A New Method for Autonomous Robot Calibration
Xiao-Lin Zhong and John M Lewis, Members, IEEE
Dept. of Mechanical, Manufacturing & Software Engineering
Napier University, Edinburgh
10 Colinton Road, Edinburgh, Scotland, EH10 5DT, UK
Abstract
A new method for autonomous robot calibration is
presented which is suitable for on-site calibration in an
industrial application environment. Using a trigger probe
as an extension of the manipulator link, robot internal
joint sensor measurements were recorded for kinematic
identification while the robot tip-point was in contact with
a constraint plane in its workspace. From the consistency
conditions of the constraint plane, the linear identification
equations were derived, from which the kinematic
parameters were extracted based on only robot internaljoint
readings without any external measurements of the end-
point locations. A recurrent neural network (RNN)
approach was applied to resolve the linear identification
problem. The RNN-based algorithm is computationally
more robust and eficient compared with conventional
numerical optimisation approaches. Both simulation and
experimental results for a six degree-of-freedom (DOF)
PUMA robot are presented in this paper.
1 Introduction
The positioning accuracy of a robot manipulator is a
major concern when using advanced off-line programming
methods in a robotic system. The desired locations of a
robot end-effector are normally specified in its workspace,
while these locations are achieved by controlling its joint
values in robot joint space. However, for a number of
reasons, including manufacturing tolerance, set-up errors,
wear and tear, transmission errors and compliance, the
internal kinematic model used in robot controller will not
accurately describe the kinematic transformation of the
actual robot. Therefore, the actual locations achieved by
controlling joint values obtained from the controller's
internal model will deviate from the desired locations. It
has been shown that as much as 95% of robot positioning
inaccuracy arises from the inaccuracy in its kinematic
model description[ll. Robot kinematic calibration is
defined as a process of improving robot positioning
accuracy through modification of its kinematic control
model without changing its hardware configurations[2].
There has been an extensive research concerning robot
kinematic calibration approaches over the past decades[11-
[ 5 ] . Kinematic calibration involves the following
procedures: i) Choose a proper kinematic model to describe
the relationship between robot joint space and its
workspace co-ordinates; ii) Take experimental
measurements of robot end-effector locations using external
IEEE International C o n f e r e n c e - 1790-
o n Robotics a n d Automation
0-78&3-1965-6/95 $4.W01995 IEEE
measuring devices; iii) Identify the parameter errors based
on the differences between the measured locations and those
predicted by the kinematic model. iv) Implement the
identified model. These procedures require extensive human
intervention and off-line processing, which precludes the
method from being used to perform on-site calibration in
an industrial environment at regular intervals. The
requirements of measurement phase are particularly
demanding, and expensive and complicated pose measuring
devices are needed for calibration (such as co-ordinate
measuring machines[51[6], theodolites[4], servo controlled
laser interferometers[7], sonic and visual sensors[3][12],
etc.). It has been reported that partial pose information is
sufficient for complete parameter identification[9]-[12].
Tang[9] utilised a special fixture to obtain partial pose
information of a robot end-effector. The fixture consists of
a flat plate with some accurately located points on it, and
an end-effector with a flat surface and some angle to the last
axis of the robot. In the 'free mode' of robot, the robot end-
effector was manually moved to the known points of the
plate and against the flat plate such that some component
of the end-effector position and orientation were 'measured.
The partial pose measurement scheme eases the
requirements for measuring devices, but it still requires
intensive human intervention during the set up and
measuring process. In addition, as pointed out by
Driels[l2], not every robot provides a 'free mode' in which
the manipulator can be moved manually while the power is
on and the joint servo system is off.
Autonomous robot calibration is defined as the
automated process of determining a robot's model by using
only its internal sensors. It has been observed that
autonomous calibrations are possible for robot
manipulators with either some a priori knowledge of the
task constraint or redundancy of the sensing systems(e.g.,
two robots gripping together)[l3]. The autonomous
calibration of hand-eye systems has also been performed by
using robot joint readings and camera co-ordinate
measurements to form the closed-loop constraints[ 131[141.
Another kind of task constraint has been proposed for robot
kinematic parameter identification which utilises laser line
tracking in robot workspace[l5]. In this paper, we present a
new robot kinematic calibration scheme which can be
implemented autonomously and is suitable for on-site
calibration in an industrial environment. Gripping a trigger
probe(Figure 2), the robot uses the probe as its extended
link to touch constraint planes in its workspace(the
locations of the constraint planes are not necessarily known
exactly). Only the robot joint readings are recorded for
identification while the tip-point of the probe is touching
the constraint plane from various configurations. A linear
identification model has been derived from the consistency
conditions of a plane, and is presented in the next section.
