Close
About
FAQ
Home
Collections
Login
USC Login
Register
0
Selected
Invert selection
Deselect all
Deselect all
Click here to refresh results
Click here to refresh results
USC
/
Digital Library
/
University of Southern California Dissertations and Theses
/
Towards socially assistive robot support methods for physical activity behavior change
(USC Thesis Other)
Towards socially assistive robot support methods for physical activity behavior change
PDF
Download
Share
Open document
Flip pages
Contact Us
Contact Us
Copy asset link
Request this asset
Transcript (if available)
Content
Towards Socially Assistive Robot Support Methods
for Physical Activity Behavior Change
by
Katelyn Swift-Spong
A Dissertation Presented to the
FACULTY OF THE GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulllment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(COMPUTER SCIENCE)
May 2019
Copyright 2019 Katelyn Swift-Spong
Acknowledgements
I would rst like to thank my advisor Prof. Maja Matari c. I am deeply grateful for
her continual support and guidance. Her mentorship has helped me to accomplish my
research goals and has shaped who I am as a researcher.
I would also like to thank my committee members, Prof. Elizabeth Zelinski and
Prof. Stefanos Nikolaidis, as well as my qualifying exam committee members, Prof.
Gaurav Sukhatme, Prof. David Traum, and Prof. Jonathan Gratch. Their feedback
and advice have helped me shape my research ideas and direction.
I am grateful to the National Science Foundation and the Southern California Clin-
ical and Translational Science Institute for supporting this research.
I would also like to thank my colleagues in the Interaction Lab who have been a joy
to work with. I have greatly enjoyed the multitude of research discussions I have had
with my fellow lab members. Their advice, humor, perspective, encouragement, and
willingness to lend a hand have made my time at USC unforgettable. I would especially
like to thank Elaine Schaertl Short for her research advice and collaboration on many
projects. I would also like to acknowledge the collaborators outside of the Interaction
Lab I have worked with and the student research assistants who I have had the pleasure
of mentoring on various projects.
ii
Finally, I would like to thank my family. My parents, who fostered my interest in
science, and Dalton Combs have supported me through every lump in the road I have
encountered in this journey.
iii
Table of Contents
Acknowledgements ii
List of Figures vii
List of Tables x
List of Algorithms xi
Abstract xii
Chapter 1: Introduction 1
1.1 Physical Activity Behavior Change . . . . . . . . . . . . . . . . . . . . . 1
1.2 Socially Assistive Robot Role in Supporting Physical Activity Behavior
Change . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.1 Behavior Change Mechanisms . . . . . . . . . . . . . . . . . . . . 4
1.2.2 SAR Support Methods . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.3 Applications to Domain-Relevant Physical Activity . . . . . . . . 6
1.3 Dissertation Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4 Dissertation Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
Chapter 2: Background and Related Work 9
2.1 Technology-Supported Physical Activity Behavior Change Interventions 9
2.1.1 Smartphone and Messaging Interventions . . . . . . . . . . . . . 9
2.1.2 Embodied Conversational Agent Interventions . . . . . . . . . . 11
2.1.3 Socially Assistive Robot Interventions . . . . . . . . . . . . . . . 11
2.2 Validation Domains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2.1 Post-Stroke Upper Extremity Rehabilitation . . . . . . . . . . . 14
2.2.2 Adolescent Physical Activity . . . . . . . . . . . . . . . . . . . . 15
2.2.3 Reducing Sedentary Behavior in Older Adults . . . . . . . . . . . 16
2.3 Behavior Change Mechanisms . . . . . . . . . . . . . . . . . . . . . . . . 17
2.3.1 Beliefs About Capabilities . . . . . . . . . . . . . . . . . . . . . . 17
2.3.2 Social In
uences . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
iv
2.3.3 Attitude Towards the Target Behavior . . . . . . . . . . . . . . . 19
2.3.4 Cueing and Reinforcement . . . . . . . . . . . . . . . . . . . . . . 20
2.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
Chapter 3: Design and Evaluation of Socially Assistive Robot Comparative Feed-
back Support 25
3.1 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.2 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.2.1 Robot and System Design . . . . . . . . . . . . . . . . . . . . . . 27
3.2.2 Study Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.2.3 Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.2.4 Hypotheses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.2.5 Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2.6 Study Population . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.3.1 Press Times . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.3.2 Delay Times . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.3.3 Self-Ecacy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.3.4 Time on Task . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.3.5 Perception of the Robot . . . . . . . . . . . . . . . . . . . . . . . 44
3.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
Chapter 4: Design and Evaluation of Socially Assistive Robot Backstory and Af-
fective Messaging Support 50
4.1 Focus Groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.2 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.2.1 Study Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.2.2 Experiment Setup . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.2.3 Robot Backstories . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.2.4 Exercise Sessions . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.2.5 Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.2.6 Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.2.7 Hypotheses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.2.8 Study Population . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
4.3.1 Backstory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
4.3.2 Physical Activity Intrinsic Motivation . . . . . . . . . . . . . . . 61
4.3.3 Physical Activity Enjoyment . . . . . . . . . . . . . . . . . . . . 62
4.3.4 Activity Levels . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
v
Chapter 5: Model of Socially Assistive Robot Habit Formation Support 66
5.1 Habit Formation Formalization . . . . . . . . . . . . . . . . . . . . . . . 68
5.2 Modeling Habit Formation Support . . . . . . . . . . . . . . . . . . . . . 69
5.2.1 Scheduling Rewards . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.2.2 Updating Automaticity Estimate . . . . . . . . . . . . . . . . . . 72
5.2.3 Scheduling Reminders . . . . . . . . . . . . . . . . . . . . . . . . 73
5.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
Chapter 6: Evaluation of Model of Socially Assistive Robot Habit Formation Sup-
port 76
6.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
6.1.1 Habit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
6.1.2 Robot and System Design . . . . . . . . . . . . . . . . . . . . . . 77
6.1.3 Robot Behavior Design . . . . . . . . . . . . . . . . . . . . . . . 79
6.1.4 Model Parameter Selection . . . . . . . . . . . . . . . . . . . . . 82
6.1.5 Study Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
6.1.6 Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
6.1.7 Data Collection and Outcome Measures . . . . . . . . . . . . . . 86
6.1.8 Study Population . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
6.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
6.2.1 Sitting Episodes with Greater than 30 Minute Duration . . . . . 89
6.2.2 All Sitting Episodes . . . . . . . . . . . . . . . . . . . . . . . . . 91
6.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
6.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
Chapter 7: Dissertation Summary 96
7.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
7.2 Future Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
Bibliography 100
vi
List of Figures
1.1 Framework for SAR physical activity behavior change support. In this
framework, the robot in the support role uses a support method to in
u-
ence a behavior change mechanism of action in order to support behavior
change for a domain-relevant physical activity. . . . . . . . . . . . . . . 3
1.2 Instantiations of the SAR physical activity behavior change support frame-
work. These instantiations include four SAR support methods, cover
three types of domain-relevant physical activity, and span three interven-
tion durations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
3.1 Experimental setup with button board and Bandit robot. . . . . . . . . 27
3.2 Overview of fully autonomous SAR coaching system. . . . . . . . . . . . 29
3.3 Example button press for request from robot to \Please press the home
buttons with both hands and then use your right hand to press button
four." . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.4 Overview of interaction sequence. . . . . . . . . . . . . . . . . . . . . . . 33
3.5 Mean press time for limb, distance, and order; and limb and distance
(***p<:001; **p<:005). Error bars depict 95% condence interval. . 38
3.6 Mean delay time for limb, distance, and order; limb and distance; and
by condition (***p<:001; **p<:005; *p<:05). Error bars depict 95%
condence interval. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.7 Participant clusters shown on a 2D plot of the rst two principal com-
ponents. Points are colored by condition and shapes correspond to the
cluster. Clusters are distinctly separated, number of participants in each
cluster ranges from two (cluster 5) to six (cluster 2); there is no apparent
relationship between cluster and condition. . . . . . . . . . . . . . . . . 40
vii
3.8 Heatmap visualization of press times and responses to self-ecacy probes
for each cluster of participants. Each row represents a cluster, and each
column represents the mean press times, mean number of self-report
probes, and mean number of button-press probes. In each heatmap, the
buttons are displayed in a grid-like pattern with the location of each cell
corresponding to the location of a button relative to the bottom center of
the heatmap. The buttons immediately adjacent to the bottom center of
the heatmap are the near buttons, and the buttons one button away from
the bottom center are the far buttons. The angles of the buttons in the
heatmap relative to the bottom center starting from the right side of the
heatmap are the same as the button angles relative to the stroke-aected
arm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.1 Robot exercise buddy experiment setup. Left: Robot performing a split
squat with an arm raise to the side. Right: Robot, step, and treadmill
setup. Participants exercised in the space between the table and treadmill. 54
4.2 Evaluations of the robot with dierent backstories. . . . . . . . . . . . . 60
4.3 Ratings of intrinsic motivation for physical activity. . . . . . . . . . . . . 62
5.1 SAR habit formation support framework. Remind: The robot provides
reminders to encourage the user to do the behavior when the cue occurs.
Reinforce: The robot provides unexpected social rewards after the user
completes the behavior. Sustain: As the habit forms the robot decreases
the reminders so that the habit remains when the robot leaves. . . . . . 67
5.2 Finite State Machine representation of habit formation for habits with
elapsed-time based cues. . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
6.1 Experimental setup components and robot internal pan-tilt structure. . 78
6.2 Relationship between estimated automaticity and reminder schedules for
selected model parameters. The estimated automaticity shown is a func-
tion of the post-cue average sitting time for the automaticity probes in
the past window. The grey and white bars correspond to the number of
reminders that were scheduled as a function of the estimated automatic-
ity. The green dot represents an in
ection point for automaticity values
greater than 0.74 where reminders no longer start until after the average
post-cue sitting time. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
viii
6.3 Sitting duration for episodes over 30 min long and estimated automaticity
for intervention episodes for each participant. Plots show when the robot
did reminder and reward behaviors based on the habit formation support
model. They also show the estimated automaticity used in the model for
each episode in the intervention. Green horizontal lines show the average
sitting duration for episodes over 30 min for the baseline, intervention,
and retention episodes. The gap in rewarded episodes for P2 was due to
technical issue with restarting the system. . . . . . . . . . . . . . . . . . 90
ix
List of Tables
3.1 Details of trail sequence, self-ecacy probe timing, and continuation
question timing for each section of the interaction. . . . . . . . . . . . . 35
3.2 Participant characteristics for the 22 participants included in the data
analysis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.3 Mean press and delay times and standard deviations for condition, button
press order, angle, distance, and limb. . . . . . . . . . . . . . . . . . . . 37
3.4 Number of additional button presses, condition, and cluster for partici-
pants who completed time on task button presses . . . . . . . . . . . . . 44
4.1 Description of both robot backstories. . . . . . . . . . . . . . . . . . . . 55
6.1 Overview of study procedure and data collection. . . . . . . . . . . . . . 87
6.2 Mean, median, and usual sitting episode duration of all sitting episodes
per participant for each study phase. All units are in minutes. . . . . . . 92
x
List of Algorithms
5.1 Habit Formation Support Algorithm . . . . . . . . . . . . . . . . . . . . 70
5.2 Reward Scheduling Algorithm . . . . . . . . . . . . . . . . . . . . . . . . 72
5.3 Reminder Scheduling Algorithm . . . . . . . . . . . . . . . . . . . . . . . 74
xi
Abstract
Socially assistive robot (SAR) systems have the potential to support the complex process
of human behavior change by providing social support such as feedback and encourage-
ment at opportune times. This dissertation presents a framework for SAR behavior
change support in the context of physical activity behavior. This framework is designed
around the goal of creating lasting behavior change that extends past the SAR inter-
action. Within this framework, the robot is equipped with one or more SAR physical
activity behavior change support methods designed to aect a specic mechanism of
behavior change.
This dissertation develops the design of SAR feedback, backstory, and messaging
support methods for physical activity behavior change. These three methods were each
designed to support a dierent mechanism of achieving behavior change by leveraging
the robot's relational and support capabilities. Feedback was designed to support a
user's beliefs about their ability to perform a physical activity task. Robot backstory
was designed to increase the robot's ability to provide social support, and messaging
was designed to increase the user's positive feelings towards the physical activity. These
three support methods are evaluated in real-world physical activity domains with a fully
autonomous SAR systems. The feedback support method is evaluated in the domain
of post-stroke rehabilitation, and the backstory and messaging support methods are
evaluated in the domain of adolescent exercise.
xii
Reminder and social reward decision making is also developed as a SAR physical
activity behavior change support method using a model of SAR habit formation support.
This model formalizes the SAR sequential decision making task of determining when to
give reminders and social rewards towards the goal of supporting the formation of a new
desired habit. Habits are formed when the occurrence of a cue is followed by a desired
behavior, and that combination is reinforced repeatedly over time. The model of habit
formation support enables a robot to intervene in this process. This model is evaluated
in the domain of reducing older adult sedentary behavior through a two-week in-home
SAR intervention. The robot was able to generate a high level of reminder adherence
in this setting.
In this work, four SAR physical activity behavior change support methods were de-
veloped and evaluated in three dierent physical activity domains with fully autonomous
SAR systems. This dissertation contributes to understanding the methods a robot could
use to support behavior change in a variety of physical activity domains both in situ
within the context of the behavior in everyday life and outside of that context.
xiii
Chapter 1
Introduction
This chapter introduces the socially assistive robot role of supporting physi-
cal activity behavior change and presents a framework for how a robot in this
role might support behavior change. Instantiations of this framework in three
domains are summarized. The chapter concludes with an enumeration of the
primary contributions of this work and an outline of this dissertation.
1.1 Physical Activity Behavior Change
Changing behaviors related to physical activity is a known challenge and improving
methods of supporting this type of change could have a large impact on health. For
example, reducing physical inactivity by 25% could prevent 1.3 million deaths per year
worldwide (Lee et al., 2012), and physical activity is associated with increased health-
related quality of life (Bize et al., 2007, Rejeski et al., 1996). In addition, sitting time
is associated with an increased risk of mortality (Katzmarzyk et al., 2009). As such,
developing techniques for supporting a person's physical activity behavior change goals
is critically important.
1
The design of many physical activity interventions has been based on theories of
motivation and behavior change. For example, interventions based on Self Determina-
tion Theory, which promotes the importance of competence, autonomy, and relatedness
in motivation (Ryan et al., 2008), have been successfully used to increase exercise be-
haviors (Chatzisarantis and Hagger, 2009, Edmunds et al., 2008). These interventions,
however, are dicult to scale up as they rely on multiple in-person interactions over
time (Chatzisarantis and Hagger, 2009, Edmunds et al., 2008, Williams et al., 2006).
In addition, interventions based on the Transtheoretical Model, which addresses ve
stages of behavior change, have also successfully increased exercise behavior, but long-
term changes post-intervention have often not been sustained (Adams and White, 2003).
Additionally, some reviews have found that many interventions have not aected phys-
ical activity behavior (Rhodes and Pfaei, 2010), and that more studies evaluating the
long-term eects of interventions for physical activity are needed (Foster et al., 2005).
These reviews have also found limitations in determining which theory-based compo-
nents of interventions have an eect on behavior change (Foster et al., 2005, Rhodes
and Pfaei, 2010).
In order to address these limitations in connecting the design components of inter-
ventions to theory, the Behavior Change Technique Taxonomy (v1) was developed by
Michie et al. (2013). This taxonomy can be used to identify and standardize the report-
ing of behavior change techniques (BCTs) used in interventions. The BCTs identied
in this taxonomy have also recently been linked to the mechanisms of action that might
mediate the eect of the the technique on behavior change (Carey et al., 2018, Connell
et al., 2018, Johnston et al., 2018).
2
1.2 Socially Assistive Robot Role in Supporting Physical
Activity Behavior Change
In the eld of socially assistive robotics (SAR) (Feil-Seifer and Matari c, 2005b, Matari c,
2017, Matari c and Scassellati, 2016), which examines how robots can assist people
through social interaction, robots have been used to support a variety of physical ac-
tivity behavior. For example, they have been used to encourage older adults to do
seated exercises (Fasola and Matari c, 2013a, 2012, Gadde et al., 2011), encourage peo-
ple post-stroke to do rehabilitation exercises (Eriksson et al., 2005, Matari c et al., 2007,
Swift-Spong et al., 2015, Tapus et al., 2008a,b), and encourage oce workers to change
walking behaviors (Reeder et al., 2011).
SAR
User
Behavior Change
Mechanism of Action
Domain-Relevant
Physical Activity
Support Method
Figure 1.1: Framework for SAR physical activity behavior change support. In this
framework, the robot in the support role uses a support method to in
uence a behavior
change mechanism of action in order to support behavior change for a domain-relevant
physical activity.
This dissertation explores the SAR role of supporting physical activity behavior
change by examining how SAR support methods might in
uence a specic behavior
change mechanism of action within the context of a physical activity task relevant to a
certain user domain. An overview of this SAR support framework is shown in Figure
1.1. In the support role, the robot could act as either a coach or a peer-level buddy.
This dissertation focuses on in
uencing behavior change mechanisms of action within a
3
specic domain in order to examine specic ways in which a robot might support the
complex process of behavior change.
We dene the goal of this SAR role as in
uencing a mechanism of action in an in-
tervention in order to create sustained long-term behavior change after the intervention
with the robot is over. As such, the SAR support role is focused long-term sustained
change rather than short-term adherence.
This dissertation addresses the following research questions in the context of using
a robot to support physical activity behavior change:
1. How should the SAR feedback, backstory, and messaging be designed in order to
support behavior change?
2. How should a robot make decisions about when to give reminders and social
rewards in order to support behavior change?
This dissertation addresses these questions by examining the eects of SAR support
methods on behavior change mechanisms of action in the context of three types of
domain-relevant physical activity, as shown in Figure 1.2. These three instantiations
of the SAR physical activity behavior change support framework each address dierent
support methods, behavior change mechanisms of action as described by Carey et al.
(2018), and domain-relevant physical activity. In addition, they vary in location of the
intervention and duration, moving from the clinic and lab settings to the home setting
and from a single session to a two-week multi-interaction intervention.
1.2.1 Behavior Change Mechanisms
Each instantiation of the SAR support framework addresses two behavior change mech-
anisms of action. The rst, in the domain of post-stroke upper extremity rehabilitation,
addresses the mechanisms of action of beliefs about capabilities and social in
uences
4
1 session,
in-clinic
4 sessions,
in-lab
2 weeks,
in-home
SAR
User
Beliefs about capabilities
Social Influences
Post-Stroke
Upper Extremity Rehabilitation
Feedback
SAR
User
Attitude Towards Behavior
Social Influences
Adolescent Exercise
Backstory and
Messaging
SAR
User
Cueing and Reinforcement
Older Adult Breaks in
Sedentary Time
Reminder and Social Reward
Decision Making
Figure 1.2: Instantiations of the SAR physical activity behavior change support frame-
work. These instantiations include four SAR support methods, cover three types of
domain-relevant physical activity, and span three intervention durations.
(Carey et al., 2018). The beliefs about capabilities mechanism is thought to relate to
self-ecacy (Carey et al., 2018), which refers to a person's belief about their ability for a
task (Bandura, 1977). Social in
uences relates to interpersonal interactions that create
change (Connell et al., 2018). The second instantiation also includes the social in
u-
ences mechanism of action as well as attitude towards the behavior mechanism (Carey
et al., 2018). The third instantiation includes the behavioral cueing and reinforcement
mechanisms of action (Carey et al., 2018). Cueing refers to the process of triggering
the behavior with an event, and reinforcement refers to the process of strengthening the
connection between the behavior and a context (Connell et al., 2018).
5
1.2.2 SAR Support Methods
The SAR support methods developed in the three instantiations of the SAR behavior
change support framework are are feedback, backstory, messaging, and reminder and
social reward decision making. These support methods were designed based on BCTs
that have been found to have connections to the mechanisms of action in each instan-
tiation. The SAR feedback is designed to use the BCTs of verbal persuasion about
capabilities and social comparison which both have strong links to beliefs about capa-
bilities and social in
uences (Johnston et al., 2018). The backstory support method is
designed to add the BCT of social support to the intervention, which targets the social
in
uences mechanism of action, and the messaging is designed to use the BCT of giv-
ing information about emotional consequences of the behavior, which has a moderate
connection to the attitude towards the behavior mechanism (Johnston et al., 2018).
Finally, a model of SAR habit formation support was developed for the reminder and
social reward decision making support method. This model is designed to use the social
reward, prompts/cues, and habit formation BCTs to target the link with the cueing and
reinforcement mechanisms of action (Johnston et al., 2018).
1.2.3 Applications to Domain-Relevant Physical Activity
In order to understand the SAR physical activity behavior change support role in mul-
tiple domains, each instantiation of the framework targets a dierent domain and type
of physical activity. The rst is post-stroke upper extremity rehabilitation. Because
the majority of stroke survivors (55-75%) have sustained upper limb impairments post-
stroke (Lai et al., 2002), and early work using SAR systems to promote recovery post-
stroke has shown promise (Feil-Seifer and Matari c, 2005a, Matari c et al., 2009, 2007),
SAR support methods targeting the behavior change aspects of upper limb rehabili-
tation post stroke could be benecial. The second instantiation focused on exercise
6
generally for overweight adolescents. We focused on this domain because many ado-
lescents engage in an insucient amount of physical activity (Hallal et al., 2012), and
social and peer support can be important in prompting exercise in adolescents (Craggs
et al., 2011, Maturo and Cunningham, 2013). We focus on the third domain of sup-
porting older adults in taking breaks in sedentary time such behavior change can be
benecial for \successful aging" (Dogra and Stathokostas, 2012).
1.3 Dissertation Contributions
The methods for SAR physical activity behavior change support developed and eval-
uated in this dissertation serve as a foundation for developing SAR behavior change
support systems. The evaluations of these methods with domain-relevant physical ac-
tivity behaviors provide insight into the eects of these methods and how they might
be improved and applied in a variety of domains.
The following are the primary contributions of this dissertation:
1. SAR Physical Activity Behavior Change Support Framework: This
framework formalizes the role of the robot in supporting physical activity be-
havior change as using support methods to in
uence mechanisms of action for
behavior change.
2. Design of SAR Physical Activity Behavior Change Support Methods:
The design of SAR feedback, messaging, backstory, and reminder and social reward
decision making to support physical activity behavior change is presented. The
designs of these methods incorporate evidence-based behavior change techniques.
3. Model of SAR Habit Formation Support: This model, used in the reminder
and social reward decision making support method, incorporates an estimate of
7
the strength of the habit and allows the robot to make decisions about when to
provide reminders and social rewards towards the goal of increasing habit strength.
4. Evaluations of Support Methods: Studies in three dierent physical activity
domains, including post-stroke rehabilitation, adolescent exercise, and older adult
breaks in sedentary time, were conducted in order to evaluate the SAR support
methods.
1.4 Dissertation Outline
The remainder of this dissertation is organized as follows:
Chapter 2 provides background and related work about behavior change inter-
ventions that are technology-supported, the mechanisms of action for behavior
change, the behavior change techniques and related theory that inform the design
of the support methods, and the validation domains.
Chapter 3 describes the design of the SAR comparative feedback support method
and an evaluation with in the post-stroke upper extremity rehabilitation domain.
Chapter 4 describes the design of the SAR messaging and backstory support meth-
ods and an evaluation in the domain of adolescent exercise.
Chapter 5 details the approach to modeling habit formation support.
Chapter 6 describes the evaluation of the SAR habit formation support model in
the domain of reducing older adult sedentary behavior.
Chapter 7 summarizes the main contributions and ndings of this dissertation and
provides directions for future work.
8
Chapter 2
Background and Related Work
This chapter reviews background and related work of physical activity behavior
change interventions supported by technology, such as interventions that are
smartphone-based, involve embodied conversational agent interactions, and in-
volve socially assistive robot interactions. Three validation domains and their
related physical activity are described, and the connections between the SAR be-
havior change support methods and the behavior change mechanisms of action
are detailed.
2.1 Technology-Supported Physical Activity Behavior
Change Interventions
There are a variety of physical activity behavior change interventions that utilize tech-
nologies ranging from smartphone applications and virtual agents.
