LeRobot is a Hugging Face robotics library built in PyTorch that aims to democratize AI for robotics by providing models, datasets, and tools for real-world robotic applications. The project seeks to lower barriers to entry so that researchers and developers can contribute to and benefit from shared datasets and pretrained models across the robotics community.
The library provides a hardware-agnostic, Python-native interface that standardizes control across diverse robotic platforms. It supports a wide range of robots including the SO100, LeKiwi, Koch, HopeJR, OMX, EarthRover, Reachy2, Unitree G1, and reBot B601, along with various teleoperation devices like gamepads, keyboards, and phones. The unified Robot class interface decouples control logic from hardware specifics, and the library is designed to be extensible so users can implement the Robot interface for custom robots while leveraging LeRobot's data collection, training, and visualization tools.
LeRobot addresses data fragmentation in robotics through its standardized LeRobotDataset format, which uses synchronized MP4 videos or images for vision data alongside Parquet files for state and action information. These datasets are hosted on the Hugging Face Hub, enabling efficient storage, streaming, and visualization of large-scale robotic datasets. The library includes tools for dataset manipulation such as deleting episodes, splitting by indices or fractions, adding or removing features, and merging multiple datasets.
The repository implements state-of-the-art policies in pure PyTorch across multiple categories. Imitation learning models include ACT, Diffusion, VQ-BeT, and Multitask DiT Policy. Reinforcement learning approaches cover HIL-SERL and TDMPC. Vision-Language-Action models include Pi0, Pi0Fast, Pi0.5, GR00T N1.7, SmolVLA, XVLA, EO-1, MolmoAct2, and WALL-OSS. The library also features world models like VLA-JEPA and reward models including SARM, TOPReward, and Robometer. Training policies is designed to be straightforward through configuration scripts.
According to GitGenius tracking data, the repository has maintained steady activity with 25,501 stargazers as of the most recent check. The project shows active community engagement with a median issue and pull request response latency of 16.8 hours across 1,252 tracked items. The most frequently used issue labels are bug with 410 occurrences, policies with 368, and question with 322. Top contributors include imstevenpmwork with 1,015 tracked events, jadechoghari with 639 events, and AdilZouitine with 449 events. The repository shares overlapping contributors with major projects including opencv/opencv, pytorch/pytorch, and flutter/flutter, indicating cross-pollination with the broader open-source ecosystem.
LeRobot supports evaluation of policies in both simulation and real hardware environments using unified evaluation scripts and standard benchmarks like LIBERO and MetaWorld. The project provides comprehensive documentation, tutorials including Chinese-language guides, community resources through Discord, and practical demonstrations such as a t-shirt folding experiment. An associated web interface called LeLab enables browser-based teleoperation, calibration, dataset recording, and training without requiring command-line interface knowledge.
