Adaptive optimal trajectory tracking control of underactuated hovercraft based on safety constraint reinforcement learning

This paper proposes a Constrained Reinforcement Learning Optimal Tracking Control (CRLOTC) strategy to address the trajectory tracking problem for underactuated hovercrafts subject to external disturbances and unknown system dynamics. A novel constraint function is designed to restrict the yaw rate and the surge speed within a safe range, thereby ensuring navigation safety. An adaptive neural network observer is employed to estimate system uncertainties, enhancing disturbance rejection and resolving the differential explosion issue inherent in dynamic surface control. To optimize energy consumption, a synergistic framework integrating backstepping control with an Actor-Critic reinforcement learning architecture is developed to solve the Hamilton-Jacobi-Bellman equation online, yielding an optimal control policy at each step. Theoretical analysis demonstrates the semi-globally uniform ultimate boundedness of all closed-loop signals and verifies the optimality of the control policy. Numerical simulations confirm the strategy's superior performance in tracking accuracy and safety assurance compared to existing methods.