LARL: LOS-guided Attention-Residual Learning for robust trajectory tracking of underactuated USVs

Accurate trajectory tracking of underactuated unmanned surface vehicles (USVs) is challenging because of nonlinear coupled dynamics, unavailable lateral actuation, model uncertainty, and environmental disturbances. This paper proposes a LOS-guided Attention-Residual Learning (LARL) framework that integrates line-of-sight (LOS) guidance with deep reinforcement learning for robust USV trajectory tracking. LOS-generated desired-heading and tracking-error information is incorporated into a structured state–reward formulation, while the 13-dimensional observation is organized into five physically meaningful tokens and processed by a multi-head attention-residual actor–critic controller based on deep deterministic policy gradient (DDPG). The framework is evaluated in simulation under disturbance-free, wind–wave–current, dynamic-parameter-perturbation, and actuator-dynamics conditions, with PID, sliding mode control, DDPG, and TD3 as baselines. Across 12 trajectory–scenario combinations, LARL achieves the lowest RMSE in 10 cases and the lowest MAE in 11 cases, and shows consistent performance across five independent training seeds. Additional tests on unseen ellipse, racetrack, and Gerono trajectories further demonstrate generalization to new path geometries. Field experiments on a self-developed USV verify practical deployability: in rectangular trajectory tracking, LARL reduces MAE by 14.21% and mean absolute rudder rate by 26.55% relative to PID, while maintaining comparable overall tracking performance. These results demonstrate the effectiveness and practical potential of the proposed structured LOS-informed attention-residual learning framework.