IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 8, NO.
1, MARCH 2003
An Exoskeletal Robot for Human Shoulder Joint
Motion Assist
Kazuo Kiguchi, Member, IEEE, Koya Iwami, Makoto Yasuda, Keigo Watanabe, Member, IEEE, and
Toshio Fukuda, Fellow, IEEE
Abstract—We have been developing exoskeletal robots in order
to assist the motion of physically weak persons such as elderly per-
sons or handicapped persons. In our previous research, a prototype
of a two degree of freedom exoskeletal robots for shoulder joint mo-
tion assist have been developed since the shoulder motion is espe-
cially important for people to take care of themselves in everyday
life. In this paper, we propose an effective fuzzy-neuro controller, a
moving mechanism of the center of rotation (CR) of the shoulder
joint of the exoskeletal robot, and intelligent interface in order to
realize a practical and effective exoskeletal robot for shoulder joint
motion assist. The fuzzy-neuro controller enables the robot to as-
sist any person’s shoulder motion. The moving mechanism of the
CR of the robot shoulder joint is used to fit the CR of the robot
shoulder joint to that of the physiological human shoulder joint
during the shoulder motion. The intelligent interface is realized by
applying a neural network and used to cancel out the effect the
human subject’s arm posture change. The effectiveness of the pro-
posed method has been evaluated by experiment.
Index Terms—Electromyogram (EMG) signals, exoskeleton,
human motion assist, soft computing.
I. INTRODUCTION
R ECENT progress in robotics and mechatronics technology
brings a lot of benefits not only in industries, but also in
welfare and medicine. We have been developing exoskeletal
robots [1]–[3] in order to assist the motion of physically weak
persons such as elderly persons or handicapped persons. It
is important that such physically weak people are able to
take care of themselves in the aging society. The exoskeletal
robots [4]–[7], which are sometimes called as exoskeletons,
power suits, man amplifiers, man magnifiers, or power assist
systems, have been mainly studied for the purpose of military
or industry use from the early 1960s. Since the design concept
is different, these robots were not suitable for physically weak
persons using in everyday life. On the other hand, active
orthotic systems [8], [9], which are similar to the exoskeletal
robots, also have been studied for the purposed of welfare
and medicine from the 1960s. In order to use these systems,
however, the users had to learn how to control the systems
because of the primitiveness of their controllers. In this paper,
we propose a two degrees of freedom (DOF) exoskeletal robot
Manuscript received November 1, 2002; revised December 21, 2002. Rec-
ommended by Technical Editor T. Nakamura. The work was supported by the
Tateisi Science and Technology Foundation, Japan under Grant 1021006.
K. Kiguchi, K. Iwami, M. Yasuda , and K. Watanabe are with the Department
of Advanced Systems Control Engineering, Saga University, Saga 840-8502,
Japan (e-mail: kiguchi@ieee.org).
T. Fukuda is with the Department of Micro System Engineering, Nagoya Uni-
versity, Chikusa-ku, Nagoya 464-8603, Japan.
Digital Object Identifier 10.1109/TMECH.2003.809168
1083-4435/03$17.00 © 2003 IEEE
125
and its control method for automatic shoulder motion assist
since human shoulder joints are involved in a lot of motion in
everyday life. The proposed exoskeletal robot is a modified
version of the previously proposed 2-DOF exoskeletal robot
prototype [3]. The architecture of the robot and the controller
are newly designed in this paper. The proposed exoskeletal
robot is automatically activated based on the human subject’s
electromyogram (EMG) signals which directly reflect the
muscle activity levels of human subject. The EMG signals are
important information to understand how the human subject
intends to move. Consequently, the EMG signals can be used
as input information for the robotic systems [10]–[12]. For
the exoskeletal robot in this study, seven kinds of the EMG
signals from the shoulder muscles of the human subject as
well as the shoulder joint angles are used as input information.
Thus the exoskeletal robot is able to assist the motion of the
human subject effectively by applying his/her EMG signals
as main input signals to the robot. Even though the EMG
signals contain very important information, it is not very easy
to predict the shoulder motion from the EMG signals in a short
time since many muscles are involved in the motion [13], [14].
Furthermore, it is difficult to obtain the same EMG signal for
the same motion even from the same person since the EMG
signal is a biologically generated signal. Moreover, the level
of the EMG signals might be much different between persons.
Therefore, the robot controller must have on-line adaptation
ability to the physiological condition of each human subject if
we apply the exoskeletal robots to several persons [15], [16].
In order to cope with this problem, a fuzzy-neuro controller,
which is able to adapt itself to the physiological condition of
each human subject on-line, is proposed for the controller of
the exoskeletal robot. The physiological control of the robot
can be realized with this control method.
