Sonar-based mapping of large-scale mobile robot environments using EM

Abstract

In this paper we present a method for learning maps with mobile robots equipped with range finders. Our method builds on an approach previously developed by the authors, which uses EM to solve the concurrent mapping and localization problem (constrained maximum likelihood estimation). In contrast to other techniques which either relied on predefined landmarks or used highly accurate sensors, our approach is able to fully exploit the rich nature of range data and to deal with noisy information coming, for example, from ultrasound sensors. During EM it uses a layered representation of maps. It operates in two stages: first, small, local maps are learned under the assumption that odometry is locally correct. EM is then applied to to estimate the positions of these local maps. Finally, the local maps are integrated into one global map using Bayes rule. Experimental results demonstrate that our approach is well suited for constructing large maps of typical indoor environments using sensors as inaccurate as sonars.