Using a physics engine to improve probabilistic object localization

Abstract

Each robot that is meant to handle objects must identify and localize the designated objects prior to any manipulation attempt. Commonly the locations of the objects are estimated separately, assuming full mutual probabilistic independence between all of them. Hereby information is lost in constellations where objects are probabilistically dependent in their state. A full state estimation could cope with this problem, although it seems infeasible due to the enormous computational demands. In this paper we present an approach that models such local dependencies and utilizes a physic engine to exploit those within a probabilistic particle filter multi object localization system, in order to improve the accuracy of the estimation results.