Capstone: Teach the Arm by Hand
No angles, no gains, no code: guide the arm through the task a few dozen times and let a policy learn the pattern. What you curate is no longer software — it is a dataset.
Builds on: 4.2 Inverse Kinematics: Aiming the Arm Backwards3.2 PID: Three Letters That Run the World
Programming by showing
Everything so far had a human translate the task into mathematics — setpoints, gains, waypoints. Imitation learning removes the translator: guide the robot’s own body through the task (grab the arm, do the motion — called teleoperation when done through a twin controller), while it records what it sensed and what its joints did. Do that a few dozen times. Then train a policy — a function from sensed situation to action — to reproduce the pattern:
The humble version is nearest-neighbour: find the recorded situations most like this one, do what they did, blended. The frontier versions are deep networks trained on the same kind of data, with the same soul. Between them lies mostly capacity, not concept.
The dataset is the program
The consequences invert habits. The policy is only defined where demonstrations were: teach only pickups from the left and the right side is terra incognita — the policy extrapolates, badly, with full confidence. Sloppy demonstrations train sloppy behaviour; ten careful, varied demos beat fifty rushed ones. Debugging changes character entirely: the fix for a failing region is rarely code — it is more demonstrations, there. Coverage, variety and care replace cleverness as the engineering virtues.
Measure it like an engineer
A learned policy has no spec sheet — it has a success rate, and you owe it an honest one: test on fresh targets it never saw (testing on the demos themselves only measures memory), count successes, note where the failures cluster, teach there, retrain, measure again. That loop — measure, find the gap, fill it with data — is the daily craft of the field, and it is exactly what this capstone has you do.
Decades of hand-coded grasp geometry lost, in the span of a few years, to “record 50 demos and train” — because contact-rich tasks (folding, inserting, wiping) resist equations but yield to examples. The open frameworks driving hobby-arm learning today are this lesson’s recipe with bigger networks: teleoperate, record, train, deploy.
⚗️ Lab — The Teaching Studio
The two-link arm, a spawn zone for targets, and a recorder. Drag the arm through pickups to record demos; flip to policy mode and let it try fresh targets alone.
- Record five quick demos on the left side only, then test everywhere — watch the right side fail with perfect confidence.
- Add demos where the failures cluster; retrain; re-measure. The success meter is your judge.
- The checklist is the capstone: work it top to bottom.