See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/256765848

Automatic Robot Joint Offset Calibration

Conference Paper · June 2012

CITATIONS READS

4 1,534

1 author:

Morten Lind
SINTEF
20 PUBLICATIONS   85 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Autocut View project

All content following this page was uploaded by Morten Lind on 22 May 2014.

The user has requested enhancement of the downloaded file.

Automatic Robot Joint Offset Calibration

Morten Lind

SINTEF Raufoss Manufacturing AS, Trondheim, Norway

Abstract A contact-based calibration technique using a robot-mounted calibration

probe, has been experimented with. The probe is to be implemented by utilizing
a wrist-mounted force sensor. The technique is based on constraining the sampled
robot joint poses to those where the probe end is in controlled contact with a planar
surface.
The paper illustrates how the technique may be used to achieve a joint offsets iden-
tification to an accuracy comparable to the accuracy with which the joint and tool
positions can be measured. The joint offsets are the main objective for calibration
since it is the main source for inaccuracy in modern industrial robots.

Keywords Robotics, Automatic Calibration, Operation Management, Maintenance,

Autonomous Production Systems

1 Introduction
High accuracy is only required by certain types of robot applications. The major-
ity of the simple teach-based industrial pick-and-place operations need only high
precision, which is inherent in industrial robots.
However, in operations where the robot is to give 3D poses with external refer-
ence it must be calibrated to the required accuracy; e.g. eye-in-hand setups based on
scanners or cameras.
For robot installations used in autonomous production systems the control soft-
ware may monitor the performance and estimate the accuracy of the robot used in
the operations. Based on this, a calibration operation may be scheduled together
with planned operations. The calibration operation should thus preferably utilize
a technique that allows it to proceed in an autonomous fashion, without requiring
manual interaction or shutting down of surrounding systems.
When targeting industrial production systems it is desirable to keep the compu-
tational complexity manageable and focus on stability. Fortunately, the mechanical
or kinematic parameters may for most industrial robots be considered of good nom-
inal accuracy, and low accuracy of an industrial robot is mainly ascribed to the joint
zero offsets.
The subject of robot calibration has been extensively treated by the scientific
community during the past three decades.
Hollerbach and Wampler [1996]; Hollerbach et al. [2008] give a thorough overview
and details of the mathematical and computational foundation of kinematic calibra-
This work has been supported by the SFI Norman and the KMB Next-Generation Robotics
research programmes; both of which are funded by the Research Council of Norway.
tion. They direct specific attention to robots, and introduce concepts of open and
closed loop calibration applicable for serial as well as parallel robot structures.
The unification of open and closed loop calibration methods described by Wampler
et al. [1995] allows for using implicit or partial pose measurement for calibration.
This effectively eliminated the need for high-cost measurement technologies tradi-
tionally required by pure open loop methods.
Regarding calibration with restriction of an end-effector point to a plane, [Zhuang
et al., 1999] stress the fact that a single planar constraint is not sufficient to guarantee
observability of all parameters to be calibrated. They further demonstrate that three
non-degenerate planar constraints can be considered equivalent to an unconstrained
position measurement device; regarding calibration.
Low-cost measurement techniques based on laser pointers have been reported
for implicit, partial closed loop calibration. Gatla et al. [2007] and Liu et al. [2009]
present automated methods where a laser pointer is mounted in the robot end-
effector. They use a camera and a position sensitive device, respectively, for identi-
fication of the precise position sampling of the laser spot.
Though most literature address calibration of extensive models of a robot, Liu
et al. [2009] single out and focus on identification of joint offset calibration. Their
focus on joint offset calibration stems from an early report by Judd and Knasinski
[1990] which claims that almost 90% of the error in the accuracy of the robots can
be ascribed to joint offset calibration.
This paper presents shortly the setup used for sampling joint space positions
for calibration; Section 2. Then the mathematical and computational techniques are
introduced in Section 3. Section 4 gives an overview of emulated experiments and
calibration results. Finally, some concluding remarks are given in Section 5.

