Multi- Criteria Optimal Design of Cable based Ankle

Rehabilitation Robot
1. Introduction
An ankle rehabilitation robot has been conceptualized and designed to realize the range
of motion, muscle strengthening and proprioception training exercises. The robotic device has
been designed to help the patient and the therapist in their cooperative efforts for the treatment of
impaired ankle joint which may be a result of some injury or stroke. After analyzing the ankle joint
anatomy and its motions, a parallel mechanism was selected for the robot. To mimic the human
ankle joint and its muscles actuation, the proposed robot uses air muscles configured in a similar
fashion to the actual muscle arrangement. The performance of a parallel robot greatly depends
on its dimensions and configuration of actuators. To explore the advantages of these robots it is
essential to obtain a set of kinematic parameters leading to optimal robot performance. The
design of robot has been evaluated on the basis of three main performance indices such as,
workspace, condition number and 2-Euclidean norm of actuator forces, under some constraints.
The performance criteria and the constraints have been discussed in detail to justify their
inclusion. The two Multi-objective Optimization Approaches (MOA), a weighted formula approach
and a pareto optimal approach have been used and the results are compared and discussed.
2. Rehabilitation and robotics
Rehabilitation in a broader sense means a practice by which any form and grade of human
physical disorder can be reinstated. The disorder could be the result of an injury or a stroke.
Conventionally, rigorous and repetitive exercises are performed under the supervision of a
therapist to restore range of motions and strength of limbs. These exercises improves motor
functions by enhancing neuro-plasticity and neuro-recovery at the affected limbs. During a
rehabilitation treatment, cooperative efforts of therapist and the patient are required over
prolonged sessions of treatment in a clinic. Moreover the patient is required to continue the
prescribed exercises at home for a speedy recovery. It has been documented that using
conventional line of treatment the recuperation is slow and sometimes continues for more than a
year. The major drawbacks of such treatments is that the patient has to travel to attend the
clinical sessions in his disability state and spend on his time and money. The workout at home is
monotonous and lacks motivation resulting in inadequate improvement. Similarly the therapist
has to perform strenuous and repetitive efforts with the patients and thus he can only attend a
limited number of patients in a day. Rehabilitation treatment, is also not well structured and
largely depends on intuitive judgment of therapist on the progress of the patient.
Robotics can play an important role in the process of rehabilitation by assisting the therapist
and the patient in reducing their efforts. Rehabilitation process can also be improved by acquiring
progressive data of patients improvement, which in turn can help the therapist to make
systematic decisions on choice of further exercises. Certain measure may be taken to make the
workouts more interesting and motivating as advised in []. There are certain other robotic aids,
which are already in use such as MIT-MANUS and LOKOMAT etc.
Rehabilitation robots are different from the industrial robots in application and use and hence
special care must be taken in their design. The industrial robots are more structured and meant
for a specific goal with minimum human interference. Moreover the operators for industrial robots
are also the trained technicians. On the other hand rehabilitation robots are generally multi-tasked
and hence should be intelligent machines. Human augmented robots should be safe to use and
must be user friendly in operation since the user could be a layperson. Thus the design and
control of these robots is a challenging task requiring multi-discipline skills and in-depth
knowledge of human joint anatomy and movements.

3. Human ankle, its problems and physiotherapy

3.1 Ankle Complex

The ankle joint [1] is a combination of two joints (Fig.1) out of which the first joint is made
of three bones: the lower end of the tibia (shinbone), the fibula (the small bone of the lower leg)
and the talus (the bone that fits into the socket formed by the tibia and the fibula). The talus sits
on top of the calcaneus (the heel bone). The talus moves mainly in one direction. It works like a
hinge to allow foot to move up (dorsiflexion) and down (plantar flexion). The second joint is
subtalar joint, in human anatomy it is also known as the talocalcaneal joint and this is a joint of
the foot. It occurs at the meeting point of the talus and the calcaneus. This joint is responsible for
the inversion and eversion of the foot, but plays no role in dorsiflexion or plantarflexion of the foot.
However it is very much a part of the ankle joint and thus can not be ignored.
There is one more joint called MTP (metatarsophalangeal) joint connecting fore and rear
with calcaneus, cuboid and Navicular bones as shown in Fig.1. The relative rotational motion
around this MTP joint realizes more natural exercises for toe raising and heel raising etc.

Fibula
Fibula

Calcaneus

Tibia
Talus
Navicular
Metatarsal
Phalange

Cuboid

Figure 1: Bones of the Ankle and sub talar joint

There are ligaments on both sides of the ankle joint that hold the bones together.
There are many tendons that cross the ankle to move the ankle and move the toes.
Ligaments connect bones to bones while tendons connect muscles to bones this has been
shown in Fig. 2. The ankle joint is capable of rotations in all three planes (sagittal, frontal and
transverse planes) as shown in Fig. 1, where sagittal plane is defined by x and z and movements
in this plane occur around the y axis; transverse plane is defined by x and y and movements in
this plane occur around the z axis; and frontal plane is defined by y and z and movements in this
plane occur around the x axis. Various ankle movements [3]can be summarized as follows (Table
1). The sub talar joint has a more constrained movement pattern owing to three articular facets
connecting the talus and calcaneus. Sub talar motion is described as pronation (lifting foot
upward) and supination (lifting foot downward). Thus the foot is the most complex bone structures
in the body since the functions of the ankle and sub talar joints are also interdependent with those
of the talonavicular (MTP) and calcaneo-cuboid joints, as these joints share bones.

Table 1. Ankle joint motions

Axes
X
Y
Z

Name of the motion

Inversion
Eversion
Dorsiflexion
Plantarflexion
Adduction
Abduction

Range of Motion

Torque Requirements (Nm)

350
250
200
400
250
200

48
34
40.7- 97.6
20.3- 36.3
30
30

3.2 Ankle injuries and physiotherapy

Ankle injuries [1] are one of the most common injuries in sports and daily active life and
on an average 600 patients per annum are being treated in New Zealand only [2]. Youngsters are
subjected to ankle injuries from sports and whilst carrying excessive load whereas children and
the elderly gets it from walking on uneven surfaces and bone weakness. Non functionality of
ankle joint is also quite common in stroke surviving patient.
Primary treatment for ankle injuries or non functionality includes rest, ice, compression
and elevation (RICE) of the affected foot along with stretching and exercise therapy, and partial
weight bearing with aids to maintain mobility in the ankle. During ankle rehabilitation the entire
foot is dorsiflexed, plantarflexed, inverted and everted considering the combined ankle and sub
talar joints.
Generally a complete rehabilitation program [4] including exercises for range of motions
(ROM), muscle strengthening, and proprioceptive training which play an important role in
speeding up recovery for normal functioning of ankle. Achilles tendon should be put in stretching
exercise as soon as possible (within 48 to 72 hrs) after the injury and the ROM must be
recovered. Once the ROM is achieved, strengthening of weakened muscles is essential for rapid
recovery and is a preventive measure against further reinjury. Manual therapy, such as toe
raises, heel walks, and toe walks, toe curls, and marble pick-ups as well as resistance exercises
to ankle joints are attempted to regain strength and coordination of the muscles. As the patient
achieves full weight bearing capability without pain, proprioceptive exercise is initiated for the
recovery of balance and postural control using wobble boards . Finally, advanced exercises using
uneven surface wobble board should be performed to regain functions specific to normal
activities.
3.3 Ankle rehabilitation robotics