2.1.1 Smartphone and Messaging Interventions
Smartphone and message-based physical activity behavior change interventions have
been eective in several domains. For example, interventions that use mobile technol-
ogy such as smartphones and activity trackers (mHealth interventions) have lead to
9
increases in physical activity (Cadmus-Bertram et al., 2015, Direito et al., 2016). In
addition, smartphone-based interventions can leverage sensors in the phone and wear-
able sensors to collect and use a variety of behavioral data related to physical activity
such as activity levels (Boulos et al., 2011, Vathsangam and Sukhatme, 2014) as well as
contextual data that can inform when to send messages such as reminders for physical
activity. For example, information about weather, which can impact physical activity
(Welch et al., 2018), and Global Positioning System (GPS) location (Boulos et al., 2011)
can be used in a smartphone application to determine when to send a reminder to take
a walk. Recently, Just-in-Time Adaptive Interventions (JITAIs) (Nahum-Shani et al.,
2017), which use mobile technologies to adapt intervention support to a user's context
and behavior, have shown promise for reducing sedentary behavior (Thomas and Bond,
2015) and increasing walking (Klasnja et al., 2018). Some of the important compo-
nents of these interventions for physical activity behavior change have been found to be
incorporating immediate access to performance data, personalization of feedback, and
including actionable behavior suggestions (Schembre et al., 2018).
Another advantage of smartphone-based interventions is that they can be person-
alized using computational models for intervention control. For example, models of
physical activity that include context and are personalized to the user are being de-
veloped (Phatak et al., 2018), and reinforcement learning has been used to personalize
notication timing based on context in a JITAI for physical activity (Gonul et al., 2018).
In addition, new methods for reducing problems associated with early behavior selec-
tion when using reinforcement learning in health intervention applications are being
developed (Tabatabaei et al., 2018). Models that use concepts from
uid dynamics are
also being developed and could be used for physical activity interventions (Spruijt-Metz
et al., 2015). These types of computational models for behavior change interventions
10
are able to leverage the sensing and continuous presence advantages of mobile devices
such as smartphones.
2.1.2 Embodied Conversational Agent Interventions
Embodied conversational agents (ECAs) (Cassell, 2000) have also been used in a variety
of behavior change interventions. For example, studies have found that older adults will
walk more when given an ECA intervention compared to a control group (Bickmore
et al., 2013, King et al., 2013), and ECAs have been used to promote walking for people
with Parkinson Disease (Ellis et al., 2013). ECAs are also being developed to use the
behavior change technique of motivational interviewing (Lisetti et al., 2015). The use
of ECAs in behavior change interventions could also lead to higher engagement rates
than text-based interventions (Lisetti et al., 2013).
2.1.3 Socially Assistive Robot Interventions
SAR systems have been used in a variety of physical activity and behavior change
interventions across many domains and multiple time scales. In the following three
sections, we discuss related work involving SAR interactions in single-session, multi-
session, and long-term interventions.
2.1.3.1 Single-Session Interventions
There have been many studies designed around single-session SAR exercise interven-
tions. For example, the role of the robot as an exercise coach for older adults in a
single exercise session has been examined (Gadde et al., 2011, G orer et al., 2013, Lot
et al., 2018). One such study found that users preferred a robot designed with relational
capabilities to a nonrelational robot (Fasola and Matari c, 2012). A study comparing
the role of the robot as partner or coach for exercise has also been done (Schneider and
11
Kummert, 2018). In addition, some work has examined interactions that involve phys-
ical contact in exercise including a robot dance partner interaction with older adults
(Chen et al., 2017). A walking companion robot has also been deployed and evaluated
in a single session walking interaction in the context of an older adult living facility
(Mucchiani et al., 2017).
Early work using SAR systems to promote recovery post-stroke has shown promise
(Matari c et al., 2009, 2007). In those SAR systems, the robot was given a coaching role
and provided instructions and feedback. The systems also utilized multiple motivational
strategies, including personalizing the robot's feedback to the user's personality (Tapus
et al., 2008a).
2.1.3.2 Multi-Session Interventions
Multi-session SAR interventions for a variety of physical activities have also been de-
signed and evaluated both in situ in real-world environments and in the in-lab context.
For example, a SAR exercise coach system for older adults doing seated exercises was
evaluated in a multi-session study over a period of two weeks (Fasola and Matari c,
2013a,b), and another robot coach system for older adult exercise was evaluated in a
ve-week multi-session study (G orer et al., 2017). A robot coach for exercise generally
was tested over four exercise sessions within two weeks (Ramgoolam et al., 2014).
SAR systems have also been evaluated in multi-session studies in the rehabilitation
domain. For example, in the post-stroke rehabilitation domain, a SAR coach provided
information about performance and encouragement during an upper-extremity move-
ment task in a three session study (Wade et al., 2011). Additionally, recent work using
an iterative in situ design process in the rehabilitation clinic setting has begun evalu-
ating a SAR system for pediatric rehabilitation over multiple sessions (Carrillo et al.,
2018).
12
Using a robot in the role of a walking companion has been studied in several do-
mains and environments such as the clinical environment with people post-stroke where
participants engaged in multiple walking interactions with a robot walking coach (Gross
et al., 2017). In another study, a robot group walking companion was deployed for a
month in a care facility and engaged in multiple walking group sessions with older adults
with dementia (Hebesberger et al., 2016).
2.1.3.3 Long-Term Interventions
SAR systems have been used to support behavior change related to physical activity in
several long-term studies in the context of the user's every day life (such as, in-oce
and in-home). Kidd and Breazeal (2008), for example, conducted a six-week in-home
robot weight loss coach intervention and found that participants interacted more with
an embodied robot weight loss coach during the intervention than with a computer
system or paper log. Additionally, Reeder et al. (2011) found that participants had
positive reactions to a social robot designed to encourage taking breaks in the oce
setting.
Outside of the domain of physical activity, recent work has shown the potential for
SAR systems to aect behavior change over long-term interactions. A one month in-
home study of a SAR system designed to increase communication skills of children with
autism spectrum disorders (ASD) showed an increase in joint attention scores from the
beginning of the end of the intervention, and caregivers reported improvements in the
child's social behavior (Scassellati et al., 2018).
Additionally, some recent work has continued to focus on long-term studies with
SAR systems designed for multiple interactions in the home context. For example, a
SAR system designed for multiple interactions around a number concepts learning game
for children with ASD was used in a one month long in-home study (Clabaugh et al.,
13
2018a,b). In a single-case analysis, a correlation was found between game level diculty
and the child's focus of attention in this study (Clabaugh et al., 2018b). Another long-
term in-home study evaluated older adults' use of a robot designed to reduce social
isolation over a month-long period (Sidner et al., 2018). The acceptance of a robot
designed to engage in multipe interactions per day by older adults was also evaluated in a
long-term in-home study using 10 day phased deployments with participants completing
between one and three phases (De Graaf et al., 2015). These recent studies among others
build on early long-term SAR interaction work such as a six-month long intervention that
used a robot, in the role of music therapist, to guide people with cognitive impairments
through a music naming task with an adaptive diculty level (Tapus et al., 2009).
2.2 Validation Domains
The SAR physical activity behavior change support methods developed in this disser-
tation are evaluated in three domains. Details of each domain and the domain-relevant
physical activity targeted in these evaluations are described in the following three sec-
tions. Opportunities for SAR intervention are also presented.
2.2.1 Post-Stroke Upper Extremity Rehabilitation
Rehabilitation post-stroke typically involves repeatedly performing exercises toward
restoring lost motor function. This process can be tedious and tiring making it dicult
to generate the necessary motivation to keep performing the rehabilitation exercises that
are crucial for recovery. SAR has the potential to support the rehabilitation process
in a variety of ways, such as providing reminders, personalizing the robot's supportive
interaction patterns for each user, and providing motivational feedback (Winkle et al.,
2018).
14
In the United States, stroke is the leading cause of serious long-term disability
(Benjamin et al., 2018) with around 2.4% of people reporting stroke-related disability
(Centers for Disease Control and Prevention, 2009). Most stroke survivors (55-75%)
have sustained upper limb impairments (Lai et al., 2002). Additionally, recent work has
shown that in-clinic improvements to upper limb motor and functional ability do not
always result in increased upper limb use in everyday life (Bailey et al., 2015, Rand and
Eng, 2015, Waddell et al., 2017), and unequal amounts of arm use have been found to
persist in everyday life (Bailey et al., 2015, Rand and Eng, 2015, Shim et al., 2014).
This phenomenon is called learned nonuse (Taub et al., 2006).
Positive social interactions post-stroke have been found to be benecial; they pre-
dict arm movement and are negatively associated with post-stroke depression (Northcott
et al., 2016, Winstein and Varghese, 2018). Unlike physically-assistive assistive reha-
bilitation robots that do not have a social interaction component, a SAR system could
leverage the benets of social interaction as a crucial part of supporting post-stroke
rehabilitation in the chronic stages of recovery. Integrating SAR into the rehabilitation
process presents a possibility to provide motivation and support through social interac-
tion for in-home and self-directed in-clinic rehabilitation exercises between interactions
with occupational or physical therapists. In order to provide such support, a SAR
system must be equipped with an eective behavior change support method that can
encourage the post-stroke patient to use the stroke-aected arm in everyday life.
2.2.2 Adolescent Physical Activity
In 2007-08, an estimated 34.7% of children aged 6-19 were overweight or obese (Og-
den et al., 2010). Despite the well-established body of literature that demonstrates
the protective eects of physical activity against obesity and diabetes risks among
15
school-age adolescents (Janssen and LeBlanc, 2010), adolescents engage in an insuf-
cient amount of physical activity (Hallal et al., 2012) that is less than 6-11 year old
children (Troiano et al., 2008). Adolescent physical activity engagement exists at various
levels (i.e., environmental, interpersonal, and individual) (Craggs et al., 2011, Lubans
et al., 2008, Maturo and Cunningham, 2013, Sallis et al., 2000), and the importance of
social and peer support in prompting physical activity in adolescents is well established
(Craggs et al., 2011, Maturo and Cunningham, 2013). While positive communication
with friends about physical activity and joint participation are associated with higher
levels of physical activity (Maturo and Cunningham, 2013), adolescents, especially those
who are overweight and obese, are at risk for social interactions, including teasing and
stigmatization, that diminish motivation for physical activity (Rukavina and Li, 2008).
Physical activity promotion eorts that harness the supportive aspects of peer and social
interactions could be eective for this age group. As such, providing peer-level support
through a SAR intervention with the robot in the role of an exercise buddy could be
help to increase exercise behavior.
2.2.3 Reducing Sedentary Behavior in Older Adults
Sedentary behavior can increase cardiovascular disease, morbidity, and mortality (Young
et al., 2016) and has been found to decrease metabolic health (Owen et al., 2010). Seden-
tary behavior has also been found to take up 65-80% older adults' waking hours (Harvey
et al., 2015). Reducing sedentary behavior is important for older adults because reduc-
ing sedentary behavior has been associated with improved functional tness, increased
\successful aging," and lower rates of disability in activities of daily living independent of
time spent exercising (Dogra and Stathokostas, 2012, Dunlop et al., 2015, Santos et al.,
2012). A SAR system that utilizes a behavior change support method could encourage
older adults to take breaks from sitting in order to reduce sedentary behavior.
16
2.3 Behavior Change Mechanisms
There are ve behavior change mechanisms of action described by Connell et al. (2018)
that are targeted in this work. These include, beliefs about capabilities, social in
uences,
attitude towards the behavior, behavioral cueing, and reinforcement (Carey et al., 2018,
Connell et al., 2018, Johnston et al., 2018). These mechanisms and their relation to the
validation domains and the SAR behavior change support methods are described in the
following four sections.
2.3.1 Beliefs About Capabilities
A person's beliefs about their capabilities to perform a task is also known as self-ecacy
(Bandura, 1997). In the domain of post-stroke rehabilitation, self-ecacy is thought to
mediate motor performance, and may underlie many facets of post-stroke outcomes,
including quality of life and performance of activities of daily living (ADLs) (Jones and
Riazi, 2011, Korpershoek et al., 2011), balance (Hellstrom et al., 2003), and upper-
extremity limb choice (Chen et al., 2013). Self-ecacy has also been found to positively
in
uence exercise behavior in people post-stroke (Shaughnessy et al., 2006). Because
of the role of self-ecacy in stroke rehabilitation and behavior change, SAR support
methods that improve self-ecacy could potentially lead to positive behavior change.
In the evaluation study described in Chapter 3, the design of the feedback SAR support
method incorporates the behavior change technique of verbal persuasion about capability,
which could impact self-ecacy (Johnston et al., 2018).
2.3.2 Social In
uences
Two BCTs that aect the social in
uences mechanism of action are used in this work
in the design of SAR behavior change support methods (Johnston et al., 2018). The
rst, is the social comparison BCT, which involves bringing the user's attention to the
17
performance of others for comparison to their own performance (Michie et al., 2011).
This technique of social comparison is used in the feedback SAR behavior change support
method evaluated in Chapter 3. Social comparative feedback has previously been shown
to improve performance on motor learning tasks (Lewthwaite and Wulf, 2010, Wulf et al.,
2012), self-ecacy for motor learning (Wulf et al., 2012), and intrinsic motivation when
given by a social agent without praise (Mumm and Mutlu, 2011). The study presented
in Chapter 3 evaluates the use of two types of social comparative feedback, self and
other, against a control of no comparative feedback in the post-stroke rehabilitation
domain.
The second BCT is social support, which involves getting support from other people
such as a peer or buddy (Michie et al., 2011). We use this technique for the backstory
SAR behavior change support method. For this method backstory is used to place
the robot in the role of a peer-level exercise buddy. Few studies exploring the use of
backstory with robots have been performed to date. For example, the Roboception-
ist characters all had a strong backstory revealed during interactions (Simmons et al.,
2011); a backstory about dragon race training was used in a study focused on children
and nutrition (Short et al., 2014). Both studies used ctional backstories, but no com-
parison of ctional vs. realistic backstories has been done in SAR. Such comparison has
been explored with virtual humans. Kang and Gratch found that people with high levels
of social anxiety disclosed more to a virtual human counselor with a human backstory
than one with a computer backstory (Kang and Gratch, 2011). A long-term study by
Bickmore et al. found that people engaged in more discussions with a virtual charac-
ter with a ctional rst person backstory than one with a backstory presented in the
3rd person (Bickmore et al., 2009). In order to explore the potential of backstory for
creating a socially supportive relationship with the robot and maintaining engagement
over multi-session studies that relate to behavior change, we conducted an evaluation
18
study comparing a ctional to a realistic robot backstory for a robot exercise buddy for
adolescents as is described in Chapter 4.
2.3.3 Attitude Towards the Target Behavior
The messaging SAR behavior change support method is targeted at in
uencing the
attitude towards the behavior mechanism of action by using the information about emo-
tional consequences BCT, which involves providing information about the emotional
consequences of the behavior (Connell et al., 2018). Studies suggest aective attitudes
or beliefs, which focus on aective outcomes such as enjoyableness, towards a new be-
havior could be more strongly connected with doing a new behavior than cognitive or
instrumental attitudes, which focus on practical outcomes such as how benecial the be-
havior is (Conner, 2013, Conner et al., 2015, Lowe et al., 2002). Development of these
aective attitudes could be an important focus of behavior change interventions. Sev-
eral intervention studies comparing the eectiveness of aective messages to cognitive
or instrumental messages have found that participants receiving aective messages had
increases in exercise behaviors (Conner et al., 2011, Morris et al., 2015). Text messaging
could be a useful way to send these aective messages in an intervention. In one study,
inactive participants who received text messages containing aective messages had a
larger increase in physical activity than those receiving instrumental messages (Sirriyeh
et al., 2010). Conner et al. (2011) suggest that aective attitudes could mediate the
connection between aective messages and behavior meaning that using aective mes-
sages in an intervention could increase aective attitudes, which could then increase
performance of the behavior. Aective messaging is a promising behavior change sup-
port method that could be used in SAR interventions. The development and evaluation
of a set of aective messages for SAR behavior change support of physical activity is
described in Chapter 4.
19
2.3.4 Cueing and Reinforcement
The behavior change mechanisms of action of behavioral cueing and reinforcement are
aected by the BCTs of habit formation, prompts/cues, and social reward (Johnston
et al., 2018). These three BCTs are all used in the SAR reminder and social reward
decision making behavior change support method.
Habit formation occurs when the process of a cue triggering a behavior is reinforced
with a reward repeatedly over time (Wood and R unger, 2016). Because habit formation
involves distinct and automatically measurable components (cues and behaviors) it can
be leveraged in SAR systems where the robot can respond to these automatically sensed
inputs. Key terms associated with the process of habit formation are detailed next.
The process of consciously forming a new habit begins with choosing a goal. It is
important in habit formation interventions that this goal be something that the user
chooses and has a desire to achieve. This is because agency is important aspect of
behavior change for health (Hekler et al., 2013). In addition, the process of forming
a habit involves moving from goal-directed to automated behavior (Wood and Neal,
2007). In this process, goals in
uence habit formation because goal-based motivation
can encourage more repetitions of a behavior which can increase the habit strength
(Wood and Neal, 2007).
Cues are a central component of habit formation. When intentions to perform a
behavior are linked to cues, the automaticity, or how automated the habit is, increases
(Orbell and Verplanken, 2010). There are several dierent types of cues that can trig-
ger a behavior. These include environmental cues where something in the environment
changes, event or activity cues where a prior event or action triggers behavior, and time
cues where a behavior occurs at a certain time or after a certain amount of time has
elapsed (Judah et al., 2013, Wood and R unger, 2016). Reminder or notication cues
are easy to implement with smartphone applications but may have disadvantages. For
20
example, the core component of many commercially available smartphone applications
targeting medication adherence rely on time-based notications that act as medication
reminders (Stawarz and Cox, 2013, Stawarz et al., 2014). These types of reminders,
however, can limit habit and automaticity increases despite increasing behavior repeti-
tion (Stawarz et al., 2015). This indicates that there might be a trade o between using
reminders or notications as cues in order to increase adherence and creating lasting
behavior change.
One way of associating a cue from everyday life with a behavior is through implemen-
tation intentions, statements that provide details of an intention to perform a behavior
upon encountering a cue (Gollwitzer, 1999, Gollwitzer and Brandst atter, 1997). For
example, an implementation intention for running might be, \When I arrive home and
see my running shoes next to the door, I will go for a run." Setting implementation in-
tentions has been found to be a successful method for turning a goal into a habit (Aarts
and Dijksterhuis, 2000). Implementation intentions have been used successfully in inter-
ventions to increase the number of times a target behavior is completed (Gollwitzer and
Brandst atter, 1997, Latimer et al., 2006). Setting implementation intentions has also
been found to increase prospective memory (Chasteen et al., 2001), a person's ability
to remember to do something after encountering a specic cue in the future (Einstein
et al., 2005, Zogg et al., 2012). Prospective memory is important for habit formation
because it increases the likelihood of implementation intentions being remembered when
the cue occurs.
There are a variety of ways that prospective memory can be strengthened in an in-
tervention. One is by providing implementation intention reminders as a person forms
a new habit. For example, implementation intention plan reminders sent via text mes-
sage have been found to increase the daily habit of walking (Prestwich et al., 2010). In
addition, reminders can increase performance on tasks that involve prospective memory
21
(Henry et al., 2012). Another method for increasing prospective memory is episodic fu-
ture thinking (EFT), or thinking about events that may occur in the future (Atance and
O'Neill, 2001). The practice of EFT can increase prospective memory for a task (Neroni
et al., 2014), and EFT ability has been shown to correlate with prospective memory in
younger adults (Terrett et al., 2015). Both implementation intention reminders and the
practice of EFT could be useful strategies for increasing prospective memory in order
to build habit strength.
Another key component of habit formation is habit reinforcement. This is the process
by which a habit is reinforced by a reward after a person completes a behavior after
encountering a cue (Wood and R unger, 2016). Rewards can be given based on a pre-
dened schedule. There are two main types of reinforcement schedules, interval and
ratio, which dene when a decision about providing a reward is made (Skinner, 1965).
In an interval schedule a reward decision is made after a certain amount of time has
elapsed and a specied number of cue-behavior pairs have been completed; in a ratio
schedule a reward decision is made after a certain number of cue-behavior pairs have
been completed (Strohacker et al., 2014). The reward decision for interval and ratio
schedules is made using a xed, variable, or random method. Using the xed method a
reward is always given when the decision point occurs (Strohacker et al., 2014). Using a
variable method a reward is not given at every decision point, but rather probabilistically
within a set number of decision points (Haw, 2008). Finally, using a random method
each decision point has some probability that a reward will be given (Haw, 2008).
A SAR could potentially provide social rewards using one of these schedules to a
person forming a new habit soon after they complete a habit behavior. Social feedback
given by a robot has been shown to to be more eective than factual feedback for energy
saving behaviors (Midden and Ham, 2009). A SAR giving feedback as a social reward
could be eective for habit reinforcement.
22
There have also been many behavior change interventions that use the concept of
habit formation. These include a
ossing intervention in which participants who
ossed
after tooth brushing formed stronger
ossing habits then those who
ossed before brush-
ing (Judah et al., 2013). Another study found that participants with spinal cord injuries
who formed intentions for when they would exercise exercised more than a control group
that did not form intentions (Latimer et al., 2006). Despite the eectiveness of many
habit formation intervention techniques most commercial smartphone applications that
relate to behavior change do not include these techniques (Stawarz et al., 2014). In ad-
dition, to our knowledge, there have not been any SAR behavior change interventions
that utilize habit formation techniques.
In this dissertation, the SAR behavior change support method of reminder and social
reward decision making uses the concept of habit formation to provide reminders and
social rewards at key times during the behavior change process in order to support the
formation of a new habit. The SAR model of habit formation support developed in
this work is described in Chapter 5 and the evaluation of this model in the domain of
reducing older adult sedentary behavior is described in Chapter 6.
2.4 Summary
In this chapter the background and work related to this dissertation was presented.
Related work using technology such as smartphones, mobile devices, ECA interactions,
and SAR interactions in physical activity behavior change interventions was discussed.
The promising results from SAR behavior change interventions suggest that SARs could
be useful for supporting behavior change. Background related to the validation domains
and the relevant physical activity needs in those domains was presented. The behavior
change mechanisms of action targeted by the SAR behavior change methods developed
in this dissertation as well as the BCTs incorporated into the design of those methods
23
were discussed. The feedback support method uses social comparative feedback towards
the goal of in
uencing self-ecacy for a task. The backstory support method uses
backstory towards the goal of creating a relatable robot buddy that can provide social
support. The messaging support method uses messages about the emotional benets
of exercise towards the goal of increasing positive attitudes towards exercise. Finally,
the reminder and social reward decision making support method uses the concept of
habit formation to provide reminders and social rewards at appropriate times towards
the goal of habit formation. In the next chapter, the design and evaluation of the SAR
comparative feedback support method are described.
24
Chapter 3
Design and Evaluation of Socially
Assistive Robot Comparative
Feedback Support
In this chapter, the design of the SAR feedback behavior change support method
is described, and an evaluation study comparing two types of SAR comparative
feedback, self and other, to a control in the domain of post-stroke upper ex-
tremity rehabilitation is detailed. The results of this study are presented and
discussed.
3.1 Related Work
Robotic systems have been used extensively for post-stroke motor rehabilitation includ-
ing both upper and lower limb rehabilitation (Hesse et al., 2003, Weber and Stein, 2018).
Upper limb rehabilitation robot systems have mostly been used for physically-assistive
contact-based interactions with either an end-eector or exoskeletal robot design in
which the robot guides the stroke-aected limb or provides resistance to its movement
25
during a specic rehabilitation exercise (Weber and Stein, 2018). Such system have the
potential to provide guided exercises more accessibly than conventional rehabilitation
therapy (Blank et al., 2014), and have been shown to be as eective as conventional
therapy and to, in some cases, improve on it (Norouzi-Gheidari et al., 2012, Veerbeek
et al., 2017). However, such systems have been developed for improving targeted motor
function, not increasing aected arm use, i.e., reducing learned nonuse.
In conventional stroke rehabilitation therapy, several methods designed specically
for reducing learned nonuse have been studied. For example, using Constraint Induced
Movement Therapy (CIMT) to target learned nonuse by limiting use of the less af-
fected arm, has shown improvements over prior conventional therapy (Wolf et al., 2006).
