Transformer-Enhanced Adaptive NMPC for Underactuated Autonomous Surface Vehicles

Abstract

The increasing demand for autonomous vehicles operating in dynamic and uncertain environments highlights the need for robust control strategies that can model and anticipate external disturbances. This paper presents a comprehensive pipeline for six-degree-of-freedom disturbance forecasting and adaptive nonlinear model predictive control (NMPC) tailored to an Autonomous Surface Vehicle (ASV), with a primary focus on the application of Transformer-based architectures in the marine domain. Departing from traditional sequence learning approaches, we evaluate the performance of transformer architectures by benchmarking them against established state-of-the-art models such as Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTMs), and Temporal Convolutional Networks (TCNs). These horizon-ahead disturbance estimates are integrated as feed-forward terms into an NMPC framework, applied over the prediction window while maintaining real-time feasibility. Validated in simulation and on a real-world ASV, our results demonstrate that the Transformer-based approach enhances trajectory-tracking accuracy and robustness, outperforming traditional architectures in compensating for environmental perturbations while maintaining acceptable computational requirements for real-time deployment.