Journal of Medical and Biological Engineering, 34(3): 284-292

284

Guidance-control-based Exoskeleton Rehabilitation Robot for

Upper Limbs: Application to Circle Drawing for
Physiotherapy and Training
Wei-Wen Wang1

Bing-Chun Tsai1

Li-Chun Hsu1

Li-Chen Fu1,2,*

Jin-Shin Lai3,4

Department of Electrical Engineering, National Taiwan University, Taipei 106, Taiwan, ROC
Department of Computer Science and Information Engineering, National Taiwan University, Taipei 106, Taiwan, ROC
3
Department of Physical Medicine and Rehabilitation, National Taiwan University, Taipei 106, Taiwan, ROC
4
Department of Physical Medicine and Rehabilitation, National Taiwan University Hospital, Taipei 106, Taiwan, ROC

Received 22 Aug 2013; Accepted 5 Feb 2014; doi: 10.5405/jmbe.1663

Abstract
Cerebral vascular disease is the leading cause of functional disability among adults. Approximately half of all
stroke survivors continue to suffer from severe neurological deficits and hemiparesis in the upper extremities as well as
many secondary complications due to immobilization. Robotics can provide highly intensive intervention in stroke
rehabilitation as well as an objective means of measuring patient progress. This study designs an upper limb
rehabilitation (Rehab) robot with multiple degrees of freedom. This design provides a wider range of motion in
3-dimentional space than that provided by an existing endpoint-fixation system. In addition, unlike cable suspension
systems that lack biofeedback, the sensors incorporated into the proposed design can be used to detect the voluntary
force produced by the stroke patient. The Rehab robot features an exoskeleton-type design with in-built redundancy, a
guidance control system, and force feedback using an electromyographic trigger. Three rehabilitation modes can be
selected by physical therapists according to the severity of the patients upper-limb impairment: passive, active, and
guidance. Guidance mode assists patients in motor training, with programs such as drawing circles, which involves
complex movements that require coordination between the shoulder and elbow joints. Such skills are ideally suited to
relearning functional tasks following a stroke. Physical experiments were conducted in this pilot study to evaluate the
performance of the Rehab robot. The results indicate that the robot could be effective. Guidance mode achieves the
desired guidance functions, informing the subject of the pose required to complete the task as well as enabling them to
reduce unnecessary muscle use.
Keywords: Rehabilitation, Exoskeleton, Kinematics, Robot-assisted therapy

1. Introduction
Most stroke patients suffer from motor dysfunction and
approximately half of all stroke survivors continue to suffer
from severe neurological deficits and hemiparesis in the upper
extremities (UEs) [1]. Many secondary complications due to
immobilization may also occur, including joint contracture,
muscle atrophy, and shoulder-hand syndrome. To prevent such
complications and regain functional motor capabilities in the
UEs, several studies have focused on the development of more
effective rehabilitation techniques for stroke patients.
Traditional rehabilitation is still limited by a number of issues.
For example, one-on-one treatment is labor-intensive and
* Corresponding author: Li-Chen Fu
Tel: +886-2-23622209; Fax: +886-2-23657887
E-mail: lichen@ntu.edu.tw

experience-dependent. Furthermore, the subjectivity involved

in most clinical scales hinders the precise quantification of
improvement following rehabilitative intervention.
As a result, interest in the use of robotic therapy for
rehabilitation is increasing [2-5]. Robotics can provide highly
intensive intervention in stroke rehabilitation as well as an
objective means of measuring patient progress. In addition, the
use of a human-computer interface enables a robot to optimize
the movement patterns required for patients with UE paresis.
Lo et al. (2010) studied the effectiveness of robot-assisted
therapy in reducing motor impairment in arms affected by
stroke [4]. They found that the robot-assisted therapy did not
significantly improve motor function over a period of 12
weeks; however, improvements in motor capability and motortask performance were observed after 36 weeks.
Three primary types of robot are used to contact or
interact with stroke patients. The first type is an endpointfixation system, such as PHANTOM [6] or MIT-Manus [2],