Construction of a suitable rehabilitation mechanism requires detailed study of human joint

movements and their control from the physiotherapy perspective. The direction, speed and range
of- joint motion are altered by injury to muscles and ligaments. There is enormous potential to
integrate medical knowledge, human movement studies and mechanism synthesis in the domain
of ankle rehabilitation devices.
Rehabilitation treatment specific to ankle joint is quite exigent for the reason that the ankle
joint is the most complex part of the human skeleton [11]. Continuing physiotherapy is required to
help recover ankle joint from the injuries or to enhance neuroplasticity for non functionality due to
stroke. Physiotherapy requires rigorous repetitive movements of limbs about their respective
joints and robotic machines are useful in such applications once appropriately programmed. For
ankle rehabilitation, typical passive devices such as elastic bands and the active units for strength
and proprioception exercises such as wobbles board and foam rollers are in use, but they allow
only simple and functional rehabilitation exercises. Robotic devices on the other hand are
advantageous owing to the fact that they are multifunctional to allow almost all types of exercises
on one machine and provides accurate motions and forces. These devices can be programmed
by a physiotherapist and exercise modes could be selected depending on the type of injury and

patients improvement. Thus, task oriented training by supervised robotic systems is helpful to
provide the patient with more diverse and useful exercises. It is realized that the parallel robots
are good choice to perform three rotational degrees of ankle movements, because they have the
benefits of high dynamic stiffness and required manipulability. So far only rigid link parallel
mechanisms [6] have been conceptualized.
However all these devices have certain common limitations and drawbacks and to overcome
them this chapter proposes an air muscle actuated soft parallel mechanism as an improvement
due to its inherent advantages over RPM. Parallel mechanisms can be defined as Soft Parallel
Mechanisms (SPM) where cables or tendons are used as links and Rigid Parallel Mechanism
(RPM) if all the links are rigid. SPM does not have any spherical joints and as a matter of fact it
has higher movability. Soft parallel devices are very light in weight and has higher payload to
weight ratio. These could be designed to be wearable and portable. SPM has simpler dynamic
model than their rigid-link counterparts because the inertias of their links (i.e., cables) can be
ignored. Soft parallel robots are being used recently in a variety of applications [1] ranging from
very sophisticated medical and manufacturing applications to less precise construction and
shipment activities. Wire flexibility of SPM poses some constraints on the workspace and an extra
variable called tensionability is required to be considered during its kinematic design. This
makes the design process more cumbersome and choice of a perfect design becomes very
difficult. It is therefore required that the design of the proposed device be optimized by carefully
considering all the constraints.

4. Cable based parallel robot for ankle rehabilitation

4.1 Existing devices based on parallel mechanisms

In order to implement ankle joint rehabilitation treatment, several devices have been
developed. Representatives of some commercial developments are Biodexs multi joint system3
and balance system, which can facilitate ROM/strengthening, and balance exercises,
respectively.
A 3-RSS/S parallel mechanism is proposed by Gengqian et al [4] and the design of
prototype has been validated using simulations. Girone et al. [5] suggested the Rutgers Ankle
that can provide active six degrees of freedom (DoF) motions at an ankle with a Stewart platform
structure to cover various exercise modes. Using this, the patients foot is fixed firmly to the
platform of the Rutgers Ankle and is synchronized to virtual reality scenarios. Recently, the
Rutgers Ankle is extended to the Dual Stewart platform with 12 (DOF) motions for gait
rehabilitation. Even though presently most of the exercise modes are possible by the use of two
Stewart platforms, balance/proprioception exercises can also be achieved with a single device.
Since ankle movement in most exercises needs less than four (DOF) motions, Dai and Zhao [6]
proposed a sprained ankle device using a three or four (DOF) parallel mechanism with a central
strut, and analyzed the orientation and stiffness of the mechanism with considerations of the
central strut. Yoon et al. [7] have designed a single platform-based reconfigurable robot
mechanism with satisfactory MTP joint motions and less than six (DOF) motions which allows
more natural foot and ankle motions to cover various exercise modes.
So far only platform type of robotic devices have been designed where foot is required to
be placed on top of a platform which is actuated from the bottom. There is a possibility of
undesirable shift in the ankle joint in all these designs. However in the proposed design care has
been taken to place actuators in a similar fashion as the actual muscle arrangement is, so that
the ankle joint stays stationary.
4.2 Cable based parallel manipulators

A cable based robot utilizes the basic idea of the Stewart platform parallel link mechanism. The
inimitable feature of this approach is to use cables as the parallel links and to use winches as the
actuators. A cable actuated robot can precisely contrive large loads with flexibility and has high
strength to weight ratio. These properties make a cable actuated robot applicable to a large
variety of applications ranging from building construction, cargo handling or inspection of an

airplane (where the workspace of the robot must be very large) to applications as small as
endoscopy, colonoscopy and surgery. Cable based Parallel robots (CBPR) are becoming a
promising alternative to the rigid-link mechanisms in many applications, such as load lifting and
positioning, coordinate measurement, aircraft testing, haptic devices [8], and robot rehabilitation
[9] etc. The reasons are quite obvious as,
1) Using wires it is possible to place the moving platform on floor because the bottom of the
platform does not have any mechanism.
2) Wire robot does not use any spherical joints and as a matter of fact it has higher movability and
lower inertia in comparison with rigid link robots. On the other hand, vibration is sometimes a
serious problem in cable drive systems. Coordinate control of multiple cables to increase the
internal force improves its mechanical rigidity and the vibration characteristics.
3) Longer Motion stroke and greater range of motion (ROM) is possible since they have larger
workspaces because their spools can reel out a large amount of cables.
4) Wire actuated parallel devices are very light in weight and are more convenient compared to
their counterpart.
CBPR have several other advantages [10] over rigid-link manipulators. First, all of their actuators
are stationary, the transmission systems can be mounted on the fixed base and thus, they have a
higher payload-to-weight ratio, which makes them attractive for high-load and high speed
applications. Second, their special designs make them potentially inexpensive, modular, and very
easy to reconfigure. Finally, they have much simpler dynamics model than their rigid-link
counterparts because the inertias of their links (i.e., cables) can be ignored.
However, the draw back is the inherent addition of wire flexibility and vibrations. Also, it should be
noted that unlike rigid links, cables are characterized by the unilateral property (can pull but
cannot push moving platforms), and therefore the formulations and results obtained for the
kinematics analysis, workspace analysis, trajectory planning, etc. of the rigid-link mechanisms
cannot be directly applied.
5. Proposed Wearable Ankle Rehabilitation Device
A parallel robot for ankle joint rehabilitation treatments has been proposed in the present chapter.
The robot is designed to provide three rotational degrees of freedom to the ankle joint for the
necessary ROM exercises. The device uses two parallel platforms, a fixed platform (FP) built-in
with a leg support structure and a moving platform (MP). Air muscles are used as actuators and
are mounted on leg support with their actuating end connected to MP through cables. These
cables pass through sleeves provided in the FP. The leg support structure is fixed using some
straps which move over the knee and fixed at thigh. Lower end of the leg support remains
approximately 100 mm above the ankle joint. MP remains below the foot and has a heel locator
and straps to fix foot. The moving platform of the robot is shaped to form a shoe of varying size
and shape. This shoe has been attached to the leg support using a special mechanism which
provides three rotational degrees of freedom to the shoe. Actuation of the moving platform or the
shoe is realized using the air muscles and the cables. Apparently the air muscles can only pull
and can not push hence to maintain the tension in all the cables it is desired to have redundant
actuation. As a matter of fact all the cable based parallel robots have redundant actuation and it
has been mentioned in the literature that to achieve n-dof motion from the manipulator the robot
should have n+1 actuators. Hence to obtain 3-dof from the robot we have used a set of four air
muscles and four cables. Coordinated and antagonistic actuations of air muscles will ensure
desired changes in the wire lengths and pose of the moving platform subsequently for a range of
ankle exercises. Coordinated and antagonistic actuations of air muscles will ensure desired
changes in the wire lengths and pose of the moving platform subsequently for a range of ankle
exercises.
6. Robot Kinematic Modeling
Robotic manipulators are required to perform specified motions in 3-D space. To
accomplish this, the end effector is moved in the workspace along a predefined trajectory. The
position and the orientation of the robot end effector at a specified workspace location can be

