update close gripper button

* add blog/guide

* add to tree

* chore(docs): rephrase rename_map docs for clarity and simplicity

---------

Co-authored-by: Steven Palma <steven.palma@huggingface.co>
Co-authored-by: Steven Palma <imstevenpmwork@ieee.org>

.github/workflows/fast_tests.yml

@@ -44,7 +44,7 @@ permissions:
 # Sets up the environment variables
 env:
   UV_VERSION: "0.8.0"
-  PYTHON_VERSION: "3.10"
+  PYTHON_VERSION: "3.12"
 
 # Ensures that only the latest commit for a PR or branch is built, canceling older runs.
 concurrency:
@@ -61,6 +61,7 @@ jobs:
       MUJOCO_GL: egl
       HF_HOME: /mnt/cache/.cache/huggingface
       HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     steps:
       - uses: actions/checkout@v6
         with:
@@ -89,5 +90,11 @@ jobs:
       - name: Install lerobot with test extras
         run: uv sync --extra "test"
 
+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          uv run hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
+          uv run hf auth whoami
+
       - name: Run pytest
         run: uv run pytest tests -vv --maxfail=10

.github/workflows/full_tests.yml

@@ -37,7 +37,7 @@ permissions:
 # Sets up the environment variables
 env:
   UV_VERSION: "0.8.0"
-  PYTHON_VERSION: "3.10"
+  PYTHON_VERSION: "3.12"
   DOCKER_IMAGE_NAME: huggingface/lerobot-gpu
 
 # Ensures that only the latest action is built, canceling older runs.
@@ -60,6 +60,7 @@ jobs:
       MUJOCO_GL: egl
       HF_HOME: /mnt/cache/.cache/huggingface
       HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     steps:
       - uses: actions/checkout@v6
         with:
@@ -87,6 +88,12 @@ jobs:
       - name: Install lerobot with all extras
         run: uv sync --extra all # TODO(Steven): Make flash-attn optional
 
+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          uv run hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
+          uv run hf auth whoami
+
       - name: Run pytest (all extras)
         run: uv run pytest tests -vv --maxfail=10
 
@@ -162,6 +169,7 @@ jobs:
       HF_LEROBOT_HOME: /home/user_lerobot/.cache/huggingface/lerobot
       TORCH_HOME: /home/user_lerobot/.cache/torch
       TRITON_CACHE_DIR: /home/user_lerobot/.cache/triton
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     container:
       image: ${{ needs.build-and-push-docker.outputs.image_tag }} # zizmor: ignore[unpinned-images]
       options: --gpus all --shm-size "16gb"
@@ -173,8 +181,13 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
+          hf auth whoami
       - name: Fix ptxas permissions
-        run: chmod +x /lerobot/.venv/lib/python3.10/site-packages/triton/backends/nvidia/bin/ptxas
+        run: chmod +x /lerobot/.venv/lib/python3.12/site-packages/triton/backends/nvidia/bin/ptxas
       - name: Run pytest on GPU
         run: pytest tests -vv --maxfail=10
       - name: Run end-to-end tests

.github/workflows/nightly.yml

@@ -28,7 +28,7 @@ on:
 # Sets up the environment variables
 env:
   UV_VERSION: "0.8.0"
-  PYTHON_VERSION: "3.10"
+  PYTHON_VERSION: "3.12"
   DOCKER_IMAGE_NAME_CPU: huggingface/lerobot-cpu:latest
   DOCKER_IMAGE_NAME_GPU: huggingface/lerobot-gpu:latest
 
@@ -119,6 +119,7 @@ jobs:
       HF_LEROBOT_HOME: /home/user_lerobot/.cache/huggingface/lerobot
       TORCH_HOME: /home/user_lerobot/.cache/torch
       TRITON_CACHE_DIR: /home/user_lerobot/.cache/triton
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     container:
       image: ${{ needs.build-docker-cpu-nightly.outputs.image_tag }} # zizmor: ignore[unpinned-images]
       options: --shm-size "16gb"
@@ -130,6 +131,11 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
+          hf auth whoami
       - name: Run pytest on CPU
         run: pytest tests -vv --maxfail=10
       - name: Run end-to-end tests
@@ -146,6 +152,7 @@ jobs:
       HF_LEROBOT_HOME: /home/user_lerobot/.cache/huggingface/lerobot
       TORCH_HOME: /home/user_lerobot/.cache/torch
       TRITON_CACHE_DIR: /home/user_lerobot/.cache/triton
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     container:
       image: ${{ needs.build-docker-gpu-nightly.outputs.image_tag }} # zizmor: ignore[unpinned-images]
       options: --gpus all --shm-size "16gb"
@@ -157,6 +164,11 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
+          hf auth whoami
       - name: Run pytest on GPU
         run: pytest tests -vv --maxfail=10
       - name: Run end-to-end tests
@@ -174,6 +186,7 @@ jobs:
       TORCH_HOME: /home/user_lerobot/.cache/torch
       TRITON_CACHE_DIR: /home/user_lerobot/.cache/triton
       CUDA_VISIBLE_DEVICES: "0,1,2,3"
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     container:
       image: ${{ needs.build-docker-gpu-nightly.outputs.image_tag }} # zizmor: ignore[unpinned-images]
       options: --gpus all --shm-size "16gb"
@@ -185,12 +198,15 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
+          hf auth whoami
       - 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
+        run: pytest -vv tests/training/

.github/workflows/quality.yml

@@ -50,7 +50,7 @@ jobs:
       - name: Set up Python
         uses: actions/setup-python@v6
         with:
-          python-version: '3.10'
+          python-version: '3.12'
 
       - name: Run pre-commit hooks
         uses: pre-commit/action@v3.0.1 # zizmor: ignore[unpinned-uses]

.github/workflows/release.yml

@@ -22,7 +22,7 @@ on:
 # Sets up the environment variables
 env:
   UV_VERSION: "0.8.0"
-  PYTHON_VERSION: "3.10"
+  PYTHON_VERSION: "3.12"
 
 jobs:
   # This job builds the Python package and publishes it to PyPI
@@ -45,7 +45,7 @@ jobs:
       - name: Set up Python
         uses: actions/setup-python@v6
         with:
-          python-version: '3.10'
+          python-version: '3.12'
 
       - name: Extract Version
         id: extract_info
@@ -83,14 +83,6 @@ jobs:
             exit 1
           fi
 
-      - name: Remove Tags with Git dependencies
-        # TODO(Steven): Temporary patch to remove pi from PyPi 0.4.0 release due to its reliance on git dependencies.
-        run: |
-          echo "::info:: Checking for Git dependencies to remove from pyproject.toml..."
-          grep -E '@ git\+https|lerobot\[pi\]' pyproject.toml | sed 's/^/::warning:: Removing line: /' || true
-          sed -E -i '/@ git\+https|lerobot\[pi\]/d' pyproject.toml
-          echo "::info:: Git dependencies removed. Proceeding with build."
-
       - name: Install build dependencies
         run: python -m pip install build
 

.github/workflows/unbound_deps_tests.yml

@@ -29,7 +29,7 @@ permissions:
 # Sets up the environment variables
 env:
   UV_VERSION: "0.8.0"
-  PYTHON_VERSION: "3.10"
+  PYTHON_VERSION: "3.12"
   DOCKER_IMAGE_NAME: huggingface/lerobot-gpu:unbound
 
 # Ensures that only the latest action is built, canceling older runs.
@@ -48,6 +48,7 @@ jobs:
       MUJOCO_GL: egl
       HF_HOME: /mnt/cache/.cache/huggingface
       HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     steps:
       - uses: actions/checkout@v6
         with:
@@ -79,7 +80,11 @@ jobs:
 
       - name: Install lerobot with all extras
         run: uv sync --extra all # TODO(Steven): Make flash-attn optional
-
+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          uv run hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
+          uv run hf auth whoami
       - name: Run pytest (all extras)
         run: uv run pytest tests -vv
 
@@ -137,6 +142,7 @@ jobs:
       HF_LEROBOT_HOME: /home/user_lerobot/.cache/huggingface/lerobot
       TORCH_HOME: /home/user_lerobot/.cache/torch
       TRITON_CACHE_DIR: /home/user_lerobot/.cache/triton
+      HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
     container:
       image: ${{ needs.build-and-push-docker.outputs.image_tag }} # zizmor: ignore[unpinned-images]
       options: --gpus all --shm-size "16gb"
@@ -148,6 +154,11 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
+      - name: Login to Hugging Face
+        if: env.HF_USER_TOKEN != ''
+        run: |
+          hf auth login --token "$HF_USER_TOKEN" --add-to-git-credential
+          hf auth whoami
       - name: Run pytest on GPU
         run: pytest tests -vv
       - name: Run end-to-end tests

.pre-commit-config.yaml

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 default_language_version:
-    python: python3.10
+    python: python3.12
 
 exclude: "tests/artifacts/.*\\.safetensors$"
 
@@ -55,7 +55,7 @@ repos:
     rev: v3.21.0
     hooks:
     -   id: pyupgrade
-        args: [--py310-plus]
+        args: [--py312-plus]
 
   ##### Markdown Quality #####
   - repo: https://github.com/rbubley/mirrors-prettier

CONTRIBUTING.md

@@ -2,7 +2,7 @@
 
 Everyone is welcome to contribute, and we value everybody's contribution. Code is not the only way to help the community. Answering questions, helping others, reaching out, and improving the documentation are immensely valuable.
 
-Whichever way you choose to contribute, please be mindful to respect our [code of conduct](./CODE_OF_CONDUCT.md) and our [AI policy](./AI_POLICY.md).
+Whichever way you choose to contribute, please be mindful to respect our [code of conduct](https://github.com/huggingface/lerobot/blob/main/CODE_OF_CONDUCT.md) and our [AI policy](https://github.com/huggingface/lerobot/blob/main/AI_POLICY.md).
 
 ## Ways to Contribute
 
@@ -32,7 +32,7 @@ git remote add upstream https://github.com/huggingface/lerobot.git
 
 ### 2. Environment Installation
 
-Please follow our [Installation Guide](./docs/source/installation.mdx) for the environment setup & installation from source.
+Please follow our [Installation Guide](https://huggingface.co/docs/lerobot/installation) for the environment setup & installation from source.
 
 ## Running Tests & Quality Checks
 
@@ -75,8 +75,8 @@ pytest -sv tests/test_specific_feature.py
 
 Use the templates for required fields and examples.
 
-- **Issues:** Follow the [ticket template](./.github/ISSUE_TEMPLATE/bug-report.yml).
-- **Pull requests:** Rebase on `upstream/main`, use a descriptive branch (don't work on `main`), run `pre-commit` and tests locally, and follow the [PR template](./.github/PULL_REQUEST_TEMPLATE.md).
+- **Issues:** Follow the [ticket template](https://github.com/huggingface/lerobot/blob/main/.github/ISSUE_TEMPLATE/bug-report.yml).
+- **Pull requests:** Rebase on `upstream/main`, use a descriptive branch (don't work on `main`), run `pre-commit` and tests locally, and follow the [PR template](https://github.com/huggingface/lerobot/blob/main/.github/PULL_REQUEST_TEMPLATE.md).
 
 One member of the LeRobot team will then review your contribution.
 

README.md

@@ -135,7 +135,7 @@ Learn how to implement your own simulation environment or benchmark and distribu
 
 ## Citation
 
-If you use LeRobot in your research, please cite:
+If you use LeRobot in your project, please cite the GitHub repository to acknowledge the ongoing development and contributors:
 
 ```bibtex
 @misc{cadene2024lerobot,
@@ -146,9 +146,26 @@ If you use LeRobot in your research, please cite:
 }
 ```
 
+If you are referencing our research or the academic paper, please also cite our ICLR publication:
+
+<details>
+<summary><b>ICLR 2026 Paper</b></summary>
+
+```bibtex
+@inproceedings{cadenelerobot,
+  title={LeRobot: An Open-Source Library for End-to-End Robot Learning},
+  author={Cadene, Remi and Alibert, Simon and Capuano, Francesco and Aractingi, Michel and Zouitine, Adil and Kooijmans, Pepijn and Choghari, Jade and Russi, Martino and Pascal, Caroline and Palma, Steven and Shukor, Mustafa and Moss, Jess and Soare, Alexander and Aubakirova, Dana and Lhoest, Quentin and Gallou\'edec, Quentin and Wolf, Thomas},
+  booktitle={The Fourteenth International Conference on Learning Representations},
+  year={2026},
+  url={https://arxiv.org/abs/2602.22818}
+}
+```
+
+</details>
+
 ## Contribute
 
-We welcome contributions from everyone in the community! To get started, please read our [CONTRIBUTING.md](./CONTRIBUTING.md) guide. Whether you're adding a new feature, improving documentation, or fixing a bug, your help and feedback are invaluable. We're incredibly excited about the future of open-source robotics and can't wait to work with you on what's next—thank you for your support!
+We welcome contributions from everyone in the community! To get started, please read our [CONTRIBUTING.md](https://github.com/huggingface/lerobot/blob/main/CONTRIBUTING.md) guide. Whether you're adding a new feature, improving documentation, or fixing a bug, your help and feedback are invaluable. We're incredibly excited about the future of open-source robotics and can't wait to work with you on what's next—thank you for your support!
 
 <p align="center">
   <img alt="SO101 Video" src="./media/readme/so100_video.webp" width="640px">

docker/Dockerfile.internal

@@ -24,7 +24,7 @@ ARG OS_VERSION=22.04
 FROM nvidia/cuda:${CUDA_VERSION}-base-ubuntu${OS_VERSION}
 
 # Define Python version argument
-ARG PYTHON_VERSION=3.10
+ARG PYTHON_VERSION=3.12
 
 # Configure environment variables
 ENV DEBIAN_FRONTEND=noninteractive \

docker/Dockerfile.user

@@ -18,8 +18,10 @@
 # docker build -f docker/Dockerfile.user -t lerobot-user .
 # docker run -it --rm lerobot-user
 
+# With USB physical access : docker run -it --device=/dev/ -v /dev/:/dev/ --rm lerobot-user
+
 # Configure the base image
-ARG PYTHON_VERSION=3.10
+ARG PYTHON_VERSION=3.12
 FROM python:${PYTHON_VERSION}-slim
 
 # Configure environment variables

docs/source/_toctree.yml

@@ -19,6 +19,8 @@
     title: Multi GPU training
   - local: peft_training
     title: Training with PEFT (e.g., LoRA)
+  - local: rename_map
+    title: Using Rename Map and Empty Cameras
   title: "Tutorials"
 - sections:
   - local: lerobot-dataset-v3

docs/source/async.mdx

@@ -48,7 +48,7 @@ python -m lerobot.async_inference.robot_client \
     --task="dummy" \ # POLICY: The task to run the policy on (`Fold my t-shirt`). Not necessarily defined for all policies, such as `act`
     --policy_type=your_policy_type \ # POLICY: the type of policy to run (smolvla, act, etc)
     --pretrained_name_or_path=user/model \ # POLICY: the model name/path on server to the checkpoint to run (e.g., lerobot/smolvla_base)
-    --policy_device=mps \ # POLICY: the device to run the policy on, on the server
+    --policy_device=mps \ # POLICY: the device to run the policy on, on the server (cuda, mps, xpu, cpu)
     --actions_per_chunk=50 \ # POLICY: the number of actions to output at once
     --chunk_size_threshold=0.5 \ # CLIENT: the threshold for the chunk size before sending a new observation to the server
     --aggregate_fn_name=weighted_average \ # CLIENT: the function to aggregate actions on overlapping portions
@@ -310,4 +310,4 @@ Asynchronous inference represents a significant advancement in real-time robotic
 - **Universal Compatibility**: Works with all LeRobot-supported policies, from lightweight ACT models to vision-language models like SmolVLA
 
 Start experimenting with the default parameters, monitor your action queue sizes, and iteratively refine your setup to achieve optimal performance for your specific use case.
-If you want to discuss this further, hop into our [Discord community](https://discord.gg/s3KuuzsPFb), or open an issue on our [GitHub repository](https://github.com/lerobot/lerobot/issues).
+If you want to discuss this further, hop into our [Discord community](https://discord.gg/s3KuuzsPFb), or open an issue on our [GitHub repository](https://github.com/huggingface/lerobot/issues).

docs/source/bring_your_own_policies.mdx

@@ -32,7 +32,7 @@ version = "0.1.0"
 dependencies = [
     # your policy-specific dependencies
 ]
-requires-python = ">= 3.11"
+requires-python = ">= 3.12"
 
 [build-system]
 build-backend = # your-build-backend
@@ -82,7 +82,7 @@ Create your policy implementation by inheriting from LeRobot's base `PreTrainedP
 # modeling_my_custom_policy.py
 import torch
 import torch.nn as nn
-from typing import Dict, Any
+from typing import Any
 
 from lerobot.policies.pretrained import PreTrainedPolicy
 from .configuration_my_custom_policy import MyCustomPolicyConfig
@@ -91,7 +91,7 @@ class MyCustomPolicy(PreTrainedPolicy):
     config_class = MyCustomPolicyConfig
     name = "my_custom_policy"
 
-    def __init__(self, config: MyCustomPolicyConfig, dataset_stats: Dict[str, Any] = None):
+    def __init__(self, config: MyCustomPolicyConfig, dataset_stats: dict[str, Any] = None):
         super().__init__(config, dataset_stats)
         ...
 ```
@@ -102,7 +102,7 @@ Create processor functions:
 
 ```python
 # processor_my_custom_policy.py
-from typing import Dict, Any
+from typing import Any
 import torch
 
 

docs/source/earthrover_mini_plus.mdx

@@ -13,7 +13,7 @@ The EarthRover Mini Plus is a fully open source mobile robot that connects throu
 ### Hardware
 
 - EarthRover Mini robot
-- Computer with Python 3.10 or newer
+- Computer with Python 3.12 or newer
 - Internet connection
 
 ### Setting Up the Frodobots SDK
@@ -170,13 +170,13 @@ Once you can drive the robot well, you can start recording data to train AI mode
 We use Hugging Face to store your data online. First, log in with your token from [Hugging Face settings](https://huggingface.co/settings/tokens):
 
 ```bash
-huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+hf auth login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 ```
 
 Store your Hugging Face username:
 
 ```bash
-HF_USER=$(huggingface-cli whoami | head -n 1)
+HF_USER=$(hf auth whoami | awk -F': *' 'NR==1 {print $2}')
 echo $HF_USER
 ```
 
@@ -204,22 +204,26 @@ Replace `your_username/dataset_name` with your Hugging Face username and a name
 
 Your dataset includes:
 
-**Your Actions (2 things)**:
+**Your Actions (2 features)**:
 
-- How much you moved forward/backward
-- How much you turned left/right
+- `linear_velocity`: How much you moved forward/backward
+- `angular_velocity`: How much you turned left/right
 
-**Robot Observations (12 things)**:
+**Robot Observations (24 features)**:
 
 - Front camera video
 - Rear camera video
 - Current speed
 - Battery level
-- Which way the robot is facing
-- GPS location (latitude, longitude, signal strength)
+- Orientation
+- GPS (latitude, longitude, signal strength)
 - Network signal strength
 - Vibration level
-- Lamp status (on/off)
+- Lamp state (on/off)
+- Accelerometer (x, y, z)
+- Gyroscope (x, y, z)
+- Magnetometer (x, y, z)
+- Wheel RPMs (4 wheels)
 
 ### Where Your Data Goes
 

docs/source/envhub.mdx

@@ -155,10 +155,10 @@ Upload your repository to Hugging Face:
 pip install huggingface_hub
 
 # Login to Hugging Face
-huggingface-cli login
+hf auth login
 
 # Create a new repository
-huggingface-cli repo create my-custom-env --type space --org my-org
+hf repo create my-org/my-custom-env
 
 # Initialize git and push
 git init

docs/source/il_robots.mdx

@@ -159,13 +159,13 @@ We use the Hugging Face hub features for uploading your dataset. If you haven't
 Add your token to the CLI by running this command:
 
 ```bash
-huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+hf auth login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 ```
 
 Then store your Hugging Face repository name in a variable:
 
 ```bash
-HF_USER=$(hf auth whoami | awk -F': *' 'NR==1 {print $2}')
+HF_USER=$(NO_COLOR=1 hf auth whoami | awk -F': *' 'NR==1 {print $2}')
 echo $HF_USER
 ```
 
@@ -327,7 +327,7 @@ You can look for other LeRobot datasets on the hub by searching for `LeRobot` [t
 You can also push your local dataset to the Hub manually, running:
 
 ```bash
-huggingface-cli upload ${HF_USER}/record-test ~/.cache/huggingface/lerobot/{repo-id} --repo-type dataset
+hf upload ${HF_USER}/record-test ~/.cache/huggingface/lerobot/{repo-id} --repo-type dataset
 ```
 
 #### Record function
@@ -491,7 +491,7 @@ If your local computer doesn't have a powerful GPU you could utilize Google Cola
 Once training is done, upload the latest checkpoint with:
 
 ```bash
-huggingface-cli upload ${HF_USER}/act_so101_test \
+hf upload ${HF_USER}/act_so101_test \
   outputs/train/act_so101_test/checkpoints/last/pretrained_model
 ```
 
@@ -499,7 +499,7 @@ You can also upload intermediate checkpoints with:
 
 ```bash
 CKPT=010000
-huggingface-cli upload ${HF_USER}/act_so101_test${CKPT} \
+hf upload ${HF_USER}/act_so101_test${CKPT} \
   outputs/train/act_so101_test/checkpoints/${CKPT}/pretrained_model
 ```
 

docs/source/installation.mdx

@@ -1,8 +1,8 @@
 # Installation
 
-This guide uses conda (via miniforge) to manage environments. If you prefer another environment manager (e.g. `uv`, `venv`), ensure you have Python >=3.10 and ffmpeg installed with the `libsvtav1` encoder, then skip ahead to [Install LeRobot](#step-3-install-lerobot-).
+This guide uses `conda` (via miniforge) to manage environments (recommended). If you prefer another environment manager (e.g. `uv`, `venv`), ensure you have Python >=3.12 and `ffmpeg` installed with the `libsvtav1` encoder, then skip ahead to [Environment Setup](#step-2-environment-setup).
 
-## Step 1: Install [`miniforge`](https://conda-forge.org/download/)
+## Step 1 (`conda` only): Install [`miniforge`](https://conda-forge.org/download/)
 
 ```bash
 wget "https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-$(uname)-$(uname -m).sh"
@@ -11,22 +11,47 @@ bash Miniforge3-$(uname)-$(uname -m).sh
 
 ## Step 2: Environment Setup
 
-Create a virtual environment with Python 3.10, using conda:
+Create a virtual environment with Python 3.12:
 
+<!-- prettier-ignore-start -->
+<hfoptions id="create_venv">
+<hfoption id="conda">
 ```bash
-conda create -y -n lerobot python=3.10
+conda create -y -n lerobot python=3.12
 ```
-
-Then activate your conda environment, you have to do this each time you open a shell to use lerobot:
-
+</hfoption>
+<hfoption id="uv">
 ```bash
+uv python install 3.12
+uv venv --python 3.12
+```
+</hfoption>
+</hfoptions>
+<!-- prettier-ignore-end -->
+
+Then activate your virtual environment, you have to do this each time you open a shell to use lerobot:
+
+<!-- prettier-ignore-start -->
+<hfoptions id="activate_venv">
+<hfoption id="conda">```bash
 conda activate lerobot
+```</hfoption>
+<hfoption id="uv">
+```bash
+# Linux/macOSsource
+source .venv/bin/activate
+# Windows PowerShell
+source .venv\Scripts\Activate.ps1
 ```
+</hfoption>
+</hfoptions>
+<!-- prettier-ignore-end -->
 
 When using `conda`, install `ffmpeg` in your environment:
 
 ```bash
 conda install ffmpeg -c conda-forge
+ffmpeg -version  # ffmpeg 8.X is not yet supported !
 ```
 
 > [!TIP]
@@ -47,6 +72,9 @@ conda install ffmpeg -c conda-forge
 > conda install evdev -c conda-forge
 > ```
 
+> [!IMPORTANT]
+> If you are using `uv` you will have to install `ffmpeg` system-wide (outside of the virtual environment). You rely on `uv` and `torchcodec` ability to dynamically link to the system `ffmpeg`.
+
 ## Step 3: Install LeRobot 🤗
 
 ### From Source
@@ -60,23 +88,45 @@ cd lerobot
 
 Then, install the library in editable mode. This is useful if you plan to contribute to the code.
 
+<!-- prettier-ignore-start -->
+<hfoptions id="install_lerobot_src">
+<hfoption id="conda">
 ```bash
 pip install -e .
 ```
+</hfoption>
+<hfoption id="uv">
+```bash
+uv pip install -e .
+```
+</hfoption>
+</hfoptions>
+<!-- prettier-ignore-end -->
 
 ### Installation from PyPI
 
 **Core Library:**
 Install the base package with:
 
+<!-- prettier-ignore-start -->
+<hfoptions id="install_lerobot_pypi">
+<hfoption id="conda">
 ```bash
 pip install lerobot
 ```
+</hfoption>
+<hfoption id="uv">
+```bash
+uv pip install lerobot
+```
+</hfoption>
+</hfoptions>
+<!-- prettier-ignore-end -->
 
 _This installs only the default dependencies._
 
 **Extra Features:**
-To install additional functionality, use one of the following:
+To install additional functionality, use one of the following (If you are using `uv`, replace `pip install` with `uv pip install` in the commands below.):
 
 ```bash
 pip install 'lerobot[all]'          # All available features
@@ -90,13 +140,10 @@ _Replace `[...]` with your desired features._
 For a full list of optional dependencies, see:
 https://pypi.org/project/lerobot/
 
-> [!NOTE]
-> For lerobot 0.4.0, if you want to install pi, you will have to do: `pip install "lerobot[pi]@git+https://github.com/huggingface/lerobot.git"`
-
 ### Troubleshooting
 
 If you encounter build errors, you may need to install additional dependencies: `cmake`, `build-essential`, and `ffmpeg libs`.
-To install these for linux run:
+To install these for Linux run:
 
 ```bash
 sudo apt-get install cmake build-essential python3-dev pkg-config libavformat-dev libavcodec-dev libavdevice-dev libavutil-dev libswscale-dev libswresample-dev libavfilter-dev
@@ -106,7 +153,7 @@ For other systems, see: [Compiling PyAV](https://pyav.org/docs/develop/overview/
 
 ## Optional dependencies
 
-LeRobot provides optional extras for specific functionalities. Multiple extras can be combined (e.g., `.[aloha,feetech]`). For all available extras, refer to `pyproject.toml`.
+LeRobot provides optional extras for specific functionalities. Multiple extras can be combined (e.g., `.[aloha,feetech]`). For all available extras, refer to `pyproject.toml`. If you are using `uv`, replace `pip install` with `uv pip install` in the commands below.
 
 ### Simulations
 

docs/source/lekiwi.mdx

@@ -279,13 +279,13 @@ We use the Hugging Face hub features for uploading your dataset. If you haven't
 Add your token to the CLI by running this command:
 
 ```bash
-huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+hf auth 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 | awk -F': *' 'NR==1 {print $2}')
 echo $HF_USER
 ```
 

docs/source/pi0.mdx

@@ -34,11 +34,6 @@ As described by Physical Intelligence, while AI has achieved remarkable success
    pip install -e ".[pi]"
    ```
 
-   > [!NOTE]
-   > For lerobot 0.4.0, if you want to install pi tag, you will have to do: `pip install "lerobot[pi]@git+https://github.com/huggingface/lerobot.git"`.
-   >
-   > This will be solved in the next patch release
-
 ## Training Data and Capabilities
 
 π₀ is trained on the largest robot interaction dataset to date, combining three key data sources:

docs/source/pi05.mdx

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

docs/source/pi0fast.mdx

@@ -43,16 +43,11 @@ This approach can transform **any existing VLM** into a VLA by training it to pr
    pip install -e ".[pi]"
    ```
 
-   > [!NOTE]
-   > For lerobot 0.4.0, if you want to install the pi tag, you will have to do: `pip install "lerobot[pi]@git+https://github.com/huggingface/lerobot.git"`.
-   >
-   > This will be solved in the next patch release
-
 ## Training a Custom FAST Tokenizer
 
 You have two options for the FAST tokenizer:
 
-1. **Use the pre-trained tokenizer**: The `physical-intelligence/fast` tokenizer was trained on 1M+ real robot action sequences and works as a general-purpose tokenizer.
+1. **Use the pre-trained tokenizer**: The `lerobot/fast-action-tokenizer` tokenizer was trained on 1M+ real robot action sequences and works as a general-purpose tokenizer.
 
 2. **Train your own tokenizer**: For maximum performance on your specific dataset, you can finetune the tokenizer on your own data.
 
@@ -114,15 +109,15 @@ lerobot-train \
 
 ### Key Training Parameters
 
-| Parameter                              | Description                                        | Default                      |
-| -------------------------------------- | -------------------------------------------------- | ---------------------------- |
-| `--policy.gradient_checkpointing=true` | Reduces memory usage significantly during training | `false`                      |
-| `--policy.dtype=bfloat16`              | Use mixed precision training for efficiency        | `float32`                    |
-| `--policy.chunk_size`                  | Number of action steps to predict (action horizon) | `50`                         |
-| `--policy.n_action_steps`              | Number of action steps to execute                  | `50`                         |
-| `--policy.max_action_tokens`           | Maximum number of FAST tokens per action chunk     | `256`                        |
-| `--policy.action_tokenizer_name`       | FAST tokenizer to use                              | `physical-intelligence/fast` |
-| `--policy.compile_model=true`          | Enable torch.compile for faster training           | `false`                      |
+| Parameter                              | Description                                        | Default                         |
+| -------------------------------------- | -------------------------------------------------- | ------------------------------- |
+| `--policy.gradient_checkpointing=true` | Reduces memory usage significantly during training | `false`                         |
+| `--policy.dtype=bfloat16`              | Use mixed precision training for efficiency        | `float32`                       |
+| `--policy.chunk_size`                  | Number of action steps to predict (action horizon) | `50`                            |
+| `--policy.n_action_steps`              | Number of action steps to execute                  | `50`                            |
+| `--policy.max_action_tokens`           | Maximum number of FAST tokens per action chunk     | `256`                           |
+| `--policy.action_tokenizer_name`       | FAST tokenizer to use                              | `lerobot/fast-action-tokenizer` |
+| `--policy.compile_model=true`          | Enable torch.compile for faster training           | `false`                         |
 
 ## Inference
 

docs/source/rename_map.mdx

@@ -0,0 +1,114 @@
+# Rename Map and Empty Cameras
+
+When you train, evaluate, or record with a robot policy, your **dataset** or **environment** provides observations under one set of keys (e.g. `observation.images.front`, `observation.images.eagle`), while your **policy** expects another (e.g. `observation.images.image`, `observation.images.image2`). The **rename map** bridges that gap without changing the policy or data source.
+
+> **Scope:** The rename map only renames **observation** keys (images and state). Action keys are not affected.
+
+## Why observation keys don't always match
+
+Policies have a fixed set of **input feature names** baked into their pretrained config. For example:
+
+- [pi0fast-libero](https://huggingface.co/lerobot/pi0fast-libero) expects `observation.images.base_0_rgb` and `observation.images.left_wrist_0_rgb`.
+- [xvla-base](https://huggingface.co/lerobot/xvla-base) expects `observation.images.image`, `observation.images.image2`, and `observation.images.image3`.
+
+Your dataset might use different names entirely (e.g. `observation.images.front`, `observation.images.eagle`, `observation.images.glove`), and your eval environment might use yet another set. Rather than editing the policy config or renaming columns in the dataset, you pass a **rename map**: a JSON dictionary that maps source keys to the keys the policy expects. Renaming happens inside the preprocessor pipeline, so the policy always sees its expected keys.
+
+## Using the rename map
+
+Pass the mapping as a JSON string on the command line. The convention is always:
+
+```
+--rename_map='{"source_key": "policy_key", ...}'
+```
+
+where **source_key** is what the dataset or environment provides, and **policy_key** is what the policy expects.
+
+Only listed keys are renamed; everything else passes through unchanged. Order of entries doesn't matter.
+
+Supported policies: **PI0**, **PI05**, **PI0Fast**, **SmolVLA**, and **XVLA**.
+
+### Training
+
+Suppose you fine-tune [lerobot/xvla-base](https://huggingface.co/lerobot/xvla-base) on a dataset with images under `observation.images.front`, `observation.images.eagle`, and `observation.images.glove`. XVLA expects `observation.images.image`, `observation.images.image2`, and `observation.images.image3`:
+
+```bash
+lerobot-train \
+  --dataset.repo_id=YOUR_DATASET \
+  --output_dir=./outputs/xvla_training \
+  --job_name=xvla_training \
+  --policy.path="lerobot/xvla-base" \
+  --policy.repo_id="HF_USER/xvla-your-robot" \
+  --policy.dtype=bfloat16 \
+  --policy.action_mode=auto \
+  --steps=20000 \
+  --policy.device=cuda \
+  --policy.freeze_vision_encoder=false \
+  --policy.freeze_language_encoder=false \
+  --policy.train_policy_transformer=true \
+  --policy.train_soft_prompts=true \
+  --rename_map='{"observation.images.front": "observation.images.image", "observation.images.eagle": "observation.images.image2", "observation.images.glove": "observation.images.image3"}'
+```
+
+### Evaluation
+
+A policy that expects `observation.images.base_0_rgb` and `observation.images.left_wrist_0_rgb` (e.g. [pi0fast-libero](https://huggingface.co/lerobot/pi0fast-libero)), but the LIBERO environment returns `observation.images.image` and `observation.images.image2`:
+
+```bash
+lerobot-eval \
+  --policy.path=lerobot/pi0fast-libero \
+  --env.type=libero \
+  ... \
+  --rename_map='{"observation.images.image": "observation.images.base_0_rgb", "observation.images.image2": "observation.images.left_wrist_0_rgb"}'
+```
+
+### Recording
+
+`lerobot-record` also supports rename maps, nested under the dataset config:
+
+```bash
+lerobot-record \ # When running inference
+  --policy.path="<user>/smolVLA_finetuned" \
+  ... \
+  --dataset.rename_map='{"observation.images.glove2": "observation.images.image"}'
+```
+
+## Alternative: edit the policy config directly
+
+If you always use the same dataset or environment, you can **edit the policy's `config.json`** so its observation keys match your data source. Then no rename map is needed.
+
+The tradeoff: modifying the policy config ties it to one data source. A rename map keeps one policy usable across many datasets and environments.
+
+## Empty cameras: fewer views than the policy expects
+
+Some policies are built for a fixed number of image inputs. If your dataset has fewer cameras, you can set **`empty_cameras`** in the policy config instead of modifying the model architecture.
+
+### How it works
+
+Setting `empty_cameras=N` adds N placeholder image features to the policy config, named:
+
+```
+observation.images.empty_camera_0
+observation.images.empty_camera_1
+...
+```
+
+At runtime, these keys have no corresponding data in the batch. The policy fills them with masked dummy tensors (padded with `-1` for SigLIP-based vision encoders, with a zero attention mask), so the extra image slots are effectively ignored during training and inference.
+
+### Example
+
+XVLA-base has three visual inputs and `empty_cameras=0` by default. Your dataset only has two cameras:
+
+1. Set `--policy.empty_cameras=1`.
+2. The config adds a third key: `observation.images.empty_camera_0`.
+3. Use the rename map for your two real cameras as usual.
+4. The third slot is masked out — no fake images needed in your dataset.
+
+## Quick reference
+
+| Goal                                      | What to do                                                                  |
+| ----------------------------------------- | --------------------------------------------------------------------------- |
+| Dataset keys ≠ policy keys                | `--rename_map='{"dataset_key": "policy_key", ...}'`                         |
+| Env keys ≠ policy keys (eval)             | `--rename_map='{"env_key": "policy_key", ...}'`                             |
+| Recording with different keys (inference) | `--dataset.rename_map='{"source_key": "policy_key", ...}'`.                 |
+| Fewer cameras than policy expects         | `--policy.empty_cameras=N` (supported by PI0, PI05, PI0Fast, SmolVLA, XVLA) |
+| Avoid passing a rename map                | Edit the policy's `config.json` so its keys match your data source          |

docs/source/unitree_g1.mdx

@@ -1,23 +1,72 @@
 # Unitree G1
 
-This guide covers the complete setup process for the Unitree G1 humanoid, from initial connection to running gr00t_wbc locomotion.
+<img
+  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/unitree_thumbnail.jpg"
+  alt="Unitree G1 locomanipulation demo"
+  style={{ width: "100%" }}
+/>
 
-## About
-
-We support both 29 and 23 DOF G1 EDU version. We introduce:
-
-- **`unitree g1` robot class, handling low level read/write from/to the humanoid**
-- **ZMQ socket bridge** for remote communication and camera streaming, allowing for remote policy deployment over wlan, eth or directly on the robot
-- **Locomotion policies** from NVIDIA gr00t and Amazon FAR Holosoma
-- **Simulation mode** for testing policies without the physical robot in mujoco
+The Unitree G1 humanoid is now supported in LeRobot! You can teleoperate, train locomanipulation policies, test in sim, and more. Both 29 and 23 DoF variants are supported.
 
 ---
 
-## Connection guide
+## Part 1: Getting Started
 
-### Step 1: Configure Ethernet Interface
+### Install the Unitree SDK
 
-Set a static IP on the same subnet as the robot:
+Follow the [unitree_sdk2_python installation guide](https://github.com/unitreerobotics/unitree_sdk2_python#installation). Tested with `unitree_sdk2py==1.0.1` and `cyclonedds==0.10.2`:
+
+```bash
+conda create -y -n lerobot python=3.12
+conda activate lerobot
+git clone https://github.com/unitreerobotics/unitree_sdk2_python.git
+cd unitree_sdk2_python
+pip install -e .
+cd ..
+```
+
+### Install LeRobot
+
+```bash
+conda install ffmpeg -c conda-forge
+conda install -c conda-forge "pinocchio>=3.0.0,<4.0.0"
+git clone https://github.com/huggingface/lerobot.git
+cd lerobot
+pip install -e '.[unitree_g1]'
+```
+
+<Tip>
+  For now, pinocchio must be installed from conda-forge (not pip) to include the
+  CasADi bindings needed for arm IK.
+</Tip>
+
+### Test the Installation (Simulation)
+
+The simulation environment has its own dependencies. Check the Simulation environment dependencies: [Unitree G1 Mujoco EnvHub](https://huggingface.co/lerobot/unitree-g1-mujoco/tree/main).
+
+```bash
+pip install mujoco loguru msgpack msgpack-numpy
+```
+
+```bash
+lerobot-teleoperate \
+  --robot.type=unitree_g1 \
+  --robot.is_simulation=true \
+  --teleop.type=unitree_g1 \
+  --teleop.id=wbc_unitree \
+  --robot.cameras='{"global_view": {"type": "zmq", "server_address": "localhost", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30, "warmup_s": 5}}' \
+  --display_data=true \
+  --robot.controller=GrootLocomotionController
+```
+
+This will launch a [MuJoCo sim instance](https://huggingface.co/lerobot/unitree-g1-mujoco/tree/main) for the G1. You can connect a gamepad to your machine before launching in order to control the robot's locomotion in sim. We support both [HolosomaLocomotionController](https://github.com/amazon-far/holosoma) and [GrootLocomotionController](https://github.com/NVlabs/GR00T-WholeBodyControl) via `--robot.controller`.
+
+- Press `9` to release the robot
+- Press `7` / `8` to increase / decrease waist height
+
+### Connect to the Physical Robot
+
+The G1's Ethernet IP is fixed at `192.168.123.164`. Your machine must have a static IP on the same subnet: `192.168.123.x` where `x ≠ 164`.
 
 ```bash
 # Replace 'enp131s0' with your ethernet interface name (check with `ip a`)
@@ -26,47 +75,23 @@ sudo ip addr add 192.168.123.200/24 dev enp131s0
 sudo ip link set enp131s0 up
 ```
 
-**Note**: The G1's Ethernet IP is fixed at `192.168.123.164`. Your computer must use `192.168.123.x` with x ≠ 164.
-
-### Step 2: SSH into the Robot
+### SSH into the Robot
 
 ```bash
 ssh unitree@192.168.123.164
 # Password: 123
 ```
 
-You should now be connected to the G1's Orin.
+### Share Internet via Ethernet
 
----
-
-## Part 2: Enable WiFi on the Robot
-
-Wlan0 is disabled by default on the G1. To enable it:
-
-### Step 1: Enable WiFi Hardware
-
-```bash
-sudo rfkill unblock wifi
-sudo rfkill unblock all
-
-# Bring up wlan0
-sudo ip link set wlan0 up
-
-# Enable NetworkManager control of wlan0
-sudo nmcli radio wifi on
-sudo nmcli device set wlan0 managed yes
-sudo systemctl restart NetworkManager
-```
-
-### Step 2: Enable Internet Forwarding
+The G1 needs internet access to clone repos and install packages. Share your laptop's connection over Ethernet:
 
 **On your laptop:**
 
 ```bash
-# Enable IP forwarding
 sudo sysctl -w net.ipv4.ip_forward=1
 
-# Set up NAT (replace wlp132s0f0 with your WiFi interface)
+# Replace wlp132s0f0 with your WiFi interface name
 sudo iptables -t nat -A POSTROUTING -o wlp132s0f0 -s 192.168.123.0/24 -j MASQUERADE
 sudo iptables -A FORWARD -i wlp132s0f0 -o enp131s0 -m state --state RELATED,ESTABLISHED -j ACCEPT
 sudo iptables -A FORWARD -i enp131s0 -o wlp132s0f0 -j ACCEPT
@@ -75,223 +100,193 @@ sudo iptables -A FORWARD -i enp131s0 -o wlp132s0f0 -j ACCEPT
 **On the G1:**
 
 ```bash
-# Add laptop as default gateway
 sudo ip route del default 2>/dev/null || true
 sudo ip route add default via 192.168.123.200 dev eth0
 echo "nameserver 8.8.8.8" | sudo tee /etc/resolv.conf
 
-# Test connection
+# Verify
 ping -c 3 8.8.8.8
 ```
 
-### Step 3: Connect to WiFi Network
+### Install the Unitree SDK on the G1
+
+Follow the [unitree_sdk2_python installation guide](https://github.com/unitreerobotics/unitree_sdk2_python#installation):
+
+```bash
+conda create -y -n lerobot python=3.12
+conda activate lerobot
+git clone https://github.com/unitreerobotics/unitree_sdk2_python.git
+cd unitree_sdk2_python
+python -m pip install -e .
+cd ..
+```
+
+### Install LeRobot on the G1
+
+```bash
+git clone https://github.com/huggingface/lerobot.git
+cd lerobot
+conda install -c conda-forge "pinocchio>=3.0.0,<4.0.0"
+python -m pip install -e '.[unitree_g1]'
+```
+
+<Tip>
+  For now, pinocchio must be installed from conda-forge (not pip) to include the
+  CasADi bindings needed for arm IK.
+</Tip>
+
+### (Optional) Enable WiFi on the Robot
+
+For wireless SSH access, you can enable WiFi on the G1 (it's blocked by default):
+
+```bash
+sudo rfkill unblock all
+sudo ip link set wlan0 up
+sudo nmcli radio wifi on
+sudo nmcli device set wlan0 managed yes
+sudo systemctl restart NetworkManager
+```
+
+**Connect to a WiFi network:**
 
 ```bash
-# List available networks
 nmcli device wifi list
 
-# Connect to your WiFi (example)
 sudo nmcli connection add type wifi ifname wlan0 con-name "YourNetwork" ssid "YourNetwork"
 sudo nmcli connection modify "YourNetwork" wifi-sec.key-mgmt wpa-psk
 sudo nmcli connection modify "YourNetwork" wifi-sec.psk "YourPassword"
 sudo nmcli connection modify "YourNetwork" connection.autoconnect yes
 sudo nmcli connection up "YourNetwork"
 
-# Check WiFi IP address
 ip a show wlan0
 ```
 
-### Step 4: SSH Over WiFi
-
-Once connected to WiFi, note the robot's IP address and disconnect the Ethernet cable. You can now SSH over WiFi:
+You can then SSH over WiFi instead of Ethernet:
 
 ```bash
-ssh unitree@<YOUR_ROBOT_IP>
+ssh unitree@<ROBOT_WIFI_IP>
 # Password: 123
 ```
 
-Replace `<YOUR_ROBOT_IP>` with your robot's actual WiFi IP address.
+---
+
+## Part 2: Teleoperation & Locomotion
+
+### Run the Robot Server
+
+On the robot (from `~/lerobot`):
+
+```bash
+cd ~/lerobot
+python src/lerobot/robots/unitree_g1/run_g1_server.py --camera
+```
+
+### Run the Locomotion Policy
+
+You can run the teleoperation client from your laptop over Ethernet, over WiFi (experimental), or directly on the robot itself. Mind potential latency introduced by your network.
+
+**From your laptop:**
+
+```bash
+lerobot-teleoperate \
+  --robot.type=unitree_g1 \
+  --robot.is_simulation=false \
+  --robot.robot_ip=<ROBOT_IP> \
+  --teleop.type=unitree_g1 \
+  --teleop.id=wbc_unitree \
+  --robot.cameras='{"global_view": {"type": "zmq", "server_address": "<ROBOT_IP>", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
+  --display_data=true \
+  --robot.controller=HolosomaLocomotionController
+```
+
+We support both [GrootLocomotionController](https://github.com/NVlabs/GR00T-WholeBodyControl) and [HolosomaLocomotionController](https://github.com/amazon-far/holosoma) via `--robot.controller`.
 
