Hierarchical Game-Theoretic and Risk-Aware Predictive Control Framework for Resilient Multi-UAV Cooperative Combat
Mingjun Tang, Renwen Chen, Junwu Zhu, Yuqi Jiang, Ran Zhang, Yu Juan Cui, Bin Zhang · IEEE Access · 2026
Multi-UAV cooperative combat systems are currently striving to close the loop from macro-tactical gaming to micro-physical execution in extreme A2/AD environments. However, they face the harsh reality of cross-timescale coupling and multi-source uncertainty. Existing divide-and-conquer strategies often fracture high-level decision-making from low-level control. Consequently, theoretical Nash equilibria easily degenerate into execution-level routs when confronted with packet loss, non-cooperative information hiding, or sudden aerodynamic disturbances. Addressing this critical pain point, we propose H-GPC (Hierarchical Game-Theoretic Predictive Control), a full-stack cooperative framework. Far from a simple stacking of algorithms, H-GPC reconstructs the agent's cognition-execution pathway. At the macro-strategic layer, we model the command hub and the swarm as a Stackelberg game, designing optimal contract menus that satisfy incentive compatibility and individual rationality. This successfully resolves resource mismatch under information asymmetry, maintaining Pareto-optimal social welfare even under severe communication conditions with 20% packet loss. At the meso-tactical layer, we introduce distributional reinforcement learning based on Implicit Quantile Networks (IQN), utilizing the CVaR criterion to dynamically adjust risk appetite. This breaks through the survival bottleneck of traditional expectation-maximization strategies in life-or-death scenarios. At the micro-control layer, we design a disturbance observer with Prescribed-Time convergence characteristics and a resilient reconfiguration mechanism based on virtual state consensus, building a final line of defense against strong nonlinear disturbances and FDI attacks. High-fidelity simulations and embedded deployment experiments on the NVIDIA Jetson Xavier NX demonstrate that H-GPC not only significantly outperforms baselines like MAPPO and CBBA in mission survival rate and loss ratio within high-risk environments but also reduces computational load by over 60% through a bi-level self-triggered mechanism. This proves the framework's engineering feasibility under strict 50Hz real-time constraints.