Reinforcement Learning-Based Prescribed Performance Tracking Control for Autonomous Vehicles With Input Saturation and State Constraints

This paper proposes a reinforcement learning-based prescribed performance control framework for the path tracking of autonomous vehicles. The Actor-Critic neural networks architecture is introduced in the design of the path tracking controller, which provides a solution for accurately approximating unknown nonlinear terms, obviating the requirement for a priori knowledge of tire cornering stiffness. Compared with traditional prescribed performance methods, the proposed control scheme can quantitatively adjust overshoot, settling time, and steady-state accuracy by integrating a set of boundary functions. Furthermore, an auxiliary function is designed to regulate the tracking error within a prescribed set regardless of the initial conditions, thereby removing their inherent feasibility constraints. To guarantee operational safety, a unified barrier function-based controller is designed to ensure state constraints (longitudinal/lateral velocity and yaw rate) are not violated, which eliminates the feasibility conditions imposed on virtual control laws. Meanwhile, a first-order filter is introduced to avoid the complexity explosion involved in the design of adaptive backstepping control strategy. Simulation results illustrate the effectiveness and superior performance of the proposed control strategy.