Collecting Social Signals in Constructive and Destructive Events during Human-robot Collaborative Tasks

Abstract

Adapting behaviors based on others' reactions is a fundamental skill for a social robot that must interact with people. In this work, we to develop a systematic method to collect ecologically plausible data of human reactions to robot behaviors and associated valence. We designed a dyadic interaction were 24 participants played a board game in a human-robot team for a chance to win a chocolate. The ''Grumpy robot'' is responsible for losing an easy-to-win game, while the ''Kind robot'' for winning a seemingly impossible-to-win game. Questionnaires show that participants recognize both robots' critical impact on the game's outcome, but show similar social attraction towards both. Videos' reactions are distinct: smiles and neutral faces to the ''Kind robot'', and laughter, confusion, or shock to the ''Grumpy robot''. Collected data will be used to teach the robot to understand human reactions.