Ph.D. Dissertation Defense Announcement

Title: Developing Agile Motor Skills on Virtual and Real Humanoids

Sehoon Ha

School of Interactive Computing

Georgia Institute of Technology

http://www.cc.gatech.edu/~sha9/

Date: Tuesday, July 28, 2015

Time: 12:00pm - 2:30pm EDT

Location: TSRB 223

Committee:

Dr. C. Karen Liu (Advisor, School of Interactive Computing, Georgia Institute of Technology)

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

Dr. Jarek Rossignac (School of Interactive Computing, Georgia Institute of Technology)

Dr. Jun Ueda (School of Mechanical Engineering, Georgia Institute of Technology)

Dr. Katsu Yamane (Disney Research Pittsburgh)

Abstract:

Demonstrating strength and agility on virtual and real humanoids has been an important goal in computer graphics and robotics. However, developing physics-based controllers for various agile motor skills requires a tremendous amount of prior knowledge and manual labor due to its

complex mechanisms. The focus of the dissertation is to develop a set of computational tools to

expedite the design process of physics-based controllers that can execute a variety of agile motor skills on virtual and real humanoids. Instead of directly deploying motions on real humanoids, this dissertation takes an approach that develops appropriate theories and models in virtual simulation and systematically transfers the solutions to hardware systems.

The algorithms and frameworks in this dissertation span various topics from specific physics-based controllers to general learning frameworks. We first present an online algorithm for controlling falling and landing motions of virtual characters. The proposed algorithm is effective and efficient enough to generate falling motions for a wide range of arbitrary initial conditions in real-time. Next, we present a robust falling strategy for real humanoids that can manage a wide range of perturbations by planning the optimal contact sequences. We then introduce an iterative learning framework to intuitively design various agile motions, which is inspired by human learning techniques. The proposed framework is followed by novel algorithms to efficiently optimize control parameters for the target tasks, especially when they have many constraints or parameterized goals. Finally, we introduce an iterative approach for exporting simulation-optimized control policies to hardware of robots to reduce the number of hardware experiments, that accompany expensive costs and labors.