Adding a \transfer package" to CIMT focused on transferring in-clinic improvements
to everyday life has also been shown to increase aected arm use over CIMT alone
(Gauthier et al., 2008). A recent approach to reducing nonuse has provided real-time
feedback about arm use in everyday life via wrist-worn devices that measure movement
and use vibration as a feedback mechanism (Wei et al., 2018). In addition, adding
movement reinforcement to a virtual reality (VR) arm reaching task by increasing the
appearance of arm movements in the virtual environment has reduced nonuse, an eect
that might be mediated by increased self-ecacy (Ballester et al., 2016). New methods
of measuring learned nonuse based on arm choice in a reaching task that can be used
relatively quickly in the clinic setting could enable further development of interventions
to increase use of the aected arm in daily life (Kim et al., 2018).
3.2 Methodology
To explore the eect of comparative feedback given by a SAR coach on post-stroke
upper-extremity reaching performance and self-ecacy, we developed an autonomous
robot coach system and conducted a between-subjects study to examine the eects of
26
three dierent types of feedback given by the robot capable of guiding a participant
through a button-pressing task and providing dierent types of performance feedback.
1
3.2.1 Robot and System Design
The experiment system we designed, show in Figure 3.1, consisted of three communi-
cating components: 1) a fully autonomous SAR system, 2) a task button board and
keypad, and 3) a set of sensors for data collection. Each component is described below.
(a) Image of Bandit robot and button board.
Keypad
Participant
Bandit
Kinect and
USB Camera
HD and USB
Cameras
(b) Diagram of experimental setup.
Figure 3.1: Experimental setup with button board and Bandit robot.
We used Bandit, a 19 degree-of-freedom (DOF) upper torso humanoid robot with 7
DOF in each arm, 2 DOF in the head, 2 DOF in the mouth, and 1 DOF in the eyebrows.
The button board consisted of ten goal buttons and two home buttons. Each goal
button was identied with a randomly assigned number from 1 to 10. There were also
two \home buttons" labeled as \L Home" and \R Home". Those buttons were used to
ensure that the participant's starting hand positions for each reach were in the desired
locations. The rest of the buttons on the board were in a semicircular layout, in two
1
This chapter describes a study which was completed in collaboration with other researchers at the
University of Southern California including Elaine Short, Eric Wade, and Maja J Matari c.
27
rows, and at ve angles relative to the longer side of the table closest to the participant
(0
, 45
, 90
, 135
, and 180
). The semicircular rows of buttons were located at 13.5 cm
and 27 cm from the center point of the semicircles for the near and far rows respectively.
The home buttons were inset 5 cm from the edge of the table, and were 5 cm oset
from the line connecting the two buttons closest to the home buttons. Additionally, a
Bluetooth keypad was placed on the button board. While the buttons on the board
were used to track the participant's performance on the task, the keypad was used to
collect the participant's responses to self-report probes.
The data collection sensors included two USB cameras, a high-denition (HD) cam-
era, and two wrist-worn accelerometers. The data from these sensors were collected for
post-hoc annotation and analysis. Data from the Microsoft Kinect RGB-D sensor were
used in real-time robot control, as described below.
3.2.1.1 System Architecture
An overview of the fully autonomous SAR system is show in Figure 3.2. A state ma-
chine was used to control the
ow of the interaction from the introduction to the end
of the session. A trial manager state was used in between trials, probes, and breaks,
to determine the next state. The robot's perceptual system included three inputs: the
button presses (from the board), self-report response entries (from the keypad), and the
participant's 3D joint positions (from the Kinect). In addition to verbal instructions,
probes, and feedback, the robot performed complementary expressive non-verbal behav-
iors, including head nods and head shakes, arm beat gestures during speech, pointing at
and looking at buttons when referring to them, opening the mouth at the beginning of a
speech phrase, and closing it at the end of a phrase. The 3D position of the participant's
head, from the Kinect data, was used as an input to the robot's head gaze controller,
allowing the robot to look at the participant during the interaction.
28
The system performed fully autonomously. The experimenter sat behind a curtain
during the experiment and could be summoned with a bell located on the table.
Robot Perception
Button Presses
Robot Behavior
Beat
Gestures
Pointing
Gestures
Head Nods
Speech Head Gaze
Mouth
Open/Close
Interaction State Machine
Trial
Manager
Trial
Break
Closing
Introduction
Button-
Press Probe
Self-Report
Probe
End
Person 3D Joint
Positions
Keypad Entries
Figure 3.2: Overview of fully autonomous SAR coaching system.
3.2.1.2 Button-Pressing Task
The layout of the button board, described above, was based on the design used by Chen
et al. in a similar seated upper-extremity reaching task designed to measure self-ecacy
in post-stroke reaching (Chen et al., 2013). A similar reaching task has also been used
to measure stroke aected arm nonuse (Han et al., 2013).
During the interaction the robot gave participants verbal instructions for completing
the button-pressing task. The task involved pressing specied buttons on the board as
instructed by the robot. A correct button press involved rst pressing both of the home
buttons, then releasing the home button of the hand specied by the robot's instruction,
using that hand to press the button specied by the robot, and then returning the hand
to again press the previously released home button, as show in Figure 3.3. A chime
sound was played after each successfully completed button press, and the robot gave
corrective feedback when the button press was not correctly/successfully completed.
29
For example, if the participant used the incorrect hand to press the indicated button,
the robot said, \"So close! You didn't use the arm I asked you to use though, so let's
try again."
Figure 3.3: Example button press for request from robot to \Please press the home
buttons with both hands and then use your right hand to press button four."
3.2.2 Study Design
The study focused on two types of comparative feedback delivered by a SAR coach,
self-comparative and other-comparative, and their impact on self-ecacy. A between-
participants design was used with three conditions: 1) no comparative feedback (no-
CF), 2) self-comparative feedback (self-CF), and 3) other-comparative feedback (other-
CF). In the no-CF condition, the control condition, participants did not receive any
comparative feedback. In the self-CF condition, they received feedback such as \During
the past few trials, you averaged a time ofs seconds faster than we would have predicted
based on your prior performance." In the other-CF condition, they received feedback
such as \You had an average time that was s seconds better than others with your
ability level." The time dierence in the feedback, s, was always 10% of the actual time
it took the participant to press the button, and the feedback was always positive; it
was not a true comparison. Only positive feedback was used because such feedback has
been shown to increase perceived competence (Badami et al., 2011).
30
3.2.3 Measures
A combination of functional assessment, questionnaire, and behavioral interaction data
were collected during the study. Occupational therapists characterized the participants'
stroke-related impairment using a functional assessment, the Fugl-Meyer Assessment of
Upper Extremity Motor Performance (Sanford et al., 1993), before the button-pressing
task. We also administered the Condence in Arm and Hand Movement (CAHM) scale
which was used by (Chen et al., 2013), a 20-item self-report measure of self-ecacy
for activities of daily living. Furthermore, we administered an interaction question-
naire based on (Lombard et al., 2000b, McCroskey and McCain, 1974b), and (Jung
and Lee, 2004b), which measures participants' perception of the robot and the inter-
action. A questionnaire with several open-ended and demographics questions was also
administered. Some additional assessments of personality, button diculty, and a
ow
questionnaire were administered, but their analysis is beyond the scope of this work.
We operationalized arm reaching self-ecacy by dening two behavioral measures
of self-ecacy, a self-report probe and button-press probe. For the self-report probe, the
participants were asked to use a keypad with their non-stroke-aected limb and enter
the number of the most dicult button they thought they could press quickly with their
stroke-aected limb. For the button-press probe, the participants were asked to press
the most challenging button they could reach quickly with their stroke-aected limb.
This operationalization of arm reaching self-ecacy was designed to quickly measure
a participant's stroke-aected arm reaching self-ecacy for the button-pressing task
by seeing which is the most dicult or challenging button participants chose to press
or self-reported they could reach quickly. This measure diers from the arm reaching
self-ecacy measure developed by Chen et al., which measured reaching self-ecacy by
asking participants to choose which of two hand-target combinations they were more
condent in reaching and repeated this process for 180 sets of hand-target combinations
31
(Chen et al., 2013). The Chen et al. measure takes about 20 minutes to complete;
it was therefore not suitable for our study, which required a self-ecacy measure that
could be acquired eciently at many points during the button-pressing task.
Input from the button board was used to collect timing information about the par-
ticipants' button presses, including press time and delay time. Press time, the time
taken to press the goal button, was measured from when the participant released the
correct home button until when the participant pressed the home button again. Delay
time, the delay before beginning a press, was measured from when both home buttons
were pressed after the robot began the button press instruction until when the par-
ticipant released the correct home button. Because the beginning of the delay time
measurement occurs after the robot beings the button press instruction and the home
buttons are pressed, it did not start at the same time for each press. For example, if
the participant pressed the home buttons before the robot began the instruction, then
the timing started at the beginning of the instruction, but if the participant pressed the
home buttons at the end of the robot's instruction, then the timing started at the end
of the instruction.
At the end of the session, participants were given the option to continue with the
button-pressing task, as a measure of engagement through time on task.
3.2.4 Hypotheses
We designed the following hypotheses to address the goals of this research:
H1: Comparative feedback conditions (self-CF and other-CF) will produce higher
self-ecacy (compared to no-CF) as measured by self-report probes, button-press
probes, and CAHM scores.
32
H2: Comparative feedback conditions (self-CF and other-CF) will produce better
performance (as compared to no-CF) in practice session, as measured by button
press times and delay times.
H3: Participants will perceive the robot coach more positively in comparative
feedback conditions (self-CF and other-CF) than in the no-CF condition, as mea-
sured by the interaction questionnaire.
3.2.5 Procedure
The participant session with the robot consisted of a set of pre-interaction question-
naires, the button pressing interaction with the robot, and post-interaction question-
naires. The pre-interaction questionnaires included the CAHM and two personality
questionnaires. The Fugl-Meyer Assessment of Upper Extremity Motor Performance
was conducted by an occupational therapist either immediately before or separately
from the session with the robot. After completing the questionnaires, the participant
was instructed to sit at the table in front of the button board. The robot rst intro-
duced the task and then asked the participant to complete a series of button press trials,
where a trial is dened as a specic instruction such as \Please press the home buttons
with both hands and then use your right hand to press button four." Each participant's
session consisted of a total of 100 trials.
Familiarization
20 Trials
Main
60 Trials
Retention
20 Trials
Time on Task
Optional up to 140 Trials
Break
Optional up
to 5 minutes
Break
Non-optional
5 minutes
Figure 3.4: Overview of interaction sequence.
33
In each of the trials, the robot asked the participant to press both home buttons by
pressing the \L home" button with the left hand and the \R home" button with the
right hand. It then asked the participant to lift either the right or left hand, press one of
the numbered goal buttons with that hand, while keeping the other hand on the home
button already being pressed, and then return the lifted hand to press the previously
pressed home button. A chime sounded when the participant returned the lifted hand
to its home button. If the participant lifted the hand that was not involved in the
button press o the home button at any point, the robot detected the error and asked
the participant to keep that hand on its home button. After each trial, the robot gave
performance feedback by telling the participant how long (in seconds) it took to press
the goal button. In the self-comparative and other-comparative conditions, comparative
feedback was given by the robot after the performance feedback for each trial. In the
control condition, no comparative feedback was given.
The 100 button press trials in the session were divided into three groups: 20 fa-
miliarization trials, 60 main trials, and 20 retention trials, as shown in Figure 3.4. An
optional 5-minute break was oered to the participant mid-way through the main tri-
als, and a mandatory 5-minute break was given before the retention trials. After the
retention task, each participant was given the option to continue in 4-trial increments,
up to 140 trials total. The button presses were designed so that during each set of 20
trials, each button was pressed once with each arm. The arm used to press the but-
ton alternated, and the button press ordering was randomized within each 20-trial set.
The order of the button presses based on the numbering on the button board was the
same for every participant, and the button presses started with the stroke-aected arm.
During the familiarization and main trials, the two self-ecacy probes (self-report and
button-press) were administered alternately, one probe every other button press (a total
of 20 times per probe). An overview of the trial sequence, self-ecacy probe timing,
34
Table 3.1: Details of trail sequence, self-ecacy probe timing, and continuation question
timing for each section of the interaction.
Section Trial Sequence Self-Ecacy
Probe
Continuation
Question
Familiar-
ization
Instruction, button
press, performance
feedback
Button-press or
self-report probe
alternating after
every two trials
None
Main Instruction, button
press, performance
feedback,
comparative
feedback (in
self-comp and
other-comp
conditions)
Button-press or
self-report probe
alternating after
every two trials
None
Retention Instruction, button
press, performance
feedback
None None
Time on
Task
Instruction, button
press, performance
feedback
None After every four
trials
and continuation question timing for each section of the interaction is shown in Table
3.1.
After the button pressing interaction, participants were asked to complete a set of
questionnaires that included the button diculty questionnaire, the CAHM, the interac-
tion questionnaire, some open ended questions, a
ow questionnaire, and a demographics
questionnaire. Finally, the participants were debriefed about the purpose of the study
and informed of the comparative feedback manipulation.
3.2.6 Study Population
According to the study inclusion criteria, participants were any time post-stroke and able
to extend the stroke-aected arm above a table while seated. There were 28 participants
35
who enrolled, and of those, four did not complete the study: two did not complete
because of technical issues (one in other-CF and one in no-CF), and two opted out of
completing the study (both in no-CF). Of the remaining 24 participants, one participant
(in no-CF) was not included in the analysis because of technical issues, and another (in
no-CF) because of a delay in completing the study. The remaining 22 participants are
included in the data analysis. Of those, eight were in self-CF, seven were in other-CF,
and seven were in no-CF. We do not have demographic data and functional assessment
data for all of the 22 participants; summary information of the available data and the
number of participants represented are shown in Table 3.2.
Table 3.2: Participant characteristics for the 22 participants included in the data anal-
ysis.
Age FMUE Gender Stroke Aected Arm
54.7 (SD = 9:6) 44.5 (SD = 14:4) 3 women, 18 men 11 Right, 11 Left
N = 20 N = 20 N = 21 N = 22
FMUE = Fugl-Meyer Assessment of Upper Extremity Motor Performance
3.3 Results
The mean press times and delay times for condition, order in which the buttons were
pressed, angle of the button with respect to the arm pressing the button, distance of the
button from the arm, and limb used to press the button are shown in Table 3.3. Due to
a typo in the retention trials, button eight was pressed twice with the stroke-unaected
arm by all participants, and button four was not pressed with the unaected arm. In
the following analysis we did not include the second press of button eight during the
retention trials.
36
Table 3.3: Mean press and delay times and standard deviations for condition, button
press order, angle, distance, and limb.
Press Time Delay Time
Condition
no-CF 1.237 (SD = 0.767) 5.737 (SD = 0.863)
self-CF 1.688 (SD = 1.002) 5.594 (SD = 1.044)
other-CF 1.487 (SD = 0.991) 6.021 (SD = 1.279)
Order
Familiarization 1.794 (SD = 1.333) 5.983 (SD = 1.361)
Main 1 1.502 (SD = 0.864) 5.744 (SD = 0.890)
Main 2 1.438 (SD = 0.817) 5.858 (SD = 0.977)
Main 3 1.378 (SD = 0.801) 5.634 (SD = 1.083)
Retention 1.280 (SD = 0.705) 5.591 (SD = 0.962)
Angle
0
1.341 (SD = 0.850) 5.708 (SD = 1.107)
45
1.483 (SD = 0.972) 5.770 (SD = 1.098)
90
1.491 (SD = 0.882) 5.784 (SD = 0.949)
135
1.577 (SD = 1.116) 5.778 (SD = 1.203)
180
1.512 (SD = 0.878) 5.780 (SD = 1.007)
Distance
Near 1.380 (SD = 0.820) 5.678 (SD = 1.033)
Far 1.583 (SD = 1.052) 5.851 (SD = 1.114)
Limb
Aected 1.792 (SD = 1.009) 5.834 (SD = 1.103)
Unaected 1.163 (SD = 0.759) 5.692 (SD = 1.045)
3.3.1 Press Times
Using a mixed-design ANOVA, we found that the time to press a button was signicantly
aected by the limb used (F (1; 19) = 42:36;p < :001), as well as by the order in
which the buttons were pressed (F (1; 19) = 20:95;p<:001), the distance to the button
(F (1; 19) = 23:90;p<:001), and the angle of the button relative to the arm pressing it
(F (1; 19) = 34:01;p<:001). A post-hoc t-test conrmed that the time to press a button
with the aected limb was higher than the time to press with the unaected limb, with
p<:001. The familiarization press took signicantly longer than all subsequent presses
(p < :001), and the rst press took signicantly longer than the retention press (p =
:005). Finally, the farther buttons took longer to press than the nearer buttons, withp<
37
Familiarization Main 1 Main 2 Main 3 Retention
Order
0.75
1.00
1.50
2.00
2.50
3.00
3.50
Press Time (Seconds)
***
***
***
***
**
Mean Press Time for Limb, Distance, and Order
Affected-Far
Affected-Near
Mean
Unaffected-Far
Unaffected-Near
Far Near
Distance
0.75
1.00
1.50
2.00
2.50
3.00
3.50
Press Time (Seconds)
Mean Press Time
for Limb and Distance
Affected
Unaffected
Figure 3.5: Mean press time for limb, distance, and order; and limb and distance
(***p<:001; **p<:005). Error bars depict 95% condence interval.
:001. Pairwise comparisons of the angles reveal a signicant dierence between 0
and
135
(p<:005). A signicant interaction eect was found between limb, distance, and
order (F (1; 19) = 10:72;p<:005)) (Figure 3.5), between distance and limb (F (1; 19) =
7:15;p < :05), and between order and limb (F (1; 19) = 13:83;p < :005). A signicant
interaction between limb, distance, order, angle and condition (F (2; 19) = 3:69;p<:05)
was also found. All post-hoc t-tests were performed with Bonferroni correction.
3.3.2 Delay Times
One participant's delay times were consistent outliers; that participant was excluded
from the following calculations for delay time. Using a mixed-model ANOVA, we
found a signicant eect of condition on delay time (F (2; 18) = 4:602;p < :05). A
post-hoc t-test showed that the other-comparative had a longer delay than both the
self-comparative condition and the no-comparative condition, both with signicance
38
p < :001, and the no-comparative condition had a longer delay time than the self-
comparative condition, with p < :05 (Figure 3.6). We again found main eects of
distance (F (1; 18) = 14:27;p < :005) and order (F (1; 18) = 17:84;p < :001) on the
delay time, with farther buttons having a longer delay than nearer buttons (p<:001).
Furthermore, there were signicant dierences between the familiarization and rst
press (p<:05), familiarization and third press (p<:001), familiarization and retention
(p < :001), second press and the third press (p < :05), and second press and reten-
tion (p < :005). Finally, we found an interaction eect between distance and limb
(F (1; 18) = 6:31;p < :05) and order and limb (F (1; 18) = 12:66;p < :005), seen in
Figure 3.6. Again, all post-hoc t-tests used Bonferroni correction.
Familiarization Main 1 Main 2 Main 3 Retention
Order
5.2
5.4
5.6
5.8
6.0
6.2
6.4
6.6
Delay Time (Seconds)
*
***
***
*
**
Mean Delay Time for Limb and Order
Affected
Mean
Unaffected
Far Near
Distance
5.2
5.4
5.6
5.8
6.0
6.2
6.4
6.6
Delay Time (Seconds)
Mean Delay Time
for Limb and Distance
Affected
Unaffected
Other-CF Self-CF No-CF
Condition
5.2
5.4
5.6
5.8
6.0
6.2
6.4
6.6
Delay Time (Seconds)
***
***
*
Mean Delay Time
for Condition
Figure 3.6: Mean delay time for limb, distance, and order; limb and distance; and
by condition (***p < :001; **p < :005; *p < :05). Error bars depict 95% condence
interval.
3.3.3 Self-Ecacy
In our rst exploratory analysis of the participant responses to the self-ecacy probes
(see (Swift-Spong et al., 2015)) we saw a heterogeneity in how participants responded to
the probes with some participants choosing a variety of buttons, some always choosing
39
more dicult buttons, and some choosing dierent buttons for the button-press and
self-report probes. We therefore sought to understand if there were any commonalities
or groupings among participants based on their self-ecacy probe responses and their
button press performance. We conducted an unsupervised machine learning analysis by
clustering participants into groups based on their self-ecacy probe choices and their
button press times. We included the button press times for button presses during the
three main trials in order to add information about each participant's aected arm
button pressing performance to the clustering analysis.
6 4 2 0 2 4 6
Principal Component 1
3
2
1
0
1
2
3
4
Principal Component 2
Self-CF
Other-CF
No-CF
Cluster 1
Cluster 2
Cluster 3
Cluster 4
Cluster 5
Figure 3.7: Participant clusters shown on a 2D plot of the rst two principal components.
Points are colored by condition and shapes correspond to the cluster. Clusters are
distinctly separated, number of participants in each cluster ranges from two (cluster 5)
to six (cluster 2); there is no apparent relationship between cluster and condition.
The analysis included the 22 participants who had complete data for the self-ecacy
probes and press times. We represented each participant with a 30-item vector that
included, for each button, the average press time for the three main trial button presses
with the stroke-aected arm, the number of times the button was chosen as a self-report
probe, and the number of times the button was chosen as a button-press probe. We then
40
standardized the data and applied Principal Component Analysis (PCA) to reduce the
30 features to ve principal component (PC) features, which together explained 71.8%
of the variance (PC1 explained 32.3%, PC2 explained 14.4%, PC3 explained 10.8%, PC4
explained 7.37%, and PC5 explained 6.97%). We then applied K-means clustering using
these ve principal component features to group participants into ve clusters. The 2D
plot of the clusters using the rst and second principal components is show in Figure
3.7. There were four participants in cluster 1, six in cluster 2, ve in cluster 3, ve in
cluster 4, and two in cluster 5. Using this visualization, we can observe that the ve
clusters did not overlap in this 2D space. We also did not nd any distinct relationship
between the clusters and condition.
Using the press times and self-ecacy probe data for participants in each cluster, we
calculated the mean press times per button for each of the main trial presses with the
stroke-aected arm, the mean number of times each button was chosen for a self-report
probe, and the mean number of times each button was chosen as a button-press probe.
We visualized these cluster means using heat maps, as show in Figure 3.8.
41
Cluster 1
2.4
2.4 2.9 2.7
2.1 2.6
2.5 2.6 2 2.2
Mean Press Times
6.8
1.5 0.75 0
0 0.25
5.2 0.75 4.5 0.25
Mean Self-Report Probes
6.8
1.2 0.5 0
0 0
1.5 0 10 0
Mean Button-Press Probes
Cluster 2
1.4
1.5 1.5 1.5
1.6 1.3
1.4 1.5 1.3 1.3
10
7.7 0.17 0.5
0 0
0.83 0.17 0 0.17
7.7
7.3 2.3 1.3
0 0
1.2 0 0 0.17
Cluster 3
1.3
1.3 1.1 1.6
1.3 1.4
1 1.1 1.2 1.1
1.2
1.6 0.4 1.2
1 1.2
0.6 4.4 6.2 2.2
0.4
0.8 0.6 0.6
0.2 1.6
1 0.6 12 2
Cluster 4
2.2
2.7 2.3 2.5
2.6 2.3
2.2 2.3 1.9 2.1
2.2
3 4.2 1.8
2.4 1.2
2 1 1.4 0.8
0.8
1 3.4 1.8
1.6 3.2
0.6 1.8 5 0.8
Cluster 5
0.78
0.87 0.78 1.1
1.1 0.92
0.67 0.72 0.73 0.68
0
1.5 0 0
0 0
14 3.5 0 0.5
0.5
0.5 0 0
0 0
18 0 0.5 0
0.0 s 1.0 s 2.0 s 2.9 s
Press Times
0.0 selections
4.0 selections
8.0 selections
12.0 selections
16.0 selections
18.5 selections
Probe Selections
Figure 3.8: Heatmap visualization of press times and responses to self-ecacy probes for
each cluster of participants. Each row represents a cluster, and each column represents
the mean press times, mean number of self-report probes, and mean number of button-
press probes. In each heatmap, the buttons are displayed in a grid-like pattern with the
location of each cell corresponding to the location of a button relative to the bottom
center of the heatmap. The buttons immediately adjacent to the bottom center of the
heatmap are the near buttons, and the buttons one button away from the bottom center
are the far buttons. The angles of the buttons in the heatmap relative to the bottom
center starting from the right side of the heatmap are the same as the button angles
relative to the stroke-aected arm.
