Hmm, You Seem Confused! Tracking Interlocutor Confusion for Situated Task-Oriented HRI

Abstract

Our research seeks to develop a long-lasting and high-quality engagement between the user and the social robot, which in turn requires a more sophisticated alignment of the user and the system than is currently commonly available. Close monitoring of interlocutors' states, and we argue their confusion state in particular, and adjusting dialogue policies based on this state of confusion is needed for successful joint activity. In this paper, we present an initial study of human-robot conversation scenarios using a Pepper robot to investigate the confusion states of users. A Wizard-of-Oz (WoZ) HRI experiment is illustrated in detail with stimuli strategies to trigger confused states from interlocutors. For the collected data, we estimated emotions, head pose, and eye gaze, and these features were analysed against the silence duration time of the speech data and the post-study self-reported confusion states that are reported by participants. Our analysis found a significant relationship between confusion states and most of these features. We see these results as being particularly significant for multimodal situated dialogues for human-robot interaction and beyond.