Human-robot interaction assessment using dynamic engagement profiles

This paper addresses the use of convolutional neural networks for image analysis resulting in an engagement metric that can be used to assess the quality of human robot interactions. We propose a method based on a pretrained convolutional network able to map emotions onto a continuous [0-1] interval, where 0 represents disengaged and 1 fully engaged. The network shows a good accuracy at recognizing the engagement state of humans given positive emotions. A time based analysis of interaction experiments between small humanoid robots and humans provides time series of engagement estimates, which are further used to understand the nature of the interaction as well as the overall mood and interest of the participant during the experiment. The method allows a real-time implementation and supports a quantitative and qualitative assessment of a human robot interaction with respect to a positive engagement and is applicable to humanoid robotics as well as other related contexts.