IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 13, NO.

3, SEPTEMBER 2005 349

A Rehabilitation Robot With Force-Position

Hybrid Fuzzy Controller: Hybrid Fuzzy
Control of Rehabilitation Robot
Ming-Shaung Ju, Chou-Ching K. Lin, Dong-Huang Lin, Ing-Shiou Hwang, and Shu-Min Chen

Abstract—The goal of this study was to design a robot system for to the elbow while the stroke patients were performing volun-
assisting in the rehabilitation of patients with neuromuscular dis- tary tracking of constant velocity movements of their elbow.
orders by performing various facilitation movements. The robot The results indicated that both assistive and resistive loads im-
should be able to guide patient’s wrist to move along planned linear
or circular trajectories. A hybrid position/force controller incorpo- proved the voluntary performance. Cozens [4], using a single
rating fuzzy logic was developed to constrain the movement in the axis robot and torque control mode, applied resistive or assistive
desired direction and to maintain a constant force along the moving torque to the elbow for rehabilitation of patients with spasticity
direction. The controller was stable in the application range of and weakness. They measured electromyographic signals of bi-
movements and forces. Offline analyses of data were used to quan- ceps and triceps as an indicator of reciprocal activation and con-
titatively assess the progress of rehabilitation. The results show that
the robot could guide the upper limbs of subjects in linear and cir- cluded that the externally imposed forces had beneficial effects.
cular movements under predefined external force levels and apply Reinkensmeyer et al. [5] showed that a counterpoise assistive
a desired force along the tangential direction of the movements. controller that compensated for gravity and elbow passive prop-
Index Terms—Fuzzy control, hybrid position/force control, re- erties could partially improve reaching movements in patients
habilitation robot. with brain injuries.
One of the major difficulties in realizing rehabilitation by
robots is the controller design. Simple linear movement with
I. INTRODUCTION simple velocity profile is relatively easy to design. Yet, manual
treatments usually involve complex maneuvers with resistive
R EHABILITATION program is the main stay of treatment
for patients suffering from trauma, stroke, or spinal cord
injury. Conventionally, these programs rely heavily on the ex-
or assistive force imposed at specific points of the movement.
Circular or more complex movements with predefined im-
perience and manual manipulation of the therapists. Since the posing force are difficult to design. While conventional position
number of patients is large and the treatment is time consuming, controllers control position, not force, conventional force con-
it is a big advance if robots can assist in performing treatment. trollers regulate force without precise control of the movement
Noritsugu et al. [1] designed an arm-like robot for treating pa- trajectory. Another is the isotropic control or impedance con-
tients with trauma, and developed four modes of linear motion trol that maintains a constant endpoint stiffness and damping.
with impedance control to control the force during the move- Hybrid position/force controllers [6], [7], controlling position
ment. They used adaptive identification method to estimate the in one direction and force in the orthogonal direction, have the
stiffness of the affected limb as an indicator of recovery. Krebs advantage of simultaneously maintaining the desired movement
et al. [2] designed a planar robot with impedance control for trajectory and force for the planar movements. An alternative
guiding patients to make movements along the specified tra- approach was adopted in the active-constrained mode of mirror
jectories. These trajectories had different stimulatory effects on image movement enabler (MIME) system [8], [9]. The robot
elbow and shoulder joints. They showed beneficial effects of the provided a viscous resistance in the direction of desired move-
robot with qualitative evaluation scales. Ju et al. [3] added dif- ment and spring-like forces in all other directions. Thus, the
ferent constant external loads, by a robot in torque control mode, system had position control in one direction and velocity-de-
pendent force control in the other direction. The reaching guide
developed by Reinkensmeyer et al. [5], in a general sense,
Manuscript received January 9, 2003; revised August 17, 2003 and August also implemented position control in one direction and force
25, 2004; accepted: December 28, 2004. This work was supported in part by the control in the other direction. Yet, the position control was
Republic of China National Science Council under Contract NSC 89-214-B-006 provided by the hardware constraint. One additional problem
–057 and Contract NSC 87-2314-B-006-125-M08.
M.-S. Ju and D.-H. Lin are with the Department of Mechanical Engineering, in designing the controller is the uncertainty about the subject
National Cheng Kung University, Tainan 70101, Taiwan, R.O.C. (e-mail: and the nonlinear dynamics of the robot such as the Coulomb
[email protected]). friction in the mechanism. In a classical design process, the
C.-C. K. Lin is with the Department of Neurology, National Cheng Kung
University Hospital, Tainan 70101, Taiwan, R.O.C. control parameters are determined according to the system
I.-S. Hwang is with the Department of Rehabilitation, National Cheng Kung model and its parameters. For a rehabilitation program assisted
University Hospital, Tainan 70101, Taiwan, R.O.C. by a robot, the subject is also part of the dynamic system and
S.-M. Chen is with the Department of Physical Therapy, National Cheng
Kung University Hospital, Tainan 70101, Taiwan, R.O.C. the dynamics model of the subject is not as clear and invari-
Digital Object Identifier 10.1109/TNSRE.2005.847354 able as the mechanical system. To solve this problem, fuzzy
1534-4320/$20.00 © 2005 IEEE

