When a policy trained with deep reinforcement learning (DRL) in simulation is deployed in the real world, its performance often deteriorates due to the Sim2Real gap. This study addresses this problem for Autonomous Surface Vessels (ASVs) by developing a robust collision-avoidance framework. We integrate a MATLAB-based ship dynamics model with ROS and Gazebo, and employ the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm. To enhance robustness and generalization, we combine domain randomization and curriculum learning. As a result, the trained agent consistently achieved a high success rate of over 90% in unseen environments, significantly outperforming a baseline TD3 agent and a conventional PID controller. This demonstrates that the proposed Sim2Real methods are highly effective for creating robust control policies for ASVs. For future work, we plan to validate the learned policy through real-world experiments.
Han et al. (Mon,) studied this question.