J. Med. Biol. Eng., Vol. 34 No. 3 2014

which fixes the distal portion of the UEs of patients in order to

guide the desired movements. This enables stroke patients to
execute tasks using only forearm support. The second type is a
cable suspension system, such as the Freebal gravity
compensation system [7]. This type of system provides
antigravity support for the UEs during rehabilitation. The third
type is an exoskeleton arm system, such as ARMin [8,9],
ULERD [10], and MAHI [11].
The proposed rehabilitation (Rehab) robot is an
exoskeleton designed specifically for UE rehabilitation. This
design provides a wider range of motion (ROM) in 3dimentional space than that provided by an endpoint-fixation
system. In addition, unlike cable suspension systems that lack
biofeedback, the sensors incorporated into the proposed design
can be used to detect the voluntary force produced by the stroke
patient for further analysis.
The proposed Rehab robot has three key elements: a
redundant design combined with selective inverse kinematics
(IK) solutions, a guidance control system, and an
electromyographic (EMG) trigger. The redundant design refers
to additional joints beyond those found in a normal human
upper limb. The IK problem arises from the redundant design;
therefore, this study investigated the geometrical relationship
between the robot arm and the human arm and devised an
effective IK solution. The guidance control system was designed
for stroke patients with mild-to-moderate UE impairment to
ensure correct joint movements. An EMG trigger modified from
MIT-Manus [12] and a proportional-integral-derivative (PID)
controller and an impedance controller, which are commonly
found in Rehab robots [13-15], are also included.
One important application of the proposed Rehab robot is
circle drawing. It is a training program that executes the
coordinated movements using the paretic arm. Virtual reality
environments and games have been developed and combined
with Rehab robots to facilitate long-term training and increase
movement in therapy [16,17]. Miyoshi et al. (2010) pointed out
that drawing a circle requires complex movements and the
coordination of multiple muscular co-contractions as well as
eccentric activity [18] in the medial/lateral and forward/
backward directions. Ju et al. (2005) used a robot to guide the
upper limbs of subjects in linear and circular movements [19].
Based on previous works [18,19], the complex circle-drawing
movement should enhance the coordination and motoring of
shoulder flexion/extension. Motions that mimic circle drawing
in the vertical plane are common in daily life, such as turning a
steering wheel or cleaning windows. The shoulder wheel has
been used for training this kind of motion. In addition, a
considerable relationship has been reported between the
activation of the motor cortex and the copying of visual
representations of particular geometrical shapes [20].

285

These modes are similar to those used in traditional programs

for the rehabilitation of stroke patients.
As with passive ROM exercises, the passive mode of the
Rehab robot provides an external force to stroke patients with
severe motor impairment. In this setting, force is provided
entirely by the Rehab robot to help execute the rehabilitation
movements designed by physical therapists. Active mode is
used for stroke patients with mild UE impairment, such that
patients can move their arm freely. Motion tracking by sensors
in the Rehab robot is used for analysis. Finally, in guidance
mode, the joint movements of patients are guided using force
feedback with an EMG trigger. For example, the affected arm
of patients can be drawn back to a plane by the motion tracking
system if it tends to deviate from the plane.
2.2 Kinematics model of human upper limbs
Generally speaking, the shoulder joint complex has 3
degrees of freedom (DOFs) in its movement, 2 DOFs at the
elbow joint, and 2 DOFs at the wrist joint. Table 1 summarizes
the normal ROM of a human.
Table 1. Movements and limitations of human upper limbs.
Movement
Shoulder flexion/extension
Shoulder abduction/adduction
Shoulder external/internal rotation
Elbow flexion/extension
Forearm pronation/supination
Wrist flexion/extension
Wrist lateral/medial deviation

Range of motion
0-180/0-60
0-180/0-70
0-34/0-97
0-150/0-0
0-90/0-90
0-80/0-70
0-30/0-20

2.3 Design of mechanical structure

The proposed exoskeleton robot arm is equipped with
more joints than those in a human arm in order to increase the
ROM to make the robot suitable for most patients. The
exoskeleton is a 9-DOF mechanical manipulator, including 6
DOFs at the shoulder joint complex, 2 DOFs at the elbow joint,
and 1 DOF at the wrist joint, as shown in Fig. 1(a). Mechanical
joints 1 through 6 of the Rehab robot are used to accommodate
motions which originate in the shoulder area; mechanical joint
7 accommodates motions in the human elbow joint
(flexion/extension); mechanical joint 9 accommodates motion
in the human forearm joint (pronation/supination); the
remaining mechanical joints accommodate motions produced
by the human wrist joint. To ensure that the robot can adapt to a
variety of patients, the mechanism for the upper arm can be
adjusted between 26 and 34 cm, while the forearm can be
adjusted between 24 and 30 cm.

2. Materials and methods

2.1 Rehabilitation modes of Rehab robot
The rehabilitation mode of the proposed Rehab robot can
be divided into three types: passive, active, and guidance.

(a)

(b)

Figure 1. (a) Photograph and (b) simplified schematic diagram of

proposed Rehab robot.

286

Rehabilitation Robot for Upper Limbs

2.4 Sensors
The proposed sensor system includes a potentiometer and
a motor encoder for each joint, as well as EMG and force
sensors for the upper limbs of the patient. The Rehab robot is
also equipped with 4 force sensors mounted at the connections
between the robot arm and the human arm, as shown in
Fig. 1(a). Each of these sensors was realized using a pair of
strain gauges to measure the interaction force between the
patient and robot. The force of the upper arm is measured by
two force sensors which monitor shoulder flexion/extension
and horizontal adduction/abduction. Elbow flexion/extension
with shoulder rotation yields an interaction force measured
from the forearm.
2.5 Safety issues
A safety system is implemented to determine which parts
of the Rehab robot may be broken or have failed. The system
includes two parts: a dual-position sensor system that obtains
two kinds of position information that validate each other and a
hand button to control the motion state by allowing state
jumping in order to prevent potential injury. An emergency
button is provided to avoid some critical circumstances that
may occur. Besides what we have just mentioned so far, our
method also asks part of the device deemed for safety guard
must be automatic in order to ensure that the system can
automatically detect the hazard before any harm could occur.
Moreover, such safety detector can prevent too frequent
artificial stop of the machine when a patient is during the
course of rehabilitation training while feeling uncomfortable,
but in fact, he can still work along with the machine well.
2.5.1 Dual-position sensor system
Although we have mentioned about the range of motion of
this rehabilitation robot previously, and have shown the design
target is to achieve the sufficient rehabilitation workspace of
the upper-limb of a human body, single failure of robot
function that may take place includes failures of components
also need to be considered during training and therapy. Given
the above-mentioned philosophy, we have to take into account
the positioning failure in particular. Here, we use the
potentiometer to detect the absolute position of each robot joint;
however, the potentiometer provides an absolute measurement
but with the lower precision. On the other hand, the encoder
usually has higher precision during measurement of the
position change, but it cannot figure out the current absolute
position due to lack of initial position information. After the
system finishes the calibration, this dual potion sensor system
will work synchronously. If one of the two sensor sub-systems
fails, the system will cut off power to avoid potential harm until
the problem is removed, and then the system will be restarted.
For the up-down linear motion joint, the dual position sensing
is set up with the same strategy to determine the upper and
lower limitations of joint position.

