Uncertainty-Aware Multimodal Inference under Sensor Degradation: A Controlled Study of Learned versus Heuristic Sensor Trust

We study a controlled synthetic state-estimation problem in which a 2D position is recovered from three sensors with different noise levels, occasional bias faults, dropout and a noisy self-reported quality signal. We ask whether learning a per-sensor uncertainty estimate and feeding it into precision-weighted fusion beats a much cheaper heuristic that uses the quality signal directly. Against a baseline ladder, with five training seeds and fixed evaluation sets, the learned model is a near-tie with the heuristic under nominal conditions (mean RMSE 0.705 vs. 0.701). A learned cross-sensor residual feature and Huber/Tukey robust reweighting did not improve on the heuristic. Under shifts that corrupt or invert the quality signal, the learned model is less affected than the heuristic, but fixed weighting is better than either adaptive method in the most severe cases; we suggest bounded predicted variance as an untested explanation. Under variance-matched heavy-tailed noise, calibration error rises more than fivefold (ECE 0.025 to 0.129) while predicted sharpness is unchanged and a moment-matched scalar recalibration does not help, which is consistent with a mismatch in tail shape. We state the limits of the study, including missing controls, and release the code.