad": (B, horizon)
         }
         """
         # Input validation.
         assert set(batch).issuperset({OBS_STATE, ACTION, "action_is_pad"})
         assert OBS_IMAGES in batch or OBS_ENV_STATE in batch
         n_obs_steps = batch[OBS_STATE].shape[1]
-        chunk_size = batch[ACTION].shape[1]
-        assert chunk_size == self.config.chunk_size
+        horizon = batch[ACTION].shape[1]
+        assert horizon == self.config.horizon
         assert n_obs_steps == self.config.n_obs_steps
 
         # Encode image features and concatenate them all together along with the state vector.

src/lerobot/policies/dsrl/configuration_dsrl.py

@@ -1,244 +0,0 @@
-# !/usr/bin/env python
-
-# Copyright 2025 The HuggingFace Inc. team.
-# All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from dataclasses import dataclass, field
-
-from lerobot.configs.policies import PreTrainedConfig
-from lerobot.configs.types import NormalizationMode
-from lerobot.optim.optimizers import MultiAdamConfig
-from lerobot.utils.constants import ACTION, OBS_IMAGE, OBS_STATE
-
-
-def is_image_feature(key: str) -> bool:
-    """Check if a feature key represents an image feature.
-
-    Args:
-        key: The feature key to check
-
-    Returns:
-        True if the key represents an image feature, False otherwise
-    """
-    return key.startswith(OBS_IMAGE)
-
-
-@dataclass
-class ConcurrencyConfig:
-    """Configuration for the concurrency of the actor and learner.
-    Possible values are:
-    - "threads": Use threads for the actor and learner.
-    - "processes": Use processes for the actor and learner.
-    """
-
-    actor: str = "threads"
-    learner: str = "threads"
-
-
-@dataclass
-class ActorLearnerConfig:
-    learner_host: str = "127.0.0.1"
-    learner_port: int = 50051
-    policy_parameters_push_frequency: int = 4
-    queue_get_timeout: float = 2
-
-
-@dataclass
-class CriticNetworkConfig:
-    hidden_dims: list[int] = field(default_factory=lambda: [256, 256])
-    activate_final: bool = True
-    final_activation: str | None = None
-
-
-@dataclass
-class ActorNetworkConfig:
-    hidden_dims: list[int] = field(default_factory=lambda: [256, 256])
-    activate_final: bool = True
-    use_layer_norm: bool = True
-
-
-@dataclass
-class NoiseActorConfig:
-    """Configuration for the noise actor in DSRL.
-    The noise actor outputs noise that gets fed to the diffusion policy.
-    """
-
-    use_tanh_squash: bool = False  # Whether to bound the noise output
-    std_min: float = 1e-5
-    std_max: float = 2.0
-    init_final: float = 0.05
-
-
-@PreTrainedConfig.register_subclass("dsrl")
-@dataclass
-class DSRLConfig(PreTrainedConfig):
-    """Diffusion Steering via Reinforcement Learning (DSRL) configuration."""
-
-    # Mapping of feature types to normalization modes
-    normalization_mapping: dict[str, NormalizationMode] = field(
-        default_factory=lambda: {
-            "VISUAL": NormalizationMode.MEAN_STD,
-            "STATE": NormalizationMode.MIN_MAX,
-            "ENV": NormalizationMode.MIN_MAX,
-            "ACTION": NormalizationMode.MIN_MAX,
-        }
-    )
-
-    # Statistics for normalizing different types of inputs
-    dataset_stats: dict[str, dict[str, list[float]]] | None = field(
-        default_factory=lambda: {
-            OBS_IMAGE: {
-                "mean": [0.485, 0.456, 0.406],
-                "std": [0.229, 0.224, 0.225],
-            },
-            OBS_STATE: {
-                "min": [0.0, 0.0],
-                "max": [1.0, 1.0],
-            },
-            ACTION: {
-                "min": [0.0, 0.0, 0.0],
-                "max": [1.0, 1.0, 1.0],
-            },
-        }
-    )
-
-    # Architecture specifics
-    # Device to run the model on (e.g., "cuda", "cpu")
-    device: str = "cpu"
-    # Device to store the model on
-    storage_device: str = "cpu"
-    # Name of the vision encoder model (Set to "helper2424/resnet10" for hil serl resnet10)
-    vision_encoder_name: str | None = None
-    # Whether to freeze the vision encoder during training
-    freeze_vision_encoder: bool = True
-    # Hidden dimension size for the image encoder
-    image_encoder_hidden_dim: int = 32
-    # Whether to use a shared encoder for actor and critic
-    shared_encoder: bool = True
-    # Number of discrete actions, eg for gripper actions
-    num_discrete_actions: int | None = None
-    # Dimension of the image embedding pooling
