Sensing-Aware Cooperative Multi-UAV Search with Reinforcement Learning

This paper presents a novel reinforcement learning framework for multi-UAV cooperative search missions with sensing imperfections. We introduce a scalable multi-agent environment that effectively models sensor uncertainty while enabling decentralized information merging between UAVs. Our reward mechanism uses reward shaping for adapting to the information flow through the mission. Using Proximal Policy Optimization (PPO), we create a training framework that achieves consistent performance across varying numbers of UAVs and targets. Our results demonstrate that the proposed method maintains high success rates even with significant sensing uncertainties, with performance improving as the number of cooperating UAVs increases. The framework shows resilience against the unlearning problem commonly encountered in uncertain environments and uses a minimal observation space. Our framework provides a foundation for deploying reliable multi-UAV search systems in real-world scenarios where sensor reliability cannot be guaranteed and computational resources are scarce.