 ---
 
-## Part 3: Robot Server Setup
+## Part 3: Loco-Manipulation with the Homunculus Exoskeleton
 
-### Step 1: Install LeRobot on the Orin
+We provide a loco-manipulation solution via the Homunculus Exoskeleton — an open-source 7 DoF exoskeleton for whole-body control. Check it out [here](https://github.com/nepyope/hmc_exo).
 
-SSH into the robot and install LeRobot:
-
-```bash
-ssh unitree@<YOUR_ROBOT_IP>
-
-conda create -y -n lerobot python=3.10
-conda activate lerobot
-git clone https://github.com/huggingface/lerobot.git
-cd lerobot
-pip install -e '.[unitree_g1]'
-git clone https://github.com/unitreerobotics/unitree_sdk2_python.git
-cd unitree_sdk2_python  && pip install -e .
-```
-
-**Note**: The Unitree SDK requires CycloneDDS v0.10.2 to be installed. See the [Unitree SDK documentation](https://github.com/unitreerobotics/unitree_sdk2_python) for details.
-
-### Step 2: Run the Robot Server
-
-On the robot:
-
-```bash
-python src/lerobot/robots/unitree_g1/run_g1_server.py
-```
-
-**Important**: Keep this terminal running. The server must be active for remote control.
-
----
-
-## Part 4: Controlling the robot
-
-With the robot server running, you can now control the robot remotely. Let's launch a locomotion policy
-
-### Step 1: Install LeRobot on your machine
-
-```bash
-conda create -y -n lerobot python=3.10
-conda activate lerobot
-git clone https://github.com/huggingface/lerobot.git
-cd lerobot
-pip install -e '.[unitree_g1]'
-git clone https://github.com/unitreerobotics/unitree_sdk2_python.git
-cd unitree_sdk2_python  && pip install -e .
-```
-
-### Step 2: Update Robot IP in Config
-
-Edit the config file to match your robot's WiFi IP:
-
-```python
-# In src/lerobot/robots/unitree_g1/config_unitree_g1.py
-robot_ip: str = "<YOUR_ROBOT_IP>"  # Replace with your robot's WiFi IP.
-```
-
-### Step 3: Run the Locomotion Policy
-
-```bash
-# Run GR00T locomotion controller
-python examples/unitree_g1/gr00t_locomotion.py --repo-id "nepyope/GR00T-WholeBodyControl_g1"
-
-# Run Holosoma locomotion controller
-python examples/unitree_g1/holosoma_locomotion.py
-
-```
-
-Press `Ctrl+C` to stop the policy.
-
----
-
-## Running in Simulation Mode (MuJoCo)
-
-You can test policies before deploying on the physical robot using MuJoCo simulation. Set `is_simulation=True` in config or pass `--robot.is_simulation=true` via CLI.
-
-### Calibrate Exoskeleton Teleoperator
+### Calibrate
 
 ```bash
 lerobot-calibrate \
-    --teleop.type=unitree_g1 \
-    --teleop.left_arm_config.port=/dev/ttyACM1 \
-    --teleop.right_arm_config.port=/dev/ttyACM0 \
-    --teleop.id=exo
+  --teleop.type=unitree_g1 \
+  --teleop.left_arm_config.port=/dev/ttyACM1 \
+  --teleop.right_arm_config.port=/dev/ttyACM0 \
+  --teleop.id=exo
 ```
 
-### Teleoperate in Simulation
+During calibration move each joint through its entire range. After fitting, move the joint in a neutral position and press `n` to advance.
 
-```bash
-lerobot-teleoperate \
-    --robot.type=unitree_g1 \
-    --robot.is_simulation=true \
-    --teleop.type=unitree_g1 \
-    --teleop.left_arm_config.port=/dev/ttyACM1 \
-    --teleop.right_arm_config.port=/dev/ttyACM0 \
-    --teleop.id=exo \
-    --fps=100
-```
-
-### Record Dataset in Simulation
+### Record a Dataset
 
 ```bash
 lerobot-record \
-    --robot.type=unitree_g1 \
-    --robot.is_simulation=true \
-    --robot.cameras='{"global_view": {"type": "zmq", "server_address": "localhost", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
-    --teleop.type=unitree_g1 \
-    --teleop.left_arm_config.port=/dev/ttyACM1 \
-    --teleop.right_arm_config.port=/dev/ttyACM0 \
-    --teleop.id=exo \
-    --dataset.repo_id=your-username/dataset-name \
-    --dataset.single_task="Test" \
-    --dataset.num_episodes=2 \
-    --dataset.episode_time_s=5 \
-    --dataset.reset_time_s=5 \
-    --dataset.push_to_hub=true \
-    --dataset.streaming_encoding=true \
-    # --dataset.vcodec=auto \
-    --dataset.encoder_threads=2
+  --robot.type=unitree_g1 \
+  --robot.is_simulation=true \
+  --robot.cameras='{"global_view": {"type": "zmq", "server_address": "localhost", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
+  --teleop.type=unitree_g1 \
+  --teleop.left_arm_config.port=/dev/ttyACM1 \
+  --teleop.right_arm_config.port=/dev/ttyACM0 \
+  --teleop.id=exo \
+  --dataset.repo_id=your-username/dataset-name \
+  --dataset.single_task="Test" \
+  --dataset.num_episodes=2 \
+  --dataset.episode_time_s=5 \
+  --dataset.reset_time_s=5 \
+  --dataset.push_to_hub=true \
+  --dataset.streaming_encoding=true \
+  --dataset.encoder_threads=2
 ```
 
-Example simulation dataset: [nepyope/teleop_test_sim](https://huggingface.co/datasets/nepyope/teleop_test_sim)
+> **Note:** Omit `--teleop.left_arm_config.port` and `--teleop.right_arm_config.port` if you're only using the joystick.
+
+Example dataset: [nepyope/unitree_box_move_blue_full](https://huggingface.co/datasets/nepyope/unitree_box_move_blue_full)
 
 ---
 
-## Running on Real Robot
+## Part 4: Training & Inference
 
-Once the robot server is running on the G1 (see Part 3), you can teleoperate and record on the real robot.
-
-### Start the Camera Server
-
-On the robot, start the ZMQ image server:
+### Train
 
 ```bash
-python src/lerobot/cameras/zmq/image_server.py
+python src/lerobot/scripts/lerobot_train.py \
+  --dataset.repo_id=your-username/dataset-name  \
+  --policy.type=pi05 \
+  --output_dir=./outputs/pi05_training \
+  --job_name=pi05_training \
+  --policy.repo_id=your-username/your-repo-id \
+  --policy.pretrained_path=lerobot/pi05_base \
+  --policy.compile_model=true \
+  --policy.gradient_checkpointing=true \
+  --wandb.enable=true \
+  --policy.dtype=bfloat16 \
+  --policy.freeze_vision_encoder=false \
+  --policy.train_expert_only=false \
+  --steps=3000 \
+  --policy.device=cuda \
+  --batch_size=32
 ```
 
-Keep this running in a separate terminal for camera streaming during recording.
+### Inference with RTC
 
-### Teleoperate Real Robot
+Once trained, we recommend deploying policies using inference-time RTC:
 
 ```bash
-lerobot-teleoperate \
-    --robot.type=unitree_g1 \
-    --robot.is_simulation=false \
-    --teleop.type=unitree_g1 \
-    --teleop.left_arm_config.port=/dev/ttyACM1 \
-    --teleop.right_arm_config.port=/dev/ttyACM0 \
-    --teleop.id=exo \
-    --fps=100
+python examples/rtc/eval_with_real_robot.py \
+  --policy.path=your-username/your-repo-id \
+  --policy.device=cuda \
+  --robot.type=unitree_g1 \
+  --robot.is_simulation=false \
+  --robot.controller=HolosomaLocomotionController \
+  --robot.cameras='{"global_view": {"type": "zmq", "server_address": "<ROBOT_IP>", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
+  --task="task_description" \
+  --duration=1000 \
+  --fps=30 \
+  --rtc.enabled=true
 ```
 
-### Record Dataset on Real Robot
-
-```bash
-lerobot-record \
-    --robot.type=unitree_g1 \
-    --robot.is_simulation=false \
-    --robot.cameras='{"global_view": {"type": "zmq", "server_address": "172.18.129.215", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
-    --teleop.type=unitree_g1 \
-    --teleop.left_arm_config.port=/dev/ttyACM1 \
-    --teleop.right_arm_config.port=/dev/ttyACM0 \
-    --teleop.id=exo \
-    --dataset.repo_id=your-username/dataset-name \
-    --dataset.single_task="Test" \
-    --dataset.num_episodes=2 \
-    --dataset.episode_time_s=5 \
-    --dataset.reset_time_s=5 \
-    --dataset.push_to_hub=true \
-    --dataset.streaming_encoding=true \
-    # --dataset.vcodec=auto \
-    --dataset.encoder_threads=2
-```
-
-**Note**: Update `server_address` to match your robot's camera server IP.
-
-Example real robot dataset: [nepyope/teleop_test_real](https://huggingface.co/datasets/nepyope/teleop_test_real)
-
 ---
 
 ## Additional Resources
@@ -300,8 +295,8 @@ Example real robot dataset: [nepyope/teleop_test_real](https://huggingface.co/da
 - [GR00T-WholeBodyControl](https://github.com/NVlabs/GR00T-WholeBodyControl)
 - [Holosoma](https://github.com/amazon-far/holosoma)
 - [LeRobot Documentation](https://github.com/huggingface/lerobot)
-- [Unitree_IL_Lerobot](https://github.com/unitreerobotics/unitree_IL_lerobot)
+- [Unitree IL LeRobot](https://github.com/unitreerobotics/unitree_IL_lerobot)
 
 ---
 
-_Last updated: December 2025_
+_Last updated: March 2026_

examples/dataset/load_lerobot_dataset.py

@@ -32,7 +32,8 @@ import torch
 from huggingface_hub import HfApi
 
 import lerobot
-from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
 
 
 def main():

examples/dataset/slurm_compute_rabc.py

@@ -0,0 +1,490 @@
+#!/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.
+
+"""
+SLURM-distributed SARM RA-BC annotation pipeline.
+
+Computes SARM progress values for all frames in a dataset, distributed across
+SLURM workers, then merges the shards into a single sarm_progress.parquet.
+
+Two subcommands, each a separate SLURM submission:
+
+  compute    – N workers, each computes progress for a subset of episodes
+  aggregate  – 1 worker, merges N shards into sarm_progress.parquet, pushes to hub
+
+Usage:
+    python slurm_compute_rabc.py compute \\
+        --repo-id user/dataset --reward-model-path user/sarm_model \\
+        --stride 10 --device cpu --workers 50 --partition cpu
+
+    python slurm_compute_rabc.py aggregate \\
+        --repo-id user/dataset --reward-model-path user/sarm_model \\
+        --partition cpu --push-to-hub
+"""
+
+import argparse
+from pathlib import Path
+
+from datatrove.executor import LocalPipelineExecutor
+from datatrove.executor.slurm import SlurmPipelineExecutor
+from datatrove.pipeline.base import PipelineStep
+
+
+class ComputeProgressShards(PipelineStep):
+    """Each worker computes SARM progress for its assigned episodes."""
+
+    def __init__(
+        self, repo_id, reward_model_path, stride=1, head_mode="sparse", device="cpu", shard_dir="rabc_shards"
+    ):
+        super().__init__()
+        if stride < 1:
+            raise ValueError(f"stride must be >= 1, got {stride}")
+        self.repo_id = repo_id
+        self.reward_model_path = reward_model_path
+        self.stride = stride
+        self.head_mode = head_mode
+        self.device = device
+        self.shard_dir = shard_dir
+
+    def run(self, data=None, rank: int = 0, world_size: int = 1):
+        import logging
+        from pathlib import Path
+
+        import numpy as np
+        import pyarrow as pa
+        import pyarrow.parquet as pq
+        import torch
+        from tqdm import tqdm
+
+        from lerobot.policies.sarm.compute_rabc_weights import (
+            generate_all_frame_indices,
+            interpolate_progress,
+            load_sarm_resources,
+        )
+        from lerobot.utils.utils import init_logging
+
+        init_logging()
+
+        dataset, reward_model, preprocess = load_sarm_resources(
+            self.repo_id,
+            self.reward_model_path,
+            self.device,
+        )
+
+        if hasattr(preprocess, "eval"):
+            preprocess.eval()
+        for step in preprocess.steps:
+            if hasattr(step, "eval"):
+                step.eval()
+
+        image_key = reward_model.config.image_key
+        state_key = reward_model.config.state_key
+        frame_gap = reward_model.config.frame_gap
+        center_idx = reward_model.config.n_obs_steps // 2
+
+        dual_mode = reward_model.config.uses_dual_heads
+        compute_sparse = self.head_mode in ("sparse", "both") or not dual_mode
+        compute_dense = self.head_mode in ("dense", "both") and dual_mode
+
+        my_episodes = list(range(dataset.num_episodes))[rank::world_size]
+        if not my_episodes:
+            logging.info(f"Rank {rank}: no episodes assigned")
+            return
+        logging.info(f"Rank {rank}: {len(my_episodes)} / {dataset.num_episodes} episodes")
+
+        all_rows = []
+
+        for ep_idx in tqdm(my_episodes, desc=f"Rank {rank}"):
+            ep = dataset.meta.episodes[ep_idx]
+            ep_start, ep_end = ep["dataset_from_index"], ep["dataset_to_index"]
+            task = dataset[ep_start].get("task", "perform the task")
+
+            all_ep_indices = generate_all_frame_indices(ep_start, ep_end, frame_gap)
+            if self.stride > 1:
+                compute_indices = [i for i in all_ep_indices if (i - ep_start) % self.stride == 0]
+                if (ep_end - 1) not in compute_indices:
+                    compute_indices.append(ep_end - 1)
+                compute_indices = sorted(set(compute_indices))
+            else:
+                compute_indices = all_ep_indices
+
+            frame_results = {}
+            for qi in tqdm(compute_indices, desc=f"  Ep {ep_idx}", leave=False):
+                try:
+                    sample = dataset[qi]
+                    batch = {
+                        image_key: sample[image_key],
+                        "task": task,
+                        "index": qi,
+                        "episode_index": ep_idx,
+                    }
+                    if state_key in sample:
+                        batch[state_key] = sample[state_key]
+
+                    with torch.no_grad():
+                        processed = preprocess(batch)
+                        vf = processed["video_features"].to(self.device)
+                        tf = processed["text_features"].to(self.device)
+                        sf = processed.get("state_features")
+                        if sf is not None:
+                            sf = sf.to(self.device)
+                        lengths = processed.get("lengths")
+
+                        sparse_val = dense_val = np.nan
+                        if compute_sparse:
+                            r = reward_model.calculate_rewards(
+                                text_embeddings=tf,
+                                video_embeddings=vf,
+                                state_features=sf,
+                                lengths=lengths,
+                                return_all_frames=True,
+                                head_mode="sparse",
+                            )
+                            sparse_val = float(r[0, center_idx] if r.ndim == 2 else r[center_idx])
+                        if compute_dense:
+                            r = reward_model.calculate_rewards(
+                                text_embeddings=tf,
+                                video_embeddings=vf,
+                                state_features=sf,
+                                lengths=lengths,
+                                return_all_frames=True,
+                                head_mode="dense",
+                            )
+                            dense_val = float(r[0, center_idx] if r.ndim == 2 else r[center_idx])
+
+                        frame_results[qi] = (sparse_val, dense_val)
+                except Exception as e:
+                    logging.warning(f"Failed frame {qi}: {e}")
+
+            if not frame_results:
+                logging.warning(f"Episode {ep_idx}: all frames failed, skipping")
+                continue
+
+            # Interpolate to all frames in this episode
+            computed_idx = np.array(sorted(frame_results.keys()))
+            all_frame_arr = np.arange(ep_start, ep_end)
+
+            sparse_vals = np.array([frame_results[i][0] for i in computed_idx]) if compute_sparse else None
+            dense_vals = np.array([frame_results[i][1] for i in computed_idx]) if compute_dense else None
+
+            if self.stride > 1 and len(computed_idx) > 1:
+                if compute_sparse:
+                    sparse_vals = interpolate_progress(computed_idx, sparse_vals, all_frame_arr)
+                if compute_dense:
+                    dense_vals = interpolate_progress(computed_idx, dense_vals, all_frame_arr)
+                output_frames = all_frame_arr
+            else:
+                # Use only successfully computed frames to avoid indexing mismatch on failures
+                output_frames = computed_idx
+
+            for i, fi in enumerate(output_frames):
+                row = {"index": int(fi), "episode_index": ep_idx, "frame_index": int(fi - ep_start)}
+                if compute_sparse:
+                    row["progress_sparse"] = float(sparse_vals[i])
+                if compute_dense:
+                    row["progress_dense"] = float(dense_vals[i])
+                all_rows.append(row)
+
+        if all_rows:
+            import pandas as pd
+
+            df = pd.DataFrame(all_rows).sort_values("index").reset_index(drop=True)
+            table = pa.Table.from_pandas(df, preserve_index=False)
+            table = table.replace_schema_metadata({b"reward_model_path": self.reward_model_path.encode()})
+            shard_dir = Path(self.shard_dir)
+            shard_dir.mkdir(parents=True, exist_ok=True)
+            out = shard_dir / f"shard_{rank:05d}.parquet"
+            pq.write_table(table, out)
+            logging.info(f"Rank {rank}: saved {len(df)} rows to {out}")
+
+
+class AggregateProgress(PipelineStep):
+    """Merge all shard parquets into final sarm_progress.parquet."""
+
+    def __init__(self, repo_id, reward_model_path, shard_dir="rabc_shards", push_to_hub=False):
+        super().__init__()
+        self.repo_id = repo_id
+        self.reward_model_path = reward_model_path
+        self.shard_dir = shard_dir
+        self.push_to_hub = push_to_hub
+
+    def run(self, data=None, rank: int = 0, world_size: int = 1):
+        import datetime
+        import logging
+        import os
+        from pathlib import Path
+
+        import pandas as pd
+        import pyarrow as pa
+        import pyarrow.parquet as pq
+
+        from lerobot.datasets.lerobot_dataset import LeRobotDataset
+        from lerobot.utils.utils import init_logging
+
+        init_logging()
+        if rank != 0:
+            return
+
+        shard_dir = Path(self.shard_dir)
+        shards = sorted(shard_dir.glob("shard_*.parquet"))
+        if not shards:
+            raise FileNotFoundError(f"No shards found in {shard_dir}")
+
+        # Log shard modification time range to help detect stale files
+        mtimes = [os.path.getmtime(s) for s in shards]
+        oldest = datetime.datetime.fromtimestamp(min(mtimes)).isoformat(timespec="seconds")
+        newest = datetime.datetime.fromtimestamp(max(mtimes)).isoformat(timespec="seconds")
+        logging.info(f"Aggregating {len(shards)} shards (oldest: {oldest}, newest: {newest})")
+
+        df = pd.concat([pd.read_parquet(s) for s in shards], ignore_index=True)
+        df = df.sort_values("index").reset_index(drop=True)
+
+        table = pa.Table.from_pandas(df, preserve_index=False)
+        table = table.replace_schema_metadata({b"reward_model_path": self.reward_model_path.encode()})
+
+        temp_ds = LeRobotDataset(self.repo_id, download_videos=False)
+        out_path = Path(temp_ds.root) / "sarm_progress.parquet"
+        out_path.parent.mkdir(parents=True, exist_ok=True)
+        pq.write_table(table, out_path)
+        logging.info(f"Saved {len(df)} rows to {out_path}")
+
+        for col in ["progress_sparse", "progress_dense"]:
+            if col in df.columns:
+                v = df[col].dropna()
+                logging.info(
+                    f"{col}: mean={v.mean():.4f} std={v.std():.4f} min={v.min():.4f} max={v.max():.4f}"
+                )
+
+        if self.push_to_hub:
+            from huggingface_hub import HfApi
+
+            api = HfApi()
+            hub_path = "sarm_progress.parquet"
+            logging.info(f"Uploading to {self.repo_id}/{hub_path}")
+            api.upload_file(
+                path_or_fileobj=str(out_path),
+                path_in_repo=hub_path,
+                repo_id=self.repo_id,
+                repo_type="dataset",
+            )
+            logging.info(f"Uploaded: https://huggingface.co/datasets/{self.repo_id}/blob/main/{hub_path}")
+
+
+def make_compute_executor(
+    repo_id,
+    reward_model_path,
+    stride,
+    head_mode,
+    device,
+    shard_dir,
+    logs_dir,
+    job_name,
+    slurm,
+    workers,
+    partition,
+    cpus_per_task,
+    mem_per_cpu,
+):
+    kwargs = {
+        "pipeline": [
+            ComputeProgressShards(repo_id, reward_model_path, stride, head_mode, device, str(shard_dir)),
+        ],
+        "logging_dir": str(logs_dir / job_name),
+    }
+
+    if slurm:
+        kwargs.update(
+            {
+                "job_name": job_name,
+                "tasks": workers,
+                "workers": workers,
+                "time": "24:00:00",
+                "partition": partition,
+                "cpus_per_task": cpus_per_task,
+                "sbatch_args": {"mem-per-cpu": mem_per_cpu},
+            }
+        )
+        return SlurmPipelineExecutor(**kwargs)
+
+    kwargs.update({"tasks": workers, "workers": 1})
+    return LocalPipelineExecutor(**kwargs)
+
+
+def make_aggregate_executor(
+    repo_id,
+    reward_model_path,
+    shard_dir,
+    logs_dir,
+    job_name,
+    slurm,
+    partition,
+    cpus_per_task,
+    mem_per_cpu,
+    push_to_hub,
+):
+    kwargs = {
+        "pipeline": [
+            AggregateProgress(repo_id, reward_model_path, str(shard_dir), push_to_hub),
+        ],
+        "logging_dir": str(logs_dir / job_name),
+    }
+
+    if slurm:
+        kwargs.update(
+            {
+                "job_name": job_name,
+                "tasks": 1,
+                "workers": 1,
+                "time": "02:00:00",
+                "partition": partition,
+                "cpus_per_task": cpus_per_task,
+                "sbatch_args": {"mem-per-cpu": mem_per_cpu},
+            }
+        )
+        return SlurmPipelineExecutor(**kwargs)
+
+    kwargs.update({"tasks": 1, "workers": 1})
+    return LocalPipelineExecutor(**kwargs)
+
+
+def _add_shared_args(p):
+    p.add_argument(
+        "--repo-id",
+        type=str,
+        required=True,
+        help="Hugging Face repository identifier, e.g. 'user/dataset'.",
+    )
+    p.add_argument(
+        "--shard-dir",
+        type=Path,
+        default=Path("rabc_shards"),
+        help="Directory to read/write per-rank parquet shards.",
+    )
+    p.add_argument(
+        "--logs-dir",
+        type=Path,
+        default=Path("logs"),
+        help="Directory for datatrove logs.",
+    )
+    p.add_argument(
+        "--job-name",
+        type=str,
+        default=None,
+        help="SLURM job name (defaults to rabc_<subcommand>).",
+    )
+    p.add_argument(
+        "--slurm",
+        type=int,
+        default=1,
+        help="1 = submit via SLURM; 0 = run locally (useful for debugging).",
+    )
+    p.add_argument(
+        "--partition",
+        type=str,
+        default=None,
+        help="SLURM partition to submit to.",
+    )
+    p.add_argument(
+        "--cpus-per-task",
+        type=int,
+        default=4,
+        help="Number of CPUs per SLURM task.",
+    )
+    p.add_argument(
+        "--mem-per-cpu",
+        type=str,
+        default="4G",
+        help="Memory per CPU, e.g. '4G' or '1950M'.",
+    )
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="SLURM-distributed SARM RA-BC annotation pipeline",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+    )
+    sub = parser.add_subparsers(dest="command", required=True)
+
+    # compute subcommand
+    cp = sub.add_parser(
+        "compute",
+        help="Distribute progress computation across SLURM workers.",
+    )
+    _add_shared_args(cp)
+    cp.add_argument(
+        "--reward-model-path",
+        type=str,
+        required=True,
+        help="Path or HF repo id of the SARM reward model.",
+    )
+    cp.add_argument(
+        "--stride",
+        type=int,
+        default=1,
+        help="Compute every Nth frame; intermediate frames are interpolated (must be >= 1).",
+    )
+    cp.add_argument(
+        "--head-mode",
+        type=str,
+        default="sparse",
+        choices=["sparse", "dense", "both"],
+        help="Which reward head(s) to compute.",
+    )
+    cp.add_argument(
+        "--device",
+        type=str,
+        default="cpu",
+        help="Device for reward model inference, e.g. 'cpu' or 'cuda'.",
+    )
+    cp.add_argument(
+        "--workers",
+        type=int,
+        default=50,
+        help="Number of parallel SLURM tasks (one shard per worker).",
+    )
+
+    # aggregate subcommand
+    ap = sub.add_parser(
+        "aggregate",
+        help="Merge per-rank shards into a single sarm_progress.parquet.",
+    )
+    _add_shared_args(ap)
+    ap.add_argument(
+        "--reward-model-path",
+        type=str,
+        required=True,
+        help="Path or HF repo id of the SARM reward model (stored in parquet metadata).",
+    )
+    ap.add_argument(
+        "--push-to-hub",
+        action="store_true",
+        help="Upload sarm_progress.parquet to the Hugging Face Hub after aggregation.",
+    )
+
+    args = parser.parse_args()
+    job_name = args.job_name or f"rabc_{args.command}"
+    kwargs = vars(args)
+    kwargs["slurm"] = kwargs.pop("slurm") == 1
+    kwargs["job_name"] = job_name
+    command = kwargs.pop("command")
+
+    executor = make_compute_executor(**kwargs) if command == "compute" else make_aggregate_executor(**kwargs)
+
+    executor.run()
+
+
+if __name__ == "__main__":
+    main()

examples/lekiwi/evaluate.py

@@ -14,8 +14,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from lerobot.datasets.feature_utils import hw_to_dataset_features
 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.processor import make_default_processors

examples/lekiwi/record.py

@@ -14,8 +14,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from lerobot.datasets.feature_utils import hw_to_dataset_features
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.utils import hw_to_dataset_features
 from lerobot.processor import make_default_processors
 from lerobot.robots.lekiwi.config_lekiwi import LeKiwiClientConfig
 from lerobot.robots.lekiwi.lekiwi_client import LeKiwiClient

examples/phone_to_so100/evaluate.py

@@ -16,15 +16,13 @@
 
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
 from lerobot.configs.types import FeatureType, PolicyFeature
+from lerobot.datasets.feature_utils import combine_feature_dicts
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
-from lerobot.datasets.utils import combine_feature_dicts
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.policies.act.modeling_act import ACTPolicy
 from lerobot.policies.factory import make_pre_post_processors
 from lerobot.processor import (
-    RobotAction,
-    RobotObservation,
     RobotProcessorPipeline,
     make_default_teleop_action_processor,
 )
@@ -40,6 +38,7 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
     InverseKinematicsEEToJoints,
 )
 from lerobot.scripts.lerobot_record import record_loop
+from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun

examples/phone_to_so100/record.py

@@ -15,11 +15,11 @@
 # limitations under the License.
 
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.datasets.feature_utils import combine_feature_dicts
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
-from lerobot.datasets.utils import combine_feature_dicts
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.processor import RobotAction, RobotObservation, RobotProcessorPipeline
+from lerobot.processor import RobotProcessorPipeline
 from lerobot.processor.converters import (
     observation_to_transition,
     robot_action_observation_to_transition,
@@ -38,6 +38,7 @@ from lerobot.scripts.lerobot_record import record_loop
 from lerobot.teleoperators.phone.config_phone import PhoneConfig, PhoneOS
 from lerobot.teleoperators.phone.phone_processor import MapPhoneActionToRobotAction
 from lerobot.teleoperators.phone.teleop_phone import Phone
+from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun

examples/phone_to_so100/replay.py

@@ -18,7 +18,7 @@ import time
 
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.processor import RobotAction, RobotObservation, RobotProcessorPipeline
+from lerobot.processor import RobotProcessorPipeline
 from lerobot.processor.converters import (
     robot_action_observation_to_transition,
     transition_to_robot_action,
@@ -27,6 +27,7 @@ from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
 from lerobot.robots.so_follower.robot_kinematic_processor import (
     InverseKinematicsEEToJoints,
 )
+from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.constants import ACTION
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import log_say

examples/phone_to_so100/teleoperate.py

@@ -16,7 +16,7 @@
 import time
 
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.processor import RobotAction, RobotObservation, RobotProcessorPipeline
+from lerobot.processor import RobotProcessorPipeline
 from lerobot.processor.converters import (
     robot_action_observation_to_transition,
     transition_to_robot_action,
@@ -31,6 +31,7 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
 from lerobot.teleoperators.phone.config_phone import PhoneConfig, PhoneOS
 from lerobot.teleoperators.phone.phone_processor import MapPhoneActionToRobotAction
 from lerobot.teleoperators.phone.teleop_phone import Phone
+from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
 

examples/port_datasets/port_droid.py

@@ -22,7 +22,8 @@ from pathlib import Path
 import numpy as np
 import tensorflow_datasets as tfds
 
-from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.utils.utils import get_elapsed_time_in_days_hours_minutes_seconds
 
 DROID_SHARDS = 2048

examples/port_datasets/slurm_upload.py

@@ -26,7 +26,7 @@ from huggingface_hub import HfApi
 from huggingface_hub.constants import REPOCARD_NAME
 from port_droid import DROID_SHARDS
 
-from lerobot.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDatasetMetadata
+from lerobot.datasets.dataset_metadata import CODEBASE_VERSION, LeRobotDatasetMetadata
 from lerobot.datasets.utils import create_lerobot_dataset_card
 from lerobot.utils.utils import init_logging
 
@@ -155,7 +155,7 @@ class UploadDataset(PipelineStep):
         from datasets.utils.tqdm import disable_progress_bars
         from huggingface_hub import CommitOperationAdd, preupload_lfs_files
 
-        from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
+        from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
         from lerobot.utils.utils import init_logging
 
         init_logging()

examples/rtc/eval_dataset.py

@@ -113,8 +113,9 @@ from lerobot.configs import parser
 from lerobot.configs.default import DatasetConfig
 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.configs.types import RTCAttentionSchedule
+from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
 from lerobot.datasets.factory import resolve_delta_timestamps
-from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.policies.factory import get_policy_class, make_pre_post_processors
 from lerobot.policies.rtc.configuration_rtc import RTCConfig
 from lerobot.policies.rtc.debug_visualizer import RTCDebugVisualizer

examples/rtc/eval_with_real_robot.py

@@ -78,10 +78,11 @@ from torch import Tensor
 
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig  # noqa: F401
 from lerobot.cameras.realsense.configuration_realsense import RealSenseCameraConfig  # noqa: F401
+from lerobot.cameras.zmq.configuration_zmq import ZMQCameraConfig  # noqa: F401
 from lerobot.configs import parser
 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.configs.types import RTCAttentionSchedule
-from lerobot.datasets.utils import build_dataset_frame, hw_to_dataset_features
+from lerobot.datasets.feature_utils import build_dataset_frame, hw_to_dataset_features
 from lerobot.policies.factory import get_policy_class, make_pre_post_processors
 from lerobot.policies.rtc.action_queue import ActionQueue
 from lerobot.policies.rtc.configuration_rtc import RTCConfig
@@ -97,6 +98,7 @@ from lerobot.robots import (  # noqa: F401
     bi_so_follower,
     koch_follower,
     so_follower,
+    unitree_g1,
 )
 from lerobot.robots.utils import make_robot_from_config
 from lerobot.utils.constants import OBS_IMAGES

examples/so100_to_so100_EE/evaluate.py

@@ -16,15 +16,13 @@
 
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
 from lerobot.configs.types import FeatureType, PolicyFeature
+from lerobot.datasets.feature_utils import combine_feature_dicts
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
-from lerobot.datasets.utils import combine_feature_dicts
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.policies.act.modeling_act import ACTPolicy
 from lerobot.policies.factory import make_pre_post_processors
 from lerobot.processor import (
-    RobotAction,
-    RobotObservation,
     RobotProcessorPipeline,
     make_default_teleop_action_processor,
 )
@@ -40,6 +38,7 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
     InverseKinematicsEEToJoints,
 )
 from lerobot.scripts.lerobot_record import record_loop
+from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun

examples/so100_to_so100_EE/record.py

@@ -16,11 +16,11 @@
 
 
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.datasets.feature_utils import combine_feature_dicts
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
-from lerobot.datasets.utils import combine_feature_dicts
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.processor import RobotAction, RobotObservation, RobotProcessorPipeline
+from lerobot.processor import RobotProcessorPipeline
 from lerobot.processor.converters import (
     observation_to_transition,
     robot_action_observation_to_transition,
@@ -35,6 +35,7 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
 )
 from lerobot.scripts.lerobot_record import record_loop
 from lerobot.teleoperators.so_leader import SO100Leader, SO100LeaderConfig
+from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
 from lerobot.utils.visualization_utils import init_rerun

examples/so100_to_so100_EE/replay.py

@@ -19,7 +19,7 @@ import time
 
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.processor import RobotAction, RobotObservation, RobotProcessorPipeline
+from lerobot.processor import RobotProcessorPipeline
 from lerobot.processor.converters import (
     robot_action_observation_to_transition,
     transition_to_robot_action,
@@ -28,6 +28,7 @@ from lerobot.robots.so_follower import SO100Follower, SO100FollowerConfig
 from lerobot.robots.so_follower.robot_kinematic_processor import (
     InverseKinematicsEEToJoints,
 )
+from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.constants import ACTION
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import log_say

examples/so100_to_so100_EE/teleoperate.py

@@ -17,7 +17,7 @@
 import time
 
 from lerobot.model.kinematics import RobotKinematics
-from lerobot.processor import RobotAction, RobotObservation, RobotProcessorPipeline
+from lerobot.processor import RobotProcessorPipeline
 from lerobot.processor.converters import (
     robot_action_observation_to_transition,
     robot_action_to_transition,
@@ -30,6 +30,7 @@ from lerobot.robots.so_follower.robot_kinematic_processor import (
     InverseKinematicsEEToJoints,
 )
 from lerobot.teleoperators.so_leader import SO100Leader, SO100LeaderConfig
+from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
 

examples/training/train_policy.py

@@ -19,8 +19,9 @@ 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.datasets.dataset_metadata import LeRobotDatasetMetadata
+from lerobot.datasets.feature_utils import dataset_to_policy_features
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
 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

examples/training/train_with_streaming.py

@@ -20,9 +20,9 @@ from pathlib import Path
 import torch
 
 from lerobot.configs.types import FeatureType
-from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
+from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
+from lerobot.datasets.feature_utils import dataset_to_policy_features
 from lerobot.datasets.streaming_dataset import StreamingLeRobotDataset
-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

examples/tutorial/act/act_training_example.py

@@ -5,8 +5,9 @@ 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.datasets.dataset_metadata import LeRobotDatasetMetadata
+from lerobot.datasets.feature_utils import dataset_to_policy_features
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
 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

examples/tutorial/act/act_using_example.py

@@ -1,7 +1,7 @@
 import torch
 
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
+from lerobot.datasets.dataset_metadata 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

examples/tutorial/diffusion/diffusion_training_example.py

@@ -5,8 +5,9 @@ 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.datasets.dataset_metadata import LeRobotDatasetMetadata
+from lerobot.datasets.feature_utils import dataset_to_policy_features
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
 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

examples/tutorial/diffusion/diffusion_using_example.py

@@ -1,7 +1,7 @@
 import torch
 
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
+from lerobot.datasets.dataset_metadata 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

examples/tutorial/pi0/using_pi0_example.py

@@ -1,7 +1,7 @@
 import torch
 
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.utils import hw_to_dataset_features
+from lerobot.datasets.feature_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

examples/tutorial/rl/hilserl_example.py

@@ -6,8 +6,8 @@ from queue import Empty, Full
 import torch
 import torch.optim as optim
 
+from lerobot.datasets.feature_utils import hw_to_dataset_features
 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

examples/tutorial/smolvla/using_smolvla_example.py

@@ -1,7 +1,7 @@
 import torch
 
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.utils import hw_to_dataset_features
+from lerobot.datasets.feature_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

pyproject.toml

@@ -25,11 +25,11 @@ discord = "https://discord.gg/s3KuuzsPFb"
 
 [project]
 name = "lerobot"
-version = "0.4.5"
+version = "0.5.1"
 description = "🤗 LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch"
 dynamic = ["readme"]
 license = { text = "Apache-2.0" }
-requires-python = ">=3.10"
+requires-python = ">=3.12"
 authors = [
     { name = "Rémi Cadène", email = "re.cadene@gmail.com" },
     { name = "Simon Alibert", email = "alibert.sim@gmail.com" },
@@ -50,7 +50,8 @@ classifiers = [
     "Intended Audience :: Education",
     "Intended Audience :: Science/Research",
     "License :: OSI Approved :: Apache Software License",
-    "Programming Language :: Python :: 3.10",
+    "Programming Language :: Python :: 3.12",
+    "Programming Language :: Python :: 3.13",
     "Topic :: Software Development :: Build Tools",
     "Topic :: Scientific/Engineering :: Artificial Intelligence",
 ]
@@ -61,26 +62,28 @@ dependencies = [
     # Hugging Face dependencies
     "datasets>=4.0.0,<5.0.0",
     "diffusers>=0.27.2,<0.36.0",
-    "huggingface-hub[hf-transfer,cli]>=0.34.2,<0.36.0",
+    "huggingface-hub>=1.0.0,<2.0.0",
     "accelerate>=1.10.0,<2.0.0",
 
     # Core dependencies
+    "numpy>=2.0.0,<2.3.0", # NOTE: Explicitly listing numpy helps the resolver converge faster. Upper bound imposed by opencv-python-headless.
     "setuptools>=71.0.0,<81.0.0",
     "cmake>=3.29.0.1,<4.2.0",
+    "packaging>=24.2,<26.0",
+
+    "torch>=2.2.1,<2.11.0",
+    "torchcodec>=0.2.1,<0.11.0; 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')",
+    "torchvision>=0.21.0,<0.26.0",
+
     "einops>=0.8.0,<0.9.0",
-    "opencv-python-headless>=4.9.0,<4.13.0",
+    "opencv-python-headless>=4.9.0,<4.14.0",
     "av>=15.0.0,<16.0.0",
     "jsonlines>=4.0.0,<5.0.0",
-    "packaging>=24.2,<26.0",
-    "pynput>=1.7.7,<1.9.0",
+    "pynput>=1.7.8,<1.9.0",
     "pyserial>=3.5,<4.0",
+
     "wandb>=0.24.0,<0.25.0",
-
-    "torch>=2.2.1,<2.11.0", # TODO: Bump dependency
-    "torchcodec>=0.2.1,<0.11.0; 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: Bump dependency
-    "torchvision>=0.21.0,<0.26.0", # TODO: Bump dependency
-
-    "draccus==0.10.0", # TODO: Remove ==
+    "draccus==0.10.0", # TODO: Relax version constraint
     "gymnasium>=1.1.1,<2.0.0",
     "rerun-sdk>=0.24.0,<0.27.0",
 
@@ -95,10 +98,14 @@ dependencies = [
 
 # Common
 pygame-dep = ["pygame>=2.5.1,<2.7.0"]
-placo-dep = ["placo>=0.9.6,<0.10.0"]
-transformers-dep = ["transformers>=4.57.1,<5.0.0"]
+placo-dep = ["placo>=0.9.6,<0.9.17"]
+transformers-dep = ["transformers>=5.3.0,<6.0.0"]
 grpcio-dep = ["grpcio==1.73.1", "protobuf>=6.31.1,<6.32.0"]
 can-dep = ["python-can>=4.2.0,<5.0.0"]
+peft-dep = ["peft>=0.18.0,<1.0.0"]
+scipy-dep = ["scipy>=1.14.0,<2.0.0"]
+qwen-vl-utils-dep = ["qwen-vl-utils>=0.0.11,<0.1.0"]
+matplotlib-dep = ["matplotlib>=3.10.3,<4.0.0", "contourpy>=1.3.0,<2.0.0"] # NOTE: Explicitly listing contourpy helps the resolver converge faster.
 
