Excitation-Supervised Closed-Loop Self-Calibration and Target Seeking for an Unknown-Pose Range-Bearing Relay

What the math gives to ML

The paper offers a constructive excitation certificate for online calibration: centered pose spread lower-bounds yaw information and therefore upper-bounds calibration variance. A supervisor compares this certificate with an accuracy-derived threshold, injects exploratory motion when the threshold is not met, and projects the nominal target-seeking command away from the excitation direction. This transfers to online neural system identification and sensor-fusion networks with unknown latent coordinate transforms, where a model should not trust a calibration head until recent data provide sufficient information. The central falsifiable prediction is that calibration variance decreases approximately inversely with accumulated excitation information, while feedback-triggered exploration remains reliable when a fixed exploration schedule decays too quickly.