Combined

Robotics and Computer-Integrated Manufacturing 35 (2015) 151168
Contents lists available at ScienceDirect
Robotics and Computer-Integrated Manufacturing

journal homepage: www.elsevier.com/locate/rcim
Geometric calibration of industrial robots using enhanced partial pose

measurements and design of experiments
Yier Wu a,b, Alexandr Klimchik a,b,n, Stphane Caro b, Benot Furet c, Anatol Pashkevich a,b
a
Ecole des Mines de Nantes, 4 rue Alfred-Kastler, Nantes 44307, France

CNRS, Institut de Recherches en Communications et et Cyberntique de Nantes, UMR CNRS 6597, 1 rue de la No, 44321 Nantes, France
c
Universit de Nantes, France
b
art ic l e i nf o
a b s t r a c t
Article history:
Received 14 November 2014
Received in revised form
9 March 2015
Accepted 12 March 2015
Available online 26 March 2015
The paper deals with geometric calibration of industrial robots and focuses on reduction of the measurement noise impact by means of proper selection of the manipulator congurations in calibration
experiments. Particular attention is paid to the enhancement of measurement and optimization techniques employed in geometric parameter identication. The developed method implements a complete
and irreducible geometric model for serial manipulator, which takes into account different sources of
errors (link lengths, joint offsets, etc). In contrast to other works, a new industry-oriented performance
measure is proposed for optimal measurement conguration selection that improves the existing
techniques via using the direct measurement data only. This new approach is aimed at nding the calibration congurations that ensure the best robot positioning accuracy after geometric error compensation. Experimental study of heavy industrial robot KUKA KR-270 illustrates the benets of the developed pose strategy technique and the corresponding accuracy improvement.
& 2015 Elsevier Ltd. All rights reserved.
Keywords:
Serial industrial robots
Geometric calibration
Enhanced partial pose measurement
Design of experiments
Industry-oriented performance measure
1. Introduction
In robotic literature, the problem of geometric calibration is
already well studied and has been in the focus of the research
community for many years [18]. As reported by a number of
authors, the manipulator geometric errors are responsible for
about 90% of the total positioning error [9]. Besides of the errors in
link lengths and joint offsets, the end-effector positioning errors
can be also caused by the non-perfect assembling of different links
and arise in shifting and/or rotation of the frames associated with
different elements, which are normally assumed to be matched
and aligned [10]. It is clear that the geometric errors do not vary
with the manipulator conguration, while their inuence on the
positioning accuracy depends on the latter. At present, there exist
various calibration techniques that are able to calibrate the manipulator geometric model using different modeling, measurement and identication methods [1116]. The identied errors can
be efciently compensated either by adjusting the controller input
(the target point) or by direct modication of the model parameters used in the robot controller.
The classical calibration procedure usually includes four steps:
n
Corresponding author at: Ecole des Mines de Nantes, 4 rue Alfred-Kastler,
Nantes 44307, France. Fax. 33 251 85 83 49.
E-mail address: alexandr.klimchik@mines-nantes.fr (A. Klimchik).
http://dx.doi.org/10.1016/j.rcim.2015.03.007
0736-5845/& 2015 Elsevier Ltd. All rights reserved.
modeling, measurement, identication and implementation. The

Modeling step focuses on the development of proper geometric
model of robotic manipulator. In the pioneer works [14], researches have used the classical DH convention for robot calibration. However, this model turned out to be discontinuous in some
cases and may lead to unacceptable identication results [17]. So,
several alternative approaches have been proposed to overcome
these difculties by means of introducing extra parameters [18,19].
Since the inclusion of additional parameters causes redundancy,
these methods raise the problem of parameter non-identiability,
which leads to the necessity of investigating the model completeness, irreducibility and continuity. For example, in [20], the
authors proposed a complete and parametrically continuous (CPC)
model and further its modied version (MCPC) for robot calibration. Besides, there have been also proposed some analytical/numerical techniques for elimination of the non-identiable parameters. For example, in [18], the authors used QR decomposition
of the identication Jacobian for model reduction and in [21], the
authors used straightforward evaluation of the Jacobian matrix
rank.
The Measurement step involves data collecting of robot link and
end-effector position/orientation. Generally, six parameters are
required to specify the manipulator end-effector location (three
translations and three rotations) [12,22], but sometimes the endeffector position is measured only [23]. Various calibration
methods based on different measurement techniques were
152
Y. Wu et al. / Robotics and Computer-Integrated Manufacturing 35 (2015) 151168
proposed, they are usually categorized as closed-loop and openloop ones. The closed-loop calibration uses physical constraints on
the manipulator end-link (point, line or plane constraints, for instance). It is claimed to be autonomous and does not require any
external device [13,21,24]. However in this case, the manipulators
must have some redundancy to perform self-motion, and the robot
conguration should be carefully selected to satisfy particular
constraints. Therefore, the open-loop methods have found wide
applications; they are based on the full or partial pose measurements of the end-effector location using external devices. In
practice, the partial pose information is often used and provides
from one to ve dimensional measurements [11,25,26] instead of
the full pose information (6-dimensional location). In general, the
lower dimensional measurement is more attractive due to simplicity of calibration experiment setup. For this so-called partial
pose measurement technique, various external devices can be applied, such as laser tracking system [23], the ball-bar system [27]
and wire potentiometer [22], etc.
The identication step in robot calibration can be treated as the
best tting of the experimental data (given input variables and
measured output variables) by corresponding models. This problem has been addressed by a number of researchers who have
used various modeling methods and identication algorithms,
such as linear least square technique, LevenbergMarquardt algorithm, Kalman ltering technique and maximum likelihood estimator etc. [16,28]. Among them, the least square technique is the
most often applied one, which aims at minimizing the sum of
squared residuals [29]. An important problem here is non-homogeneity of the residual errors (distances and angles, for instance).
To solve this problem, usually a straightforward solution is applied: assigning weights or normalization, but this weight assigning procedure is very non-formal and not rigorous (while
being essential for the nal results). To solve the corresponding
optimization problem, there exist various numerical algorithms
such as gradient search [27,30], heuristic search and the others
[31]. However, these numerical techniques are often difcult to
apply due to large number of parameters to be tuned, that often
lead to low convergence. Nevertheless, for the case of geometric
calibration, the errors in the parameters are relatively small, so the
linearization technique can be successfully applied. In this case,
the solution of a linear least square problem can be found

straightforwardly (i.e., via the pseudo-inverse of MoorePenrose)
[32,33]. It should be mentioned that in some particular cases, for
instance, when the geometric errors are relatively large, the solution can only be found iteratively [15].
The most essential works on the above mentioned calibration
methods in robotics literature are summarized in Table 1. Among
these publications, limited number of works directly addresses the
problem of parameter identication accuracy and reduction of the
impact of measurement errors. Although the calibration accuracy
may be improved by straightforwardly increasing the number of
experiments [27], the measurement congurations may also affect
the robot calibration [34]. It has been shown that the latter may
signicantly improve the identication accuracy [35]. Intuitively,
using diverse manipulator congurations for different experiments seems perfectly corresponds to the basic idea of the classical experiment design theory, which intends to spread the
measurements as much distinct as possible [15]. However, the
classical results are mostly obtained for very specic models (such
as the linear regression) and cannot be applied directly due to
non-linearity of the relevant expressions of robot geometric
model.
At present, there are few works where the problem of optimal
pose selection for robot calibration has been discussed [39,40]. In
these works, in order to compare the plans of experiments, several
quantitative performance measures have been proposed and used
as the objectives of the optimization problem associated with the
optimal sets of measurement poses. In dening the objectives, the
authors in [35,4042] proposed some observability indices, which
are based on the singular values of the identication Jacobian
(condition number, for instance). These indices have been examined and compared in [38,39,43,44], where the authors paid
more attention to developing efcient numerical algorithms, such
as genetic algorithm, Tabu search, DETMAX and also hybrid
methods in order to obtain the optimal measurement congurations. However, these approaches deal with rather abstract notions
that are not directly related to the robot accuracy and may lead to
some unexpected results, for example, when the condition number is good, but the parameter estimation errors are rather high.
Besides, it usually requires very intensive and time consuming
Table 1
Summary of related works for geometric calibration
Application (Manipulator) Number of model
parameters
Number of measurement
congurations
Measurement device
Identication algorithm
Achieved accuracy,
[mm]
6-dof parallel robot [25]
35
80(1)
Two inclinometers(a)
0.40
Stewart platform [36]

PUMA 560 [23]
PUMA 560 [27]
PUMA 560 [22]
PUMA 560 [13]
Schilling Titan II [37]
Stubli TX90 [15]
SCARA robot [38]
Gough platform [39]
42
27
36
24
23
42
23
30
42
15(1)
25(1)
800(1)
48(1)
100(3)
800(2)
100(2)
10(4)
18(5)
Single theodolite(a)
Laser tracking system(a)
Ball-bar system(a)
Wire potentiometer(a)
(b)
(b)
Touching probe(b)
(b)
Vision system(c)
LevenbergMarquardt
method
Non-linear LS
Gradient search method

Non-linear LS
Non-linear LS
Linear LS
Weighted pseudo inverse
Genetic algorithm
Heuristic search
Selection of measurement congurations:

1
Random congurations.
Well distributed congurations.
Noise amplication index.
4
Minimum condition number.
5
Several observability indices.
Measurement technique:
2
3
a
b
c
Open-loop measurement.
Closed-loop measurement.
Simulation.
0.50
0.10
0.08
0.50
0.25
5.70
0.22
3.60
1.30
computations caused by a poor convergence and high dimension

of the search space (number of calibration experiments multiplied
by the manipulator joint number). Therefore, for the industrial
applications, existing approaches should be essentially revised.
The primary goal of this work is to achieve the desired robot
positioning accuracy using minimum number of experiments.
Here it is proposed to introduce an additional step to the classical
calibration procedure, the design of experiments, which is performed before measurements and is aimed at obtaining the set of
measurements poses that ensures good calibration results (robot
accuracy after error compensation). It allows us to improve the
efciency of error compensation and to estimate the robot accuracy, which is important for industrial applications.
To address the above mentioned problems, the remainder of
the paper is organized as follows. Section 2 presents the problem
of geometric calibration in general. Section 3 describes a suitable
manipulator geometric model for calibration purposes (complete,
irreducible model). Section 4 contains one of the main contributions: an enhanced partial pose measurement method. Section 5
describes a dedicated identication algorithm for manipulator
geometric parameters. Section 6 proposes a new approach for
optimal measurement pose selection and evaluates the calibration
efciency improvement. Section 7 presents the experimental results obtained for the geometric calibration of a KUKA KR-270
robot. Finally, Section 8 summarizes the main results and contributions of this paper.
153
2. Problem of geometric calibration

In robotics, calibration is a process that allows us to estimate
the manipulator geometric parameters, which are employed in
robot controller. In practice, the nominal values of these parameters are different from the real ones, so they should be identied for each particular manipulator using data from calibration
experiments. As it was mentioned before, the conventional calibration procedure includes four sequential steps. In the scope of
this paper, an additional step is introduced that deals with the
design of calibration experiments, in order to improve the calibration accuracy. A relevant enhanced robot calibration procedure
is presented in Fig. 1. The particularities of each step of this procedure are described below.
Step 1: This step deals with manipulator modeling and is aimed
at developing a geometric model that is suitable for calibration
(complete and irreducible), i.e. which is good enough from physical point of view and does not create any numerical problems
during identication. This model should allow us to compute the
end-effector location for any given values of the actuated joint
coordinates q (provided that the manipulator parameters are
known). However, for calibration purposes, it is usually required a
linearized version of this model allowing to evaluate the inuence
of the small variations of q and . So let us assume that the manipulator links are rigid and corresponding geometric model can
be written as the vector function t=g (q, ), where t=(p, )T denes
the manipulator end-effector location (position and orientation),
vector q contains all actuated joint coordinates, and vector
= 0 + collects all geometric parameters and their
Fig. 1. Enhanced robot calibration procedure.
154
deviations. Under this assumption, the actual location of the manipulator end-effector, which incorporates the geometric errors is
expressed as t=g (q, 0 + ). In practice, the geometric errors
are usually relatively small, therefore the following linerized
model can be used
t=t 0 + J
(1)
where t0 = g (q, 0 ) is the end-effector location computed using

the nominal geometric parameters, J is the identication Jacobian
matrix, which can be computed using the derivative g (q, 0 )/ .
More details concerning the computation of identication Jacobian
can be found in [45].
Step 2: This is an additional step (design of experiments) that is
introduced here in the calibration procedure. It is aimed at
choosing optimal measurement congurations for calibration experiments. It should rely on an appropriate performance measure,
which takes into account the particularities of the technological
process (robotic-based machining, for instance). It should be also
able to obtain solution within the work-cell constraints, and to
adjust the number of experiments with respect to the measurement system precision. In practice, the inuence of the geometric
errors on the end-effector position varies from one conguration
to another and essentially differs throughout the workspace. So,
the desired accuracy is usually required to be achieved for rather
limited workspace area (for example, where the workpieces are
located in the robotic cell). For this reason, in this paper, it is
proposed to limit the benchmark manipulator congurations by a
single one (the machining conguration, for instance), which will
be further referred to as the manipulator test-pose. To develop a
new approach of calibration experiments design that utilizes the
above proposed ideas, let us introduce several basic denitions:
Denition 1. The plan of experiments is a set of manipulator
congurations {qi , i = 1, m} that are used for the measurements of
the end-effector positions {pij , i = 1, m , j = 1, n} and for further
identication of the desired parameters .
Denition 2. The manipulator test-pose is a particular robot conguration q0 (that is usually specied in relevant technological
process), for which it is required to achieve the best compensation
of the end-effector positioning errors.
Denition 3. The quality of the plan of experiments is dened by
the efciency of manipulator positioning error compensation at
the test-pose, which is the root-mean-square distance 0 between
the desired manipulator end-effector position and the position
obtained after error compensation.
Step 3: This step (measurements) deals with carrying out calibration experiments using the obtained congurations. Depending
Fig. 2. Measurement tool with several reference points.
on the measurement methods (measurement tools and devices,

reference point locations, see Fig. 2, where a typical manipulator
mounting ange is shown), it may provide different experimental
data (the end-effector position/location, etc.). For the conventional
full-pose measurement technique that is frequently used in robot
calibration, the corresponding optimization problem allowing us
to compute the desired parameters is expressed as
m
ti Ji 2
i=1
min
(2)
However, the residual components of this system of identication equations are non-homogeneous (millimeters and radians,
for instance). In some cases, these components are normalized
before computing the squared sum, but it is a non-trivial step that
affects the identication accuracy. To overcome this difculty, it is
proposed to enhance the partial pose measurement method that
uses directly and only the positioning coordinates, but for several
reference points for each manipulator conguration. More details
of this method and its advantages will be presented in Section 4.
Step 4: This step deals with the identication and is aimed at
estimating the geometric parameters by using the corresponding
model and proper identication algorithm. Usually, the identication algorithms are based on the minimization of the leastsquare objectives that are derived assuming that the measurement
tool has a single reference point (see Eq.(2)), while the proposed
measurement technique operates with several of them. For this
reason, it is required to revise the existing identication techniques, taking into account both modication of the objective
function and increasing of the number of parameters (since each
reference point introduces additional parameters).
In addition, this step includes the evaluation of the parameter
identication accuracy. In practice, different sources of error may
affect the identication precision. They include the measurement
errors of the external device providing the end-effector position
coordinates (laser tracker in our case), the errors in the actuator
encoders (internal measurement devices) giving the manipulator
joint coordinates that depend on encoder resolution, etc. Besides,
the assumption concerning the manipulator model (the link rigidity, for instance) may also affect the identication accuracy. It is
clear that, all these sources of error can be hardly taken into account in calibration. For this reason, only the most signicant of
the above mentioned sources of error should be considered in the
accuracy analysis. As follows from our experience, the inaccuracy
of external measurement system has the most signicant impact
on the robot positioning accuracy, comparing to other sources of
error that can be assumed negligible in the frame of geometric
calibration.
Step 5: At the last step (implementation), the geometric errors
are compensated by modication of the geometric parameter values embedded in the robot controller. In the case when some
errors cannot be entered in the controller directly, an off-line error
compensation technique is required. This technique should compensate the manipulator errors via modication of the target
trajectory that becomes slightly different from the desired one
[46].
It is clear that the proposed scheme of robot calibration procedure allows us to improve the calibration accuracy for given
number of experiments (or to minimize the number of experiments for given accuracy). The steps 1, 4 and 5 in the calibration
procedure have been already well studied [9,47], while the steps
2 and 3 still require some revision in terms of the applicability to
particular manufacturing process where the robot is used. Hence,
the goal of this work is the enhancement of calibration technique
for manipulator geometric parameters using enhanced partial
pose measurement and design of experiments. Particular
problems that should be considered are the following:

(1) Development of an industry-oriented performance measure
which has clear physical meaning that is related to the robot
accuracy after geometric error compensation.
(2) Enhancement of partial pose measurement method that allows us to avoid the problem of non-homogeneity in the
identication equations.
(3) Experimental validation of the developed approach for geometric calibration of an industrial KUKA KR-270 robot.
155
number of techniques that allows us to obtain the manipulator

model of such type, which is denitely complete but includes redundant parameters to be eliminated (methods of Hayati, WhitneyLosinski, etc.). In this work, we will use the model generation
technique that is based on dedicated analytical elimination rules
and includes the following steps:
Step 1. Construction of the complete and reducible model in
the form of homogeneous matrices product.
The base transformation Tbase = TxTyTz R xR yR z

The joint and link transformations TJoint, jTLink, j b
These problems will be considered in more details in the following sections.
(1) For revolute joint TJoint, jTLink, j = R e, j (q j , qj )[TuTvR uR v]Lj

(2) For prismatic joint TJoint, jTLink, j = Te, j (q j , qj )[R uR v]Lj
where e j is the joint axis, u j and v j are the axes orthogonal to e j .
3. Manipulator geometric modeling

To be suitable for robot calibration, the manipulator geometric
model must satisfy certain requirements. In particular, it should be
complete, i.e. is able to describe all possible errors in link/joint
geometry, but not redundant (i.e. does not contain parameters that
inuence the end-effector position/orientation in the same way
for any manipulator conguration). In previous works [48] [49], it
was shown that the conventional D-H model may produce problems for parameter identication because its incompleteness. To
avoid this difculty, some modications have been proposed that
however introduce some redundancy, which may cause nonidentiability of certain parameters. This redundancy can be
eliminated by applying either numerical or analytical techniques
[20,21,50] that allow us to obtain an appropriate model, which is
usually referred to as complete, irreducible and continuous one.
Let us apply one of these techniques [51] to generate the desired
model for heavy industrial manipulator KUKA KR-270 that will be
used in experimental study.
In the frame of the above dened notations and assuming that
the manipulator links are rigid enough and the non-geometric
factors are negligible in this level of calibration, the general expression of the geometric model for a n-dof serial manipulator can
be described as a sequence of homogeneous transformations
T(q) = Tbase ( b )TJoint (q1, q1)TLink ( L1) ...
TJoint (qn, qn )TLink ( Ln ) Ttool (t )
The tool transformation Ttool = TxTyTz R xR yR z t

Step 2. Elimination of non-identiable and semi-identiable
parameters in accordance with specic rules for different nature
and structure of consecutive joints.
For the case of consecutive revolute joint R e, j (q j , qj )
(1) if e j e j 1, eliminate the term R u, L j 1 or R v, L j 1 that corresponds to R e, j ;

(2) if e j e j 1, eliminate the term Tu, L j k or Tv, L j k that denes the
translation orthogonal to the joint axes, for which k is
minimum (k 1).
For the case of consecutive prismatic joint Te, j (q j , qj )
(1) if e j e j 1, eliminate the term Tu, L j 1 or Tu, L j 1 that corresponds to Te, j ;
(2) if e j e j 1, eliminate the term Tu, L j k or Tv, L j k that denes the
translation in the direction of axis e j , for which k is minimum (k 1).
Let us apply the above presented technique to the industrial

robot KUKA KR-270 (See Fig. 3), which is used in experimental
validations of this paper. For this manipulator that includes six
revolute joints, the complete (but redundant) model contains 42
parameters and can be presented as
(3)
where T with different indices denote the relevant transformation

matrices of size 4 4 , q is the vector of the actuated joint coordinates, while the vectors b , t , Lj and the scalars qj are the
manipulator geometric parameters corresponding to the base,
tool, links and joints, respectively. In the literature, there are a
Fig. 3. The industrial serial robot KUKA KR-270 and its geometric parameters. (a) Industrial robot KUKA KR270 and (b) the manipulator architecture.
156
T=TxTyTz R x R y R z R z (q1 + q1)TxTyR x R y

b
L1
where q j is the joint offset, p xj , p yj , p zj and xj , yj , zj are the

relevant translational and rotational parameters, and j indicates
the joint/link number. For these parameters, the corresponding
nominal values are
R y (q2 + q2 )[TxTz R x R z ]L 2
R y (q3 + q3 )[TxTz R x R z ]L 3 R x (q4 + q4 )
0 = { d2 0 0 0 d 3 0 0 0 d 4 d5 0 0 0 0 0 0 0 0 }
TyTz R y R z R y (q5 + q5 )[TxTz R x R z ]L5

L4
R x (q6 + q6 )TyTz R y R z TxTyTz R x R y R z
L6
t
(4)
It should be mentioned that the nominal values of some

parameters can be found in the manufacturer datasheets, but the
remaining ones are assumed to be equal to zero. Applying the
elimination rules for the case of consecutive revolute joints, the
parameters {R y, L1, Tz, L2, R x, L 3, R y, L 4, Tx, L5, R x, L5} are sequentially
eliminated from the redundant model (4). Here, it is worth making
the following remarks:
Remark 1. In the redundant model (4), it has been already taken
into account that the nominal geometric parameter d1 (shift of the
robot base along z-axis) cannot be identied separately from the
base transformation.
Remark 2. For the rst and the last joints, which are connected to
the robot base and tool respectively, the offsets q1 and q6 are
treated as semi-identiable parameters. So, they are eliminated
from the manipulator geometric model and are incorporated in
the base and tool parameters. However, the actuated joint variables q1 and q6 must retain in the model.
Remark
3. The geometric parameters of the last link

cannot be identied separately from the tool
transformation. So, it is reasonable to include these parameters in
the tool transformation.
{Ty, Tz, R y, R z}L6
Remark 4. In the case when only position measurements are

available, the tool orientations are not known. So, the parameters
of the rotational transformations {R x, R y, R z } corresponding to the
t
tool are treated as non-identiable.
Remark 5. Six parameters describing the base transformation
three parameters {Tx, Ty, Tz } that dene
t
tool transformation can be treated as known (there are dedicated
techniques to identify them separately).
{Tx, Ty, Tz, R x, R y, R z}band
This nally allows us to obtain the complete and irreducible

geometric model for the considered manipulator that includes 18
principle parameters to be identied.1
Trobot = R z (q1)TxTyR x
L1
R y (q2 + q2 )[TxR x R z ]L 2 R y (q3 + q3 )[TxTz R z ]L 3

