Learning Calibrated Uncertainty Margins for an Explainable 4D-Radar UAV Warning Policy

Reliable collision warning from 4D millimeter-wave radar is hindered by the evaluation of deterministic distance–speed boundaries on uncertain track states. Directly learning the alarm decision may improve flexibility, but obscures safety-critical logic. To reconcile robustness with interpretability, this paper develops a hybrid uncertainty-aware warning framework in which a lightweight multilayer perceptron (MLP) estimates deviations from nominal radar tracks, whereas the final binary decision remains governed by explicit geometric and kinematic conditions. Validation-residual calibration transforms the instantaneous estimates into empirical upper margins, and a projection-bounded leakage recursion preserves recent uncertainty during brief observation gaps. The margins act only on the time-based entry branch; the hard-distance override, motion-direction gate, confirmation logic, and release rule remain deterministic. The framework is evaluated using 36 fixed-radar, single-reflector physical recordings and 216 trial-isolated controlled-degradation realizations derived from them. The calibrated margins attain 95.24% range-error coverage and 94.04% speed-error coverage relative to the nominal tracks. Among 198 realizations with a reference entry, robust-margin methods reduce late or missed warnings from 15 to 2. At this protected operating point, the calibrated MLP decreases alarm disagreement from 4.94% for a fixed margin to 3.90%, compared with 4.16% for a calibrated linear model. The results identify a condition-dependent margin-allocation benefit while retaining an auditable warning policy for the tested recordings.