Authorized licensed use limited to: Istanbul Medipol Universitesi. Downloaded on June 19,2022 at 20:09:00 UTC from IEEE Xplore. Restrictions apply.
350 IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 13, NO. 3, SEPTEMBER 2005

logic is incorporated in the current position and force control

algorithms. Fuzzy control [10]–[12] is known for coping with
nonlinear systems and systems that have uncertainty in its
parameters. Yet, a main problem associated with fuzzy control
is to demonstrate the global stability [13]. Many researchers
have tried to maintain the stability by fine tuning or modifying
a stable linear proportional, integral, and derivative (PID)
controller [14]–[16]. In the past, we successfully applied fuzzy
control to the above-knee prostheses and cycling exercise
by functional electrical stimulation [17]–[19]. Designing a
controller for rehabilitation robot is more difficult, because the
external disturbance itself is subjected to another unresolved
controller (the human control).
The main purpose of the current study is to construct a
robotic system specialized for rehabilitation of elbow and
shoulder. The robot, incorporating hybrid position/force control
and fuzzy logic, shall be able to assist the subject in performing
both passive and active movements along the designed trajec-
tories with specified loads. The actual position and force data
are collected simultaneously for offline analyses. The specific
goals of current work include robot construction with control
algorithm implementation, trajectory planning, preliminary
clinical evaluation, and performance analyses with subjects.

II. METHODS
A. Experimental Setup
The experimental setup consisted of a robot mechanism,
motion controller, position and force–torque sensors, and a
personal computer. With a wide strip attached to the robot
mechanism, the upper limb of the subject was suspended at a
horizontal position and the hand in pronation. The subject’s
wrist was attached to the force–torque transducer with a clamp
that allowed free pronation/supination. An elastic wrist–hand
orthosis maintained the wrist in a pronation position. A strap
supported the proximal forearm with the shoulder in 90
of abduction. For the command generation, the commands
calculated by the personal computer were converted into Fig. 1. Photograph of the robot-aided rehabilitation system with (a) a subject
and (b) a schematic drawing of robot mechanism.
analog signals through a D/A card (PCL726, Advantech Inc.,
http://www.advantech.com) and sent to drivers of two ac
motors of the robot. The inputs to the controller consisted of main advantage of the design was that when the physical config-
endpoint force and position information. Endpoint force was uration of the four links satisfied the relation of (1), the dynamic
detected by a force sensor (MC3A-S1000, Advanced Mechan- equation of the robot could be theoretically decoupled as (2)
ical Technology Inc., Watertown, MA) attached on the robot
between the wrist-fixing part and Link 4 [Fig. 1(b)]. The signal (1)
from the force sensor, amplified by a custom made amplifier,
was digitized by an A/D board (PCL818HD, Advantech Inc.,
http://www.advantech.com). The rotation of each alternating (2)
current (ac) motor was detected by an encoder attached to
the motor and counted by using a counter board (PCL833, where , , and were the masses and lengths to
Advantech Inc., http://www.advantech.com). From the joint the centers of mass of third and fourth links, respectively,
angles of the robot, one can calculate Cartesian coordinates of and were lengths of first and second links, respectively,
the end effector. The force and position data were used both for and were the input torques of the robot joints, respectively,
closed-loop control and for later offline analyses. and were constants, and and were the angles
As shown in Fig. 1(b), we adopted the five-bar-link drive of first and second links relative to the horizontal axis ( ),
mechanism [20] for constructing the robot. The motors were lo- respectively. Equation (2) meant that the two axes were inde-
cated at the base (origin of – coordinate system), in order pendent and it made the controller design process of simpler.
to minimize the inertial loading of the whole robot system. The The mechanism was driven by two ac motors (MSM022A1,