R x (q4 + q4 )TyTz R z R y (q5 + q5 )[Tz R z ]L5 R x (q6 )
L4
(5)
This model will be further used for geometric calibration of the

industrial robot KUKA KR-270 and for the optimal selection of the
measurement poses. Let us collect these parameters in the following vector
= px1 p y1 x1 q2 px2 x2 z2 q3 px3 pz 3 z 3 q4 p y4 pz 4 z 4

q5 pz5 z5
(6)
1
It should be stressed that 6 parameters related to the base transformation
and 6 parameters describing the tool transformation are not included in this expression (see Remarks 1 and 5), so it is in good agreement with common expression of Zhuang [52] that for robot with 6 rotational joints yields 30 independent
parameters.
(7)
where the geometric meaning of d2, ... , d5 is illustrated in Fig. 3. In

the following sections, this model will be used for computing the
end-effector location of the KUKA KR-270 robot required for some
numerical routines employed in parameter identication algorithms.
4. Enhanced partial pose measurement method

In industrial applications, it is often used the partial pose
measurement method that requires obtaining the end-effector
position coordinates only (without orientation). On the other
hand, this simplication does not allow the user to identify certain
manipulator parameters that can be estimated via the end-effector
orientation. For this reason, this section presents an intermediate
technique, where the orientation is not computed directly but is
incorporated in the identication equations via the Cartesian coordinates of several reference points.
For the conventional full pose measurement technique, the desired parameters are identied from the full-scale linearized geometric model (1), which can be rewritten as
ti = Ji , i = 1, 2, ... m
(8)
where t i = (p xi , p yi , p zi , xi , yi , zi )T is the pose deviation caused by small variation in the model parameters . It is
clear that the corresponding system of linear equations can be
solved with respect to if the number of experiments m is sufciently high and the manipulator congurations {qi , i = 1, m} are
different to ensure non-singularity of relevant observation matrix
used in the identication procedure. For this technique, each
conguration qi produces six scalar equations to be used for the
identication. Corresponding optimization problem (2) whose
solution leads to the desired parameters is often solved without paying attention to the non-homogeneity of the residual
components. In some cases, the weighted least-square technique
is used to resolve this problem, but the weighting coefcients are
usually dened intuitively, which may affect essentially the
identication accuracy.
The main difculty of this conventional technique (full-pose) is
that the orientation components ( xi . yi . zi )T cannot be measured directly. So, these angles are computed using excessive
number of measurements for the same conguration qi , which
produce Cartesian coordinates {(p xij , p yij , p zij )T | j = 1, n ; n 3} for
several reference points of the measurement tool attached to the
manipulator mounting ange (Fig. 2). Hence, instead of using 3mn
scalar equations, that can be theoretically obtained from the
measurement data, this conventional approach uses only 6m scalar
equations for the identication. This may obviously lead to some
loss of the parameter estimation accuracy.
To overcome this difculty, the proposed technique is based on
reformulation of the optimization problem (2) using only the data
directly available from the measurement system, i.e. the Cartesian
coordinates of all reference points pij (see Fig. 2). This idea allows
us to obtain homogeneous identication equations where each
residual has the same unit (mm, for instance), and the optimization problem is rewritten as follows
m
pij Jji(p) 2
i=1 j=1
min
(9)
157
Fig. 4. Difference between conventional full-pose approach and enhanced partial pose approache.
Here, the matrix J j i(p) with the superscripts (p) denotes the
position rows of the corresponding identication Jacobian J j i , the
index i denes the manipulator conguration number, and the
index j denotes the reference point number. An obvious advantage of this formulation is its simplicity and clarity of the residual vector norm denition (conventional Euclidian norm can be
applied here reasonably, the normalization is not required). So, the
problem of the weighting coefcient selection does not exist in
this case. In fact, under the assumption that measurement errors
are modeled as a set of independent and identically distributed (i.i.
d.) random values (similar for all directions x, y, z and for all
measurement congurations), the optimal linear estimator should
operate with equal weights for all equations. Besides, the most
important issue is related to the potential benets in the identication accuracy, since the total number of scalar equations incorporated in the least-square objective increases from 6m to 3mn.
To compare the efciency of the presented approach with the
conventional one, a simulation study has been carried out, which
dealt with geometric calibration of a 3-link spatial manipulator
(Fig. 4). Detail description of this example can be found in [53],
where it has been proved that the enhanced technique based on
partial pose information ensures essential improvement of parameter identication accuracy. Using these identied geometric
parameters, it is possible to evaluate the manipulator end-effector
positioning accuracy throughout the workspace. Corresponding
results are shown in Fig. 4, in which the achieved robot accuracy
has been compared for two techniques. As follows from this gure,
using the proposed approach, the maximum positioning error has
not even reached the minimum one by using conventional technique. Moreover, the minimum positioning error has been reduced
by a factor of 4. Fig. 5
Therefore, the partial pose technique is rather promising and
will be further used for calibration experiments in this work. In
contrast to the conventional methods, this technique allows us to
avoid the problem of non-homogeneity of the relevant optimization objective and does not require any normalization (which
arises in the case when full pose residuals are used).
5. Identication of manipulator geometric parameters

5.1. Identication algorithm for the enhanced partial pose method
Let us assume that the measurement tool has n reference
points (n 3) that are used to estimate relevant vectors of the
Cartesian coordinates pij = (p xij , p yij , p zij )T for m manipulator congurations qi . In this notation, the subscript i and subscript j
denote the experiment number and reference point number respectively. Correspondingly, the manipulator geometric model (3)
can be rewritten as
j
T ij = TbaseTrobot (qi , )T tool
; i = 1, m, j = 1, n
(10)
Fig. 5. Improvement of manipulator positioning accuracy after calibration due to the enhanced partial pose technique: (a) Conventional technique and (b) proposed
technique.
158
where the vectors pij are incorporated in the fourth column of the
homogenous transformation matrix T ij , the matrix Tbase denes the
j
robot base location, the matrices T tool
, j = 1, n describe the locations of the reference points that are observed by the measurement system (see Fig. 2). Here, the matrix function Trobot ( qi , )
describes the manipulator geometry and depends on the current
values of the actuated coordinates qi and the parameters to be
estimated. Taking into account that any homogeneous transformation matrix T ba can be split into the rotational R ba and translational
pba components and presented as
j
pij = pbase + R baseprobot (qi , ) + R baseR robot (qi , )ptool
;
i = 1, m, j = 1, n.
(12)
This allows us to obtain 3mn scalar equations for the calibration

purposes, where n 3 and m is high enough to ensure identiability of the desired parameters.
Applying the least-square method, the corresponding optimization problem can be presented as
(16)
that can also be rewritten in a matrix form as
pirobot
p base
i

T
i
+ I ~probot R robot base
uj
tool
(17)
and
j
j
utool
= R base ptool
(18)
j
Here the vectors pbase , base and utool
, j = 1, n are treated as
unknowns.
Applying to the linear system (17) the linear least-square
technique, the desired vectors dening the base and tool transformation parameters can be expressed as follows
m
1 m
p ; ; u1 ; ... un = A j T A j A j T p
tool
base base tool
i
i
i
i
i = 1
i = 1
(19)
where
j
pij pbase R baseprobot (qi , ) R baseR robot (qi , )ptool
2
i=1 j=1
j
= pbase + pirobot pirobot [~ base ] + R irobot utool
(11)
the vector of the reference point positions pij , j = 1, n (that are

measured in the calibration experiments) can be expressed in the
following form
(15)
This leads to the following simplied expression of Eq. (12)
pij
pij
b b
R p
T ba = a a ,
0 1
0
z
y
0 x
[~] = z
0
y x
min
j , }
{p base,R base,ptool
(13)
j
where the vectors/matrices pbase , R base , ptool
, j = 1, n and are
treated as unknowns.
The main difculty with this optimization problem is that some
of the unknowns are included in the objective function in highly
non-linear way. So, to solve this problem, numerical optimization
technique is required. However in practice, the deviations in the
model parameters are relatively small, which allows us to linearize
the manipulator geometric model (12). This leads to a linear leastsquare problem, whose solution can be obtained straightforwardly
with the matrix pseudo-inverse. However, to simplify computations, here it is proposed to apply the linearization technique sequentially and separately with respect to two different subsets of
the model parameters (corresponding to the base/tool transformations and the manipulator geometry). Consequently, the identication procedure is split into two steps. In the frame of this
approach, the rst step deals with the estimation of pbase , R base ,
j
, which are related to the base and tool transformations (asptool
suming that the manipulator parameters are known). The second
step focuses on the estimation of under the assumption that the
base and tool components have been already identied. In order to
ensure that the desired identication accuracy can be achieved,
these two steps are repeated iteratively.
Step 1. For the rst step, taking into account that the errors in
the base orientation are relatively small, the matrix R base is presented in the following form
R base = [~ base ]+I
(14)
where I is a 3 3 identity matrix, vector base includes the deviations in the base orientation angles, and the operator [~] transforms the vector = (x , y , z )T into the skew symmetric matrix
as
j
Ai = I
...
~pi
T i
robot R robot
~pi
T
robot
...
0
...
R irobot
...
~pi
T
robot
...
0
...
...
... R irobot
...
(20)
and the residuals are integrated in a single vector

pi = (p1i , ... , pin )T . Finally, the variables dening the location of
the reference points are computed using Eq. (18) as
j
j
. This allows us to nd the homogeneous transforptool
= R Tbaseutool
j
mation matrices Tbase and T tool
that are contained in Eq. (10).
Step 2. On this step, the manipulator geometric parameters
are estimated. For this purpose, the principal system (10) is linearized and rewritten in the form
pij = Jj i(p)
pij
pij
(21)
i
probot
where
is the residual vector corresponding to the
=
jth reference point for the ith manipulator conguration, is

the vector of geometric errors, the matrix J j i is the identication
Jacobian computed for the conguration qi with respect to the
reference point j . Applying to this system the least-square technique, the desired vectors of geometric errors can be obtained
as
m n
1 m n
= Jj i(p) T Jj i(p) Jj i(p) T pij

i=1 j=1
i=1 j=1
(22)
It should be noted that, to achieve the desired accuracy for the

original non-linear problem (13), the steps 1 and 2 should be repeated iteratively.
Another particularity may arise here is related to the property
of measurement noise. In the above expressions, it was explicitly
assumed that the measurement errors are similar for all directions.
However, for some measurement systems, the errors in the longitudinal and transversal directions may essentially differ. In this
case, the Eqs. (19) and (22) should be slightly modied by
reasonable to investigate the rst approach that deals with optimization of the measurement congurations for limited number of
experiments.
In more general case when the measurement errors differ from
direction to direction, the expectation E i Ti can be expressed as
including weighting coefcients
p ; ; u1 ; ... un
tool
base base tool
m
1 m
T
T
= A ij W ij2A ij A ij W ij2pi
i=1
i=1
(23)
and
m n
1 m n
j (p) T
j2 j (p)
= Ji W i Ji Jj i(p) T W ij2pij
i=1 j=1
i=1 j=1
(24)
5.2. Inuence of the measurement errors on the identication

accuracy
Under the assumption that measurement noise has the most
signicant impact on the robot positioning accuracy and the other
sources of error are negligible, the basic equation of calibration (21) should integrate the measurement errors and is expressed as
pij = Jj i(p) + ij ; i = 1, m, j = 1, n
(xij ,
yij ,
ij kjT
) = diag (
xij2,
yij2,
zij2
), if i = k
(28)
(see our previous study on this issue presented in [54]). So, the
covariance matrix dening the calibration accuracy can be rewritten in the following form
where the weighting coefcients matrix W ij is computed using a

technique proposed in our previous work [54].
Hence, the above presented identication algorithm is able to
provide the estimation of the manipulator geometric parameters
as well as the matrices of the base and tool transformations.
However, the obtained identication results usually include some
dispersion due to measurement errors. So, in order to achieve
desired identication accuracy, the inuence of these errors
should be evaluated and reduced as much as possible, which will
be in the focus of the following subsection.
ij
159
(25)
m n
1
^) = 2 J j (p) T W j2J j (p)
cov(
i
i=1 j=1
(29)
where the weighting coefcient matrix W ij can be computed as in

Eq. (24). It is clear that here the optimization of measurement
congurations is also promising. However, in practice for the
design of calibration experiments, the measurement errors are
assumed to follow the i.i.d assumption.
In the literature, the problem of optimal pose selection for calibration experiments have been studied in a number of works
[36,3944], where several scalar criteria were proposed to dened
this type of optimality in formal way. The main drawback of these
approaches is that the relevant optimization objectives are not
directly related to the manipulator positioning accuracy and its
targeted industrial application (they usually focus on the parameter identication accuracy). Hence, in order to achieve simultaneously high accuracy both for the manipulator parameters
and for the end-effector position (or to nd reasonable trade-off),
it is required to revise the existing techniques and to dene a
proper objective for measurement pose selection, taking into account the specicities of the application area studied in this work.
This issue is in the focus of the next section.
zij )T
where the vectors

denote the additive random
=
errors, which are usually assumed to be unbiased and i.i.d. with
the standard deviation . Then, using Eq. (22), the estimates of the
desired parameters can be presented as
m n
1 m n
^ = + J j (p) T J j (p) J j (p) T j
i=1 j=1
i=1 j=1
(26)
where the second term describes the stochastic component. As

follows from this expression, the considered identication algorithm provides the unbiased estimate of the desired parameters,
^) = where, E () denotes the mathematical expectation
i.e., E (
of the random value. Taking into account the statistical properties
of the measurement errors (which are assumed to be similar for all
reference points, all manipulator congurations and all directions,
in accordance with expression E i Ti = 2I), the desired covar-
iance matrix of , which denes the identication accuracy, can

be computed as
m n
1
^) = 2 J j (p) T J j (p)
cov(
i
i
i=1 j=1
(27)
Hence, the impact of the measurement errors on the identied

values of the geometric parameters is dened by the matrix sum
n
j (p) T j (p)
m
J i that in literature is also called the information
i=1 j=1 J i
matrix. It is clear that to achieve the best accuracy, the elements of
covariance matrix (27) should be as small as possible. This requirement can be satised by proper selection of the experiment
input data (i.e., the measurement congurations {qi , i = 1, m} ) as
well as by increasing the number of experiments m. Since increasing of the measurements is rather time consuming, it is
6. Optimal selection of measurement congurations

This section proposes a new approach for calibration experiments design that has two distinct features: (i) optimization based
on a new industrial-oriented performance measure that evaluates
the manipulator positioning accuracy after calibration; (ii) utilization of experimental data obtained by means of the enhanced
partial pose measurement method.
6.1. Test-pose based approach for calibration experiments design
In robot calibration, the desired manipulator parameters are
estimated using experimental data corrupted by the measurement
noise. For this reason, the parameters estimates are not equal to
the true values, they vary from one set of experiments to another
and can be treated as random ones. As follows from previous
section, relevant identication algorithms provide unbiased estimates (i.e. their expectation is equal to the true values) but their
dispersion essentially depends on the set of measurement congurations that provides the experimental data for the identication. Hence, it is reasonable to select the measurement congurations in the best way, in order to ensure the lowest impact of
the measurement errors on the parameter estimates. In the literature, this problem is known as the calibration experiments
design. However, existing approaches focus on the accuracy of the
parameter estimation (dened by the relevant covariance matrix (27)), while the considered industrial application motivates us
to focus on the manipulator positioning accuracy after calibration.
In more details, the notion of manipulator positioning accuracy
after calibration is illustrated in Fig. 6. It is assumed here that the
160
pd , as shown in Fig. 6.
The dispersion of these points can be evaluated by the variance
E pk T pk which in accordance with the above denition is equal
to the square of the performance measure 0. This yields the following expression
02 = E T J(p0) T J(p0)
(32)
which can be rewritten

pT p trace( p pT ) as
02 = trace J(p0) E ( T )J(p0) T
Fig. 6. Dispersion of the manipulator positioning errors after calibration and performance measure for selection of measurement congurations (for given single
target point).
desired end-effector position is pd , but without calibration, the

end-effector is located at the point p0 = g (q0 , ), which can be
computed using the nominal geometric model. Here, the joint
coordinate vector q0 is obtained from equation pd = g (q0 , 0 ) via
the inverse kinematics. Using calibration, for each set of experi^ k that
mental data, it is possible to nd the parameter estimates
allow us to compensate partially the positioning errors by computing another joint coordinate vector qk from the equation
^ k ) and to relocate the end-effector at the point
pd = g (qk ,
pk = g (qk , ), which is closer to the desired position pd . Evaluating

the distribution of Cartersian coordinates of points pk , it should be
mentioned that those points are concentrated around the desired
position pd in such way that:
E (pk ) = pd
(30)
So, the target position can be treated as the center. To evaluate

their dispersion with respect to the desired position, relevant
distances k = dist(pk , pd ) can be used. This leads to the following
statistical performance measure
0 =
E (pk pd )T (p k p d )
(31)
which is the root-mean-square distance between the target position and the end-effector position after calibration. This indicator
is used below to describe the geometric errors compensation efciency. It is clear that the performance measure 0 is directly
related to the manipulator accuracy in an engineering viewpoint.
It is clear that the positioning error scattering and relevant
performance measure 0 highly depend on the target point position and varies throughout the workspace. In the frame of this
work, it is assumed that the manipulator accuracy can be evaluated for so-called test-pose that is specied in the relevant
technological process. This idea allows us to use the above mentioned performance measure 0 as an objective in the calibration
experiments design.
Using the adopted notations and assuming that the manipulator geometric model is linearized, the distance k can be computed as the Euclidean norm of the vector pk = J(p0) k , where the
subscript0 in the identication Jacobian J(p0) is related to the test
^ k is the difference between the estimated
pose q0 and k =
and true values of the robot geometric parameters respectively.
Further, taking into account expression (26) and the assumptions
concerning the measurement errors that are treated as unbiased
and i.i.d. random variables, it can be easily proved that the expectation E ( pk ) = 0. Therefore, the points pk that the end-effector
attains after compensation are located around the desired position
using
the
identity
equation
(33)
Further, by applying Eq. (27) and considering that the term

E ( T ) is the covariance matrix of the geometrical error esti^), the desired expression can be premates, i.e., E ( T ) = cov(
sented in the nal form as
02
2trace J(p0)
i = 1
j=1
1
(p) T
J 0
(34)
02
can be treated as the
J j i(p) T J j i(p)
As follows from this expression,
^), where the weighting

weighted trace of the covariance matrix cov(
coefcients are computed using the test-pose joint coordinates q0 .
Hence, the proposed performance measure has obvious advantage
compared to the existing ones [40], which operate with pure
trace of this matrix and involve straightforward summing of its
diagonal elements (which may be of different units). Based on this
performance measure, the calibration experiments design can be
reduced to the following optimization problem
m n
trace J(p0) J j i(p) T J j i(p) J(p0) T min

{q i ...q m}
i = 1 j = 1
(35)
whose solution gives a set of the desired measurement congurations {q1, .. . , qm } .

Hence, in the frame of the proposed approach, the calibration
quality (evaluated via the error compensation accuracy 0 ) is
completely dened by the set of Jacobian matrices J(p1) , ... , J(pm)
that depend on the manipulator congurations {q1, .. . , qm } , while

the Jacobian matrix J(p0) corresponding the test-pose q0 denes the
weighting coefcients. It is worth mentioning that test-pose based
approach can be also extended for calibration of the manipulator
elasto-static parameters (see [55], for more details). The advantages of the proposed approach will be illustrated in the following subsections.
6.2. Comparison analysis of the proposed and conventional
approaches
Let us illustrate the advantages of the test-pose based approach
by an example of the geometrical calibration of a two-link planar
manipulator. It is assumed that the nominal link lengths {l1, l2 }
differ from the real ones, and these deviations {l1, l2 } should be
identied by means of calibration. In this case, the manipulator
end-effector position can be expressed as
px = (l1 + l1)cos q1 + (l2 + l2 )cos(q1 + q2 )

py = (l1 + l1)sin q1 + (l2 + l2 )sin(q1 + q2 )
(36)
where px and py dene the end-effector position, q1, q2 are the joint
coordinates that dene the manipulator conguration. It can be
proved that in this case the parameter covariance matrix does not
depend on the angles q1i and is expressed as
2
cov() =
2
m
2
m ( i = 1 cos q2i )
cos q2i
i=1
(37)
m
m
m
i=1
cos q2i
where the vector = (l1, l2 ) denotes the parameters deviations to be identied, m is the number of experiments and is the
standard deviation. of the measurement noise.
For comparison purposes, the plans of experiments were obtained using three different strategies:
(1) the measurement congurations were generated randomly;
(2) the measurement congurations were obtained using the
conventional approach based on D-optimality principle;
(3) the measurement congurations were obtained using the
proposed test-pose based approach (see Section 6.1).
For the rst approach (i), the measurement congurations were
found in a trivial way, using a uniform random number generator
scaled within the joint limits. For the conventional approach
(ii), where the D-optimality principle was used (that has been
proved to be efcient in many applications), the performance
measure is equal of the covariance matrix determinant (37), which
yields
det(cov()) =
m2
m
i = 1
cos q2i )
(38)
As follows from this expression, this criterion requires mini2

mization of the term ( m
i=1 cos q2i ) . So, the determinant minimum value is equal to 2/m2 and it is reached when
m
cos q2i = 0, i = 1, ... , m
(39)
i=1
It should be mentioned that this optimality condition also satises the A- and G-optimality principles. More details concerning
the calibration experiment planning using the above conditions
can be found in [56].
For the proposed approach (iii), it is assumed that the calibration quality is evaluated in the predened manipulator test conguration (q10 , q20 ). In this case, the performance measure 02 (34)
can be computed as
02 = 2 2
m cos q20 im=1 cos q2i

