You are on page 1of 11

156 IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 17, NO.

2, APRIL 2009

The Development of Two Mobile Gait


Rehabilitation Systems
Kap-Ho Seo, Member, IEEE, and Ju-Jang Lee, Fellow, IEEE

Abstract—The ability to walk without the help of a caretaker en- Devices for gait rehabilitation require specific weight un-
hances the quality of life for those who are bed-ridden or confined loading and walking speed adjustment with respect to the
to a wheelchair. At present, most of the available gait rehabilitation degree of disability [5]–[7]. In medical centers or rehabilitation
robot systems have been designed to support the body weight ex-
ternally. For gait training to be effective, a mobile body weight sup- facilities, nurses and physiotherapists assist gait training for
port (BWS) mechanism is needed. In mobile gait training robot sys- subjects with the traditional tools for weight support such
tems, functions such as patient path following and constant BWS as canes and crutches in a static condition to improve their
are important issues, particularly in dynamic environments. In the muscular activation capability and their sense of balance. Al-
present study, two types of robotic systems were developed for gait though these efforts improve the general condition of patients,
rehabilitation. The first is known as the mobile manipulator type
and the second the mobile vehicle type. The differences between
some gait anomalies nonetheless remain after treatment. With
the two systems in design and control are discussed. A control al- the traditional tools, the amount of weight support is neither
gorithm based on a neural network was used to compensate for constant nor quantifiable because it depends on the strength
dynamic interactions, unmodeled dynamics, and disturbances by of the patient as well as their degree of control of their upper
the user on the system. Both electrical and pneumatic BWS mech- limbs and trunk.
anisms were built and compared. The proposed BWS systems were
At present, walking rehabilitation systems with body weight
tested experimentally for their effectiveness in gait rehabilitation
while maximizing the therapeutic outcome. support (BWS) have been recommended particularly to suit dy-
namic conditions. In these systems, the subject walks on a tread-
Index Terms—Body weight support (BWS), gait rehabilitation,
mill at a comfortable speed with partially unloaded body weight
mobile manipulator, RBF neural network.
[7]–[14]. This method has been tested with healthy subjects with
up to 70% of their body weight being supported. Patients who
I. INTRODUCTION receive this therapy have shown a significant increase in walking
ability and speed. Treadmill systems may not be effective in re-
HE AVERAGE lifetime of humans has been increasing, as
T a consequence the demand for devices to support elderly
and disabled people is also increasing. Recent improvements in
habilitation because walking on a moving belt is not ecological.
There is a need to support actual overground walking with a re-
habilitation robot [15]–[20]. Here we focus on the function of
living conditions have increased the population of elderly and weight unloading during gait training [21].
disabled people [1]. Among many daily activities, walking is In 1985, BWS was proposed as a means to facilitate the
the most basic and fundamental activity which enables a human expression of gait patterns as a therapeutic tool in the gait
being to enjoy his life. Previous efforts to help partially immo- retraining of neurologically impaired patients [22]. In the last
bilized people to walk without aid has led to the development years, the BWS method has gained more reputation within the
of gait rehabilitation devices. Immobilization causes profound rehabilitation community as a convenient means of coping with
changes in muscle and bone masses, particularly in the lower the difficulties of walking reeducation. The currently available
limbs. Patients thus afflicted, especially the elderly, are likely to BWS systems can be subdivided into four groups based on
incur additional risk of injury upon reambulation [2]. Early mo- the technology used to exert the assistance force. They can be
bilization has been proven useful in preventing contracture of driven electrically, pneumatically, hydraulically, or mechan-
the lower limbs [3]. This approach may avoid or reduce several ically [8]. A popular design by Barbeau et al. [7] allowed
drawbacks associated with pathologies, including osteoporosis, the patients to walk on a treadmill while being mechanically
spasticity, and bedsores. As these subjects are often unable to supported using a harness to compensate for the body weight.
exert the muscular force necessary to maintain their posture and The gait training robot AID-1 [5] reduces the body weight
ambulate, it is essential to provide them with partial BWS and by supporting the patient at the trunk using pneumatic lifting
protect them from falling [4]. force. The weight relieving device REHABOT was designed for
patients who suffer from orthopedic or central nervous system
disorders. These patients have difficulties when undergoing gait
Manuscript received March 05, 2008; revised July 02, 2008, October 02, training on parallel bars; thus, this design allows them to walk
2008; accepted October 28, 2008. First published February 18, 2009; current
version published April 08, 2009.
on a circular path. It has also been used to examine the gait per-
The authors are with the Department of Electrical Engineering and Computer formance of healthy subjects with up to 60% BWS [4]. Norman
Science, Korea Advanced Institute of Science and Technology, Daejon 305-701, et al. [6] constructed a treadmill-based apparatus in which both
Korea (e-mail: neoworld@kaist.ac.kr).
Color versions of one or more of the figures in this paper are available online
the treadmill and the BWS device were hydraulically driven.
at http://ieeexplore.ieee.org. Most researchers have pursued their designs assuming a
Digital Object Identifier 10.1109/TNSRE.2009.2015179 constant unloading force despite the vertical movement of the
1534-4320/$25.00 © 2009 IEEE
SEO AND LEE: THE DEVELOPMENT OF TWO MOBILE GAIT REHABILITATION SYSTEMS 157

