Game Theoretic Methods for Human-Robot Parallel Play
Author(s)
Bansal, Shray
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
When humans perform parallel play- complete tasks in the presence of other humans- it is
often necessary to balance self-interest and cooperation. This dissertation focuses on activi-
ties that involve two agents with separate goals who interact due to shared space. The main
thesis of this work is: Parallel play can model many kinds of human-robot interactions,
admitting efficient and effective game-theoretic techniques for decision-making in a wide
variety of coordination tasks with humans and human-like agents. This thesis is based on
the observation that human activities involve coordinating with other individuals interested
in achieving their own goals, this often results in some amount of goal conflict but does not
usually lead to pure competition. By using game theory- the study of decision-making in
rational, interacting agents- to model the self-interested nature of the agents, we can derive
methods that are more compatible with human behavior.
Although the game-theoretic solutions to complex sequential games can be infeasible
in general, we develop efficient solutions using the inherent structure of parallel play by
defining them as a subset of general-sum games. We perform experiments in three inter-
action scenarios. First, we model driving as a common payoff game and show that giving
similar importance to agent goals helps us achieve coordination between them. Second,
we consider parallel manipulation between a human and a robot on a shared table. Here,
we introduce interaction-supporting actions to reduce the likelihood of collisions. We also
develop a method for interactive decision-making that computes and infers Nash equilibria
to coordinate the agent behaviors. Last, we consider a purely cooperative setting where
agents coordinate complex actions to prepare dishes in a shared kitchen. Here, we extend
our Nash equilibrium selection approach to using reinforcement learning and show how
our method applies to human-AI interaction settings even beyond parallel play.
Sponsor
Date
2023-04-30
Extent
Resource Type
Text
Resource Subtype
Dissertation