To improve the robustness and efficiency of the
identification algorithm, a Hopfield type recurrent neural
network[ 161 was applied instead of the conventional
numerical least square methods. The neural network
algorithm for identification is presented in section 3. In
section 4 we describe the data collection method used and
the experimental set-up. Both simulation and experimental
results for a PUMA 560 robot are given in section 5 ,
assuming a special case in which we place the constraint
plane so that it aligns with the robot base axis.
Discussions and conclusions are given in the final section.
2 Formulation of the Kinematic
Identification Model
The objective of kinematic calibration is to identify the
kinematic parameter errors Ap which are assumed to be a
small perturbations from its nominal values specified by
the robot manufacturer. Without external measurements,
the actual position Pi = [niyizilT of robot end-point
remains unknown but must be near the nominal position
Pi0 predicted by the robot controller. Using the Taylor
series to the first order, we have:
Pi = Pio+ JiAp = [x? y? z ? ] ~+ [Ji"JiY JtITAp
(1)
Where Jiis the robot special Jacob which can be calculated
with robot joint readings and nominal kinematic
parameters[8]; and i = 1, 2, ..., m is the subscript index
representing different touch points of the robot.
Figure 1. Constraint conditions for co-planar points
Although the exact locations of touch points are
unknown, they are constrained to lie on a plane. The
consistency condition of a plane leads to the construction
of the identification model (Figure 1).The difference vector
between two consecutive touch points is:
APi = Pi - Pi.1
= [AxiO Ay? Az?]T + [AJi" AJiy AJ:lTAp
where
Ax? = xiO - ~ i . 1 ' ; Ayio = yio - yi-1'; Az? = zi0 - zi-1' ;
AJi" z z Ji" - Ji-1' ;AJiY = JiY - Ji.lY ; AJ? = Jt - Ji-1' .
Every three touch points will specify a norm to the
constraint plane, which is the cross product between two
consecutive difference vectors. Ignoring the second order
terms of .4p, we have:
Z, = W i + l x APi = [U@ bOi coilT+ [ ~ l bli
i clilTAp (3)
where
aOi = A\yioAZi.,lo- A Y ~ + ~ O A Z ~ " ;
bOi = AzioAXi.+lo- A Z ~ + ~ O A X ~ ~ ;
c O ~ AxiOAyi+io - AXi+loAyiO;
a l i = (Ay?AJi+lz c Azi+loAJiY)-(Ayi+loAJ~+ Az,OAJi+,Y);
b l i = (AZ?AJi+l" t,A n i + ~ o A J ~ ) - ( A z i + l o A+J iAx x ? A J ~ + ~);'
cli = (Ax~OAJ~+~Y + Ayi+l0AJiX)-(Ani+1OAJiY+ AyiOAJi,,').
Similarly, we have the expression of the next normal
vector of the Same plane:
&+I = [aOi+lbQ+l cOi+lIT+ [ a l i + l b l i + l cli+1lTAp.(4)
The consistency condition for one plane is that its
norms must biz parallel, i.e, the cross product between each
pair of norms must be a zero vector. Similar to (3), we
have:
&+I x Zi = [AOi BOi CoilT+ [Ali Bli CliITAp
=o (5)
where
AOi = bOi~Oi+l- bOi+lcOi;
BOi = coiaoi+1- cOi*lUOi;
COi = UOibOi+l - uOi+lbOi;
A l i = (bOicli+l + ~ O i + l b l i-) (bOi+lcli + cOibli+l);
Bli = (COioUli.,l+ UOi+lCli ) - (COi+lUli + uOicli+l);
Cli = (UOibli,l + bOi+lUli) - (UOi+lbli + bOiUli+l).
Denoting H =: [AIl B11 Cll ...AIm B1, CImIT; and
AX = [A01 BO1 CO1 ... AO, BO, CO,lT,
We have a linear system:
In ( 6 ) , the coeffcicient matrix H and AX can be
calculated based on the difference of nominal positions
predicted by controller, and the difference of special Jacob
for each pair of consecutive touch points(on1y joint
readings are required for the computations). The only
unknown remaining is the kinematic error to be identified.
Eqn.(6) serves as tJhe linear identification model of the new
method.
