You are on page 1of 11

Industrial Robot: An International Journal

Absolute accuracy analysis and improvement of a hybrid 6-DOF medical robot


Ahmed Joubair Long Fei Zhao Pascal Bigras Ilian Bonev
Article information:
To cite this document:
Ahmed Joubair Long Fei Zhao Pascal Bigras Ilian Bonev , (2015),"Absolute accuracy analysis and improvement of a hybrid 6-
DOF medical robot", Industrial Robot: An International Journal, Vol. 42 Iss 1 pp. 44 - 53
Permanent link to this document:
http://dx.doi.org/10.1108/IR-09-2014-0396
Downloaded on: 25 February 2016, At: 08:39 (PT)
References: this document contains references to 23 other documents.
To copy this document: permissions@emeraldinsight.com
The fulltext of this document has been downloaded 319 times since 2015*
Users who downloaded this article also downloaded:
Haixia Wang, Xiao Lu, Zhanyi Hu, Yuxia Li, (2015),"A vision-based fully-automatic calibration method for hand-eye serial robot",
Downloaded by RMIT University At 08:39 25 February 2016 (PT)

Industrial Robot: An International Journal, Vol. 42 Iss 1 pp. 64-73 http://dx.doi.org/10.1108/IR-06-2014-0352


Bingxi Jia, Shan Liu, Yi Liu, (2015),"Visual trajectory tracking of industrial manipulator with iterative learning control", Industrial
Robot: An International Journal, Vol. 42 Iss 1 pp. 54-63 http://dx.doi.org/10.1108/IR-09-2014-0392
Robert Bogue, (2015),"Robotic exoskeletons: a review of recent progress", Industrial Robot: An International Journal, Vol. 42 Iss
1 pp. 5-10 http://dx.doi.org/10.1108/IR-08-2014-0379

Access to this document was granted through an Emerald subscription provided by emerald-srm:393177 []
For Authors
If you would like to write for this, or any other Emerald publication, then please use our Emerald for Authors service
information about how to choose which publication to write for and submission guidelines are available for all. Please visit
www.emeraldinsight.com/authors for more information.
About Emerald www.emeraldinsight.com
Emerald is a global publisher linking research and practice to the benefit of society. The company manages a portfolio of
more than 290 journals and over 2,350 books and book series volumes, as well as providing an extensive range of online
products and additional customer resources and services.
Emerald is both COUNTER 4 and TRANSFER compliant. The organization is a partner of the Committee on Publication Ethics
(COPE) and also works with Portico and the LOCKSS initiative for digital archive preservation.

*Related content and download information correct at time of download.


Absolute accuracy analysis and improvement
of a hybrid 6-DOF medical robot
Ahmed Joubair, Long Fei Zhao, Pascal Bigras and Ilian Bonev
École de technologie supérieure, Montreal, QC, Canada

Abstract
Purpose – The purpose of this paper is to describe a calibration method developed to improve the accuracy of a six degrees-of-freedom medical
robot. The proposed calibration approach aims to enhance the robot’s accuracy in a specific target workspace. A comparison of five observability
indices is also done to choose the most appropriate calibration robot configurations.
Design/methodology/approach – The calibration method is based on the forward kinematic approach, which uses a nonlinear optimization
model. The used experimental data are 84 end-effector positions, which are measured using a laser tracker. The calibration configurations are chosen
through an observability analysis, while the validation after calibration is carried out in 336 positions within the target workspace.
Findings – Simulations allowed finding the most appropriate observability index for choosing the optimal calibration configurations. They also
showed the ability of our calibration model to identify most of the considered robot’s parameters, despite measurement errors. Experimental tests
confirmed the simulation findings and showed that the robot’s mean position error is reduced from 3.992 mm before calibration to 0.387 mm after,
and the maximum error is reduced from 5.957 to 0.851 mm.
Originality/value – This paper presents a calibration method which makes it possible to accurately identify the kinematic errors for a novel medical
Downloaded by RMIT University At 08:39 25 February 2016 (PT)

robot. In addition, this paper presents a comparison between the five observability indices proposed in the literature. The proposed method might
be applied to any industrial or medical robot similar to the robot studied in this paper.
Keywords Robot calibration, Kinematic calibration, Robot accuracy, Manipulator accuracy, Kinematic parameters, Observability
Paper type Research paper.

1. Introduction only kinematic calibration is considered and is referred to here


simply as robot calibration.
Parallel robots are often considered more precise than serial
This paper describes a full kinematic calibration of a 6-DOF
robots and have a greater load-to-weight ratio because their hybrid robot which is dedicated to medical applications
structure is stiffer without being bulkier (Joubair et al., 2013a). (echography). Few research studies have been dedicated to
The poor accuracy of serial robots is mainly attributed to error enhancing the accuracy of medical robots. Mavroidis et al.
accumulation, but serial robots offer larger workspaces. (1998) performed a positioning error analysis on a 6-DOF
Consequently, the so-called hybrid serial–parallel manipulators manipulator, considering the kinematic and non-kinematic
are considered as a compromise solution. They offer a stiff errors of the robot’s parameters. The robot studied was
structure and relatively large workspace. Therefore, greater designed for the accurate positioning of patients under
repeatability is obtained, but to ensure good accuracy, the treatment using particle beams, and its repeatability and
robot must be calibrated. accuracy were evaluated at ⫾ 0.15 and ⫾ 5 mm, respectively,
Two categories of robot calibration are used: kinematic and and no calibration was carried out to improve this robot’s
non-kinematic. The first method is based on the identification accuracy. Janvier et al. (2008) evaluated the performance of a
of and compensation for the errors related to the robot’s medical robotic three-dimensional (3D) ultrasound imaging
kinematic parameters (e.g. link lengths and active joint offsets, system, dedicated to lower limb blood vessels. The evaluated
as well as the base and the tool frame locations). robot’s repeatability was around 0.20 mm, while the robot’s
Non-kinematic calibration considers additional parameters accuracy was improved to 0.57 and 1.6 mm, for the mean and
related to other sources of errors, such as thermal and stiffness maximum values, respectively. Szep et al. (2009) analyzed, by
effects. For more details about the different categories of robot simulation, the accuracy of a 2-DOF (RPRPR) medical
calibration, interested readers are referred to Roth et al. parallel robot, with respect to the kinematic errors. In (Song
et al., 2009), the position accuracy of a medical robotic system
(1987); Elatta et al. (2004); and Joubair (2012). In this paper,
designed for spine surgery was evaluated and improved to
1.1418 and 0.8878 mm for the maximum and the mean
values, respectively. The robot’s accuracy was improved by a
The current issue and full text archive of this journal is available on kinematic calibration. More recently, Torres et al. (2011)
Emerald Insight at: www.emeraldinsight.com/0143-991X.htm

