Date of Award

Fall 2018

Project Type

Thesis

Program or Major

Computer Science

Degree Name

Master of Science

First Advisor

Momotaz Begum

Second Advisor

Philip J Hatcher

Third Advisor

Marek Petrik

Abstract

The presence of robots in society is becoming increasingly common, triggering the need to learn reliable policies to automate human-robot interactions (HRI). Manually developing policies for HRI is particularly challenging due to the complexity introduced by the human component. The aim of this thesis is to explore the benefits of leveraging temporal reasoning to learn policies for HRIs from demonstrations. This thesis proposes and evaluates two distinct temporal reasoning approaches. The first one consists of a temporal-reasoning-based learning from demonstration (TR-LfD) framework that employs a variant of an Interval Temporal Bayesian Network to learn the temporal dynamics of an interaction. TR-LfD exploits Allen’s interval algebra (IA) and Bayesian networks to effectively learn complex temporal structures. The second approach consists of a novel temporal reasoning model, the Temporal Context Graph (TCG). TCGs combine IA, n-grams models, and directed graphs to model interactions with cyclical atomic actions and temporal structures with sequential and parallel relationships. The proposed temporal reasoning models are evaluated using two experiments consisting of autonomous robot-mediated behavioral interventions. Results indicate that leveraging temporal reasoning can improve policy generation and execution in LfD frameworks. Specifically, these models can be used to limit the action space of a robot during an interaction, thus simplifying policy selection and effectively addressing the issue of perceptual aliasing.

Download

Share

COinS