-    image_embedding_pooling_dim: int = 8
-
-    # Name of the action policy
-    action_policy_name: str = "pi0"
-    action_policy_weights: str | None = "lerobot/pi0_base"
-
-    # Training parameter
-    # Number of steps for online training
-    online_steps: int = 1000000
-    # Number of steps for offline training
-    offline_steps: int = 100000
-    # Capacity of the online replay buffer
-    online_buffer_capacity: int = 100000
-    # Capacity of the offline replay buffer
-    offline_buffer_capacity: int = 100000
-    # Whether to use asynchronous prefetching for the buffers
-    async_prefetch: bool = False
-    # Number of steps before learning starts
-    online_step_before_learning: int = 100
-    # Frequency of policy updates
-    policy_update_freq: int = 1
-
-    # SAC algorithm parameters
-    discount: float = 0.99
-    # Initial temperature value
-    temperature_init: float = 1.0
-    # Number of critics in the ensemble
-    num_critics: int = 2
-    # Number of subsampled critics for training
-    num_subsample_critics: int | None = None
-    # Learning rate for the critic network
-    critic_lr: float = 3e-4
-    # Learning rate for the actor network
-    actor_lr: float = 3e-4
-    # Learning rate for the temperature parameter
-    temperature_lr: float = 3e-4
-    # Weight for the critic target update
-    critic_target_update_weight: float = 0.005
-    # Update-to-data ratio for the UTD algorithm (If you want enable utd_ratio, you need to set it to >1)
-    utd_ratio: int = 1
-    # Hidden dimension size for the state encoder
-    state_encoder_hidden_dim: int = 256
-    # Dimension of the latent space
-    latent_dim: int = 256
-    # Target entropy for the SAC algorithm
-    target_entropy: float | None = None
-    # Whether to use backup entropy for the SAC algorithm
-    use_backup_entropy: bool = True
-    # Gradient clipping norm for the SAC algorithm
-    grad_clip_norm: float = 40.0
-
-    # Network configuration
-    # Configuration for the critic network architecture
-    critic_network_kwargs: CriticNetworkConfig = field(default_factory=CriticNetworkConfig)
-    # Configuration for the noise critic network architecture
-    noise_critic_network_kwargs: CriticNetworkConfig = field(default_factory=CriticNetworkConfig)
-    # Configuration for the noise actor network architecture
-    noise_actor_network_kwargs: ActorNetworkConfig = field(default_factory=ActorNetworkConfig)
-    # Configuration for the noise actor specific parameters
-    noise_actor_kwargs: NoiseActorConfig = field(default_factory=NoiseActorConfig)
-    # Configuration for actor-learner architecture
-    actor_learner_config: ActorLearnerConfig = field(default_factory=ActorLearnerConfig)
-    # Configuration for concurrency settings (you can use threads or processes for the actor and learner)
-    concurrency: ConcurrencyConfig = field(default_factory=ConcurrencyConfig)
-
-    # Optimizations
-    use_torch_compile: bool = True
-
-    def __post_init__(self):
-        super().__post_init__()
-
-    def get_optimizer_preset(self) -> MultiAdamConfig:
-        return MultiAdamConfig(
-            weight_decay=0.0,
-            optimizer_groups={
-                "critic_action": {"lr": self.critic_lr},
-                "critic_noise": {"lr": self.critic_lr},
-                "noise_actor": {"lr": self.actor_lr},
-                "temperature": {"lr": self.temperature_lr},
-            },
-        )
-
-    def get_scheduler_preset(self) -> None:
-        return None
-
-    def validate_features(self) -> None:
-        has_image = any(is_image_feature(key) for key in self.input_features)
-        has_state = OBS_STATE in self.input_features
-
-        if not (has_state or has_image):
-            raise ValueError(
-                "You must provide either 'observation.state' or an image observation (key starting with 'observation.image') in the input features"
-            )
-
-        if ACTION not in self.output_features:
-            raise ValueError("You must provide 'action' in the output features")
-
-    @property
-    def image_features(self) -> list[str]:
-        return [key for key in self.input_features if is_image_feature(key)]
-
-    @property
-    def observation_delta_indices(self) -> list:
-        return None
-
-    @property
-    def action_delta_indices(self) -> list:
-        return None
-
-    @property
-    def reward_delta_indices(self) -> None:
-        return None

