Deep reinforcement learning-based fast hybrid planning for multi-type on-orbit service under mass variation dynamics

This study proposes a deep reinforcement learning-based fast planning method for the hybrid planning of multi-type tasks of GEO spacecraft under ”service station-service satellite” on-orbit service mode, considering mass variation caused by service. The hierarchical architecture achieves joint fast estimation of transfer duration and service sequence: the lower layer solves orbital transfer planning via supervised learning, introducing residual modules for high-precision prediction of Lambert optimal transfer duration to ensure computational efficiency for upper-layer planning; the upper layer handles service sequence planning with deep reinforcement learning, proposing a categorical-encoded Transformer attention model that ensures target position order invariance while satisfying start-end position constraints. To mitigate the impacts of hard constraints on model training, this study simultaneously proposes a bidirectional calculation of fuel consumption. Finally, simulations show that the method reduces the feasible solution planning time from the order of hours to the order of seconds for 10-target missions, significantly lowering computational resource consumption.