LeRobot — Hugging Face's Open-Source Robotics Framework

2026-09-10T00:00:00Z 2026-09-10T00:00:00Z robotics ai huggingface machine-learning imitation-learning reinforcement-learning pytorch

LeRobot — Making AI for Robotics Accessible

LeRobot is Hugging Face's open-source library for real-world robotics, focused on imitation learning and reinforcement learning with PyTorch. It covers the entire robot learning stack: simulation, teleoperation, dataset collection, model training, and real-world deployment.

Why it's interesting

Feature What it means
End-to-end One framework from data collection → training → deployment on real hardware
Pretrained models Download policies from Hugging Face Hub and run them on your robot
Datasets Thousands of robot demonstration datasets on the Hub, standardized format (Parquet + MP4)
Simulation support Train in sim, zero-shot transfer to real hardware
Hardware abstraction Unified Robot class interface — swap robots without changing your code
Plugin ecosystem Third-party robots and teleoperators auto-discovered by package name

Core Components

1. Datasets (LeRobotDataset) - Standardized format: Parquet for features + MP4 for video - Stored on Hugging Face Hub — browse thousands of existing datasets - Tools to delete episodes, split by indices, add/remove features, merge datasets - Efficient streaming from cloud storage

from lerobot.common.datasets import LeRobotDataset
dataset = LeRobotDataset("lerobot/aloha_mobile_cabinet")

2. Policies (State-of-the-art models in pure PyTorch) - Imitation Learning: ACT, Diffusion Policy, pi0 - Reinforcement Learning: PPO, SAC, TD3, HIL-SERL - Vision-Language-Action (VLA): SmolVLA, pi0 VLA - All policies share a unified interface — swap between them with a config change

Training example:

python lerobot/scripts/train.py \
    --dataset.repo_id=lerobot/aloha_mobile_cabinet \
    --policy.type=diffusion

3. Hardware Support (built-in and extensible) | Robot Type | Examples | |------------|----------| | Arms | SO-100, Koch v1.2/v1.5, Aloha, etc. | | Mobile | LEKIWI, LoCoBot, Stretch | | Hands | LEAP Hand, Inspire, etc. | | Teleoperators | Phone-based, SpaceMouse, VR, etc. |

Plugins auto-discover by package name prefix: - lerobot_robot_<name> — robot plugin - lerobot_teleoperator_<name> — teleoperation plugin

4. Simulation Environments - Gymnasium-based API with standard benchmarks - Train in simulation, deploy to real hardware with zero-shot transfer - Common benchmarks: PushT, Aloha, LIBERO

5. Evaluation

python lerobot/scripts/eval.py \
    --policy.path=lerobot/pi0_libero_finetuned

Getting Started

pip install lerobot

Then browse datasets on the Hugging Face Hub: https://huggingface.co/lerobot

Learning resources: - Robot Learning Tutorial — free hands-on course - T-Shirt Folding Experiment — end-to-end demo - LeLab — browser-based web interface (teleoperate, record, train SO arm from browser, no CLI) - Chinese tutorials for SO-ARM101

Why this is fun

  1. Train a robot to fold t-shirts — end-to-end example shipped by HF
  2. Phone teleoperation — use your phone to collect demonstration data for training
  3. Browser-based training — LeLab lets you do everything from a web browser
  4. Zero-shot sim-to-real — train in simulation, deploy to real robot without changes
  5. Plugins extend everything — community builds robot drivers and teleoperation tools as pip-installable packages

You Can Build One — The SO-ARM100 / SO-ARM101

The LeRobot ecosystem is built around the SO-ARM100 and SO-ARM101 — fully open-source, 3D-printable robot arms designed to work with LeRobot out of the box.

What it costs

Option Cost Where
DIY (print + source parts) ~$100 Print the STL files yourself, buy servos + controller
DIY Kit (no print) ~$160 - $200 WowRobo, Partabot, AliExpress
3D printed parts only ~$35 AliExpress, Seeed Studio
Fully assembled ~$250-350 WowRobo, AliExpress

The Build

  • Design: 6-axis robot arm, ~400mm reach, uses standard hobby servos (LD-1501MG or similar)
  • Control: USB/serial to a simple servo controller board
  • Dual-arm teleoperation: Build a "leader" arm (you move it) and a "follower" arm (it copies you) — this is how data collection works
  • Required for training: A pair of arms (leader + follower), or one arm + phone/VR teleoperator
  • Sim-only option: No hardware needed — train and evaluate entirely in simulation

Where to start

  1. SO-ARM100 — the original, ~$100 DIY build
  2. SO-ARM101 — newer dual-arm kit, improved assembly
  3. LeRobot docs: Install → collect dataset → train → deploy
  4. Chinese tutorials with full video build guide for SO-ARM101
  5. YouTube assembly videos — 67-minute SO-ARM101 guide available

Citation

@misc{lerobot,
    author = {LeRobot Contributors},
    title = {{LeRobot}: State-of-the-art Machine Learning for Real-World Robotics in Pytorch},
    howpublished = {\\url{https://github.com/huggingface/lerobot}}
}

Related