LeRobot: A Library for Real-World Robot Learning

LeRobot is Hugging Face's answer to the fragmentation in robot learning codebases. Instead of every lab maintaining its own implementation of Diffusion Policy, ACT, or Octo, LeRobot provides a single, well-documented PyTorch library with pretrained weights, standard datasets (ALOHA, Push-T, etc.), and training scripts that work out-of-the-box on real hardware. The library is built around four pillars: datasets (the `LeRobotDataset` format stores raw sensor data on disk and only loads what's needed), models (pretrained weights for ACT, Diffusion Policy, and Octo on Hugging Face Hub), examples (training scripts for imitation and RL), and simulation (basic MuJoCo and habitat integration). The key design philosophy is composability: you can swap out the backbone, the policy head, or the data augmentation pipeline without rewriting the training loop. LeRobot also introduces `OpenPi0`, a community replication of Physical Intelligence's π₀ VLA model, available entirely in PyTorch. This makes it possible to run a state-of-the-art policy on a consumer GPU without needing proprietary infrastructure. Perhaps the most impactful feature is the `lerobot` CLI: with a one-liner (`pip install lerobot`), researchers can download pretrained models, visualize datasets in 3D, and fine-tune policies on their own robot data. The accompanying dataset viewer and hardware integration guides make it one of the most accessible entry points for newcomers to robot learning.