*********************************
There is now a CONTENT FREEZE for Mercury while we switch to a new platform. It began on Friday, March 10 at 6pm and will end on Wednesday, March 15 at noon. No new content can be created during this time, but all material in the system as of the beginning of the freeze will be migrated to the new platform, including users and groups. Functionally the new site is identical to the old one. webteam@gatech.edu
*********************************
Title: Efficient and Principled Robot Learning: Theory and Algorithms
Ching-An Cheng
Robotics PhD student
School of Interactive Computing
Georgia Institute of Technology
Date: Tuesday, April 30, 2019
Time: 1:00pm - 3:00pm (EST)
Location: CCB 340
Committee:
Dr. Byron Boots (advisor), School of Interactive Computing, Georgia Institute of Technology
Dr. Seth Hutchinson, School of Interactive Computing, Georgia Institute of Technology
Dr. Karen Liu, School of Interactive Computing, Georgia Institute of Technology
Dr. Evangelos Theodorou, School of Aerospace Engineering, Georgia Institute of Technology
Dr. Geoff Gordon, Microsoft Research and Machine Learning Department, Carnegie Mellon University
Abstract:
Roboticists have long envisioned fully-automated robots that can operate reliably in unstructured environments. This is an exciting but extremely difficult problem; in order to succeed, robots must reason about sequential decisions and their consequences in face of uncertainty. As a result, in practice, the engineering effort required to build reliable robotic systems is both demanding and expensive. This research aims to provide a set of techniques for efficient and principled robot learning. It tackles this challenge from a theoretical perspective, in order to provide a panoramic perspective on the underlying difficulties and indicate systematic directions for improvement. A better theoretical formulation for system design, that more closely integrates analysis and practical needs, is a key to address current deficiencies. Here these theoretical principles are applied to design better algorithms in three important aspects of robot learning: policy optimization, uncertainty modeling, and development of structural policies. It uses and extends online learning, optimization, and control theory, and is demonstrated on applications including reinforcement and imitation learning, Gaussian process learning, and integrated motion planning and control. A shared feature across this research is the reciprocal interaction between the development of practical algorithms and the advancement of abstract analyses. Real-world challenges force the rethinking of proper theoretical formulations, which in turn lead to refined analyses and new algorithms that can rigorously leverage these insights to achieve better performance.