human body. This effort was motivated by the needs of ortho- TABLE I
pedic patients for whom the loading on the injured skeletal SYSTEM CONFIGURATION OF TWO SYSTEMS
structure must necessarily be under control. Cavagna et al.
[23] demonstrated that the main mechanism to contain energy
consumption during locomotion is the exchange between the
kinetic and gravitational potential energy within each stride.
This mechanism would be wholly accepted if the applied
unloading force was constant over the stride period, as if the
patient walked in a reduced gravity field. However, in reality,
the unloading force is not constant; thus, there is a nonzero
amount of external work that can affect relearning of the correct
motor pattern. In most of the BWS systems available, a constant
unloading force is not always guaranteed and the adjustment of
the BWS device is purely mechanical [7]–[16].
In the present research, two types of gait rehabilitation robot
systems are designed, a mobile manipulator type and a mobile
vehicle type. Additionally, an evaluation of their performance is
conducted. The paper is arranged in the following order. Sec-
tion II describes the system architecture and the difference be-
tween the two systems. Section III describes the controller de-
sign. Section IV presents the experiments and results of the two
systems. Concluding remarks and further developmental issues
are discussed in Section V.

II. GAIT REHABILITATION ROBOT SYSTEMS


Two types of gait rehabilitation robot systems were devel-
oped. They have been termed “WHERE-I” and “WHERE-II.”
WHERE-I is a mobile manipulator type, whereas WHERE-II is
a mobile vehicle type. Each system consists of four main sub-
systems: a mobile base subsystem, a BWS subsystem, an in-
tention analysis subsystem, and a safety subsystem. The main
difference between the two systems is that WHERE-I is based
on a mobile manipulator and an electrical BWS mechanism,
whereas WHERE-II is based on a mobile vehicle with a one-link
manipulator and a pneumatic BWS mechanism. There are six
axes of motion at the pelvis that are involved in walking. As
shown in Table I, three axes are inhibited. Therefore, WHERE-I
and WHERE-II can move forward/backward and can rotate in
place. The up-and-down motion generates the force, which is
unloaded/controlled by BWS system.

A. WHERE-I
The WHERE-I gait rehabilitation system is shown in Fig. 1. It
has a mobile manipulator with a one-link robot arm, an electrical
BWS mechanism, an intention analysis system using a linear
potentiometer, and a safety system with an ultrasonic sensor.
1) Mobile Manipulator: The mobile platform has two steer- Fig. 1. WHERE-I.
able driving wheels and two castors at the rear. The driving
wheels are connected to two 165 W ac motors that are differ-
entially driven by an onboard CPU. The steering mechanism is supervision of a medical doctor. The user has to walk the mo-
actuated by a 150 W dc motor. It is possible to adjust the height tion of the system. In the working mode, the mobile base fol-
of the system to that of the user. The mobile base is meant to lows the motion of the user. The one-link rotary arm mounted
track the motion of the user; thus, there are two linear poten- on the mobile platform supports the weight of the user while
tiometers to sense the motion of the user and feed it back as stabilizing the user during walking. The joint is actuated by a
the desired motion for the differential drive of the mobile base. dc motor through a torque-enhancing 150:1 gear. The control
There are two operational modes: the training mode and the scheme for mobile manipulator is introduced in Section III-A.
working mode. In the training mode, the main computer gen- 2) Electrical BWS Mechanism: The one-link robotic arm
erates the reference trajectory (position and speed) under the supports the body weight under the shoulder. In Fig. 1, two
158 IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 17, NO. 2, APRIL 2009

Fig. 2. WHERE-I: Sensing mechanism of force on the system using load cell,
and user’s motion intention using potentiometers.

fingers of a belt hold the trunk. As each finger can mechani-


cally move in a side, the distance between two fingers can be
adjusted to a different trunk size. The force on the robot arm
is measured as shown in Fig. 2. As the load cell feels the force
through a lever, it is possible to determine the following linear
relationship:

(1)

