CA-MAPPO: Collaborative UAV Search via Convolutional Attention and Adaptive Reward

Le résumé fourni par la source

Multi-UAV cooperative search plays a vital role in large-scale surveillance, emergency response, and environmental monitoring. However, existing decision-making frameworks face two challenges: redundant paths caused by repeated area revisits during execution, and limited extraction of global spatial patterns from target probability maps (TPMs), which weakens coordination in partially observable environments. To overcome these issues, we propose CA-MAPPO, a multi-agent reinforcement learning framework that integrates convolutional attention with adaptive reward optimization. The adaptive reward mechanism dynamically adjusts exploration incentives according to residual targets and task progress, thereby mitigating late-stage redundancy. Meanwhile, the convolutional attention module combines convolutional layers with spatial attention to emphasize target-dense or uncertain regions while preserving global context. The extracted global features are fused with local observations to guide decentralized policy learning across UAVs. Extensive experiments across diverse scenarios demonstrate that CA-MAPPO significantly outperforms state-of-the-art baselines in target discovery, coverage, and task completion, while exhibiting strong scalability and generalization under varying environmental complexities.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.