8

th
World Congress on Structural and Multidisciplinary Optimization
June 1 - 5, 2009, Lisbon, Portugal
1
Isotropy design and optimization of a planar parallel manipulator with Combination of
Fuzzy Logic and Genetic Algorithm
Hamidreza Kordjazi
1
, Alireza Akbarzadeh-T
2
, Ali Shokouhi Rostami
3
1
M.Sc.student, Mechanical Engineering Department, Ferdowsi University oI Mashhad, Iran, Kordjazigmail.com
2
Assistant proIessor, Mechanical Engineering Department, Ferdowsi University oI Mashhad, Iran,
AliAkbarzadehTyahoo.com
3
M.Sc. Electrical Engineering, Mapna group oI iran, Alishokohirostamigmail.com
1. Abstract
In an isotropic conIiguration, the sensitivity oI a manipulator in both velocity and torque errors is at a minimum
and the manipulator can be controlled equally well in all directions. II the Jacobian matrix is isotropic throughout
the entire workspace, manipulator is Iully-isotropic and thereIore the condition number oI the Jacobian matrix is
one. Manipulator design can be expressed as a Iunction oI workspace requirements. This work presents a hybrid
Iuzzy logic - genetic algorithm (FL-GA) method Ior optimization and dimensional synthesis oI a 3PRR (prismatic-
revolute- revolute) planar parallel manipulator Ior a prescribed workspace. The algorithm is made oI classical
genetic algorithm coupled with Iuzzy logic. The Iuzzy logic controller monitors the variation oI genetic algorithm
variables during the Iirst run oI GA and modiIies the initial bounding intervals to restart the next run oI the
algorithm. Links oI robot are minimized while desired workspace is achieved.
2. Keywords: Genetic Algorithm, Fuzzy Logic, Isotropy, Workspace, Parallel Robot
3. Introduction
A parallel robot is deIined as a mechanism having at least two kinematics chains connecting the base to the end
eIIector. The perIormance oI a machine tool with parallel kinematics can be evaluated by its kinematic, static and
dynamic properties. Parallel manipulators have received great attention due to their properties oI increased
accuracy, high stiIIness, high Rigidity, high payload capability, high speed, good dynamic characteristics and
precise positioning capability. But one oI their drawbacks is that their perIormances depend heavily on their
geometry |1]. Due to their parallel topology, including limited workspace, diIIiculties in their analysis, synthesis,
control and trajectory planning |2|. Serial manipulators have disadvantages oI low precision, low stiIIness and low
power. Also, they are generally operated at low velocity to avoid excessive vibration and deIlection. Parallel robots
have many applications, such as industrial automation, telescopes, Iine positioning devices, Iast packaging,
machine tool, medical, Ilight simulators, micromanipulators, and parallel machine tools |3|. However the
optimization oI parallel robot can be challenging since it can involve many parameters, like workspace, physical
size oI robot, stiIIness, accuracy, singularity, isotropy |4|. Optimization oI parallel manipulators is an important
and challenging problem. There are two primary important issues in the Iield oI parallel manipulator: one is
mechanical architecture design and the other is optimum design oI the manipulator.
An isotropic manipulator has optimum dexterity when it reaches an isotropic position, so it is desirable to develop
manipulators that can reach more isotropic positions. With many isotropic positions to choose Irom, we can choose
a preIerable position Ior some speciIic applications. For example, an isotropic position with smaller singular
values provides high resolution Ior Iine position control oI a manipulator. The isotropic manipulators are obtained
by solving a system oI nonlinear equations developed by the isotropy conditions.
In this paper we Iirst introduce the isotropy problem Ior a planar parallel manipulator and then given speciIic area
as a desired workspace, Iind the minimum dimension Ior parallel robot where the workspace contains the desired
workspace |5, 6|.
The old techniques Ior optimization are all gradient based search methods and hence require the calculation oI
derivatives. This characteristic makes these techniques very demanding in computation time and in some cases
they may even Iail to converge. New methods based on artiIicial intelligence or probabilistic approach has
emerged. The most Iamous is genetic algorithm. Genetic algorithm is a robust method Ior searching the optimum
solution to a complex problem, although it may not necessarily lead to the best possible solution. GA generally
represents a solution using strings (also reIerred to as chromosomes) oI variables that represent the problem.
Choosing a large bounding interval could lead to the same problem because the limited number oI individuals oI
the population are scattered randomly over a large interval. Choosing large number oI individual could cover more
regions oI the interval, but will slow the optimization without guaranteeing the global optimum, thus we improve
the perIormance oI GA search ability through the adaptive search range mechanism through Iuzzy logic.
2
A combined Iuzzy logic - genetic algorithm method (FL-GA) is proposed. This algorithm has the capability to
adjust its starting population to avoid local minimum and to obtain accuracies that classical genetic algorithms Iail
to obtain. The process is started with an initial population chosen within the initial bounding intervals. The Iuzzy
logic controller monitors the evolution oI the diIIerent variables during the optimization and adjusts the bounding
intervals Ior each design variable. These new intervals are then used to start a second round oI optimization in
order to improve the Iinal result |7|.
4. Description of the manipulator
In this section, we introduce 3PRR planar parallel manipulator. The manipulator consists oI a base plate, a
movable platIorm, and three links, each oI which has a prismatic joint and two consecutive revolute joints. The
detailed description about the manipulator can be Iound in Figure 1. Only the prismatic joints are actuated. Three
degrees-oI-Ireedom (DOF) oI the PRR manipulator are the translations along the X and Y axis and the rotation
about the Z axis. P

