LeRobot: An Open-Source Library for End-to-End Robot Learning

1 min read

LeRobot is Hugging Face’s open-source library covering the whole robot learning stack: low-level motor control middleware, dataset collection, storage and streaming, implementations of learning algorithms from several families, and an asynchronous inference stack. It targets real robots on low-cost hardware and is built to take new embodiments. The authors favour learned methods that improve with data and compute over hand-built pipelines.

Findings

  • The abstract reports no quantitative results; it describes the library.
  • The public repository ships policies including ACT, Diffusion Policy, π0 and SmolVLA, plus wrappers for benchmarks such as LIBERO, Meta-World and RoboTwin.

Limits

  • Abstract and repository README only.

Relevance

A shared dataset format and policy zoo make it cheap to compare learned policies across arms, the precondition for any fair robustness benchmark. Open question for robust robotics: none of the bundled benchmarks inject sensor faults or adversarial inputs, so who adds that layer? See foundation-models-for-robotics.