The authors thank the Canada Research Chairs program, the Canadian
Foundation for Innovation and NSERC for funding this work.
Industrial Robot: An International Journal
42/1 (2015) 44 –53 Received 30 September 2014
© Emerald Group Publishing Limited [ISSN 0143-991X] Revised 3 November 2014
[DOI 10.1108/IR-09-2014-0396] Accepted 10 November 2014

44
Hybrid 6-DOF medical robot Industrial Robot: An International Journal
Ahmed Joubair, Long Fei Zhao, Pascal Bigras and Ilian Bonev Volume 42 · Number 1 · 2015 · 44 –53

presented the calibration of an ultrasound acquisition system, Figure 1 The MedRUE robot with the tool part
in which the probe is attached to a 6-DOF serial robot. The
probing of a human femur shows the capability of the system
to measure the corresponding dimensions with an error of
6.0553 mm.
In all the works presented above, the obtained accuracy (i.e.
the maximum position error) after calibration is higher than 1
mm and, in some works, only simulations are performed. In
our work, a calibration framework is developed and applied to
significantly improve the accuracy of an ultrasound imaging
medical robot. This robot is expected to be remotely
controlled by a doctor or any specialized operator. For this
reason, good accuracy and therefore a calibration of the robot
is required. Our calibration approach is based on reducing the
residual of the end-effector position errors by using forward
kinematic equations (forward calibration method). The
calibration configurations are optimally chosen through an
observability analysis. Figure 2 One five-bar mechanism of the MedRUE robot
The observability analysis determines the most
appropriate poses inside the robot workspace that provide
Downloaded by RMIT University At 08:39 25 February 2016 (PT)

the best identification of the robot’s parameter errors. This


analysis is based on the so-called observability indices,
which allow the quantifying of the efficiency of the
parameters’ identification process. Five observability
indices (O1, O2, O3, O4 and O5) are presented in the
literature (Sun and Hollerbach, 2008; Borm and Menq,
1989; Nahvi and Hollerbach, 1996; Driels and Pathre,
1990). Comparisons of the efficiency of these indices were
achieved by Sun and Hollerbach (2008); Joubair et al.
(2013b, 2013c), and most conclusions showed that the
most appropriate index for the kinematic calibration is O1.
A verification of this claim and a detailed explanation of this
index are presented in the section Observability analysis.
The end-effector position measurements are usually carried
out by means of 3D measurement devices, such as a probing symmetrically assembled on the robot base, which, in turn, is
coordinate measuring machine (CMM) (Joubair and Bonev, fixed to a linear guide actuated through a servomotor M1. We
2013; Joubair et al., 2012), an optical CMM (Nubiola et al., note that the two five-bar mechanisms are considered to be
2013), a probing articulated arm (Joubair et al., 2012) or a perfectly parallel to each other and perpendicular to the xB,
laser tracker (Joubair et al., 2013b; Xueyou and Shenghua, which is the x axis of the base reference frame (FB). Each
2007). Some one-dimensional measurement instruments are five-bar mechanism i (i ⫽ 1 [. . .] 2) has five links: the distance
also used: a ballbar (Kim, 2005; Huang et al., 2006; Nubiola di between the anchor points of the two proximal links and the
and Bonev, 2014) or a laser interferometer (Alici and four mobile links having Lij (j ⫽ 1 [. . .] 4) as lengths. The five
Shirinzadeh, 2003). In our measurement process, we use a links connect five revolute joints (Ai, Bi, Ci, Di and Ei), among
laser tracker having an uncertainty of ⫾ 40 ␮m. which only two (Ai and Ci) are actuated by using rotary motors
This paper is organized as follows: a description of the Mi⫻2 and Mi⫻2⫹1, which involve two angles, qi⫻2 and qi⫻2⫹1,
kinematic calibration model is given in the next section, respectively. A total of five angles of active joints are
followed by a presentation of the parameter identification considered (q2 [. . .], q6), while the variable attributed to the
approach and an observability analysis, as well as an linear motion is denoted q1.
explanation of the identifiable parameters. The next sections To close the kinematic chain, joints E1 and E2 are linked
present a simulation study followed by an experimental through the probe support (Figure 1), which has a universal
calibration, including an analysis of the results. Finally, joint at each extremity Gi. The x coordinate of Gi with respect
conclusions are drawn in the last section. to the base frame is denoted by d4i. When the robot is at its
home (zero) position, the vector rE1E2 is supposed to be parallel
2. Robot calibration model to rG1G2 and 㛳rG1G2㛳 is equal to d41⫹d42; otherwise, it will be the
2.1 The robot description and the main reference difference between these two distances. This difference is
frames compensated by a passive prismatic joint located between G2
The MedRUE robot (Medical Robot for vascular Ultrasound and the probe support. Finally, the origin of the last reference
Examination), shown in Figure 1, is a 6-DOF hybrid serial– frame F6 is located midway between G1 and G2; its x axis (x6)
parallel robot with five active revolute joints and one linear is defined to align with rG1G2, and z6 is pointing toward the
guide. It is composed of two five-bar mechanisms (Figure 2) probe center. The probe is actuated – to have a rotation