As a check of the constraint conditions of (3,assuming
we have the knowledge that the constraint plane is aligned
with the robot base x-y plane, the z component of AP, will
be zero. From (2), we have:
AJ,"Ap + AZ? = 0 (7)
Using the above equation and ignoring the second order
terms of Ap, it is easy to show that the x, y components
of (3) are zero (which agrees with the physical implication
that the normal to the plane is perpendicular to the base n-y
(2) plane) and the relationship of (5) holds.
Although the exact location of the constraint plane is
not necessarily known, care must be exercised in placing
1791-
the constraint plane in the workspace. Considerations
include:
the robot configurations enabling desirable and safe
touch on the plane;
* the workspace in which accuracy is critical;
0 the optimal identification configurations of the robot;
the workcell layout and the kind of constraint plane
available.
As robot base axes are always set up to align with
some reference planes in the workcell, the assumption
made earlier, in which the constraint plane is aligned with
the robot base co-ordinates, has practical significance. In
the case that external constraint planes are used, it is also
easy to align the plane with the robot base axis using the
probe and VAL I1 axis motion function. To maximise the
robot movement ranges, it is desirable to have the robot
touch the constraint planes separately which lie
perpendicular to the robot's base axes. The linear
identification model in this case can be constructed
according to (7), using the correct component of (2) at each
point.
3 RNN- b ased Identi fication A 1 gorithm
The linear identification model (6) can be resolved by
using any standa-d linear least square method such as
Levenberg-Marquart algorithm. It is known that the
Hopfield type recurrent neural networks@") can be
employed to solve a class of optimisation problems
[16][18]. The advantages of RNN method over the
conventional numerical approaches are its computational
efficiency and robustness, due to the parallel computation
in hardware implementation, and the use of differential
equations in simulation instead of difference equations as
used in numerical optimisation algorithms[ 181. In our
earlier work[l7], we developed a l2"-based calibration
compensation algorithm which exhibited the numerical
robustness near robot singular configurations. We now
present a similar algorithm for the kinematic identification
problem, which might be ill-conditioned due to the limited
range of robot movement during data collection. The
resolution of (6) is equivalent to the minimisation of the
following energy function in the least square sense:
E=+[HAp+AX]T[HAp+AX]+iApThAp (8)
Where is a positive diagonal weight matrix for
regulation. Writing this in the form of Hopfield network
energy, we have:
n n in
E = - + ~ T ~ A P ~- A
i,j=l
c lP , ~~ p ~
j=1
+ + C ( A(9)X ~ ) ~
i=l
wheiE
l . For UNIMATION PUMA robot in this case
- 1792
m
and Ij = - c H i A & , which determine the network
i=l
connection weights and input currents respectively based on
the nominal kinematic model and parameters. The Apj is
thej-th component of kinematic parameters to be identified
which corresponds to the j-th neuron state. n is the
dimension of the vector of kinematic error while m is the
row number of coefficient matrix H .
The energy function gradient with respect to the i-th
neuron state can be obtained from (9):
(10)
The time evolution of neuron states should follow' the
opposite direction of the energy function gradient, so we
have the neuron circuit dynamics equation:
d ( A p i )- p j ( $ T v A p j
-- +Ii)
dt j=l
where p i is the z-th diagonal element of the positive
diagonal coefficient matrix p which is chosen to ensure the
stability and convergence speed of the neuron circuit. Given
the initial condition of the neuron states(Apj= 0, i = 1, ...,
n ) , the above ordinary differential equation (ODE)
determines the neuron state trajectories, and hence the
kinematic errors to be identified (the stable states of the
neurons). Eq. (11) is actually a linear differential equation
of high gain which can be solved by any standard ODE
method. To improve the robustness against measurement
noise and numerical perturbations, some non-linearity can
also be incorporated into (1 1) through the introduction of
the neuron "sigmoid" to the network[l8]. Another
advantage of the R"-based method is the potential
implementation of parallel computation which leads to on-
line identification of kinematic parameters, which is
independent of the robot DOF and the number of
parameters to be identified.
After kinematic errors have been obtained through (1l),
the kinematic parameters are updated in the forward
kinematic model and the neuron inputs and weight
connections are updated correspondingly. The above
procedure is repeated until the algorithm converges to the
prescribed accuracy. The Hopfield type recurrent neural
network can converge to the global optimal given the
specific initial conditions which are in the vicinity of its
optimal solutions[l6]. Since the kinematic errors are only
small perturbation from the nominal parameters, the above
procedures are guaranteed to converge provided the
kinematic errors are small. Normally three iterations are
sufficient to get the desirable results.