src/lerobot/policies/dsrl/processor_dsrl.py

@@ -1,89 +0,0 @@
-# !/usr/bin/env python
-
-# Copyright 2025 The HuggingFace Inc. team.
-# All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""
-Processor for DSRL policy.
-
-DSRL uses a similar processing pipeline as SAC since it operates on
-state-action transitions. The main difference is that internally it
-also works with noise, but that's handled within the policy itself.
-"""
-
-from typing import Any
-
-import torch
-
-from lerobot.policies.dsrl.configuration_dsrl import DSRLConfig
-from lerobot.processor import (
-    AddBatchDimensionProcessorStep,
-    DeviceProcessorStep,
-    NormalizerProcessorStep,
-    PolicyAction,
-    PolicyProcessorPipeline,
-    RenameObservationsProcessorStep,
-    UnnormalizerProcessorStep,
-)
-from lerobot.processor.converters import (
-    policy_action_to_transition,
-    transition_to_policy_action,
-)
-from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
-
-
-def make_dsrl_pre_post_processors(
-    config: DSRLConfig,
-    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
-) -> tuple[
-    PolicyProcessorPipeline[dict, dict],
-    PolicyProcessorPipeline[PolicyAction, PolicyAction],
-]:
-    """Create preprocessor and postprocessor pipelines for DSRL policy.
-
-    Args:
-        config: DSRL policy configuration
-        dataset_stats: Optional dataset statistics for normalization
-
-    Returns:
-        Tuple of (preprocessor, postprocessor) pipelines
-    """
-    input_steps = [
-        RenameObservationsProcessorStep(rename_map={}),
-        AddBatchDimensionProcessorStep(),
-        DeviceProcessorStep(device=config.device),
-        NormalizerProcessorStep(
-            features={**config.input_features, **config.output_features},
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-        ),
-    ]
-    output_steps = [
-        UnnormalizerProcessorStep(
-            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
-        ),
-        DeviceProcessorStep(device="cpu"),
-    ]
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )

src/lerobot/policies/factory.py

@@ -30,9 +30,8 @@ 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.dsrl.configuration_dsrl import DSRLConfig
+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
@@ -59,7 +58,7 @@ 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", "dsrl".
+              "vqbet", "pi0", "pi05", "sac", "reward_classifier", "smolvla".
 
     Returns:
         The policy class corresponding to the given name.
@@ -83,10 +82,6 @@ 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
 
@@ -107,10 +102,10 @@ def get_policy_class(name: str) -> type[PreTrainedPolicy]:
         from lerobot.policies.smolvla.modeling_smolvla import SmolVLAPolicy
 
         return SmolVLAPolicy
-    elif name == "dsrl":
-        from lerobot.policies.dsrl.modeling_dsrl import DSRLPolicy
+    elif name == "groot":
+        from lerobot.policies.groot.modeling_groot import GrootPolicy
 
-        return DSRLPolicy
+        return GrootPolicy
     else:
         raise NotImplementedError(f"Policy with name {name} is not implemented.")
 
