LeRobot

HuggingFace ML library for real-world robotics in PyTorch

License

Apache-2.0

Overview

HuggingFace's open-source machine learning library for real-world robotics in PyTorch. Provides models, datasets, and tools to democratize robotics development with emphasis on imitation learning and reinforcement learning. Features pre-built affordable robot platforms (SO-101 at €114, HopeJR humanoid arm, LeKiwi mobile robot), pretrained models on HuggingFace Hub, human-collected demonstration datasets, and simulation environments (ALOHA, PushT, SimXArm). Enables training robots in minutes with demonstration data and supports real-world transfer.

LeRobot is Hugging Face open-source platform for real-world robotics, focused on making state-of-the-art AI and machine learning accessible to the robotics community. Building on Hugging Face expertise in democratizing AI, LeRobot provides pre-trained models, datasets, and tools specifically designed for robotic manipulation and control tasks. The platform emphasizes ease of use, reproducibility, and community collaboration, enabling researchers and developers to quickly deploy advanced AI capabilities on physical robots. LeRobot bridges the gap between cutting-edge AI research and practical robotic applications.

Why it matters

LeRobot is doing for robot learning what Hugging Face Transformers did for NLP — providing a unified, accessible framework that lowers the barrier to entry. With pre-built affordable hardware, shared datasets, and pre-trained models on the Hub, it is rapidly becoming the standard starting point for robot learning research.

Problem it solves

Robot learning research is fragmented across dozens of incompatible codebases, making it impossible to reproduce results or compare methods. LeRobot standardizes the training and evaluation pipeline so researchers can focus on algorithms rather than infrastructure.

Key takeaways

• Unified API for training robot policies across algorithms (ACT, Diffusion Policy, TDMPC)
• Pre-trained models and datasets shared through HuggingFace Hub
• Affordable hardware platforms: SO-101 (114 EUR), HOPEJr, LeKiwi
• Works with both simulated environments and real robots
• Rapidly growing community contributing models, datasets, and hardware integrations

Who it is for

• Robot learning researchers at all levels
• ML engineers building manipulation policies
• Hobbyists with low-cost arms (SO-101, Koch)
• Companies evaluating imitation learning approaches

Limitations

• Focus is on imitation learning — less support for RL-based approaches
• Real robot support depends on available hardware interfaces
• Rapidly evolving API means documentation can lag behind releases
• Compute requirements for training can be significant (GPU recommended)