2 Calibration Setup
The calibration setup uses a robot, a calibration plate, a calibration probe tool, and
a force sensing device. The experimental setup is modelled and emulated in a phys-
ically realistic manner, providing a platform for the real calibration application sys-
tem to be developed and tested.

2.1 Calibration Plate and Probe

The setup used for developing a calibration technique suitable for integration with
production operations is illustrated in Figure 1. Two types of setups are illustrated.
Figure 1b illustrates a setup where a large separate plate is set up within reach of
the robot end-effector and Figure 1a illustrates a smaller plate mounted fixed with
the robot base. These two different types of plate setups both have advantages and
disadvantages.
The base mounted plate is inviting due to the mechanically fixed mounting
method with the robot base. This allows for accurate mounting and full a priori
knowledge of the calibration plate surface for calibration analysis. However, since
the plate is considered a fixed installation with the robot itself, a major drawback is
that the plate needs to be quite small to not interfere with workcell design.
 (a) Mounting of a small plate at the (b) Mounting of a standing plate
robot base. within reach of the robot.

Fig. 1 Calibration setup showing a Universal Robots UR5 robot, contact probing tool,
and two different calibration plate setups.

A calibration plate mounted on a separate stand in the workcell area, within

reach of the robot end-effector, is more tempting with respect to getting a larger
sampling space for the robot. This will improve the quality of the calibration analy-
sis and hence give better results. However, the plate, or possibly several plates, can
not be regarded as a simple element belonging to the robot. It must be given separate
concern with every robot installation in every workcell setup for all manufacturing
operations.
The calibration tool used is a precision-machined peg, which is adequate for
mounting either directly in the robot tool flange, as illustrated in Figure 1, or on
the output flange of a wrist mounted force sensor. It is a cylindrical steel rod, of
radius 0.01 m with a hemispherical surface of the same radius at the probe end and
a threaded tap for mounting, of size M6, at the robot end-effector end.
The focus on establishing mechanical contact between probe and calibration
plate by pure force sensing is based on a conjectured pre-adaptation. It is assumed
that most robots in future flexibly robotized manufacturing cells will incorporate
accessible force measurement capabilities.

2.2 Emulated System Setup

The system of robot, calibration probe tool, calibration plate, and force-based me-
chanical contact sensor has been set up in the Blender Game Engine, giving a con-
trol platform for operating the robot and identifying contact between the calibration
probe and the calibration plate. The setup used is the one illustrated in Figure 1a.
The use of the Blender Game Engine for real-time, realistic emulation of robot-
based workcell setups has been demonstrated by Lind and Skavhaug [2011].
The basic real-time robot control is realistically emulated by using a model from
the PyMoCo robot control framework. The calibration operation application, fully
dissociated and distributed from the emulated setup, also uses the PyMoCo frame-
work for Ethernet-based communication and control of the emulated or physical
setup.
The calibration probe sensor for mechanical contact identification is emulated by
a very simple computation of the tool centre point and its relation to the calibration
plate surface. A priori exact knowledge of the calibration plate and the calibration
probe tool makes this emulation very simple based on computation. A more complex
geometrical surface of the calibration plate or a priori unknowns or uncertainties in
the geometrical setup would require the emulation to change to a model for sensing
based on geometric collision detection in the Blender Game Engine.

2.3 Calibration Application Logic

Working either with the real robot, force sensor and physical setup, or with the em-
ulated setup, the calibration application operates the robot to continually sample the
joint space position at random tool spaces poses of contact between the calibration
probe and the calibration plate. From the a priori known pose and size of the cali-
bration plate, random, collision free tool space poses of the robot may be generated.
These tool space poses must be generated with adequate limits on the tool frame
position and orientation, such that the tool may be moved at low speed towards, and
to contact with the, calibration plate, without any other collisions.
The calibration sampling operation of the robot then consists in generating a
predefined number of joint space positions of random tool frame pose where the
calibration probe is in contact with the calibration plate. The calibration sample set
of joint space positions is then the input to the calibration analysis computation.