On the other hand, the mechanism of the prototype of the ex-
oskeletal robot in our previous study was too simple in com-
parison with that of the human shoulder. In addition to this,
it is impossible to set the center of rotation (CR) of the robot
shoulder joint is the same as that of the human shoulder joint
(glenohumeral joint) since it is located inside of the human body.
Furthermore, human shoulder complex provides 7-DOF for the
arm movement since shoulder complex consists of the scapula,
clavicle, and humerus and moves conjointly [17]. Consequently,
the CR of the human shoulder joint is dislocated according to
the shoulder motion. Since the upper arm of the human sub-
ject is almost fixed to the arm holder of the exoskeletal robot
in our system, the subject must move his/her body instead of
his/her upper arm to adjust the location of the CR of the shoulder
126
Fig. 1. Architecture of the exoskeletal robot.
joint. Consequently, the human body in the exoskeletal robot
was sometimes forced to move back and forth or left and right
according to the shoulder motion since the CR of the robot
shoulder joint was fixed and different from that of the human
shoulder joint.
In this paper, we also propose a mechanism of the moving CR
of the shoulder joint of the 2-DOF exoskeletal robot for shoulder
motion assist in order to cancel out the ill effects caused by
the position difference of the CR between the robot shoulder
joint and the human shoulder joint during the shoulder motion.
The proposed mechanism makes the CR of the robot shoulder
joint mechanically move in accordance with the shoulder joint
motion. The linkwork mechanism has been applied to realize
the proposed mechanism. By the effect of the proposed shoulder
mechanism, the generated shoulder motion of the human subject
becomes smooth like the physiological shoulder motion.
In our previous study [3], the arm posture of the human sub-
ject was supposed to be the same at all times. Even though the
human subject tries to perform the same shoulder motion, how-
ever, the amount of the EMG signals from the shoulder muscles
varies [20] if the arm posture is changed since the disposition
of the shoulder muscles are changed [18], [19]. Therefore, the
arm posture has to be taken into account in order to carry out the
reliable shoulder motion assist. In this paper, we propose intel-
ligent interface between the human subject and the fuzzy-neuro
controller to cancel out the effect caused by subject’s arm pos-
ture difference. In this method, the fuzzy-neuro controller is ad-
justed instantly by the intelligent interface in accordance with
the human subject’s arm posture. The intelligent interface is re-
alized by applying a neural network.
The effectiveness of the proposed exoskeletal robot and its
control method has been evaluated by experiment with human
subjects.
II. EXOSKELETAL ROBOT
The architecture of the exoskeletal robot is shown in Fig. 1.
The exoskeletal robot consists of a frame, two main links, an
arm holder, two dc motor [Harmonic Drive System Company],
IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 8, NO. 1, MARCH 2003
drive wires, wire tension sensors (strain gauges), and the mech-
anism of the moving CR of the shoulder joint. The exoskeletal
robot worn by a human subject is supposed to assist the subject’s
shoulder joint motion (flexion-extension and abduction-adduc-
tion motions as shown in Fig. 2) by manipulating the subject’s
upper arm with the arm holder, which is fixed on the slider on
the link-2. The manipulation of the subject’s upper arm is car-
ried out by controlling the arm holder motion with dc motors
via driving wires. The inside of the arm holder is covered by an
air cushion in which air pressure is adjustable to fit any size of
upper arm. The flexibility of the air cushion softens the motion
difference of the robot and the human subject caused by the dif-
ference of the CR of the shoulder joints. Considering the fact
that many physically weak persons use a wheel chair, the heavy
parts of the proposed exoskeletal robot (i.e., the dc motors) are
installed in the chair, and the other parts of the exoskeletal robot
are directly attached to the human subject.
Human shoulder joint (glenohumeral joint) consists of
many muscles such as deltoid, biceps, triceps, pectoralis
major, infraspinatus, and teres major, and moves in 3-DOF
(flexion-extension, abduction-adduction, and internal-external
rotation). However, human shoulder complex provides 7-DOF
for the arm movement since shoulder complex consists of the
scapula, clavicle, and humerus and moves conjointly [17]. In
the case of shoulder flexion motion, the clavicle rotates about an
anteroposterior axis at the sternoclavicular joint. The rotation
results in the elevation and translation (to medial direction) of
the acromioclavicular joint with regard to the sternoclavicular
joint [21]. The translation of the sternoclavicular joint results in
the translation of the glenohumeral joint to the same direction.
Therefore, the CR of the glenohumeral joint is dislocated
according to the shoulder motion.
Since the human upper arm is almost fixed in the arm holder
of the robot, the relative distance between the arm holder and the
CR of the human shoulder joint is almost constant. Therefore,
the distance between the arm holder and the CR of the robot
shoulder joint must be moderately adjusted in accordance with
the shoulder motion, in order to cancel out the ill effects caused
by the position difference of the CR between the robot shoulder
and the human shoulder.