Fig. 3. Operation examples of WHERE-I’s intention analysis. (a) v = 0, ! =


where is the applied force on the load cell, is the force 0. (b) v > 0, ! = 0. (c) v < 0, ! = 0. (d) v = 0, ! > 0. (E) v = 0, ! < 0.
on the sliding bar, is the distance from the static axis to the
point where the force is applied, and is the distance between
the static axis and the position of the load cell. A load cell trans-
B. WHERE-II
ducer with a maximum capacity of 500 Kg feeds back the user’s
force on the manipulator to the force control unit. The WHERE-II system is shown in Fig. 4. It consists of a mo-
3) Intention Analysis System: Two linear potentiometers bile vehicle with a one-link robot arm, a pneumatic BWS mech-
with a 500 mm stroke length are attached to the top plate of the anism, a user intention analysis system using an accelerometer
manipulator, as shown in Fig. 1. Each sensor measures the body and an electronic compass, and a safety system with an ultra-
movement of either side. The signals are used to determine sonic sensor.
walking speed and the goal direction. After lowpass filtering 1) Mobile Vehicle: The mobile platform has two driving
the raw signals of the linear potentiometers, the deviation from wheels and two castors at the front corners. Two 250 W dc
neutral position is calculated to obtain the velocity. The differ- motors were used to drive the two wheels, which are located at
ence between them is used as the desired direction of motion. the middle of the mobile base, differentially. A single onboard
In Figs. 2 and 3, the linear velocity and the angular velocity computer was used to drive the two wheels differentially. The
were directly calculated from potentiometers mobile base is meant to follow the user’s motion, which is
sensed by two accelerometers and the electronic compass. The
sensor feedback is used as the desired motion of the mobile
base. WHERE-II also has the aforementioned two operational
(2) modes. Two one-link rotary arms are mounted on either side
of the mobile platform and are synchronously controlled to
where and are constants for the conversion from dis- adjust the height of the system for various users. A 130 W dc
tance to velocity, and are the displacements of the po- servomotor actuates the rotational joint through a torque-en-
tentiometer from the center point. hancing 100:1 gear. It also supports the body weight of the
4) Safety System: Four ultrasonic range sensors were used user at the initial system setup. After the setup of the system
to detect objects around the system. As the users are either el- for gait training, pneumatic cylinders are used for the active
derly or disabled people, they are unable to avoid collision with BWS. The control scheme for the mobile vehicle is introduced
nearby objects; thus, a range sensor feedback was used to steer in Section III-B.
the system away from nearby objects. In this application, ultra- 2) Pneumatic BWS Mechanism: During human walking, the
sonic sensors are used for “stop-and-go” operations. The sim- alternate up and down motions on either side are not identical. In
plest collision avoidance operation is to stop when confronted particular, for a disabled person whose gait style is not natural,
with an object. When the object is cleared, the robot continues the differences in the two sides can be very different. Thus, it
to move along its predetermined path (in the training mode) or is necessary to have independent control of each pneumatic ac-
according to the user’s motion (in the working mode). tuator. Therefore, two pneumatic actuators on the left and right
SEO AND LEE: THE DEVELOPMENT OF TWO MOBILE GAIT REHABILITATION SYSTEMS 159

and are not constants; instead, they vary as a function of ,


usually following a linearly increasing law. In the position con-
trol of the pneumatic actuators, the control performance tends
to be unsatisfactory and unpredictable due to the change of pa-
rameters and the variable friction force. In this study, a radial
basis function (RBF) neural network was designed in order to
compensate for the errors, .
3) Intention Analysis System: WHERE-I can determine mo-
tion intention of a user directly through its linear potentiometers.
In contrast, WHERE-II uses two accelerometers and an elec-
tronic compass to measure linear and angular velocities. How-
ever, these sensors have intrinsic measurement errors. These er-
rors are compensated using an ultrasonic sensor mounted at the
front of the system.
4) Safety System: Similar to WHERE-I, WHERE-II is
equipped with four ultrasonic sensors to detect nearby objects.
The simplest operation, “stop-and-go,” is used for collision
avoidance.

III. SYSTEM CONTROL


The developed systems can be described as a mobile manip-
ulator (WHERE-I) and mobile vehicle (WHERE-II). The ob-
jective of controlling this system is two-fold: 1) controlling the
mobile base to follow the trajectories generated by the reha-
bilitation program, or the path the user intends to walk, and
2) controlling the robot arm to adjust the height of the system to
support part of the body weight specified by a medical doctor.
Since there may be dynamic interactions between the two con-
trols, and disturbances from the user or rough terrain, the control
Fig. 4. WHERE-II. problem of this system is challenging.

A. WHERE-I: Mobile Manipulator


side were used for BWS mechanism. Pneumatic actuators are Since it is very difficult to model dynamics of this system due
driven by the pressure difference. It can be described as a system to its complex structure, and also due to disturbances or varying
with differential pressure as input and position as output. As far loads on it when the user exerts his ambulation on it we incorpo-
as vertical forces are concerned, a walking person may be mod- rated neural network based control methodology. The dynamics
eled as a simple body of mass , subjected to the action of ac- of a mobile manipulator subject to kinematic constraints can be
celeration, gravity and a vertical supporting force , expressed obtained using Lagrangian approach as follows [24]:
as the fraction of the body weight. Thus, the dynamics of mo-
tion of the actuator can be expressed as (6)

