Kevin Xie

University of Toronto, Nvidia (United Kingdom)

Papers

Total Citations

105

H-Index

About

Kevin Xie is a researcher whose work sits at the compelling intersection of physics-based simulation, human motion analysis, and reinforcement learning. His most recognized contribution, "Physics-based Human Motion Estimation and Synthesis from Videos" (2021, 77 citations), addresses a fundamental challenge in computer graphics, gaming, and robotics simulation: generating realistic human motion without relying on costly motion capture data. By proposing a framework for training generative models of physically plausible motion directly from video, Xie helped democratize high-quality motion synthesis and opened new pathways for scalable data generation in embodied AI research. Complementing this work, Xie has made notable strides in model-based reinforcement learning through his research on continual learning with hypernetworks (accumulating over 26 citations across versions). This work tackles the critical problem of dynamics model accuracy in non-stationary environments, moving beyond the conventional approach of retraining from scratch and instead enabling agents to adapt continuously — a key capability for real-world deployment. Together, these contributions reflect Xie's broader ambition to build intelligent systems that are both physically grounded and adaptable, making his work particularly relevant to researchers in robotics, animation, and autonomous systems.