Examples

Nine robots, four controllers, one motion each. Every pane below plays the same held-out motion under a different set of joint gains; the faint ghost is where the joint was told to be. Links turn toward red as that joint falls behind its command. Drag a pane to orbit, scroll or pinch to zoom, and pick a robot to load it.

What the four columns are

• Factory gains are the drive's as-found settings: the soft defaults a servo ships with, before anyone tunes it.
• Analytic start is computed from the robot description alone (masses, inertias, gear ratios), before the robot moves at all.
• Tuned is what pidtuna delivers: it identifies each joint, sweeps the gains to the edge of stability, refines gains and feed-forward, and then checks the result on motion it did not tune on.
• Oracle (best possible) is the best admissible controller a brute-force search found on the true simulated plant. It is a ceiling, not something a real tuner can know: the tuner never sees the plant the oracle searched.

All four play the same held-out motion through the tuner's own simulators: a single-axis servo simulator for the lever and the conveyor, MuJoCo for the rest. The replay is precomputed; your browser only redraws it.

Tuned does not always beat the analytic start on tracking. On the lever, the analytic start tracks better but holds the drive at its torque limit for part of the held-out motion, and the tuner refuses a controller that does that. The score table says so when it happens, in the "At torque limit" column.