Reinforcement Learning for Joint Radar and Communication Enabled Multi-UAV Cooperative Dynamic Target Detection and Tracking

Joint radar and communication (JRC)-enabled multiple unmanned aerial vehicles (UAVs) are gaining attention due to their potential to improve spectrum utilization and provide enhanced flexibility in target detection and tracking. However, most existing studies focus solely on a single UAV or static environment, neglecting the benefits of UAV cooperation and the complexity of the dynamic environment. In this paper, we consider a JRCenabled multi-UAV cooperative dynamic target detection and tracking scenario, in which each UAV simultaneously performs both detection and communication. Given the constraints of JRC power and spectrum, a trade-off between detection and communication performance is required such that sensing power, communication channel allocation, and UAV trajectories should be jointly optimized. Considering the challenge of integer and non-convex constraints, we propose a reinforcement learningbased cooperative target detection and tracking (RLCTDT) algorithm in which the optimal strategy of each UAV is iteratively derived through the interaction between UAV swarm and environment. In particular, the reward function, mainly consisting of sensing rewards and guidance rewards, is newly designed to guide UAVs in exploring and exploiting environment information efficiently. Recurrent neural networks (RNNs) are introduced into RLCTDT to abstract history observations for improving estimation accuracy of state value and utilize history actions for better decision-making. Numerical experiments indicate that our approach achieves a 16% higher stabilized system reward than the best baseline, RLCTDT-KL, which reflects improvements in target tracking and environment sensing