Resilient Object Perception for Robotics

Abstract

A broad array of applications, ranging from search and rescue to self-driving vehicles, requires robots to perceive and understand the geometry of objects in the environment. Object perception needs to reliably work in a variety of scenarios and preserve a desired level of performance in the face of outliers and shifts from the training domain. Obtaining such a level of performance requires robust estimation algorithms that are able to identify and reject outliers, as well as techniques to continually improve performance of learningbased perception modules during test-time. In this thesis, we address these challenges by proposing (1) certifiably optimal solvers and a graph-theoretic framework that together help achieve state-of-the-art pose estimation performance even under high outlier rates, (2) self-supervised object pose estimators that can improve performance during test-time with accuracy comparable to state-of-the-art supervised methods, and (3) a test-time adaptation method for both object shape reconstruction and pose estimation without the need for CAD models. Throughout the thesis, we demonstrate that by using a variety of tools from optimization and learning, we can develop resilient object perception systems that perform reliably in a wide range of conditions.