UBC Computer Science alum wins Best Paper Award at SIGGRAPH Asia 2025
Winning paper combines machine learning with character animation methods to make game design more accessible
UBC Computer Science MSc alum Ruiyu Gou won a Best Paper Award at computer graphics conference SIGGRAPH Asia 2025, which took place from December 15 - 18 in Hong Kong. The paper, Control Operators for Interactive Character Animation, was deemed to be in the top 2% of the 301 accepted papers at the conference.
Gou, a former MSc student supervised by UBC Computer Science Professor Emeritus Michiel van de Panne and a current PhD student in computer graphics at SFU, worked on this project during her internship at Epic Games. For this paper, she collaborated with Dr. van de Panne and Principal Animation Programmer Daniel Holden from Epic Games to create a machine-learning based framework based on “Control Operators,” which makes it considerably easier for game designers to develop and train machine-learning models that directly generate the realistic motions of in-game characters.
“There has been so much research progress in using neural networks to model motion, but it’s not really making its way into industry,” says Gou. “We started this project with the idea of building a system, including a framework, so that someone who doesn’t understand the full technical underpinnings of machine learning can still get the characters to behave the way they want to and respond to the controls the way they want.”
In the last decade, researchers have been developing neural networks, a type of machine learning model, to help game designers create character controllers. However, these neural networks still required extensive knowledge of machine learning and training pipelines in order to customize them as needed for high-quality, real-time generation of animated motion.
To make these technical challenges easier for developers, the new framework makes the design process more intuitive by automating many processes, including building the neural network and processing of the available training. The framework uses recent generative models, including flow-matching models, to produce natural and realistic looking motions that meet desired game-player goals, such as joy-stick control over locomotion speed and direction or a target chair to sit on.
The researchers also conducted a user study to test their framework on a group of game industry practitioners, including animators, technical designers and game developers. In the user study, the participants followed a tutorial introducing the system’s workflow and created and tested an animation controller. Compared to other controller development systems, the participants rated Control Operators competitive in metrics such as controllability, accessibility, and quality of animation and rated it top in scalability.
“This is a real step forward in being able to bring the power of modern machine learning methods into the game engine in an accessible way,” says Dr. van de Panne.