MLFrameworksModels

LeRobot Documentation — Real-World Robot Learning with PyTorch

by Hugging Face

IntermediateDocumentationFree~4-6 hours for the core tutorials without hardware; days to weeks once a real arm is in the loop

The teleoperate-record-train-deploy loop for real robots, in the same PyTorch idiom you already use for LLMs.

Start LearningAdded Sep 20, 2026 · Updated Sep 20, 2026

Overview

The LeRobot docs are organised around one workflow — Teleoperate, Record, Train, Deploy — and every section maps onto a stage of it. Get Started covers installation (`pip install lerobot`) and branches three ways: you have a robot, you have no hardware yet, or you want to contribute. The hardware track documents supported platforms including the flagship low-cost SO-101 arm, LeKiwi mobile base, Koch v1.1, HopeJR, OpenARM, Reachy 2 and the Unitree G1 humanoid, plus keyboard, gamepad and phone teleoperation. The no-hardware track is genuinely complete: you can train on any LeRobotDataset streamed from the Hugging Face Hub, evaluate against the LIBERO (130+ VLA tasks) and Meta-World benchmarks, or run the free Colab notebooks. Policy pages document imitation learning (ACT, Diffusion Policy, VQ-BeT, Multitask DiT), reinforcement learning (HIL-SERL, TDMPC), vision-language-action models (π₀, π₀.5, GR00T N1.7, SmolVLA, XVLA, EO-1, WALL-OSS) and world models (VLA-JEPA, FastWAM). Supporting guides carry the parts tutorials usually omit — a CLI cheat sheet, a Compute & Hardware guide telling you which policy fits your GPU and how long training actually takes, the LeRobotDataset v3 format (chunked episodes, streaming, Parquet metadata), camera troubleshooting, and LeLab, a browser GUI for the same workflow. Apache-2.0, ~27.6k GitHub stars, maintained by the Hugging Face robotics team.

At a Glance

Topic
ML
Level
Intermediate
Format
Documentation
Cost
Free
Duration
~4-6 hours for the core tutorials without hardware; days to weeks once a real arm is in the loop
Provider
Hugging Face
Hands-on
Yes — code/exercises
Certificate
None

What You’ll Learn

  • Run the full teleoperate, record, train and deploy loop on a physical robot arm
  • Record synchronised camera video and action data into the standard LeRobotDataset v3 format
  • Train an ACT imitation-learning policy from your own demonstrations as a first end-to-end project
  • Fine-tune and deploy vision-language-action models including SmolVLA, π₀, π₀.5 and GR00T N1.7
  • Evaluate policies in simulation against the LIBERO and Meta-World manipulation benchmarks
  • Choose a policy that fits your GPU budget using the Compute and Hardware guide's training-time tables
  • Stream large robot datasets from the Hugging Face Hub instead of downloading hundreds of gigabytes
  • Diagnose the failure modes that actually bite beginners, starting with camera lighting and calibration

Highlights

  • The no-hardware path is first-class, not an afterthought — train and benchmark policies before buying an arm
  • Hardware-agnostic Python interface spans a $100 SO-101 arm and a Unitree G1 humanoid with the same CLI
  • Policy coverage is unusually current: π₀.5, GR00T N1.7, XVLA and world models like VLA-JEPA are documented, not just referenced
  • LeRobotDataset v3 handles 400GB+ datasets with chunked episodes and video streaming, which is where most robot-learning tutorials stop
  • 27.6k stars, Apache-2.0, and a plugin system that lets you add custom hardware without forking the core library

Who It’s For

Best For

  • ML engineers with PyTorch experience who want to move from LLMs into embodied AI
  • Researchers evaluating vision-language-action policies against a common benchmark suite
  • Hobbyists assembling a low-cost SO-101 or LeKiwi robot and needing a working training pipeline
  • Teams standardising robot data collection on a shareable dataset format

Prerequisites

  • Comfortable Python and working familiarity with PyTorch training loops
  • A CUDA GPU for training policies (Colab notebooks work for the tutorials)
  • Optional: a supported robot arm — everything else runs in simulation or on Hub datasets

FAQ

What is LeRobot Documentation — Real-World Robot Learning with PyTorch?

LeRobot is Hugging Face's open-source robotics library and its documentation is the fastest route from transformer engineering into vision-language-action models on physical hardware. It is written for Python and PyTorch developers who have never touched a robot, and it walks the whole loop: teleoperate an arm, record demonstrations as a versioned dataset, train an imitation-learning or VLA policy on it, then deploy that policy back onto the robot. After working through it you can train ACT or SmolVLA on a Hub dataset and evaluate it in simulation without owning any hardware at all.

Is LeRobot Documentation — Real-World Robot Learning with PyTorch free?

LeRobot Documentation — Real-World Robot Learning with PyTorch is free to access.

What level is LeRobot Documentation — Real-World Robot Learning with PyTorch for?

LeRobot Documentation — Real-World Robot Learning with PyTorch is aimed at a intermediate audience. Recommended background: Comfortable Python and working familiarity with PyTorch training loops, A CUDA GPU for training policies (Colab notebooks work for the tutorials), Optional: a supported robot arm — everything else runs in simulation or on Hub datasets.

How long does LeRobot Documentation — Real-World Robot Learning with PyTorch take?

Expect roughly ~4-6 hours for the core tutorials without hardware; days to weeks once a real arm is in the loop. Most learners work through it at their own pace.

What will I learn from LeRobot Documentation — Real-World Robot Learning with PyTorch?

You'll learn: Run the full teleoperate, record, train and deploy loop on a physical robot arm; Record synchronised camera video and action data into the standard LeRobotDataset v3 format; Train an ACT imitation-learning policy from your own demonstrations as a first end-to-end project; Fine-tune and deploy vision-language-action models including SmolVLA, π₀, π₀.5 and GR00T N1.7; Evaluate policies in simulation against the LIBERO and Meta-World manipulation benchmarks; Choose a policy that fits your GPU budget using the Compute and Hardware guide's training-time tables; Stream large robot datasets from the Hugging Face Hub instead of downloading hundreds of gigabytes; Diagnose the failure modes that actually bite beginners, starting with camera lighting and calibration.

Topics

roboticsvision-language-actionimitation-learningpytorchlerobotembodied-ai

Sources

This page was written from 3 sources, 1 on domains other than huggingface.co.

  1. 1.huggingface.coindexvendor
  2. 2.github.comlerobot
  3. 3.huggingface.colerobot release v040vendor