University of Southern California Dissertations and Theses

Design of adaptive automated robotic task presentation system for stroke rehabilitation

(USC Thesis Other)

Design of adaptive automated robotic task presentation system for stroke rehabilitation

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content

DESIGN OF ADAPTIVE AUTOMATED ROBOTIC
TASK PRESENTATION SYSTEM FOR STROKE REHABILITATION

by

Younggeun Choi

________________________________________________________________________

A Dissertation Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(COMPUTER SCIENCE)

August 2010

Copyright 2010 Younggeun Choi

ii
Acknowledgements

First of all, I would like to thank my parents for their patience, endless support,
and love throughout my Phd life. My most gratitude should be also given to the
wonderful people at the University of Southern California. This thesis and my life in
academia would never been possible without motivational input, feedback with
various error signals, guidance with active intervention, and patience with high
stability from Dr. Nicolas Scweighofer. He is a great advisor, a heartful friend and I
am honored to be his first student. Next, I am grateful to Dr. James Gordon, not only
for serving on my dissertation committee, but also for all the advice, direction and
support he provided in almost every phase of my research. Special thanks are due to
Dr. Stefan Schaal for his serving on my dissertation committee, and he provided
excellent courses, from which I obtained many ideas for my research. I also would
like to thank Dr. Michael Arbib for his serving on my dissertation committee, Dr.
Carolee Winstein for making valuable suggestions, Matthew Sandusky, Chad Louie
for their technical assistance, and Huiting Goh, Shuya Chen, Jihye Lee for their help
in recruiting subjects and running experiment. Another thanks goes to all my lab
members, Cheol Han, Feng Qi, Jeongyoon Lee, Yukikazu Hidaka, Sungshin Kim,
Hyeshin Park, and Yupeng Xiao for sharing wonderful life in the basement of CHP
building, having exciting discussion, stimulating quasistatic ‘research neurons’ for
faster graduation, trading our own bodies for pilot experiment subjects, and enjoying
unforgettable parties that turned out to be neurophysiologically very forgettable. I
would like to extend my best wishes to all of them. I also have to greatly appreciate
my sister, Youngjin Choi, for the many meals she cooked and left for me, and my

iii
lovely niece, Jiwoo Seo for being so cute.
This work was supported in part by the Division of Biokinesiology and Physical
Therapy, University of Southern California, and in part by the National Institutes of
Health under Grant R03 HD050591-02.

iv
Table of Contents

Acknowledgements

List of Tables

List of Figures

Abstract

Chapter 1. Introduction
1.1. Motivation
1.2. Principles for stroke rehabilitation
1.3. Existing robotic systems

Chapter 2. Design of ADAPT
2.1. Introduction
2.2. Intelligent control architecture of ADAPT
2.3. Robot and devices
2.4. Safety

Chapter 3. Adaptive Task Schedule:
Performance-based adaptive schedule enhance motor learning
3.1. Introduction
3.2. Methods
3.2.1. Subjects and design
3.2.2. Experimental procedures
3.3. Results
3.4. Conclusion and Discussion

Chapter 4. Functional Task Model:
Capturing the dynamics of functional tasks
4.1. Introduction
4.2. Modelling with RFWR
4.2.1. RFWR Algorithm
4.2.2. Karnopp Friction Model with RFWR
4.3. Simulation
4.4. Experimental setup
4.5. Trajectory generation from the functional task model
4.6. Results
4.7. Conclusion

Chapter 5. Functional Task Display:
Model Based Lower-level Admittance Controller for Haptic Display
ii

vi

vii

ix

1
1
3
7

12
12
14
15
21

23

23
26
26
27
32
38

42

42
45
45
47
50
54
54
55
59

61

v
5.1. Introduction
5.2. The architecture of the low-level admittance controller
5.3. The performance of the low-level admittance controller

Chapter 6. The Feasibility of ADAPT:
Feasibility tests of ADAPT
to a healthy subject and patients with chronic stroke
6.1. Introduction
6.2. The experiment with a healith subject
6.2.1. Method
6.2.2. Results
6.3. The experiment with participants with chronic stroke
6.3.1. Method
6.3.1.1. Participants
6.3.1.2. Rehabilitation robot
6.3.1.3. Functional tasks
6.3.1.4. Adaptive task scheduler
6.3.1.5. Experimental procedure
6.3.2. Results
6.3.2.1. Functional task trajectory
6.3.2.2. Performance adaptation
6.3.2.3. Movement time improvement
6.3.2.4. Acceptance of ADAPT
6.3.3. Discussion
6.3.4. Conclusion

Chapter 7. Conclusion and Future work
7.1. Contributions
7.1.1. Design of ADAPT for stroke rehabilitation
7.1.2. A novel modeling method for adaptive difficulty of
functional task
7.1.3. A novel admittance controller with a force model and a
motion model
7.1.4. Performance based adaptive schedule for motor learning
7.1.5. A feasibility test of ADAPT with participants with
chronic stroke
7.2. Future work
7.2.1. Improvement of ADAPT design
7.2.2. Modeling the task dynamics and haptic controller design
7.2.3. Model-based adaptive schedules, and Bayesian
statistical models of performance
7.2.4. Randomized Clinical Trials

Bibliography

61
63
65

69

69
69
70
71
74
74
74
75
76
76
77
82
82
85
87
89
90
92

94
94
94
95

96

96
97

97
98
99
100

102

104

vi
List of Tables

Table 6.1. Functional task list.

Table 6.2. Subscale findings of the IMI questionnaire administered after
training

Table 6.3. Patients’ acceptance for the training session with ADAPT

80

81

81

vii
List of Figures

Figure 1.1. The MIT-MANUS trains elbow and shoulder flexion and extension
and shoulder abduction and adduction.

Figure 1.2. The Stanford MIME robotic device assists the movement of the
affected limb

Figure 1.3. The Northwestern ARM-Guide makes and assists the reaching
movements in one direction at a time.

Figure 1.4. The AutoCITE presents 8 functional tasks in semi-automated way.

Figure 1.5. ADLER presents functional tasks with multiple artifacts, which are
interachageable.

Figure 2.1. The principles of stroke rehabilitation and implementation them in
ADAPT.

Figure 2.2. The conceptual design of the ADAPT system

Figure 2.3. Robot and Devices for ADAPT

Figure 2.4. Different functional task tools in use.

Figure 2.5. Tool-changing process of ADAPT.

Figure 3.1 Average performance and standard error on the four tasks on the
delayed retention test on day 4 for each condition.

Figure 3.2. Mean performance in the five trials of the delayed retention test
given on day 4 for each task and each schedule.

Figure 3.3. Examples of adaptive number of trials schedule in the AdapTr
condition for the four tasks

Figure 3.4. Example of adaptive difficulty schedule for the four tasks.

Figure 4.1. The modified Karnopp model.

Figure 4.2. (A) Simulated training data, and possible local models for the
training data. (B) The model learned from the simulated data by RFWR

Figure 4.3. Experimental setup for capturing dynamics.
7

8

9

10

11

13

15

18

19

20

33

34

35

36

50

53

54

viii

Figure 4.4. Doorknob dynamics model.

Figure 4.5. Jar dynamics model

Figure 5.1. Low-level admittance controller for ADAPT

Figure 5.2. Dynamics models of the doorknob and trajectories simulated by
ADAPT for three different difficulty levels.

Figure 5.3 Dynamics models of the jar and trajectories simulated by ADAPT
for three different difficulty levels

Figure 6.1. Illustration of the function of the high-level adaptive task scheduler

Figure 6.2. Experiment procedure with ADAPT.

Figure 6.3. Sample trajectories of functional tasks of participants with chronic
stroke

Figure 6.4. Progress of Performance & Difficulty of participants with chronic
stroke

Figure 6.5. Movement times in pre and post test sessions

57

58

63

67

68

73

79

84

86

88

ix
Abstract
Robotic technology has the potential to deliver therapy activities for rehabilitation of
arm and hand function after stroke more efficiently and effectively than conventional
rehabilitation, as it can objectively dose the prescribed intensive amount of therapy in
automated design with less cost and effort, and can provide highly reliable measurement
of patients’ progress. The primary goal of this dissertation is to develop a robotic
rehabilitation system that fulfills current guidelines for the stroke rehabilitation: motor
training focus on realistic tasks that require reaching and manipulation and engage
patients post stroke intensively, actively, and adaptively.
Firstly, we presented a novel robotic task-practice system, i.e., adaptive and
automatic presentation of tasks (ADAPT), which was designed according to the
guidelines. A modular and reconfigurable robot with the configuration of a 3 degree-of-
freedom (DOF) wrist mounted on a 1-DOF linear actuator simulates the dynamics of
functional tasks and presents the functional tasks to patients post stroke. A novel tool-
changing system enables ADAPT to automatically switch between the tools
corresponding to the functional tasks. The control architecture of ADAPT is composed of
three main components: a high-level task scheduler, a functional task model, and a low-
level admittance controller. The high-level task scheduler adaptively selects the task to
practice and sets the task difficulty based on the previous performance of the patients.
The functional task model generates desired trajectories based on learned models of task
dynamics. Tasks dynamics are modeled with receptive field weighted regression (RFWR),
such that the feel of the task tools is accurately modeled, and the task difficulty can be

x
easily adjusted. The low-level admittance controller, which is also learned with RFWR,
implements the selected task trajectory for robot–patient interaction.
Secondly, we introduced new adaptive schedules for the high level adaptive task
scheduler of ADAPT in an attempt to maximize the relearning of multiple functional
tasks and balancing learning among tasks in a limited training time. Although random
scheduling of several tasks has been shown to enhance learning more than blocked
scheduling does, the advantages of random scheduling may be limited because it does not
take into account the nominal difficulty of each task, the difference in difficulty between
tasks, and the skill level of the learner in that type of schedule. We proposed two new
algorithms for adaptively determining the nominal difficulty and the number of trials for
each task on the basis of both current and delayed performance of the learner (N = 48).
We tested the adaptive algorithms in a 2 × 2 factorial design, and they show that the
algorithms outperform random scheduling when performance is measured on a delayed
retention test.
Finally, we investigated the feasibility of ADAPT to patients post stroke by
evaluating safety, system utility, fidelity of simulated tasks and patient acceptance. Five
patients with chronic stroke participated in approximately one hour training session with
an adaptive difficulty schedule. Additional pre and post test sessions lasted approximately
10 minutes each, and questionnaire were administered after all the sessions. All
participants completed the presented sessions, functional measurements, and
questionnaires without any adverse event or report from the participants. ADAPT
provided adaptive training tailored to patients’ performance by modulating task difficulty
in the training session. The results from this study validate the feasibility of ADAPT for

xi
rehabilitation of arm and hand function after stroke, and provide justification for
continued investigation of clinical efficacy.

１
Chapter 1. Introduction

1.1. Motivation
Over 50 million Americans have a disability that affects one or more of their major
life activities (Field & Jette, 2007; McGinnis & Moore, 2006) and only 30% of adults
with disabilities are employed (Cooper et al., 2008). The economic cost of the associated
loss of functioning is high. In particular, about 700,000 people suffer from a new or
recurrent stroke each year, with about 4,700,000 Americans living with the effects of
stroke (American Heart Association, 2003). The estimated annual burden from stroke-
related disability is $53.6 billion, of which $20.6 billion is in indirect costs due to lost
productivity and income (American Heart Association, 2006). Over 80% of first-time
strokes (infarctions only) involve acute hemi-paresis of the upper limb (Desrosiers et al.,
2003), for which this proposal has a direct and immediate audience. Though the incidence
of disability directly linked to upper extremity hemiparesis resulting from Cerebral palsy
or Traumatic head injury is smaller than for stroke, the need for effective treatments in
these chronically disabled populations who are aging with these disabilities is equally
high (Kemp & Moaqueda, 2004). Cerebral palsy affects 1.2 in 1000 children; 17% of
children who have cerebral palsy and were preterm at birth, and 56% who were term
experience a hemiparesis (Menkes, 1995). Stroke and other neurologic insults in
childhood add to these numbers. The population of adult with CP hemiparesis who are
aging provides yet additional challenges. Center for Disease Control (CDC) estimates
that at least 5.3 million Americans, approximately 2% of the U.S. population, currently
have a long-term or lifelong need for help to perform activities of daily living as a result

２
of a TBI (Thurman, Alverson, Dunn, Guerrero, & Sniezek, 1999). TBI can cause a wide
range of functional changes affecting thinking, sensation, language, and/or emotions. It
can also cause epilepsy and increase the risk for conditions such as Alzheimer’s disease,
Parkinson’s disease, and other brain disorders that become more prevalent with age
(National Institute of Neurological Disorders and Stroke, 2002).
Recent technological advances in computation, machine learning, communication,
and miniaturization of sensors bring us closer to a future where intelligent devices and
technology-embedded environments can contribute to reducing disparities in employment
and full community participation for individuals aging with disabilities (Cooper, 2008).
Automated rehabilitation is a particularly promising and active area of research, given the
potential cost savings of utilizing a rehabilitation system available in a person's home
without a need for costly trained professionals.
Despite the tremendous promise and the large effort involved in developing and
testing of automated rehabilitation systems in the last two decades, actual deployment in
the clinics or in the patient’s home has so far been limited. Therefore, the primary
motivation for this dissertation study is to develop an automated rehabilitation system
that is designed according to the latest evidence-based guidelines of stroke rehabiliation,
overcomes the drawbacks of exising automated rehabilitation systems, and is readily
exploitable in the current medical practice environment.

３
1.2. Principles for stroke rehabilitation
Retraining upper extremity functions in the neurological population with
hemiparesis is clinically important because a substantial number of activities of daily
living involve use of the upper extremities (Coster, et al., 2004). Rehabilitation of reach
and grasp skills has been shown to be critical for patients in their attempts to return to a
reasonable quality of life (Carr & Shepherd, 2003) and improved participation (Duncan,
et al., 1999). Therefore, we focus on the development of a robotic rehabilitation system
for the rehabilitation of arm and hand function after stroke.
To be effective, our robotic rehabilitation system should be designed based on
current rehabilitation principles derived from basic and clinical rehabilitation research.
We describe four principles with recent basic research evidences and clinical study
reports.
Fisrtly, intensive practice improves performance. There is now definite (Phase III
RCT) evidence that intensive task-specific practice, in which participants actively engage
in repeated attempts to produce motor behaviors beyond their present capabilities, is
effective for improving upper extremity function and use after stroke (Butefisch, 1995; G.
Kwakkel, Wagenaar, Twisk, Lankhorst, & Koetsier, 1999; Wolf, Blanton, Baer, Breshears,
& Butler, 2002; Wolf et al., 2006; Wolf et al., 2008). However, current medical practice
does not adequately allow for the required training intensity. Although experiment with
animals and clinical trials show that effective rehabilitation requires thousands of
movements over weeks, (Doyon et al., 1997; Karni, 1995; Nudo, Wise, SiFuentes, &
Milliken, 1996; Pavlides, Miyashita, & Asanuma, 1993) patients receive only a small
fraction of this dose (Keith & Cowell, 1987; Lang et al., 2009; Lincoln, Willis, Philips,

４
Juby, & Berman, 1996; Mackey, Ada, Heard, & Adams, 1996). Such low levels of
practice in the clinics are in part because conventional medical rehabilitation models are
constrained by economic considerations, at least in the US, and are therefore not adequate
to maximize functional outcomes. In addition, because of the belief that therapy is only
marginally effective, health insurance companies often reject requests for rehabilitation in
the chronic period of disability. This reality is perhaps most evident in the population of
adults with hemiparetic CP who received regular rehabilitation services until the age of
21. A critical means to improve functional outcomes in patients with neurological
disability is to increase dramatically the amount of task practice, while maintaining
overall costs at a reasonable level. One possible method to increase practice time in a
cost-effective manner is to supplement the patient's one-to-one interaction with a physical
therapist with sessions on automated, adaptive, and low-cost task practice systems that
can be used in the clinic or at home to largely augment the actual dose of therapy needed
to improve outcomes.
Secondly, task-specific practice is more beneficial than non functional task practice.
A principal role of the physical or occupational therapist in rehabilitation of upper
extremities after stroke is to encourage re-learning of motor tasks. For this purpose, the
therapist selects salient tasks for the patient to work on, adjusts task parameters by
increasing difficulty and complexity as learning progresses, and provides informative and
motivational feedback (Gordon, 2000). What type of tasks should the therapist set-up?
First, meaningful, tasks should be used (Bayona, Bitensky, Salter, & Teasell, 2005; Fisher
& Sullivan, 2001; Van Peppen et al., 2004; Winstein et al., 2004). Second, the
manipulation of real and functional objects is beneficial (van Vliet, Kerwin, Sheridan, &

５
Fentem, 1995; C. Wu, Trombly, Lin, & Tickle-Degnen, 2000). Thus, upper extremity
rehabilitation involving manipulation of real objects in a variety of functional and
meaningful tasks can improve function.
Thirdly, patients should be actively performing the motor tasks. Interventions with
active participation and repetitive practice have been shown to be effective (Duncan,
1997). When patients are too weak to move voluntarily after stroke, passive movements
are often used as part of the rehabilitation program. However effectiveness data is lacking
(Jeffery & Good, 1995). Furthermore, although passive movements elicits activity in the
motor cortex (Carel et al., 2000), actively generated movements are more effective in
eliciting both performance improvements and cortical reorganization (Conner, Culberson,
Packowski, Chiba, & Tuszynski, 2003; Karni, 1995; Lotze, Braun, Birbaumer, Anders, &
Cohen, 2003; Muellbacher, Ziemann, Boroojerdi, Cohen, & Hallett, 2001; Nudo, et al.,
1996). Motor adaptation is evoked significantly more by unhindered movement than by
haptic guidance (Kluzik, Diedrichsen, Shadmehr, & Bastian, 2008). Along these lines, it
has recently been questioned whether assistive robotic approaches have the potential to
produce greater benefits than is possible with simpler techniques, such as unassisted
practice; instead, repetitive movement attempts by the patient, rather than the assistance
from the robot, may be the effective ingredient (Kahn, Lum, Rymer, & Reinkensmeyer,
2006; G. Kwakkel, Kollen, & Krebs, 2008; Marchal-Crespo & Reinkensmeyer, 2009).
Finally, the tasks should be always “optimally challenging” to maximize both
relearning and motivation. Animal research suggests that challenging tasks that elicit
motor learning, but not repetitive use, enhance motor cortical reorganization (Conner, et
al., 2003; Kleim et al., 2002; Nudo, et al., 1996; Plautz, 2000). Thus, the tasks should not

６
be trivially simple. However, tasks should not be too hard either, as learning failure may
occur (Sanger, 2004). Thus, the tasks should have the “just-right” (Ayres, 1972) or
“optimal” (Vygotsky, 1978) challenge. Besides enhancing learning, or re-learning after
neurological insult, the optimal challenge has also been linked to motivation. This is
important because effective motor rehabilitation in general, and training on our simple
devices in particular, could be excruciatingly tedious and dull. Thus, for our systems to be
truly effective in the clinic or patient’s home, we need to pay careful attention to
maximize and maintain the patient’s motivation. Initial engagement in performing
activities such as learning new motor skills, or re-learning skills in case of individuals
with hemiparesis, depends on the value placed on the skill (presumably high in the case
of re-learning motor skills neurological insult in a number of individuals) and on the
expectancy to reach rehabilitation goals. Subsequent sustained engagement in the
learning process, however, largely depends on the activities themselves and their
associated feedback (Sansone & Harackiewicz, 2002). In particular, motivation research
suggests that, to enhance and maintain high levels of intrinsic motivation: (i) the
challenges must be in accord with one’s capability (Deci & Ryan, 1985; Hebb, 1955), and
(ii) one needs to receive feedback of progress towards the goal (Csikszentmihalyi, 1979).
The above principles are well addressed in the definition of stroke rehabilitation
from Ontario Stroke Rehabilitation Consensus Panel 2007: “Stroke rehabilitation is a
progressive, dynamic, goal-oriented process aimed at enabling a person with an
impairment to reach his or her optimal physical, cognitive, emotional, communicative
and/or social functional level”. Thus, implementing those principles in a robotic system
will define the robotic stroke rehabilitation, and we aims at developing such robotic

７
rehabilitation system to enhance the recovery of upper extremity functions in patients
with stroke.

