huggingface/lerobot

What it solves

LeRobot addresses the high barrier to entry in real-world robotics by providing a standardized, open-source ecosystem. It solves the problem of data fragmentation and hardware incompatibility by offering a unified interface for controlling diverse robots and a scalable dataset format for sharing and training models.

How it works

LeRobot operates as a PyTorch-native library that decouples control logic from hardware specifics through a unified Robot class. It uses the LeRobotDataset format (combining Parquet files for state/action data and MP4/images for vision) to enable efficient streaming and visualization via the Hugging Face Hub. The library includes a suite of state-of-the-art policies for imitation learning, reinforcement learning, and Vision-Language-Action (VLA) models, which can be trained and evaluated in simulation or on real hardware.

Who it’s for

It is designed for robotics researchers, developers, and hobbyists who want to build, train, and deploy AI-driven robot policies without needing to write hardware-specific code from scratch.

Highlights

• Hardware Agnostic: Supports a wide range of hardware from low-cost arms like SO-100 to humanoids and third-party plugins.
• Standardized Datasets: Integration with Hugging Face Hub for scalable storage and streaming of robotic data.
• Diverse Model Library: Implements a wide array of policies including ACT, Diffusion, Pi0, and GR00T.
• Unified Evaluation: Built-in tools to evaluate policies using standard benchmarks like LIBERO and MetaWorld.