 # Motors
 feetech = ["feetech-servo-sdk>=1.0.0,<2.0.0"]
@@ -112,34 +119,35 @@ 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"]
 unitree_g1 = [
+    # "unitree-sdk2==1.0.1",
     "pyzmq>=26.2.1,<28.0.0",
     "onnxruntime>=1.16.0,<2.0.0",
-    "pin>=3.0.0,<4.0.0",
+    "onnx>=1.16.0,<2.0.0",
     "meshcat>=0.3.0,<0.4.0",
-    "matplotlib>=3.9.0,<4.0.0",
-    "casadi>=3.6.0,<4.0.0",
+    "lerobot[matplotlib-dep]",
+    "lerobot[pygame-dep]",
 ]
 reachy2 = ["reachy2_sdk>=1.0.15,<1.1.0"]
 kinematics = ["lerobot[placo-dep]"]
 intelrealsense = [
     "pyrealsense2>=2.55.1.6486,<2.57.0 ; sys_platform != 'darwin'",
-    "pyrealsense2-macosx>=2.54,<2.55.0 ; sys_platform == 'darwin'",
+    "pyrealsense2-macosx>=2.54,<2.57.0 ; sys_platform == 'darwin'",
 ]
-phone = ["hebi-py>=2.8.0,<2.12.0", "teleop>=0.1.0,<0.2.0", "fastapi<1.0"]
+phone = ["hebi-py>=2.8.0,<2.12.0", "teleop>=0.1.0,<0.2.0", "fastapi<1.0", "lerobot[scipy-dep]"]
 
 # Policies
 wallx = [
-    "transformers==4.49.0",
-    "peft==0.17.1",
-    "scipy==1.15.3",
-    "torchdiffeq==0.2.5",
-    "qwen_vl_utils==0.0.11"
+    "lerobot[transformers-dep]",
+    "lerobot[peft]",
+    "lerobot[scipy-dep]",
+    "torchdiffeq>=0.2.4,<0.3.0",
+    "lerobot[qwen-vl-utils-dep]",
 ]
-pi = ["transformers @ git+https://github.com/huggingface/transformers.git@fix/lerobot_openpi", "scipy>=1.10.1,<1.15"]
+pi = ["lerobot[transformers-dep]", "lerobot[scipy-dep]"]
 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"]
 groot = [
     "lerobot[transformers-dep]",
-    "peft>=0.13.0,<1.0.0",
+    "lerobot[peft]",
     "dm-tree>=0.1.8,<1.0.0",
     "timm>=1.0.0,<1.1.0",
     "safetensors>=0.4.3,<1.0.0",
@@ -148,13 +156,13 @@ groot = [
     "ninja>=1.11.1,<2.0.0",
     "flash-attn>=2.5.9,<3.0.0 ; sys_platform != 'darwin'"
 ]
-sarm = ["lerobot[transformers-dep]", "faker>=33.0.0,<35.0.0", "matplotlib>=3.10.3,<4.0.0", "qwen-vl-utils>=0.0.14,<0.1.0"]
+sarm = ["lerobot[transformers-dep]", "faker>=33.0.0,<35.0.0", "lerobot[matplotlib-dep]", "lerobot[qwen-vl-utils-dep]"]
 xvla = ["lerobot[transformers-dep]"]
 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"]
-peft = ["lerobot[transformers-dep]", "peft>=0.18.0,<1.0.0"]
+async = ["lerobot[grpcio-dep]", "lerobot[matplotlib-dep]"]
+peft = ["lerobot[transformers-dep]", "lerobot[peft-dep]"]
 
 # Development
 dev = ["pre-commit>=3.7.0,<5.0.0", "debugpy>=1.8.1,<1.9.0", "lerobot[grpcio-dep]", "grpcio-tools==1.73.1", "mypy>=1.19.1"]
@@ -162,13 +170,19 @@ 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>=0.1.2,<0.2.0"]
+# NOTE: Explicitly listing scipy helps flatten the dependecy tree.
+aloha = ["gym-aloha>=0.1.2,<0.2.0", "lerobot[scipy-dep]"]
 pusht = ["gym-pusht>=0.1.5,<0.2.0", "pymunk>=6.6.0,<7.0.0"] # TODO: Fix pymunk version in gym-pusht instead
-libero = ["lerobot[transformers-dep]", "hf-libero>=0.1.3,<0.2.0"]
-metaworld = ["metaworld==3.0.0"]
+libero = ["lerobot[transformers-dep]", "hf-libero>=0.1.3,<0.2.0; sys_platform == 'linux'", "lerobot[scipy-dep]"]
+metaworld = ["metaworld==3.0.0", "lerobot[scipy-dep]"]
 
 # All
 all = [
+    # NOTE(resolver hint): scipy is pulled in transitively via lerobot[scipy-dep] through
+    # multiple extras (aloha, metaworld, pi, wallx, phone). Listing it explicitly
+    # helps pip's resolver converge by constraining scipy early, before it encounters
+    # the loose scipy requirements from transitive deps like dm-control and metaworld.
+    "scipy>=1.14.0,<2.0.0",
     "lerobot[dynamixel]",
     "lerobot[gamepad]",
     "lerobot[hopejr]",
@@ -176,8 +190,8 @@ all = [
     "lerobot[reachy2]",
     "lerobot[kinematics]",
     "lerobot[intelrealsense]",
-    # "lerobot[wallx]",
-    # "lerobot[pi]", TODO(Pepijn): Update pi to transformers v5
+    "lerobot[wallx]",
+    "lerobot[pi]",
     "lerobot[smolvla]",
     # "lerobot[groot]", TODO(Steven): Gr00t requires specific installation instructions for flash-attn
     "lerobot[xvla]",
@@ -189,10 +203,11 @@ all = [
     "lerobot[aloha]",
     "lerobot[pusht]",
     "lerobot[phone]",
-    "lerobot[libero]",
+    "lerobot[libero]; sys_platform == 'linux'",
     "lerobot[metaworld]",
     "lerobot[sarm]",
     "lerobot[peft]",
+    # "lerobot[unitree_g1]", TODO: Unitree requires specific installation instructions for unitree_sdk2
 ]
 
 [project.scripts]
@@ -221,7 +236,7 @@ lerobot = ["envs/*.json"]
 where = ["src"]
 
 [tool.ruff]
-target-version = "py310"
+target-version = "py312"
 line-length = 110
 exclude = ["tests/artifacts/**/*.safetensors", "*_pb2.py", "*_pb2_grpc.py"]
 
@@ -313,7 +328,7 @@ default.extend-ignore-identifiers-re = [
 # Uncomment [tool.mypy] first, then uncomment individual module overrides as they get proper type annotations
 
 [tool.mypy]
-python_version = "3.10"
+python_version = "3.12"
 ignore_missing_imports = true
 follow_imports = "skip"
 # warn_return_any = true
@@ -397,85 +412,3 @@ ignore_errors = false
 # [[tool.mypy.overrides]]
 # module = "lerobot.scripts.*"
 # ignore_errors = false
-
-[tool.uv]
-# wallx requires transformers==4.49.0 which conflicts with other extras that need >=4.53.0
-conflicts = [
-    [
-        { extra = "wallx" },
-        { extra = "transformers-dep" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "pi" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "smolvla" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "groot" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "xvla" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "sarm" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "hilserl" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "libero" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "peft" },
-    ],
-    [
-        { extra = "wallx" },
-        { extra = "all" },
-    ],
-    # pi uses custom branch which conflicts with transformers-dep
-    [
-        { extra = "pi" },
-        { extra = "transformers-dep" },
-    ],
-    [
-        { extra = "pi" },
-        { extra = "smolvla" },
-    ],
-    [
-        { extra = "pi" },
-        { extra = "groot" },
-    ],
-    [
-        { extra = "pi" },
-        { extra = "xvla" },
-    ],
-    [
-        { extra = "pi" },
-        { extra = "sarm" },
-    ],
-    [
-        { extra = "pi" },
-        { extra = "hilserl" },
-    ],
-    [
-        { extra = "pi" },
-        { extra = "libero" },
-    ],
-    [
-        { extra = "pi" },
-        { extra = "peft" },
-    ],
-    [
-        { extra = "pi" },
-        { extra = "all" },
-    ],
-]

requirements-macos.txt

@@ -1,76 +1,73 @@
 #
-# This file is autogenerated by pip-compile with Python 3.10
+# This file is autogenerated by pip-compile with Python 3.12
 # by the following command:
 #
 #    pip-compile --output-file=requirements-macos.txt requirements.in
 #
 -e .[all]
     # via -[all]
-absl-py==2.3.1
+absl-py==2.4.0
     # via
     #   dm-control
     #   dm-env
     #   dm-tree
     #   labmaze
     #   mujoco
-    #   tensorboard
-accelerate==1.11.0
+accelerate==1.13.0
     # via
     #   lerobot
     #   peft
 aiohappyeyeballs==2.6.1
     # via aiohttp
-aiohttp==3.13.1
+aiohttp==3.13.3
     # via fsspec
 aiosignal==1.4.0
     # via aiohttp
+annotated-doc==0.0.4
+    # via
+    #   fastapi
+    #   typer
 annotated-types==0.7.0
     # via pydantic
-antlr4-python3-runtime==4.9.3
-    # via
-    #   hydra-core
-    #   omegaconf
-anyio==4.11.0
+anyio==4.12.1
     # via
+    #   httpx
     #   starlette
     #   watchfiles
-asttokens==3.0.0
+asttokens==3.0.1
     # via stack-data
-async-timeout==5.0.1
-    # via aiohttp
 attrs==25.4.0
     # via
     #   aiohttp
     #   dm-tree
     #   jsonlines
-    #   jsonschema
-    #   referencing
     #   rerun-sdk
 av==15.1.0
-    # via lerobot
-bddl==1.0.1
-    # via libero
-certifi==2025.10.5
     # via
+    #   lerobot
+    #   qwen-vl-utils
+certifi==2026.2.25
+    # via
+    #   httpcore
+    #   httpx
     #   requests
     #   sentry-sdk
 cffi==2.0.0
     # via pymunk
-cfgv==3.4.0
+cfgv==3.5.0
     # via pre-commit
-charset-normalizer==3.4.4
+charset-normalizer==3.4.5
     # via requests
-click==8.3.0
+click==8.3.1
     # via
+    #   typer
     #   uvicorn
     #   wandb
-cloudpickle==3.1.1
-    # via
-    #   gymnasium
-    #   libero
-cmake==4.1.0
+cloudpickle==3.1.2
+    # via gymnasium
+cmake==4.1.3
     # via lerobot
-cmeel==0.57.3
+cmeel==0.59.0
     # via
     #   cmeel-assimp
     #   cmeel-boost
@@ -108,15 +105,17 @@ cmeel-zlib==1.3.1
     # via cmeel-assimp
 coal-library==3.0.1
     # via pin
-contourpy==1.3.2
-    # via matplotlib
-coverage[toml]==7.11.0
+contourpy==1.3.3
+    # via
+    #   lerobot
+    #   matplotlib
+coverage[toml]==7.13.4
     # via pytest-cov
 cycler==0.12.1
     # via matplotlib
-datasets==4.1.1
+datasets==4.6.1
     # via lerobot
-debugpy==1.8.17
+debugpy==1.8.20
     # via lerobot
 decorator==5.2.1
     # via ipython
@@ -130,7 +129,7 @@ dill==0.4.0
     #   multiprocess
 distlib==0.4.0
     # via virtualenv
-dm-control==1.0.34
+dm-control==1.0.37
     # via gym-aloha
 dm-env==1.6
     # via dm-control
@@ -138,69 +137,55 @@ 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==3.8.4
     # 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
-    #   libero
+einops==0.8.2
+    # via lerobot
 eiquadprog==1.2.9
     # via placo
-etils[epath,epy]==1.13.0
+etils[epath,epy]==1.14.0
     # via mujoco
-exceptiongroup==1.3.0
-    # via
-    #   anyio
-    #   ipython
-    #   pytest
 executing==2.2.1
     # via stack-data
+faker==34.0.2
+    # via lerobot
 farama-notifications==0.0.4
     # via gymnasium
-fastapi==0.119.1
-    # via teleop
-fastjsonschema==2.21.2
-    # via nbformat
+fastapi==0.135.1
+    # via
+    #   lerobot
+    #   teleop
 feetech-servo-sdk==1.0.0
     # via lerobot
-filelock==3.20.0
+filelock==3.25.0
     # via
     #   datasets
     #   diffusers
     #   huggingface-hub
+    #   python-discovery
     #   torch
-    #   transformers
     #   virtualenv
-fonttools==4.60.1
+fonttools==4.61.1
     # via matplotlib
 frozenlist==1.8.0
     # via
     #   aiohttp
     #   aiosignal
-fsspec[http]==2025.9.0
+fsspec[http]==2026.2.0
     # via
     #   datasets
     #   etils
     #   huggingface-hub
     #   torch
-future==1.0.0
-    # via libero
 gitdb==4.0.12
     # via gitpython
-gitpython==3.1.45
+gitpython==3.1.46
     # via wandb
 glfw==2.10.0
     # via
@@ -212,7 +197,6 @@ grpcio==1.73.1
     #   lerobot
     #   reachy2-sdk
     #   reachy2-sdk-api
-    #   tensorboard
 grpcio-tools==1.73.1
     # via
     #   lerobot
@@ -223,71 +207,67 @@ gym-hil==0.1.13
     # via lerobot
 gym-pusht==0.1.6
     # via lerobot
-gymnasium==1.2.1
+gymnasium==1.2.3
     # via
     #   gym-aloha
     #   gym-hil
     #   gym-pusht
     #   lerobot
-    #   libero
     #   metaworld
 h11==0.16.0
-    # via uvicorn
-h5py==3.15.1
-    # via robomimic
+    # via
+    #   httpcore
+    #   uvicorn
 hebi-py==2.11.0
     # via lerobot
-hf-transfer==0.1.9
-    # via huggingface-hub
-hf-xet==1.1.10
+hf-xet==1.3.2
     # via huggingface-hub
 hidapi==0.14.0.post4
     # via
     #   gym-hil
     #   lerobot
+httpcore==1.0.9
+    # via httpx
 httptools==0.7.1
     # via uvicorn
-huggingface-hub[cli,hf-transfer]==0.35.3
+httpx==0.28.1
+    # via
+    #   datasets
+    #   huggingface-hub
+huggingface-hub==1.6.0
     # via
     #   accelerate
     #   datasets
     #   diffusers
     #   lerobot
     #   peft
-    #   timm
     #   tokenizers
     #   transformers
-hydra-core==1.3.2
-    # via libero
-identify==2.6.15
+identify==2.6.17
     # via pre-commit
 idna==3.11
     # via
     #   anyio
+    #   httpx
     #   requests
     #   yarl
-imageio[ffmpeg]==2.37.0
+imageio[ffmpeg]==2.37.2
     # via
     #   gym-aloha
     #   gym-hil
     #   lerobot
     #   metaworld
-    #   robomimic
     #   scikit-image
 imageio-ffmpeg==0.6.0
-    # via
-    #   imageio
-    #   robomimic
-importlib-metadata==8.7.0
+    # via imageio
+importlib-metadata==8.7.1
     # via diffusers
-importlib-resources==6.5.2
-    # via etils
 iniconfig==2.3.0
     # via pytest
-inquirerpy==0.3.4
-    # via huggingface-hub
-ipython==8.37.0
+ipython==9.11.0
     # via meshcat
+ipython-pygments-lexers==1.1.1
+    # via ipython
 ischedule==1.2.7
     # via placo
 jedi==0.19.2
@@ -296,44 +276,24 @@ jinja2==3.1.6
     # via torch
 jsonlines==4.0.0
     # via lerobot
-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
+lazy-loader==0.5
     # via scikit-image
-libero @ git+https://github.com/huggingface/lerobot-libero.git@main
-    # via lerobot
-llvmlite==0.45.1
-    # via numba
+librt==0.8.1
+    # via mypy
 lxml==6.0.2
     # via dm-control
-markdown==3.9
-    # via tensorboard
 markdown-it-py==4.0.0
-    # via
-    #   jupytext
-    #   mdit-py-plugins
+    # via rich
 markupsafe==3.0.3
-    # via
-    #   jinja2
-    #   werkzeug
-matplotlib==3.10.7
-    # via
-    #   lerobot
-    #   libero
+    # via jinja2
+matplotlib==3.10.8
+    # via lerobot
 matplotlib-inline==0.2.1
     # via ipython
-mdit-py-plugins==0.5.0
-    # via jupytext
 mdurl==0.1.2
     # via markdown-it-py
 mergedeep==1.3.4
@@ -346,41 +306,35 @@ mock-serial==0.0.1
     # via lerobot
 mpmath==1.3.0
     # via sympy
-mujoco==3.3.7
+mujoco==3.5.0
     # via
     #   dm-control
     #   gym-aloha
     #   gym-hil
-    #   libero
     #   metaworld
-    #   robosuite
-multidict==6.7.0
+multidict==6.7.1
     # via
     #   aiohttp
     #   yarl
-multiprocess==0.70.16
+multiprocess==0.70.18
     # via datasets
+mypy==1.19.1
+    # via lerobot
 mypy-extensions==1.1.0
-    # via typing-inspect
-nbformat==5.10.4
-    # via jupytext
-networkx==3.4.2
     # via
-    #   bddl
+    #   mypy
+    #   typing-inspect
+networkx==3.6.1
+    # via
     #   scikit-image
     #   torch
-ninja==1.13.0
-    # via lerobot
-nodeenv==1.9.1
+nodeenv==1.10.0
     # 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
@@ -389,16 +343,14 @@ numpy==2.2.6
     #   dm-env
     #   dm-tree
     #   gymnasium
-    #   h5py
     #   hebi-py
     #   imageio
     #   labmaze
-    #   libero
+    #   lerobot
     #   matplotlib
     #   meshcat
     #   metaworld
     #   mujoco
-    #   numba
     #   opencv-python
     #   opencv-python-headless
     #   pandas
@@ -406,26 +358,18 @@ numpy==2.2.6
     #   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
+opencv-python==4.13.0.92
     # via
     #   gym-pusht
-    #   libero
     #   reachy2-sdk
-    #   robosuite
 opencv-python-headless==4.12.0.88
     # via lerobot
 orderly-set==5.5.0
@@ -435,97 +379,87 @@ packaging==25.0
     #   accelerate
     #   datasets
     #   huggingface-hub
-    #   hydra-core
-    #   jupytext
     #   lazy-loader
     #   lerobot
     #   matplotlib
     #   peft
     #   pytest
+    #   qwen-vl-utils
     #   reachy2-sdk
     #   scikit-image
-    #   tensorboard
-    #   tensorboardx
     #   transformers
     #   wandb
 pandas==2.3.3
     # via
     #   datasets
     #   lerobot
-parso==0.8.5
+parso==0.8.6
     # via jedi
-peft==0.17.1
+pathspec==1.0.4
+    # via mypy
+peft==0.18.1
     # via lerobot
 pexpect==4.9.0
     # via ipython
-pfzy==0.3.4
-    # via inquirerpy
-pillow==12.0.0
+pillow==12.1.1
     # via
     #   diffusers
     #   imageio
-    #   lerobot
     #   matplotlib
     #   meshcat
+    #   qwen-vl-utils
     #   rerun-sdk
-    #   robosuite
     #   scikit-image
-    #   tensorboard
     #   torchvision
 pin==3.4.0
     # via placo
-placo==0.9.14
+placo==0.9.16
     # via lerobot
-platformdirs==4.5.0
+platformdirs==4.9.4
     # via
-    #   jupyter-core
+    #   python-discovery
     #   virtualenv
     #   wandb
 pluggy==1.6.0
     # via
     #   pytest
     #   pytest-cov
-pre-commit==4.3.0
+pre-commit==4.5.1
     # via lerobot
 prompt-toolkit==3.0.52
-    # via
-    #   inquirerpy
-    #   ipython
+    # via ipython
 propcache==0.4.1
     # via
     #   aiohttp
     #   yarl
-protobuf==6.31.0
+protobuf==6.31.1
     # via
     #   dm-control
     #   grpcio-tools
     #   lerobot
     #   reachy2-sdk
     #   reachy2-sdk-api
-    #   tensorboard
-    #   tensorboardx
     #   wandb
-psutil==7.1.1
+psutil==7.2.2
     # via
     #   accelerate
     #   imageio
     #   peft
-    #   robomimic
 ptyprocess==0.7.0
     # via pexpect
 pure-eval==0.2.3
     # via stack-data
-pyarrow==21.0.0
+pyarrow==23.0.1
     # via
     #   datasets
     #   rerun-sdk
-pycparser==2.23
+pycparser==3.0
     # via cffi
-pydantic==2.12.3
+pydantic==2.12.5
     # via
     #   fastapi
     #   wandb
-pydantic-core==2.41.4
+pydantic-core==2.41.5
     # via pydantic
 pygame==2.6.1
     # via
@@ -535,33 +469,35 @@ pygame==2.6.1
 pygments==2.19.2
     # via
     #   ipython
+    #   ipython-pygments-lexers
     #   pytest
+    #   rich
 pymunk==6.11.1
     # via
     #   gym-pusht
     #   lerobot
-pyngrok==7.4.1
+pyngrok==7.5.1
     # via meshcat
 pynput==1.8.1
     # via
     #   gym-hil
     #   lerobot
-pyobjc-core==12.0
+pyobjc-core==12.1
     # via
     #   pyobjc-framework-applicationservices
     #   pyobjc-framework-cocoa
     #   pyobjc-framework-coretext
     #   pyobjc-framework-quartz
-pyobjc-framework-applicationservices==12.0
+pyobjc-framework-applicationservices==12.1
     # via pynput
-pyobjc-framework-cocoa==12.0
+pyobjc-framework-cocoa==12.1
     # via
     #   pyobjc-framework-applicationservices
     #   pyobjc-framework-coretext
     #   pyobjc-framework-quartz
-pyobjc-framework-coretext==12.0
+pyobjc-framework-coretext==12.1
     # via pyobjc-framework-applicationservices
-pyobjc-framework-quartz==12.0
+pyobjc-framework-quartz==12.1
     # via
     #   pynput
     #   pyobjc-framework-applicationservices
@@ -570,13 +506,13 @@ pyopengl==3.1.10
     # via
     #   dm-control
     #   mujoco
-pyparsing==3.2.5
+pyparsing==3.3.2
     # via
     #   dm-control
     #   matplotlib
 pyquaternion==0.9.9
     # via reachy2-sdk
-pyrealsense2-macosx==2.54.2
+pyrealsense2-macosx==2.56.5
     # via lerobot
 pyserial==3.5
     # via
@@ -585,7 +521,6 @@ pyserial==3.5
     #   lerobot
 pytest==8.4.2
     # via
-    #   bddl
     #   lerobot
     #   pytest-cov
     #   pytest-timeout
@@ -596,11 +531,14 @@ pytest-timeout==2.4.0
     # via lerobot
 python-dateutil==2.9.0.post0
     # via
+    #   faker
     #   matplotlib
     #   pandas
-python-dotenv==1.1.1
+python-discovery==1.1.1
+    # via virtualenv
+python-dotenv==1.2.2
     # via uvicorn
-pytz==2025.2
+pytz==2026.1.post1
     # via pandas
 pyyaml==6.0.3
     # via
@@ -609,13 +547,10 @@ pyyaml==6.0.3
     #   draccus
     #   hebi-py
     #   huggingface-hub
-    #   jupytext
-    #   omegaconf
     #   peft
     #   pre-commit
     #   pyngrok
     #   pyyaml-include
-    #   timm
     #   transformers
     #   uvicorn
     #   wandb
@@ -625,15 +560,13 @@ pyzmq==27.1.0
     # via
     #   lerobot
     #   meshcat
-reachy2-sdk==1.0.14
+qwen-vl-utils==0.0.14
+    # via lerobot
+reachy2-sdk==1.0.15
     # via lerobot
 reachy2-sdk-api==1.0.21
     # via reachy2-sdk
-referencing==0.37.0
-    # via
-    #   jsonschema
-    #   jsonschema-specifications
-regex==2025.10.23
+regex==2026.2.28
     # via
     #   diffusers
     #   transformers
@@ -642,184 +575,150 @@ requests==2.32.5
     #   datasets
     #   diffusers
     #   dm-control
-    #   huggingface-hub
+    #   qwen-vl-utils
     #   teleop
-    #   transformers
     #   wandb
-rerun-sdk==0.26.1
+rerun-sdk==0.26.2
     # via lerobot
 rhoban-cmeel-jsoncpp==1.9.4.9
     # via placo
-robomimic==0.2.0
-    # via libero
-robosuite==1.4.0
-    # via libero
-rpds-py==0.28.0
-    # via
-    #   jsonschema
-    #   referencing
-safetensors==0.6.2
+rich==14.3.3
+    # via typer
+safetensors==0.7.0
     # via
     #   accelerate
     #   diffusers
     #   lerobot
     #   peft
-    #   timm
     #   transformers
 scikit-image==0.25.2
     # via
     #   gym-pusht
     #   lerobot
-scipy==1.15.3
+scipy==1.17.1
     # via
     #   dm-control
+    #   lerobot
     #   metaworld
-    #   robosuite
     #   scikit-image
-sentry-sdk==2.42.1
+    #   torchdiffeq
+sentry-sdk==2.54.0
     # via wandb
 shapely==2.1.2
     # via gym-pusht
+shellingham==1.5.4
+    # via typer
 six==1.17.0
     # via
     #   pynput
     #   python-dateutil
-smmap==5.0.2
+smmap==5.0.3
     # via gitdb
-sniffio==1.3.1
-    # via anyio
 stack-data==0.6.3
     # via ipython
-starlette==0.48.0
+starlette==0.52.1
     # via fastapi
 sympy==1.14.0
     # via torch
-teleop==0.1.2
+teleop==0.1.4
     # via lerobot
-tensorboard==2.20.0
-    # via robomimic
-tensorboard-data-server==0.7.2
-    # via tensorboard
-tensorboardx==2.6.4
-    # via robomimic
-termcolor==3.1.0
-    # via
-    #   lerobot
-    #   robomimic
-thop==0.1.1.post2209072238
-    # via libero
-tifffile==2025.5.10
+termcolor==3.3.0
+    # via lerobot
+tifffile==2026.3.3
     # via scikit-image
-timm==1.0.20
-    # via lerobot
-tokenizers==0.22.1
+tokenizers==0.22.2
     # via transformers
 toml==0.10.2
     # via draccus
-tomli==2.3.0
-    # via
-    #   cmeel
-    #   coverage
-    #   jupytext
-    #   pytest
-torch==2.7.1
+torch==2.10.0
     # via
     #   accelerate
     #   lerobot
     #   peft
-    #   robomimic
-    #   thop
-    #   timm
+    #   torchdiffeq
     #   torchvision
-torchcodec==0.5
+torchcodec==0.10.0
     # via lerobot
-torchvision==0.22.1
-    # via
-    #   lerobot
-    #   robomimic
-    #   timm
-tornado==6.5.2
+torchdiffeq==0.2.5
+    # via lerobot
+torchvision==0.25.0
+    # via lerobot
+tornado==6.5.4
     # via meshcat
-tqdm==4.67.1
+tqdm==4.67.3
     # via
     #   datasets
     #   dm-control
     #   huggingface-hub
     #   peft
-    #   robomimic
     #   transformers
 traitlets==5.14.3
     # via
     #   ipython
-    #   jupyter-core
     #   matplotlib-inline
-    #   nbformat
-transformers==4.57.1
+transformers==5.3.0
     # via
     #   lerobot
-    #   libero
     #   peft
 transforms3d==0.4.2
     # via teleop
+typer==0.24.1
+    # via
+    #   huggingface-hub
+    #   transformers
 typing-extensions==4.15.0
     # via
     #   aiosignal
     #   anyio
     #   etils
-    #   exceptiongroup
+    #   faker
     #   fastapi
     #   gymnasium
     #   huggingface-hub
-    #   ipython
-    #   multidict
+    #   mypy
     #   pydantic
     #   pydantic-core
-    #   referencing
     #   rerun-sdk
     #   starlette
     #   torch
     #   typing-inspect
     #   typing-inspection
-    #   uvicorn
-    #   virtualenv
     #   wandb
 typing-inspect==0.9.0
     # via draccus
 typing-inspection==0.4.2
-    # via pydantic
-tzdata==2025.2
+    # via
+    #   fastapi
+    #   pydantic
+tzdata==2025.3
     # via pandas
 u-msgpack-python==2.8.0
     # via meshcat
-urllib3==2.5.0
+urllib3==2.6.3
     # via
     #   requests
     #   sentry-sdk
-uvicorn[standard]==0.38.0
+uvicorn[standard]==0.41.0
     # via teleop
 uvloop==0.22.1
     # via uvicorn
-virtualenv==20.35.3
+virtualenv==21.1.0
     # via pre-commit
-wandb==0.21.4
-    # via
-    #   lerobot
-    #   libero
+wandb==0.24.2
+    # via lerobot
 watchfiles==1.1.1
     # via uvicorn
-wcwidth==0.2.14
+wcwidth==0.6.0
     # via prompt-toolkit
 websocket-client==1.9.0
     # via teleop
-websockets==15.0.1
+websockets==16.0
     # via uvicorn
-werkzeug==3.1.3
-    # via tensorboard
-wrapt==2.0.0
+wrapt==2.1.2
     # via dm-tree
 xxhash==3.6.0
     # via datasets
-yarl==1.22.0
+yarl==1.23.0
     # via aiohttp
 zipp==3.23.0
     # via

requirements-ubuntu.txt

@@ -1,12 +1,12 @@
 #
-# This file is autogenerated by pip-compile with Python 3.10
+# This file is autogenerated by pip-compile with Python 3.12
 # by the following command:
 #
 #    pip-compile --output-file=requirements-ubuntu.txt requirements.in
 #
 -e .[all]
     # via -[all]
-absl-py==2.3.1
+absl-py==2.4.0
     # via
     #   dm-control
     #   dm-env
@@ -14,30 +14,33 @@ absl-py==2.3.1
     #   labmaze
     #   mujoco
     #   tensorboard
-accelerate==1.11.0
+accelerate==1.13.0
     # via
     #   lerobot
     #   peft
 aiohappyeyeballs==2.6.1
     # via aiohttp
-aiohttp==3.13.1
+aiohttp==3.13.3
     # via fsspec
 aiosignal==1.4.0
     # via aiohttp
+annotated-doc==0.0.4
+    # via
+    #   fastapi
+    #   typer
 annotated-types==0.7.0
     # via pydantic
 antlr4-python3-runtime==4.9.3
     # via
     #   hydra-core
     #   omegaconf
-anyio==4.11.0
+anyio==4.12.1
     # via
+    #   httpx
     #   starlette
     #   watchfiles
-asttokens==3.0.0
+asttokens==3.0.1
     # via stack-data
-async-timeout==5.0.1
-    # via aiohttp
 attrs==25.4.0
     # via
     #   aiohttp
@@ -47,30 +50,35 @@ attrs==25.4.0
     #   referencing
     #   rerun-sdk
 av==15.1.0
-    # via lerobot
-bddl==1.0.1
-    # via libero
-certifi==2025.10.5
     # via
+    #   lerobot
+    #   qwen-vl-utils
+bddl==1.0.1
+    # via hf-libero
+certifi==2026.2.25
+    # via
+    #   httpcore
+    #   httpx
     #   requests
     #   sentry-sdk
 cffi==2.0.0
     # via pymunk
-cfgv==3.4.0
+cfgv==3.5.0
     # via pre-commit
-charset-normalizer==3.4.4
+charset-normalizer==3.4.5
     # via requests
-click==8.3.0
+click==8.3.1
     # via
+    #   typer
     #   uvicorn
     #   wandb
-cloudpickle==3.1.1
+cloudpickle==3.1.2
     # via
     #   gymnasium
-    #   libero
-cmake==4.1.0
+    #   hf-libero
+cmake==4.1.3
     # via lerobot
-cmeel==0.57.3
+cmeel==0.59.0
     # via
     #   cmeel-assimp
     #   cmeel-boost
@@ -108,20 +116,24 @@ cmeel-zlib==1.3.1
     # via cmeel-assimp
 coal-library==3.0.1
     # via pin
-contourpy==1.3.2
-    # via matplotlib
-coverage[toml]==7.11.0
+contourpy==1.3.3
+    # via
+    #   lerobot
+    #   matplotlib
+coverage[toml]==7.13.4
     # via pytest-cov
+cuda-bindings==12.9.4
+    # via torch
+cuda-pathfinder==1.4.1
+    # via cuda-bindings
 cycler==0.12.1
     # via matplotlib
-datasets==4.1.1
+datasets==4.6.1
     # via lerobot
-debugpy==1.8.17
+debugpy==1.8.20
     # via lerobot
 decorator==5.2.1
     # via ipython
-decord==0.6.0
-    # via lerobot
 deepdiff==8.6.1
     # via lerobot
 diffusers==0.35.2
@@ -132,7 +144,7 @@ dill==0.4.0
     #   multiprocess
 distlib==0.4.0
     # via virtualenv
-dm-control==1.0.34
+dm-control==1.0.37
     # via gym-aloha
 dm-env==1.6
     # via dm-control
@@ -140,7 +152,6 @@ dm-tree==0.1.9
     # via
     #   dm-control
     #   dm-env
-    #   lerobot
 docopt==0.6.2
     # via num2words
 draccus==0.10.0
@@ -148,66 +159,60 @@ draccus==0.10.0
 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
+    # via hf-libero
+egl-probe==1.0.2
+    # via robomimic
 eigenpy==3.10.3
     # via coal-library
-einops==0.8.1
+einops==0.8.2
     # via
-    #   flash-attn
+    #   hf-libero
     #   lerobot
-    #   libero
 eiquadprog==1.2.9
     # via placo
-etils[epath,epy]==1.13.0
+etils[epath,epy]==1.14.0
     # via mujoco
-evdev==1.9.2
+evdev==1.9.3
     # via pynput
-exceptiongroup==1.3.0
-    # via
-    #   anyio
-    #   ipython
-    #   pytest
 executing==2.2.1
     # via stack-data
+faker==34.0.2
+    # via lerobot
 farama-notifications==0.0.4
     # via gymnasium
-fastapi==0.119.1
-    # via teleop
+fastapi==0.135.1
+    # via
+    #   lerobot
+    #   teleop
 fastjsonschema==2.21.2
     # via nbformat
 feetech-servo-sdk==1.0.0
     # via lerobot
-filelock==3.20.0
+filelock==3.25.0
     # via
     #   datasets
     #   diffusers
     #   huggingface-hub
+    #   python-discovery
     #   torch
-    #   transformers
     #   virtualenv
-flash-attn==2.8.3
-    # via lerobot
-fonttools==4.60.1
+fonttools==4.61.1
     # via matplotlib
 frozenlist==1.8.0
     # via
     #   aiohttp
     #   aiosignal
-fsspec[http]==2025.9.0
+fsspec[http]==2026.2.0
     # via
     #   datasets
     #   etils
     #   huggingface-hub
     #   torch
 future==1.0.0
-    # via libero
+    # via hf-libero
 gitdb==4.0.12
     # via gitpython
-gitpython==3.1.45
+gitpython==3.1.46
     # via wandb
 glfw==2.10.0
     # via
@@ -230,50 +235,60 @@ gym-hil==0.1.13
     # via lerobot
 gym-pusht==0.1.6
     # via lerobot
-gymnasium==1.2.1
+gymnasium==1.2.3
     # via
     #   gym-aloha
     #   gym-hil
     #   gym-pusht
+    #   hf-libero
     #   lerobot
-    #   libero
     #   metaworld
 h11==0.16.0
-    # via uvicorn
-h5py==3.15.1
+    # via
+    #   httpcore
+    #   uvicorn
+h5py==3.16.0
     # via robomimic
 hebi-py==2.11.0
     # via lerobot
-hf-transfer==0.1.9
-    # via huggingface-hub
-hf-xet==1.1.10
+hf-egl-probe==1.0.2
+    # via hf-libero
+hf-libero==0.1.3
+    # via lerobot
+hf-xet==1.3.2
     # via huggingface-hub
 hidapi==0.14.0.post4
     # via
     #   gym-hil
     #   lerobot
+httpcore==1.0.9
+    # via httpx
 httptools==0.7.1
     # via uvicorn
-huggingface-hub[cli,hf-transfer]==0.35.3
+httpx==0.28.1
+    # via
+    #   datasets
+    #   huggingface-hub
+huggingface-hub==1.6.0
     # via
     #   accelerate
     #   datasets
     #   diffusers
     #   lerobot
     #   peft
-    #   timm
     #   tokenizers
     #   transformers
 hydra-core==1.3.2
-    # via libero
-identify==2.6.15
+    # via hf-libero
+identify==2.6.17
     # via pre-commit
 idna==3.11
     # via
     #   anyio
+    #   httpx
     #   requests
     #   yarl
-imageio[ffmpeg]==2.37.0
+imageio[ffmpeg]==2.37.2
     # via
     #   gym-aloha
     #   gym-hil
@@ -285,16 +300,14 @@ imageio-ffmpeg==0.6.0
     # via
     #   imageio
     #   robomimic
-importlib-metadata==8.7.0
+importlib-metadata==8.7.1
     # via diffusers
-importlib-resources==6.5.2
-    # via etils
 iniconfig==2.3.0
     # via pytest
-inquirerpy==0.3.4
-    # via huggingface-hub
-ipython==8.37.0
+ipython==9.11.0
     # via meshcat
+ipython-pygments-lexers==1.1.1
+    # via ipython
 ischedule==1.2.7
     # via placo
 jedi==0.19.2
@@ -303,40 +316,41 @@ jinja2==3.1.6
     # via torch
 jsonlines==4.0.0
     # via lerobot
-jsonschema==4.25.1
+jsonschema==4.26.0
     # via nbformat
 jsonschema-specifications==2025.9.1
     # via jsonschema
 jupyter-core==5.9.1
     # via nbformat
-jupytext==1.18.1
+jupytext==1.19.1
     # via bddl
 kiwisolver==1.4.9
     # via matplotlib
 labmaze==1.0.6
     # via dm-control
-lazy-loader==0.4
+lazy-loader==0.5
     # via scikit-image
-libero @ git+https://github.com/huggingface/lerobot-libero.git@main
-    # via lerobot
-llvmlite==0.45.1
+librt==0.8.1
+    # via mypy
+llvmlite==0.46.0
     # via numba
 lxml==6.0.2
     # via dm-control
-markdown==3.9
+markdown==3.10.2
     # via tensorboard
 markdown-it-py==4.0.0
     # via
     #   jupytext
     #   mdit-py-plugins
+    #   rich
 markupsafe==3.0.3
     # via
     #   jinja2
     #   werkzeug
-matplotlib==3.10.7
+matplotlib==3.10.8
     # via
+    #   hf-libero
     #   lerobot
-    #   libero
 matplotlib-inline==0.2.1
     # via ipython
 mdit-py-plugins==0.5.0
@@ -353,36 +367,38 @@ mock-serial==0.0.1
     # via lerobot
 mpmath==1.3.0
     # via sympy
-mujoco==3.3.7
+mujoco==3.5.0
     # via
     #   dm-control
     #   gym-aloha
     #   gym-hil
-    #   libero
+    #   hf-libero
     #   metaworld
     #   robosuite
-multidict==6.7.0
+multidict==6.7.1
     # via
     #   aiohttp
     #   yarl
-multiprocess==0.70.16
+multiprocess==0.70.18
     # via datasets
+mypy==1.19.1
+    # via lerobot
 mypy-extensions==1.1.0
-    # via typing-inspect
+    # via
+    #   mypy
+    #   typing-inspect
 nbformat==5.10.4
     # via jupytext
-networkx==3.4.2
+networkx==3.6.1
     # via
     #   bddl
     #   scikit-image
     #   torch
-ninja==1.13.0
-    # via lerobot
-nodeenv==1.9.1
+nodeenv==1.10.0
     # via pre-commit
 num2words==0.5.14
     # via lerobot
-numba==0.62.1
+numba==0.64.0
     # via robosuite
 numpy==2.2.6
     # via
@@ -391,7 +407,6 @@ numpy==2.2.6
     #   cmeel-boost
     #   contourpy
     #   datasets
-    #   decord
     #   diffusers
     #   dm-control
     #   dm-env
@@ -399,9 +414,10 @@ numpy==2.2.6
     #   gymnasium
     #   h5py
     #   hebi-py
+    #   hf-libero
     #   imageio
     #   labmaze
-    #   libero
+    #   lerobot
     #   matplotlib
     #   meshcat
     #   metaworld
@@ -426,49 +442,51 @@ numpy==2.2.6
     #   torchvision
     #   transformers
     #   transforms3d
-nvidia-cublas-cu12==12.6.4.1
+nvidia-cublas-cu12==12.8.4.1
     # via
     #   nvidia-cudnn-cu12
     #   nvidia-cusolver-cu12
     #   torch
-nvidia-cuda-cupti-cu12==12.6.80
+nvidia-cuda-cupti-cu12==12.8.90
     # via torch
-nvidia-cuda-nvrtc-cu12==12.6.77
+nvidia-cuda-nvrtc-cu12==12.8.93
     # via torch
-nvidia-cuda-runtime-cu12==12.6.77
+nvidia-cuda-runtime-cu12==12.8.90
     # via torch
-nvidia-cudnn-cu12==9.5.1.17
+nvidia-cudnn-cu12==9.10.2.21
     # via torch
-nvidia-cufft-cu12==11.3.0.4
+nvidia-cufft-cu12==11.3.3.83
     # via torch
-nvidia-cufile-cu12==1.11.1.6
+nvidia-cufile-cu12==1.13.1.3
     # via torch
-nvidia-curand-cu12==10.3.7.77
+nvidia-curand-cu12==10.3.9.90
     # via torch
-nvidia-cusolver-cu12==11.7.1.2
+nvidia-cusolver-cu12==11.7.3.90
     # via torch
-nvidia-cusparse-cu12==12.5.4.2
+nvidia-cusparse-cu12==12.5.8.93
     # via
     #   nvidia-cusolver-cu12
     #   torch
-nvidia-cusparselt-cu12==0.6.3
+nvidia-cusparselt-cu12==0.7.1
     # via torch
-nvidia-nccl-cu12==2.26.2
+nvidia-nccl-cu12==2.27.5
     # via torch
-nvidia-nvjitlink-cu12==12.6.85
+nvidia-nvjitlink-cu12==12.8.93
     # via
     #   nvidia-cufft-cu12
     #   nvidia-cusolver-cu12
     #   nvidia-cusparse-cu12
     #   torch
-nvidia-nvtx-cu12==12.6.77
+nvidia-nvshmem-cu12==3.4.5
+    # via torch
+nvidia-nvtx-cu12==12.8.90
     # via torch
 omegaconf==2.3.0
     # via hydra-core
-opencv-python==4.12.0.88
+opencv-python==4.13.0.92
     # via
     #   gym-pusht
-    #   libero
+    #   hf-libero
     #   reachy2-sdk
     #   robosuite
 opencv-python-headless==4.12.0.88
@@ -487,6 +505,7 @@ packaging==25.0
     #   matplotlib
     #   peft
     #   pytest
+    #   qwen-vl-utils
     #   reachy2-sdk
     #   scikit-image
     #   tensorboard
@@ -497,21 +516,21 @@ pandas==2.3.3
     # via
     #   datasets
     #   lerobot
-parso==0.8.5
+parso==0.8.6
     # via jedi
-peft==0.17.1
+pathspec==1.0.4
+    # via mypy
+peft==0.18.1
     # via lerobot
 pexpect==4.9.0
     # via ipython
-pfzy==0.3.4
-    # via inquirerpy
-pillow==12.0.0
+pillow==12.1.1
     # via
     #   diffusers
     #   imageio
-    #   lerobot
     #   matplotlib
     #   meshcat
+    #   qwen-vl-utils
     #   rerun-sdk
     #   robosuite
     #   scikit-image
@@ -519,28 +538,27 @@ pillow==12.0.0
     #   torchvision
 pin==3.4.0
     # via placo
-placo==0.9.14
+placo==0.9.16
     # via lerobot
-platformdirs==4.5.0
+platformdirs==4.9.4
     # via
     #   jupyter-core
+    #   python-discovery
     #   virtualenv
     #   wandb
 pluggy==1.6.0
     # via
     #   pytest
     #   pytest-cov
-pre-commit==4.3.0
+pre-commit==4.5.1
     # via lerobot
 prompt-toolkit==3.0.52
-    # via
-    #   inquirerpy
-    #   ipython
+    # via ipython
 propcache==0.4.1
     # via
     #   aiohttp
     #   yarl
-protobuf==6.31.0
+protobuf==6.31.1
     # via
     #   dm-control
     #   grpcio-tools
@@ -550,7 +568,7 @@ protobuf==6.31.0
     #   tensorboard
     #   tensorboardx
     #   wandb
-psutil==7.1.1
+psutil==7.2.2
     # via
     #   accelerate
     #   imageio
@@ -560,17 +578,17 @@ ptyprocess==0.7.0
     # via pexpect
 pure-eval==0.2.3
     # via stack-data
-pyarrow==21.0.0
+pyarrow==23.0.1
     # via
     #   datasets
     #   rerun-sdk
-pycparser==2.23
+pycparser==3.0
     # via cffi
-pydantic==2.12.3
+pydantic==2.12.5
     # via
     #   fastapi
     #   wandb
-pydantic-core==2.41.4
+pydantic-core==2.41.5
     # via pydantic
 pygame==2.6.1
     # via
@@ -580,12 +598,14 @@ pygame==2.6.1
 pygments==2.19.2
     # via
     #   ipython
+    #   ipython-pygments-lexers
     #   pytest
+    #   rich
 pymunk==6.11.1
     # via
     #   gym-pusht
     #   lerobot
-pyngrok==7.4.1
+pyngrok==7.5.1
     # via meshcat
 pynput==1.8.1
     # via
@@ -595,7 +615,7 @@ pyopengl==3.1.10
     # via
     #   dm-control
     #   mujoco
-pyparsing==3.2.5
+pyparsing==3.3.2
     # via
     #   dm-control
     #   matplotlib
@@ -621,13 +641,16 @@ pytest-timeout==2.4.0
     # via lerobot
 python-dateutil==2.9.0.post0
     # via
+    #   faker
     #   matplotlib
     #   pandas
-python-dotenv==1.1.1
+python-discovery==1.1.1
+    # via virtualenv
+python-dotenv==1.2.2
     # via uvicorn
 python-xlib==0.33
     # via pynput
-pytz==2025.2
+pytz==2026.1.post1
     # via pandas
 pyyaml==6.0.3
     # via
@@ -642,7 +665,6 @@ pyyaml==6.0.3
     #   pre-commit
     #   pyngrok
     #   pyyaml-include
-    #   timm
     #   transformers
     #   uvicorn
     #   wandb
@@ -652,7 +674,9 @@ pyzmq==27.1.0
     # via
     #   lerobot
     #   meshcat
-reachy2-sdk==1.0.14
+qwen-vl-utils==0.0.14
+    # via lerobot
+reachy2-sdk==1.0.15
     # via lerobot
 reachy2-sdk-api==1.0.21
     # via reachy2-sdk
@@ -660,7 +684,7 @@ referencing==0.37.0
     # via
     #   jsonschema
     #   jsonschema-specifications
-regex==2025.10.23
+regex==2026.2.28
     # via
     #   diffusers
     #   transformers
@@ -669,60 +693,62 @@ requests==2.32.5
     #   datasets
     #   diffusers
     #   dm-control
-    #   huggingface-hub
+    #   qwen-vl-utils
     #   teleop
-    #   transformers
     #   wandb
-rerun-sdk==0.26.1
+rerun-sdk==0.26.2
     # via lerobot
 rhoban-cmeel-jsoncpp==1.9.4.9
     # via placo
+rich==14.3.3
+    # via typer
 robomimic==0.2.0
-    # via libero
+    # via hf-libero
 robosuite==1.4.0
-    # via libero
-rpds-py==0.28.0
+    # via hf-libero
+rpds-py==0.30.0
     # via
     #   jsonschema
     #   referencing
-safetensors==0.6.2
+safetensors==0.7.0
     # via
     #   accelerate
     #   diffusers
     #   lerobot
     #   peft
-    #   timm
     #   transformers
 scikit-image==0.25.2
     # via
     #   gym-pusht
     #   lerobot
-scipy==1.15.3
+scipy==1.17.1
     # via
     #   dm-control
+    #   lerobot
     #   metaworld
     #   robosuite
     #   scikit-image
-sentry-sdk==2.42.1
+    #   torchdiffeq
+sentry-sdk==2.54.0
     # via wandb
 shapely==2.1.2
     # via gym-pusht
+shellingham==1.5.4
+    # via typer
 six==1.17.0
     # via
     #   pynput
     #   python-dateutil
     #   python-xlib
-smmap==5.0.2
+smmap==5.0.3
     # via gitdb
-sniffio==1.3.1
-    # via anyio
 stack-data==0.6.3
     # via ipython
-starlette==0.48.0
+starlette==0.52.1
     # via fastapi
 sympy==1.14.0
     # via torch
-teleop==0.1.2
+teleop==0.1.4
     # via lerobot
 tensorboard==2.20.0
     # via robomimic
@@ -730,46 +756,38 @@ tensorboard-data-server==0.7.2
     # via tensorboard
 tensorboardx==2.6.4
     # via robomimic
-termcolor==3.1.0
+termcolor==3.3.0
     # via
     #   lerobot
     #   robomimic
 thop==0.1.1.post2209072238
-    # via libero
-tifffile==2025.5.10
+    # via hf-libero
+tifffile==2026.3.3
     # via scikit-image
-timm==1.0.20
-    # via lerobot
-tokenizers==0.22.1
+tokenizers==0.22.2
     # via transformers
 toml==0.10.2
     # via draccus
-tomli==2.3.0
-    # via
-    #   cmeel
-    #   coverage
-    #   jupytext
-    #   pytest
-torch==2.7.1
+torch==2.10.0
     # via
     #   accelerate
-    #   flash-attn
     #   lerobot
     #   peft
     #   robomimic
     #   thop
-    #   timm
+    #   torchdiffeq
     #   torchvision
-torchcodec==0.5
+torchcodec==0.10.0
     # via lerobot
-torchvision==0.22.1
+torchdiffeq==0.2.5
+    # via lerobot
+torchvision==0.25.0
     # via
     #   lerobot
     #   robomimic
-    #   timm
-tornado==6.5.2
+tornado==6.5.4
     # via meshcat
-tqdm==4.67.1
+tqdm==4.67.3
     # via
     #   datasets
     #   dm-control
@@ -783,26 +801,29 @@ traitlets==5.14.3
     #   jupyter-core
     #   matplotlib-inline
     #   nbformat
-transformers==4.57.1
+transformers==5.3.0
     # via
+    #   hf-libero
     #   lerobot
-    #   libero
     #   peft
 transforms3d==0.4.2
     # via teleop
-triton==3.3.1
+triton==3.6.0
     # via torch
+typer==0.24.1
+    # via
+    #   huggingface-hub
+    #   transformers
 typing-extensions==4.15.0
     # via
     #   aiosignal
     #   anyio
     #   etils
-    #   exceptiongroup
+    #   faker
     #   fastapi
     #   gymnasium
     #   huggingface-hub
-    #   ipython
-    #   multidict
+    #   mypy
     #   pydantic
     #   pydantic-core
     #   referencing
@@ -811,46 +832,46 @@ typing-extensions==4.15.0
     #   torch
     #   typing-inspect
     #   typing-inspection
-    #   uvicorn
-    #   virtualenv
     #   wandb
 typing-inspect==0.9.0
     # via draccus
 typing-inspection==0.4.2
-    # via pydantic
-tzdata==2025.2
+    # via
+    #   fastapi
+    #   pydantic
+tzdata==2025.3
     # via pandas
 u-msgpack-python==2.8.0
     # via meshcat
-urllib3==2.5.0
+urllib3==2.6.3
     # via
     #   requests
     #   sentry-sdk
-uvicorn[standard]==0.38.0
+uvicorn[standard]==0.41.0
     # via teleop
 uvloop==0.22.1
     # via uvicorn
-virtualenv==20.35.3
+virtualenv==21.1.0
     # via pre-commit
-wandb==0.21.4
+wandb==0.24.2
     # via
+    #   hf-libero
     #   lerobot
-    #   libero
 watchfiles==1.1.1
     # via uvicorn
-wcwidth==0.2.14
+wcwidth==0.6.0
     # via prompt-toolkit
 websocket-client==1.9.0
     # via teleop
-websockets==15.0.1
+websockets==16.0
     # via uvicorn
-werkzeug==3.1.3
+werkzeug==3.1.6
     # via tensorboard
-wrapt==2.0.0
+wrapt==2.1.2
     # via dm-tree
 xxhash==3.6.0
     # via datasets
-yarl==1.22.0
+yarl==1.23.0
     # via aiohttp
 zipp==3.23.0
     # via

requirements.in

@@ -1,9 +1,9 @@
 # requirements.in
 
-# 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-macos.txt was generated on macOS and is platform-specific (macOS 26.3.1 25D2128 arm64).
+# Darwin MacBook-Pro.local 25.3.0 Darwin Kernel Version 25.3.0: Wed Jan 28 20:54:55 PST 2026; root:xnu-12377.91.3~2/RELEASE_ARM64_T8132 arm64
 
