SIGGRAPH 2022 Outstanding Doctoral Dissertation Award: Peng
Awardee(s):
Award:
-
Outstanding Doctoral Dissertation Award
Dissertation Title:
- Acquiring Motor Skills Through Motion Imitation and Reinforcement Learning
Description:
ACM SIGGRAPH is pleased to announce Xue-Bin “Jason” Peng, PhD, as the 2022 recipient of the Outstanding Doctoral Dissertation Award. In his dissertation, Jason presents significant advances in character animation with deep reinforcement learning to produce highly-controllable naturalistic physics-based motion.
Humans and animals are capable of awe-inspiring feats of agility produced by drawing from a vast repertoire of motor skills. In sharp contrast, artificial agents in robotics or computer-generated animation are often stiff and awkward—or else very limited in repertoire—despite decades of progress in the design of controllers. Jason Peng has developed a series of motion imitation and reinforcement learning techniques which upend a decades-long line of research in mocap-based motion synthesis: his work allows agents to learn a large spectrum of highly dynamic and athletic behaviors by mimicking demonstrations. Instead of designing controllers or reward functions for each skill of interest, the agent need only to be provided with a few example motion clips of the desired skill to synthesize a controller, that not only closely replicates the target behavior but can also be robust to perturbations — such as standing up after a fall or reacting to being hit with an object. The resulting controllers thus reconciliate natural motion and interactivity/response to unpredicted events, marking a significant improvement over previous approaches. His work on adversarial motion priors further shows how to generalize these ideas to large training sets.
In addition to his accomplishments in character animation, Jason has applied his methods to real-world quadrupedal and bipedal robots. He demonstrated that a wide range of learned locomotion gaits can be executed on robot hardware, even enabling a quadruped to chase its own tail. This work beautifully illustrates how to go from real-world data (mocap) to a simulated world for learning the control policies, and then back to the real world for deployment. This kind of interdisciplinary progress is emblematic of the diversity and impact of work developed in the graphics community.
Dr. Peng has advanced the field of controller design through motion imitation and reinforcement learning with a doctoral dissertation in which each chapter develops innovative ideas and progresses towards a steadily increasing level of capability. The SIGGRAPH community thus recognizes Jason Peng for these extraordinary achievements with the 2022 ACM SIGGRAPH Doctoral Dissertation Award, and looks forward to his future contributions.
