Third "Coffee with EUMR"

Physical modeling of sonar is complex, and often infeasible without detailed data on the modeled subsea environment. Simulation methods therefore tend to simplify sonar modeling in the interest of practical simulations. Fusion of sonar data into high-definition models face similar obstacles. In particular, sonars with wide beam angles require significant simplifications in order to construct bathymetry or object models. We present a line of research that replaces classical sonar models by black box neural networks.  By relying on learnt models rather than hand-crafted ones, we hope to enhance the model quality by supplying more data, rather than manually fine-tuning for each environment or object. This has the potential of greatly enhancing statistical and visual fidelity. We investigate neural representation learning as well as generative adversarial models (GANs) as representations of sonar data. We showcase applications of each method to sidescan simulation and to bathymetry reconstruction from sidescan. Both GANs and neural representations have contributed to significant advances in visual sensing, and our preliminary results demonstrate their practicality in the subsea domain.

AUVs have been successfully used as mobile sensors in oceanographic applications for over a decade. However, marine vehicles typically have relatively low speed and endurance, making it interesting to use the energy in the currents to simultaneously reduce commuting times and power consumption. This is particularly true in coastal operations with small portable AUVs, where the magnitude of the water velocity can exceed the vehicles' maximum speed. In this talk a method for AUV trajectory optimization based on dynamic programming and high resolution model forecasts of the water velocity will be presented, as well as its implementation. The discussion will be illustrated by numerical simulations and experimental results in the Sado estuary in Portugal.