Explainable deep reinforcement learning for three-axis robust attitude control of flying wing aircraft

Owing to the absence of horizontal and vertical tails and the reliance on drag rudders for yaw control, flying wing aircraft exhibit distinct aerodynamic coupling and three-axis motion-coupling characteristics. Consequently, the flight control law design method must simultaneously address external turbulence, model parameter uncertainty, and strong coupling. In this paper, a neural network (NN)-based robust three-axis attitude control method utilizing the proximal policy optimization (PPO) algorithm is proposed. First, the state space of the agent is augmented by incorporating coupling states to eliminate the three-axis coupling and time-lagged states, which compensates for the servo-actuator delay. Second, a hierarchical-reward shaping strategy comprising subgoals for tracking accuracy, stability, and actuator-rate limits induces a piecewise control mechanism, thereby increasing robustness. Third, a progressive three-axis assisted training scheme is introduced to sequentially train the longitudinal, lateral, and directional networks and ensure accurate tracking and effective decoupling. Finally, kernel SHapley Additive exPlanations (Kernel SHAP) is employed to visualize the attributions of the state features. The time histories of the Shapley attribution, interpreted within the flight dynamics framework, elucidate the control mechanism, the compensation mechanism, and the sources of robustness for the NN controllers. The simulation results indicate that the proposed NN controllers achieve rapid and highly precise command tracking while maintaining stable and robust performance, even under severe atmospheric turbulence and significant model parameter perturbations.