Reinforcement Learning Control for AUV-Tow-Load Systems Under Safety-Critical Condition

In this paper, an safety-critical reinforcement learning control solution is developed for cable subsystems of autonomous underwater vehicle-tow-load systems with unmodeled dynamics. To achieve predetermined tracking performance guarantees, a nonlinear mapping is designed to transform the system dynamics into a new one complemented by predefined boundary functions. Based on this, a system identifier is constructed to provide online estimates for the controller, which releases the demand of exact modeling of cable subsystems. By employing experience reply technique, an update law is designed for a single critic neural network, where activation functions involved with system dynamics and reference trajectory is employed, allowing for an enhanced robustness of the closed loop using released persistent of excitation (PE) condition. Numerical simulation is provided to validate the effectiveness of the proposed control algorithm.