45
Hybrid 6-DOF medical robot Industrial Robot: An International Journal
Ahmed Joubair, Long Fei Zhao, Pascal Bigras and Ilian Bonev Volume 42 · Number 1 · 2015 · 44 –53

around x6 – by a small rotary actuator (M6) attached to the rOiBi ⫽ rOiAi ⫹ 关Li1 cos 共qi⫻2 ⫹ ␦qi⫻2兲 Li1 sin 共qi⫻2 ⫹ ␦qi⫻2兲 0兴T
link, having L14 as its length. In our calibration process, the (4)
probe is replaced with a specially designed part on which a
measurement target is attached, and the corresponding center
is considered to be the origin of the tool reference frame (Ftool). rOiDi ⫽ rOiCi ⫹ 关Li3 cos 共qi⫻2⫹1 ⫹ ␦qi⫻2⫹1兲 Li3 sin 共qi⫻2⫹1
We also use a world reference frame (Fworld), which is ⫹ ␦qi⫻2⫹1兲 0兴T (5)
associated with the robot’s work-cell.
Fworld(with axes xw, yw and zw) is chosen to be on the robot’s where, Li1 and Li3 are the lengths of the four swinging links as
work table and to have approximately the same orientation as shown in Figure 2, and ␦qi is the offset of active joint i. While,
FB, which is located on the robot’s base at OB. As shown in the vector rOiAi is calculated as follows:
Figure 1, the axis xB is aligned with the axis of the linear guide,
and zB is normal to the plane defined by the platform of the
robot’s base. The translation TB ⫽ [xb, yb, zb]T of FB with
rOiAi ⫽ 冋兹 共C1y ⫺ A1y兲2 ⫹ 共C1z ⫺ A1z兲2 0 0
2
册 T
(6)
respect to Fworld, and the orientation (␾bx, ␾by, ␾bz), described
in XYZ fixed Euler angles, are expected to be identified by the and:
calibration process.
Knowing that the end-effector’s orientation is not used in rOiCi ⫽ ⫺rOiAi (7)
our calibration process, therefore, only the translation
6
Ttool ⫽ 关xt yt zt 兴T of Ftool with respect to the robot’s last The coordinates of Ei with respect to a frame Fi are obtained
reference frame F6 is identified. The origin of Ftool is described as follows:
Downloaded by RMIT University At 08:39 25 February 2016 (PT)

to be the center of the end-effector’s probe (replaced in our


experiments by a measurement target), and its orientation is rOiEi ⫽ rOiDi ⫹ rDiSi ⫹ rSiEi (8)
the same as that of F6.
where:
2.2 Kinematic model of the robot
Given the vector ␺ ⫽ [q1, q2 [. . .], q6]T of the active joint rDiBi Li4
冉⫺ Li2

2 2
rDiSi ⫽ ⫹1 (9)
variables, the end-effector’s pose with respect to the world 2 㛳rD B 㛳2 i i
frame is represented by using homogeneous matrices as
follows:
rSiEi ⫽ 兹L 2
⫺ 㛳rDiSi㛳2Rz 共 ␲2 兲 r̂
Oi
(10)
冋 册
i4 DiBi
world world
R T
world
Atool (␺) ⫽ tool tool
⫽ ABworldAB6 Atool
6
(1)
0 1 and Rz is the rotation matrix around the z axis, and r̂DiBi is the
unit vector along rDiBi.
where:
The coordinates of Ei (Figures 1 and 2) with respect to a

冋R0 T1 册
frame FB are obtained as follows:
ABworld ⫽
B B
(2)
rOBEi ⫽ AiBrOiEi (11)

冤 冥
1 0 0 xt where:
0 1 0 yt
A 6
⫽ (3) 共⫺1兲id3i ⫹ q1 ⫹ ␦q1

冤 冥
tool
0 0 1 zt 0 0 1
0 0 0 1 Aiy ⫹ Ciy
⫺sin 共␪i兲 cos 共␪i兲 0
2
Ai0 ⫽
b
with A denoting the homogeneous matrix representing a Aiz ⫹ Ciz
a ⫺cos 共␪i兲 ⫺sin 共␪i兲 0
frame a with respect to a frame b. The calculation of AB6 is 2
presented in the following paragraphs. 0 0 0 1
The forward kinematic equations of MedRUE are obtained (12)
by solving the direct kinematics of the two five-bar
mechanisms (Figure 2), i.e. identify the position of each Ei and ␪i, which is the angle between rAiCi and the normal of the
with respect to FB, and then find the coordinates of Otool (the xByB plane, is calculated as follows:
origin of the tool reference frame) with respect to FB. Finally,
the Otool coordinates are transformed so that they are with ␪i ⫽ atan2共yAi ⫺ yCi, zCi ⫺ zAi兲 (13)
respect to Fworld. Hence, the position of the last joint Ei of each
five-bar mechanism is calculated with respect to its own The orientation of F6 with respect to FB is obtained from the
reference frame Fi having Oi as origin, and then the obtained corresponding rotation matrix RB6 ⫽ 关Rx共␥兲Ry共␤兲Rz共␣兲兴, where
positions are transformed so that they are with respect to FB. ␥, ␣ and ␤ are the fixed XYZ Euler angles. The angle ␥ can
The origin Oi of each five-bar mechanism is defined to be be calculated directly as ␥ ⫽ qDE1 ⫹ q6., while ␣ and ␤ are
midway on di; xi is aligned with rCiAi, and zi is parallel to xB. obtained from the first column of the rotation matrix, which,
The coordinates of Bi and Di are expressed with respect to in turn, corresponds to the unit vector r̂G1G2 ⫽ 关ux uy uz 兴T. As
the local frame Fi as follows: a result:

46
Hybrid 6-DOF medical robot Industrial Robot: An International Journal
Ahmed Joubair, Long Fei Zhao, Pascal Bigras and Ilian Bonev Volume 42 · Number 1 · 2015 · 44 –53

␣ ⫽ sin⫺1共uy cos ␥ ⫹ uz sin ␥兲 (14) (positions only in our case) with respect to the parameters
expected to be identified (i.e. the 36 kinematic parameters):
␤ ⫽ a tan 2共uysin ␥ ⫺ uzcos ␥, 兹共cos ␣兲2 ⫺ 共uysin ␥ ⫺ uzcos ␥兲2兲 (15) J ⫽ 关J 1 · · · J m 兴 T (18)

The translation of F6 with respect to FB is calculated as where:


follows:
⭸xi共␺i, pnom兲 ⭸xi共␺i, pnom兲

冋 册 冋 册
冤 冥
d31 ⫺ d41 d41 ⫹ d5 · · ·
¡ ⭸p1 ⭸pn
TB6 ⫽ OBE1 ⫹ 0 ⫹ R共␥, ␤, ␣兲 0 (16) ⭸yi共␺i, pnom兲 ⭸yi共␺i, pnom兲
0 0 Ji ⫽ · · · (19)
⭸p1 ⭸pn
⭸zi共␺i, pnom兲 ⭸yi共␺i, pnom兲
Finally, the pose of the last joint with respect to FB is · · ·
⭸p1 ⭸pn
expressed as follows:

