Towards Accurate 3D Positioning in Large-Scale Underwater Environments Daniele Evangelista * , Ivano Donadi * , Daniel Fusaro * , Emilio Olivastri * and Alberto Pretto * * Department of Information Engineering, University of Padua, Padua, Italy Email: evangelista@dei.unipd.it, ivano.donadi@dei.unipd.it, fusarodani@dei.unipd.it, emilio.olivastri@phd.unipd.it, alberto.pretto@dei.unipd.it Abstract—An effective and reliable AUV (autonomous under- water vehicle) should be capable to carry out complex underwa- ter maintenance and exploration tasks in complete autonomy. Unfortunately, the hostility of the underwater environment, combined with the lack of any communication infrastructure and global positioning, introduce important scientific challenges. In this paper, we introduce an underwater global positioning system for AUVs suitable for large-scale environments sparsely populated by man-made infrastructure. Our system relies on sonar and stereo camera data, by means of a novel deep descriptor for sonar images and a two-view pixel-wise voting network for 6DoF (Six degrees of freedom) object pose estimation, both trained on synthetically generated datasets. The global AUV position is then estimated by exploiting a probabilistic framework that fuses on-line the available sensor readings. Index Terms—Marine Robotics, Range Sensing, Vision-Based Navigation, Localization I. I NTRODUCTION Underwater robots (UID, Underwater Inspec- tion/Intervention Drones) represent essential tools for carrying out complex assembly and maintenance operations of plants in underwater environments. The growing interest in highly sustainable offshore energy hubs will make these tools even more important since they can carry out heavy but necessary operations in underwater environments in a totally safe way. AUVs could carry out such tasks also in a fully autonomous way, without the need for remote piloting and possibly without the need for a support vessel, with undoubted advantages from an economic, environmental and personnel safety point of view. One of the biggest challenges in autonomous underwater navigation is the capability of the AUV to localize itself, since the GPS signal is obviously not available, while the USBL (Ultra-short baseline) acoustic positioning system, when available, can be unstable and very noisy. To address this task, we are developing innovative deep-learning-based perception front-ends designed to process data acquired by a multibeam sonar and a stereo camera. In particular, we are developing a global and compact descriptor computed on the incoming, real sonar images that can be compared with a database of descriptors, each one computed on synthetic sonar images acquired at known locations in the working area, to assess the current position estimate. Furthermore, we are developing an object pose estimation module that, using Fig. 1. Example data extracted from Gazebo simulator with the plugin proposed in [1]. The top row shows an example of a waypoint for inspecting an underwater asset (left) that requires multiple viewpoints to be inspected completely (right). The bottom row shows an example of sonar reading from the simulator. only a CAD model of the object and the AUV stereo camera input, is able to estimate the pose of an object w.r.t. (with respect to) the AUV in extremely challenging conditions. Acquiring and annotating large amounts of underwater data is extremely expensive and time-consuming. To overcome this limitation, both modules are trained by using synthetic datasets, possibly tailored to the target environment with domain adaptation techniques. The information coming from the perception front-ends is finally fused with the information coming from the INS/DVL (Inertial Navigation System/Doppler Velocity Log) sensor which equips the AUV by exploiting a graph-based positioning framework, to estimate and update the global pose of the AUV w.r.t. a provided map of a large-scale environment sparsely populated by man-made assets. This system is developed in the context of a research project carried out by the University of Padua. It has the ambitious goal of providing AUVs with an advanced autonomous per- ception system capable of increasing their productivity and effectiveness. In this short paper, we first introduce the use-case we are addressing, specifically the large-scale autonomous monitoring of underwater assets 1 , then we will briefly introduce the 1 An asset is any man-made underwater structure. 2022 I-RIM Conference October 13-14, Milan, Italy ISBN: 9788894580532 DOI:10.5281/zenodo.7531274 85