2
m2 (im=1 cos q2i )
(40)
As follows from relevant analysis, the minimum value of 02 is

equal to
02 min =
cos2 q20
2
m 1 sin q20
(41)
and it is achieved when the measurement congurations satisfy

the equation
m
i=1
cos q2i = m
1 sin q20
cos q20
(42)
which essentially differ from (39). It should be mentioned that

general solution of Eq. (42) for m congurations can be replaced by
the decomposition of the whole conguration set by the subsets of
2 and 3 congurations (while providing the same identication
accuracy). This essentially reduces computational complexity and
allows user to reduce number of different measurement congurations without loss of accuracy.
161
Table 2
Accuracy comparison of the proposed and conventional approaches
Test-pose (q20 ), [deg.]
30
60
90
120
150
180
02 /c2
0.5
0.75
0.83
0.83
0.75
0.5
Accuracy improvement (%)
41
15
10
10
15
41
Using the above presented expressions for the robot accuracy

after calibration, the proposed and conventional approaches can
be compared analytically and numerically. In particular, for the test
pose ( , q20 ), the conventional approach (ii) ensures the positioning accuracy after compensation c2 = 2 2/m, while for the proposed approach (iii), similar performance measure is equal to
02 min = 2/mcos2 q20 /(1 sin q20 ). Corresponding values are
compared in Table 2, which proves that using the proposed approach for the calibration experiment design allows us to improve
the positioning accuracy up to 41%.
To illustrate advantages of the proposed approach, Fig. 7 presents simulation results for manipulator positioning errors after
compensation corresponding to three different sets of measurement congurations employed in calibration. It is also assumed
that the manipulator parameters are l1 = 1 m, l2 = 0.8 m ; the
number of measurement congurations m = 2; the test conguration is dened by the vector q0 = ( 45 , 20), and the s.t.d. of
the measurement errors is = 1 mm . For comparison purposes,
the following plans of experiments (measurement congurations)
have been considered:
(1) Random plan: q1 = (0 , 10) and q2 = (0 , 10) , which has
been generated randomly;
(2) Conventional plan: q1 = (0 , 90) and q2 = (0 , 90) , which
satises D-optimality principle [34];
(3) Proposed plan: q1 = (0 , 46) and q2 = (0 , 46), which satises the test-pose based approach.
To obtained reliable statistics, the calibration experiments have
been repeated 100 times. Corresponding results presented in Fig. 7
show that the proposed approach allows us to increase accuracy of
the end-effector position on average by 18% comparing to the
D-optimal plan and by 48% comparing to the randomly generated
plan.
Hence, this simple example conrms that the proposed performance measure is attractive for practicing engineers and allows
us to avoid the multi-objective optimization problem that arises
while minimizing all elements of the covariance matrix (27) simultaneously. In addition, using this approach, it is possible to nd
a balance between the accuracy of different geometrical parameters whose inuence on the nal robot accuracy is unequal.
Another example conrming this conclusion is presented in our
previous work [57].
It is worth mentioning that the proposed approach allows essential improvement of the calibration efciency and to achieve
the best manipulator positioning accuracy for the user-dened test
congurations related to the manufacturing task (in contrast to the
conventional approaches that are targeted at the best parameter
identication accuracy). However, for typical industrial robots
whose model includes very high number of parameters, relevant
optimization becomes extremely time consuming. For this reason,
the next subsection focuses on simplication of numerical routines employed in the selection of optimal measurement
congurations.
162
Fig. 7. Dispersion of manipulator positioning errors after calibration for different plans of experiments: (a) Random plan, (b) conventional plan and (c) proposed plan.
6.3. Simplication of the optimal pose selection procedure

It is clear that analytical solutions of relevant optimization
problems (35) can hardly be obtained (for example, when the
number of parameters to be identied is very high, the analytical
computations of the matrix inversion in these expressions are
hardly possible). So, applying a numerical optimization technique
is the only reasonable way, but the convergence rate, the total
computational time and ability to attain the global minimum become key issues. For this reason, several conventional optimization techniques were examined. In order to improve their efciency, two techniques that are adapted to the test-pose based
approach have been proposed: (i) application of parallel and hybrid computations; (ii) generation of quasi-optimal solutions using
lower-dimensional calibration plans.
In order to obtain the global optimum for the considered problem, it is required numerous repetitions of the optimization with
different starting points. As follows from our experiences, even
using thousands of them may be not enough for nding the global
optimum but the required computational time could overcome
hundreds of hours. So, it is reasonable to apply parallel computing
technique to speed up the design process and to take advantage of
multi-core architecture in modern computers. Relevant computations in this work were carried out on a workstation with 12 cores,
which allowed us to decrease the computational time by the factor
of 1012. However, it is not enough yet to solve the problem of real
industrial size, where several dozen of parameters should be
identied. To overcome this difculty, it has been proposed a hybrid approach that combines advantages of the genetic algorithm
and the gradient search. The idea behind this technique is to
modify the starting point selection strategy for the gradient search
in order to improve the algorithm efciency. To ensure better
convergence to the global minimum, it has been proposed to use
the best half of nal solutions obtained from GA as the starting
points for gradient search. This hybrid approach has been proved
to be quite efcient in terms of computational time (improved by a
factor of 5, in addition to parallel computing) and allows us to
avoid convergence to the local minima.
On the other hand, as follows from our experiences, the diversity of the measurement poses does not contribute signicantly
to the accuracy improvement if m (the total number of measurements) is high enough. This allows us to propose an alternative
technique, which uses the same measurement congurations several times (allowing to simplify and speed up the measurements).
This approach can be also referred to as the reduction of problem
dimension.
To explain the proposed approach in more details, let us assume that the problem of the optimal pose selection has been
solved for m different congurations and the obtained calibration

plan ensures the positioning accuracy 0m . Using these notations,
let us evaluate the calibration accuracy for two alternative strategies that employ total number of experiments k m:
Strategy #1 (conventional): the measurement congurations
are found from the full-scale optimization of size k m.
Strategy #2 (proposed): the measurement congurations are
obtained by simple repetition of the congurations got from the
low-dimensional optimization problem of size m (i.e. at each
conguration, the measurements are repeated k times).
It is clear that the calibration accuracy 0km for the strategy #1 is
better than the accuracy corresponding to the strategy #2 that can
be expressed as 0m / k . However, as follows from our study, this
difference is quite small if the total number of measurements is
high enough, while the number of different congurations m is
larger than 3. This allows us to essentially reduce the size of the
optimization problem employed in the optimal selection of measurement poses without signicant impact on the positioning
accuracy.
To demonstrate the validity of the proposed approach, a
benchmark example that deals with geometric calibration of a
6-dof manipulator has been solved using strategies #1 and #2,
assuming that the total number of measurements is equal to 12
(i.e. using different factorizations such as 12 1, 6 2, 4 3,
3 4 ). Relevant results are presented in Table 3, where the rst
four lines give the accuracy 0 and the last line shows corresponding computational time. It is noteworthy that the factorization 12 1, where all measurement poses are different, is only 6%
better comparing to the factorization 3 4 where measurements
are repeated 4 times in 3 different congurations. At the same
time, the factorizations 6 2 and 4 3 give almost the same results as the factorization 12 1. On the other hand, the computational time of the optimal pose generation for m = 3 is much
lower than for m = 12. Hence, as follows from these results, repeating experiments with optimal plans obtained for the lower
number of congurations provides almost the same performance
as full-dimensional optimal plan. Obviously, this reduction of the
measurement pose number is very attractive from an engineering
viewpoint. This technique will be used in the application example
presented in the following section.
7. Experimental results: geometric calibration of KUKA KR-270

To conrm the applicability of the proposed calibration techniques and demonstrate their benets from engineering point of
view, this section presents the experimental procedure, the identication results as well as the accuracy analysis for geometric
Table 3
Comparison of the optimal and quasi-optimal solutions for measurement congurations in calibration experiments, evaluated via 0min , [mm] (case of a 6-dof
manipulator, repetitions of measurements k times for m different congurations)

Total number of
measurements
Number of different congurations
m=3
km = 3
km = 12
Computational time
m=6

m = 12
0.0637
(3 1)
0.0521
(4 1)
km = 4
km = 6
m=4
0.0450
(3 2)
0.0319
(3 4)
38 min
0.0301
(4 3)
45 min
0.0426
(6 1)
0.0301
(6 2)
56 min

0.0301
(12 1)
1.6 h
calibration of the industrial KUKA KR-270 robot.

7.1. Experimental environment and measurement setup
The manufacturing cell where the examined robot has been
installed is presented in Fig. 8. The work-cell includes a 6-dof industrial KUKA KR-270 robot with six revolute joints, a machining
table, a vertical frame for mounting the pieces. It should be noted
that for geometric calibration, the above mentioned equipment
(that can be also treated as the obstacles) cause some limits for
placement of the external measurement device. Taking into account particularities of the technological process considered in
this work, the manipulator test-pose (congurationq0 ) has been
dened in the location where the best robot positioning accuracy
should be achieved: q0 = (76.7, 56.9, 89.3, 45.1, 76, 57.2) . It
is worth mentioning that similar conguration has been used in
previous works [58,59] to evaluate the quality of the robot-based
machining.
To identify the desired geometric parameters, the manufacturing cell is equipped with some additional measuring devices
that provide us with Cartesian coordinates of the references points
for each manipulator conguration. Besides, the manipulator joint
angles required for the identication procedure are obtained from
the robot control system. So, entire experimental setup includes
the following units:
6-dof KUKA KR-270 robotic manipulator whose geometric

parameters should be identied (repeatability of this robot is
163
60 mm [60], details concerning its kinematics are presented in

Section 3);
Robot control system KR-C2, which is used for changing the
manipulator congurations and measuring the corresponding
joint angles with a precision equal to 70.0001;
Special measurement tool with three reference points located
on the circle of radius 104 mm, this tool is attached to the
manipulator mounting ange;
Laser tracker Leica AT-901 that is used to measure the Cartesian
coordinates of the reference point with a precision of 10 mm
[61];
Laser tracker reector that is sequentially attached to the reference points (with precision about 1 mm), it allows the measurement device to estimate the distances and compute the
required Cartesian coordinates;
The experimental setup for manipulator geometric calibration

is shown in Fig. 9. It is worth mentioning that the calibration experiments are carried out in a limited area (smaller than the robot
entire workspace) caused by the work-cell size limitation and
some obstacles. For this reason, some of the manipulator congurations cannot be reached during the experiments (As a consequence, they are not included in the optimal plan).
7.2. Optimal measurement pose selection
While selecting the minimum number of measurement congurations, it is necessary to keep in mind that each manipulator
pose produces 6 independent equations only that are used for
identication. On the other hand, the set of geometric parameters
to be identied includes 33 unknowns:
(1) 18 principal parameters of the KUKA KR-270 robot;
(2) 6 parameters describing the laser tracker location with respect
to the robot base frame (both position and orientation);
(3) 9 parameters describing locations of the end-effector reference
points with respect to the manipulator mounting ange (positions only for three points).
Therefore, at least six different measurement congurations are
required to ensure non-singularity of the identication Jacobian
and ability to estimate the desired values. For this reason, relevant
optimization problem aiming at determining optimal measurement poses has been solved for the conguration number m = 6.
To take into account the manipulator joint limits and the workcell constraints, the optimization problem for measurement
Fig. 8. The experimental work-cell environment: (a) general view; (b) typical machining conguration (test-pose).
164
Fig. 9. Experimental setup for manipulator geometric calibration.
Table 4
The joint limits of robot KUKA KR-270
Table 6
Comparison of calibration plans with different diversity of measurement
congurations.
q1
q2
q3
q4
q5
q6
qmin , [deg.]
180
145
110
180
125
180
qmax ,
180
155
180
125
180
Calibration plans
[deg.]
Table 5
The work-cell space boundaries with respect to the robot base frame
px
py
pz
pmin , [mm]
1400
3000
300
pmax , [mm]
1800
2200
3500
Robot accuracy 0 , Computational time

[mm]
(i)
{Sol. #1 3}
{Sol. #2 3}
{Sol. #3 3}
(ii) {Sol. #1 2, Sol. #2}
{Sol. #1 2, Sol. #3}
{Sol. #1, Sol. #2 2}
{Sol. #2 2, Sol. #3}
{Sol. #1, Sol. #3 2}
{Sol. #2, Sol. #3 2}
(iii) {Sol. #1, Sol. #2, Sol. #3}
Random congurations (for
comparison)
7.85
7.84
7.83
7.84
7.84
7.83
7.83
7.83
7.83
7.83
17.33
56 min
56 min
56 min
1.9 h
1.9 h
1.9 h
1.9 h
1.9 h
1.9 h
2.8 h
0.05 s
congurations selection (see Eq.(35)) should be solved subject to

qmin qi qmax and pmin g (qi ) pmax . In particular, {qi i = 1, 6}
denote the measurement congurations, the function g (qi ) describes the manipulator geometric model and returns the endeffector position coordinates of the current conguration. These
constraints take into account the manipulator joint limits
[qmin , qmax ] and the work-cell boundaries [pmin , pmax ], whose values are given in Table 4 and 0, respectively. Table 5
This optimization problem has been solved using the MATLAB
software with the built-in optimization functions ga and
fmincon, which are required for the proposed hybrid approach
that employs the genetic algorithm and the gradient search. Corresponding solution minimizes the objective function 0 (the
Table 7
Identication results for manipulator tool transformations
Reference point #1 (P1) Reference point #2 (P2) Reference point #3 (P3)
Value, [mm] CI
Value, [mm] CI
Value,
[mm]
CI
px
277.23
70.05
276.49
7 0.05
278.44
70.05
py
46.53
70.04
48.25
7 0.04
103.73
70.05
pz
93.87
70.04
94.05
7 0.05
2.17
70.05
Table 8
Identication results for manipulator geometric parameters
Parameter Unit
Value
Condence interval
Estimated using covariance matrix
Estimated using Gibbs

sampling
px1 d2
[mm]
0.353
7 0.086
7 0.102
py1
[mm]
0.426
7 0.272
7 0.421
x1
[deg.]
0.015
7 0.005
7 0.005
q2
[deg.]
0.007
7 0.005
7 0.004
7 0.060
px2 d3
[mm]
0.458
7 0.082
x2
[deg.]
0.022
7 0.014
7 0.022
z2
[deg.]
0.023
7 0.005
7 0.005
q3
[deg.]
0.023
7 0.019
7 0.013
px3 d4
[mm]
0.214
7 0.089
7 0.093
pz3 d5
[mm]
0.508
7 0.363
7 0.259
z3
[deg.]
0.011
7 0.017
7 0.022
q4
[deg.]
0.001
7 0.008
7 0.009
py4
[mm]
0.167
7 0.113
7 0.044
pz4
[mm]
0.018
7 0.073
7 0.044
z4
[deg.]
0.025
7 0.015
7 0.010
q5
[deg.]
0.011
7 0.027
7 0.009
pz5
[mm]
0.016
7 0.104
7 0.041
z5
[deg.]
0.008
7 0.018
7 0.007
Table 9
Evaluation of the manipulator accuracy improvement based on residual analysis
Criterion
Coordinate-based
residuals, [mm]
Distance-based residuals, [mm]
max
RMS
max
RMS
Before
calibration
After
calibration
Improvement
factor
1.25
0.54
1.31
0.94
0.32
0.10
0.39
0.17
4.0
5.3
3.5
5.5
proposed performance measure), which describes the manipulator

positioning accuracy after calibration. The measurement noise
parameter has been taken from the technical specication of the
laser tracker and is equal to 10 mm. It should be mentioned that, in
order to reduce the computational efforts and, to pay more attention to the parameters that can be tuned in the robot controller,
only nine the most essential geometric parameters were considered
while computing the Jacobian matrices J(p0) and J(pi ) . They include the
link lengths {d2, ... , d5 } whose nominal values are known and the
joint offsets {q1, ... , q5 } that are nominally equal to zero.
In order to nd a solution as close as possible to the global
minimum, the optimization problem has been solved several times
with different starting points. Nevertheless, three different
165
solutions have been obtained that ensure almost the same value of
the considered performance measure 0 (13.6mm). Corresponding
solutions (measurement congurations) are presented in [45]. For
comparison purposes, these solutions have been evaluated both
separately and in different combinations, assuming that the
measurements are performed 18 times in the following way: (i)
repeating three times the measurements in congurations from a
single set; (ii) using twice congurations from one set and only
once from the second set; (iii) using all congurations from three
sets simultaneously (but only once). Corresponding values of 0
are presented in Table 6. As follows from this table, the diversity of
manipulator congurations has almost negligible contribution to
the improvement of robot accuracy (it is about 7.85 mm, the difference is less than 0.2%). This conrms the results from Subsection 6.3, which claims that using simple repetition of the optimal
plan with lower number of measurement congurations essentially reduces the experimental complexity while the same calibration accuracy can be achieved.
7.3. Identication of geometric parameters
The obtained measurement congurations have been used for
the calibration experiments for KUKA KR-270 industrial robot. It is
worth mentioning that each manipulator conguration provides
27 values of the position coordinates. These coordinates have been
obtained using two different locations of the laser tracker (see more
details in [45]). However, at certain congurations, some of the
reference points were not visible for both laser tracker locations.
This problem can be solved by increasing the number of laser
tracker locations, but in practice such solution is limited by the
experimental time as well as the work-cell constraints. On the
other hand, since the calibration experiment employs two laser
tracker placements, 6 additional parameters describing the second
laser tracker location should be also identied. In total, the system
of identication equations contains 432 expressions that can be
used to identify the whole set of 39 geometric parameters. To
achieve the highest identication accuracy, here it is proposed to
use all measurements corresponding to 18 manipulator congurations simultaneously for calibration of the geometric
parameters.
Using the obtained measurement data, the two-step identication procedure has been applied (see Section 5). On the rst
step, the base and tool transformations have been computed,
corresponding results are presented in Table 7. On the second step,
these transformations have been used for the identication of the
manipulator geometric parameters, which are presented in Table 8. It should be mentioned that in order to increase the identication accuracy, this two-step procedure has been repeated
iteratively (280 iterations, computing time was less than two
Fig. 10. Histograms of residual distribution along X-, Y-, and Z-directions after geometric calibration: (a) X-direction, (b) Y-direction and Z-direction.
166
Fig. 11. Residual distribution after geometric calibration for different measurement congurations.
minutes). As follows from the results, the desired geometric

parameters have been identied with high accuracy, which has
been evaluated using two different techniques (based on the statistical properties extracted from the covariance matrix and using
the Gibbs sampling).
The results presented in Table 8 include 18 parameters, some of
which cannot be modied in the robot control software. So, it is
useful to examine the effect of reducing the number of these
parameters by setting them to their nominal values. Relevant
analysis shows that the manipulator end-effector positioning error
impact because of such a simplication essentially differs from one
parameter to another, and they can be split into the following
groups:
Parameters

{px1, p y1, px2 , px3 , pz 3 , p y4 , q2, q3, z2, z 3, z 4 } ,

whose neglecting leads to the loss of accuracy from 0.10 mm to
1.03 mm;
Parameters {pz 4 , pz5 , q5, x1, x2, z5 } , whose neglecting leads
to the loss of accuracy from 0.02 mm to 0.09 mm;
Parameter q4 , whose neglecting leads to the loss of accuracy
is about 4 mm.
Comparing to the machining accuracy required for the considered milling process (0.050.25 mm), the above listed positioning error impacts are not negligible for the most of the geometric parameters. So, their deviations should be compensated
either in the geometric model embedded in the robot controller or
at the step of generation of the machining trajectory.
For comparison purposes, the manipulator accuracy improvement due to calibration has been studied based on the residual
analysis before and after calibration (computed using the nominal
and identied values of geometric parameters respectively). Here,
two types of residuals have been examined, the coordinate-based
and distance-based ones. Corresponding results are presented in
Table 9, which includes the maximum and root mean square
(RMS) values of the relevant residuals. As follows from the results,
both types of the residuals have been essentially reduced after
calibration. In particular, the maximum values have been reduced
by a factor of 4 and 3.5, while the RMS values have been decreased
by a factor of 5.3 and 5.5, respectively.
Hence, the obtained results allow us to improve essentially the
manipulator accuracy for the measurement congurations that
were used in the identication. So, it is reasonable to expect that
using the geometric model, which integrates the identied parameters, the desired positioning accuracy for the given test conguration can be also achieved. A more detailed analysis concerning the parameter identication accuracy and its impact on
the robot positioning accuracy are discussed in the next

subsection.
7.4. Analysis of the identication results
In order to evaluate the calibration results, let us rst analyze
the residuals computed from the identication equations for each
coordinate separately. Their histograms are shown in Fig. 10 and
corresponding distributions for each conguration are presented
in Fig. 11. As follows from the analysis, the residuals tend to follow
the normal probability distributions with zero mean and almost
the same parameter , which is equal to 0.10 mm, 0.09 mm, and
0.11 mm for X-, Y-, and Z- direction, respectively. The latter justies
the utilization of ordinary least square technique (with equal
weights) for the parameter identication and allows us to conclude that the measurement noise parameter in our experiment
is about 0.1 mm.
It is worth mentioning that the noise parameter estimated
from the residual analysis is essentially higher than the precision
of the laser tracker measurement system, which is dened in the
technical specications as 0.01 mm. This difference may be due to
the limitations of the geometric model, which does not take into
account a number of essential features such as the elastostatic
deformations due to gravity forces, the friction/backlash in joints
and other factors that affect the robot repeatability (70.06 mm, as
specied in the data sheets). Nevertheless, the geometric calibration ensures essential improvement of the robot accuracy in the
unloaded mode. Relevant computations show that for the considered test-pose, it is possible to achieve a positioning accuracy of
about 0.04 mm that is acceptable for the considered technological
process. On the other hand, this issue motivates further research
devoted to modeling of non-geometric factors and estimation of
relevant parameters.
8. Conclusions
This paper presents a new approach for calibration experiments design for serial industrial robots. This approach employs a
new industry-oriented performance measure, which evaluates the
quality of calibration plan via the manipulator positioning accuracy after geometric error compensation, and considers the industrial requirements associated with the prescribed manufacturing task. It is proved that the proposed performance measure
can be presented as the weighted trace of the relevant covariance
matrix, where the weighting coefcients are dened by the corresponding test-pose. Such an approach allows us to nd the
optimal measurement congurations for calibration experiments