where kinematic constraints are described by


(3)
(7)
where is the resulting ground reaction and is the vertical
coordinate oriented upwards. In practice, it is necessary to add and denotes the generalized coordinates,
the error term in the second term of (3) to account for the is the symmetric and positive definite in-
effects of the mass, friction and strain specific to the technical ertia matrix, represents the centripetal and
solution adopted to realize the supporting device. Equation (5) Coriolis matrix, represents the friction and
is a general form of , which accounts for the effect of pure gravitational vector, is the constraint matrix,
friction , stiffness , viscous drag , and moving masses is the Lagrange multiplier which denotes the vector of
constraint force, denotes bounded unknown distur-
(4) bances including unstructured dynamics, is
the input transformation matrix, and is the torque
(5) input vector.
The generalized coordinates may be separated into two sets
where is the cross-section area of the cylinder and is the , where describes the generalized co-
differential pressure in the cylinder. In (5), the coefficients ordinates that appear in the constraint (7), and are the
160 IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 17, NO. 2, APRIL 2009

free generalized coordinates; . Therefore, (7) can be Assuming that there exist a Lyapunov function
simplified as , a positive continuous function, , a reference
smooth feedback velocity , such that

(8)
(15)
with . Assume that the robot is fully actuated
then (6) can be rewritten as
Taking the time derivative of (14), we obtain

(16)
(9)
Performing detailed mathematical derivation with (15) yields

Dynamics of mobile base can be expressed as


(17)

(10) Once simplified

Let denote a full rank formed by


columns that span the null space of defined in (8). Multi-
plying both sides of (9) by , applying kinematic constraints,
, and substituting steering system,
yields (18)
(11) with the unknown nonlinear terms which should be estimated
through the learning algorithm with neural networks
where , ,
, , and .
Dynamics of robot manipulator can also be expressed as (19)

(12)
(20)
In this stage, to design a controller, we consider a Lyapunov
function candidate for mobile manipulator as
Neural networks such as an RBF network has been shown
to have a universal approximation ability to approximate any
(13) smooth function on a compact set. Two RBF networks with a
sufficiently many number of hidden units such that
where is not a position error in the world coordinate space but
a position error in the robot coordinate space, velocity tracking (21)
error, where is a reference velocity , (22)
is a filtered tracking error in manipulator joints. The first term
in (13) was considered to maintain the user’s position within where is the input pattern to the neural networks defined as
the clearance area in robot. The second term for tracking error, .
and the third term for joint position and velocity error enable The ideal and unknown weights and
us to control robot manipulator which should adjust the prede- in that respect are assumed to be constant and bounded
termined height and support the specified portion of the user’s by
weight.
Now, take a Lyapunov function candidate of the following (23)
form [18], [19], [25]:
where and are known positive constants. The num-
bers of hidden-layer neurons of the two RBF neural networks
(14) are and . The approximation errors ,
are bounded by and , with and
In our system, the dynamic coupling term introduced in [25] is being two positive constants; The properly chosen RBFs
not required, because linkage between the mobile base and robot for the hidden-layer neurons are
arm is fixed. and .
SEO AND LEE: THE DEVELOPMENT OF TWO MOBILE GAIT REHABILITATION SYSTEMS 161

The basis functions of neural networks were defined to be where , , ,


Gaussian , , , and
. Applying weight update law,

(24) (34)

From the matrix theory, the following property holds [24]:


The centers and widths are all chosen a priori and kept
fixed for simplicity. Therefore, only the weights and are (35)
adjustable during the learning process.
The estimates of and are given by From (34) and (35),
(25)
(26)

Theorem 1: By choosing the control law as


(36)
(27)
(28) which is guaranteed to be negative provided

where is the tracking error, and is filtered tracking error for


the manipulator, and the weight update laws for the two neural (37)
networks are
or
(29)
(30) (38)

where is control gains, is pos- By Lemma 4.1.1 in [24], for all in the compact set
itive constants representing the learning rates of the two neural , since is asymptotically stable, there exist
networks, and is small positive design parameter. By computable positive constants , and , such that
choosing them properly, the tracking errors of error dynamics
and the neural network estimation weights are all guaranteed to (39)
be uniformly ultimately bounded.
Proof: Substitute the control law (27), (28) into the deriva- where , and , are constants rep-
tive of Lyapunov function (18), so that resenting the boundness, .
To ensure that the approximation property of the two neural
network on-line estimators holds throughout, the tracking error-
vector should be always kept in the compact set
. This may be achieved by selecting
the minimum control gain to satisfy

(31)
(40)
where .
Let a new Lyapunov function candidate with neural network
Thus, if the parameters are appropriately chosen, the uniform
learning dynamics be defined as
ultimate boundedness of the tracking errors and , and the
neural network weights can be guaranteed.
(32)
B. WHERE-II: Mobile Vehicle
Differentiating (32) yields
Where-II system is a nonholonomic mobile vehicle system.
Dynamics of nonholonomic mobile robot system is similar to
that of a mobile manipulator.
In the case of mobile vehicle, since the parameters for robot
manipulator are zero, the dynamics of mobile vehicle is easily
modified from (9) as follows:
(33) (41)
162 IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 17, NO. 2, APRIL 2009