42
We see that cluster 5 tended to have faster button presses (M = 0:832;SD =
0:334) than the other four clusters with clusters one and four having slower button
presses (M = 2:464;SD = 0:870; M = 2:317;SD = 0:745, respectively), and clusters
2 and 3 having a medium speed (M = 1:431;SD = 0:431; M = 1:256;SD = 0:499,
respectively). Participants in cluster 1 tended to choose buttons at near 0
(M = 10
times) and far 90
(M = 6:8 times) the most when pressing the button for the button-
press probes. When choosing a button on the keypad for the self-report probe they
tended to choose far 180
(M = 5:2 times) more frequently than for the button-press
probes (M = 1:5 times). On average participants in cluster 1 chose the button at far
90
the same amount for both probe types (M = 6:8 times). Participants in cluster
2 tended to choose buttons at far 135
and far 90
the most, choosing each button
slightly more for the keypad-probes (M = 7:7 times; M = 10 times, respectively) than
the button-press probes (M = 7:3 times; M = 7:7 times, respectively). This was the
only group that did not choose the button at near 0
at all for both probes.
Cluster 3 participants tended to choose the button at near 0
the most for the button-
press probes (M = 12 times). For the self-report probes, they chose near 0
on average
less (M = 6:2 times) than for the button-press probes, and they choose the button at
near 180
more on average (M = 4:4 times) than they did for the self-report probes
(M = 0:6 times). In cluster 4 participants tended to choose a variety of buttons for
both probes choosing near 0
the most on average for the button-press probes (M = 5
times) and far 90
the most on average for the self-report probes (M = 4:2 times). The
two participants in cluster 5 tended to choose the far 180
button the most, choosing
it slightly more (M = 18 times) for the button-press probes than the self-report probes
(M = 14 times).
43
There were 15 participants with complete pre- and post-interaction CAHM ques-
tionnaire data. Using a mixed-design ANOVA, we did not nd any signicant main
eects or interaction eects of condition on the pre- and post-interaction CAHM scores.
3.3.4 Time on Task
Seven participants completed additional button presses during the optional time on task
section, as show in Table 3.4. Of those, two were in the other-CF condition, two were
in self-CF, and three were in no-CF. Relative to the clusters described in the previous
section, two of the participants were in cluster 2, three were in cluster 4, and two were
in cluster 5.
Table 3.4: Number of additional button presses, condition, and cluster for participants
who completed time on task button presses
Button Presses 4 4 4 4 8 20 24
Cluster 4 2 4 4 5 2 5
Condition other-CF self-CF self-CF no-CF no-CF no-CF other-CF
3.3.5 Perception of the Robot
Fifteen participants had complete interaction questionnaire data. Using a one-way
ANOVA, we did not nd any signicant main eects of condition on ratings of the
robot as measured by subscales from the interaction questionnaire.
3.4 Discussion
This work examined the eects of comparative feedback given by an autonomous socially
assistive robot coach on self-ecacy, performance, and perceptions of the robot. Our
rst hypothesis, H1, that comparative feedback would lead to increased self-ecacy,
44
was not supported by either the self-ecacy probes or the CAHM scores. The second
hypothesis, H2, that comparative feedback would lead to better performance, was also
not supported by the button press times, but was partially supported by the delay times,
since the self-comparative and no-comparative conditions had signicantly lower delay
times than the other-comparative condition, and the no-comparative condition had a
signicantly longer delay time than the self-comparative condition. This indicates that
receiving other-comparative feedback about a motor rehabilitation task could lead to
having increased hesitation before beginning the task. In addition, the lower delay
times with self-comparative feedback compared to no-comparative could indicate that
this type of feedback, comparing to one's own performance, could lead to a decrease in
hesitation before beginning a press. This result expands what has been found in previous
work studying the eects of other-comparative feedback given by either a person or an
on-screen agent (Lewthwaite and Wulf, 2010, Mumm and Mutlu, 2011, Wulf et al.,
2012). Finally, the third hypothesis, H3, that participants in the comparative feedback
conditions would have higher ratings of the robot, was not supported.
The press time results suggest that, for this button pressing task, the relative di-
culty of the button presses for near, far, unaected, and aected arm button presses is
as expected; far buttons took longer to press than near, and pressing the button with
the aected arm took longer than with the unaected arm. The results also suggest
that press time performance can improve for this task over repeated button presses.
The results, however, do not provide much insight on the relative diculty of buttons
based on angle because the only signicant dierence in press times among the dier-
ent button angle positions was between 0
and 135
, with 135
taking longer to press.
This pattern of diculty is consistent with Chen et al.'s nding that, for a non-stroke
aected control group, self-ecacy for reaches on the opposite side of the body (135
45
and 180
) was lower than for reaches on the same side of the body (0
and 45
) (Chen
et al., 2013).
We clustered the participant results and characterized the ve resulting groupings.
Cluster 5 seemed to include participants who were high performers; they had the lowest
average press times, and tended to choose far 180
, which might be a more dicult
button given that it is at the far position and requires an across-body reach. It is possible
that this task was too easy for the participants in this cluster. Cluster 2 participants
seemed condent and tended to challenge themselves with their probe choices. They had
medium press time performance, but tended to choose more dicult buttons at the far
and 135
and 90
positions. They also did not choose any buttons at the easier position
of near 0
. Cluster 3 participants, on the other hand, seemed less condent and did
not tend to challenge themselves. They had a similar average press time performance
as cluster 2, but tended to choose the the easier button at near 0
for the button-press
probes. Participants in cluster 1 had a high average press times, and seemed to choose
probes that were both more challenging (far 90
) and less challenging (near 0
). Cluster
4 participants seemed to explore the space of probe choice options. They had relatively
high average press times, and tended to choose a variety of buttons in response to
the probes. We also found an interesting pattern for participants in clusters 1 and 3;
they choose more dicult buttons for the self-report probes and easier buttons for the
button-press probes. This discrepancy could suggest that for some people the probes
measure slightly dierent aspects of their self-ecacy for this task.
Overall, the cluster groupings seem to indicate that participants varied in terms
of their performance, exploration of dierent buttons, condence, and willingness to
challenge themselves with pressing more dicult buttons. We also saw that none of the
participants who completed additional time on task button presses were in clusters 1 or
3, the clusters that tended to press the easy button of near 0
the most frequently. The
46
lack of any discernible pattern between these cluster groupings and study condition could
indicate that personalizing robot feedback to variables such as performance relative to
a task, self-ecacy, and motivation could be important.
One limitation of this study is that baseline measurement of self-ecacy for each
button per participant were not available. The lack of a baseline makes the probe
choices dicult to interpret because each person's perception of their ability to press
each button is not known. Therefore, we can only interpret the probe choices with
respect to the objective performance metric of press times, which might not correlate
with an individual's subjective perception of their ability. An extension of this work
could involve using the operationalization of arm reaching self-ecacy from (Chen et al.,
2013) to get baseline self-ecacy data. These data could then be used as a prior in
personalized computational models of SAR decision making for increasing self-ecacy.
Another factor that could be important to include in personalized SAR coaching models
is concordance between the dominant arm and the stroke aected arm, which might
aect self-ecacy (Chen et al., 2013).
An additional limitation of this work is the way delay time was measured. Because
the measurement time started when both home buttons were pressed after the robot
began the instruction, the delay could have started at dierent times during the robot
instruction. A better measurement procedure would standardize the start time of the
measurement relative to the robot's instruction.
Another limitation of the autonomous SAR system used in this study is the error
correction behavior of the robot. The robot corrected participant pressing errors (e.g.,
holding the home buttons down incorrectly or pressing the wrong button) whenever
they occurred; since participants varied in the amount of errors made, those who made
many errors became annoyed with the robot. Future work could examine the role of
frequency of error correction behaviors more closely.
47
Finally, because there was only one session with the robot, we cannot derive insight
into how comparative feedback given by a SAR coach might aect a person's self-ecacy
over repeated interactions or how it might aect real world arm nonuse. Future work
in this area could include multiple sessions, focusing on self-comparative feedback and
personalizing the interaction based a person's response to self-ecacy probes. Future
work could also include an accelerometer-based measures of arm nonuse in daily life such
as those used in (Bailey et al., 2015, Rand and Eng, 2015, Shim et al., 2014) would also
expand the usefulness of the approach. Finally, while this work used a robot coach to
motivate people to perform an exercise task, it did not provide feedback on the specic
reaching motions; future work could focus on providing feedback about motion quality
as well as reach time.
3.5 Summary
We presented a fully autonomous SAR system for coaching a person post-stroke in an
upper extremity reaching task. We showed that people have improved performance on
the task over time when guided by the SAR coach, a promising result for future devel-
opment of SARs. We also found that people had a higher delay time before beginning
a reaching task when they received other-comparative feedback from a SAR coach than
when they received self-comparative or no comparative feedback, and they had a lower
delay time with self-comparative feedback than no-comparative, suggesting that fu-
ture systems should focus on self-comparative rather than other-comparative feedback.
We also grouped participants using an unsupervised learning clustering algorithm and
characterized the press time performance and response to self-ecacy probes of those
groups. We suggest that the characteristics of those groupings should be taken into
48
account by researchers designing personalized SAR feedback algorithms to improve self-
ecacy over time. In the next chapter, the design and evaluation of the SAR backstory
and messaging support method are described.
49
Chapter 4
Design and Evaluation of Socially
Assistive Robot Backstory and
Aective Messaging Support
In this chapter, the design of the SAR support methods of backstory and mes-
saging are detailed and evaluated. A study comparing a ctional to a realistic
robot backstory for a robot exercise buddy for adolescents is presented. This
study additionally presents the development and validation of a set of aective
messages for exercise encouragement. The backstories and aective messages
were developed with feedback from focus groups described in this chapter. The
methodology and results of this study are described, and a discussion is pre-
sented.
In order to develop a set of aective messages and to compare a ctional to a
realistic robot backstory a set of focus groups and a study were conducted.
1
The
study consisted of a four-session SAR-based physical activity (PA) intervention with
1
The focus groups and study described in this chapter were completed in collaboration with other
researchers at the University of Southern California including Cheng K. Fred Wen, Gillian O'Reilly,
Donna Spruijt-Metz, and Maja J Matari c.
50
11-14 year old overweight adolescents, and it compared a ctional to a realistic robot
backstory (Swift-Spong et al., 2016). Focus groups conducted before the study were
also used for developing the initial set of aective messages. These messages were
then used in the four-session evaluation study. The role of the robot in this context
is dened as a buddy that performs a task co-actively with the user when possible and
has a similar ability level at the task as the user. Results from the study suggest
that participants responded positively to the robot exercise buddy. Some responses to
qualitative questions also suggest that participants found the aective messages from
the robot to be supportive. In addition, a trend was found towards increased intrinsic
motivation for PA post-intervention.
4.1 Focus Groups
A series of two-session focus groups were conducted with a total of 12 participants aged
11-14 years old in three separate groups. The focus group sessions were performed to
inform the design of the interaction and centered around aspects such as the idea of
using the Nao robot as an exercise buddy, possible backstories for the robot, exercises,
aective messages, and using music during the exercise sessions. The Nao robot was
selected for this study because it is the best available robot platform based on its size,
safety, cost, and ability to do some exercises. Ideas and feedback collected from these
focus groups were used to develop and nalize the backstory of the robot, the exercises,
and the motivational messages.
Past work has shown that aective messages, which focus on the emotional outcome
of a behavior, such as how it aects happiness enjoyment, increase physical activity
levels in inactive older adolescents more than instrumental messages, which focus on
the benecial or harmful outcomes such as the eects of exercise on health (Sirriyeh
et al., 2010). Consequently, to design the aective messages for the SAR system, focus
51
group participants provided feedback on a list of aective messages in four categories
{ congratulatory, encouraging, feedback, and testimonial. Examples of those aective
messages include: \Let's keep up the good work. Exercising will make us feel energetic
and happy." and \I am so happy that we nished that exercise." This feedback was
then used to select a set of aective messages for the four-session intervention study.
4.2 Methodology
4.2.1 Study Design
The study compared two robot backstories using a four-session within-participants study
design with 11-14 year-old overweight adolescents. A within-participants design was
used because the study did not include enough participants for a between-participants
comparison. The backstories were designed to portray the robot as a peer-level exercise
buddy. Each participant took part in four exercise sessions, two with each robot back-
story. The order of backstory presentation was randomized using block randomization.
In both backstory conditions, the robot presented information about its characteristics
at analogous interaction points at the beginning, during, and at the end of each ex-
ercise session. Four sessions were used so that each backstory would be seen in two
exercise sessions in order to minimize novelty. Questionnaire data related to enjoyment
of physical activity, perceived stress, and intrinsic motivation to exercise were collected
before the rst exercise session and after the fourth, and questionnaire data related to
perceptions of the robot were collected after the last session with each robot backstory
(the second and fourth exercise session).
52
4.2.2 Experiment Setup
The experiment setup is seen in Figure 4.1. The humanoid robot, a Nao Next Gen V4,
was placed on a table so as to be close to or slightly below eye level of the participant.
A monitor next to the robot displayed the exercise sequence for the session, including
the name and number of repetitions or duration for each exercise. It was also used
to show demonstrations of each exercise. A keypad was placed between the robot
and the monitor. Using this keypad participants could respond to questions from the
robot using keys labeled \Yes" and \No", inform the robot of being done with a set
of exercises using a key labeled \Done", and stop or pause the interaction with a key
labeled \End". The robot autonomously responded to these inputs based on a pre-
determined script. A 3D vision sensor (Microsoft Kinect) hanging from the ceiling
and an Augmented Reality (AR) tag on top of the robot's head were used to capture
the position of the robot. The robot used this position information to autonomously
move back to its starting position after exercises that involved foot movement. An
exercise step and treadmill were also in the exercise area. The robot wore either green
or orange wristbands, and the color of these was switched between the two backstories
to indicate that the backstory was dierent. The pre-determined scripts for what the
robot with each backstory said were the same except for some of the exercise names
and characteristics of the backstory, as described in Table 4.1. In addition, the robot
performed scripted non-verbal behaviors such as head nods and shakes, pointing, idle
behaviors while talking, and other expressive gestures.
A 3D vision sensor (Microsoft Kinect) was placed behind the robot and monitor and
used to record the participant's body pose information. Two USB cameras and one high
denition camera were used to record audio and video of the interaction from dierent
angles. The experimenters were in the same room as the participant, but behind an
53
Figure 4.1: Robot exercise buddy experiment setup. Left: Robot performing a split
squat with an arm raise to the side. Right: Robot, step, and treadmill setup. Partici-
pants exercised in the space between the table and treadmill.
opaque black curtain, in case the participants had any questions or problems, and the
participants were aware of the experimenters' presence behind the curtain.
4.2.3 Robot Backstories
Based on the ideas and insights from the focus groups, two robot backstories were
developed, one ctional and one realistic. Table 4.1 illustrates the dierences in how
the robot used each backstory to describe itself.
4.2.4 Exercise Sessions
The exercise sessions were designed around the idea of circuit training, which has been
shown to be eective (Davis et al., 2011), and included strength training exercises,
treadmill, and step up exercises. The exercise sessions were designed to be around 30
minutes long. The exact duration of each session was determined by the participant's
pace in performing the exercises. Strength training exercises were chosen that the robot
and participant could do together. Due to motion limitations, the robot could not
perform the following exercises: walking, running on the treadmill, and step ups. The
strength training exercises included upper body arm movement exercises such as bicep
54
Table 4.1: Description of both robot backstories.
Character-
istic
Fictional Backstory Realistic Backstory
Name Mio Nao
Identity I am from another planet
called Kepler 186f. I came
to Earth because my home
planet is unhealthy and I want
to prevent Earth from becom-
ing unhealthy too.
I am a robot. There are many
robots like me around the
world, playing soccer, helping
kids, and getting involved in
research. I came to Los Ange-
les to be part of research here
at USC.
Reason for
exercising
I need to exercise more on
Earth because Kepler 186f has
low gravity and I need to get
stronger in order to live here.
I have been programmed with
many exercises, and I am here
to share some of my knowl-
edge with you.
Exercise
ability
After all, I am from another
planet. Kepler 186f has lower
gravity than Earth, so my bal-
ance isn't so great here.
My motors can only move in
certain ways and I am not pro-
grammed to do complicated
balancing.
Source of
demo
videos
A friend of mine made videos
of these exercises so you will
know how to do them.
Researchers at USC gave me
some demo videos of these ex-
ercises so you will know how
to do them.
Speech
under-
standing
and vision
I don't see or hear so well be-
cause my home planet is super
dark and loud.
I've also been programmed to
speak English but can not un-
derstand it, and I don't have
any algorithms to help me see
what you are doing.
Source of
knowledge
Here is the second set of exer-
cises that I learned during my
space travel training on Ke-
pler 186f.
Here is the second set of exer-
cises that I was programmed
to do by researchers at USC.
curls and raising the arms to the front, side, and over the head as well as lower body
exercises such as squats, split squats, and side steps. Combinations of upper and lower
body exercises, and sequences of two exercises were also included in the circuit training
design.
55
Each session contained three step up exercises (right leg, left leg, and both legs in a
quick jumping manner), one ve minute walk or run on the treadmill, two sequences of
exercises, four exercises that involved the lower body, and two upper body only exercises.
Each of the four sessions therefore had a slightly dierent set of exercises but the total
number of exercises was the same across sessions, as was the ratio of upper body, lower
body, sequence, treadmill, and step up exercises. The exercise sessions were presented
in the same order for all participants.
To explain how each exercise (except the treadmill) should be done, a pre-recorded
demonstration video of a human doing and explaining the exercise was played on the
monitor before the exercise began. For the treadmill exercise, an explanation, demon-
stration, and participant practice of using the treadmill was done by the experimenter
before the rst exercise session. In addition, participants were allowed to choose the
speed of the treadmill and were told that they could either walk or run on the treadmill.
Participants were also told that they could do the strength training exercises either at
the same pace as the robot or at their own pace. Participants were asked to complete
a specied number of repetitions of each of the strength training exercises (10 or 15
depending on the exercise), and they were asked to do each of the treadmill and step
up exercises for a certain amount of time displayed using a countdown timer on the
monitor.
Participants were instructed to push the \Done" button after completing the repeti-
tions for each strength training exercise, to let the robot know they were nished. The
press of the \Done" button was the only input the robot had to know if the participants
completed the exercises. As such it would have been possible for participants to press
the \Done" button and continue with the session without actually completing the exer-
cise. After the \Done" button was pushed, the robot gave feedback by saying one of the
56
aective messages described in Section 4.1. The aective messages were pre-scripted
and were the same for both robot backstories.
4.2.5 Measures
A collection of measures was used, including questionnaires and behavioral measures.
Questionnaires administered both pre-intervention (immediately before the rst exercise
session) and post-intervention (immediately after the fourth exercise session) included a
modied version of the Physical Activity Enjoyment Scale (PACES) (Motl et al., 2001),
which measures enjoyment of physical activity, a modied version of the Perceived Stress
Visual Analog Scale (Cohen et al., 1983), which measures perceived stress levels, and
the Exercise Self-Regulation Questionnaire (SRQ-E) (Ryan and Connell, 1989), which
measures intrinsic motivation for physical activity. In addition, the Big Five Inventory,
which measures personality traits along ve dimensions was given immediately before
the rst interaction session (Benet-Mart nez and John, 1998, John et al., 1991, 2008)
along with a set of demographics questions. Height and weight (using a Tanita Body
Composition Analyzer) were also taken immediately before the rst exercise session and
after the last exercise session, and BMI was calculated.
Four questionnaires were used to assess the participants' reactions to the two robot
backstories. These were given immediately after the last session for each backstory (the
second and fourth exercise sessions). The rst three of these questionnaires included
the social richness subscale of the Temple Presence Inventory (TPI) (Lombard et al.,
2000a), interaction-related questions modied from the TPI (Lombard et al., 2000a),
and a modied version of the social attraction subscale of the Interpersonal Attraction
Scale (McCroskey and McCain, 1974a). Modied versions of these three questionnaires
were previously successfully used by Jung and Lee in a study with social robots (Jung
and Lee, 2004a). The fourth questionnaire was a Robot Expectations Questionnaire
57
(Lohse, 2011). In addition, several open-ended questions about the interaction with
the robot and the robot's backstory were asked after the second and fourth sessions.
Finally, after the fourth exercise session, participants were also asked an open-ended
question about which robot (i.e., backstory) was a more enjoyable exercise buddy and
why.
Additional data collected during each exercise session included video, audio, skeletal
tracking (using a Kinect), and heart rate (using a Mio Alpha 1 wrist-worn heart rate
monitor). Accelerometer data were collected during waking hours (using an Actigraph,
Inc. GT3X-BT accelerometer) before the rst exercise session and after the fourth
exercise session. Analysis of the video, audio, skeletal tracking, heart rate, Perceived
Stress Visual Analog Scale, Big Five Inventory, and Robot Expectations Questionnaire
data is outside the scope of this dissertation proposal.
4.2.6 Procedure
At the beginning of each exercise session, participants were told that the robot was
currently sleeping. They were also told the name of the robot and that they would
exercise with a dierent robot the following week. Participants were instructed to press
the \Yes" button to wake up the robot. When they did so, the autonomous robot in-
teraction began. After participants completed each exercise session and questionnaires,
they received compensation of $20 in cash (an amount typical for the type of study and
geographic area).
4.2.7 Hypotheses
H1: Participants will have more positive reactions to the robot with a ctional backstory
than the robot with a realistic backstory.
H2: Intrinsic motivation for physical activity as measured by the SRQ-E will increase
58
from before the rst exercise session to after the fourth exercise session.
H3: Physical activity enjoyment as measured by the PACES will increase from before
the rst exercise session to after the fourth exercise session.
H4: Activity levels as measured using the ActiGraph activity monitor will increase from
before the rst exercise session to after the fourth exercise session.
4.2.8 Study Population
There were a total of 22 participants, of whom 18 completed the study (two of the
18 also participated in the focus groups). Participants were recruited from events and
organizations in the local area, such as churches and health fairs, and from a participant
pool from other health-related studies at USC. Of the four who did not complete the
study, two completed the rst session only, and the other two completed the rst two
sessions. Only data from the 18 participants who completed all four sessions were used
in the analysis. This set of 18 participants was composed of 12 females and 6 males,
with an average age of 12.11 (SD = 1:28;Range : 11 15) and an average Body
Mass Index of 27.75 (SD = 5:54;Range : 23:00 46:22) at baseline. Participants were
11-14 years old at the time of recruitment. Fourteen of the 18 participants were self-
identied as Hispanic, 2 as mixed ethnicity (1 as Mexican and African American and 1 as
African American and French), 1 as White, and 1 as other. The average number of days
between the rst exercise session and the fourth was 13.17 (SD = 4:18;Range : 1022).
Participants were told that if there were any questions they did not want to answer on
the questionnaires, those could be skipped. Thus, some of the questionnaire scales had
missing data.
59
4.3 Results
4.3.1 Backstory
Using a paired t-test no statistically signicant dierences were found between the
participants' ratings of the ctional and realistic robot backstories when comparing
ratings of the social richness (t(15) = 0:11;p = 0:91;M
F
= 4:85;SD
F
= 0:72;M
R
=
4:86;SD
R
= 0:48), interaction (t(0:07) = 17;p = 0:95;M
F
= 4:82;SD
F
= 0:94;M
R
=
4:85;SD
R
= 1:13) and social attraction (t(17) = 0:17;p = 0:87;M
F
= 5:40;SD
F
=
1:25;M
R
= 5:33;SD
R
= 1:17) subscales as seen in Figure 4.2. These ratings were on a
seven-point scale, with two participants who did not respond to all the questions used
in the social richness subscale rating the ctional backstory.
1
2
3
4
5
6
7
Social Richness Interaction Social Attraction
Robot Measure
Rating
Robot Backstory
Fictional
Realistic
Perceptions of Robot Backstories
Figure 4.2: Evaluations of the robot with dierent backstories.
In response to open-ended question about which robot backstory was more enjoyable,
17 participants responded; 7 reported that they either liked both equally or that they
were the same, 9 participants chose a single backstory, and one mentioned liking both
and also mentioned one of the backstories. Several themes emerged in response to the
question of what participants liked about exercising with the robot with each backstory.
Some participants enjoyed feeling comfortable and not pressured exercising with the
robot. For example, one said: \I feel more comfortable, I can correct my mistake
60
without being shy or embarrassed." Several participants also mentioned liking that the
robot did the exercises along with them and encouraged them to continue doing the
exercises. When asked what they disliked, participants responded that the robot could
not do all the exercises with them, that it was slow, and that it asked the same questions
multiple times.