@@ -124,8 +119,8 @@ 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",
-                     "reward_classifier", "dsrl".
+                     "diffusion", "act", "vqbet", "pi0", "pi05", "sac", "smolvla",
+                     "reward_classifier".
         **kwargs: Keyword arguments to be passed to the configuration class constructor.
 
     Returns:
@@ -142,8 +137,6 @@ 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":
@@ -154,8 +147,8 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
         return SmolVLAConfig(**kwargs)
     elif policy_type == "reward_classifier":
         return RewardClassifierConfig(**kwargs)
-    elif policy_type == "dsrl":
-        return DSRLConfig(**kwargs)
+    elif policy_type == "groot":
+        return GrootConfig(**kwargs)
     else:
         raise ValueError(f"Policy type '{policy_type}' is not available.")
 
@@ -213,6 +206,27 @@ 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,
@@ -267,14 +281,6 @@ 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
 
@@ -314,10 +320,11 @@ def make_pre_post_processors(
             config=policy_cfg,
             dataset_stats=kwargs.get("dataset_stats"),
         )
-    elif isinstance(policy_cfg, DSRLConfig):
-        from lerobot.policies.dsrl.processor_dsrl import make_dsrl_pre_post_processors
 
-        processors = make_dsrl_pre_post_processors(
+    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"),
         )
@@ -332,6 +339,7 @@ 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.
@@ -348,6 +356,8 @@ 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.
@@ -389,8 +399,10 @@ 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
 
     if cfg.pretrained_path:
@@ -407,4 +419,21 @@ def make_policy(
 
     # policy = torch.compile(policy, mode="reduce-overhead")
 
+    if not rename_map:
+        expected_features = set(cfg.input_features.keys()) | set(cfg.output_features.keys())
+        provided_features = set(features.keys())
+        if expected_features and provided_features != expected_features:
+            missing = expected_features - provided_features
+            extra = provided_features - expected_features
+            # TODO (jadechoghari): provide a dynamic rename map suggestion to the user.
+            raise ValueError(
+                f"Feature mismatch between dataset/environment and policy config.\n"
+                f"- Missing features: {sorted(missing) if missing else 'None'}\n"
+                f"- Extra features: {sorted(extra) if extra else 'None'}\n\n"
+                f"Please ensure your dataset and policy use consistent feature names.\n"
+                f"If your dataset uses different observation keys (e.g., cameras named differently), "
+                f"use the `--rename_map` argument, for example:\n"
+                f'  --rename_map=\'{{"observation.images.left": "observation.images.camera1", '
+                f'"observation.images.top": "observation.images.camera2"}}\''
+            )
     return policy

src/lerobot/policies/groot/README.md

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

src/lerobot/policies/groot/__init__.py

@@ -0,0 +1,21 @@
+#!/usr/bin/env python
+
+# Copyright 2025 Nvidia and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .configuration_groot import GrootConfig
+from .modeling_groot import GrootPolicy
+from .processor_groot import make_groot_pre_post_processors
+
+__all__ = ["GrootConfig", "GrootPolicy", "make_groot_pre_post_processors"]

src/lerobot/policies/groot/action_head/__init__.py

@@ -1,18 +1,14 @@
-#!/usr/bin/env python
-
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
-#     http://www.apache.org/licenses/LICENSE-2.0
+# http://www.apache.org/licenses/LICENSE-2.0
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-
-from .config_widowx import WidowXConfig
-from .widowx import WidowX

src/lerobot/policies/groot/action_head/action_encoder.py

@@ -0,0 +1,54 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+
+
+def swish(x):
+    return x * torch.sigmoid(x)
+
+
+class SinusoidalPositionalEncoding(nn.Module):
+    """
+    Produces a sinusoidal encoding of shape (B, T, w)
+    given timesteps of shape (B, T).
+    """
+
+    def __init__(self, embedding_dim):
+        super().__init__()
+        self.embedding_dim = embedding_dim
+
+    def forward(self, timesteps):
+        # timesteps: shape (B, T)
+        # We'll compute sin/cos frequencies across dim T
+        timesteps = timesteps.float()  # ensure float
+
+        b, t = timesteps.shape
+        device = timesteps.device
+
+        half_dim = self.embedding_dim // 2
+        # typical log space frequencies for sinusoidal encoding
+        exponent = -torch.arange(half_dim, dtype=torch.float, device=device) * (
+            torch.log(torch.tensor(10000.0)) / half_dim
+        )
+        # Expand timesteps to (B, T, 1) then multiply
+        freqs = timesteps.unsqueeze(-1) * exponent.exp()  # (B, T, half_dim)
+
+        sin = torch.sin(freqs)
+        cos = torch.cos(freqs)
+        enc = torch.cat([sin, cos], dim=-1)  # (B, T, w)
+
+        return enc