and to improve essentially the robot positioning accuracy for a
desired manipulator test-pose.
Dedicated algorithm for geometric parameter identication is
based on an enhanced partial pose measurement method, which
uses only direct position measurements from an external device
for several end-effector reference points. It allows the user to increase essentially the parameter identication accuracy and to
avoid additional computations of the end-effector orientation
components, which may cause non-homogeneity in relevant
identication equations.
The obtained theoretical results have been validated via experimental study that deals with geometric calibration of a KUKA KR-270
industrial robot. The manipulator geometric parameters have been
identied with accuracy equal to 0.15 mm and 0.01 for linear and
angular ones respectively (in average). These results allowed us to
achieve a manipulator positioning accuracy equals to 0.17 mm, which
is 5.5 times better compared to the non-calibrated robot.
Acknowledgments
The authors would like to acknowledge the nancial support of the
French Agence Nationale de la Recherche (Project ANR-2010-SEGI003-02-COROUSSO), France and FEDER ROBOTEX project, France.
References
[1] S. Aguado, D. Samper, J. Santolaria, J.J. Aguilar, Identication strategy of error
parameter in volumetric error compensation of machine tool based on laser
tracker measurements, Int. J. Mach. Tools Manuf. 53 (2012) 160169.
[2] G. Du, P. Zhang, D. Li, Online robot calibration based on hybrid sensors using
Kalman lters, Robot. Comput.-Integr. Manuf. 31 (2015) 91100.
[3] A. Joubair, M. Slamani, I.A. Bonev, A novel XY-Theta precision table and a
geometric procedure for its kinematic calibration, Robot. Comput.-Integr.
Manuf. 28 (2012) 5765.
[4] S. Marie, E. Courteille, P. Maurine, Elasto-geometrical modeling and calibration
of robot manipulators: application to machining and forming applications,
Mech. Mach. Theory 69 (2013) 1343.
[5] A. Nubiola, I.A. Bonev, Absolute robot calibration with a single telescoping
ballbar, Precis. Eng. 38 (2014) 472480.
[6] Z.S. Roth, B. Mooring, B. Ravani, An overview of robot calibration, IEEE J. Robot.
Autom. 3 (1987) 377385.
[7] J. Santolaria, F.-J. Brosed, J. Velzquez, R. Jimnez, Self-alignment of on-board
measurement sensors for robot kinematic calibration, Precis. Eng. 37 (2013)
699710 7//.
[8] J. Santolaria, M. Gins, Uncertainty estimation in robot kinematic calibration,
Robot. Comput.-Integr. Manuf. 29 (2013) 370384.
[9] A. Elatta, L.P. Gen, F.L. Zhi, Y. Daoyuan, L. Fei, An overview of robot calibration,
Inform. Technol. J. 3 (2004) 7478.
[10] A. Klimchik, A. Pashkevich, D. Chablat, G. Hovland, Compliance error compensation technique for parallel robots composed of non-perfect serial chains,
Robot. Comput.-Integr. Manuf. 29 (2013) 385393.
[11] D. Daney and I. Z. Emiris, Robust parallel robot calibration with partial information, in: Proceedings of the IEEE International Conference on, Robotics
and Automation, ICRA, 2001, pp. 32623267.
[12] M.R. Driels, Full-pose calibration of a robot manipulator using a coordinatemeasuring machine, Int. J. Adv. Manuf. Technol. 8 (1993) 3441.
[13] M. Ikits and J. M. Hollerbach, Kinematic calibration using a plane constraint,
in: Proceedings of the IEEE International Conference on Robotics and Automation, 1997, pp. 31913196.
[14] W. Veitschegger and C.-h. Wu, A method for calibrating and compensating
robot kinematic errors, in: Proceedings of the IEEE International Conference
on Robotics and Automation, 1987, pp. 3944.
[15] H. Hage, P. Bidaud, and N. Jardin, Practical consideration on the identication
of the kinematic parameters of the Stubli TX90 robot, in: Proceedings of the
13th World Congress in Mechanism and Machine Science, Guanajuato, Mexico,
2011, p. 43.
[16] J.-M. Renders, E. Rossignol, M. Becquet, R. Hanus, Kinematic calibration and
geometrical parameter identication for robots, Robot. Autom. IEEE Trans. 7
(1991) 721732.
[17] B. Mooring The effect of joint axis misalignment on robot positioning accuracy,
in: Proceedings of the Engineering Conference on ASME International Computers. 1983, pp. 151156.
[18] W. Khalil, E. Dombre, Modeling, Identication and Control of Robots, Butterworth-Heinemann, Oxford, 2004.
167
[19] S. A. Hayati, Robot arm geometric link parameter estimation, in: Proceedings
of the 22nd IEEE Conference on, Decision and Control 1983, pp. 14771483.
[20] H. Zhuang, Z.S. Roth, F. Hamano, A complete and parametrically continuous
kinematic model for robot manipulators, Robot. Autom. IEEE Trans. 8 (1992)
451463.
[21] M. A. Meggiolaro and S. Dubowsky, An analytical method to eliminate the
redundant parameters in robot calibration, in: Proceedings of the IEEE International Conference on, Robotics and Automation ICRA'00 2000, pp. 3609
3615.
[22] M.R. Driels, W. Swayze, Automated partial pose measurement system for
manipulator calibration experiments, Robot. Autom. IEEE Trans. 10 (1994)
430440.
[23] Y. Bai, H. Zhuang, and Z. S. Roth, Experiment study of PUMA robot calibration
using a laser tracking system, in: Proceedings of the IEEE International
Workshop on Soft Computing in Industrial Applications SMCia/03. 2003, pp.
139144.
[24] A. Rauf and J. Ryu, Fully autonomous calibration of parallel manipulators by
imposing position constraint, in: Proceedings of the IEEE International Conference on, Robotics and Automation ICRA. 2001, pp. 23892394.
[25] S. Besnard and W. Khalil, Calibration of parallel robots using two inclinometers, in: Proceedings of the IEEE International Conference on, Robotics
and Automation 1999, pp. 17581763.
[26] A. Rauf, A. Pervez, J. Ryu, Experimental results on kinematic calibration of
parallel manipulators using a partial pose measurement device, Robot. IEEE
Trans. 22 (2006) 379384.
[27] A. Goswami, A. Quaid, and M. Peshkin, Complete parameter identication of a
robot from partial pose information, in: Proceedings of the IEEE International
Conference on, Robotics and Automation 1993, pp. 168173.
[28] B.W. Mooring, Z.S. Roth, M.R. Driels, Fundamentals of Manipulator Calibration,
Wiley, New York, 1991.
[29] C. Rao, H. Toutenburg, Linear Models and Generalizations: Least Squares and
Alternatives, Springer, New York, 1999.
[30] H. Zhuang, J. Yan, O. Masory, Calibration of Stewart platforms and other parallel manipulators by minimizing inverse kinematic residuals, J. Robot. Syst. 15
(1998) 395405.
[31] R. Fletcher, Practical Methods of optimization, John Wiley & Sons, Hoboken,
NJ, Wiley, 2013 452 p.
[32] E. Moore, H., On the reciprocal of the general algebraic matrix, Bull. Am. Math.
Soc. 26 (1920) 394395.
[33] R. Penrose, A generalized inverse for matrices, Proc. Cambridge Philos. Soc
(1955) 406413.
[34] J.-H. Borm, C.-H. Meng, Determination of optimal measurement congurations
for robot calibration based on observability measure, Int. J. Robot. Res. 10
(1991) 5163.
[35] M.R. Driels, U.S. Pathre, Signicance of observation strategy on the design of
robot calibration experiments, J. Robot. Syst. 7 (1990) 197223.
[36] H. Zhuang, O. Masory, and J. Yan, Kinematic calibration of a Stewart platform
using pose measurements obtained by a single theodolite, in: Proceedings of
the IEEE/RSJ International Conference on Intelligent Robots and Systems 95.
Human Robot Interaction and Cooperative Robots 1995, pp. 329334.
[37] M. A. Meggiolaro, G. Scrifgnano, and S. Dubowsky, Manipulator calibration
using a single endpoint contact constraint, in: Proceedings of ASME Design
Engineering Technical Conference, Baltimore, USA, 2000.
[38] H. Zhuang, J. Wu, and W. Huang, Optimal planning of robot calibration experiments by genetic algorithms, in: Proceedings of the IEEE International
Conference on, Robotics and Automation 1996, pp. 981986.
[39] D. Daney, Y. Papegay, B. Madeline, Choosing measurement poses for robot
calibration with the local convergence method and Tabu search, Int. J. Robot.
Res. v24 (2005) 501518.
[40] Y. Sun and J. M. Hollerbach, Observability index selection for robot calibration,
in: Proceedings of the IEEE International Conference on, Robotics and Automation ICRA 2008, pp. 831836.
[41] A. Nahvi and J. M. Hollerbach, The noise amplication index for optimal pose
selection in robot calibration, in: Proceedings. of the IEEE International Conference on Robotics and Automation 1996, pp. 647654.
[42] A. Nahvi, J. M. Hollerbach, and V. Hayward, Calibration of a parallel robot using
multiple kinematic closed loops, in: Proceedings of the IEEE International
Conference on Robotics and Automation 1994, pp. 407412.
[43] D. Daney, Optimal measurement congurations for Gough platform calibration, in: Proceedings of the IEEE International Conference on, Robotics and
Automation ICRA'02. 2002, pp. 147152.
[44] H. Zhuang, K. Wang, and Z. S. Roth, Optimal selection of measurement congurations for robot calibration using simulated annealing, in: Proceedings of
the IEEE International Conference on Robotics and Automation 1994, pp. 393
398.
[45] Y. Wu, Optimal pose selection for the identication of geometric and elastostatic parameters of machining robots, Ecole des Mines de Nantes, PhD Thesis
(2014).
[46] Y. Chen, J. Gao, H. Deng, D. Zheng, X. Chen, R. Kelly, Spatial statistical analysis
and compensation of machining errors for complex surfaces, Precis. Eng. 37
(2013) 203212 1//.
[47] Z. Roth, B. Mooring, B. Ravani, An overview of robot calibration, Robot. Autom.
IEEE J. 3 (1987) 377385.
[48] H. Stone, A. Sanderson, and C. Neuman, Arm signature identication, in:
Proceedings of the IEEE International Conference on, Robotics and Automation
1986, pp. 4148.
168
[49] J. Denavit, R. Hartenberg, A kinematic notation for lower-pair mechanisms

based on matrices, Trans. ASME J. Appl. Mech. 23 (1955) 215221.
[50] K. Schrer, S.L. Albright, M. Grethlein, Complete, minimal and model-continuous kinematic models for robot calibration, Robot. Comput.-Integr. Manuf.
13 (1997) 7385.
[51] A. Pashkevich, Computer-aided generation of complete irreducible models for
robotic manipulators, in: Proceedings of the 3rd International Conference of
Modellimg and Simulation. University of Technology of Troyes, France, 2001,
pp. 293298.
[52] H. Zhuang, F. Adviser-Hamano, and Z. S. Adviser-Roth, Kinematic modeling,
identication and compensation of robot manipulators, PhD Dissertation,
University Boca Raton, FL, USA, 1989.
[53] A. Klimchik, Y. Wu, S. Caro, B. Furet, A. Pashkevich, Geometric and elastostatic
calibration of robotic manipulator using partial pose measurements, Adv.
Robot. 28 (2014) 14191429.
[54] A. Klimchik, Y. Wu, G. Abba, S. Ganrnier, B. Furet, and A. Pashkevich, Robust
algorithm for calibration of robotic manipulator model, in: Proceedings of the
IFAC Conference on Manufacturing Modeling, Management and Control, Saint
Petersburg, Russia, 2013, pp. 838842.
[55] A. Klimchik, Y. Wu, A. Pashkevich, S. Caro, B. Furet, Optimal Selection of

Measurement Congurations for Stiffness Model Calibration of Anthropomorphic Manipulators, Appl. Mech. Mater. 162 (2012) 161170.
[56] A. Klimchik, S. Caro, A. Pashkevich, Optimal pose selection for calibration of
planar anthropomorphic manipulators, Precis. Eng. 40 (2015) 214229.
[57] A. Klimchik, A. Pashkevich, Y. Wu, B. Furet, S. Caro , In: C.-Y. Su, S. Rakheja,
H. Liu (Ed.), Optimization of measurement congurations for geometrical
calibration of industrial robot 2012. Intelligent Robotics and Applications,
Springer, 132143, http://dx.doi.org/10.1007/978-3-642-33509-9_13.
[58] S. Caro, C. Dumas, S. Garnier, and B. Furet, Workpiece placement optimization
for machining operations with a KUKA KR270-2 robot, in: Proceedings of the
IEEE International Conference on Robotics and Automation (ICRA). 2013, pp.
29212926.
[59] C. Dumas, S. Caro, M. Cherif, S. Garnier, B. Furet, X. Zha, et al., Joint stiffness
identication of industrial serial robots, Robotica 30 (2012) 649659.
[60] Kuka, http://www.kuka-robotics.com.
[61] Leica-geosystems. http://metrology.leica-geosystems.com/en/index.htm.
Journal of Mechanical Science and Technology 28 (3) (2014) 963~969

www.springerlink.com/content/1738-494x
DOI 10.1007/s12206-013-1167-7
Optimal combination of design parameters for improving the kinematics

characteristics of a midsize truck through design of experiment
Bo Myung Kim1, Jae Won Kim2, Il Dong Moon3 and Chae Youn Oh4,*
1
KDC, 205 GNATC, 676 Bangji-ri, Sanam, Sachen, Gyungnam, 664-942, Korea
Dept. of Bio-Nano System Engineering, Chonbuk National University, 664-14 1-Ga, Deokjin-Dong, Deokjin-Gu, Jeonju, Jeonbuk, 561-756, Korea
3
Vehicle Advanced Team, Corporate R&D Division, Hyundai-Kia Motors, 772-1 Jangduk-Dong, Hwaseong, Gyeonggi, 445-706, Korea
4,*
Div. of Mechanical System Engineering, Chonbuk National University, 664-14 1-Ga, Deokjin-Dong, Deokjin-Gu, Jeonju, Jeonbuk, 561-756, Korea
2
(Manuscript Received July 27, 2012; Revised August 9, 2013; Accepted September 16, 2013)
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Abstract
This paper optimizes the combination of design parameters for improving the kinematic characteristics of a midsize truck using both
design of experiment and computer simulation. A computational model of the front suspension and steering system of a midsize truck is
developed for analyzing kinematic and compliance characteristics. A taper leaf spring is modeled as a flexible body using finite elements.
A bump mode test is performed to validate the reliability of the developed computational model. Mean absolute values of the toe angle
and wheel base change are used as objective functions. Modifiable hard points are selected as design parameters. An optimal combination of design parameters for improving kinematic characteristics is suggested based on analyses of variance and factor effects using a
table of orthogonal arrays.
Keywords: Design of experiment; Kinematics analysis; Toe angle; Wheel base; Suspension parameter measuring device (SPMD); Table of orthogonal
array; Analysis of variance (ANOVA); Analysis of factor effect
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
1. Introduction
Since the online purchase of goods has become more common, the demands on midsize trucks have increased rapidly.
In order to safely and quickly deliver goods, midsize trucks
must perform reliably. Part of the process of improving the
performance of midsize trucks is improving straight-ahead
driving stability.
The kinematic and compliance characteristics of the front
suspension and steering systems of midsize trucks influence
the wheel alignment and wheel center locus. Changes in wheel
alignment and wheel center locus influence straight-ahead
driving stability. The joints and links that comprise the front
suspension and steering system influence kinematic characteristics significantly. The springs and bushings used for connecting links significantly influence compliance characteristics. In addition, the left and right asymmetry of the steering
system influences the kinematic and compliance characteristics.
In the early stage of the development of a midsize truck,
many vehicle and part tests are performed to find an optimal
combination of design parameters to influence straight-ahead
*
Corresponding author. Tel.: +82 63 270 2377, Fax.: +82 63 270 2388
E-mail address: ohcy@jbnu.ac.kr
Recommended by Associate Editor Gang-Won Jang

KSME & Springer 2014
driving stability. A computational model can be used to find

an optimal combination of design parameters, and thus design
purpose can be achieved with less time and cost.
Recent studies have analyzed the dynamic characteristics of
commercial vehicles using computational models [1-6]. In
order to improve the reliability of the computational model of
a commercial vehicle, the leaf spring and frame were modeled
with finite elements [7-9]. An appropriate computational
model for analyzing kinematic and compliance characteristics
has also been developed [10]. A computational model was
used to find an optimal combination of design parameters for
improved maneuverability [11]. In addition, a computational
model was used to optimize cab structure in order to improve
vehicle ride comfort [12].
This paper describes the development of the front suspension and steering system computational model of a midsize
truck. In order to validate the reliability of the computational
model, a test is performed with a suspension parameter measurement device (SPMD). We also propose a scheme to optimize the design parameters for the improvement of kinematic
characteristics.
2. Computational modeling
The computational model of the front suspension and steer-
964
B. M. Kim et al. / Journal of Mechanical Science and Technology 28 (3) (2014) 963~969
Fig. 1. Finite element model of the taper leaf spring.
ing system of a midsize truck was developed with

MSCADAMS [13]. The suspension system of the midsize
truck used in this paper consists of a front axle, shock absorber,
shackle, and taper leaf spring. The taper leaf spring is composed of three leaves. Since the taper leaf spring not only deforms like a flexible body but also moves like a rigid body, the
taper leaf spring significantly influences the kinematic and
compliance characteristics. This paper modeled the taper leaf
spring as a flexible body. The flexible body model of the taper
leaf spring was developed with finite elements. The finite
element model of the taper leaf spring is composed of 20,360
nodes, 14,704 hexahedral elements, and 26 multi-point constraints. The density, modulus of elasticity, and Poisson ratio
of the taper leaf spring's finite element model were set to 7.8E006 kg/mm3, 210000 N/mm2, and 0.3 respectively. Fig. 1
shows the finite element model of the taper leaf spring.
A dummy part was attached at the fore end of the upper finite element leaf spring model to connect the fore end of the
taper leaf spring with the frame. Then, the dummy part was
joined to the frame, a ground part, with a bushing element. A
dummy part was attached at the rear end of the upper finite
element leaf spring model to connect the rear end of the taper
leaf spring with the shackle. Then, the dummy part was joined
to the shackle with a bushing element. The shackle was connected to the frame with a bushing element. To connect the
upper and the middle leaf spring, dummy parts were attached
at the front, middle and rear of the bottom of the upper finite
element leaf spring model as shown in Fig. 1. Also, dummy
parts were attached at the front, middle and rear of the top of
the middle finite element leaf spring model. Two dummy parts
which were attached at the front of the upper and the middle
finite element leaf spring model were joined using an in-plane
joint. Two dummy parts which were attached at the rear of the
upper and the middle finite element leaf spring model were
joined using an in-plane joint. Two dummy parts which were
attached at the middle of the upper and the middle finite element leaf spring model were joined using a fixed joint. The
middle and the lower finite element leaf spring model were
connected in the same manner as the connection of the upper
and the middle finite element leaf spring model. To connect
the taper leaf spring with the front axle, a dummy part was
attached at the middle of the bottom of the lower finite ele-
Fig. 2. Computational model of the front suspension and steering system of a mid-sized truck on a test rig.
ment leaf spring model. The dummy part was joined to the
front axle with a bushing element. The upper end of a shock
absorber was connected to the frame with a universal joint.
The lower end of the shock absorber was connected to the
front axle with a spherical joint.
The steering system of the midsize truck used in this paper
is composed of a steering wheel, steering column, steering
column housing, steering shaft, steering rack, drop arm, drag
link, knuckle, and tie rod. The steering wheel was connected
to the steering column with a revolute joint. The steering column was connected to the steering column housing with a
cylindrical joint. Also, it was connected to the steering shaft
with a universal joint. The steering column housing was connected to the ground part with a fixed joint. The steering shaft
was connected to the steering rack with a revolute joint. The
steering rack was connected to a drop arm with a revolute
joint. The drop arm was connected to a drag link with a spherical joint. The drag link was connected to the left knuckle with
a universal joint. The left end of a tie rod was connected to the
left knuckle with a spherical joint, and the right end of a tie
rod was connected to the right knuckle with a universal joint.
Fig. 2 shows the computational model of the front suspension
and steering system of a midsize truck on a test rig.
3. Validation of the developed computational model

In order to validate the computational model, a bump mode
test was performed with a SPMD. In the bump mode test, the
left and right tires were actuated in in-phase mode. Jounce and
rebound strokes for the test were -55 mm, and 55 mm. Factors
such as wheel rate, toe angle, caster angle, and wheel base
changes representing kinematic characteristics were measured.
A simulation was performed with the developed computational model under the same conditions as the test.
Figs. 3-6 compare the wheel rate, toe angle, caster angle,
and wheel base changes acquired from a simulation with those
acquired from the test. Fig. 3 shows a comparison of the
Fig. 3. Comparison of the changes in wheel rate from the simulation

vs. the SPMD test.
changes in the left and right wheel rates. The simulation results predict the trends of the test results over the whole range
of the stroke. Fig. 4 shows a comparison of the changes in the
left and right toe angles. The trends seen in the simulation
show a little difference from the results of the test at strokes
between -55 mm ~ -30 mm for both the left and right toe angles. However, the trends seen in the simulation are in good
agreement with those from the test for the remainder of the
stroke. Fig. 5 shows a comparison of the changes in the left
and right caster angles. For the left caster angle, the trend seen
in the simulation show a little difference with those from the
test stroke between -55 mm ~ -30 mm, but the simulation
result provides a good estimate of the trend seen in the test
result for the remainder of the stroke. For the right caster angle,
the simulation result provides a good estimate of the trend
seen in the test result over the whole range of stroke. Fig. 6
shows a comparison of the changes in the left and right wheel
base. The simulation results predict the test results over the
whole range of the stroke. From these comparisons, we can
conclude that the wheel rate change, toe angle change, caster
angle change, and wheel base change considered as major
factors representing the kinematic characteristic acquired from
the simulation provide a good estimate of those acquired from
the test in general.
4. Optimal combination of design parameters

This paper suggests an optimal combination of design parameters for improving kinematic characteristics using the
validated computational model and design of experiment. The
965
Fig. 4. Comparison of the changes in toe angle from the simulation vs.
the SPMD test.
Fig. 5. Comparison of the changes in caster angle from the simulation

vs.the SPMD test.
toe angle and wheel base change are considered to be the two
most important factors influencing straight-ahead driving stability. Therefore, we sought to optimize the design parameters
in order to minimize toe angle and wheel base changes.
To design an experiment, we need to define an objective
function, select proper design parameters, and select the num-
966
Table 1. Design parameters and design values at each level.

