Abstract
To address the high complexity and varying scale of swarm-to-swarm orbital pursuit–evasion games, this paper proposes a hierarchical decision-making framework consisting of a progressive spatial-coverage target assignment method and a scalable strategy computation method. At the global decision level, the proposed target assignment method successively assigns pursuers to evaders by jointly considering the distance cost and spatial coverage benefit, thereby decomposing the swarm-to-swarm game into multiple multi-to-one local pursuit–evasion games. At the local decision level, the Multi-Agent Twin Delayed Deep Deterministic Policy Gradient (MATD3) algorithm is adopted for strategy training. A cross-attention mechanism is further incorporated to encode the states of varying numbers of pursuers into a unified representation, facilitating strategy computation across different game scales. The relationship between attention weights and relative distances is analyzed to examine the spatial information captured by the cross-attention mechanism. Extensive simulations are conducted in three representative swarm-to-swarm scenarios (5-vs-2, 8-vs-3, and 10-vs-4), with the evaders governed by a fixed-parameter artificial potential field (APF) model. The results show that the proposed method achieves pursuit success rates exceeding 90% across the tested scenarios, demonstrating its effectiveness and scalability under the considered conditions.