Path and View Planning for Underwater Surface Inspection

We discuss the problem of active inspection of a marine structure by an autonomous underwater vehicle (AUV). The AUV must plan a path of sensing views that estimates the difference between two surfaces: one reconstructed from prior data, and one reconstructed online during inspection. We propose modeling the structure’s surface using non-parametric Bayesian regression, which allows for a number of novel cost functions for planning the path of the AUV to minimize some metric of change detection between reconstructions. We analyze the resulting cost functions using the theoretical tools of adaptivity gaps and (adaptive) submodularity, and we draw connections to the more general active classification problem. In addition, we discuss experimental trials that utilize profiling sonar data from a ship hull inspection. This work moves towards formal analysis of active sensing for inspection tasks.