Intention-Aware Model Predictive Control for Orbital Pursuit-Evasion Games

Accurately understanding the maneuver intentions of non-cooperative spacecraft in orbital pursuit-evasion games remains challenging, as most existing approaches rely on configuration-based intention classification or assume a predefined control law, limiting their ability to predict future behaviors under unknown policies. This paper proposes a learning-based intention-aware trajectory prediction and MPC-based maneuvering strategy for orbital pursuit-evasion games with impulsive maneuvers. A diverse maneuvering dataset is first generated through self-play reinforcement learning, which enables the modeling of realistic and generalizable adversarial policies. An Informer-based neural predictor is then proposed to forecast long-horizon 3D trajectories of the evader spacecraft, implicitly capturing its maneuver intentions while achieving accurate, real time prediction with low computational load. The predicted trajectories are subsequently incorporated into an MPC-based maneuvering strategy, which allows responsive motion planning against the evolving behavior of the evader. The proposed strategy advances intention understanding from configuration level reasoning to trajectory-level, intention-aware prediction, and is effective for targets following unknown but fixed policies. High-fidelity simulation results across multiple pursuit-evasion scenarios verify the accuracy, generalizability, and practical feasibility of the proposed strategy.