Engagement Detection During Deictic References in Human-Robot Interaction

Humans are typically skilled interaction partners and detect even small problems during an interaction. In contrast, interactive robot systems often lack the basic capabilities to sense the engagement of their interaction partners and keep a common ground. This becomes even more problematic if humanoid robots with human-like behavior are used as they build up high expectations in terms of their cognitive capabilities. This paper contributes an approach for analyzing human engagement during object references in an explanation scenario based on time series alignment. An experimental guide scenario in a smart home environment was used to collect a training and test dataset where the engagement classification is carried out by human operators. The experiments already performed on the dataset give deeper insights into the presented task and motivate an incremental, mixed modality approach to engagement classification. While some of the results rely on external sensors they give an outlook on the requirements and possibilities for HRI scenarios with next-gen social robots.