Title: Game Theory for Human-Robot Parallel Play
Date: Friday, March 17, 2023.
Time: 9:00 - 11:00 AM (EST)
Location: (Hybrid) Coda C1115
Zoom: https://gatech.zoom.us/j/96635227323
Shray Bansal
Computer Science Ph.D. Candidate
School of Interactive Computing
Georgia Institute of Technology
Committee:
Dr. Charles L. Isbell (Advisor) – College of Interactive Computing, Georgia Institute of Technology
Dr. Ayanna Howard (Co-advisor) – College of Engineering, The Ohio State University
Dr. Mark Riedl – School of Interactive Computing, Georgia Institute of Technology
Dr. Sonia Chernova – School of Interactive Computing, Georgia Institute of Technology
Dr. Michael Littman – Department of Computer Science, Brown University
Abstract:
When humans perform parallel play -complete tasks in the presence of other humans- it is often necessary to balance self-interest and cooperation due to the conflict caused by shared space. This is common in everyday activities like driving, buying groceries, and sharing meals. As robots become increasingly integrated into human environments, they will inevitably share space with humans and have to balance coordinating with humans and achieving their own goals.
Although parallel play comes naturally to humans, the inherent complexity and uncertainty of this interaction are challenging for robots. Our work explores how a robot can achieve similar levels of coordination and performance in tasks where a human and a robot share a common space but have separate independent goals. Game theory is helpful in understanding human decision-making processes and we use it to guide the robot's decision-making. Our main thesis is: parallel play models many kinds of human-robot interactions and admits game theoretic methods for decision-making to enable coordination with humans.
We make the following contributions to human-robot parallel play to support this thesis. (1) Propose a cooperative planning method for driving as parallel play and show how the distributions of agent rewards affect coordination outcomes, (2) formalize actions to support interaction in close-proximity manipulation and study the trade-off in task performance and human preference when employing them, (3) develop a planning method that uses a Nash equilibrium selection strategy to coordinate with agents with different personalities in parallel manipulation, and (4) generate diverse multi-agent reinforcement learning policies using suboptimal partners and leverage the Nash equilibria to coordinate in a cooperative cooking task.