Heterogeneous UAVs Trajectory Optimization for Post-Disaster Target Search Based on MARL with Graph Attention Network

Large-scale post-disaster target search missions pose significant challenges for Unmanned Aerial Vehicles (UAVs) due to the complex and unpredictable environments. Heterogeneous swarms that combine the benefits of multiple UAVs can enhance adaptability and search efficiency. Therefore, a leader-follower strategy is employed for heterogeneous multi-UAVs in this paper, where fixed-wing UAVs serve as leaders to provide dynamic relay communication coverage, while multi-rotor UAVs act as followers for multi-target search tasks. We construct a cooperative graph to model relationships among neighboring UAVs and design importance indices for both cooperation and search targets to optimize this graph, streamlining state observation vectors. We then introduce Graph Attention Networks (GAT) to aggregate information from adjacent nodes in the cooperative graph. Additionally, we employ Multi-Agent Reinforcement Learning (MARL) to optimize flight trajectories separately for fixed-wing and multi-rotor UAVs, enhancing cooperation efficiency in multi-target searches. Based on the above methods, we propose a Graph Attention Network multi-agent Actor-Critic (GATAC) algorithm, which effectively addresses the dimension explosion problem that arises as the number of agents increases. Simulation results demonstrate that the GATAC algorithm outperforms existing methods in terms of target search, energy consumption and broken link duration. Finally, we establish a multi-UAV cooperative validation platform, and experimental trajectory plots confirm the stable performance of the GATAC algorithm in real-world environments.