4.3.2 Physical Activity Intrinsic Motivation
The subscales of the SRQ-E were calculated, as seen in Figure 4.3, measured on a
seven-point scale and listed from most internally and autonomously motivated, on
the left, to least, on the right. Several participants did not answer all of the ques-
tions for each subscale (one for pre-intervention Intrinsic Motivation, two for post-
intervention Intrinsic Motivation, three for pre-intervention Identied Regulation, two
for pre-intervention Introjected Regulation, one for post-intervention Introjected Reg-
ulation, and two for pre-intervention External Regulation). Using paired t-tests, no
signicant dierences between the participants' ratings pre- and post-intervention for
any of the subscales were found: Intrinsic Motivation (t(14) =1:64;p = 0:12;M
PRE
=
5:32;SD
PRE
= 1:69;M
POST
= 5:59;SD
POST
= 1:11), Identied Regulation (t(14) =
0:80;p = 0:44;M
PRE
= 5:67;SD
PRE
= 1:60;M
POST
= 5:67;SD
POST
= 1:14), Intro-
jected Regulation (t(14) = 0:93;p = 0:37;M
PRE
= 3:25;SD
PRE
= 1:45;M
POST
=
2:87;SD
POST
= 1:49), and External Regulation (t(15) = 1:03;p = 0:32;M
PRE
=
2:14;SD
PRE
= 0:95;M
POST
= 1:99;SD
POST
= 1:14). The Relative Autonomy In-
dex (RAI) was calcualted as a weighted sum of the four subscales that gives the weights
of 2; 1; 1; and 2 to the Intrinsic Motivation, Identied Regulation, Introjected Regula-
tion, and External Regulation subscales respectively for a maximum rating of +24 and a
minimum of24 (Grolnick and Ryan, 1989). Using a paired t-test, an increase in RAI
from pre- to post-intervention was found that was approaching signicance (t(12) =
61
1
2
3
4
5
6
7
Intrinsic Motivation
Identified Regulation
Introjected Regulation
External Regulation
SRQ−E Subscale
Rating
Time
Pre
Post
Intrinsic Motivation for Physical Activity
Figure 4.3: Ratings of intrinsic motivation for physical activity.
1:85;p = 0:09;M
PRE
= 8:95;SD
PRE
= 5:00;M
POST
= 9:64;SD
POST
= 4:92) with
three participants who did not respond to all the questions used to calculate the pre-
intervention measure, one for the post-intervention measure, and one for both pre- and
post-intervention measures.
4.3.3 Physical Activity Enjoyment
Using a paired t-test, no statistically signicant dierence was found between partici-
pants' ratings of physical activity enjoyment before the rst exercise session and after
the fourth, using the PACES measure (t(14) = 0:97;p = 0:35;M
PRE
= 4:01;SD
PRE
=
0:48;M
POST
= 4:04;SD
POST
= 0:35). Participants had high measures of physical ac-
tivity enjoyment both before and after the intervention on a ve-point scale, with one
participant who did not respond to all the questions in the pre-intervention measure
and two for the post-intervention measure.
62
4.3.4 Activity Levels
Participants were asked to wear a waist-worn accelerometer during waking hours, except
during activities that involve water, for seven days both before the rst exercise session
and after the last exercise session. The accelerometers were set to collect data using a
30 second epoch.
ActiLife 6.11 was used for downloading, validating, and processing the accelerometer
data. For data processing non-wear time was dened as a consecutive period of 60
minutes with zero activity count. Within this period an allowance of two minutes of
spikes between 0 and 100 was allowed, and a valid wear day was dened as 600 minutes or
more of wear time. Cut points dened in Evenson et al. (2008) for calculating sedentary
behavior and moderate-to-vigorous physical activity (MVPA) were used.
Eleven participants had two or more valid wear days during the pre-intervention
measure before the rst exercise session, and ve participants had two or more valid wear
days during the post-intervention measure after the last exercise session. Using paired
t-tests no signicant dierences were found between pre- and post-intervention measures
of sedentary behavior (t(2) = 1:07;p = 0:40) and MVPA (t(2) = 2:20;p = 0:16).
4.4 Discussion
No dierence between the ratings of the two robot backstories were found for the social
richness, interaction, or social attraction subscales, therefore failing to support the rst
hypothesis, H1. In addition, 7 of the 17 participants who responded to the open-ended
question about which robot backstory they enjoyed more liked them both equally or
saw them as the same. These results could indicate that the dierences between the
backstories were not salient enough to change the nature of the interaction. This lack
of salience could in part be due to the fact that several aspects of the robot did not
63
change between the backstories. Specically, the non-verbal behaviors, the sound of
the voice, and the appearance only changed slightly when the color of the wristbands
was changed between backstories. This could indicate that backstory dierences need
to be accompanied by salient changes in robot behavior and appearance in order to be
perceived as meaningful.
Participants had positive (above neutral) reactions to the robot with the two back-
stories on the social richness, interaction, and social attraction subscales. They also
mentioned feeling comfortable exercising with the robot, liking that the robot did ex-
ercises along with them, and liking the encouragement from the robot. These positive
reactions to the robot exercise buddy could indicate this type of exercise buddy with a
backstory could be useful in behavior change interventions. In addition, the responses to
qualitative questions that suggest that some participants liked the encouragement from
the robot could indicate that the set of aective messages used to provide encouragement
is useful.
The second hypothesis, H2, was not fully supported by the SQR-E, but there was
a trend approaching signicance that RAI increased from pre- to post-intervention.
Because this robot exercise buddy intervention, however, was not compared to a control
with no robot there is no way to know if this change is due to the interaction with the
robot.
No support for H3 was found as there were no signicant dierences between physical
activity enjoyment ratings before and after the intervention. This could in part be due
to a ceiling eect because the physical activity enjoyment ratings before the intervention
were already high.
The fourth hypothesis, H4, was not supported by pre-post dierences in activity
levels. This lack of change could be due to the short duration of the intervention. In
64
addition, the intervention did not include strategies or encouragement for participants
to continue exercising outside of the experiment sessions.
4.5 Summary
In this chapter, a SAR system capable of serving as an peer-level exercise buddy was
evaluated with overweight adolescents comparing ctional and realistic robot backsto-
ries. Although there were no dierences in participants' ratings of the robot characters,
participants did react positively to the robot exercise buddy. This suggests that robots
in multi-session or long-term behavior change interventions with a realistic or ctional
backstory might be accepted by users. In addition, a set of aective messages was de-
veloped using feedback from three focus groups, and these messages were used in the
four-session study. Some participants indicated that they liked receiving encouragement
from the robot. This could indicate that the aective messages were encouraging and
accepted. In the next chapter, a model of SAR habit formation support is presented
which formalises the SAR behavior change support method of reminder and social re-
ward decision making.
65
Chapter 5
Model of Socially Assistive Robot
Habit Formation Support
In this chapter, a model of habit formation support is presented. This model
formalises the SAR behavior change support method of reminder and social
reward decision making. A framework for how a robot could support the process
of habit formation through reminders and social rewards, a formalization of the
types of habits the support model is designed around, and algorithms for socially
assistive robot support decision making are detailed in this chapter. The design
decisions behind these model components are discussed.
This work contributes a model of habit formation support given by a robot that
aims to balance providing support to increase repetition of the habit with minimizing
the amount of intervention from the robot. The robot's goal is to make the habit an
automatic process that does not rely on willpower alone for adherence. We dene a habit
as consisting of a cue that occurs in everyday life and a desired behavior. We additionally
include the constraint that the cue must occur before the desired behavior. The habit is
considered completed when the desired behavior is done after the cue occurs. We dene
automaticity as a property of a habit that represents the strength of the unidirectional
66
connection from the cue and the behavior, or the propensity to perform the behavior
after the cue occurs. We also call this the strength of the cue-behavior pair. We dene a
habit as having formed when the automaticity increases to the point that the behavior
occurs at the desired frequency and time period after the cue of the user forming the
habit.
Time
Cue Behavior Reward Cue Behavior Reward
Reminders
when cue occurs
Time
Cue Behavior Reward Cue Behavior Reward
Social
rewards
Time
Cue Behavior Reward Cue Behavior Reward
Social
rewards
Reminders
when cue occurs
Remind
Reinforce
Sustain
Figure 5.1: SAR habit formation support framework. Remind: The robot provides
reminders to encourage the user to do the behavior when the cue occurs. Reinforce:
The robot provides unexpected social rewards after the user completes the behavior.
Sustain: As the habit forms the robot decreases the reminders so that the habit remains
when the robot leaves.
67
We model SAR habit formation support as a sequential decision making task in
which the robot decides when to remind the user to complete the desired behavior and
when to reinforce the completed habit with social rewards. The goal of this sequential
decision making task is to increase the automaticity of the habit so that when the robot
leaves the habit is sustained. Habit formation is the process of a cue being followed by
the desired behavior, and that cue-behavior pair being reinforced repeatedly over time.
By making decisions about when to remind and reinforce, the robot becomes a part of
the habit formation paradigm, shown in Figure 5.1.
Using this framework we developed our model of habit formation support to speci-
cally focus on supporting habits with elapsed-time based cues. For these types of habits,
the cue is considered to have occurred when a specied period of time goes by without
the desired behavior occurring. Examples of habits with elapsed-time based cues are
taking a break from working or watching TV every thirty minutes to go for a walk,
taking a break from looking at a computer screen every twenty minutes to look away
from the screen at an object in the distance, and drinking water every hour. These
habits address reducing sedentary behavior, reducing eye strain, and reducing dehydra-
tion. These dier from habits with event-based cues such as
ossing after the event of
tooth brushing occurs and habits with time-based cues such as exercising at seven in
the morning.
5.1 Habit Formation Formalization
We formalize the concept of habit formation for habits with elapsed-time based cues
using a Finite State Machine (FSM) as shown in Figure 5.2. In this FSM, we dene
three states, the state of the behavior being absent for less than or equal to a specied
time (S1), the state of the behavior being absent for greater than a specied time (S2),
and the state of the behavior being present (S3). S2 represents the occurrence of the
68
cue, and S3 represents the occurrence of the desired behavior. Transitions between the
states occur when the desired behavior begins, when it ends, and when the behavior
has been absent for a specied period of time.
Behavior
absent x
minutes
Behavior
present
Behavior ends
Behavior begins
Behavior
absent > x
minutes
Behavior begins
Behavior absent
x minutes
S1 S2 S3
≤
Figure 5.2: Finite State Machine representation of habit formation for habits with
elapsed-time based cues.
The goal of supporting the formation of a habit represented by this FSM is to
strengthen the unidirectional connection between S2 and S3 (represented by a bold
dashed line). This connection represents the automaticity of the habit, or the probability
of the behavior occurring after the cue. We dene an episode of the the habit to have
been completed when a user returns to state S1 after starting in state S1, moving state
S2, and then moving to S3.
5.2 Modeling Habit Formation Support
The habit formation support model is designed around the goal of making the habit
automatic so that when the robot leaves the user continues to do the habit without
the continued support of the robot. It is designed to support building a sustainable
habit, not to achieve perfect adherence in the short term. With this goal in mind,
the schedule of reminders is designed to increase probability of the user completing the
69
desired behavior sooner than they would otherwise after the cue occurs. Social rewards
are then given in order to reinforce the connection between the cue and the behavior.
Algorithm 5.1 Habit Formation Support Algorithm
n = 1, current episode number
b, number of episodes in reward block
a, number of episodes from one automaticity probe to the next
Generate reward schedule (Algorithm 5.2)
while Operating do
if Cue occurred and no behavior and (n 1) mod (a)6= 0 then
while Behavior has not occurred do
Give reminders on schedule
end while
else if Cue occurred and behavior occurred then
Give reward on schedule
if n mod (b) = 0 then
Generate new reward schedule (Algorithm 5.2)
end if
if (n 1) mod (a) = 0 then
Update automaticity using Equation 5.1
end if
Generate new reminder schedule (Algorithm 5.3)
Increment n
else
Monitor for cue and behavior
end if
end while
The robot makes decisions about when to provide reminders and social rewards using
Algorithm 5.1. While the robot is operating in a support role, it completes reminder and
reward actions. During operation, the robot continually monitors for the occurrence of
the cue. Once the cue has occurred the robot sometimes give reminders to the user to
complete the desired behavior based on a predetermined reminder schedule. Then once
the desired behavior has occurred the robot sometimes does a rewarding social behavior
based on a predetermined reward schedule. Next, the robot updates its reward schedule
(Algorithm 5.2) and estimate of the user's automaticity for the habit (Equation 5.1)
70
if it is the correct episode for updating. Finally, the robot generates a new reminder
schedule (Algorithm 5.3).
There are two main assumptions we make in the design of this support model. The
rst is the assumption that the robot's behavior is most reinforcing when the reward
is unexpected and no reminders are given. The second assumption is that the time
between the cue and behavior in the absence of reminders contains meaning about the
current automaticity. Because of this assumption we set aside one episode every a
episodes as automaticity probes where no reminders are given in order to update the
estimated automaticity. The impact of these two assumptions on determining when
to give reminders and how to estimate automaticity is described in Sections 5.2.3 and
5.2.2.
5.2.1 Scheduling Rewards
Algorithm 5.2 details how rewards are scheduled in the habit formation model. We use
a ratio schedule in which rewards are given based on the number of times the behavior
has been completed (P erez et al., 2016, Strohacker et al., 2014). Because unexpected
rewards are more reinforcing than expected rewards (Watabe-Uchida et al., 2017), we
randomize the timing of when rewards are given using a block randomization method.
This is done by randomly selecting from a specied number of future episodes, b, which
r episodes to schedule a reward behavior. We use block randomization instead of a
variable ratio schedule (varies the number of episodes after which a reward is given)
or a random ratio schedule (randomly gives a reward with some probability after each
episode) in order to keep the reward schedule somewhat unpredictable, but also maintain
regular rewarding interactions with the robot (P erez et al., 2016).
71
Algorithm 5.2 Reward Scheduling Algorithm
b, number of episodes in reward block
r, number of rewarded episodes in reward block
function schedule rewards
return Randomly select r episodes from the next b episodes
end function
5.2.2 Updating Automaticity Estimate
The estimated automaticity, A, for the habit is dened as follows,
A(n;C) = e
p
p
X
i=0
C
nai
)
(5.1)
where n is the current episode,C is a vector of the elapsed times between the cue and
the behavior for all past episodes, a is the number of episodes from one automaticity
probe to the next,p is a past window of automaticity probe episodes, and is a constant.
At the startup phase of robot support before a(p 1) episodes have occurred, p in this
equation is set to (n 1)=a + 1 because at startup the past window of episodes has not
yet occurred. Using this equation, the automaticity range is between zero and one.
We use an average of the time dierence between the cue and behavior of the past p
automaticity probes in order to avoid a single episode having an overly large eect on the
automaticity estimate. This average is then used as the input to an exponential function.
We use this exponential function because we assume that the time dierence between the
cue and the behavior has a relationship with automaticity where shorter time dierences
correspond to higher automaticity, and longer time dierences correspond to a much
lower automaticity. The exponential function models this assumed relationship.
72
5.2.3 Scheduling Reminders
Because we assume that the rewards are most reinforcing when reminders are not given,
the reminder scheduling algorithm is designed to balance the trade-o between giving
reminders too frequently and reducing the potential increase in automaticity and giving
them too infrequently and reducing the likelihood that the user does the behavior soon
after the cue. In order to balance this trade-o, reminders are given more frequently
and sooner after the cue when the automaticity is low than when the automaticity is
high. We prioritize reminders when automaticity is low in order to increase the user's
awareness of the elapsed-time-cue. Then as automaticity increases, we assume that this
awareness has also increased and give fewer reminders.
Reminders are scheduled using Algorithm 5.3. This schedule is determined using
the user's current estimated automaticity for the habit. The algorithm also includes
a set time period over which reminders can be scheduled after the cue occurs, and a
maximum number of reminders that can be given over that time period. These limits
allow for avoiding giving reminders too frequently or for too long of a period of time. In
this algorithm is rst calculated based on the current automaticity and the maximum
number of reminders. then corresponds to the number of reminders that are given for
that schedule.
The algorithm iterates through each reminder and returns a scheduled time post-cue
for each. The equation we use in this algorithm for scheduling reminders is based on the
reminder scheduling equation used by Hou et al. which was used to give users reminders
with increasing frequency before the time at which a task needed to be competed (Hou
et al., 2016a,b). We use a modied version, however, that reverses the order such that
reminders are given with decreasing frequency after the cue occurs. Reminders are only
given after the cue occurs because the goal is to make doing the behavior after the cue
more automatic. Giving reminders before the cue increases the automaticity between
73
Algorithm 5.3 Reminder Scheduling Algorithm
T , time period over which reminders are given
m, maximum number of reminders
, current automaticity
, constant
function schedule reminders
=bma +me
for i = 1 to m do
R
i
=T
1 2
(i)
+ 2
+T
Increment i
end for
return R
end function
the reminder and the behavior because of a phenomenon called \forward blocking" in
which a cue between an earlier cue and an outcome does not become as associated with
that outcome as the earlier cue (Mitchell et al., 2006).
The equation in this algorithm for scheduling reminders allows for the time span
between reminders to increase over time. We schedule reminders this way because we
make the assumption that it is perceived as less annoying than giving reminders at a
set or increasing frequency. This design is analogous to algorithms used in computer
networks when a message from one node to another node returns with an error. In these
systems, it is common to use an exponential backo algorithm, which exponentially
increases the time between messages, to determine when to retry sending the message
(Kwak et al., 2005). Our reminder schedules have a similar property of progressively
increasing the time span between reminders.
5.3 Summary
In this chapter, we presented a model of SAR habit formation support that includes a
framework for the actions a robot can take to support habit formation, a formalization
of process of habit formation for habits with elapsed-time based cues, and the algorithms
74
and equation used for SAR decision making. We additionally described the goal of habit
formation support, which is to make the connection between the cue and the behavior
more automatic so that the habit is sustained after the robot is no longer present,
and we described how this goal shaped the design decisions we made in developing the
model. In the next chapter, we evaluate this model in the domain of reducing older
adult sedentary behavior.
75
Chapter 6
Evaluation of Model of Socially
Assistive Robot Habit Formation
Support
In this chapter, a study designed to evaluate the habit formation support model
with the specic habit of getting up after sitting for 30 minutes is presented.
The design of a fully autonomous SAR system developed for this study is de-
scribed. The three phases of the study, including a two-week in-home robot
intervention with two older adult participants, as well as the outcome measures
are detailed. The results of the study are described, and a discussion of those
results is presented.
6.1 Methodology
In order to evaluate the habit formation model detailed in Chapter 5 we developed
an autonomous SAR system designed to support older adults in forming a habit of
76
getting up after sitting for 30 minutes (min). We conducted a single-case design study
to evaluate the model and the SAR system.
6.1.1 Habit
In this study, we evaluated our model of habit formation support with the habit of
getting up after 30 min of sitting. For this habit, the elapsed-time based cue is sitting for
30 min, and the desired behavior is getting up from sitting. We chose this habit because
it is measurable with commonly used commercially available sensors, and it could lead
to reduced sedentary behavior, which can be benecial for older adults (Dogra and
Stathokostas, 2012, Dunlop et al., 2015, Santos et al., 2012). We chose 30 min because
it has been recommended as a goal for maximum sitting time (Owen et al., 2011), and
has been used as a goal in multiple behavior change intervention studies (Atlas and
Deyo, 2001, Guitar et al., 2018).
6.1.2 Robot and System Design
The fully autonomous SAR system we created for this study was designed to allow
the robot to immediately detect when the user began and ended a bout of sitting and
interact with the user based on that sensed sitting information. This setup included
two main components, the robot setup on a small table (shown in Figure 6.1b) and the
chair sensor (shown in Figure 6.1a).
The robot, referred to as \Kiwi," was designed to operate continuously in the par-
ticipant's home so that it could initiate interactions without needing to switched on.
We designed a two degree-of-freedom internal structure for the robot, shown in Figure
6.1c, that consisted of a bottom (pan) motor that rotated the body of the robot from
left to right and a top (tilt) motor that rotated the head of the robot forward and
backwards (Deng et al., 2019, in preparation). The top also had a physical stop to keep
77
(a) Kiwi robot and chair sensor connected to Arduino Uno R3.
(b) Kiwi robot and table
setup.
(c) Robot internal pan-tilt struc-
ture.
Figure 6.1: Experimental setup components and robot internal pan-tilt structure.
the head of the robot from falling all the way forward or all the way backwards when
the motors were not engaged. The Co-Robot Dialog system (CoRDial) (Short et al.,
2017a) was used to synchronize robot motions, facial expressions, speech, and visemes,
and an updated version of the expressive browser-based robot face from this system was
displayed on a ve inch LCD screen (Deng et al., 2019, in preparation). The Amazon
78
Polly \Justin" text-to-speech voice was used to generate speech for the robot. The ex-
ternal covering of the robot consisted of a bird-like design that has previously been used
for SAR interactions with older adults (Short et al., 2017b). This external covering was
designed using an iterative approach incorporating feedback from six older adults who
were show early design drawings in focus group-like interviews.
The Kiwi robot was attached to a small table that also housed the computer for
controlling the robot, speakers, a USB camera, a power strip, and the power adaptors
and cables needed for powering and connecting the LCD screen, the computer, and the
motors. The chair sensor part of the experimental setup consisted of a Recora Chair
Occupancy Sensor connected to an Arduino Uno R3 with a Bluetooth module. In order
to avoid detecting small changes in position as standing up or sitting down, a wait time
of two seconds was used before registering a change in the state of the chair sensor as a
stand or sit event.
6.1.3 Robot Behavior Design
The interactions with the robot consisted of two types of behaviors: reminders and
rewards. The robot made decisions about when to give reminders and rewards using
the habit formation support model described in Chapter 5. At the beginning of each
interaction, the robot \woke up" by lifting its head and opening its eyes, and at the
end \fell asleep" by lowering it's head forward and closing its eyes. The participant's
name was used in some of the greeting and praise statements in order to personalize the
interaction as was done by Fasola and Matari c (2012).
6.1.3.1 Reminder behaviors
Reminder behaviors included three statements: a greeting, sitting time feedback, and
a question. For example, \Hi [participant's name]. You've been sitting for 49 minutes.
79
Want to take a walking break?" The robot did expressive movements during the greeting
and question statements. In order to keep the reminders from becoming repetitive,
we randomly selected seven behaviors and corresponding movements for each possible
reminder time. Seven behaviors were selected for each reminder time based on our
estimate that participants were unlikely to receive a reminder for the same reminder
time more than seven times in a day. Thus, having seven variations for each reminder
time kept the reminders variable within a day. These seven reminders for each reminder
time were chosen from a set of six greetings, two phrasings of sitting time feedback, 11
questions, four greeting movements, and three question movements for use in the study.
These options were chosen because they were the largest numbers of variations on the
statements and movements we could create while keeping the variations comparable.
6.1.3.2 Reward behaviors
We designed three types of reward behaviors: dancing, telling a joke, and giving praise.
Our past research (Short et al., 2017b) indicated that robot dancing could be an appeal-
ing reward behavior. In addition, work studying the eects of comedy has found that it
can be inherently rewarding (Franklin and Adams, 2011), and past research has found
that praise can increase intrinsic motivation when given by a social agent and increase
motor performance when given by two robots (Mumm and Mutlu, 2011, Okumura et al.,
2017).
We created 33 of each type of reward behavior for a total of 99 reward behaviors
in order to have enough unique behaviors to avoid repetition during the study based
on a conservative estimated need of seven reward behaviors per day for 14 days for a
total of 98. One additional reward behavior was included in order to make the number
of behaviors for each type of reward equal. For each type of reward, these 33 behav-
iors were randomly selected from sets of statements, movements, and expressions and
80
detailed below. As with the reminder behaviors, the options for these sets were chosen
because they were the largest numbers of variations on the statements, movements, and
expressions we could create while keeping the variations comparable.
6.1.3.2.1 Dance Reward Behaviors
The dance reward behaviors consisted of two statements (praise and a dance introduc-
tion) followed by a dance routine. An example of these two statements together is,
\Good going: [participant's name]! Time to dance!" A short expressive movement
was done at the beginning of the praise statement, and an expressive facial expression
was done at the beginning of the dance introduction statement. Next, one of two four-
second dancing routines was randomly selected and performed by the robot while one
of 27 lyric-free eight-second music clips played with a fade-in and out. We created two
dancing routines in order to add variation to the dancing reward behavior. The low
number of degrees-of-freedom of the robot, however, limited the number of unique rou-
tines we could create. The routines were kept to four seconds long in order to keep the
total behavior time similar to the joke and praise reward behaviors. The music clips
were cut to eight seconds to t with the timing of the robot's movements. Twenty seven
music clips were used in order to add variation. The dance behaviors used in the study
were created by randomly selecting components from a set of 12 praise statements, eight
dance introduction statements, two expressive movement, and two facial expressions.
