Keywords: Robotics, Probabilistic Inference, Belief Propagation, Graph Algorithms, Pose Tracking, Uncertainty
TL;DR: This paper investigates the potential for Counter-Hypothetical reasoning to be introduced within belief propagation to produce the most likely posterior samples in real-world, continuous robot pose tracking tasks.
Abstract: For safe and efficient collaborative environments, co-located robots must be able to visually estimate and track the movements of surrounding robots. Nonparametric belief propagation provides a probabilistic framework for autonomous systems to reason about uncertainty and multiple hypotheses at once, as well as leverage portions of the robot that are better observed. Due to its computational constraints on the sample size, belief propagation often causes the filter to converge to incorrect estimates, because the particles are ineffectively representing the true belief. Our work seeks to maintain particle diversity by re-initializing the particles from a variety of proposal distributions as needed. We extend promising work of adaptive particle reinvigoration from a single distribution in the particle filter domain, which introduced Counter-Hypothetical reasoning to independently estimate when the filter was in failure mode. Our proposed framework explores the usefulness of this reasoning within belief propagation to manage sampling from multiple distributions and propagating its information through the standard graphical model. We present preliminary qualitative results for this method on tracking 21 links on a humanoid robot, Digit.