Scalable Pursuit–Evasion Game for Multi-Fixed-Wing UAV Based on Dynamic Target Assignment and Hierarchical Reinforcement Learning

The unmanned aerial vehicle (UAV) pursuit–evasion game is the fundamental framework for promoting autonomous decision-making and collaborative control of multi-UAV systems. Faced with the limitations of current deep reinforcement learning methods in terms of transferability and generalization for scalable multi-fixed-wing UAV pursuit–evasion game scenarios, this paper proposes a hierarchical collaborative pursuit–evasion game framework based on target allocation and hierarchical reinforcement learning. The framework comprises three layers: target allocation layer, maneuver decision-making layer, and flight control layer. The target allocation layer employs a dynamic target assignment method based on a dynamic value adjustment mechanism, decomposing the multi-vs.-multi pursuit–evasion game into several one-vs.-one confrontations. The maneuver decision-making layer utilizes a maneuver decision-making method based on trajectory prediction and hierarchical reinforcement learning to generate adversarial maneuver commands. The flight control layer adopts a stable gradient-assisted reinforcement learning flight controller to ensure stable UAV flight. Comparisons with other algorithms across 3V3, 6V6, 9V9, and 12V12 scenarios demonstrate that the proposed method achieves high win rates in diverse game scales. The comparison results also demonstrate the advantages of the framework proposed in this paper in terms of training efficiency and large-scale scalability.