An Information-Theoretic Localization Criterion for Robot Map Building

Abstract

In the process of robot map building, many different aspects of the problem must be taken into account, like the environment the robot moves in, the type of sensors it uses, the dimensionality of the measurements, etc. One fundamental issue becomes the extraction of relevant features from the raw measurements to be used by the robot for self-localization. Recently, a Bayesian localization error formula was proposed to be used as an optimization criterion in a neural-network based nonlinear feature extraction scheme. We derive in this paper an alternative criterion for robot localization that is based on information-theoretic concepts and that is more efficient to compute, and we compare with the localization error criterion. We discuss the case of linear projection with PCA and carry out an analysis thereof. Finally, we show experimental results for supervised data obtained by a real robot in a typical office environment. Keywords: Robot map building, robot localization, information t...