What is it about?
Reproducing human behaviors in a realistic way is a fundamental research goal in computer animation research. In this work, we present a way of producing human motions in a physically believable manner by using physically simulated characters, conditional VAEs, and reinforcement learning.
Featured Image
Photo by Aideal Hwa on Unsplash
Why is it important?
Learning controllers for physically simulated characters is very time-consuming and computationally expensive. Training a backbone model that can be reused for many different downstream tasks is a promising way of resolving those problems. We introduce a new way of training a backbone model for physically simulated characters and demonstrate that many challenging downstream tasks can efficiently be solved by using our model.
Perspectives
I hope the method will be a stepping stone to build a general purpose backbone model that can handle large and diverse behaviors. I hope this article inspires not only researchers in computer graphics but also researchers in other domains such as robotics and machine learning.
Jungdam Won
Facebook Inc
Read the Original
This page is a summary of: Physics-based character controllers using conditional VAEs, ACM Transactions on Graphics, July 2022, ACM (Association for Computing Machinery),
DOI: 10.1145/3528223.3530067.
You can read the full text:
Resources
Contributors
The following have contributed to this page