Multiplying both sides of (41) by , applying kinematic con- In the same manner with Theorem 1, by choosing the control
straints, , and substituting , law as
we obtain
(51)

(42) where is vehicle tracking error, the weight update law for the
neural network can be described by
where , , ,
, and . (52)
Similar to a mobile manipulator, the stability of the tracking
system for mobile robot can be proven by choosing the fol- where , , and . By choosing them properly,
lowing Lyapunov function candidate: the tracking errors of error dynamics and the neural network
estimation weights are all guaranteed to be uniformly ultimately
(43) bounded.
Since the proof is similar to that of mobile manipulator, it is
where is a position error in the vehicle coordinate space, ve- omitted.
locity tracking error, where is a reference velocity.
The first term in (43) was considered to maintain the user’s po-
sition within the clearance area in the vehicle. The second term IV. EXPERIMENT AND DISCUSSION
in (43) considers tracking error. The human gait cycle consists of two phases, the stance phase
Let us take a Lyapunov function candidate in the following when the foot is in contact with the ground, and the swing phase
form: when the foot is in the air [26]. Within these phases, there are
four events: a foot strike (or heel strike), an opposite toe off,
(44) an opposite foot strike, and a toe off. During the stance phase,
which makes up about 62% of the gait cycle, single limb stance
Assuming that there exist a Lyapunov function occurs for approximately 35%–40% of the total gait cycle, with
, a positive continuous function, , a reference double support stance for the balance of the cycle. It is presumed
smooth feedback velocity , such that that the gait pattern of the subject is similar to that of a healthy
person.
(45)
A. Results
Taking the time derivative of (44), we obtain Ten male (height: , weight: ) and five
female (height: , weight: ) patients took part
(46) in the experiment. They were in the stage of gait rehabilitation
after suffering minor leg injuries.
Performing detailed mathematical derivation with (45) yields Before applying the RBF controller to the system, the effec-
tiveness of this controller was compared to a well-tuned PID
controller. The mobile BWS system is nonlinear, implying that
(47) there are unknown dynamics and disturbances. A RBF neural-
network-based method was developed for the velocity control
with the unknown nonlinear terms which should be estimated with unknown dynamics and disturbances. Neural networks are
through learning algorithm with RBF neural networks. used for the online estimations of unknown nonlinear dynamics.
In order to control the system with a controller such as a PID,
(48) the system parameters should be known precisely. As shown in
Fig. 5, the system parameters are disturbed when a user inter-
A RBF network with a sufficient high number of hidden units acts with the system. This causes a performance degradation of
such that the PID controller.
1) WHERE-I: An experiment was conducted to find the force
(49) on the load cell according to the displacement of the sliding bar.
As the load cell takes forces through lever action, this relation-
may be used to approximate the unknown nonlinear terms. The ship should be identified before forces can be directly measured
inputs for the neural network of the mobile vehicle are desired with the load cell. Fig. 6 shows the experimental result of this
position, position errors, reference velocity and acceleration, relationship using an articulated steel of 50 Kg. Since there are
and velocity tracking error. some deviations from the ideal linear relationship, the load cell
The estimate of is given by reading was compensated in the control experiment using this
data. This process is also applied to the coefficient tuning in the
(50) load cell of WHERE-II.
SEO AND LEE: THE DEVELOPMENT OF TWO MOBILE GAIT REHABILITATION SYSTEMS 163

Fig. 7. Control of manipulator for weight unloading.

Fig. 5. Control performance when system parameters are changed. In inter-


acting with patient, system dynamics are varying over time.

Fig. 8. Control performance of weight unloading while walking.

Fig. 6. Experimental result of the load cell.

Fig. 7 shows the experimental result for weight unloading due


to the position control of the manipulator. It was found that the
quantity of the unloaded weight can be controlled by adjusting
height of the manipulator.
Fig. 8 shows the control performance of weight unloading
while walking. When a human walks, the center of gravity
moves up and down, resulting a change in the forces on the
manipulator. If this fact is not considered, the force on the ma-
nipulator changes as the subject walks. The proposed control
algorithm, however, is capable of supporting variable weights.
Fig. 9 shows the sensor feedback of the user’s trajectory
through the use of the two linear potentiometers. The first Fig. 9. Motion detection while walking.
figure shows the displacement of the sensors to the plate, and
the phase difference indicates the direction of motion. Each
sensor measures the body movement of each side. According goal direction of the user. The second figure shows the control
to (2), their signals are used to find the walking speed and the amount of the supported weights while walking.
164 IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 17, NO. 2, APRIL 2009

Fig. 10. Acceleration on knees while walking. Fig. 12. Velocity tracking performance of mobile vehicle.

Fig. 11. Control performance of weight unloading while walking. Fig. 13. Tracking errors of mobile vehicle.