is the length oI ith prismatic joint.

Figure 1. 3PRR Planar Parallel Manipulator
5. Isotropy
Generation oI smooth and bounded joint velocities greatly depends on the rank preservation and conditioning oI
the Jacobian matrix |8|. Isotropy oI a manipulator is related to condition number oI its Jacobian matrix, which can
be calculated as the ratio oI the largest and the smallest singular values. Jacobian matrix oI a robotic manipulator is
the matrix mapping the actuated joint velocity space and the end eIIector velocity space, and also the static load on
the end eIIector and the actuated joint Iorces or torques. For parallel manipulator we can write:
[t +Kp = u (1)
Where [ and K are Jacobian matrices, p is the vector oI actuated joint rates and t deIines as twist vector such that:
t = |x y ]
1
(2)
p = |p
1
p
2
p
3
]
1
(S)
The condition number oI the Jacobian matrix is not only a measure oI how accurate the Iorce and motion transIer
between the actuators and the end eIIector, but also the measure oI the ease with which the manipulator can
arbitrarily change its position and orientation and apply Iorces in arbitrary directions. Because oI this Iact, the
condition number oI the manipulator Jacobian matrix has been recognized as an optimization criterion in structural
synthesis oI manipulators even though it is not the only synthesis criterion |2|. The condition number is based on a
concept common to all matrices, whether square or not, i.e., their singular values. For an m n matrix A, with
mn, we can deIine its m singular values as the non-negative square roots oI the non-negative eigenvalues oI the
matrix AA
T
.
k =
o
m
o
s
(4)
Here o
m
and o
s
aie largest and smallest singular values. It is apparent that a singular matrix has a minimum
singular value oI zero, and hence, its condition number becomes inIinite and iI the singular values oI a matrix are
identical, then the condition number oI the matrix attains a minimum value oI unity. The reason why isotropic
matrices are desirable is that they can be inverted at no cost because the inverse oI an isotropic matrix or the
generalized inverse oI a rectangular isotropic matrix Ior that matter is proportional to its transpose, the
proportionality Iactor being the reciprocal oI its multiple singular values.
From the above discussion, and considering that the Jacobian matrices are conIiguration dependent, it is apparent
that the condition number oI the Jacobian matrices oI a manipulator is conIiguration-dependent as well. We deIine
a parallel manipulator as isotropic iI both [ and K are being the proportional to an identity matrix such that:
3
[ [
1
= o
2
I
33
(S)
K K
1
=
2
I
33
(6)
For the velocity oI point P (center oI mobile triangle) we can write:
P

= I

A
+ (I

B
- I

A
) + [P

- I

B
(7)
Also we have:
I

B
- I

A
= 0

F c (8)
Where E is rotation matrix and deIined as:
E = j
u -1
1 u
[ (9)
And [ and K obtained such that:
[ =
l
l
l
l
l
A
1
B
1

