Coverage Path Planning for Unmanned Surface Vehicles in Unknown Environments Based on Reinforcement Learning

In marine resource exploration and development, unmanned surface vehicles (USVs) often need to plan a safe and collision-free exploration path. When the environment is unknown, offline planning methods cannot provide an optimal solution; thus, online path planning must be conducted while exploring the environmental map. In this paper, we investigated how reinforcement learning can address this issue. Considering USVs with sensing capabilities, we constructed coverage and obstacle maps using simulated 2D marine radar data to update the known environment state. Furthermore, we proposed a method for constructing multi-scale frontier maps based on line-of-sight angles using egocentric maps, along with a new set of reward functions to facilitate coverage path planning. Through extensive experiments, we demonstrated that our method significantly improves coverage while reducing redundant path generation.