Generating Robotic Emotional Body Language of Targeted Valence and Arousal with Conditional Variational Autoencoders

Abstract

Non-verbal communication that encompasses emotional body language is a crucial aspect of social robotics applications. Deep learning models for the generation of robotic expressions of bodily affect gain more and more ground recently over the hand-coded methods. In this work, we present a Conditional Variational Autoencoder network that generates emotional body language animations of targeted valence and arousal for a Pepper robot, and we conduct a user study to evaluate the interpretability of the generated animations.