lerobot
LeRobot

Open-source AI for robots that live in the real world: record, train, deploy, share.

Hugging Face Hugging Face lerobot v0.6.1 · huggingface/lerobot
Two arms, one policy, folding a shirt
Hugging Face Hugging Face

The AI community building the future

Datasets

Thousands of curated datasets for ML tasks

Saves time in data collection and pre-processing

Some include a dataset viewer for interacting with data

Models

Pre-trained ML models

Can be used to perform diverse ML tasks including text, audio, and image

Supports dozens of libraries, including Transformers

Spaces

Build, host, and share ML apps in a single, organized workspace

Thousands of ML-powered demo apps

Hosted with built-in Python frameworks (Streamlit or Gradio) or JS and HTML (static)

Buckets

S3-like object storage on the Hub

Mutable and unversioned: fast storage for checkpoints, logs and raw data

Train LeRobot policies by streaming straight from a bucket

LeRobot

One library for the whole robot learning loop

episode 01Robots episode 02Teleoperate episode 03Datasets episode 04Models episode 05Inference episode 06Evaluate failures and new episodes become training data

The same Python API from a $100 arm to a humanoid, and the same dataset format from recording to training.

EPISODE 01 / 06

Robots

task: Build a robot, or plug in the one you have
Open hardware

Real robots you can 3D-print and assemble yourself

SO-101
A 3D-printed arm. You move the leader by hand and the follower copies it.
LeKiwi
The same arm on a mobile base, so the robot can drive to the task.

Parts lists, print files and assembly guides are open source.

SO-101 arm
LeKiwi
Robot-agnostic

Same API, from a desk arm to a humanoid

Arms

  • SO-100 / SO-101 · Koch · OMX
  • OpenArm · reBot B601

Mobile

  • LeKiwi
  • EarthRover Mini+

Humanoids

  • Unitree G1 · Reachy 2
  • HopeJR · any robot via plugins
Mobile manipulator with two arms and a LeRobot sticker Orange and black humanoid robot Quadruped robot walking out of a lift Small robot with a camera head on a table
EPISODE 02 / 06

Teleoperation

task: Show the robot how it's done, and record it
Collecting demonstrations with a VR headset
Control

Drive the robot with whatever is on your desk

leader arm
A twin arm you move by hand
phone
Point and tilt to steer
gamepad
Mobile bases and quick tests
keyboard
No extra hardware at all
VR · glove
Headsets and the Homunculus glove for hands

Every session is saved as an episode, ready to train on.

EPISODE 03 / 06

Datasets

task: Share every episode with everyone
The ChatGPT moment of robotics?

Robot data is growing like text data once did

80k+LeRobot datasets on the Hub
200+ TBof robot data hosted
2 PBmicroAGI has committed to using 2 PB of HF Buckets
Cumulative LeRobot datasets on the Hub, from near zero in mid-2024 to over 80,000 in mid-2026
Community datasets

Thousands of tasks, kitchens, desks and robots

They all load with the same line of code.

from lerobot.datasets import LeRobotDataset
ds = LeRobotDataset("lerobot/svla_so101_pickplace")
LeRobotDataset v3

One format for every robot

meta/info.json # features, fps meta/stats.json # normalisation meta/episodes/ # offsets, tasks data/*.parquet # state, action videos/*.mp4 # every camera
stream
Train without downloading: Hub or Buckets
video
Cameras stored as MP4, encoded while recording
scale
Many episodes per file: millions of episodes
The dataset visualizer: every camera, state and action, in sync
EPISODE 04 / 06

Models

task: Learn a policy from the data
Two generations of robot policies

From one task per model to one model for many tasks

ACT architecture: transformer encoder and decoder predicting a chunk of actions
ACT · Action Chunking with Transformers

Small, trained from scratch on one task. Fast to train on a single GPU.

GR00T N1.7 architecture: vision-language model feeding a diffusion transformer action head
GR00T N1.7 · a Vision-Language-Action model

Pretrained on many robots, follows language instructions, fine-tuned on yours.

