A hierarchical dual-stream POMDP–IPSO framework for multi-UAV cooperative inspection scheduling and control

Abstract

Large-scale power-transmission-line inspection requires coordinated task allocation and safe local motion control. This study presents a hierarchical dual-stream partially observable Markov decision process (POMDP)–Improved Particle Swarm Optimization (IPSO) framework for multi-UAV inspection. At the lower level, a centralized-training/decentralized-execution Multi-Agent Proximal Policy Optimization (MAPPO) algorithm trains a dual-stream actor that separates short-horizon collision perception from communication-topology coordination; a sigmoid curriculum fuses the streams and a differentiable safety filter enforces soft separation, boundary, velocity, and acceleration constraints. At the upper level, IPSO jointly optimizes task-to-UAV assignment and visit order under time, energy, priority-coverage, battery, and time-window criteria. Closed-loop feedback triggers replanning when tasks are completed, deadlines are threatened, energy reserves are low, or a route segment becomes infeasible. In addition to the restricted operational scenarios, we provide a reproducible benchmark with 19 scenarios, 574 tasks, 30 scheduling seeds, and 30 controller test seeds. Under equal evaluation budgets, IPSO reduced mean fitness by 4.8% relative to standard PSO, while the reconstructed dual-stream MAPPO controller achieved 89.3% ± 10.1% task success. Across the evaluated cases, the proposed framework achieved better average performance in most scenarios rather than uniformly dominating every baseline. The code, generated scenarios, raw run-level data, and 600-dpi figures are supplied for independent reproduction. Clinical trial number Not applicable.