RoboEarth Web-Enabled and Knowledge-Based Active Perception

Abstract

In this paper we explore how a visual SLAM system and a robot knowledge base can mutually benefit from each other. The object recognition and mapping methods are used for grounding abstract knowledge and for creating a semantically annotated environment map that is available for reasoning. The knowledge base allows to reason about which object types are to be expected while exploring an environment and where to search for novel objects given a partial environment map. Prior information like task descriptions and object models is loaded from RoboEarth, a web-based knowledge base for exchanging knowledge between robots, and the created maps are again uploaded to RoboEarth. We show that by exploiting knowledge about common objects in a room and about the co-occurrence of objects, both efficacy and efficiency of the perception can be boosted.