Avoiding robot faux pas

Abstract

Contextual cues strongly in influence the behavior of people in social environments, and people are very adept at interpreting and responding to these cues. While robots are becoming increasingly present in these spaces, they do not yet share humans' essential sense of contextually-bounded social propriety. However, it is essential for robots to be able to modify their behavior depending on context so that they operate in an appropriate manner across a variety of situations. In our work, we are building models of context for social robots, that operate on real-world, naturalistic, noisy data, across multi-context and multi-person settings. In this paper, we discuss one aspect of this work, which concerns teaching a robot an appropriateness function for interrupting a person in a public space. We trained a support-vector machine (SVM) to learn an association between contextual cues and the reaction of people being interrupted by a robot across three different contexts. Overall, our results are promising, and further work on integrating context models into social robots could lead to interesting and impactful findings across the HRI community.