obtained by controlling the displacements of its links. This calls for a mathematical model of the
robot which can define the relationship between end effector motions and link displacements in
the domains of time and space. This mathematical model is called kinematic model and the
derivatives of this model describes mechanics of the motion without taking forces into account.
When the forces and/or torques are considered with the mechanics of motion the analysis is
termed as dynamics. The kinematic and the dynamic study are essential tools for the design of
robotic manipulators. Kinematic model helps us to completely describe the position and
orientation of the end effector. The kinematic parameters of joints are of two types, fixed
parameters and variables parameters and the kinematic model can be defined by providing the
information about both types of kinematic parameters of the links.
Thus the kinematic model in general, describes the spatial motion of the end effector in
time domain, about a fixed reference frame (or fixed platform). The kinematic study includes two
types of models namely Forward Kinematic (FK) model and Inverse Kinematic (IK) model. The
forward kinematic model provides position and orientation of the end effector as a function of
variable and constant kinematic parameters of all the links. Similarly the inverse kinematic model
is constructed to find the set of joint parameters that would bring the end effector in a desired
location in the workspace. The desired task of the end effector is specified in terms of its position
and orientation in the workspace. The joint variables to accomplish this task are found using IK
analysis. Joint variables, in turn are used to find the instantaneous coordinates of the end
effector using FK analysis.
It has been well established [] that for serial robots the inverse kinematic analysis is
difficult and provides more than one feasible solutions. However the forward kinematic analysis is
simple and close form solution is possible. Conversely for parallel robots the IK solution exists in
close form but the FK solution is not possible. While doing a FK analysis for parallel robots one
ends up with a set of highly coupled nonlinear equations for which unique solution is not possible.
Both types of kinematic models are normally required to study manipulator differential
motion, its statics and to implement a desired control scheme for the end effector. Kinematic
analysis is also essential to estimate the feasible workspace of the robot and to perform a
singularity analysis. A brief discussion on the kinematic modelling is presented in the following
section.
6.1 Inverse Kinematic Analysis

As with other parallel mechanisms, the inverse kinematics of a cable-driven system is

relatively simple and provides a unique solution of cable lengths for given end effector pose. The
analytic wire lengths have been determined in terms of the pose of the moving platform.
YO
O

bi

bi

FP

ZO
XO
O

PE

YE

O E,

ai

ai

MC
XE

ZE

Figure 2. A sketch of cable and position vectors

of connection points on FP and MP

The diagram presented in Figure 2 shows the motion of the proposed SPM. The connection
points on MP and FP are denoted by ais and bis respectively. The connection points on fixed
platform are in the same plane (ZO = 0) and are placed at a radial distance rb from the coordinate
system which is located at O. The position vectors of point bis on the FP are defined by
O

rb cos i
bi = rb sin i
0

(1)

where, rb is the radial distance from the base coordinate frame O. The variable i denotes the
angular position of point bi on FP with respect to axis X0. The moving platform similarly has a set
of connection points located at the corners of a rectangle and the coordinate frame attached to
OE is about 60mm above the center of mass (MC) of the MP (with reference to the position of
ankle joint which is approximately 60mm above moving plate level). The position vectors of the
four connection points on the moving platform ca be given as follows:

Xi
ai = Yi
60

(2)

Variable X i and Yi are the coordinates of each connection point as shown in Figure 2. The
position vectors of the link lengths in terms of end poses can be expressed as a system of four
equations described below:
O

Li = O PE + O RE E ai O bi

i = 1,...4.

(3)
O

where PE represents the position vector of point OE with respect to O. RE is the rotational
transformation matrix of MP with respect to FP using a fixed axis rotation sequence of , and
about X0, Y0, and Z0 axes, respectively, and can be written as below.

cos cos
RE = sin cos
sin

sin cos + cos sin sin

cos cos + sin sin sin
cos sin

sin sin + cos sin cos

cos sin + sin sin cos (4)

cos cos

6.2 Forward Kinematics

Forward kinematic (FK) mapping for parallel manipulators is difficult compared to serial
manipulators as it involves highly coupled nonlinear equations and their closed-form solution
derivation is a matter of interest to the researchers [55]. It is important to solve forward kinematics
since it is a key element in closed loop position and force control of parallel manipulators. It is an
essential block in the trajectory control of a manipulator and hence accurate prediction of end
pose from the available link lengths is essential. In the development of the proposed robot a
Sugeno based fuzzy logic model has been constructed to implement FK and it has been found
that this model is significantly8 accurate and is also time efficient.
6.3 Geometric Modeling

The kinematic models establish the correlation between the joint displacements and the
position and orientation of end effector of a robot. This correlation can only be used for the static
control of manipulator in the workspace. For a robotic manipulator the final desired position is
important and at the same time the velocity by which it has traversed to reach to the final location
is also equally important. Thus it is essential to obtain a mapping between joint velocities and end
effector velocity. This mapping can be defined by a matrix, which is called the robot Jacobian

matrix. The Jacobian matrix depends on robot configuration and linearly maps the cartesian
velocity in to joint velocities. It is interesting to note that the Jacobian matrix defined for the
parallel robots corresponds to the inverse Jacobian of the serial robots. To determine the
Jacobian matrix of parallel robots two approaches, namely geometric approach and analytical
approach can be used. In the present chapter we have used a geometric approach as discussed
below:
To determine geometric Jacobian matrix using robot geometry, initially a relation between
link lengths and end effector pose is formulated. For the subject robot this relation is given by
equation (3). Later on this relation is derived as described below to finally provide the Jacobian
matrix.

q& = J (q )t

(5)

It is evident from the equation (3) that link lengths can be easily calculated for a given set
of end effector orientation. The magnitude of each
O

L2i = OLTi

Li vector is
O

(6)

Li

Using the time derivative of the kinematic constraint equations, the relation between the cable
velocity vector q and the twist of MP i.e. t can be related by the Jacobian matrix as follows:

2 Li L& i = ( P& + O R& E E a i ) T ( Pi + O R E E a i o bi ) + ( Pi + O R E E a i O bi ) T ( P& + O R& E E a i )

(7)

Using linear algebra identity, if c and d are column vectors

c T d + d T c = 2c T d = 2cd T

(8)

which implies following

2 Li L& i = 2( O P + O R E E a i O bi ) T ( O P& + O R& E E a i )

Since the end effector is constrained to only rotational motion the time derivative of

(9)
O

P shall

become zero. Setting P& = 0 and further simplifying we get,

O

Li L& i =( O Li ) T ( O R& E E a i )

(10)

R& E = O E O RE

(11)

Li L& i =( O Li ) T ( O E O R E E a i )

(12)

ai = O R E E ai

(13)

Li L& i =( O Li ) T ( O E O a i )
L L& =( O L ) T ( O O a )

(14)

Li L& i =( a i Li ) ( E )

(16)

but since
O

also
O

(15)

if t is the twist vector it is defined as

0
t = O
E

(17)

equation 12 can be arranged and compared with

Xq& = Yt

(18)

q& = Jt

(19)

J = X 1 Y

(20)

where the inverse Jacobian can be defined as

or
Here, ( E ) is the angular velocity vector of the MP with respect to frame O. The inverse
Jacobian matrix of the cable robot can be written such that its ith row is
O

L
J i = O ai i
Li

where i = 1,2,....,4 .

(21)

6.4 Workspace Analysis

The workspace analysis of parallel robots is different from that of serial robots.
Translational and orientation workspace of serial robots can be assigned to different sets of joints
and can be studied separately. However for the parallel robots both kinds of workspaces are
achieved through coupled motion of its links or joints and thus the total workspace can not be
decoupled. For parallel mechanisms with rigid links, workspace is the space where the inverse or
forward kinematic solutions exist. However for a cable based robot, its workspace is the space
where sets of positive cable tensions exist because they are needed to constrain the moving
platform all the time regardless of any external wrench. In other words, if there is at least one set
of positive cable tensions at a specific pose forming a force closure, then this pose belongs to the
force-closure workspace. Generally, force-closure workspace is a set of poses that the forceclosure condition is satisfied.
Thus in mechanisms, it is not only necessary to solve the closure equations but it is also
essential to verify that equilibrium can be achieved with non-negative actuator (cable) forces. The
design and workspace of a 66 cable-suspended parallel robot has been discussed in [70-75],
and workspace volume is characterized as the set of points where the centroid of the moving
platform can reach with tensions in all suspension cables at a constant orientation. The main
contribution of [71] is in establishing that for any geometry of platforms the largest workspace
volume occurs when the moving platform (MP) is the same size as the base platform (BP).
Workspace evaluation is generally performed [73] in the context of rigidity, tensionability
and force/torque capacity, these terms are explained below:
6.4.1 Rigidity
A is rigid at a given pose under applied external forces and spine force only if all the
cables are in tension. Since the rigidity of a depends on the external load and hence for rigidity
evaluation, dynamic forces should be also considered. Thus, rigidity analysis requires the motion,
inertia and all other externally applied forces apart from kinematic constraints to be considered.
For simplicity, another property called tensionability can be considered which only depends on
the kinematics and at the same time fulfills the criteria of rigidity of the manipulator.
6.4.2 Tensionability
Tensionability is defined as an essential property of and a manipulator is said to be
tensionable provided that there exists a finite spine force when the external loads are applied. On

a separate note tensionability and sufficient spine force together fulfills the criteria for rigidity.
However the reverse is not true i.e. if a manipulator is rigid but not tensionable and is in
equilibrium then this equilibrium may be an unstable one, meaning that a small change in the
pose of end effector may result in instability due to loss of tensionability. The stability of can also
be explained in terms of stiffness of the device which will be discussed in later sections.
6.4.3 Pre-tension
A positive spine force exerts a finite tension in the cables which is called Pre-tension. This helps
in keeping all the cables tensionable in the workspace. However a smaller pre-tension may not
yield tensionability and at the same time, excessive pre-tension may result in instability of the
manipulator. There exists an optimum value for this pre-tension in the cables for which the
manipulator shall be rigid and tensionable. The proposed device basically has four actuated links
and accommodates patients ankle joint (which acts as a central strut) in the parallel device. The
air muscles which are the actuating links are all in their half contracted positions initially to
facilitate the antagonistic actuation of moving platform. The cables connecting both platforms are
given some Pre-tension, by adjusting the cable lengths. Force experienced by the central strut
due to this pre-tension in the cables is called spine force and is denoted by FS here. This helps in
keeping all the cables tensionable in the workspace. The static force and moment balance on the
MP are:
4

FS + U i = 0

(22)

i =1

( m
o

i =1

Ui ) + M ext = 0

(23)

Though there are no external moments applied to the MP, due to ankle stiffness, finite
moments are required to move ankle joint passively. To realize the ROM exercises air
muscles works antagonistically and applies the required moments at the end effector. To
find corresponding tensions in individual cable the dual relationship between kinematics
and statics can be used as follows:
(24)
M ext = J T U

M ext = M x0

M y0

M z0

(25)

Where U is the vector of cable forces present in each cable, J is the Geometric
Jacobian matrix of the robot and M ext is a 3 dimensional vector containing the required moments
given by (5). The rows in Eq. (5) represent the moments required to orient ankle joint about the
x0 , y0 & z0 directions, respectively. Now, to obtain the equations for the force in each of the
cable, Eq. (4) is rearranged in the following manner.

U = J M ext

(26)

J = ( J T ) 1

(27)

where

Next, at each point within the possible workspace, the equation describing the force in
each cable is used to see if tension is obtainable. A Matlab program is written to search the entire
workspace and check for the condition of tensionability of cables and link length constraint.

A novel approach to check for the condition of tensionability has been used. The
proposed robot is redundantly actuated, i.e. to achieve three degrees of freedom four actuators
are used. Initially, using equation 7 the force vector for all the cable to achieve flexion trajectory is
computed. This force vector has two pull and two push forces, since cables can not push these
compressive forces are superimposed to the two pull forces. Similar treatment is given to other
trajectories such as inversion-eversion and adduction-abduction. Now using force superposition
principle, force vector can be obtained for a trajectory which has all three simultaneous rotations.
The pretension in the cables takes care that all the cables remain under tension at all the times.
Sets of end effector orientation in the workspace
600
400

T1

Forces in cables 1,2,3 and 4 (N)

200
0
0

50

100

150

200

250

300

350

400

450

500

800
600

T2

400
200
0

50

100

150

200

250

300

350

400

450

500

600
400

T3

200
0
0

50

100

150

200

250

300

350

400

450

500

1000

T4

500
0
0

50

100

150

200

250

300

350

400

450

500

Figure 3. Static cable forces at different end effector orientations

7. Design criterions of Cable based Ankle Rehabilitation Device (CARD)

In the light of unidirectional nature of the cable forces, the design of cable-based
manipulator is more complicated than the rigid-link devices. There are certain criterias specific to
the cable based parallel robots which require more attention. These design criterion are
discussed below:
1. Maximum workspace criteria
2. Near unity condition number criteria
3. Singular value based criteria
4. Minimum force norm based criteria
5. Other criterias
An explanation of these measures is important to state their importance.
7.1 Workspace criterion

Workspace is vital parameter in the domains of kinematic analysis and should be given
proper attention. The feasible workspace volume depends on the geometrical configuration of the
robot such as the size of the platforms and placement of connection points on them apart from
other constraints discussed in the previous section. Thus by changing the geometrical
parameters it is possible to change the feasible workspace. It is desired that the workspace of a
robot, under given constraints should be as large as possible for greater maneuverability.
Further, unlike serial robots, workspace of parallel robots is unevenly shaped due to their
complex kinematics. This further contributes in lowering the size of the feasible workspace. Apart
from the size, the quality of the workspace is also important and it is desired that the workspace

should be free from singularities. In the present study feasible workspace has been represented
by an index as given below.