Authorized licensed use limited to: Istanbul Medipol Universitesi. Downloaded on June 19,2022 at 20:09:00 UTC from IEEE Xplore. Restrictions apply.
JU et al.: REHABILITATION ROBOT WITH FORCE-POSITION HYBRID FUZZY CONTROLLER 351

Fig. 2. (a) Definitions of the tangential direction and normal direction of movement. (b) Block diagram of the position/force hybrid control with fuzzy logics.

http://www.motionautomation.com) with a maximal torque empirically [22]. For the overall control scheme, we adopted
output of 96 Nm after adding a gear speed reducer (50:1). the hybrid position/force control [6], i.e., controlling force
During the design process, we found that friction had signif- in the tangential direction of the movement and controlling
icant effects on the robot performance and the friction torque position in the orthogonal direction [Fig. 2(a)]. The main goal
could be added to (2), as of control was to keep the hand of the subject in a predefined
track and to impose an assistant or resistant force along the
(3) moving direction. Since force was the control goal along the
moving direction, position could not be controlled and the
when subject could move with an individually preferred speed. In the
(4)
when scheme, the direction of force control ( ) was determined first
where and were torques to be compensated, was by a selection matrix ( ) and the direction of position control
a constant for viscous friction term and was a static for ( ) was determined accordingly by another selection matrix
Coulomb friction torque [21]. and were determined ( ), where is the identity matrix. [Fig. 2(b)]

Authorized licensed use limited to: Istanbul Medipol Universitesi. Downloaded on June 19,2022 at 20:09:00 UTC from IEEE Xplore. Restrictions apply.
352 IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 13, NO. 3, SEPTEMBER 2005

B. Position Control
Both the position and force controls were implemented in dis-
crete time and the sampling rate was 50 Hz. The end position
error along was calculated first and transformed into the
equivalent end force error ( ) by a predefined stiffness con-
stant ( ), i.e.,

(5)

where is the end position error along in one time step.

was transformed to the needed torque acting on joints by
the Jacobian matrix ( )

where (6)

and were the inputs to the fuzzy logic and were

regarded as error and error change in the logic, respectively.
In the fuzzification process, inputs ( and ) were scaled
and mapped to the five triangle-shaped membership func-
tions [Fig. 3(a)], namely, large negative (LN), small negative
(SN), zero (ZE), small positive (SP), and large positive (LP),
respectively. One membership degree was obtained for each
scaled input and membership function combination. Then, each
combination of mapped inputs (E and EC) activated one control
action ( , where , standing for LN, SN, ZE, SP, and LP,
was th activated control function) according to the inference
rule table [Fig. 3(b)]. The control rule table was determined
based on the step response of a presumed second-order system,
one of standard procedures for fuzzy logic control table de-
termination [23]. We set as the value at the center of base Fig. 3. (a) Membership functions and (b) the inference rule table for (c) the
position controller, the fuzzy inference, or decision process. Five membership
of corresponding membership function (e.g., for SN) and functions: LN: large negative (left thick line); SN: small negative (left dashed
as the minimum of the membership degrees of E and EC line); ZE: zero (solid line); SP: small positive (right dashed line); and LP:
large positive (right thick line) were constructed, where E was the input. Error
[Fig. 3(c)]. The output of the fuzzy logic was obtained in the
defuzzification process by using the center-of-gravity method
1
( ) and error change (  ) were projected to the membership function axis,
respectively. Projected results (E and E ) and the corresponding membership
degrees were obtained. Then, the inference rule determined B (the output of
=
the fuzzification process), where i LN, SN, ZE, SP, and LP.

(7) The force error along direction ( ) was transformed into

joint torque error ( ) by . and change of it ( ) were
fed to the fuzzy PI tuner. The fuzzy PI tuner, having five mem-
where was the final output of the position fuzzy controller bership functions in both input and output similar to the position
and indexed all the combinations of activated control actions controller described above, had separate fuzzy logics for P and
from the rule table [Fig. 3(b)]. I parts, and each operated in a similar way as in the position
controller. The outputs from the fuzzy logic ( for P part and
C. Force Control for I part) were linearly combined with the conventional PI
controller
Force control was implemented in the tangent direction of ac-
tual movement. By this convention, the robot was able to pro-
vide a desired resistive or assistive force to the subjects [24]. (8)
Since the properties of the environment (and/or the subject)
were unknown a priori, it was impossible to determine the op- where and were the parameters for the conventional PI
timal control parameters exactly. Therefore, we combined a con- controller, respectively, and was the final output of the force
ventional linear PI (proportional–integral) controller and a fuzzy control. The maximal values of and were set to be
PI tuner as the force controller. The purpose of the fuzzy PI tuner equal to and , respectively.
was to compensate for the nonlinear dynamics (joint friction) of The whole controller was realized by using C language in
the robot and the unknown disturbing force from the subject. DOS environment in a 486 personal computer.