3 Computational Technique
The computational techniques used for analysing and calibrating the sampled robot
configurations generated with the emulated system setup rests on the techniques
described by Ikits and Hollerbach [1997] and Zhuang et al. [1999].
The kinematic computation system is set up to compute with the centre of the
calibration probe hemisphere as the controlled tool centre point. This controlled
point, at a given joint space position q, and computed from the forward kinematics,
is denoted p̃(q).
The calibration plane is offset by the probe hemisphere radius from the cali-
bration plate surface in the direction of the plate outward normal. When contact is
established between probe and plate, it is known that the real tool centre position,
p, lies in the calibration plane. For a joint space position in the sample set, qi , the
real probe centre position is denoted by pi . The corresponding position based on
the joint space position and forward kinematics is denoted by p̃i = p̃(qi ).
For representing the calibration plane, a plane vector is used; as also used by
Ikits and Hollerbach [1997]. A plane vector is a linear operator, p n, such that for
any point, p, on the plane p np = 1. For points off the plane, p np gives an error
measure along the normal direction of the plane.

3.1 Calibration Equations and Procedure

The calibration analysis addressed here assumes that the major part of the error
between real and computed tool centre positions under constraint of the calibration
plate may be ascribed to joint offset errors of the robot. It is assumed that the joint
zero offsets, qoff , may be found to minimize the error between real and computed
tool centre positions:
( )
X 2
min (p̃(qi + qoff ) − pi ) (1)
qoff
i

Given the known plane vector, p n, it is known that the identities p npi = 1}
holds. Given the forward kinematics it is possible to compute the plane vector errors,
ei , from the samples by the defining identity p np̃i = ei . By subtraction the plane
vector error and the plane vector identity gives the difference equation for each
sample:
p
n(pi − p̃i ) = 1 − ei (2)
Under the assumption that small joint offsets corrects the kinematics, it may be
approximated that pi ≈ p̃(qi + qoff ). Using the position part of the manipulator
Jacobian, Jpi = Jp (q), a Taylor expansion to first order in qoff may be set up as
p̃(qi + qoff ) − p̃(qi ) ≈ Jpi qoff . Together this transforms Equation 2 to
p
nJpi qoff = 1 − ei (3)

Stacking Equation 3 over all samples will, for a sufficient number of samples,
result in an over-determined system of the linear form Ax = b, which solves the
minimization problem given in Equation 1. It is of a suitable form to be solved by
linear least squares with Newton-Raphson, and this method has been implemented.

3.2 Noise and Pose Set Selection

Parameter observability and noise sensibility are central concepts of statistical ap-
proaches to robot calibration. Typically a singular value-based analysis for the con-
dition of the lefthand side operator in the stacked version of Equation 3 is used, with
different observability measures. Hollerbach and Wampler [1996] list three differ-
ent observability measures that have been reported in the literature. In the work
presented here the minimal singular value has been chosen because it resulted in
clear and consistent correlation with the real error of the resulting calibration.
Input noise is, in the case of robot control, the error or noise in reading the joint
encoders. An upper bound of the level of the input noise for a given robot may
be estimated by the tool space precision, commonly called repeatability. A robot
with a reach of 1 m and a precision of 0.0001 m would have an upper bound of
the joint precision around 0.0001 rad. It would thus be reasonable to presume the
input noise, qnoise , as uniform or Gaussian distribution in the approximate interval
[−0.0005 rad, 0.0005 rad].
The output noise is the precision of the external metrological system. In the pre-
sented work, the precision of a force sensor based approach to establish a controlled
mechanical contact, is currently unknown. It is presumed that a positional error of
the tool centre point in relation to the calibration plate will be attainable within the
range between 0.0005 m and 0.002 m in the direction normal to the calibration plate.
Hence, it is not unreasonable to assume that the noise, pnoise , may follow a uniform
or Gaussian distribution in the interval of approximately [−0.001 m; 0.001 m]. The
achieved precision will probably depend heavily on the speed of operation, and it
is assumed that it will be possible to tune the method and implementation of the
contact system to target some positional precision within this range.

