Deep reinforcement learning based trajectory real-time planning for hypersonic gliding vehicles

To overcome the shortcomings of traditional NLP methods for trajectory planning problems, an intelligent trajectory real-time planning method is designed for hypersonic gliding vehicles (HGVs), which is composed of two stages: the agent training stage and the real-time trajectory generation stage. During the training stage, the HGV model is considered as an agent, and an environment containing flight information and relative information is constructed. Given the trajectory planning problem possessing continuous state-action space, the twin delayed deep deterministic policy gradient (TD3) is employed, based on which the HGV agent is trained in the environment. To match the real flight environment for HGVs, the process and terminal constraints are taken into consideration, such as the limit of dynamic pressure, overload, and the terminal miss distance, etc. The reward shaping technique is adopted to tackle the multiple constraints. The second stage is the real-time trajectory generation stage, during which a trajectory satisfying the multiple constraints is generated online by the TD3-based method. The simulation results verify the effectiveness of the proposed method.