Large language model-augmented model predictive control for marine vessels in uncertain marine environments

This paper presents a novel control framework that integrates Large Language Models (LLMs) with Model Predictive Control (MPC) to improve ship navigation performance under uncertain and variable marine conditions. The LLM is fine-tuned using synthetic datasets generated from a 2D ship model with injected environmental disturbances, enabling it to learn residual dynamics not captured by the nominal model. During operation, the LLM estimates real-time model error as a function of the current vessel state, control input, and exogenous factors such as wave height and wind speed. This estimated uncertainty is incorporated into the MPC prediction model, resulting in enhanced state forecasting and more robust control decisions. Simulation results demonstrate that the LLM-MPC framework significantly reduces steady-state tracking errors, improves convergence rates, and smoothens control actions compared to MPC only without LLM estimation. This approach offers a scalable, data-efficient solution for enhancing autonomy in intelligent marine systems operating in complex ocean environments.