Dynamic control of robot perception using multi-property inference grids

Abstract

An approach to dynamic planning and control of the perceptual activities of an autonomous mobile robot equipped with multiple sensor systems is considered. The robot is conceptually seen as an experimenter. The author discusses the explicit characterization of task-specific information requirements, the use of stochastic sensor models to determine the utility of sensory actions and perform sensor selection, and the application of information-theoretic models to measure the extent, accuracy, and complexity of the robot's world model. It is shown how the loci of interest of relevant information and the corresponding loci of observation can be computed, allowing the robot to servo on the information required to solve a given task. The use of these models is outlined in the development of strategies for perception control, and in the integration of perception and locomotion. Some illustrations of the methodology are provided.