Autonomous Exploration with Virtual Map using Unmanned Surface Vehicles

Autonomous exploration on the sea surface poses significant challenges, particularly due to the lack of quantitative criteria to guide USVs in establishing exploration objectives that simultaneously maximize valuable environmental data collection and ensure map stability. To address this limitation, this paper implements an exploration strategy based on expectation maximization (EM). In this approach, unobserved potential landmarks are modeled as latent variables, and their uncertainties are predicted and systematically managed through a virtual map. This enables the USV to generate and pursue exploration objectives by integrating information from both the virtual map and an occupancy grid map. Candidate paths are initially generated using the Rapidly Exploring Random Tree (RRT) algorithm, with the path utility function being influenced by the covariance associated with potential landmarks and localization. The proposed EM exploration framework is implemented in the Robot Operating System (ROS) and validated in a simulated ocean environment featuring static obstacles and a single USV. Experimental results, including recorded candidate targets and navigation trajectories, demonstrate that the EM method provides improved exploration stability and higher map coverage. Specifically, compared to the traditional Next-Best-View (NBV) method, the EM approach achieves a map coverage improvement of 0.27.