冋 册
and ␺i ⫽ 关qi1, qi2, · · · qi6 兴T is the vector of the joint
RB6 TB6
A ⫽
B
6 (17) variables for the calibration configuration i (i ⫽ 1 [. . .] m),
0 1
and n is the total number of the kinematic parameters
(n ⫽ 36).
To summarize, the considered robot’s parameters are:
The number of the independent (i.e. identifiable)
● The lengths of the ten links of the two five-bar mechanisms
parameters corresponds to the rank rJ of J. The rank of our
Downloaded by RMIT University At 08:39 25 February 2016 (PT)

are: L11, L12, L13, L14, L21, L22, L23 and L24.
Jacobian matrix is 28, which means that eight parameters are
● The y and z coordinates of the anchor points of the two
not identifiable. To determine these eight parameters, we use
proximal links of the five-bar mechanisms are: A1y, A1z,
an iterative algorithm, in which each iteration allows us to find
C1y, C1z, A2y, A2z, C2y and C2z.
the parameter that, if eliminated from J, leads to the maximum
● The offsets of the six active joints are: ␦q1, ␦q2, ␦q3, ␦q4,
reduction of the condition number while retaining the rank rJ
␦q5, ␦q6.
at its initial value (i.e. rJ ⫽ 28). Considering that the number
● The offset parameters for the tool part are: d31, d32, d41,
of columns of J is denoted by m, the used column elimination
d42, and d5.
algorithm is explained below:
● The parameters defining the base with respect to Fworld are:
While rJ ⬍ m, do:
xb, yb, zb, ␾by, ␾bx and ␾bz
For i ⫽ 1 to m
● The position of the tool frame with respect to the last ● Remove the column i from J, and calculate both the new
frame is: xt, yt and zt. ⴱ ⴱ
rank rJ and the new condition number cJ for the new
Eight parameters – among all 36 considered parameters – are Jacobian matrix.
ⴱ ⴱ ⴱ
redundant, which means that we need to reduce the number ● If rJ ⫽ rJ: Save i, rJ and cJ as a line in a (m ⫻ 3) result
of the identifiable parameters to 28 (see section Identifiable matrix M.
parameters).
End for
● Execute an ascending sort on M, based on the third

3. Parameter identification column (i.e. cJ ). The sorted matrix is denoted by Mⴱ. If the
To identify the values of the robot’s parameters, we decided to separated elimination of some parameters gives the same

use forward kinematic calibration, which is based on cJ , the priority in the sort process – for these parameters –
minimizing the errors of the end-effector poses. However, only is given randomly.
positions are considered in our identification process, as they ● Extract the column number i from the first line of Mⴱ (The

are easier to measure. To achieve efficient parameter line having the lowestrJ ) and remove the corresponding
identification, excluding the non-identifiable parameters from column from J. The obtained Jacobian matrix is denoted
the identification process is appropriate. In addition, the by Jⴱ, and its number of columns is mⴱ ⫽ m-1.
efficiency of this identification might be considerably ● Replace J by Jⴱ.
improved through suitable selection of the calibration ● Replace m by mⴱ.
configurations. More details about the identification process
End while
and the most appropriate calibration configurations are given
After executing the above algorithm, which is programmed
in subsequent sections.
in Matlab, the found non-identifiable parameters are: A2y,
A2z, d31, d32, d41, d5, xb and zb. Consequently, the
3.1 Identifiable parameters corresponding columns of these parameters are removed from
As mentioned previously, our robot has 36 kinematic matrix J, and the number of identifiable parameters becomes
parameters. Some of these parameters are not identifiable, n ⫽ 28, which correspond to the rank of J.
mainly because of their redundancy with other parameters.
Therefore, the redundant parameters should be excluded from 3.2 Target workspace and identification process
the identification process. To determine the redundant Our robot will be calibrated in a specific area of its workspace.
parameters, we use the identification Jacobian matrix J, which This volume is the workspace covering the equivalent of the
is composed of the derivative of the calibration poses patient’s leg. Figure 3 shows the target workspace, which has

47
Hybrid 6-DOF medical robot Industrial Robot: An International Journal
Ahmed Joubair, Long Fei Zhao, Pascal Bigras and Ilian Bonev Volume 42 · Number 1 · 2015 · 44 –53

Figure 3 The robot’s world, base and tool reference frames, with only a few seconds, we choose to use 84 calibration
the target workspace, as well as calibration positions and trajectory configurations.
After deciding on the number of calibration configurations,
an optimal selection of these is desirable to achieve efficient
parameter identification. Usually, selecting calibration
configurations which are uniformly distributed inside the
robot workspace allow good parameter identification.
However, to get even better results, an optimization process –
called an observability analysis – can be used to select the best
configurations for the identification process. This analysis
allows us to find the most suitable robot configurations in
which the parameter errors are precisely identifiable.
The observability analysis uses the so-called observability
indices. These indices are calculated by using the singular
values ␴i (i ⫽ 1 [. . .] n) obtained from the diagonal matrix ⌺
of the singular value decomposition of the sensitivity matrix,
which is also called the identification Jacobian matrix J. For n
independent parameters, ⌺ is presented as follows:

␴1

冤 冥
0 · · · 0
␴2
Downloaded by RMIT University At 08:39 25 February 2016 (PT)

0 · · · 0
É É Ì É
been defined as a half-cylinder, 300 mm in diameter and 700 ⌺⫽ 0 0 · · · ␴n (22)
mm in length. Therefore, the aim of the robot calibration is to 0 0 · · · 0
improve its accuracy inside this target workspace. É É É É
The parameters are identified by using the least-square 0 0 · · · 0
minimization method, which minimizes the residual of the
end-effector positions: Five observability indices have been presented in the literature
(O1 [. . .] O5). According to Sun and Hollerbach (2008);
m