Authorized licensed use limited to: Istanbul Medipol Universitesi. Downloaded on June 19,2022 at 20:09:00 UTC from IEEE Xplore. Restrictions apply.
JU et al.: REHABILITATION ROBOT WITH FORCE-POSITION HYBRID FUZZY CONTROLLER 353

G
Fig. 4. Block diagram of the force control model. represented the joint one
of the robot and other symbols were explained in the text.

Fig. 6. Planned trajectories including linear, circular, and figure eight

trajectories in both clockwise and counter-clockwise directions.

and adding the fuzzy logic, as well as by adding Gaussian dis-

turbance to excite the model. The Gaussian noises were used to
represent all the disturbances, including the friction.
The stability of the model was also investigated by the con-
ventional describing function analyses. According to Aracil and
Gordillo [25], the describing function of the fuzzy force PI-con-
troller could be defined as

Fig. 5. Root locus plot of the one joint model of the robot. Solid squares (10)
showed the corresponding root loci of the system with the chosen control
parameters.
where was assumed to a simple sinusoidal function
, both and were constants, and
D. Controller Stability Analyses
In order to analyze the stability of torque controller part, we (11)
employed both model simulation and conventional describing
function analyses. From (3), one decoupled robot joint with the
(12)
associated controller was modeled, as in Fig. 4. The transfer
function of the decoupled robot joint 1 ( ) could be expressed
as (13)

(9)

where and were defined in (3) and (4). It was assumed

that, during the robot movements, the robot links were in con- (14)
tinuous movement and could be neglected. was a con-
stant. First, we worked on the controller without the fuzzy logic and were the saturation limits of fuzzy logic for inte-
part. Since the contact stiffness ( ) depended on the interac- gration and proportional parts, respectively. As explained in the
tion with a subject, it was not a constant and could not be deter- above, and were equal to and , respectively. The
mined precisely. Accordingly, the optimal conventional PI con- condition for the first harmonic balance was
troller parameters could not be determined. We set as . The stability of the fuzzy controller was
and determined empirically. Fig. 5 showed the effects then visualized by drawing a series of curves of
of changing , when was assumed to be 1, on the system against .
root loci. The system was stable as long as was greater than
zero. The thick dots showed the roots of the system at the chosen E. Trajectory Planning
that we used for the rest of this article. Then the fuzzy logic We designed three types of movement trajectories, namely,
was added. The model was simulated by changing and linear, circular, and figure eight (Fig. 6). The normal linear

Authorized licensed use limited to: Istanbul Medipol Universitesi. Downloaded on June 19,2022 at 20:09:00 UTC from IEEE Xplore. Restrictions apply.
354 IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 13, NO. 3, SEPTEMBER 2005

Fig. 9. Results of describing function analyses. Dots showed the trajectories

of the controller due to the errors with frequencies ranging from 0.01 to 10 Hz
and amplitudes from 0.1 to 100 N. Solid lines showed the trajectory of G 0
with different K values.

G
Fig. 7. Step responses of the simplified robot ( ) with different control
parameters. Upper plot showed the effects of changing K and K , where
K =K = 0:06=0=006 (dashed–dotted line), 0.6/0/06 (solid line), 6/0.6
(dotted line) and 60/6 (dashed line), respectively. Lower plot demonstrated the
effect of adding the fuzzy logic, with K =K = 0:6=0=06.

Fig. 10. Robot-alone without carrying a subject (left column) and passive
(right column) linear movements of a normal subject. First row shows the
movement trajectories. Upper solid line is shifted 5 cm to the right and the lower
solid line is shifted 5 cm to the left for clarity. Dotted line is the target trajectory.
Upper solid line represents the results with the conventional PI controller.
Lower solid line represents the results using the hybrid fuzzy controller. Solid
lines in the second and third rows are the corresponding velocity trajectories
under hybrid fuzzy (second) and PI (third) controller respectively. Dotted line
is the target velocity.
Fig. 8. Step responses of the controllers under different magnitudes of
Gaussian disturbance. Solid lines: Classical PI controller plus the fuzzy logic. was a circular movement with a diameter of 28 cm. The figure
Dotted lines: Classical PI controller without the fuzzy logic.
eight movement trajectory consisted of two ellipses (14 28
cm) with their short axes in alignment in radial direction and
movement trajectory was an outward movement from the chest with opposite moving directions. Since the movement velocity
of 28 cm in length and the oblique linear movement trajectory was not controlled, in the design process, a trajectory was
was an oblique (45 ) outward movement from the chest of partitioned into 4000 segments by length along the trajectory
21.2 ( ) cm in length. The circular movement trajectory instead of the lapsed time [26].

