Abstract
With the rapid development of the low-altitude economy, there is a growing demand for large-scale uncrewed aerial vehicle (UAV) swarms to support diverse applications. These low-altitude operational scenarios require precise and efficient UAV swarm coordination, making autonomous flight control a key challenge. In this work, we investigate a distributed UAV swarm flight control problem in which each UAV autonomously makes decisions based on partial observations of neighboring UAVs and the surrounding environment. This problem involves multi-objective optimization, as the goal is to simultaneously minimize the cost of UAV control and maximize the speed of UAV service activation. To address the inefficiency of traditional reinforcement learning for this multi-objective problem when confronted with fluctuating low-altitude environmental conditions, we propose a distributed autonomous control method based on evolutionary multi-agent multi-objective reinforcement learning (EMAMORL). This approach integrates distributed intelligence with multi-objective reinforcement learning to generate numerous policies for distinct preferences for different objectives in a single iteration while reducing communication and control overheads. During the online inference phase, agents use these trained policies for real-time decision-making. Simulation results show that EMAMORL outperforms baseline evolutionary algorithms in generating high-quality Pareto-optimal policies and maintains robust performance in dynamic low-altitude environments with wind disturbances.