Level
Parameter
Fig. 6. Comparison of the changes in wheel base from the simulation

vs. the SPMD test.
ber of levels. This paper defines objective functions to minimize the mean absolute values of toe angle and wheel base
change as shown in Eqs. (1) and (2).
n
T (j )
i
Minimize
F (j ) =
i =1
(1)
(2)
D (d )
i
Minimize
D (d ) =
i =1
In Eqs. (1) and (2), n indicates the total number of data

points, Ti() is a toe angle at time ti in Fig. 4, and Di() is a
wheel base change at time ti in Fig. 6.
Kinematic hard points, which are considered to be important parameters in the early stages of design, were selected as
candidate parameters. Design parameters which can be easily
modified among candidate parameters are selected for optimization. These were the hard points connecting the drag link to
the knuckle, and connecting the drop arm to the drag link.
The number of levels of all design parameters was set at
three. The level range for all design parameters was set to 20
mm, which is a modifiable range. Table 1 shows the design
parameters and design values at each level.
A table of orthogonal arrays was created as described by Eq. (3).
3m -1
L3m (3 2 ) .
(3)
Level 0
Level 1
Level 2
x-coordinate of a hard point

joining drag link and knuckle
Current
-20
Current
Current
+20
y-coordinate of a hard point

Current
-20
Current
Current
+20
z-coordinate of a hard point

Current
-20
Current
Current
+20
x-coordinate of a hard point

B joining drop arm and drag
link
Current
-20
Current
Current
+20
y-coordinate of a hard point

C joining drop arm and drag
link
Current
-20
Current
Current
+20
z-coordinate of a hard point

D joining drop arm and drag
link
Current
-20
Current
Current
+20
In Eq. (3), 3m is the number of rows representing the total

number of experiments (N) and (3m-1)/2 is the number of columns in the table of orthogonal arrays. Since the number of
levels is 3 (m = 3), the total number of experiments is 27, and
the number of columns is 13. Table 2 shows the table of orthogonal arrays generated in this paper. In the table, the numbers 0, 1, and 2 refer to level index. We run a simulation for
each trial defined in the table. The value of the objective function for each trial was computed. The right column of the table
contains the computed value of the objective function. In Table 2, "T" is the total sum of objective functions. "CT" is a
corrective term, and was computed by the equation, CT =
T2/N.
Analysis of variance (ANOVA) was performed to analyze
the influence of design parameters on the toe angle and wheel
base change using Table 2 with MINITAB [14]. Table 3
shows the result of analysis of variance on the toe angle. We
can confirm from Table 3 that parameters D, B, and F were
primary factors influencing the objective function. Factor effect analysis was performed with Table 3. Fig. 7 shows the
factor effect analysis result in the graph. As Fig. 7 shows, in
order to minimize the change in the toe angle, parameters B, D,
G, and C need to be increased, and parameters F and A need
to be decreased. Therefore, the optimal combination of design
parameters to minimize the change in the toe angle is
A0G2F0B2C2D2.
Table 4 shows the result of analysis of variance of the wheel
base change using Table 2. Fig. 8 shows the factor effect analysis result of the wheel base change. As Table 4 and Fig. 8
show, the wheel base change and the toe angle both have the
same result.
A simulation was performed with the suggested optimal
combination of design parameters. Table 5 compares object
function values acquired from the simulation with the current
combination of design parameters and those acquired from the
967
Table 2. Orthogonal array.
Table 4. Result of analysis of variance of wheel base change.

Objective function
value
Design parameter
Trial
1 2 3 4 5 6 7 8 9 10 11 12
13 Toe angle
Wheel
base
0 0 0 0 0 0 0 0 0 0
0.4899
5.2583
0 0 0 0 1 1 1 1 1 1
0.4648
5.2204
0 0 0 0 2 2 2 2 2 2
0.4377
5.1794
0 1 1 1 0 0 0 1 1 1
0.1578
4.5886
0 1 1 1 1 1 1 2 2 2
0.5917
5.4122
0 1 1 1 2 2 2 0 0 0
0.6319
5.4729
0 2 2 2 0 0 0 2 2 2
0.1926
4.8014
0 2 2 2 1 1 1 0 0 0
0.2354
4.8744
0 2 2 2 2 2 2 1 1 1
0.7356
5.6297
10 1 0 1 2 0 1 2 0 1 2
0.2834
4.9468
11 1 0 1 2 1 2 0 1 2 0
0.2541
4.9016
12 1 0 1 2 2 0 1 2 0 1
0.9164
5.9023
13 1 1 2 0 0 1 2 0 1 2
0.2578
4.9081
14 1 1 2 0 1 2 0 1 2 0
0.2305
4.8663
15 1 1 2 0 2 0 1 2 0 1
0.8295
5.7714
16 1 2 0 1 0 1 2 0 1 2
0.2363
4.8758
17 1 2 0 1 1 2 0 1 2 0
0.2109
4.8370
18 1 2 0 1 2 0 1 2 0 1
0.7577
5.6632
19 2 0 2 1 0 2 1 0 2 2
0.1708
4.6439
20 2 0 2 1 1 0 2 1 0 2
0.7177
5.6024
21 2 0 2 1 2 1 0 2 1 0
0.7150
5.5981
22 2 1 0 2 0 2 1 0 2 1
0.1556
4.6332
23 2 1 0 2 1 0 2 1 0 2
0.6529
5.5049
24 2 1 0 2 2 1 0 2 1 0
0.6478
5.4984
25 2 2 1 0 0 2 1 0 2 1
0.1429
4.6243
26 2 2 1 0 1 0 2 1 0 2
0.5989
5.4235
27 2 2 1 0 2 1 0 2 1 0
0.5937
5.4155
AG e e F e e e B C
e T = 12.310 140.05
Factor
Sum of
square
Degree of
freedom
Mean
square
(%)
Contribution
0.0143
0.0072
0.260
0.0684
0.0342
1.230
2.7376
1.3688
49.13
1.7260
0.8630
30.98
0.0003
0.0001
0.001
1.0256
0.5128
18.41
Table 5. Comparison of mean absolute values of toe angle and wheel

base change before and after optimization.
Current
combination
Optimal
combination
Toe
angle
A1G1F1B1C1D1
A0G2F0B2C2D2
0.44
0.15
Wheel
base
A1G1F1B1C1D1
A0G2F0B2C2D2
5.19 mm
4.54 mm
65.9% Down
12.5% Down
Fig. 7. Result of factor effect analysis of toe angle.
CT = 5.613 726.49
Table 3. Result of analysis of variance of toe angle.
Factor
Sum of
square
Degree of
freedom
Mean
square
(%)
Contribution
0.0144
0.0072
0.53
0.0314
0.0157
1.16
0.9738
0.4869
35.90
0.6587
0.3294
24.28
0.0087
0.0044
0.32
1.0256
0.5128
37.81
simulation with the suggested optimal combination of design

parameters. As Table 5 shows, the suggested optimal combination of design parameters reduced the toe angle change by
65.9% and reduced the wheel base change by 12.5%. Fig. 9
compares the changes in toe angle and wheel base acquired
Fig. 8. Result of factor effect analysis of wheel base change.
from a simulation performed with the current combination of

design parameters with those acquired from a simulation performed with the suggested optimal combination of design
parameters.
968
and design parameters B, D, G, and C need to be increased in

order to minimize changes in toe angle and wheel base. Toe
angle change was reduced by 65.9% and wheel base change
was reduced by 12.5% in a simulation using the suggested
optimal combination of design parameters. There may be the
interaction effects among the design factors on the kinematic
characteristics. This paper considered only the main effects.
We are going to study the interaction effects in future.
References
Fig. 9. Comparison of the changes in toe angle and wheel base from
the current combination of design parameters vs. the suggested optimal
combination of design parameters.
5. Conclusions
This paper described the development of a computational
model of the front suspension and steering system of a midsize truck for the purpose of analyzing the kinematic characteristics. A taper leaf spring was modeled as a flexible body
using finite elements. A bump mode test was performed with
a SPMD and a simulation was performed under the same conditions as the test. The trends in both the left and right toe
angles and left caster angle acquired from the simulation
showed a little difference from those acquired from the test at
strokes between -55 mm ~ -30 mm. However, the trends in the
wheel rate, toe angle, caster angle, and wheel base change
acquired from the simulation were good estimates of those
acquired from the test in general. Design of experiment and
computational simulation were both used to seek an optimal
combination of design parameters for minimizing the change
in toe angle and wheel base. Mean absolute values of toe angle and wheel base change were used as objective functions.
Changes in the x, y, and z directions of hard points connecting
the drag link to the knuckle, and connecting the drop arm to
the drag link were selected as design parameters for optimization. A table of orthogonal array having 13 columns and 27
rows was generated. The result of analysis of variance showed
that x and z coordinate of a hard point joining drop arm and
drag link, and z coordinate of a hard point joining drag link
and knuckle were primary factors influencing toe angle and
wheel base change. The result of factor effect analysis suggested that design parameters A and F need to be decreased
[1] A. Ichikawa, H. Shinjo, T. Shima and Y. Susuki, Practical

application of CAE for truck controllability and stability,
SAE paper 912530 (1991) 673-681.
[2] I. D. Moon, H. J. Kwon and C. Y. Oh, Development of a
computer model for the turning maneuver analysis of a
heavy truck, J. of KSAE, 8 (4) (2000) 121-129.
[3] I. D. Moon and C. Y. Oh, A study on the estimation of the
ride quality of a large-sized truck using a computer model, J.
of KSME, 25 (12) (2001) 2048-2055.
[4] I. D. Moon and C. Y. Oh, A study on the effects of hysteretic
characteristics of leaf springs on handling of a large-sized
truck, J. of KSAE, 9 (5) (2001) 157-164.
[5] S. L. Haas, Ranking tires for heavy truck steering feel performance using a simulation model, Tire Science and Technology, 34 (1) (2006) 64-82.
[6] T. W. Park, H. J. Yim, G. H. Lee, C. J. Park and I. H. Jeong,
Development of mini-bus ride analysis method, J. of KSAE,
7 (1) (1999) 149-154.
[7] I. D. Moon and C. Y. Oh, Development of a computer model
of a large-sized truck considering the frame as a flexible
body, J. of KSAE, 11 (6) (2003) 197-204.
[8] I. D. Moon, H. S. Yoon and C. Y. Oh, A flexible multi-body
dynamic model for analyzing the hysteretic characteristics
and the dynamic stress of a taper leaf spring, J. of Mechanical Science and Technology, 20 (10) (2006) 1638-1645.
[9] I. D. Moon, C. Y. Oh and S. H. Oh, The effects of the
mounted method of frame of a large truck on handling performance, J. of KSPE, 21 (8) (2004) 112-119.
[10] I. D. Moon and C. Y. Oh, Computational model for analyzing the kinematics and compliance characteristics of a
commercial vehicle's front suspension system, International Journal of Automotive Technology, 13 (2) (2012)
279-284.
[11] I. D. Moon, D. H. Lee and C. Y. Oh, A study on optimal
combination of design parameters for improving handling
performance of a large truck using design of experiments, J.
of KSME, 28 (6) (2004) 799-806.
[12] S. H. Ryu and K. Choi, A study on the optimum design if
the front structure of heavy duty truck cab using sensitivity
analysis and taguchi method, KSAE 2006 Spring Conference,
(2006) 2006-2012.
[13] MSC.ADAMS, MacNeal-Schwendler, Inc., Oakdale, CA,
U.S.A. (2001).
[14] MINITAB, Minitab, Inc., State College, PA, U.S.A. (2005).
Bo Myung Kim received his B.S. degree in Aerospace Engineering and M.S.
degree in Precision Mechanical Engineering from Chonbuk National University, Korea, in 2003 and 2009, respectively. He is currently a researcher at
KDC Co. in Sacheon, Korea. Mr. KIMs
research interests are in the area of vehicle dynamics.
969
Chae Youn Oh received his B.S. degree in Mechanical Engineering from

Chonbuk National University, Korea, in
1982. He then received his M.S. degree
and Ph.D. in Mechanical Engineering
from Iowa State University, U.S.A., in
1987 and 1990, respectively. He is currently a Professor at the Division of
Mechanical System Engineering at Chonbuk National University in Jeonju, Korea. Dr. OHs research interests are in the
area of vehicle dynamics and haptics.
404
ISSN 13921207. MECHANIKA. 2015 Volume 21(5): 404413
Optimal kinematic design of a multi-link steering system for a bus

independent suspension: An application of response surface
methodology
M.M. Topa*, U. Deryal**, E. Bahar***, G. Yavuz****
*Faculty of Engineering, Dokuz Eyll University, 35397 Izmir, Turkey, E-mail: murat.topac@deu.edu.tr
**Faculty of Engineering, Dokuz Eyll University, 35397 Izmir, Turkey, E-mail: ugur.deryal@ogr.deu.edu.tr
***Faculty of Engineering, Dokuz Eyll University, 35397 Izmir, Turkey, E-mail: egemen.bahar@ogr.deu.edu.tr
****Hexagon Studio, TOSB 1.Cadde 15.Yol No: 7, 41420, Kocaeli, Turkey, E-mail: Guven.Yavuz@hexagonstudio.com.tr
http://dx.doi.org/10.5755/j01.mech.21.5.11964
Nomenclature
F - steering error; L - steering angle of the wheel;
S - steering wheel angle; V - toe angle; CCD - central
composite design; DSA - design sensitivity analysis;
eL - unit vector of the steering axis; eYR - unit vector of the
wheel rotation axis; FFD - full factorial design;
LF - wheelbase, mm; MSE - maximum steering error, ;
O - centre of the bend; SA - centre of gravity of the vehicle
body; sR - track width, mm; SS - sweep study; zA8 - wheel
displacement, mm
1. Introduction
Through their design simplicity and ease of manufacturing, Ackerman trapezoidal linkage has a broad application area as the steering system of heavy commercial
vehicles equipped with solid axles. On the other hand, as a
result of the comfort and control requirements, one of the
main targets to be reached in the design of vehicle suspensions is to keep the unsprung mass as small as possible. In
order to satisfy these demands, independent front suspensions (IFS) are applied increasingly on busses and trucks
by the heavy commercial vehicle manufacturers [1]. In this
case, more sophisticated systems are demanded to meet the
sufficient steering and independent wheel travel functions
simultaneously. Because of its design advantages, multilink steering linkage (or opposed four-bar linkage [2, 3])
is used in the majority of the passenger busses equipped
with IFS. This mechanism basically consists of two relay
levers, one track rod, two tie rods and two steering arms as
seen in Fig. 1.
Kinematic model of a typical bus IFS including
Fig. 1 General view of a bus IFS with the multi-link

steering linkage
the multi-link steering linkage is also seen in Fig. 2. Here,

A8 is the centre of mass of the wheel assembly. A7 is the
intersection point of steering axis and wheel rotation axis
which are described by the unit vectors eL and eYR respectively. The co-ordinate system x-y-z is described at the
vehicle body centre of gravity SA. As can be seen from this
model, the relay levers are attached to the vehicle body via
revolute joints A14. Transmission of the rotational motion
between the relay levers is provided by a track rod which is
mounted to the relay levers with spherical joints A13. In
majority of the busses, assembly of the relay levers and the
track rod is planar. Tie rod is also connected to steering
arm and relay lever spatially via spherical joints A9 and A10
respectively.
Fig. 2 Kinematic model of the IFS and steering linkage

Generally, one of the main requirements of a vehicle steering mechanism is to give to the steerable wheels
a correlated L such that, the intersection of the wheel axes
should meet at the centre of the bend, O [3, 4]. This rule
which can also be seen for a two axle vehicle in Fig. 3 is
known as Ackermann Principle.
Fig. 3 Ackermann steering geometry and V angle of the

wheel 1
405
(1)
where La is the steering angle of the outer wheel and LaA

is the ideal turning angle which is obtained from Eq. (1)
for a given steering angle Li of the inner wheel. The deviation F between the ideal steering angle and real turning
angle of the wheel which is caused by the steering mechanism geometry is called the steering error or Ackermann error. Basically, a steering mechanism should satisfy Ackermann principle for given steering error tolerances.
F can be written as:
F Li La Li LaA Li .
(2)
The spatial position of the tie rod of a multi-link

steering system also affects the toe (V) deviation of the
wheel (Fig. 3) as well as MSE. Hence, during the kinematic design process of a multi-link steering mechanism, the
deviation of V caused by the wheel travel should also be
taken into account.
In literature, there are some published works on
kinematic optimisation of the steering systems. Zhou et al.
optimised a rack-and-pinion steering mechanism by combining MATLAB genetic algorithm toolbox with MSC.
Adams to improve steering and V characteristics [5].
Hanzaki et al. performed the combined kinematic and sensitivity optimisation of a rack-and-pinion steering linkage
used in passenger cars [6]. Oz et al. presented a model validation methodology and the optimisation study on the
hardpoints of solid axle suspension and steering systems of
a heavy commercial vehicle by using Design of Experiments DOE approach [7]. Liang and Xin optimised the toe
deviation of a double wishbone suspension during vertical
wheel travel via Adams/View [8]. An interesting paper was
published by Kim et al. on the effect of the drag link hardpoints on V angle and the deviation of the wheelbase for a
solid front axle. In their work, they used a L27(313) type
orthogonal array and ANOVA to obtain the optimal combination of design parameters [9]. In open literature, to
date and to the best of the authors' knowledge there have
been a few works published on the kinematic optimisation
the multi-link steering mechanisms for independent suspensions. Bian et al. established the multibody model of
the steering mechanism based on the R-W (Robertson and
Wittenburg) method for the MacPherson strut [10] and
double wishbone suspension [11] for automobiles. In these
two works, the suggested models are identical. In summary, all of these works mainly focalise on optimisation of
the partial kinematic parameters of a steering system rather
than presenting a complete optimization procedure.
In the present study, a response surface-based design procedure to build a multi-link steering system for a
passenger bus IFS which satisfies optimum tolerances of
V and F deviations is proposed. A brief summary of the
method used in this work is seen in Fig. 4. In stage 1, a
multibody model of a bus IFS including the steering mechanism was performed by using MSC. Adams commercial
software. In order to compose the IFS model, the hard-
Stage 1 (Multibody modelling)

Adams/CarTM
Stage 2 (Toe optimisation)
Stage 3 (MSE optimisation)
Parameter selection (DSA)
Sweep study
Threshold value
MSE
Standardised effect
Iteration number
DOE-RSM (CCD)
DOE-RSM (CCD)
Response
surface
Factor 1
j
LF
Factor 3
Factor 2
Factor 3
Factor 1
Optimum toe deviation
Initial
Optimised
Toe angle
Factor 2
Factor 2
Optimum steering error

Ackermann error
cot Li
Term
Wheel travel
LaA Li tan 1
points A1 to A8 were drawn from a produced intercity bus.

In this model, primary position of the tie rod and the position tolerances of the hardpoints A9 and A10 were chosen
by considering the design limitations such as the brake
system and the knuckle design. In stage 2, optimal positions of A9 and A10 which directly affect the V angle of the
front wheels during the wheel travel were determined by
using DOE approach via Adams/Insight multi-objective
optimisation tool. To this end, firstly the most important
factors among the Degrees of Freedom (DOF) of A9 and
A10 on V deviation were chosen via a factorial designbased DSA. Results obtained from FFD study were utilised
in MINITAB, a practical stastistical software package
[12]. In order to find out the exact optimum locations of A9
and A10, CCD was also utilised. By using the results, the
vertical position of the multi-link mechanism plane was
also determined. In stage 3, which is the final phase, MSE
was optimised in the L range of the front wheels by taking
Ackermann principle into account. For this reason, an Hshaped parallel arm mechanism which is the most general
from was chosen as the base model for optimisation study.
Optimal positions of the hardpoints A10, A13 and A14 which
constitutes the kinematic shape of the relay lever and directly affect the steering error is studied by using a composite method including SS and CCD.
Factor 1
Mathematically, Ackermann principle can be expressed as:
Base
Optimised
Steering angle
Fig. 4 Summary of the optimisation process
406
2. Multibody model of the double wishbone suspension
3. Methodology
A three dimensional multibody model of the bus

IFS including the full steering system is composed by using Adams/Car module of MSC. Adams . In this model
which is seen in Fig. 5, Adams/Car co-ordinate axis
convention was applied [13]. Model consists of 15 elements. Kinematic constraint elements are also shown in
Fig. 6. In this model, steering wheel 1 and steering column
2 are directly connected with the upper part of the intermediate shaft 3 via universal joint (U). A translational joint
(T) which has single sliding DOF is defined between the
upper and lower parts of the intermediate shaft. Lower part
is also connected to the rack of the steering box 4 by a universal joint. Pitman arm 5 is fixed to the output shaft and
also connected to the drag link 6. The other end of drag
link and track rod 8 is mounted to the relay lever 7 via
spherical joints (S). In order to reduce the DOF and simplify the model, Convel homokinetic joint (C) which has 2
DOF is utilised instead of a 3 DOF spherical joint for relay
lever-tie rod 9 connection (A10). Steering arm 10 - tie rod
connection is provided by a spherical joint.
In this study DOE-RSM methods are utilised to

determine the optimum values of the parameters providing
the desired ranges of V deviation and F. Optimisation
process was carried out by Adams/Insight which includes DOE and RSM tools. The DOE approach is used for
understanding the correlation between the design parameters of the system and its performance [17]. Essentially,
RSM is one of the extended DOE methods which uses a
polinomial type regression model y(x) [18]. Principal target
of the response surface experiments is to obtain an appropriate model to estimate and analyse the relationship between design variables and system response. For a second
order response surface model, the regression model is defined in general form as [19]:
Direction
14
10
15
7
6
12
IIx
1 : Steering wheel
6 : Drag link
11: Upright
Fig.
5 Multibody model
of the bus IFS and
steering system
R
S
Lower wishbone
Upper wishbone
AirIIzspring
Shock absorber
IIy
U
U: Universal
T: Translational
R: Revolute
S: Spherical
C: Homokinetic
IIy
Direction R
Eq. (3) may also be written in terms of e.g. two

variables as:
y ,
12:
13:
14:
IIx
15:
IIx
(4)
(5)
(6)
where
y y1 ,y2 ,...,yM ;
T
1 x11
1 x12

... ...
1 x1M
x21
x22
...
x2 M
Fig. 6 Kinematic constraints of the steering mechanism

Steering axis is defined as the rotation axis of a
revolute joint (R) which is located on the upright 11. Upright is also connected with upper 13 and lower 12 wishbones via revolute and spherical joints. Air spring 15 is
directly mounted on upright instead of a wishbone. According to [14, 15] spring rate of the suspension can be
assumed as iF 1 (-) for this design type. PAC2002 tyre
model [16] was used for tyres with dimensions of 295/80 R
22.5 which is similar those fitted on the bus axles.
(7)
x51
... x52
;
... ...
... x5 M
...
0 , 1 , 2 , 3 , 4 , 5 ;
T
1 , 2 ,..., M ,
T
(3)
This model can also be expressed in matrix form

for M experiments as:
IIz
7 : Relay lever
8 : Track rod
9 : Tie rodR R
10 : Steering arm
i j
y 0 1 x1 2 x2 3 x3 4 x4 5 x5 .
IIy
2 : Steering column
3 : Intermediate shaft
Direction
4 : Steering box
5 : Pitman arm R
i 1
In order to linearise the regression model, new

variables may be expressed as x3 = x12, x4 = x22 and
x5 = x1x2. Hence, Eq. (4) can be written as:
y 0 1 x1 2 x2 3 x12 4 x22 5 x1 x2 .
11
13
y 0 i xi ij xi x j .
(8)
(9)
(10)
here y is vector of observations, X is the model matrix, is

the vector which includes the interception parameter 0 and
the partial regression coefficients and is the vector of
random errors [20], the estimated value of which minimises can be expressed as:
T T y .
1
(11)
ADAMS/Insight uses the method of least

squares to estimate the coefficients in the regression
model [21]. In this study, CCD type which is offered in the
design specification table of Adams/Insight was utilised
407
for this purpose. The CCD involves the use of a two-level
factorial or fraction combined with 2k axial or star points.
Hence, the design includes factorial points, 2k axial points,
and total nc centre runs, yielding a total number of
2k+2k+nc runs are carried out to achieve experimental
data. A comparison of the two level FFD and the CCD for
three factors is seen in Fig. 7.
Fig. 7 Comparison of FFD and CCD (according to [17])