Authorized licensed use limited to: Istanbul Medipol Universitesi. Downloaded on June 19,2022 at 20:09:00 UTC from IEEE Xplore. Restrictions apply.
JU et al.: REHABILITATION ROBOT WITH FORCE-POSITION HYBRID FUZZY CONTROLLER 355

Fig. 11. Results of a normal subject performing active linear movements. Left column is the movement trajectories and the right column shows the force
trajectories. Dotted lines are target paths and preset loads. Solid lines are actual trajectories and exerted forces.

F. Safety harder to keep a constant force level. The fuzzy logic helped to
reduce the errors.
Safety measures were implemented both in software and
Fig. 9 showed the results of analyses by describing function
hardware levels. The robot stopped when the movement mag-
method. The points in the right upper side of the solid curve
nitude or velocity exceeded the predefined limits and the
experimenter could stop the robot with an interruption com- ( ) were the disturbances that did not jeopardize the system
stability. The analyses showed that, with the implemented fuzzy
mand at any time. Four optical limit switches were employed
controller, the robot arm was stable in face of disturbance of fre-
to define the range of movement for both joints of the robot.
quencies from 10 to 0.01 Hz and magnitudes from 0.1 to 100 N.
The robot could also be stopped by shutting down the power
supply with an emergency button by the patient or the physical In other words, the robot arm would be stable in the expected
therapeutic maneuver ranges. The figure also showed that, as the
therapist.
contact stiffness ( ) increased, the range of stability would be
reduced.
III. RESULTS
B. Linear Passive and Active Movements
A. Advantages and Stability of Fuzzy Force Control
Fig. 10 showed the comparison of fuzzy and conventional PI
Fig. 7(a) showed the simulated step responses of four sets of controllers in robot-alone and passive left-oblique linear move-
and for the conventional PI controller in the absence of ments. Both the PI and the fuzzy controller could achieve the
fuzzy logic part. We chose and for the real desired trajectory. The performance of the fuzzy controller was
robot. Fig. 7(b) demonstrated the contribution of the fuzzy logic. even better in both conditions. In the robot-alone condition, the
For the combination of and that we chose, the fuzzy means and standard deviations of the mean path errors for re-
logic made the response faster while reduced the magnitude of peated trials were cm and cm, for
oscillation. the fuzzy and PI controllers, respectively. The difference was
In addition, the fuzzy controller also contributed to the re- statistically significant by -test ( ). On one normal
sistance to the external disturbance. Fig. 8 showed the step re- subject in passive linear movements, the means and standard de-
sponses of the robot in face of Gaussian disturbances of different viations of the mean path errors for repeated trials were
magnitudes. As the magnitude increased, the controllers became cm and cm for the fuzzy and PI controllers

Authorized licensed use limited to: Istanbul Medipol Universitesi. Downloaded on June 19,2022 at 20:09:00 UTC from IEEE Xplore. Restrictions apply.
356 IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 13, NO. 3, SEPTEMBER 2005

Fig. 12. Results of a normal subject performing circular and figure eight movements. First column (from left) shows the movement trajectories, the second
velocity, the third the force in the tangential direction, and the fourth the absolute value of force in the orthogonal direction of the movement. Arrows in the first
column indicate the direction of movement. Arrows in the fourth column mark the locations where the subjects need to change direction in the y axis.