KIGUCHI et al.: EXOSKELETAL ROBOT FOR HUMAN SHOULDER MOTION ASSIST
(a)
(b)
(c)
Fig. 2. Assisted shoulder joint motion.
The proposed mechanism of the moving CR of the robot
shoulder, which consists of links and a slider as shown in Fig. 3,
is installed between the link-1 and the link-2 of the robot.
The motion of the proposed mechanism is depicted in Fig. 4.
The joint between the link-1 and the link-2 (i.e., the shoulder
joint of the exoskeletal robot) is supposed to be located at
just behind the armpit of the human subject. The proposed
mechanism makes the CR of the robot shoulder joint move
behind (farther position from the arm holder) in accordance
with the shoulder vertical flexion angle in the case of vertical
flexion motion, and move inward (closer position to the arm
holder) in accordance with the shoulder horizontal extension
127
Fig. 3. Proposed linkwork mechanism (abducted position).
(a)
(b)
Fig. 4. Motion of the proposed linkwork mechanism.
angle in the case of horizontal extension motion. The linkwork
mechanism has been applied to realize the proposed mecha-
nism. In the case of shoulder vertical flexion-extension motion,
the link-2 is vertically rotated with respect to the joint between
the link-1 and link-2. As the link-2 rotates vertically, the ad-
ditional link (the link for the slider) is rotated with respect to
another joint (joint-2). Note that the joint-2 is a universal joint.
The other end of the link for the slider is attached on the slider
on the link-2. Since the radius of the link-2 and the link for the
128 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 8, NO. 1, MARCH 2003

Fig. 5. Location of electrodes.

slider are different, the slider moves along the link-2 according
to the shoulder flexion angle. In the case of shoulder horizontal
flexion-extension motion, the link-1 is rotated about its axis ac-
cording to the horizontal flexion-extension angle. As the link-1
rotates, joint-3 is rotated with respect to the axis of the link-1.
The rotation of the joint-3 causes the movement of the posi-
tion of the joint-2 along the lateral-medial direction as shown in
Fig. 4. As the position of the joint-2 moves along the lateral-me-
dial direction, the slider moves along the link-2 since the link for
the slider is connected to the joint-2.
Usually, the limitation of human shoulder movable range is
180 in flexion, 60 in extension, 180 in abduction, and 75 Fig. 6. Architecture of the fuzzy-neuro controller.
in adduction. Considering the practical application to everyday
life, the shoulder motion limitation of the proposed robot is 0
in extension and adduction, 90 in flexion, and 90 in abduc-
tion this system. The maximum angular velocity of the motor is
limited by the hardware for safety. The maximum torque of the
robot (i.e., the maximum current of the motor) is also limited by
both the hardware and software for safety. Furthermore, there is
an emergency stop switch beside the robot.

III. CONTROL OF THE ROBOT

By adjusting the amount of force generated by the shoulder
muscles, the shoulder motion can be moderately controlled. The
muscle activity level can be described by the EMG signal. Con-
sequently, human intention of shoulder motion can be estimated
by observing the EMG signals of the shoulder muscles.
Fuzzy-neuro control, combination of fuzzy control and
adaptive neuro control, is applied to control the exoskeletal
robot. The initial fuzzy IF-THEN control rules of the fuzzy-neuro
control are designed based on the analyzed human subject’s
shoulder motion patterns in the experiment [3] and the experi-
mental results in another research [13], [14] assuming that the
arm posture of the human subject is in standard posture (i.e.,
shoulder rotation angle is neutral [0 ], elbow flexion/extension
angle is neutral [0 ], and arm pronation/supination angle is
neutral [0 ]).
The skin surface EMG signals of shoulder muscles, which
imply the human subject’s intention, and the shoulder joint an-
gles are used as input signals of the robot controller in order
to control the robot as the human subject intended. The loca-
tion of electrodes on shoulder muscles is shown in Fig. 5. The
electrodes are located on the anterior, posterior and middle part
Fig. 7. Teaching equipment.
of deltoid, biceps, triceps, pectoralis major (lateral part), teres
major, pectoralis major (clavicular part), and trapezius and those
are connected to ch.1, ch.2, ch.3, ch.4, ch.5, ch.6, and ch.7, re-
spectively.
The input variables of the fuzzy-neuro control are the mean
absolute value (MAV) [22] of EMG of seven kinds of muscles.
The equation of the MAV is written as
(1)
where is the voltage value at th sampling and is the
number of samples in a segment. The number of samples is set
to be 100 and the sampling time is set to be 0.5 ms in this study.
