Orbital Interception Pursuit Strategy for Random Evasion Using Deep Reinforcement Learning

Aiming at the interception problem of noncooperative evader spacecraft adopting random maneuver strategy in one-to-one orbital pursuit–evasion problem, an interception strategy with decision-making training mechanism for the pursuer based on deep reinforcement learning is proposed. Its core purpose is to improve the success rate of interception in the environment with high uncertainty. First of all, a multi-impulse orbit transfer model of pursuer and evader is established, and a modular deep reinforcement learning training method is built. Second, an effective reward mechanism is proposed to train the pursuer to choose the impulse direction and impulse interval of the orbit transfer and to learn the successful interception strategy with the optimal fuel and time. Finally, with the evader taking a random maneuver decision in each episode of training, the trained decision-making strategy is applied to the pursuer, the corresponding interception success rate of which is further analyzed. The results show that the pursuer trained can obtain universal and variable interception strategy. In each round of pursuit–evasion, with random maneuver strategy of the evader, the pursuer can adopt similar optimal decisions to deal with high-dimensional environments and thoroughly random state space, maintaining high interception success rate.