Characterization and tuning

Tune from the inside out. Changing an earlier layer invalidates later gains.

1. Verify mechanics and units

Lift the drive. Positive voltage must increase both reported wheel distances. Pushing straight must increase both equally. A counter-clockwise turn must increase heading. Measure effective wheel circumference under robot load and effective track width from repeated multi-turn tests, not CAD alone.

2. Identify each side's feedforward

Log voltage, wheel velocity, and wheel acceleration during slow quasistatic ramps and faster steps. Fit volts = kS*sign(v) + kV*v + kA*a independently for left and right. Validate on held-out runs. Use a voltage ceiling below nominal while first testing.

3. Tune wheel velocity feedback

Start with ki = kd = 0. Increase kp until measured velocity follows steps without chatter. Add only enough filtered kd to damp a repeatable overshoot. Use ki only for a persistent loaded bias that feedforward did not capture; set a tight integralLimit.

4. Establish conservative trajectory limits

Measure sustainable wheel speed and use 70–80% initially. Increase acceleration until tracking error rises or wheels slip, then back off. Set deceleration separately. Lower centripetal acceleration until tight turns no longer scrub or tip. A good controller cannot recover traction that the plan already demanded.

5. Tune pose feedback

Keep kd around 0.7–0.9. Raise kp until injected 5–10 cm pose errors recover promptly. If it snakes on straights, reduce kp or fix delayed/noisy localization. Keep heading tolerance realistic; demanding less than sensor noise guarantees dithering.

6. Validate adversarially

Test low battery, added payload, reverse paths, starts offset in every direction, tight S-curves, and endpoint disturbances. Log reference/measured pose, both wheel targets/measurements, voltages, saturation, status, and actual loop dt. Do not tune based only on the final pose.