1.3. Existing robotic systems
In view of the shortcomings of the conventional medical practice model for
rehabilitation, there is a growing interest in employing robotic rehabilitation technology
for rehabilitation of upper extremity functions. We reviewed several pioneer works that
provided valuable ideas and lessons to be addressed for the development of our robotic
system.

Figure 1.1. The MIT-MANUS trains elbow and shoulder flexion and extension and
shoulder abduction and adduction.

MIT-MANUS is a 2 DOF planar device that assists shoulder and elbow movement
by moving a patient’s affected hand as shown in Figure 1.1 (Aisen, Krebs, Hogan,
McDowell, & V olpe, 1997). Patient’s lower arm and wrist are strapped into a brace that is
attached to the handle of the robot. Most exercises are visually guided games that require
the movement of the handle. Altough it is widely used for retraining reaching movements,

８
and has been extensively tested in clinical trials using automated intensive training with
motivating games (Fasoli, Krebs, Stein, Frontera, & Hogan, 2003; MacClellan et al.,
2005), there are several limitations that do not address the latest principles for stroke
rehabilitation. The tasks of the MANUS are not real functional tasks based on daily living
activities but virtual tasks based on games. Because the robot mostly assists the
movements of the affected arm with the hand constrained on the handle, it does not
require active reaching movement, but limited active engagement in the manipulation of
the handle.

Figure 1.2. The Stanford MIME robotic device assists the movement of the affected limb.
It also can execute mirror movements with the affected and unaffected limbs
simultaneously using a second robotic arm.

Stanford MIME uses a PUMA 560 robot manipulator to assist the movement of
forearm in a large range of three-dimensional space unlike the MIT-MANUS (Burgar,
Lum, Shor, & Machiel Van der Loos, 2000). The force-base assistance is triggered when
the measured patient’s force is large enough against the robot. A unique modality of the

９
MIME is bimanual exercise, where active movement of the unaffected arm is mirrored by
simultaneous passive movement of the affected arm by the robotic device (Figure 1.2).
However, the MIME has limitations in task-specificity and active engagement similar to
the MIT-MANUS.
ARM-guide system assists straight-line reaching movement of the affected arm
(Reinkensmeyer, Emken, & Cramer, 2004). The ARM-guide uses ‘guided force training’
where patient's reaching movement is halted if the patient pushes with an abnormally
large force perpendicular to the straight-line. If the ARM-guide’s reaching movement,
which is crucial to the patient’s active engagement in daily living activities, is combined
with daily living activities, its effectiveness might be enhanced.

Figure 1.3. The Northwestern ARM-Guide makes and assists the reaching movements in
one direction at a time.

There are also other similar robotic systems such as Bi-Manu-Track (Hesse, Schulte-
Tigges, Konrad, Bardeleben, & Werner, 2003), GENTLE/S (Coote, Murphy, Harwin, &
Stokes, 2008), Neurorehabilitation Robot (NeReBot) (Rosati, Gallina, & Masiero, 2007),
REHAROB (Fazekas, Horvath, & Toth, 2006), Arm Coordination Training 3-D (ACT3D)
(Sukal, Krosschell, & Dewald, 2007), and ARMin (Nef et al., 2007) that assists the robot

１０
assists the movements of the affected limb. Recognizing that pure assistance may not be
entirely beneficial for recovery of arm and hand function (Kahn, Lum, Rymer, &
Reinkensmeyer, 2006; Marchal-Crespo & Reinkensmeyer, 2009), later developments
included balancing assistance provided by the robot with active movement by the patient,
e.g. (Kahn, Rymer, & Reinkensmeyer, 2004), EMG triggered robots (Dipietro, et al.,
2005), and rehabilitation of wrist (Krebs, et al., 2007) and hand (Dovat, et al., 2008)
functions. These systems have been shown to be effective to some extent (Krebs, Aisen,
Volpe, & Hogan, 1999; Krebs, Hogan, et al., 1999; Kwakkel, Kollen, & Krebs, 2008;
Lum, Burgar, Shor, Majmundar, & Van der Loos, 2002), and can be used with
participants that have no or little residual movement capabilities. However, most of them
could not fully address the latest principles for stroke rehabilitation.

Figure 1.4. The AutoCITE presents 8 functional tasks in semi-automated way.

There are few rehabilitation systems, which are most relevant to the principles, but
not all. AutoCite (Lum et al., 2004), a semi-automated (non-robotic) system that allows
patients to engage in the practice of functional tasks, has been shown to be as effective as
standard Constraint Induced (CI) therapy (Taub, Lum, Hardin, Mark, & Uswatte, 2005).
In AutoCite, however, the number of tasks cannot be increased easily, task selection is

１１
manual, and task difficulty is not adjusted automatically.
A smart training system with a robotic arm (M. P. Dijkers et al., 1991; Erlandson,
deBear, Kristy, Dijkers, & Wu, 1990) was developed to re-train reaching movements with
adaptive difficulty. This system however does not contain automatic task selection and is
only designed for reaching.
ADLER (Johnson, Wisneski, J., Nathan, & R., 2006) presents multiple functional
tasks for activities of daily living with several artifacts. Although it provides realistic
functional tasks, ADLER cannot implement adaptive task schedule, because it does not
have an automatic artifact changing system.

Figure 1.5. ADLER presents functional tasks with multiple artifacts, which are
interachageable.

１２
Chapter 2. Design of ADAPT

2.1. Introduction
No previous robotic approach could fully address the latest principles of stroke
rehabilitation. Not many of them could exactly parallel the role of rehabilitation
therapists, who typically expend considerable effort to set up functional tasks that require
a patient to actively engage in challenging reach-and-grasp practice (Gordon, 2000).
While computational motor control and learning principles such as context, task
generalization and training schedule have been expected to extend the efficacy of robotic
neurorehabilitation (Huang & Krakauer, 2009), the design of previous robotic systems
could not support such principles. Notably, no previous system has been designed to
manipulate task schedules using an automatic tool changer to optimize re-learning of
multiple functional tasks.
We thus propose a new system, ADAPT (ADaptive and Automatic Presentation of
Tasks), that automatically presents functional tasks that require reaching and
manipulation, that can accommodate an expanding number of such tasks, and that allows
the implementation of performance-based adaptive task scheduling and adaptive
modification of task difficulty. ADAPT simulates the dynamics of daily living functional
tasks, such as opening a doorknob, opening a jar, turning a key, etc. Unlike most other
robotic systems, ADAPT does not move the patients. Instead, it presents tasks adaptively,
such that each patient can perform doable, but constantly challenging tasks. Therefore,
the system is designed for patients with some volitional motor capability of the arm and
hand, as those patients benefit the most from intensive rehabilitation (Carr & Shepherd,

１３
2003). More specifically, we designed ADAPT in accordance with the principles
addressed in Chapter 1 as shown Figure 2.1.
The primary aims of this paper are to present the design and the control architecture
of ADAPT, and to show the successful operation of our approach. The utility of ADAPT
for rehabilitation with stroke patients will be tested in future clinical trials.

Figure 2.1. The principles of stroke rehabilitation and implementation them in ADAPT.

１４
2.2. Intelligent control architecture of ADAPT
Figure 2.2 presents the overall conceptual design of ADAPT system with the control
architecture. The high-level adaptive task scheduler aims at maximizing the re-learning
of multiple functional tasks, and balancing learning among tasks in a limited training time:
at each trial, it selects both the task to practice within a task bank and the task difficulty
based on the patient’s previous performance, input from the therapist, and constraints
given by the adaptive schedules. The high-level adaptive task scheduler also sends a
command to the tool changing system that picks up the tool corresponding to the selected
task. The functional task model generates a desired trajectory to simulate the dynamics of
the task with the difficulty specified by the high-level adaptive task scheduler. The low-
level admittance controller computes the control current corresponding to the desired
trajectory. The computed control current is then applied to a general-purpose robot to
simulate the task, and the patient feels the simulated dynamics by reaching and grasping
the selected tool. The safety surveillance module continuously monitors the status of the
robot and the controllers in the ADAPT system, and provides appropriate actions to
minimize any risk to patients.

１５

Figure 2.2. The conceptual design of the ADAPT system. At each trial, the high-level
adaptive task scheduler adaptively chooses a task and its difficulty based on subject’s
performance, prior practice records, and rehabilitation therapist’s input. Rehabilitation
therapist’s input is not implemented in the current version of ADAPT. The tool changing
system automatically selects the tool corresponding to the selected task. The low-level
admittance controller computes the control current needed to simulate the desired task
dynamics from the functional task model during robot-patient interaction. The control
architecture of ADAPT controller is implemented in Linux-operated computer

2.3. Robot and devices
The robot used in ADAPT is a modular and reconfigurable robot from AMTEC
Robotics (AMTEC) (Figure 2.3). The current configuration consists of a 3-degree-of-
freedom (DOF) wrist mounted on a 1-DOF linear actuator. This configuration allows the
system to present the end-effector at different linear vertical locations and to rotate the
end-effector in almost any orientation. Although it provides less haptic fidelity than
backdrivable haptic devices, our general-purpose robot can generate higher torque, which
is needed for a number of functional tasks (eg. opening a jar). The low backdrivability
also makes it easier for the robot to automatically pick up new tools with our tool

１６
changing system. Furthermore, due to both the simplified computations of kinematics and
dynamics and the small workspace compared to traditional multiple DOFs robotic arms,
our robot provides safe task presentation to patients. Finally, note that the current
configuration can be relatively easily extended with additional AMTEC linear and
angular modules to increase the number of possible tasks and ranges of motion.
A 6-DOF ATI Force/Torque (F/T) sensor (MINI SI-580-20) is attached to the end-
effector to measure interaction forces between the subject and the robot. The tool changer
connects a functional task tool, such as a doorknob, to the F/T sensor. Encoders in the
motor module provide position data of each joint. A Pentium-4 3.4Ghz PC with a Linux
operating system (SUSE 10) receives the position data via a CAN bus, receives the
interaction force data via a National Instrument A/D converter, and sends control
commands via the CAN bus.
Functional task tools such as a doorknob, a screwdriver, a jar, a faucet, keys, etc. are
arranged in a tool rack. After the high-level adaptive task scheduler chooses a task, the 4
DOF robot positions each joint so that it can pick up the appropriate tool in the tool rack.
As shown in Figure 2.4, the robot’s end effector is equipped with a master plate, and each
tool is equipped with an interface plate from ATI Corp. A pneumatic system ensures the
locking and unlocking of the tools by a 4/2 way pneumatic valve (V5A-3341-BX1,
MEAD corp), which is computer-controlled via RS-232C, a serial communication. The
PC sends a lock command to a 24 VDC relay circuit, which sets the direction of the air
flow in the pneumatic valve such that the master plate of the tool changer docks with the
tool plate. After the two plates are docked, the robot repositions to present the new tool to
the subject. In the current version, four tools are arranged in the rack, and up to six tools

１７
can be included. The tool changing process of these tools in use is demonstrated in Figure
2.5.
The subject is seated in a chair facing the robot with an in-between table, lays
his/her hands on the table, and practices the tasks simulated by the robot with the more
affected upper limb. As shown in Figure 2.4, with different task tools, patients can
practice a variety of different reaching and grasping tasks.

１８

Figure 2.3 Robot and Devices for ADAPT. ADAPT controller program in Linux PC
controls the robot via CAN and the tool changer controller via serial communication (RS-
232C). Tool changer controller controls the master plate pneumatically to lock and unlock
a tool equipped with a tool plate . Two kinds of emergency buttons: one is on top of the
power box, the other is in hand of an unaffected arm. Two sensors: torque sensor is
attached between the master plate and robot’s endeffector, encoder is embedded in each
module of the robot. (Thick line: signal path, thin solid line: functional blocks, dotted line:
highlighted component)

１９

Figure 2.4. Different functional task tools in use. Six different grasping tasks with four
different functional task tools (doorbell: bell pushing, jar: jar opening, doorknob: knob
turning, door locking, door opening, and screwdriver: screwdriving) are developed. Two
emergency stop buttons is in the range of both unaffected arm and affected arm. The
home position pad has an on-off switch to detect the hand.

２０

Figure 2.5. Tool-changing process of ADAPT.

２１
2.4. Safety
Safety was a crucial issue in the design of ADAPT. From the initial robot design
process, we made special efforts to maximize operational safety. Our choice of design
makes our robot safer than a traditional multi-DOF robotic arm because of the small
overall workspace. The linear DOF is only used for tool positioning, not for task
dynamics simulation. Furthermore, the patient is not strapped to the robot.
For simplicity and safety reasons, we chose functional tasks that require movements
around a single DOF during robot-subject interactions. Because many functional tasks in
daily living (such as turning a key or door handle, steering, opening a jar, turning a water
faucet, wrist supination/pronation, etc.) need only a single rotary DOF, this configuration
does not overly restrict the number of tasks. After a task is set up for presentation to the
subject, the magnetic brakes that are built into the robotic articulations are engaged on the
other three DOFs during subject–robot interactions. This single DOF method simplifies
kinematics and dynamics computation; the robot thus never falls into the wrist-singular
posture, which can occur in PUMA-like manipulators (Yokokohji, Hollis, & Kanade).
Several surveillance routines are implemented to limit the maximal torque output
and cap the maximum velocity of the linear and rotational motors. Watchdog routines,
which continuously check for failure of the position and force sensors, computer crashes,
and electrical failures, can automatically freeze the robot by engaging the magnetic
breaks in all DOFs at any time.
Furthermore, two emergency stop buttons can stop all robot operation and turn on
magnetic brakes to disable any movement of all 4 DOF of the robot. The main emergency
red stop button of the power box is accessible to the therapist. The patient holds the

２２
second emergency stop button at all times with his/her less affected hand (see Figure 2.3
and 2.4).
Finally, to limit possible patient-robot collisions, subjects are seated with their trunk
fastened to the back of the chair by a seatbelt. This seatbelt has the additional advantage
of limiting compensatory movements with the trunk.

２３
Chapter 3. Adaptive Task Schedule: Performance-based adaptive schedule
enhance motor learning

3.1. Introduction
As described in the conceptual design of ADAPT (Figure 2.2), the control
architecture consists of three main components, high-level adaptive task scheduler,
functional task model, and low-level admittance controller. The high-level adaptive task
scheduler has two adaptive components: the first determines the task schedules, the
second sets the difficulty of the task selected at each trial. For each component, we
propose two algorithms based on subjects’ performance to enhance the learning of the
functional tasks, and test their effectiveness using a visuo-motor task. The preliminary
experiment with the visuo-motor task will provide supporting theoretical background for
the scheduling method to enhance learning multiple functional tasks.
Learning multiple motor skills is ubiquitous in activities such as sports, music,
professional skill development, or rehabilitation after brain injury. The positive
effects of constant versus variable practice schedules on learning, which emerged
directly from the predictions of the Schema Theory (Schmidt, 1975), have been
shown repeatedly (McCraken & Stemach, 1977) (Pigott & Shapiro, 1984) (Catalano
& Kleiner, 1984). Lee et al. (Lee, Magill, & Weeks, 1985) reviewed these literatures
and determined that for variable practice to be most effectively utilized (relative to
constant practice) it should be randomized, rather than blocked. Such positive
effect of random practice on learning has been found in a number of studies (Boyce
& Del Rey, 1990; R. A. Carlson, Sullivan, & Schneider, 1989; R.A. Carlson &

２４
Yaure, 1990; Goode & Magill, 1986; Osu, Hirai, Yoshioka, & Kawato, 2004; Shea &
Morgan, 1979). Thus, although individual skills can be scheduled sequentially, the
contextual interference effect, in which random scheduling of several tasks enhances
performance measured in delayed retention tests compared to sequential, or blocked,
scheduling. Note however, that not all investigators have been able to replicate the
contextual interference effect, and many of these studies use a transfer tests with few
trials and not retention tests.
Despite its successes, however, random scheduling suffers from at least two
serious limitations. First, it does not take into account the two components of task
difficulty (Guadagnoli & Lee, 2004), that is, the difficulty of the task itself (nominal
task difficulty), which is constant, and the skill level of the learner (functional task
difficulty), which improves as learning progresses. Second, it does not take into
account the difference in difficulty between tasks, that is, all tasks in the learning
program are treated equally for all learners at all time.
Matching task difficulty to the skill level of the learner, both initially and as
learning progresses, has the potential to enhance learning efficacy of each task. On
one hand, the tasks should not be too easy to perform, as mere repetitions do not lead
to change in performance, and do not seem to induce cortical reorganization, which
is produced by mastering challenging tasks (Kleim, et al., 2002; Nudo, et al., 1996;
Plautz, 2000). On the other hand, the tasks should not elicit too large initial errors; in
such cases, failure of learning can occur despite repeated practice (Sanger, 2004). In
motor learning, (Bernstein, 1967) has suggested that people attempting to learn to
perform a difficult motor task try to ameliorate the degrees-of-freedom problem

２５
through a developmental progression. In a similar vein, but in an educational context,
(Vygotsky, 1978), proposed that task difficulty should be maintained near the
“optimal” point (or the “just-right” challenge (Ayres, 1972)) to enhance learning
effectiveness. In artificial neural networks, a few studies have shown that adaptive
scheduling of task difficulty can enhance learning. (Sanger, 2004) showed that
training a neural network to learn the dynamics of a multi-joint arm was much faster
when the speed of desired movements was slowly decreased compared to training the
network on fast movements. Note that by keeping difficulty constant as learning
progress, the learning curve became relatively flat. (Ivanchenko & Jacobs, 2003)
showed that a network whose training included developmental progressions
outperformed fixed training on motor tasks that were relatively difficult to learn.
When initial difficulty is too large however, that is when the combined high nominal
and functional task difficulty create large errors, because of the lack of availability
of appropriate training examples, learning can fail, and performance does not
improve (Sanger, 2004).
Taking into account the difference in difficulty between tasks has the potential
to enhance learning performance globally. The total amount of practice is the single
most important variable for skill acquisition (Schmidt & Lee, 1999). Further,
performance improvement as a function of practice trials can be well modeled by
negatively accelerated monotonically increasing functions, such as exponentials or
power functions (Liu, Mayer-Kress, & Newell, 2003; Schmidt & Lee, 1999).
According to such models, for each new practice trial, the gain in performance is less
than the gain of the previous trial; in the limit, the gain tends to zero. Furthermore,

２６
depending on nominal difficulty and on the skill of the learner, the rate of
performance improvement varies from task to task. Thus, in simple random
scheduling, it may be that for an easy task, each new trial induces a “labor in vain
effect” (Nelson & Leonesio, 1988) as performance improvement levels off. For a
harder task, however, performance may still be relatively poor, but still improving.
The effectiveness of adaptive scheduling in the number of trials has repeatedly been
shown in artificial neural network learning with “active input selection” methods. In
these methods, the input (“tasks”) to the neural network that potentially led to large
decrement in error (Zhang, 1994), or that are maximally informative (MacKay, 1992)
are selected for neural network training. These method have been shown to reduce
largely the training time of the neural network, e.g. (Zhang, 1994)

3.2. Method
3.2.1. Subjects and design
48 college students, including 30 females, participated in this experiment that took
place in four consecutive days. Informed consent was obtained from all subjects, and the
study approved by the USC Human Subjects Institutional Review Board.
The subjects were required to learn four visuo-motor transformations over the
course of three learning sessions, each separated by 24 hours. A learning session
consisted of 120 trials. At each trial, subjects were instructed to move a cursor shown on
a computer screen from an initial position to a target within a limited time, the allocated
movement time (AMT), using a sort of force feedback joystick, a “spaceball” (HP 5000).
Each visuo-motor transformation was defined by a color-coded target and an associated

２７
angular relationship between the spaceball movement and the cursor movement. The four
angles (-30 deg, 60 deg, -90 deg and 120 deg) were randomly assigned to the four color
coded targets for each subject. For each task, nominal difficulty was controlled by the
amount of time allocated to reach the target, AMT.
The subjects were randomly assigned to one of four conditions, with 12 subjects per
condition, in a two-by-two factorial design: 1) fixed number of trial and fixed difficulty
(Fix), 2) adaptive number of trials and fixed difficulty (AdapTr), 3) adaptive difficulty
and fixed number of trials (AdapDif), and 4) adaptive number of trials and adaptive
difficulty (AdapTrDif). In the Fix and AdapDif conditions, the number of trials was
constant, with 40 trials per task per session. In the AdapTr and AdapTrDif conditions, the
number of trials per task was allowed to vary between 10 and 90 in sessions 2 and 3. In
the Fix and AdapTr conditions, the difficulty was constant, with AMT = 0.6 sec. In the
AdapDif and AdapTrDif, AMT was allowed to vary between 2.5 and 0.6 sec in all
sessions, and was set at 2.5 seconds at the beginning of session 1 for all tasks.