-# requirements-ubuntu.txt was generated on Linux and is platform-specific (Ubuntu 24.04.3 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
+# requirements-ubuntu.txt was generated on Linux and is platform-specific (Ubuntu 24.04.4 LTS x86_64).
+# Linux lerobot-linux 6.17.0-14-generic #14~24.04.1-Ubuntu SMP PREEMPT_DYNAMIC Thu Jan 15 15:52:10 UTC 2 x86_64 x86_64 x86_64 GNU/Linux
 
 -e .[all]

src/lerobot/async_inference/helpers.py

@@ -23,7 +23,7 @@ from typing import Any
 import torch
 
 from lerobot.configs.types import PolicyFeature
-from lerobot.datasets.utils import build_dataset_frame, hw_to_dataset_features
+from lerobot.datasets.feature_utils import build_dataset_frame, hw_to_dataset_features
 
 # NOTE: Configs need to be loaded for the client to be able to instantiate the policy config
 from lerobot.policies import (  # noqa: F401

src/lerobot/async_inference/policy_server.py

@@ -39,15 +39,13 @@ import grpc
 import torch
 
 from lerobot.policies.factory import get_policy_class, make_pre_post_processors
-from lerobot.processor import (
-    PolicyAction,
-    PolicyProcessorPipeline,
-)
+from lerobot.processor import PolicyProcessorPipeline
 from lerobot.transport import (
     services_pb2,  # type: ignore
     services_pb2_grpc,  # type: ignore
 )
 from lerobot.transport.utils import receive_bytes_in_chunks
+from lerobot.types import PolicyAction
 
 from .configs import PolicyServerConfig
 from .constants import SUPPORTED_POLICIES

src/lerobot/async_inference/robot_client.py

@@ -49,9 +49,14 @@ import torch
 
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig  # noqa: F401
 from lerobot.cameras.realsense.configuration_realsense import RealSenseCameraConfig  # noqa: F401
-from lerobot.robots import (
-    RobotConfig,  # noqa: F401
+from lerobot.robots import (  # noqa: F401
+    Robot,
+    RobotConfig,
+    bi_so_follower,
+    koch_follower,
     make_robot_from_config,
+    omx_follower,
+    so_follower,
 )
 from lerobot.transport import (
     services_pb2,  # type: ignore

src/lerobot/cameras/zmq/camera_zmq.py

@@ -181,7 +181,7 @@ class ZMQCamera(Camera):
         try:
             message = self.socket.recv_string()
         except Exception as e:
-            # Check for ZMQ timeout (EAGAIN/Again) without requiring global zmq import
+            # zmq is lazy-imported in connect(), so check by name to avoid a top-level import
             if type(e).__name__ == "Again":
                 raise TimeoutError(f"{self} timeout after {self.timeout_ms}ms") from e
             raise

src/lerobot/cameras/zmq/image_server.py

@@ -23,6 +23,7 @@ import base64
 import contextlib
 import json
 import logging
+import threading
 import time
 from collections import deque
 
@@ -42,10 +43,57 @@ def encode_image(image: np.ndarray, quality: int = 80) -> str:
     return base64.b64encode(buffer).decode("utf-8")
 
 
+class CameraCaptureThread:
+    """Background thread that continuously captures and encodes frames from a camera."""
+
+    def __init__(self, camera: OpenCVCamera, name: str):
+        self.camera = camera
+        self.name = name
+        self.latest_encoded: str | None = None  # Pre-encoded JPEG as base64
+        self.latest_timestamp: float = 0.0
+        self.frame_lock = threading.Lock()
+        self.running = False
+        self.thread: threading.Thread | None = None
+
+    def start(self):
+        """Start the capture thread."""
+        self.running = True
+        self.thread = threading.Thread(target=self._capture_loop, daemon=True)
+        self.thread.start()
+
+    def stop(self):
+        """Stop the capture thread."""
+        self.running = False
+        if self.thread:
+            self.thread.join(timeout=1.0)
+
+    def _capture_loop(self):
+        """Continuously capture and encode frames at the camera's native rate."""
+        while self.running:
+            try:
+                frame = self.camera.read()  # Blocks at camera's native rate
+                timestamp = time.time()
+                # Encode immediately in capture thread (this is the slow part)
+                encoded = encode_image(frame)
+                with self.frame_lock:
+                    self.latest_encoded = encoded
+                    self.latest_timestamp = timestamp
+            except Exception as e:
+                logger.warning(f"Camera {self.name} capture error: {e}")
+                time.sleep(0.01)
+
+    def get_latest(self) -> tuple[str | None, float]:
+        """Get the latest encoded frame and its timestamp."""
+        with self.frame_lock:
+            return self.latest_encoded, self.latest_timestamp
+
+
 class ImageServer:
     def __init__(self, config: dict, port: int = 5555):
+        # fps controls the publish loop rate (how often frames are sent over ZMQ), not the camera capture rate
         self.fps = config.get("fps", 30)
         self.cameras: dict[str, OpenCVCamera] = {}
+        self.capture_threads: dict[str, CameraCaptureThread] = {}
 
         for name, cfg in config.get("cameras", {}).items():
             shape = cfg.get("shape", [480, 640])
@@ -61,6 +109,10 @@ class ImageServer:
             self.cameras[name] = camera
             logger.info(f"Camera {name}: {shape[1]}x{shape[0]}")
 
+            # Create capture thread for this camera
+            capture_thread = CameraCaptureThread(camera, name)
+            self.capture_threads[name] = capture_thread
+
         # ZMQ PUB socket
         self.context = zmq.Context()
         self.socket = self.context.socket(zmq.PUB)
@@ -73,6 +125,18 @@ class ImageServer:
     def run(self):
         frame_count = 0
         frame_times = deque(maxlen=60)
+        last_published_ts: dict[str, float] = {}
+
+        # Start all capture threads
+        for capture_thread in self.capture_threads.values():
+            capture_thread.start()
+
+        # Wait for first frames to be captured and encoded
+        logger.info("Waiting for cameras to start capturing...")
+        for name, capture_thread in self.capture_threads.items():
+            while capture_thread.get_latest()[0] is None:
+                time.sleep(0.01)
+            logger.info(f"Camera {name} ready (capture + encode in background)")
 
         try:
             while True:
@@ -80,10 +144,12 @@ class ImageServer:
 
                 # Build message
                 message = {"timestamps": {}, "images": {}}
-                for name, cam in self.cameras.items():
-                    frame = cam.read()  # Returns RGB
-                    message["timestamps"][name] = time.time()
-                    message["images"][name] = encode_image(frame)
+                for name, capture_thread in self.capture_threads.items():
+                    encoded, timestamp = capture_thread.get_latest()
+                    if encoded is not None and timestamp > last_published_ts.get(name, 0.0):
+                        message["timestamps"][name] = timestamp
+                        message["images"][name] = encoded
+                        last_published_ts[name] = timestamp
 
                 # Send as JSON string (suppress if buffer full)
                 with contextlib.suppress(zmq.Again):
@@ -102,6 +168,8 @@ class ImageServer:
         except KeyboardInterrupt:
             pass
         finally:
+            for capture_thread in self.capture_threads.values():
+                capture_thread.stop()
             for cam in self.cameras.values():
                 cam.disconnect()
             self.socket.close()

src/lerobot/configs/default.py

@@ -36,6 +36,16 @@ class DatasetConfig:
     video_backend: str = field(default_factory=get_safe_default_codec)
     streaming: bool = False
 
+    def __post_init__(self) -> None:
+        if self.episodes is not None:
+            if any(ep < 0 for ep in self.episodes):
+                raise ValueError(
+                    f"Episode indices must be non-negative, got: {[ep for ep in self.episodes if ep < 0]}"
+                )
+            if len(self.episodes) != len(set(self.episodes)):
+                duplicates = sorted({ep for ep in self.episodes if self.episodes.count(ep) > 1})
+                raise ValueError(f"Episode indices contain duplicates: {duplicates}")
+
 
 @dataclass
 class WandBConfig:
@@ -47,6 +57,7 @@ class WandBConfig:
     notes: str | None = None
     run_id: str | None = None
     mode: str | None = None  # Allowed values: 'online', 'offline' 'disabled'. Defaults to 'online'
+    add_tags: bool = True  # If True, save configuration as tags in the WandB run.
 
 
 @dataclass

src/lerobot/configs/policies.py

@@ -30,8 +30,8 @@ from lerobot.configs.types import FeatureType, PolicyFeature
 from lerobot.optim.optimizers import OptimizerConfig
 from lerobot.optim.schedulers import LRSchedulerConfig
 from lerobot.utils.constants import ACTION, OBS_STATE
+from lerobot.utils.device_utils import auto_select_torch_device, is_amp_available, is_torch_device_available
 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__)

src/lerobot/configs/train.py

@@ -50,6 +50,9 @@ class TrainPipelineConfig(HubMixin):
     # `seed` is used for training (eg: model initialization, dataset shuffling)
     # AND for the evaluation environments.
     seed: int | None = 1000
+    # Set to True to use deterministic cuDNN algorithms for reproducibility.
+    # This disables cudnn.benchmark and may reduce training speed by ~10-20 percent.
+    cudnn_deterministic: bool = False
     # Number of workers for the dataloader.
     num_workers: int = 4
     batch_size: int = 8

src/lerobot/data_processing/sarm_annotations/subtask_annotation.py

@@ -746,7 +746,8 @@ def save_annotations_to_dataset(
     dataset_path: Path, annotations: dict[int, SubtaskAnnotation], fps: int, prefix: str = "sparse"
 ):
     """Save annotations to LeRobot dataset parquet format."""
-    from lerobot.datasets.utils import DEFAULT_EPISODES_PATH, load_episodes
+    from lerobot.datasets.io_utils import load_episodes
+    from lerobot.datasets.utils import DEFAULT_EPISODES_PATH
 
     episodes_dataset = load_episodes(dataset_path)
     if not episodes_dataset or len(episodes_dataset) == 0:
@@ -840,7 +841,7 @@ def generate_auto_sparse_annotations(
 
 def load_annotations_from_dataset(dataset_path: Path, prefix: str = "sparse") -> dict[int, SubtaskAnnotation]:
     """Load annotations from LeRobot dataset parquet files."""
-    from lerobot.datasets.utils import load_episodes
+    from lerobot.datasets.io_utils import load_episodes
 
     episodes_dataset = load_episodes(dataset_path)
     if not episodes_dataset or len(episodes_dataset) == 0:

src/lerobot/datasets/aggregate.py

@@ -24,7 +24,16 @@ import pandas as pd
 import tqdm
 
 from lerobot.datasets.compute_stats import aggregate_stats
-from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
+from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
+from lerobot.datasets.feature_utils import get_hf_features_from_features
+from lerobot.datasets.io_utils import (
+    get_file_size_in_mb,
+    get_parquet_file_size_in_mb,
+    to_parquet_with_hf_images,
+    write_info,
+    write_stats,
+    write_tasks,
+)
 from lerobot.datasets.utils import (
     DEFAULT_CHUNK_SIZE,
     DEFAULT_DATA_FILE_SIZE_IN_MB,
@@ -32,14 +41,7 @@ from lerobot.datasets.utils import (
     DEFAULT_EPISODES_PATH,
     DEFAULT_VIDEO_FILE_SIZE_IN_MB,
     DEFAULT_VIDEO_PATH,
-    get_file_size_in_mb,
-    get_hf_features_from_features,
-    get_parquet_file_size_in_mb,
-    to_parquet_with_hf_images,
     update_chunk_file_indices,
-    write_info,
-    write_stats,
-    write_tasks,
 )
 from lerobot.datasets.video_utils import concatenate_video_files, get_video_duration_in_s
 
@@ -289,7 +291,9 @@ def aggregate_datasets(
 
     logging.info("Find all tasks")
     unique_tasks = pd.concat([m.tasks for m in all_metadata]).index.unique()
-    dst_meta.tasks = pd.DataFrame({"task_index": range(len(unique_tasks))}, index=unique_tasks)
+    dst_meta.tasks = pd.DataFrame(
+        {"task_index": range(len(unique_tasks))}, index=pd.Index(unique_tasks, name="task")
+    )
 
     meta_idx = {"chunk": 0, "file": 0}
     data_idx = {"chunk": 0, "file": 0}

src/lerobot/datasets/backward_compatibility.py

@@ -1,56 +0,0 @@
-# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-import packaging.version
-
-V30_MESSAGE = """
-The dataset you requested ({repo_id}) is in {version} format.
-
-We introduced a new format since v3.0 which is not backward compatible with v2.1.
-Please, update your dataset to the new format using this command:
-```
-python -m lerobot.datasets.v30.convert_dataset_v21_to_v30 --repo-id={repo_id}
-```
-
-If you already have a converted version uploaded to the hub, then this error might be because of
-an older version in your local cache. Consider deleting the cached version and retrying.
-
-If you encounter a problem, contact LeRobot maintainers on [Discord](https://discord.com/invite/s3KuuzsPFb)
-or open an [issue on GitHub](https://github.com/huggingface/lerobot/issues/new/choose).
-"""
-
-FUTURE_MESSAGE = """
-The dataset you requested ({repo_id}) is only available in {version} format.
-As we cannot ensure forward compatibility with it, please update your current version of lerobot.
-"""
-
-
-class CompatibilityError(Exception): ...
-
-
-class BackwardCompatibilityError(CompatibilityError):
-    def __init__(self, repo_id: str, version: packaging.version.Version):
-        if version.major == 2 and version.minor == 1:
-            message = V30_MESSAGE.format(repo_id=repo_id, version=version)
-        else:
-            raise NotImplementedError(
-                "Contact the maintainer on [Discord](https://discord.com/invite/s3KuuzsPFb)."
-            )
-        super().__init__(message)
-
-
-class ForwardCompatibilityError(CompatibilityError):
-    def __init__(self, repo_id: str, version: packaging.version.Version):
-        message = FUTURE_MESSAGE.format(repo_id=repo_id, version=version)
-        super().__init__(message)

src/lerobot/datasets/compute_stats.py

@@ -15,7 +15,7 @@
 # limitations under the License.
 import numpy as np
 
-from lerobot.datasets.utils import load_image_as_numpy
+from lerobot.datasets.io_utils import load_image_as_numpy
 
 DEFAULT_QUANTILES = [0.01, 0.10, 0.50, 0.90, 0.99]
 

src/lerobot/datasets/dataset_metadata.py

@@ -0,0 +1,517 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from pathlib import Path
+
+import numpy as np
+import packaging.version
+import pandas as pd
+import pyarrow as pa
+import pyarrow.parquet as pq
+from huggingface_hub import snapshot_download
+
+from lerobot.datasets.compute_stats import aggregate_stats
+from lerobot.datasets.feature_utils import _validate_feature_names, create_empty_dataset_info
+from lerobot.datasets.io_utils import (
+    get_file_size_in_mb,
+    load_episodes,
+    load_info,
+    load_stats,
+    load_subtasks,
+    load_tasks,
+    write_info,
+    write_json,
+    write_stats,
+    write_tasks,
+)
+from lerobot.datasets.utils import (
+    DEFAULT_EPISODES_PATH,
+    DEFAULT_FEATURES,
+    INFO_PATH,
+    check_version_compatibility,
+    flatten_dict,
+    get_safe_version,
+    is_valid_version,
+    update_chunk_file_indices,
+)
+from lerobot.datasets.video_utils import get_video_info
+from lerobot.utils.constants import HF_LEROBOT_HOME
+
+CODEBASE_VERSION = "v3.0"
+
+
+class LeRobotDatasetMetadata:
+    def __init__(
+        self,
+        repo_id: str,
+        root: str | Path | None = None,
+        revision: str | None = None,
+        force_cache_sync: bool = False,
+        metadata_buffer_size: int = 10,
+    ):
+        self.repo_id = repo_id
+        self.revision = revision if revision else CODEBASE_VERSION
+        self.root = Path(root) if root is not None else HF_LEROBOT_HOME / repo_id
+        self.writer = None
+        self.latest_episode = None
+        self.metadata_buffer: list[dict] = []
+        self.metadata_buffer_size = metadata_buffer_size
+
+        try:
+            if force_cache_sync:
+                raise FileNotFoundError
+            self.load_metadata()
+        except (FileNotFoundError, NotADirectoryError):
+            if is_valid_version(self.revision):
+                self.revision = get_safe_version(self.repo_id, self.revision)
+
+            (self.root / "meta").mkdir(exist_ok=True, parents=True)
+            self.pull_from_repo(allow_patterns="meta/")
+            self.load_metadata()
+
+    def _flush_metadata_buffer(self) -> None:
+        """Write all buffered episode metadata to parquet file."""
+        if not hasattr(self, "metadata_buffer") or len(self.metadata_buffer) == 0:
+            return
+
+        combined_dict = {}
+        for episode_dict in self.metadata_buffer:
+            for key, value in episode_dict.items():
+                if key not in combined_dict:
+                    combined_dict[key] = []
+                # Extract value and serialize numpy arrays
+                # because PyArrow's from_pydict function doesn't support numpy arrays
+                val = value[0] if isinstance(value, list) else value
+                combined_dict[key].append(val.tolist() if isinstance(val, np.ndarray) else val)
+
+        first_ep = self.metadata_buffer[0]
+        chunk_idx = first_ep["meta/episodes/chunk_index"][0]
+        file_idx = first_ep["meta/episodes/file_index"][0]
+
+        table = pa.Table.from_pydict(combined_dict)
+
+        if not self.writer:
+            path = Path(self.root / DEFAULT_EPISODES_PATH.format(chunk_index=chunk_idx, file_index=file_idx))
+            path.parent.mkdir(parents=True, exist_ok=True)
+            self.writer = pq.ParquetWriter(
+                path, schema=table.schema, compression="snappy", use_dictionary=True
+            )
+
+        self.writer.write_table(table)
+
+        self.latest_episode = self.metadata_buffer[-1]
+        self.metadata_buffer.clear()
+
+    def _close_writer(self) -> None:
+        """Close and cleanup the parquet writer if it exists."""
+        self._flush_metadata_buffer()
+
+        writer = getattr(self, "writer", None)
+        if writer is not None:
+            writer.close()
+            self.writer = None
+
+    def __del__(self):
+        """
+        Trust the user to call .finalize() but as an added safety check call the parquet writer to stop when calling the destructor
+        """
+        self._close_writer()
+
+    def load_metadata(self):
+        self.info = load_info(self.root)
+        check_version_compatibility(self.repo_id, self._version, CODEBASE_VERSION)
+        self.tasks = load_tasks(self.root)
+        self.subtasks = load_subtasks(self.root)
+        self.episodes = load_episodes(self.root)
+        self.stats = load_stats(self.root)
+
+    def pull_from_repo(
+        self,
+        allow_patterns: list[str] | str | None = None,
+        ignore_patterns: list[str] | str | None = None,
+    ) -> None:
+        snapshot_download(
+            self.repo_id,
+            repo_type="dataset",
+            revision=self.revision,
+            local_dir=self.root,
+            allow_patterns=allow_patterns,
+            ignore_patterns=ignore_patterns,
+        )
+
+    @property
+    def url_root(self) -> str:
+        return f"hf://datasets/{self.repo_id}"
+
+    @property
+    def _version(self) -> packaging.version.Version:
+        """Codebase version used to create this dataset."""
+        return packaging.version.parse(self.info["codebase_version"])
+
+    def get_data_file_path(self, ep_index: int) -> Path:
+        if self.episodes is None:
+            self.episodes = load_episodes(self.root)
+        if ep_index >= len(self.episodes):
+            raise IndexError(
+                f"Episode index {ep_index} out of range. Episodes: {len(self.episodes) if self.episodes else 0}"
+            )
+        ep = self.episodes[ep_index]
+        chunk_idx = ep["data/chunk_index"]
+        file_idx = ep["data/file_index"]
+        fpath = self.data_path.format(chunk_index=chunk_idx, file_index=file_idx)
+        return Path(fpath)
+
+    def get_video_file_path(self, ep_index: int, vid_key: str) -> Path:
+        if self.episodes is None:
+            self.episodes = load_episodes(self.root)
+        if ep_index >= len(self.episodes):
+            raise IndexError(
+                f"Episode index {ep_index} out of range. Episodes: {len(self.episodes) if self.episodes else 0}"
+            )
+        ep = self.episodes[ep_index]
+        chunk_idx = ep[f"videos/{vid_key}/chunk_index"]
+        file_idx = ep[f"videos/{vid_key}/file_index"]
+        fpath = self.video_path.format(video_key=vid_key, chunk_index=chunk_idx, file_index=file_idx)
+        return Path(fpath)
+
+    @property
+    def data_path(self) -> str:
+        """Formattable string for the parquet files."""
+        return self.info["data_path"]
+
+    @property
+    def video_path(self) -> str | None:
+        """Formattable string for the video files."""
+        return self.info["video_path"]
+
+    @property
+    def robot_type(self) -> str | None:
+        """Robot type used in recording this dataset."""
+        return self.info["robot_type"]
+
+    @property
+    def fps(self) -> int:
+        """Frames per second used during data collection."""
+        return self.info["fps"]
+
+    @property
+    def features(self) -> dict[str, dict]:
+        """All features contained in the dataset."""
+        return self.info["features"]
+
+    @property
+    def image_keys(self) -> list[str]:
+        """Keys to access visual modalities stored as images."""
+        return [key for key, ft in self.features.items() if ft["dtype"] == "image"]
+
+    @property
+    def video_keys(self) -> list[str]:
+        """Keys to access visual modalities stored as videos."""
+        return [key for key, ft in self.features.items() if ft["dtype"] == "video"]
+
+    @property
+    def camera_keys(self) -> list[str]:
+        """Keys to access visual modalities (regardless of their storage method)."""
+        return [key for key, ft in self.features.items() if ft["dtype"] in ["video", "image"]]
+
+    @property
+    def names(self) -> dict[str, list | dict]:
+        """Names of the various dimensions of vector modalities."""
+        return {key: ft["names"] for key, ft in self.features.items()}
+
+    @property
+    def shapes(self) -> dict:
+        """Shapes for the different features."""
+        return {key: tuple(ft["shape"]) for key, ft in self.features.items()}
+
+    @property
+    def total_episodes(self) -> int:
+        """Total number of episodes available."""
+        return self.info["total_episodes"]
+
+    @property
+    def total_frames(self) -> int:
+        """Total number of frames saved in this dataset."""
+        return self.info["total_frames"]
+
+    @property
+    def total_tasks(self) -> int:
+        """Total number of different tasks performed in this dataset."""
+        return self.info["total_tasks"]
+
+    @property
+    def chunks_size(self) -> int:
+        """Max number of files per chunk."""
+        return self.info["chunks_size"]
+
+    @property
+    def data_files_size_in_mb(self) -> int:
+        """Max size of data file in mega bytes."""
+        return self.info["data_files_size_in_mb"]
+
+    @property
+    def video_files_size_in_mb(self) -> int:
+        """Max size of video file in mega bytes."""
+        return self.info["video_files_size_in_mb"]
+
+    def get_task_index(self, task: str) -> int | None:
+        """
+        Given a task in natural language, returns its task_index if the task already exists in the dataset,
+        otherwise return None.
+        """
+        if task in self.tasks.index:
+            return int(self.tasks.loc[task].task_index)
+        else:
+            return None
+
+    def save_episode_tasks(self, tasks: list[str]):
+        if len(set(tasks)) != len(tasks):
+            raise ValueError(f"Tasks are not unique: {tasks}")
+
+        if self.tasks is None:
+            new_tasks = tasks
+            task_indices = range(len(tasks))
+            self.tasks = pd.DataFrame({"task_index": task_indices}, index=pd.Index(tasks, name="task"))
+        else:
+            new_tasks = [task for task in tasks if task not in self.tasks.index]
+            new_task_indices = range(len(self.tasks), len(self.tasks) + len(new_tasks))
+            for task_idx, task in zip(new_task_indices, new_tasks, strict=False):
+                self.tasks.loc[task] = task_idx
+
+        if len(new_tasks) > 0:
+            # Update on disk
+            write_tasks(self.tasks, self.root)
+
+    def _save_episode_metadata(self, episode_dict: dict) -> None:
+        """Buffer episode metadata and write to parquet in batches for efficiency.
+
+        This function accumulates episode metadata in a buffer and flushes it when the buffer
+        reaches the configured size. This reduces I/O overhead by writing multiple episodes
+        at once instead of one row at a time.
+
+        Notes: We both need to update parquet files and HF dataset:
+        - `pandas` loads parquet file in RAM
+        - `datasets` relies on a memory mapping from pyarrow (no RAM). It either converts parquet files to a pyarrow cache on disk,
+          or loads directly from pyarrow cache.
+        """
+        # Convert to list format for each value
+        episode_dict = {key: [value] for key, value in episode_dict.items()}
+        num_frames = episode_dict["length"][0]
+
+        if self.latest_episode is None:
+            # Initialize indices and frame count for a new dataset made of the first episode data
+            chunk_idx, file_idx = 0, 0
+            if self.episodes is not None and len(self.episodes) > 0:
+                # It means we are resuming recording, so we need to load the latest episode
+                # Update the indices to avoid overwriting the latest episode
+                chunk_idx = self.episodes[-1]["meta/episodes/chunk_index"]
+                file_idx = self.episodes[-1]["meta/episodes/file_index"]
+                latest_num_frames = self.episodes[-1]["dataset_to_index"]
+                episode_dict["dataset_from_index"] = [latest_num_frames]
+                episode_dict["dataset_to_index"] = [latest_num_frames + num_frames]
+
+                # When resuming, move to the next file
+                chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, self.chunks_size)
+            else:
+                episode_dict["dataset_from_index"] = [0]
+                episode_dict["dataset_to_index"] = [num_frames]
+
+            episode_dict["meta/episodes/chunk_index"] = [chunk_idx]
+            episode_dict["meta/episodes/file_index"] = [file_idx]
+        else:
+            chunk_idx = self.latest_episode["meta/episodes/chunk_index"][0]
+            file_idx = self.latest_episode["meta/episodes/file_index"][0]
+
+            latest_path = (
+                self.root / DEFAULT_EPISODES_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
+                if self.writer is None
+                else self.writer.where
+            )
+
+            if Path(latest_path).exists():
+                latest_size_in_mb = get_file_size_in_mb(Path(latest_path))
+                latest_num_frames = self.latest_episode["episode_index"][0]
+
+                av_size_per_frame = latest_size_in_mb / latest_num_frames if latest_num_frames > 0 else 0.0
+
+                if latest_size_in_mb + av_size_per_frame * num_frames >= self.data_files_size_in_mb:
+                    # Size limit is reached, flush buffer and prepare new parquet file
+                    self._flush_metadata_buffer()
+                    chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, self.chunks_size)
+                    self._close_writer()
+
+            # Update the existing pandas dataframe with new row
+            episode_dict["meta/episodes/chunk_index"] = [chunk_idx]
+            episode_dict["meta/episodes/file_index"] = [file_idx]
+            episode_dict["dataset_from_index"] = [self.latest_episode["dataset_to_index"][0]]
+            episode_dict["dataset_to_index"] = [self.latest_episode["dataset_to_index"][0] + num_frames]
+
+        # Add to buffer
+        self.metadata_buffer.append(episode_dict)
+        self.latest_episode = episode_dict
+
+        if len(self.metadata_buffer) >= self.metadata_buffer_size:
+            self._flush_metadata_buffer()
+
+    def save_episode(
+        self,
+        episode_index: int,
+        episode_length: int,
+        episode_tasks: list[str],
+        episode_stats: dict[str, dict],
+        episode_metadata: dict,
+    ) -> None:
+        episode_dict = {
+            "episode_index": episode_index,
+            "tasks": episode_tasks,
+            "length": episode_length,
+        }
+        episode_dict.update(episode_metadata)
+        episode_dict.update(flatten_dict({"stats": episode_stats}))
+        self._save_episode_metadata(episode_dict)
+
+        # Update info
+        self.info["total_episodes"] += 1
+        self.info["total_frames"] += episode_length
+        self.info["total_tasks"] = len(self.tasks)
+        self.info["splits"] = {"train": f"0:{self.info['total_episodes']}"}
+
+        write_info(self.info, self.root)
+
+        self.stats = aggregate_stats([self.stats, episode_stats]) if self.stats is not None else episode_stats
+        write_stats(self.stats, self.root)
+
+    def update_video_info(self, video_key: str | None = None) -> None:
+        """
+        Warning: this function writes info from first episode videos, implicitly assuming that all videos have
+        been encoded the same way. Also, this means it assumes the first episode exists.
+        """
+        if video_key is not None and video_key not in self.video_keys:
+            raise ValueError(f"Video key {video_key} not found in dataset")
+
+        video_keys = [video_key] if video_key is not None else self.video_keys
+        for key in video_keys:
+            if not self.features[key].get("info", None):
+                video_path = self.root / self.video_path.format(video_key=key, chunk_index=0, file_index=0)
+                self.info["features"][key]["info"] = get_video_info(video_path)
+
+    def update_chunk_settings(
+        self,
+        chunks_size: int | None = None,
+        data_files_size_in_mb: int | None = None,
+        video_files_size_in_mb: int | None = None,
+    ) -> None:
+        """Update chunk and file size settings after dataset creation.
+
+        This allows users to customize storage organization without modifying the constructor.
+        These settings control how episodes are chunked and how large files can grow before
+        creating new ones.
+
+        Args:
+            chunks_size: Maximum number of files per chunk directory. If None, keeps current value.
+            data_files_size_in_mb: Maximum size for data parquet files in MB. If None, keeps current value.
+            video_files_size_in_mb: Maximum size for video files in MB. If None, keeps current value.
+        """
+        if chunks_size is not None:
+            if chunks_size <= 0:
+                raise ValueError(f"chunks_size must be positive, got {chunks_size}")
+            self.info["chunks_size"] = chunks_size
+
+        if data_files_size_in_mb is not None:
+            if data_files_size_in_mb <= 0:
+                raise ValueError(f"data_files_size_in_mb must be positive, got {data_files_size_in_mb}")
+            self.info["data_files_size_in_mb"] = data_files_size_in_mb
+
+        if video_files_size_in_mb is not None:
+            if video_files_size_in_mb <= 0:
+                raise ValueError(f"video_files_size_in_mb must be positive, got {video_files_size_in_mb}")
+            self.info["video_files_size_in_mb"] = video_files_size_in_mb
+
+        # Update the info file on disk
+        write_info(self.info, self.root)
+
+    def get_chunk_settings(self) -> dict[str, int]:
+        """Get current chunk and file size settings.
+
+        Returns:
+            Dict containing chunks_size, data_files_size_in_mb, and video_files_size_in_mb.
+        """
+        return {
+            "chunks_size": self.chunks_size,
+            "data_files_size_in_mb": self.data_files_size_in_mb,
+            "video_files_size_in_mb": self.video_files_size_in_mb,
+        }
+
+    def __repr__(self):
+        feature_keys = list(self.features)
+        return (
+            f"{self.__class__.__name__}({{\n"
+            f"    Repository ID: '{self.repo_id}',\n"
+            f"    Total episodes: '{self.total_episodes}',\n"
+            f"    Total frames: '{self.total_frames}',\n"
+            f"    Features: '{feature_keys}',\n"
+            "})',\n"
+        )
+
+    @classmethod
+    def create(
+        cls,
+        repo_id: str,
+        fps: int,
+        features: dict,
+        robot_type: str | None = None,
+        root: str | Path | None = None,
+        use_videos: bool = True,
+        metadata_buffer_size: int = 10,
+        chunks_size: int | None = None,
+        data_files_size_in_mb: int | None = None,
+        video_files_size_in_mb: int | None = None,
+    ) -> "LeRobotDatasetMetadata":
+        """Creates metadata for a LeRobotDataset."""
+        obj = cls.__new__(cls)
+        obj.repo_id = repo_id
+        obj.root = Path(root) if root is not None else HF_LEROBOT_HOME / repo_id
+
+        obj.root.mkdir(parents=True, exist_ok=False)
+
+        features = {**features, **DEFAULT_FEATURES}
+        _validate_feature_names(features)
+
+        obj.tasks = None
+        obj.subtasks = None
+        obj.episodes = None
+        obj.stats = None
+        obj.info = create_empty_dataset_info(
+            CODEBASE_VERSION,
+            fps,
+            features,
+            use_videos,
+            robot_type,
+            chunks_size,
+            data_files_size_in_mb,
+            video_files_size_in_mb,
+        )
+        if len(obj.video_keys) > 0 and not use_videos:
+            raise ValueError(
+                f"Features contain video keys {obj.video_keys}, but 'use_videos' is set to False. "
+                "Either remove video features from the features dict, or set 'use_videos=True'."
+            )
+        write_json(obj.info, obj.root / INFO_PATH)
+        obj.revision = None
+        obj.writer = None
+        obj.latest_episode = None
+        obj.metadata_buffer = []
+        obj.metadata_buffer_size = metadata_buffer_size
+        return obj

src/lerobot/datasets/dataset_tools.py

@@ -38,19 +38,22 @@ from tqdm import tqdm
 
 from lerobot.datasets.aggregate import aggregate_datasets
 from lerobot.datasets.compute_stats import aggregate_stats
-from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
+from lerobot.datasets.io_utils import (
+    get_parquet_file_size_in_mb,
+    load_episodes,
+    write_info,
+    write_stats,
+    write_tasks,
+)
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.datasets.utils import (
     DATA_DIR,
     DEFAULT_CHUNK_SIZE,
     DEFAULT_DATA_FILE_SIZE_IN_MB,
     DEFAULT_DATA_PATH,
     DEFAULT_EPISODES_PATH,
-    get_parquet_file_size_in_mb,
-    load_episodes,
     update_chunk_file_indices,
-    write_info,
-    write_stats,
-    write_tasks,
 )
 from lerobot.datasets.video_utils import encode_video_frames, get_video_info
 from lerobot.utils.constants import HF_LEROBOT_HOME, OBS_IMAGE
@@ -89,8 +92,8 @@ def delete_episodes(
     Args:
         dataset: The source LeRobotDataset.
         episode_indices: List of episode indices to delete.
-        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.
+        output_dir: Root directory where the edited dataset will be stored. If not specified, defaults to $HF_LEROBOT_HOME/repo_id. Equivalent to new_root in EditDatasetConfig.
+        repo_id: Edited dataset identifier. Equivalent to new_repo_id in EditDatasetConfig.
     """
     if not episode_indices:
         raise ValueError("No episodes to delete")
@@ -152,7 +155,7 @@ def split_dataset(
         dataset: The source LeRobotDataset to split.
         splits: Either a dict mapping split names to episode indices, or a dict mapping
                 split names to fractions (must sum to <= 1.0).
-        output_dir: Base directory for output datasets. If None, uses default location.
+        output_dir: Root directory where the split datasets will be stored. If not specified, defaults to $HF_LEROBOT_HOME/repo_id.
 
     Examples:
       Split by specific episodes
@@ -243,8 +246,8 @@ def merge_datasets(
 
     Args:
         datasets: List of LeRobotDatasets to merge.
-        output_repo_id: Repository ID for the merged dataset.
-        output_dir: Directory to save the merged dataset. If None, uses default location.
+        output_repo_id: Merged dataset identifier.
+        output_dir: Root directory where the merged dataset will be stored. If not specified, defaults to $HF_LEROBOT_HOME/output_repo_id.
     """
     if not datasets:
         raise ValueError("No datasets to merge")
@@ -288,8 +291,8 @@ def modify_features(
         dataset: The source LeRobotDataset.
         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.
+        output_dir: Root directory where the edited dataset will be stored. If not specified, defaults to $HF_LEROBOT_HOME/repo_id. Equivalent to new_root in EditDatasetConfig.
+        repo_id: Edited dataset identifier. Equivalent to new_repo_id in EditDatasetConfig.
 
     Returns:
         New dataset with features modified.
@@ -390,8 +393,8 @@ def add_features(
     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.
+        output_dir: Root directory where the edited dataset will be stored. If not specified, defaults to $HF_LEROBOT_HOME/repo_id. Equivalent to new_root in EditDatasetConfig.
+        repo_id: Edited dataset identifier. Equivalent to new_repo_id in EditDatasetConfig.
 
     Returns:
         New dataset with all features added.
@@ -427,8 +430,8 @@ def remove_feature(
     Args:
         dataset: The source LeRobotDataset.
         feature_names: Name(s) of features 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.
+        output_dir: Root directory where the edited dataset will be stored. If not specified, defaults to $HF_LEROBOT_HOME/repo_id. Equivalent to new_root in EditDatasetConfig.
+        repo_id: Edited dataset identifier. Equivalent to new_repo_id in EditDatasetConfig.
 
