A Deep Reinforcement Learning-Based Multi-UAV Global Path Planning

To address the problem of global path planning for multiple unmanned aerial vehicles (UAVs), the Hindsight Prioritized Experience Replay-Multi-Agent Deep Deterministic Policy Gradient (HPER-MADDPG) algorithm is proposed. Firstly, the kinematic model of UAVs and the three-dimensional environment model are established and transformed into a partially observable Markov game problem. Secondly, to address the issue of low training efficiency of random experience replay, the HPER mechanism is designed based on MADDPG. After the UAVs complete the interaction with the environment, all experience tuples are stored in the experience buffer pool as trajectory sequences, and the sampling priorities are calculated. If a sampled experience is unsuccessful, it is mapped into a new experience for training. Experimental results show that this algorithm achieves the global path planning of multiple UAVs under multiple constraints, and the maximum time difference in arrival among the UAVs is less than 0.2 seconds.