I=

f
T

where
and

T = ( max min )( max min )( max min )

f = 1 2

1 = workspace points satisfying tensionability criteria

2 = workspace points which are reachable with the

(28)

restricted stroke length of the

actuator.
In the proposed work robot design has been optimized to achieve maximum permissible
workspace volume.

7.2 Condition Number criterion

As discussed in the previous sections, the Jacobian matrix J of a robot maps joint rates
to the cartesian velocities of the manipulator. The condition number of this matrix is a measure of
its sensitivity to changes in the kinematic variables of the robot. A robot design with near unity
condition number is desirable since it minimizes the error in the end effector wrench due to input
errors in joint torques. The condition number can also be used to evaluate the workspace
singularities. This number also reveals as how far a robot is from its present configuration to the
nearest singular configuration. Stiffness of the end effector due to joint stiffnesses can also be
obtained using condition number. Thus it is evident from the above discussion that the condition
number is a vital design parameter and the robot configuration should be optimally designed to
produce a minimum condition number close to unity. The condition number k is defined as the
ratio of the largest singular value l to the smallest singular value s of the matrix J for a
fixed orientation of the manipulator . The singular value can further be defined as the square root
of the eigenvalues of JJ T and J T J . The range of condition number is described as below.

1 k <

(29)

When the condition number approaches unity, the matrix J is said to be well conditioned or far
from singularities. On the contrary, if the condition number is higher, the matrix is said to be ill
conditioned. An ill conditioned Jacobian matrix will further magnify the kinematic or dynamic error
present in the robot motion. Sometimes to avoid an infinite right hand side bound, an inverted
form of the condition number referred to as the conditioning index ( C.I .. )can also be used.
1
(30)
C.I . =
k
Where 0 C.I . 1
The condition number is obtained at each individual point in the feasible workspace region, for a
fixed orientation of the end effector. Though it is a useful information but to analyze the behavior
of the condition number over the entire workspace volume a Global Conditioning Number (GCN)
is generally used. In the present work robot performance has been defined by Global Condition
Number (GCN) given as,

GCN =

(k )dw

dw

(31)

where k is the condition number at a given orientation and W is the feasible workspace. Since
it is difficult to calculate the exact solution to the integrals mentioned above, GCN is discretely
defined and expressed as
n

GCN =

(k )
i =1

(32)

where W is the total number of n discrete feasible points constituting the workspace and the
numerator is the sum of condition numbers obtained at different points in the feasible workspace
volume grid. The GCN is bounded by the range

1 GCN

(33)

Here, when the GCN is a large number the entire workspace tends to be ill conditioned and when
the GCN is near one the entire workspace is said to be well conditioned. GCN further depends on
the robot configuration which is defined by arrangements of connection points at both the
platforms and the link lengths. Hence there exists an optimum robot configuration for a good GCN
and performance thereof,. To ensure that all the points in the workspace provides a condition
number within certain range, the maximum value of the condition number for a particular robot
design over the entire workspace can be obtained and minimized. Once the maximum GCN is
minimized, it can be ensured that
1. The final GCN represents the average behavior of condition number over the feasible
workspace.
2. The condition number all over the feasible workspace is always less than the minimum GCN
value.
7.3 Singular value based criterion

Singular values are important measure of kinematic behavior of the robot and provides
an assessment on its controllability. The manipulator loses or gains extra degrees of freedom
when it enters in to a singular configuration. A kinematic singularity occurs when the determinant
of the Jacobian matrix becomes zero in the workspace.

det( J ) = 0

(34)

Referring to equation (14), it is apparent that when J is singular and its null space is non-zero,
there will be certain non-zero cartesian vectors (t) resulting in zero joint vectors ( q& ). This further
means that, despite the joints being locked, the end effector can still have some infinitesimal
motion in a particular direction and gains one or more degrees of freedom. Hence it is essential to
optimize the design parameters of the robot, in order to minimize singular points in the
workspace. The condition number provides a fraction of maximum and minimum singular values
and not the magnitude of these values in the workspace. Thus a criteria based on minimization of
maximum singular values shall yield a workspace free from singularities. However in the present
optimization problem when GCN is minimized (which is the ratio of maximum singular value to the
minimum singular value), the maximum singular value also get minimized.

7.4 Minimum actuator force based criteria

Since we intend to design a wearable robot for ankle joint rehabilitation treatment, it is
desired to keep the length of the robotic device similar to the length of the patients leg. The
length of the robot is governed by the length of its actuators hence the actuators should be kept
as small as possible. Further the size of these actuators depends on the cable forces calculated
in section 6.4. Longer air muscles are required for higher forces in the individual links. The
lengths of the actuators can be minimized by lowering the actuator force requirements. Once
again it is apparent that the forces in the actuators are the function of robots geometrical
parameters. By selecting connection points farther from the axis of rotation, the forces can be
greatly reduces. To minimize the actuators force vector it is convenient to summarize the set
values of force vector in a single number. Vector norms are generally used to represent the
vector in a single value. Three types of vector norms are generally used such as, 1-norm, 2-norm
or -norm. 2-norm or euclidean length is preferred over the other two norms because it is more
sensitive towards changes in larger force components. 1-norm is equally sensitive to all the force
components and -norm is only sensitive to the changes in the largest force component. In the
present study, the higher forces in the actuators are required to be minimized to prevent the
cables from breaking. Moreover, the higher actuator forces may also cause undesired cable
elongation affecting the positional accuracy adversely. 2-Norm or the Euclidean distance of the
actuator forces can be given as below:

= 0 U

(35)

7.5 Miscellaneous criterion

There are certain other design criteria which are specific to the application of the robot.
Such criteria are size of the manipulator and the fixed platform, range of motions of the robot,
material selection based on strength criteria, etc.
Despite their many advantages over serial robots, the parallel manipulators have some
innate problems owing to their close chain mechanism. A parallel robot basically has several
serial robots connected in parallel and its feasible workspace is the intersection of the individual
workspace contribution of its joints. Hence the feasible workspace of the parallel robots is small in
size compared to serial robots. Moreover due to the complex kinematics of parallel robots, the
shape of their workspace is quite irregular. Apart from the size and the shape, the workspace is
also affected by the kinematic singularities. The serial robots have only one kind of singularity
whereas the parallel robots due to their complex mechanism have three kinds of singularities
such as inverse kinematic singularity, direct kinematic singularity and the combined singularities.
The design of cable based parallel robots is again more challenging since they have an
added constraint on the workspace in the form of tensionability. The feasible workspace of these
robots is the euclidean space where the robot manipulator can reach with positive tension in all its
cables. Thus apart from the singularities the condition of tensionability is also required to be
checked to evaluate a feasible workspace for the manipulator.
The proposed robot is to be used for ankle joint rehabilitation treatments and the subjects
are supposed to fix their foot and the ankle into the moving platform. It is evident that the device
has to be robust to take care of the physical differences in the shape and size of different patients
foot and ankle. Differences in the foot of different subjects also amounts to the small variations in
the robot kinematic and dynamic parameters. Furthermore, when in actual use, the effective
location of the ankle joint in the robotic device also changes to some extent. This is due to the
fact that the ankle joint is made up of two joints (as explained in section 3.1) and different ankle
motions are realized by one of these joints.

