What is lerobot?

LeRobot is a Python library from Hugging Face aimed at making real world robotics research and building easier to get into, by providing shared models, datasets, and tools built on PyTorch. The core idea is to lower the barrier for people who want to work with robots, whether that means training a robot to perform a task or contributing data and models that others can reuse. The library gives you a single, hardware agnostic way to control very different robots through a common Python interface, ranging from inexpensive robotic arms to more advanced humanoid robots. It supports a long list of hardware out of the box, including SO100, LeKiwi, Koch, HopeJR, several other robot arms, gamepads, keyboards, and phones, and it is built to be extended so you can plug in your own custom robot using the same interface. A major part of the project is the LeRobotDataset format, a standardized way of storing robotics data using video files for what the robot sees and separate data files for its state and actions, all hosted on the Hugging Face Hub so people can browse and stream large robotics datasets without downloading everything up front. The library also includes tools for editing these datasets, such as deleting episodes or merging datasets together. On the modeling side, LeRobot implements a wide range of policies in PyTorch, covering imitation learning, reinforcement learning, vision language action models, world models, and reward models, with training handled through a single command line script once a dataset and policy type are chosen. Installation is done through pip, and after that a single command reports information about your setup. The project is released under the Apache 2.0 license and has an active Discord community for support.