respectively. The difference was also statistically significant by less than 1.25 cm during the whole trajectory in all four types
-test ( ). The velocity profile derived from the fuzzy of movements. Though the robot system did not regulate the
controller showed more frequent and smaller amplitude adjust- moving velocity, we instructed the subject to keep the move-
ments, in contrast to the coarser and less frequent adjustments ment at constant velocity as possible by projecting the Cartesian
in the case of conventional PI control. Though the adjustment coordinates against the desired movement on a monitor. The ve-
seemed very busy, the movement of the robot was relatively locity profile showed that the velocity was relatively constant at
smooth because of the natural damping effects of mechanical 10 cm/s. The control system kept a resistive load of 10 N. In
devices. the fourth column, the force along the direction normal to the
Fig. 11 showed the results of linear active movements of a movement, showed that the subject could follow the trajectory
normal subject against a load of ten Newtons along three direc- well, except when the subject needed to change direction in the
tions. The robot with the fuzzy controller could guide the sub- left-right axis (arrows).
ject to perform the movement well in all three directions and,
except at the beginning, kept the load constant as specified. On IV. DISCUSSION
one normal subject moving the right upper limb, the means and
There are several existing robot systems for rehabilitation, in-
standard deviations of the mean path errors for consecutive nine
cluding MIT-MANUS [2], MIME [8], and reaching guide [5].
trials in left-oblique, straight-forward, and right-oblique direc-
The basic construction principle of our robot and MIT-MANUS
tions were cm, cm, and is similar. Impedance control, implemented in MIT-MANUS,
cm, respectively. is stable and robust to the uncertainties due to physical contact.
Yet, the controller cannot control the position and force indepen-
C. Circular and Figure Eight Active Movements
dently. When the preset impedance is small, subjects can more
Fig. 12 showed the results of circular and figure eight ac- successfully complete movements at the expense of larger po-
tive movements of a normal subject against a load of ten New- sition errors. On the other hand, when the preset impedance is
tons. The trajectory was well maintained with the position error large, the position error is small but the subject is subjected to a

Authorized licensed use limited to: Istanbul Medipol Universitesi. Downloaded on June 19,2022 at 20:09:00 UTC from IEEE Xplore. Restrictions apply.
JU et al.: REHABILITATION ROBOT WITH FORCE-POSITION HYBRID FUZZY CONTROLLER 357