     Returns:
         New dataset with features removed.
@@ -915,7 +918,8 @@ def _write_parquet(df: pd.DataFrame, path: Path, meta: LeRobotDatasetMetadata) -
 
     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
+    from lerobot.datasets.feature_utils import get_hf_features_from_features
+    from lerobot.datasets.io_utils import embed_images
 
     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")
@@ -1475,7 +1479,9 @@ def modify_tasks(
 
     # Collect all unique tasks and create new task mapping
     unique_tasks = sorted(set(episode_to_task.values()))
-    new_task_df = pd.DataFrame({"task_index": list(range(len(unique_tasks)))}, index=unique_tasks)
+    new_task_df = pd.DataFrame(
+        {"task_index": list(range(len(unique_tasks)))}, index=pd.Index(unique_tasks, name="task")
+    )
     task_to_index = {task: idx for idx, task in enumerate(unique_tasks)}
 
     logging.info(f"Modifying tasks in {dataset.repo_id}")
@@ -1529,7 +1535,7 @@ def modify_tasks(
 
 def convert_image_to_video_dataset(
     dataset: LeRobotDataset,
-    output_dir: Path,
+    output_dir: Path | None = None,
     repo_id: str | None = None,
     vcodec: str = "libsvtav1",
     pix_fmt: str = "yuv420p",
@@ -1548,8 +1554,8 @@ def convert_image_to_video_dataset(
 
     Args:
         dataset: The source LeRobot dataset with images
-        output_dir: Directory to save the new video dataset
-        repo_id: Repository ID for the new dataset (default: original_id + "_video")
+        output_dir: Root directory where the edited dataset will be stored. If not specified, defaults to $HF_LEROBOT_HOME/repo_id. Equivalent to new_root in EditDatasetConfig.
+        repo_id: Edited dataset identifier. Equivalent to new_repo_id in EditDatasetConfig.
         vcodec: Video codec (default: libsvtav1)
         pix_fmt: Pixel format (default: yuv420p)
         g: Group of pictures size (default: 2)
@@ -1600,6 +1606,7 @@ def convert_image_to_video_dataset(
             # Video info will be updated after episodes are encoded
 
     # Create new metadata for video dataset
+    output_dir = Path(output_dir) if output_dir is not None else HF_LEROBOT_HOME / repo_id
     new_meta = LeRobotDatasetMetadata.create(
         repo_id=repo_id,
         fps=dataset.meta.fps,

src/lerobot/datasets/factory.py

@@ -20,11 +20,9 @@ import torch
 
 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.configs.train import TrainPipelineConfig
-from lerobot.datasets.lerobot_dataset import (
-    LeRobotDataset,
-    LeRobotDatasetMetadata,
-    MultiLeRobotDataset,
-)
+from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.multi_dataset import MultiLeRobotDataset
 from lerobot.datasets.streaming_dataset import StreamingLeRobotDataset
 from lerobot.datasets.transforms import ImageTransforms
 from lerobot.utils.constants import ACTION, OBS_PREFIX, REWARD

src/lerobot/datasets/feature_utils.py

@@ -0,0 +1,552 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from pprint import pformat
+from typing import Any
+
+import datasets
+import numpy as np
+from PIL import Image as PILImage
+
+from lerobot.configs.types import FeatureType, PolicyFeature
+from lerobot.datasets.utils import (
+    DEFAULT_CHUNK_SIZE,
+    DEFAULT_DATA_FILE_SIZE_IN_MB,
+    DEFAULT_DATA_PATH,
+    DEFAULT_FEATURES,
+    DEFAULT_VIDEO_FILE_SIZE_IN_MB,
+    DEFAULT_VIDEO_PATH,
+)
+from lerobot.utils.constants import ACTION, OBS_ENV_STATE, OBS_STR
+from lerobot.utils.utils import is_valid_numpy_dtype_string
+
+
+def get_hf_features_from_features(features: dict) -> datasets.Features:
+    """Convert a LeRobot features dictionary to a `datasets.Features` object.
+
+    Args:
+        features (dict): A LeRobot-style feature dictionary.
+
+    Returns:
+        datasets.Features: The corresponding Hugging Face `datasets.Features` object.
+
+    Raises:
+        ValueError: If a feature has an unsupported shape.
+    """
+    hf_features = {}
+    for key, ft in features.items():
+        if ft["dtype"] == "video":
+            continue
+        elif ft["dtype"] == "image":
+            hf_features[key] = datasets.Image()
+        elif ft["shape"] == (1,):
+            hf_features[key] = datasets.Value(dtype=ft["dtype"])
+        elif len(ft["shape"]) == 1:
+            hf_features[key] = datasets.Sequence(
+                length=ft["shape"][0], feature=datasets.Value(dtype=ft["dtype"])
+            )
+        elif len(ft["shape"]) == 2:
+            hf_features[key] = datasets.Array2D(shape=ft["shape"], dtype=ft["dtype"])
+        elif len(ft["shape"]) == 3:
+            hf_features[key] = datasets.Array3D(shape=ft["shape"], dtype=ft["dtype"])
+        elif len(ft["shape"]) == 4:
+            hf_features[key] = datasets.Array4D(shape=ft["shape"], dtype=ft["dtype"])
+        elif len(ft["shape"]) == 5:
+            hf_features[key] = datasets.Array5D(shape=ft["shape"], dtype=ft["dtype"])
+        else:
+            raise ValueError(f"Corresponding feature is not valid: {ft}")
+
+    return datasets.Features(hf_features)
+
+
+def _validate_feature_names(features: dict[str, dict]) -> None:
+    """Validate that feature names do not contain invalid characters.
+
+    Args:
+        features (dict): The LeRobot features dictionary.
+
+    Raises:
+        ValueError: If any feature name contains '/'.
+    """
+    invalid_features = {name: ft for name, ft in features.items() if "/" in name}
+    if invalid_features:
+        raise ValueError(f"Feature names should not contain '/'. Found '/' in '{invalid_features}'.")
+
+
+def hw_to_dataset_features(
+    hw_features: dict[str, type | tuple], prefix: str, use_video: bool = True
+) -> dict[str, dict]:
+    """Convert hardware-specific features to a LeRobot dataset feature dictionary.
+
+    This function takes a dictionary describing hardware outputs (like joint states
+    or camera image shapes) and formats it into the standard LeRobot feature
+    specification.
+
+    Args:
+        hw_features (dict): Dictionary mapping feature names to their type (float for
+            joints) or shape (tuple for images).
+        prefix (str): The prefix to add to the feature keys (e.g., "observation"
+            or "action").
+        use_video (bool): If True, image features are marked as "video", otherwise "image".
+
+    Returns:
+        dict: A LeRobot features dictionary.
+    """
+    features = {}
+    joint_fts = {
+        key: ftype
+        for key, ftype in hw_features.items()
+        if ftype is float or (isinstance(ftype, PolicyFeature) and ftype.type != FeatureType.VISUAL)
+    }
+    cam_fts = {key: shape for key, shape in hw_features.items() if isinstance(shape, tuple)}
+
+    if joint_fts and prefix == ACTION:
+        features[prefix] = {
+            "dtype": "float32",
+            "shape": (len(joint_fts),),
+            "names": list(joint_fts),
+        }
+
+    if joint_fts and prefix == OBS_STR:
+        features[f"{prefix}.state"] = {
+            "dtype": "float32",
+            "shape": (len(joint_fts),),
+            "names": list(joint_fts),
+        }
+
+    for key, shape in cam_fts.items():
+        features[f"{prefix}.images.{key}"] = {
+            "dtype": "video" if use_video else "image",
+            "shape": shape,
+            "names": ["height", "width", "channels"],
+        }
+
+    _validate_feature_names(features)
+    return features
+
+
+def build_dataset_frame(
+    ds_features: dict[str, dict], values: dict[str, Any], prefix: str
+) -> dict[str, np.ndarray]:
+    """Construct a single data frame from raw values based on dataset features.
+
+    A "frame" is a dictionary containing all the data for a single timestep,
+    formatted as numpy arrays according to the feature specification.
+
+    Args:
+        ds_features (dict): The LeRobot dataset features dictionary.
+        values (dict): A dictionary of raw values from the hardware/environment.
+        prefix (str): The prefix to filter features by (e.g., "observation"
+            or "action").
+
+    Returns:
+        dict: A dictionary representing a single frame of data.
+    """
+    frame = {}
+    for key, ft in ds_features.items():
+        if key in DEFAULT_FEATURES or not key.startswith(prefix):
+            continue
+        elif ft["dtype"] == "float32" and len(ft["shape"]) == 1:
+            frame[key] = np.array([values[name] for name in ft["names"]], dtype=np.float32)
+        elif ft["dtype"] in ["image", "video"]:
+            frame[key] = values[key.removeprefix(f"{prefix}.images.")]
+
+    return frame
+
+
+def dataset_to_policy_features(features: dict[str, dict]) -> dict[str, PolicyFeature]:
+    """Convert dataset features to policy features.
+
+    This function transforms the dataset's feature specification into a format
+    that a policy can use, classifying features by type (e.g., visual, state,
+    action) and ensuring correct shapes (e.g., channel-first for images).
+
+    Args:
+        features (dict): The LeRobot dataset features dictionary.
+
+    Returns:
+        dict: A dictionary mapping feature keys to `PolicyFeature` objects.
+
+    Raises:
+        ValueError: If an image feature does not have a 3D shape.
+    """
+    # TODO(aliberts): Implement "type" in dataset features and simplify this
+    policy_features = {}
+    for key, ft in features.items():
+        shape = ft["shape"]
+        if ft["dtype"] in ["image", "video"]:
+            type = FeatureType.VISUAL
+            if len(shape) != 3:
+                raise ValueError(f"Number of dimensions of {key} != 3 (shape={shape})")
+
+            names = ft["names"]
+            # Backward compatibility for "channel" which is an error introduced in LeRobotDataset v2.0 for ported datasets.
+            if names[2] in ["channel", "channels"]:  # (h, w, c) -> (c, h, w)
+                shape = (shape[2], shape[0], shape[1])
+        elif key == OBS_ENV_STATE:
+            type = FeatureType.ENV
+        elif key.startswith(OBS_STR):
+            type = FeatureType.STATE
+        elif key.startswith(ACTION):
+            type = FeatureType.ACTION
+        else:
+            continue
+
+        policy_features[key] = PolicyFeature(
+            type=type,
+            shape=shape,
+        )
+
+    return policy_features
+
+
+def combine_feature_dicts(*dicts: dict) -> dict:
+    """Merge LeRobot grouped feature dicts.
+
+    - For 1D numeric specs (dtype not image/video/string) with "names": we merge the names and recompute the shape.
+    - For others (e.g. `observation.images.*`), the last one wins (if they are identical).
+
+    Args:
+        *dicts: A variable number of LeRobot feature dictionaries to merge.
+
+    Returns:
+        dict: A single merged feature dictionary.
+
+    Raises:
+        ValueError: If there's a dtype mismatch for a feature being merged.
+    """
+    out: dict = {}
+    for d in dicts:
+        for key, value in d.items():
+            if not isinstance(value, dict):
+                out[key] = value
+                continue
+
+            dtype = value.get("dtype")
+            shape = value.get("shape")
+            is_vector = (
+                dtype not in ("image", "video", "string")
+                and isinstance(shape, tuple)
+                and len(shape) == 1
+                and "names" in value
+            )
+
+            if is_vector:
+                # Initialize or retrieve the accumulating dict for this feature key
+                target = out.setdefault(key, {"dtype": dtype, "names": [], "shape": (0,)})
+                # Ensure consistent data types across merged entries
+                if "dtype" in target and dtype != target["dtype"]:
+                    raise ValueError(f"dtype mismatch for '{key}': {target['dtype']} vs {dtype}")
+
+                # Merge feature names: append only new ones to preserve order without duplicates
+                seen = set(target["names"])
+                for n in value["names"]:
+                    if n not in seen:
+                        target["names"].append(n)
+                        seen.add(n)
+                # Recompute the shape to reflect the updated number of features
+                target["shape"] = (len(target["names"]),)
+            else:
+                # For images/videos and non-1D entries: override with the latest definition
+                out[key] = value
+    return out
+
+
+def create_empty_dataset_info(
+    codebase_version: str,
+    fps: int,
+    features: dict,
+    use_videos: bool,
+    robot_type: str | None = None,
+    chunks_size: int | None = None,
+    data_files_size_in_mb: int | None = None,
+    video_files_size_in_mb: int | None = None,
+) -> dict:
+    """Create a template dictionary for a new dataset's `info.json`.
+
+    Args:
+        codebase_version (str): The version of the LeRobot codebase.
+        fps (int): The frames per second of the data.
+        features (dict): The LeRobot features dictionary for the dataset.
+        use_videos (bool): Whether the dataset will store videos.
+        robot_type (str | None): The type of robot used, if any.
+
+    Returns:
+        dict: A dictionary with the initial dataset metadata.
+    """
+    return {
+        "codebase_version": codebase_version,
+        "robot_type": robot_type,
+        "total_episodes": 0,
+        "total_frames": 0,
+        "total_tasks": 0,
+        "chunks_size": chunks_size or DEFAULT_CHUNK_SIZE,
+        "data_files_size_in_mb": data_files_size_in_mb or DEFAULT_DATA_FILE_SIZE_IN_MB,
+        "video_files_size_in_mb": video_files_size_in_mb or DEFAULT_VIDEO_FILE_SIZE_IN_MB,
+        "fps": fps,
+        "splits": {},
+        "data_path": DEFAULT_DATA_PATH,
+        "video_path": DEFAULT_VIDEO_PATH if use_videos else None,
+        "features": features,
+    }
+
+
+def check_delta_timestamps(
+    delta_timestamps: dict[str, list[float]], fps: int, tolerance_s: float, raise_value_error: bool = True
+) -> bool:
+    """Check if delta timestamps are multiples of 1/fps +/- tolerance.
+
+    This ensures that adding these delta timestamps to any existing timestamp in
+    the dataset will result in a value that aligns with the dataset's frame rate.
+
+    Args:
+        delta_timestamps (dict): A dictionary where values are lists of time
+            deltas in seconds.
+        fps (int): The frames per second of the dataset.
+        tolerance_s (float): The allowed tolerance in seconds.
+        raise_value_error (bool): If True, raises an error on failure.
+
+    Returns:
+        bool: True if all deltas are valid, False otherwise.
+
+    Raises:
+        ValueError: If any delta is outside the tolerance and `raise_value_error` is True.
+    """
+    outside_tolerance = {}
+    for key, delta_ts in delta_timestamps.items():
+        within_tolerance = [abs(ts * fps - round(ts * fps)) / fps <= tolerance_s for ts in delta_ts]
+        if not all(within_tolerance):
+            outside_tolerance[key] = [
+                ts for ts, is_within in zip(delta_ts, within_tolerance, strict=True) if not is_within
+            ]
+
+    if len(outside_tolerance) > 0:
+        if raise_value_error:
+            raise ValueError(
+                f"""
+                The following delta_timestamps are found outside of tolerance range.
+                Please make sure they are multiples of 1/{fps} +/- tolerance and adjust
+                their values accordingly.
+                \n{pformat(outside_tolerance)}
+                """
+            )
+        return False
+
+    return True
+
+
+def get_delta_indices(delta_timestamps: dict[str, list[float]], fps: int) -> dict[str, list[int]]:
+    """Convert delta timestamps in seconds to delta indices in frames.
+
+    Args:
+        delta_timestamps (dict): A dictionary of time deltas in seconds.
+        fps (int): The frames per second of the dataset.
+
+    Returns:
+        dict: A dictionary of frame delta indices.
+    """
+    delta_indices = {}
+    for key, delta_ts in delta_timestamps.items():
+        delta_indices[key] = [round(d * fps) for d in delta_ts]
+
+    return delta_indices
+
+
+def validate_frame(frame: dict, features: dict) -> None:
+    expected_features = set(features) - set(DEFAULT_FEATURES)
+    actual_features = set(frame)
+
+    # task is a special required field that's not part of regular features
+    if "task" not in actual_features:
+        raise ValueError("Feature mismatch in `frame` dictionary:\nMissing features: {'task'}\n")
+
+    # Remove task from actual_features for regular feature validation
+    actual_features_for_validation = actual_features - {"task"}
+
+    error_message = validate_features_presence(actual_features_for_validation, expected_features)
+
+    common_features = actual_features_for_validation & expected_features
+    for name in common_features:
+        error_message += validate_feature_dtype_and_shape(name, features[name], frame[name])
+
+    if error_message:
+        raise ValueError(error_message)
+
+
+def validate_features_presence(actual_features: set[str], expected_features: set[str]) -> str:
+    """Check for missing or extra features in a frame.
+
+    Args:
+        actual_features (set[str]): The set of feature names present in the frame.
+        expected_features (set[str]): The set of feature names expected in the frame.
+
+    Returns:
+        str: An error message string if there's a mismatch, otherwise an empty string.
+    """
+    error_message = ""
+    missing_features = expected_features - actual_features
+    extra_features = actual_features - expected_features
+
+    if missing_features or extra_features:
+        error_message += "Feature mismatch in `frame` dictionary:\n"
+        if missing_features:
+            error_message += f"Missing features: {missing_features}\n"
+        if extra_features:
+            error_message += f"Extra features: {extra_features}\n"
+
+    return error_message
+
+
+def validate_feature_dtype_and_shape(
+    name: str, feature: dict, value: np.ndarray | PILImage.Image | str
+) -> str:
+    """Validate the dtype and shape of a single feature's value.
+
+    Args:
+        name (str): The name of the feature.
+        feature (dict): The feature specification from the LeRobot features dictionary.
+        value: The value of the feature to validate.
+
+    Returns:
+        str: An error message if validation fails, otherwise an empty string.
+
+    Raises:
+        NotImplementedError: If the feature dtype is not supported for validation.
+    """
+    expected_dtype = feature["dtype"]
+    expected_shape = feature["shape"]
+    if is_valid_numpy_dtype_string(expected_dtype):
+        return validate_feature_numpy_array(name, expected_dtype, expected_shape, value)
+    elif expected_dtype in ["image", "video"]:
+        return validate_feature_image_or_video(name, expected_shape, value)
+    elif expected_dtype == "string":
+        return validate_feature_string(name, value)
+    else:
+        raise NotImplementedError(f"The feature dtype '{expected_dtype}' is not implemented yet.")
+
+
+def validate_feature_numpy_array(
+    name: str, expected_dtype: str, expected_shape: list[int], value: np.ndarray
+) -> str:
+    """Validate a feature that is expected to be a numpy array.
+
+    Args:
+        name (str): The name of the feature.
+        expected_dtype (str): The expected numpy dtype as a string.
+        expected_shape (list[int]): The expected shape.
+        value (np.ndarray): The numpy array to validate.
+
+    Returns:
+        str: An error message if validation fails, otherwise an empty string.
+    """
+    error_message = ""
+    if isinstance(value, np.ndarray):
+        actual_dtype = value.dtype
+        actual_shape = value.shape
+
+        if actual_dtype != np.dtype(expected_dtype):
+            error_message += f"The feature '{name}' of dtype '{actual_dtype}' is not of the expected dtype '{expected_dtype}'.\n"
+
+        if actual_shape != expected_shape:
+            error_message += f"The feature '{name}' of shape '{actual_shape}' does not have the expected shape '{expected_shape}'.\n"
+    else:
+        error_message += f"The feature '{name}' is not a 'np.ndarray'. Expected type is '{expected_dtype}', but type '{type(value)}' provided instead.\n"
+
+    return error_message
+
+
+def validate_feature_image_or_video(
+    name: str, expected_shape: list[str], value: np.ndarray | PILImage.Image
+) -> str:
+    """Validate a feature that is expected to be an image or video frame.
+
+    Accepts `np.ndarray` (channel-first or channel-last) or `PIL.Image.Image`.
+
+    Args:
+        name (str): The name of the feature.
+        expected_shape (list[str]): The expected shape (C, H, W).
+        value: The image data to validate.
+
+    Returns:
+        str: An error message if validation fails, otherwise an empty string.
+    """
+    # Note: The check of pixels range ([0,1] for float and [0,255] for uint8) is done by the image writer threads.
+    error_message = ""
+    if isinstance(value, np.ndarray):
+        actual_shape = value.shape
+        c, h, w = expected_shape
+        if len(actual_shape) != 3 or (actual_shape != (c, h, w) and actual_shape != (h, w, c)):
+            error_message += f"The feature '{name}' of shape '{actual_shape}' does not have the expected shape '{(c, h, w)}' or '{(h, w, c)}'.\n"
+    elif isinstance(value, PILImage.Image):
+        pass
+    else:
+        error_message += f"The feature '{name}' is expected to be of type 'PIL.Image' or 'np.ndarray' channel first or channel last, but type '{type(value)}' provided instead.\n"
+
+    return error_message
+
+
+def validate_feature_string(name: str, value: str) -> str:
+    """Validate a feature that is expected to be a string.
+
+    Args:
+        name (str): The name of the feature.
+        value (str): The value to validate.
+
+    Returns:
+        str: An error message if validation fails, otherwise an empty string.
+    """
+    if not isinstance(value, str):
+        return f"The feature '{name}' is expected to be of type 'str', but type '{type(value)}' provided instead.\n"
+    return ""
+
+
+def validate_episode_buffer(episode_buffer: dict, total_episodes: int, features: dict) -> None:
+    """Validate the episode buffer before it's written to disk.
+
+    Ensures the buffer has the required keys, contains at least one frame, and
+    has features consistent with the dataset's specification.
+
+    Args:
+        episode_buffer (dict): The buffer containing data for a single episode.
+        total_episodes (int): The current total number of episodes in the dataset.
+        features (dict): The LeRobot features dictionary for the dataset.
+
+    Raises:
+        ValueError: If the buffer is invalid.
+        NotImplementedError: If the episode index is manually set and doesn't match.
+    """
+    if "size" not in episode_buffer:
+        raise ValueError("size key not found in episode_buffer")
+
+    if "task" not in episode_buffer:
+        raise ValueError("task key not found in episode_buffer")
+
+    if episode_buffer["episode_index"] != total_episodes:
+        # TODO(aliberts): Add option to use existing episode_index
+        raise NotImplementedError(
+            "You might have manually provided the episode_buffer with an episode_index that doesn't "
+            "match the total number of episodes already in the dataset. This is not supported for now."
+        )
+
+    if episode_buffer["size"] == 0:
+        raise ValueError("You must add one or several frames with `add_frame` before calling `add_episode`.")
+
+    buffer_keys = set(episode_buffer.keys()) - {"task", "size"}
+    if not buffer_keys == set(features):
+        raise ValueError(
+            f"Features from `episode_buffer` don't match the ones in `features`."
+            f"In episode_buffer not in features: {buffer_keys - set(features)}"
+            f"In features not in episode_buffer: {set(features) - buffer_keys}"
+        )

src/lerobot/datasets/image_writer.py

@@ -13,6 +13,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+import logging
 import multiprocessing
 import queue
 import threading
@@ -22,6 +23,8 @@ import numpy as np
 import PIL.Image
 import torch
 
+logger = logging.getLogger(__name__)
+
 
 def safe_stop_image_writer(func):
     def wrapper(*args, **kwargs):
@@ -31,7 +34,7 @@ def safe_stop_image_writer(func):
             dataset = kwargs.get("dataset")
             image_writer = getattr(dataset, "image_writer", None) if dataset else None
             if image_writer is not None:
-                print("Waiting for image writer to terminate...")
+                logger.warning("Waiting for image writer to terminate...")
                 image_writer.stop()
             raise e
 
@@ -89,8 +92,7 @@ def write_image(image: np.ndarray | PIL.Image.Image, fpath: Path, compress_level
             PIL.Image.Image object.
 
     Side Effects:
-        Prints an error message to the console if the image writing process
-        fails for any reason.
+        Logs an error message if the image writing process fails for any reason.
     """
     try:
         if isinstance(image, np.ndarray):
@@ -101,7 +103,7 @@ def write_image(image: np.ndarray | PIL.Image.Image, fpath: Path, compress_level
             raise TypeError(f"Unsupported image type: {type(image)}")
         img.save(fpath, compress_level=compress_level)
     except Exception as e:
-        print(f"Error writing image {fpath}: {e}")
+        logger.error("Error writing image %s: %s", fpath, e)
 
 
 def worker_thread_loop(queue: queue.Queue):

src/lerobot/datasets/io_utils.py

@@ -0,0 +1,342 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import json
+from pathlib import Path
+from typing import Any
+
+import datasets
+import numpy as np
+import pandas
+import pandas as pd
+import pyarrow.dataset as pa_ds
+import pyarrow.parquet as pq
+import torch
+from datasets import Dataset
+from datasets.table import embed_table_storage
+from PIL import Image as PILImage
+from torchvision import transforms
+
+from lerobot.datasets.utils import (
+    DEFAULT_DATA_FILE_SIZE_IN_MB,
+    DEFAULT_EPISODES_PATH,
+    DEFAULT_SUBTASKS_PATH,
+    DEFAULT_TASKS_PATH,
+    EPISODES_DIR,
+    INFO_PATH,
+    STATS_PATH,
+    flatten_dict,
+    serialize_dict,
+    unflatten_dict,
+)
+from lerobot.utils.utils import SuppressProgressBars
+
+
+def get_parquet_file_size_in_mb(parquet_path: str | Path) -> float:
+    metadata = pq.read_metadata(parquet_path)
+    total_uncompressed_size = 0
+    for row_group in range(metadata.num_row_groups):
+        rg_metadata = metadata.row_group(row_group)
+        for column in range(rg_metadata.num_columns):
+            col_metadata = rg_metadata.column(column)
+            total_uncompressed_size += col_metadata.total_uncompressed_size
+    return total_uncompressed_size / (1024**2)
+
+
+def get_hf_dataset_size_in_mb(hf_ds: Dataset) -> int:
+    return hf_ds.data.nbytes // (1024**2)
+
+
+def load_nested_dataset(
+    pq_dir: Path, features: datasets.Features | None = None, episodes: list[int] | None = None
+) -> Dataset:
+    """Find parquet files in provided directory {pq_dir}/chunk-xxx/file-xxx.parquet
+    Convert parquet files to pyarrow memory mapped in a cache folder for efficient RAM usage
+    Concatenate all pyarrow references to return HF Dataset format
+
+    Args:
+        pq_dir: Directory containing parquet files
+        features: Optional features schema to ensure consistent loading of complex types like images
+        episodes: Optional list of episode indices to filter. Uses PyArrow predicate pushdown for efficiency.
+    """
+    paths = sorted(pq_dir.glob("*/*.parquet"))
+    if len(paths) == 0:
+        raise FileNotFoundError(f"Provided directory does not contain any parquet file: {pq_dir}")
+
+    with SuppressProgressBars():
+        # We use .from_parquet() memory-mapped loading for efficiency
+        filters = pa_ds.field("episode_index").isin(episodes) if episodes is not None else None
+        return Dataset.from_parquet([str(path) for path in paths], filters=filters, features=features)
+
+
+def get_parquet_num_frames(parquet_path: str | Path) -> int:
+    metadata = pq.read_metadata(parquet_path)
+    return metadata.num_rows
+
+
+def get_file_size_in_mb(file_path: Path) -> float:
+    """Get file size on disk in megabytes.
+
+    Args:
+        file_path (Path): Path to the file.
+    """
+    file_size_bytes = file_path.stat().st_size
+    return file_size_bytes / (1024**2)
+
+
+def embed_images(dataset: datasets.Dataset) -> datasets.Dataset:
+    """Embed image bytes into the dataset table before saving to Parquet.
+
+    This function prepares a Hugging Face dataset for serialization by converting
+    image objects into an embedded format that can be stored in Arrow/Parquet.
+
+    Args:
+        dataset (datasets.Dataset): The input dataset, possibly containing image features.
+
+    Returns:
+        datasets.Dataset: The dataset with images embedded in the table storage.
+    """
+    # Embed image bytes into the table before saving to parquet
+    format = dataset.format
+    dataset = dataset.with_format("arrow")
+    dataset = dataset.map(embed_table_storage, batched=False)
+    dataset = dataset.with_format(**format)
+    return dataset
+
+
+def load_json(fpath: Path) -> Any:
+    """Load data from a JSON file.
+
+    Args:
+        fpath (Path): Path to the JSON file.
+
+    Returns:
+        Any: The data loaded from the JSON file.
+    """
+    with open(fpath) as f:
+        return json.load(f)
+
+
+def write_json(data: dict, fpath: Path) -> None:
+    """Write data to a JSON file.
+
+    Creates parent directories if they don't exist.
+
+    Args:
+        data (dict): The dictionary to write.
+        fpath (Path): The path to the output JSON file.
+    """
+    fpath.parent.mkdir(exist_ok=True, parents=True)
+    with open(fpath, "w") as f:
+        json.dump(data, f, indent=4, ensure_ascii=False)
+
+
+def write_info(info: dict, local_dir: Path) -> None:
+    write_json(info, local_dir / INFO_PATH)
+
+
+def load_info(local_dir: Path) -> dict:
+    """Load dataset info metadata from its standard file path.
+
+    Also converts shape lists to tuples for consistency.
+
+    Args:
+        local_dir (Path): The root directory of the dataset.
+
+    Returns:
+        dict: The dataset information dictionary.
+    """
+    info = load_json(local_dir / INFO_PATH)
+    for ft in info["features"].values():
+        ft["shape"] = tuple(ft["shape"])
+    return info
+
+
+def write_stats(stats: dict, local_dir: Path) -> None:
+    """Serialize and write dataset statistics to their standard file path.
+
+    Args:
+        stats (dict): The statistics dictionary (can contain tensors/numpy arrays).
+        local_dir (Path): The root directory of the dataset.
+    """
+    serialized_stats = serialize_dict(stats)
+    write_json(serialized_stats, local_dir / STATS_PATH)
+
+
+def cast_stats_to_numpy(stats: dict) -> dict[str, dict[str, np.ndarray]]:
+    """Recursively cast numerical values in a stats dictionary to numpy arrays.
+
+    Args:
+        stats (dict): The statistics dictionary.
+
+    Returns:
+        dict: The statistics dictionary with values cast to numpy arrays.
+    """
+    stats = {key: np.array(value) for key, value in flatten_dict(stats).items()}
+    return unflatten_dict(stats)
+
+
+def load_stats(local_dir: Path) -> dict[str, dict[str, np.ndarray]] | None:
+    """Load dataset statistics and cast numerical values to numpy arrays.
+
+    Returns None if the stats file doesn't exist.
+
+    Args:
+        local_dir (Path): The root directory of the dataset.
+
+    Returns:
+        A dictionary of statistics or None if the file is not found.
+    """
+    if not (local_dir / STATS_PATH).exists():
+        return None
+    stats = load_json(local_dir / STATS_PATH)
+    return cast_stats_to_numpy(stats)
+
+
+def write_tasks(tasks: pandas.DataFrame, local_dir: Path) -> None:
+    path = local_dir / DEFAULT_TASKS_PATH
+    path.parent.mkdir(parents=True, exist_ok=True)
+    tasks.to_parquet(path)
+
+
+def load_tasks(local_dir: Path) -> pandas.DataFrame:
+    tasks = pd.read_parquet(local_dir / DEFAULT_TASKS_PATH)
+    tasks.index.name = "task"
+    return tasks
+
+
+def load_subtasks(local_dir: Path) -> pandas.DataFrame | None:
+    """Load subtasks from subtasks.parquet if it exists."""
+    subtasks_path = local_dir / DEFAULT_SUBTASKS_PATH
+    if subtasks_path.exists():
+        return pd.read_parquet(subtasks_path)
+    return None
+
+
+def write_episodes(episodes: Dataset, local_dir: Path) -> None:
+    """Write episode metadata to a parquet file in the LeRobot v3.0 format.
+    This function writes episode-level metadata to a single parquet file.
+    Used primarily during dataset conversion (v2.1 → v3.0) and in test fixtures.
+
+    Args:
+        episodes: HuggingFace Dataset containing episode metadata
+        local_dir: Root directory where the dataset will be stored
+    """
+    episode_size_mb = get_hf_dataset_size_in_mb(episodes)
+    if episode_size_mb > DEFAULT_DATA_FILE_SIZE_IN_MB:
+        raise NotImplementedError(
+            f"Episodes dataset is too large ({episode_size_mb} MB) to write to a single file. "
+            f"The current limit is {DEFAULT_DATA_FILE_SIZE_IN_MB} MB. "
+            "This function only supports single-file episode metadata. "
+        )
+
+    fpath = local_dir / DEFAULT_EPISODES_PATH.format(chunk_index=0, file_index=0)
+    fpath.parent.mkdir(parents=True, exist_ok=True)
+    episodes.to_parquet(fpath)
+
+
+def load_episodes(local_dir: Path) -> datasets.Dataset:
+    episodes = load_nested_dataset(local_dir / EPISODES_DIR)
+    # Select episode features/columns containing references to episode data and videos
+    # (e.g. tasks, dataset_from_index, dataset_to_index, data/chunk_index, data/file_index, etc.)
+    # This is to speedup access to these data, instead of having to load episode stats.
+    episodes = episodes.select_columns([key for key in episodes.features if not key.startswith("stats/")])
+    return episodes
+
+
+def load_image_as_numpy(
+    fpath: str | Path, dtype: np.dtype = np.float32, channel_first: bool = True
+) -> np.ndarray:
+    """Load an image from a file into a numpy array.
+
+    Args:
+        fpath (str | Path): Path to the image file.
+        dtype (np.dtype): The desired data type of the output array. If floating,
+            pixels are scaled to [0, 1].
+        channel_first (bool): If True, converts the image to (C, H, W) format.
+            Otherwise, it remains in (H, W, C) format.
+
+    Returns:
+        np.ndarray: The image as a numpy array.
+    """
+    img = PILImage.open(fpath).convert("RGB")
+    img_array = np.array(img, dtype=dtype)
+    if channel_first:  # (H, W, C) -> (C, H, W)
+        img_array = np.transpose(img_array, (2, 0, 1))
+    if np.issubdtype(dtype, np.floating):
+        img_array /= 255.0
+    return img_array
+
+
+def hf_transform_to_torch(items_dict: dict[str, list[Any]]) -> dict[str, list[torch.Tensor | str]]:
+    """Convert a batch from a Hugging Face dataset to torch tensors.
+
+    This transform function converts items from Hugging Face dataset format (pyarrow)
+    to torch tensors. Importantly, images are converted from PIL objects (H, W, C, uint8)
+    to a torch image representation (C, H, W, float32) in the range [0, 1]. Other
+    types are converted to torch.tensor.
+
+    Args:
+        items_dict (dict): A dictionary representing a batch of data from a
+            Hugging Face dataset.
+
+    Returns:
+        dict: The batch with items converted to torch tensors.
+    """
+    for key in items_dict:
+        first_item = items_dict[key][0]
+        if isinstance(first_item, PILImage.Image):
+            to_tensor = transforms.ToTensor()
+            items_dict[key] = [to_tensor(img) for img in items_dict[key]]
+        elif first_item is None:
+            pass
+        else:
+            items_dict[key] = [x if isinstance(x, str) else torch.tensor(x) for x in items_dict[key]]
+    return items_dict
+
+
+def to_parquet_with_hf_images(
+    df: pandas.DataFrame, path: Path, features: datasets.Features | None = None
+) -> None:
+    """This function correctly writes to parquet a panda DataFrame that contains images encoded by HF dataset.
+    This way, it can be loaded by HF dataset and correctly formatted images are returned.
+
+    Args:
+        df: DataFrame to write to parquet.
+        path: Path to write the parquet file.
+        features: Optional HuggingFace Features schema. If provided, ensures image columns
+                  are properly typed as Image() in the parquet schema.
+    """
+    # TODO(qlhoest): replace this weird synthax by `df.to_parquet(path)` only
+    ds = datasets.Dataset.from_dict(df.to_dict(orient="list"), features=features)
+    ds.to_parquet(path)
+
+
+def item_to_torch(item: dict) -> dict:
+    """Convert all items in a dictionary to PyTorch tensors where appropriate.
+
+    This function is used to convert an item from a streaming dataset to PyTorch tensors.
+
+    Args:
+        item (dict): Dictionary of items from a dataset.
+
+    Returns:
+        dict: Dictionary with all tensor-like items converted to torch.Tensor.
+    """
+    for key, val in item.items():
+        if isinstance(val, (np.ndarray | list)) and key not in ["task"]:
+            # Convert numpy arrays and lists to torch tensors
+            item[key] = torch.tensor(val)
+    return item

src/lerobot/datasets/lerobot_dataset.py

@@ -23,526 +23,52 @@ from pathlib import Path
 
 import datasets
 import numpy as np
-import packaging.version
 import pandas as pd
 import PIL.Image
-import pyarrow as pa
 import pyarrow.parquet as pq
 import torch
 import torch.utils
 from huggingface_hub import HfApi, snapshot_download
 from huggingface_hub.errors import RevisionNotFoundError
 
-from lerobot.datasets.compute_stats import aggregate_stats, compute_episode_stats
-from lerobot.datasets.image_writer import AsyncImageWriter, write_image
-from lerobot.datasets.utils import (
-    DEFAULT_EPISODES_PATH,
-    DEFAULT_FEATURES,
-    DEFAULT_IMAGE_PATH,
-    INFO_PATH,
-    _validate_feature_names,
+from lerobot.datasets.compute_stats import compute_episode_stats
+from lerobot.datasets.dataset_metadata import CODEBASE_VERSION, LeRobotDatasetMetadata
+from lerobot.datasets.feature_utils import (
     check_delta_timestamps,
-    check_version_compatibility,
-    create_empty_dataset_info,
-    create_lerobot_dataset_card,
-    embed_images,
-    flatten_dict,
     get_delta_indices,
-    get_file_size_in_mb,
     get_hf_features_from_features,
-    get_safe_version,
-    hf_transform_to_torch,
-    is_valid_version,
-    load_episodes,
-    load_info,
-    load_nested_dataset,
-    load_stats,
-    load_subtasks,
-    load_tasks,
-    update_chunk_file_indices,
     validate_episode_buffer,
     validate_frame,
+)
+from lerobot.datasets.image_writer import AsyncImageWriter, write_image
+from lerobot.datasets.io_utils import (
+    embed_images,
+    get_file_size_in_mb,
+    hf_transform_to_torch,
+    load_episodes,
+    load_nested_dataset,
     write_info,
-    write_json,
-    write_stats,
-    write_tasks,
+)
+from lerobot.datasets.utils import (
+    DEFAULT_EPISODES_PATH,
+    DEFAULT_IMAGE_PATH,
+    create_lerobot_dataset_card,
+    get_safe_version,
+    is_valid_version,
+    update_chunk_file_indices,
 )
 from lerobot.datasets.video_utils import (
     StreamingVideoEncoder,
-    VideoFrame,
     concatenate_video_files,
     decode_video_frames,
     encode_video_frames,
     get_safe_default_codec,
     get_video_duration_in_s,
-    get_video_info,
     resolve_vcodec,
 )
 from lerobot.utils.constants import HF_LEROBOT_HOME
 
-CODEBASE_VERSION = "v3.0"
-
-
-class LeRobotDatasetMetadata:
-    def __init__(
-        self,
-        repo_id: str,
-        root: str | Path | None = None,
-        revision: str | None = None,
-        force_cache_sync: bool = False,
-        metadata_buffer_size: int = 10,
-    ):
-        self.repo_id = repo_id
-        self.revision = revision if revision else CODEBASE_VERSION
-        self.root = Path(root) if root is not None else HF_LEROBOT_HOME / repo_id
-        self.writer = None
-        self.latest_episode = None
-        self.metadata_buffer: list[dict] = []
-        self.metadata_buffer_size = metadata_buffer_size
-
-        try:
-            if force_cache_sync:
-                raise FileNotFoundError
-            self.load_metadata()
-        except (FileNotFoundError, NotADirectoryError):
-            if is_valid_version(self.revision):
-                self.revision = get_safe_version(self.repo_id, self.revision)
-
-            (self.root / "meta").mkdir(exist_ok=True, parents=True)
-            self.pull_from_repo(allow_patterns="meta/")
-            self.load_metadata()
-
-    def _flush_metadata_buffer(self) -> None:
-        """Write all buffered episode metadata to parquet file."""
-        if not hasattr(self, "metadata_buffer") or len(self.metadata_buffer) == 0:
-            return
-
-        combined_dict = {}
-        for episode_dict in self.metadata_buffer:
-            for key, value in episode_dict.items():
-                if key not in combined_dict:
-                    combined_dict[key] = []
-                # Extract value and serialize numpy arrays
-                # because PyArrow's from_pydict function doesn't support numpy arrays
-                val = value[0] if isinstance(value, list) else value
-                combined_dict[key].append(val.tolist() if isinstance(val, np.ndarray) else val)
-
-        first_ep = self.metadata_buffer[0]
-        chunk_idx = first_ep["meta/episodes/chunk_index"][0]
-        file_idx = first_ep["meta/episodes/file_index"][0]
-
-        table = pa.Table.from_pydict(combined_dict)
-
-        if not self.writer:
-            path = Path(self.root / DEFAULT_EPISODES_PATH.format(chunk_index=chunk_idx, file_index=file_idx))
-            path.parent.mkdir(parents=True, exist_ok=True)
-            self.writer = pq.ParquetWriter(
-                path, schema=table.schema, compression="snappy", use_dictionary=True
-            )
-
-        self.writer.write_table(table)
-
-        self.latest_episode = self.metadata_buffer[-1]
-        self.metadata_buffer.clear()
-
-    def _close_writer(self) -> None:
-        """Close and cleanup the parquet writer if it exists."""
-        self._flush_metadata_buffer()
-
-        writer = getattr(self, "writer", None)
-        if writer is not None:
-            writer.close()
-            self.writer = None
-
-    def __del__(self):
-        """
-        Trust the user to call .finalize() but as an added safety check call the parquet writer to stop when calling the destructor
-        """
-        self._close_writer()
-
-    def load_metadata(self):
-        self.info = load_info(self.root)
-        check_version_compatibility(self.repo_id, self._version, CODEBASE_VERSION)
-        self.tasks = load_tasks(self.root)
-        self.subtasks = load_subtasks(self.root)
-        self.episodes = load_episodes(self.root)
-        self.stats = load_stats(self.root)
-
-    def pull_from_repo(
-        self,
-        allow_patterns: list[str] | str | None = None,
-        ignore_patterns: list[str] | str | None = None,
-    ) -> None:
-        snapshot_download(
-            self.repo_id,
-            repo_type="dataset",
-            revision=self.revision,
-            local_dir=self.root,
-            allow_patterns=allow_patterns,
-            ignore_patterns=ignore_patterns,
-        )
-
-    @property
-    def url_root(self) -> str:
-        return f"hf://datasets/{self.repo_id}"
-
-    @property
-    def _version(self) -> packaging.version.Version:
-        """Codebase version used to create this dataset."""
-        return packaging.version.parse(self.info["codebase_version"])
-
-    def get_data_file_path(self, ep_index: int) -> Path:
-        if self.episodes is None:
-            self.episodes = load_episodes(self.root)
-        if ep_index >= len(self.episodes):
-            raise IndexError(
-                f"Episode index {ep_index} out of range. Episodes: {len(self.episodes) if self.episodes else 0}"
-            )
-        ep = self.episodes[ep_index]
-        chunk_idx = ep["data/chunk_index"]
-        file_idx = ep["data/file_index"]
-        fpath = self.data_path.format(chunk_index=chunk_idx, file_index=file_idx)
-        return Path(fpath)
-
-    def get_video_file_path(self, ep_index: int, vid_key: str) -> Path:
-        if self.episodes is None:
-            self.episodes = load_episodes(self.root)
-        if ep_index >= len(self.episodes):
-            raise IndexError(
-                f"Episode index {ep_index} out of range. Episodes: {len(self.episodes) if self.episodes else 0}"
-            )
-        ep = self.episodes[ep_index]
-        chunk_idx = ep[f"videos/{vid_key}/chunk_index"]
-        file_idx = ep[f"videos/{vid_key}/file_index"]
-        fpath = self.video_path.format(video_key=vid_key, chunk_index=chunk_idx, file_index=file_idx)
-        return Path(fpath)
-
-    @property
-    def data_path(self) -> str:
-        """Formattable string for the parquet files."""
-        return self.info["data_path"]
-
-    @property
-    def video_path(self) -> str | None:
-        """Formattable string for the video files."""
-        return self.info["video_path"]
-
-    @property
-    def robot_type(self) -> str | None:
-        """Robot type used in recording this dataset."""
-        return self.info["robot_type"]
-
-    @property
-    def fps(self) -> int:
-        """Frames per second used during data collection."""
-        return self.info["fps"]
-
-    @property
-    def features(self) -> dict[str, dict]:
-        """All features contained in the dataset."""
-        return self.info["features"]
-
-    @property
-    def image_keys(self) -> list[str]:
-        """Keys to access visual modalities stored as images."""
-        return [key for key, ft in self.features.items() if ft["dtype"] == "image"]
-
-    @property
-    def video_keys(self) -> list[str]:
-        """Keys to access visual modalities stored as videos."""
-        return [key for key, ft in self.features.items() if ft["dtype"] == "video"]
-
-    @property
-    def camera_keys(self) -> list[str]:
-        """Keys to access visual modalities (regardless of their storage method)."""
-        return [key for key, ft in self.features.items() if ft["dtype"] in ["video", "image"]]
-
-    @property
-    def names(self) -> dict[str, list | dict]:
-        """Names of the various dimensions of vector modalities."""
-        return {key: ft["names"] for key, ft in self.features.items()}
-
-    @property
-    def shapes(self) -> dict:
-        """Shapes for the different features."""
-        return {key: tuple(ft["shape"]) for key, ft in self.features.items()}
-
-    @property
-    def total_episodes(self) -> int:
-        """Total number of episodes available."""
-        return self.info["total_episodes"]
-
-    @property
-    def total_frames(self) -> int:
-        """Total number of frames saved in this dataset."""
-        return self.info["total_frames"]
-
-    @property
-    def total_tasks(self) -> int:
-        """Total number of different tasks performed in this dataset."""
-        return self.info["total_tasks"]
-
-    @property
-    def chunks_size(self) -> int:
-        """Max number of files per chunk."""
-        return self.info["chunks_size"]
-
-    @property
-    def data_files_size_in_mb(self) -> int:
-        """Max size of data file in mega bytes."""
-        return self.info["data_files_size_in_mb"]
-
-    @property
-    def video_files_size_in_mb(self) -> int:
-        """Max size of video file in mega bytes."""
-        return self.info["video_files_size_in_mb"]
-
-    def get_task_index(self, task: str) -> int | None:
-        """
-        Given a task in natural language, returns its task_index if the task already exists in the dataset,
-        otherwise return None.
-        """
-        if task in self.tasks.index:
-            return int(self.tasks.loc[task].task_index)
-        else:
-            return None
-
-    def save_episode_tasks(self, tasks: list[str]):
-        if len(set(tasks)) != len(tasks):
-            raise ValueError(f"Tasks are not unique: {tasks}")
-
-        if self.tasks is None:
-            new_tasks = tasks
-            task_indices = range(len(tasks))
-            self.tasks = pd.DataFrame({"task_index": task_indices}, index=tasks)
-        else:
-            new_tasks = [task for task in tasks if task not in self.tasks.index]
-            new_task_indices = range(len(self.tasks), len(self.tasks) + len(new_tasks))
-            for task_idx, task in zip(new_task_indices, new_tasks, strict=False):
-                self.tasks.loc[task] = task_idx
-
-        if len(new_tasks) > 0:
-            # Update on disk
-            write_tasks(self.tasks, self.root)
-
-    def _save_episode_metadata(self, episode_dict: dict) -> None:
-        """Buffer episode metadata and write to parquet in batches for efficiency.
-
-        This function accumulates episode metadata in a buffer and flushes it when the buffer
-        reaches the configured size. This reduces I/O overhead by writing multiple episodes
-        at once instead of one row at a time.
-
-        Notes: We both need to update parquet files and HF dataset:
-        - `pandas` loads parquet file in RAM
-        - `datasets` relies on a memory mapping from pyarrow (no RAM). It either converts parquet files to a pyarrow cache on disk,
-          or loads directly from pyarrow cache.
-        """
-        # Convert to list format for each value
-        episode_dict = {key: [value] for key, value in episode_dict.items()}
-        num_frames = episode_dict["length"][0]
-
-        if self.latest_episode is None:
-            # Initialize indices and frame count for a new dataset made of the first episode data
-            chunk_idx, file_idx = 0, 0
-            if self.episodes is not None and len(self.episodes) > 0:
-                # It means we are resuming recording, so we need to load the latest episode
-                # Update the indices to avoid overwriting the latest episode
-                chunk_idx = self.episodes[-1]["meta/episodes/chunk_index"]
-                file_idx = self.episodes[-1]["meta/episodes/file_index"]
-                latest_num_frames = self.episodes[-1]["dataset_to_index"]
-                episode_dict["dataset_from_index"] = [latest_num_frames]
-                episode_dict["dataset_to_index"] = [latest_num_frames + num_frames]
-
-                # When resuming, move to the next file
-                chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, self.chunks_size)
-            else:
-                episode_dict["dataset_from_index"] = [0]
-                episode_dict["dataset_to_index"] = [num_frames]
-
-            episode_dict["meta/episodes/chunk_index"] = [chunk_idx]
-            episode_dict["meta/episodes/file_index"] = [file_idx]
-        else:
-            chunk_idx = self.latest_episode["meta/episodes/chunk_index"][0]
-            file_idx = self.latest_episode["meta/episodes/file_index"][0]
-
-            latest_path = (
-                self.root / DEFAULT_EPISODES_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
-                if self.writer is None
-                else self.writer.where
-            )
-
-            if Path(latest_path).exists():
-                latest_size_in_mb = get_file_size_in_mb(Path(latest_path))
-                latest_num_frames = self.latest_episode["episode_index"][0]
-
-                av_size_per_frame = latest_size_in_mb / latest_num_frames if latest_num_frames > 0 else 0.0
-
-                if latest_size_in_mb + av_size_per_frame * num_frames >= self.data_files_size_in_mb:
-                    # Size limit is reached, flush buffer and prepare new parquet file
-                    self._flush_metadata_buffer()
-                    chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, self.chunks_size)
-                    self._close_writer()
-
-            # Update the existing pandas dataframe with new row
-            episode_dict["meta/episodes/chunk_index"] = [chunk_idx]
-            episode_dict["meta/episodes/file_index"] = [file_idx]
-            episode_dict["dataset_from_index"] = [self.latest_episode["dataset_to_index"][0]]
-            episode_dict["dataset_to_index"] = [self.latest_episode["dataset_to_index"][0] + num_frames]
-
-        # Add to buffer
-        self.metadata_buffer.append(episode_dict)
-        self.latest_episode = episode_dict
-
-        if len(self.metadata_buffer) >= self.metadata_buffer_size:
-            self._flush_metadata_buffer()
-
-    def save_episode(
-        self,
-        episode_index: int,
-        episode_length: int,
-        episode_tasks: list[str],
-        episode_stats: dict[str, dict],
-        episode_metadata: dict,
-    ) -> None:
-        episode_dict = {
-            "episode_index": episode_index,
-            "tasks": episode_tasks,
-            "length": episode_length,
-        }
-        episode_dict.update(episode_metadata)
-        episode_dict.update(flatten_dict({"stats": episode_stats}))
-        self._save_episode_metadata(episode_dict)
-
-        # Update info
-        self.info["total_episodes"] += 1
-        self.info["total_frames"] += episode_length
-        self.info["total_tasks"] = len(self.tasks)
-        self.info["splits"] = {"train": f"0:{self.info['total_episodes']}"}
-
-        write_info(self.info, self.root)
-
-        self.stats = aggregate_stats([self.stats, episode_stats]) if self.stats is not None else episode_stats
-        write_stats(self.stats, self.root)
-
-    def update_video_info(self, video_key: str | None = None) -> None:
-        """
-        Warning: this function writes info from first episode videos, implicitly assuming that all videos have
-        been encoded the same way. Also, this means it assumes the first episode exists.
-        """
-        if video_key is not None and video_key not in self.video_keys:
-            raise ValueError(f"Video key {video_key} not found in dataset")
-
-        video_keys = [video_key] if video_key is not None else self.video_keys
-        for key in video_keys:
-            if not self.features[key].get("info", None):
-                video_path = self.root / self.video_path.format(video_key=key, chunk_index=0, file_index=0)
-                self.info["features"][key]["info"] = get_video_info(video_path)
-
-    def update_chunk_settings(
-        self,
-        chunks_size: int | None = None,
-        data_files_size_in_mb: int | None = None,
-        video_files_size_in_mb: int | None = None,
-    ) -> None:
-        """Update chunk and file size settings after dataset creation.
-
-        This allows users to customize storage organization without modifying the constructor.
-        These settings control how episodes are chunked and how large files can grow before
-        creating new ones.
-
-        Args:
-            chunks_size: Maximum number of files per chunk directory. If None, keeps current value.
-            data_files_size_in_mb: Maximum size for data parquet files in MB. If None, keeps current value.
-            video_files_size_in_mb: Maximum size for video files in MB. If None, keeps current value.
-        """
-        if chunks_size is not None:
-            if chunks_size <= 0:
-                raise ValueError(f"chunks_size must be positive, got {chunks_size}")
-            self.info["chunks_size"] = chunks_size
-
-        if data_files_size_in_mb is not None:
-            if data_files_size_in_mb <= 0:
-                raise ValueError(f"data_files_size_in_mb must be positive, got {data_files_size_in_mb}")
-            self.info["data_files_size_in_mb"] = data_files_size_in_mb
-
-        if video_files_size_in_mb is not None:
-            if video_files_size_in_mb <= 0:
-                raise ValueError(f"video_files_size_in_mb must be positive, got {video_files_size_in_mb}")
-            self.info["video_files_size_in_mb"] = video_files_size_in_mb
-
-        # Update the info file on disk
-        write_info(self.info, self.root)
-
-    def get_chunk_settings(self) -> dict[str, int]:
-        """Get current chunk and file size settings.
-
-        Returns:
-            Dict containing chunks_size, data_files_size_in_mb, and video_files_size_in_mb.
-        """
-        return {
-            "chunks_size": self.chunks_size,
-            "data_files_size_in_mb": self.data_files_size_in_mb,
-            "video_files_size_in_mb": self.video_files_size_in_mb,
-        }
-
-    def __repr__(self):
-        feature_keys = list(self.features)
-        return (
-            f"{self.__class__.__name__}({{\n"
-            f"    Repository ID: '{self.repo_id}',\n"
-            f"    Total episodes: '{self.total_episodes}',\n"
-            f"    Total frames: '{self.total_frames}',\n"
-            f"    Features: '{feature_keys}',\n"
-            "})',\n"
-        )
-
-    @classmethod
-    def create(
-        cls,
-        repo_id: str,
-        fps: int,
-        features: dict,
-        robot_type: str | None = None,
-        root: str | Path | None = None,
-        use_videos: bool = True,
-        metadata_buffer_size: int = 10,
-        chunks_size: int | None = None,
-        data_files_size_in_mb: int | None = None,
-        video_files_size_in_mb: int | None = None,
-    ) -> "LeRobotDatasetMetadata":
-        """Creates metadata for a LeRobotDataset."""
-        obj = cls.__new__(cls)
-        obj.repo_id = repo_id
-        obj.root = Path(root) if root is not None else HF_LEROBOT_HOME / repo_id
-
-        obj.root.mkdir(parents=True, exist_ok=False)
-
-        features = {**features, **DEFAULT_FEATURES}
-        _validate_feature_names(features)
-
-        obj.tasks = None
-        obj.subtasks = None
-        obj.episodes = None
-        obj.stats = None
-        obj.info = create_empty_dataset_info(
-            CODEBASE_VERSION,
-            fps,
-            features,
-            use_videos,
-            robot_type,
-            chunks_size,
-            data_files_size_in_mb,
-            video_files_size_in_mb,
-        )
-        if len(obj.video_keys) > 0 and not use_videos:
-            raise ValueError()
-        write_json(obj.info, obj.root / INFO_PATH)
-        obj.revision = None
-        obj.writer = None
-        obj.latest_episode = None
-        obj.metadata_buffer = []
-        obj.metadata_buffer_size = metadata_buffer_size
-        return obj
+logger = logging.getLogger(__name__)
 
 
 def _encode_video_worker(
@@ -596,7 +122,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
               the dataset from that address and load it, pending your dataset is compliant with
               codebase_version v3.0. If your dataset has been created before this new format, you will be
               prompted to convert it using our conversion script from v2.1 to v3.0, which you can find at
-              lerobot/datasets/v30/convert_dataset_v21_to_v30.py.
+              lerobot/scripts/convert_dataset_v21_to_v30.py.
 
