Multimodal Analysis of User Engagement with a Recommender Robot in Cafe Settings

Predicting user engagement is essential to ensure natural human-robot interactions (HRI) and to enhance social acceptance of robots. Motivated by this, this paper introduces a framework for predicting human engagement during HRI in retail settings. Specifically, we (1) introduced novel annotations for the LISI-HRI dataset, a new in-the-wild dataset comprising interactions between a recommender robot and customers of a cafeteria, to model levels of human engagement during HRI, (2) analysed the effectiveness of modelling engagement state with our selected multimodal features, and (3) designed and evaluated a Support Vector Machine (SVM) based user engagement prediction approach. The obtained results aim to serve as a baseline on this dataset for engagement prediction in real-world human-robot interaction scenarios.