Why this unit exists: The most common reason people abandon a robotics tutorial is not difficulty — it is getting stuck mid-session because a dependency is missing, their Python version is wrong, or they do not know who to ask. This unit exists so that never happens to you.

What Is LeRobot?

LeRobot is HuggingFace's open-source robot learning library. It gives you a unified dataset format, pre-trained VLA models (ACT, Diffusion Policy, Pi0Fast, SmolVLA), and a hardware abstraction layer that works identically across more than 10 robot platforms. It is the de facto standard for open-source imitation learning in 2025–2026.

This path is framework-focused, not hardware-specific. The workflow you learn here applies to any supported robot and transfers directly to new hardware as you add it.

What You Will Build

By the end of this path you will have a trained ACT pick-and-place policy that runs on your robot (or in simulation). You will record 50–100 demonstrations, train the policy, evaluate it, and share both the dataset and the model on HuggingFace Hub. The same workflow scales to any task — once you have done it once, you can apply it to anything.

Supported Hardware

This path works with any of the following. You only need one.

SO-100 / SO-101
Koch v1.1
OpenArm 101
DK1 Bimanual Kit
HopeJR
Reachy2
MuJoCo Sim (no hardware needed)

No physical robot? You can complete the entire path in MuJoCo simulation — Unit 1 covers simulation setup explicitly.

Software Checklist

Verify each item before starting Unit 1.

  • Python 3.10 or higher — run python3 --version to check. Install via sudo apt install python3.10 or pyenv.
  • pip — run pip3 --version. Install via sudo apt install python3-pip.
  • Git — run git --version. Install via sudo apt install git.
  • A free HuggingFace account — needed in Unit 3 to push your dataset and in Unit 6 to share your model. Create one at huggingface.co — it takes 2 minutes.
  • ~10 GB free disk space — for LeRobot, MuJoCo, and your dataset. Datasets grow quickly; 20 GB is comfortable.
  • CUDA GPU (optional) — ACT training on CPU works for small datasets but takes 8–12 hours instead of 1–2 hours. An NVIDIA GPU with 8GB+ VRAM is recommended for Unit 4. Cloud GPU options are covered in Unit 4.

Time Estimates

These are realistic estimates including setup and troubleshooting time. Training in Unit 4 can run unattended — start it before sleep.

Unit What You Do Time
0This orientation20 min
1Install and verify LeRobot~1 h
2Explore dataset format~1.5 h
3Record 50+ demonstrations~2 h
4Train ACT policy~3 h
5Evaluate and diagnose~1 h
6Improve and share~1.5 h
Total ~10 h 20 min

How to Get Help

If you get stuck, use these resources in this order:

  1. Check the completion check at the bottom of the unit you are in — it often identifies exactly what is missing.
  2. Search the LeRobot GitHub Issues — most environment and hardware config errors are documented there.
  3. Post in the SVRC forum — include your Python version, OS, robot type, the exact error message, and which unit you are in.
  4. Join HuggingFace Discord #lerobot — the LeRobot maintainers are active and respond quickly during PST business hours.

Do not skip ahead while blocked. The path is sequential by design — if you are stuck in Unit 3, it usually means Unit 1 or Unit 2 is not fully complete.

Orientation Complete When...

You have checked every item in the software checklist, you have a HuggingFace account, you know your path to the forum if you get stuck, and you have set aside your first 1-hour session for Unit 1.