Learning-Based Model Predictive Control for Cooperative Spacecraft Swarm Reconfiguration

This article presents a reference-free and decentralized algorithm for spacecraft swarm reconfiguration that simultaneously ensures collision avoidance and fuel efficiency. The objective is to transfer a large number of capability-constrained spacecraft from arbitrary initial formations to a desired target configuration. To enhance autonomy and real-time responsiveness, a learning-based model predictive control (MPC) framework is developed, in which reinforcement learning (RL) is integrated into the MPC structure. By leveraging the full-horizon maneuver schedule generated by the RL agent, the MPC algorithm can compute maneuver strategies directly toward the target configuration at the terminal time, without relying on a predefined reference trajectory. In contrast to prior works that employ continuous-thrust control with a fixed discretization interval $\Delta t$, the proposed method utilizes impulsive maneuvers with an adaptively adjusted $\Delta t$. This design significantly improves the flexibility of the algorithm and reduces the online computation frequency when handling large-scale reconfiguration scenarios. To address the computational complexity of collision-avoidance constraints in large-scale swarms, a fast-constructible collision-avoidance region is developed based on reachable sets. The recursive feasibility and asymptotic stability of the overall system are guaranteed through a Lyapunov-based analysis incorporating artificial constraints. Finally, the proposed approach is validated through extensive Monte Carlo simulations and benchmark comparisons with existing state-of-the-art methods.