兺 共共x
Joubair et al. (2013b, 2013c), the index O1 seems to be the
minimize est,i ⫺ xmeas,i兲2 ⫹ 共yest,i ⫺ ymeas,i兲2 most appropriate index for the kinematic calibration.
i⫽1
Furthermore, this claim is verified in our simulation study,
⫹ 共zest,i ⫺ zmeas,i兲2兲 (20) and the obtained results have been experimentally confirmed.
The index O1 is calculated as follows:
Xest, i is the vector of the estimated coordinates of the
calibration configuration i, calculated by using the 1
共␴1␴2 . . . ␴n兲 n
corresponding set of joint values ␺i ⫽ 关qi1, qi2, · · · qi6 兴T
O1 ⫽ (23)
and the parameter vector p. Vector p is initialized with the 兹m
parameters’ nominal values and is updated at each iteration of
the identification algorithm. Thus, Xest, i is expressed as where, m is the number of calibration configurations, and ␴1
follows: [. . .] ␴n are the singular values of the identification matrix J
for the n ⫽ 28 identifiable parameters.
Xest,i(␺i, p) ⫽ 关xest,i yest,i zest,i 兴T (21) To select the appropriate calibration configurations, we use
an add–subtract algorithm (Joubair et al., 2013b, 2013c) based
Finally, Xmeas, i represents the measured coordinates of the on the DETMAX method, which was originally presented by
position i. It is expressed as the coordinates on the Mitchell (1974). A set of N ⫽ 420 configurations, uniformly
end-effector’s position with respect to the world reference distributed in the robot’s Cartesian workspace (Figure 3), is
frame. selected. The 420 configurations are considered as candidates
The calibration model equation (20) is solved by using the from which the M ⫽ 84 calibration configurations will be
function lsqnonlin from the Matlab optimization toolbox. chosen. The initial set of calibration configurations, denoted
by ␻, is uniformly selected among the N candidates and is
3.3 Observability analysis updated at each iteration of the algorithm. The update of ␻ is
As explained in the previous section, 28 parameters are achieved by the so-called addition and subtraction steps. The
expected to be identified in our calibration process. addition consists of adding the configuration among the
Theoretically, only 28 equations are needed to achieve this remaining N–M, leading to the highest value of the used
parameter identification. These equations can be obtained by observability index Oi, and the subtraction allows us to
using ten calibration configurations, which give 30 equations, eliminate the configuration having the minimum contribution
as each configuration implies three equations (for x, y and z). to the enhancement of Oi. These two steps (i.e. addition and
However, because of the measurement noise, an subtraction) are executed successively until the added
over-constraining of the identification system is necessary. configuration is itself removed in the following subtraction
Knowing that collecting the data for a specific position takes step.

48
Hybrid 6-DOF medical robot Industrial Robot: An International Journal
Ahmed Joubair, Long Fei Zhao, Pascal Bigras and Ilian Bonev Volume 42 · Number 1 · 2015 · 44 –53

The pose selection algorithm is separately executed for each 关qi1, qi2, · · · qi6 兴T corresponding to each calibration
of the five observability indices to compare their effectiveness, configuration. In addition, as shown in equation (24), a vector
by examining the obtained accuracy after calibration for each Ei of the simulated measurement noise is added to the
index. The results of this comparison are presented in the next corresponding position i:
section and in the experimental tests. For illustration
purposes, the improvement of the value of the most Xmeas,i(␺i, pact) ⫽ 关xmeas,i ymeas,i zmeas,i 兴T ⫹ Ei (24)
appropriate index O1 with respect to the successive iterations
is shown in Figure 4. The parameters’ nominal and actual values are presented in
Table I, as well as the identified values with and without
4. Simulation study measurement errors.
After the parameters have been identified, the position
The aim of this study is to evaluate the efficiency of our errors (i.e. the absolute position accuracy) are assessed in the
calibration process. Hence, the impact of the measurement robot’s target workspace. As mentioned earlier, 420 positions
noise on the parameter identification is evaluated, and the were considered in our calibration process. Only 84 positions
effectiveness of the chosen calibration configurations is also were used in the parameter identification, while the remaining
verified. Furthermore, the calibration results from a set of 336 positions were dedicated to validation. The position
calibration configurations uniformly distributed inside the errors are calculated for each validation position, by
robot workspace are compared with the calibration calculating the difference between the actual (i.e. real
configurations obtained by an observability analysis which position) and the desired positions. An actual position is
involves the five indices presented in the literature. calculated by substituting the corresponding joint values and
Downloaded by RMIT University At 08:39 25 February 2016 (PT)

For simulation purposes, the actual parameters’ values are the actual parameter values in equation (21), while a desired
simulated by introducing randomly generated errors into the position is obtained by using the identified parameters’ values.
ranges ⫾ 2 mm and ⫾ 1.5 ° for the distances and angles, The parameter identification and the accuracy assessment
respectively. These errors lead to poor robot accuracy that will are repeated both for the five sets of calibration configurations
be improved through the calibration process, by identifying obtained by separately using the five observability indices (O1,
the actual values of the robot’s parameters as closely as O2, O3 O4 and O5), and for the 84 calibration configurations
possible, despite the presence of measurement noise. The uniformly distributed in the robot’s target workspace. The
measurement errors are generated according to a normal composed xyz position errors, before and after calibration, are
distribution with a range of ⫾ 120 ␮m, which is an summarized in Table II, which shows that the index O1 is the
exaggeration of the uncertainty of our measurement best observability index for our robot’s calibration. Figure 5
instrument (i.e. a laser tracker with an uncertainty of ⫾ 40 illustrates the histograms representing the distribution of the
␮m). A total of 420 positions uniformly distributed inside the xyz composed errors before and after calibration, by using the
robot’s workspace are used (Figure 3). Of these, 84 positions observability index O1.
are dedicated to the parameter identification, and the 336 In summary, we can conclude that our calibration model is
remaining positions are used in the validation process. able to effectively identify the kinematic parameter errors and,
The parameters’ values are identified as explained in therefore, to significantly enhance the robot’s accuracy,
the previous section (Identification process). However, the regardless of the measurement errors. In addition, our
measured positions are simulated by substituting in the simulation has demonstrated that the observability index O1
forward kinematic equations vector pact of the actual yields the best improvement of the robot’s accuracy. However,
parameters’ values (Table I) and vector ␺i ⫽ we notice that the indices O3 and O5 also give results which are
close to the accuracy obtained by O1. Therefore, more
Figure 4 Evolution of the observability index O1 with respect to investigations were carried out in the experimental analysis to
the addition–subtract algorithm iterations shed more light on this matter. These will be discussed in the
next section.