2.5.2 Hand button for motion suspension

A hand button panel was designed for the patients use. It
prevents the Rehab robot from imposing harm to the patient
during the course of training. When the patient cannot
accomplish a training step due to pain, this mechanism
immediately moves the robot back to the previous step. The
break posture is recorded and given to a therapist to evaluate
the difficulties encountered during the designed course of
rehabilitation and to identify possible modification of the
rehabilitation program later on.
2.5.3 Emergency button
The robot arm uses AC 110 V for power. Two emergency
buttons can be used to cut off the power system immediately.
One is a manual control and the other is for the pedal break.
When the power is shut down, it is critical for the robot arm not
to fall quickly. A high-ratio reduction gear is thus used for the
transmissions and vertical joints as a lock of mechanical
structure.
2.6 Inverse kinematics of Rehab robot
This section explores all feasible motions of the
exoskeleton-type Rehab robot arm through the study of IK
solutions. A simplified schematic diagram of the robot structure
is presented in Fig. 1(b), in which a total of 12 coordinate
frames are assigned to the base and appropriate locations on the
11 joint axes using Denavit-Hartenberg (D-H) notation. The
origins of coordinate frames are referred to as joint pivots for
convenience. Note that joint V1 (z8) and joint V2 (zV1) are
stationary; hence, the associated rotation angles V1 and V2 are
constant, leading to a U-shaped link which connects the arm at
the origin of the coordinate frame {XYZ7} to the handle for
grasping at the origin of the coordinate frame {XYZ9}.
Notations d1 (with sliding joint) and 2 to 9 (with revolute
joints) are treated as variables corresponding to various joint
motions. The associated D-H parameters are listed in Table 2.
Table 2. Denavit-Hartenberg parameters of Rehab robot.
Joint

d (cm)

a (cm)

(rad.)

1
2
3
4
5
6
7
8
V1
V2
9

2
3
4
5
6
7
8
V1
V2
9

d1
0
0
0
0
d6
0
0
0
0
d9

0
a2
a3
a4
0
0
a7
a8
aV1
0
0

0
0
0
-/2
-/2
/2
0
/2
0
-/2
0

Home
(rad.)
/2
0
0
0
-/2
0
/2
0
/2
0
0

The parallel motion principle is generally employed when

the relationship between the exoskeleton robot arm and the
human arm is sought, which in turn helps solving IK solutions
of the robot arm. In short, the problem is simplified by finding
only the solutions to which both the Rehab robot and the
human arm will conform since the robot is exoskeleton-type.
The position trajectories of the essential joint pivots of the
Rehab robot are first determined by basing these trajectories on

J. Med. Biol. Eng., Vol. 34 No. 3 2014

those found in various joints of the human arm. More

specifically, the positions of the three joint pivots of the robot
arm which respectively correspond to the human shoulder joint,
elbow joint, and wrist joint are found. This facilitates solving
the pose of the Rehab robot using geometry information
associated with these three joint pivot positions.
Suppose that the positions of the human shoulder joint Ohs,
elbow joint Ohe, and wrist joint Ohw, as well as the length of
upper arm lhse, the length of forearm lhew, and the parallel
distance lhr between the Rehab robot and the human arm are
known. Denote the pivots of joints 1 to 9 of the robot arm as O1
to O9, and rename the pivots of joints 5, 7, and 8 as robot
shoulder joint Ors( = O5), elbow joint Ore( = O7), and wrist joint
Orw( = O8), respectively. In accordance with the formerly
mentioned principle, the latter three robot joint pivots should
correspond to joints Ohs, Ohe, and Ohw of the human arm,
respectively. If the lengths of the upper arm and the forearm of
the robot are denoted as lrse( = d6) and lrew( = a7), respectively,
the relationship between the human arm and the Rehab robot
can be characterized as shown in Fig. 1. The mentioned IK
calculations are presented in greater detail in our previous
paper [21].
2.7 Control system
Figure 2 presents the control system for the four
rehabilitation modes. The control system comprises an
impedance controller, an EMG trigger, and a guidance
controller. The impedance controller is a torque controller that
minimizes the torque/force interaction between the Rehab robot
and the human arm as much as possible, thereby enabling the
Rehab robot to follow human volition in its movement. The
EMG trigger is used to check whether the human muscles have
contracted. Finally, the guidance controller provides the local
target of the human pose based on the current pose and hand
position (or pose) of the target. Passive mode is executed using
zero gain in the impedance controller and 100% gain in the
guidance controller, whereas active mode is executed using
positive gain in the impedance controller and zero gain in the
guidance controller. Guidance mode is executed using positive
gain in the impedance controller and 10%~90% gain in the
guidance controller.