Four kinds of fuzzy linguistic variables (ZO: zero, PS: posi-
tive small, PM: positive medium, and PB: positive big) are pre-
pared for each MAV of EMG of main muscles (ch. 2, 4-6). Three
kinds of fuzzy linguistic variables (ZO: zero, PS: positive small,
KIGUCHI et al.: EXOSKELETAL ROBOT FOR HUMAN SHOULDER MOTION ASSIST 129

Fig. 8. Change of membership function.

and PB: positive big) are prepared for each MAV of EMG of
the other muscles (ch. 1, 3, and 7). Another three kinds of fuzzy
linguistic variables (LA: low angle, MA: medium angle, HA:
high angle) are prepared for each shoulder joint angle. The out-
puts of the fuzzy-neuro control are the torque command to gen-
erate the desired shoulder motion of the exoskeletal robot. The
torque command for the exoskeletal robot joints is then trans-
ferred to the force command for each driving wire. The relation
between the torque command for the exoskeletal robot joints and
the force command for driving wires is written as the following
equation:

(2)

where is the torque command vector for the exoskeletal robot Fig. 9. Effect of the arm posture change.
joints, is the force command vector for the driving wires,
and is the Jacobian which relates the exoskeletal robot joint
where is a threshold value and is a weight. The process
velocity to the driving wire velocity. Force control is carried out
of the fuzzy-neuro controller is the same as that of ordinal sim-
to realize the desired force ( ) in driving wires by the driving
plified fuzzy controllers. Consequently, the output of the fuzzy-
motors.
neuro controller is calculated with the following equation:
In the fuzzy-neuro controller, 32 kinds of fuzzy IF-THEN rules
are prepared to generate the desired torque of the exoskeletal
robot. The architecture of the fuzzy-neuro controller is depicted (7)
in Fig. 6. Here means sum of the inputs, means multi-
plication of the inputs. Two kinds of nonlinear functions (
and ) are applied to express the membership function of the where represents the output vector, denotes the degree of
fuzzy-neuro controller. fitness of the rule, and is the weight for the rule.
When the exoskeletal robot is attached to another human sub-
(3) ject, or when physiological condition of the human subject is
changed a lot, on-line adaptation of fuzzy-neuro controller is
(4) carried out by adjusting each weight of the fuzzy-neuro to min-
(5) imize the amount of muscle activity and motion error which is
given by the teaching equipment shown in Fig. 7. The angle of
(6) link-1 of the teaching equipment is supposed to correspond to
130 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 8, NO. 1, MARCH 2003

Fig. 10. Architecture of the intelligent interface.

that of link-1 of the exoskeletal robot, and the angle of link-2

of the teaching equipment is supposed to correspond to that
of link-2 of the exoskeletal robot. In the adaptation process,
the human subject indicates his/her desired shoulder motion by
demonstrating the same motion with the teaching equipment
using his/her wrist. In this study, both the antecedent part and the
consequent part of the fuzzy IF-THEN control rules are supposed
to be adjusted to fit physiological condition of each human sub-
ject by using the back-propagation learning algorithm in online
manner. The evaluation function for the fuzzy-neuro controller
training is written as:

(8)

where is the desired shoulder angle indicated by the teaching

equipment, is the measured shoulder angle, is a coefficient
which changes the degree of consideration of the muscle activity
minimization, is the desired muscle activity level in
ch.i, and is the measured muscle activity level in ch.i.
The assistance level of the robot can be moderately adjusted by
changing the desired muscle activity levels. Note that certain de-
sired muscle activity levels are prepared for each shoulder mo-
tion (i.e., vertical shoulder flexion/extension motion and hori-
zontal shoulder flexion/extension motion) considering physio-
logical muscle allocation.
The inputs to the fuzzy-neuro controller are instantly adjusted
by the modification coefficients outputted from the intelligent
interface, which is explained in the next section, in accordance
with the human subject’s arm posture. The definition of mem-
bership functions of input variables to the fuzzy-neuro controller
is adjusted immediately by multiplying the input variables by
the modification coefficients. This operation makes the same
Fig. 11. Experimental setup.
effect as changing the membership functions wider or narrower
[23] as shown in Fig. 8.
IV. INTELLIGENT INTERFACE
Human shoulder joint consists of many muscles and moves
in 3-DOF (flexion-extension, abduction-adduction, and in-
ternal-external rotation). The muscle activity level can be
described by the EMG signal. Even though the human subject
tries to perform the same shoulder motion, the amount of
the EMG signals from the shoulder muscles varies if the arm
posture is changed because of physiological reason [20]. The
displacement of the shoulder muscles is geometrically changed
if the arm posture is changed.
KIGUCHI et al.: EXOSKELETAL ROBOT FOR HUMAN SHOULDER MOTION ASSIST
(a)
(b)
Fig. 12. Experimental results for Trajectory 1 with assist of the exoskeletal
robot (standard arm posture).