3.2.2. Experimental procedures
Subjects were comfortably seated at a table in front of a computed display, and were
holding the spaceball in their right hand. The position of the subjects was normalized
both across subjects and across days: we measured the distance between the spine (at the
C3-C4 level) and the acromion process of the scapula. The subjects were positioned such
that the mid-line was aligned with the middle of the computer display. The spaceball was
positioned and attached on the table with Velcro at 15 cm in front of the subject’s body
and to the right of the midline at the spine-scapula distance.

２８
At the beginning of each trial, a colored target was shown as a 3 cm diameter disk
located 10 cm above the initial position, indicated as a white fixation cross. After one
second, a change of the fixation cross color from white to the color of the target,
indicated the “GO” signal. After a variable reaction time (RT), the subjects started to
move the spaceball; the subjects had AMT seconds to move the cursor onto the target.
The cursor movement was visible on the screen at any time t < = AMT. At t = AMT, the
cursor movement was stopped. The cursor and the target were visible during the
intertrial interval, which lasted a minimum 1 sec. The variable intertrial interval was
adjusted such that the trial duration was always 5.5 seconds (Go signal duration + RT +
AMT + intertrial interval = 5.5 second). If the RT was larger than 1 second, the trial was
aborted and “Next time move faster” was displayed until the next trial started.
Performance error was defined as the distance between the cursor position and the
target position at AMT. If the cursor was on target for a duration of 100 ms before the
end of trial, the performance error was zero, the trial was terminated and a bird chirping
sound was played. The maximum performance error was 16 cm. In the adaptive difficulty
conditions, because the task difficulty was manipulated to keep performance relatively
constant (see below), it is hard to perceive any progress in performance directly from the
task. Thus, to indicate progress in performance, a display above the target showed the
score of each task at all time in all conditions. The score was given in the same color as
the task. Performance feedback was computed by scaling AMT (given by Equation 4, see
below) between 0 and 100 (AMT was computed in all schedules, but used only for
feedback in the Fix and AdapTr schedules).

２９
Each session began with a pre-test and ended with a post-test, with four pseudo-
randomly distributed trials per task in each test. A fourth test was given 24 hours after the
last session. We note here that the pre-tests on day 2 and 3, and the test on day 4, can be
considered as 24-hour delayed retention performance tests
Fixed condition (Fix)
The FIX schedule was the control condition. Tasks were scheduled pseudo-randomly,
such that each task was scheduled once in a block of four trials. Thus, there were 30 trials
per task per session, and the difficulty was kept constant at AMT: 0.6 sec.
Adaptive number of trials condition (AdapTr)
In this condition, we adjusted the number of trials per task in sessions 2 and 3 by
distributing the total number of trials among the tasks:
) ( ) ( task PerfE rials totalNoOfT task NoOfTrial
norm
 
(3.1)
where rials totalNoOfT is the total number of trials in one learning session, and
) (task PerfE
norm
is the normalized performance error for the task. To obtain an estimate
of performance for each task, we measured performance error directly with a test before
each training session. However, because this test was given at fixed (high) difficulty level,
it must be short (or the purpose of adaptive difficulty schedules would be defeated); as
performance was estimated in a few trials, it was likely to be noisy. Therefore, we also
used past performance error. Although performance during practice is often not a good
indicator or long-term retention (Cahill, McGaugh, & Weinberger, 2001; Winstein, 1991),
it reflects a measure of task difficulty. We computed the normalized performance error by
combining these two (imperfect) measures in the following parameter-free Equation:

３０
() ()
()
(( ) ( ))
norm
ii
i
PerfEc task PerfEp task
PerfE task
PerfcE task PerfEp task





(3.2)
where ) (task PerfEc is the performance error obtained at the pre-test preceding practice
in the current session (which can be considered as a 24-hour delayed retention
performance test), ) (task PerfEp the performance error in the test immediately
following the previous practice session, and the denominator is a normalizing factor with
the sum over all tasks0F
1
. As tasks compete for the number of trials, we provided a range
of number of trials to avoid the possibility that one task takes all the available trials in a
session (maximum: 75 and minimum: 15). Once the numbers of trials were determined
(after the pre-test in sessions 2 and 3), the trials were scheduled in a pseudo-random
manner to minimize consecutive trials of the same task: only if one task has is allocated
more than 60 trials, consecutive trials can occur.
Adaptive difficulty condition (AdapDif)
The adaptive schedules attempted to maintain an optimal challenge for each task by
manipulating the allocated movement time ) (t AMT at each trial t. If the current
performance error is above the reference error, then task difficulty is too high, and
allocated movement time is increased. If on the contrary, current performance error is
below the reference error, then difficulty is too low, and allocated movement time is
decreased. To accomplish this, we used an error-reduction learning rule, in which the
change in difficulty is based on the squared difference between performance error and the
reference performance error, which determines the optimal challenge:

1
Our choice of formula to compute normalized performance error is motivated by the ‘predict’ and ‘correct’ steps in
Kalman filter (Welch & Bishop, 2004), which uses the Bayesian rule: the posterior estimation of the performance error
value is obtained by multiplying the prior estimation PerfEp(task) by the likelihood PerfEc(task), obtained by current
measurements (Thrun, 2000), divided by a normalization term.

３１
2
) ) ( ( ) (
ref
PerfE t PerfE t E   , where t is the trial index, ) (t PerfE is the performance
error on trial t, which is small when performance is good,
ref
PerfE is a reference
performance error.
Our difficulty update algorithm is derived from the assumption that performance
curve of motor learning can be relatively well modeled with exponential functions. This
model allows us to compute the derivative of the squared difference between the subject
performance and the reference performance error. Through this mathematical
minimization, we can derive flat performance curve near the reference performance error.
At each trial, we adjusted the task difficulty ) (t AMT by subtracting a fraction of the
derivative of this difference with respect to the number of trials. Thus, the allocated
movement time ) (t AMT at trial t is given by:
()
(1) ()
dE t
AMT t AMT t
dt
   (3.3)
where is a small positive constant. By assuming that performance is well modeled by an
exponential function of the number of trials t, with the task difficulty controlling the
learning rate: ) ) ( exp( ) ( t t AMT A t PerfE    , we obtained the following update rule:
(1) () (1 ( () ))
ref
AMT t AMT t PerfE t PerfE     
(3.4)
where  is a “learning rate”. Because the optimal challenge point is an unknown
parameter, the reference performance error
ref
PerfE was set at 4 cm after pilot testing.
Adaptive difficulty for each task started at the onset of session 1 and continued in
sessions 2 and 3. AMT was allowed to vary between 2.5 and 0.6 sec, and was initially
set at 2.5 seconds at the beginning of session 1 for all tasks. Tasks were scheduled

３２
pseudo-randomly as in the Fix condition, with 30 trials per task.
Adaptive difficulty and number of trials (AdapTrDif)
In this condition, the difficulty is adapted as in the AdapDif condition, and the
number of trials as in theAdapTr condition.

3.3. Results
We first report results for the performance error for the average of the four tasks, and
then for individual tasks. There was no significant difference between conditions in pre-
test of day 1 (one way ANOV A, F(3,44) = 5.21, p = 0.22). In the delayed retention test on
day 4, there was an effect of condition on performance error (Brown-Forsythe test,
F(3,22.9) = 4.11, p = 0.018). Subjects performed better in the AdapTrDif condition
(performance error: 2.31 ± 0.42 cm) than in the control Fix condition (5.58 ± 1.18 cm, p
= 0.016, see Figure 3.1.A). The effect size was ES = 1.11, which is considered a large
effect size. Similarly, subjects performed better in the AdapTr condition (2.63 ± 0.42 cm),
than in the Fix condition (p = 0.038, see Figure 3.1.A). The effect size was ES = 0.96,
which is also a large effect size. Although results did not reach significance, there was a
trend when comparing performance error in the adaptive difficulty condition (3.45 ± 0.60
cm) with performance error in the control condition (p = 0.045, non-corrected for
multiple comparisons, ES = 0.64, see Figure 3.1.A). There was no significant difference
for the other comparisons (p = 1). Note that in the fixed schedule, some learners could not
learn the harder tasks, and exhibited “failure of learning” (Figure 3.1.B, top panel).

３３
A B

Figure 3.1.A: Average performance and standard error on the four tasks on the delayed
retention test on day 4 for each condition. The stars indicate significantly smaller errors in
the adaptive number of trial condition (AdapTr) and in the adaptive number of trial and
adaptive difficulty condition (AdapTrDif) compared to the Fix (control) condition (p <
0.05, Bonferroni corrected for multiple comparisons) B. Histogram showing performance
error on each trial in the session 4 test for each condition. Note the relatively large
number of trials with poor performance in the fixed schedule condition Fix – including a
number of trials where performance error was maximum (16 cm), indicating failure of
learning.

Performance error in the fixed schedule on day 2 can be considered as an index of task
difficulty (in days 3 and 4, there may be floor effects, and in the first day, since the
subjects are using the joystick for the first time, performance was very noisy). Mean
performance error in day 2 for task 1 (-30 degrees) was 7.90 cm; performance for task 2
(60 degrees), 8.18 cm; for task 3 (-90 degrees) 10.30 cm, and for task 4 (120 degrees):
10.35 cm.

３４

Figure 3.2: Mean performance in the five trials of the delayed retention test given on day
4 for each task and each schedule. The performance of the Fix group (squares and solid
line) is repeated on each panel for comparison with the other three adaptive groups,
which show better performance (smaller errors) for all trials and tasks.

Figure 3.2 shows the mean performance across subjects for each task and each
condition on day 4. As is apparent in the figure, there is a trend in worsening performance
for task 2, 3 and 4 in the fix schedules compared to task 1. As is also apparent in the
figure, the adaptive schedules reduce the differences in performance between tasks.
Considering all the tasks and schedules on day 4, there was no main effect of task on
performance (Repeated ANOVA F(3, 132) = 1.85, p = 0.141), but there was a significant
interaction task x condition effect (F(9, 132) = 0.26; p = 0.026). Further, there was no
effect of condition for task 1 (Brown-Forsythe test, p = 0.10), and task 4 (120 degrees,

３５
Brown-Forsythe test, p = 0.259). However, there were significant effects of condition for
task 2 (60 degrees; Brown-Forsythe test, p = 0.037) and task 3 (-90 degrees; Brown-
Forsythe test, p = 0.002). Thus, a possible floor effect limited the effects of adaptive
schedules for task 1. For task 4, a possible ceiling effect, or a slow rate of learning, also
limited the effect of the adaptive schedules.

Figure 3.3: Examples of adaptive number of trials schedule in the AdapTr condition for
the four tasks. The arrows on the left indicate the angles of the visuomotor transformation
for each task (-30, 60, -90, and 120 degrees). The dots show the performance for each
trials, for the three practice sessions (from left to right), for tasks 1 to 4 (from top to
bottom). Performance error during training is fitted with exponentials for comparison of
the four tasks. Filled circles: median performance in the tests before training. Plus marks:
median performance in the tests immediately following training. The numbers of trials for
each task are shown left of performance for each session. For task 3, performance was
relatively poor in the post-test in the pre-test in session 2; thus, a larger number of trials
(37) were scheduled for this task in session 2. On the contrary, this learner’s performance
was good in task 2 before session 2; hence the relatively small number of trials for this
task in this session. In each session, the total number of trials was kept constant at 120
trials.

３６

Figure 3.4. Example of adaptive difficulty schedule for the four tasks. The arrows on
the left indicate the angles of the visuomotor transformation for each task (-30, 60, -90,
and 120 degrees). A: Performance error as a function of trials (cm). Note that for the four
tasks, the average performance error was quickly brought near the challenge point (4 cm).
B: Change in the difficulty parameter AMT (seconds) in each trial. As performance
improved, the task difficulty, controlled by AMT, decreased. Note that because tasks 1
and 2 were relatively easy for this subject in session 1, AMT decreased rapidly in the first
session in order to maintain performance around “optimal challenge”, as can be seen in
the corresponding graphs in A.
A
B

３７
The adaptive number of trials schedules produced different number of trials for each
task, and more trials were allocated to harder tasks than easier tasks, as desired. Figure
3.3 shows an example of a task schedule in the AdapTr condition. In this condition, in
day 2, the mean number of trials change between task 1 and task 2 was 47.7%, between
task 1 and task 3, 87.5%, and between task 4 and task 1 67.3% In day 3, the mean number
of trials change between task 1 and task 2 was 63.5%, between task 1 and task 3 was:
67.8%, and between task 1 and task 4 was 69.9%. The number of trials for task 1 was
significantly less than 30 trials in the AdapTr condition, which is the number of trials for
each task in the fixed schedule condition (t-tests, day 1: p = 0.003, and day 2: p = 0.0045).
We then compared the number of trials between tasks in the AdapTr condition. An one-
way ANOV A showed a main effect of task (p = 0.0064). In day 2, task 1 was significantly
different from all the others (p < 0.05, Bonferroni corrections) and in day 3: task 1 is
different from task 3 (p < 0.05, Bonferroni corrections).
Although, our adaptive number of trial algorithm was designed to minimize the
number of consecutive trials, consecutive trials are inevitable if more than 60 trials are
allocated to a single task. The different number of trials for each task between the control
group and the AdapTr-groups is potentially confounding number of trials with number of
consecutive trials (If there are more trials for task A, two or even three consecutive trials
for A are more likely). This is unlikely to be the case however, because only two
schedules out of the 48 schedules with adaptive number of trials generated (12 subjects in
the two adaptive number of trial conditions, AdapTr and AdapTrDiff, in both the second
day and the third day) had consecutive trials: one schedule at day 2 in AdapTr (with 15,
15, 75, and 15 trials per tasks) and the other at day 3 in AdapTrDif (17. 20, 15, and 68

３８
trials per tasks).
In the adaptive difficulty schedule, performance error was maintained near the reference
performance error, as desired. Figure 3.4 shows an example of a performance and
adaptive difficulty in the AdapDif condition. In the Fix condition, average performance
over subjects and all 30 trials in sessions 1 and 2 was 10.44 ± 0.13 cm and 6.75 ± 0.14 cm
respectively. In the AdapDif condition, average performance in sessions 1 and 2 was 6.73
± 0.16 cm and 3.84 ± 0.11 cm respectively. Performance during practice in the AdapDif
condition was better than that in the Fix condition in both sessions (two-tail unbalanced
Satterthwaite test , p < 0.001 for both sessions) and closer to the reference performance
error in the AdapDif condition (one sample t-test with 4 cm as the test value: AdapDif
condition : p = 0.15, and Fix condition p < 0.001).
In our study, the task difficulty is manipulated with the movement duration. Thus,
adapting the task difficulty is potentially confounded with changing the duration of
overall time the learners are actively performing the tasks, and the results improvement in
performance could be simply due to an increase in total practice time. To study this
possibility, we computed the actual total practice time for the subjects in the AdapDif
group and correlated it with performance in day 4. We found no correlation between total
practice time and performance in this group (r = -0.053, p = 0.80).

3.4. Conclusion and discussion
The results of these experiments demonstrate that adaptive scheduling can
improve performance in multi-session multi-tasks learning program. Adapting the

３９
number of trials as a function of performance, largely improved retention compared
to fixed, random scheduling. Varying difficulty also improved learning although to a
lesser extent. Furthermore, although not significantly different from performance in
the adaptive number of trial, probably because of a floor effect, performance in the
adaptive number of trial and difficulty condition was superior to that in the fixed
condition. In all adaptive conditions performance on all tasks improved; in the
fixed schedule however, some learners could not learn the harder tasks, and exhibited
“failure of learning” (Figure 3.1.B, top panel). Nominal task difficulty was crucial in
determining positive effects of the adaptive schedules however: the easiest and most
difficult task did not show significant effects of schedules, possibly due to a floor
and a ceiling effect; the effects of adaptive schedules were large in tasks of
intermediate and high difficulty, however.
Given the practical importance in enhancing the efficacy and efficiency in motor
learning in a number of domains such as sports, professional skill development, and
rehabilitation, it is surprising that only a few behavioral studies have looked at
adaptive practice schedules to enhance learning. Two types of adaptive scheduling
methods have been designed, that can be divided in “learner controlled” and
“computer controlled” (as in the present study). In learner controlled methods, the
learner determines the practice schedules. (Titzer, Shea, & Romack, 1993) showed
that a learner-determined schedule had the same beneficial effect as blocked during
acquisition and was equivalent to random practice in retention, thus facilitating both
performance and learning. (W. F. Wu, Magill, & Foto, 2005) showed that the
performance in a learner-determined schedule group was superior to that of a yoked-

４０
control group. A possible limitation of these methods, however, is that the chosen
schedules are based on imperfect meta-cognitive judgments: because learners suffer
from illusions-of-competence during practice (Simon & Bjork, 2001), these
schedules may not optimally enhance long-tem retention. Besides our study, we are
aware of only one other motor learning scheduling study that does not rely on meta-
cognitive judgments, but in which the schedule is “computer controlled” and based
on measured performance: the “win-shift/lose-stay” method (Simon, Cullen, & Lee,
2002). In this method, switching to another task occurs only after the learner has
achieved a criterion level of success. Although beneficial to learning, a potential
limitation of this method is that it does not distinguish between learning and
performance (Cahill, et al., 2001; Winstein, 1991), as it is performance during
practice that is used to tailor the schedule. The method we proposed here adapts the
number of trials based on immediate and delayed retention performance.
The notion of optimal challenge has been studied in motivation research, and
has notably been linked to intrinsic motivation, which reflects one’s propensity to
engage in a task for its own sake, and, in doing so, seek out and master “optimal”
challenges, i.e. challenges in accordance with one’s capability (Deci & Ryan, 1985;
Hebb, 1955; White, 1959). Intrinsic motivation has been shown to be sustained if the
“optimal” challenge is itself sustained, and if reception of informed feedback shows
one’s progress towards the goal (Csikszentmihalyi, 1979). Thus, future work should
aim at dissecting the effects of optimal challenge in enhancing learning (as has been
shown here) and in enhancing intrinsic motivation. Further work should also aim at
determining the “optimal challenge point” meta-parameter. In the present study, a

４１
reference performance error was empirically determined and taken as an approximate
challenge point. New methods should be developed to automatically adjust this
parameter, perhaps based both on user preference and on performance. Because
different practice schedules yielded large differences in delayed performance error,
we believe that our findings are robust, and warrant the development of similar
methods outside the laboratory. Thanks to the recent availability of relatively cheap
and simple motion capture systems, adaptive task scheduling as described in the
present paper could be used in real applications, such as sports, or rehabilitation of
hand function after brain injury.
Our next goals with this research are to apply the proposed algorithms to the high-
level adaptive task scheduler in Figure 2.2 to select the right task at the right difficulty
based on patients’ performance, and to test if they enhance the recovery of upper
extremity functions in patients with stroke.

