Available online at www.sciencedirect.

com

ScienceDirect
Procedia Engineering 69 (2014) 459 – 463

24th DAAAM International Symposium on Intelligent Manufacturing and Automation, 2013

Calibration of an Industrial Robot using a Stereo Vision System

Marko Švaco*, Bojan Šekoranja, Filip Šuligoj, Bojan Jerbić
University of Zagreb , Faculty of Mechanical Engineering and Naval Architecture, Department of Robotics and Production System Automation,
Ivana Lučića 5, 10000 Zagreb, Croatia

Abstract

Industrial robots have very good repeatability but still lack good absolute accuracy. The main reason is difference between the
ideal robot kinematic model integrated in the robot controller and actual robot parameters. A method for identifying certain
parameters of the robot model has been proposed. A noncontact method using a stereovision system attached to the robot arm is
utilized for providing measurements of calibration points in space. Points are represented as spheres which localized by the stereo
vision system project a circle in two image capture planes independent of the viewing angle. Spatial coordinates of each sphere
center are acquired in different robot configurations. From these readings errors of robot absolute positioning are measured. The
standard Denavit-Hartenberg (DH) notation is used when the modified model parameters containing joint encoder offset values
are directly input to the robot controller. Calibration experiments carried out on a KUKA KR 6 R900 industrial robot show
improved accuracy results. The maximum positioning error around calibration points was decreased from 3.63 mm prior to
calibration, to 1.29 mm after the calibration procedure.
© 2014 The Authors. Published by Elsevier Ltd.
© 2014 The Authors. Published by Elsevier Ltd.
Selection and peer-review under responsibility of DAAAM International Vienna.
Selection and peer-review under responsibility of DAAAM International Vienna
Keywords: Robot calibration; absolute accuracy

1. Introduction

Industrial robots have a very broad field of application ranging from assembly, manufacturing, and other
technologies. Industrial robots mostly deal with processes in known and structured environments where robot motion
is preprogrammed and executed, continuously repeating the same paths and motions. For this reason industrial
robots are designed and manufactured to provide very high repeatability for successful execution of predefined tasks.

* Corresponding author. Tel.: +385 1 6168 380; fax: + 385 1 6156 940.
E-mail address: [email protected]

1877-7058 © 2014 The Authors. Published by Elsevier Ltd.

Selection and peer-review under responsibility of DAAAM International Vienna
doi:10.1016/j.proeng.2014.03.012
460 Marko Švaco et al. / Procedia Engineering 69 (2014) 459 – 463

The aspect, in which industrial robots are still not able to provide satisfactory results, when delivered to end
customers, is absolute accuracy. In 6 degree of freedom (DOF) serial link manipulators [1] this problem is
emphasized. As every industrial robot is manufactured in a certain tolerance range there are no two identical
mechanical units. On the other hand each robot controller utilizes the same control model with ideal DH [2]
parameters (link lengths, offsets, twist, etc.). For this reason there always exists a certain error describing the
difference between the ideal position in the robot model and the actual position of the robot in some reference frame.
Research in the field of robot calibration shows different calibration methods and algorithms which evaluate DH
parameters of the robot model or describe novel kinematic robot models with additional parameters for improving
accuracy results. In [3] a calibration method is proposed utilizing a virtual closed kinematic chain method. A laser
spot is projected on a distant plane and small robot motions are magnified. A calibration method for identification of
36 independent parameters of a robot arm is described in [4]. Authors apply a principle of moving the robot endpoint
around spherical shells centered on a fixed point, collect data and perform identification of robot kinematic
parameters. Another calibration approach utilizing cooperative coevolutionary networks and pseudoerror theory is
described in [5]. Measurement methods in robot calibration range from theodolite [6], laser beam projection [3], 2D
machine vision [7], portable coordinate measuring machines [8] and other.
From the robot kinematic model overview industrial robot calibration methods can roughly be distinguished in
three categories:

x Error registration – the robot is viewed as a black box where compensation factors and values are given for every
robot pose
x Static calibration – identification of time invariant robot parameters
x Dynamic calibration – identification of robot motion parameters

Duelen and Schroer [9] suggest that all parameters of the kinematic robot model could be identified only by
measuring the position of a certain point of the robot end effector. Applying this principle, in this research a static
noncontact closed chain calibration method using stereovision is presented for evaluating robot kinematic parameters
that can be embedded directly in KUKA robot controllers. With this approach the control hierarchy and structure of
a robotic system is left intact and the usability of an industrial robot is not altered. In following sections the
experimental setup is described, methods and algorithms are presented and preliminary results are discussed.

