A Multi-UAV Cooperative Coverage Method Based on Sparse Dual-Attention Reinforcement Learning

Abstract

This letter addresses the problem of cooperative multiple unmanned aerial vehicle (multi-UAV) coverage in dynamic environments, where UAVs aim to maximize target coverage while avoiding obstacles in real-time. Existing approaches often face difficulties in real-time decision-making for complex coverage tasks and exhibit limited adaptability. To overcome this problem, we propose a dual-attention mechanism with dynamic sparse activation, integrated into a multi-agent deep reinforcement learning (MADRL) framework under centralized training and decentralized execution (CTDE). Our method hierarchically processes the observation set-including targets, obstacles, and neighboring UAVs-while employing dynamic sparse activation to improve the accuracy and efficiency of key feature learning during training. This mechanism enables agents to selectively focus on critical environmental cues, reducing redundant computations and improving decision-making in real-time multi-UAV coordination. Furthermore, to enhance cooperative behavior among UAVs, an interaction attention mechanism is designed for each UAV, enabling them to coordinate coverage strategies through information exchange. Experimental results demonstrate that our approach outperforms existing methods in complex coverage tasks and exhibits strong scalability in large-scale UAV coordination scenarios.