larger off-trajectory resistance. The force/position hybrid con- [6] M. H. Raibert and J. J. Craig, “Hybrid position/force control of manipu-
trol scheme developed in this work enables the robot to impose lators,” Trans. ASME J. Dyn. Syst., Meas. Control, vol. 102, pp. 126–133,
1981.
an assistive or resistive force to the subject, independent of the [7] H. Suh, J. H. Hong, S.-R. Oh, and K. B. Kim, “Fuzzy rule based
lateral position error. The fuzzy logic is intended to cope with position/force control of industrial manipulators,” in Proc. IEEE/RSJ
the nonlinear dynamics of the robot mechanism and to ensure Int. Workshop Intelligent Robots Syst., vol. 3, Nov. 3–5, 1991, pp.
1617–1622.
that the robot can operate for different subjects. The problem [8] P. S. Lum, C. G. Burgar, P. C. Shor, M. Majmundar, and H. F. M. Van der
of our control scheme is that stability is difficult to prove or Loos, “Robot-assisted movement training compared with conventional
predict. We did many model analyses to compare the perfor- therapy techniques for the rehabilitation of upper-limb motor function
after stroke,” Arch. Phys. Med. Rehab., vol. 83, no. 7, pp. 952–959, 2002.
mance and stability of the fuzzy controller with the original PI [9] C. G. Burgar, P. S. Lum, P. Shor, and H. F. M. Van der Loos, “Devel-
controller that the fuzzy logic is added. The fuzzy controller is opment of robots for rehabilitation therapy: The Palo Alto VA/Standord
experience,” J. Rehabil. Res. Develop., vol. 37, no. 6, pp. 663–673, 2000.
stable and has a faster response. MIME system is very flexible [10] L.-X. Wang, Adaptive Fuzzy Systems and Control: Design and Stability
in programming and can be operated in two-side or one-side Analysis. Englewood Cliffs, NJ: Prentice Hall, 1994.
mode. While the two-side mode, working by position control, [11] W. J. Rugh, “Analytical framework for gain scheduling,” IEEE Control
Syst. Mag., pp. 79–84, 1991.
is similar to our robot in providing assistive force, the one-side [12] S. H. Lee, S. G. Kim, and J. T. Lim, “Fuzzy-logic-based fast gain-sched-
mode in active-constrained mode is similar to our robot in pro- uling control for nonlinear suspension systems,” IEEE Trans. Ind. Elec-
viding resistive force. The other similar point is that both sys- tron., vol. 45, no. 6, pp. 953–955, Dec. 1998.
[13] K. Tanaka and H. O. Wang, Fuzzy Control Systems Design and Anal-
tems do not control velocity vigorously. The main difference ysis. New York: Wiley, 2001.
is that our robot provides a constant force and MIME provides [14] A.-A. S. Ibrahim, “Nonlinear PID controller design using fuzzy logic,”
a viscous-like resistance. As in MIME system, our robot can in IEEE MELECON 2002, Cairo, Egypt, May 7–9, 2002.
[15] U.-C. Moon and K. Y. Lee, “Hybrid algorithm with fuzzy system and
also be programmed to operate in other operation modes but we conventional PI control for the temperature control of TV glass furnace,”
have not fully explored these directions yet. Reaching guide has IEEE Trans. Control Syst. Technol., vol. 11, no. 4, pp. 548–554, Jul.
advantages in that the system is simpler and the hardware con- 2003.
[16] J. Deskur, R. Muszynski, and D. Sarnowski, “Tuning and investigation
straint is more solid. The disadvantage is that the trajectory is of combined fuzzy controller,” in Proc. 7th Int. Conf. Power Electron.
rigid and it may need different tracks for different treatment tra- Variable Speed Drives, Sep. 21–23, 1998.
jectories. [17] M. S. Ju, S. H. Yi, Y. G. Tsuei, and Y. L. Chou, “Fuzzy control of elec-
trohydraulic above-knee prostheses,” JSME Int. J., ser. C, vol. 38, pp.
The applicability of our robot to the patient with spasticity is 78–86, 1995.
currently unknown. The robot will be safer as compared with [18] M. S. Ju and D. L. Yang, “Design of adaptive fuzzy controls based on
the robots implemented with a pure position control. Stability natural control laws,” Fuzzy Sets Syst., vol. 81, pp. 191–204, 1995.
[19] J. J. J. Chen, Y. N. Yu, D. G. Hung, B. T. Ann, and G. C. Chang, “Ap-
of the force control part is a potential problem. A model of a plying fuzzy logic to control of cycling movement induced by functional
human subject with spasticity is necessary for performing ex- electrical stimulation,” IEEE Trans. Rehabil. Eng., vol. 5, pp. 158–169,
tensive simulations to test the stability of the man–robot system. 1997.
[20] H. Asada and K. Youcef-Toumi, “Analysis and design of a direct-drive
Initially, we will choose patients with progressively more severe arm with a five- bar-link parallel drive mechanism,” Trans. ASME J. Dyn.
spasticity and observe the performance. Syst. Meas. Control, vol. 106, pp. 255–230, Sep. 1984.
[21] S. M. Phillips and K. R. Ballou, “Friction modeling and compensation
for an industrial robot,” J. Robot Syst., vol. 10, pp. 947–971, 1993.
[22] Z. Lu, K. B. Shimoga, and A. A. Goldenberg, “Experimental determi-
V. CONCLUSION nation of dynamic parameters of robotic arms,” J. Robot Syst., vol. 10,
pp. 1009–1029, 1993.
We designed an arm-like robot with hybrid position/force [23] C. C. Lee, “Fuzzy logic in control systems: Fuzzy logic controller—Part
control scheme incorporating fuzzy logic. The controller was 1,” IEEE Trans Syst. Man Cybern., vol. 20, no. 2, pp. 402–417, Mar./Apr.
stable in the application range of movements and forces. The 1990.
[24] Z. Y. Chou, “Contact force control in soft boundary condition,” M.S.
preliminary results using normal and stroke subjects showed thesis (in Chinese), Nat. Yung-Lin Technol. College, Dept. Mech. Eng.
that the robot successfully guided the subject through linear and Technol., Toliu, Taiwan, 1997.
[25] J. Aracil and F. Gordillo, “Describing function method for stability anal-
circular movements. The robot could also impose a constant ysis of PD and PI fuzzy controllers,” Fuzzy Sets Syst., vol. 143, pp.
force on the subject during the movement. The next step is to 233–249, 2004.
evaluate the effectiveness of this robot in assisting patients in a [26] T. Yoshikawa, Foundations of Robotics: Analysis and Control. Cam-
bridge, MA: MIT Press, 1990.
larger scale rehabilitation program.