The kinematic and geometric parameter variations in the joint space results in the
corresponding cartesian parameters variations. Owing to this fact, the Jacobian matrix of the
robot, which relates the joint and cartesian rates is required to be well conditioned. It is apparent
that if the Jacobian matrix is not well conditioned then the small changes in the joint variables will
result in very large changes in the cartesian variables. Which will further lead to difficulty in
manipulator control and large trajectory following errors. Nevertheless, by choosing optimal
geometric parameters of the robot, the condition number of the Jacobian matrix can be improved
and the design can be made robust to parametric variations.
Another problem with cable based parallel robots is that the cables can not be subjected
to very high forces. Higher forces in the cables may cause breakage or undesired elongation of
the cable which will adversely affect the positional accuracy. Moreover to achieve higher forces,
the required length of the air muscles also increases proportionally (as discussed in section 7.4)
which is not desirable. Optimizing the geometrical parameters of the robot, the force requirements
can be minimized.
It can be concluded from the above discussion that the design of the robot is required to
be optimized to maximize its workspace, and minimize the condition number and the actuator
force requirements. The dimensional analysis and optimization of the robot design may include
the key design parameters such as the shapes and sizes of the parallel platforms, locations of the
connection points of the cables on both these platforms and the lengths of cables. Such a design
optimization is essential to reduce the above mentioned shortcoming of the robot while
maintaining its inherent merits.
8. Design Optimization
Though the parallel robots have been extensively used recently in a wide range of
applications, their potential has not been completely exploited. The reasons are obvious and have
been discussed in the previous section. To address the issues of the workspace, singularity and
the robust design of parallel robots, normally trial and error approach is used. This approach is
based on rigorous experiments or simulation runs and intuitive judgments on the results
thereafter. The main drawback of this approach is that with an increase in number of tunable
parameters the required number of simulation runs increases exponentially. Moreover tuning of
all the design parameters simultaneously is difficult and time consuming. Some of the
researchers have tried to optimize one or more of the design objectives using some numerical
methods. Several performance indices such as manipulability, isotropy, dexterity index,
conditioning index, global conditioning index and global isotropy index have been defined by
different researchers (Khatami & Sassani, 2002) to compare the performances of competing
manipulator designs.
In a recent study (Hassan & Khajepour, 2008) the actuator forces in a cable based
parallel manipulator have been optimized using a minimum norm solution. Author have used
Dykstras projection method to optimally distribute the actuator forces among the cables and the
redundant limbs. Though the force distribution among links have successfully been optimized to
provide minimum norm solution, the geometrical parameters have not been tuned to minimize the
actuator forces. Genetic algorithms (GA) have been used by (Sergiu et al.,2006 ) to optimize the
workspace of a 2-DoF parallel robot using a mono-objective function. A novel kinematic design
method has been implemented by (Liu, 2005) and various performance indices such as global
conditioning workspace, global conditioning index and global stiffness index have been used to
obtain the design parameters. A multi-criteria optimization based on conventional weighted
average approach has been performed by (Lemay & Notash, 2003). The authors have proposed
a combination of GA and simulated annealing algorithms to optimize workspace, dexterity and
mass & size of the manipulator simultaneously. Workspace and stiffness of a modular parallel
robot have been studied and optimized by (Merlet, 2003). The author has proposed a branch and
prune type algorithm to improve the robot performance. A new performance index called space
utilization has been proposed by (Stock & Miller, 2003) to evaluate the optimal kinematic design
of a linear Delta robot. Exhaustive search minimization method has been used to optimize
mobility, workspace and manipulability. Global conditioning index has been optimized as a result
of altering the length of links by (Khatami & Sassani, 2002). The authors have used a nested

implementation of two GAs to obtain a mini-max genetic solution. GA have been used with
constraints defined as penalty functions by (Su et al., 2001) to minimize the minimum condition
numbers in the entire workspace. To validate and verify the algorithm, results obtained from GA
have further been compared with the Quasi-Newton method. Architectural optimization of a 3-dof
parallel robot has been performed by (Tsai & Joshi, 2000) to maximize the global conditioning
index.
8.1 Mathematical Formulation

The optimization problem is formulated as below.

(1) Minimize Global Condition Number (GCN)

f (q1 , q 2 ,....q8 ) = GCN (i ) + P(i, j )

GCN (i ) =

(k )dW

(36)
(37)

dW

Here, qs are geometrical variables, i is the number of robot designs and j W , k is the
condition number at a given orientation and W is the feasible workspace. Further P (i, j ) is
a penalty term defined as

( J (i, j ))

P(i, j ) = max

(
J
(
i
,
j
))
min
max
where

min

the smallest singular value and

robots jacobian matrix

(38)

max is defined as largest singular value of the

J.

(2) Maximize the workspace volume

f (i, , , ) = [( max min )( max min )( max min )]

(39)

where , , are roll, pitch and yaw angles for the moving platform
(3) Minimize Tension norm in the workspace

U = J M ext

(40)

J = ( J T ) 1

(41)

where

and

M ext = M x0

M y0

M z0

(42)

and U is the vector of cable forces present in each cable, J is the Geometric Jacobian
matrix of the robot and M ext is a 3 dimensional vector containing the required moments.

8.2 Kinematic Constraints

8.2.1 Workspace Constraints

0 q1 , q 4

, 0 q 2 , q3

and 0 q5 ,..., q8

(43)

2
1 r1 ,...., r8 2

(44)

l mmin l m l mmax

(45)

stroke length check for mth link

8.2.2 Link Interference Constraints

Link interference restriction between any two adjacent links

Dist ( Lm , Ln ) ( Rm + Rn )

(46)

Um 0

(47)

[U m ] max 1

(48)

8.2.3 Tensionability Constraints

Tensionability constraint

Checking maximum tension in a link

1 , 2 and 1

are constants

r1 ,...., r8 are the radial distances of the connection points on both platforms from their
respective origins.

Lm and Ln are adjacent links connecting two parallel platforms.

