A Feature‐Decoupled and Gated‐Interaction‐Enhanced Deep Reinforcement Learning for Path‐Following of Large‐Inertia Vessels

ABSTRACT Path‐following control for large‐inertia surface vessels remains challenging due to slow yaw dynamics and environmental uncertainties. This paper proposes a hybrid framework integrating Line‐of‐Sight (LOS) guidance, drift‐angle compensation, and an Adaptive Gated Interaction Twin Delayed Deep Deterministic Policy Gradient (AGI‐TD3) controller. The primary innovation lies in the hierarchical feature‐processing architecture of the AGI‐TD3 agent. By implementing decoupled encoding for heterogeneous state‐action subspaces, the network isolates interference from disparate physical semantics. Furthermore, an adaptive gated interaction mechanism is introduced to selectively modulate information flows, reinforcing the causal relationship between heading errors and control actions. Simulation results and full‐scale experiments on a 40 m‐class vessel demonstrate that the proposed method significantly improves path following performance and robustness under realistic operating conditions.