DCG-UPUP-Away : automatic symbol acquisition through grounding to unknowns

Abstract

Research in automatic natural language grounding, in which robots understand how phrases relate to real-world objects or actions, offers a compelling reality in which untrained humans can operate highly sophisticated robots. Current techniques for training robots to understand natural language, however, assume that there is a fixed set of phrases or objects that the robot will encounter during deployment. Instead, the real world is full of confusing jargon and unique objects that are nearly impossible to anticipate and therefore train for. This thesis presents a model called the Distributed Correspondence Graph - Unknown Phrase, Unknown Percept - Away (DCG-UPUP-Away) that augments the state of the art Distributed Correspondence Graph by recognizing unknown phrases and objects as unknown, as well as reasoning about objects that are not currently perceived. Furthermore, experimental results in simulation, as well as a trial run on a turtlebot platform, validate the effectiveness of DCG-UPUP-Away in grounding phrases and learning new phrases.