Rm and Rn are the radius of cables of adjacent links.
U m is the vector of cable tension in link
Here, qis are geometrical variables, i is the number of binary string in GA and j W . Both
moving and fixed platforms provide eight geometrical variables (qis) as shown in the Figure 2.
Sizes and separation of the two platforms have been kept constant thereby fixing r1, r2, and r3.
Though the device is left-right symmetric but we have evaluated all the kinematic parameters

independently to see whether the points are required to be placed symmetrically. It seems that
due to the specific trajectory, the manipulator has to follow, q5 and q8 are not symmetric.
This is a multivariable constrained optimization problem where location of points at platforms or
geometrical parameters are varied to obtain a desired GCN. To solve such problem
conventionally direct search or gradient decent methods are used but these techniques become
less efficient when the search space is large and is finely discretized. Further these methods
sometimes get trapped into a local minima and fail to provide a global optimum solution. It has
been observed that when the objective function does not change over certain points in
succession, these traditional methods become less effective. We propose to use a modified
Genetic Algorithms in the present optimization problem. GA works with population of points and
process them simultaneously and hence is more likely to give a global solution. As is evident from
Figure 4-6, that values of condition number are quite similar in the neighborhood of any given
point and the variation is very smooth. Thus discretizing the workspace does not lead to any
information loss and this fact further supports our choice of GA.
It has been observed [3] that GA is a global optimizer and converges to the optimum
solution very quickly but its local search capabilities are not equally good. Though it converges
very fast to the possible solution but then to fine tune the solution it takes time and sometimes
might not converge to the desired accuracy. To overcome this problem we propose to use one of
the gradient descent algorithms (for local optimization) in the selection operator of GA. In the
present work after a fixed accuracy is achieved from GA cycles, the selection process shifts to the
gradient based selection to fine tune results.
8.2. Multi-criterion Optimization Techniques
8.2.1 The Conventional Weighted-Formula Approach: Transforming a Multi-objective Problem into a
Single-objective Problem

In the literature, by far the most used approach to cope with a multi-objective problem
consists of transforming it into a single-objective problem. This is typically done by assigning a
numerical weight to each objective (evaluation criterion) and then combining the values of the
weighted criteria into a single value by either adding or multiplying all the weighted criteria. That
is, the quality Q of a given candidate model is typically given by one of the two kinds of formula:
Q = w 1 x c 1 + w 2 x c 2 + . . . + wn x c n
(49)
or
Q = c1 W1 x c2 W2 x . . . x cn Wn
(50)
where wi, i=1,n, denotes the weight assigned to criteria ci, and n is the number of
evaluation criteria. Let us mention some examples of this approach in the context of the two
multiobjective scenarios discussed in the Introduction. The first scenario involves rule induction
algorithms for classification, where it is common to evaluate the quality of a candidate rule by
measuring two or more criteria. An example is:
Q = w1xAccuracy + w2xcost
(51)
An instance of the general formula structure (1) combining cost and accuracy into a
single measure of rule quality. This formula and its variations are used in several rule induction
algorithms [Bruha & Tkadlec 2003], [Furnkranz & Flach 2003]. Another example is:
Q = costx w + accuracyx (1-w)

(52)

An instance of the general formula structure (2) used in [Kaufmann & Michalski 1999] to
produce one of the evaluation criteria used in a lexicographic approach for rule evaluation. The
second scenario involves attribute selection algorithms for classification, where it is also common
to evaluate the quality of a candidate attribute subset by measuring two or more criteria. An
example is:
Q = x Acc + x (1 (S + F)/2)
(53)
This formula was used by [Cherkauer & Shavlik 1996] to measure the quality of a
candidate attribute subset in an attribute selection method following the wrapper approach, where

Acc (accuracy) was measured by the validation-set accuracy of a decision tree built with the
selected attributes, S was the decision tree size and F was the number of attributes (features) in
the candidate attribute subset. Other examples of attribute selection methods following the
wrapper approach and combining two or more attribute subset quality criteria in a weighted
formula can be found in [Liu & Motoda 1998].
8.2.2 The Lexicographic Approach

The basic idea of this approach is to assign different priorities to different objectives, and
then focus on optimizing the objectives in their order of priority. Hence, when two or more
candidate models are compared with each other to choose the best one, the first thing to do is to
compare their performance measure for the highest-priority objective. If one candidate model is
significantly better than the other with respect to that objective, the former is chosen. Otherwise
the performance measure of the two candidate models is compared with respect to the second
objective. Again, if one candidate model is significantly better than the other with respect to that
objective, the former is chosen, otherwise the performance measure of the two candidate models
is compared with respect to the third criterion. The process is repeated until one finds a clear
winner or until one has used all the criteria. In the latter case, if there was no clear winner, one
can simply select the model optimizing the highest-priority objective. A well-known algorithm
using this approach is the AQ18 rule induction algorithm [Kaufmann & Michalski 1999] and its
[SIGKDD Explorations. Volume 6,Issue 2 - Page 78] variants. AQ18 uses a lexicographic
evaluation functional (LEF) defined by a sequence of pairs <(c1,t1), (c2,t2), , (cn,tn)>, where
each ci, i=1,,n, represents the value of a performance criterion for a given candidate model,
and each ti, i=1,,n, represents the tolerance associated with ci. This tolerance is specified by a
threshold indicating the maximum value that the performance criterion ci for a given candidate
model is allowed to deviate from the value of ci for the best current candidate model. More
precisely, let M1 and M2 be two candidate models being compared, and assume, without loss of
generality, that M1 has a better value of ci than M2. If the difference between M1s ci value and
M2s ci value is greater than ti, then M1 is immediately considered better than M2, without the
need to check lowerpriority objectives. Otherwise the difference between M1 and M2 is not
considered significant, and in order to select the best candidate model one has to use the
remaining lower-priority objectives. In [Kaufmann & Michalski 1999] this approach is used with a
LEF where a predictive accuracy-related measure (a combination of completeness and
consistency gain) is used as the highest-priority criterion (c1), and rule description simplicity as
the next criterion (c2); but of course other criteria could be used.
8.2.3 The Pareto Approach

The basic idea of the Pareto approach is that, instead of transforming a multi-objective
problem into a single-objective problem and then solving it by using a single-objective search
method, one should use a multi-objective algorithm to solve the original multi-objective problem.
Intuitively, this approach makes sense. One should adapt the algorithm to the problem being
solved, rather than the other way around. In any case, this intuition needs to be presented in
more formal terms, which is done in the following. Let us start with a definition of Pareto
dominance. A solution s1 is said to dominate (in the Pareto sense) a solution s2 if and only if s1 is
strictly better than s2 with respect to at least one of the criteria (objectives) being optimized and
s1 is not worse than s2 with respect to all the criteria being optimized. Mathematically, assuming
without loss of generality that all criteria ci, i=1,k, are to be maximized, a solution s1
dominates a solution s2
iff there exists ci such that s1(ci) > s2(ci) and for all ci, i=1,k, s1(ci) >= s2(ci),
where s1(ci) denotes the quality of solution s1 with respect to criteria ci and k is the
number of criteria being optimized. A solution si is said to be non-dominated if and only if there is
no solution sj that dominates si. Note that the Pareto approach never mixes different criteria into a
single formula all criteria are treated separately.

