Data-driven Approaches to Uncertainty Modelling for SLAM in the Open Sea
Time: Thu 2022-12-15 14.00
Location: Kollegiesalen, Brinellvägen 8, Stockholm
Video link: https://kth-se.zoom.us/j/69562986585
Subject area: Computer Science
Doctoral student: Ignacio Torroba , Robotik, perception och lärande, RPL
Opponent: Professor John Leonard, Massachusetts Institute of Technology
Supervisor: Associate Professor John Folkesson, Robotik, perception och lärande, RPL; Prof. Patric Jensfelt, Robotik, perception och lärande, RPL
Autonomous underwater vehicles (AUVs) equipped with multibeam echo-sounders have become indispensable tools for bathymetric mapping due to their ability to reach seabed regions inaccessible to surface vessels. However, the closer proximity to the survey area comes at the expense of a growing error in the AUV global pose estimate due to the lack of prior maps or a geo-referencing system underwater, such as GPS. This limitation, together with the changing environment dynamics in deep sea waters and the scale of the areas to map, makes simultaneous localization and mapping (SLAM) a necessary enabler for long-range, reliable and safe AUV navigation in open sea missions.SLAM has allowed the safe deployment of self-driving cars on the streets and service robots in our homes, but remains a challenge in the deep sea domain. This is due to the constrained sensing capabilities available underwater and the scarcity of distinguishable features in the seabed. As a result of these, successful place recognition is infrequent, yielding loop closure (LC) detections more sparse and therefore more crucial. To adequately factor in each LC constraint in a SLAM back-end, their uncertainties need to be carefully parameterized to weight their influence in the final AUV trajectory estimate. Thus, this thesis is concerned with modelling these uncertainties, in particular when analytical models cannot be derived, focusing instead in data-driven methods.
We present our contributions in three key SLAM areas targeting this problem. First, our work on inferring the uncertainties in the bathymetric submap registration process shows how deep learning techniques can be successfully applied to learning noise models directly from raw data and without ground truth position information. We further show how the predicted uncertainties improve the convergence of submap-based graph-SLAM solutions in AUV surveys.Secondly, we introduce a methodology to construct terrain representations with Stochastic Variational Gaussian processes (SVGP) propagating the AUV localization and sensors uncertainties into the final maps. The proposed approach is not limited to any GP kernel or noise model in the data and can handle datasets of millions of training points. The experiments demonstrate how the learned terrain models yield improved particle filter estimates in AUV localization problems.Finally, we adapt the previous SVGP mapping approach to online bathymetric learning and demonstrate its scalability and flexibility in a Rao-Blackwellized SLAM framework. The presented RBPF-SVGP solution is capable of maintaining up to 100 particles in parallel, each with a single SVGP map capable of regressing entire surveys. Our results show how the RBPF-SVGP can perform in real time in an embedded platform and can be executed live in an AUV.
Additionally, all the implementations proposed have been made publicly available to promote further research in underwater SLAM and the adoption of common open-source frameworks, datasets and benchmarks in the field.