Unsupervised learning of sensory primitives from optical flow fields

Abstract

Adaptive behaviour of animats largely depends on the processing of their sensory information. In this paper, we examine the estimation of robot egomotion from visual input by unsupervised online learning. The input is a sparse optical flow field constructed from discrete motion detectors. The global flow field properties depend on the robot motion, the spatial distribution of motion detectors with respect to the robot body and the visual environment. We show how online linear Principal Component Analysis can be applied to this problem to enable a robot to continuously adapt to a changing environment. © 2014 Springer International Publishing Switzerland.