1
-[A
1
B
1

1
F(B
1
P

)
A
2
B
2

1
-[A
2
B
2

1
F(B
2
P

)
A
3
B
3

1
-[A
3
B
3

1
F(B
3
P

)
1
1
1
1
1
(1u)
K = _
A
1
B
1

u u
u A
2
B
2

u
u u A
3
B
3

_ (11)
For normalizing matrix [, as needed to deIine its condition number, each term oI the third column oI [ is divided
by the characteristic length L.
[ =
l
l
l
l
l
A
1
B
1

1
-[A
1
B
1

1
F(B
1
P

)I
A
2
B
2

1
-[A
2
B
2

1
F(B
2
P

)I
A
3
B
3

1
-[A
3
B
3

1
F(B
3
P

)I
1
1
1
1
1
(12)
By replacing (11) and (12) in Eq.(5) we have:
(A

)
2
+ ((A

) F (B

P)I)
2
= o
2
(1S)
A
i
B
i

1
A
]
B
]

+(( A
i
B
i

1
F (B
i
P

) ( A
]
B
]

1
F (B
]
P

))I)
2
= u i = ] (14)
ThereIore Irom relation (13) and (14) we should have:
A
1
B
1

= A
2
B
2

= A
3
B
3

(1S)
B
1
P

= B
2
P

= B
3
P

(16)
A
1
B
1
1
A
2
B
2
= A
2
B
2
1
A
3
B
3
= A
3
B
3
1
A
1
B
1
(17)
Also Ior isotropy oI matrix K we have:
(A
1
B
1
)
2
= (A
2
B
2
)
2
= (A
3
B
3
)
2
(18)
In summary, the constraints deIined in the Eq.(15) to (18 ) are:
Bi should be placed at the vertices oI an equilateral triangle, segments A

Iorman equilateral triangle; the axes oI

the prismatic pairs deIine an equilateral triangle and point P placed in the center oI the base and movable triangles.
ThereIore A
1
, A
2
and A
3
should be vertices oI an equilateral triangle. Also B

should made an equilateral triangle

as the vertices places on A
I
B
I
. We can see that there are inIinite numbers oI equilateral triangles Ior A
1
A
2
A
3
and
B
1
B
2
B
3
(Figure 2). This means that robot has many isotropic conIigurations.
Figure 2. Equilateral triangle Ior A
1
A
2
A
3
anu B
1
B
2
B
3
4
6. Workspace
The workspace oI a robot is deIined as the set oI all end eIIector conIigurations which can be reached by some
choice oI joint coordinates. Compared with the serial manipulators, parallel manipulators have relatively small
workspaces. Thus the workspace oI a parallel manipulator is one oI the most important aspects to reIlect its
working capacity, and it is necessary to analyze the shape and volume oI the workspace Ior enhancing applications
oI parallel manipulators |9|. ThereIore Ior design a manipulator it is oI particular interest to determine how the
workspace varies with diIIerent values oI the architectural parameters.
There are diIIerent types oI workspaces namely maximal workspace or reachable workspace, constant orientation
workspace, inclusive orientation workspace, and total orientation workspace. The maximal workspace or
reachable workspace is the set oI locations oI the end eIIector that may be reached with at least one orientation oI
the platIorm. The reachable workspace oI a 3PRR manipulator presented in Figure 3.