In this study, intelligent interface is proposed to take into
account the subject’s arm posture. The intelligent interface
modifies the controller inputs by multiplying modification co-
efficients according to the subject’s arm posture. A neural net-
work is used to realize the intelligent interface. Preliminary
experiment, which investigates the effect of subject’s arm pos-
ture (the effect of shoulder vertical flexion/extension angle,
shoulder horizontal flexion/extension angle, shoulder internal
rotation angle, elbow flexion/extension angle, and forearm
pronation/supination angle) with respect to the amount of EMG
signals of shoulder muscles, was performed to prepare the
training data of the neural network. An example of the effect
the arm posture change is shown in Fig. 9. The neural network
makes a nonlinear mapping between subject’s arm posture and
modification coefficients for the controller inputs by off-line
learning. The architecture of the neural network is depicted in
131
(a)
(b)
Fig. 13. Experimental results for Trajectory 1 without assist of the exoskeletal
robot (standard arm posture).
Fig. 10. This neural network realizes the intelligent interface
between the human subject and the fuzzy-neuro controller.
The neural network consists of three layers (input layer,
hidden layer, and output layer). The input layer consists of
five neurons, the hidden layer 50 neurons, and the output
layer eight neurons. Sigmoid function is used for neurons
in the hidden layer and the output layer. The input variables
to the neural network are shoulder flexion/extension angle,
shoulder horizontal flexion/extension angle, shoulder internal
rotation angle, elbow flexion/extension angle, and forearm
pronation/supination angle. The outputs from the neural net-
work are the modification coefficients for input variables to
the fuzzy-neuro controller (i.e., EMG signals of the shoulder
muscles).
The neural network has been trained with the training data
(91,000 data set) obtained from the preliminary experimental
132
(a)
(b)
Fig. 14. Experimental results for Trajectory 2 with assist of the exoskeletal
robot (standard arm posture).
results using the back-propagation learning algorithm in off-line
manner.
V. EXPERIMENT
Experiment has been carried out with a health human
male subject (22 years old) to evaluate the effectiveness of
the proposed exoskeletal robot and its control method. The
experimental setup is depicted in Fig. 11. The amplified EMG
signals are sampled at a rate of 2 kHz and the signals from the
wire tension sensors are also sampled at a rate of 2 kHz and
low-pass filtered at 8 Hz.
For the first experiment, the target following experiments
have been carried out with and without assist of the exoskeletal
robot in order to verify the controllability of the exoskeletal
robot. When the experiment without assist of the exoskeletal
robot was performed, the human subject wore the exoskeletal
robot to measure the shoulder angle. We had experimentally
verified that wearing the robot did not affect the EMG signals
IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 8, NO. 1, MARCH 2003
(a)
(b)
Fig. 15. Experimental results for Trajectory 2 without assist of the exoskeletal
robot (standard arm posture).
of the human subject. This experiment has been performed
with the standard arm posture of the human subject (without
changing the arm posture of the human subject). The initial
control rules of the fuzzy-neuro controller were designed
assuming the arm posture of the human subject was in standard.
In the experiment, the target trajectory (Trajectory 1: both
vertical flexion-extension and horizontal flexion-extension
trajectory, Trajectory 2: vertical flexion-extension trajectory at
the horizontal flexion angle 30 ) of shoulder is displayed on the
monitor, and a human subject is supposed to make his shoulder
angles follow it. The generated shoulder trajectory is supposed
to be very close to the target trajectory if the exoskeletal robot
is well controlled, and the EMG levels of shoulder muscles are
supposed to be lower if the exoskeletal robot effectively assists
the shoulder motion of the human subject. The experimental
results (EMG signals at the anterior and middle part of deltoid)
of the human subject for the Trajectory 1 with and without assist
of the exoskeletal robot are shown in Figs. 12 and 13, and those
for the Trajectory 2 with and without assist of the exoskeletal
KIGUCHI et al.: EXOSKELETAL ROBOT FOR HUMAN SHOULDER MOTION ASSIST
(a)
(b)
Fig. 16. Experimental results for Trajectory 2 without assist of the exoskeletal
robot (changed arm posture).
robot are shown in Figs. 14 and 15, respectively. These results
show that the human subject can follow the target trajectories
with and without support of the exoskeletal robot. This means
the exoskeletal robot does not constrain the subject’s shoulder
motion. One can also see that the EMG levels of shoulder
muscles become lower when the shoulder motion of the subject
is assisted by the exoskeletal robot. Here, muscle activity level
(MAV) at the anterior part of deltoid was reduced to 18% and
15% for the Trajectory 1 and 2, respectively. Muscle activity
level (MAV) at the middle part of deltoid was reduced to 15%
and 60% for the Trajectory 1 and 2, respectively. These results
show the effectiveness of the proposed robot in human shoulder
motion assist in the case when the arm posture of the human
subject is in standard.
