Learning Path · Intermediate · Any Supported Robot

LeRobot: Open-Source Robot Learning from Zero to Deployment

A structured, sequential path through HuggingFace's LeRobot framework. Install, record demonstrations, train an ACT policy, and deploy it — on any supported robot or entirely in simulation.

Total Time ~10 hours
Difficulty Intermediate
Hardware Any supported robot or MuJoCo simulation
Prerequisites Python basics, Linux terminal basics
Simulation Option Yes — full path completable in MuJoCo
You Will Build A trained ACT pick-and-place policy, runnable on your robot

Before You Start — Check These Prerequisites

  • Comfortable with a Linux terminal (cd, ls, pip install)
  • Python 3.10 or higher installed
  • A supported robot (SO-100, Koch v1.1, OpenArm, DK1, HopeJR, Reachy2) or a simulation environment
  • A free HuggingFace account (for dataset and model hosting)
  • About 10 hours of total time across multiple sessions

Not sure if you qualify? Start with Unit 0 — it is specifically designed to answer that question.

Your Path at a Glance

Complete the units in order. Each unit ends with a completion check — do not move on until you pass it.

0

Before You Begin: LeRobot Setup Required

What LeRobot is, supported hardware, software requirements, time estimates, and where to get help. No code — pure orientation.

~20 min · No code
1

Install and Verify LeRobot Required

pip vs source install, hardware config setup, and running a pre-trained ACT policy in MuJoCo simulation to verify your environment.

~1 hour
2

Understand the LeRobot Dataset Format Required

Dataset structure, episode schema, loading and visualizing existing datasets from HuggingFace Hub. Understand what goes in before you record your own.

~1.5 hours
3

Record Your Own Demonstrations Required

Connect your robot or simulation, record demonstrations with the LeRobot record command, and push your dataset to HuggingFace Hub.

~2 hours
4

Train a Policy Required

Choose between ACT, Diffusion Policy, and SmolVLA. Train ACT on your dataset, read training logs, and save a checkpoint ready for evaluation.

~3 hours
5

Evaluate Your Policy Required

Evaluate in simulation or on your real robot, measure success rate, and diagnose the top failure mode before iterating.

~1 hour
6

Improve and Share Required

Run the data flywheel, mix datasets, share your model and dataset on HuggingFace Hub, and plan your next steps with Pi0Fast and SmolVLA.

~1.5 hours

Time Breakdown

Unit Topic Time
Unit 0Orientation~20 min
Unit 1Install & Verify~1 h
Unit 2Dataset Format~1.5 h
Unit 3Record Demos~2 h
Unit 4Train Policy~3 h
Unit 5Evaluate~1 h
Unit 6Improve & Share~1.5 h
Total ~10 h 20 min
Start Unit 0 — Orientation →

Takes about 20 minutes. No technical content — makes sure you have everything ready before the path begins.