lerobot

LeRobot is an open-source robot AI framework released by Hugging Face, with the goal of making robot AI easier to get started with, adopting an end-to-end learning approach. The project is based on PyTorch, targets real-world robots, and provides models, datasets, and tools, hoping to lower the barrier to entry so that everyone can contribute to and benefit from shared datasets and pretrained models.

In terms of use, LeRobot aims to provide a unified training and deployment framework for real robots, solve the problem of data fragmentation in the robotics field, and promote the democratization of physical AI. Through standardized interfaces and dataset formats, it allows developers to reuse data collection, training, and visualization tools across different hardware and policies.

Main capabilities include:

1. A hardware-agnostic, Python-native unified Robot class interface that decouples control logic from hardware details, standardizes cross-platform control, and covers everything from low-cost robotic arms (SO-100) to humanoid robots. Natively supported hardware includes SO100, LeKiwi, Koch, HopeJR, OMX, EarthRover, Reachy2, gamepads, keyboards, phones, OpenARM, Unitree G1, and reBot B601. The library itself is designed to be extensible; users can implement the Robot interface to use LeRobot's data collection, training, and visualization tools for their own custom robots.
2. A standardized LeRobotDataset format that uses synchronized MP4 videos (or images) to store visual information and Parquet files to store state and action data, hosted on the Hugging Face Hub, supporting efficient storage, streaming, and visualization of massive robot datasets. Accompanying tools support seamlessly deleting episodes, splitting by index or ratio, adding or removing features, and merging multiple datasets.
3. State-of-the-art policies implemented in pure PyTorch, covering imitation learning, reinforcement learning, vision-language-action (VLA) models, world models, and reward models, and existing policies have been proven transferable to the real world and can be directly trained and deployed. Training policies only requires running script configurations; users can also implement their own policies, reuse LeRobot's collection, training, and visualization tools, and share models to the HF Hub.
4. Inference and evaluation: provides a unified evaluation script that can evaluate policies in simulation or on real hardware, supports standard benchmarks such as LIBERO and MetaWorld, and allows implementing and distributing custom simulation environments or benchmarks through EnvHub documentation.
5. Third-party hardware ecosystem: In addition to natively supported hardware, the community maintains a growing plugin ecosystem covering other robots, teleoperation devices, cameras, and sensors, including UFACTORY xArm, Universal Robots UR5e, Franka, AgileX Piper, Trossen WidowX, ARX5, I2RT YAM, GELLO, SpaceMouse, Meta Quest, ROS 2 bridge, tactile and depth cameras, etc. Plugins are automatically discovered by package name: LeRobot will import any installed package prefixed with lerobot_robot_, lerobot_teleoperator_, or lerobot_camera_, and after installation, its registered types can be used directly in the CLI.

In terms of usage, LeRobot can be installed directly from PyPI; see the installation documentation for detailed installation instructions. Training a policy only requires running a script configuration. Inference and evaluation use a unified evaluation script. The documentation provides complete tutorials and APIs, and there is also a LeRobot+SO-ARM101 Chinese tutorial (Tongji Zihao Xiong), covering assembly, teleoperation, datasets, training, and deployment, validated by Seed Studio and 5 global hackathon participants. The community also provides LeLab - a web interface for LeRobot that allows teleoperation, calibration, dataset recording, playback, and training of the SO robot arm in the browser without the CLI. In addition, there are resources such as a free hands-on robot learning course and an end-to-end T-shirt folding demonstration. Community communication is available via Discord and X.

Applicable scenarios: data collection, teleoperation, imitation learning and reinforcement learning training for real robots (especially low-cost robot arms and humanoid robots), vision-language-action model deployment, simulation and real-machine benchmark evaluation, and robot learning teaching and research.

In terms of limitations, the materials do not provide specific fee or pricing information (the project is open source, but the materials do not clarify the details of the fee model); GPU/RAM requirements and the expected training time for each policy should refer to the computing hardware guide; some hardware and third-party plugins are maintained by the community and are not officially natively supported. Details such as the specific list of supported policies and the number of datasets are not fully listed in the materials.