When the arm posture of the human subject is changed
(shoulder horizontal flexion/extension angle: 30 , shoulder in-
ternal rotation angle: 0 , elbow flexion angle: 90 , and forearm
pronation/supination angle: 0 ) from the standard posture, the
133
(a)
(b)
Fig. 17. Experimental results for Trajectory 2 with assist of the exoskeletal
robot (changed arm posture) without the proposed interface.
EMG levels of shoulder muscles during the shoulder vertical
flexion-extension motion at the horizontal flexion angle 30
(Trajectory 2) are changed as shown in Fig. 16. If we apply
the exoskeletal robot for motion assist with the fuzzy-neuro
controller designed for the standard arm posture without the
proposed interface (neural network) to this case, the exoskeletal
robot does not work very efficiently as shown in Fig. 17. In
this case, the EMG levels of shoulder muscles are not so much
improved and the target following becomes worse than those
without assist of the exoskeletal robot. Here, muscle activity
levels (MAV) at the anterior and middle part of deltoid were
reduced to only 57% and 58%, respectively. However, if the
fuzzy-neuro controller is modified by the proposed interface,
the results become better as well as those in Fig. 14 as shown
in Fig. 18. Here, muscle activity levels (MAV) at the anterior
and middle part of deltoid were reduced to 18% and 31%, re-
spectively. These results show the importance and effectiveness
of the proposed interface.
134
posture and modification coefficients for the fuzzy-neuro con-
troller inputs was examined in the preliminary experiment and
then off-line trained by a neural network to realize the intelli-
gent interface. The effectiveness of the proposed method has
been verified by the experiment with healthy human subjects.
We would like to apply the proposed system to elderly persons
and handicapped persons for the future research.
(a)
(b)
Fig. 18. Experimental results for Trajectory 2 with assist of the exoskeletal
robot (changed arm posture) with the proposed interface.
VI. CONCLUSION
In order to realize a practical and effective exoskeletal robot
for shoulder joint motion assist, an effective fuzzy-neuro con-
troller, a moving mechanism of the CR of the shoulder joint of
the exoskeletal robot, and intelligent interface have been pro-
posed in this paper. A moving mechanism of the CR of the
shoulder joint of the exoskeletal robot has been proposed to
fit the CR of the robot shoulder joint to that of the physio-
logical human shoulder joint during the shoulder motion. An
effective fuzzy-neuro controller has been also proposed to au-
tomatically control the robot with EMG signals of the human
shoulder muscles. Furthermore, an intelligent interface between
the human subject and the fuzzy-neuro controller has been pro-
posed to cancel out the ill effects caused by subject’s arm pos-
ture in this paper. The effective control of the exoskeletal robot
for human shoulder motion assist can be expected by realizing
the intelligent interface. A nonlinear map between subject’s arm
IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 8, NO. 1, MARCH 2003
REFERENCES
[1] K. Kiguchi, S. Kariya, K. Watanabe, K. Izumi, and T. Fukuda, “An ex-
oskeletal robot for human elbow motion support – Sensor fusion, adap-
tation, and control,” IEEE Trans. Syst., Man, Cybern., pt. B, vol. 31, pp.
353–361, June 2001.
[2] , “Application of multiple fuzzy-neuro controllers of an exoskeletal
robot for human elbow motion support,” in Proc. 32nd Int. Symp.
Robotics (ISR2001), 2001, pp. 1563–1568.
[3] K. Kiguchi, K. Iwami, T. Saza, S. Kariya, K. Watanabe, K. Izumi, and
T. Fukuda, “A study of an exoskeletal robot for human shoulder mo-
tion support,” Proc. IEEE/RSJ Int. Conf. Intelligent Robots and Systems
(IROS2001), pp. 2111–2116, 2001.
[4] R. S. Mosher and B. Wendel, “Force-reflecting electrohydraulic servo-
manipulator,” Electro-Technol., pp. 138–141, Dec. 1960.
[5] W. Cloud, “Man amplifiers: Machines that let you carry a ton,” Popular
Sci., vol. 187, no. 5, pp. 70–73, 1965.
[6] R. S. Mosher, “Handyman to Hardiman,” Soc. Auto. Eng., publication
MS670 088, 1967.
[7] H. Kazerooni and S. L. Mahoney, “Dynamics and control of robotic sys-
tems worn by humans,” Trans. ASME, J. Dyn. Syst., Meas. Control, vol.
113, no. 3, pp. 379–387, 1991.
[8] V. L. Nickel, A. Karchak Jr., and J. R. Allen, “Electrically powered
orthotic systems,” J. Bone and Joint Surgery, vol. 51-A, no. 2, pp.
343–351, 1969.
[9] N. Benjuya and S. B. Kenney, “Hybrid arm orthosis,” J. Prosthetics and
Orthotics, vol. 2, no. 2, pp. 155–163, 1990.
