Coupling calibrated aerial perception to adaptive UAV emergency logistics scheduling: a YOLOv8–reinforcement learning closed-loop proof of concept

Unmanned aerial vehicles can survey an affected area and deliver relief, but most scheduling studies assume that demand is already known. We examine what happens when demand must instead be inferred from imperfect aerial detection. The framework combines an empirically characterized YOLOv8-Nano observation model, a disaster-situation state map, and a confidence-aware PPO scheduler in a simulated re-observe–re-plan loop. The detector achieved mAP50 = 0.251 and an end-to-end CPU rate of approximately 22 frames per second. Its global ECE was 0.0104, whereas the people-group ECE increased to 0.192 at the operational confidence threshold. Across seven independent training seeds, confidence-awareness significantly increased the composite training reward (Wilcoxon p = 0.016), but delivery rate, coverage, and unmet demand did not differ from the confidence-blind controller, and average delay was longer under the confidence-aware policy. This reward difference is interpreted as a training diagnostic because the two controllers use different reward definitions. A single-replication 2 × 2 factorial analysis indicated that either component individually recorded higher delivery than the baseline, whereas their combination did not; this exploratory pattern has not been confirmed across multiple training seeds. Within the tested operating range, the coupled system remained below the perfect-information reference. An independent recall–FDR sweep indicated that delivery was more sensitive to false-discovery rate than to recall for the fixed policy evaluated. These findings are limited to a controlled simulation with temporally independent detection errors, a synthetic demand proxy derived from a non-disaster benchmark, and a two-dimensional grid. Under these conditions, perception strongly limits logistics performance, while the effect of confidence-awareness on relief outcomes remains unestablished.