Methodology of the study is summarised in
Fig. 8.
objective (V or F) is defined. Except for some differences, optimisation procedures for V and F are similar.
Absolute value of the maximum deviation of the objective
obtained from the primary simulation is imported to Adams/Insight. The factors and design targets are also defined. In the light of the design constraints, the variation
ranges of the factors are chosen. Effective and noneffective parameters are identified by screening experiments. Results obtained from these experiments are also
used for the SS of F optimisation. For the optimisation
processes, investigation strategy is chosen as DOE-RSM.
Number of the runs is determined according to the design
type. The design space and workspace which contain the
full set of the design trials and the results of their analyses
are generated. Optimum set of the factors which gives the
target value of the design objective is obtained by fitting
the results to a polynomial or a response surface. In order
to control the estimation results of the regression analysis,
a multibody model which contains the optimum values of
the factors is also carried out. Results obtained from this
model are compared with the target value. Some steps of
this flow chart have similar characteristics with the methodology given by [22] for the optimisation of the suspension parameters to improve impact harshness (IH) of road
vehicles.
4. Toe optimisation
Fig. 9 shows the six total translational DOF of the
hardpoints A9 and A10 which determine the position of the
tie rod and directly affect the V in case of wheel travel.
Since tie rod is connected to the relay lever by the spherical joint A10, initial position of the hardpoints A13 and A14
does not have any remarkable effect on V for a given value of L. Hence, only the hardpoints of A9 and A10 are chosen as factors in stage 2. A summary of the design limitations are also given in Fig. 10. Here OV, the midpoint of
the front axle was chosen as the reference point for this
work. Appropriate positions of the hardpoints A 9 and A10
were searched in the design volumes Cube 1 and Cube 2.
Initial positions of the cubes and their edge lengths were
chosen according to the design limitations which are summarised below. Position of A9 in x axis should render possible enough space for wheel-end and brake system (Volume B) to eliminate any penetration of the mechanical elements. Initial value of A10x co-ordinate is chosen such that
the mechanism should not be blocked in the L range of the
front wheels. As a design rule, the angles and between
steering arm and tie rod should not be lower than 15 [2] in
Fig. 8 Flow chart for the optimisation process

At the first stage, a primary simulation (parallel
wheel travel or steering) is carried out by using the initial
kinematic model generated in Adams/Car. The design
Fig. 9 Factors for A9 and A10
408
course of the maximum steering position of the front
wheels as seen in Fig. 11.
Steel wheel limits the -y co-ordinate of A9 because of the installation issues. A gap e is necessary
which limits the -y co-ordinate of A10 because of the installation issues. In this design, gap e is assumed as 60 mm by
considering the physical diameter of the spherical joint
A10.
ing experiments of the DSA, since it is the most proper

type [22]. In this design type, only the lowest and the highest values of each factor are taken into account. As a rule,
full factorial design provides 2n runs for a single screen of
experiments where n is the number of the factors. For
n = 6, FFD provides 64 trials (runs) which is considered as
a reasonable experiment number. The results of the trials
were fitted to a first order polynomial, whose general form
is given in Eq. (5). Schematic of the design model can be
seen in Fig. 13. Here, A9 and A10 (the red line) represent
the chosen initial position of the tie rod. In Fig. 13, A9i and
A10i also stand for the i-th observation as an example (the
green line). All of the possible design combinations which
connect A9 and A10 were generated by Adams/Insight
with the use of the cube edges. Successive simulations for
every design combination were carried out. Length of the
cube edges were chosen as 80 mm.
Fig. 13 Model for the tie rod position (schematic)
Fig. 10 Summary of design limitations (schematic): a - end

view, b - plan view
A9 15
IIy
A14
A14
IIx
Li
A7
A13
A13
A10
15
A9
La
A7
A10
j
Fig. 11 Full left turn simulation

Because of the design issues, position of the steering arm - knuckle connection is lower than the wheel center A8 as seen from Fig. 12. Moreover, lower wishbone
limits the vertical position of the tie rod. Hence, it is assumed that the vertical (z) co-ordinates of A9 and A10
should be in the range of z = (0 - 80) mm. For this assumption, brake system geometry is also taken into account.
Design criterion was chosen such as the V should not exceed 0.3 in the wheel travel range of zA8 = 100 mm.
In order to determine the effects of the factors on

V clearly, results obtained from the 64 trials given in the
design space were imported to MINITAB , statistical
software. Design type was defined as custom factorial design. V was also defined as the response of the analysis.
The level of significance was chosen as 0.01. Fig. 14
shows the Pareto chart of the standardised effects obtained
from MINITAB for the total 6 factors of A9 and A10. In
this chart, only the main effects of the DOFs are taken
into account. For = 0.01, the threshold value was calculated as u = 2.665. Since their standardised effects are
greater that this limit, A9z and A10z are predicted as the factors which directly affect the V during jounce and rebound. In order to test this result, the co-ordinate A9x
which is the most effective factor under the u limit was
solely altered in the range of 40 mm. It was found that
this alteration changed V about 2.5% for zA8 = 100 mm
which can be considered as negligible effect.
Since it uses a first order polinomial type regression model, the two level factorial experiments based DSA
Fig. 12 Design limitations for brake system and wishbone

Before the optimisation process, DSA was applied
to determine the effective and non-effective parameters
among the tie rod factors on the response of the system.
The two-level FFD type was chosen to generate the screen-
Fig. 14 Pareto chart of the standardised effects ( = 0.01)
409
Table 1
Initial and optimised hardpoint co-ordinates for tie rod
Factor, mm
A9x
A9y
A9z
A10x
A10y
A10z
Initial
-268
-795
-80
200
-100
0
shows the contour plots obtained from the software for the
interactions of the design parameters where, the first term
indicates the ordinate and the second is for abscissa.
100
zA8, (mm)
does not give information about the possible curvature

characteristic of the response. Hence, a second order polynomial model whose general form is given in Eq. (3) was
utilised. CCD type that is offered in the design specification table of Adams/Insight was used to determine the
optimum hardpoint locations of the tie rod. Total 50 runs
were generated for the CCD process by the software. Optimal hardpoint co-ordinates of the tie rod obtained from
Adams/Insight are compared with the initial model in
Table 1 (for wheel 1). Parallel wheel travel simulation example of the IFS [23] for zA8=100 mm is seen in
Fig. 15.
Initial
Optimised
-100
-2
-1
0
V, ()
Fig. 16 Comparison of the V deviation for initial and optimised designs
Optimised CCD
-308
-755
-55.25
240
-140
-42.4
c
Fig. 15 Parallel wheel travel simulation of the IFS: zA8=:
a - +100 mm; b - 0 mm; c - 100 mm
Deviation characteristics of initial and optimised
models are given in Fig. 16. Maximum V values were obtained for + 100 mm bump and -100 mm rebound of the
wheel as -0.26 and -0.32 respectively.
In order to evaluate the interaction effects of the
tie rod factors on V deviation, the design matrix obtained
from CCD was also imported to MINITAB. Fig. 17
Fig. 17 Contour plots for V

In order to obtain these plots, firstly, all of the
factors were adjusted to their optimal values (hold values)
obtained from CCD (Table 1). Then, only two of the factors were varied in the range of 40 mm. As can be clearly
seen from the contour plots, V deviates strongly along the
A9z and A10z axes. The response surface given in Fig. 18
which is identical to contour plot (A10z*A9z) also shows
that A9z has a greater effect on the V than A10z. This result
is compatible with the Pareto chart shown in Fig. 14.
410
ing iS ratio which can be defined as:
iS
2.5
V 1.5
-75
-50
-50
-25
-25
0
35
A9z
25
iS, ()
A10z
Fig. 18 V response as a function of A9z and A10z
Basic dimensions of the passenger bus used in

this study are seen in Table 2. F and MSE values were
obtained by taking these values as reference. Since a lower
range of L is used in most of the driving conditions of a
passenger bus, minimising the F for this range is acceptable [10, 11].
Table 2
Basic dimensions of the passenger bus (mm)
LV
3957
sRV
2096
15
L1 : 440 mm
5. Steering error optimisation
LF
6050
sRH
1825
j
1844
In this study, design criterion was chosen such as

maximum steering error which is defined as:
MSE max . La Li LaA Li
(13)
where, S is the steering wheel angle. In this study, possible range of iS was chosen in the range of 18-23 (-) due to
the manufacturers demand. Effect of relay lever length on
iS of the base mechanism can be seen as a function of S in
Fig. 20. L1 was chosen as 540 mm.
0.5
-75
S
,
L
-750
-500
-250
L2 : 540 mm
0
S, ()
250
L3 : 640 mm
500
750
Fig. 20 Effect of the relay lever length on iS

Design schematic of the relay lever-vehicle body
connection is illustrated upside down in Fig. 21. Here, the
longitudinal rails M are welded to the lateral rail N on
which the bearing of A14 is located. In order to find out the
appropriate design range for optimisation of the relay lever, a pre-study includes the SS option of Adams/Insight
was carried out. In this type of design study, the possible
range of a factor is determined by taking the design limitations into account. The chosen number of runs specifies
how the factor interval will be divided [21]. Results of the
runs give an estimation for the deviation characteristic of
the MSE20 response.
(12)
should not be higher than 0.5 in the range of Li = 20.

MSE20 is the maximum steering error obtained in
Li = 20 range. Fig. 19 shows installation region of the
steering system on the bus framework structure (area C)
and the base model used in the third stage of this work
respectively. In order to optimise MSE, an H-shaped planar
parallel arm mechanism which is consisted of two relay
levers and a track rod was chosen as the base model.
Fig. 21 Design detail of the vehicle body-relay lever connection
Fig. 19 Bus framework structure and the schematic

H-shaped base model for the optimisation study
In this model, initial position of the track rod was
selected in the middle of the relay lever. The length L of
the relay lever in x axis is directly affects the overall steer-
In this work, SS is performed in two successive

steps: Primary position of A10 was assumed as fixed in x-y
plane. Position of hardpoint A14 was changed in y axis in
the range of 50 mm. Dimensions of the rails on which
relay lever attached via A14 bearings were taken into account to determine this design limit. In Fig. 21, u represents the initial trial point of the SS1. Step size was chosen
as 5 mm for the total s1 = 100 mm design length. 21 trials
were generated. As can be seen from the convergence history of SS1 shown in Fig. 22, minimum value of MSE20
was obtained at trial 19 as MSE20 = 0.525. Here, trial 11
represents the base model. In the subsequent step, hard-
411
6
MSE20, ()
point A14 was adjusted to the minimum value obtained

from SS1. Then, the planar position of track rod was altered only in x axis in the range of 50 mm. Step size was
also chosen as 5 mm for the total s2 = 100 mm design
length. For SS2, 21 trials were generated where v represents the initial trial point. Convergence history of the SS2
shows that (Fig. 22) trial 10 provides an estimated optimum value of MSE20 which is calculated as 0.524. All of
the trials of SS1 and SS2 were analysed for S = + 420
which correponds to iS 21 and Li 20.
4
2
15
30
45
60
Iteration number
75
90
Fig. 23 Convergence history of CCD (estimations)

Table 3
Hardpoint co-ordinates of base and optimised models
Factor (mm)
A9x
A9y
A10x
A10y
A13x
A13y
A14x
A14y
Base
-308
-755
240
-140
-28.5
-140
-300
-140
Optimised CCD
-270.7
-746.49
-240.83
-140.3
-1.9
-140.41
-296.32
-182.77
Fig. 22 Convergence histories for: a - SS1; b - SS2

Since the main target of the SS1 and SS2 is to predict the reasonable design range of the hardpoints for the
multi-link mechanism, results obtained from these studies
are rough. These results were used to perform a final optimisation process using CCD. According to research [23],
angle (Fig. 11) also affects the steering error. Because of
this, A9x and A9y should also be taken into account for the
CCD study. Hence, A9, A10, A13 and A14 were chosen as
the design variables. On the other hand, in order not to
increase the V deviation during the wheel travel, all of the
factors cannot be used for the optimisation process. As can
be seen from the Pareto chart in Fig. 14, A10x and A10y
which have lower standardised effects than u = 2.665 can
be chosen as design variables for the relay lever. Because
of the planarity of the mechanism, A13z and A14z were assumed as equal to A10z obtained from CCD. Variation
range of the factors was chosen in x-y plane as 10 mm
except A9x and A13x. In order to shorten the design length
of the steering arm, design constraint for A9x was selected
as (0-40) mm. It is known from Pareto chart that this alteration does not have any remarkable effect on the V angle.
A13x range was also selected as (0- 30) mm to increase the
rigidity of the frame. By using CCD option of Adams/Insight 88 trials were generated by the software for
8 factors. Convergence history can be seen in Fig. 23
where minimum MSE20 was obtained as 0.2. By using
the optimal hardpoint co-ordinates obtained from
Adams/Insight which are compared with the base model
in Table 3, a final multibody model was composed and
analysed via Adams/Car.
Comparison of the base and optimal multi-link
model geometries is seen in Fig. 24. It was found that, for
the optimised model, angle between tie rod and steering
arm satisfies 15 condition. was obtained as 20
for Li = 55 which portrayes the extreme steering condition. For the optimal design, the angle between the line
|A10-A14| and the x axis was calculated as opt 4.5. MSE20
was achieved as 0.34.
Fig. 24 Comparison of the base and optimised geometries

Obtained minimum values and percent reductions
of MSE20 after every step of the optimisation study are
given in Table 4 and in Fig. 25 respectively. In Fig. 25, a.
the estimated MSE20 values for SS1, SS2 and CCD are
compared with F (Li) = 2.22 that is also obtained from
the base model for Li = 20.
Table 4
Estimated and obtained MSE20 values (Li 20)
Estimations
SS1
SS2
0.525 0.524
CCD
0.2
Final model
Design target
0.34
0.5
In order to calculate the percent reduction, MSE20

of the base model was assumed as 100%. As can be seen
from Fig. 25, a, MSE20 dramatically decreases after SS1
where the reduction was calculated as 80.1% in comparison with the base model. MSE reduction after SS2 and
CCD were also calculated as 0.04% and 12.3% respectively. Hence, it can be concluded that the most effective factor on MSE20 reduction is A14y or the angle. Additionally,
it should be noted that the calculated MSE20 obtained from
SS1, SS2 and CCD are the estimated values. MSE20 provided from the multibody analyses of base and final optimised
models are also given in Fig. 25, b. Total reduction of
MSE20 was obtained as 84.8%.
Ackermann error deviations of the base and the
optimised models as a function of Li are given in Fig. 26.
As also can be seen that MSE achieved from the final op-
412
MSE20 (%)
timised model is 0.46 at Li = 30 in the range of

Li = (0-44) which exceedingly satisfies the design target.
For instance, by using Eq. (1) Li is calculated as 33.14 for
R0min = 10.4 m, the minimum turning radius of the passenger bus and the dimensions given in Table 2. Steering error
of the base mechanism was obtained for Li = 30 as 4.4
which means the reduction by 89.6%.
19.9
19.9
15.2
7.6
Base
SS1
SS2
CCD
Base Optimal
F, ()
Fig. 25 Reduction of MSE20: a - estimations during the

optimisation stages; b - comparison of base and optimised models
5
4
3
2
1
0
-1
Base
Optimal
and A10z. For this design example, percent effects of A9z

and A10z were calculated as 44.71% and 31.76% respectively. CCD gives the optimum values of these coordinates more precisely than the FFD with a fewer number of design trials.
2. By using CCD, V deviation during the wheel
travel was reduced up to 85.4% in comparison with the
initial design. Maximum V values were obtained for +100
mm bump and -100 mm rebound of the wheel as -0.26 and
-0.32 respectively.
3. By optimizing the shape of the relay lever,
MSE of the initial parallel arm (H-shaped) base mechanism
was reduced up to 89.6% via SS1, SS2 and CCD in the
range of Li = 44. It is observed that the most effective
part of the estimation analyses on MSE reduction is SS1.
By this way it can be concluded that co-ordinate A14y (or
the angle ) has the greatest effect on MSE. In this design,
effect of SS2 can be assumed as negligible.
4. The obtained optimal mechanism satisfies the
design target of MSE 0.5 in the range of Li = 44.
Moreover, the mechanism can perform its function up to
Li = 55 without any lock up effect.
Acknowledgements
The authors are grateful for the suggestions of
Prof. N. S. Kuralay, Ph.D. of Dokuz Eyll University, Department of Mechanical Engineering. The licensed software support of Hexagon Studio is also acknowledged.
References
10
15
20
25
Li, ()
30
35
40
45
Fig. 26 Comparison of the F (Li) curves for base and optimised models
6. Conclusions
In this work, a DOE-RSM based design application to obtain a multi-link steering mechanism which gives
optimum deviation of V and steering error was developed
and applied on an MSC.Adams multibody model of a bus
IFS. In order to carry out the optimisation of the V, the
most effective parameters among the tie rod co-ordinates
on V deviation were first identified via DSA by using Adams/Insight multi-objective optimisation tool. The FFD
was used to determine the rank of importance of the coordinates of the tie rod hardpoints on V angle. Results
were evaluated by using MINITAB a practical statistical
software package. Since the FFD merely uses the high and
low values of the factors, it is not adequate to determine
the possible curvature of the response. In order to find out
the intermediate values of the parameters which give the
optimal tie rod position, CCD was also applied. In the final
stage of the study, geometry of the steering trapezoid
which gives the optimum MSE was determined via SS and
CCD. In order to do that the co-ordinate A10z was assumed
as the design constraint which determines the vertical position of the multi-link mechanism plane. Results obtained
from this study are summarised as follows:
1. Results of the DSA showed that for a multi-link
steering mechanism, the most effective factors among the
tie rod co-ordinates on V are the vertical components A9z
1. Timoney, E.; Timoney S. 2003. A review of the development of independent suspension for heavy vehicles, 2003 SAE International Truck and Bus Meeting
and Exhibition, SAE Technical Paper 2003-01-3433.
http://dx.doi.org/10.4271/2003-01-3433.
2. Reimpell, J. 1974. Fahrwerktechnik, Bd. 3, Wrzburg:
Vogel-Verlag, 177p.
3. Reimpell, J.; Stoll, H.; Betzler, J.W. 2002. The Automotive Chassis: Engineering Principles, Warrendale,
PA.: Society of Automotive Engineers, Inc.
4. Simionescu, P.A.; Beale, D. 2002. Optimum synthesis
of the four-bar function generator in its symmetric embodiment: the Ackermann steering linkage, Mechanism
and Machine Theory 37(12): 1487-1504.
http://dx.doi.org/10.1016/S0094-114X(02)00071-X.
5. Zhou, B.; Li, D.; Yang, F. 2009. Optimization design
of steering linkage in independent suspension based on
genetic algorithm, IEEE 10th International Conference
on Computer-Aided Industrial Design & Conceptual
Design: CAID&CD. Wenzhou, 45-48p.
http://dx.doi.org/10.1109/CAIDCD.2009.5374895.
6. Hanzaki, A.R.; Rao, P.V.M.; Saha, S.K. 2009. Kinematic and sensitivity analysis and optimization of planar rack-and-pinion steering linkages, Mechanism and
Machine Theory 44(1) 42-56.
http://dx.doi.org/10.1016/j.mechmachtheory.2008.02.0
14.
7. Oz, Y.; Ozan, B.; Uyanik, E. 2012. Steering system
optimization of a Ford heavy-commercial vehicle using
kinematic & compliance analysis, SAE Technical Paper 2012-01-1937.
http://dx.doi.org/10.4271/2012-01-1937.
413
8. Liang J.; Xin, L. 2012. Simulation analysis and optimization design of front suspension based on ADAMS,
Mechanika 18(3): 337-340.
http://dx.doi.org/10.5755/j01.mech.18.3.1873.
9. Kim, B.; M., Kim, J.W.; Moon, I.D.; Oh, C.Y. 2014.
Optimal combination of design parameters for improving the kinematics characteristics of a midsize truck
through design of experiment, Journal of Mechanical
Science and Technology 28(3): 963-969.
http://dx.doi.org/10.1007/s12206-013-1167-7.
10. Bian, X.L.; Song, B.A.; Becker, W. 2003. The optimisation design of the McPherson strut and steering
mechanism for automobiles, Forschung im Ingenieurwesen 68(1): 60-65.
http://dx.doi.org/10.1007/s10010-003-0107-6.
11. Bian, X.L.; Song, B.A.; Walter, R. 2004. Optimization of steering linkage and double wishbone suspension via R-W multibody dynamic analysis, Forschung
im Ingenieurwesen 69(1): 38-43.
http://dx.doi.org/10.1007/s10010-004-0136-9.
12. Minitab User's Guide 1: Data Graphics and Macros
(Release 13 for Windows), 1999. Pennsylvania State
University, USA.
13. MSC.Adams. 2002. Product Catalog. MSC. Software
Corporation.
14. Blundell, M.; Harty, D. 2006. The Multibody Systems
Approach to Vehicle Dynamics, London: Elsevier Butterworth Heinemann, 172p.
15. Matschinsky, W. 2007. Radfhrungen der Straenfahrzeuge. 3. aktualisierte und erweiterte Auflage, Berlin Heidelberg: Springer-Verlag, 96, 97p.
16. Kuiper, E.; Van Ooster, J.J.M. 2007. The PAC2002
advanced handling tire model, Vehicle System Dynamics: International Journal of Vehicle Mechanics and
Mobility, 45(1): 153-167.
http://dx.doi.org/ 10.1080/00423110701773893.
17. Montgomery, D.C. 2000. Design and analysis of experiments. Fifth Edition, Hoboken, New Jersey: John
Wiley & Sons, Inc. 275p.
18. Park, K.; Heo, S.J.; Kang, D.O.; Jeong, J.I.; Yi,
J.H.; Lee, J.H.; Kim, K.W. 2013. Robust design optimization of suspension system considering steering
pull reduction, International Journal of Automotive
Technology 14(6): 927-933.
http://dx.doi.org/10.1007/s12239-013-0102-3.
19. Han, H.; Park, T. 2004. Robust optimal design of
multi-body systems, Multibody system dynamics
11(2): 167-183.
http://dx.doi.org/10.1023/B:MUBO.0000025414.28789
.34.
20. Myers, R.H.; Montgomery, D.C.; Anderson-Cook,
C.M. 2009. Response Surface Methodology: Process
and Product Optimization Using Design of Experiments, Third Edition. Hoboken, New Jersey: John
Wiley & Sons, Inc. 704 p.

