Improved D3QN with graph augmentation for enhanced multi-UAV cooperative path planning in urban environments

Path planning of multi-unmanned aerial vehicles (multi-UAVs) in complex urban environments is a challenging task, which suffers from low autonomous decision-making capability and execution efficiency. Traditional methods are struggling to cope with the increasing task demands. Deep reinforcement learning (DRL) offers a more flexible and efficient solution for path planning through offline training and online reasoning, which is expected to better adapt to the complexity of urban environments. Thus, a multi-UAV path planning method based on improved Dueling Double DQN (D3QN) and graph augmentation techniques is proposed in this paper, which effectively addresses the collaborative path planning problem in multi-UAV systems by combining DRL and Graph Neural Networks (GNN). In the proposed method, the D3QN algorithm is adopted to enhance the interaction between UAVs and the environment using a Graph Convolutional Network (GCN) communication model based on historical trajectory information. Furthermore, a task-driven dynamic priority adjustment mechanism is proposed to balance the relationship between UAV paths and energy consumption. Meanwhile, a dynamic incentive adjustment mechanism combining complex real-time environment information and UAV states is proposed to enhance the performance and adaptability of the path planning algorithm. Simulation experiments were conducted to validate the algorithm and compare it with the general DRL methods such as DQN and D3QN, and the results demonstrate the effectiveness and superiority of the proposed method.