Multi-Model Safe Neuro-Optimal Output Tracking Control of Autonomous Surface Vessels with Explainable AI

This paper presents a safety-aware deep reinforcement learning (DRL)-based trajectory tracking control of autonomous surface vessels (ASVs). A multilayer neural network (MNN) observer estimates the ASV's state and uncertain dynamics. By utilizing the estimate state vector from the observer, a safety-aware DRL-based optimal policy is formulated using control barrier function (CBF) and Karush-Kuhn-Tucker (KKT) conditions. An actor-critic MNN with singular value decomposition (SVD)-based update mitigates vanishing gradients. To enhance adaptability, an online safe lifelong learning (SLL) scheme counters catastrophic forgetting across varying ASV dynamics. The Shapley Additive Explanations (SHAP) method identifies key features influencing the control policy. Simulations on an underactuated ASV show that SLL-based control reduces cumulative costs by 17 % and RMS error by 32 % compared to a baseline without SLL.