４２
Chapter 4. Functional Task Model: Capturing the dynamics of functional
tasks

4.1. Introduction
The simplest way for stroke patients to practice functional tasks is to practice with
real functional task tools such as a door knob, a jar with glass bottle, a gear shift etc. The
problem with using such passive tools however is that the difficulty of the tools cannot be
adapted to the individual, and (presumably) improving, patient’s performance. Thus,
some patients with severe impairments might not be able to practice at all with real task
tools. Therefore, our ADAPT system is designed to simulate those tasks using the
dynamics models of them and present the functional tasks to stroke patients with adaptive
difficulty based on their skill. In this chapter, we discuss how to identify and model the
dynamics of functional task tools so that the dynamics can be simulated in ADAPT to
present multiple functional tasks with adaptive difficulty for stroke rehabilitation. More
specifically, we address how the functional task model in Figure 2.2 generates the desired
trajectories for the selected task and difficulty, if the task to practice and the task
difficulty have been determined by the high-level adaptive task scheduler.
Accurate modeling of the dynamics of nonlinear passive tools is the first and
essential step for realistic haptic simulation of them. The dynamics of those tools are not
so simple to be modeled in a simple linear Equation, but it may include many nonlinear
physical properties such as friction, detents, backlash and other nonlinearities that are
difficult to model for use in haptic simulation.
Several previous efforts have focused on the challenging of modeling the dynamics

４３
of those tools to the fidelity acceptable for haptic simulation. One of most typical
methods is to formulate an ideal dynamics Equation from the device design and adjust the
parameters of the dynamics Equation (Allotta, Colla, & Bioli, 1999; Hayward &
Armstrong, 2000). However, the complexity of device design and the inability of the
formulated dynamics Equation to model it may decrease the fidelity of the haptic display.
One popular solution to this limitation is to measure the physical properties directly by
actuating the device (Colton, et al., 2005; Hasser & Cutkosky, 2002; K. E. MacLean &
Durfee, 1995; Miller & Colgate, 1998; Swindells & MacLean, 2007). Some works with
this method just record the haptic profile (e.g. Force-position relationship) and play it in
the haptic interface without fitting to a parameterized model (Angerilli, Frisoli, Salsedo,
Marcheschi, & Bergamasco, 2001). Even though it is easy to implement without
identification process, most of these methods are limited to specific devices and it is
difficult to generalize to different devices or the different feel of even the same device.
The others fit the measured data to a parameterized model with special technique of
parameter estimation (Colton, et al., 2005; K. M. MacLean, 1996). Still, their application
is limited to specific devices.
Our goal in this research is to develop the method for the modeling of nonlinear
passive devices such that (1) it should be easy to generalize the modeling process to other
nonlinear passive devices, (2) the feel (or the difficulty of operation) of the device should
be adjustable, (3) the parameterized model should have meaningful physical properties,
and (4) the model should be easily simulated in the haptic device. Colton and Hollerbach
(Colton, et al., 2005) proposed a modeling approach most relevant to the four criteria is
from. They used exponentially-weighted least-squares (EWLS), a recursive regression

４４
algorithm to estimate the parameters of their general nonlinear model with the measured
data. The parameters for the general nonlinear model of a nonlinear spring at each data
point were computed with EWLS, meaning that each data point has a different parameter
set. To predict the force from the general nonlinear model with the estimated parameters,
it should preserve all parameter sets at each data point or should interpolate the
parameters at specific increment. Preserving all the varying parameters complicates the
implementation of the estimated model in the haptic display, and the interpolation with
the estimated parameters needs to be developed not to make any fitting problem.
To meet the criteria and simplify the model structure, we first propose to formulate
the modeling problem with receptive field weighted regression (RFWR) algorithm
(Schaal & Atkeson, 1998), which is a constructive incremental version of locally
weighted regression. RFWR constructs multiple local linear models (receptive fields)
adaptively depending on the measured data. Receptive fields are created adaptively when
more accuracy is needed and deleted when the model is over-fitted. Second, we propose
to use Karnopp’s friction model as an essential local model for a receptive field, because
friction is all but absent in all human interaction with passive devices. Depending on the
physical properties of a target device, other linear component such as stiffness, damping
or inertia can be added to the essential local model. Therefore, multiple of local
Karnopp’s friction models (or with additional physical properties) are combined to model
the target device with RFWR algorithm. Friction occurs in all passive devices while a
user interacts with them, and the Karnopp’s friction model was shown to be a good
choice in haptic display (Richard, 2000). Therefore, with Karnopp’s friction model as
an essential component for the local model in RFWR, the proposed modeling method can

４５
be generalized to many different devices. Further, each local model represents
meaningful physical properties so that we can easily adjust the properties of the simulated
dynamics.
In the next section, we present how to model the devices with RFWR. RFWR is
modified to use Karnopp’s friction model that is nonlinear. We also test our approach
with simulation, and compare with other approaches.

4.2. Modelling with RFWR
4.2.1. RFWR algorithm
RFWR is one of the regression algorithms that learn the function describing the
relationship from input to target. In RFWR, the function values are approximated by a
combination of N individually weighted locally linear models. Then, these values are
normalized by the sum of all weights (Schaal & Atkeson, 1998). Thus, the predicted
output value for a query point x

is given by





K
k
k
K
k
k k
w
y w
y
1
1
ˆ
ˆ
(4.1)
T T
k k
T
k k
T
k k
c x x x b b c x y ) 1 , ) ((
~
, ) ( ˆ
, 0
  

 
       (4.2)
where
k
y ˆ is the individual prediction from each receptive field,
k
 is the linear model
parameter, the weights
k
w corresponds to the activation strength of each receptive field
and
k
c is the center of k-th linear model. The weights are characterized by a kernel
function, a variety of which have been used in (Atkeson, Moore, & Schaal, 1997). For

４６
example, a Gaussian kernel is used to compute the weights as
k
T
k k k k
T
k k
M M D c x D c x w      )), ( ) (
2
1
exp(
   
(4.3)
where
k
w is a positive definite distance matrix. While learning with RFWR, the
parameter
k
 for each local linear model and the distance matrix for the weights to
determine the shape and size of the receptive field are updated.
The parameter
k
 can be computed by the least squares for locally weighted
regression (Atkeson, et al., 1997). Given the input matrix
T
p
x x x X )
~
,...,
~
,
~
(
2 1
 with all p
training data points, similarly the output matrix
T
p
y y y Y ) ,..., , (
2 1
  
 , and the
corresponding diagonal weight matrix
T
w w w diag W ) ,..., , (
2 2 1
  
 , the parameter vector
 can be given as:
WY X WX X
T T 1
) (

  (4.4)
The recursive version for the above least squares can compute exactly the same result
(Ljung & Söderström, 1986):
)
~
( ,
~ ~
~ ~
1
,
~ 1 1 1
x y e
x P x
w
P x x P
P P e x wP
n
cv
n
n T n
n n T
cv
n n n



   













   
  


(4.5)
where  is a forgetting factor to gradually cancel the contributions from previous data.
This update rule guarantees the global minimum of the weighted squared error with
avoiding the explicit matrix inversion. While updating the parameter
k
 , the RFWR
algorithm should adjust the shape and size of the local model, which represent how each
data point is weighted depending on the distance from the center of each local model, and

４７
therefore how all local models are combined to construct the regression with the RFWR.
The distance matrix in Equation (4.3) specifies the shape and size, and is updated by
incremental gradient method:
M
J
M M
n n


 


1
(4.6)
The cost function for the update:
 
 




n
j i
ij
p
i
i i i
i i i
D
x P x w
y y w
W
J
1 ,
2
1
2
2
)
~ ~
1 (
ˆ
1
 (4.7)
The left term of the cost function J is the locally weighted leave-one-out cross
validation error for function approximation to become asymptotically unbiased, and the
right penalty term avoids an ever increasing number of receptive fields. See (Schaal &
Atkeson, 1998) for more details about the RFWR algorithm.

4.2.2. Karnopp friction model with RFWR
Here, we show how we included the Karnopp model within the RFWR framework
to model the dynamics of any 1 DOF rotary tool with friction.
Friction exists in all passive devices, and it is highly nonlinear that there exist a
variety of different models for it (Olsson, Astrom, de Wit, Gafvert, & Lischinsky, 1998).
(Richard, 2000) characterized friction with Karnopp model, and simulated the estimated
dynamics successfully with the haptic interface. We propose to use the Karnopp friction
model as the local model (receptive field) in the modeling with RFWR as the friction is
essential in modeling the dynamics of the passive devices. In (Richard, 2000), the
Karnopp model was formulated as a linear model with extra input dimensions using the

４８
signum function of velocity, and the parameter was computed with least squares. Figure
4.1 demonstrates the Karnopp friction model used in (Richard, 2000). The corresponding
model Equation is expressed as:







  
  
 
 

v q for q b q C
v q for D
v q for D
v q for q b q C
n n
other n
other p
p p
friction
  


  
) sgn(
0 ) , max(
0 ) , max(
) sgn(



(4.8)
where
other
 is the non-frictional external torque applied to the passive device. In v q   ,
the only way to predict the
friction
 is to measure the
other
 . In v q   , the Equation (4.8)
can be represented in linear form:
     














  
n
n
p
p
n n p p friction measured
b
C
b
C
q q q q     ) sgn( ) sgn(
(4.9)

or, in matrix from with all measured data points as:
 



























 










 
n
n
p
p
N N N N frictionN
friction
measured
b
C
b
C
q q q q
q q q q
   
   
) sgn( ) sgn(
... ... ... ...
) sgn( ) sgn(
...
1 1 1 1
1

(4.10)
         X
friction measured

(4.11)
where  is a noise. In this way, the Karnopp friction model of Equation (4.9) can be
used as the local model in Equation (4.2). The local model can have other component

４９
such as inertia, stiffness, or damping depending on the properties of the passive device.
For example, the door knob might have stiffness and inertia properties. Then Equation
(4.9) will have stiffness term:
        
stiffness inertia friction measured

   
    
  

























   
x
k
k
m
b
C
b
C
q q q q q q q
n
p
n
n
p
p
n p n n p p
stiffness friction inertia measured
      ) sgn( ) sgn(

(4.12)

Therefore, we can formulate the local model for RFWR with general passive device
dynamics as in the case of Equation (4.12).
After formulation of the local model with the Karnopp friction model based function,
the next step is how to weight the local models, more specifically, in which dimension
the local model is changing. In (Colton, et al., 2005), the experiment with a linear spring
demonstrated the position dependence of the parameters (stiffness, damping, inertia,
friction). The fact that changing position means more or less the deformation of most
passive devices may validate the position dependent weighting in RFWR. We modified
RFWR algorithm for only position to affect the weighting of the individual local model.
Therefore, the distance matrix in Equation (4.3) should be only one dimension with

５０
position, and the center of each local model in Equation (4.3) should be specified by only
position.

Figure 4.1. The modified Karnopp model. C
p
and C
n
are the positive and negative values
of the dynamic friction, b
p
and b
n
are the positive and negative values of the viscous
friction, velocity is the relative velocity between two surfaces, D
p
and D
n
are the positive
and negative values of the static friction, The width of shaded area defined by v is where
the velocity is considered to be zero.

4.3. Simulation
In this section, we test the proposed approach with numerically simulated data.
For the simulation, a swept sine waveform trajectory (Swindells & MacLean, 2007) was
used to generate the training data set:
) )
) (
( sin(
2 2
2
a a
d
t a b
a b
c q  



 

(4.13)

５１
)
) (
(
2
) )
) (
( cos(
2 2
a
d
t a b
d
a a
d
t a b
a b
c q 

   



 
 
 (4.14)
where q is angle, q  angular velocity, ttime from 1 to 2 by 0.0002 second, and
) , , , ( d c b a constants set to (0, 2, 1, 3). The torque data set from the perfect model with
noisy torque is generated with the assumptions that the passive device is composed of
stiffness and friction properties, and they are changing depending on angle. The perfect
model with the assumptions is:





















  
n
p
n
n
p
p
n p n n p p stiffness friction Model
k
k
b
C
b
C
q q q q q q     ) sgn( ) sgn(   

(4.15)
The model with noisy torque is given by:
) 1 , 0 ( * 3 . 0 N
Model noisy Model measured
       
(4.16)
where ) 1 , 0 (  is random number selected from Gaussian distribution with mean zero
and variance one. The training data for simulation are generated by the following
position-dependent parameters for Equation (4.15):
2 2 2 2
2 2
4 4
4 . 1 , 6 . 0 , 6 . 1 , 4 . 0
) 0 ( 8 . 0 , 0 . 1
) 0 ( 8 . 0 , 6 . 0
q b q C q b q C
q if q k q k
q if q k q k
n n p p
n p
n p
        
  
  

(4.17)
The training data with the parameters of Equation (4.17) and three position
dependent local models are illustrated in Figure 4.2.A The Figure 4.2.A explains
graphically how local models (Karnopp friction-based models) are different depending on

５２
position value.
Figure 4.2.B is the approximated model trained by RFWR with position-dependent
weighting. In more nonlinear range of small angles, more receptive fields are created, but
in more linear range, less receptive fields are created. The learning performance from
simulation is measured by a normalized mean square error (nMSE). The resulting nMSE
for the simulation was 0.015. Therefore, we expect the proposed modeling method to be
applied to real functional task modeling.

５３

Figure 4.2. (A) Simulated training data, and possible local models for the training data.
(B) The model learned from the simulated data by RFWR

５４
4.4. Experimental setup
The picture in Fig 4.3 shows how the training data for the modeling of functional task
are recorded in case of door knob turning task. We obtained motion and torque data to
train the above model with the F/T sensor introduced in section II and a magnetic motion
sensor (MINI bird; Ascension Technology Corporation) attached to the passive tool to be
modeled. The tool is attached to the torque sensor, which is positioned not to move. The
subject then manipulates the passive tool, and the torque and angle data are
simultaneously recorded. The position data from the magnetic motion sensor (MINI bird)
and torque values from Force/Torque sensor are transmitted to the PC and recorded. The
training data is post-processed to be used for the modeling with RFWR.

Figure 4.3. Experimental setup for capturing dynamics. The robot is not moving with the
magnetic brake on.

4.5. Trajectory generation from the functional task model
After the model of Equation (4.1), which uses with the local linear model of Equation
(4.12), is learned from the training data, the dynamics model for the functional task tsk
and difficulty Diff can be assumed to be:
6 DOF
Force/Torque
Sensor
Magnetic
Motion Sensor
Functional
tool
Linux PC
Subject
Interaction
Torque
Serial Communincation
NI A/D converting card

５５
1
1
0,
(, )
,
(,1)
K
kk
k
K
k
k
TT
kk k k
TT
wy
y tsk Diff
w
y xb b x Diff
xx




 






(4.18)
where we can assume that  of Equation (4.11) (the modified Karnopp model) is equal to
each 
k
. Then, by inserting (4.18) into (4.12) with assumption that I
tool
is the inertia of the
tool and the current state vector ) , , ( q q q    is (ang, vel, acc), the desired trajectory can be
derived:
t vel ang
t
I
Diff tsk y
vel vel
I
Diff tsk y
acc acc
Diff tsk y acc I
d d
tool
measured
d
tool
measured
d
tool measured inertia measured
 


 

 
    
) , (
) , (
) , (


  
(4.19)

where (ang, vel, acc) is a current state vector, (ang
d
, vel
d
, acc
d
) is a desired state vector,
and t is a sampling time of the robot. The current state vector is measured by the
encoder of the robot at each time step.

4.6. Results
Training data to model doorknob turning and jar closing-opening were obtained as
explained in section 4.4. Figure 4.4.A demonstrates the sample trajectory of turning and
releasing a real doorknob in a torque versus angle plane. One important aspect of

５６
doorknob dynamics is that turning (positive velocity) movements require more torque
than releasing (negative velocity) movements because of friction: this is clearly shown
near zero velocity in Figure 4.4.C. This confirms that the Karnopp model is well suited
for modeling doorknob dynamics. The important aspect of doorknob dynamics is that
both the stiffness and the friction magnitude are changing with position (angle). For
example, the slope near -2 rad is larger than the slope near -0.5 rad, and the torque
difference between turning and releasing at angle -2 rad is larger than the difference at -
0.5 rad (Figure 4.4.A). Position (angle)-dependent local Karnopp models are thus
necessary. We used local Karnopp models with the state input vector and the parameters
of Equation (4.15). Figure 4.4.B shows the collected training data, the doorknob
dynamics model built on the training data, and Gaussian weight functions. The final
learned dynamics model contains fifteen local Karnopp models, which are combined after
learning with RFWR to generate the model. The combined model corresponds to y in
Equation (4.1) and Equation (4.18), and the weight function is defined by Equation (4.3).
In Figure 4.4.C, the dots represent the collected the training data at angles between -1.55
rad and -1.45 rad, and the solid lines are the torque values predicted from the model with
inputs of the velocity range between -10 rad/s and 10 rad/s and at angle -1.5 rad. As can
be seen in the figure, the model has the same shape as the modified Karnopp model (c.f.
Figure 4.1), and fits the data very well. A sample trajectory and the training data of jar
closing-opening are plotted in Figure 4.5.A and 4.5.B, respectively. The jar dynamics is
almost a pure friction problem, with no stiffness, with the friction magnitude changing
depending on angle (higher friction with more closing).

５７

Figure 4.4. Doorknob dynamics model. (A) Torque versus angle trajectory of a real
doorknob turned and released three times. (B) Training data (dots) and the doorknob
dynamics model (surface) after training by RFWR with the modified Karnopp friction
model. Solid lines on the bottom are fifteen weight functions given by equation (2). (C)
Local torque versus velocity relation by the trained model (solid lines) and the training
data (dots) around angle = -1.5 rad (-1.55 < angle < -1.45).

５８

Figure 4.5. Jar dynamics model. (A) Torque versus angle trajectory of a real jar closed
and opened three times. (B) Training data (dots) and the jar dynamics model (surface)
after training by RFWR with the modified Karnopp friction model. Solid lines on the
bottoms are eighteen weight functions corresponding to (2). (C) The local torque versus
velocity relation by the trained model (solid lines) and the training data (dots) around
angle = 0.2 rad (0.19 < angle < 0.21).

５９
As shown in Figure 4.5, the jar does not move until the exerted torque exceeds a
positive threshold for jar closing or is less than a negative threshold for jar opening.
Because it was hard for the subject to generate high velocity for jar closing, too few
training data were collected to obtain reasonable viscosity values (b
p
and b
n
in Equation
(4.15)). We therefore set these viscosity values to zero. Figure 4.5.C shows that this
assumption is reasonable, as the velocity range of the training data is narrow. With
assumption that the current state vector ) , , ( q q q    is (ang, vel, acc), the input state vector
of Equation (4.12) without stiffness and with zero viscosity is then given by:
 

T
n p
n n p p
C C
vel vel vel vel x
0 0
) sgn( ) sgn(




(4.20)
As for the doorknob, we applied the local Karnopp models with the state input vector
and the parameters of Equation (4.20) to the RFWR regression. Figure 4.5.B shows the
collected training data, the jar dynamics model built with the training data with our
proposed method and the eighteen weight functions given by RFWR after training. The
local training data and the model at angle 0.2 rad in the Figure 4.5.C illustrate the exact
shape of the modified Karnopp model with zero viscosity (c.f. Figure 4.1). The good
model fit validates the use of the local Karnopp friction model for the jar as well.

4.7. Conclusion
The results in Figure 4.4.BC and Figure 4.5.BC clearly shows that our novel
modeling methodology can be used to closely predicts the dynamics of passive tools with
different physical properties. Furthermore, from this modeling method, we could easily
extract parameters related with stiffness, damping and friction for the generalization of

６０
task difficulties, because each local linear model corresponds to a typical linear dynamics.
The captured dynamics will be simulated by ADAPT using the low-level admittance
controller in Figure 2.1.

