Date of Award

1-1-2018

Document Type

Dissertation

Degree Name

Ph.D.

Organizational Unit

Daniel Felix Ritchie School of Engineering and Computer Science, Computer Science

First Advisor

Mohammad H. Mahoor, Ph.D.

Second Advisor

Kateri McRae

Third Advisor

Kimon Valavanis

Fourth Advisor

Mathew Rutherford

Fifth Advisor

Timothy Sweeny

Keywords

Affect perception, Empathic robot, Facial expressions, Human robot interaction, Rear projected robot, Social robots

Abstract

Social (or Sociable) robots are designed to interact with people in a natural and interpersonal manner. They are becoming an integrated part of our daily lives and have achieved positive outcomes in several applications such as education, health care, quality of life, entertainment, etc. Despite significant progress towards the development of realistic social robotic agents, a number of problems remain to be solved. First, current social robots either lack enough ability to have deep social interaction with human, or they are very expensive to build and maintain. Second, current social robots have yet to reach the full emotional and social capabilities necessary for rich and robust interaction with human beings. To address these problems, this dissertation presents the development of a low-cost, flexible, affect-aware rear-projected robotic agent (called ExpressionBot), that is designed to support verbal and non-verbal communication between the robot and humans, with the goal of closely modeling the dynamics of natural face-to-face communication.

The developed robotic platform uses state-of-the-art character animation technologies to create an animated human face (aka avatar) that is capable of showing facial expressions, realistic eye movement, and accurate visual speech, and then project this avatar onto a face-shaped translucent mask. The mask and the projector are then rigged onto a neck mechanism that can move like a human head. Since an animation is projected onto a mask, the robotic face is highly flexible research tool, mechanically simple, and low-cost to design, build and maintain compared with mechatronic and android faces. The results of our comprehensive Human-Robot Interaction (HRI) studies illustrate the benefits and values of the proposed rear-projected robotic platform over a virtual-agent with the same animation displayed on a 2D computer screen. The results indicate that ExpressionBot is well accepted by users, with some advantages in expressing facial expressions more accurately and perceiving mutual eye gaze contact.

To improve social capabilities of the robot and create an expressive and empathic social agent (affect-aware) which is capable of interpreting users' emotional facial expressions, we developed a new Deep Neural Networks (DNN) architecture for Facial Expression Recognition (FER). The proposed DNN was initially trained on seven well-known publicly available databases, and obtained significantly better than, or comparable to, traditional convolutional neural networks or other state-of-the-art methods in both accuracy and learning time. Since the performance of the automated FER system highly depends on its training data, and the eventual goal of the proposed robotic platform is to interact with users in an uncontrolled environment, a database of facial expressions in the wild (called AffectNet) was created by querying emotion-related keywords from different search engines. AffectNet contains more than 1M images with faces and 440,000 manually annotated images with facial expressions, valence, and arousal. Two DNNs were trained on AffectNet to classify the facial expression images and predict the value of valence and arousal. Various evaluation metrics show that our deep neural network approaches trained on AffectNet can perform better than conventional machine learning methods and available off-the-shelf FER systems.

We then integrated this automated FER system into spoken dialog of our robotic platform to extend and enrich the capabilities of ExpressionBot beyond spoken dialog and create an affect-aware robotic agent that can measure and infer users' affect and cognition. Three social/interaction aspects (task engagement, being empathic, and likability of the robot) are measured in an experiment with the affect-aware robotic agent. The results indicate that users rated our affect-aware agent as empathic and likable as a robot in which user's affect is recognized by a human (WoZ).

In summary, this dissertation presents the development and HRI studies of a perceptive, and expressive, conversational, rear-projected, life-like robotic agent (aka ExpressionBot or Ryan) that models natural face-to-face communication between human and emapthic agent. The results of our in-depth human-robot-interaction studies show that this robotic agent can serve as a model for creating the next generation of empathic social robots.

Publication Statement

Copyright is held by the author. User is responsible for all copyright compliance.

Rights Holder

Ali Mollahosseini

Provenance

Received from ProQuest

File Format

application/pdf

Language

en

File Size

214 p.

Discipline

Computer engineering, Computer science

Download



Share

COinS