Evolving body kinematics for virtual characters

This source preferred by Wen Tang

Authors: Gatzoulis, C., Tang, W. and Stoddart, W.

Start date: 20 June 2006

DOI: 10.2312/LocalChapterEvents/TPCG/TPCG06/211-217

This data was imported from Scopus:

Authors: Gatzoulis, C., Tang, W. and Stoddart, W.J.

Journal: Theory and Practice of Computer Graphics 2006, TPCG 2006 - Eurographics UK Chapter Proceedings

Pages: 203-210

ISBN: 9783905673593

Physically-based character animation systems often require complex knowledge of the underlying equations of motion. Hence, producing physically-realistic animations can be time consuming with these systems. In this paper, we present an approach that automatically searches for kinematics solutions for virtual characters. Characters learn their locomotion by evolving body kinematics. We designed two different control architectures for the character's learning process with predefined motion data sets and a feedback system. The first system is based on a layer of genetic algorithms (GA) and the second is based on a Reinforcement Learning (RL) approach. Animation systems based on these control architectures require little knowledge of the physics equations of motions, but can generate physically-feasible motions in real-time through observations of available motion data sets, such as previous animations or motion capture data. This animation approach allows animators to construct easily realistic body kinematics motion for computer game characters. Embedded with simulated musculature of human body, the system also has applications in sports and physiotherapy for motion visualization. The test data also demonstrates the advantages and drawbacks of the two types of control methods. © The Eurographics Association 2006.

The data on this page was last updated at 04:55 on June 17, 2019.