[10] O. Fukuda, T. Tsuji, A. Ohtsuka, and M. Kaneko, “EMG-based human-
robot interface for rehabilitation aid,” in Proc. IEEE Int. Conf. Robotics
and Automation, 1998, pp. 3942–3947.
[11] D. Nishikawa, W. Yu, H. Yokoi, and Y. Kakazu, “EMG prosthetic hand
controller using real-time learning method,” in Proc. IEEE Int. Conf.
Systems, Man, and Cybernetics, 1999, pp. I-153–I-158.
[12] H. Huang and C. Chen, “Development of a myoelectric discrimination
system for a multi-degree prosthetic hand,” in Proc. IEEE Int. Conf.
Robotics and Automation, 1999, pp. 2392–2397.
[13] B. Laursen, B. R. Jensen, G. Nemeth, and G. Sjogaad, “A model pre-
dicting individual shoulder muscle forces based on relationship between
electromyographic and 3D external forces in static position,” J. Biome-
chanics, vol. 31, no. 8, pp. 731–739, 1998.
[14] A. T. C. Au and R. F. Kirsch, “EMG-based prediction of shoulder and
elbow kinematics in able-bodied and spinal cord injured individuals,”
IEEE Trans. Rehab. Eng., vol. 8, pp. 471–480, Dec. 2000.
[15] T. Tsuji, Y. Kawaguchi, and M. Kaneko, “An adaptive training method
for human-robot systems using neural networks” (in Japanese), J. Robot.
Soc. Jpn., vol. 18, no. 5, pp. 683–389, 2000.
[16] S. Fujii, D. Nishikawa, and H. Yokoi, “Development of prosthetic hand
using adaptive control method for human characteristics,” in Intell.
Auton. Syst.. Amsterdam, The Netherlands: IOS Press, 1998, pp.
360–367.
[17] V. M. Zatsiorsky, Kinematics of Human Motion. Champaign, IL:
Human Kinetics, 1998.
[18] H. G. Coury, S. Kumar, and Y. Narayan, “An electromyographic study of
upper limb adduction force with varying shoulder and elbow postures,”
J. Electromyography and Kinesiol., vol. 8, pp. 157–168, 1998.
[19] H. Graichen, K.-H. Englmeier, M. Reiser, and F. Eckstein, “An in vivo
technique for determining 3-D muscular moment arms in different joint
positions and during muscular activation,” Clinical Biomechan., vol. 16,
pp. 389–394, 2001.
[20] R. L. Lieber, “Skeletal muscles is a biological example of a linear
electro-active actuator,” presented at the SPIE 6th Annu. Int. Symp.
Smart Structures and Materials, San Diego, CA, Mar. 1999.
[21] Z. Dvir and N. Berme, “The shoulder complex in elevation of the arm:
A mechanism approach,” J. Biomechan., vol. 11, pp. 219–225, 1978.
KIGUCHI et al.: EXOSKELETAL ROBOT FOR HUMAN SHOULDER MOTION ASSIST
[22] B. Hudgins, P. Parker, and R. N. Scott, “A new strategy for multifunction
myoelectric control,” IEEE Trans. Biomed. Eng., vol. 40, pp. 82–94, Jan.
1993.
[23] K. Kiguchi and T. Fukuda, “Intelligent position/force controller for in-
dustrial robot manipulators application of fuzzy neural networks,” IEEE
Trans. Ind. Electron., vol. 44, pp. 753–761, Dec. 1997.
Kazuo Kiguchi (S’92–M’93) received the B.E.
degree in mechanical engineering from Niigata
University, Niigata, Japan in 1986, the Master of
Applied Science degree in mechanical engineering
from the University of Ottawa, Ottawa, ON, Canada,
in 1993, and the Ph.D. degree in engineering from
Nagoya University, Nagoya, Japan, in 1997.
From 1986 to 1989, he was a Research Engineer
with Mazda Motor Company and from 1989 to 1991,
with MHI Aerospace Systems Company. From 1994
to 1999, he was with the Department of Industrial and
Systems Engineering, Niigata College of Technology, Niigata, Japan. Currently,
he is an Associate Professor in the Department of Advanced Systems Control
Engineering, Graduate School of Science and Engineering, Saga University,
Saga, Japan. His research interests include biorobotics, intelligent robots, ma-
chine learning, application of soft computing for robot control, and application
of robotics in medicine.
