Proactive multimodal perception for feature based anchoring of complex objects

Abstract

Every artificial embodied system that act in a physical environment has to connect perceptual data to abstract symbols. This is called symbol anchoring. It is a nontrivial task due to uncertainty of measurements and the virtually unlimited number of distinguishable objects that can be perceived. We review a commonly used anchoring framework and use our mobile service robot to give an example for complex multisensory anchoring of semantically important objects. In a detailed discussion we point out four major problems of the reviewed framework that preclude its application to complex multimodal systems. We propose a novel feature based strategy to avoid the arising computational problems adopting an intelligent scheduler for perceptual activities. We show that scheduling perception is crucial to reduce computational costs for anchoring.