Leveraging Online Virtual Agents to Crowdsource Human-Robot Interaction

Abstract

Robots require a broad range of interaction skills in order to work effectively alongside humans. They must have the ability to detect and recognize the actions and intentions of a person, produce functionally valid and situationally appropriate actions, and engage in social interactions through physical cues and dialog. However, social interactions with one of today’s robots will quickly become one-sided and repetitive, even after just a few minutes due to its shallow depth of knowledge and experience. This problem exposes weaknesses in the underlying traditional approaches that aim to pre-code responses to a limited number of inputs. We propose the use of crowdsourcing as a tool for the development of social robots that allow for rich, diverse and natural human-robot interaction. To enable crowdsourcing at a massive scale, we describe a newly implemented system that uses online virtual agents to collect data and then leverages the resulting corpus to train our robot behavior system for use on a real world task.