6.1.3.3 Joke Reward Behaviors
The joke reward behaviors consisted of four statements: praise, joke preface, question,
and punchline. Question and answer style jokes comprised the question and punchline
statements. An example of a complete joke statement is, \Well done! My jokes are
pretty corny, but here goes. What did the left eye say to the right eye? Between you and
81
me, something smells." The joke preface statement was included in order to establish
that the robot acknowledges that the joke is cheesy or corny. This type of preface can
improve the performance of formulaic jokes (Bird, 2008). An expressive movement was
done at the beginning of the praise statement, and additional expressions were done
at the end of the preface and beginning of the question. Each of the question and
answer style jokes were combined with randomly selected praise (from 12 statements),
preface (from 12 statements), expressive movement (from four movements), post-preface
expression (from two expressions), and pre-question expression (from three expressions)
to create the joke reward behaviors.
6.1.3.4 Praise Reward Behaviors
Finally, the praise reward behaviors consisted of a single praise statement (chosen from
15 options) and an expressive movement at the beginning of that statement (chosen
from three options). The praise reward behaviors were randomly selected from these
options for use in the study. An example praise reward statement was \Nice work!"
6.1.4 Model Parameter Selection
The following parameter values were used in model of habit formation support for the
habit of getting up after 30 min of sitting.
Number of episodes from one automaticity probe to the next, a = 3
Number of episodes in the past window of automaticity probe episodes, p = 3
Number of episodes in a reward block, b = 6
Number of rewarded episodes in a reward block, r = 4
Maximum number of reminders for an episode, m = 13
82
Time period over which reminders can be given, T = 3 hours
Constant in Equation 5.1, = 0:015
Constant in Algorithm 5.3,
= 0:25
We recruited participants who typically sit three or more hours per day at home
in the same chair. Based on this recruitment criteria, we estimated a lower bound on
the number of sitting episodes we expect participants to have in a day to be six, or the
number of episodes that t within three hours. Based on this lower bound estimate,
we chose three as the number of episodes from one automaticity probe to the next in
order to have an estimated two automaticity updates per day. We chose this in order
to keep the robot's behavior adaptive within the time frame of a day to any changes
in automaticity. We also chose three as the number of episodes in the past window
of automaticity probe episodes used to update the automaticity estimate in order to
incorporate an estimated day or two of past automaticity probe data into the current
automaticity estimate. This was done to keep any single sitting episode from greatly
increasing or decreasing the automaticity estimate since sitting time in any one episode
could be aected by factors other than automaticity for the habit.
We chose six as the number of episodes in a reward block in order to cycle through
one reward block per day given our lower bound estimate of six episodes per day. We
chose four as the number of rewarded episodes in each block so that the robot would
have multiple interactions with the person per day with the lower bound of six episodes
even if no reminders were given.
Using these parameters the reminder schedules for giving 1-13 reminders were cal-
culated. Figure 6.2 shows these reminder schedules. The dots in this gure represent
the time post-cue when a reminder was given. This gure also shows the estimated
83
0 20 40 60 80 100 120 140 160 180
Time Post-Cue (Minutes)
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
Estimated Automaticity
0
1
2
3
4
5
6
7
8
9
10
11
12
13
Number of Reminders
Estimated Automaticity and Reminder Schedules
Estimated Automaticity
Scheduled Reminders
Inflection Point
Figure 6.2: Relationship between estimated automaticity and reminder schedules for
selected model parameters. The estimated automaticity shown is a function of the
post-cue average sitting time for the automaticity probes in the past window. The
grey and white bars correspond to the number of reminders that were scheduled as a
function of the estimated automaticity. The green dot represents an in
ection point for
automaticity values greater than 0.74 where reminders no longer start until after the
average post-cue sitting time.
automaticity for dierent post-cue sitting time averages using Equation 5.1. The alter-
nating grey and white horizontal bars correspond to the step function for determining
how many reminders to give based on the estimated automaticity. For example, an esti-
mated automaticity between 0 and 0.038 corresponds to 13 reminders. For automaticity
values greater than 0.74 reminders were given after the average post-cue sitting time.
This means that for the estimated automaticity 0.74, if the user were to sit during future
episodes less than their average sitting times during the last three automaticity probes
(< 50:3 min), then they did not receive any reminders from the robot. This point is
represented with a green dot on the automaticity curve, and is called the in
ection
point.
84
We chose the maximum number of reminders per episode as 13, the time period
over which reminders would be given as three hours, and the constant
as 0.25 in
order to create reminder schedules that would include frequent at low automaticity
levels, without overwhelming the user with reminders. With these parameter values,
the maximum number of reminders that could be given in an hour of sitting was ve.
We chose the constant value of 0.015 so that the in
ection point for these reminder
schedules would be located around 75% of the maximum automaticity. This value was
chosen so that the last 25% of automaticity increase could be aected mainly by social
rewards rather than reminders.
6.1.5 Study Design
In order to evaluate the eects of the habit formation support system on sedentary be-
havior we used a mixed methods three-phase single-case study design. The rst phase
of the study was a baseline phase in which we collected data for one week prior to the
introduction of the robot. The second phase was a two week long intervention phase
where the SAR system was placed in the participant's home for two weeks. The third
phase was a one-week long retention phase in which the data collection of the rst phase
was repeated. Because the goal of the intervention is to support the development of
a sustained habit, the study was not designed with the expectation of any changes to
sedentary behavior being reversible. We used a combination of quantitative behavioral
data, subjective questionnaire measures, and qualitative data to evaluate the interven-
tion.
6.1.6 Procedure
The study took place entirely in the participants' homes. There were between 4 and
6 scheduled home visits to administer questionnaires, set up equipment, check system
85
functionality, remove equipment, and administer a semi-structured interview. The chair
sensor, Arduino Uno R3 with Bluetooth module, and computer were set up at the
beginning of the study and removed at the end. The chair sensor was placed in the
chair that the participant sat in most of the time while at home. The robot was set up
at the beginning of the second week and removed after the third week. The robot was
placed in the same room near the chair that the participant sat in most of the time while
at home. Participants were introduced to the robot during a short tutorial after the
robot was set up in the home at the beginning of the second week. During this tutorial
participants were shown a demonstration of the reminder and reward behaviors. At the
end of the fourth week participants received compensation of a $200 Amazon gift card
($50 for each week of the study).
6.1.7 Data Collection and Outcome Measures
A combination of subjective questionnaire measures, behavioral data, and qualitative
interview data were collected in order to evaluate the habit formation intervention with
the robot. An overview of study procedure and data collection is shown in Table 6.1.
Questionnaires were given at the beginning of the rst week, beginning of the second
week, end of the third week, and end of the fourth week of the study. We also asked
participants to complete a daily questionnaire, the 4-item Self-Report Behavioural Au-
tomaticity Index (SRBAI), which uses a 5-point Likert scale to measure subjective
habit automaticity (Gardner et al., 2012). The pre-week-1 questionnaires included a
demographics questionnaire and the Big Five Inventory, a 44-item questionnaire which
measures personality along ve dimensions using a 5-point Likert scale (Benet-Mart nez
and John, 1998, John et al., 1991, 2008). It also included a shortened 8-item version
of the Physical Activity Enjoyment Scale (PACES-8) that uses a 7-point bipolar rat-
ing scale to measure physical activity enjoyment (Mullen et al., 2011) and the 14-item
86
Table 6.1: Overview of study procedure and data collection.
Pre-Week-1 Questionnaires
Personality (Big-5)
Demographics
Physical activity enjoyment (PACES-8)
Attitudes towards robots (NARS)
Week 1 No robot
Pre-Week-2 Questionnaires
Robot acceptance (Almere)
Physical activity enjoyment (PACES-8)
Attitudes towards robots (NARS)
Chair sitting time
Week 2
Robot
Interaction audio and video
Accelerometer data
Week 3
Post-Week-3 Questionnaires
Robot acceptance (Almere)
Physical activity enjoyment (PACES-8)
Attitudes towards robots (NARS)
Open-ended questions
Daily habit
automaticity
(SRBAI)
Week 4 No robot
Post-Week-4 Questionnaires
Physical activity enjoyment (PACES-8)
Attitudes towards robots (NARS)
Semi-structured interview
Negative Attitude toward Robots Scale (NARS) that measures attitudes towards robots
using a 5-point Likert scale (Nomura et al., 2006). The PACES-8 and the NARS ques-
tionnaires were also given at three additional time points (immediately after the tutorial
at the beinning of the second week, after the third week, and after the fourth week). Ad-
ditionally, the Almere Model, a modied version of the Unied Theory of Acceptance
and Use of Technology (UTAUT) Model specically designed to measure acceptance
of social agents by older adults, was given at the beginning of the second week after
the tutorial and after the third week (Heerink et al., 2010). The Almere is a 41-item
87
questionnaire that uses a 5-point Likert scale. Four open-ended questions about the
interaction with the robot were given at the end of week three.
Behavioral data related to sedentary behavior and activity levels were also collected
during all four weeks of the study. Sitting time was measured using the chair sensor in
the chair the participant sat in most at home connected to the computer over Bluetooth
with the Arduino Uno R3. This recorded the times at which the participant began and
stopped sitting in the chair. Accelerometer data were also collected using a wrist-worn
GENEActiv Original accelerometer initialized to collect data at a sampling frequency
of 20 Hz. Participants were asked to wear the accelerometer during waking hours. In
addition, audio and video of the participant were recorded during reminder and reward
interactions for 20 seconds following the start of each interaction.
Robot and model behaviors were also recorded including the timing of reminders
and rewards, the reminder schedules for each episode, and the estimated automaticity
for each episode. Finally, a semi-structured interview was given at the end of the
fourth week of the study in order to collect subjective qualitative feedback about the
participants' experience with the robot. This interview was audio and video recorded
and transcribed.
6.1.8 Study Population
Inclusion criteria for the study were for participants to be 65-80 years old, live alone with
no pets, typically sit on one chair at home that is close to a wall outlet for three or more
hours per day, be able to hear the robot's speech, and be able to stand up regularly
and walk around. Participants who were not
uent in English or who had cognitive
impairments were not included in the study. Two participants were enrolled in the
study. One participant (P1) completed the study, and one participant (P2) completed
all but the last three days of the study as well as the post-week-4 questionnaires and
88
the semi-structured interview. The participants were 72 and 78 years old, both were
female, and both rated their experience working with robots as none on a 4-point scale
from 1 (none) to 4 (a lot).
6.2 Results
The SAR system operated fully autonomously in participants' homes during the two
week intervention portion of the study. The research team was contacted by participants
three times during the robot interventions related to problems with the robot. One time
was to conrm that the robot was operating correctly. Another was because the browser-
based robot face had crashed, which was xed the next day. The third was because of
a power outage which caused the robot's computer to shut down. The system was
restarted later that day. Because of a technical issue with restarting the system no
rewards were given during the reward block that began one episode before the power
outage.
6.2.1 Sitting Episodes with Greater than 30 Minute Duration
In order to understand the eect of the intervention on the habit of standing up after
sitting for 30 min, we conducted an exploratory data analysis. We visually inspected
the sitting time data by plotting the durations of all sitting episodes greater than 30
min long in order for each of the three phases of the study as shown in Figure 6.3. In
addition to the duration of each sitting episode over 30 min, the reminders given by the
robot, the rewarded episodes, the estimated automaticity, and the mean sitting duration
for each phase of the study are included in this gure.
P1 had a total of 27 sitting episodes longer than 30 min in the baseline phase with
mean sitting duration 49.2 min (SD = 14:9 min, Range: 30:0 92:4 min), 53 in the
intervention (M = 45:5 min, SD = 16:3 min, Range: 30:2 126:3 min), and 33 in the
89
1 4 7 10 13 16 19 22 25 1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 2 5 8 11 14 17 20 23 26 29 32
Baseline Episode Number Intervention Episode Number Retention Episode Number
0
50
100
150
200
250
Sitting Duration (Minutes)
0
0.2
0.4
0.6
0.8
1.0
Estimated Automaticity
P1 Sitting Episodes Greater than 30 Minutes and Estimated Automaticity
1 4 7 10131619222528313437 1 4 7 101316192225283134374043464952555861646770737679 1 4 7 10131619
Baseline Episode Number Intervention Episode Number Retention Episode Number
0
50
100
150
200
250
Sitting Duration (Minutes)
Epidode with no Reward
Episode with Reward
Reminder
Estimated Automaticity
Start/End Intervention
Baseline Mean Sitting Duration
Intervention Mean Sitting Duration
Retention Episode Number
0
0.2
0.4
0.6
0.8
1.0
Estimated Automaticity
P2 Sitting Episodes Greater than 30 Minutes and Estimated Automaticity
Figure 6.3: Sitting duration for episodes over 30 min long and estimated automaticity
for intervention episodes for each participant. Plots show when the robot did reminder
and reward behaviors based on the habit formation support model. They also show
the estimated automaticity used in the model for each episode in the intervention.
Green horizontal lines show the average sitting duration for episodes over 30 min for
the baseline, intervention, and retention episodes. The gap in rewarded episodes for P2
was due to technical issue with restarting the system.
90
retention (M = 50:5 min,SD = 20:6 min,Range: 31:9 132:1 min). P2 had a total of
38 sitting episodes longer than 30 min in the baseline phase with mean sitting duration
86.7 min (SD = 51:4 min, Range: 30:6 230:2 min), 82 in the intervention (M = 55:0
min,SD = 30:8 min,Range: 30:3 91:2 min), and 20 in the retention (M = 91:2 min,
SD = 59:6 min, Range: 38:7 217:3 min).
P1 received a total of 11 reminders from the robot, and always stood up after
receiving a reminder. P2 received a total of 44 reminders from the robot, and always
stood up after receiving a reminder except for one episode in which two reminders
were given before the participant stood up. P1's estimated automaticity began at
0.51 and ended at 0.72 with an average of 0.68 (SD = 0:14;Range: 0:45 0:92), and
P2's estimated automaticity began at 0.54 and ended at 0.25 with an average of 0.48
(SD = 0:16;Range: 0:25 0:94).
6.2.2 All Sitting Episodes
In order to examine the eect of the intervention on sitting behavior including sitting
episodes shorter than 30 min, we calculated descriptive statistics based on all of the
sitting episodes measured by the chair sensor for each participant in each of the three
phases of the study as shown in Table 6.2. The usual sitting duration was calculated
based on the \usual bout duration," or the duration of the sitting bout at which half of
the sitting time has occurred, described by Stephens et al. (2014). In order to calculate
the usual sitting duration, we calculated the cumulative (ordered by duration) percent
contribution to total sitting time of each sitting episode. We then linearly interpolated
between the two episodes closest to 50% in order to estimate the sitting episode duration
at 50% of the cumulative sitting time.
91
Table 6.2: Mean, median, and usual sitting episode duration of all sitting episodes per
participant for each study phase. All units are in minutes.
Mean Median Usual Sitting Duration
P1
Baseline 18.8 (SD = 20.6) 10.2 41.8
Intervention 13.4 (SD = 17.6) 5.9 36.0
Retention 20.4 (SD = 22.5) 11.1 40.0
P2
Baseline 47.2 (SD = 53.0) 26.1 93.7
Intervention 30.2 (SD = 31.0) 21.0 42.4
Retention 46.3 (SD = 57.5) 21.9 119.5
6.3 Discussion
The SAR habit formation support model was able to personalize the robot's reminder
timing to individual dierences. The two participants diered in their sitting behavior,
and that dierence aected the model in several notable ways.
First, they diered in their average baseline sitting duration for sitting episodes over
30 min. P1 had a much lower baseline at 49.2 min than P2 at 86.7 min. These initial
dierences seem related to the estimated automaticity. For example, P2 had a lower
average automaticity over all of the intervention episodes than P1 suggesting that there
may be an inverse relationship between the baseline average sitting duration and the
estimated automaticity during the intervention.
Additional dierences in estimated automaticity include a dierence in variation
and a dierence in the net change from the beginning to the end of the intervention.
P1's estimated automaticity varies within a smaller range than P2's, and P1 had a net
increase in estimated automaticity from the beginning to the end of the intervention.
P2, on the other hand, despite having a maximum estimated automaticity slightly
higher than P1's maximum, ultimately had a net decrease from the beginning to the
end of the intervention. This decrease in estimated automaticity could indicate that
P2's automaticity for the habit did not increase.
92
P1 also had a larger percentage of episodes with automaticity above the in
ection
point during the intervention with 18 (34.0%) episodes than P2 with 6 (7.3%) episodes.
The lower number of reminders given to P1 could indicate that the model parameters
used for this study should be changed in order to move the in
ection point to a lower
average time post-cue. With the model parameters selected in this study, P1 could be
considered to have \maxed out" of receiving reminders for much of the intervention.
Both participants had lower mean sit durations over 30 min during the intervention
than during the baseline and retention phases with P2 having a greater dierence than
P1. This dierence in how much each participant responded to the intervention is
not surprising given that P2 started with a higher average sitting duration than P1.
Neither participant, however, had a reduction in average sitting duration for sitting
episodes greater than 30 min from baseline to retention. This could indicate that their
automaticity for the habit did not increase. In addition, neither participant had a large
decrease from baseline to retention in mean, median, or usual sitting duration for all
sitting episodes. This indicates that the intervention may not have had a sustained
eect lasting longer than the intervention on sitting behavior.
Both participants had high adherence to the reminders from the robot, meaning
that they almost always got up immediately after a reminder was given. There was
only one episode for which two reminders were needed. The reminders seemed to have
more of an impact on sitting behavior than the rewards. For example, the automaticity
probe episodes for P2, which had no reminders, had an average sitting duration of 83.0
min, which is only a small decrease from the baseline of 86.7 min suggesting that the
sitting episodes with no reminders were relatively unaected by the intervention for this
participant. The high adherence to reminders also suggests that the scheduled timing
of additional reminders for a single episode of sitting might be less important than the
timing of the rst reminder because multiple reminders are usually not needed.
93
There were several limitations of this study. The largest limitation was the small
number of participants. Although we were able to examine how each participant re-
sponded to the intervention, the limited number of participants means that we cannot
generalize these results. Another limitation is in the design of the study phases. First,
because the intervention was always introduced after one week of baseline measurement,
this study design does not conform to the multiple-baseline single-case design in which
the intervention is introduced at a random time point (Byiers et al., 2012). This ad-
ditionally limits the generalizability of the study. In addition, the intervention period
may not have been suciently long enough for a habit to form. While the one-week
post-intervention retention phase is informative, an additional retention measurement
at a later time point would be useful to see if the sitting behavior changed well past the
intervention.
We saw from this study that using the habit formation support model the robot was
able to personalize its behavior to each participant's sitting behavior. Future work could
further personalize other aspects of the robot's behavior such as providing reminders
of dierent salience and rewards personalized to what the participant enjoys. We also
saw that participants had a high adherence to reminders in this study. Future work
could examine how that adherence might change over a longer intervention period.
Although we did not see evidence of a sustained habit formation in this study, future
work could explore dierent ways of supporting habit formation with a robot. For
example, incorporating an activity prediction model (Minor et al., 2015) for predicting
when the user would stand up without intervention could inform when the robot should
provide reminders to avoid providing a reminder close to when the participant would
otherwise do the behavior.
94
6.4 Summary
This chapter presented the design and evaluation of a fully autonomous SAR system
designed to support older adults in forming a habit of getting up after sitting for 30
min. This system provided real-time in situ feedback and encouragement using the
habit formation support model to participants in their homes. Although we did not
see any evidence of sustained habit formation from baseline to retention, we did see
that participants had high adherence to reminders given by the robot to take a break
from sitting. We also saw that participant variation in baseline average sitting duration
aected the estimated automaticity and the number of reminders given by the robot.
95
Chapter 7
Dissertation Summary
7.1 Summary
This dissertation examined the SAR role of supporting physical activity behavior
change. The role of a robot in supporting physical activity behavior change was de-
ned as using a SAR support method towards the goal of aecting specic behavior
change mechanisms of action within a physical activity domain. This dissertation intro-
duced a framework for SAR physical activity behavior change support that incorporated
this role of the robot. The framework was designed with the goal of enabling a robot
to provide support during the behavior change process in order to create sustained
long-term change that persists after the intervention with the robot. The framework
was instantiated in three physical activity domains, and developed four SAR support
methods including SAR feedback, backstory, messaging, and reminder and social reward
decision making. These support methods were evaluated with fully autonomous SAR
systems.
The rst instantiation of the framework centered around the SAR feedback support
method. This feedback was designed to be comparative with the goal of aecting the
96
beliefs about capabilities, which relates to self-ecacy, and social in
uences behavior
change mechanisms of action (Carey et al., 2018). This support method was evaluated
in the domain of post-stroke upper extremity rehabilitation, and found that participants
improved at an upper extremity reaching task over time when guided by a SAR coach.
This work additionally found that participants had the shortest delay time before be-
ginning an upper extremity reaching task when they received self-comparative feedback
from the robot compared to other- or no-comparative feedback, suggesting that self-
comparative feedback should used rather than other-comparative for the SAR feedback
support method. This work also grouped participants using an unsupervised machine
learning clustering algorithm based on their performance and response to self-ecacy
probes.
The second instantiation centered around the SAR backstory and messaging sup-
port methods. The design of these support methods was targeted towards the social
in
uences and attitude towards the behavior mechanisms of action for behavior change
(Carey et al., 2018). These support methods were evaluated in the domain of adolescent
exercise using the robot in the role of an exercise buddy. Two robot backstories, ctional
and realistic, and a set of aective messages were designed based on feedback from focus
groups. Participants rated the robot positively, and some participants indicated that
they liked receiving encouragement from the robot.
The nal instantiation of the SAR physical activity behavior change support frame-
work developed the reminder and social reward decision making support method tar-
geting the behavioral cueing and reinforcement behavior change mechanisms of action
(Carey et al., 2018). This method was formalized through a model of SAR habit for-
mation support designed to enable a robot to determine when to provide reminders and
social rewards in order to support habit formation. This model was evaluated in the
domain of reducing older adult sedentary behavior by supporting the habit of getting
97
up after sitting for 30 minutes. The robot provided reminders and gave social rewards
in real-time in participants' homes during a two-week intervention. This work found
that participants had high adherence to the reminders from the robot.
These three instantiations together contribute the design of SAR feedback, back-
story, and messaging support methods and a model of SAR habit formation support as
well as evaluations of these methods in three physical activity domains.
7.2 Future Directions
Behavior change for physical activity is a complex process. The SAR support methods
presented in this dissertation have identied several important areas for future work in
supporting this complex process. This dissertation presented methods for automatically
gathering information about a user's self-ecacy for a specic task and estimating au-
tomaticity for a specic habit, but models for quantifying and estimating these abstract
concepts (self-ecacy and automaticity) in more generalizable ways over time with little
or no additional input from the user are needed. The subjective and behavioral measures
that have previously been developed for measuring these psychological concepts should
serve as the basis and sometimes baseline for developing continuous measures, but be-
cause they take time for the user to complete, other methods are needed that minimize
queries to the user and enable a robot to maintain a continuous estimate. Probabilistic
methods such as Bayesian knowledge tracing have shown promise in the domain of SAR
tutoring systems for estimating various types of user knowledge (Clabaugh and Matari c,
2019), and could be a promising model for this application.
Models for SAR decision making that take into account the current context of the
user could also be benecial especially in interventions that include frequent and im-
mediate feedback, praise, and encouragement. For example, in this dissertation, the
habit formation support model did not take into account potentially important context
98
variables such as the time of day, day of week, or state of the activity the user was
engaged in such as watching TV. These variables could possibly aect the user's behav-
ior and could be important for the robot to take into account when making decisions
about when and how to intervene. Probabilistic representations that incorporate these
variables into the state space such as Markov decision processes could be benecial.
More work towards the personalization of SAR support behaviors is also needed.
In this dissertation, SAR behaviors were designed for the context of each application,
but not for each individual. There could be a variety of individual dierences that were
not taken into account that might be important. For example, dierent people might
enjoy one type of social reward from the robot more than another. Additionally, as SAR
behavior change interventions continue to move more into real world environments such
as the home, oce, clinic, hospital, assisted living, and skilled nursing facilities, and
involve many interactions in these real world environments, the relationship dynamics
between the user and the robot over time could be important to model in future work
related to SAR behavior change support.
