Auto-Docking for Leisure Boats Using Deep Reinforcement Learning With Action Masking

Abstract The demand for auto-docking technology is growing to alleviate the burden on inexperienced operators. Previous guidance law-based approaches have primarily focused on ensuring robustness in specific scenarios and marina layouts. This study investigates the enhancement of docking performance using deep reinforcement learning across various scenarios, including different initial positions and heading angles, dock shapes, and docking modes. The study integrates the Proximal Policy Optimization (PPO) with action masking algorithm to prevent invalid actions that might cause collision or move the boat away from the target. Comparative analyses reveal that PPO with action masking outperforms traditional line-of-sight (LOS) guidance laws, demonstrating superior adaptability to changing initial positions and headings. The results suggest that deep reinforcement learning with action masking can enhance safety, efficiency, and reliability in auto-docking under diverse maritime conditions.