*********************************
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: Learning Motion Policies for Dexterous Manipulation with Geometric Fabrics
Date: Wednesday, May 4th, 2022
Time: 4:00 PM – 6:00 PM EST
Location: Zoom Meeting (https://gatech.zoom.us/j/96014725301?pwd=SlpXbUhMNFptbzUvRng4RzYvVHhjUT09)
Man Xie
Ph.D. Student in Computer Science
School of Interactive Computing
Georgia Institute of Technology
Committee:
Dr. Frank Dellaert (Advisor) – School of Interactive Computing, Georgia Institute of Technology
Dr. Harish Ravichandar (Co-advisor) – School of Interactive Computing, Georgia Institute of Technology
Dr. Seth Hutchinson – School of Interactive Computing, Georgia Institute of Technology
Dr. Byron Boots – School of Computer Science & Engineering, University of Washington
Dr. Nathan Ratliff – Seattle Robotics Lab, Nvidia
Summary:
Given that our world is designed by and for humans with impressive dexterity, it is natural to posit that robots with dexterous manipulation skills can effectively operate in a variety of human environments. However, dexterous manipulation in multi-fingered robots has been a long-standing challenge in robotics due to the fact that dexterity requires complex skills, such as coordinating numerous degrees of freedom, balancing contact forces, and breaking and reestablishing contacts with objects. Existing methods are often limited to simulation, or rely on extensive interactions with the environment and incur considerable computational burden. This thesis makes contributions that will help robots efficiently learn and reliably execute dexterous manipulation skills. First, we develop and investigate a new class of behavioral dynamical systems, we call Geometric Fabrics, that can encode complex behaviors in high dimensions by introducing strong structural inductive biases while retaining expressivity. Second, to circumvent the need for painstaking human effort required to manually design these models, we introduce Neural Geometric Fabrics (NGF) that can be used to efficiently learn generalizable manipulation skills. Third, we propose a policy learning architecture that can encode nominal behaviors that are shared across a wide variety of dexterous manipulation tasks into a NGF model and expose a high-level action space to enable efficient policy learning for a novel task with commonly used RL methods.