Figure 2. Diagram of control system.

2.7.1 Impedance controller

Geometry is first used to examine the relationship between
force sensing and various motions of the human arm. Figure 3
shows the relationship between the exerted forces due to
movement of the upper arm and shoulder, which

287

mathematically can be expressed in terms of the equivalent

shoulder torques:

hs1 = ru sin(hs2 + F1 coshs3 + F2 sinhs3)

sin(hs2 + coshs3(F3lu + F3rf cos(he )))

+ sinhs3 F4 rf + lu cos(he )

F3rf cos(hs2 + ) sin(he )

(1)

hs2 = F4 coshs3(rf + cos(he ) lu)

F3sinhs3 (lu + cos(he ) rf

+ ru(F2 coshs3 F1 sinhs3) (Gu (hs2)

(2)

(a)

(b)

(c)

Figure 3. Forces (F1~F4) sensed from upper arm and forearm. (a) Front,
(b) right, and (c) top views.

Similarly, the exerted forces due to movement of the

forearm result in the corresponding equivalent shoulder and
elbow torques are mathematically expressed as follows:

hs3 = F3rf sin (he ) Gf (hs2,hs3,he)

(3)

he = F4rf Gf (hs2,hs3,he)

(4)

In the above equations, the variables are defined as:

hs1 represents equivalent shoulder torque causing the
exoskeleton to move under shoulder abduction/ adduction;
hs2 represents equivalent shoulder torque causing the
exoskeleton to move under shoulder flexion/extension;
hs3 represents equivalent shoulder torque causing the
exoskeleton to move under shoulder rotation;
he represents equivalent elbow torque causing the exoskeleton
to move for elbow flexion/extension;
ru represents the distance between the shoulder and force
sensor;
rf represents the distance between the elbow and force sensor of
forearm;
lu represents the distance between the shoulder and elbow;
hs1 represents the angle of shoulder abduction/adduction;
hs2 represents the angle of shoulder flexion/extension;
hs3 represents the angle of shoulder rotation;
he represents angle of elbow flexion/extension;
F1, F2 represent the forces sensed from the upper arm;
F3, F4 represent the forces sensed from the forearm;
Gu(hs2) represents the influence of gravity on the upper arm;
Gf/hs2/hs3/he represents the influence of gravity on the
forearm.

Rehabilitation Robot for Upper Limbs

288

Impedance control is used to imitate mechanical

impedance between the pose and torque of the human upper
limb. The general form is:
M m (d ) Bm (d ) Km (d )

(5)

where:
hs1 hs2 hs3 he] is vector of equivalent torques based on
Eqs. (1) to (4);
hs1 hs2 hs3 he] is the current pose of the human upper
limb, where hs1, hs2, hs3, and he are as previously
defined;
d is the desired pose of the human upper limb;
Mm, Bm, and Km are inertia, damping, and stiffness matrices,
respectively.
Since the motion of the Rehab robot for stroke patients is
typically slow, the inertia and velocity effects are relatively
small, and in most cases can be ignored. Therefore, for
simplicity, it is assumed below that the imitated inertia Mm and
damping Bm of the mechanical impedance are zero. Under these
assumptions, Eq. (5) can be simplified as follows:

d + K m -1 = +
qd q q

Lr-1Lh(

(6)
Jr-1(q)

Jh(

(7)

where:
= Km-1 is the compensated pose of the human upper limb;
x=Lh() is the direct kinematics model of the human upper
limb;
q= Lr-1(x) is the IK model of the Rehab robot;
Jh( is the Jacobian matrix of the human upper limb such that
x = Jh(;
Jr-1(q) is the inverse Jacobian matrix of the Rehab robot such
that q = Jr-1(qx.
2.7.2 Force feedback system with EMG trigger
EMG sensors are an efficient way to record and evaluate
electrical activities in skeletal muscles. Many studies have
shown that the mechanical force exerted during muscle
contractions is directly related to EMG amplitude [22]. Stroke
patients with paretic upper limbs are unable to execute
functional tasks independently; thus, a Rehab robot is often
required to provide assistance for the completion of tasks. In
this study, the Rehab robot incorporates a force feedback
system with an EMG trigger.
EMG equipment is used to detect the weak EMG signals
produced by stroke patients. The force sensor is a transducer
that converts a mechanical input force into electrical output
signals. Therefore, comparing EMG signals and electrical
signals from the mechanical force sensors enables the proposed
system to identify muscular contractions. The Rehab robot then
assists the patient with appropriate external force to complete
the designated task. The system is detailed as follows:
(1) EMG pre-processing: EMG signals mch(t) are recorded
using a band-pass filter from 20 Hz to 450 Hz. Myoelectric
activity Ech(t) is defined as follows:
Ech (t )

t T mch

(t )dt

(8)

where ch = 1, 2, , 8 indicates the specific muscle.

