Date: Wednesday, April 12, 2023 

Time: 3:00 PM - 4:30 PM EST 

Location: https://gatech.zoom.us/j/99939049932 

Yunbo Zhang 

Computer Science Ph.D. Student 

School of Interactive Computing 

Georgia Institute of Technology 

Committee: 

Dr. Greg Turk (Advisor) - School of Interactive Computing, Georgia Institute of Technology 

Dr. Sehoon Ha - School of Interactive Computing, Georgia Institute of Technology 

Dr. Jie Tan - School of Interactive Computing, Georgia Institute of Technology 

Dr. Charles C. Kemp – Coulter Department of Biomedical Engineering, Georgia Institute of Technology 

Dr. Yuting Ye – Reality Labs Research, Meta 

Abstract: 

The proposed thesis focuses on using deep reinforcement learning to generate physics-based character animations involving complex physical interactions between characters and environments. We present three pieces of work each generating animations for a different type of physical interaction. In the first piece of work, we present a training framework for controlling tools to manipulate general amorphous materials such as sweep and gathering viscous fluid on a table surface or tossing and flipping meat patties using a flat pan. In the second piece of work, we develop a curriculum learning pipeline to generate animations for dexterous in-hand manipulation with various rigid objects. In the last piece of work, we develop a reward formulation, namely “Interaction Graph” that measures general physical interactions between characters and objects. We use the reward to produce animations for complex multi-character physical exercises. To conclude my study, I propose one additional project that extends the idea of “interaction graph” and applies that to the observation space and solves more general in-hand manipulation problems. In addition, we will combine the outcome of the general control policy with a pre-trained kinematic motion generator to produce a variety of in-hand manipulation animations with different hand models and manipulated objects.