5. Actual calibration and validation


As mentioned earlier, MedRUE is dedicated to the ultrasound
imaging of lower limb blood vessels. Therefore, the target
workspace has been precisely selected to replicate the area in
which the patient’s leg is reached by the ultrasound probe
(Figure 3).
The measurement instrument used in our calibration is a
laser tracker, which has an uncertainty of ⫾ 40␮m. This
uncertainty is reduced in our data acquisition process by
taking the measurement for each position five times, and the
average of each of these five measurements is considered to be
the measured value of the corresponding position. The target
used for the position measurement is a 1.5 inch spherically
mounted retroreflector (SMR) placed on a magnetic nest,
which, in turn, is attached to the robot’s end-effector

49
Hybrid 6-DOF medical robot Industrial Robot: An International Journal
Ahmed Joubair, Long Fei Zhao, Pascal Bigras and Ilian Bonev Volume 42 · Number 1 · 2015 · 44 –53

Table I Results of the simulated parameter identification


Identified parameter values, Identified parameter values,
Parameters Nominal values Actual parameter values without measurement errors with measurement errors
A1y [mm] ⫺233.000 ⫺232.034 ⫺232.034 ⫺231.815
A1z [mm] 178.000 179.412 179.412 178.643
C1y [mm] ⫺83.000 ⫺82.595 ⫺82.595 ⫺82.813
C1z [mm] 438.000 438.250 438.250 438.110
A2y [mm]ⴱ ⫺233.000 ⫺233.000 ⫺233.000 ⫺233.000
A2z [mm]ⴱ 178.000 178.000 178.000 178.000
C2y [mm] ⫺83.000 ⫺82.156 ⫺82.156 ⫺82.602
C2z [mm] 438.000 436.199 436.199 436.574
L11 [mm] 400.000 400.930 400.930 400.924
L12 [mm] 520.000 520.221 520.221 519.831
L13 [mm] 400.000 399.285 399.285 399.301
L14 [mm] 520.000 520.332 520.332 519.842
L21 [mm] 400.000 399.799 399.799 399.692
L22 [mm] 520.000 520.664 520.664 520.835
L23 [mm] 400.000 399.959 399.959 399.896
L24 [mm] 520.000 520.235 520.235 520.223
Downloaded by RMIT University At 08:39 25 February 2016 (PT)

d31 [mm]ⴱ 166.800 166.800 166.800 166.800


d32 [mm]ⴱ 166.800 166.800 166.800 166.800
d41 [mm]ⴱ 41.500 41.500 41.500 41.500
d42 [mm] 41.500 41.952 41.952 41.994
d5 [mm]ⴱ 0.000 0.000 0.000 0.000
␦q1 [mm] ⫺0.657 0.102 0.102 0.060
␦q2[°] 1.770 0.503 0.503 0.822
␦q3[°] ⫺0.766 ⫺0.601 ⫺0.601 ⫺0.410
␦q4[°] 1.735 ⫺0.389 ⫺0.389 ⫺0.360
␦q5[°] ⫺0.178 0.350 0.350 0.539
␦q6[°] 1.122 0.825 0.825 0.935
xb [mm]ⴱ ⫺109.000 ⫺109.000 ⫺109.000 ⫺109.000
yb [mm] ⫺139.000 ⫺137.110 ⫺137.110 ⫺137.654
zb [mm]ⴱ 31.000 31.000 31.000 31.000
⌽bx[°] 0.000 ⫺0.234 ⫺0.234 0.250
⌽by[°] 0.000 ⫺0.513 ⫺0.513 ⫺0.511
⌽bz[°] 0.000 ⫺1.252 ⫺1.252 ⫺1.249
xt [mm] 3.000 2.244 2.244 2.312
yt [mm] ⫺55.000 ⫺54.408 ⫺54.408 ⫺54.363
zt [mm] 138.000 139.732 139.732 139.840
Note: The non-identifiable parameters are denoted by ‘ⴱ’

Table II Position errors, before and after calibration, with different sets of calibration configurations
After calibration by using calibration configurations based on:
Position errors [mm] Before calibration O1 O2 O3 O4 O5 Uniform distribution
Mean 11.474 0.014 0.034 0.020 0.029 0.012 0.048
Maximum 20.168 0.034 0.100 0.055 0.111 0.038 0.166
STD 4.253 0.006 0.023 0.011 0.022 0.006 0.040
Note: The best accuracy values, after calibration, are in bold

(Figure 6). Finally, to identify the robot’s world frame with (20). The identified parameters’ values are presented in
respect to the laser tracker frame, three additional nests are Table III, where the non-identifiable parameters are denoted
permanently attached to the robot’s frame. by “ⴱ”.
The position measurements were carried out at the same The accuracy after calibration was evaluated in 336
420 configurations used in our simulation study. A total of 84 positions (i.e. the remaining positions among the 420 initially
positions were used in the parameters’ identification process, measured). The obtained results for the robot’s accuracy (i.e.
which used the minimization method presented in equation the xyz composed errors of the end-effector’s position with

50
Hybrid 6-DOF medical robot Industrial Robot: An International Journal
Ahmed Joubair, Long Fei Zhao, Pascal Bigras and Ilian Bonev Volume 42 · Number 1 · 2015 · 44 –53

