An Improved DRL Algorithm for Reliable Path Planning

In recent years, deep reinforcement learning (DRL) is an innovative way applied to target tracking of unmanned aerial vehicle (UAV). However, the instability of neural networks and the difficulty of designing reward functions have hindered its application in practice. To end these, this article proposes a novel method called LVFSAC, which combines the Lyapunov-vector-field (LVF) and soft actor-critic (SAC) of DRL, improving the stability of DRL and the efficiency of designing sophisticated reward functions. First of all, a UAV autonomous motion control (AMC) system is designed based on the framework of offline training- online decision-making, and the maximum entropy DRL soft actor-critic (SAC) is used for global path planning to guide real-time motion decisions. Secondly, to ensure the closed-loop stability of the AMC system, the reference paths are derived by using LVF theory based on the UAV situational information, which is fed back to adjust system training. Finally, the simulation platform with generalized scenarios is designed to validate the proposed method based on the UAV autonomous motion control application. The results show that our method not only achieves better stability during training but also obtains a higher success rate while ensuring path safety and reliability.