 
         2. Your dataset doesn't already exists (either on local disk or on the Hub): you can create an empty
@@ -1326,7 +852,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
                             temp_path = future.result()
                             results[video_key] = temp_path
                         except Exception as exc:
-                            logging.error(f"Video encoding failed for {video_key}: {exc}")
+                            logger.error(f"Video encoding failed for {video_key}: {exc}")
                             raise exc
 
                 for video_key in self.meta.video_keys:
@@ -1365,7 +891,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         if end_episode is None:
             end_episode = self.num_episodes
 
-        logging.info(
+        logger.info(
             f"Batch encoding {self.batch_encoding_size} videos for episodes {start_episode} to {end_episode - 1}"
         )
 
@@ -1375,7 +901,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         episode_df = pd.read_parquet(episode_df_path)
 
         for ep_idx in range(start_episode, end_episode):
-            logging.info(f"Encoding videos for episode {ep_idx}")
+            logger.info(f"Encoding videos for episode {ep_idx}")
 
             if (
                 self.meta.episodes[ep_idx]["data/chunk_index"] != chunk_idx
@@ -1605,7 +1131,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
 
     def start_image_writer(self, num_processes: int = 0, num_threads: int = 4) -> None:
         if isinstance(self.image_writer, AsyncImageWriter):
-            logging.warning(
+            logger.warning(
                 "You are starting a new AsyncImageWriter that is replacing an already existing one in the dataset."
             )
 
@@ -1683,7 +1209,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         if image_writer_processes or image_writer_threads:
             obj.start_image_writer(image_writer_processes, image_writer_threads)
 
-        # TODO(aliberts, rcadene, alexander-soare): Merge this with OnlineBuffer/DataBuffer
         obj.episode_buffer = obj.create_episode_buffer()
 
         obj.episodes = None
@@ -1717,184 +1242,3 @@ class LeRobotDataset(torch.utils.data.Dataset):
             obj._streaming_encoder = None
 
         return obj
-
-
-class MultiLeRobotDataset(torch.utils.data.Dataset):
-    """A dataset consisting of multiple underlying `LeRobotDataset`s.
-
-    The underlying `LeRobotDataset`s are effectively concatenated, and this class adopts much of the API
-    structure of `LeRobotDataset`.
-    """
-
-    def __init__(
-        self,
-        repo_ids: list[str],
-        root: str | Path | None = None,
-        episodes: dict | None = None,
-        image_transforms: Callable | None = None,
-        delta_timestamps: dict[str, list[float]] | None = None,
-        tolerances_s: dict | None = None,
-        download_videos: bool = True,
-        video_backend: str | None = None,
-    ):
-        super().__init__()
-        self.repo_ids = repo_ids
-        self.root = Path(root) if root else HF_LEROBOT_HOME
-        self.tolerances_s = tolerances_s if tolerances_s else dict.fromkeys(repo_ids, 0.0001)
-        # Construct the underlying datasets passing everything but `transform` and `delta_timestamps` which
-        # are handled by this class.
-        self._datasets = [
-            LeRobotDataset(
-                repo_id,
-                root=self.root / repo_id,
-                episodes=episodes[repo_id] if episodes else None,
-                image_transforms=image_transforms,
-                delta_timestamps=delta_timestamps,
-                tolerance_s=self.tolerances_s[repo_id],
-                download_videos=download_videos,
-                video_backend=video_backend,
-            )
-            for repo_id in repo_ids
-        ]
-
-        # Disable any data keys that are not common across all of the datasets. Note: we may relax this
-        # restriction in future iterations of this class. For now, this is necessary at least for being able
-        # to use PyTorch's default DataLoader collate function.
-        self.disabled_features = set()
-        intersection_features = set(self._datasets[0].features)
-        for ds in self._datasets:
-            intersection_features.intersection_update(ds.features)
-        if len(intersection_features) == 0:
-            raise RuntimeError(
-                "Multiple datasets were provided but they had no keys common to all of them. "
-                "The multi-dataset functionality currently only keeps common keys."
-            )
-        for repo_id, ds in zip(self.repo_ids, self._datasets, strict=True):
-            extra_keys = set(ds.features).difference(intersection_features)
-            if extra_keys:
-                logging.warning(
-                    f"keys {extra_keys} of {repo_id} were disabled as they are not contained in all the "
-                    "other datasets."
-                )
-                self.disabled_features.update(extra_keys)
-
-        self.image_transforms = image_transforms
-        self.delta_timestamps = delta_timestamps
-        # TODO(rcadene, aliberts): We should not perform this aggregation for datasets
-        # with multiple robots of different ranges. Instead we should have one normalization
-        # per robot.
-        self.stats = aggregate_stats([dataset.meta.stats for dataset in self._datasets])
-
-    @property
-    def repo_id_to_index(self):
-        """Return a mapping from dataset repo_id to a dataset index automatically created by this class.
-
-        This index is incorporated as a data key in the dictionary returned by `__getitem__`.
-        """
-        return {repo_id: i for i, repo_id in enumerate(self.repo_ids)}
-
-    @property
-    def fps(self) -> int:
-        """Frames per second used during data collection.
-
-        NOTE: Fow now, this relies on a check in __init__ to make sure all sub-datasets have the same info.
-        """
-        return self._datasets[0].meta.info["fps"]
-
-    @property
-    def video(self) -> bool:
-        """Returns True if this dataset loads video frames from mp4 files.
-
-        Returns False if it only loads images from png files.
-
-        NOTE: Fow now, this relies on a check in __init__ to make sure all sub-datasets have the same info.
-        """
-        return self._datasets[0].meta.info.get("video", False)
-
-    @property
-    def features(self) -> datasets.Features:
-        features = {}
-        for dataset in self._datasets:
-            features.update({k: v for k, v in dataset.hf_features.items() if k not in self.disabled_features})
-        return features
-
-    @property
-    def camera_keys(self) -> list[str]:
-        """Keys to access image and video stream from cameras."""
-        keys = []
-        for key, feats in self.features.items():
-            if isinstance(feats, (datasets.Image | VideoFrame)):
-                keys.append(key)
-        return keys
-
-    @property
-    def video_frame_keys(self) -> list[str]:
-        """Keys to access video frames that requires to be decoded into images.
-
-        Note: It is empty if the dataset contains images only,
-        or equal to `self.cameras` if the dataset contains videos only,
-        or can even be a subset of `self.cameras` in a case of a mixed image/video dataset.
-        """
-        video_frame_keys = []
-        for key, feats in self.features.items():
-            if isinstance(feats, VideoFrame):
-                video_frame_keys.append(key)
-        return video_frame_keys
-
-    @property
-    def num_frames(self) -> int:
-        """Number of samples/frames."""
-        return sum(d.num_frames for d in self._datasets)
-
-    @property
-    def num_episodes(self) -> int:
-        """Number of episodes."""
-        return sum(d.num_episodes for d in self._datasets)
-
-    @property
-    def tolerance_s(self) -> float:
-        """Tolerance in seconds used to discard loaded frames when their timestamps
-        are not close enough from the requested frames. It is only used when `delta_timestamps`
-        is provided or when loading video frames from mp4 files.
-        """
-        # 1e-4 to account for possible numerical error
-        return 1 / self.fps - 1e-4
-
-    def __len__(self):
-        return self.num_frames
-
-    def __getitem__(self, idx: int) -> dict[str, torch.Tensor]:
-        if idx >= len(self):
-            raise IndexError(f"Index {idx} out of bounds.")
-        # Determine which dataset to get an item from based on the index.
-        start_idx = 0
-        dataset_idx = 0
-        for dataset in self._datasets:
-            if idx >= start_idx + dataset.num_frames:
-                start_idx += dataset.num_frames
-                dataset_idx += 1
-                continue
-            break
-        else:
-            raise AssertionError("We expect the loop to break out as long as the index is within bounds.")
-        item = self._datasets[dataset_idx][idx - start_idx]
-        item["dataset_index"] = torch.tensor(dataset_idx)
-        for data_key in self.disabled_features:
-            if data_key in item:
-                del item[data_key]
-
-        return item
-
-    def __repr__(self):
-        return (
-            f"{self.__class__.__name__}(\n"
-            f"  Repository IDs: '{self.repo_ids}',\n"
-            f"  Number of Samples: {self.num_frames},\n"
-            f"  Number of Episodes: {self.num_episodes},\n"
-            f"  Type: {'video (.mp4)' if self.video else 'image (.png)'},\n"
-            f"  Recorded Frames per Second: {self.fps},\n"
-            f"  Camera Keys: {self.camera_keys},\n"
-            f"  Video Frame Keys: {self.video_frame_keys if self.video else 'N/A'},\n"
-            f"  Transformations: {self.image_transforms},\n"
-            f")"
-        )

src/lerobot/datasets/multi_dataset.py

@@ -0,0 +1,210 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import logging
+from collections.abc import Callable
+from pathlib import Path
+
+import datasets
+import torch
+import torch.utils
+
+from lerobot.datasets.compute_stats import aggregate_stats
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.video_utils import VideoFrame
+from lerobot.utils.constants import HF_LEROBOT_HOME
+
+logger = logging.getLogger(__name__)
+
+
+class MultiLeRobotDataset(torch.utils.data.Dataset):
+    """A dataset consisting of multiple underlying `LeRobotDataset`s.
+
+    The underlying `LeRobotDataset`s are effectively concatenated, and this class adopts much of the API
+    structure of `LeRobotDataset`.
+    """
+
+    def __init__(
+        self,
+        repo_ids: list[str],
+        root: str | Path | None = None,
+        episodes: dict | None = None,
+        image_transforms: Callable | None = None,
+        delta_timestamps: dict[str, list[float]] | None = None,
+        tolerances_s: dict | None = None,
+        download_videos: bool = True,
+        video_backend: str | None = None,
+    ):
+        super().__init__()
+        self.repo_ids = repo_ids
+        self.root = Path(root) if root else HF_LEROBOT_HOME
+        self.tolerances_s = tolerances_s if tolerances_s else dict.fromkeys(repo_ids, 0.0001)
+        # Construct the underlying datasets passing everything but `transform` and `delta_timestamps` which
+        # are handled by this class.
+        self._datasets = [
+            LeRobotDataset(
+                repo_id,
+                root=self.root / repo_id,
+                episodes=episodes[repo_id] if episodes else None,
+                image_transforms=image_transforms,
+                delta_timestamps=delta_timestamps,
+                tolerance_s=self.tolerances_s[repo_id],
+                download_videos=download_videos,
+                video_backend=video_backend,
+            )
+            for repo_id in repo_ids
+        ]
+
+        # Disable any data keys that are not common across all of the datasets. Note: we may relax this
+        # restriction in future iterations of this class. For now, this is necessary at least for being able
+        # to use PyTorch's default DataLoader collate function.
+        self.disabled_features = set()
+        intersection_features = set(self._datasets[0].features)
+        for ds in self._datasets:
+            intersection_features.intersection_update(ds.features)
+        if len(intersection_features) == 0:
+            raise RuntimeError(
+                "Multiple datasets were provided but they had no keys common to all of them. "
+                "The multi-dataset functionality currently only keeps common keys."
+            )
+        for repo_id, ds in zip(self.repo_ids, self._datasets, strict=True):
+            extra_keys = set(ds.features).difference(intersection_features)
+            if extra_keys:
+                logger.warning(
+                    f"keys {extra_keys} of {repo_id} were disabled as they are not contained in all the "
+                    "other datasets."
+                )
+                self.disabled_features.update(extra_keys)
+
+        self.image_transforms = image_transforms
+        self.delta_timestamps = delta_timestamps
+        # TODO(rcadene, aliberts): We should not perform this aggregation for datasets
+        # with multiple robots of different ranges. Instead we should have one normalization
+        # per robot.
+        self.stats = aggregate_stats([dataset.meta.stats for dataset in self._datasets])
+
+    @property
+    def repo_id_to_index(self):
+        """Return a mapping from dataset repo_id to a dataset index automatically created by this class.
+
+        This index is incorporated as a data key in the dictionary returned by `__getitem__`.
+        """
+        return {repo_id: i for i, repo_id in enumerate(self.repo_ids)}
+
+    @property
+    def fps(self) -> int:
+        """Frames per second used during data collection.
+
+        NOTE: Fow now, this relies on a check in __init__ to make sure all sub-datasets have the same info.
+        """
+        return self._datasets[0].meta.info["fps"]
+
+    @property
+    def video(self) -> bool:
+        """Returns True if this dataset loads video frames from mp4 files.
+
+        Returns False if it only loads images from png files.
+
+        NOTE: Fow now, this relies on a check in __init__ to make sure all sub-datasets have the same info.
+        """
+        return self._datasets[0].meta.info.get("video", False)
+
+    @property
+    def features(self) -> datasets.Features:
+        features = {}
+        for dataset in self._datasets:
+            features.update({k: v for k, v in dataset.hf_features.items() if k not in self.disabled_features})
+        return features
+
+    @property
+    def camera_keys(self) -> list[str]:
+        """Keys to access image and video stream from cameras."""
+        keys = []
+        for key, feats in self.features.items():
+            if isinstance(feats, (datasets.Image | VideoFrame)):
+                keys.append(key)
+        return keys
+
+    @property
+    def video_frame_keys(self) -> list[str]:
+        """Keys to access video frames that requires to be decoded into images.
+
+        Note: It is empty if the dataset contains images only,
+        or equal to `self.cameras` if the dataset contains videos only,
+        or can even be a subset of `self.cameras` in a case of a mixed image/video dataset.
+        """
+        video_frame_keys = []
+        for key, feats in self.features.items():
+            if isinstance(feats, VideoFrame):
+                video_frame_keys.append(key)
+        return video_frame_keys
+
+    @property
+    def num_frames(self) -> int:
+        """Number of samples/frames."""
+        return sum(d.num_frames for d in self._datasets)
+
+    @property
+    def num_episodes(self) -> int:
+        """Number of episodes."""
+        return sum(d.num_episodes for d in self._datasets)
+
+    @property
+    def tolerance_s(self) -> float:
+        """Tolerance in seconds used to discard loaded frames when their timestamps
+        are not close enough from the requested frames. It is only used when `delta_timestamps`
+        is provided or when loading video frames from mp4 files.
+        """
+        # 1e-4 to account for possible numerical error
+        return 1 / self.fps - 1e-4
+
+    def __len__(self):
+        return self.num_frames
+
+    def __getitem__(self, idx: int) -> dict[str, torch.Tensor]:
+        if idx >= len(self):
+            raise IndexError(f"Index {idx} out of bounds.")
+        # Determine which dataset to get an item from based on the index.
+        start_idx = 0
+        dataset_idx = 0
+        for dataset in self._datasets:
+            if idx >= start_idx + dataset.num_frames:
+                start_idx += dataset.num_frames
+                dataset_idx += 1
+                continue
+            break
+        else:
+            raise AssertionError("We expect the loop to break out as long as the index is within bounds.")
+        item = self._datasets[dataset_idx][idx - start_idx]
+        item["dataset_index"] = torch.tensor(dataset_idx)
+        for data_key in self.disabled_features:
+            if data_key in item:
+                del item[data_key]
+
+        return item
+
+    def __repr__(self):
+        return (
+            f"{self.__class__.__name__}(\n"
+            f"  Repository IDs: '{self.repo_ids}',\n"
+            f"  Number of Samples: {self.num_frames},\n"
+            f"  Number of Episodes: {self.num_episodes},\n"
+            f"  Type: {'video (.mp4)' if self.video else 'image (.png)'},\n"
+            f"  Recorded Frames per Second: {self.fps},\n"
+            f"  Camera Keys: {self.camera_keys},\n"
+            f"  Video Frame Keys: {self.video_frame_keys if self.video else 'N/A'},\n"
+            f"  Transformations: {self.image_transforms},\n"
+            f")"
+        )

src/lerobot/datasets/online_buffer.py

@@ -1,382 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""An online buffer for the online training loop in train.py
-
-Note to maintainers: This duplicates some logic from LeRobotDataset and EpisodeAwareSampler. We should
-consider converging to one approach. Here we have opted to use numpy.memmap to back the data buffer. It's much
-faster than using HuggingFace Datasets as there's no conversion to an intermediate non-python object. Also it
-supports in-place slicing and mutation which is very handy for a dynamic buffer.
-"""
-
-import os
-from pathlib import Path
-from typing import Any
-
-import numpy as np
-import torch
-
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-
-
-def _make_memmap_safe(**kwargs) -> np.memmap:
-    """Make a numpy memmap with checks on available disk space first.
-
-    Expected kwargs are: "filename", "dtype" (must by np.dtype), "mode" and "shape"
-
-    For information on dtypes:
-    https://numpy.org/doc/stable/reference/arrays.dtypes.html#arrays-dtypes-constructing
-    """
-    if kwargs["mode"].startswith("w"):
-        required_space = kwargs["dtype"].itemsize * np.prod(kwargs["shape"])  # bytes
-        stats = os.statvfs(Path(kwargs["filename"]).parent)
-        available_space = stats.f_bavail * stats.f_frsize  # bytes
-        if required_space >= available_space * 0.8:
-            raise RuntimeError(
-                f"You're about to take up {required_space} of {available_space} bytes available."
-            )
-    return np.memmap(**kwargs)
-
-
-class OnlineBuffer(torch.utils.data.Dataset):
-    """FIFO data buffer for the online training loop in train.py.
-
-    Follows the protocol of LeRobotDataset as much as is required to have it be used by the online training
-    loop in the same way that a LeRobotDataset would be used.
-
-    The underlying data structure will have data inserted in a circular fashion. Always insert after the
-    last index, and when you reach the end, wrap around to the start.
-
-    The data is stored in a numpy memmap.
-    """
-
-    NEXT_INDEX_KEY = "_next_index"
-    OCCUPANCY_MASK_KEY = "_occupancy_mask"
-    INDEX_KEY = "index"
-    FRAME_INDEX_KEY = "frame_index"
-    EPISODE_INDEX_KEY = "episode_index"
-    TIMESTAMP_KEY = "timestamp"
-    IS_PAD_POSTFIX = "_is_pad"
-
-    def __init__(
-        self,
-        write_dir: str | Path,
-        data_spec: dict[str, Any] | None,
-        buffer_capacity: int | None,
-        fps: float | None = None,
-        delta_timestamps: dict[str, list[float]] | dict[str, np.ndarray] | None = None,
-    ):
-        """
-        The online buffer can be provided from scratch or you can load an existing online buffer by passing
-        a `write_dir` associated with an existing buffer.
-
-        Args:
-            write_dir: Where to keep the numpy memmap files. One memmap file will be stored for each data key.
-                Note that if the files already exist, they are opened in read-write mode (used for training
-                resumption.)
-            data_spec: A mapping from data key to data specification, like {data_key: {"shape": tuple[int],
-                "dtype": np.dtype}}. This should include all the data that you wish to record into the buffer,
-                but note that "index", "frame_index" and "episode_index" are already accounted for by this
-                class, so you don't need to include them.
-            buffer_capacity: How many frames should be stored in the buffer as a maximum. Be aware of your
-                system's available disk space when choosing this.
-            fps: Same as the fps concept in LeRobot dataset. Here it needs to be provided for the
-                 delta_timestamps logic. You can pass None if you are not using delta_timestamps.
-            delta_timestamps: Same as the delta_timestamps concept in LeRobotDataset. This is internally
-                converted to dict[str, np.ndarray] for optimization purposes.
-
-        """
-        self.set_delta_timestamps(delta_timestamps)
-        self._fps = fps
-        # Tolerance in seconds used to discard loaded frames when their timestamps are not close enough from
-        # the requested frames. It is only used when `delta_timestamps` is provided.
-        # minus 1e-4 to account for possible numerical error
-        self.tolerance_s = 1 / self.fps - 1e-4 if fps is not None else None
-        self._buffer_capacity = buffer_capacity
-        data_spec = self._make_data_spec(data_spec, buffer_capacity)
-        Path(write_dir).mkdir(parents=True, exist_ok=True)
-        self._data = {}
-        for k, v in data_spec.items():
-            self._data[k] = _make_memmap_safe(
-                filename=Path(write_dir) / k,
-                dtype=v["dtype"] if v is not None else None,
-                mode="r+" if (Path(write_dir) / k).exists() else "w+",
-                shape=tuple(v["shape"]) if v is not None else None,
-            )
-
-    @property
-    def delta_timestamps(self) -> dict[str, np.ndarray] | None:
-        return self._delta_timestamps
-
-    def set_delta_timestamps(self, value: dict[str, list[float]] | None):
-        """Set delta_timestamps converting the values to numpy arrays.
-
-        The conversion is for an optimization in the __getitem__. The loop is much slower if the arrays
-        need to be converted into numpy arrays.
-        """
-        if value is not None:
-            self._delta_timestamps = {k: np.array(v) for k, v in value.items()}
-        else:
-            self._delta_timestamps = None
-
-    def _make_data_spec(self, data_spec: dict[str, Any], buffer_capacity: int) -> dict[str, dict[str, Any]]:
-        """Makes the data spec for np.memmap."""
-        if any(k.startswith("_") for k in data_spec):
-            raise ValueError(
-                "data_spec keys should not start with '_'. This prefix is reserved for internal logic."
-            )
-        preset_keys = {
-            OnlineBuffer.INDEX_KEY,
-            OnlineBuffer.FRAME_INDEX_KEY,
-            OnlineBuffer.EPISODE_INDEX_KEY,
-            OnlineBuffer.TIMESTAMP_KEY,
-        }
-        if len(intersection := set(data_spec).intersection(preset_keys)) > 0:
-            raise ValueError(
-                f"data_spec should not contain any of {preset_keys} as these are handled internally. "
-                f"The provided data_spec has {intersection}."
-            )
-        complete_data_spec = {
-            # _next_index will be a pointer to the next index that we should start filling from when we add
-            # more data.
-            OnlineBuffer.NEXT_INDEX_KEY: {"dtype": np.dtype("int64"), "shape": ()},
-            # Since the memmap is initialized with all-zeros, this keeps track of which indices are occupied
-            # with real data rather than the dummy initialization.
-            OnlineBuffer.OCCUPANCY_MASK_KEY: {"dtype": np.dtype("?"), "shape": (buffer_capacity,)},
-            OnlineBuffer.INDEX_KEY: {"dtype": np.dtype("int64"), "shape": (buffer_capacity,)},
-            OnlineBuffer.FRAME_INDEX_KEY: {"dtype": np.dtype("int64"), "shape": (buffer_capacity,)},
-            OnlineBuffer.EPISODE_INDEX_KEY: {"dtype": np.dtype("int64"), "shape": (buffer_capacity,)},
-            OnlineBuffer.TIMESTAMP_KEY: {"dtype": np.dtype("float64"), "shape": (buffer_capacity,)},
-        }
-        for k, v in data_spec.items():
-            complete_data_spec[k] = {"dtype": v["dtype"], "shape": (buffer_capacity, *v["shape"])}
-        return complete_data_spec
-
-    def add_data(self, data: dict[str, np.ndarray]):
-        """Add new data to the buffer, which could potentially mean shifting old data out.
-
-        The new data should contain all the frames (in order) of any number of episodes. The indices should
-        start from 0 (note to the developer: this can easily be generalized). See the `rollout` and
-        `eval_policy` functions in `eval.py` for more information on how the data is constructed.
-
-        Shift the incoming data index and episode_index to continue on from the last frame. Note that this
-        will be done in place!
-        """
-        if len(missing_keys := (set(self.data_keys).difference(set(data)))) > 0:
-            raise ValueError(f"Missing data keys: {missing_keys}")
-        new_data_length = len(data[self.data_keys[0]])
-        if not all(len(data[k]) == new_data_length for k in self.data_keys):
-            raise ValueError("All data items should have the same length")
-
-        next_index = self._data[OnlineBuffer.NEXT_INDEX_KEY]
-
-        # Sanity check to make sure that the new data indices start from 0.
-        assert data[OnlineBuffer.EPISODE_INDEX_KEY][0].item() == 0
-        assert data[OnlineBuffer.INDEX_KEY][0].item() == 0
-
-        # Shift the incoming indices if necessary.
-        if self.num_frames > 0:
-            last_episode_index = self._data[OnlineBuffer.EPISODE_INDEX_KEY][next_index - 1]
-            last_data_index = self._data[OnlineBuffer.INDEX_KEY][next_index - 1]
-            data[OnlineBuffer.EPISODE_INDEX_KEY] += last_episode_index + 1
-            data[OnlineBuffer.INDEX_KEY] += last_data_index + 1
-
-        # Insert the new data starting from next_index. It may be necessary to wrap around to the start.
-        n_surplus = max(0, new_data_length - (self._buffer_capacity - next_index))
-        for k in self.data_keys:
-            if n_surplus == 0:
-                slc = slice(next_index, next_index + new_data_length)
-                self._data[k][slc] = data[k]
-                self._data[OnlineBuffer.OCCUPANCY_MASK_KEY][slc] = True
-            else:
-                self._data[k][next_index:] = data[k][:-n_surplus]
-                self._data[OnlineBuffer.OCCUPANCY_MASK_KEY][next_index:] = True
-                self._data[k][:n_surplus] = data[k][-n_surplus:]
-        if n_surplus == 0:
-            self._data[OnlineBuffer.NEXT_INDEX_KEY] = next_index + new_data_length
-        else:
-            self._data[OnlineBuffer.NEXT_INDEX_KEY] = n_surplus
-
-    @property
-    def data_keys(self) -> list[str]:
-        keys = set(self._data)
-        keys.remove(OnlineBuffer.OCCUPANCY_MASK_KEY)
-        keys.remove(OnlineBuffer.NEXT_INDEX_KEY)
-        return sorted(keys)
-
-    @property
-    def fps(self) -> float | None:
-        return self._fps
-
-    @property
-    def num_episodes(self) -> int:
-        return len(
-            np.unique(self._data[OnlineBuffer.EPISODE_INDEX_KEY][self._data[OnlineBuffer.OCCUPANCY_MASK_KEY]])
-        )
-
-    @property
-    def num_frames(self) -> int:
-        return np.count_nonzero(self._data[OnlineBuffer.OCCUPANCY_MASK_KEY])
-
-    def __len__(self):
-        return self.num_frames
-
-    def _item_to_tensors(self, item: dict) -> dict:
-        item_ = {}
-        for k, v in item.items():
-            if isinstance(v, torch.Tensor):
-                item_[k] = v
-            elif isinstance(v, np.ndarray):
-                item_[k] = torch.from_numpy(v)
-            else:
-                item_[k] = torch.tensor(v)
-        return item_
-
-    def __getitem__(self, idx: int) -> dict[str, torch.Tensor]:
-        if idx >= len(self) or idx < -len(self):
-            raise IndexError
-
-        item = {k: v[idx] for k, v in self._data.items() if not k.startswith("_")}
-
-        if self.delta_timestamps is None:
-            return self._item_to_tensors(item)
-
-        episode_index = item[OnlineBuffer.EPISODE_INDEX_KEY]
-        current_ts = item[OnlineBuffer.TIMESTAMP_KEY]
-        episode_data_indices = np.where(
-            np.bitwise_and(
-                self._data[OnlineBuffer.EPISODE_INDEX_KEY] == episode_index,
-                self._data[OnlineBuffer.OCCUPANCY_MASK_KEY],
-            )
-        )[0]
-        episode_timestamps = self._data[OnlineBuffer.TIMESTAMP_KEY][episode_data_indices]
-
-        for data_key in self.delta_timestamps:
-            # Note: The logic in this loop is copied from `load_previous_and_future_frames`.
-            # Get timestamps used as query to retrieve data of previous/future frames.
-            query_ts = current_ts + self.delta_timestamps[data_key]
-
-            # Compute distances between each query timestamp and all timestamps of all the frames belonging to
-            # the episode.
-            dist = np.abs(query_ts[:, None] - episode_timestamps[None, :])
-            argmin_ = np.argmin(dist, axis=1)
-            min_ = dist[np.arange(dist.shape[0]), argmin_]
-
-            is_pad = min_ > self.tolerance_s
-
-            # Check violated query timestamps are all outside the episode range.
-            assert (
-                (query_ts[is_pad] < episode_timestamps[0]) | (episode_timestamps[-1] < query_ts[is_pad])
-            ).all(), (
-                f"One or several timestamps unexpectedly violate the tolerance ({min_} > {self.tolerance_s=}"
-                ") inside the episode range."
-            )
-
-            # Load frames for this data key.
-            item[data_key] = self._data[data_key][episode_data_indices[argmin_]]
-
-            item[f"{data_key}{OnlineBuffer.IS_PAD_POSTFIX}"] = is_pad
-
-        return self._item_to_tensors(item)
-
-    def get_data_by_key(self, key: str) -> torch.Tensor:
-        """Returns all data for a given data key as a Tensor."""
-        return torch.from_numpy(self._data[key][self._data[OnlineBuffer.OCCUPANCY_MASK_KEY]])
-
-
-def compute_sampler_weights(
-    offline_dataset: LeRobotDataset,
-    offline_drop_n_last_frames: int = 0,
-    online_dataset: OnlineBuffer | None = None,
-    online_sampling_ratio: float | None = None,
-    online_drop_n_last_frames: int = 0,
-) -> torch.Tensor:
-    """Compute the sampling weights for the online training dataloader in train.py.
-
-    Args:
-        offline_dataset: The LeRobotDataset used for offline pre-training.
-        online_drop_n_last_frames: Number of frames to drop from the end of each offline dataset episode.
-        online_dataset: The OnlineBuffer used in online training.
-        online_sampling_ratio: The proportion of data that should be sampled from the online dataset. If an
-            online dataset is provided, this value must also be provided.
-        online_drop_n_first_frames: See `offline_drop_n_last_frames`. This is the same, but for the online
-            dataset.
-    Returns:
-        Tensor of weights for [offline_dataset; online_dataset], normalized to 1.
-
-    Notes to maintainers:
-        - This duplicates some logic from EpisodeAwareSampler. We should consider converging to one approach.
-        - When used with `torch.utils.data.WeightedRandomSampler`, it could completely replace
-          `EpisodeAwareSampler` as the online dataset related arguments are optional. The only missing feature
-          is the ability to turn shuffling off.
-        - Options `drop_first_n_frames` and `episode_indices_to_use` can be added easily. They were not
-          included here to avoid adding complexity.
-    """
-    if len(offline_dataset) == 0 and (online_dataset is None or len(online_dataset) == 0):
-        raise ValueError("At least one of `offline_dataset` or `online_dataset` should be contain data.")
-    if (online_dataset is None) ^ (online_sampling_ratio is None):
-        raise ValueError(
-            "`online_dataset` and `online_sampling_ratio` must be provided together or not at all."
-        )
-    offline_sampling_ratio = 0 if online_sampling_ratio is None else 1 - online_sampling_ratio
-
-    weights = []
-
-    if len(offline_dataset) > 0:
-        offline_data_mask_indices = []
-        for start_index, end_index in zip(
-            offline_dataset.meta.episodes["dataset_from_index"],
-            offline_dataset.meta.episodes["dataset_to_index"],
-            strict=True,
-        ):
-            offline_data_mask_indices.extend(range(start_index, end_index - offline_drop_n_last_frames))
-        offline_data_mask = torch.zeros(len(offline_dataset), dtype=torch.bool)
-        offline_data_mask[torch.tensor(offline_data_mask_indices)] = True
-        weights.append(
-            torch.full(
-                size=(len(offline_dataset),),
-                fill_value=offline_sampling_ratio / offline_data_mask.sum(),
-            )
-            * offline_data_mask
-        )
-
-    if online_dataset is not None and len(online_dataset) > 0:
-        online_data_mask_indices = []
-        episode_indices = online_dataset.get_data_by_key("episode_index")
-        for episode_idx in torch.unique(episode_indices):
-            where_episode = torch.where(episode_indices == episode_idx)
-            start_index = where_episode[0][0]
-            end_index = where_episode[0][-1] + 1
-            online_data_mask_indices.extend(
-                range(start_index.item(), end_index.item() - online_drop_n_last_frames)
-            )
-        online_data_mask = torch.zeros(len(online_dataset), dtype=torch.bool)
-        online_data_mask[torch.tensor(online_data_mask_indices)] = True
-        weights.append(
-            torch.full(
-                size=(len(online_dataset),),
-                fill_value=online_sampling_ratio / online_data_mask.sum(),
-            )
-            * online_data_mask
-        )
-
-    weights = torch.cat(weights)
-
-    if weights.sum() == 0:
-        weights += 1 / len(weights)
-    else:
-        weights /= weights.sum()
-
-    return weights

src/lerobot/datasets/pipeline_features.py

@@ -17,8 +17,9 @@ from collections.abc import Sequence
 from typing import Any
 
 from lerobot.configs.types import PipelineFeatureType
-from lerobot.datasets.utils import hw_to_dataset_features
-from lerobot.processor import DataProcessorPipeline, RobotAction, RobotObservation
+from lerobot.datasets.feature_utils import hw_to_dataset_features
+from lerobot.processor import DataProcessorPipeline
+from lerobot.types import RobotAction, RobotObservation
 from lerobot.utils.constants import ACTION, OBS_IMAGES, OBS_STATE, OBS_STR
 
 
@@ -43,11 +44,11 @@ def create_initial_features(
     return features
 
 
-# Helper to filter state/action keys based on regex patterns.
-def should_keep(key: str, patterns: tuple[str]) -> bool:
+# Helper to filter state/action keys based on compiled regex patterns.
+def should_keep(key: str, patterns: tuple[re.Pattern] | None) -> bool:
     if patterns is None:
         return True
-    return any(re.search(pat, key) for pat in patterns)
+    return any(pat.search(key) for pat in patterns)
 
 
 def strip_prefix(key: str, prefixes_to_strip: tuple[str]) -> str:
@@ -88,6 +89,8 @@ def aggregate_pipeline_dataset_features(
     Returns:
         A dictionary of features formatted for a Hugging Face LeRobot Dataset.
     """
+    compiled_patterns = tuple(re.compile(p) for p in patterns) if patterns is not None else None
+
     all_features = pipeline.transform_features(initial_features)
 
     # Intermediate storage for categorized and filtered features.
@@ -119,7 +122,7 @@ def aggregate_pipeline_dataset_features(
             # 2. Apply filtering rules.
             if is_image and not use_videos:
                 continue
-            if not is_image and not should_keep(key, patterns):
+            if not is_image and not should_keep(key, compiled_patterns):
                 continue
 
             # 3. Add the feature to the appropriate group with a clean name.

src/lerobot/datasets/push_dataset_to_hub/utils.py

@@ -1,73 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-import datasets
-import torch
-
-
-# TODO(aliberts): remove
-def calculate_episode_data_index(hf_dataset: datasets.Dataset) -> dict[str, torch.Tensor]:
-    """
-    Calculate episode data index for the provided HuggingFace Dataset. Relies on episode_index column of hf_dataset.
-
-    Parameters:
-    - hf_dataset (datasets.Dataset): A HuggingFace dataset containing the episode index.
-
-    Returns:
-    - episode_data_index: A dictionary containing the data index for each episode. The dictionary has two keys:
-        - "from": A tensor containing the starting index of each episode.
-        - "to": A tensor containing the ending index of each episode.
-    """
-    episode_data_index = {"from": [], "to": []}
-
-    current_episode = None
-    """
-    The episode_index is a list of integers, each representing the episode index of the corresponding example.
-    For instance, the following is a valid episode_index:
-      [0, 0, 0, 1, 1, 1, 1, 2, 2, 2, 2, 2]
-
-    Below, we iterate through the episode_index and populate the episode_data_index dictionary with the starting and
-    ending index of each episode. For the episode_index above, the episode_data_index dictionary will look like this:
-        {
-            "from": [0, 3, 7],
-            "to": [3, 7, 12]
-        }
-    """
-    if len(hf_dataset) == 0:
-        episode_data_index = {
-            "from": torch.tensor([]),
-            "to": torch.tensor([]),
-        }
-        return episode_data_index
-    for idx, episode_idx in enumerate(hf_dataset["episode_index"]):
-        if episode_idx != current_episode:
-            # We encountered a new episode, so we append its starting location to the "from" list
-            episode_data_index["from"].append(idx)
-            # If this is not the first episode, we append the ending location of the previous episode to the "to" list
-            if current_episode is not None:
-                episode_data_index["to"].append(idx)
-            # Let's keep track of the current episode index
-            current_episode = episode_idx
-        else:
-            # We are still in the same episode, so there is nothing for us to do here
-            pass
-    # We have reached the end of the dataset, so we append the ending location of the last episode to the "to" list
-    episode_data_index["to"].append(idx + 1)
-
-    for k in ["from", "to"]:
-        episode_data_index[k] = torch.tensor(episode_data_index[k])
-
-    return episode_data_index

src/lerobot/datasets/sampler.py

@@ -13,10 +13,13 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+import logging
 from collections.abc import Iterator
 
 import torch
 
+logger = logging.getLogger(__name__)
+
 
 class EpisodeAwareSampler:
     def __init__(
@@ -39,13 +42,35 @@ class EpisodeAwareSampler:
             drop_n_last_frames: Number of frames to drop from the end of each episode.
             shuffle: Whether to shuffle the indices.
         """
+        if drop_n_first_frames < 0:
+            raise ValueError(f"drop_n_first_frames must be >= 0, got {drop_n_first_frames}")
+        if drop_n_last_frames < 0:
+            raise ValueError(f"drop_n_last_frames must be >= 0, got {drop_n_last_frames}")
+
         indices = []
         for episode_idx, (start_index, end_index) in enumerate(
             zip(dataset_from_indices, dataset_to_indices, strict=True)
         ):
             if episode_indices_to_use is None or episode_idx in episode_indices_to_use:
+                ep_length = end_index - start_index
+                if drop_n_first_frames + drop_n_last_frames >= ep_length:
+                    logger.warning(
+                        "Episode %d has %d frames but drop_n_first_frames=%d and "
+                        "drop_n_last_frames=%d removes all frames. Skipping.",
+                        episode_idx,
+                        ep_length,
+                        drop_n_first_frames,
+                        drop_n_last_frames,
+                    )
+                    continue
                 indices.extend(range(start_index + drop_n_first_frames, end_index - drop_n_last_frames))
 
+        if not indices:
+            raise ValueError(
+                "No valid frames remain after applying drop_n_first_frames and drop_n_last_frames. "
+                "All episodes were either filtered out or had too few frames."
+            )
+
         self.indices = indices
         self.shuffle = shuffle
 

src/lerobot/datasets/streaming_dataset.py

@@ -13,7 +13,8 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-from collections.abc import Callable, Generator, Iterator
+from collections import deque
+from collections.abc import Callable, Generator, Iterable, Iterator
 from pathlib import Path
 
 import datasets
@@ -21,16 +22,13 @@ import numpy as np
 import torch
 from datasets import load_dataset
 
-from lerobot.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDatasetMetadata
+from lerobot.datasets.dataset_metadata import CODEBASE_VERSION, LeRobotDatasetMetadata
+from lerobot.datasets.feature_utils import get_delta_indices
+from lerobot.datasets.io_utils import item_to_torch
 from lerobot.datasets.utils import (
-    Backtrackable,
-    LookAheadError,
-    LookBackError,
     check_version_compatibility,
     find_float_index,
-    get_delta_indices,
     is_float_in_list,
-    item_to_torch,
     safe_shard,
 )
 from lerobot.datasets.video_utils import (
@@ -40,6 +38,164 @@ from lerobot.datasets.video_utils import (
 from lerobot.utils.constants import HF_LEROBOT_HOME, LOOKAHEAD_BACKTRACKTABLE, LOOKBACK_BACKTRACKTABLE
 
