Collecting Data for Human Confusion When Following Simulated Robot’s Instructions
Open Access
Author:
Anderson, Caroline
Area of Honors:
Computer Science
Degree:
Bachelor of Science
Document Type:
Thesis
Thesis Supervisors:
Alan Richard Wagner, Thesis Supervisor John Joseph Hannan, Thesis Honors Advisor
Keywords:
computer vision human robot interaction convolutional neural networks computer science unity simulation human computer interaction
Abstract:
This thesis aimed to create a publicly available human-robot interaction confusion dataset generated from interactions with a simulated robot that elicits confusion through specific stimuli. This simulation was created in Unity and was accessed by subjects through the Internet. We gathered a dataset size of 10,237 images from 32 participants. From this data set, we built several classifiers to detect if a subject is being exposed to confusion stimuli or not. Our best classifier was a 36-layer network that achieved an accuracy of 55.6% on our testing data. Areas of future work include expanding our data set and building a transfer learning model that has an emotion detector as the base model.