Model zoo

20+ models, one API: same dataset, same lerobot-train

Vision-Language-Action 11

Pretrained generalists you fine-tune on your own robot data
π0.5Physical Intelligence
SmolVLAHugging Face
π0Physical Intelligence
π0-FASTPhysical Intelligence
GR00T N1.7NVIDIA
MolmoAct2Ai2
WALL-OSSX Square Robot
X-VLA
EO-1
EVO1
VLA-JEPA

Imitation learning 4

Task-specific policies trained from scratch
ACT
Diffusion Policy
VQ-BeT
Multitask DiT

Reward models 4

Score progress and success from video
SARM
ROBOMETER
TOPReward
Reward classifier

World action models 2

Predict future video and actions together
FastWAM
LingBot-VA

Reinforcement learning 2

Learn from trial, error and human corrections
HIL-SERL
TD-MPC
Coming next release: DM0.5 · FineART-VLA · FLUX 3 Action · LaWAM
Plus Real-Time Chunking: smooth motion for large VLAs despite slow inference
LeRobot v0.6.1
SARM progress estimates on two real tasks
Reward models

How close is the robot to done?

SARM watches the camera and the task text and predicts progress for every frame, from 0 to 1.

EPISODE 05 / 06

Inference

task: Run the policy on the real robot
Inference

The trained policy drives the real robot

Remote inference

The policy runs on a remote GPU. The robot keeps moving while the next chunk of actions is computed.

See it on X · @LeRobotHF ↗
Robot computer laptop or single-board computer cameras · motors · teleop · recording action buffer executes the next action every tick Policy server GPU workstation, cloud or HF Jobs π0.5 · SmolVLA · GR00T · ACT … observation + task next chunk of actions
EPISODE 06 / 06

Evaluation

task: Measure, then improve
LIBERO demonstrations in simulation
Benchmarks

Compare policies on the same tasks

LIBEROMeta-WorldRoboCasa RoboTwinVLABenchRoboMME

Simulation gives a cheap, repeatable score. On the real robot, rollouts are recorded so failures become new training data.

AFTER THE LAST EPISODE

Why LeRobot?

The community, and what it builds.
What makes LeRobot special?

Its community

LeRobot worldwide hackathon
Community

Builders on every continent

World map with LeRobot community groups across North America, Europe, Africa, Asia, South America and Oceania
LeRobot community locations
Two-arm robot folding shirts · 2× speed
Endless possibilities

Folding laundry is a dataset away

Record a few dozen demonstrations, fine-tune a pretrained policy, deploy. The same recipe works for your task.

Who uses LeRobot

Labs, startups and big tech build on it

NVIDIA, Unitree, Scale, ROBOTIS, Physical Intelligence, Stanford, Microsoft, University of Oxford, Shanghai AI Laboratory, rerun.io, AgiBot, Qualia, Enactic, Ai2, X Square Robot, intrinsic, ByteDance, Hexagon, RobCo
LeRobot v0.6 · July 2026

Imagine, evaluate, improve

imagine
Policies that imagine the future before acting (world action models)
evaluate
Reward models that tell you when the robot succeeds, and six new simulation benchmarks
improve
A deployment CLI that turns failures into training data
also
Depth sensing · VLM dataset annotation · custom video encoding · cloud training on HF Jobs · a leaner install
Get started

From zero to a trained policy in three commands

# install
pip install lerobot

# 1. record demonstrations with a leader arm
lerobot-record --robot.type=so101_follower --teleop.type=so101_leader \
  --dataset.repo_id=$HF_USER/my_task --dataset.single_task="Grab the black cube"

# 2. train a policy (add --job.target=a10g-small to train on a Hugging Face GPU)
lerobot-train --dataset.repo_id=$HF_USER/my_task --policy.type=act

# 3. run it on the robot
lerobot-rollout --policy.path=$HF_USER/my_policy --robot.type=so101_follower

Robot port and camera flags omitted. Full guide: huggingface.co/docs/lerobot

More than LeRobot

Open-source robots from the same community

Reachy Mini, a small white desktop robot with antennas
Microduck, a small orange and white walking robot
Thank you

Questions, robots, datasets: come and say hi.

LeRobot LeRobot
QR code to LeRobot on GitHub
GitHub
QR code to LeRobot on X
X
QR code to the LeRobot Discord
Discord
← → to move · F for full screen · T for theme