Mobile robot localization using active sensing based on Bayesian network inference

In this paper we propose a novel method of sensor planning for a mobile robot localization problem. We represent the conditional dependence relation between local sensing results, actions, and belief of the global localization using a Bayesian network. Initially, the structure of the Bayesian network is learned from the complete data of the environment using the K2K2 algorithm combined with a genetic algorithm (GA). In the execution phase, when the robot is kidnapped to some place, it plans an optimal sensing action by taking into account the trade-off between the sensing cost and the global localization belief, which is obtained by inference in the Bayesian network. We have validated the learning and planning algorithm by simulation experiments in an office environment.