Committee Chair: Odest Chadwicke Jenkins

Abstract:
This dissertation proposes various techniques for creating articulated object representations in a scalable manner that require little human labeling. These representations are shown to be useful through experiments in articulated object pose and configuration estimation, as well as performance via comparisons to other methods on common articulated object datasets. We posit that these scalable representations can be used in the future to improve robot manipulation capabilities directly or as augmentations to data-driven techniques. The approach proposed in this work addresses the dimensionality and data scale issues originally described as difficulties in articulated object representation, which should improve the robotics communities' ability to ingest, share, and utilize articulated object data within subsequent research efforts.