When Sensor Fusion Loses

On a maneuvering target inside its configured model bank, an interacting multiple-model filter reduces OSPA localization error from 26.92 to 14.40. That is the useful result. The boundary is equally useful: the same bank loses on straight motion and on a 12-degree-per-second turn outside the configured plus-or-minus 6-degree-per-second hypotheses.

More models do not create universal robustness. They trade mismatch against mode probability, mixing, and additional variance. The appropriate question is not whether IMM is better, but whether the expected motion regime is represented.

In the recorded light-clutter crossing, JPDA localizes slightly better yet performs worse on total OSPA: 39.55 versus 28.73 for GNN. It maintains roughly 2.7 times as many false tracks. Under 80 false alarms, the ordering reverses narrowly—47.75 versus 49.09—but the JPDA run costs 17.08 seconds and coverage falls to 0.737.

This is why SIGNALROOM shows truth, detections, estimates, and residuals as separate visual layers. A smoother-looking track can coexist with a worse population-level metric.

The VIO benchmark reaches 0.041 meters ATE over a 40-second repeating figure-eight while dead reckoning reaches 2.106 meters. During hover, the ranking collapses: MSCKF reaches 1.202 meters while dead reckoning remains at 0.094. With no translational parallax, depth becomes unobservable and 12,086 feature measurements are rejected.

No UI control can repair missing information. The interface therefore labels hover as an observability loss, not a tuning problem.

Phase-gradient autofocus reduces a four-radian defocus impulse width from 1.1587 meters to 0.3317, close to the 0.3326-meter theoretical reference. Random high-order errors, however, may barely improve or can make the width metric misleading.

A responsible imaging demonstration keeps the phase-error family, point-target assumption, and measurement definition beside the improvement ratio.