Animate your avatar: learning conditional human motion prior

Abstract

This research project focuses on the concept of virtual humans and aims to enable natural language control of 3D avatars, allowing them to perform human-like movements that are coherent with their surrounding environment. To achieve this goal, the project proposes to learn a "conditional" human motion prior that takes into account scene information and/or language descriptions. This approach is intended to generate more coherent and meaningful human-like motions that are better suited to specific scenarios and objectives. The potential applications of this research include gaming, virtual avatars, and the Metaverse. The project employs a two-stage approach that combines language-conditioned human motion generation and physics-based character control techniques to generate diverse and physically plausible human motions in a physical world from language descriptions. The results of the project demonstrate the effectiveness of this approach in terms of motion diversity, faithfulness to the language description, and physical plausibility.