关键词: PA-MADDPG算法, 固定翼无人机集群, 物理感知注意力, 自适应门控复合稠密奖励函数, 进化策略同步, 多智能体强化学习

Abstract: To address the problems of formation maintenance and cooperative obstacle avoidance for fixed-wing unmanned aerial vehicle (UAV) swarms in complex and unknown threat environments, a physics-aware attention multi-agent deep deterministic policy gradient (PA-MADDPG) algorithm is proposed in this paper. First, a physics-aware attention module is designed, which explicitly integrates the artificial potential field (APF) method into the attention network as physical prior knowledge. This significantly enhances the collision avoidance robustness and cross-scale topological generalization capability of the swarm under complex dynamic threats. Second, an adaptive gated composite dense reward function is formulated to effectively alleviate the insufficient gradient signals caused by sparse rewards and multi-objective conflicts in long-horizon tasks. Furthermore, an evolutionary strategy synchronization mechanism is introduced. By periodically screening high-fitness agents and synchronizing their policy parameters to other agents via soft updates, coupled with prioritized experience replay (PER), the training convergence is substantially accelerated. Simulation results demonstrate that in both static and dynamic threat scenarios, the proposed algorithm achieves inter-UAV collision avoidance and environment threat avoidance rates of over 92%, significantly outperforming baseline algorithms in comprehensive task effectiveness. In dynamic topological generalization scenarios, the avoidance rates stably remain above 82%. These results confirm that the PA-MADDPG algorithm not only realizes efficient obstacle avoidance and formation maintenance for fixed-wing UAV swarms but also exhibits strong robustness for flexible transfer in complex reconfiguration tasks.

Key words: PA-MADDPG, Fixed-wing UAV swarm, Physics-aware attention, Adaptive gated composite dense reward, Evolutionary strategy synchronization, Multi-agent reinforcement learning