Field trials of Multi-vehicle Adaptive Robotic Surveys for Seafloor Mapping and Characterization
Oscar Pizarro

Multiple lines of research have to come together to demonstrate habitat mapping and benthic cover estimation using adaptive robotic surveys. In this paper we report on a few of these strands including: high resolution seafloor imaging from traditional AUVs and Lagrangian drifters, automated interpretation of seafloor imagery, and adaptive sampling guided by the uncertainty in predictions of benthic cover.

High resolution seafloor imaging from mobile autonomous platforms has become a valuable tool for habitat classification, stock assessment and seafloor exploration. This abstract addresses the concept of joint seafloor survey planning using both navigable and drifting platforms, and presents results from an experiment using a bottom surveying AUV and a drifting Lagrangian camera float. We consider two classes of vehicles; one which is able to self propel and execute structured surveys, and one which is Lagrangian and moves only with the currents. The navigable vehicle is the more capable and the more expensives asset of the two. The Lagrangian platforms is a low cost imaging tool that can actively control its altitude above the seafloor to obtain high quality images but can not otherwise control its trajectory over the bottom. When used together the vehicles offer several scenarios for joint operations. When used in an exploratory manner the Lagrangian float is an inexpensive way to collect images from an unknown area. Depending on the collected images, a follow on structured survey with the navigable AUV can collect additional information. When used simultaneously the drifting float can guide the AUV trajectory over an area. When both platforms are equipped with acoustic tracking and communications the AUV trajectory can be automatically redirected to follow the Lagrangian float using one of many patterns. This capability allows for surveys that are potentially more representative of the near bottom oceanographic conditions at the desired location. Results where both platforms were used as part of a coral habitat monitoring project are included.

Machine learning research offers flexible and powerful approaches that can use observations of different scales and modalities to construct predictive models with meaningful representations of uncertainty. Beyond providing a sense of the quality of the models, these representations can guide further collection of observations to improve predictive capabilities. The traditionally-resource constrained problem of generating habitat maps from full coverage acoustic multibeam data and targeted optical surveys can be viewed through the lens of machine learning and adaptive sampling. We have investigated the use of techniques in machine learning such as deep learning methods, Gaussian Processes and Dirichlet-Multinomial regressors to generate habitat maps and to suggest where further sampling would be most useful. We present results based on surveys performed in Australia using ship-borne multibeam sonar and precisely georeferenced imagery collected with Autonomous Underwater Vehicles (AUVs).

Based on these experiences we reflect on successes, ongoing challenges, future research directions for advanced autonomy for robots operating near the seafloor. We also discuss areas potentially ripe to transition from research to operational capabilities.