Incremental Safe Reinforcement Learning for Flight Control with Control Barrier Function

In this paper, an incremental safe reinforcement learning algorithm, namely Incremental Safe Dual Heuristic Programming (ISDHP), is proposed for flight control systems with state constraints. Firstly, a recursive least squares (RLS) method is employed for online identification of incremental system dynamics to achieve model-free real-time adaptation without offline training. Secondly, the cost function is augmented with the control barrier function (CBF) to ensure that state constraints are satisfied. Thirdly, an actor-critic network structure with experience replay is designed to approximate the optimal control policy, where network weights are updated via gradient descent to minimize the temporal difference error. Finally, numerical simulations on a aircraft longitudinal dynamic model validate that the proposed ISDHP algorithm achieves effective tracking of the reference command while strictly confining the angle of attack within the safe range, demonstrating its superiority in both optimality and safety.