2) WHERE-II: Fig. 10 shows the sensing of user’s motion such that the ground reaction forces exceed those required for
using the two accelerometers, together with an electronic com- static equilibrium. If the control action is not applied to the pneu-
pass and an ultrasonic sensor. Accelerometers cannot measure matic cylinder, the signal of the load cell fluctuates along with
the absolute displacement between user and the robot system the motion of human body. This means that constant BWS is not
due to intrinsic characteristics and measurement error. These realized. However, with the proposed controller, the unloading
errors are amassed during system operation and can eventually weight could be kept at a constant value.
result in a faulty operation. To compensate for these errors, the Fig. 12 shows the control performance of velocity tracking
ultrasonic sensor measures the absolute distance between the while walking. From the accelerometer measurements, the de-
robot system and the subject. The first figure in Fig. 10 shows the sired reference velocity and the trajectory for the mobile vehicle
horizontal signal of the two accelerometers. The second figure can be calculated. The first figure in Fig. 12 shows the tracking
shows the vertical signal. Fig. 10 indicates that the acceleration result for the linear velocity, and the second figure shows that of
of the human body on the right and left side of the system are angular velocity. The proposed neural network controller learns
actually different due to the leg injuries of the subject. This indi- the system dynamics online. The proposed controller is capable
cates that the BWS mechanism should be performed separately of compensating any sudden change of robot dynamics due to
on each side. Fig. 11 shows the control performance of weight its learning ability. Fig. 13 shows the velocity tracking error of
unloading while walking. During the gait cycle, the mass of the Fig. 12.
entire body moves up and down in a sinusoidal pathway of low 3) Rehabilitation Performance Comparison:
amplitude. Tables II and III show the temporal parameters of the
The sinusoidal nature of the motion (with its changing direc- affected and unaffected limbs according to body weight
tions) results in accelerations of the center of mass of the body, unloading. A 5-min walking trial (Ten male and five female
SEO AND LEE: THE DEVELOPMENT OF TWO MOBILE GAIT REHABILITATION SYSTEMS 165

TABLE II
TEMPORAL PARAMETERS OF AFFECTED LIMB ACCORDING TO BODY WEIGHT UNLOADING

TABLE III
TEMPORAL PARAMETERS OF UNAFFECTED LIMB ACCORDING TO BODY WEIGHT UNLOADING

subjects) was performed by each patient. The mean and • The BWS mechanism is based on a pneumatic actuator
deviation of the results are given in the Tables II and III. that has proven to be effective at maintaining a con-
WHERE-II can support the distributed control scheme on stant force unloading. This force should be held constant
each pneumatic actuator; hence, the more effective rehabilita- under various BWS conditions, and it appears to be im-
tion performance is expected compared to what would be ob- portant for a correct motor pattern.
tained by a centralized control scheme. These tables indicate • The system was designed such a way that the rotational
that WHERE-I has a difference between the stances of affected axis of the human body is located on the line between
and unaffected limb of nearly 5% when 40% of the body weight the driving wheels. This enables the system to realize
is unloaded. On the other hand, WHERE-II has nearly a 2% dif- three-degree-of-freedom motion conveniently.
ference and its relearning gait pattern is more similar to a normal • If at times the subject loses his balance during training,
gait pattern compared to that of WHERE-II. Detailed discussion the body support mechanism with the harness can still
on this is provided in the next section. support him autonomously. In contrast, in WHERE-I,
in emergency situations, the user could impose unin-
B. Discussion tentional force to the system that can cause undesirable
commands. On the other hand, the WHERE-II system in
The following discussion focuses on the design and control emergency situations does not take any forward move-
aspect of the proposed robot system. The two proposed robot ment commands; thus, safety is assured.
systems WHERE-I and WHERE-II have several advantages as • The patient feels the contact force from the robot when
follows. he moves up and down (when the BWS device uploads
1) WHERE-I the patient’s weight). The proposed systems can be con-
• The height of this system is too low, signifying that the sidered as a spring, mass, and damper system, as shown
entire system stands below the subject’s shoulder, which in (6) and (41). The BWS device was designed to hold a
makes the user feel comfortable. It’s a known fact that constant portion of the patient’s weight upwards. There-
when the height of the system is higher than the height fore, the interaction between the patient and the robot
of the eyes, the user commonly feels uneasy. Therefore, system can be explained in the terms of the “stiffness.”
it is a design guideline that the system height be lowered. Stiffness is important factor in the detection the contact
• The user’s intention can be directly and easily calculated force. If the force from objects is same, an object with
using two potentiometers. This leads to a simple struc- low stiffness has a softer feeling compared to an object
ture and convenient operation. that is stiffer. Usually, the stiffness of a pneumatic actu-
• An electrical BWS mechanism was adopted for this ator is lower than that of an electrical motor. Therefore,
system. A stable control algorithm was proposed for the the patient in WHERE-II feels a low contact force.
mobile manipulator using the RBF neural network. In terms of system control, a RBF neural controller that
2) WHERE-II tracks the patient and supports the body weight constantly
• The most important feature in this system is the separate is proposed. For the control of the robot system, two neural
BWS mechanism on the left and right sides of the human network controllers are proposed to control the vehicle and
body. The gait patterns of the left and right sides are manipulator independently. A neural-network-based method-
somewhat different and are determined by the nature of ology was developed for the velocity control with unknown
the disability. Thus, a distributed design and control for dynamics and disturbances. Neural networks are used for online
the BWS is more effective. estimations of unknown nonlinear dynamics.
166 IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 17, NO. 2, APRIL 2009

