Stylistic Locomotion Modeling with Conditional Variational Autoencoder

Abstract

We propose a novel approach to create generative models for distinctive stylistic locomotion synthesis. The approach is inspired by the observation that human styles can be easily distinguished from a few examples. However, learning a generative model for natural human motions which display huge amounts of variations and randomness would require a lot of training data. Furthermore, it would require considerable efforts to create such a large motion database for each style. We propose a generative model to combine the large variation in a neutral motion database and style information from a limited number of examples. We formulate the stylistic motion modeling task as a conditional distribution learning problem. Style transfer is implicitly applied during the model learning process. A conditional variational autoencoder (CVAE) is applied to learn the distribution and stylistic examples are used as constraints. We demonstrate that our approach can generate any number of natural-looking human motions with a similar style to the target given a few style examples and a neutral motion database.