Information fractals for approximate reasoning in sensor-based robot grasp control

Abstract

This research is concerned with the intelligent control of autonomous robot systems in unstructured, highly uncertain environments. We show that the theory of fractal sets is a useful tool for approximate machine reasoning at multiple levels of precision. We first derive an algorithm for the classification and combination of incomplete and imprecise patterns of evidence. We also derive a new algorithm for belief propagation across incompatible frames. These algorithms are then applied to the sensorimotor control of robotic systems. The first application is to a sensor fusion problem, involving the recognition of 3D objects from a combination of vision and touch data. The second application is to sensor based grasping with multifingered hands. These applications integrate high and low level representations of (i) irregular and sparse sensory patterns and (ii) preshaped grasp control under uncertainty.