Simplicity

A
C

accuracy
Figure: Pareto dominance illustration
For an understanding of the concept of Pareto dominance, let us say that the two
objectives to be maximized are the accuracy and the simplicity of a model. In the figure, model A
is dominated by model B and by model D; model C is dominated by model D; and model E is
dominated by model F. Models B, D and F are non-dominated solutions. They form the so-called
Pareto front. Once Pareto dominance has been defined, the next step is to understand the crucial
differences between a multi-objective algorithm based on Pareto dominance and a singleobjective algorithm. There are two related crucial differences. First, there is a difference in the
kind of output expected from each of these two kinds of algorithm. A multi-objective algorithm
should return to the user a set of non-dominated solutions, rather than just a single solution as in
a single-objective algorithm. simplicity accuracy Second, and related to the first difference, the
search performed by a multi-objective algorithm should explore a considerably wider area of the
search space and keep track of all non-dominated solutions found so far, in order to find as many
solutions in the Pareto front as possible. Of course, this makes a Pareto-based multi-objective
optimization algorithm more complex than its single-objective counterpart. Some Pareto-based
multi-objective optimization data mining algorithms are discussed in [Kim et al. 2000],
[Bhattacharyya 2000], [Pappa et al. 2002].

9. Genetic Algorithm Methodology

The fitness function and parameter selection have been discussed here and for a detailed
reading on GA [7,10] is recommended.
9.1 Generation

Initial population of 100 binary coded strings of 96 bits was generated using Knuths random
number generator [7]. Eight variables q1,q8 as shown below, one for each connection point
location have been defined. The binary string of 96 bits has 12 bits assigned for each of the eight
variables thus the solution accuracy of the order of 10-05 can be achieved.

1101..1
1231010..0
123 0110..1
1231011..0
1231010..1
123 0101..0
1231010..1
1231001..0
123
q1

q2

q3

q4

q5

q6

q7

q8

numbers and value of GCN and penalty term for each string (each string corresponds to a robot
design) is evaluated from the feasible workspace to further find the fitness function.
9.2 Fitness Evaluation

In order to evaluate each string in the population, its fitness must be calculated. Generally GA is
used for maximization problems and hence for present case of minimization of GCN a special
fitness function (16) has been used. The binary strings are decoded to real numbers and value of
GCN and penalty term for each string (each string corresponds to a robot design) is evaluated
from the feasible workspace to further determine the value of the fitness function.
9.3 Reproduction

The population initially selected may not have all the good strings in terms of their fitness values.
Thus good strings have been analyzed and their multiple copies based on individual fitness have
been selected in the mating pool. The Roulette-Wheel Selection method [3] has been used for
reproduction. After an assumed accuracy is achieved and a near global optimal point has been
identified the selection process shifts to the gradient based selection and multiple copies of binary
strings in proportion to their gradient values are stored in the mating pool. Thus strings with
larger gradients have more copies in the mating pool. This enables the algorithm to fine tune the
global optimal point and perform local search more effectively.
9.4 Crossover and Mutation

A four point crossover with 0.95 probability has been used and mutation has been performed with
0.01 probability. Generally crossover probability is kept close to one so that all the parent strings
may get a chance to crossover. Mutation helps in finding a global optimum solution but to avoid a
random search its probability is kept low.

3
r2

q2 q3

3
2
r6

r3

r5

A0
r1
1

q1

q4

r4

q6

q7

r7
r8

q5

q8

Figure 2. Rectangular end platform and circular fixed

platform with their respective geometrical
parameters

Condition number

3
2.8
2.6
2.4
2.2
0

500

1000

1500

2000

2500

3000

Workspace points
Figure 3: Condition number behavior over the
workspace

10. Results and discussions

After four initial GA iterations, GCN value is converged to 3.385 which is a near
optimal value and later on using modified GA, it is further reduced to 2.5572 in 18 iterations.
Further reduction in GCN was not observed, possibly because five out of eight variables
approached their limiting values.

Global Condition Number

2nd Norm of Cable Forces

8000

6000

4000

2000

0
0

10

15

20

25

30

Figure 3. Static cable forces at different end effector orientations

Cable Force Norm

8000

6000

4000

2000

0
1
0.9

Fe
as
ibl
e

0.8
0.7

Wo
rks
pa
ce

10

15

20

25

er
Numb
dition
n
o
C
l
globa

Figure 4. Static cable forces at different end effector orientations

The geometrical parameters obtained were [36, 36, 4, 6, 75, 74, 75, 52.4] and as can be
seen that except for q3, q4 and q8 all other variables are in their extremities. The feasible
workspace was once again checked randomly with optimum robot configuration for condition
number. It was found (Figure 3) that the range of condition number variation in the entire feasible
workspace was from 2.3322 to 2.9241, which is within permissible limits. This also shows that the
robot configuration obtained is stable and has better manipulability in its feasible workspace for
given range of orientations. The condition number was checked at each point in the workspace,
which is represented by a set of three angles, however due to difficulty in plotting results with
respect to three orientations, results have been provided for three fixed yaw orientations ()
shown in Figures 4-6. These results again confirms that the condition number is well within
desired limits and varies smoothly in the vicinity for the entire feasible workspace. It appears from
the results that the optimum variables obtained are symmetric and hence they could be grouped
together to reduce complexity.

0
Figure 5. Condition number vs orientation at ( = 20 )

0
Figure 6. Condition number vs orientation at = 20

0
Figure 4. Condition number vs orientation at ( = 0 )

Global Condition Number

30

25

20

15

10

1000

2000

3000

4000

5000

6000

7000

Cable Force Norm

Figure 4. Static cable forces at different end effector orientations

1
0.95

Permissible Workspace

0.9
0.85
0.8
0.75
0.7
0.65
0.6
0.55
0.5
0

1000

2000

3000

4000

5000

6000

Cable Force Norm

Figure 4. Permissible workspace vs static cable forces

Table 1. Final kinematic design parameters

q1
q2
q3
q4
q5
q6
q7
q8
r1
r2
r3
r4
r5
r6
r7
r8

44.7415 deg
65.7147 deg
59.9703 deg
70.9323 deg
41.6373 deg
29.9708 deg
32.7772 deg
29.9708 deg
90 mm
114.8 mm
86 mm
94.5 mm
85 mm
104.5 mm
77 mm
86.4 mm

7000

11. CONCLUSIONS
A new soft parallel mechanism (SPM) for ankle joint rehabilitation has been proposed in the
chapter, after carefully studying the complexities of human ankle joint and its motions. The
proposed device is an improvement in terms of simplicity, rigidity and payload performance. The
proposed device is very light in weight (total weight is less than 1 Kg excluding the weight of
support mechanism) and is quite inexpensive. The workspace study is carried out for future
applications of task space trajectory control. Main objective of this work was to optimize the
kinematic design so that the performance of this robotic prototype can be enhanced. An
optimization problem was formulated wherein the optimum GCN was obtained (constrained by a
penalty term) by altering the configuration of connection points at the two parallel platforms.
Genetic algorithm is slightly modified to perform local search efficiently and for this a gradient
based selection operator is used in place of Roulette-Wheel Selection after a near optimal value
for GCN is obtained. GA finds the near optimum global solution very quickly and thereafter
gradient descent selection operator further narrows down the result. Simulation run at 3000
random points in the feasible workspace shows (Figure 3) that the resulting robot configuration is
quite robust and variation as well as range of condition number is within acceptable limits. Since it
is not possible to obtain GCN value equal to unity, the final values obtained for GCN and P (
which are 2.5572 and 2.9241) are reasonably acceptable.