4 Data Collection
Renishaw Probes were originally used for accurate
workpiece set up and workpiece measurement for CNC
lathes. The probe is in effect an omni-directional switch
that triggers when the probe contacts the workpiece from
any direction[l9]. For our application, a special tool
changer was made to hold the probe (Figure 2). The trigger
signal is transmitted as an input to control the robot. The
switch is kept on while the probe is in contact with the
object. The tip-point of the stylus is a ruby ball so that the
contact point from any direction is a constant distance from
the centre of the tip. The stylus has some amount of over-
travel in each direction allowing a certain probing speed.
The data collection procedure written in pseudo-code is as
foollow s:

Repeate until max. number of touch points reached{

MoveTo(START);
Point = RandomConfiguration(MAX,MIN);
MoveTo (Point);
while(ProbeSignal == OFF){
ProbeBy(XSTEP,YSTEP,Z S E P ) ;
1
while(ProbeSignal==ON){
ProbeBy(-XS TEP/ 10,-YSTEP/l 0,-ZSTEP/lO>;

RecordData(JointValues, CartesianValues);
1 Figure 2. Robot touch constraint plane using a Probe
The robot moves from a start point to a point above the 5 Results for a Puma 560 Robot
constraint plane where its configuration is randomly
generated within the robot movement ranges, from which a
The PUMA 560 robot is a six DOF manipulator with
desirable and safe touch on the constraint plane is ensured.2 six revolute joints. There are in total 24 kinematic
From that point, the robot probes the plane by moving in parameters, using Denavit-Hartenberg (D-H) notation, to
small steps toward the constraint plane until the probe describe the kinemtaic model. In general another six
signal is on. As there is some over-travel of the stylus, the parameters including three each of rotation{roll-pitch-yaw }
stop point is not the point of the first touch due to the and translatjon parameters are needed for TOOL
probing speed. Therefore fine tuning is needed to retract the transfoi-mation.The nominal parameters are listed in Table
first touch point. Whilst still in contact with the plane, the 1. The x-direction offset of the TOOL(probe) is
fine tuning process begins by moving the tip-point away intentionally built to make the robot wrist parameter
from the plane in steps oP one tenth of the probing steps. identifiable. \Ye also noted that there exists dependence
Then the joint values of the robot and the corresponding between the TOOL parameters and the last link parameters.
Cartesian co-ordinates are recorded for post-processing. The To eliminate the parameter redundancy, the TOOL
above process is repeated until the desired number of touch parameters are incorporated into the last link parameters,
points has been reached. regarding the probe as the extended last link of the robot.
The probing and fine tuning direction is along the The total number of kinematic parameters to be identified
normal to the constraint plane. The probing steps are is reduced to 24.
normally set less than 1 mm hence the fine tuning steps
are less than 0.1 "(the repeatability of' the Renishaw
Table 1 Nominal Parameters of a Puma 560 Robot
probe is in the order of pm), thus achievcs a measurement
accuracy in the order of 0.1 mm which is sufficient for LinkNo. a;h"m ddmm) ahd) TOOL
robot calibration(if more accurate measuremcnt is required, 1 0 0 -~ 1 2 O.O(rad)
smaller probing steps can be set). In the special case when 2 431.8 149.09 0 O.O(rad)
the constraint planes are aligned with robot base planes, the 3 -28.32 0 7d2 O.O(rad)
robot only probe in one axis direction, the movements 4 0 433.07 -7~12 5(")
along the other two axis direclions are set to zero. 5 0 0 rc/2 O.O(mm)
6 - 0 56.25 0 302.34(mm)

The simul,ation program was built to test the proposed

2 . This can be done either on-line using VAL I1 random
generator, given the intervals of Cartesian co-ordinates, or
through off-line planning process using robot simulation
package.
calibration algorithm. Without loss of generality, we
assumed a speciall case that the constraint planes were
placed so that they aligned with robot base planes. To
maximise the robot joint movement and make all the

1793
parameters "observable" from end-point positions, three level of robot repeatability after calibration. The cross-
constraint planes were placed perpendicular to each base evaluation has also been performed by collecting 20 test
axis. Forty random configurations in Cartesian space for points across robot workspace using precision external co-
each constraint plane were generated satisfying the ordinate measuring machine (CMM). The residual root
constraint conditions. Then the nominal inverse kinematic mean square errors for the test points are improved to 0.68
model was used to find the corresponding joint values for mm for position and 0.89 degree for orientation. The
each of Cartesian configurations. If there was no error in calibration results are consistent with the results of other
the kinematic model, it would be found that the positions calibration methods using sophisticated external
achieved by controlling those joint values would perfectly measurements.
