一种时限可达域引导的航天器博弈决策学习方法

To address the issues of real-time decision-making limitations in impulse-thrust-driven spacecraft pursuit-evasion games and the incapability of traditional reward functions to adapt to long-distance high-dynamic adversarial learning environments, this paper investigates intelligent maneuver decision-making and fuel optimization for spacecraft game confrontations. Firstly, the orbital game dynamics and maneuver constraint model are established. Secondly, a time-constrained single-impulse reachable domain solving method for spacecraft is proposed, and neural networks are integrated to perform quantitative fitting of orbital danger zones. Furthermore, a hierarchical reinforcement learning control framework is designed based on a distributed system architecture, and the proximal policy optimization (PPO) algorithm is employed to carry out red-blue adversarial learning training. Finally, the proposed maneuver strategies are validated. Simulation results demonstrate that in the two-body dynamics orbital game scenario, the danger zone strategy reduces average fuel consumption by 33.81%, and the game strategies improve the hit rate by an average of 38.41% compared with traditional strategies.