Multi-UAV Path Planning Based on Multi-Agent Deep Reinforcement Learning

As the widespread adoption of unmanned aerial vehicles (UAVs) in various fields, the UAVs path planning serves as a prerequisite for ensuring the safety and efficiency of the flight. Existing methods rely on grid airspace and detailed modeling for a specific environment, limiting portability and inflexibility responding to diverse tasks and environments. They utilize discrete actions space, neglecting the dynamics characteristics of UAVs, which leads to planning relatively rough paths. Furthermore, the current Studies lack sufficient consideration of reliance on spatiotemporal data and disregard multi-UAV collisions and unstable training environments, which hinder the realization of collaboration among UAVs. In response to these challenges, this paper propose an Enhanced Multi-Agent Twin Delayed Deep Deterministic Policy Gradient (Enhanced-MATD3). We model a state space according to the 3D scene and design a continuous action space. Further, we devise reward function based on obstacle and collision avoidance principles. By enhancing the actor and critic network from both temporal and spatial dimensions, we broaden UAVs perception range and sharpen their responsiveness to dynamic environmental changes. Our work use different scenarios during the training process, and the experimental results demonstrate the advantages of our algorithm in terms of average path length, planning time, and convergence speed over the baseline.