V. CONCLUSION [17] C. Lee and J. Lee, “Walking-support robot system for walking reha-
bilitation: Design and control,” J. Artif. Life Robot., vol. 4, no. 4, pp.
Here we presented two mobile gait rehabilitation systems that 206–211, 2000.
enable overground gait training. Experimental results suggest [18] C. Lee, K. Seo, C. Oh, and J. Lee, “A system for gait rehabilitation
that different levels of assistance are needed to compensate for system with body weight support: Mobile manipulator approach,” J.
HWRS-ERC, vol. 2, no. 3, pp. 16–21, 2001.
asymmetrical gait and impairment. This implies that the BWS [19] C. Lee, K. Seo, C. Kim, S. Oh, and J. Lee, “A system for gait reha-
mechanism should be performed independently at each side. bilitation: Mobile manipulator approach,” in Proc. IEEE Conf. Robot.
Our systems were developed to provide constant BWS. The pro- Automat., 2002, vol. 3, pp. 3254–3259.
[20] C. Lee, K. Seo, C. Oh, and J. Lee, “Newly designed rehabilitation
posed RBF controller was designed to provide such a constant robot system for walking-aid with peumatic actuator,” Int. J. Human-
unloading capability. This enables the patient to retrain at a con- Friendly Welfare Robot. Syst., vol. 4, pp. 42–46, 2003.
stant unloading weight. Clinical and experimental trials were [21] K. Chun, K. Cho, and B. Kim, “Influence of body weight unloading on
hemiplegic gait,” (in Korean) J. Korean Acad. Rehabil. Med., vol. 23,
carried out and the effectiveness of the proposed systems in gait no. 2, pp. 371–376, 1999.
rehabilitation demonstrated. Future research includes the com- [22] L. Finch and H. Barbeau, “Influences of partial weight bearing on
bination of the mobile BWS system with a telemetric system normal human gait: The development of a gait retraining strategy,”
Can. J. Neurolog. Sci., vol. 12, p. 183, 1985.
and improvements to afford outdoor operation. [23] G. A. Cavagna, N. C. Heuglund, and C. R. Taylor, “Mechanical work in
terrestrial locomotion: Two basic mechanisms for minimizing energy
REFERENCES expenditure,” Amer. J. Physiol., vol. 233, no. 5, pp. R243–61, 1977.
[1] The 2007 Census of South Korea, , 2007 [Online]. Available: http:// [24] F. L. Lewis, S. Jagannathan, and A. Yesildirek, Neural Network Control
of Robot Manipulators and Nonlinear Systems. London, U.K.: Taylor
www.stat.go.kr
[2] S. A. Bloomfield, “Changes in musculoskeletal structure and function Francis, 1999.
[25] S. Lin and A. A. Goldenberg, “Neural-network control of mobile
with prolonged bed rest,” Med. Sci. Sports Exer., vol. 29, no. 2, pp.
197–206, 1997. manipulators,” IEEE Trans. Neural Networks, vol. 12, no. 5, pp.
1121–1133, Sep. 2001.
[3] L. Walfson, R. Whipple, P. Amerman, and A. Kleinberg, “Stressing the
posture response – A quantitative method for testing balance,” J. Amer. [26] Human Walking, J. Rose and J. G. Gamble, Eds., 2nd ed. Baltimore,
MD: Williams Wilkins, 1994.
Geriat. Soc., vol. 34, pp. 845–850, 1986.
[4] N. Siddiqi, F. Gazzani, J. D. Jardins, and E. Y. S. Chao, “The use of a
robotic device for gait training and rehabilitation,” in Medicine Meets Kap-Ho Seo (M’09) was born in Pusan, Korea,
Virtual Reality, K. S. Morgan, Ed. et al. Amsterdam, The Nether- in 1974. He received the B.S. degree from Korea
lands: IOS Press, 1997. University, Seoul, Korea, in 1999, and the M.S.
[5] T. Ide, N. Siddiqi, and N. Akamatsu, “Expectation for medical and degree in electrical engineering and the Ph.D. degree
healthcare robotics,” Adv. Robot., vol. 7, no. 2, pp. 189–200, 1993. from Korea Advanced Institute of Science and
[6] K. E. Norman, A. Pepin, M. Ladouceur, and H. Barbeau, “A treadmill Technology (KAIST), Daejeon, Korea, and 2001
apparatus and harness support for evaluation and rehabilitation of gait,” and 2009, respectively.
Arch. Phys. Med. Rehab., vol. 76, pp. 772–778, 1995. He has worked in developing robotic system for the
[7] H. Barbeau, M. Wainberg, and L. Finch, “Description of a system for elderly and disabled. He and his fellows developed
