Editorial: Navigation and Perception for Autonomous Surface Vessels

Fausto Ferreira, Alberto Quattrini Li, Ørnulf Jan Rødseth

Abstract: To generalize obstacle avoidance from ground mobile robots to marine vehicles and to bridge the gap between simulation and reality, "Robust ASV Navigation Through Ground to Water Cross-Domain Deep Reinforcement Learning" (Lambert et al, 2021) proposes a navigation method based on a deep reinforcement learning framework for high-level control, integrated with low-level controllers specific to the vehicle. The paper demonstrates the cross-domain generalizability of the proposed method by training a Deep Q Network for a simulated Autonomous Ground Vehicle (AGV) with navigation tasks. It then successfully implements and tests the proposed DRL method on a real AGV as well as an Autonomous Surface Vehicle (ASV) in different scenarios without any re-training. While the paper can be expanded to handle more complex scenarios, e.g., with dynamic obstacles, the paper provides a good methodology for training reinforcement-learning based approaches in domains that are comparatively easier and lower cost for deployments compared to challenging target domains, as the marine one.Deep RL are not the only techniques that can be used for autonomous navigation. The complexity of the issue requires a deep study of the field. Indeed, collision avoidance and safe navigation in environments where other crewed or uncrewed ships sail, are perhaps the biggest challenges for safe operation of autonomous surface vessels. In particular, the coexistence with other crewed ships require that autonomous ve...