match those constraint conditions. However, by inducing Table 2 Identified Errors of a Puma 560 Robot
small errors in the parameters(the induced errors were Link No. Aai(mm> Adi(") Acl,(rad) Aqi(rad)
randomly generated in the range of +2 mm for linear 1 0.761 0.031 0.0002 -0.0005
parameters and k0.02 radian for angular parameters), the 2 0.459 0.068 -0.0019 -0.0047
achieved positions by the 'actual' robot will deviate from 3 0.351 -0.091 0.0030 0.0024
the constraint planes(the dashed lines of Figure 3, the first 4 0.194 -0.445 0.0043 0.0176
40 points are the x component deviations of the 40 touch 5 0.026 -0.655 -0.0053 0.0036
points on the constraint plane perpendicular to x -axis from 6 0.075 -0.372 -0.0006 0.0017
the x-axis constraint plane; the next 40 points are the y
component deviations of the 40 points on the y-axis plane
from the y-axis plane and the last 40 are z component
deviations of the 40 points on the z-axis plane from the z-
axis plane). After calibrating the kinematic model using the
approach as presented in the previous sections, the
deviations from the constraint conditions are dramatically .......... I

reduced(the solid line in Figure 3). The identified errors are Before
almost identical to those induced. Calibration
To test the simulation results, the experimental set-up
was used for data collection(Figure 2). Three calibrated flat
planes were placed perpendicular to each of robot base axis After
for data collection. One hundred touch points on each of the Calibration
plane were collected. Sixty randomly chosen from each of
the 100 point were used for calibration and the remaining
points for independent test. Table 2 lists the identified
kinematic errors based on the calibration data. Figure 4 plot
-15 I I I I I I I
0 20 40 60 80 100 120
the x,y, and z component of the differences between each
two consecutive touch points on the same constraint Calibration Points
plane(the first 39 points are the difference of x components
obtained from the 40 test points on the constraint plane Figure 3. Simulation result with induced kinematic errors
perpendicular to the x-axis, and the next 39 are y
component differences obtained from the 40 test points on
the y-axis plane and the last 39 are z component differences 4
obtained from the 40 test points on the z-axis plane). The
dashed lines represent the differences of positions predicted
I I
by the un-calibrated model in the robot controller while the
solid lines represent the differences of positions predicted
by the updated model using the identified errors. The
symmetry of the graph is due to the use of the differences
between consecutive points. It is shown that the calibrated
model works well for test data points as well. Therefore it
reflects the true geometry of the robot rather than simply
the best fit of the calibration data. Using the reference
positions of the constraint planes perpendicular to the base Calibration
axis, we can evaluate directly the positioning accuracy
achieved by this calibration. Considering the positions
reported by the calibrated model as the reference positions I I I I I I
0 20 40 60 80 100 120
of the constraint planes, the accuracy comparisons based on
Test Points
the test points are given in Table 3. We can see that the
position accuracy of the robot has been improved to the
Figure 4. Test Result with experimental data

1794
 Table 3 Accuracy Comparisons Based on Test Points
Before Calibration After Calibration
(mm) dx 4 & d x dv d?
avg. 1.210 4.545 1.313 0.198 0.175 0.210
stdev. 0.196 0.737 0.213 0.032 0.028 0.034
max. 2.662 5.357 3.250 0.579 0.361 0.525
6 Conclusions and Discussions
A new autonomous robot calibration scheme has been
developed in this paper. Renishaw probes were originally
used for workpiece setup and measurements for CNC
lathes. We applied these cost effective sensors successfully
for robot on-site calibration in an industrial application
environment. Neither external measurements nor accurate
fixture set-up are needed for such a calibration. The
recurrent neural network-based parameter identification
algorithm is computationally more robust and efficient
than conventional least square methods. Both simulation
and experimental results for a Puma robot show that robot
positioning accuracy can be improved to the level of robot
repeatability.
To identify all kinematic parameters, normally more
than one constraint plane is needed for data collection.
Proper tool offsets are also needed to make the robot wrist
parameters identifiable. In the special case when constraint
planes are aligned with robot base planes, three constraint
planes peipendicular to each base co-ordinate axis are used
for full parameter identifications. Our future work will
investigate the optimal placement of the constraint planes
in the constrained environments so that the robot has the
optimal identification configurations. Constraint surfaces
other than planes may also be suitable for the proposed
calibration method (such as a spherical ball) as long as the
surfaces have known shapes and are suitable for touching
with the probe.