2. System setup

The system consists of a 6 DOF KUKA KR 6 R900 robot arm with declared repeatability of ±0.03 mm and
payload of 6 kg. DH parameters of the robot are depicted in Table 1.

Table 1. Original values of DH robot parameters.

d (mm) θ (deg) a (mm) α (deg)
Link 1 400 θ1 25 90
Link 2 0 θ2 455 0
Link 3 0 θ3 35 90
Link 4 420 θ4 0 -90
Link 5 0 θ5 0 90
Link 6 80 θ6 0 180

The robot has a tool changing device which allows flexible automatic tool exchange. The stereo vision system (as
shown in Fig.1 a) consists of two perpendicular CCD cameras which form a virtual robot tool center point (TCP).
The two CCD cameras have macro lenses and allow high magnification of the viewed object. For the calibration
purpose a black polymer sphere was used. One pixel is roughly mapped to 0.05mm and the vision system allows
high resolution and accurate readings of sphere center.
 Marko Švaco et al. / Procedia Engineering 69 (2014) 459 – 463 461

Fig. 1. (a) Schematic overview of the stereovision system; (b) Calibration procedure experimental setup.

2.1. Calibration procedure

The virtual TCP is manually guided to desired positions and orientations where the calibration sphere is located
(Fig 1. b). The machine vision algorithm adjusts robot position with respect to the calibration sphere in the camera
reference frame. Robot joint states are recorded for later use in the parameter optimization algorithm. After an
adequate number of measurements are made the recorded joint states for each calibration point are compared.
The transformation matrix as shown in equation (1), with six joint states (θ) as variables, is composed of six
respective transformation matrices containing the DH parameters previously noted in Table 1. and information about
the robot TCP (tool).

T A1 ˜ A2 ˜ A3 ˜ A4 ˜ A5 ˜ A6 ˜ tool (1)

The maximum error for each calibration sphere position is calculated. The iterative algorithm minimizes the
maximum errors and applies new values for joint offset values. The algorithm searches the state space of vector Δθ
to find the minimum of the function around calibration points (ni) containing different robot configurations (ck) as
shown in equation (2):

ni  ck : max(Tik )  min(Tik ) o min (2)

Purpose of the algorithm is to find the best match of parameters Δθ for all calibration points c. A least square
(LS) method was not used mainly because it yields results that might have higher maximum errors opposed to the
iterative method. The least squares method yields better results over the average error than the iterative algorithm.
The main purpose of the developed algorithm was to minimize the overall maximum error around each calibration
point. In order to improve overall accuracy more calibration points provide a larger sample and therefore ensure
more reliable and accurate correction factors.

3. Results

Optimized joint offset values Δθ were acquired from 24 calibration points and are shown in Table 2. Constant
offset values compensate for the difference between the ideal kinematic model and the actual robot and are added to
the joint vector θ.
462 Marko Švaco et al. / Procedia Engineering 69 (2014) 459 – 463

Table 2. Calibrated joint offset values.

Δθ 1 (deg) Δθ 2 (deg) Δθ 3 (deg) Δθ 4 (deg) Δθ 5 (deg) Δθ 6 (deg)

-0.538 +1.077 -0.163 +1.090 -0.286 +0.559

The maximum error prior to calibration was reduced after optimizing the offset values of the initial kinematic
model as shown in Table 3. A comparison between errors in the robot reference frame around calibration points is
presented.

Table 3. Absolute positioning errors prior and after calibration.

x (mm) y (mm) z (mm) E (mm)
Initial kinematic model 2.01 1.25 2.75 3.63
Optimized kinematic model 0.80 0.66 0.76 1.29
Δ (initial – optimized) 1.21 0.59 1.99 2.34

Absolute positioning errors take into account translational displacements (x, y, z) and Euclidian distances (E) of
the maximum deviation around each calibration point. Angular displacements of ideal robot pose and actual robot
pose will be part of future research.

4. Conclusion and future work

The proposed approach for robot calibration to increase absolute accuracy has proven good results. Maximum
error prior to calibration was 3.63 mm and 1.29 mm after optimizing the offset values of the initial robot kinematic
model.
The optimizing procedure of offset values, which can be directly imported in the KUKA robot controller, has a
limited possibility of minimizing the absolute positioning error. In Table 4. preliminary simulation results are shown
where 30 robot kinematic parameters were optimized. Six new parameters were introduced, an encoder scale factor
for each robot joint reading.