Figure 5 Histograms of position errors in the target workspace Table III Identified values of the robot’s parameters by using the
observability index O1
Parameters Nominal values Identified values
A1y [mm] ⫺233 ⫺224.895
A1z [mm] 178 172.092
C1y [mm] ⫺83 ⫺75.641
C1z [mm] 438 433.902
A2y [mm]ⴱ ⫺233 ⫺233.000
A2z [mm]ⴱ 178 178.000
C2y [mm] ⫺83 ⫺79.920
C2z [mm] 438 438.724
L11 [mm] 400 400.342
L12 [mm] 520 516.704
L13 [mm] 400 401.443
L14 [mm] 520 520.762
L21 [mm] 400 401.732
L22 [mm] 520 524.236
L23 [mm] 400 401.282
L24 [mm] 520 525.544
Downloaded by RMIT University At 08:39 25 February 2016 (PT)

d31 [mm]ⴱ 166.800 166.800


d32 [mm]ⴱ 166.800 166.800
d41 [mm]ⴱ 41.5 41.500
d42 [mm] 41.5 39.195
d5 [mm]ⴱ 0 0
␦q1 [mm] 0 8.521
␦q2[°] 0 ⫺0.464
␦q3[°] 0 ⫺0.351
␦q4[°] 0 ⫺0.433
␦q5[°] 0 ⫺0.313
␦q6[°] 0 5.727
Figure 6 The calibration setup xb [mm]ⴱ ⫺109 ⫺109
yb [mm] ⫺139 ⫺134.354
zb [mm]ⴱ 31 31
⌽bx[°] 0 ⫺0.568
⌽by[°] 0 ⫺0.115
⌽bz[°] 0 0.274
xt [mm] 3 ⫺7.402
yt [mm] ⫺55 ⫺45.224
zt [mm] 138 140.046
Note: The non-identifiable parameters are denoted by ‘ⴱ’

the simulation study, the only difference being that the actual
position values were measured by the laser tracker.
The results summarized in Table IV confirm that O1 is the
most appropriate index for calibrating our robot; it gave the
best accuracy after calibration. For more details about
respect to Fworld) before and after calibration are presented in the accuracy before and after calibration, Table V shows the
Figure 7 and summarized in Table IV, while the position errors according to each coordinate separately.
corresponding distributions are presented as histograms in By analyzing Table V, we notice that the x coordinate makes
Figure 8. To confirm the effectiveness of the observability the highest contribution to the robot’s position errors in both
index O1, the parameter identification was carried out before and after calibration assessments. The fact that this high
separately, by using six calibration sets composed of 84 contribution remains even after calibration leads us to attribute
configurations each: those obtained by using the five these position errors to non-modelled parameter errors.
observability indices and a set of configurations uniformly Knowing that no load was applied to the robot’s end-effector,
distributed inside the target workspace. The results of this during our tests, and the temperature was controlled around
comparison are presented in Table IV. We note that the robot’s 24°C in our laboratory, the only remaining and significant source
position errors were calculated in the same fashion as presented in of error is that of backlash. In addition, the movement along

51
Hybrid 6-DOF medical robot Industrial Robot: An International Journal
Ahmed Joubair, Long Fei Zhao, Pascal Bigras and Ilian Bonev Volume 42 · Number 1 · 2015 · 44 –53

Figure 7 Position errors, before and after calibration Figure 8 Histograms of the position errors

the x axis is mainly achieved by the linear motor, which allows the
displacement of the whole structure of the robot according to the
x coordinate. As a conclusion, we can attribute the high errors on
Downloaded by RMIT University At 08:39 25 February 2016 (PT)

the x axis mainly to the backlash, and partially to the lead errors
of the linear guide, which are specified by the manufacturer to be
50 ␮m/300 mm. The lead errors provided by the manufacturer
were initially considered in the robot controller; however, an
accurate assessment of these errors might only marginally
enhance the accuracy along the x axis, as the displacement along
this axis is only 700 mm. The compensation for the lead errors
requires a more accurate measuring instrument than a laser
tracker (ideally a laser interferometer).
The backlash impact cannot be compensated by the
kinematic calibration. To reduce backlash errors, an
improvement of the gearbox quality of the robot’s actuators is
suggested (mainly the linear guide). In addition, we emphasize
that it is possible to model and compensate the backlash
Table V Position errors, according to x, y and z coordinates, by using
errors; it forms part of the so-called non-kinematic calibration,
the observability index O1
which is not considered in this paper.
We recall that in our calibration model, we neglected some Before calibration After calibration
kinematic errors, such as the angle between the links of the two Position errors x y z x y z
five-bar mechanisms and the yBzB plane (i.e. the links were Mean [mm] 3.159 1.645 1.195 0.340 0.070 0.124
supposed to be perfectly perpendicular to xB). By considering these Max [mm] 4.860 4.319 3.577 0.819 0.195 0.411
errors, the accuracy after calibration might be better, but the STD [mm] 0.407 1.044 0.989 0.186 0.045 0.079
kinematic model would be too complicated, which, in turn, would
significantly increase the delays of the robot’s controller calculations
(e.g. trajectory generation). Besides, the obtained accuracy after
are the classical kinematic parameters (i.e. distances and link
calibration is sufficient for the considered application.
lengths), the active joint offsets, the parameters defining the tool
reference frame with respect to the last joint and the robot’s base
6. Conclusions reference frame with respect to the world frame. A simulation
In this paper, we have presented a full kinematic calibration of a study showed that the used calibration model, despite
novel medical robot which is dedicated to the 3D ultrasound measurement errors, was effective in identifying 28 among the 36
imaging of lower limb blood vessels. The identified parameters considered parameters, as eight parameters were redundant. The

Table IV Position xyz composed errors, before and after calibration, obtained by using the six sets of calibration configurations
Position After calibration by using calibration configurations based on:
errors Before calibration O1 O2 O3 O4 O5 Uniform distribution
Mean [mm] 3.992 0.387 0.765 0.355 0.623 0.373 0.757
Max [mm] 5.957 0.851 1.885 1.387 1.394 0.860 1.580
STD [mm] 0.639 0.170 0.360 0.213 0.324 0.182 0.392
Note: The best accuracy values, after calibration, are in bold