６１
Chapter 5. Functional Task Display: Model Based Lower-level Admittance
Controller for Haptic Display

5.1. Introduction
Haptic devices are increasingly used in a large range of applications such as
teleoperation (Park & Khatib, 2006), virtual reality (Gomez, Burdea, & Langrana, 1995),
robotic surgery (Cavusoglu, Tendick, & Sastryand, 2001), and rehabilitation therapy (H. I.
Krebs et al., 2004). In many of these applications, haptic system development has
focused on the design of new special purpose haptic devices. Special purpose haptic
devices have many advantages because the design of the system can be optimized for the
application purpose. Even though they have shown successful performance, they might
have several drawbacks. For example, it is quite difficult to reapply them to other
applications. Further, it takes relatively long time to develop such a special-purpose
design, and thus the development costs could be too high to be used in a wide range of
fields.
One possible alternative could be off-the-shelf, general-purpose robots such as
PUMA-like manipulators (Burgar, et al., 2000; Clover, 1999; Yokokohji, et al., 1999) or
modular robots (Y. G. Choi & Schweighofer, 2008). These robots can significantly
reduce development time, mostly can provide wider range of operation, and can generate
higher torque. Furthermore, as shown in (Y. G. Choi & Schweighofer, 2008), we can
easily install extra equipments for additional functionalities. However, most general-
purpose robots have relatively poor haptic performance compared with special-purpose
haptic interfaces, because they are highly geared, lowly back-drivable, and have

６２
relatively high friction and inertia. Although several researchers have successfully
minimized these effects to use general-purpose robots in haptic applications (Burgar, et
al., 2000; Y. G. Choi & Schweighofer, 2008; Clover, 1999; Yokokohji, et al., 1999),
feedback force bandwidth, and stability could not be achieved as good as in special-
purpose haptic interfaces. Moreover, there has been relatively less research in improving
the haptic performance of general-purpose robots due to the above mechanical
deficiencies.
The primary aim of this chapter is to develop a novel controller that can simulate
functional tasks with a general-purpose robot. The controller should be able to display the
functional task dynamics modeled from the previous chapter through ADAPT, which is a
4-DOF general-purpose robot with low back-drivability. The captured functional task
dynamics will be displayed as a practicing task to stroke patients by ADAPT with the
controller, as the low-level admittance controller in Figure 2.1. More specifically, the low
level controller receives the desired trajectory from the functional task model captured in
Chapter 4, and controls ADAPT to simulate the capture dynamics of the selected task
with different levels of difficulties. In addition to proposing the novel controller for
ADAPT, we address how to improve the haptic performance of a general-purpose robot
in terms of high damping display.

６３

Figure 5.1. Low-level admittance controller for ADAPT. The controller generates the
desired motion in response to the external force by the subject. The measured external
force is input to the functional task model that defines the dynamics of the functional
tasks. Based on the dynamics of the task, the functional task model generates the ideal
desired motion. f1 computes the control current for the desired motion, and f2 computes
the control current to compensate against the external force. The desired state vector q
d

and the current state vector q corresponds to (ang
d
, vel
d
, acc
d
) and (ang, acc, vel) in the
text.

5.2. The architecture of the low-level admittance controller
The low level controller (Figure 5.1) receives the desired trajectory from the
functional task model. Because our general-purpose robot has low back-drivability, we
used an admittance control strategy to compute the control signal for the desired motion
response to external force. The control signal is the motor current, and is computed from
the outputs of two feedforward control modules and a typical PD feedback controller. The
first module is a controller for motion without interaction force (motion model f1), and
the second is a controller for interaction force without motion (force model f2). We
trained both modules off-line with RFWR. Such modular design of the low-level
admittance controller with RFWR has three main advantages: First, learning the control
currents directly allows accurate control even when, as in our system, the robot’s
Feedback
Controller
(PD)
Robot
q
d
cur
fb
cur
ff
q
+
−
+
+
cur
motion
+
cur
force
τ
human
Force Model f2
(Feedforward
Controller)
Motion Model f1
(Feedforward
Controller)
-
Low-level
Admittance Controller
Functional
Task Model

６４
dynamics are unknown, and when the control current is (presumably) not proportional to
the motor torque. Second, compared to a combined controller, our modular controller
simplifies the training of the controllers, because separation of the force and motion
models allows us to train each model with fewer training data points. Third, the force
model can be used to simulate infinite stiffness or damping, which is needed to
implement a virtual wall or tasks with high friction. The motion model is trained with a
typical sinusoidal excitation without any human-robot interaction. Training data for the
force model are recorded from the F/T sensor while a subject exerts force on the tool with
the robot immobilized with position control. We assume the dynamics of our robot to be:
measured d d d robot motor
vel ang n acc I      ) , (
(5.1)
where (ang
d
, vel
d
, acc
d
) are the desired angles, angular velocity, and angular acceleration,
respectively, τ
motor
is the control torque, τ
measured
is the external torque measured by F/T
sensor, I
robot
is the inertia of the haptic interface, and n(ang
d
, vel
d
) models possible
nonlinear effects. As τ
measured
in Equation (5.1) is the interaction torque exerted by the
subject, the other terms in Equation (5.1) contribute to τ
motor
only via motion. Thus, from
the dynamics Equation for motion only (with no human-robot interaction), we assume
that the current-output function for motion only is given by the motion model f1 with
desired trajectory as an input:
) , , ( 1 ) (
) , (
d d d motor motion
d d d robot motor
acc vel ang f f cur
vel ang n acc I
  
 


(5.2)
Similarly, we assume that the current-output function for interaction only is given by f2
with τ
measured
as an input:

６５
) ( 2
measured force
f cur    (5.3)
From Equations (5.1), (5.2), (5.3), the real control current to the motor is given by:
) ( 2 ) , , ( 1
measured d d d motor
f acc vel ang f cur    (5.4)
Finally, to drive the errors between the desired motion and the actual motion to zero,
feedback terms are added to Eqaution (5.4) as:
) ( ) (
) ( 2 ) , , ( 1
ang ang K vel vel K
f acc vel ang f cur
d p d d
measured d d d motor
   
  
(5.5)
The desired trajectory (ang
d
, vel
d
, acc
d
) is numerically integrated from (4.19), as
described in the previous Chapter.

5.3. The performance of the low-level admittance controller
Next, we tested 1. the performance of the low-level admittance controller with these
two tasks, and 2. the feasibility of the proposed modeling method for adjustment of task
difficulty. Figure 5.2 shows results of the robot simulation of the doorknob task for three
levels of difficulty. In Figure 5.2.A the model of Figure 4.4.B was used. Although some
errors are present because of sensor noise and inherent limitations of our robot due to its
low backdrivability, the trajectory closely follows the learned model (surface). In Figure
5.2.BC, the difficulty was adjusted by multiplying 
k
in Equation (4.18) by a constant,
Diff. The surfaces in Figure 5.2.BC show the desired model after scaling all terms of the
model with the difficulty, Diff = 0.5, and Diff = 2, respectively. High stiffness simulation
as in Figure 5.2.C showed good trajectory. As shown in the releasing trajectory of Figure

６６
5.2.B, very low stiffness seems to generate more errors, which we believe are due to the
inherent limitations of our robot. However, the trajectory itself does not deviate much
from the model, and the subject reported good “door-knob feel” in all difficulty
conditions.
Figure 5.3 shows the robot performance for jar closing-opening for the three
difficulty levels. Although the overall shape of the trajectory follows the model, the jar-
opening trajectory has a relatively higher starting velocity than the trajectory of the
training data in Figure 4.5.B, which may be due to our assumption of zero viscosity
parameters in the Karnopp model. Nonetheless, the subject reported that the feeling of
opening a jar with ADAPT was very similar to real jar opening. As shown in Figure
5.3.BC, the simulations with different difficulties successfully followed the desired
model; in both cases, the subject also reported good feeling of jar closing-opening. In
sum, the low-level admittance controller successfully displayed the two representative
functional tasks with adaptive difficulties in ADAPT.

６７

Figure 5.2. Dynamics models of the doorknob and trajectories simulated by ADAPT for
three different difficulty levels. (A) The doorknob dynamics model (surface) after
training by RFWR with the modified Karnopp friction model and the trajectory (dots)
simulated by ADAPT during interaction with a subject. The difficulty of this model is
used as a reference. (B) The difficulty was decreased by half and the trajectories
simulated by ADAPT during interaction with a subject. (C) The difficulty was increased
by twice and the trajectories simulated by ADAPT during interaction with a subject.

６８

Figure 5.3. Dynamics models of the jar and trajectories simulated by ADAPT for three
different difficulty levels. (A) The jar dynamics model (surface) after training by RFWR
with the modified Karnopp friction model and the trajectory (dots) simulated by ADAPT
during interaction with a subject. The difficulty of this model is used as a reference. (B)
The difficulty was decreased by half and the trajectories simulated by ADAPT during
interaction with a subject. (C) The difficulty was increased by twice and the trajectories
simulated by ADAPT during interaction with a subject.

６９
Chapter 6. The Feasibility of ADAPT: Feasibility tests of ADAPT to a
healthy subject and patients with chronic stroke

6.1. Introduction
ADAPT should be extensively tested to be used in clinical trials to help the recovery
of upper extremity functions in patients with stroke. We need to validate the functionality,
reliability, and safety of ADAPT. The preliminary experiments with subjects are required
to estabilish the feasibility of ADAPT. Because safety is a crucial issue in the robotic
systems with human interaction, we need to perform long-term, intensive, and systematic
tests with a greater number of subjects, both healthy and with stroke.
In this chapter, we address two feasibility studies for ADAPT. First study is a
preliminary experiment with a healthy subject to test mainly the functionality and safety
of ADAPT. Second one is an experiment with five participants with chronic stroke to test
the acceptance, and safety of ADAPT. The primary aim of these studies is to estabilishthe
feasibility for a Phase I randomized clinical trials (RCT) to test the efficacy of ADAPT in
future work.

6.2. The experiment with a healith subject
The primary aim of this study is to test mainly the scheduling properties of the high-
level adaptive task scheduler of Figure 2.1 and overall functionality of ADAPT with the
adaptive scheduling of three tasks: doorknob turning, jar closing-opening, doorbell

７０
pushing. Additionally, the reliability and safety of ADAPT will be tested, and the reported
problems will be addressed in the next feasibility study with participants with stroke.

6.2.1. Method
Only one healthy subject participated. The experiment contained two practice
sessions, each consisting of 90 trials and separated by one hour (in future therapeutic
applications of ADAPT, sessions may be separated by days or weeks). In the first session,
we only adjusted the task difficulty, keeping the number of trials equal for all tasks. In the
second session, we adjusted both difficulty and the number of trials based on
performance. Specifically, in the first practice session, the three tasks were randomly
scheduled: each task was presented thirty times, and the difficulty for each task was
adaptively updated based on performance. At each trial, the instructions (ready, go, back,
feedback) were shown sequentially on a computer screen along with corresponding audio
signals. Feedback was displayed as "failure" or "success". For the doorknob turning task,
the trial is a success, if the subject turns the doorknob more than a target angle. For the jar
closing-opening task, the trial is a success, if the subject closes the jar lid more than a
target angle and opens it again. For the doorbell pushing task, the trial is a success, if the
pushing force is higher than a threshold value.
Here, we explain how to compute the two adaptive components of high-level
adaptive task scheduler: the number of trials for the task schedule and the difficulty of the
task at each trial. The algorithm used to adaptively determine the number of trials is
based on a performance test following training on the previous session and on a
performance test preceding the current session (Y. Choi, Qi, Gordon, & Schweighofer,

７１
2008). We thus gave one post-test session of 30 trials (10 trials per task) immediately
after first training session (Figure 6.1.A) and, one later hour, another 30 trial test. The
number of trials for each task in the second practice session is determined with
 


i
i pos s i pre s
pos s pre s
tsk PE tsk PE
tsk PE tsk PE
tsk NumOfTrial
) ( * ) (
) ( * ) (
) (
, 1 ,
, 1 ,

(6.1)
where PE
s,pre
(tsk) is the performance error for the task tsk in the pre-test session of the
second practice session, and PE
s-1,pos
(tsk) is the performance error for the task tsk in the
post-test session of the first practice session. In this experiment, we assumed that the
performance error is the number of failure trials at the specific difficulty in the test
session. 10 and 0 are the maximum for the performance error and the minimum each.
Thus, the number of trials for each task is proportional to the multiplication of the
performance errors at the two test sessions.
We now discuss how difficulty is automatically adjusted based on patient’s
performance in the high-level adaptive task scheduler. The difficulty update function is
similar to (Y . Choi, et al., 2008) and given by
)) ( ) ( ( 1 ( ) ( ) (
1
tsk P tsk P tsk D tsk D
ref t t t
   

 (6.2)
where D
t
(tsk) is the difficulty for the current trial t and the task tsk, α is learning rate,
P
t
(tsk) is the performance, and P
ref
is the reference challenging performance.

6.2.2. Results
The number of trials for each was adapted based on the subject’s performance using
the algorithm of Equation (6.1). The number of trials in the second session was 56 for the

７２
more difficult jar closing-opening task, 25 for the doorknob turning task, and 9 for the
easier doorbell pushing task. Figure 6.1.B illustrates the randomly distributed schedule
with adjusted number of trials for each task.
The results for the adaptive difficulty algorithm are demonstrated in Figure 6.1.CDE.
In doorknob turning (Figure 6.1.C) and jar closing-opening (Figure 6.1.D)), the
performance is measured by the range of motion (angle), that is, how far the subject
turned the knob or the jar in a trial. The performance in doorbell pushing (Figure 6.1.E) is
measured by the force exerted on the doorbell. Note that in future work, the reference
challenging performance will be set after pilot testing. Here, we simply took values close
to maximum possible performance that the subject could exert in the three tasks. Thus,
the subject had to learn how to increase the force exerted in the three tasks. Here, the
difficulty, D = 1 means that the simulated dynamics by ADAPT reflects that of the real
passive tool. For doorbell pushing, our subject exerted typical pushing forces of around
15 N, which we used as threshold force for D = 1. Note that the update rate α must be
negative for doorbell pushing, and positive for the other two tasks.
In sum, the functionalities and adaptive difficulty of the high-level adaptive task
scheduler were successfully implemented. The adaptive algorithms worked similarly to
the visuo-motor task experiment in Chapter 3.
All the operations of ADAPT including the tool changing process for 180 task trials
worked without any problem. The preliminary experiment with a healthy subject does not
validate the efficacy of ADAPT for the rehabilitation of stroke patients. Therefore, after
fully testing ADAPT with healthy subjects, we will begin pilot studies to test the safety
and feasibility of our system with stroke patients.

７３
0 5 10 15 20 25 30
0
50
100
150
200
Force [N]
trial
0 5 10 15 20 25 30
0
2
4
6
8
Diffculty
Bell
Difficulty
Force
Reference
0 5 10 15 20 25 30
0
0.25
0.5
0.75
1
Angle [rad]
trial
0 5 10 15 20 25 30
0
1
2
3
4
Diffculty
Jar
Difficulty
Angle
Reference
0 5 10 15 20 25 30
1
1.5
2
2.5
3
Angle [rad]
trial
0 5 10 15 20 25 30
0
5
10
15
20
Diffculty
Knob
Difficulty
Angle
Reference

Figure 6.1. Illustration of the function of the high-level adaptive task scheduler. (A)
Random task schedule of the first practice session with three tasks (jar closing-opening,
doorknob turning, doorbell pushing) with thirty trials per each task. (B) Task schedule of
the second practice session with adaptive number of trials for each task. The more
difficult task (jar closing-opening) was allocated more trials than the others tasks. The
difficulty of doorknob turning (C) and (D) jar closing-opening was adapted such that the
performance (angle) remained near the challenging point (reference). (E) Adaptive task
difficulty of doorbell pushing in the first training session. In this case, the performance
(force) increased as the difficulty increased.

７４
6.3. The experiment with participants with chronic stroke
The primary aim of this experiment is to establish the feasibility of ADAPT for the
training of participants post stroke. We evaluated safety, overall functionality of ADAPT
for the participants with stroke, fidelity of simulated tasks and patient acceptance. In
addition, we tested how adaptive difficulty algorithm from (Y . Choi, et al., 2008) could be
used for modulation of task difficulty in the training with ADAPT.

6.3.1. Method
6.3.1.1. Participants
Four males Four males and one female (age 61.5 ± 3.4 years) with chronic stroke
(time since stroke 6.3 ± 2.3 years) participate in the study, which was approved by the
IRB at the University of Southern California. All participants received written and verbal
information about the study, and signed an informed consent form. All participants had a
single episode of stroke at least 9 months ago, and could be classified to have “moderate
to mild” impairments. The average baseline Fugl Meyer (FM) score was 49.2 ± 5.6, in the
range 42 to 56, and all participants could produce 90 degree of shoulder flexion and 30
degree of elbow extension, as required to perform the easiest task in the study.
Participants were excluded if they have had more than one stroke, have serious
uncontrolled medical conditions, show excessive pain in any joint of the more affected
extremity that could influence participation in the tasks, have a score of less than 25 on
the Folstein Mini Mental State Examination, have a score less than 40 on the upper
extremity Fugl-Meyer, do not demonstrate a thorough understanding of the instructions,
or could not perform at least four tasks at their lowest difficulty in the pre-test with

７５
ADAPT (see below)

6.3.1.2. Rehabilitation Robot
ADAPT with the configuration of a 3-DOF wrist mounted on a 1-DOF linear
actuator presents a 1-DOF functional task at a desired position after picking up the tool
corresponding to the selected task as described in Chapter 2. Four task tools (door knob,
door bell, jar, screw driver) for all six tasks are arranged in a tool rack, from which
ADAPT picks up a tool for a selected functional task. These tasks with different tools
require different grasping and arm motion (Table 6.1). The high-level adaptive task
scheduler aims to maximize the relearning of multiple functional tasks and balancing
learning among tasks in a limited training time; at each trial, it selects both the task to
practice within a task bank and the task difficulty based on the patient’s previous
performance, input from the therapist, and constraints given by the adaptive schedules.
The high-level adaptive task scheduler also sends a command to the tool-changing
system that picks up the tool corresponding to the selected task. The functional task
model generates a desired trajectory to simulate the dynamics of the task with the
difficulty specified by the high-level adaptive task scheduler. The low-level admittance
controller computes the control current corresponding to the desired trajectory. The
computed control current is then applied to a general-purpose robot to simulate the task,
and the patient feels the simulated dynamics by reaching and grasping the selected tool.
The safety surveillance module continuously monitors the status of the robot and the
controllers in the ADAPT system and provides appropriate actions to minimize any risk
to patients.

７６
6.3.1.3. Functional tasks
ADAPT tasks are based on activities of daily living, and at present we have
developed six functional tasks (doorknob turning, door opening, door locking, door bell
pushing, jar opening, screw driving) that require reaching and manipulation (Figure 2.4,
Table 6.1). With the novel modelling method of Chapter 4 for the dynamics of the
functional tasks, the feel or difficulty of the tasks by ADAPT is adjustable to the
individual and (presumably) improving performance. If we reformulate the method for
the study of this experiment, it can express the dynamics of a functional task as
(, , ) Diff f q q q    
(6.3)
where  is the torque exerted by a subject, Diff is the adjustable difficulty, q , q  , q 
are angle, velocity, acceleration respectively of task motion, and (, , ) f qqq  is the
dynamics model of an original functional task constructed by the combination of multiple
local linear models (Y Choi, Gordon, Kim, & Schweighofer, 2009). The dynamics
models for all six tasks are programmed as the functional task model in the control
architecture (Figure 2.2), and ADAPT can adaptively change the task difficulty by
controlling the Diff in equation (6.3). In addition to generating the captured dynamics
with the novel model, presenting real functional tools as an end-effecting handle provides
genuine grasping feeling that is difficult in universal haptic devices.

6.3.1.4. Adaptive task scheduler
The high-level adaptive task scheduler in Figure 2.2 determines the task to practice
and the difficulty of the selected task to enhance relearning of multiple functional tasks. It

７７
makes ADAPT an ideal platform for delivering various training schedules based on motor
learning principles and rehabilitation guidelines. In this study, we only used the adaptive
difficulty algorithm in the high-level adaptive scheduler. The update function for the
adaptive algorithm is similar to Chapter 3 and is given by
,1, , ,
(1 ( ))
tk t k tk ref k
Diff Diff Perf Perf 

 
(6.3)
where Diff
t,k
is the difficulty for the current trial t and the task k, α is learning rate, Perf
t,k

is the performance, and Perf
ref,k
is the reference challenging performance. The difficulty
in Equation (6.3) is updated to constantly maintain performance around a “challenging
point”, Perf
ref,k
. For example, if a task k at a current trial t is so easy that the performance
Perf
t,k
is over Perf
ref,k
, then the difficulty Diff
t,k
will be increased progressively and
adaptively until the performance Perf
t,k
is below Perf
ref,k
. If the performance Perf
t,k
is far
below Perf
ref,k
, then the task at the current trial is too difficult, and so the difficulty will be
decreased adaptively again by the above update function. Although the current ADAPT
system allows a variety of different adaptive schedule as in (Y. Choi, et al., 2008), here
we used a pseudo-random schedule for 12 three-trial blocks of 4 tasks with the adaptive
difficulty algorithm.

