Robot Control Explained for ML People: PID, FK/IK, and Why ACT Outputs Joint Angles

kamathsblog · x · 2026-09-19

jarothschild published an explainer on robot arm control aimed at ML practitioners with no robotics background, using the SO-100 arm (as used by LeRobot) with animations. It covers what a policy output does to a servo, PID and why cheap arms overshoot, how teleop data is recorded, joint vs task space plus FK/IK, why the SO-100 reaches 0% of gripper poses within arm's length, and why ACT outputs joint angles while OpenVLA doesn't. First in a series; the next piece will cover ACT.

Original post →

More from Embodied

Embodied channel →