99
Bibliography
H. Aarts and A. Dijksterhuis. Habits as knowledge structures: automaticity in goal-
directed behavior. Journal of personality and social psychology, 78(1):53, 2000.
J. Adams and M. White. Are activity promotion interventions based on the transthe-
oretical model eective? a critical review. British Journal of Sports Medicine, 37(2):
106{114, 2003.
C. M. Atance and D. K. O'Neill. Episodic future thinking. Trends in cognitive sciences,
5(12):533{539, 2001.
S. J. Atlas and R. A. Deyo. Evaluating and managing acute low back pain in the primary
care setting. Journal of general internal medicine, 16(2):120{131, 2001.
R. Badami, M. VaezMousavi, G. Wulf, and M. Namazizadeh. Feedback after good
versus poor trials aects intrinsic motivation. Research Quarterly for Exercise and
Sport, 82(2):360{364, 2011.
R. R. Bailey, J. W. Klaesner, and C. E. Lang. Quantifying real-world upper-limb activity
in nondisabled adults and adults with chronic stroke. Neurorehabilitation and neural
repair, 29(10):969{978, 2015.
B. R. Ballester, M. Maier, R. M. S. S. Mozo, V. Casta~ neda, A. Du, and P. F. Verschure.
Counteracting learned non-use in chronic stroke patients with reinforcement-induced
movement therapy. Journal of neuroengineering and rehabilitation, 13(1):74, 2016.
A. Bandura. Self-ecacy: toward a unifying theory of behavioral change. Psychological
review, 84(2):191, 1977.
A. Bandura. Self-ecacy: The exercise of control. W.H. Freeman & Company, 1997.
V. Benet-Mart nez and O. P. John. Los Cinco Grandes across cultures and ethnic
groups: multitrait multimethod analyses of the Big Five in Spanish and English. J.
of personality and Social Psychology, 75(3):729{50, sep 1998.
100
E. J. Benjamin, S. S. Virani, C. W. Callaway, A. M. Chamberlain, A. R. Chang,
S. Cheng, S. E. Chiuve, M. Cushman, F. N. Delling, R. Deo, et al. Heart disease
and stroke statistics2018 update: a report from the american heart association. Cir-
culation, 137(12):e67{e492, 2018.
T. Bickmore, D. Schulman, and L. Yin. Engagement vs. deceit: Virtual humans with
human autobiographies. In In Proc. of the 9th Int. Conf. on Intelligent Virtual Agents
(IVA), pages 6{19, 2009.
T. W. Bickmore, R. A. Silliman, K. Nelson, D. M. Cheng, M. Winter, L. Henault, and
M. K. Paasche-Orlow. A randomized controlled trial of an automated exercise coach
for older adults. Journal of the American Geriatrics Society, 61(10):1676{1683, 2013.
C. Bird. it's not very funny: Heightened performance through prefaces of formulaic jokes
in interaction. In Proceedings of the Sixteenth Annual Symposium About Language
and Society-Austin. Citeseer, 2008.
R. Bize, J. A. Johnson, and R. C. Plotniko. Physical activity level and health-related
quality of life in the general adult population: a systematic review. Preventive
medicine, 45(6):401{415, 2007.
A. A. Blank, J. A. French, A. U. Pehlivan, and M. K. O'Malley. Current trends in robot-
assisted upper-limb stroke rehabilitation: promoting patient engagement in therapy.
Current physical medicine and rehabilitation reports, 2(3):184{195, 2014.
M. N. Boulos, S. Wheeler, C. Tavares, and R. Jones. How smartphones are changing the
face of mobile and participatory healthcare: an overview, with example from ecaalyx.
Biomedical engineering online, 10(1):24, 2011.
B. J. Byiers, J. Reichle, and F. J. Symons. Single-subject experimental design for
evidence-based practice. American journal of speech-language pathology, 2012.
L. A. Cadmus-Bertram, B. H. Marcus, R. E. Patterson, B. A. Parker, and B. L. Morey.
Randomized trial of a tbit-based physical activity intervention for women. American
journal of preventive medicine, 49(3):414{418, 2015.
R. N. Carey, L. E. Connell, M. Johnston, A. J. Rothman, M. de Bruin, M. P. Kelly, and
S. Michie. Behavior change techniques and their mechanisms of action: a synthesis of
links described in published intervention literature. Annals of Behavioral Medicine,
2018.
F. M. Carrillo, J. Butchart, S. Knight, A. Scheinberg, L. Wise, L. Sterling, and C. Mc-
Carthy. Adapting a general-purpose social robot for paediatric rehabilitation through
in situ design. ACM Transactions on Human-Robot Interaction (THRI), 7(1):12,
2018.
101
J. Cassell. Embodied conversational agents. MIT press, 2000.
Centers for Disease Control and Prevention. Prevalence and most common causes of
disability among adults{united states, 2005. MMWR: Morbidity and Mortality weekly
report, 58(16):421{426, 2009.
A. L. Chasteen, D. C. Park, and N. Schwarz. Implementation intentions and facilitation
of prospective memory. Psychological Science, 12(6):457{461, 2001.
N. L. Chatzisarantis and M. S. Hagger. Eects of an intervention based on self-
determination theory on self-reported leisure-time physical activity participation. Psy-
chology and Health, 24(1):29{48, 2009.
S. Chen, R. Lewthwaite, N. Schweighofer, and C. J. Winstein. Discriminant validity of a
new measure of self-ecacy for reaching movements after stroke-induced hemiparesis.
Journal of Hand Therapy, 26(2):116{123, 2013.
T. L. Chen, T. Bhattacharjee, J. M. Beer, L. H. Ting, M. E. Hackney, W. A. Rogers,
and C. C. Kemp. Older adults' acceptance of a robot for partner dance-based exercise.
PloS one, 12(10):e0182736, 2017.
C. Clabaugh and M. Matari c. Escaping oz: autonomy in socially assistive robotics.
Annu. Rev. Control Robot. Auton. Syst., 2:3361, 2019.
C. Clabaugh, D. Becerra, E. Deng, G. Ragusa, and M. Matari c. Month-long, in-home
case study of a socially assistive robot for children with autism spectrum disorder.
In Companion of the 2018 ACM/IEEE International Conference on Human-Robot
Interaction, pages 87{88. ACM, 2018a.
C. Clabaugh, S. Jain, B. Thiagarajan, Z. Shi, L. Mathur, G. Ragusa, and M. Matari c.
Attentiveness of children with diverse needs to a socially assistive robot in the home.
In International Symposium on Experimental Robotics, 2018b.
S. Cohen, T. Kamarck, and R. Mermelstein. A global measure of perceived stress. J.
of Health and Social Behavior, pages 385{396, 1983.
L. E. Connell, R. N. Carey, M. de Bruin, A. J. Rothman, M. Johnston, M. P. Kelly,
and S. Michie. Links between behavior change techniques and mechanisms of action:
an expert consensus study. Annals of Behavioral Medicine, 2018.
M. Conner. Health cognitions, aect and health behaviors. The European Health Psy-
chologist: Bulletin of the European Health Psychology Society, 15(2):33{39, 2013.
M. Conner, R. E. Rhodes, B. Morris, R. McEachan, and R. Lawton. Changing exercise
through targeting aective or cognitive attitudes. Psychology and Health, 26(2):133{
149, 2011.
102
M. Conner, R. McEachan, N. Taylor, J. O'Hara, and R. Lawton. Role of aective
attitudes and anticipated aective reactions in predicting health behaviors. Health
Psychology, 34(6):642, 2015.
C. Craggs, K. Corder, E. M. van Sluijs, and S. J. Grin. Determinants of change in
physical activity in children and adolescents: a systematic review. American J. of
Rreventive Medicine, 40(6):645{658, 2011.
J. N. Davis, L. E. Gyllenhammer, A. A. Vanni, M. Meija, A. Tung, E. T. Schroeder,
D. Spruijt-Metz, and M. I. Goran. Startup circuit training program reduces metabolic
risk in latino adolescents. Medicine and Sci. in Sports and Exercise, 43(11):2195, 2011.
M. M. De Graaf, S. B. Allouch, and T. Klamer. Sharing a life with harvey: Exploring
the acceptance of and relationship-building with a social robot. Computers in human
behavior, 43:1{14, 2015.
E. Deng, N. S. Dennler, K. Swift-Spong, and M. J. Matari c. Mosir: A modular, open-
source interactive robotplatform for in-situ experimentation. Department of Computer
Science, University of Southern California. Tech Report, 2019, in preparation.
A. Direito, E. Carra ca, J. Rawstorn, R. Whittaker, and R. Maddison. mhealth tech-
nologies to in
uence physical activity and sedentary behaviors: behavior change tech-
niques, systematic review and meta-analysis of randomized controlled trials. Annals
of behavioral medicine, 51(2):226{239, 2016.
S. Dogra and L. Stathokostas. Sedentary behavior and physical activity are indepen-
dent predictors of successful aging in middle-aged and older adults. Journal of aging
research, 2012, 2012.
D. D. Dunlop, J. Song, E. K. Arntson, P. A. Semanik, J. Lee, R. W. Chang, and J. M.
Hootman. Sedentary time in us older adults associated with disability in activities of
daily living independent of physical activity. Journal of Physical Activity and Health,
12(1):93{101, 2015.
J. Edmunds, N. Ntoumanis, and J. L. Duda. Testing a self-determination theory-based
teaching style intervention in the exercise domain. European Journal of Social Psy-
chology, 38(2):375{388, 2008.
G. O. Einstein, M. A. McDaniel, R. Thomas, S. Mayeld, H. Shank, N. Morrisette,
and J. Breneiser. Multiple processes in prospective memory retrieval: factors deter-
mining monitoring versus spontaneous retrieval. Journal of Experimental Psychology:
General, 134(3):327, 2005.
103
T. Ellis, N. K. Latham, T. R. DeAngelis, C. A. Thomas, M. Saint-Hilaire, and T. W.
Bickmore. Feasibility of a virtual exercise coach to promote walking in community-
dwelling persons with parkinson disease. American journal of physical medicine &
rehabilitation/Association of Academic Physiatrists, 92(6):472, 2013.
J. Eriksson, M. J. Matari c, and C. Winstein. Hands-o assistive robotics for post-
stroke arm rehabilitation. In Proc. IEEE International Conference on Rehabilitation
Robotics (ICORR'05), pages 21{24, 2005.
K. R. Evenson, D. J. Catellier, K. Gill, K. S. Ondrak, and R. G. McMurray. Calibration
of two objective measures of physical activity for children. J. of Sports Sciences, 26
(14):1557{1565, 2008.
J. Fasola and M. Matari c. A socially assistive robot exercise coach for the elderly.
Journal of Human-Robot Interaction, 2(2):3{32, 2013a.
J. Fasola and M. J. Matari c. Using socially assistive human{robot interaction to moti-
vate physical exercise for older adults. Proc. of the IEEE, 100(8):2512{2526, 2012.
J. Fasola and M. J. Matari c. Socially assistive robot exercise coach: Motivating older
adults to engage in physical exercise. In Experimental Robotics, pages 463{479.
Springer, 2013b.
D. Feil-Seifer and M. J. Matari c. Dening socially assistive robotics. In in Proc. of
the 9th IEEE International Conference on Rehabilitation Robotics (ICORR), pages
465{468, 2005a.
D. Feil-Seifer and M. J. Matari c. Dening socially assistive robotics. In in Proc. of
the 9th IEEE International Conference on Rehabilitation Robotics (ICORR), pages
465{468, 2005b.
C. Foster, M. Hillsdon, M. Thorogood, A. Kaur, and T. Wedatilake. Interventions
for promoting physical activity. The Cochrane database of systematic reviews, (1):
CD003180, 2005.
R. G. Franklin and R. B. Adams. The reward of a good joke: neural correlates of
viewing dynamic displays of stand-up comedy. Cognitive, Aective, & Behavioral
Neuroscience, 11(4):508{515, 2011.
P. Gadde, H. Kharrazi, H. Patel, and K. F. MacDorman. Toward monitoring and in-
creasing exercise adherence in older adults by robotic intervention: a proof of concept
study. J. of Robotics, 2011, 2011.
B. Gardner, C. Abraham, P. Lally, and G.-J. de Bruijn. Towards parsimony in habit
measurement: Testing the convergent and predictive validity of an automaticity sub-
scale of the self-report habit index. International Journal of Behavioral Nutrition and
Physical Activity, 9(1):102, 2012.
104
L. V. Gauthier, E. Taub, C. Perkins, M. Ortmann, V. W. Mark, and G. Uswatte.
Remodeling the brain: plastic structural brain changes produced by dierent motor
therapies after stroke. Stroke, 39(5):1520{1525, 2008.
P. M. Gollwitzer. Implementation intentions: strong eects of simple plans. American
psychologist, 54(7):493, 1999.
P. M. Gollwitzer and V. Brandst atter. Implementation intentions and eective goal
pursuit. Journal of personality and social psychology, 73(1):186, 1997.
S. Gonul, T. Namli, M. Baskaya, A. A. Sinaci, A. Cosar, and I. H. Toroslu. Optimization
of just-in-time adaptive interventions using reinforcement learning. In International
Conference on Industrial, Engineering and Other Applications of Applied Intelligent
Systems, pages 334{341. Springer, 2018.
B. G orer, A. A. Salah, and H. L. Akn. A robotic tness coach for the elderly. In
International Joint Conference on Ambient Intelligence, pages 124{139. Springer,
2013.
B. G orer, A. A. Salah, and H. L. Akn. An autonomous robotic exercise tutor for elderly
people. Autonomous Robots, 41(3):657{678, 2017.
W. S. Grolnick and R. M. Ryan. Parent styles associated with children's self-regulation
and competence in school. J. of Educational Psychology, 81(2):143, 1989.
H.-M. Gross, A. Scheidig, K. Debes, E. Einhorn, M. Eisenbach, S. Mueller, T. Schmiedel,
T. Q. Trinh, C. Weinrich, T. Wengefeld, et al. Roreas: robot coach for walking and ori-
entation training in clinical post-stroke rehabilitationprototype implementation and
evaluation in eld trials. Autonomous Robots, 41(3):679{698, 2017.
N. Guitar, A. MacDougall, D. Connelly, and E. Knight. Fitbit activity trackers interrupt
workplace sedentary behavior: A new application. Workplace health & safety, 66(5):
218{222, 2018.
P. C. Hallal, L. B. Andersen, F. C. Bull, R. Guthold, W. Haskell, U. Ekelund, L. P. A.
S. W. Group, et al. Global physical activity levels: surveillance progress, pitfalls, and
prospects. The lancet, 380(9838):247{257, 2012.
C. E. Han, S. Kim, S. Chen, Y.-H. Lai, J.-Y. Lee, R. Osu, C. J. Winstein, and
N. Schweighofer. Quantifying arm nonuse in individuals poststroke. Neurorehabil-
itation and Neural Repair, 27(5):439{447, 2013.
J. A. Harvey, S. F. Chastin, and D. A. Skelton. How sedentary are older people? a
systematic review of the amount of sedentary behavior. Journal of aging and physical
activity, 23(3):471{487, 2015.
105
J. Haw. Random-ratio schedules of reinforcement: The role of early wins and unrein-
forced trials. Journal of Gambling Issues, pages 56{67, 2008.
D. Hebesberger, C. Dondrup, T. Koertner, C. Gisinger, and J. Prip
. Lessons learned
from the deployment of a long-term autonomous robot as companion in physical
therapy for older adults with dementia: A mixed methods study. In The Eleventh
ACM/IEEE International Conference on Human Robot Interaction, pages 27{34.
IEEE Press, 2016.
M. Heerink, B. Kr ose, V. Evers, and B. Wielinga. Assessing acceptance of assistive
social agent technology by older adults: the almere model. International journal of
social robotics, 2(4):361{375, 2010.
E. B. Hekler, P. Klasnja, V. Traver, and M. Hendriks. Realizing eective behavioral
management of health: The metamorphosis of behavioral science methods. Pulse,
IEEE, 4(5):29{34, 2013.
K. Hellstrom, B. Lindmark, B. Wahlberg, and A. R. Fugl-Meyer. Self-ecacy in relation
to impairments and activities of daily living disability in elderly patients with stroke:
a prospective investigation. Journal of rehabilitation medicine, 35(5):202{207, 2003.
J. D. Henry, P. G. Rendell, L. H. Phillips, L. Dunlop, and M. Kliegel. Prospective
memory reminders: A laboratory investigation of initiation source and age eects.
The Quarterly Journal of Experimental Psychology, 65(7):1274{1287, 2012.
S. Hesse, H. Schmidt, C. Werner, and A. Bardeleben. Upper and lower extremity
robotic devices for rehabilitation and for studying motor control. Current opinion in
neurology, 16(6):705{710, 2003.
J. Hou, C. Miao, and Y. Liu. Prospective memory aid reminder system design for group
tasks. In Software, Knowledge, Information Management & Applications (SKIMA),
2016 10th International Conference on, pages 433{438. IEEE, 2016a.
J. Hou, Z. Zeng, C. Miao, and Y. Liu. Prospective memory aid: A reminding model
based on fuzzy cognitive maps. In Fuzzy Systems (FUZZ-IEEE), 2016 IEEE Inter-
national Conference on, pages 170{177. IEEE, 2016b.
I. Janssen and A. G. LeBlanc. Review systematic review of the health benets of physical
activity and tness in school-aged children and youth. Int. J. of Behavioral Nutrition
and Physical Activity, 7(40):1{16, 2010.
O. P. John, E. M. Donahue, and R. L. Kentle. The Big Five Inventory: Versions 4a
and 54. University of California, Berkeley, CA, Institute of Personality and Social
Research, 1991.
106
O. P. John, L. P. Naumann, and C. J. Soto. Paradigm shift to the integrative Big Five
Trait taxonomy. In Handbook of Personality: Theory and Research, pages 114{158.
2008.
M. Johnston, R. Carey, L. C. Bohlen, D. W. Johnston, A. Rothman, M. de Bruin,
M. P. Kelly, H. Groarke, and S. Michie. Linking behavior change techniques and
mechanisms of action: triangulation of ndings from literature synthesis and expert
consensus. 2018.
F. Jones and A. Riazi. Self-ecacy and self-management after stroke: a systematic
review. Disability and rehabilitation, 33(10):797{810, 2011.
G. Judah, B. Gardner, and R. Aunger. Forming a
ossing habit: An exploratory
study of the psychological determinants of habit formation. British journal of health
psychology, 18(2):338{353, 2013.
Y. Jung and K. Lee. Eects of physical embodiment on social presence of social robots.
In Proc. of Presence, pages 80{87, 2004a.
Y. Jung and K. M. Lee. Eects of physical embodiment on social presence of social
robots. In Proc. of the 7th International Workshop on Presence, pages 80{87, 2004b.
S.-H. Kang and J. Gratch. People like virtual counselors that highly-disclose about
themselves. The Annu. Review of CyberTherapy and Telemedicine, 167:143{148, 2011.
P. T. Katzmarzyk, T. S. Church, C. L. Craig, and C. Bouchard. Sitting time and
mortality from all causes, cardiovascular disease, and cancer. Medicine & Science in
Sports & Exercise, 41(5):998{1005, 2009.
C. D. Kidd and C. Breazeal. Robots at home: Understanding long-term human-robot
interaction. In Intelligent Robots and Systems, 2008. IROS 2008. IEEE/RSJ Inter-
national Conference on, pages 3230{3235. IEEE, 2008.
S. Kim, H. Park, C. E. Han, C. Winstein, and N. Schweighofer. Measuring habitual
arm use post-stroke with a bilateral time-constrained reaching task. Frontiers in
Neurology, 9:883, 2018.
A. C. King, T. W. Bickmore, M. I. Campero, L. A. Pruitt, and J. L. Yin. Employing
virtual advisors in preventive care for underserved communities: results from the
compass study. Journal of health communication, 18(12):1449{1464, 2013.
P. Klasnja, S. Smith, N. J. Seewald, A. Lee, K. Hall, B. Luers, E. B. Hekler, and
S. A. Murphy. Ecacy of contextually tailored suggestions for physical activity: a
micro-randomized optimization trial of heartsteps. Annals of Behavioral Medicine,
2018.
107
C. Korpershoek, J. van der Bijl, and T. B. Hafsteinsd ottir. Self-ecacy and its in
uence
on recovery of patients with stroke: a systematic review. Journal of Advanced Nursing,
67(9):1876{1894, 2011.
B.-J. Kwak, N.-O. Song, and L. E. Miller. Performance analysis of exponential backo.
IEEE/ACM Transactions on Networking (TON), 13(2):343{355, 2005.
S.-M. Lai, S. Studenski, P. W. Duncan, and S. Perera. Persisting consequences of stroke
measured by the stroke impact scale. Stroke, 33(7):1840{1844, 2002.
A. E. Latimer, K. A. M. Ginis, and K. P. Arbour. The ecacy of an implementation
intention intervention for promoting physical activity among individuals with spinal
cord injury: a randomized controlled trial. Rehabilitation Psychology, 51(4):273, 2006.
I.-M. Lee, E. J. Shiroma, F. Lobelo, P. Puska, S. N. Blair, P. T. Katzmarzyk, L. P. A.
S. W. Group, et al. Eect of physical inactivity on major non-communicable diseases
worldwide: an analysis of burden of disease and life expectancy. The lancet, 380
(9838):219{229, 2012.
R. Lewthwaite and G. Wulf. Social-comparative feedback aects motor skill learning.
The Quarterly Journal of Experimental Psychology, 63(4):738{749, 2010.
C. Lisetti, R. Amini, U. Yasavur, and N. Rishe. I can help you change! an empathic
virtual agent delivers behavior change health interventions. ACM Transactions on
Management Information Systems (TMIS), 4(4):19, 2013.
C. Lisetti, R. Amini, and U. Yasavur. Now all together: overview of virtual health as-
sistants emulating face-to-face health interview experience. KI-K unstliche Intelligenz,
29(2):161{172, 2015.
M. Lohse. Bridging the gap between users' expectations and system evaluations. In
Proceedings of the 20th IEEE Int. Symp. on Robot and Human Interactive Commu-
nication (Ro-Man), pages 485{490, Atlanta, Georgia, 2011.
M. Lombard, T. B. Ditton, D. Crane, B. Davis, G. Gil-Egui, K. Horvath, J. Rossman,
and S. Park. Measuring presence: A literature-based approach to the development
of a standardized paper-and-pencil instrument. In 3rd Int. Workshop on Presence,
volume 240, Delft, The Netherlands, 2000a.
M. Lombard, T. B. Ditton, D. Crane, B. Davis, G. Gil-Egui, K. Horvath, J. Rossman,
and S. Park. Measuring presence: A literature-based approach to the development
of a standardized paper-and-pencil instrument. In Proc. of the 3rd International
Workshop on Presence, 2000b.
A. Lot, C. Langensiepen, and S. W. Yahaya. Socially assistive robotics: Robot exercise
trainer for older adults. Technologies, 6(1):32, 2018.
108
R. Lowe, F. Eves, and D. Carroll. The in
uence of aective and instrumental beliefs on
exercise intentions and behavior: A longitudinal analysis. Journal of Applied Social
Psychology, 32(6):1241{1252, 2002.
D. R. Lubans, C. Foster, and S. J. Biddle. A review of mediators of behavior in in-
terventions to promote physical activity among children and adolescents. Preventive
Medicine, 47(5):463{470, 2008.
M. Matari c, A. Tapus, C. Winstein, and J. Eriksson. Socially assistive robotics for
stroke and mild tbi rehabilitation. Advanced Technologies in Rehabilitation, 145:249{
262, 2009.
M. J. Matari c. Socially assistive robotics: Human augmentation versus automation.
Science Robotics, 2(4):eaam5410, 2017.
M. J. Matari c and B. Scassellati. Socially Assistive Robotics, pages 1973{1994. Springer
International Publishing, 2016.
M. J. Matari c, J. Eriksson, D. J. Feil-Seifer, and C. J. Winstein. Socially assistive
robotics for post-stroke rehabilitation. Journal of NeuroEngineering and Rehabilita-
tion, 4(1):5, 2007.
C. C. Maturo and S. A. Cunningham. In
uence of friends on children's physical activity:
a review. American J. of Public Health, 103(7):e23{e38, 2013.