(2) Triggered signal: Threshold Tch is determined
according to the myoelectric activity of the relaxed muscle. The
triggered signal is defined as follows:
1, if Ech (t ) Tch
TCch (t )
otherwise
0,

(9)

The control strategy (6) is then modified to reflect the

incorporation of the EMG trigger:

d t (t) TCch t K

(10)

2.7.3 Guidance controller

The IK of the Rehab robot described in Section 2 does not
consider the problem of redundancy if the desired pose of a
human arm is known. However, we cannot determine which
pose is appropriate if we only know the target position of the
human hand in daily life. A previous study [23] proposed an
optimal feedback control system to minimize movement in
order to address the issue of redundancy. Thus, this study
adopts a straightforward geometric method to resolve this
challenge.
As outlined in [24], when a healthy human subject moves
their hand from one point to another, the trajectory of the hand
tends to follow the dotted line shown in Fig. 4. In other words,
the dotted line in Fig. 4 can be used to identify the various
reference positions of the human wrist. The explicit formula is:
Xhwr = Xhw + KG (Xhwt Xhw)

(11)

where:

Xhwr = [xhwr yhwr zhwr] is the reference position of the wrist;

Xhwt = [xhwt yhwt zhwt] is the target position of the wrist;

Xhw = [xhw yhw zhw] is the current position of the wrist;
KG represents guidance gain.

Figure 4. Two desired paths: line and circle. This figure shows the
relationship between the reference pose, current pose, and
desired path.

The position of the elbow is crucial in determining the

complete reference pose of an upper limb; however, solutions
related to the elbow suffer from the problem of redundancy (as
shown in Fig. 4). Based on a minimum movement policy, this
study finds the reference elbow position with the minimum
distance from the current elbow position. Thus, all solutions to
the elbow reference position are found when the shoulder and
reference wrist positions (referring to Fig. 5(a)) are known. In

J. Med. Biol. Eng., Vol. 34 No. 3 2014

(a)

(b)

Figure 5. Two spherical surfaces. The center of the first spherical

surface is the position of shoulder Xhs and the radius is the
length of the upper arm lhes. The center and radius of the
second spherical surface are the reference position of wrist
Xhwr and the length of forearm lhew, respectively. The double
line represents the solutions for the position of the elbow,
comprising two spherical surfaces. (a) The solution is a circle
whose radius and center are respectively r and P. (b)
Relationship between circle C, current elbow position Xhe, and
projection Q.

other words, once the shoulder position is known, it is clear that

elbow position Xher = [xher yher zher] sits on the spherical

surface:
Xher Xhs = lhes

(12)

where Xhs = [0 0 0] is the shoulder position and lhes

denotes the distance between the shoulder and the elbow. Next,
given this wrist position, the elbow position Xher is situated on
the spherical surface:
Xher Xhwr lhew

(13)

where lhew represents the distance between the elbow and the
wrist. Now, all possible reference elbow positions which lie on
a circle that intersects the spheres from (12) and (13) and are
located within a plane referred to as the E-plane can be found.
The solution circle of all possible Xher satisfies the following
plane (E-plane) equation:

(Xhwr Xhs)T(Xher Xhs) 2 XhwrTXher 2

(14)

where = lhsw2 + lhes2 lhew2 with lhsw = Xhwr Xhs. Using Eq.
(14), the plane in space can be rewritten as follows:

xhwr xher yhwr yher zhwr zher 2

(15)

The normal vector of plane (15) is the vector of the reference

position of the wrist:
Xhwr xhwr

yhwr

zhwr

Next, center P and radius r of the circle, representing all

possible elbow reference solutions, are found. By employing
the law of cosine given all edge lengths of the triangle
XherXhwr Xhs, the following is obtained:
P = Xh

Xhwr Xhs
l cos
Xhwr Xhs hes

r lhes sin
1
where X hwr X hs X her cos (

(16)
(17)

2
2
2
lhsw
lhes
lhew
).
2lhes lhsw

289

All possible elbow reference positions comprise a circle;

therefore, redundancy in the solution remains. To compensate
for this issue, an additional constraint is imposed to maintain
the distance between the candidate elbow solution and the
current position of the human elbow minimum. According to
Fig. 5(b), the distance between the current elbow position Xhe
and the solution circle is:
D hdisQ,C

(18)

where h is the distance between the current elbow position; the

E-plane is derived from Eq. (15); and Q is the projection of the
current elbow position onto the E-plane (15). Q is calculated
according to the current position of wrist Xhe following the
direction of the normal vector Xhwr /Xhwr of the E-plane. This
allows us to move the distance between the E-plane and the
current position of wrist dis(E-plane, Xhe). The value of Q is
defined as follows:
Xhwr
Q = Xhe dis(E-plane, Xhe)

Xhwr
xhwr xhe+ yhwr yhe+ zhwr zhe
Xhwr
= Xhe

Xhwr
Xhwr

(19)

Additionally, C is an arbitrarily position on the solution circle.