locomotor rehabilitation,” Med. Bio. Eng. Comp., vol. 25, pp. 341–344, three gait training systems and intelligent bed robotic
1987. system. His current research interests include neural
[8] F. Gazzani, A. Fadda, M. Torre, and V. Macellari, “WARD: A pneu- network, vision-based mobile robot, and rehabilitation robot system.
matic system for body weight relief in gait rehabilitation,” IEEE Trans.
Rehabil. Eng., vol. 8, no. 4, pp. 506–513, Dec. 2000.
[9] R. B. Knight, J. He, M. R. Carhart, and J. Koeneman, “Design and
development of a simple, low cost gait training assistive device,” in Ju-Jang Lee (M’86–SM’99–F’08) was born in
Proc. 25th Annu. Int. Conf. IEEE EMBS, 2003, pp. 1724–1727. Seoul, Korea, in 1948. He received the B.S. and
[10] J. Lee, P. E. Allaire, A. C. Wolovick, W. Jiang, P. O. Riley, and D. M.S. degrees in electrical engineering from Seoul
C. Kerrigan, “Control system for partial body weight support device National University, Seoul, Korea, in 1973 and 1977,
for human gait training,” in Proc. 2004 IEEE Int. Conf. Control Appl., respectively, and the Ph.D. degree in electrical engi-
2004, pp. 1723–1728. neering from the University of Wisconsin, Madison,
[11] L. Lünenberger, G. Colombo, R. Riener, and V. Dietz, “Clinical as- in 1984.
sessments performed during robotic rehabilitation by the gait training From 1977 to 1978, he was a Research Engineer
robot lokomat,” in Proc. 2005 IEEE Int. Conf. Rehabil. Robot., 2005, at the Korean Electric Research and Testing Institute,
pp. 345–348. Korea. From 1978 to 1979, he was a Design and Pro-
[12] M. Frey, G. Colombo, M. Vaglio, R. Bucher, M. Jorg, and R. Riener, “A cessing Engineer at GTE Automatic Electric Com-
novel mechatronic body weight support system,” IEEE Trans. Neural pany, North Lake, IL. In 1983, he joined briefly the project engineer in the
Syst. Rehabil. Eng., vol. 14, no. 3, pp. 311–321, Sep. 2006. Research and Development Department of the Wisconsin Electric Power Com-
[13] D. Aoyagi, W. E. Ichinose, S. J. Harkema, D. J. Reinkensmeyer, and pany, Milwaukee, WI. Since 1984, he has been with the Department of Elec-
J. E. Bobrow, “A robot and control algorithm that can synchronously trical Engineering and Computer Science, KAIST, Daejeon, Korea, where he is
assist in naturalistic motion during body weight supported gait training currently a Professor. In 1987, he was a Visiting Professor at the Robotics Labo-
following neurologic injury,” IEEE Trans. Neural Syst. Rehabil. Eng., ratory of Imperial College Science and Technology, London, U.K. From 1991 to
vol. 15, no. 3, pp. 387–400, Sep. 2007. 1992, he was a visiting scientist at the Robotics of Carnegie Mellon University,
[14] J. Zitzewitz, M. Bernhardt, and R. Riener, “A novel method for auto- Pittsburgh, PA. His current research interests include intelligent and variable
matic treadmill speed adaptation,” IEEE Trans. Neural Syst. Rehabil. structure control of mobile robot, evolutionary emotional robot, service robot
Eng., vol. 15, no. 3, pp. 401–409, Sep. 2007. for the disabled, space flexible manipulators, intelligent transportation system,
[15] Y. Nemoto, S. Egawa, A. Koseki, S. Hattori, T. Ishii, and M. Fujie, and power system stabilizer. He developed several mobile robots for walking
“Power-assisted walking support system for elderly,” in Proc. 20th training and intelligent bed robotic system for the elderly and disabled. He is
Annu. Int. Conf. IEEE Eng. Med. Biol. Soc., 1998, vol. 20, pp. an Associate Editor of Artificial Life and Robotics, and Machine Intelligence
2693–2695. and Robotics Control. He is an Editor of the International Journal of Assistive
[16] M. Peshkin, D. A. Brown, J. J. Santos-Munné, A. Makhlin, E. Lewis, Robotics and Mechatronics.
J. E. Colgate, J. Patton, and D. Schwandt, “KineAssist: A robotic over- He is a Fellow of SICE and ICROS He is an Associate Editor of IEEE
ground gait and balance training device,” in Proc. 2005 IEEE 9th Int. TRANSACTIONS ON INDUSTRIAL ELECTRONICS and IEEE TRANSACTIONS ON
Conf. Rehabil. Robot., 2005, pp. 241–246. INDUSTRIAL INFORMATICS.

You might also like