 
+class LookBackError(Exception):
+    """
+    Exception raised when trying to look back in the history of a Backtrackable object.
+    """
+
+    pass
+
+
+class LookAheadError(Exception):
+    """
+    Exception raised when trying to look ahead in the future of a Backtrackable object.
+    """
+
+    pass
+
+
+class Backtrackable[T]:
+    """
+    Wrap any iterator/iterable so you can step back up to `history` items
+    and look ahead up to `lookahead` items.
+
+    This is useful for streaming datasets where you need to access previous and future items
+    but can't load the entire dataset into memory.
+
+    Example:
+    -------
+    ```python
+    ds = load_dataset("c4", "en", streaming=True, split="train")
+    rev = Backtrackable(ds, history=3, lookahead=2)
+
+    x0 = next(rev)  # forward
+    x1 = next(rev)
+    x2 = next(rev)
+
+    # Look ahead
+    x3_peek = rev.peek_ahead(1)  # next item without moving cursor
+    x4_peek = rev.peek_ahead(2)  # two items ahead
+
+    # Look back
+    x1_again = rev.peek_back(1)  # previous item without moving cursor
+    x0_again = rev.peek_back(2)  # two items back
+
+    # Move backward
+    x1_back = rev.prev()  # back one step
+    next(rev)  # returns x2, continues forward from where we were
+    ```
+    """
+
+    __slots__ = ("_source", "_back_buf", "_ahead_buf", "_cursor", "_history", "_lookahead")
+
+    def __init__(self, iterable: Iterable[T], *, history: int = 1, lookahead: int = 0):
+        if history < 1:
+            raise ValueError("history must be >= 1")
+        if lookahead <= 0:
+            raise ValueError("lookahead must be > 0")
+
+        self._source: Iterator[T] = iter(iterable)
+        self._back_buf: deque[T] = deque(maxlen=history)
+        self._ahead_buf: deque[T] = deque(maxlen=lookahead) if lookahead > 0 else deque()
+        self._cursor: int = 0
+        self._history = history
+        self._lookahead = lookahead
+
+    def __iter__(self) -> "Backtrackable[T]":
+        return self
+
+    def __next__(self) -> T:
+        # If we've stepped back, consume from back buffer first
+        if self._cursor < 0:  # -1 means "last item", etc.
+            self._cursor += 1
+            return self._back_buf[self._cursor]
+
+        # If we have items in the ahead buffer, use them first
+        item = self._ahead_buf.popleft() if self._ahead_buf else next(self._source)
+
+        # Add current item to back buffer and reset cursor
+        self._back_buf.append(item)
+        self._cursor = 0
+        return item
+
+    def prev(self) -> T:
+        """
+        Step one item back in history and return it.
+        Raises IndexError if already at the oldest buffered item.
+        """
+        if len(self._back_buf) + self._cursor <= 1:
+            raise LookBackError("At start of history")
+
+        self._cursor -= 1
+        return self._back_buf[self._cursor]
+
+    def peek_back(self, n: int = 1) -> T:
+        """
+        Look `n` items back (n=1 == previous item) without moving the cursor.
+        """
+        if n < 0 or n + 1 > len(self._back_buf) + self._cursor:
+            raise LookBackError("peek_back distance out of range")
+
+        return self._back_buf[self._cursor - (n + 1)]
+
+    def peek_ahead(self, n: int = 1) -> T:
+        """
+        Look `n` items ahead (n=1 == next item) without moving the cursor.
+        Fills the ahead buffer if necessary.
+        """
+        if n < 1:
+            raise LookAheadError("peek_ahead distance must be 1 or more")
+        elif n > self._lookahead:
+            raise LookAheadError("peek_ahead distance exceeds lookahead limit")
+
+        # Fill ahead buffer if we don't have enough items
+        while len(self._ahead_buf) < n:
+            try:
+                item = next(self._source)
+                self._ahead_buf.append(item)
+
+            except StopIteration as err:
+                raise LookAheadError("peek_ahead: not enough items in source") from err
+
+        return self._ahead_buf[n - 1]
+
+    def history(self) -> list[T]:
+        """
+        Return a copy of the buffered history (most recent last).
+        The list length ≤ `history` argument passed at construction.
+        """
+        if self._cursor == 0:
+            return list(self._back_buf)
+
+        # When cursor<0, slice so the order remains chronological
+        return list(self._back_buf)[: self._cursor or None]
+
+    def can_peek_back(self, steps: int = 1) -> bool:
+        """
+        Check if we can go back `steps` items without raising an IndexError.
+        """
+        return steps <= len(self._back_buf) + self._cursor
+
+    def can_peek_ahead(self, steps: int = 1) -> bool:
+        """
+        Check if we can peek ahead `steps` items.
+        This may involve trying to fill the ahead buffer.
+        """
+        if self._lookahead > 0 and steps > self._lookahead:
+            return False
+
+        # Try to fill ahead buffer to check if we can peek that far
+        try:
+            while len(self._ahead_buf) < steps:
+                if self._lookahead > 0 and len(self._ahead_buf) >= self._lookahead:
+                    return False
+                item = next(self._source)
+                self._ahead_buf.append(item)
+            return True
+        except StopIteration:
+            return False
+
+
 class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
     """LeRobotDataset with streaming capabilities.
 

src/lerobot/datasets/video_utils.py

@@ -37,6 +37,8 @@ import torchvision
 from datasets.features.features import register_feature
 from PIL import Image
 
+logger = logging.getLogger(__name__)
+
 # List of hardware encoders to probe for auto-selection. Availability depends on the platform and FFmpeg build.
 # Determines the order of preference for auto-selection when vcodec="auto" is used.
 HW_ENCODERS = [
@@ -94,7 +96,7 @@ def detect_available_hw_encoders() -> list[str]:
             av.codec.Codec(codec_name, "w")
             available.append(codec_name)
         except Exception:  # nosec B110
-            pass  # nosec B110
+            logger.debug("HW encoder '%s' not available", codec_name)  # nosec B110
     return available
 
 
@@ -103,14 +105,14 @@ def resolve_vcodec(vcodec: str) -> str:
     if vcodec not in VALID_VIDEO_CODECS:
         raise ValueError(f"Invalid vcodec '{vcodec}'. Must be one of: {sorted(VALID_VIDEO_CODECS)}")
     if vcodec != "auto":
-        logging.info(f"Using video codec: {vcodec}")
+        logger.info(f"Using video codec: {vcodec}")
         return vcodec
     available = detect_available_hw_encoders()
     for encoder in HW_ENCODERS:
         if encoder in available:
-            logging.info(f"Auto-selected video codec: {encoder}")
+            logger.info(f"Auto-selected video codec: {encoder}")
             return encoder
-    logging.info("No hardware encoder available, falling back to software encoder 'libsvtav1'")
+    logger.info("No hardware encoder available, falling back to software encoder 'libsvtav1'")
     return "libsvtav1"
 
 
@@ -118,7 +120,7 @@ def get_safe_default_codec():
     if importlib.util.find_spec("torchcodec"):
         return "torchcodec"
     else:
-        logging.warning(
+        logger.warning(
             "'torchcodec' is not available in your platform, falling back to 'pyav' as a default decoder"
         )
         return "pyav"
@@ -208,7 +210,7 @@ def decode_video_frames_torchvision(
     for frame in reader:
         current_ts = frame["pts"]
         if log_loaded_timestamps:
-            logging.info(f"frame loaded at timestamp={current_ts:.4f}")
+            logger.info(f"frame loaded at timestamp={current_ts:.4f}")
         loaded_frames.append(frame["data"])
         loaded_ts.append(current_ts)
         if current_ts >= last_ts:
@@ -244,7 +246,7 @@ def decode_video_frames_torchvision(
     closest_ts = loaded_ts[argmin_]
 
     if log_loaded_timestamps:
-        logging.info(f"{closest_ts=}")
+        logger.info(f"{closest_ts=}")
 
     # convert to the pytorch format which is float32 in [0,1] range (and channel first)
     closest_frames = closest_frames.type(torch.float32) / 255
@@ -348,7 +350,7 @@ def decode_video_frames_torchcodec(
         loaded_frames.append(frame)
         loaded_ts.append(pts.item())
         if log_loaded_timestamps:
-            logging.info(f"Frame loaded at timestamp={pts:.4f}")
+            logger.info(f"Frame loaded at timestamp={pts:.4f}")
 
     query_ts = torch.tensor(timestamps)
     loaded_ts = torch.tensor(loaded_ts)
@@ -374,7 +376,7 @@ def decode_video_frames_torchcodec(
     closest_ts = loaded_ts[argmin_]
 
     if log_loaded_timestamps:
-        logging.info(f"{closest_ts=}")
+        logger.info(f"{closest_ts=}")
 
     # convert to float32 in [0,1] range
     closest_frames = (closest_frames / 255.0).type(torch.float32)
@@ -408,14 +410,14 @@ def encode_video_frames(
     imgs_dir = Path(imgs_dir)
 
     if video_path.exists() and not overwrite:
-        logging.warning(f"Video file already exists: {video_path}. Skipping encoding.")
+        logger.warning(f"Video file already exists: {video_path}. Skipping encoding.")
         return
 
     video_path.parent.mkdir(parents=True, exist_ok=True)
 
     # Encoders/pixel formats incompatibility check
     if (vcodec == "libsvtav1" or vcodec == "hevc") and pix_fmt == "yuv444p":
-        logging.warning(
+        logger.warning(
             f"Incompatible pixel format 'yuv444p' for codec {vcodec}, auto-selecting format 'yuv420p'"
         )
         pix_fmt = "yuv420p"
@@ -508,7 +510,7 @@ def concatenate_video_files(
     output_video_path = Path(output_video_path)
 
     if output_video_path.exists() and not overwrite:
-        logging.warning(f"Video file already exists: {output_video_path}. Skipping concatenation.")
+        logger.warning(f"Video file already exists: {output_video_path}. Skipping concatenation.")
         return
 
     output_video_path.parent.mkdir(parents=True, exist_ok=True)
@@ -693,7 +695,7 @@ class _CameraEncoderThread(threading.Thread):
                 self.result_queue.put(("ok", None))
 
         except Exception as e:
-            logging.error(f"Encoder thread error: {e}")
+            logger.error(f"Encoder thread error: {e}")
             if container is not None:
                 with contextlib.suppress(Exception):
                     container.close()
@@ -819,7 +821,7 @@ class StreamingVideoEncoder:
             count = self._dropped_frames[video_key]
             # Log periodically to avoid spam (1st, then every 10th)
             if count == 1 or count % 10 == 0:
-                logging.warning(
+                logger.warning(
                     f"Encoder queue full for {video_key}, dropped {count} frame(s). "
                     f"Consider using vcodec='auto' for hardware encoding or increasing encoder_queue_maxsize."
                 )
@@ -841,7 +843,7 @@ class StreamingVideoEncoder:
         # Report dropped frames
         for video_key, count in self._dropped_frames.items():
             if count > 0:
-                logging.warning(f"Episode finished with {count} dropped frame(s) for {video_key}.")
+                logger.warning(f"Episode finished with {count} dropped frame(s) for {video_key}.")
 
         # Send sentinel to all queues
         for video_key in self._frame_queues:
@@ -851,7 +853,7 @@ class StreamingVideoEncoder:
         for video_key in self._threads:
             self._threads[video_key].join(timeout=120)
             if self._threads[video_key].is_alive():
-                logging.error(f"Encoder thread for {video_key} did not finish in time")
+                logger.error(f"Encoder thread for {video_key} did not finish in time")
                 self._stop_events[video_key].set()
                 self._threads[video_key].join(timeout=5)
                 results[video_key] = (self._video_paths[video_key], None)
@@ -863,7 +865,7 @@ class StreamingVideoEncoder:
                     raise RuntimeError(f"Encoder thread for {video_key} failed: {data}")
                 results[video_key] = (self._video_paths[video_key], data)
             except queue.Empty:
-                logging.error(f"No result from encoder thread for {video_key}")
+                logger.error(f"No result from encoder thread for {video_key}")
                 results[video_key] = (self._video_paths[video_key], None)
 
         self._cleanup()
@@ -1071,13 +1073,13 @@ class VideoEncodingManager:
         elif self.dataset.episodes_since_last_encoding > 0:
             # Handle any remaining episodes that haven't been batch encoded
             if exc_type is not None:
-                logging.info("Exception occurred. Encoding remaining episodes before exit...")
+                logger.info("Exception occurred. Encoding remaining episodes before exit...")
             else:
-                logging.info("Recording stopped. Encoding remaining episodes...")
+                logger.info("Recording stopped. Encoding remaining episodes...")
 
             start_ep = self.dataset.num_episodes - self.dataset.episodes_since_last_encoding
             end_ep = self.dataset.num_episodes
-            logging.info(
+            logger.info(
                 f"Encoding remaining {self.dataset.episodes_since_last_encoding} episodes, "
                 f"from episode {start_ep} to {end_ep - 1}"
             )
@@ -1094,7 +1096,7 @@ class VideoEncodingManager:
                     episode_index=interrupted_episode_index, image_key=key, frame_index=0
                 ).parent
                 if img_dir.exists():
-                    logging.debug(
+                    logger.debug(
                         f"Cleaning up interrupted episode images for episode {interrupted_episode_index}, camera {key}"
                     )
                     shutil.rmtree(img_dir)
@@ -1105,8 +1107,8 @@ class VideoEncodingManager:
             png_files = list(img_dir.rglob("*.png"))
             if len(png_files) == 0:
                 shutil.rmtree(img_dir)
-                logging.debug("Cleaned up empty images directory")
+                logger.debug("Cleaned up empty images directory")
             else:
-                logging.debug(f"Images directory is not empty, containing {len(png_files)} PNG files")
+                logger.debug(f"Images directory is not empty, containing {len(png_files)} PNG files")
 
         return False  # Don't suppress the original exception

src/lerobot/envs/libero.py

@@ -29,7 +29,7 @@ from gymnasium import spaces
 from libero.libero import benchmark, get_libero_path
 from libero.libero.envs import OffScreenRenderEnv
 
-from lerobot.processor import RobotObservation
+from lerobot.types import RobotObservation
 
 
 def _parse_camera_names(camera_name: str | Sequence[str]) -> list[str]:

src/lerobot/envs/metaworld.py

@@ -25,7 +25,7 @@ import metaworld.policies as policies
 import numpy as np
 from gymnasium import spaces
 
-from lerobot.processor import RobotObservation
+from lerobot.types import RobotObservation
 
 # ---- Load configuration data from the external JSON file ----
 CONFIG_PATH = Path(__file__).parent / "metaworld_config.json"

src/lerobot/envs/utils.py

@@ -29,7 +29,7 @@ from torch import Tensor
 
 from lerobot.configs.types import FeatureType, PolicyFeature
 from lerobot.envs.configs import EnvConfig
-from lerobot.processor import RobotObservation
+from lerobot.types import RobotObservation
 from lerobot.utils.constants import OBS_ENV_STATE, OBS_IMAGE, OBS_IMAGES, OBS_STATE, OBS_STR
 from lerobot.utils.utils import get_channel_first_image_shape
 

src/lerobot/motors/dynamixel/dynamixel.py

@@ -122,14 +122,12 @@ class DynamixelMotorsBus(SerialMotorsBus):
         port: str,
         motors: dict[str, Motor],
         calibration: dict[str, MotorCalibration] | None = None,
-        protocol_version: int = PROTOCOL_VERSION,
     ):
         super().__init__(port, motors, calibration)
         import dynamixel_sdk as dxl
 
         self.port_handler = dxl.PortHandler(self.port)
-        self.packet_handler = dxl.PacketHandler(protocol_version)
-        print(f"Using protocol version {protocol_version}")
+        self.packet_handler = dxl.PacketHandler(PROTOCOL_VERSION)
         self.sync_reader = dxl.GroupSyncRead(self.port_handler, self.packet_handler, 0, 0)
         self.sync_writer = dxl.GroupSyncWrite(self.port_handler, self.packet_handler, 0, 0)
         self._comm_success = dxl.COMM_SUCCESS

src/lerobot/motors/dynamixel/tables.py

@@ -33,58 +33,6 @@
 # 2. We can change the value of the MyControlTableKey enums without impacting the client code
 
 
-# {data_name: (address, size_byte)}
-# https://emanual.robotis.com/docs/en/dxl/ax/{MODEL}/#control-table
-AX_SERIES_CONTROL_TABLE = {
-    # EEPROM Area
-    "Model_Number": (0, 2),
-    "Firmware_Version": (2, 1),
-    "ID": (3, 1),
-    "Baud_Rate": (4, 1),
-    "Return_Delay_Time": (5, 1),
-    "CW_Angle_Limit": (6, 2),
-    "CCW_Angle_Limit": (8, 2),
-    "Temperature_Limit": (11, 1),
-    "Min_Voltage_Limit": (12, 1),
-    "Max_Voltage_Limit": (13, 1),
-    "Max_Torque": (14, 2),
-    "Status_Return_Level": (16, 1),
-    "Alarm_LED": (17, 1),
-    "Shutdown": (18, 1),
-    # RAM Area
-    "Torque_Enable": (24, 1),
-    "LED": (25, 1),
-    "CW_Compliance_Margin": (26, 1),
-    "CCW_Compliance_Margin": (27, 1),
-    "CW_Compliance_Slope": (28, 1),
-    "CCW_Compliance_Slope": (29, 1),
-    "Goal_Position": (30, 2),
-    "Moving_Speed": (32, 2),
-    "Torque_Limit": (34, 2),
-    "Present_Position": (36, 2),
-    "Present_Speed": (38, 2),
-    "Present_Load": (40, 2),
-    "Present_Voltage": (42, 1),
-    "Present_Temperature": (43, 1),
-    "Registered": (44, 1),
-    "Moving": (46, 1),
-    "Lock": (47, 1),
-    "Punch": (48, 2),
-}
-
-# https://emanual.robotis.com/docs/en/dxl/ax/{MODEL}/#baud-rate4
-AX_SERIES_BAUDRATE_TABLE = {
-    9_600: 207,
-    19_200: 103,
-    57_600: 34,
-    115_200: 16,
-    200_000: 9,
-    250_000: 7,
-    400_000: 4,
-    500_000: 3,
-    1_000_000: 1,
-}
-
 # {data_name: (address, size_byte)}
 # https://emanual.robotis.com/docs/en/dxl/x/{MODEL}/#control-table
 X_SERIES_CONTROL_TABLE = {
@@ -166,14 +114,6 @@ X_SERIES_ENCODINGS_TABLE = {
     "Present_Velocity": X_SERIES_CONTROL_TABLE["Present_Velocity"][1],
 }
 
-# {data_name: size_byte}
-AX_SERIES_ENCODINGS_TABLE = {
-    "Goal_Position": AX_SERIES_CONTROL_TABLE["Goal_Position"][1],
-    "Moving_Speed": AX_SERIES_CONTROL_TABLE["Moving_Speed"][1],
-    "Present_Position": AX_SERIES_CONTROL_TABLE["Present_Position"][1],
-    "Present_Speed": AX_SERIES_CONTROL_TABLE["Present_Speed"][1],
-}
-
 MODEL_ENCODING_TABLE = {
     "x_series": X_SERIES_ENCODINGS_TABLE,
     "xl330-m077": X_SERIES_ENCODINGS_TABLE,
@@ -182,8 +122,6 @@ MODEL_ENCODING_TABLE = {
     "xm430-w350": X_SERIES_ENCODINGS_TABLE,
     "xm540-w270": X_SERIES_ENCODINGS_TABLE,
     "xc430-w150": X_SERIES_ENCODINGS_TABLE,
-    "ax_series": AX_SERIES_ENCODINGS_TABLE,
-    "ax-12a": AX_SERIES_ENCODINGS_TABLE,
 }
 
 # {model: model_resolution}
@@ -196,8 +134,6 @@ MODEL_RESOLUTION = {
     "xm430-w350": 4096,
     "xm540-w270": 4096,
     "xc430-w150": 4096,
-    "ax_series": 1024,
-    "ax-12a": 1024,
 }
 
 # {model: model_number}
@@ -209,7 +145,6 @@ MODEL_NUMBER_TABLE = {
     "xm430-w350": 1020,
     "xm540-w270": 1120,
     "xc430-w150": 1070,
-    "ax-12a": 12,
 }
 
 # {model: available_operating_modes}
@@ -231,8 +166,6 @@ MODEL_CONTROL_TABLE = {
     "xm430-w350": X_SERIES_CONTROL_TABLE,
     "xm540-w270": X_SERIES_CONTROL_TABLE,
     "xc430-w150": X_SERIES_CONTROL_TABLE,
-    "ax_series": AX_SERIES_CONTROL_TABLE,
-    "ax-12a": AX_SERIES_CONTROL_TABLE,
 }
 
 MODEL_BAUDRATE_TABLE = {
@@ -243,8 +176,6 @@ MODEL_BAUDRATE_TABLE = {
     "xm430-w350": X_SERIES_BAUDRATE_TABLE,
     "xm540-w270": X_SERIES_BAUDRATE_TABLE,
     "xc430-w150": X_SERIES_BAUDRATE_TABLE,
-    "ax_series": AX_SERIES_BAUDRATE_TABLE,
-    "ax-12a": AX_SERIES_BAUDRATE_TABLE,
 }
 
 AVAILABLE_BAUDRATES = [

src/lerobot/motors/motors_bus.py

@@ -29,7 +29,7 @@ from dataclasses import dataclass
 from enum import Enum
 from functools import cached_property
 from pprint import pformat
-from typing import Protocol, TypeAlias
+from typing import Protocol
 
 import serial
 from deepdiff import DeepDiff
@@ -38,8 +38,8 @@ from tqdm import tqdm
 from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
 from lerobot.utils.utils import enter_pressed, move_cursor_up
 
-NameOrID: TypeAlias = str | int
-Value: TypeAlias = int | float
+type NameOrID = str | int
+type Value = int | float
 
 logger = logging.getLogger(__name__)
 
@@ -1277,4 +1277,4 @@ class SerialMotorsBus(MotorsBusBase):
 
 
 # Backward compatibility alias
-MotorsBus: TypeAlias = SerialMotorsBus
+MotorsBus = SerialMotorsBus

src/lerobot/optim/optimizers.py

@@ -23,7 +23,8 @@ import draccus
 import torch
 from safetensors.torch import load_file, save_file
 
-from lerobot.datasets.utils import flatten_dict, unflatten_dict, write_json
+from lerobot.datasets.io_utils import write_json
+from lerobot.datasets.utils import flatten_dict, unflatten_dict
 from lerobot.utils.constants import (
     OPTIMIZER_PARAM_GROUPS,
     OPTIMIZER_STATE,

src/lerobot/optim/schedulers.py

@@ -23,7 +23,7 @@ import draccus
 from torch.optim import Optimizer
 from torch.optim.lr_scheduler import LambdaLR, LRScheduler
 
-from lerobot.datasets.utils import write_json
+from lerobot.datasets.io_utils import write_json
 from lerobot.utils.constants import SCHEDULER_STATE
 from lerobot.utils.io_utils import deserialize_json_into_object
 

src/lerobot/policies/factory.py

@@ -18,15 +18,14 @@ from __future__ import annotations
 
 import importlib
 import logging
-from typing import Any, TypedDict
+from typing import Any, TypedDict, Unpack
 
 import torch
-from typing_extensions import Unpack
 
 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.configs.types import FeatureType
-from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
-from lerobot.datasets.utils import dataset_to_policy_features
+from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
+from lerobot.datasets.feature_utils import dataset_to_policy_features
 from lerobot.envs.configs import EnvConfig
 from lerobot.envs.utils import env_to_policy_features
 from lerobot.policies.act.configuration_act import ACTConfig
@@ -44,13 +43,14 @@ from lerobot.policies.utils import validate_visual_features_consistency
 from lerobot.policies.vqbet.configuration_vqbet import VQBeTConfig
 from lerobot.policies.wall_x.configuration_wall_x import WallXConfig
 from lerobot.policies.xvla.configuration_xvla import XVLAConfig
-from lerobot.processor import PolicyAction, PolicyProcessorPipeline
+from lerobot.processor import PolicyProcessorPipeline
 from lerobot.processor.converters import (
     batch_to_transition,
     policy_action_to_transition,
     transition_to_batch,
     transition_to_policy_action,
 )
+from lerobot.types import PolicyAction
 from lerobot.utils.constants import (
     ACTION,
     POLICY_POSTPROCESSOR_DEFAULT_NAME,

src/lerobot/policies/groot/eagle2_hg_model/image_processing_eagle2_5_vl_fast.py

@@ -4,17 +4,16 @@
 # Licensed under The MIT License [see LICENSE for details]
 # --------------------------------------------------------
 
+from __future__ import annotations
 
 # copy from https://github.com/huggingface/transformers/blob/main/src/transformers/models/llava_onevision/image_processing_llava_onevision_fast.py
-from typing import Optional
-
 from transformers.image_processing_utils import (
     BatchFeature,
     get_patch_output_size,
 )
 from transformers.image_processing_utils_fast import (
     BaseImageProcessorFast,
-    DefaultFastImageProcessorKwargs,
+    ImagesKwargs,
     group_images_by_shape,
     reorder_images,
 )
@@ -77,7 +76,7 @@ def crop(img: torch.Tensor, left: int, top: int, right: int, bottom: int) -> tor
     return img[:, top:bottom, left:right]
 
 
-class Eagle25VLFastImageProcessorKwargs(DefaultFastImageProcessorKwargs):
+class Eagle25VLFastImageProcessorKwargs(ImagesKwargs):
     max_dynamic_tiles: int | None
     min_dynamic_tiles: int | None
     use_thumbnail: bool | None
@@ -165,11 +164,11 @@ class Eagle25VLImageProcessorFast(BaseImageProcessorFast):
 
     def _resize_for_patching(
         self,
-        image: "torch.Tensor",
+        image: torch.Tensor,
         target_resolution: tuple,
-        interpolation: "F.InterpolationMode",
+        interpolation: F.InterpolationMode,
         input_data_format: ChannelDimension,
-    ) -> "torch.Tensor":
+    ) -> torch.Tensor:
         """
         Resizes an image to a target resolution while maintaining aspect ratio.
 
@@ -219,8 +218,8 @@ class Eagle25VLImageProcessorFast(BaseImageProcessorFast):
         return best_ratio
 
     def _pad_for_patching(
-        self, image: "torch.Tensor", target_resolution: tuple, input_data_format: ChannelDimension
-    ) -> "torch.Tensor":
+        self, image: torch.Tensor, target_resolution: tuple, input_data_format: ChannelDimension
+    ) -> torch.Tensor:
         """
         Pad an image to a target resolution while maintaining aspect ratio.
         """
@@ -236,15 +235,15 @@ class Eagle25VLImageProcessorFast(BaseImageProcessorFast):
 
     def _get_image_patches(
         self,
-        image: "torch.Tensor",
+        image: torch.Tensor,
         min_num: int,
         max_num: int,
         size: tuple,
         tile_size: int,
         use_thumbnail: bool,
-        interpolation: "F.InterpolationMode",
+        interpolation: F.InterpolationMode,
         pad_during_tiling: bool,
-    ) -> list["torch.Tensor"]:
+    ) -> list[torch.Tensor]:
         image_size = get_image_size(image, channel_dim=ChannelDimension.FIRST)
         orig_height, orig_width = image_size
         aspect_ratio = orig_width / orig_height
@@ -305,8 +304,8 @@ class Eagle25VLImageProcessorFast(BaseImageProcessorFast):
 
     def _pad_for_batching(
         self,
-        pixel_values: list["torch.Tensor"],
-    ) -> list["torch.Tensor"]:
+        pixel_values: list[torch.Tensor],
+    ) -> list[torch.Tensor]:
         """
         Pads images on the `num_of_patches` dimension with zeros to form a batch of same number of patches.
 
@@ -327,14 +326,14 @@ class Eagle25VLImageProcessorFast(BaseImageProcessorFast):
 
     def _preprocess(
         self,
-        images: list["torch.Tensor"],
+        images: list[torch.Tensor],
         do_resize: bool,
         size: SizeDict,
         max_dynamic_tiles: int,
         min_dynamic_tiles: int,
         use_thumbnail: bool,
         pad_during_tiling: bool,
-        interpolation: Optional["F.InterpolationMode"],
+        interpolation: F.InterpolationMode | None,
         do_center_crop: bool,
         crop_size: SizeDict,
         do_rescale: bool,

src/lerobot/policies/groot/processor_groot.py

@@ -49,7 +49,7 @@ from lerobot.processor.converters import (
     policy_action_to_transition,
     transition_to_policy_action,
 )
-from lerobot.processor.core import EnvTransition, TransitionKey
+from lerobot.types import EnvTransition, TransitionKey
 from lerobot.utils.constants import (
     ACTION,
     HF_LEROBOT_HOME,

src/lerobot/policies/pi0/modeling_pi0.py

@@ -15,16 +15,16 @@
 # limitations under the License.
 
 import builtins
+import copy
 import logging
 import math
 from collections import deque
 from pathlib import Path
-from typing import TYPE_CHECKING, Literal, TypedDict
+from typing import TYPE_CHECKING, Literal, TypedDict, Unpack
 
 import torch
 import torch.nn.functional as F  # noqa: N812
 from torch import Tensor, nn
-from typing_extensions import Unpack
 
 from lerobot.utils.import_utils import _transformers_available
 
@@ -32,13 +32,21 @@ from lerobot.utils.import_utils import _transformers_available
 if TYPE_CHECKING or _transformers_available:
     from transformers.models.auto import CONFIG_MAPPING
     from transformers.models.gemma import modeling_gemma
-    from transformers.models.gemma.modeling_gemma import GemmaForCausalLM
-    from transformers.models.paligemma.modeling_paligemma import PaliGemmaForConditionalGeneration
+
+    from lerobot.policies.pi_gemma import (
+        PaliGemmaForConditionalGenerationWithPiGemma,
+        PiGemmaForCausalLM,
+        _gated_residual,
+        layernorm_forward,
+    )
 else:
     CONFIG_MAPPING = None
     modeling_gemma = None
-    GemmaForCausalLM = None
-    PaliGemmaForConditionalGeneration = None
+    PiGemmaForCausalLM = None
+    _gated_residual = None
+    layernorm_forward = None
+    PaliGemmaForConditionalGenerationWithPiGemma = None
+
 
 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.policies.pi0.configuration_pi0 import DEFAULT_IMAGE_SIZE, PI0Config
@@ -191,7 +199,7 @@ def resize_with_pad_torch(  # see openpi `resize_with_pad_torch` (exact copy)
     if images.dtype == torch.uint8:
         resized_images = torch.round(resized_images).clamp(0, 255).to(torch.uint8)
     elif images.dtype == torch.float32:
-        resized_images = resized_images.clamp(-1.0, 1.0)
+        resized_images = resized_images.clamp(0.0, 1.0)
     else:
         raise ValueError(f"Unsupported image dtype: {images.dtype}")
 
@@ -202,7 +210,7 @@ def resize_with_pad_torch(  # see openpi `resize_with_pad_torch` (exact copy)
     pad_w1 = pad_w0 + remainder_w
 
     # Pad
-    constant_value = 0 if images.dtype == torch.uint8 else -1.0
+    constant_value = 0 if images.dtype == torch.uint8 else 0.0
     padded_images = F.pad(
         resized_images,
         (pad_w0, pad_w1, pad_h0, pad_h1),  # left, right, top, bottom
@@ -221,14 +229,14 @@ def resize_with_pad_torch(  # see openpi `resize_with_pad_torch` (exact copy)
 def compute_layer_complete(
     layer_idx, inputs_embeds, attention_mask, position_ids, adarms_cond, paligemma, gemma_expert
 ):
-    models = [paligemma.language_model, gemma_expert.model]
+    models = [paligemma.model.language_model, gemma_expert.model]
     query_states = []
     key_states = []
     value_states = []
     gates = []
     for i, hidden_states in enumerate(inputs_embeds):
         layer = models[i].layers[layer_idx]
-        hidden_states, gate = layer.input_layernorm(hidden_states, cond=adarms_cond[i])  # noqa: PLW2901
+        hidden_states, gate = layernorm_forward(layer.input_layernorm, hidden_states, adarms_cond[i])
         gates.append(gate)
         input_shape = hidden_states.shape[:-1]
         hidden_shape = (*input_shape, -1, layer.self_attn.head_dim)
@@ -254,10 +262,10 @@ def compute_layer_complete(
         query_states, key_states, cos, sin, unsqueeze_dim=1
     )
     batch_size = query_states.shape[0]
-    scaling = paligemma.language_model.layers[layer_idx].self_attn.scaling
+    scaling = paligemma.model.language_model.layers[layer_idx].self_attn.scaling
     # Attention computation
     att_output, _ = modeling_gemma.eager_attention_forward(
-        paligemma.language_model.layers[layer_idx].self_attn,
+        paligemma.model.language_model.layers[layer_idx].self_attn,
         query_states,
         key_states,
         value_states,
@@ -265,7 +273,7 @@ def compute_layer_complete(
         scaling,
     )
     # Get head_dim from the current layer, not from the model
-    head_dim = paligemma.language_model.layers[layer_idx].self_attn.head_dim
+    head_dim = paligemma.model.language_model.layers[layer_idx].self_attn.head_dim
     att_output = att_output.reshape(batch_size, -1, 1 * 8 * head_dim)
     # Process layer outputs
     outputs_embeds = []
@@ -277,15 +285,15 @@ def compute_layer_complete(
             att_output = att_output.to(layer.self_attn.o_proj.weight.dtype)
         out_emb = layer.self_attn.o_proj(att_output[:, start_pos:end_pos])
         # first residual
-        out_emb = modeling_gemma._gated_residual(hidden_states, out_emb, gates[i])  # noqa: SLF001
+        out_emb = _gated_residual(hidden_states, out_emb, gates[i])
         after_first_residual = out_emb.clone()
-        out_emb, gate = layer.post_attention_layernorm(out_emb, cond=adarms_cond[i])
+        out_emb, gate = layernorm_forward(layer.post_attention_layernorm, out_emb, adarms_cond[i])
         # Convert to bfloat16 if the next layer (mlp) uses bfloat16
         if layer.mlp.up_proj.weight.dtype == torch.bfloat16:
             out_emb = out_emb.to(dtype=torch.bfloat16)
         out_emb = layer.mlp(out_emb)
         # second residual
-        out_emb = modeling_gemma._gated_residual(after_first_residual, out_emb, gate)  # noqa: SLF001
+        out_emb = _gated_residual(after_first_residual, out_emb, gate)
         outputs_embeds.append(out_emb)
         start_pos = end_pos
     return outputs_embeds
@@ -358,7 +366,7 @@ class PaliGemmaWithExpertModel(
         vlm_config_hf.text_config.num_hidden_layers = vlm_config.depth
         vlm_config_hf.text_config.num_key_value_heads = vlm_config.num_kv_heads
         vlm_config_hf.text_config.hidden_activation = "gelu_pytorch_tanh"
-        vlm_config_hf.text_config.torch_dtype = "float32"
+        vlm_config_hf.text_config.dtype = "float32"
         vlm_config_hf.text_config.vocab_size = 257152
         vlm_config_hf.text_config.use_adarms = use_adarms[0]
         vlm_config_hf.text_config.adarms_cond_dim = vlm_config.width if use_adarms[0] else None
@@ -366,7 +374,7 @@ class PaliGemmaWithExpertModel(
         vlm_config_hf.vision_config.intermediate_size = 4304
         vlm_config_hf.vision_config.projection_dim = 2048
         vlm_config_hf.vision_config.projector_hidden_act = "gelu_fast"
-        vlm_config_hf.vision_config.torch_dtype = "float32"
+        vlm_config_hf.vision_config.dtype = "float32"
 
         action_expert_config_hf = CONFIG_MAPPING["gemma"](
             head_dim=action_expert_config.head_dim,
@@ -377,13 +385,13 @@ class PaliGemmaWithExpertModel(
             num_key_value_heads=action_expert_config.num_kv_heads,
             vocab_size=257152,
             hidden_activation="gelu_pytorch_tanh",
-            torch_dtype="float32",
+            dtype="float32",
             use_adarms=use_adarms[1],
             adarms_cond_dim=action_expert_config.width if use_adarms[1] else None,
         )
 
-        self.paligemma = PaliGemmaForConditionalGeneration(config=vlm_config_hf)
-        self.gemma_expert = GemmaForCausalLM(config=action_expert_config_hf)
+        self.paligemma = PaliGemmaForConditionalGenerationWithPiGemma(config=vlm_config_hf)
+        self.gemma_expert = PiGemmaForCausalLM(config=action_expert_config_hf)
         self.gemma_expert.model.embed_tokens = None
 
         self.to_bfloat16_for_selected_params(precision)
@@ -398,10 +406,11 @@ class PaliGemmaWithExpertModel(
         else:
             raise ValueError(f"Invalid precision: {precision}")
 
+        # Keep full vision path in float32 so we never toggle (toggle causes optimizer
+        # "same dtype" error). Align with PI05.
         params_to_keep_float32 = [
-            "vision_tower.vision_model.embeddings.patch_embedding.weight",
-            "vision_tower.vision_model.embeddings.patch_embedding.bias",
-            "vision_tower.vision_model.embeddings.position_embedding.weight",
+            "vision_tower",
+            "multi_modal_projector",
             "input_layernorm",
             "post_attention_layernorm",
             "model.norm",
@@ -413,8 +422,8 @@ class PaliGemmaWithExpertModel(
 
     def _set_requires_grad(self):
         if self.freeze_vision_encoder:
-            self.paligemma.vision_tower.eval()
-            for param in self.paligemma.vision_tower.parameters():
+            self.paligemma.model.vision_tower.eval()
+            for param in self.paligemma.model.vision_tower.parameters():
                 param.requires_grad = False
         if self.train_expert_only:
             self.paligemma.eval()
@@ -424,15 +433,23 @@ class PaliGemmaWithExpertModel(
     def train(self, mode: bool = True):
         super().train(mode)
         if self.freeze_vision_encoder:
-            self.paligemma.vision_tower.eval()
+            self.paligemma.model.vision_tower.eval()
         if self.train_expert_only:
             self.paligemma.eval()
 
     def embed_image(self, image: torch.Tensor):
-        return self.paligemma.model.get_image_features(image)
+        # Vision tower and multi_modal_projector are kept in float32 (params_to_keep_float32). Align with PI05.
+        out_dtype = image.dtype
+        if image.dtype != torch.float32:
+            image = image.to(torch.float32)
+        image_outputs = self.paligemma.model.get_image_features(image)
+        features = image_outputs.pooler_output * self.paligemma.config.text_config.hidden_size**0.5
+        if features.dtype != out_dtype:
+            features = features.to(out_dtype)
+        return features
 
     def embed_language_tokens(self, tokens: torch.Tensor):
-        return self.paligemma.language_model.embed_tokens(tokens)
+        return self.paligemma.model.language_model.embed_tokens(tokens)
 
     def forward(
         self,
@@ -446,7 +463,7 @@ class PaliGemmaWithExpertModel(
         if adarms_cond is None:
             adarms_cond = [None, None]
         if inputs_embeds[1] is None:
-            prefix_output = self.paligemma.language_model.forward(
+            prefix_output = self.paligemma.model.language_model.forward(
                 inputs_embeds=inputs_embeds[0],
                 attention_mask=attention_mask,
                 position_ids=position_ids,
@@ -470,7 +487,7 @@ class PaliGemmaWithExpertModel(
             prefix_output = None
             prefix_past_key_values = None
         else:
-            models = [self.paligemma.language_model, self.gemma_expert.model]
+            models = [self.paligemma.model.language_model, self.gemma_expert.model]
             num_layers = self.paligemma.config.text_config.num_hidden_layers
 
             # Check if gradient checkpointing is enabled for any of the models
@@ -510,7 +527,7 @@ class PaliGemmaWithExpertModel(
             def compute_final_norms(inputs_embeds, adarms_cond):
                 outputs_embeds = []
                 for i, hidden_states in enumerate(inputs_embeds):
-                    out_emb, _ = models[i].norm(hidden_states, cond=adarms_cond[i])
+                    out_emb, _ = layernorm_forward(models[i].norm, hidden_states, adarms_cond[i])
                     outputs_embeds.append(out_emb)
                 return outputs_embeds
 
@@ -576,29 +593,19 @@ class PI0Pytorch(nn.Module):  # see openpi `PI0Pytorch`
             # Also compile the main forward pass used during training
             self.forward = torch.compile(self.forward, mode=config.compile_mode)
 
-        msg = """An incorrect transformer version is used, please create an issue on https://github.com/huggingface/lerobot/issues"""
-
-        try:
-            from transformers.models.siglip import check
-
-            if not check.check_whether_transformers_replace_is_installed_correctly():
-                raise ValueError(msg)
-        except ImportError:
-            raise ValueError(msg) from None
-
     def gradient_checkpointing_enable(self):
         """Enable gradient checkpointing for memory optimization."""
         self.gradient_checkpointing_enabled = True
-        self.paligemma_with_expert.paligemma.language_model.gradient_checkpointing = True
-        self.paligemma_with_expert.paligemma.vision_tower.gradient_checkpointing = True
+        self.paligemma_with_expert.paligemma.model.language_model.gradient_checkpointing = True
+        self.paligemma_with_expert.paligemma.model.vision_tower.gradient_checkpointing = True
         self.paligemma_with_expert.gemma_expert.model.gradient_checkpointing = True
         logging.info("Enabled gradient checkpointing for PI0Pytorch model")
 
     def gradient_checkpointing_disable(self):
         """Disable gradient checkpointing."""
         self.gradient_checkpointing_enabled = False
-        self.paligemma_with_expert.paligemma.language_model.gradient_checkpointing = False
-        self.paligemma_with_expert.paligemma.vision_tower.gradient_checkpointing = False
+        self.paligemma_with_expert.paligemma.model.language_model.gradient_checkpointing = False
+        self.paligemma_with_expert.paligemma.model.vision_tower.gradient_checkpointing = False
         self.paligemma_with_expert.gemma_expert.model.gradient_checkpointing = False
         logging.info("Disabled gradient checkpointing for PI0Pytorch model")
 
@@ -760,7 +767,7 @@ class PI0Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         suffix_embs, suffix_pad_masks, suffix_att_masks, adarms_cond = self.embed_suffix(state, x_t, time)
 
         if (
-            self.paligemma_with_expert.paligemma.language_model.layers[0].self_attn.q_proj.weight.dtype
+            self.paligemma_with_expert.paligemma.model.language_model.layers[0].self_attn.q_proj.weight.dtype
             == torch.bfloat16
         ):
             suffix_embs = suffix_embs.to(dtype=torch.bfloat16)
@@ -834,7 +841,7 @@ class PI0Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         prefix_position_ids = torch.cumsum(prefix_pad_masks, dim=1) - 1
 
         prefix_att_2d_masks_4d = self._prepare_attention_masks_4d(prefix_att_2d_masks)
-        self.paligemma_with_expert.paligemma.language_model.config._attn_implementation = "eager"  # noqa: SLF001
+        self.paligemma_with_expert.paligemma.model.language_model.config._attn_implementation = "eager"  # noqa: SLF001
 
         _, past_key_values = self.paligemma_with_expert.forward(
             attention_mask=prefix_att_2d_masks_4d,
@@ -908,6 +915,7 @@ class PI0Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         full_att_2d_masks_4d = self._prepare_attention_masks_4d(full_att_2d_masks)
         self.paligemma_with_expert.gemma_expert.model.config._attn_implementation = "eager"  # noqa: SLF001
 
+        past_key_values = copy.deepcopy(past_key_values)
         outputs_embeds, _ = self.paligemma_with_expert.forward(
             attention_mask=full_att_2d_masks_4d,
             position_ids=position_ids,
@@ -997,14 +1005,12 @@ class PI0Policy(PreTrainedPolicy):
         # Check if dataset_stats were provided in kwargs
         model = cls(config, **kwargs)
 
-        # Now manually load and remap the state dict
+        # Load state dict (expects keys with "model." prefix)
         try:
-            # Try to load the pytorch_model.bin or model.safetensors file
             print(f"Loading model from: {pretrained_name_or_path}")
             try:
                 from transformers.utils import cached_file
 
-                # Try safetensors first
                 resolved_file = cached_file(
                     pretrained_name_or_path,
                     "model.safetensors",
@@ -1012,7 +1018,7 @@ class PI0Policy(PreTrainedPolicy):
                     force_download=kwargs.get("force_download", False),
                     resume_download=kwargs.get("resume_download"),
                     proxies=kwargs.get("proxies"),
-                    use_auth_token=kwargs.get("use_auth_token"),
+                    token=kwargs.get("token"),
                     revision=kwargs.get("revision"),
                     local_files_only=kwargs.get("local_files_only", False),
                 )
@@ -1025,7 +1031,7 @@ class PI0Policy(PreTrainedPolicy):
                 print("Returning model without loading pretrained weights")
                 return model
 
-            # First, fix any key differences # see openpi `model.py, _fix_pytorch_state_dict_keys`
+            # First, fix any key differences (see openpi model.py, _fix_pytorch_state_dict_keys)
             fixed_state_dict = model._fix_pytorch_state_dict_keys(original_state_dict, model.config)
 
             # Then add "model." prefix for all keys that don't already have it
@@ -1070,7 +1076,7 @@ class PI0Policy(PreTrainedPolicy):
                 print("All keys loaded successfully!")
 
         except Exception as e:
-            print(f"Warning: Could not remap state dict keys: {e}")
+            print(f"Warning: Could not load state dict: {e}")
 
         return model
 
@@ -1120,6 +1126,14 @@ class PI0Policy(PreTrainedPolicy):
                 # Some checkpoints might have this, but current model expects different structure
                 logging.warning(f"Vision embedding key might need handling: {key}")
 
+            if (
+                key == "model.paligemma_with_expert.paligemma.lm_head.weight"
+                or key == "paligemma_with_expert.paligemma.lm_head.weight"
+            ):
+                fixed_state_dict[
+                    "model.paligemma_with_expert.paligemma.model.language_model.embed_tokens.weight"
+                ] = value.clone()
+
             fixed_state_dict[new_key] = value
 
         return fixed_state_dict

src/lerobot/policies/pi05/modeling_pi05.py

@@ -15,16 +15,16 @@
 # limitations under the License.
 
