Controllable Part-wise Motion Generation

Loading...
Thumbnail Image

URL

Journal Title

Journal ISSN

Volume Title

School of Science | Master's thesis

Date

2024-12-31

Department

Major/Subject

Computer Science

Mcode

Degree programme

Master's Programme in Computer, Communication and Information Sciences

Language

en

Pages

35

Series

Abstract

Recent advances in generative models show promise in generating diverse and expressive character animations. Most existing models, however, rely on full-body representations, which can restrict the manipulation of character motions when control and adjustments are needed for specific body parts, such as hands, to avoid collisions. Building on existing autoregressive conditional variational autoencoder architectures, this work explores the use of part-wise generative models to address this limitation. By dividing the character into upper and lower body parts and treating them as separate modalities, we introduce more control over specific body parts, enabling finer manipulation of character motions. To handle the interdependencies between body parts, we leverage a multimodal mixture-of-experts prior, allowing information sharing between the different parts. This approach enables the generation of diverse and controllable animations, improving the adaptability of characters in collision avoidance scenarios. We demonstrate the effectiveness of our model in two collision avoidance tasks: dancing in narrow corridors and avoiding collisions between two characters. Our model outperforms the baseline method by minimizing collisions between the character's body parts and obstacles, showcasing its potential for adaptable animation generation in dynamic environments.

Description

Supervisor

Hämäläinen, Perttu

Thesis advisor

Kim, Nam Hee

Keywords

generative models, character animation, conditional variational autoencoders, collision avoidance, machine learning, games

Other note

Citation