Demonstrating MoveAE: Modifying affective robot movements using classifying variational autoencoders

Abstract

We developed a method for modifying emotive robot movements with a reduced dependency on domain knowledge by using neural networks. We use hand-crafted movements for a Blossom robot and a classifying variational autoencoder to adjust affective movement features by using simple arithmetic in the network's learned latent embedding space. We will demonstrate the workflow of using a graphical interface to modify the valence and arousal of movements. Participants will be able to use the interface themselves and watch Blossom perform the modified movements in real time.