src/lerobot/policies/groot/action_head/cross_attention_dit.py

@@ -0,0 +1,370 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import torch
+import torch.nn.functional as F  # noqa: N812
+from diffusers import ConfigMixin, ModelMixin
+from diffusers.configuration_utils import register_to_config
+from diffusers.models.attention import Attention, FeedForward
+from diffusers.models.embeddings import (
+    SinusoidalPositionalEmbedding,
+    TimestepEmbedding,
+    Timesteps,
+)
+from torch import nn
+
+
+class TimestepEncoder(nn.Module):
+    def __init__(self, embedding_dim, compute_dtype=torch.float32):
+        super().__init__()
+        self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=1)
+        self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim)
+
+    def forward(self, timesteps):
+        dtype = next(self.parameters()).dtype
+        timesteps_proj = self.time_proj(timesteps).to(dtype)
+        timesteps_emb = self.timestep_embedder(timesteps_proj)  # (N, D)
+        return timesteps_emb
+
+
+class AdaLayerNorm(nn.Module):
+    def __init__(
+        self,
+        embedding_dim: int,
+        norm_elementwise_affine: bool = False,
+        norm_eps: float = 1e-5,
+        chunk_dim: int = 0,
+    ):
+        super().__init__()
+        self.chunk_dim = chunk_dim
+        output_dim = embedding_dim * 2
+        self.silu = nn.SiLU()
+        self.linear = nn.Linear(embedding_dim, output_dim)
+        self.norm = nn.LayerNorm(output_dim // 2, norm_eps, norm_elementwise_affine)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        temb: torch.Tensor | None = None,
+    ) -> torch.Tensor:
+        temb = self.linear(self.silu(temb))
+        scale, shift = temb.chunk(2, dim=1)
+        x = self.norm(x) * (1 + scale[:, None]) + shift[:, None]
+        return x
+
+
+class BasicTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout=0.0,
+        cross_attention_dim: int | None = None,
+        activation_fn: str = "geglu",
+        attention_bias: bool = False,
+        upcast_attention: bool = False,
+        norm_elementwise_affine: bool = True,
+        norm_type: str = "layer_norm",  # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single', 'ada_norm_continuous', 'layer_norm_i2vgen'
+        norm_eps: float = 1e-5,
+        final_dropout: bool = False,
+        attention_type: str = "default",
+        positional_embeddings: str | None = None,
+        num_positional_embeddings: int | None = None,
+        ff_inner_dim: int | None = None,
+        ff_bias: bool = True,
+        attention_out_bias: bool = True,
+    ):
+        super().__init__()
+        self.dim = dim
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        self.dropout = dropout
+        self.cross_attention_dim = cross_attention_dim
+        self.activation_fn = activation_fn
+        self.attention_bias = attention_bias
+        self.norm_elementwise_affine = norm_elementwise_affine
+        self.positional_embeddings = positional_embeddings
+        self.num_positional_embeddings = num_positional_embeddings
+        self.norm_type = norm_type
+
+        if positional_embeddings and (num_positional_embeddings is None):
+            raise ValueError(
+                "If `positional_embeddings` type is defined, `num_positional_embeddings` must also be defined."
+            )
+
+        if positional_embeddings == "sinusoidal":
+            self.pos_embed = SinusoidalPositionalEmbedding(dim, max_seq_length=num_positional_embeddings)
+        else:
+            self.pos_embed = None
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Self-Attn
+        if norm_type == "ada_norm":
+            self.norm1 = AdaLayerNorm(dim)
+        else:
+            self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
+
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=cross_attention_dim,
+            upcast_attention=upcast_attention,
+            out_bias=attention_out_bias,
+        )
+
+        # 3. Feed-forward
+        self.norm3 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)
+        self.ff = FeedForward(
+            dim,
+            dropout=dropout,
+            activation_fn=activation_fn,
+            final_dropout=final_dropout,
+            inner_dim=ff_inner_dim,