J. McCroskey and T. McCain. The Measurement of Interpersonal Attraction. Speech
Monographs, 41(1961):261{266, 1974a.
J. C. McCroskey and T. A. McCain. The measurement of interpersonal attraction.
Speech Monographs, 41(3):261{266, 1974b.
S. Michie, S. Ashford, F. F. Sniehotta, S. U. Dombrowski, A. Bishop, and D. P. French. A
rened taxonomy of behaviour change techniques to help people change their physical
activity and healthy eating behaviours: the calo-re taxonomy. Psychology & health,
26(11):1479{1498, 2011.
S. Michie, M. Richardson, M. Johnston, C. Abraham, J. Francis, W. Hardeman, M. P.
Eccles, J. Cane, and C. E. Wood. The behavior change technique taxonomy (v1)
of 93 hierarchically clustered techniques: building an international consensus for the
reporting of behavior change interventions. Annals of behavioral medicine, 46(1):
81{95, 2013.
C. Midden and J. Ham. Using negative and positive social feedback from a robotic agent
to save energy. In Proceedings of the 4th International Conference on Persuasive
Technology, page 12. ACM, 2009.
109
B. Minor, J. R. Doppa, and D. J. Cook. Data-driven activity prediction: Algorithms,
evaluation methodology, and applications. In Proceedings of the 21th ACM SIGKDD
International Conference on Knowledge Discovery and Data Mining, pages 805{814.
ACM, 2015.
C. J. Mitchell, P. F. Lovibond, E. Minard, and Y. Lavis. Forward blocking in human
learning sometimes re
ects the failure to encode a cue{outcome relationship. The
Quarterly Journal of Experimental Psychology, 59(05):830{844, 2006.
B. Morris, R. Lawton, R. McEachan, R. Hurling, and M. Conner. Changing self-reported
physical activity using dierent types of aectively and cognitively framed health
messages, in a student population. Psychology, health & medicine, pages 1{10, 2015.
R. W. Motl, R. K. Dishman, R. Saunders, M. Dowda, G. Felton, and R. R. Pate. Mea-
suring enjoyment of physical activity in adolescent girls. American J. of Preventive
Medicine, 21(2):110{117, 2001.
C. Mucchiani, S. Sharma, M. Johnson, J. Sefcik, N. Vivio, J. Huang, P. Cacchione,
M. Johnson, R. Rai, A. Canoso, et al. Evaluating older adults' interaction with a
mobile assistive robot. In Intelligent Robots and Systems (IROS), 2017 IEEE/RSJ
International Conference on, pages 840{847. IEEE, 2017.
S. P. Mullen, E. A. Olson, S. M. Phillips, A. N. Szabo, T. R. W ojcicki, E. L. Mailey,
N. P. Gothe, J. T. Fanning, A. F. Kramer, and E. McAuley. Measuring enjoyment of
physical activity in older adults: invariance of the physical activity enjoyment scale
(paces) across groups and time. International Journal of Behavioral Nutrition and
Physical Activity, 8(1):103, 2011.
J. Mumm and B. Mutlu. Designing motivational agents: the role of praise, social
comparison, and embodiment in computer feedback. Computers in Human Behavior,
27(5):1643{1650, 2011.
I. Nahum-Shani, S. N. Smith, B. J. Spring, L. M. Collins, K. Witkiewitz, A. Tewari,
and S. A. Murphy. Just-in-time adaptive interventions (jitais) in mobile health: key
components and design principles for ongoing health behavior support. Annals of
Behavioral Medicine, 52(6):446{462, 2017.
M. A. Neroni, N. Gamboz, and M. A. Brandimonte. Does episodic future thinking
improve prospective remembering? Consciousness and cognition, 23:53{62, 2014.
T. Nomura, T. Suzuki, T. Kanda, and K. Kato. Measurement of negative attitudes
toward robots. Interaction Studies, 7(3):437{454, 2006.
N. Norouzi-Gheidari, P. S. Archambault, and J. Fung. Eects of robot-assisted therapy
on stroke rehabilitation in upper limbs: systematic review and meta-analysis of the
literature. Journal of Rehabilitation Research & Development, 49(4), 2012.
110
S. Northcott, B. Moss, K. Harrison, and K. Hilari. A systematic review of the impact
of stroke on social support and social networks: associated factors and patterns of
change. Clinical Rehabilitation, 30(8):811{831, 2016.
C. L. Ogden, M. D. Carroll, L. R. Curtin, M. M. Lamb, and K. M. Flegal. Prevalence
of high body mass index in us children and adolescents, 2007-2008. Jama, 303(3):
242{249, 2010.
S. Okumura, M. Kimoto, M. Shiomi, T. Iio, K. Shimohara, and N. Hagita. Do social
rewards from robots enhance oine improvements in motor skills? In International
Conference on Social Robotics, pages 32{41. Springer, 2017.
S. Orbell and B. Verplanken. The automatic component of habit in health behavior:
habit as cue-contingent automaticity. Health Psychology, 29(4):374, 2010.
N. Owen, G. N. Healy, C. E. Matthews, and D. W. Dunstan. Too much sitting: the
population-health science of sedentary behavior. Exercise and sport sciences reviews,
38(3):105, 2010.
N. Owen, T. Sugiyama, E. E. Eakin, P. A. Gardiner, M. S. Tremblay, and J. F. Sallis.
Adults' sedentary behavior: determinants and interventions. American journal of
preventive medicine, 41(2):189{196, 2011.
O. D. P erez, M. R. Aitken, P. Zhukovsky, F. A. Soto, G. P. Urcelay, and A. Dickinson.
Human instrumental performance in ratio and interval contingencies: A challenge for
associative theory. The Quarterly Journal of Experimental Psychology, pages 1{13,
2016.
S. S. Phatak, M. T. Freigoun, C. A. Mart n, D. E. Rivera, E. V. Korinek, M. A. Adams,
M. P. Buman, P. Klasnja, and E. B. Hekler. Modeling individual dierences: A case
study of the application of system identication for personalizing a physical activity
intervention. Journal of biomedical informatics, 79:82{97, 2018.
A. Prestwich, M. Perugini, and R. Hurling. Can implementation intentions and text
messages promote brisk walking? a randomized trial. Health Psychology, 29(1):40,
2010.
D. Ramgoolam, E. Russell, and A. B. Williams. Towards a social and mobile humanoid
exercise coach. In Proc. of the 9th ACM/IEEE Int. Conf. on Human-Robot Interaction
(HRI), pages 274{275, 2014.
D. Rand and J. J. Eng. Predicting daily use of the aected upper extremity 1 year after
stroke. Journal of Stroke and Cerebrovascular Diseases, 24(2):274{283, 2015.
111
S. Reeder, B. Kechavarzi, Z. Zimmerman, et al. Designing a robot through prototyp-
ing in the wild. In Proceedings of the 6th ACM/IEEE Int. Conf. on Human-Robot
Interaction (HRI), pages 239{240. ACM, 2011.
W. J. Rejeski, L. R. Brawley, and S. A. Shumaker. Physical activity and health-related
quality of life. Exercise and sport sciences reviews, 24(1):71{108, 1996.
R. E. Rhodes and L. A. Pfaei. Mediators of physical activity behaviour change among
adult non-clinical populations: a review update. International Journal of Behavioral
Nutrition and Physical Activity, 7(1):37, 2010.
P. Rukavina and W. Li. School physical activity interventions: do not forget about
obesity bias. Obesity Reviews, 9(1):67{75, 2008.
R. M. Ryan and J. P. Connell. Perceived locus of causality and internalization: exam-
ining reasons for acting in two domains. J. of Personality and Social Psychology, 57
(5):749, 1989.
R. M. Ryan, H. Patrick, E. L. Deci, and G. C. Williams. Facilitating health behaviour
change and its maintenance: Interventions based on self-determination theory. Euro-
pean Health Psychologist, 10(1):2{5, 2008.
J. F. Sallis, J. J. Prochaska, W. C. Taylor, et al. A review of correlates of physical
activity of children and adolescents. Medicine and Sci. in Sports and Exercise, 32(5):
963{975, 2000.
J. Sanford, J. Moreland, L. R. Swanson, P. W. Stratford, and C. Gowland. Reliability of
the fugl-meyer assessment for testing motor performance in patients following stroke.
Physical Therapy, 73(7):447{454, 1993.
D. A. Santos, A. M. Silva, F. Baptista, R. Santos, S. Vale, J. Mota, and L. B. Sardinha.
Sedentary behavior and physical activity are independently related to functional t-
ness in older adults. Experimental gerontology, 47(12):908{912, 2012.
B. Scassellati, L. Boccanfuso, C.-M. Huang, M. Mademtzi, M. Qin, N. Salomons, P. Ven-
tola, and F. Shic. Improving social skills in children with asd using a long-term,
in-home social robot. Science Robotics, 3(21):eaat7544, 2018.
S. M. Schembre, Y. Liao, M. C. Robertson, G. F. Dunton, J. Kerr, M. E. Haey,
T. Burnett, K. Basen-Engquist, and R. S. Hicklen. Just-in-time feedback in diet and
physical activity interventions: Systematic review and practical design framework.
Journal of medical Internet research, 20(3), 2018.
S. Schneider and F. Kummert. Comparing the eects of social robots and virtual agents
on exercising motivation. In International Conference on Social Robotics, pages 451{
461. Springer, 2018.
112
M. Shaughnessy, B. M. Resnick, and R. F. Macko. Testing a model of post-stroke
exercise behavior. Rehabilitation nursing, 31(1):15{21, 2006.
S. Shim, H. Kim, and J. Jung. Comparison of upper extremity motor recovery of stroke
patients with actual physical activity in their daily lives measured with accelerome-
ters. Journal of physical therapy science, 26(7):1009{1011, 2014.
E. Short, K. Swift-Spong, J. Greczek, A. Ramachandran, A. Litoiu, E. C. Grigore,
D. Feil-Seifer, S. Shuster, J. J. Lee, S. Huang, et al. How to train your dragonbot:
Socially assistive robots for teaching children about nutrition through play. In Proc.
of the 23rd IEEE Int. Symp. on Robot and Human Interactive Communication (Ro-
Man), pages 924{929, 2014.
E. Short, D. Short, Y. Fu, and M. J. Matari c. Sprite: Stewart platform robot for
interactive tabletop engagement. Department of Computer Science, University of
Southern California. Tech Report, 2017a.
E. S. Short, K. Swift-Spong, H. Shim, K. M. Wisniewski, D. K. Zak, S. Wu, E. Zelinski,
and M. J. Matari c. Understanding social interactions with socially assistive robotics
in intergenerational family groups. In Robot and Human Interactive Communication
(RO-MAN), 2017 26th IEEE International Symposium on, pages 236{241. IEEE,
2017b.
C. L. Sidner, T. Bickmore, B. Nooraie, C. Rich, L. Ring, M. Shayganfar, and L. Var-
doulakis. Creating new technologies for companionable agents to support isolated
older adults. ACM Transactions on Interactive Intelligent Systems (TiiS), 8(3):17,
2018.
R. Simmons, M. Makatchev, R. Kirby, M. K. Lee, I. Fanaswala, B. Browning, J. Forlizzi,
and M. Sakr. Believable robot characters. AI Magazine, 32(4):39{52, 2011.
R. Sirriyeh, R. Lawton, and J. Ward. Physical activity and adolescents: an exploratory
randomized controlled trial investigating the in
uence of aective and instrumental
text messages. British J. of Health Psychology, 15(4):825{840, 2010.
B. F. Skinner. Science and human behavior. Simon and Schuster, 1965.
D. Spruijt-Metz, E. Hekler, N. Saranummi, S. Intille, I. Korhonen, W. Nilsen, D. E.
Rivera, B. Spring, S. Michie, D. A. Asch, et al. Building new computational models
to support health behavior change and maintenance: new opportunities in behavioral
research. Translational behavioral medicine, 5(3):335{346, 2015.
K. Stawarz and A. L. Cox. How technology supporting daily habits could help women re-
member oral contraception. In Proceedings of the 27th International British Computer
Society Human Computer Interaction Conference Workshop on Habits in Human-
Computer Interaction, 2013.
113
K. Stawarz, A. L. Cox, and A. Blandford. Don't forget your pill!: Designing eective
medication reminder apps that support users' daily routines. In Proceedings of the
SIGCHI Conference on Human Factors in Computing Systems, pages 2269{2278.
ACM, 2014.
K. Stawarz, A. L. Cox, and A. Blandford. Beyond self-tracking and reminders: designing
smartphone apps that support habit formation. In Proceedings of the 33rd Annual
ACM Conference on Human Factors in Computing Systems, pages 2653{2662. ACM,
2015.
S. K. Stephens, E. A. Winkler, S. G. Trost, D. W. Dunstan, E. G. Eakin, S. F. Chastin,
and G. N. Healy. Intervening to reduce workplace sitting time: how and when do
changes to sitting time occur? Br J Sports Med, 48(13):1037{1042, 2014.
K. Strohacker, O. Galarraga, and D. M. Williams. The impact of incentives on exercise
behavior: a systematic review of randomized controlled trials. Annals of Behavioral
Medicine, 48(1):92{99, 2014.
K. Swift-Spong, E. Short, E. Wade, and M. J. Matari c. Eects of comparative feedback
from a socially assistive robot on self-ecacy in post-stroke rehabilitation. In Proc. of
the 14th IEEE Int. Conf. on Rehabilitation Robotics (ICORR), pages 764{769, 2015.
K. Swift-Spong, C. K. F. Wen, D. Spruijt-Metz, and M. J. Matari c. Comparing back-
stories of a socially assistive robot exercise buddy for adolescent youth. In To appear
in the Proceedings of the 25th IEEE International Symposium on Robot and Human
Interactive Communication (Ro-Man). ACM, 2016.
S. A. Tabatabaei, M. Hoogendoorn, and A. van Halteren. Narrowing reinforcement
learning: Overcoming the cold start problem for personalized health interventions. In
International Conference on Principles and Practice of Multi-Agent Systems, pages
312{327. Springer, 2018.
A. Tapus, C. T apu s, and M. J. Matari c. Userrobot personality matching and assistive
robot behavior adaptation for post-stroke rehabilitation therapy. Intelligent Service
Robotics, 1(2):169, 2008a.
A. Tapus, C. T apu s, and M. J. Matari c. Userrobot personality matching and assistive
robot behavior adaptation for post-stroke rehabilitation therapy. Intelligent Service
Robotics, 1(2):169{183, 2008b.
A. Tapus, C. Tapus, and M. J. Mataric. The use of socially assistive robots in the design
of intelligent cognitive therapies for people with dementia. In Rehabilitation Robotics,
2009. ICORR 2009. IEEE International Conference on, pages 924{929. IEEE, 2009.
E. Taub, G. Uswatte, V. Mark, and D. Morris. The learl1ed nonuse phenomenon:
implications for rehabilitation. Eura Medicophys, 42:241{55, 2006.
114
G. Terrett, N. S. Rose, J. D. Henry, P. E. Bailey, M. Altgassen, L. H. Phillips, M. Kliegel,
and P. G. Rendell. The relationship between prospective memory and episodic future
thinking in younger and older adulthood. The Quarterly Journal of Experimental
Psychology, pages 1{14, 2015.
J. G. Thomas and D. S. Bond. Behavioral response to a just-in-time adaptive inter-
vention (jitai) to reduce sedentary behavior in obese adults: Implications for jitai
optimization. Health Psychology, 34(S):1261, 2015.
R. P. Troiano, D. Berrigan, K. W. Dodd, L. C. Masse, T. Tilert, M. McDowell, et al.
Physical activity in the united states measured by accelerometer. Medicine and Sci.
in Sports and Exercise, 40(1):181, 2008.
H. Vathsangam and G. S. Sukhatme. Using phone-based activity monitors to promote
physical activity in older adults: a pilot study. In Healthcare Innovation Conference
(HIC), 2014 IEEE, pages 42{47. IEEE, 2014.
J. M. Veerbeek, A. C. Langbroek-Amersfoort, E. E. Van Wegen, C. G. Meskers, and
G. Kwakkel. Eects of robot-assisted therapy for the upper limb after stroke: a
systematic review and meta-analysis. Neurorehabilitation and neural repair, 31(2):
107{121, 2017.
K. J. Waddell, M. J. Strube, R. R. Bailey, J. W. Klaesner, R. L. Birkenmeier, A. W.
Dromerick, and C. E. Lang. Does task-specic training improve upper limb perfor-
mance in daily life poststroke? Neurorehabilitation and neural repair, 31(3):290{300,
2017.
E. Wade, A. Parnandi, R. Mead, and M. Matari c. Socially assistive robotics for guiding
motor task practice. Paladyn, 2(4):218{227, 2011.
M. Watabe-Uchida, N. Eshel, and N. Uchida. Neural circuitry of reward prediction
error. Annual review of neuroscience, 40:373{394, 2017.
L. M. Weber and J. Stein. The use of robots in stroke rehabilitation: A narrative review.
NeuroRehabilitation, 43(1):99{110, 2018.
W. X. Wei, K. N. Fong, R. C. Chung, H. K. Cheung, and E. S. Chow. remind-to-move
for promoting upper extremity recovery using wearable devices in subacute stroke:
a multi-center randomized controlled study. IEEE Transactions on Neural Systems
and Rehabilitation Engineering, 2018.
W. A. Welch, B. Spring, S. M. Phillips, and J. Siddique. Moderating eects of weather-
related factors on a physical activity intervention. American journal of preventive
medicine, 54(5):e83{e89, 2018.
115
G. C. Williams, H. A. McGregor, D. Sharp, C. Levesque, R. W. Kouides, R. M. Ryan,
and E. L. Deci. Testing a self-determination theory intervention for motivating to-
bacco cessation: supporting autonomy and competence in a clinical trial. Health
Psychology, 25(1):91, 2006.
K. Winkle, P. Caleb-Solly, A. Turton, and P. Bremner. Social robots for engagement in
rehabilitative therapies: Design implications from a study with therapists. In Proceed-
ings of the 2018 ACM/IEEE International Conference on Human-Robot Interaction,
pages 289{297. ACM, 2018.
C. Winstein and R. Varghese. Been there, done that, so what's next for arm and hand
rehabilitation in stroke? NeuroRehabilitation, (Preprint):1{15, 2018.
S. L. Wolf, C. J. Winstein, J. P. Miller, E. Taub, G. Uswatte, D. Morris, C. Giuliani,
K. E. Light, D. Nichols-Larsen, E. Investigators, et al. Eect of constraint-induced
movement therapy on upper extremity function 3 to 9 months after stroke: the excite
randomized clinical trial. Jama, 296(17):2095{2104, 2006.
W. Wood and D. T. Neal. A new look at habits and the habit-goal interface. Psycho-
logical review, 114(4):843, 2007.
W. Wood and D. R unger. Psychology of habit. Annual Review of Psychology, 67:
289{314, 2016.
G. Wulf, S. Chiviacowsky, and R. Lewthwaite. Altering mindset can enhance motor
learning in older adults. Psychology and Aging, 27(1):14, 2012.
D. R. Young, M.-F. Hivert, S. Alhassan, S. M. Camhi, J. F. Ferguson, P. T. Katzmarzyk,
C. E. Lewis, N. Owen, C. K. Perry, et al. Sedentary behavior and cardiovascular
morbidity and mortality: a science advisory from the american heart association.
Circulation, 134(13):e262{e279, 2016.
J. B. Zogg, S. P. Woods, J. A. Sauceda, J. S. Wiebe, and J. M. Simoni. The role of
prospective memory in medication adherence: a review of an emerging literature.
Journal of Behavioral Medicine, 35(1):47{62, 2012.
116
Abstract (if available)
Abstract
Socially assistive robot (SAR) systems have the potential to support the complex process of human behavior change by providing social support such as feedback and encouragement at opportune times. This dissertation presents a framework for SAR behavior change support in the context of physical activity behavior. This framework is designed around the goal of creating lasting behavior change that extends past the SAR interaction. Within this framework, the robot is equipped with one or more SAR physical activity behavior change support methods designed to affect a specific mechanism of behavior change. ❧ This dissertation develops the design of SAR feedback, backstory, and messaging support methods for physical activity behavior change. These three methods were each designed to support a different mechanism of achieving behavior change by leveraging the robot's relational and support capabilities. Feedback was designed to support a user's beliefs about their ability to perform a physical activity task. Robot backstory was designed to increase the robot's ability to provide social support, and messaging was designed to increase the user's positive feelings towards the physical activity. These three support methods are evaluated in real-world physical activity domains with a fully autonomous SAR systems. The feedback support method is evaluated in the domain of post-stroke rehabilitation, and the backstory and messaging support methods are evaluated in the domain of adolescent exercise. ❧ Reminder and social reward decision making is also developed as a SAR physical activity behavior change support method using a model of SAR habit formation support. This model formalizes the SAR sequential decision making task of determining when to give reminders and social rewards towards the goal of supporting the formation of a new desired habit. Habits are formed when the occurrence of a cue is followed by a desired behavior, and that combination is reinforced repeatedly over time. The model of habit formation support enables a robot to intervene in this process. This model is evaluated in the domain of reducing older adult sedentary behavior through a two-week in-home SAR intervention. The robot was able to generate a high level of reminder adherence in this setting. ❧ In this work, four SAR physical activity behavior change support methods were developed and evaluated in three different physical activity domains with fully autonomous SAR systems. This dissertation contributes to understanding the methods a robot could use to support behavior change in a variety of physical activity domains both in situ within the context of the behavior in everyday life and outside of that context.
Linked assets
University of Southern California Dissertations and Theses
Conceptually similar
PDF
Physical and social adaptation for assistive robot interactions
PDF
Multiparty human-robot interaction: methods for facilitating social support
PDF
Managing multi-party social dynamics for socially assistive robotics
PDF
On virtual, augmented, and mixed reality for socially assistive robotics
PDF
Nonverbal communication for non-humanoid robots
PDF
Coordinating social communication in human-robot task collaborations
PDF
Socially assistive and service robotics for older adults: methodologies for motivating exercise and following spatial language instructions in discourse
PDF
Modeling dyadic synchrony with heterogeneous data: validation in infant-mother and infant-robot interactions
PDF
Situated proxemics and multimodal communication: space, speech, and gesture in human-robot interaction
PDF
Efficiently learning human preferences for proactive robot assistance in assembly tasks
PDF
Quality diversity scenario generation for human robot interaction
PDF
Decision support systems for adaptive experimental design of autonomous, off-road ground vehicles
PDF
AI-driven experimental design for learning of process parameter models for robotic processing applications
PDF
Robust loop closures for multi-robot SLAM in unstructured environments
PDF
Motion coordination for large multi-robot teams in obstacle-rich environments
PDF
Socially-informed content analysis of online human behavior
PDF
Modeling and regulating human interaction with control affine dynamical systems
PDF
Parasocial consensus sampling: modeling human nonverbal behaviors from multiple perspectives
PDF
Identifying Social Roles in Online Contentious Discussions
PDF
Contingency handling in mission planning for multi-robot teams
Asset Metadata
Creator
Swift-Spong, Katelyn Elizabeth
(author)
Core Title
Towards socially assistive robot support methods for physical activity behavior change
School
Viterbi School of Engineering
Degree
Doctor of Philosophy
Degree Program
Computer Science
Publication Date
04/28/2019
Defense Date
02/19/2019
Publisher
University of Southern California
(original),
University of Southern California. Libraries
(digital)
Tag
behavior change,human-robot interaction,OAI-PMH Harvest,robotics,socially assistive robotics
Format
application/pdf
(imt)
Language
English
Contributor
Electronically uploaded by the author
(provenance)
Advisor
Mataric, Maja J. (
committee chair
), Nikolaidis, Stefanos (
committee member
), Zelinski, Elizabeth (
committee member
)
Creator Email
kate.swift.spong@gmail.com,swiftspo@usc.edu
Permanent Link (DOI)
https://doi.org/10.25549/usctheses-c89-153735
Unique identifier
UC11660715
Identifier
etd-SwiftSpong-7312.pdf (filename),usctheses-c89-153735 (legacy record id)
Legacy Identifier
etd-SwiftSpong-7312.pdf
Dmrecord
153735
Document Type
Dissertation
Format
application/pdf (imt)
Rights
Swift-Spong, Katelyn Elizabeth
Type
texts
Source
University of Southern California
(contributing entity),
University of Southern California Dissertations and Theses
(collection)
Access Conditions
The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...
Repository Name
University of Southern California Digital Library
Repository Location
USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA
Tags
behavior change
human-robot interaction
robotics
socially assistive robotics