4 Experiments and Results

A calibration application operating the emulated robot using the emulated probe
tool on the a priori known calibration plate was implemented to make samples at
different noise levels of the probe tool sensor system. Calibration operations of sam-
pling 300 configurations at positional noises with un-biased uniform distribution
in [−δpnoise ; δpnoise ] for δpnoise ∈ {0.0005 rad, 0.001 rad, 0.002 rad} has been
conducted. Over the data sets for different noise levels, an orthogonal cross of data
sets for calibration analysis are generated with input noises. The input noises are
added to all joint values and are also taken from an un-biased uniform distribution
in [−δqnoise ; δqnoise ] for δqnoise ∈ {0.0005 m, 0.001 m, 0.002 m}.
All experiments were done with setting a joint offsets vector to some realistic
fixed values. The calibration analysis of the data sampled on the emulated setup may
then compute the root-mean-square (RMS) error of the identified qoff with respect
to the real joint offsets with which the emulated robot was operating.
The calibration analysis modifies the raw sample sets of 300 pose measurements
with different position noise, to generate the full matrix of sample sets of 300 sam-
ples for each combination of position and joint reading noise. Over this set of data
samples, for each noise level, is drawn random sample sub-sets of various sizes to
determine the quality of calibration (the RMS error), the correlation of number of
pose samples used with the condition number, and the correlation of the condition
number with the RMS error. For each combination of sub-sample size, joint read-
ing noise, and position noise, a number of calibrations are performed to generate
statistics. The presented results performs 10 such calibration computations under
identical circumstances to generate a statistic base.
Figure 2 shows plots for calibration sample size, condition number, and RMS
error of the calibration data set of 300 samples with δqnoise = 0.0005 rad and
δpnoise = 0.001 m. These position measurement and joint reading noise levels are
considered to be realistically achievable.
Three important observations are obtained by analysing the calibration results,
and this is consistent over all computational results of all of the position and joint
reading noise data sets:
– The condition number has an dominant monotonic behaviour with respect to
sample size used for calibration; see Figure 2a.
– The RMS error has a dominant behaviour of a decreasing upper bound with
increasing condition number; see Figure 2b.
– The rise in condition number with sample size seen in Figure 2a and the de-
creasing upper bound on RMS error with increasing condition number seen in
Figure 2b, though flattening out, does not end within the 300 maximum sub-
sample sizes used in this analysis.
 Condition vs. sample size for p_noise=0.0010 and q_noise=0.0005 0.06 RMS error vs. condition for p_noise=0.0010 and q_noise=0.0005

0.7
0.05
0.6
0.04
Condition number

RMS Error [rad]

0.5
0.03
0.4

0.02
0.3

0.2 0.01

0.1 0.000.0
50 100 150 200 250 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Sample size Condition number

(a) Plot of condition number versus sam- (b) Plot of RMS error in the identification
ple sizes. Average error and standard of real joint offsets versus condition num-
deviation in red. ber.

Fig. 2 Illustration of calibration results in the assumed realistic case of realistic noise
levels: δqnoise = 0.0005 m and δpnoise = 0.001 m

In addition, for the pertinent case of realistic noise levels underlying Figure 2b,
the best achievable upper bound of the RMS error is on the lower side of 0.005 rad.
This corresponds realistically to a worst case 0.004 m end-effector error on a UR5
robot; with the worst case being all error on one of the shoulder joints.
Figure 2b suggests that keeping the condition number beyond 0.4 results in an
upper bound to the RMS error of approximately 0.01 rad. Figure 2a indicates that
for achieving a condition number beyond 0.4 a sub-sample of approximately 100 is
necessary.
Taking this as a rough rule of thumb, an error analysis can now be set up for
conveying a rough impression of the achievable accuracy of calibration at various
noise levels, when restricting to a sub-sample size of 100. For generating good statis-
tics, for a fixed noise pair, viz input and output, the worst case calibration accuracy
from the groups of sub-sample sizes in the interval of [70; 130] will be averaged
to represent an estimate of the achievable upper bound to the RMS error around a
sub-sample size of 100. The results are displayed in Table 1.