Because h is a constant, the amplitude of distance D is
determined by dis(Q,C) regardless of where C is positioned on
the circle. Thus, the minimum dis(Q,C) can be found if C is
selected as the intersection of the line connecting the center of
circle P and projection Q (as shown in Fig. 5(b)). Thus, the
reference elbow position leading to the minimum movement
can be expressed as:
X her arg min | X he C | P r
C

QP
QP

(20)

Thus far, all of the reference elbow locations have been

determined point-wise, but this is insufficient to constrain or
guide patients through a desired target path for the hand. A
method capable of guiding the patient to return their hand to the
desired path when the hand deviates from that path is thus
proposed. The method is suggested by Fig. 4, which illustrates
two desired paths: a linear path and a circular path in free space.
In this method, the system controller first chooses the best
direction for the wrist to return to the desired path. The
controller then determines the optimal reference elbow pose
according to the previous strategy in order to guide the patient
to complete the wrist movement successfully.

3. Results
Drawing circles in the frontal plane is a training program
in which the rehabilitation mode (passive, active, or guidance)
can be selected. The experiments conducted in this work were
approved by the Institutional Review Board of National Taiwan
University Hospital.
Prior to initiating circle drawing, the Rehab robot must be
calibrated. The shoulder and elbow joints of the Rehab robot
are set to initial positions and the lengths of the Rehab robot
upper arm and forearm are adjusted. The subject sits with their

290

Rehabilitation Robot for Upper Limbs

upper arm and forearm attached to the support base,

immobilized by straps. The hand of the subject grasps the
handle. Surface EMG electrodes are then attached to the skin
surface of the subject, namely at the deltoid muscle (anterior,
middle, and posterior), the bicep brachii muscle, and the tricep
brachii muscle. These muscles are responsible for shoulder
flexion/extension, abduction, and elbow flexion/extension. The
EMG noise level is set at the threshold value during the resting
state. During the experiment, an EMG amplitude which
exceeded this threshold value indicated muscle activation.
Following setup, the task of circle drawing is executed
with visual feedback from the computer screen in each mode.
Up to shoulder level, the diameter of the target circle can
be adjusted from 0 to 30 cm in the clockwise direction. The
distance between the circle center and the acromion of the right
shoulder is approximately 30 cm. The speed of circle drawing
can also be adjusted (30 to 60 seconds per circle). The subject
follows the circle track shown on the computer in front of
them.
Different designs can be substituted depending on the
needs of individual patients. A training program which involves
circle drawing in the frontal plane is suitable for stroke patients
who have adequate ROM in their right upper limbs (i.e., the
flexion ROM of the right shoulder joint is over 90, and the
extension ROM of the right elbow joint is not limited). The
ROM and muscle strength of the shoulder joint play important
roles in executing daily activities and also emphasize
movement of the shoulder complex.
Two healthy subjects from the Rehab robot design group
(male, ages = 24 and 25 years) participated in the circle
drawing experiment. In the experiment, the subjects drew
circles 24 cm in diameter. Motion speed was set sufficiently
slow (60 seconds per circle) to ensure stability in the movement.
The three rehabilitation modes were used in sequence.
Experimental data were recorded for further analysis.
Figure 6 shows the trajectory of the wrist in the xz and yz
planes under the three modes. Errors in wrist trajectory are
shown in Fig. 7. The angle of the upper limb and the force of
the joint under the three modes are shown in Figs. 8 and 9,
respectively. Table 3 presents the trajectory error of each joint
using the Rehab robot in passive mode.

effectively guide or limit the trajectory of the human wrist.

Comparing the trajectory errors of the wrist under guidance
mode and active mode, the former is obviously reduced,
particularly in the y axis (as shown in Figs. 7(b) and 7(c)).

(a)

(b)

(c)
Figure 6. Trajectory of human wrist in xz and yz planes for subject 1 in
(a) passive, (b) active, and (c) guidance modes.

(a)

4. Discussion
The trajectory error for every joint of the Rehab robot is
bound by a small value in passive mode. The trajectory of the
wrist tries to follow the circle path (shown in Figs. 6 and 7). It
appears that the PID controller provides better tracking results
in the experiment when the system is set to passive mode.
Figures 6(b) and 6(c) clearly show that the circular
trajectory in active mode is not as smooth as that of guidance
mode. This is because moving the wrist along a pre-defined
path according to a low resolution screen is not particularly
easy for patients. Specifically, the depth value (y axis) of the
wrist appears to drift a great deal in active mode (as shown in
Fig. 6(b)). However, this defect is overcome in guidance mode
(as shown in Fig. 6(c)), which demonstrates that this mode can

(b)

(c)
Figure 7. Error trajectory in position of human wrist in (a) passive
(subject 1), (b) active (subjects 1 and 2), and (c) guidance
modes (subjects 1 and 2).

J. Med. Biol. Eng., Vol. 34 No. 3 2014

291

Table 3. Trajectory error of each Rehab robot joint in passive mode.

(a)

(b)

(c)
Figure 8. Angle of upper limb versus time in (a) passive (subject 1), (b)
active (subjects 1 and 2), and (c) guidance modes (subjects 1
and 2).

(a)

(b)

(c)
Figure 9. Internal force between human upper limb and Rehab robot in
(a) passive (subject 1), (b) active (subjects 1 and 2), and (c)
guidance modes (subjects 1 and 2).