Table 4. Calibrated values of all robot DH parameters with supplied encoder scale values.
d (mm) θ (deg) a (mm) α (deg) encoder scale
Link 1 400.51 θ1-0.753 23.65 90.43 0.99908
Link 2 0.16 θ2+0.969 456.95 -0.28 0.99871
Link 3 0.41 θ3-0.235 35.95 89.81 0.99883
Link 4 420.28 θ4+1.090 0.25 -89.97 1.00025
Link 5 0.05 θ5-0.286 -0.95 89.97 1.0000
Link 6 80.09 θ6+0.404 0.80 180.19 0.99908

With the optimized model the accuracy around calibration points has been further improved, with a maximum
error of 0.74 mm. Parameters from Table 4. cannot be directly imported to the robot controller. A kinematic solver
needs to be developed for processing the new model definition. The error of 0.74 mm was obtained as a simulation
result from MATLAB software and still needs to be validated on the actual setup. In Fig. 2. an overview of all
calibration accuracy results is presented.
 Marko Švaco et al. / Procedia Engineering 69 (2014) 459 – 463 463

Fig. 2. Calibration results overview (values in mm).

As calibration requires manual adjustment of different poses it is a time consuming process. Future development
of the calibration procedure will include automatic calibration in a multiagent robot system [10,11,12]. The
calibration sphere will be manipulated by another robot and the calibration procedure will be done autonomously.

Acknowledgements

The authors would like to acknowledge the support of Croatian Institute of Technology, Croatian Ministry of
Science, Education and Sports for funding the project grant TP-11/0120-56: Robotic application in neurosurgery and
the Croatian Ministry of Science, Education and Sports for funding the project No.: 012-1201948-1941:
Autonomous multiagent automatic assembly.

References

[1] B. Siciliano, L. Sciavicco, L. Villani, G. Oriolo: Robotics: Modeling, Planning and Control, Springer, London, 2010.
[2] R.P. Paul: Robot Manipulators: Mathematics, Programming and Control, The MIT Press, 1981.
[3] C.S. Gatla, R. Lumia, J. Wood, G. Starr: Calibration of Industrial Robots by Magnifying Errors on a Distant Plane, Proceedings of the 2007
IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 3834-3841.
[4] A. Goswami, A. Quaid, M.Peshkin: Complete Parameter Identification of a Robot From Partial Pose Information, International Conference on
Robotics and Automation,pp. 68 - 173, IEEE, (1993)..
[5] D. Zhang, Z. Gao: Optimal Kinematic Calibration of Parallel Manipulators With Pseudoerror Theory and Cooperative Coevolutionary
Network: Transactions on Industrial Electronics, Vol.59, No. 8., (2012), IEEE, pp. 3221-3231
[6] D.E. Whitney, C.A. Lozinski, J.M. Rourke: Industrial Robot Forward Calibration Method and Results, Journal of Dynamics Systems,
Measurement, and Control,Vol.108/3 (1986), pp. 1-8.
[7] K. Radkhah, T. Hemker, O.V. Stryk: Self-Calibration for Industrial Robots with Rotational Joints: International Journal of Mechatronics and
Manufacturing Systems, Vol. 3, No. 3/4, (2010), pp. 187-209 .
[8] J. Liu, Y. Zhang, Z. Li: Improving Positioning Accuracy of a Neurosurgical Robot System: Transactions on Mechatronics, Vol. 12, No. 5.,
(2007), IEEE, pp. 527- 533.
[9] G. Duelen, K. Schroer: Robot Calibration – Methods and Results. Robotics and Computer-Integrated Manufacturing, Vol. 8 (No. 4), pp. 223-
231, (1991).
[10] M. Svaco, B. Sekoranja, B. Jerbic: A Multiagent Framework for Industrial Robotic Applications, Cihan H. Dagli (ed.), Procedia Computer
Science, Vol 6, pp. 291-296, Elsevier B.V. (2011).
[11] M. Svaco, B. Sekoranja, B. Jerbic: Autonomous Planning Framework for Distributed Multiagent Robotic Systems, Proc. of The 2nd Doc.
Conf. on Computing, Electrical and Industrial Syst., Caparica, Portugal, Luis M. Camarinha-Matos (Ed.), pp. 147–154, Springer, (2011).
[12] T. Stipancic, B. Jerbic, P. Curkovic: Context-Aware System Applied in Industrial Assembly Environment, International Journal of Advanced
Robotic Systems, Antonio Visioli (Ed.), Vol 9. ISBN: 1729-8806, InTech, DOI: 10.5772/53235. (2012).