Gradient-enhanced PINN-based hybrid attitude control for reentry vehicle under uncertainties

Reentry vehicle attitude control faces critical challenges from model uncertainties, strong nonlinear dynamics, and fast response requirements. To address the limitations of traditional methods, which suffer from strong parameter sensitivity and lack of optimality guarantees in physical models, this paper combines physical modeling with data-driven approaches. It proposes a hybrid attitude tracking control framework based on gradient-enhanced physical information neural network (PINN). Gradient-enhanced PINN employs a dual-branch neural network architecture to explicitly output both the value function and its gradient, thereby eliminating costly numerical differentiation during rapid inference. Concurrently, a hybrid strategy is established in network training that aligns the physical information residual from the Hamilton-Jacobi-Isaacs equation with the gradient supervision loss, enhancing computational efficiency and convergence speed. Furthermore, the explicit gradient outputs are integrated into the dual-loop attitude control structure, driving stable outer-loop attitude tracking and inner-loop angular velocity stabilization without computational overhead. Considering practical multi-actuator coordination challenges, a gradient-enhanced torque allocation algorithm is proposed, dynamically balancing control authority between aerodynamic surfaces and thrust vectoring using state sensitivity information. Extensive comparative simulations and Monte Carlo robustness analysis validate the framework’s superior performance over existing methods in tracking accuracy, computational efficiency, and adaptive resource allocation under uncertainties.