Title: On the Use of Samples for Effortless Certifiable Optimization
Abstract: A possibly surprising fact is that most optimization problems in robotics, although non-convex, can be solved to global optimality, sometimes with speeds on par with efficient local solvers. Indeed, many instances of problems such as simultaneous localization and mapping, continuous-time state estimation, and optimal control problems, are polynomial optimization problems with Markov-like sparsity. For such problems, one can use a relatively standard recipe for deriving easy-to-solve convex relaxations with which one can identify global solutions. I will give an introduction to these methods, and then I will present two ongoing works: First, an extension of our previously published method AutoTight that allows to globally solve any problem given only a couple of randomly generated feasible samples — thus avoiding the high burden of error-prone problem formulations. Second, I will present a novel algorithm that extends these methods beyond polynomial models, and could be used, for example, to improve sample efficiency in reinforcement learning by more optimally exploring the search space.
Bio: Frederike Dümbgen is an incoming assistant professor at Carnegie Mellon University. She is currently a researcher in the WILLOW team of Inria Paris, where she has been working on optimization for robotics since May 2024. Before joining Inria, she spent two years as a postdoctoral fellow at the Robotics Institute of the University of Toronto, collaborating with Prof. Timothy D. Barfoot on certifiable optimization. Frederike holds a Ph.D. in Computer and Communication Scienes from École Polytechnique Fédérale de Lausanne (EPFL), Switzerland, where her research focused on developing systems for non-visual spatial perception using diverse sensor modalities. She earned her B.Sc. and M.Sc. in Mechanical Engineering from EPFL in 2013 and 2016, respectively, with a minor in Computational Science and Engineering. She performed her Master’s thesis at the Autonomous Systems Lab of ETH Zürich and gained experience in industry as an intern of Disney Research and ABB, among others. She was recognized as a R:SS Pioneer in 2024, as Google’s Women Techmarker in 2020, and has served as a co-chair of the RAS technical committee on model-based optimization since 2024.
Speaker: Frederike Dümbgen, CMU
Tuesday, May 27, 2025 – 10:00 to 11:00am EST
Montgomery Knight – 317
Host: Lu Gan