Your first arm

This is the module the Hardware Pool exists for. You can do the reading without a robot and the building with one.

Do this

• The SO-ARM100 repository (The Robot Studio). The arm's home: bill of materials, print files, assembly. Read it even if your arm is already built; you'll be fixing it later.
• LeRobot documentation: SO-101 (Hugging Face). Assembly, motor setup, calibration, teleoperation, recording. Follow it exactly the first time. Improvise the second time.
• The Hugging Face Robotics Course. The free, approachable walk through the ideas behind LeRobot: datasets, imitation learning, policies. Do it alongside the docs.
• Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware (Zhao et al., 2023). The ACT paper. Read the introduction and the method section; it's the policy you'll train first.
• Diffusion Policy (Chi et al., 2023) and SmolVLA (Hugging Face, 2025). Read after your first ACT policy works, to see where it goes next.

Build

1. Calibrate both arms and teleoperate. Pick up a block ten times.
2. Record fifty episodes of one task with a camera. Upload the dataset to the LeRobot organisation on Hugging Face or your own account.
3. Train ACT on it. Run the policy. Film the first success and the first failure. Post both.

Skip

• Training on your laptop CPU. Use a free Colab GPU or a cloud GPU for the training step.
• Changing hyperparameters before you've reproduced the defaults.

Done when: a policy you trained does a task you demonstrated, at least some of the time, and you can say why it fails when it fails.