 import builtins
+import copy
 import logging
 import math
 from collections import deque
 from pathlib import Path
-from typing import TYPE_CHECKING, Literal, TypedDict
+from typing import TYPE_CHECKING, Literal, TypedDict, Unpack
 
 import torch
 import torch.nn.functional as F  # noqa: N812
 from torch import Tensor, nn
-from typing_extensions import Unpack
 
 from lerobot.utils.import_utils import _transformers_available
 
@@ -32,14 +32,20 @@ from lerobot.utils.import_utils import _transformers_available
 if TYPE_CHECKING or _transformers_available:
     from transformers.models.auto import CONFIG_MAPPING
     from transformers.models.gemma import modeling_gemma
-    from transformers.models.gemma.modeling_gemma import GemmaForCausalLM
-    from transformers.models.paligemma.modeling_paligemma import PaliGemmaForConditionalGeneration
+
+    from lerobot.policies.pi_gemma import (
+        PaliGemmaForConditionalGenerationWithPiGemma,
+        PiGemmaForCausalLM,
+        _gated_residual,
+        layernorm_forward,
+    )
 else:
     CONFIG_MAPPING = None
     modeling_gemma = None
-    GemmaForCausalLM = None
-    PaliGemmaForConditionalGeneration = None
-
+    PiGemmaForCausalLM = None
+    _gated_residual = None
+    layernorm_forward = None
+    PaliGemmaForConditionalGenerationWithPiGemma = None
 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.policies.pi05.configuration_pi05 import DEFAULT_IMAGE_SIZE, PI05Config
 from lerobot.policies.pretrained import PreTrainedPolicy, T
@@ -92,10 +98,11 @@ def create_sinusoidal_pos_embedding(  # see openpi `create_sinusoidal_pos_embedd
 
 
 def sample_beta(alpha, beta, bsize, device):  # see openpi `sample_beta` (exact copy)
-    alpha_t = torch.as_tensor(alpha, dtype=torch.float32, device=device)
-    beta_t = torch.as_tensor(beta, dtype=torch.float32, device=device)
+    # Beta sampling uses _sample_dirichlet which isn't implemented for MPS, so sample on CPU
+    alpha_t = torch.tensor(alpha, dtype=torch.float32)
+    beta_t = torch.tensor(beta, dtype=torch.float32)
     dist = torch.distributions.Beta(alpha_t, beta_t)
-    return dist.sample((bsize,))
+    return dist.sample((bsize,)).to(device)
 
 
 def make_att_2d_masks(pad_masks, att_masks):  # see openpi `make_att_2d_masks` (exact copy)
@@ -189,7 +196,7 @@ def resize_with_pad_torch(  # see openpi `resize_with_pad_torch` (exact copy)
     if images.dtype == torch.uint8:
         resized_images = torch.round(resized_images).clamp(0, 255).to(torch.uint8)
     elif images.dtype == torch.float32:
-        resized_images = resized_images.clamp(-1.0, 1.0)
+        resized_images = resized_images.clamp(0.0, 1.0)
     else:
         raise ValueError(f"Unsupported image dtype: {images.dtype}")
 
@@ -200,7 +207,7 @@ def resize_with_pad_torch(  # see openpi `resize_with_pad_torch` (exact copy)
     pad_w1 = pad_w0 + remainder_w
 
     # Pad
-    constant_value = 0 if images.dtype == torch.uint8 else -1.0
+    constant_value = 0 if images.dtype == torch.uint8 else 0.0
     padded_images = F.pad(
         resized_images,
         (pad_w0, pad_w1, pad_h0, pad_h1),  # left, right, top, bottom
@@ -219,14 +226,14 @@ def resize_with_pad_torch(  # see openpi `resize_with_pad_torch` (exact copy)
 def compute_layer_complete(
     layer_idx, inputs_embeds, attention_mask, position_ids, adarms_cond, paligemma, gemma_expert
 ):
-    models = [paligemma.language_model, gemma_expert.model]
+    models = [paligemma.model.language_model, gemma_expert.model]
     query_states = []
     key_states = []
     value_states = []
     gates = []
     for i, hidden_states in enumerate(inputs_embeds):
         layer = models[i].layers[layer_idx]
-        hidden_states, gate = layer.input_layernorm(hidden_states, cond=adarms_cond[i])  # noqa: PLW2901
+        hidden_states, gate = layernorm_forward(layer.input_layernorm, hidden_states, adarms_cond[i])
         gates.append(gate)
         input_shape = hidden_states.shape[:-1]
         hidden_shape = (*input_shape, -1, layer.self_attn.head_dim)
@@ -252,10 +259,10 @@ def compute_layer_complete(
         query_states, key_states, cos, sin, unsqueeze_dim=1
     )
     batch_size = query_states.shape[0]
-    scaling = paligemma.language_model.layers[layer_idx].self_attn.scaling
+    scaling = paligemma.model.language_model.layers[layer_idx].self_attn.scaling
     # Attention computation
     att_output, _ = modeling_gemma.eager_attention_forward(
-        paligemma.language_model.layers[layer_idx].self_attn,
+        paligemma.model.language_model.layers[layer_idx].self_attn,
         query_states,
         key_states,
         value_states,
@@ -263,7 +270,7 @@ def compute_layer_complete(
         scaling,
     )
     # Get head_dim from the current layer, not from the model
-    head_dim = paligemma.language_model.layers[layer_idx].self_attn.head_dim
+    head_dim = paligemma.model.language_model.layers[layer_idx].self_attn.head_dim
     att_output = att_output.reshape(batch_size, -1, 1 * 8 * head_dim)
     # Process layer outputs
     outputs_embeds = []
@@ -275,15 +282,15 @@ def compute_layer_complete(
             att_output = att_output.to(layer.self_attn.o_proj.weight.dtype)
         out_emb = layer.self_attn.o_proj(att_output[:, start_pos:end_pos])
         # first residual
-        out_emb = modeling_gemma._gated_residual(hidden_states, out_emb, gates[i])  # noqa: SLF001
+        out_emb = _gated_residual(hidden_states, out_emb, gates[i])
         after_first_residual = out_emb.clone()
-        out_emb, gate = layer.post_attention_layernorm(out_emb, cond=adarms_cond[i])
+        out_emb, gate = layernorm_forward(layer.post_attention_layernorm, out_emb, adarms_cond[i])
         # Convert to bfloat16 if the next layer (mlp) uses bfloat16
         if layer.mlp.up_proj.weight.dtype == torch.bfloat16:
             out_emb = out_emb.to(dtype=torch.bfloat16)
         out_emb = layer.mlp(out_emb)
         # second residual
-        out_emb = modeling_gemma._gated_residual(after_first_residual, out_emb, gate)  # noqa: SLF001
+        out_emb = _gated_residual(after_first_residual, out_emb, gate)
         outputs_embeds.append(out_emb)
         start_pos = end_pos
     return outputs_embeds
@@ -356,7 +363,7 @@ class PaliGemmaWithExpertModel(
         vlm_config_hf.text_config.num_hidden_layers = vlm_config.depth
         vlm_config_hf.text_config.num_key_value_heads = vlm_config.num_kv_heads
         vlm_config_hf.text_config.hidden_activation = "gelu_pytorch_tanh"
-        vlm_config_hf.text_config.torch_dtype = "float32"
+        vlm_config_hf.text_config.dtype = "float32"
         vlm_config_hf.text_config.vocab_size = 257152
         vlm_config_hf.text_config.use_adarms = use_adarms[0]
         vlm_config_hf.text_config.adarms_cond_dim = vlm_config.width if use_adarms[0] else None
@@ -364,7 +371,7 @@ class PaliGemmaWithExpertModel(
         vlm_config_hf.vision_config.intermediate_size = 4304
         vlm_config_hf.vision_config.projection_dim = 2048
         vlm_config_hf.vision_config.projector_hidden_act = "gelu_fast"
-        vlm_config_hf.vision_config.torch_dtype = "float32"
+        vlm_config_hf.vision_config.dtype = "float32"
 
         action_expert_config_hf = CONFIG_MAPPING["gemma"](
             head_dim=action_expert_config.head_dim,
@@ -375,13 +382,13 @@ class PaliGemmaWithExpertModel(
             num_key_value_heads=action_expert_config.num_kv_heads,
             vocab_size=257152,
             hidden_activation="gelu_pytorch_tanh",
-            torch_dtype="float32",
+            dtype="float32",
             use_adarms=use_adarms[1],
             adarms_cond_dim=action_expert_config.width if use_adarms[1] else None,
         )
 
-        self.paligemma = PaliGemmaForConditionalGeneration(config=vlm_config_hf)
-        self.gemma_expert = GemmaForCausalLM(config=action_expert_config_hf)
+        self.paligemma = PaliGemmaForConditionalGenerationWithPiGemma(config=vlm_config_hf)
+        self.gemma_expert = PiGemmaForCausalLM(config=action_expert_config_hf)
         self.gemma_expert.model.embed_tokens = None
 
         self.to_bfloat16_for_selected_params(precision)
@@ -396,10 +403,11 @@ class PaliGemmaWithExpertModel(
         else:
             raise ValueError(f"Invalid precision: {precision}")
 
+        # Keep full vision path in float32 so we never toggle (toggle causes optimizer
+        # "same dtype" error). Saves memory vs full float32; more memory than only 3 params.
         params_to_keep_float32 = [
-            "vision_tower.vision_model.embeddings.patch_embedding.weight",
-            "vision_tower.vision_model.embeddings.patch_embedding.bias",
-            "vision_tower.vision_model.embeddings.position_embedding.weight",
+            "vision_tower",
+            "multi_modal_projector",
             "input_layernorm",
             "post_attention_layernorm",
             "model.norm",
@@ -411,8 +419,8 @@ class PaliGemmaWithExpertModel(
 
     def _set_requires_grad(self):
         if self.freeze_vision_encoder:
-            self.paligemma.vision_tower.eval()
-            for param in self.paligemma.vision_tower.parameters():
+            self.paligemma.model.vision_tower.eval()
+            for param in self.paligemma.model.vision_tower.parameters():
                 param.requires_grad = False
         if self.train_expert_only:
             self.paligemma.eval()
@@ -422,15 +430,23 @@ class PaliGemmaWithExpertModel(
     def train(self, mode: bool = True):
         super().train(mode)
         if self.freeze_vision_encoder:
-            self.paligemma.vision_tower.eval()
+            self.paligemma.model.vision_tower.eval()
         if self.train_expert_only:
             self.paligemma.eval()
 
     def embed_image(self, image: torch.Tensor):
-        return self.paligemma.model.get_image_features(image)
+        # Vision tower and multi_modal_projector are kept in float32 (params_to_keep_float32).
+        out_dtype = image.dtype
+        if image.dtype != torch.float32:
+            image = image.to(torch.float32)
+        image_outputs = self.paligemma.model.get_image_features(image)
+        features = image_outputs.pooler_output * self.paligemma.config.text_config.hidden_size**0.5
+        if features.dtype != out_dtype:
+            features = features.to(out_dtype)
+        return features
 
     def embed_language_tokens(self, tokens: torch.Tensor):
-        return self.paligemma.language_model.embed_tokens(tokens)
+        return self.paligemma.model.language_model.embed_tokens(tokens)
 
     def forward(
         self,
@@ -444,7 +460,7 @@ class PaliGemmaWithExpertModel(
         if adarms_cond is None:
             adarms_cond = [None, None]
         if inputs_embeds[1] is None:
-            prefix_output = self.paligemma.language_model.forward(
+            prefix_output = self.paligemma.model.language_model.forward(
                 inputs_embeds=inputs_embeds[0],
                 attention_mask=attention_mask,
                 position_ids=position_ids,
@@ -468,7 +484,7 @@ class PaliGemmaWithExpertModel(
             prefix_output = None
             prefix_past_key_values = None
         else:
-            models = [self.paligemma.language_model, self.gemma_expert.model]
+            models = [self.paligemma.model.language_model, self.gemma_expert.model]
             num_layers = self.paligemma.config.text_config.num_hidden_layers
 
             # Check if gradient checkpointing is enabled for any of the models
@@ -508,7 +524,7 @@ class PaliGemmaWithExpertModel(
             def compute_final_norms(inputs_embeds, adarms_cond):
                 outputs_embeds = []
                 for i, hidden_states in enumerate(inputs_embeds):
-                    out_emb, _ = models[i].norm(hidden_states, cond=adarms_cond[i])
+                    out_emb, _ = layernorm_forward(models[i].norm, hidden_states, adarms_cond[i])
                     outputs_embeds.append(out_emb)
                 return outputs_embeds
 
@@ -573,29 +589,19 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
             # Also compile the main forward pass used during training
             self.forward = torch.compile(self.forward, mode=config.compile_mode)
 
-        msg = """An incorrect transformer version is used, please create an issue on https://github.com/huggingface/lerobot/issues"""
-
-        try:
-            from transformers.models.siglip import check
-
-            if not check.check_whether_transformers_replace_is_installed_correctly():
-                raise ValueError(msg)
-        except ImportError:
-            raise ValueError(msg) from None
-
     def gradient_checkpointing_enable(self):
         """Enable gradient checkpointing for memory optimization."""
         self.gradient_checkpointing_enabled = True
-        self.paligemma_with_expert.paligemma.language_model.gradient_checkpointing = True
-        self.paligemma_with_expert.paligemma.vision_tower.gradient_checkpointing = True
+        self.paligemma_with_expert.paligemma.model.language_model.gradient_checkpointing = True
+        self.paligemma_with_expert.paligemma.model.vision_tower.gradient_checkpointing = True
         self.paligemma_with_expert.gemma_expert.model.gradient_checkpointing = True
         logging.info("Enabled gradient checkpointing for PI05Pytorch model")
 
     def gradient_checkpointing_disable(self):
         """Disable gradient checkpointing."""
         self.gradient_checkpointing_enabled = False
-        self.paligemma_with_expert.paligemma.language_model.gradient_checkpointing = False
-        self.paligemma_with_expert.paligemma.vision_tower.gradient_checkpointing = False
+        self.paligemma_with_expert.paligemma.model.language_model.gradient_checkpointing = False
+        self.paligemma_with_expert.paligemma.model.vision_tower.gradient_checkpointing = False
         self.paligemma_with_expert.gemma_expert.model.gradient_checkpointing = False
         logging.info("Disabled gradient checkpointing for PI05Pytorch model")
 
@@ -737,7 +743,7 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         suffix_embs, suffix_pad_masks, suffix_att_masks, adarms_cond = self.embed_suffix(x_t, time)
 
         if (
-            self.paligemma_with_expert.paligemma.language_model.layers[0].self_attn.q_proj.weight.dtype
+            self.paligemma_with_expert.paligemma.model.language_model.layers[0].self_attn.q_proj.weight.dtype
             == torch.bfloat16
         ):
             suffix_embs = suffix_embs.to(dtype=torch.bfloat16)
@@ -808,7 +814,7 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         prefix_position_ids = torch.cumsum(prefix_pad_masks, dim=1) - 1
 
         prefix_att_2d_masks_4d = self._prepare_attention_masks_4d(prefix_att_2d_masks)
-        self.paligemma_with_expert.paligemma.language_model.config._attn_implementation = "eager"  # noqa: SLF001
+        self.paligemma_with_expert.paligemma.model.language_model.config._attn_implementation = "eager"  # noqa: SLF001
 
         _, past_key_values = self.paligemma_with_expert.forward(
             attention_mask=prefix_att_2d_masks_4d,
@@ -880,6 +886,7 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         full_att_2d_masks_4d = self._prepare_attention_masks_4d(full_att_2d_masks)
         self.paligemma_with_expert.gemma_expert.model.config._attn_implementation = "eager"  # noqa: SLF001
 
+        past_key_values = copy.deepcopy(past_key_values)
         outputs_embeds, _ = self.paligemma_with_expert.forward(
             attention_mask=full_att_2d_masks_4d,
             position_ids=position_ids,
@@ -969,14 +976,12 @@ class PI05Policy(PreTrainedPolicy):
         # Check if dataset_stats were provided in kwargs
         model = cls(config, **kwargs)
 
-        # Now manually load and remap the state dict
+        # Load state dict (expects keys with "model." prefix)
         try:
-            # Try to load the pytorch_model.bin or model.safetensors file
             print(f"Loading model from: {pretrained_name_or_path}")
             try:
                 from transformers.utils import cached_file
 
-                # Try safetensors first
                 resolved_file = cached_file(
                     pretrained_name_or_path,
                     "model.safetensors",
@@ -984,7 +989,7 @@ class PI05Policy(PreTrainedPolicy):
                     force_download=kwargs.get("force_download", False),
                     resume_download=kwargs.get("resume_download"),
                     proxies=kwargs.get("proxies"),
-                    use_auth_token=kwargs.get("use_auth_token"),
+                    token=kwargs.get("token"),
                     revision=kwargs.get("revision"),
                     local_files_only=kwargs.get("local_files_only", False),
                 )
@@ -997,7 +1002,7 @@ class PI05Policy(PreTrainedPolicy):
                 print("Returning model without loading pretrained weights")
                 return model
 
-            # First, fix any key differences # see openpi `model.py, _fix_pytorch_state_dict_keys`
+            # First, fix any key differences (see openpi model.py, _fix_pytorch_state_dict_keys)
             fixed_state_dict = model._fix_pytorch_state_dict_keys(original_state_dict, model.config)
 
             # Then add "model." prefix for all keys that don't already have it
@@ -1009,8 +1014,6 @@ class PI05Policy(PreTrainedPolicy):
                     new_key = f"model.{key}"
                     remapped_state_dict[new_key] = value
                     remap_count += 1
-                    if remap_count <= 10:  # Only print first 10 to avoid spam
-                        print(f"Remapped: {key} -> {new_key}")
                 else:
                     remapped_state_dict[key] = value
 
@@ -1044,7 +1047,7 @@ class PI05Policy(PreTrainedPolicy):
                 print("All keys loaded successfully!")
 
         except Exception as e:
-            print(f"Warning: Could not remap state dict keys: {e}")
+            print(f"Warning: Could not load state dict: {e}")
 
         return model
 
@@ -1098,6 +1101,14 @@ class PI05Policy(PreTrainedPolicy):
                 # Some checkpoints might have this, but current model expects different structure
                 logging.warning(f"Vision embedding key might need handling: {key}")
 
+            if (
+                key == "model.paligemma_with_expert.paligemma.lm_head.weight"
+                or key == "paligemma_with_expert.paligemma.lm_head.weight"
+            ):
+                fixed_state_dict[
+                    "model.paligemma_with_expert.paligemma.model.language_model.embed_tokens.weight"
+                ] = value.clone()
+
             fixed_state_dict[new_key] = value
 
         return fixed_state_dict

src/lerobot/policies/pi05/processor_pi05.py

@@ -23,7 +23,6 @@ import torch
 
 from lerobot.configs.types import PipelineFeatureType, PolicyFeature
 from lerobot.policies.pi05.configuration_pi05 import PI05Config
-from lerobot.policies.pi05.modeling_pi05 import pad_vector
 from lerobot.processor import (
     AddBatchDimensionProcessorStep,
     DeviceProcessorStep,
@@ -37,7 +36,7 @@ from lerobot.processor import (
     UnnormalizerProcessorStep,
 )
 from lerobot.processor.converters import policy_action_to_transition, transition_to_policy_action
-from lerobot.processor.core import EnvTransition, TransitionKey
+from lerobot.types import EnvTransition, TransitionKey
 from lerobot.utils.constants import (
     OBS_STATE,
     POLICY_POSTPROCESSOR_DEFAULT_NAME,
@@ -68,9 +67,6 @@ class Pi05PrepareStateTokenizerProcessorStep(ProcessorStep):
         # TODO: check if this necessary
         state = deepcopy(state)
 
-        # Prepare state (pad to max_state_dim)
-        state = pad_vector(state, self.max_state_dim)
-
         # State should already be normalized to [-1, 1] by the NormalizerProcessorStep that runs before this step
         # Discretize into 256 bins (see openpi `PaligemmaTokenizer.tokenize()`)
         state_np = state.cpu().numpy()

src/lerobot/policies/pi0_fast/configuration_pi0_fast.py

@@ -54,7 +54,7 @@ class PI0FastConfig(PreTrainedConfig):
 
     tokenizer_max_length: int = 200  # see openpi `__post_init__`
     text_tokenizer_name: str = "google/paligemma-3b-pt-224"
-    action_tokenizer_name: str = "physical-intelligence/fast"
+    action_tokenizer_name: str = "lerobot/fast-action-tokenizer"
     temperature: float = 0.0
     max_decoding_steps: int = 256
     fast_skip_tokens: int = 128

src/lerobot/policies/pi0_fast/modeling_pi0_fast.py

@@ -19,13 +19,12 @@ import logging
 import math
 from collections import deque
 from pathlib import Path
-from typing import TYPE_CHECKING, Literal, TypedDict
+from typing import TYPE_CHECKING, Literal, TypedDict, Unpack
 
 import numpy as np
 import torch
 import torch.nn.functional as F  # noqa: N812
 from torch import Tensor, nn
-from typing_extensions import Unpack
 
 from lerobot.utils.import_utils import _scipy_available, _transformers_available
 
@@ -38,11 +37,16 @@ else:
 if TYPE_CHECKING or _transformers_available:
     from transformers import AutoTokenizer
     from transformers.models.auto import CONFIG_MAPPING
-    from transformers.models.paligemma.modeling_paligemma import PaliGemmaForConditionalGeneration
+
+    from lerobot.policies.pi_gemma import (
+        PaliGemmaForConditionalGenerationWithPiGemma,
+        PiGemmaModel,
+    )
 else:
     CONFIG_MAPPING = None
-    PaliGemmaForConditionalGeneration = None
     AutoTokenizer = None
+    PiGemmaModel = None
+    PaliGemmaForConditionalGenerationWithPiGemma = None
 
 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.policies.pi0_fast.configuration_pi0_fast import PI0FastConfig
@@ -121,7 +125,7 @@ def resize_with_pad_torch(  # see openpi `resize_with_pad_torch` (exact copy)
     if images.dtype == torch.uint8:
         resized_images = torch.round(resized_images).clamp(0, 255).to(torch.uint8)
     elif images.dtype == torch.float32:
-        resized_images = resized_images.clamp(-1.0, 1.0)
+        resized_images = resized_images.clamp(0.0, 1.0)
     else:
         raise ValueError(f"Unsupported image dtype: {images.dtype}")
 
@@ -132,7 +136,7 @@ def resize_with_pad_torch(  # see openpi `resize_with_pad_torch` (exact copy)
     pad_w1 = pad_w0 + remainder_w
 
     # Pad
-    constant_value = 0 if images.dtype == torch.uint8 else -1.0
+    constant_value = 0 if images.dtype == torch.uint8 else 0.0
     padded_images = F.pad(
         resized_images,
         (pad_w0, pad_w1, pad_h0, pad_h1),  # left, right, top, bottom
@@ -206,16 +210,22 @@ class PI0FastPaliGemma(nn.Module):
         vlm_config_hf.text_config.num_hidden_layers = vlm_config.depth
         vlm_config_hf.text_config.num_key_value_heads = vlm_config.num_kv_heads
         vlm_config_hf.text_config.hidden_activation = "gelu_pytorch_tanh"
-        vlm_config_hf.text_config.torch_dtype = "float32"
+        vlm_config_hf.text_config.dtype = "float32"
         vlm_config_hf.text_config.vocab_size = 257152
         vlm_config_hf.text_config.use_adarms = use_adarms[0]
         vlm_config_hf.text_config.adarms_cond_dim = vlm_config.width if use_adarms[0] else None
         vlm_config_hf.vision_config.intermediate_size = 4304
         vlm_config_hf.vision_config.projection_dim = 2048
         vlm_config_hf.vision_config.projector_hidden_act = "gelu_fast"
-        vlm_config_hf.vision_config.torch_dtype = "float32"
+        vlm_config_hf.vision_config.dtype = "float32"
 
-        self.paligemma = PaliGemmaForConditionalGeneration(config=vlm_config_hf)
+        self.paligemma = PaliGemmaForConditionalGenerationWithPiGemma(config=vlm_config_hf)
+
+        # Use PI Gemma (AdaRMS) as language model when use_adarms[0] is True so that
+        # forward(..., adarms_cond=...) is supported (same as pi0/pi05).
+        if use_adarms[0]:
+            text_config = self.paligemma.config.text_config
+            self.paligemma.model.language_model = PiGemmaModel(text_config)
 
         self.to_bfloat16_for_selected_params(precision)
 
@@ -228,10 +238,11 @@ class PI0FastPaliGemma(nn.Module):
         else:
             raise ValueError(f"Invalid precision: {precision}")
 
+        # Keep full vision path in float32 so we never toggle (toggle causes optimizer
+        # "same dtype" error). Align with PI05.
         params_to_keep_float32 = [
-            "vision_tower.vision_model.embeddings.patch_embedding.weight",
-            "vision_tower.vision_model.embeddings.patch_embedding.bias",
-            "vision_tower.vision_model.embeddings.position_embedding.weight",
+            "vision_tower",
+            "multi_modal_projector",
             "input_layernorm",
             "post_attention_layernorm",
             "model.norm",
@@ -242,10 +253,18 @@ class PI0FastPaliGemma(nn.Module):
                 param.data = param.data.to(dtype=torch.float32)
 
     def embed_image(self, image: torch.Tensor):
-        return self.paligemma.model.get_image_features(image)
+        # Vision tower and multi_modal_projector are kept in float32 (params_to_keep_float32). Align with PI05.
+        out_dtype = image.dtype
+        if image.dtype != torch.float32:
+            image = image.to(torch.float32)
+        image_outputs = self.paligemma.model.get_image_features(image)
+        features = image_outputs.pooler_output * self.paligemma.config.text_config.hidden_size**0.5
+        if features.dtype != out_dtype:
+            features = features.to(out_dtype)
+        return features
 
     def embed_language_tokens(self, tokens: torch.Tensor):
-        return self.paligemma.language_model.embed_tokens(tokens)
+        return self.paligemma.model.language_model.embed_tokens(tokens)
 
     def forward(
         self,
@@ -259,7 +278,7 @@ class PI0FastPaliGemma(nn.Module):
         if adarms_cond is None:
             adarms_cond = [None, None]
         if inputs_embeds[1] is None:
-            prefix_output = self.paligemma.language_model.forward(
+            prefix_output = self.paligemma.model.language_model.forward(
                 inputs_embeds=inputs_embeds[0],
                 attention_mask=attention_mask,
                 position_ids=position_ids,
@@ -306,24 +325,14 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`
             self.sample_actions_fast = torch.compile(self.sample_actions_fast, mode=config.compile_mode)
             self.forward = torch.compile(self.forward, mode=config.compile_mode)
 
-        msg = """An incorrect transformer version is used, please create an issue on https://github.com/huggingface/lerobot/issues"""
-
-        try:
-            from transformers.models.siglip import check
-
-            if not check.check_whether_transformers_replace_is_installed_correctly():
-                raise ValueError(msg)
-        except ImportError:
-            raise ValueError(msg) from None
-
     def gradient_checkpointing_enable(self):
         """Enable gradient checkpointing for memory optimization."""
         self.gradient_checkpointing_enabled = True
         # Call the proper gradient_checkpointing_enable() method with use_reentrant=False for better memory efficiency
-        self.paligemma_with_expert.paligemma.language_model.gradient_checkpointing_enable(
+        self.paligemma_with_expert.paligemma.model.language_model.gradient_checkpointing_enable(
             gradient_checkpointing_kwargs={"use_reentrant": False}
         )
-        self.paligemma_with_expert.paligemma.vision_tower.gradient_checkpointing_enable(
+        self.paligemma_with_expert.paligemma.model.vision_tower.gradient_checkpointing_enable(
             gradient_checkpointing_kwargs={"use_reentrant": False}
         )
         logging.info("Enabled gradient checkpointing for PI0FastPytorch model")
@@ -332,8 +341,8 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`
         """Disable gradient checkpointing."""
         self.gradient_checkpointing_enabled = False
         # Call the proper gradient_checkpointing_disable() method
-        self.paligemma_with_expert.paligemma.language_model.gradient_checkpointing_disable()
-        self.paligemma_with_expert.paligemma.vision_tower.gradient_checkpointing_disable()
+        self.paligemma_with_expert.paligemma.model.language_model.gradient_checkpointing_disable()
+        self.paligemma_with_expert.paligemma.model.vision_tower.gradient_checkpointing_disable()
         logging.info("Disabled gradient checkpointing for PI0FastPytorch model")
 
     def _apply_checkpoint(self, func, *args, **kwargs):
@@ -523,7 +532,7 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`
 
         # Convert embeddings to bfloat16 if needed
         if (
-            self.paligemma_with_expert.paligemma.language_model.layers[0].self_attn.q_proj.weight.dtype
+            self.paligemma_with_expert.paligemma.model.language_model.layers[0].self_attn.q_proj.weight.dtype
             == torch.bfloat16
         ):
             prefix_embs = prefix_embs.to(dtype=torch.bfloat16)
@@ -616,7 +625,7 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`
         )
 
         if (
-            self.paligemma_with_expert.paligemma.language_model.layers[0].self_attn.q_proj.weight.dtype
+            self.paligemma_with_expert.paligemma.model.language_model.layers[0].self_attn.q_proj.weight.dtype
             == torch.bfloat16
         ):
             prefix_embs = prefix_embs.to(dtype=torch.bfloat16)
@@ -714,7 +723,7 @@ class PI0FastPytorch(nn.Module):  # see openpi `PI0Pytorch`
 
         # Ensure correct precision (bfloat16/float32)
         if (
-            self.paligemma_with_expert.paligemma.language_model.layers[0].self_attn.q_proj.weight.dtype
+            self.paligemma_with_expert.paligemma.model.language_model.layers[0].self_attn.q_proj.weight.dtype
             == torch.bfloat16
         ):
             prefix_embs = prefix_embs.to(dtype=torch.bfloat16)
@@ -897,14 +906,12 @@ class PI0FastPolicy(PreTrainedPolicy):
         # Check if dataset_stats were provided in kwargs
         model = cls(config, **kwargs)
 
-        # Now manually load and remap the state dict
+        # Load state dict (expects keys with "model." prefix)
         try:
-            # Try to load the pytorch_model.bin or model.safetensors file
             print(f"Loading model from: {pretrained_name_or_path}")
             try:
                 from transformers.utils import cached_file
 
-                # Try safetensors first
                 resolved_file = cached_file(
                     pretrained_name_or_path,
                     "model.safetensors",
@@ -912,7 +919,7 @@ class PI0FastPolicy(PreTrainedPolicy):
                     force_download=kwargs.get("force_download", False),
                     resume_download=kwargs.get("resume_download"),
                     proxies=kwargs.get("proxies"),
-                    use_auth_token=kwargs.get("use_auth_token"),
+                    token=kwargs.get("token"),
                     revision=kwargs.get("revision"),
                     local_files_only=kwargs.get("local_files_only", False),
                 )
@@ -925,8 +932,9 @@ class PI0FastPolicy(PreTrainedPolicy):
                 print("Returning model without loading pretrained weights")
                 return model
 
-            # First, fix any key differences # see openpi `model.py, _fix_pytorch_state_dict_keys`
+            # First, fix any key differences (see openpi model.py, _fix_pytorch_state_dict_keys)
             fixed_state_dict = model._fix_pytorch_state_dict_keys(original_state_dict, model.config)
+
             # Then add "model." prefix for all keys that don't already have it
             remapped_state_dict = {}
             remap_count = 0
@@ -936,8 +944,6 @@ class PI0FastPolicy(PreTrainedPolicy):
                     new_key = f"model.{key}"
                     remapped_state_dict[new_key] = value
                     remap_count += 1
-                    if remap_count <= 10:  # Only print first 10 to avoid spam
-                        print(f"Remapped: {key} -> {new_key}")
                 else:
                     remapped_state_dict[key] = value
 
@@ -971,7 +977,7 @@ class PI0FastPolicy(PreTrainedPolicy):
                 print("All keys loaded successfully!")
 
         except Exception as e:
-            print(f"Warning: Could not remap state dict keys: {e}")
+            print(f"Warning: Could not load state dict: {e}")
 
         return model
 

src/lerobot/policies/pi0_fast/processor_pi0_fast.py

@@ -23,7 +23,6 @@ import torch
 
 from lerobot.configs.types import PipelineFeatureType, PolicyFeature
 from lerobot.policies.pi0_fast.configuration_pi0_fast import PI0FastConfig
-from lerobot.policies.pi0_fast.modeling_pi0_fast import pad_vector
 from lerobot.processor import (
     ActionTokenizerProcessorStep,
     AddBatchDimensionProcessorStep,
@@ -38,7 +37,7 @@ from lerobot.processor import (
     UnnormalizerProcessorStep,
 )
 from lerobot.processor.converters import policy_action_to_transition, transition_to_policy_action
-from lerobot.processor.core import EnvTransition, TransitionKey
+from lerobot.types import EnvTransition, TransitionKey
 from lerobot.utils.constants import (
     OBS_STATE,
     POLICY_POSTPROCESSOR_DEFAULT_NAME,
@@ -69,9 +68,6 @@ class Pi0FastPrepareStateAndLanguageTokenizerProcessorStep(ProcessorStep):
         # TODO: check if this necessary
         state = deepcopy(state)
 
-        # Prepare state (pad to max_state_dim)
-        state = pad_vector(state, self.max_state_dim)
-
         # State should already be normalized to [-1, 1] by the NormalizerProcessorStep that runs before this step
         # Discretize into 256 bins (see openpi `PaligemmaTokenizer.tokenize()`)
         state_np = state.cpu().numpy()

src/lerobot/policies/pi_gemma.py

@@ -0,0 +1,363 @@
+# Copyright 2025 Physical Intelligence and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+import torch
+from torch import nn
+
+from lerobot.utils.import_utils import _transformers_available
+
+if TYPE_CHECKING or _transformers_available:
+    from transformers.cache_utils import DynamicCache
+    from transformers.masking_utils import create_causal_mask
+    from transformers.modeling_layers import GradientCheckpointingLayer
+    from transformers.modeling_outputs import BaseModelOutputWithPast
+    from transformers.models.gemma.modeling_gemma import (
+        GemmaAttention,
+        GemmaConfig,
+        GemmaForCausalLM,
+        GemmaMLP,
+        GemmaModel,
+    )
+    from transformers.models.paligemma.modeling_paligemma import (
+        PaliGemmaForConditionalGeneration,
+        PaliGemmaModel,
+    )
+else:
+    GemmaAttention = None
+    GemmaConfig = None
+    GemmaForCausalLM = None
+    GemmaMLP = None
+    GemmaModel = None
+    PaliGemmaModel = None
+    PaliGemmaForConditionalGeneration = None
+    DynamicCache = None
+    GradientCheckpointingLayer = None
+    BaseModelOutputWithPast = None
+    create_causal_mask = None
+
+
+def _gated_residual(
+    x: torch.Tensor | None,
+    y: torch.Tensor | None,
+    gate: torch.Tensor | None,
+) -> torch.Tensor | None:
+    """Gated residual: x + y when gate is None, else x + y * gate."""
+    if x is None and y is None:
+        return None
+    if x is None or y is None:
+        return x if x is not None else y
+    if gate is None:
+        return x + y
+    return x + y * gate
+
+
+def layernorm_forward(
+    layernorm: nn.Module,
+    x: torch.Tensor,
+    cond: torch.Tensor | None = None,
+):
+    """
+    call layernorm and return hidden states and gate
+    if cond is not None, use conditional norm
+    otherwise, use normal gemma norm
+    """
+    if cond is not None:
+        return layernorm(x, cond=cond)
+    else:
+        return layernorm(x)
+
+
+class PiGemmaRMSNorm(nn.Module):
+    """
+    Adaptive RMSNorm for PI Gemma (AdaRMS).
+    When cond_dim is set, uses cond to modulate scale/shift/gate; otherwise behaves like standard GemmaRMSNorm.
+    forward(x, cond=None) returns (output, gate) for use with _gated_residual.
+    """
+
+    def __init__(self, dim: int, eps: float = 1e-6, cond_dim: int | None = None):
+        super().__init__()
+        self.eps = eps
+        self.dim = dim
+        self.cond_dim = cond_dim
+        if cond_dim is not None:
+            self.dense = nn.Linear(cond_dim, dim * 3, bias=True)
+            nn.init.zeros_(self.dense.weight)
+        else:
+            self.weight = nn.Parameter(torch.zeros(dim))
+            self.dense = None
+
+    def _norm(self, x):
+        # Compute variance in float32 (like the source implementation)
+        var = torch.mean(torch.square(x.float()), dim=-1, keepdim=True)
+        # Compute normalization in float32
+        normed_inputs = x * torch.rsqrt(var + self.eps)
+        return normed_inputs
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        cond: torch.Tensor | None = None,
+    ) -> tuple[torch.Tensor, torch.Tensor | None]:
+        dtype = x.dtype
+        normed = self._norm(x)
+        if cond is None or self.dense is None:
+            normed = normed * (1.0 + self.weight.float())
+            return normed.type_as(x), None
+        if cond.shape[-1] != self.cond_dim:
+            raise ValueError(f"Expected cond dim {self.cond_dim}, got {cond.shape[-1]}")
+        modulation = self.dense(cond)
+        if len(x.shape) == 3:
+            modulation = modulation.unsqueeze(1)
+        scale, shift, gate = modulation.chunk(3, dim=-1)
+        normed = normed * (1 + scale.float()) + shift.float()
+        return normed.to(dtype), gate.to(dtype)
+
+    def extra_repr(self) -> str:
+        if self.dense is not None:
+            return f"dim={self.dim}, eps={self.eps}, adaptive=True, cond_dim={self.cond_dim}"
+        return f"dim={self.dim}, eps={self.eps}"
+
+
+def _get_pi_gemma_decoder_layer_base():
+    """base for PiGemmaDecoderLayer"""
+
+    class _PiGemmaDecoderLayerBase(GradientCheckpointingLayer):
+        """Decoder layer that uses PiGemmaRMSNorm and _gated_residual, compatible with v5 Gemma."""
+
+        def __init__(self, config: GemmaConfig, layer_idx: int):
+            super().__init__()
+            self.hidden_size = config.hidden_size
+            self.self_attn = GemmaAttention(config=config, layer_idx=layer_idx)
+            self.mlp = GemmaMLP(config)
+            cond_dim = (
+                getattr(config, "adarms_cond_dim", None) if getattr(config, "use_adarms", False) else None
+            )
+            self.input_layernorm = PiGemmaRMSNorm(
+                config.hidden_size, eps=config.rms_norm_eps, cond_dim=cond_dim
+            )
+            self.post_attention_layernorm = PiGemmaRMSNorm(
+                config.hidden_size, eps=config.rms_norm_eps, cond_dim=cond_dim
+            )
+
+        def forward(
+            self,
+            hidden_states: torch.Tensor,
+            attention_mask: torch.Tensor | None = None,
+            position_ids: torch.LongTensor | None = None,
+            past_key_values=None,
+            use_cache: bool = False,
+            cache_position: torch.LongTensor | None = None,
+            position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
+            adarms_cond: torch.Tensor | None = None,
+            **kwargs,
+        ) -> torch.Tensor:
+            residual = hidden_states
+            hidden_states, gate = self.input_layernorm(hidden_states, cond=adarms_cond)
+            hidden_states, _ = self.self_attn(
+                hidden_states,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_values=past_key_values,
+                use_cache=use_cache,
+                cache_position=cache_position,
+                position_embeddings=position_embeddings,
+                **kwargs,
+            )
+
+            hidden_states = _gated_residual(residual, hidden_states, gate)
+
+            residual = hidden_states
+            hidden_states, gate = self.post_attention_layernorm(hidden_states, cond=adarms_cond)
+            hidden_states = self.mlp(hidden_states)
+            hidden_states = _gated_residual(residual, hidden_states, gate)
+            return hidden_states
+
+    return _PiGemmaDecoderLayerBase
+
+
+class PiGemmaModel(GemmaModel):  # type: ignore[misc]
+    """
+    GemmaModel extended with AdaRMS (adaptive RMSNorm) and gated residuals when config.use_adarms is True.
+    """
+
+    def __init__(self, config: GemmaConfig, **kwargs):
+        super().__init__(config, **kwargs)
+        # if not getattr(config, "use_adarms", False):
+        #     return
+        cond_dim = getattr(config, "adarms_cond_dim", None)
+        pi_gemma_decoder_layer_base = _get_pi_gemma_decoder_layer_base()
+        self.layers = nn.ModuleList(
+            [pi_gemma_decoder_layer_base(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.norm = PiGemmaRMSNorm(config.hidden_size, eps=config.rms_norm_eps, cond_dim=cond_dim)
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: DynamicCache | None = None,
+        inputs_embeds: torch.FloatTensor | None = None,
+        use_cache: bool | None = None,
+        output_attentions: bool | None = None,
+        output_hidden_states: bool | None = None,
+        cache_position: torch.LongTensor | None = None,
+        adarms_cond: torch.Tensor | None = None,
+        **kwargs,
+    ) -> BaseModelOutputWithPast:
+        """
+        adarms_cond (`torch.Tensor` of shape `(batch_size, cond_dim)`, *optional*):
+            Condition for ADARMS.
+        """
+        output_attentions = (
+            output_attentions if output_attentions is not None else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if self.gradient_checkpointing and self.training and use_cache:
+            import logging
+
+            logging.warning(
+                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
+            )
+            use_cache = False
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache()
+
+        if cache_position is None:
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            cache_position = torch.arange(
+                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
+            )
+
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        causal_mask = create_causal_mask(
+            config=self.config,
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            cache_position=cache_position,
+            past_key_values=past_key_values,
+            position_ids=position_ids,
+        )
+
+        # embed positions
+        hidden_states = inputs_embeds
+        # Convert to bfloat16 if the first layer uses bfloat16
+        if len(self.layers) > 0 and self.layers[0].self_attn.q_proj.weight.dtype == torch.bfloat16:
+            hidden_states = hidden_states.to(torch.bfloat16)
+
+        # create position embeddings to be shared across the decoder layers
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+
+        # normalized
+        # Gemma downcasts the below to float16, causing sqrt(3072)=55.4256 to become 55.5
+        # See https://github.com/huggingface/transformers/pull/29402
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+
+        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            layer_outputs = decoder_layer(
+                hidden_states,
+                attention_mask=causal_mask,
+                position_ids=position_ids,
+                past_key_values=past_key_values,
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+                cache_position=cache_position,
+                position_embeddings=position_embeddings,
+                adarms_cond=adarms_cond,
+                **kwargs,
+            )
+
+            hidden_states = layer_outputs
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+        hidden_states, _ = self.norm(hidden_states, adarms_cond)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values if use_cache else None,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+
+
+class PiGemmaForCausalLM(GemmaForCausalLM):  # type: ignore[misc]
+    """
+    Causal LM wrapper using PiGemmaModel as the backbone, for consistency with GemmaForCausalLM
+    and the language model used in pi0_fast. Use this for the action expert in pi0/pi05.
+    """
+
+    def __init__(self, config: GemmaConfig, **kwargs):
+        super().__init__(config, **kwargs)
+        self.model = PiGemmaModel(config)
+
+
+class PaliGemmaModelWithPiGemma(PaliGemmaModel):
+    """PaliGemmaModel whose language_model is PiGemmaModel (custom decoder with PiGemmaRMSNorm and gated residuals)."""
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.language_model = PiGemmaModel(config.text_config)
+
+
+class PaliGemmaForConditionalGenerationWithPiGemma(PaliGemmaForConditionalGeneration):
+    """PaliGemmaForConditionalGeneration using PiGemma decoder for the language model."""
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = PaliGemmaModelWithPiGemma(config)
+
+    # Make modules available through conditional class for BC
+    @property
+    def language_model(self):
+        return self.model.language_model
+
+
+__all__ = [
+    "PiGemmaModel",
+    "PiGemmaForCausalLM",
+    "PiGemmaRMSNorm",
+    "_gated_residual",
+    "layernorm_forward",
+    "PaliGemmaModelWithPiGemma",
+    "PaliGemmaForConditionalGenerationWithPiGemma",
+]

src/lerobot/policies/pretrained.py

@@ -19,7 +19,7 @@ import os
 from importlib.resources import files
 from pathlib import Path
 from tempfile import TemporaryDirectory
-from typing import TypedDict, TypeVar
+from typing import TypedDict, TypeVar, Unpack
 
 import packaging
 import safetensors
@@ -28,7 +28,6 @@ from huggingface_hub.constants import SAFETENSORS_SINGLE_FILE
 from huggingface_hub.errors import HfHubHTTPError
 from safetensors.torch import load_model as load_model_as_safetensor, save_model as save_model_as_safetensor
 from torch import Tensor, nn
-from typing_extensions import Unpack
 
 from lerobot.configs.policies import PreTrainedConfig
 from lerobot.configs.train import TrainPipelineConfig

src/lerobot/policies/sac/reward_model/configuration_classifier.py

@@ -33,7 +33,7 @@ class RewardClassifierConfig(PreTrainedConfig):
     latent_dim: int = 256
     image_embedding_pooling_dim: int = 8
     dropout_rate: float = 0.1
-    model_name: str = "helper2424/resnet10"
+    model_name: str = "helper2424/resnet10"  # TODO: This needs to be updated. The model on the Hub doesn't call self.post_init() in its __init__, which is required by transformers v5 to set all_tied_weights_keys. The from_pretrained call fails when it tries to access this attribute during _finalize_model_loading.
     device: str = "cpu"
     model_type: str = "cnn"  # "transformer" or "cnn"
     num_cameras: int = 2