21. MSC.ADAMS. 2013. ADAMS/Insight User Guide.
MSC. Software Corporation.
22. Aydn, M.; nlsoy, S. 2012. Optimization of suspension parameters to improve impact harshness of road
vehicles, The International Journal of Advanced Manufacturing Technology 60(5-8): 743-754.
http://dx.doi.org/10.1007/s00170-011-3589-7.
23. Genta, G. 1997. Motor Vehicle Dynamics, Singapore:
World Scientific Publishing Co. Pte. Ltd., Regal Pres
(S) Pte. Ltd. 216p.
M. M. Topa, U. Deryal, E. Bahar, G. Yavuz
OPTIMAL KINEMATIC DESIGN OF A MULTI-LINK
STEERING SYSTEM FOR A BUS INDEPENDENT
SUSPENSION: AN APPLICATION OF RESPONSE
SURFACE METHODOLOGY
Summary
A response surface-based design application to
obtain an optimum multi-link steering mechanism is presented. Design problem is essentially established on two
main goals: minimum deviation of toe angle (V) during
the wheel travel and optimum steering error during the
steering angle (L) range of the wheel. In the first stage, a
complete multibody model of the suspension system including the steering mechanism was composed by using
MSC.Adams software. In order to identify the most effective parameters among the tie rod co-ordinates on V deviation, a Full Factorial Design (FFD) - based Design Sensitivity Analysis (DSA) was carried out via Adams/Insight
multi-objective optimisation tool. Central Composite Design (CCD) was also implemented to find out the optimum
position of the tie rod. In the final stage, optimum hardpoint positions of the steering mechanism were searched
by a combination of sweep study (SS) and CCD to provide
the minimum deviation of Ackermann error. The optimisation results show that it is possible to reduce the maximum
steering error (MSE) of the system up to 89.6 % in comparison with the parallel arm base mechanism by using the
proposed methodology.
Keywords: Ackermann steering, central composite design
(CCD), design of experiments (DOE), independent front
suspension (IFS), maximum steering error (MSE), multilink steering system, optimisation, regression, response
surface methodology (RSM).
Received April 17, 2015
Accepted June 23, 2015
ARTICLE IN PRESS
Robotics and Computer-Integrated Manufacturing 24 (2008) 239248

www.elsevier.com/locate/rcim
Parametric design optimization of 2-DOF RR planar manipulator

A design of experiment approach
B.K. Rout, R.K. Mittal
Mechanical Engineering Group, Birla Institute of Technology and Science, Pilani 333031, Rajasthan, India
Received 20 February 2005; received in revised form 5 September 2006; accepted 10 October 2006
Abstract
This work illustrates simulation approach for optimizing the parametric design and performance of a 2-DOF RR planar
manipulator. Using dynamic and kinematic models of a manipulator different performance measures for the manipulator are obtained
for different combination of parameters with effect of noise incorporated to imitate the real time performance of the manipulator.
A novel approach has been proposed to model, the otherwise difcult to model, noise effects. The data generated during simulation for
various parameter combinations are utilized to analyze the statistical signicance of kinematic and dynamic parameters on performance
of manipulator using ANOVA technique. The parameter combinations, which give optimum performance measures obtained for
different points in workspace, are compared and reported.
r 2006 Elsevier Ltd. All rights reserved.
Keywords: Parametric design; Control factors; Noise factors; Positional error; Reliability
1. Introduction
Optimal design of manipulators, whose importance in
manufacturing is increasing day by day, has been a
challenge. Little success has been obtained in this area
because of difculties associated with geometrical constraints and complex models of the manipulator. Mostly
industrial manipulators are required to perform tasks with
a higher precision and speed than human beings. To
perform a task, a robot is commanded to move its endeffector to a specied position but the actual position
reached may be quite different from the desired one. This
difference in the actual and desired position for the endeffector is termed as positional error of a manipulator and
the average precision with which the manipulator moves its
arm to the commanded position is termed as its positional
accuracy [1]. The positional error for an industrial
manipulator may be 0.1 mm and repeatability as high as
10 mm [2]. Some studies relevant to the stochastic analysis
Corresponding author. Tel.: +91 1596 245073 225.
E-mail addresses: rout@bits-pilani.ac.in (B.K. Rout),

rkm@bits-pilani.ac.in (R.K. Mittal).
0736-5845/$ - see front matter r 2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.rcim.2006.10.008
of positional error have been carried out. An overview of

existing work on robot calibration and identied basic
issues for robot precision was presented by Roth et al. [3] in
1987 but many developments have taken place since then.
Whitney et al. [4] proposed a method for improving
orientation and/or location accuracy of a programmable
robot with respect to a target object. Their method consists
of calibrating the position of a terminal control frame
associated with a robot end-effector which is coupled to a
robot distal link. Benhabib et al. [5] introduced direct and
inverse error analyses while Jang et al. [6] had reported two
types of positional errors, geometric and non-geometric for
industrial robots. Geometric errors result from imprecise
manufacturing and non-geometric errors result from
gravity, joint compliances and gear transmission errors.
Azadivar [7] investigated the effect of joint position errors
on accuracy of performance in a process of inserting a pin
into a hole. Bhatti and Rao [8] presented a concept of
reliability as a probabilistic measure of manipulator
performance. Menq and Borm [9] proposed ve error
measure indices to show the quantitative distribution of the
positional error. Manoochehri and Seireg [10] developed
a computer programme for form synthesis and optimal
ARTICLE IN PRESS
240
B.K. Rout, R.K. Mittal / Robotics and Computer-Integrated Manufacturing 24 (2008) 239248
design of robot manipulator. The optimization strategy for

synthesis and control utilized a dynamic programming
approach that made it possible to select the optimum
parameters in stage-wise manner without sacricing the
interactions inherent in the highly coupled nonlinear
robotic systems. Offodile and Ugwu [11] investigated effect
of various process variables such as speed of the tool and
payload on the robot repeatability. They observed experimentally that speed of robot travel in work envelope has
dominating effects on its repeatability.
Approaches incorporating statistical method in design
exist but few have been included here. Parkinson et al. [12]
and Chen et al. [13] indicated that the early design phase of
a product has greatest impact on life cycle and its quality of
performance. Welch et al. [14] observed that presence of
systematic error rather than random error in statistical
testing is inappropriate. Chen et al. [15] proposed robust
concept exploration method, which facilitates quick
evaluation of different design alternatives and helps in
generation of top-level design specication in the early
stages of design. The concept of robust design was applied,
to the design of a high-speed civil transport by Chen et al.
[16], to the design of family of general aviation aircraft by
Simpson et al. [17], to manufacturing simulation by
Peplinski et al. [18] and to the design of a turbine lift
engine by Koch et al. [19,20]. Guinta et al. [21] used
metamodelling techniques for deterministic computer
experiments containing numerical noise. An extensive
review of the methodologies to obtain a robust design that
has less performance variation due to the variations of
control factors and noise factors is given by Rout and
Mittal [22]. The Taguchi method has been applied by Rout
and Mittal [23] for nding the optimum parameters for
reduction of performance variation, thereby increasing
positional accuracy of a manipulator. Rout and Mittal [24]
investigated the statistical signicance of manipulator
kinematic parameters using design of experiment approach. Parametric tolerance design of a manipulator
using full factorial design of experiment approach has been
attempted by Rout and Mittal [25] without considering
noise factor. Lastly, Rout and Mittal [26] had attempted to
design and optimize the performance of 2-degree of
freedom manipulator using articial neural network
technique.
As it is very tedious, time consuming and uneconomical
to build large number of prototypes and conduct physical
experiments on manipulators by varying values of factors
which can easily be changed by designer, i.e. control
factors, a novel simulation method is proposed using
control factors. Developed methodology helps in incorporating effect of noises in dynamic model of manipulator to
have real time performance. Current work focuses on
parametric design of manipulator using design of experiment (DOE) approach, to select a combination of control
factors of a product or process in such a way that its
performance will become insensitive to noise factors, that
are difcult to control. DOE techniques are fairly standard
approaches and commonly used in statistical quality

control, but its application to robust robotic parameter
design is rare. Experiments conducted using above
technique helps in studying the combined effect of factors
on performance. It allows the effect of a factor to be
estimated at several levels of the other factors yielding
conclusions that are valid over a range of experimental
conditions [30]. This paper illustrates procedure to apply
DOE technique to manipulator parameter design, to
identify parameters responsible for performance variation
and to nd optimal combination, which deliver optimal
performance. Discussed study is an ofine strategy, which
is novel and helps designer to select the parameters
beforehand to reduce the performance variation, prior to
actual manufacturing is carried out.
The rest of this paper is organized in six sections. In
Section 2 steps required to apply DOE technique to
manipulator design is discussed. In Section 3, developed
novel simulation method to incorporate effect of noise
factors and a search technique for simulating the performance is discussed. Data used for simulation and analysis
of results of experiment are discussed in Sections 4 and 5.
Finally, the investigation is concluded in Section 6.
2. Application of DOE technique to manipulator design

The robot parameters like its conguration, link dimensions, inertia, and actuators, etc. play vital role in its
performance. Robotic system designer is required to make
decisions regarding these parameters. Except in few specic
applications, designer chooses a particular parameter or
particular parameter combination by convenience, as no
tools are available overlooking alternative solutions
that may give optimal performance. By using the DOE
approach for parametric design, it is possible to nd the
best combination of parameters for an optimal performance of the robotic manipulator. The DOE technique has
been used successfully to optimize processes and designs
for diverse non-robotic systems. A technique for application of DOE approach to a manipulator parameter design
has been developed and presented next. A 2-DOF RR
planar manipulator is considered as an example to
establish the application of parameter design technique to
robot manipulators.
2.1. Statement of the problem

The main function for a robot manipulator is to
accurately reach the commanded position. For the
2-DOF RR planar manipulator, the target position is in
the work-plane of the manipulator and is described by
point Px; y assuming xy-plane is the work-plane of the
manipulator. The mathematical model to simulate the
performance of manipulator and compute the position
reached is presented in Section 2.1.1.
ARTICLE IN PRESS
2.1.1. Kinematic and dynamic models of 2-DOF RR planar

manipulator
The kinematic and dynamic models of 2-DOF RR
planar manipulator are developed considering the manipulator shown in Fig. 1. The two links have lengths l1, l2,
respectively. Let Px; y be the target position of endeffector in the Cartesian workspace of the manipulator, the
point of interest, for which the performance in terms of
positional accuracy has to be modeled and optimized. The
coordinates of P in Cartesian coordinate for joint angles
y1 and y2 are given by [27]
x l 1 C 1 l 2 C 12 ,
(1)
y l 1 S 1 l 2 S 12 ,
(2)
where C i cos yi ; Si sin yi ; C ij cosyi yj and Sij

sinyi yj with i, j 1, 2 for the two links, link 1 and link
2, respectively.
From Eqs. (1) and (2), joint variables y1 and y2 are
obtained as

1 yl 1 l 2 C 2 xl 2 S 2
y1 tan
(3)
xl 1 l 2 C 2 yl 2 S 2
and
y2 cos
1

x2 y2 l 21 l 22
.
2l 1 l 2
(4)
Differentiating Eqs. (1) and (2) and solving for joint

velocities y_ 1 and y_ 2 gives
_ 12
_ 12 yS
xC
y_ 1
l 1 S2
(5)
and
x_ x y_ y
y_ 2
,
l 1 l 2 S2
Assuming link masses are m1, m2 and joint torques are

t1 ; t2 , respectively, and links as rigid members with mass
concentrated at center of gravity, the dynamic model of
2-DOF RR planar manipulator is [27]
hm

i
m2 2
1
m2 l 21
l 2 m2 l 1 l 2 C 2 y 1
t1
3
3
2

l2 l1
l 2 C 2 y 2 m2 l 1 l 2 S2 y_ 1 y_ 2
m2
3 2
m

m2
m2
2
1
l 1 l 2 S2 y_ 2
m2 gl 1 C 1
gl 2 C 12 ,

7
2
2
2
hm
i
m2
m2 2
2 2
t2
l2
l 1 l 2 C 2 y 1
l y2
3
2
3 2
m2
m2
2
l 1 l 2 S2 y_ 1
gl 2 C 12 .
8
2
2
Eqs. (7) and (8) are solved for y 1 and y 2 and are written in a
compact form as
be cd
y 1
ad b2
and
bc ae
y 2
,
ad b2
where a; b; c; d and e; are given as
m

1
1
m2 l 21 m2 l 22 m2 l 1 l 2 C 2 ,
a
3
3
2

l
1
b m2 2 l 1 l 2 C 2 ,
3 2
m

m2
1
m2 gl 1 C 1
gl 2 C 12 m2 l 1 l 2 S2 y_ 1 y_ 2
c
2
2
m2
2
l 1 l 2 S 2 y_ 2 t1 ,

2
(6)
d
_ y
_ represent end-effector velocity ~
where x;
ve with
x_ ve cos a and y_ ve sin a, and a is angle made by ~
ve
with positive x-axis of base frame.
vey
ve
vex
241
m2 2
l ,
3 2
(9)
(10)
(11)
(12)
13
(14)
m2
m2
2
l 1 l 2 S2 y_ 1
gl 2 C 12 t2 .
(15)
2
2
Eqs. (1)(15) are used to identify signicant factors
and compute the performance measures for the robotic
manipulator.
2.2. Identification of factors and levels
P(x,y)
m2
l2
Y
2
m1
1+2
l1
Various parameters inuencing the working of manipulator are identied with the help of a parameter diagram
(P-Diagram) for manipulator as shown in Fig. 2. The
parameters other than input and output are classied as
control factor (CF) and noise factor (NF).
1
X
Fig. 1. 2-DOF RR planar manipulator and its parameters.
2.2.1. The control factors

A control factor is a product or process parameter whose
values can be selected and controlled by the design or
manufacturing engineer. The CFs for manipulator are
identied from the kinematic and dynamic equations.
ARTICLE IN PRESS
242
Noise Factors
Environmental Conditions,
Eletrical Noise, Joint Friction,
Manufacturing and Assembly Errors
Input
Command for
work to be done
Response
Positional Error
Robot
Control Factors
Link Length,
Link Mass, Joint Torque
Fig. 2. Parameter diagram for robot performance.
From Eqs. (9) and (10), it is observed that y 1 and y 2 depend

on following six independent parameters apart from the
variables y1 ; y2 ; y_ 1 and y_ 2 :
1.
2.
3.
4.
5.
6.
length of link 1, l1;

length of link 2, l2;
mass of link 1, m1;
mass of link 2, m2;
torque applied at joint 1, t1 ;
torque applied at joint 2, t2 ;
These six parameters are identied as CFs because

designer can choose their values for optimal performance.
For design optimization using DOE technique, parameters can be considered at several levels. In this study,
control factors are considered at two levels: high and low.
Therefore, six control factors at two levels, results in a set
of 26 (64) different combinations of six control factors.
2.2.2. Sources of noise in the manipulatorthe noise factors
The function of the manipulator is to move its endeffector to a desired point accurately. However, the
discrepancy exists between the desired and actual point
reached because of noise factors. The NFs cause the endeffector to deviate from its target point. NFs are those
factors that are difcult, expensive, or hard to control
during production or operation. Some of the NFs that have
direct inuence on the performance of the manipulator are:
(a) environmental conditions in which manipulator operates,
(b) errors in manufacture and assembly,
(c) uctuations in electricity supply, causing deviation in
the joint actuator torque,
(d) friction at the manipulator joints, and
(e) joint compliance between the joint encoder and the
actual angular output.
There can be several other noise factors, which may have
inuence on performance of the manipulator. As these
noise factors are very difcult to quantify, their effects on

performance are difcult to compute, a novel approach has
been proposed to include the effects of NFs in the DOE to
obtain a robust design of robotic manipulator. This
approach is explained in Section 2.3.
2.3. Selection of performance measures for manipulator
the response variable
To investigate the impact of different parameters on
performance variation of manipulator, several performance measures have been proposed by researchers.
A performance measure for a manipulator is a parameter
dened on the space of all postures of the manipulator that
measures some general property of the manipulator that
allows choosing the optimal combination of parameters
[3,4]. For a robotic manipulator positional performance
parameters are accuracy, repeatability, and resolution.
A combination of these parameters is dened as positional
error ei as distance between the actual point reached by the
end-effector Pi xi ; yi ; zi in the ith experiment and desired
point Px; y; z, that is

1=2
.
(16)
i xi x2 yi y2 zi z2
The following four performance measures have been
considered in this work.
(a) Positional error: For the 2-DOF RR planar
manipulator with xy-plane as the workplane positional
error i from Eq. (16) will reduce to

1=2
.
(17)
i xi x2 yi y2
The objective function to be optimized is to minimize the
positional error considering the uncertainty.
(b) Mean positional error: Mean positional error is
dened as

n
1X
i ,
n i1
(18)
ARTICLE IN PRESS
where n is number of experiments or replications and i is

positional error for ith experiment.
(c) Signal to noise ratio: The signal to noise (SN) ratio
proposed by Taguchi, is used as the data transformation
method to consolidate the repetitive data into one value,
which reects the mean value and amount of variation
present in the data. For the robotic manipulator, the design
objective is to minimize the positional error hence, it is
always desired that it should be as small as possible.
Therefore, as per Taguchi method, the target is of smallerthe-better type and SN ratio for the smaller-the-better case
is given as [28]
!
n
1X
SN 10 log10
2
(19)
n i1 i
SN ratio is an essential indicator of the ability of the system
to perform well in relation to the effect of noise and
measure to carry out the analysis of experiment.
(d) Reliability: The reliability (R) of a manipulator is
dened by Bhatti and Rao [8] as the probability of the
end-effector reaching a point or in a close vicinity of it
within specied range. If the end-effector reaches a point
within the specied range, it is considered as a successful
experiment. The reliability R is given as
R
Number of successful experiments s

.
Total number of experiments
n
(20)
The specied range around a target point is called the

permissible error region and its shape and size depends on
the intended use of the manipulator. The reliability as
performance measure has been used to evaluate the overall
performance of all control factor combinations. The results
obtained are used for validating the outcome of DOE
technique to manipulator design.
2.4. Design the experimentchoice of experimental design
Experiment is dened as a test in which purposeful
changes are made to the input variables of a system
so that reasons for changes in the output response observed
can be identied. There are several factors of interest
in an experiment therefore to deal with these factors a
factorial experiment is used. Factorial experiment is an
experimental strategy in which design factors are varied
together and effects of all possible combinations of the
levels of factors for the experiment are investigated. For a
general case where there are a levels of factor A, b levels of
factor B, c levels of factor C, arranged in a factorial
experiment, there will be a b c n total observations,
where n is number of replications of the complete
experiment. One of the special cases is that of k number
of factors, each at two levels such a design requires 2k
observations and is called a 2k factorial design. Hence,
factorial design of six control factors, each at two levels
leads to 26 runs and design matrix of these factors, are
presented in Appendix A.
243
2.5. Performing and analyzing the experiment

To conduct the experiments, strategies adopted to
simulate performances are discussed below in Section 3.
To investigate the effect of parametric variation on
performance of manipulator, an experiment is conducted
with the help of design matrix. Each factorial combination
is run for nite number of replications to capture the effect
of noise. The developed methodology and data utilized to
simulate the performance is provided in Sections 4 and 5.
It is important to mention that while simulating the
performance for the experiment no constraints such as
maximum velocity and acceleration of links of manipulator
are imposed.
From each replication positional error as outcome of the
experiment is obtained and thereafter the performance
measure SN ratio and mean positional error are computed
for each CF combinations. For parameter optimization
problem, optimal CF combination should have lowest
performance variation and nearer to target mean performance. When both the conditions are satised SN ratio
becomes highest. Therefore to obtain optimal parameter
combination, SN ratio values are compared after the
conduct of experiment.
To compute reliability of a particular parameter
combination, simulation is conducted for more number
of cycles. The Cartesian coordinates obtained from each
simulation determine whether the point lies within the
specied region or not. The reliability of the parameter
combination is computed using Eq. (20).
Statistical analyses of performance of experiment are
carried out using analysis of variance (ANOVA) technique,
which is a powerful tool for understanding complex
physical phenomenon. ANOVA is used to subdivide the
total variation into variation due to main factors, variation
due to interacting factors and variation due to error. After
this statistical tests like F-test are used to study statistically
signicant CF and interacting factors, which helps in
screening many factors to discover the vital few and how
they interact. For current study statistically signicant CF
and interacting factors are determined using ANOVA, and
its results are discussed in Section 6.
3. Strategy to incorporate effects of noise for parameter
design
For given set of control factor values and target point
Px; y, following six parameters are computed from the
kinematic and dynamic model equations (3)(6) and
(9)(10)
1. angular displacements y1 , y2 ,
2. angular velocities y_ 1 , y_ 2 ,
3. angular accelerations y 1 , y 2 .
The computed values of these six parameters are free
from the effects of noise. In order to incorporate the effect
ARTICLE IN PRESS
244
of noise in the above parameters, the effect of noise, in

form of individual errors for the six control factors is
generated randomly. The randomly generated errors are
assumed to follow normal distribution with zero mean
and specied standard deviation. Using set of values of
control factors with noise, the above six parameters with
noise incorporated are obtained as: angular displacements
n
n
(yn1 , yn2 ) from Eqs. (3) and (4); angular velocities (y_ 1 , y_ 2 )
n
n
from Eqs. (5) and (6); and angular accelerations (y 1 , y 2 )
from Eqs. (9) and (10). To compute the point actually
reached by end-effector with the presence of noise in
control factors, a search technique has been developed and
is described in Section 3.1.
3.1. Search technique
n
n
Taking computed y 1 , y 2 as input and control factors at
nominal values (without noise) the angular displacements
of links ya1 and ya2 are obtained using Eqs. (9) and (10),
where superscript a indicates actual values with noise
present. Since Eqs. (9) and (10) are nonlinear, transcedental
equations, the values of ya1 and ya2 are obtained using a
heuristic search algorithm.
The steps of the algorithm are given below:
Algorithm 1. Search algorithm

