A Learned Approach to Adaptive Sampling for Seabed Identification with Autonomous Vehicles

We propose a sampling approach for guiding an autonomous underwater vehicle to adaptively collect ambient acoustic sound, with the goal of partitioning the seabed according to its geoacoustic properties. Existing approaches to adaptive sampling are based on heuristic policies that aim to balance exploration of the entire region with refinement of the existing partition estimates. We utilize behavioral cloning, which trains a neural network to imitate an expert policy from simulated data, training the policy to maximize the area under the curve of F1 score versus distance traveled. Results on synthetic and real-world sediment data show our approach outperforms existing methods in terms of sampling efficiency.