+            bias=ff_bias,
+        )
+        if final_dropout:
+            self.final_dropout = nn.Dropout(dropout)
+        else:
+            self.final_dropout = None
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor | None = None,
+        encoder_hidden_states: torch.Tensor | None = None,
+        encoder_attention_mask: torch.Tensor | None = None,
+        temb: torch.LongTensor | None = None,
+    ) -> torch.Tensor:
+        # 0. Self-Attention
+        if self.norm_type == "ada_norm":
+            norm_hidden_states = self.norm1(hidden_states, temb)
+        else:
+            norm_hidden_states = self.norm1(hidden_states)
+
+        if self.pos_embed is not None:
+            norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+        attn_output = self.attn1(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states,
+            attention_mask=attention_mask,
+            # encoder_attention_mask=encoder_attention_mask,
+        )
+        if self.final_dropout:
+            attn_output = self.final_dropout(attn_output)
+
+        hidden_states = attn_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        # 4. Feed-forward
+        norm_hidden_states = self.norm3(hidden_states)
+        ff_output = self.ff(norm_hidden_states)
+
+        hidden_states = ff_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+        return hidden_states
+
+
+class DiT(ModelMixin, ConfigMixin):
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        num_attention_heads: int = 8,
+        attention_head_dim: int = 64,
+        output_dim: int = 26,
+        num_layers: int = 12,
+        dropout: float = 0.1,
+        attention_bias: bool = True,
+        activation_fn: str = "gelu-approximate",
+        num_embeds_ada_norm: int | None = 1000,
+        upcast_attention: bool = False,
+        norm_type: str = "ada_norm",
+        norm_elementwise_affine: bool = False,
+        norm_eps: float = 1e-5,
+        max_num_positional_embeddings: int = 512,
+        compute_dtype=torch.float32,
+        final_dropout: bool = True,
+        positional_embeddings: str | None = "sinusoidal",
+        interleave_self_attention=False,
+        cross_attention_dim: int | None = None,
+    ):
+        super().__init__()
+
+        self.attention_head_dim = attention_head_dim
+        self.inner_dim = self.config.num_attention_heads * self.config.attention_head_dim
+        self.gradient_checkpointing = False
+
+        # Timestep encoder
+        self.timestep_encoder = TimestepEncoder(
+            embedding_dim=self.inner_dim, compute_dtype=self.config.compute_dtype
+        )
+
+        all_blocks = []
+        for idx in range(self.config.num_layers):
+            use_self_attn = idx % 2 == 1 and interleave_self_attention
+            curr_cross_attention_dim = cross_attention_dim if not use_self_attn else None
+
+            all_blocks += [
+                BasicTransformerBlock(
+                    self.inner_dim,
+                    self.config.num_attention_heads,
+                    self.config.attention_head_dim,
+                    dropout=self.config.dropout,
+                    activation_fn=self.config.activation_fn,
+                    attention_bias=self.config.attention_bias,
+                    upcast_attention=self.config.upcast_attention,
+                    norm_type=norm_type,
+                    norm_elementwise_affine=self.config.norm_elementwise_affine,
+                    norm_eps=self.config.norm_eps,
+                    positional_embeddings=positional_embeddings,
+                    num_positional_embeddings=self.config.max_num_positional_embeddings,
+                    final_dropout=final_dropout,
+                    cross_attention_dim=curr_cross_attention_dim,
+                )
+            ]
+        self.transformer_blocks = nn.ModuleList(all_blocks)
+
+        # Output blocks
+        self.norm_out = nn.LayerNorm(self.inner_dim, elementwise_affine=False, eps=1e-6)
+        self.proj_out_1 = nn.Linear(self.inner_dim, 2 * self.inner_dim)
+        self.proj_out_2 = nn.Linear(self.inner_dim, self.config.output_dim)
+        print(
+            "Total number of DiT parameters: ",
+            sum(p.numel() for p in self.parameters() if p.requires_grad),