Step 1: Read nominal level values of six control factors
l1, l2, m1, m2, t1 and t2 .
Step 2: Read standard deviations sl 1 ; sl 2 ; sm1 ; sm2 ; st1 and
st2 for the six control factors.
Step 3: Read the range and step size for y1 , y2 variations
and permissible error e.
Step 4: Read the manipulator target point Px; y.
Step 5: Obtain y1 , y2 , y_ 1 , y_ 2 , y 1 and y 2 for input (nominal
values) of control factors from Eqs. (3)(6) and
(9)(10).
Step 6: Generate random errors based on standard
deviations for six control factors and obtain
control factor values with noise. Using these
n
n
values of control factors, compute yn1 , yn2 , y_ 1 , y_ 2 ,
n
n
y 1 , and y 2 from Eqs. (3)(6) and (9)(10).
Step 7: Make starting guess for ya1 and ya2 , which is equal
to y1 and y2 in Step 5
Step 8: Compute y a ; y a from Eqs. (9) and (10).
1
2
Step 9: Compare the values of y a ; y a obtained in Step 8
1 2
n n
with values of y 1 ; y 2 obtained in Step 6. If
differences are with in the specied permissible
error e go to Step 10, else increment ya1 and ya2 for
chosen step size within the range and go to Step 8.
Step 10: Terminate the search and return ya1 and ya2 .
For each factorial combination simulations are performed to obtain the individual performance measure of
the experiment. The simulation is also run for required
number of replications to compute combined performance
measure. Likewise for other factorial combinations and

replications, simulations are carried out to compute
performance measure discussed.
4. Simulation
The numerical values used to simulate the performance
are given below:
(a) Number of levels for each control factor 2.
(b) Nominal values of six control factors at two levels and
standard deviations are given in Table 1.
(c) Number of combinations in factorial design 26 64.
(d) Design matrix containing 64 CF combinations is
provided in Appendix A.
(e) Chosen number of replications for each combination 8
(for factorial design) 100 (for reliability).
(f) Coordinates of target point in workspace: case I: x
0:40 m; y 0:30 m and case II: x 0:50 m, y 0:40 m.
(g) The step size of increment for search, range of search
and permissible error value e for the search algorithm
were chosen as: y1incr 0:01y1 , y2incr 0:01y2 0:5 y1
py1 p1:5 y1 and 0:5 y2 py2 p1:5 y2 , e 0:05.
(h) Tolerance selected for target point for computation of
reliability Dx 0:0005 m; Dy 0:0005 m:
Using the above numerical values simulation were
carried out. To simulate the performance for each factor
combination is run for eight replications for cases I and II.
The simulated performances are analyzed for statistical
analysis and its results are provided in Tables 2 and 3,
respectively. The performance measures, i.e. positional
error and SN ratio are computed for each parameter
combination and its trend are displayed in Figs. 3(a) and
3(b) for case I and Figs. 4(a) and 4(b) for case II. To
validate the results of factorial design, performance
measure reliability is computed after running the simulation for 100 cycles. The results for each combination are
displayed in Figs. 3(c) and 4(c) for the two cases,
respectively. After comparing the values of performance
measure, optimal combinations having optimal performance are presented in Tables 4 and 5.
Table 1
Values for control factors at two different levels
Control factors
Low level
High level
Standard deviation
l1 (m)
l2 (m)
m1 (kg)
m2 (kg)
t1 (N m)
t2 (N m)
0.40
0.25
5.5
4.0
500
100
0.50
0.35
6.5
5.0
800
105
0.0001
0.0001
0.01
0.01
0.1
0.1
ARTICLE IN PRESS
Table 2
Analysis of variance of factorial design for case I
Source
Sum of
squares
DF
11.52
1
l1
47.33
1
l2
m1
3.154 105 1
m2
30.19
1
1.02
1
t1
t2
2.56
1
l1l2
1.13
1
m1 t1
3.53
1
4.03
1
m2 t1
l 1 l 2 m1 t1 t2
3.56
1
Residual
385.27
478
Corrected 506.35
511
sum total
Mean
square
F-value
Remark
11.52
3.67
Signicant
47.33
58.72
Signicant
3.154 105 3.913 105
30.19
37.46
Signicant
1.02
1.26
2.56
3.18
1.13
1.040
3.53
4.38
Signicant
4.03
5.00
Signicant
3.56
4.41
Signicant
0.81
Table 3
Analysis of variance of factorial design for case II
Source
Sum of
squares
DF
Mean
square
F-value
Remark
l1
l2
m2
t1
l1l2
l 1 t1
l 2 t1
l 1 l 2 t1
Residual
Corrected
sum total
193.54
415.76
44.59
39.04
47.08
147.30
11.03
80.47
2106.19
3085.01
1
1
1
1
1
1
1
1
503
511
193.54
415.76
44.59
39.04
47.08
147.30
11.03
80.47
4.19
46.22
99.29
10.65
9.32
11.24
35.18
2.63
19.22
Signicant
Signicant
Signicant
Signicant
Signicant
Signicant
Signicant
5. Analysis of performance results

Analysis of the experimental results has been done using
ANOVA technique and for computational work Design
Expert (DX) Version-6 software [29] is used. Since each
control factor has two levels, 8 replications of six-factor
experiment required 512 simulations to obtain performance
measure. Analysis of test cases mentioned is carried out
and results are discussed below.
(i) The performance measure utilized for analysis of
experiment is positional error i . For test case I, it
is observed that control factors l 1 ; l 2 ; m2 ; and interacting factors m1 t1 ; m2 t1 ; l 1 l 2 m1 t1 t2 are statistically signicant by comparing computed F statistic value with
tabulated F statistic value for given factor and
interaction factor combination. The results of ANOVA are presented in Table 2 and indicate that all
control factors have statistically signicant role to play
in performance measure apart from the interactions.
245
None of the control factors go out of contention for

parameter design. Observing this SN ratio and
reliability, are used to nd the suitable control factor
combinations for which performance is optimal. For
case II, it is observed that control factors l 1 ; l 2 ; m2 ; t1
and interacting factors l 1 l 2 ; l 1 t1 ;l 1 l 2 t1 are signicant.
The results of ANOVA are given in Table 3.
(ii) From the simulation results for mean positional error,
SN ratio and reliability, for case I, in Figs. 3(a), 3(b)
and 3(c), respectively, for 64 parameter combinations
optimal parameters are selected. The basis for selection
of optimum parameter combination is that it should
have maximum value of SN ratio and reliability and
minimum mean positional error. The optimum parameter combination using above performance measure
are shown in Table 4. Similarly, from simulation of
performance measures for case II as given in Figs. 4(a),
4(b) and 4(c), respectively, for 64 parameter combinations, the optimum parameter combinations are
identied and presented in Table 5.
(iii) For case I, it is observed from Fig. 3(a) that mean
positional error is maximum 2:2003 102 m at
combination number 42 and is minimum 0:07355
102 m at combination number 29. For case II mean
positional error is maximum 6:685 102 m at
combination number 4 and is minimum 0:88964
102 m at combination number 55 as observed in
Fig. 4(a). It is also observed that mean positional
error is less than 2 102 m from combination
numbers 47 to 64.
It is observed that statistically signicant factors are
different for different target points in workspace indicating
that different factors have different contribution to
performance variation as target position changes. It is
important to notice that statistically signicant factors are
major contributor to performance variation of manipulator and even there is mathematical relationship and all
parameters are utilized in simulating the performance. In
addition to this optimum combination of control factors
required to perform task are different for both the cases,
which indicate that one set of parameters will behave
differently for different type of task. It is seen that SN
ratios of different combinations are increasing from
combination no. 21st to 33rd and 53rd to 64th and
maximum reliability appear at 16th and 33rd combination
number of control factors for case I. Except few, all
combination show poorer performance for case II as
compared to case I.
The selection of suitable tolerance range for reliability
computation is done by trial and error, because it is
observed that a wider tolerance range gives higher
reliability and a tighter tolerance gives poorer reliability
for all factor combinations. Finally, the optimum factor
combinations obtained using SN ratio and reliability do
not agree in the both the cases but trends and performance
measure peaks observed for different combination numbers
ARTICLE IN PRESS
Mean positinal error in

cm
246
2.5
2
1.5
1
0.5
0
1
7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64
Combination number
2
64
61
58
55
52
49
46
43
40
37
34
31
28
25
22
19
16
13
10
S/N ratio
0
-2
-4
-6
-8
-10
Combination number
0.08
Reliability
0.06
0.04
0.02
61
57
53
49
45
41
37
33
29
25
21
17
13
Combination number
Fig. 3. Performance measures for case I: (a) mean positional error; (b) SN ratio; (c) reliability.
8
6
4
2
Combination number
58
55
46
43
40
37
34
31
28
25
22
19
16
13
10
0
1
-10
-15
0.08
Reliability
-20
0.06
Combination number
0.04
0.02
Combination number
Fig. 4. Performance measures for case II.
64
61
58
55
52
49
46
43
40
37
34
31
28
22
19
16
13
10
0
25
S/N ratio
-5
64
61
55
52
58
52
49
49
46
43
40
37
34
31
28
25
22
19
16
13
10
0
1
Mean positional error in

cm
ARTICLE IN PRESS
Table 4
Optimum parameters for different performance measure for case I
Control factor
Target at x 0.40 m, y 0.30 m

SN ratio
Value
(combination no.)
l1 (m)
l2 (m)
m1 (kg)
m2 (kg)
t1 (N m)
t2 (N m)
Reliability
Mean positional
error (m)
0.446933 (29)
0.071 (49)
0.00739 (29)
0.40
0.35
6.5
5
500
105
0.50
0.35
5.5
4
500
100
0.40
0.35
6.5
5
500
105
Table 5
Optimum parameters for different performance measure for case II
Control factor
Target at x 0.50 m, y 0.40 m

SN ratio
Reliability
Mean positional
error (m)
Value
(combination no.)
1.66023 (48)
0.072 (22)
0.00889 (55)
l1 (m)
l2 (m)
m1 (kg)
m2 (kg)
t1 (N m)
t2 (N m)
0.50
0.25
6.5
5
800
105
0.40
0.35
5.5
5
500
105
0.50
0.35
5.5
5
800
100
are equally comparable in both the cases. Possible reason

for disagreement in optimal solution could be due to
the transformation of positional error into the Taguchis
SN ratio which may not be same as the untransformed
result obtained from reliability computation. It is
observed that most of the combinations for cases I and II
show poor simulation performance, possible reason
may be attributed to the assumed retarding torque not
being able to satisfy the requirement. Other reason may be
the inherent round off error present in search solution
computation.
6. Conclusions
Present work tries to look at the performance variation
problem of manipulator in robot manufacturer and
designers perspective and its suitable solutions. It gives
an insight in use of simulation method for modeling and
optimizing the performance of robot manipulators. It
illustrates a novel approach, i.e. search (heuristic) based
method to obtain performance measures of manipulator.
Followed by use of ANOVA technique to determine
statistically signicant parameter and parameter combinations. While simulating performance of manipulator no
constraints on angular velocity and accelerations are
247
applied. The approach presented will help in determining

signicant control factors responsible for performance
variation and identication of the optimum parameters
rather than spending effort in controlling performance of
Table A1
Combination
number
l1 (m)
l2 (m)
m1 (kg)
m2 (kg)
t1
(N m)
t2
(N m)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
0.40
0.40
0.40
0.40
0.40
0.40
0.40
0.40
0.40
0.40
0.40
0.40
0.40
0.40
0.40
0.40
0.40
0.40
0.40
0.40
0.40
0.40
0.40
0.40
0.40
0.40
0.40
0.40
0.40
0.40
0.40
0.40
0.50
0.50
0.50
0.50
0.50
0.50
0.50
0.50
0.50
0.50
0.50
0.50
0.50
0.50
0.50
0.50
0.50
0.50
0.50
0.50
0.50
0.50
0.50
0.50
0.25
0.25
0.25
0.25
0.25
0.25
0.25
0.25
0.25
0.25
0.25
0.25
0.25
0.25
0.25
0.25
0.35
0.35
0.35
0.35
0.35
0.35
0.35
0.35
0.35
0.35
0.35
0.35
0.35
0.35
0.35
0.35
0.25
0.25
0.25
0.25
0.25
0.25
0.25
0.25
0.25
0.25
0.25
0.25
0.25
0.25
0.25
0.25
0.35
0.35
0.35
0.35
0.35
0.35
0.35
0.35
5.5
5.5
5.5
5.5
5.5
5.5
5.5
5.5
6.5
6.5
6.5
6.5
6.5
6.5
6.5
6.5
5.5
5.5
5.5
5.5
5.5
5.5
5.5
5.5
6.5
6.5
6.5
6.5
6.5
6.5
6.5
6.5
5.5
5.5
5.5
5.5
5.5
5.5
5.5
5.5
6.5
6.5
6.5
6.5
6.5
6.5
6.5
6.5
5.5
5.5
5.5
5.5
5.5
5.5
5.5
5.5
4
4
4
4
5
5
5
5
4
4
4
4
5
5
5
5
4
4
4
4
5
5
5
5
4
4
4
4
5
5
5
5
4
4
4
4
5
5
5
5
4
4
4
4
5
5
5
5
4
4
4
4
5
5
5
5
500
500
800
800
500
500
800
800
500
500
800
800
500
500
800
800
500
500
800
800
500
500
800
800
500
500
800
800
500
500
800
800
500
500
800
800
500
500
800
800
500
500
800
800
500
500
800
800
500
500
800
800
500
500
800
800
100
105
100
105
100
105
100
105
100
105
100
105
100
105
100
105
100
105
100
105
100
105
100
105
100
105
100
105
100
105
100
105
100
105
100
105
100
105
100
105
100
105
100
105
100
105
100
105
100
105
100
105
100
105
100
105
ARTICLE IN PRESS
248
Table A1 (continued )
Combination
number
l1 (m)
l2 (m)
m1 (kg)
m2 (kg)
t1
(N m)
t2
(N m)
57
58
59
60
61
62
63
64
0.50
0.50
0.50
0.50
0.50
0.50
0.50
0.50
0.35
0.35
0.35
0.35
0.35
0.35
0.35
0.35
6.5
6.5
6.5
6.5
6.5
6.5
6.5
6.5
4
4
4
4
5
5
5
5
500
500
800
800
500
500
800
800
100
105
100
105
100
105
100
105
manipulator. This approach provides a powerful design

tool for selecting combination of parameters for optimal
performance of manipulator.
Appendix A. Design matrix for control factor combinations
The design matrix for control factor combinations is
given in Table A1.
References
[1] Driels MR, Swayze WE. Automated partial pose measurement
system for manipulator calibration experiments. IEEE Trans Robot
Automat 1993;10:43040.
[2] Chen J, Chao LM. Positioning error analysis for robot manipulators
with all rotary joints. IEEE J Robot Automat 1987;3:53545.
[3] Roth ZS, Mooring BW, Ravani B. An overview of robot calibration.
IEEE J Robot Automat 1987;3:37785.
[4] Whitney DE, Lozinski CA, Rourke JM. Industrial robot forward
calibration method and results. J Dyn Syst Meas Control 1986;
108:110.
[5] Benhabib B, Fenton RG, Goldenberg AA. Computer aided
joint error analysis of robot joints. IEEE J Robot Automat
1987;4:31723.
[6] Jang H, Kim SH, Kwak YK. Calibration of geometric and nongeometric errors of an industrial robot. Robotica 2001;19:31121.
[7] Azadivar S. The effect of joint position errors of industrial robots on
their performance in manufacturing operations. IEEE J Robot
Automat 1987;2:10915.
[8] Bhatti PK, Rao SS. Reliability analysis of robot manipulators.
ASME J Mech Transmissions Automat Des 1988;110:17581.
[9] Menq CH, Borm JH. Statistical characterization of position errors of
an ensemble of robots and its applications. J Mech Transmissions
Automat Des 1989;111:21522.
[10] Manoochehri S, Seireg AA. A computer based methodology for the
form synthesis and optimal design of robot manipulators. J Mech Des
1990;112:5017.
[11] Offodile OF, Ugwu K. Evaluating the effect of speed and payload on
robot repeatability. Robot Comput Integr Manuf 1991;8:2733.
[12] Parkinson A, Sorenson S, Pourhassan N. A general approach for
robust optimal design. ASME J Mech Des 1993;115:7480.
[13] Chen W, Allen JK, Tsui KL, Mistree F. A procedure for robust
design: minimizing variation caused by noise factors and control
factors. Trans ASME J Mech Des 1996;118:47885.
[14] Welch WJ, Buck RJ, Sacks J, Wynn HP, Mitchell TJ, Morris MD.
Screening, predicting and computer experiments. Technometrics
1992;34(1):1525.
[15] Chen W, Allen JK, Mavris D, Mistree F. A concept exploration
method for determining robust top level specications. Eng Optim
1996;26:13758.
[16] Chen W, Simpson TW, Allen JK, Mistree F. Use of design capability
indices to satisfy the range set of design requirements. ASME Design
Automation Conference, Irvine, CA, ASME paper 96-DETC/DAC1090.
[17] Simpson TW, Chen W, Allen JK, Mistree F. Conceptual design of a
family of products through the use of the robust concept exploration
method. 6th AIAA/USAF/NASA/ISSMO symposium on multidisciplinary analysis and optimization Bellevue, WA, 1996, AIAA,
2. p. 153545.
[18] Peplinski JD, Allen JK, Mistree F. Integrating product design with
manufacturing process design using robust concept exploration
method. Design theory and methodology conference, Irvine, CA,
ASME, Paper No.96-DETC/DTM-1502, 1996.
[19] Koch PN, Allen JK, Barlow A, Mistree F. Conguring turbine
propulsion systems using robust concept exploration. In: ASME
design automation conference. Irvine CA: ASME; 1996; Paper No.
96-DETC/DAC-1285.
[20] Koch PN, Allen JK, Mistree F, Mavris D. The problem of size in
robust design. ASME design automation conference, Sacramont, CA,
ASME, paper No. DETC97DAC3983, 1997.
[21] Guinta AA, Balabanov V, Haim D, Grossman B, Mason WH,
Watson LT. Wing design for a high speed civil transport using a
design of experiments methodology. 6th AIA/USAF/NASA/ISSMO
symposium on multidisciplinary analysis and optimization, Bellevue,
WA, AIAA vol. 1, 1996. p. 16883.
[22] Rout BK, Mittal RK. A review on minimizing variability in product
performance with the help of computational tools by controlling
active and noise factor for robust design. Proceedings of national
conference on current applications of computers in design engineering, J.N.V. University, Jodhpur, India, 2001. p. 1323.
[23] Rout BK, Mittal RK. Parametric robot manipulator design
optimization using taguchi method. Proceedings of 48th congress,
Indian Society for Theoretical and Applied Mechanics, IIT Guahati,
India, 2003. p. 1527.
[24] Rout BK, Mittal RK. Manipulator kinematic parameter selection
design of experiment approach. Conference proceeding of emerging
trends of concurrent engineering, N.I.T Rourkela, Orissa, India,
2005. p. 1209.
[25] Rout BK, Mittal RK. Parametric tolerance design of manipulator.
Proceedings of conference on emerging trends in mechanical
engineering, N.I.T. Kurukshetra, Haryana, India, 2005. p. 5561.
[26] Rout BK, Mittal RK. Modeling and optimization of A 2 DOF RR
planar manipulator using ANN. 14th ISME international conference,
Delhi College of Engineering, New Delhi, India, 2005. p. 4017.
[27] Mittal RK, Nagrath IJ. Robotics and control. New Delhi: Tata
McGraw Hill Publication; 2003.
[28] Park SH. Robust design and analysis for quality engineering.
London: Chapman & Hall Publication; 1998.
[29] Free down load software. Design Expert (DX) Version 6: http://
www.statease.com.
[30] Montgomery DC. Design and analysis of experiments, 5th Ed. New
York: Wiley; 2001.

Combined

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Combined

Uploaded by

Copyright:

Available Formats

Robotics and Computer-Integrated Manufacturing 35 (2015) 151168

Contents lists available at ScienceDirect

Robotics and Computer-Integrated Manufacturing

Geometric calibration of industrial robots using enhanced partial pose

Ecole des Mines de Nantes, 4 rue Alfred-Kastler, Nantes 44307, France

modeling, measurement, identication and implementation. The

Y. Wu et al. / Robotics and Computer-Integrated Manufacturing 35 (2015) 151168

the solution of a linear least square problem can be found

6-dof parallel robot [25]

Stewart platform [36]

Gradient search method

Selection of measurement congurations:

Y. Wu et al. / Robotics and Computer-Integrated Manufacturing 35 (2015) 151168

computations caused by a poor convergence and high dimension

2. Problem of geometric calibration

Fig. 1. Enhanced robot calibration procedure.

Y. Wu et al. / Robotics and Computer-Integrated Manufacturing 35 (2015) 151168

where t0 = g (q, 0 ) is the end-effector location computed using

Fig. 2. Measurement tool with several reference points.

on the measurement methods (measurement tools and devices,

Y. Wu et al. / Robotics and Computer-Integrated Manufacturing 35 (2015) 151168

problems that should be considered are the following:

number of techniques that allows us to obtain the manipulator

The base transformation Tbase = TxTyTz R xR yR z

These problems will be considered in more details in the following sections.

(1) For revolute joint TJoint, jTLink, j = R e, j (q j , qj )[TuTvR uR v]Lj

3. Manipulator geometric modeling

T(q) = Tbase ( b )TJoint (q1, q1)TLink ( L1) ...

TJoint (qn, qn )TLink ( Ln ) Ttool (t )

The tool transformation Ttool = TxTyTz R xR yR z t

For the case of consecutive revolute joint R e, j (q j , qj )

(1) if e j e j 1, eliminate the term R u, L j 1 or R v, L j 1 that corresponds to R e, j ;

Let us apply the above presented technique to the industrial

where T with different indices denote the relevant transformation

Y. Wu et al. / Robotics and Computer-Integrated Manufacturing 35 (2015) 151168

T=TxTyTz R x R y R z R z (q1 + q1)TxTyR x R y

where q j is the joint offset, p xj , p yj , p zj and xj , yj , zj are the

TyTz R y R z R y (q5 + q5 )[TxTz R x R z ]L5

It should be mentioned that the nominal values of some

3. The geometric parameters of the last link

{Ty, Tz, R y, R z}L6

Remark 4. In the case when only position measurements are

{Tx, Ty, Tz, R x, R y, R z}band

This nally allows us to obtain the complete and irreducible

R y (q2 + q2 )[TxR x R z ]L 2 R y (q3 + q3 )[TxTz R z ]L 3

This model will be further used for geometric calibration of the

= px1 p y1 x1 q2 px2 x2 z2 q3 px3 pz 3 z 3 q4 p y4 pz 4 z 4

where the geometric meaning of d2, ... , d5 is illustrated in Fig. 3. In

4. Enhanced partial pose measurement method

Y. Wu et al. / Robotics and Computer-Integrated Manufacturing 35 (2015) 151168

5. Identication of manipulator geometric parameters

Y. Wu et al. / Robotics and Computer-Integrated Manufacturing 35 (2015) 151168

This allows us to obtain 3mn scalar equations for the calibration

that can also be rewritten in a matrix form as

the vector of the reference point positions pij , j = 1, n (that are

This leads to the following simplied expression of Eq. (12)

R base = [~ base ]+I

and the residuals are integrated in a single vector

jth reference point for the ith manipulator conguration, is

= Jj i(p) T Jj i(p) Jj i(p) T pij

It should be noted that, to achieve the desired accuracy for the

Y. Wu et al. / Robotics and Computer-Integrated Manufacturing 35 (2015) 151168

including weighting coefcients

5.2. Inuence of the measurement errors on the identication