Figure 3. Workspace OI 3PRR Planar Parallel Manipulator
7. Genetic Algorithm
Genetic algorithm is a probabilistic technique that uses a population oI designs rather than a single design at a time
and utilizes the concept oI biological structure to natural selection and survival oI the Iittest. This method requires
no previous experience on the problem. An initial randomized population that consists oI a group oI chromosomes
and represent the problem variables, produces new populations through successive iterations, using various
genetic operators. The common operators are selection, mutation, crossover and elitism.GA use objective Iunction
inIormation, not derivatives or other auxiliary knowledge, use probabilistic transition rules, not deterministic rules.
Coding components oI possible solutions into a chromosome is the Iirst part oI a GA Iormulation. Each
chromosome is a potential solution and is comprised oI a series oI substrings or genes, representing components or
variables that either Iorm or can be used to evaluate the objective Iunction oI the problem. Chromosomes in the
population with high Iitness values have a high probability oI being selected Ior combination with other
chromosomes oI high Iitness. Combination is achieved through the crossover oI pieces oI genetic material between
selected chromosomes. Mutation allows Ior the random mutations oI bits oI inIormation in individual genes.
Through successive generations, Iitness should progressively improve.
8. Algorithm of the optimization
The aim oI this section is to develop and to solve the optimization problem oI selecting the geometric design
variables Ior the 3PRR planar parallel manipulator having a prescribed workspace. Given a speciIied area deIines
by some points and we want to Iind the parameters oI the 3PRR robot having a workspace that includes the
speciIied area. To deIine the objective Iunction, we note this theorem that the total distance oI a point in the
equilateral triangle Irom 3 lines oI the equilateral triangle is equal to the height oI equilateral triangle. The
algorithm is as Iollow:
Select 2 points A(x
1
, y
1
) and B(x
2
, y
2
) and compose 2 equilateral triangles with these 2 points (Figure 4). Then
we determine the distance oI each point oI prescribed workspace Irom 3 lines oI 2 triangles. II the total distance in
each part is equal to the height, this triangle can be the workspace oI 3PRR planar parallel manipulator. We want to
Iind the minimum area Ior this triangle.
A penalty Iunction method is used to handle the constraints. The objective Iunction is accordingly constructed as:
F = F
1
+ F
2
(19)
Where:
F
1
= (x
1
- x
2
)
2
+ (y
1
- y
2
)
2
(2u)
And F
2
is a penalty Iunction deIined as Iollows:
5
F
2
= C. I (21)
Here, C is a large positive constant and V is a number that relate to the minimum number oI point oI desired
workspace where outside oI 2 triangles.
Figure 4. Objective Function deIinition
Figure 5 shows the combined FL-GA algorithm. The process, which is the classical genetic algorithm
optimization, is started with initial population chosen randomly within the initial bounding intervals. The FL
monitors the evolution oI the diIIerent variables during the optimization process and adjusts the bounding intervals
oI each variable Ior the next round oI optimization process. These new intervals are then used to start a second
round oI the genetic algorithm.
Figure 5. The modiIied FL- GA optimization scheme
The knowledge base is built during the Iirst run oI the genetic algorithm optimization. By analyzing the Iinal
result, a parameter
x
is determined and the bounding interval Ior the design parameter x is corrected using the
Iollowing equation:
x
mn
-
= x
uc
-

x
2
(x
mux
- x
mn
) (22)
x
mux
-
= x
uc
+

x
2
(x
mux
- x
mn
) (2S)
Where x
uc
is the average value oI the design variable x oI all the individuals oI the last generation and
6
|x
mn
x
mux
] is the initial bounding interval. The coeIIicient
x
(the output variable) is obtained Irom the
knowledge oI the two inputs, variables E and K
x
where E is the error Iound aIter the Iirst run oI the genetic
algorithm optimization and K
x
is a counter oI the variation oI each parameter during the last 30 generations. K
x
is a counter Ior each one oI the variables and it is ranging between 0 and 30. It starts at 0 and during the last 30
generations it is incremented by 1 each time the variable changes.
E =
x
Iust
- x
mn
x
mn
(24)
FuzziIication oI the input variables is the Iirst step in the design oI a Iuzzy logic controller. FuzziIication oI the
input variables involves quantizing the universes into a number oI Iuzzy sets. The output variables also need to be
quantized in a similar manner. Quantization involves breaking up a Iuzzy input (and also output) variable into
several Iuzzy subsets. Figure 6 shows the membership Iunctions chosen Ior 2 Iuzzy input variables and the output
variable and Table 1 contains the deIinition oI the linguistic parameters.
Figure 6. Membership Iunction Ior two inputs and the output
.
Table 1.The linguistic parameters
Very Small VS
Small S
Medium M
High H
Very High VH
Table 2. Rule matrix
D
x
K
x