Axis
1
2
3
4
5
6
7
8
9

Root mean square of error (rad)

0.0824 (cm)
0.0022 (rad)
0.0032 (rad)
0.0010 (rad)
0.0224 (rad)
0.0137 (rad)
0.0049 (rad)
0.0000 (rad)
0.0000 (rad)

Range of error
-0.0784~0.1673 (cm)
-0.0047~0.0041 (rad)
-0.0083~0.0048 (rad)
-0.0144~0.0222 (rad)
-0.0333~0.0322 (rad)
-0.0237~0.0251 (rad)
-0.0072~0.0073 (rad)
-0.0000~0.0000 (rad)
-0.000~0.0000 (rad)

Figure 8 illustrates that the tendencies of the shoulder and

elbow flexion are similar; however, the tendencies of shoulder
abduction and internal rotation differ among the three modes.
This may be attributed to the different strategies used to resolve
redundancy. The strategy in passive mode is to adopt minimal
elbow movement, whereas that in active mode is to give control
to the subject. In contrast, the strategy in guidance mode is
based on the trajectory error of the wrist in which 1) minimal
elbow movement is selected if the trajectory error of the wrist
is too high; 2) the subject is given control if the trajectory error
of the wrist is sufficiently small; 3) or a mix of 1) and 2) if the
trajectory error is moderate. In the second case, the control
system of the Rehab robot will not intervene, thereby allowing
the subject to continuously control their upper limb provided
that the wrist position does not deviate too far from the circle
path (as shown in Figs. 8(a) and 8(b)).
In passive mode, the subject does not provide any force to
the Rehab robot; therefore, the force sensors sense only the
gravity affecting the upper limb. In practical terms, the force
value is not smooth (Fig. 9) because the Rehab robot is not a
completely rigid body. Active and guidance modes also
experience the same problem, but the control system
incorporating the EMG trigger is able to filter out some of the
structural oscillations. Generally, EMG signals reflect the
volition of the subject. In Figs. 6(c) and 7(a), the subjects
elbow is constantly working (biceps and triceps) because the
subject is trying to control the depth (y axis) value of the wrist
by manipulating their elbow joint. It is worth mentioning that
the amplitude of the EMG signal in guidance mode is smaller
than that in active mode, which implies that the subject is able
to control their wrist to trace the circular path more easily. To
summarize, guidance mode achieves the desired guidance
function, informing the subject of the pose required to complete
the task as well as enabling them to reduce unnecessary muscle
use.

5. Conclusion
This study proposed a Rehab robot that can detect
voluntary movements using force sensors. The robot includes a
redundant design combined with IK solutions, a guidance
control system, and force feedback with an EMG trigger. The
rehabilitation modes (passive, active, and guidance) of this
Rehab robot are similar to techniques used in traditional
rehabilitation training programs. The circle drawing exercise
was used to demonstrate coordination training of multiple
muscles and joints. Circle drawing with a robot also emulates
the concept of motor learning, which emphasizes intensity and

Rehabilitation Robot for Upper Limbs

292

task-specificity to enable effective motor recovery following a

stroke. It has also been reported that circle drawing is
associated with cognitive effort. In other words, relearning
motor control promotes the formation of internal models.

Acknowledgments
This research was sponsored by the National Science
Council, Taiwan, under grants NSC 96-2218-E-002-008 and
NSC 100-2321-B-002-076, National Taiwan University,
Taiwan, under grant NTU-CESRP-103R7617, and National
Taiwan University Hospital, Taiwan, under grant NTUH99P08.

References
[1]

[2]

[3]

[4]

[5]

[6]

[7]

[8]

[9]

H. C. Fischer, K. Stubblefield, T. Kline, X. Luo, R. V. Kenyon

and D. G. Kamper, Hand rehabilitation following stroke: a pilot
study of assisted finger extension training in a virtual
environment, Top Stroke Rehabil., 14: 1-12, 2007.
H. I. Krebs, N. Hogan, B. T. Volpe, M. L. Aisen, L. Edelstein
and C. Diels, Overview of clinical trials with MIT-MANUS: a
robot-aided neuro-rehabilitation facility, Technol. Health Care,
7: 419-423, 1999.
S. Hesse, H. Schmidt, C. Werner, C. Rybski, U. Puzich and A.
Bardeleben, A new mechanical arm trainer to intensify the
upper limb rehabilitation of severely affected patients after
stroke: design, concept and first case series, Eura. Medicophys.,
43: 463-468, 2007.
A. C. Lo, P. D. Guarino, L. G. Richards, J. K. Haselkorn, G. F.
Wittenberg, D. G. Federman, R. J. Ringer, T. H. Wagner, H. I.
Krebs, B. T. Volpe, C. T. Bever, D. M. Bravata, P. W. Duncan, B.
H. Corn, A. D. Maffucci, S. E. Nadeau, S. S. Conroy, J. M.
Powell, G. D. Huang and P. Peduzzi, Robot-assisted therapy for
long-term upper-limb impairment after stroke, N. Engl. J. Med.,
362: 1772-1783, 2010.
P. C. Kung, C. C. K. Lin, M. S. Ju and S. M. Chen, Reducing
abnormal synergies of forearm, elbow and shoulder joints in
stroke patients with neuro-rehabilitation robot treatment and
assessment, J. Med. Biol. Eng., 32: 139-146, 2012.
A. Bardorfer, M. Munih, A. Zupan and A. Primozic, Upper
limb motion analysis using haptic interface, IEEE-ASME Trans.
Mechatron., 6: 253-260, 2001.
M. Bergamasco, B. Allotta, L. Bosio, L. Ferretti, G. Parrini, G.
Prisco, F. Salsedo and G. Sartini, An arm exoskeleton system
for teleoperation and virtualenvironments applications, Proc.
IEEE Int. Conf. Robotics & Automation, 2: 1449-1454, 1994.
T. Nef, M. Mihelj, G. Colombo and R. Riener, ARMin-robot for
rehabilitation of the upper extremities, Proc. IEEE Int. Conf.
Robotics & Automation, 3152-3157, 2006.
M. Mihelj, T. Nef and R. Riener, ARMin II-7 DoF
rehabilitation robot: mechanics and kinematics, Proc. IEEE Int.
Conf. Robotics & Automation, 4120-4125, 2007.

