Incremental Knowledge Acquisition for Human-Robot Collaboration

Abstract. Robots require a significant amount of domain knowledge to collaborate with humans in complex domains. Since it is difficult to provide accurate and complete domain knowledge, active learning algo-rithms have been developed to enable robots to acquire relevant infor-mation by posing questions when necessary. Human participants may, however, lack the time and expertise to provide elaborate and accurate responses. Success of active learning in human-robot interaction thus depends on robots posing questions that enable faster learning using limited interaction with non-expert humans. Towards this objective, this paper presents an architecture for incremental active learning. Robots equipped with this architecture construct candidate questions using lo-cal and global contextual cues. These queries are ranked based on utility, which is is measured as a combination of measures for ambiguity and in-formation gain. Human responses to the top-ranked questions are used to update the robot’s knowledge. This paper illustrates and evaluates the architecture’s capabilities in a simulated domain, significantly reducing the number of questions posed in comparison with algorithms that use the individual measures or a strategy for randomly selecting questions