REFERENCES Ming-Shaung Ju was born in 1956. He received

the B.Sc. and M.Sc. degrees in mechanical en-
[1] T. Noritsugu and T. Tanaka, “Application of rubber artificial muscle ma- gineering from National Cheng Kung University
nipulator as a rehabilitation robot,” IEEE/ASME Trans. Mechatronics, (NCKU), Tainan, Taiwan, R.O.C., in 1978 and 1982,
vol. 2, no. 4, pp. 259–267, Dec. 1997. respectively, and the Ph.D. degree in mechanical
[2] H. I. Krebs, N. Hogan, M. L. Aisen, and B. T. Volpe, “Robot-aided neu-
engineering from Case Western Reserve University,
rorehabilitation,” IEEE Trans. Rehab. Eng., vol. 6, no. 1, pp. 75–87, Mar.
1998. Cleveland, OH, in 1986.
[3] M.-S. Ju, C.-C. K. Lin, J.-R. Chen, H.-S. Cheng, and C.-W. Lin, “Per- In 1987, he joined the Faculty of Mechanical
formance of elbow tracking under constant torque disturbance in stroke Engineering Department, NCKU, where he was an
patients and normal subjects,” Clinical Biomech., vol. 17, pp. 640–649, Associate Professor and became a Professor in 1993.
2002. He has worked in biomechanics, biocybernetics,
[4] J. A. Cozens, “Robotic assistance of an active upper limb exercise in mechatronics, and microelectromechanical systems for biomedical engineering
neurological impaired patients,” IEEE Tran. Rehab. Eng., vol. 7, no. 4, applications. He has served on many scientific committee boards in Taiwan,
pp. 254–256, Jun. 1999. namely, the Chinese Society of Mechanical Engineering, the Chinese Society
[5] D. Reinkensmeyer, C. Takahashi, and W. Timoszyk, “Evaluation of an of Biomedical Engineering and the Formosa Society of Biomechanics. He is
assistive controller for reaching following brain injury,” in Proc. 1st Joint also a member of the Man–Machine System Committee of the International
BMES/EMBS Conf., 1999, p. 631. Federation for the Theory of Machines and Mechanisms.

Authorized licensed use limited to: Istanbul Medipol Universitesi. Downloaded on June 19,2022 at 20:09:00 UTC from IEEE Xplore. Restrictions apply.
358 IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 13, NO. 3, SEPTEMBER 2005

Chou-Ching K. Lin received the B.S. degree in Ing-Shiou Hwang received the B.S. degree in phys-
medicine from National Yang-Ming University, ical therapy from National Taiwan University, Taipei,
Taipei, Taiwan, R.O.C., in 1988, and the M.Sc. and Taiwan, R.O.C., in 1988, the M.S. degree in biomed-
Ph.D. degrees in biomedical engineering from Case ical engineering from National Cheng Kung Univer-
Western Reserve University, Cleveland, OH, in 1994 sity, Tainan, Taiwan, R.O.C., in 1993, and the Ph.D.
and 1997, respectively. degree in biomedical engineering from the University
Since passing the Taiwan Neurology and Geron- of Texas, Austin, in 1998.
tology Boards in 1999 and 2003, he joined the In 1999, he was with the School of Physical
Department of Neurology, Medical Center, National Therapy, National Cheng Kung University as an
Cheng Kung University, Tainan, Taiwan, R.O.C., as Assistant Professor. He was also an Executive Editor
a Clinical Assistant Professor. His areas of interest of the Formosan Journal of Physical Therapy,
include functional electrical stimulation, neuro-rehabilitation, neural electro- official journal of the Physical Therapy Association of the R.O.C. (Taiwan),
physiology, and nerve regeneration. from 1999 to 2001. Since 2002, he has been an Associate Professor in the
Department of Rehabilitation, National Cheng Kung University. His major
research interests are in biosignal processing, clinical electrophysiology, and
rehabilitative neuroscience.

Shu-Min Chen received the M.D. degree from the

Department of Medicine, National Cheng Kung Uni-
versity (NCKU), Tainan, Taiwan, R.O.C., in 1992.
Dong-Huang Lin was born in 1974. He received She has served in the Department of Physical
the B.Sc. degree in mechanical engineering from Medicine and Rehabilitation, Medical Center,
National Central University, Taipei, Taiwan, R.O.C., NCKU since 1992. She has more than 10 years
in 1997, and the M.Sc. degree in mechanical en- of clinic experience in the field of rehabilitation.
gineering from National Cheng Kung University, In 1995, she served on the National Board of the
Tainan, Taiwan, R.O.C., in 2001. Rehabilitation Medicine Association, Taiwan. She
He has been worked on robotics for rehabilitation has been an Assistant Professor of Physical Medicine
engineering applications. and Rehabilitation, Medical Center, NCKU, since
2002. Her major interests include myofascial pain syndrome, neuromuscular
rehabilitation, and musculoskeletal sonography.

Authorized licensed use limited to: Istanbul Medipol Universitesi. Downloaded on June 19,2022 at 20:09:00 UTC from IEEE Xplore. Restrictions apply.