52
Hybrid 6-DOF medical robot Industrial Robot: An International Journal
Ahmed Joubair, Long Fei Zhao, Pascal Bigras and Ilian Bonev Volume 42 · Number 1 · 2015 · 44 –53

calibration configurations were chosen through an observability Joubair, A., Slamani, M. and Bonev, I.A. (2013a), “A novel
analysis, which demonstrated that the index O1 was the most XY-Theta precision table and a geometric procedure for its
appropriate observability index, allowing the best accuracy kinematic calibration”, Journal of Robotics and Computer
improvement of the robot. The experimental tests confirmed the Integrated Manufacturing, Vol. 28 No. 1, pp. 57-65.
effectiveness of our calibration process, by improving the robot’s Joubair, A., Slamani, M. and Bonev, I.A. (2013c), “Kinematic
mean positioning errors from 3.992 to 0.387 mm and the calibration of a five-bar planar parallel robot using all
maximum positioning errors from 5.957 to 0.851 mm. The working modes”, Journal of Robotics and Computer Integrated
accuracy after calibration (i.e. 0.851 mm) is sufficient for Manufacturing, Vol. 29 No. 4, pp. 15-25.
the requirements of the application to which the calibrated robot Kim, H.S. (2005), “Kinematic calibration of a Cartesian
is dedicated. Furthermore, we note that the improvement in parallel manipulator”, International Journal of Control,
accuracy was limited by the robot’s repeatability, which, in turn, Automation and Systems, Vol. 3 No. 3, pp. 453-460.
was evaluated to be around 0.1 mm. The main reason for the
Mavroidis, C., Flanz, J., Dubowsky, S., Drouet, P. and Goitein, M.
repeatability errors was attributed to the backlash in the robot’s
(1998), “High performance medical robot requirements and
linear joint.
accuracy analysis”, Robotics and Computer-Integrated
Manufacturing, Vol. 14 Nos 5/6, pp. 329-338.
References Mitchell, T.J. (1974), “An algorithm for the construction of
Alici, G. and Shirinzadeh, B. (2003), “Laser interferometry D-optimal experimental designs”, Technometrics, Vol. 42
based robot position error modelling for kinematic No. 1, pp. 203-210.
calibration”, Proceedings of the 2003 IEEE/RSJ International Nahvi, J. and Hollerbach, J. (1996), “The noise amplification
Downloaded by RMIT University At 08:39 25 February 2016 (PT)

Conference on Intelligent Robots and Systems (IROS 2003), index for optimal pose selection in robot calibration”,
Las Vegas, NV, pp. 27-31. Proceedings of the 1996 IEEE International Conference on
Borm, J.H. and Menq, C.H. (1989), “Experimental study of Robotics and Automation, Minneapolis.
observability of parameter errors in robot calibration”, Nubiola, A. and Bonev, I.A. (2014), “Absolute robot
Proceedings of the 1989 IEEE International Conference on calibration with a single telescoping ballbar”, Precision
Robotics and Automation, Scottsdale. Engineering, Vol. 38 No. 3, pp. 472-480.
Driels, M.R. and Pathre, U.S. (1990), “Significance of observation Nubiola, A., Slamani, M., Joubair, A. and Bonev, I.A. (2013),
strategy on the design of robot calibration experiments”,
“Comparison of two calibration methods for a small
Journal of Robotic Systems, Vol. 7 No. 2, pp. 197-223.
industrial robot based on an optical CMM and a laser
Elatta, A.Y., Gen, L.P., Zhi, F.L., Daoyuan, Y. and Fei, L.
tracker”, Robotica, Vol. 32 No. 3, pp. 447-466.
(2004), “An overview of robot calibration”, Information
Roth, Z., Mooring, B. and Ravani, B. (1987), “An overview of
Technology Journal, Vol. 3 No. 1, pp. 74-78.
robot calibration”, IEEE Journal of Robotics and Automation,
Huang, T., Hong, Z.Y., Mei, J.P. and Chetwynd, D.G. (2006),
“Kinematic calibration of the 3-DOF module of a 5-DOF Vol. 3 No. 5, pp. 377-385.
reconfigurable hybrid robot using a double-ball-bar system”, Song, Y., An, G. and Zhang, J. (2009), “Positioning accuracy
Proceedings of the 2006 IEEE/RSJ International Conference on of a medical robotic system for spine surgery”, 2nd
Intelligent Robots and Systems, Peking, pp. 508-512. International Conference on Biomedical Engineering and
Janvier, M.A., Durand, L.J., Roy Cardinal, M.H., Renaud, I., Informatics, BMEI ’09, Tianjin, pp. 1-5.
Chayer, B., Bigras, P., de Guise, J., Soulez, G. and Cloutier, Sun, Y. and Hollerbach, J. (2008), “Observability index selection
G. (2008), “Performance evaluation of a medical robotic for robot calibration”, Proceedings of the 2008 IEEE
3D-ultrasound imaging system”, Medical Image Analysis, International Conference on Robotics and Automation, Pasadena.
Vol. 12 No. 3, pp. 275-290. Szep, C., Stan, S., Csibi, V., Manic, M. and Balan, R. (2009),
Joubair, A. (2012), “Contribution à l’amélioration de la “Kinematics, workspace, design and accuracy analysis of
précision absolue des robots parallèles”, PhD thesis De RPRPR medical parallel robot”, Human System Interactions,
Doctorat Électronique, École de Technologie Supérieure, 2009, HSI ’09, IEEE Press Piscataway, New Jersey, pp. 75-80.
Montréal. Torres, P.M.B., Gonçalves, P.J.S. and Martins, J.M.M.
Joubair, A. and Bonev, I.A. (2013), “Comparison of the (2011), “Robot calibration for precise ultrasound image
efficiency of five observability indices for robot calibration”, acquisition”, Proceedings of International Conference On
Mechanism and Machine Theory, Vol. 70 No. 1, pp. 254-265. Innovations, Recent Trends And Challenges In Mechatronics,
Joubair, A., Nubiola, A. and Bonev, I.A. (2013b), “Calibration Mechanical Engineering and New High-Tech Products
efficiency analysis based on five observability indices and two Development, MECAHITECH’11, Vol. 3, pp. 63-70.
calibration models for a six-axis industrial robot”, SAE
International Journal of Aerospace, Vol. 6 No. 1, pp. 161-168.
Corresponding author
Joubair, A., Slamani, M. and Bonev, I.A. (2012), “Kinematic
calibration of a 3-DOF planar parallel robot”, Industrial Ahmed Joubair can be contacted at: ahmedjoubair@
Robot: An International Journal, Vol. 39 No. 4, pp. 392-400. hotmail.com

For instructions on how to order reprints of this article, please visit our website:
www.emeraldgrouppublishing.com/licensing/reprints.htm
Or contact us for further details: permissions@emeraldinsight.com

53

You might also like