Acknowledgements
The first author would like to acknowledge the CVCP
and Napier University for financial support for this
research. Our thanks are due to Mr Bill Campbell for his
contributions during the experiments and Prof. A R Young
for his encouragement. We are also grateful to the
anonymous reviewers for useful comments.
References
[l] Judd, R.P. and Knasinski, A.B., 1991, "A technique to
calibrate industrial robots with experimental verification",
IEEE Trans. on Robotics & Automation, 6(1):20-30
[2] Mooring,B. W., Roth, Z.S. and Driels, M.R. 1991,
"Fundamentals of Manipulator Calibration", John Wiley &
Sons, Inc., PP. 221-225
[3] Stone, H.W., Sanderson,A.C. and Neuman,C.P., 1986,
"Arm Signature Identification", Proc. IEEE Int'l Con. on
Robotics and Automation, San Francisco, CA, PP. 41-48
- 1795
[4] Whitney, D., Lozinsky and Rourke,J., 1986, "Industrial
robot fonwartl calibration method and results", ASME J. of
Dynamic Systems, Measur. and Control, 108, PP. 1-8
151 Driels, lM.R., Swayze, LW and Potter L.S. 1993,
"Full-pose calibration of a robot manipulator using a
coordinate measuring machine", Int. J. of Adv Manuf
Technol 8: E'P. 34-41
[6] Zhong, X L , Lewis, J.M., and Rea, H., 1994, "Neuro-
accuracy Compensator for Industrial Robots", Proc. of
IEEE Int'l Con5 on Neural Networks, WCC1'94, Vol. 5 ,
PP. 2797-2802, Orlando, Florida
[7] Prenninger, J.R., Vincze, M. and Gander, H, 1993,
"Contactless jpositjon and orientation measurement of robot
end-effector", IEEE Int. Con$ Robotics and Automation,
Atlanta, Vol.1: PF'. 180-185
[8] Mirnnan C.R. and Gupta K.C., 1992, "Compensation
of Robot Joint Variables using Special Jacobian Matrices",
J. of Robotic System 9(1), PP. 113-137
[9] Tang G.-R and Mooring B W, 1992, "Plane-Motion
Approach to Manipulator Calibration", Int J. of Adv
Manuf Technol, Vol. 7 PP 21-28
[lo] Veitscheggei, W K and Wu, C-H, 1988, "Robot
Calibration and Compensation", IEEE Int'l J of Robotics
and Automation, Vol. 4 No 6, PP. 643-656
[ l l ] Goswami, A, Quaid, A, and Peshkin, M, 1993,
"Complete parameter identification of a robot from partial
pose informa.tion", Proc. IEEE Int'l Con5 on Robotics
and Automation, Vol. 1: PP. 168-173
[12] Driels, M.R. and Swayze,W., 1994, "Automated
partial pose measurement system for manipulator
calibration experiments", IEEE Trans. on Robotics and
Automation, Vol. 10, No. 4, PP. 430-440
[13] Bennett :D. J., Geiger, D. and J.M. Hollerbach, 1991,
"Autonomous Robot calibration for hand-eye co-
ordination", h t ' l J. of Robotics Research, VoL: 10, No.
5,PP:550-559
[14] Zlhuang, H, Wang, L and Roth, Z.S, 1993,
Simultaneous calibration of a robot and a hand-mounted
camera", Proc. IEEE Int'l Conf. on Robotics and
Automation, 'Vol. 2: PP. 149-154
[15] Newman, W.S. and Osbom, D.W., 1993, " A new
method for kinematic parameter calibration via laser line
tracking", Proc. IEEE Int'l Con$ on Robotics and
Automation, 'Vol. 2: PP. 160-165
[16] D. Tank and J. Hopfield, 1986, "Simple 'Neural'
Optimisation Networks: An A/D Converter, Signal
Decision Circuit, and Linear Programming Circuit", IEEE
Trans. on Circuits and System, Vol. CAS-33, PP.533-544
[17] Zhong, X.L. and Lewis, J.M., 1994, "Kinematic
identification and compensation using neural optimisation
networks", Third Int'l con$ on Automation, robotics and
computer vision, Singapore
[18] A. Cichocki and R. Unbehauen, 1993, "Neural
Networks for Optimization and Signal Processing", John
Wiley & Sons
[19] Renishaw Metrology Ltd., 1983, "User's Guide: LP2
and LP2H Probes with hard wired or induction
transmission "