Overview

LeRobot is a unified, end-to-end stack for robot learning developed by Hugging Face. It integrates across the entire robotics pipeline — from low-level motor control to large-scale dataset management and policy deployment.

Core Architecture

Unified Robot Integration

Python-based middleware API for real-world motor control across diverse platforms. Supports:

  • SO-100/101 arms (~€225), ALOHA-2 bimanual manipulator, Koch-v1.1
  • Mobile platforms: Stretch-3, LeKiwi

Standardized Datasets (LeRobotDataset)

  • Multimodal format for recording, storing, and streaming high frame-rate sensory and image data
  • As of September 2025: 16,000+ datasets from 2,200+ contributors, 3.9+ million episodes
  • StreamingLeRobotDataset: fetches frames on-demand, bounded memory usage

Optimized Inference Stack

  • Decouples action prediction from action execution
  • Policies can run remotely while maintaining real-time control via asynchronous action streaming

Algorithm Implementations (PyTorch)

  • Reinforcement learning: SERL, TD-MPC
  • Behavioral cloning: ACT (52M params, ~5ms inference), Diffusion Policy
  • Foundation models: π₀ (3.5B params), SmolVLA (450M params)

Impact

Reduces barriers by unifying hardware interfaces, data management, and algorithm implementations — addressing the fragmentation that historically slowed robot learning development.