Intelligent Reentry Guidance with Dynamic No-Fly Zones Based on Deep Reinforcement Learning

Aimed at avoiding multiple dynamic no-fly zones and satisfying path constraints and terminal constraints in the reentry process of hypersonic glide vehicles, intelligent reentry guidance based on deep reinforcement learning is developed. Firstly, the guidance is decoupled as longitudinal guidance and lateral guidance. The lateral guidance provides the sign of the bank angle to adjust the heading direction while the longitudinal guidance outputs the magnitude of the bank angle through the artificial intelligence interface. Then, the reentry guidance simulation is mapped to a Markov Decision Process, in which the essential elements including state, action, and reward are defined or designed adaptively. Finally, the policy neural network is trained by the twin delayed deep deterministic policy gradient (TD3) algorithm. By selecting proper hyperparameters and network architecture, the policy neural network is able to converge. Simulations imply that under the influence of dynamic no-fly zones, initial state errors, and kinds of online dispersion, the proposed guidance can avoid all the no-fly zones and reach the target accurately with all the satisfied path constraints.