VS S M H VH
VS
VS S S M M
S
VS S M H M
M
S S H M H
H
M M M H H
VH
H H M M M
Table 3. Coordinates oI the points oI desired workspace
Point number 1 2 3 4 5 6 7
x (cm) 5 12 26 30 20 10 2
y (cm) 5 2 10 17 23 25 14
7
Figure 7. Desired workspace Ior 3PRR robot
Table 4. GA parameters
30 The Size OI The Population
80 Maximum Number OI Generations
4 Number OI Variables
8 Number OI Bits
0.7 Crossover Rate
0.02 Mutation
9. Results
Figure 8 shows the comparison oI two methods Ior 80 generations and table 5 shows the initial and corrected
boundary Ior each parameter oI GA. Results shows the eIIiciency oI combination oI FL with GA.
Figure 8. Comparison oI GA and FL-GA
Table 5. Parameters oI the FL-GA method
X1 Y1 X2 Y2
Initial boundary |0 80 | |0 80 | |0 80 | |0 80 |
Kx 12 14 10 16
E 0.0029
Dx 1.4940 1.3812 1.1921 1.4116
Corrected boundary |18.6 54.6| |-1.5 34.7| |-12.3 18.5| |12.7 47.8|
F (GA) 40.47
F (FL-GA) 37.84
8
ThereIore:
H
1
H
2

= H
2
H
3

= H
1
H
3

= S7.84 cm (25)
A
i
B
i

+ B
i
P

=
V3
2
H
1
H
2

= S2.77 cm (26)
10. Conclusion
In this paper we Iirst discuss about isotropy oI the Jacobian matrices oI parallel manipulator and used this concept
to design a 3PRR planar parallel manipulator. We obtained the relation between the parameters oI manipulator.
Then we presented a combined Iuzzy logic - genetic algorithm method Ior workspace optimization oI a 3PRR
planar parallel manipulator. The proposed method is made oI a classical genetic algorithm coupled with a Iuzzy
logic controller. This controller monitors the variation oI the variables during the Iirst run oI the genetic algorithm
and modiIies the initial bounding intervals to restart a second round oI the genetic algorithm. The desired
workspace oI the robot was obtained. The links Ior this robot are all minimized. Using this method showed that
these results are always better than those obtained by genetic algorithm lonely.
11. References
|1| X.J. Liu, J. Wang and J. Kim, Determination oI the Link Lengths Ior a Spatial 3-DOF Parallel Manipulator,
Mechanical Design, 128 (365), 2006.
|2| G. Alici and B. Shirinzadeh, Optimum synthesis oI planar parallel manipulators based on kinematic isotropy
and Iorce balancing , Robotica, 22, 97108, 2004.
|3| J. P. Merlet, Jacobian, Manipulability, Condition Number, and Accuracy oI Parallel Robots, Mechanical
Design, 128 (199), 2006.
|4| S. Stan, M. Manic, V. Maties , And R.Balan, Evolutionarv Approach To Optimal Design Of 3 DOF
Translation Exoskeleton And Medical Parallel Robots, HSI 2008, Krakow, Poland, May 25-27
|5| M.A. Laribi, L.Romdhane, S.Zeghloul, Analysis and dimensional synthesis oI the DELTA robot Ior a
prescribed workspace", Mechanism and Machine Theorv, 42, 859870, 2007.
|6| Z. AIIi , L. Romdhane , A. Maalej, Dimensional synthesis oI a 3-translational-DOF in-parallel manipulator
Ior a desired workspace", European Journal of Mechanics, 23, 311324, 2004.
|7| M.S. Osman, Mahmoud A. Abo-Sinna, A.A. Mousa, A combined genetic algorithm-Iuzzy logic controller
(GAFLC) in nonlinear programming, Applied Mathematics and computation, 170, 821840, 2005.
|8| K.Y. Tsai and K.D. Huang, The design oI isotropic 6-DOF parallel manipulators using isotropy generators,
Mechanism and Machine Theorv, 38, 11991214, 2003.
|9| Y. Li, Kinematic Analysis and Design oI a New 3-DOF Translational Parallel Manipulator, Mechanical
Design, 128 (729), 2006.