عن المشروع

LeRobot is a comprehensive open-source library designed to lower the barrier to entry for real-world robotics by providing models, datasets, and tools built on PyTorch. It serves as a unified interface for controlling diverse robotic platforms, ranging from low-cost arms like the SO-100 to humanoids such as the Unitree G1 and Reachy2. The library emphasizes hardware agnosticism, allowing developers to decouple control logic from specific hardware details through a standardized `Robot` class interface. A core feature of LeRobot is its approach to solving data fragmentation in robotics through the LeRobotDataset format. This standardized format combines synchronized MP4 videos (or images) for visual data with Parquet files for state and action data. It is deeply integrated with the Hugging Face Hub, enabling efficient storage, streaming, and visualization of massive robotic datasets. Users can easily load datasets, manipulate episodes, and merge data sources using simple Python APIs. LeRobot implements a wide array of state-of-the-art policies covering Imitation Learning, Reinforcement Learning, and Vision-Language-Action (VLA) models. Notable supported models include ACT, Diffusion Policy, Pi0, GR00T N1.5, and SmolVLA. The library also supports emerging world models like VLA-JEPA and reward models such as SARM and Robometer. Training these policies is streamlined through script configurations, allowing users to launch training jobs with simple command-line arguments specifying the policy type and dataset repository. Inference and evaluation are supported through unified scripts that work with both simulation environments and real hardware. LeRobot integrates with standard benchmarks like LIBERO and MetaWorld, and provides tools for users to implement and distribute their own simulation environments via EnvHub. The project fosters an open-source ecosystem with extensive documentation, community support via Discord, and resources for custom hardware and policy integration.