Two-Level-Attention-Based Continuous Trajectory Design and Computation Offloading for Multi-UAV Cooperative Target Search

Unmanned Aerial Vehicles (UAVs) are widely used for region monitoring and target search. However, existing UAV swarm search methods usually suffer from UAV motion models' low accuracy and inflexibility, resulting in low search efficiency, and there is a gap between the UAV motion model design and reality. Moreover, current methods ignore the impact of search task offloading locations and computational resource allocation on search efficiency. Finally, existing UAV swarm search methods lack scalability and generalization as the number of UAVs increases. This paper explores the cooperative search problem of UAV swarms in three-dimensional continuous space. We comprehensively consider UAV partial observability, search area delay constraint, collision avoidance condition, energy collection and consumption, etc. We jointly optimize the UAV flight trajectory, charging decision, search task offloading, and computing resource allocation to reduce the uncertainty of the search area and improve target detection rates. Specifically, we propose a heuristic embedded Multi-Agent Proximal Policy Optimization (MAPPO) algorithm. This method significantly reduces the action space dimension by using a heuristic energy-efficient task offloading and computational resource allocation strategy, and uses MAPPO with a two-level attention mechanism, a Beta distribution, and a safe flying control strategy to efficiently obtain the optimal flight trajectory and charging strategy. Two-level attention can handle a dynamically sized agent network, significantly improving algorithm performance and scalability. Furthermore, through the Curriculum Learning (CL) mechanism, we extend our algorithm to a large-scale UAV swarm cooperative search scenario. Simulation results show that compared with other benchmarks, our method can significantly reduce the search area's uncertainty and improve the target detection rate.