ECE Seminar: David M. Rosen

Title: Provably Sound Perception for Reliable Autonomy

David M. Rosen

Location: Zoom Link

Abstract:  Machine perception — the ability to construct accurate models of the world from raw sensor data — is an essential capability for mobile robots, supporting such fundamental functions as planning, navigation, and control.  However, the development of algorithms for robotic perception that are both *practical* and *reliable* presents a formidable challenge: such methods must be capable of solving complex estimation tasks in real-time on resource-limited mobile platforms, while remaining robust to challenges such as sensor noise, uncertain or misspecified perceptual models, and potentially contaminated measurements. In this talk, I show how one can meet these challenges through the design of practical perception methods that are both *computationally efficient* and *provably sound*, focusing on the foundational problem of spatial perception.  I begin with a brief introduction to pose-graph optimization (PGO): this problem lies at the core of many fundamental spatial perception tasks (including robotic mapping, sensor network localization, and 3D visual reconstruction), but is high-dimensional and nonconvex, and therefore challenging to solve in general. Nevertheless, I show how one can leverage convex relaxation to efficiently recover *exact, certifiably optimal* PGO solutions in a noise regime that encompasses most practical robotics and computer vision applications.  Our algorithm, SE-Sync, is the first practical method provably capable of recovering correct (globally optimal) PGO solutions. Next, I address the design of machine learning methods for spatial perception, focusing on the fundamental problem of rotation estimation.  I show that topological obstructions can actually prevent deep neural networks (DNNs) employing common rotation parameterizations (e.g. quaternions) from learning to estimate widely-dispersed rotation targets, as is required in (for example) object pose estimation. I then describe a novel parameterization of 3D rotations that overcomes this obstruction, and that supports an explicit notion of uncertainty in our DNNs’ predictions.  Experiments confirm that (as predicted by theory) DNNs employing this representation achieve superior accuracy and reliability when applied to object pose estimation, and that their predicted uncertainties enable the reliable identification of out-of-distribution test examples (including corrupted inputs). Finally, I will conclude with a discussion of future directions that aim to unify provably sound estimation and learning methods, thereby enabling the creation of perception systems with both the *robustness* and *adaptability* necessary to support reliable long-term autonomy in the real world.

Speaker Bio:  David M. Rosen is a postdoctoral associate in the Laboratory for Information and Decision Systems at the Massachusetts Institute of Technology.  His research addresses the design of practical provably robust methods for machine perception, using a combination of tools from optimization, geometry, algebra, and probabilistic inference.  He holds the degrees of BS in Mathematics from the California Institute of Technology (2008), MA in Mathematics from the University of Texas at Austin (2010), and ScD in Computer Science from the Massachusetts Institute of Technology.  Prior to joining LIDS, he was a Research Scientist at Oculus Research (now Facebook Reality Labs) in Seattle.His work has been recognized with a Best Paper Award at the 2016 International Workshop on the Algorithmic Foundations of Robotics, an RSS Pioneer Award at Robotics: Science and Systems 2019, and a Best Student Paper Award at Robotics: Science and Systems 2020.