Dr. Kiguchi is a Member of the Robotics Society of Japan, IEEE Robotics
and Automation Society, IEEE Systems, Man, and Cybernetics Society, IEEE
Engineering in Medicine and Biology Society, IEEE Industrial Electronics So-
ciety and IEEE Computer Society, the Japan Society of Mechanical Engineers,
the Society of Instrument and Control Engineers, the Japan Society of Computer
Aided Surgery, International Neural Network Society, Japan Neuroscience So-
ciety, the Virtual Reality Society of Japan, the Japanese Society of Prosthetics
and Orthotics, and the Japanese Society for Clinical Biomechanics and Related
Research. He received the J. F. Engelberger Best Paper Award at WAC2000.
Koya Iwami was born on November 23, 1978. He
received the B.E. degree from Saga University, Saga,
Japan, in 2001, where he is currently pursuing the
M.E. degree in the Department of Advanced Systems
Control Engineering.
Makoto Yasuda was born on February 3, 1979. He
received the B.E. degree from Saga University, Saga,
Japan, in 2002, where he is currently pursuing the
M.E. degree in the Department of Advanced Systems
Control Engineering.
135
Keigo Watanabe (S’83–M’90) received the B.E.
and M.E. degrees in mechanical engineering from
the University of Tokushima, Tokushima, Japan, in
1976 and 1978, respectively, and the D.E. degree in
aeronautical engineering from Kyushu University,
Fukuoka, Japan, in 1984.
From 1980 to March 1985, he was a Research
Associate at Kyushu University. From April 1985
to March 1990, he was an Associate Professor at
the College of Engineering, Shizuoka University,
Shizuoka, Japan. From April 1990 to March 1993, he
was an Associate Professor, and from April 1993 to March 1998, he was a Full
Professor in the Department of Mechanical Engineering, Saga University, Saga,
Japan. Since April 1998, he has been with the Department of Advanced Sys-
tems Control Engineering, Graduate School of Science and Engineering, Saga
University. He has published more than 360 technical papers in transactions,
journals, and international conference proceedings, and is the author or editor
of 18 books, including Adaptive Estimation and Control (Englewood Cliffs,
NJ: Prentice-Hall, 1991), Stochastic Large-Scale Engineering Systems (New
York: Marcel Dekker, 1992) and Intelligent Control Based on Flexible Neural
Networks (Norwell, MA: Kluwer, 1999). He is an Active Reviewer of many
journals and transactions, and an Editor-in-Chief of Machine Intelligence and
Robotic Control, and an Editorial Board Member of the Journal of Intelligent
and Robotic Systems and the Journal of Knowledge-Based Intelligent Engi-
neering Systems. His research interests are in stochastic adaptive estimation
and control, robust control, neural network control, fuzzy control, and genetic
algorithms and their applications to machine intelligence and robotic control.
Dr. Watanabe is a Member of the Society of Instrument and Control Engi-
neers, Japan Society of Mechanical Engineers, Japan Society for Precision En-
gineering, Institute of Systems, Control and Information Engineers, the Japan
Society for Aeronautical and Space Sciences, Robotics Society of Japan, and
Japan Society for Fuzzy Theory and Systems.
Toshio Fukuda (M’83–SM’93–F’95) graduated
from Waseda University, Tokyo, Japan, in 1971 and
received the M.S. and Dr. Eng. degrees from the
University of Tokyo, Tokyo, Japan, in 1973 and
1977, respectively. He also studied at the Graduate
School of Yale University, New Haven, CT, from
1973 to 1975.
In 1977, he joined the National Mechanical
Engineering Laboratory, Tsukuba, Japan, and from
1979 to 1980, became a Visiting Research Fellow
at the University of Stuttgart, Stuttgart, Germany. In
1982, he joined the Science University of Tokyo, Tokyo, Japan, and in 1989,
joined Nagoya University where he is currently a Professor in the Department
of Micro System Engineering, Department of Mechano-Informatics and
Systems, engaging in the research of intelligent robotic systems, cellular
robotic systems, mechatronics and micro- and nanorobotics. He is an author of
six books, editor of five books, and has published over 1,000 technical papers
in micro- and nanosystems, robotics, mechatronics, and automation.
Dr. Fukuda has been a Fellow of the Society of Instrument and Control
Engineers (SICE), since 1995. He received the IEEE Eugene Mittlemann
Award in 1997, the Banki Donat Medal from the Polytechnic University of
Budapest, Budapest, Hungary in 1997, the Medal from the City of Sartillo,
Sartillo, Mexico in 1998, and the IEEE Millennium Medal, in 2000. He was
the Vice President of the IEEE Industrial Electronics Society (IES) from 1990
to 1999, IEEE Neural Network Council Secretary since 1992, Vice President
of the International Fuzzy Systems Association (IFSA) since 1997, President
of IEEE Robotics and Automation Society President from 1998 to 1999,
Editor-in-Chief of the IEEE/ASME TRANSACTIONS ON MECHATRONICS from
2000 to 2002, Director of the IEEE Division X from 2001 to 2002, and the
Founding President of the IEEE Nanotechnology Council since 2002.