How Should a Robot Interrupt a Conversation Between Multiple Humans

This paper addresses the question of how and when a robot should interrupt a meeting-style conversation between humans. First, we observed one-to-one human-human conversations. We then employed raters to estimate how easy it was to interrupt each participant in the video. At the same time, we gathered behavioral information about the collocutors (presence of speech, head pose and gaze direction). After establishing that the raters’ ratings were similar, we trained a neural network with the behavioral data as input and the interruptibility measure as output of the system. Once we validated the similarity between the output of our estimator and the actual interruptiblitiy ratings, we proceeded to implement this system on our desktop social robot, CommU. We then used CommU in a human-robot interaction environment, to investigate how the robot should barge-in into a conversation between multiple humans. We compared different approaches to interruption and found that users liked the interruptibility estimation system better than a baseline system which doesn’t pay attention to the state of the speakers. They also preferred the robot to give advance non-verbal notifications of its intention to speak.