LeRobot (Hugging Face)
Hugging Face's open-source robot learning library: shared datasets, ready policies (ACT, Diffusion, SmolVLA, pi0, GR00T) and low-cost robot arms
Overview
LeRobot is Hugging Face's open-source (Apache-2.0) library for real-world robot learning in PyTorch. It gives one standard dataset format (LeRobotDataset: video plus Parquet state and action data, shared on the Hugging Face Hub), training and evaluation tools, and a long list of ready policies: imitation learning (ACT, Diffusion, VQ-BeT), reinforcement learning (HIL-SERL, TD-MPC), and vision-language-action models including SmolVLA, Physical Intelligence's pi0 and pi0.5, and NVIDIA GR00T N1.7. It drives low-cost hardware such as the SO-100 and SO-101 arms, Koch, LeKiwi and HopeJR, as well as Reachy 2 and Unitree G1, and evaluates in simulation with LIBERO and MetaWorld. Launched in May 2024, it reached v0.7.0 in 2026. Also from Hugging Face: the Hugging Face Hub, TRL, Transformers, and Argilla, which joined Hugging Face in 2024.
The Verdict
Who Should Use LeRobot?
Best For
- Hobbyists, students and labs starting robot learning on cheap arms
- Researchers comparing policies on one shared dataset format
- Teams fine-tuning open VLAs (SmolVLA, pi0, GR00T) on their own demos
Not Ideal For
- Photorealistic simulation or large-scale RL (use a full simulator)
- Industrial robots that need ROS 2 integration and safety certification
- Heavy-payload manipulation (SO-101 class arms are light-duty)
What's Great
- Apache-2.0, with community robot datasets shared on the Hugging Face Hub
- One library for many policies, from ACT to open VLAs
- Works with affordable DIY arms such as the SO-101
- Fast release cadence (v0.6.0 in July 2026, v0.7.0 since)
Watch Out For
- Low-cost arms such as the SO-101 carry small payloads (about 500 g reported)
- Built-in simulation covers benchmark environments (LIBERO, MetaWorld), not a photorealistic simulator
- Few independent reviews or comparisons exist yet
Pricing
View all features & details
Policies
- Imitation: ACT, Diffusion, VQ-BeT, Multitask DiT
- RL: HIL-SERL, TD-MPC
- VLAs: SmolVLA, pi0, pi0-FAST, pi0.5, GR00T N1.7
- World models and reward models
Hardware
- SO-100 and SO-101 arms
- Koch, LeKiwi, HopeJR
- Reachy 2, Unitree G1
Data and Simulation
- LeRobotDataset format on the Hugging Face Hub
- LIBERO and MetaWorld evaluation
- EnvHub for custom environments
How It Compares
| Feature | LeRobot | Isaac Sim / Isaac Lab | openpi |
|---|---|---|---|
| What it is | Library, datasets and real-robot tooling | Simulator plus RL framework | VLA models and code |
| License | Apache-2.0 | Apache-2.0 / BSD-3-Clause | Apache-2.0 |
| Real low-cost robots | Yes (SO-100/101, Koch, LeKiwi) | See vendor | ALOHA, DROID examples |
| Simulation | LIBERO, MetaWorld | Full photorealistic sim | LIBERO |
| Price | Free | Free | Free |
Is this your tool? Add a badge