Chair: Chad Jenkins

Abstract:
General purpose robots must reliably estimate the pose of high degree-of-freedom objects within their environment despite the presence of heavy clutter and partial observability. While probabilistic inference offers a diagnosable framework to reason about these poses over time, it currently lacks any explicit modeling of ambiguity or doubt. We introduce deep evidential reasoning to nonparametric Bayesian inference by representing ambiguity and doubt alongside the filter’s belief distribution.

This dissertation focuses on alleviating particle deprivation within nonparametric Bayesian inference for high degree-of-freedom pose estimation and tracking from monocular video. We present the counter-hypothetical particle filter, which independently quantifies the disqualifying evidence against its particle set to detect failure mode and redistribute its samples. Our work then extends differentiable nonparametric belief propagation to model ambiguity as well as doubt—creating a hierarchy of the usefulness of the observations at each node. This presented approach, Weighted And Graphical Evidential Reasoning (WAGER-DNBP), is validated on our contributed Progress LUMBER Dataset. Our 44k frame dataset includes pose annotations for a moving 29-link biped robot, Digit, and is unique for its feature of heavy occlusion caused by static and dynamic obstacles.