Investigating Deep Learning Approaches for Human-Robot Proxemics
Yuan Gao, Sebastian Wallkötter, Mohammad Obaid, Ginevra Castellano · 2018
In this paper, we investigate the applicability of deep learning methods to adapt and predict comfortable human-robot proxemics. Proposing a network architecture, we experiment with three different layer configurations, obtaining three different end-to-end trainable models. Using these, we compare their predictive performances on data obtained during a human-robot interaction study. We find that our long short-term memory based model outperforms a gated recurrent unit based model and a feed-forward model. Further, we demonstrate how the created model can be used to create customized comfort zones that can help create a personalized experience for individual users.