DSAC-T-Based Dynamic Positioning for a Vessel Actuated Only by Dual Stern Azimuth Thrusters Under Directional Wind

Deep reinforcement learning (DRL) enables model-free adaptive control for marine dynamic positioning (DP) under changing disturbances. However, its application remains limited for vessels with strict mechanical constraints, such as the common stern-only dual-azimuth configuration. This paper applies DSAC-T (Distributional Soft Actor-Critic with Three Refinements), a distributional off-policy deep reinforcement learning algorithm, to DP of an 83 m offshore supply vessel under this stern-only configuration. DSAC-T is compared with proximal policy optimization (PPO) and soft actor–critic (SAC) using a shared reward and observation design and an equal simulation budget. All DRL controllers, together with a fixed-gain proportional–derivative (PD) baseline, are evaluated in simulations covering eight wind directions at two wind speeds, together with a calm-water case. DSAC-T achieves the lowest aggregate position-keeping error among the three learning-based methods with lower propulsor-command activity than PPO. We also find that a tighter Gaussian position tolerance in the reward function substantially reduces steady-state position error with a small increase in propulsor-command activity, identifying this tolerance as an influential design parameter for stern-only DP.