6.3.1.5. Experimental procedure
Participants sat on a chair in front of ADAPT and was instructed to reach and
manipulate a functional task tool, which was selected by the adaptive scheduler, picked
up and presented by the tool changer of ADAPT. The experiment started with a pre-test
followed by approximately one and half hour training session, and ended with a pos-ttest

７８
after one hour break. In pre-test and post-test, participants practiced all six tasks in each
block of three trials (totally 18 trials) at easiest difficulty without feedback. Then, four
tasks were selected randomly from the six tasks of pre-test for the practice in the training
session. If a participant could not complete a specific task, the task was not selected. One
participant who could not complete more than two tasks in pre-test, was excluded. In the
training session, the participants had to practice the selected four tasks in adaptive
difficulty mode, in which the task difficulty was updated by the Equation (6.3). The
difficulty was set to be lowest initially, and supposed to increase progressively. We could
increase quickly the difficulty with high learning rate in the Equation (6.3) so that the
performance leads suddenly to the challenging point. However, the initial easiest
difficulty means the high success probability of initial trials, which is important for
motivation. Furthermore, the gradual and progressive adaptation promtes longer-lasting
effect (van Asseldonk, Wessels, Stienen, van der Helm, & van der Kooij, 2009). The
training session consisted of 48 blocks of three trials (12 blocks per task), which are
pseudo-randomly scheduled so that each task occurred once in any four consecutive
blocks.
At the beginning of each trial, we instructed the participants to put their affected
hand on the home position (Figure 2.4), and an auditory instruction corresponding to a
specific functional task together with an instruction on screen signalled the start of the
selected functional task (Figure 6.2). 11 seconds were allowed for the completion of one
trial, and then the participants received an auditory feedback for success or failure.
After every block of three trials for each task, the feedback of the participants’ progress
for all practiced trials was displayed (Figure 6.2). The next trial was not presented, until

７９
the participants’ hand was back on the home position after the feedback display allowing
for some rest time in case of being fatigued. After all the sessions, two questionnaires
were administered to evaluate the acceptance of ADAPT for patients post stroke. Firstly,
we used Intrinsic Motivation Inventory (IMI) questionnaire that is designed to assess the
participants’ subjective experience related to a target activity in laboratory experiments
(Deci & Ryan, 1985; Ryan & Deci, 2000). IMI has been successfully used to measure
stroke patients’ experience in robotic training (Colombo et al., 2007), and we measured
six categories about our participants’ experience in the training with ADAPT (Table 6.2).
Secondly, we designed more specific questions about the acceptance of ADAPT
including safety, fidelity of simulated tasks, and system utility (Table 6.3). We designed
this questionnaire based on other robot acceptance studies such as in (M. Dijkers et al.,
1991; H. Krebs, Hogan, Aisen, & V olpe, 1998).

Figure 6.2. Experiment procedure with ADAPT. Instruction is displayed together with
human voice reading of the instruction. Feedback of the progress for each task is
displayed only after a block of three trials for each task.

８０

Task Description Performance
metric
Tool
Doorknob turning Turn a door knob with power grasping
to the end of turning range, and release
it.
Angle Knob
Doorbell pushing Push a door bell with a finger over a
threshold force.
Force Doorbell
Jar opening Turn a jar with power grasping up to
the end of turning range.
Angle Jar
Door opening Move a door knob horizontally with
power grasping. Mostly elbow and
shoulder motion.
Angle Knob
Door locking Turn a door knob button with pinch
grasping to the end of turning range.
Angle Knob
Screwdriving Turn a screwdriver with power
grasping to the end of turning range.
Angle Screw
driver

Table 6.1. Functional task list

８１

Subscale Score (Mean±SD)
Interest/Enjoyment 6.17 ± 1.25
Perceived Competence 4.90 ± 1.88
Effort/Importance 6.32 ± 0.75
Value/Usefulness 6.34 ± 1.43
Pressure/Tension 1.84 ± 0.94
Perceived Choice 6.43 ± 1.40

Table 6.2. Subscale findings of the IMI questionnaire administered after training.
(subscale range = 1 – 7)

Question Score (Mean±SD)
Comfortable with sound &
appearance of ADAPT
6.90 ± 0.32
Safety 6.80 ± 0.45
Comfortable with seatbelt &
chair

6.10 ± 1.52
Similar to real functional tasks 5.90 ± 1.62
Fatigue or Frustration 3.40 ± 1.95
Clear instructions from ADAPT 7.00 ± 0.00

Table 6.3. Patients’ acceptance for the training session with ADAPT. (subscale
range = 1 – 7)

８２
6.3.2. Results
All participants completed the presented sessions, functional measurements, and
questionnaires. No adverse event occurred and no adverse response to the training with
ADAPT was reported. In general, ADAPT could lead all participants to train in the near
challenging point by progressively increasing task difficulty with help of adaptive task
difficulty algorithm and progressive feedback report, even though tasks are repetitively
presented.

6.3.2.1. Functional task trajectory
To show the physical properties of functional tasks and how difficulty changing
affects the dynamics of the tasks, we plotted the torque versus angle trajectories for three
representative tasks in Figure 6.3. The difficulties were modulated based on participants’
performance using Equation (6.4), and ADAPT simulated the task dynamics
corresponding to the modulated difficulties using Equation (6.2). The sample trajectories
at two different levels of task difficulty were selected from trials in the training session of
one participant (FM score: 45). Door locking task requires spring-like torque, which
increases proportionally to rotated angle from the initial point until it passes the lock
clicking point (Figure 6.3.A). If the angle is over the lock clicking point, rotation for door
locking can be easily completed by opposite directional spring-like torque, which is
proportional to the angle from the end point. ADAPT maintains these physical properties
while modulating difficulties by scaling Diff of Equation (6.2). Doorknob turning task
also has the property of a typical spring whose torque is proportional to the rotated angle,

８３
and jar opening task mainly consists of static friction torque which serves as an initial
large threshold and small dynamic friction torque during rotation (Figure 6.3.BC). Other
realistic properties such as position-dependent stiffness captured from real task tools were
also implemented by our previously proposed method to increase the fidelity of the
simulated tasks (Y Choi, et al., 2009). The difficulty of the other three tasks was also
manipulated exactly in the same method. Door bell pushing task was implemented by
playing a bell sound when the pushing force reached a specific threshold in limited time.
Door opening task’s dynamics was assumed to be a pure damping, and this task used a
different DOF of the ADAPT robot, which required movement from left to right in the
plane horizontal to Earth’s surface. Screw driving task was implemented with the friction
dynamics similar to jar opening task. In general, ADAPT could present successfully
realistic functional tasks of different difficulties, and most participants reported that the
simulated tasks felt similar to real tasks. More details about the participants’ response are
discussed in later section.

８４
-20 0 20 40 60 80 100 120 140
-1.8
-1.4
-1.0
-0.6
-0.2
0.2
Angle (deg)
Torque (Nm)
Doorknob turning trajectory
Starting
angle
open
release
A
lock
-10 0 10 20 30 40 50
-0.5
0
0.5
1
1.5
2
2.5
3
Angle (deg)
Torque (Nm)
Jar opening trajectory
Easy
Difficult
Starting
angle
open
B
0 10 20 30 40 50 60
-0.05
0
0.05
0.1
0.15
0.2
Angle (deg)
Torque (Nm)
Door locking trajectory
Door lock clicking
angle
Starting
angle
C

Figure 6.3. Sample trajectories of functional tasks. Torque versus position trajectories for
three functional tasks in the training session. Blue solid line is the trajectory of each
task at low difficulty, which is presented at early trial of the training session. Red dot line
is the trajectory of each task at high difficulty, which is presented near the end of the
training session.

８５
6.3.2.2. Performance adaptation
The adaptive difficulty schedule in the training session aims to enhance the learning
effectiveness of each task by maintaining participants’ performance near the challenging
performance, where the tasks are not too easy and not too difficult. Figure 6.4 shows
examples of performance and adaptive difficulty in the training session of a participant
(FM score: 45). In Figure 6.4A, the algorithm progressively increased the task difficulty
based on initial successful performance, and adaptively responded to participant’s
performance by increasing or decreasing the difficulty by the right amount of change to
maintain the performance near the challenging point. In case of Figure 6.4B, the
algorithm also adaptively modulated the difficulty based on performance, and at near the
maximum difficulty, the participant started to fail. After several trials of failure and
success with the modulation of the difficulty, the participant could succeed at later trials,
meaning that the performance increased at near the maximum difficulty. . In Figure 6.4C,
the participant succeeded all 36 trials of door locking task. Therefore, the initially lowest
difficulty kept increasing adaptively according to the adaptive difficulty algorithm
Equation (6.4) until it reached the maximum limit, as current performance is always
better than the challenging performance

８６
0 10 20 30
0
5
10
Trial
Difficulty
10 20 30
0
30
60
90
Performance (deg)
Door locking training progress
Starting Angle
Challenging
Performance
0 10 20 30
0
5
10
Difficulty
Jar opening training progress
0 10 20 30
-20
10
40
70
100
Trial
Performance (deg)
Difficulty
Performance
Starting Angle
Challenging
Performance
0 10 20 30
0
5
10
Difficulty
0 10 20 30
0
100
200
Trial
Performance (deg)
Doorknob turning training progress
Challenging
Performance
Starting Angle
B
C
A

Figure 6.4. Progress of Performance & Difficulty. Progress of performance and difficulty
is plotted for a representative participant (FM score: 45) from the training session.
Difficulty is expected to be adapted such that the performance remains near the
challenging performance. Initial difficulty is set at lowest value, and increases depending
on performance. In case of the doorknob turning (A), the participant performance was
maintained near the challenging point with the adaptive difficulty. Door locking task (C)
is so easy for the participant that the difficulty is maxed out after being increased. In jar
opening task (B), the participant failed near the highest difficulty, and he could learn how
to complete trials near the end after several failure trials.

８７
6.3.2.3. Movement time improvement
To analyze the movement time in pre and post test sessions, we first computed the
median for each task, and then performed paired sample t-test to analyze the effect of
training. We report results as means ± standard deviations. Our significance level was p
< .05. While all six tasks were given in pre and post sessions, only four tasks out of six
tasks were practiced during the training session. Therefore, 20 samples out of 30 samples
(five participants by six tasks) were the movement times of trained tasks, and 10 samples
were the movement times of untrained tasks in each of pre and post sessions. Figure 6.5A
demonstrates movement times for trained and untrained tasks in pre and post sessions. In
both cases, the movement time decreased from pre to post: trained tasks (pre: 5.18 ±
2.040, post: 4.12 ± 1.004), non-trained tasks (4.17 ± 0.836, 3.96 ± 0.741). The paired t-
test analysis produced a statistically significant difference (pre-post: 1.06 ± 1.916, p =
0.022) for trained tasks, but non-trained tasks did not show a significant difference (pre-
post: 0.21 ± 0.370, p = 0.105). Figure 6.5B demonstrates movement times for each of six
tasks in pre and post sessions. For all six tasks, the movement time decreased from pre to
post (Figure 6.5B): doorknob turning, (pre: 6.87 ± 2.556, post: 5.30 ± 0.393), doorbell
pushing (4.40 ± 0.401, 2.82 ± 1.564), jar opening (3.89 ± 0.756, 3.64 ± 0.468), door
opening (6.50 ± 3.899, 3.80 ± 0.046), door locking (4.30 ± 0.448, 3.96 ± 0.027),
screwdriving (4.74 ± 0.451, 4.63 ± 0.480).

８８
0
2
4
6
8
10
movement time (sec)
Knob turning
Movement time change
Bell pushing
Jar opening
Door opening
Door locking
Screwdriving
Pretest
Posttest
0
2
4
6
8
10
Movement time (sec)
Trained tasks
Movement time change
Non-trained tasks
Pretest
Posttest

A
B

Figure 6.5. Movement times in pre and post test sessions. Movement times were
measured in pre and post test session and were plotted with mean and standard deviation.
In (A), we compared movement times between pre-test and post-test to measure the effect
of training session. In (B), we compared movement time between pre-test and post-test
for each task.

８９
6.3.2.4. Acceptance of ADAPT
The high score of the interest/enjoyment of IMI implies that our participants found
the training session with ADAPT very interesting (Table 6.2). Relatively low score of the
perceived/competence might be due to the different levels of disability of our participants,
corresponding to the results from the previous study (Colombo, et al., 2007). The high
score of effort/importance and value/usefulness subscales indicates that participants were
highly motivated and satisfied with the experience and the result from the training with
ADAPT. The low score of pressure/tension subscale means that participants did not feel
much pressure or tension during the training with ADAPT
Results from the second questionnaire also, in general, suggest that the training with
ADAPT was well accepted and tolerated by participants. While there were a few
complaints about the seatbelt and the chair, participants mostly felt comfortable
interacting with ADAPT. Safety concerns were strongly addressed from the initial design
process of ADAPT, and no adverse event occurred during direct interaction between
participants and ADAPT. The questionnaire result showed that participants felt very safe
and they could clearly understand how to interact with ADAPT through auditory and
visual instructions. Participant reported that the simulated functional tasks were fairly
similar to real tasks. One participant verbally reported that pushing bell required too
much force, suggesting that a task trial with high difficulty could be unrealistic.
Because task difficulty was mostly related to torque, later trials with high difficulty made
participant exert high torque and feel fatigue.

９０
6.3.3. Discussion
In this study, we tested and evaluated the feasibility of ADAPT to five participants
with chronic stroke during one day training session. The main purpose of our study was
to test whether ADAPT was feasible to rehabilitation training for patients post stroke.
Specifically, we evaluated safety, overall functionality of ADAPT, fidelity of simulated
tasks and patient acceptance. Furthermore, we showed how adaptive practice schedules
(Y. Choi, et al., 2008) could be used for modulation of task difficulty in the training with
ADAPT, and we could observe potential effectiveness for clinical use from the results of
movement time analysis.
Although the feasibility of ADAPT was well established from the results of this
study, there are several limitations or weaknesses to be discussed or be addressed in
future work. First, the difficulty of current tasks was mostly proportional to torque or
force to be exerted for manipulation in ADAPT. Even though motor strengthening is
beneficial for stroke recovery (Ada, Dorsch, & Canning, 2006), there are many functional
tasks, which do not require much motor strength but fine motor coordination such as
radio tuning or threading. This strength-oriented difficulty algorithm also caused the task
of high difficulty to be felt less similar to the real functional task, as some participants
verbally reported. Second, we could observe from post recorded-video analysis that some
task tools were not manipulated as instructed, even though participants did know the
correct way of manipulation. In difficult trials of door locking task which requires pinch-
grasping, one participant (FM score: 42) happened to use power grasping, which could be
seen as compensatory movements to succeed in the trial. Therefore, tools which can
constrain hand or finger postures will be helpful in administering more task-specific

９１
training. Third, while seat-belting enhanced the safety feature, some participants reported
that it made reaching movement uncomfortable. Sometimes, too tight seat-belting
prevented participants from any compensatory behaviour, so not being able to be even
close to success of the task. It has been argued that stroke recovery therapy does not have
to focus solely on perfect movement of normal healthy (G Kwakkel, 2009), but should
take some compensatory strategy into account to be more effective. Therefore, we need
more research on how to trade off between safety and compensation in ADAPT.
While we previously showed that adaptive difficulty algorithms outperform fixed
difficulty in motor learning (Y. Choi, et al., 2008), it was not trivial to apply those
algorithms to stroke rehabilitation, because it’s hard to compare the performance between
different tasks, and challenging point might be different for each participant. There have
been several attempts to make robotic training adaptive. In reaching tasks with assistance
of MIT-MANUS robot, speeed, time or EMG were modulated adaptively based on
patient’s performance (H. Krebs, Palazzolo, Dipietro, Volpe, & Hogan, 2003).
Optimization theory was used to derive adaptive update algorithm of assistance force by
minimizing a weighted error sum from desired trajectory and assistance force (Emken,
Benitez, & Reinkensmeyer, 2007). However, those algorithms could be applied to only
limited range of tasks, and it is hard to apply them to our functional tasks, which have
different physical properties. One advantage of our adaptive algorithm is that difficulty
can be used irrespective of task as it is a ratio between the simulated task dynamics and
original task dynamics as shown in Equation (1). Therefore, it does not have to be a
physical quantity such as movement time or assistance force as in (H. Krebs, et al., 2003).
However, one limitation of our algorithm is that the performance value is a physical

９２
quantity such as rotated angle or pushing force. It also makes not easy to compare
performances between different tasks, although such comparison is necessary to select a
task to practice from a specific adaptive task schedule. Success rate as performance
metric is one possible solution to this problem, and we are underway to develop statistical
models of performance for each task of each subject. Specifically, Bayesian logistic
regression models (Bishop, 2006) that receive a difficulty value as an input and generates
expected probability of success as an output. The model will be trained with success data:
if the user can complete a task trial within limited time at a difficulty value, the difficulty
will be assigned 1; if not, the difficulty will be assigned 0. If we set the challenging
performance as 80 percent, then the above model for each task of each participant can
predict the difficulty that will lead the participant to perform at the challenging
performance, the success rate of 80 percent. Then, our adaptive difficulty algorithm can
more readily modulate task difficulty toward the difficulty corresponding to the
challenging performance from the model, even though the participant’s skill level for a
specific task has been changed due to training. More details about the performance model
will be introduced in our next work. In the present study, we empirically determined the
challenge performance, but future work should focus on how to determine this meta-
parameter.

6.3.4. Conclusion
The results from this study validate the feasibility of ADAPT for rehabilitation of
arm and hand function after stroke, and provide justification for continued investigation
of clinical efficacy. ADAPT provided adaptive training tailored to patients’ performance

９３
by modulating task difficulty. Safe automatic presentation of functional tasks with
ADAPT showed the potential to engage any motor learning schedule in stroke
rehabilitation. This design feature of ADAPT sheds light on the future work of
developing hypothesis-driven adaptive schedules which choose the right functional task
and difficulty for optimal stroke rehabilitation. Next steps of this study will be to test the
efficacy of ADAPT with a simple random schedule in a Phase I randomized clinical trials
(RCT), and to enhance the efficacy with adaptive schedules based on motor learning
models, which predicts long-term retention from the latest training information.

９４
Chapter 7. Contributions and Future Work

7.1. Contributions
This thesis describes the design and development of the robotic rehabilitation system,
ADAPT (ADaptive and Automatic Presentation of Tasks) for enhancing the recovery of
upper extremity functions in patients with stroke. The primary contributions from the
thesis work are described in the following sections.

7.1.1. Design of ADAPT for stroke rehabilitation
We presented a novel robotic task practice system, ADAPT (ADaptive and
Automatic Presentation of Tasks), designed to enhance the recovery of upper extremity
functions in patients with stroke. We designed ADAPT in accordance to current training
principles for stroke rehabilitation: ADAPT engages the patient intensively, actively, and
adaptively in a variety of realistic functional tasks that require reaching and manipulation.
A general-purpose robot simulates the dynamics of the functional tasks, and presents
these functional tasks to the patient. A novel tool changing system enables ADAPT to
automatically switch between the tools corresponding to the functional tasks. The control
architecture of ADAPT is composed of three main components, a high-level task
scheduler, a functional task model, and a low-level admittance controller. Thanks to our
novel functional task model and low-level admittance controller, the robot precisely
simulated the desired task dynamics of a door knob and jar with the difficulties specified
by the high-level adaptive task scheduler. We believe that the design of ADAPT is not

９５
only oriented by the latest principles of stroke rehabilitation, but also is optimized to test
recent and upcoming motor learning theories.

