Cross-Layer Co-Design and Physics-Grounded Evaluation of Hybrid Reinforcement Learning and Federated ANN-PID Protocol Suites for Dense Urban UAV Delivery Networks

Abstract

Urban unmanned aerial vehicle (UAV) fleets require coordinated optimization of communication, mobility, and energy management. This study evaluates conventional and artificial intelligence (AI)-enabled cross-layer protocol suites in a common five-hour simulation of 7,500 UAVs, 240 base stations, and 40 vertiports across 144 km². They include AODV-based conventional routing, reinforcement learning-based UAV-AI-EDRL, federated ANN-PID-based UAV-AI-FHAP, reduced learning baselines, and a composite. Methodological contributions comprise a mission-constrained cross-layer objective, CMA-ES calibration benchmarks, physics-grounded training, paired evaluation over six environmental realizations, and packet-level NS-3 checks. In fleet-scale runs, proactive routing reduced mean end-to-end delay from 66.0 to 25.3 ms (61.6%), an advantage NS-3 reproduced at every transmit power only with pre-resolved link-layer addresses. UAV-AI-EDRL reduced power by 8.7% and near-miss events by 18.1% on the principal training seed, although safety gains varied across seeds. UAV-AI-FHAP reduced power by 4.6% and near-miss events by 30.9%, but completed 15.9% fewer missions because a charging threshold calibrated on a reduced-area oracle interrupted long delivery legs; geometry-preserving recalibration reduced the deficit to 2.2%, and a leg-aware charging rule produced a 13.9% gain (2.3% if the conventional suite also uses it). Ablations with equivalence tests attributed the gains to proactive forwarding, throttling, spread control, and reactive avoidance, finding no measurable suite-level benefit from DQN, PPO, path planning, ANN-PID, or federated aggregation; suites without learned power control preserved superior raw radio metrics. These findings, limited to the simulated conditions, shifts, and fleet sizes, reveal trade-offs among latency, energy, safety, and mission completion, emphasizing operationally weighted selection and field validation.