Tesca Fitzgerald
Computer Science PhD Student
School of Interactive Computing
College of Computing
Georgia Institute of Technology

Date: Friday, December 16, 2016
Time: 2:00 pm to 4:00 pm (EST)
Location: TBD

Committee
----------------
Dr. Ashok Goel (Advisor), School of Interactive Computing, Georgia Institute of Technology
Dr. Andrea Thomaz (Advisor), Department of Electrical and Computer Engineering, University of Texas at Austin, and School of Interactive Computing, Georgia Institute of Technology
Dr. Sonia Chernova, School of Interactive Computing, Georgia Institute of Technology
Dr. Henrik Christensen, Department of Computer Science, University of California at San Diego, and School of Interactive Computing, Georgia Institute of Technology
Dr. Brian Scassellati, Department of Computer Science, Yale University

Abstract
---------------
As robots become more commonplace, they will need to address a wide variety of problems. Since a robot cannot be programmed to complete every task, it is necessary for robots to learn new tasks by interacting with a human teacher. Learning from Demonstration is an effective method for this; however, current methods require that the robot receive many demonstrations of a task, or they are limited to completing tasks which are similar to previous demonstrations.

​I propose that the differences between a source environment (in which the task is demonstrated) and a target environment (containing new objects) lie on a spectrum of similarity. At one end of the spectrum, the source and the target tasks are identical so that the memory of previously learned skills directly supplies the answer for the target problem. In the middle of the spectrum, differences between the source and target tasks may be limited to minor modifications to object features or positions. Further along the spectrum, the differences between the source and target tasks may include significant modifications to object features and configurations, necessitating new action models to address the target problem.

I propose that a robot may address this range of transfer tasks by (i) analyzing the similarities and differences between the source and target problems, (ii) identifying the level of knowledge abstraction appropriate for transfer for the given type of similarity, and (iii) collaborating with a human teacher to ground the knowledge abstractions in the transfer task.

I further propose a cognitive system based on case-based analogical learning that may enable a robot to collaborate with a human teacher to transfer task knowledge to a range of target problems. Given task demonstrations in source domains, this system stores the task knowledge as individual cases in memory. When given a target problem, it retrieves a similar case from memory, and then identifies the level of abstraction at which knowledge from the source case should be transferred to the target problem. At present, this cognitive system focuses on addressing the problem of transferring knowledge from a small number of examples, with proposed work addressing the problem of creative transfer.

----------------------