7.1.2. A novel modeling method for adaptive difficulty of functional task
We presented a novel modeling method for functional tasks such that the feel or
difficulty of the task tools is adjustable. The functional task model in ADAPT design
generates desired trajectories based on learned models of task dynamics. Tasks dynamics
are modeled with Receptive Field Weighted Regression (RFWR), such that the feel of the
task tools are accurately modeled and the task difficulty can be easily adjusted. The
captured task model is then embedded as a functional task model in the ADAPT control
architecture to generate the desired trajectory for the selected task with specified
difficulty. State-dependent nonlinear dynamics (e.g. position-dependent stiffness or
damping) of passive tools can be modeled by the combination of multiple local linear
models (Colton & Hollerbach, 2005). We formulated the modeling of the functional tasks
as locally weighted regression problem with Receptive Field Weighted Regression
(RFWR) (Schaal & Atkeson, 1998). RFWR has two main advantages. First, compared to
the method used in (Colton & Hollerbach, 2005), it provides a computationally efficient
way of modeling position-dependent dynamics, because it adaptively creates and prunes
the local models. Second, as only few local linear models are created, it allows good
generalization.

９６
7.1.3. A novel admittance controller with a force model and a motion model
We developed a novel controller based on the estimation of two separate feed-
forward modules: the first for motion without interaction force (motion model) and the
second for interaction force without motion (force model). As the low-level admittance
controller in the ADAPT control architecture, the outputs of these two modules are
combined to compute the desired control signal (current) for the desired motion with the
measured interaction force. To estimate both modules, we used Locally Weighted
Regression (LWR) (Schaal & Atkeson, 1998), a machine learning technique that has been
successfully tested in the approximation of the inverse dynamics of robots [21]. Our
computational approach is practical and useful when, as in our system, the dynamics
model of the haptic device is not known and the control current is not proportional to the
motor torque. Moreover, the controller can implement virtual walls because the force
model module is estimated based on the assumption of no motion. Implementation of a
doorknob task with a virtual wall demonstrates the feasibility of our approach in realistic
simulating the dynamics of functional tasks. The acceptance of the task fidelity simulated
by ADAPT to healthy subjects and participants with chronic stroke was successful as
shown in Chapter 6.

7.1.4. Performance based adaptive schedule for motor learning
In the design of ADAPT, the high-level adaptive task scheduler is a crucial
component for the stroke rehabilitation by ADAPT to be more effective than others. The
high-level adaptive task scheduler has two adaptive components: the first determines the
task schedules, the second sets the difficulty of the task selected at each trial. In the

９７
process of finding the most effective practice schedule for stroke rehabilitation, we tested
two adaptive algorithms that adaptively determine the nominal difficulty and the number
of trials for each task, based on both current and delayed performance of the learner. We
tested these adaptive algorithms in a two-by-two factorial design, and show that these
algorithms outperform random scheduling, when performance is measured on a delayed
retention test. Although a number of other adaptive task scheduling algorithms can be
used in ADAPT, we use this algorithm in our current high-level adaptive task scheduler.

7.1.5. A feasibility test of ADAPT with participants with chronic stroke
We tested and evaluated the feasibility of ADAPT to five participants with chronic
stroke during one day training session. The main purpose was to test whether ADAPT
was feasible to rehabilitation training for patients post stroke. Specifically, we evaluated
safety, overall functionality of ADAPT, fidelity of simulated tasks and patient acceptance.
All participants completed the presented sessions, functional measurements, and
questionnaires. No adverse event occurred and no adverse response to the training with
ADAPT was reported. Furthermore, we showed how adaptive practice schedules (Y . Choi,
et al., 2008) could be used for modulation of task difficulty in the training with ADAPT,
and we could observe potential effectiveness for clinical use from the results of
movement time analysis.

7.2. Future work
Possible directions and topics for future work in the research areas associated with
this thesis are described in the sections below

９８
7.2.1. Improvement of ADAPT design
At this stage, we have implemented only six unimanual tasks (doorknob turning, jar
opening, doorbell pushing, door opening, door locking, screw driving). Using the current
configuration of a 3-DOF wrist mounted on a 1-DOF, we will in future versions develop
additional tasks that require active manipulation of concrete objects (e.g., turning a key or
door handle, steering, turning a water faucet, lifting jugs, etc.), and will allow a larger
number of possible types of grasps (e.g. overhand, precision, lateral pinch, power) and
individual finger motions. We will also develop bimanual tasks, as a growing number of
studies have shown the potential of bilateral training on the recovery of the paretic limb
post stroke (Cauraugh & Kim, 2002; Whitall, Waller, Silver, & Macko, 2000). The
difficulty of current tasks was mostly proportional to torque or force to be exerted for
manipulation in ADAPT. Even though motor strengthening is beneficial for stroke
recovery (Ada, et al., 2006), there are many functional tasks, which do not require much
motor strength but fine motor coordination such as radio tuning or threading. Because
those tasks are also largely necessary in daily living activities, adaptively practice them in
the frame of ADAPT will help relearning the fine motor coordination for upper
extremities. In the feasibility experiment with participants with chronic stroke, some task
tools were not manipulated as instructed, even though participants did know the correct
way of manipulation. At high difficulty of door locking task which requires pinch-
grasping, one participant happened to use power grasping, which could be seen as
compensatory movements to succeed in the trial. Therefore, tools which can constrain
hand or finger postures will be helpful in administering more task-specific training.
In the current version, a seat belt is used for safety reason, but also to prevent

９９
compensatory trunk movements, which are known to interfere with recovery of reaching
movements (Cirstea, Leduc, & Levin, 1998). However, even with the belt, patients may
still be able to generate abnormal movements, such as twisting at the elbow for tasks that
require twisting at the wrist, or rotating the back for tasks that require rotation at the
shoulder. If future testing shows that these types of compensatory movements are
important, we will track shoulder, elbow, and wrist movements with motion tracking
sensors to identify, and then try to correct by providing feedback, these abnormal
movement patterns.
While the realism of functional task is a crucial factor in task-oriented training for
stroke rehabilitation, training on the robot should also be motivating. In this respect, our
system can be largely improved. For instance, performing the current tasks even with
adaptive difficulty was not motivating to some subjects. Developing game-based task
trainings will be one possible solution,
Safety concerns were strongly addressed in the initial design process of ADAPT, and
the subject did not feel threatened from the robot movement. However, to maximize
safety, we need to perform long-term, intensive, and systematic tests with a greater
number of subjects, both healthy and with stroke.

7.2.2. Modeling the task dynamics and haptic controller design
Although we successfully modeled the task dynamics with help of Karnopp friction
model and RFWR, The true test of the model’s performace should be based on
psychophysical user testing. We only provided partially it in the feasibility test of ADAPT.
Another issue is that the quality of the modeling of these task dynamics depends on the

１００
performance of haptic rendering devices. As our robot is not built for the purpose of
haptic rendering devices, the psychophysical user test might not be feasible.We need to
test the modeling in pure haptic devices, and compare it with other modeling methods. In
the current version of the modeling method, we assumed that the one dimensional torque
is the only external force or torque. For example, the normal pushing force of the screw
driving task was not considered to be an external force. However, if the normal pushing
force of the screw driving task is high, the friction torque will increase largely. Two
dimensional friction will, furthermore, allows for more realistic modeling of functional
tasks, which have different frictional properties in different directions. It would greatly
improve the realism of our upcoming new functional tasks.
Even though our haptic controller, the low-level admittance controller in the ADAPT
control architecture showed acceptable task fidelity, it does not mean that our robot
outperforms other commercially available haptic interfaces. Indeed, most commercial
haptic interfaces have special hardware structures that are designed for specific human-
robot applications. Thus, it is required to further test experimentally the controller in the
environment of human interaction, or to develop a stability analysis for the range of
frequency that humans can generate during interaction, and study how the model behaves
when back-drivability varies.

7.2.3. Model-based adaptive schedules and Bayesian models of performance
Because our proposed adaptive schedules in the Chapter 5 relied on performance on
delayed retention tests however, the number of trials for each task could only be adjusted
in training sessions following delayed retention tests. If we could adapts the number of

１０１
trials before any delayed retention test data is available by using predictions of delayed
performance, the schedule will reflect more precise information of upcoming learning. To
predict the delayed performance, we need a motor learning model that can take care of
contextual interference effect for learning multiple tasks, can predict not only the current
level of learned skills but also how fast the learned skill will be forgotten. Even though
there have been several attempts to model human motor learnings, there is a long way to
go. The two-process memory model of (Smith et al. 2006), which contains both a fast-
learning and fast-forgetting module and a slow-learning and slow-forgetting module can
predict long-term performance with simple linear state space model. Lee et al (2009) also
showed that dual adaptation with contextual cues can be modeled by a parallel structure
of the fast and slow modules using the simple linear state space model. While there are
still several limitations in these models to be applied to the training of multiple functional
tasks for stroke rehabilitation, the adaptive tasks scheduling based on such neural models
of motor learning shed a light on how to optimize stroke rehabilitation. The model of Han
et al (2008) computes an optimal dose of therapy for stroke rehabilitation. For example, if
motor retraining after stroke takes the amount of therapy based on the model, then the
stroke patients will enter a virtuous circle in which spontaneous arm use and motor
performance reinforce each other. While it is only a simulation study based on
computational models, building such models based on training data of long term clinical
trials such as EXCITE (Winstein 200) will provide cost-effective rehabilitation methods
tailored to individuals poststroke.
One limitation of our algorithm is that the performance value is a physical quantity
such as rotated angle or pushing force. It also makes not easy to compare performances

１０２
between different tasks, although such comparison is necessary to select a task to practice
from a specific adaptive task schedule. Success rate as performance metric is one possible
solution to this problem, and we are underway to develop statistical models of
performance for each task of each subject. Specifically, Bayesian logistic regression
models (Bishop, 2006) that receive a difficulty value as an input and generates expected
probability of success as an output. The model will be trained with success data: if the
user can complete a task trial within limited time at a difficulty value, the difficulty will
be assigned 1; if not, the difficulty will be assigned 0. If we set the challenging
performance as 80 percent, then the above model for each task of each participant can
predict the difficulty that will lead the participant to perform at the challenging
performance, the success rate of 80 percent. Then, our adaptive difficulty algorithm can
more readily modulate task difficulty toward the difficulty corresponding to the
challenging performance from the model, even though the participant’s skill level for a
specific task has been changed due to training.

7.2.4. Randomized Clinical Trials
We successfully conducted a pilot study to test the safety and feasibility of our
system with stroke patients. The right next step will be to test the efficacy of ADAPT
with a simple random schedule in a Phase I randomized clinical trials (RCT). We will
compare pre- and post-training scores on a number of instruments, with a focus on those
measuring arm and hand function and use (e.g., the Wolf Motor Function Test (Wolf et al.,
2001) and the Motor Activity Log (Uswatte, Taub, Morris, Vignolo, & McCulloch,

１０３
2005)). Furthermore, we will test the possibility of using ADAPT as an automated tool
for measuring arm and hand function in patients with stroke. In parallel with these
clinical trials, we will use ADAPT to conduct experiments to determine highly effective
adaptive task schedules based on motor learning models of Section 7.2.3.

１０４
Bibliography

Ada, L., Dorsch, S., & Canning, C. (2006). Strengthening interventions increase strength
and improve activity after stroke: a systematic review. Aust J Physiother, 52(4), 241-248.
Allotta, B., Colla, V., & Bioli, G. (1999). A Mechatronic Device for Simulation Push-
buttons and Knobs. Paper presented at the IEEE International Conference on Multimedia
Computing and Systems.
AMTEC. http://www.amtec-robotics.com/produkte_en.html.
Angerilli, M., Frisoli, A., Salsedo, F., Marcheschi, S., & Bergamasco, M. (2001,
September 18-21 2001). Haptic simulation of an automotive manual gearshift. Paper
presented at the Proceedings of Roman2001, International Workshop on Robot-Human
Communication, Bordeaux-Paris, France.
Atkeson, C. G., Moore, A. W., & Schaal, S. (1997). Locally weighted learning. Artificial
Intelligence Review, 11(1-5), 11-73.
Ayres, A. J. (1972). Sensory integration and learning disorders. Los Angeles: Western
Psychological Services.
Bayona, N. A., Bitensky, J., Salter, K., & Teasell, R. (2005). The role of task-specific
training in rehabilitation therapies. Top Stroke Rehabil, 12(3), 58-65.
Bernstein, N. (1967). The coordination and regulation of movements. New York:
Pergamon. .
Bishop, C. M. (2006). Pattern recognition and machine learning: Springer.
Boyce, B. A., & Del Rey, P. (1990). Designing applied research in a naturalistic setting
using a contextual interference paradigm. Journal of Human Movement Studies, 18, 189-
200.
Burgar, C. G., Lum, P. S., Shor, P. C., & Machiel Van der Loos, H. F. (2000).
Development of robots for rehabilitation therapy: the Palo Alto VA/Stanford experience.
J Rehabil Res Dev, 37(6), 663-673.

１０５
Butefisch, C., Hummelsheim, H, Denzler, P, Mauritz, KH. (1995). Repetitive training of
isolated movements improves the outcome of motor rehabilitation of the centrally paretic
hand. J Neurol Sci, 130, 59-68.
Cahill, L., McGaugh, J. L., & Weinberger, N. M. (2001). The neurobiology of learning
and memory: some reminders to remember. Trends Neurosci, 24(10), 578-581.
Carel, C., Loubinoux, I., Boulanouar, K., Manelfe, C., Rascol, O., Celsis, P., et al. (2000).
Neural substrate for the effects of passive training on sensorimotor cortical representation:
a study with functional magnetic resonance imaging in healthy subjects. J Cereb Blood
Flow Metab, 20(3), 478-484.
Carlson, R. A., Sullivan, M. A., & Schneider, W. (1989). Practice and working memory
effects in building procedural skill. Journal of Experimental Psychology, Learning,
Memory, and Cognition, 15( ), 517–526.
Carlson, R. A., & Yaure, R. G. (1990). Practice schedules and the use of component skills
in problem solving. Journal of Experimental Psychology: Learning, Memory, and
Cognition , , 16, 484-496.
Carr, J. H., & Shepherd, R. B. (2003). Stroke Rehabilitation - Guidelines for Exercise
and Training to Optimize Motor Skill: Butterworth-Heinemann.
Catalano, J. F., & Kleiner, B. M. (1984). Distant transfer in coincident timing as a
function of variability of practice. Perceptual & Motor Skills, 58, 851-856., 58, 851-856.
Cauraugh, J. H., & Kim, S. (2002). Two coupled motor recovery protocols are better than
one - Electromyogram-triggered neuromuscular stimulation and bilateral movements.
Stroke, 33(6), 1589-1594.
Cavusoglu, M. C., Tendick, F., & Sastryand, S. (2001). Telesurgery and surgical
simulation: Haptic interfaces to real and virtual surgical environments. Paper presented
at the In: Touch in Virtual Environment, McLaughlin, M. L., Hespanha, J. P. and
Sukhatme, G. (Eds.) IMSC Series in Multimedia. Prentice Hall. .
Choi, Y., Gordon, J., Kim, D., & Schweighofer, N. (2009). An Adaptive Automated
Robotic Task-Practice System for Rehabilitation of Arm Functions After Stroke. IEEE
TRANSACTIONS ON ROBOTICS, 25(3), 556-568.

１０６
Choi, Y., Qi, F., Gordon, J., & Schweighofer, N. (2008). Performance-based adaptive
schedules enhance motor learning. J Mot Behav, 40(4), 273-280.
Choi, Y. G., & Schweighofer, N. (2008). ADAPT – Adaptive Automated Robotic Task
Practice System for Stroke Rehabilitation. Paper presented at the 2008 IEEE Int. Conf.
Robotics and Automation,.
Cirstea, M. C., Leduc, B., & Levin, M. (1998). Hemiparetic patients recruit additional
degrees of freedom to compensate lost motor function. . Paper presented at the
Proceedings of the Twelfth Congress of the International Society of Electrophysiology
and Kinesiology, Montreal, Quebec, Canada.
Clover, C. L. (1999). A control-system architecture for robots used to simulate dynamic
force and moment interaction between humans and virtual objects. Ieee Transactions on
Systems Man and Cybernetics Part C-Applications and Reviews, 29(4), 481-493.
Colombo, R., Pisano, F., Mazzone, A., Delconte, C., Micera, S., Carrozza, M., et al.
(2007). Design strategies to improve patient motivation during robot-aided rehabilitation.
J Neuroeng Rehabil, 4, 3.
Colton, B. M., & Hollerbach, J. M. (2005). Identification of Nonlinear Passive Devices
for Haptic Simulations. Paper presented at the Symposium on Haptic Interfaces for
Virtual Environment and Teleoperator Systems.
Conner, J. M., Culberson, A., Packowski, C., Chiba, A. A., & Tuszynski, M. H. (2003).
Lesions of the Basal forebrain cholinergic system impair task acquisition and abolish
cortical plasticity associated with motor skill learning. Neuron, 38(5), 819-829.
Cooper, R. A. (2008). Quality-of-life technology. A human-centered and holistic design.
IEEE Eng Med Biol Mag, 27, 10-11.
Cooper, R. A., Dicianno, B. E., Brewer, B., LoPresti, E., Ding, D., & Simpson, R. (2008).
A perspective on intelligent devices and environments in medical rehabilitation. Med Eng
Phys, 30, 1387-1398.
Coote, S., Murphy, B., Harwin, W., & Stokes, E. (2008). The effect of the GENTLE/s
robot-mediated therapy system on arm function after stroke. Clinical Rehabilitation,
22(5), 395-405.

１０７
Csikszentmihalyi, M. (1979). Intrinsic rewards and emergent motivation. In M. Lepper &
D. Green (Eds.), The hidden costs of reward: New perspectives on the psychology of
human motivation . New York: Erlbaum.
Deci, E., & Ryan, R. M. (1985). Intrinsic motivation and self-determination in human
behavior. New York: Plenum.
Desrosiers, J., Malouin, F., Bourbonnais, D., Richards, C. L., Rochette, A., & Bravo, G.
(2003). Arm and leg impairments and disabilities after stroke rehabilitation: relation to
handicap. Clin Rehabil, 17(6), 666-673.
Dijkers, M., deBear, P., Erlandson, R., Kristy, K., Geer, D., & Nichols, A. (1991). Patient
and staff acceptance of robotic technology in occupational therapy: a pilot study. J
Rehabil Res Dev, 28(2), 33-44.
Dijkers, M. P., deBear, P. C., Erlandson, R. F., Kristy, K., Geer, D. M., & Nichols, A.
(1991). Patient and staff acceptance of robotic technology in occupational therapy: a pilot
study. J Rehabil Res Dev, 28(2), 33-44.
Doyon, J., Gaudreau, D., Laforce, R., Jr., Castonguay, M., Bedard, P. J., Bedard, F., et al.
(1997). Role of the striatum, cerebellum, and frontal lobes in the learning of a visuomotor
sequence. Brain Cogn, 34(2), 218-245.
Duncan, P. W. (1997). Synthesis of intervention trials to improve motor recovery
following stroke. Tpoics in Stroke Rehabilitation, 3, 1-20.
Emken, J., Benitez, R., & Reinkensmeyer, D. (2007). Human-robot cooperative
movement training: learning a novel sensory motor transformation during walking with
robotic assistance-as-needed. J Neuroeng Rehabil, 4, 8.
Erlandson, R. F., deBear, P., Kristy, K., Dijkers, M., & Wu, S. (1990). A robotic system to
provide movement therapy. Paper presented at the Proc. 5th Int. Service Robot Conf,
Detroit, MI,.
Fasoli, S. E., Krebs, H. I., Stein, J., Frontera, W. R., & Hogan, N. (2003). Effects of
robotic therapy on motor impairment and recovery in chronic stroke. Archives of Physical
Medicine and Rehabilitation, 84(4), 477-482.

