Cooperative Collision Avoidance Method for Multi-UAV Systems Based on MAD3QN

To address the dynamic coupling problem in cooperative collision avoidance among multiple unmanned aerial vehicles (UAVs) operating in low-altitude environments, where individual decision-making is simultaneously influenced by spatial constraints from static obstacles and multi-source disturbances from other dynamic agents, a cooperative decision-making method is developed based on a multi-agent dueling double deep Q-network (MAD3QN). A local observation model based on sector-based detection is employed, in which the states of nearby dynamic targets and static obstacles are uniformly encoded into observation vectors with a fixed structure, resulting in a consistent observation space for each agent. A multi-factor composite reward function is designed, and combined with a distributed execution framework and shared policy architecture to enhance the robustness of the policy and accelerate training convergence. Simulation results across various complex scenarios demonstrate that the proposed method consistently achieves a collision avoidance success rate exceeding 98%, significantly outperforming existing approaches. This research provides an effective technical solution and theoretical foundation for the safe and coordinated operation of multi-UAV systems in low-altitude airspace.