Abstract
In this talk, I will discuss some of the recent work in my group on learning-aided geometric methods for solving a range of exciting problems in robotics, such as state estimation, localization, and point cloud registration. The emergence of deep learning has provided the opportunity to leverage data for creating a form of memory for online deployment as well as estimating hard-to-measure quantities in real-time. On the other hand, when assumptions are satisfied, geometric methods can provide performance guarantees and interpretable outcomes in real-world applications. I will highlight how tight integrations of these two seemingly apart approaches are possible and will enable new opportunities for advancing robotics and autonomous systems.