[10] Z. Song and S. Guo, Design process of exoskeleton

rehabilitation device and implementation of bilateral upper limb
motor movement, J. Med. Biol. Eng., 32: 323-330, 2012.
[11] A. Gupta and M. K. O'Malley, Design of a haptic arm
exoskeleton for training and rehabilitation, IEEE-ASME Trans.
Mechatron., 11: 280-289, 2006.
[12] L. Dipietro, M. Ferraro, J. J. Palazzolo, H. I. Krebs, B. T. Volpe
and N. Hogan, Customized interactive robotic treatment for
stroke: EMG-triggered therapy, IEEE Trans. Neural Syst.
Rehabil. Eng., 13: 325-334, 2005.
[13] Y. Yong, W. Lan, T. Jie and Z. Lixun, Arm rehabilitation robot
impedance control and experimentation, Proc. IEEE Int. Conf.
Robotics & Biomimetics, 914-918, 2006.
[14] D. Formica, L. Zollo and E. Guglielmelli, Torque-dependent
compliance control in the joint space for robot-mediated motor
therapy, J. Dyn. Syst. Meas. Control-Trans. ASME, 128: 152158, 2006.
[15] H. I. Krebs, J. J. Palazzolo, L. Dipietro, B. T. Volpe and N. Hogan,
Rehabilitation robotics: Performance- based progressive
robot-assisted therapy, Auton. Robot., 15: 7-20, 2003.
[16] J. Furusho and T. Kikuchi, A 3-D rehabilitation system for
upper limbs EMUL, and a 6-DOF rehabilitation system
Robotherapist, and other rehabilitation system with high
safety, in: Sashi S Kommu (Ed.), Rehabilitation Robotics,
Vienna: I-Tech Education and Publishing, 115-136, 2007.
[17] H. Peralta, E. Monacelli, C. Riman, M. Baklouti, F. B. Ouezdou,
I. Mougharbel, I. Laffont and J. Bouteille, PLEIA: A
reconfigurable platform for evaluation of HCI acting, in: Sashi
S Kommu (Ed.), Rehabilitation Robotics, Vienna: I-Tech
Education and Publishing, 413-430, 2007.
[18] T. Miyoshi, Y. Takahashi, H. Lee, T. Suzuki and T. Komeda,
Upper limb neurorehabilitation in patients with stroke using
haptic device system: Reciprocal bi-articular muscle activities
reflect as a result of improved circle-drawing smoothness,
Disabil. Rehabil. Assist Technol., 5: 370-375, 2010.
[19] M. S. Ju, C. C. K. Lin, D. H. Lin, I. S. Hwang and S. M. Chen,
A rehabilitation robot with force-position hybrid fuzzy
controller: Hybrid fuzzy control of rehabilitation robot, IEEE
Trans. Neural Syst. Rehabil. Eng., 13: 349-358, 2005.
[20] S. M. Lewis, T. A. Jerde, C. Tzagarakis, M. A. Georgopoulos, N.
Tsekos, B. Amirikian, S. G. Kim, K. Ugurbil and A. P.
Georgopoulos, Cerebellar activation during copying geometrical
shapes, J. Neurophysiol., 90: 3874-3887, 2003.
[21] B. C. Tsai, W. W. Wang, L. C. Hsu, L. C. Fu and J. S. Lai, An
articulated rehabilitation robot for upper limb physiotherapy and
training, Proc. IEEE Int. Conf. Intelligent Robots & Systems,
1470-1475, 2010.
[22] K. M. Calder, M. J. Agnew, D. W. Stashuk and L. McLean,
Reliability of quantitative EMG analysis of the extensor carpi
radialis muscle, J. Neurosci. Methods, 168: 483-493, 2008.
[23] E. Todorov and M. I. Jordan, Optimal feedback control as a
theory of motor coordination, Nat. Neurosci., 5: 1226-1235,
2002.
[24] T. Flash and N. Hogan, The coordination of arm movements:
an experimentally confirmed mathematical model, J. Neurosci.,
5: 1688-1703, 1985.