Using particle filter and machine learning for accuracy estimation of robot localization

Abstract

Robot localization is a fundamental capability of all mobile robots. Because of uncertainties in acting and sensing and environmental factors such as people flocking around robots there is always the risk that a robot loses its localization. Very often behaviors of robots rely on a valid position estimation. Thus, for dependability of robot systems it is of great interest for the system to know the state of its localization component. In this paper we present an approach that allows a robot to asses if the localization is still valid. The approach assumes that the underlying localization approach is based on a particle filter. We use deep learning to identify temporal patterns in the particles in the case of losing/lost localization in combination with weak classifiers from the particle set and perception for boosted learning of a localization monitor. The approach is evaluated in a simulated transport robot environment where a degraded localization is provoked by disturbances cased by dynamic obstacles.