Persistent Multi-UAV Road Network Monitoring via MAPPO with Rolling-Horizon PUCT Search and Distillation

Abstract

Persistent road-network monitoring with multiple UAVs requires continuously refreshing timesensitive traffic knowledge under battery constraints. Unlike one-shot coverage, monitoring quality depends on both where and when UAVs revisit road segments, making long-term coordination a central challenge. In this paper, we formulate persistent multi-UAV road-network monitoring as an asynchronous graph decision problem with temporally decaying road knowledge. We develop a decentralized monitoring framework in which UAVs repeatedly select road segments to maintain network-wide awareness while operating under charging constraints. To improve decision quality without sacrificing scalability, we combine learned decentralized control based on multiagent proximal policy optimization (MAPPO) with short-horizon planning based on predictor + upper confidence bounds for trees (PUCT) search. The framework also operates in the presence of background delivery UAV traffic, whose motion affects coordination in the shared road network. Experiments in an urban road-network environment show that the framework provides an effective and practical solution for persistent monitoring and improves decentralized monitoring performance in dynamic road networks.