LeRobot, developed by Hugging Face, is a comprehensive Python library designed to democratize access to and accelerate progress in the field of robotics through end-to-end learning. Its primary purpose is to lower the barrier to entry for researchers, developers, and enthusiasts, enabling them to contribute to and benefit from shared datasets, pre-trained models, and standardized tools. The project aims to make AI for robotics more accessible and facilitate real-world applications.

At its core, LeRobot provides a hardware-agnostic interface built on PyTorch, streamlining control across a wide range of robotic platforms. This unified `Robot` class simplifies interaction with diverse hardware, from low-cost arms like the SO-100 to more complex systems like humanoids. This abstraction allows users to focus on the core robotics tasks rather than getting bogged down in hardware-specific details. The library supports a growing list of robots and teleoperation devices, and it is designed to be extensible, allowing users to integrate their own custom robots seamlessly.

A key feature of LeRobot is its standardized data format, the LeRobotDataset. This format addresses the common problem of data fragmentation in robotics by providing a consistent structure for storing and managing robotic datasets. The LeRobotDataset format uses synchronized MP4 videos (or images) for visual data and Parquet files for state and action data. This format is hosted on the Hugging Face Hub, enabling efficient storage, streaming, and visualization of massive robotic datasets. Users can easily explore, download, and utilize thousands of robotics datasets available on the Hub. The library also provides tools for manipulating datasets, such as deleting episodes, splitting datasets, adding/removing features, and merging multiple datasets.

LeRobot also offers a collection of state-of-the-art (SoTA) policies implemented in pure PyTorch. These policies cover various areas, including Imitation Learning, Reinforcement Learning, and Vision-Language-Action (VLA) models. The library provides tools to instrument and inspect the training process. Users can train policies using simple configuration scripts. The library supports a variety of models, including ACT, Diffusion, VQ-BeT for Imitation Learning, HIL-SERL, TDMPC, and QC-FQL (coming soon) for Reinforcement Learning, and Pi0Fast, Pi0.5, GR00T N1.5, SmolVLA, and XVLA for VLAs. The library also allows users to implement their own policies and leverage LeRobot's data collection, training, and visualization tools, and share their models on the Hugging Face Hub.

Furthermore, LeRobot provides tools for inference and evaluation. Users can evaluate their policies in simulation or on real hardware using a unified evaluation script. The library supports standard benchmarks like LIBERO and MetaWorld, with more benchmarks planned for the future. The project also offers comprehensive documentation, tutorials, and a supportive community through Discord and other channels. The project encourages community contributions and provides resources for getting started, including detailed documentation and a contributing guide. The project is actively maintained by the LeRobot team at Hugging Face and is intended to be a collaborative effort to advance the field of robotics.