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,  # Shape: (B, T, D)
+        encoder_hidden_states: torch.Tensor,  # Shape: (B, S, D)
+        timestep: torch.LongTensor | None = None,
+        encoder_attention_mask: torch.Tensor | None = None,
+        return_all_hidden_states: bool = False,
+    ):
+        # Encode timesteps
+        temb = self.timestep_encoder(timestep)
+
+        # Process through transformer blocks - single pass through the blocks
+        hidden_states = hidden_states.contiguous()
+        encoder_hidden_states = encoder_hidden_states.contiguous()
+
+        all_hidden_states = [hidden_states]
+
+        # Process through transformer blocks
+        for idx, block in enumerate(self.transformer_blocks):
+            if idx % 2 == 1 and self.config.interleave_self_attention:
+                hidden_states = block(
+                    hidden_states,
+                    attention_mask=None,
+                    encoder_hidden_states=None,
+                    encoder_attention_mask=None,
+                    temb=temb,
+                )
+            else:
+                hidden_states = block(
+                    hidden_states,
+                    attention_mask=None,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=None,
+                    temb=temb,
+                )
+            all_hidden_states.append(hidden_states)
+
+        # Output processing
+        conditioning = temb
+        shift, scale = self.proj_out_1(F.silu(conditioning)).chunk(2, dim=1)
+        hidden_states = self.norm_out(hidden_states) * (1 + scale[:, None]) + shift[:, None]
+        if return_all_hidden_states:
+            return self.proj_out_2(hidden_states), all_hidden_states
+        else:
+            return self.proj_out_2(hidden_states)
+
+
+class SelfAttentionTransformer(ModelMixin, ConfigMixin):
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        num_attention_heads: int = 8,
+        attention_head_dim: int = 64,
+        output_dim: int = 26,
+        num_layers: int = 12,
+        dropout: float = 0.1,
+        attention_bias: bool = True,
+        activation_fn: str = "gelu-approximate",
+        num_embeds_ada_norm: int | None = 1000,
+        upcast_attention: bool = False,
+        max_num_positional_embeddings: int = 512,
+        compute_dtype=torch.float32,
+        final_dropout: bool = True,
+        positional_embeddings: str | None = "sinusoidal",
+        interleave_self_attention=False,
+    ):
+        super().__init__()
+
+        self.attention_head_dim = attention_head_dim
+        self.inner_dim = self.config.num_attention_heads * self.config.attention_head_dim
+        self.gradient_checkpointing = False
+
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    self.inner_dim,
+                    self.config.num_attention_heads,
+                    self.config.attention_head_dim,
+                    dropout=self.config.dropout,
+                    activation_fn=self.config.activation_fn,
+                    attention_bias=self.config.attention_bias,
+                    upcast_attention=self.config.upcast_attention,
+                    positional_embeddings=positional_embeddings,
+                    num_positional_embeddings=self.config.max_num_positional_embeddings,
+                    final_dropout=final_dropout,
+                )
+                for _ in range(self.config.num_layers)
+            ]
+        )
+        print(
+            "Total number of SelfAttentionTransformer parameters: ",
+            sum(p.numel() for p in self.parameters() if p.requires_grad),
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,  # Shape: (B, T, D)
+        return_all_hidden_states: bool = False,
+    ):
+        # Process through transformer blocks - single pass through the blocks
+        hidden_states = hidden_states.contiguous()
+        all_hidden_states = [hidden_states]
+
+        # Process through transformer blocks
+        for _idx, block in enumerate(self.transformer_blocks):
+            hidden_states = block(hidden_states)
+            all_hidden_states.append(hidden_states)
+
+        if return_all_hidden_states:
+            return hidden_states, all_hidden_states
+        else:
+            return hidden_states