Table 1 RMS calibration error for realistic levels of input (qnoise ) and output (pnoise ) noises.
Error values are computed as averages of worst case errors for the sub-sample sizes in the
interval [70; 130].
aa
aa pnoise 0.0005 m 0.0010 m 0.0020 m
qnoise aa
a
0.0005 rad 0.00570 rad 0.00983 rad 0.01627 rad
0.0010 rad 0.00922 rad 0.01215 rad 0.01910 rad
0.0020 rad 0.01791 rad 0.01670 rad 0.02248 rad
 5 Conclusions and Future Work
It has been demonstrated that, at realistic ranges of input and output noise, the
method described in this paper may bring a UR5 robot to an accuracy of about
0.005 m. As the current accuracy of robots from the manufacturer is presumed to be
some cms, the achieved accuracy is a great improvement.
Activity is proceeding to set up a real system for calibration by the method
described in this paper. The most critical challenge is the development of a system
for identifying mechanical contact between probe and calibration plate.
The presumption, based on the account by Judd and Knasinski [1990], may not
be entirely true. Hence, it may not be sufficient to simply perform a joint offsets cal-
ibration, but also calibration of some identified critical kinematic parameters needs
to be incorporated with the method.
The assumption of uniform or Gaussian noise may not hold, especially for the
output noise; i.e. the probe measurement. Biased noise in sampling may have to
be addressed in the real system, by incorporating techniques such as described by
Hollerbach and Wampler [1996].

References
Gatla, C.S., Lumia, R., Wood, J., Starr, G., (Dec 2007). An Automated Method to Calibrate
Industrial Robots Using a Virtual Closed Kinematic Chain. Transactions on Robotics
vol. 6, no. 23, pp. 1105 –1116
Hollerbach, J.M., Khalil, W., Gautier, M., Model Identification. In: Siciliano, B., Khatib, O.
(eds.) Springer Handbook of Robotics, pp. 321–344. Springer Berlin Heidelberg (2008),
http://www.springerlink.com/content/r6m219/
Hollerbach, J.M., Wampler, C.W., (1996). The Calibration Index and Taxonomy for Robot
Kinematic Calibration Methods. International Journal of Robotics Research vol. 6, no. 15,
pp. 573–591
Ikits, M., Hollerbach, J.M., (Apr 1997). Kinematic Calibration Using a Plane Constraint. In:
International Conference on Robotics and Automation. vol. 4, pp. 3191–3196. IEEE
Judd, R.P., Knasinski, A.B., (Feb 1990). A Technique to Calibrate Industrial Robots with
Experimental Verification. IEEE Transactions on Robotics and Automation vol. 1, no. 6,
pp. 20–30
Lind, M., Skavhaug, A., (2011). Using the blender game engine for real-time emulation of
production devices. International Journal of Production Research vol. 0, no. 0, pp. 1–17,
online available, iFirst
Liu, Y., Xi, N., Zhang, G., Li, X., Chen, H., Zhang, C., Jeffery, M.J., Fuhlbrigge, T.A., (Oct
2009). An Automated Method to Calibrate Industrial Robot Joint Offset Using Virtual
Line-based Single-point Constraint Approach. In: International Conference on Intelligent
Robots and Systems. pp. 715–720. IEEE/RSJ
Wampler, C.W., Hollerbach, J.M., Arai, T., (Oct 1995). An Implicit Loop Method for Kine-
matic Calibration and its Application to Closed-Chain Mechanisms. Transactions on
Robotics and Automation vol. 5, no. 11, pp. 710–724
Zhuang, H., Motaghedi, S.H., Roth, Z.S., (May 1999). Robot calibration with planar con-
straints. In: International Conference on Robotics and Automation. vol. 1, pp. 805–810.
IEEE

View publication stats