PUB - Publikationen an der Universität Bielefeld

Social robots’ capabilities are advancing and so is their deployment in numerous domains - as language tutors, workout partners or social companions. An increase in social robots’ abilities to autonomously choose behaviors inevitably augments the likelihood that these behaviors deviate from what a user might expect and approve of. One useful tool to increase transparency and enhance trust and acceptance in human-robot interaction are behavior explanations. Yet, sufficient models of how to enable self-explanations of autonomous social robot behavior in order to adequately inform users and prevent potential negative effects are missing. This thesis investigated effects of providing human-inspired self-explanations for robot behavior in social human-robot interaction with differing content (Paper A) and at different times (Paper B). Based on these findings, a dialogic model of a social robot’s self-explanations was developed, implemented as part of an interaction architecture and evaluated in a user study (Paper C). More concretely: a first study (A) aimed at investigating what effects a social robot’s self explanations produce on user’s perception of the behavior and how these effects differ as a function of different explanation content. An explanation type model was developed based on humans’ behavior explanations and conceptually grounded in the robot’s behavior generation process. Results demonstrated that verbal self-explanations could increase understandability and desirability of robot behaviors. Positive effects were higher for causally structured than simpler explanations and varied for different types of behavior. Evaluation of (B) the timing of a robot’s self-explanations surprisingly revealed negative effects of explaining undesirable behavior before, as compared to after execution of the behavior. These contextual differences highlight the importance of considering the socio-interactive context for deciding when to give what kind of explanation. The gained insights were transferred to an interaction setting (C): a dialogue-based, socio-interactive model for behavior explanations was proposed and requirements for explainable architectures for social robots were postulated. The explanation model was implemented as part of an explainable interaction architecture and tested in an acquaintance scenario, demonstrating both, successful behavior and explanation generation. Overall, positive effects of an autonomous robot’s self-explanations were shown to vary as a function of explanandum desirability and explanation content and timing, and explanation generation based on the proposed socio-interactive framework enabled the robot to autonomously and coherently self-explain its behavior.