Safe-Reinforcement-Learning-Aided Lightweight Cooperation for Multi-AAV Data Collection in Random WSN

Unmanned aerial vehicle (UAV) data collection problems in wireless sensor networks (WSNs) under random and uncertain environments are critical challenging, due to massive burden of real-time communication among UAVs for aligning with observation and state information, ect. For this sake, this work investigates the UAV cooperation problem of WSN data collection by jointly maximizing collected data amount while minimizing cooperation cost. We formulate the problem as a constrained partially observable Markov decision process (CPO-MDP), which stimulates the design of a novel safe reinforcement learning aided lightweight cooperation (SRL-LC) framework for multi-UAV data collection. Speficially, the self-conscious cooperative communication scheme is developed to assist the optimization of the UAV trajectory decision making. Additionally, a safety module embedded in the decision network integrates a relaxed artificial potential field (APF) algorithm, enabling UAVs to maintain safety distance constraints during training. Simulation results demonstrate that the proposed SRL-LC framework achieves data collection performance comparable to the full-cooperation scheme for various settings of prior information, while reducing communication cost by approximately 85%. Moreover, the SRL-LC framework ensures zero violations of safety constraints throughout the training process.