Skip to main content

Listed

, home
FREN

LeRobot

Definition

A Python library from Hugging Face devoted to machine learning for real robots. It provides a common interface to control robots, from small low-cost arms to humanoids, along with tools to record, store and share demonstration datasets. Its documentation details imitation learning from teleoperated demonstrations.

Checked on against the references below. · Written with AI · checked against the sources Source: Hugging Face, LeRobot documentation.

Concrete examples

Hello Robot describes its stretch4_lerobot repository as a LeRobot integration for Stretch 4: record teleoperated demonstrations, train policies and run them on the robot, using stock LeRobot commands. LeRobot’s documentation states that Unitree G1 is supported, in both 29 and 23 degree-of-freedom versions. It also says Reachy 2 is fully supported by its recording features.

What it changes for you

To integrate your own robot, the documentation requires a way to read sensor data and send motor commands programmatically, for example through the manufacturer’s SDK. Support also depends on the variant: on the 23-DoF G1, the SONIC controller is not supported and MuJoCo simulation is not supported yet (a native model is planned). For Reachy 2, VR teleoperation goes through Pollen’s application, which is separate from LeRobot: the application drives the robot while LeRobot handles recording; LeRobot also supports robot-to-robot teleoperation, with one Reachy 2 controlling another.

Not to be confused with

LeRobot is not a policy but the library that implements several, from the lightweight imitation-learning model ACT to the vision-language-action models π₀ and SmolVLA, all trainable and deployable with the same commands.

Additional material checked on .

Related terms

References

Back to the glossary