１０８
Fazekas, G., Horvath, M., & Toth, A. (2006). A novel robot training system designed to
supplement upper limb physiotherapy of patients with spastic hemiparesis. International
Journal of Rehabilitation Research, 29(3), 251-254.
Field, M. J., & Jette, A. M. (2007). The Future of Disability in America. Washington, DC:
The National Academies Press.
Fisher, B. E., & Sullivan, K. J. (2001). Activity-dependent factors affecting poststroke
functional outcomes. Top Stroke Rehabil, 8(3), 31-44.
Gomez, D., Burdea, G., & Langrana, N. (1995). Integration of the Rutgers. Master II in a
virtual reality simulation,. Paper presented at the Proc. VRAIS.
Goode, S., & Magill, R. (1986). Contextual interference effects in learning three
badminton serves. Research Quarterly for Exercise and Sport, 53, 308-314.
Gordon, J. (2000). Assumptions underlying physical therapy intervention: Theoretical
and historical perspectives. In C. J.H. & S. R.B. (Eds.), Movement science: Foundations
for physical therapy in rehabilitation, 2nd edition (pp. 1-31). Rockville, MD.: Aspen
Publishers.
Guadagnoli, M. A., & Lee, T. D. (2004). Challenge point: a framework for
conceptualizing the effects of various practice conditions in motor learning. J Mot Behav,
36(2), 212-224.
Hasser, C. J., & Cutkosky, M. R. (2002). System identification of the human hand
grasping a haptic knob. Paper presented at the IEEE Symposium on Haptic Interfaces for
Virtual Environment and Teleoperator Systems.
Hayward, V., & Armstrong, B. (2000). A new computational model of friction applied to
haptic rendering. Experimental Robotics Vi, 250, 403-412.
Heart and Stroke Statistics, 2006 Update. (2006). Dallas, TX: American Heart
Association.
Heart Disease and Stroke Statistics, 2003 Update:. (2003). American Heart Association.
Hebb, D. O. (1955). Drives and the CNS. Psychol Rev, 62, 243-254.

１０９
Huang, V. S., & Krakauer, J. W. (2009). Robotic neurorehabilitation: a computational
motor learning perspective. J Neuroeng Rehabil, 6, 5.
Ivanchenko, V ., & Jacobs, R. A. (2003). A developmental approach AIDS motor learning.
Neural Comput, 15(9), 2051-2065.
Jeffery, D. R., & Good, D. C. (1995). Rehabilitation of the stroke patient. Curr Opin
Neurol, 8(1), 62-68.
Johnson, M. J., Wisneski, K. J., J., A., Nathan, D., & R., S. (2006). Development of
ADLER:The Activities of Daily Living Exercise Robot. Paper presented at the IEEE-
EMBS Biomedical Robotics (BioRob 2006), Pisa, Italy.
Kahn, L. E., Lum, P. S., Rymer, W. Z., & Reinkensmeyer, D. J. (2006). Robot-assisted
movement training for the stroke-impaired arm: Does it matter what the robot does? J
Rehabil Res Dev, 43(5), 619-630.
Karni, A. (1995). When practice makes perfect. Lancet, 345(8946), 395.
Keith, R. A., & Cowell, K. S. (1987). Time use of stroke patients in three rehabilitation
hospitals. Soc Sci Med, 24(6), 529-533.
Kemp, B., & Moaqueda, L. (2004). Aging with a Disability: What the Clinician Needs to
Know . . Baltimore, MD.: Johns Hopkins University Press.
Kleim, J. A., Barbay, S., Cooper, N. R., Hogg, T. M., Reidel, C. N., Remple, M. S., et al.
(2002). Motor learning-dependent synaptogenesis is localized to functionally reorganized
motor cortex. Neurobiol Learn Mem, 77(1), 63-77.
Kluzik, J., Diedrichsen, J., Shadmehr, R., & Bastian, A. J. (2008). Reach adaptation: what
determines whether we learn an internal model of the tool or adapt the model of our arm?
J Neurophysiol, 100(3), 1455-1464.
Krebs, H., Hogan, N., Aisen, M., & Volpe, B. (1998). Robot-aided neurorehabilitation.
IEEE Trans Rehabil Eng, 6(1), 75-87.
Krebs, H., Palazzolo, J., Dipietro, L., Volpe, B., & Hogan, N. (2003). Rehabilitation
robotics: Performance-based progressive robot-assisted therapy. AUTONOMOUS
ROBOTS, 15(1), 7-20.

１１０
Krebs, H. I., Ferraro, M., Buerger, S. P., Newbery, M. J., Makiyama, A., Sandmann, M.,
et al. (2004). Rehabilitation robotics: pilot trial of a spatial extension for MIT-Manus. J
Neuroengineering Rehabil, 1(1), 5.
Kwakkel, G. (2009). Towards integrative neurorehabilitation science. Physiother Res Int,
14(3), 137-146.
Kwakkel, G., Kollen, B. J., & Krebs, H. I. (2008). Effects of robot-assisted therapy on
upper limb recovery after stroke: a systematic review. Neurorehabil Neural Repair, 22(2),
111-121.
Kwakkel, G., Wagenaar, R. C., Twisk, J. W., Lankhorst, G. J., & Koetsier, J. C. (1999).
Intensity of leg and arm training after primary middle-cerebral artery stroke: a
randomized trial. Lancet, 354, 191-196.
Lang, C. E., Macdonald, J. R., Reisman, D. S., Boyd, L., Jacobson Kimberley, T.,
Schindler-Ivens, S. M., et al. (2009). Observation of amounts of movement practice
provided during stroke rehabilitation. Arch Phys Med Rehabil, 90(10), 1692-1698.
Lee, T. D., Magill, R. A., & Weeks, D. J. (1985). Influence of practice schedule on testing
schema theory predictions in adults. Journal of Motor Behavior, 17, 283–299.
Lincoln, N. B., Willis, D., Philips, S. A., Juby, L. C., & Berman, P. (1996). Comparison of
rehabilitation practice on hospital wards for stroke patients. Stroke, 27(1), 18-23.
Liu, Y. T., Mayer-Kress, G., & Newell, K. M. (2003). Beyond curve fitting: a dynamical
systems account of exponential learning in a discrete timing task. J Mot Behav, 35(2),
197-207.
Ljung, L., & Söderström, T. (1986). Theory and practice of recursive identification:
Cambridge, MIT Press.
Lotze, M., Braun, C., Birbaumer, N., Anders, S., & Cohen, L. G. (2003). Motor learning
elicited by voluntary drive. Brain, 126(Pt 4), 866-872.
Lum, P. S., Taub, E., Schwandt, D., Postman, M., Hardin, P., & Uswatte, G. (2004).
Automated Constraint-Induced Therapy Extension (AutoCITE) for movement deficits
after stroke. J Rehabil Res Dev, 41(3A), 249-258.

１１１
MacClellan, L. R., Bradham, D. D., Whitall, J., V olpe, B., Wilson, P. D., Ohlhoff, J., et al.
(2005). Robotic upper-limb neurorehabilitation in chronic stroke patients. Journal of
Rehabilitation Research and Development, 42(6), 717-722.
MacKay, D. J. C. (1992). Information-based objective function for active data selection.
Neural Comput, 4, 590-604.
Mackey, F., Ada, L., Heard, R., & Adams, R. (1996). Stroke rehabilitation: are highly
structured units more conducive to physical activity than less structured units? Arch Phys
Med Rehabil, 77(10), 1066-1070.
MacLean, K. E., & Durfee, W. K. (1995). An apparatus to study the emulation of haptic
feedback. Paper presented at the Proc. ASME Dynamic Systems and Control Division.
MacLean, K. M. (1996). The Haptic Camera: A Technique for Characterizing and
Playing Back Haptic Properties of Real Environments. Paper presented at the Haptic
Interfaces for Virtual Environments and Teleoperator Systems (HAPTICS),
ASME/IMECE.
Marchal-Crespo, L., & Reinkensmeyer, D. J. (2009). Review of control strategies for
robotic movement training after neurologic injury. J Neuroeng Rehabil, 6, 20.
McCraken, H. D., & Stemach, G. E. (1977). A test of the schema theory of discrete motor
learning. Journal of Motor Behavior, 9, 193-201.
McGinnis, S. L., & Moore, J. (2006). The impact of the aging population on the health
workforce in the United States--summary of key findings. Cah Social Demogr Med, 46,
193-220.
Menkes, J. M. (1995). Textbook of Child Neurology. Baltimore, MD: : Williams &
Wilkins.
Miller, B. E., & Colgate, J. E. (1998). Using a Wavelet Network to Characterize Real
Environment for Haptic Display. Paper presented at the ASME Dynamic Systems and
Control Division.
Muellbacher, W., Ziemann, U., Boroojerdi, B., Cohen, L., & Hallett, M. (2001). Role of
the human motor cortex in rapid motor learning. Exp Brain Res, 136(4), 431-438.

１１２
Nef, T., Mihelj, M., Kiefer, G., Perndl, C., Mueller, R., & Riener, R. (2007). ARMin -
Exoskeleton for arm therapy in stroke patients. 2007 Ieee 10th International Conference
on Rehabilitation Robotics, Vols 1 and 2, 68-74 1135.
Nelson, T. O., & Leonesio, R. J. (1988). Allocation of self-paced study time and the
"labor-in-vain effect". J Exp Psychol Learn Mem Cogn, 14(4), 676-686.
Nudo, R. J., Wise, B. M., SiFuentes, F., & Milliken, G. W. (1996). Neural substrates for
the effects of rehabilitative training on motor recovery after ischemic infarct. Science,
272(5269), 1791-1794.
Olsson, H., Astrom, K., de Wit, C., Gafvert, M., & Lischinsky, P. (1998). Friction models
and friction compensation. Eur. J. Control, 4(3), 176-195.
Osu, R., Hirai, S., Yoshioka, T., & Kawato, M. (2004). Random presentation enables
subjects to adapt to two opposing forces on the hand. Nat Neurosci, 7(2), 111-112.
Park, J., & Khatib, O. (2006). A haptic teleoperation approach based on contact force
control. International Journal of Robotics Research, 25(5-6), 575-591.
Pavlides, C., Miyashita, E., & Asanuma, H. (1993). Projection from the sensory to the
motor cortex is important in learning motor skills in the monkey. J Neurophysiol, 70(2),
733-741.
Pigott, R. E., & Shapiro, D. C. (1984). Motor schema: The structure of the variability
session. Research Quarterly for Exercise and Sport, 55, 41-45.
Plautz, E., Milliken GW, Nudo RJ. (2000). Effects of repetitive motor training on
movement representations in adult squirrel monkeys: role of use versus learning.
Neurobiol Learn Mem, 74, 27-55.
Richard, C. (2000). On the identification and haptic display of friction. DAI, 61(11B),
6103.
Rosati, G., Gallina, P., & Masiero, S. (2007). Design, implementation and clinical tests of
a wire-based robot for neurorehabilitation. Ieee Transactions on Neural Systems and
Rehabilitation Engineering, 15(4), 560-569.

１１３
Ryan, R., & Deci, E. (2000). Self-determination theory and the facilitation of intrinsic
motivation, social development, and well-being. Am Psychol, 55(1), 68-78.
Sanger, T. D. (2004). Failure of motor learning for large initial errors. Neural Comput,
16(9), 1873-1886.
Sansone, C., & Harackiewicz, J. M. (2002). Intrinsic and extrinsic motivation. San Diego,
CA: Academic Press.
Schaal, S., & Atkeson, C. G. (1998). Constructive incremental learning from only local
information. Neural Computation, 10(8), 2047-2084.
Schmidt, R. A. (1975). A schema theory of discrete motor skill learning. Psychological
Review, 82, 225-260.
Schmidt, R. A., & Lee, T. D. (1999). Motor control and learning: a behavioral emphasis.
Champaign, Ill: Human Kinetics.
Shea, J. B., & Morgan, R. (1979). Contextual interference effects on the acquisition,
retention, and transfer of motor skills. J. exp. psychol. [ Hum. Learn.], 3, 179-187.
Simon, D. A., & Bjork, R. A. (2001). Metacognition in motor learning. J Exp Psychol
Learn Mem Cogn, 27(4), 907-912.
Simon, D. A., Cullen, J. D., & Lee, T. D. (2002). Win-shift/lose-stay: Contingent
switching as an alternative to random practice? Journal of Sport and Exercise Psychology,
24, S116.
Sukal, T. M., Krosschell, K. J., & Dewald, J. P. A. (2007). Use of the ACT3D system to
evaluate synergies in children with spastic hemiparetic cerebral palsy: A pilot study. 2007
Ieee 10th International Conference on Rehabilitation Robotics, Vols 1 and 2, 964-
967 1135.
Swindells, C., & MacLean, K. E. (2007). Capturing the Dynamics of Mechanical Knobs.
Paper presented at the World Haptics Conference
Taub, E., Lum, P. S., Hardin, P., Mark, V. W., & Uswatte, G. (2005). AutoCITE:
automated delivery of CI therapy with reduced effort by therapists. Stroke, 36(6), 1301-
1304.

１１４
Thrun, S. (2000). Probabilistic algorithms in robotics. AI Magazine, 21, 93-109.
Thurman, D., Alverson, C., Dunn, K., Guerrero, J., & Sniezek, J. (1999). Traumatic brain
injury in the United States: a public health perspective. Journal of Head Trauma and
Rehabilitation 14, 602–615.
Titzer, R., Shea, J. B., & Romack, J. (1993). The effect of learner control on the
acquisition and retention of a motor task. Journal of Sport and Exercise Psychology, 13,
S84.
Traumatic brain injury: hope through research. (2002). Bethesda (MD): National
Institutes of Health.
Uswatte, G., Taub, E., Morris, D., Vignolo, M., & McCulloch, K. (2005). Reliability and
validity of the upper-extremity motor activity Log-14 for measuring real-world arm use.
Stroke, 36(11), 2493-2496.
van Asseldonk, E. H., Wessels, M., Stienen, A. H., van der Helm, F. C., & van der Kooij,
H. (2009). Influence of haptic guidance in learning a novel visuomotor task. J Physiol
Paris, 103(3-5), 276-285.
Van Peppen, R. P., Kwakkel, G., Wood-Dauphinee, S., Hendriks, H. J., Van der Wees, P.
J., & Dekker, J. (2004). The impact of physical therapy on functional outcomes after
stroke: what's the evidence. Clin Rehabil, 18, 833-862.
van Vliet, P., Kerwin, D. G., Sheridan, M. R., & Fentem, P. (1995). The influence of
goals on the kinematics of reaching forllowing stroke. Neurology Report, 19, 11-16.
Vygotsky, L. S. (1978). Mind in Society. Cambridge, MA: Harvard University Press.
Welch, G., & Bishop, G. (2004). An Introduction to the Kalman Filter (No. TR 95-041):
UNC-Chapel Hill.
Whitall, J., Waller, S. M., Silver, K. H. C., & Macko, R. F. (2000). Repetitive bilateral
arm training with rhythmic auditory cueing improves motor function in chronic
hemiparetic stroke. Stroke, 31(10), 2390-2395.
White, R. H. (1959). Motivation reconsidered: the concept of competence. Psychol Rev,
66, 297-333.

１１５
Winstein, C. J. (1991). Knowledge of results and motor learning--implications for
physical therapy. Phys Ther, 71(2), 140-149.
Winstein, C. J., Rose, D. K., Tan, S. M., Lewthwaite, R., Chui, H. C., & Azen, S. P.
(2004). A randomized controlled comparison of upper-extremity rehabilitation strategies
in acute stroke: A pilot study of immediate and long-term outcomes. Arch Phys Med
Rehabil, 85(4), 620-628.
Wolf, S. L., Blanton, S., Baer, H., Breshears, J., & Butler, A. J. (2002). Repetitive Task
Practice: A critical review of constraint induced therapy in stroke. The Neurologist, 8,
325-338.
Wolf, S. L., Catlin, P. A., Ellis, M., Archer, A. L., Morgan, B., & Piacentino, A. (2001).
Assessing Wolf Motor function Test as outcome measure for research in patients after
stroke. Stroke, 32(7), 1635-1639.
Wolf, S. L., Winstein, C. J., Miller, J. P., Taub, E., Uswatte, G., Morris, D., et al. (2006).
Effect 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.
Wolf, S. L., Winstein, C. J., Miller, J. P., Thompson, P. A., Taub, E., Uswatte, G., et al.
(2008). Retention of upper limb function in stroke survivors who have received
constraint-induced movement therapy: the EXCITE randomised trial. Lancet Neurol, 7(1),
33-40.
Wu, C., Trombly, C. A., Lin, K., & Tickle-Degnen, L. (2000). A kinematic study of
contextual effects on reaching performance in persons with and without stroke: influences
of object availability. Arch Phys Med Rehabil, 81(1), 95-101.
Wu, W. F., Magill, R. A., & Foto, J. G. (2005). Allowing learners to choose: Self-
regulated practice schedules for learning multiple movement patterns. Journal of Sport &
Exercise Psychology, 27, S161.
Yokokohji, Y., Hollis, R. L., & Kanade, T. (1999). WYSIWYF display: A visual/haptic
interface to virtual environment. Presence-Teleoperators and Virtual Environments, 8(4),
412-434.

１１６
Zhang, B. T. (1994). Accelerated learning by active example selection. International
Journal of Neural Systems, 5, 67-75.

Abstract (if available)

Abstract Robotic technology has the potential to deliver therapy activities for rehabilitation of arm and hand function after stroke more efficiently and effectively than conventional rehabilitation, as it can objectively dose the prescribed intensive amount of therapy in automated design with less cost and effort, and can provide highly reliable measurement of patients‟ progress. The primary goal of this dissertation is to develop a robotic rehabilitation system that fulfills current guidelines for the stroke rehabilitation: motor training focus on realistic tasks that require reaching and manipulation and engage patients post stroke intensively, actively, and adaptively.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Computational model of stroke therapy and long term recovery

PDF

Modeling motor memory to enhance multiple task learning

PDF

Machine learning of motor skills for robotics

PDF

The task matrix: a robot-independent framework for programming humanoids

PDF

Modeling human reaching and grasping: cortex, rehabilitation and lateralization

PDF

Learning reaching skills in non-disabled and post-stroke individuals

PDF

Experimental and computational explorations of different forms of plasticity in motor learning and stroke recovery

PDF

Bayesian methods for autonomous learning systems

PDF

Virtual surgeries as a tool for studying motor learning

PDF

Data-driven robotic sampling for marine ecosystem monitoring

PDF

The representation, learning, and control of dexterous motor skills in humans and humanoid robots

PDF

Development and implementation of a modular muscle-computer interface for personalized motor rehabilitation after stroke

PDF

A framework for automated administration of post stroke assessment test

PDF

Data-driven autonomous manipulation

PDF

Relationship between brain structure and motor behavior in chronic stroke survivors

PDF

Intelligent robotic manipulation of cluttered environments

PDF

Iterative path integral stochastic optimal control: theory and applications to motor control

PDF

Rethinking perception-action loops via interactive perception and learned representations

PDF

Estimation-based control for humanoid robots

PDF

Deficits and rehabilitation of upper-extremity multi-joint movements in individuals with chronic stroke

Asset Metadata

Creator Choi, Younggeun (author)

Core Title Design of adaptive automated robotic task presentation system for stroke rehabilitation

School Viterbi School of Engineering

Degree Doctor of Philosophy

Degree Program Computer Science

Publication Date 06/04/2010

Defense Date 04/06/2010

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag haptics,machine learning,motor learning,OAI-PMH Harvest,Rehabilitation,robotics,stroke

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Schweighofer, Nicolas (committee chair), Gordon, James (committee member), Schaal, Stefan (committee member)

Creator Email younggch@gmail.com,younggch@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-m3110

Unique identifier UC161870

Identifier etd-Choi-3764 (filename),usctheses-m40 (legacy collection record id),usctheses-c127-337922 (legacy record id),usctheses-m3110 (legacy record id)

Legacy Identifier etd-Choi-3764.pdf

Dmrecord 337922

Document Type Dissertation

Rights Choi, Younggeun

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Repository Name Libraries, University of Southern California

Repository Location Los Angeles, California

Repository Email cisadmin@lib.usc.edu