ROBOT CONTROL USING NEURAL NEIWORRS

G. Jaein, D. Cluuney and D. White

Neural Systems Incorporated
2827 West 43rd Avenue
Vancouver, B. C. CANADA V6N 3H9

ABSTRACT

Actual robotic kinematics may differ from the control algorithms derived
from theoretical kinematic models in different application situations. The reason for
the difference can be natural or unforeseen changes in the mechanical structures. The
difference between the theoretical models and the actual kinematics is almost always
complex and unknown apriori. This suggests the idea of additional adaptive
transformations to compensate for the unknown differences between the theoretical
models and the actual kinematics to improve the accuracy of control. Algorithmic
transformations do exist to maintain a minimal difference between the theoretical
transformations and the derived algorithmic model of a robot., however they are too
complex or computationally intense for practical realizations.

A neural network theory is applied to theoretical robot kinematics to learn

accuracy transforms. The network is trained on accuracy data that characterize the
actual robot kinematics. The network learns the differences in the joint angles to
improve the accuracy between the effector endpoint resulting from the theoretically
calculatedjoint angles and the desired endpoint.

The trained networkgenemkm a stationary vector field of accuracy data in a

two-dimensional planar region. Results show that a neural network, as presented
here, can increase both the accuracy and positional repeatability of robots.

Application of a neural network reduces required computational power,

calibration time, maintenance cost and engineering time when developing controllers
for new robots by it’s emergent genemlization [l],fmrlt tolerant and self-organization
properties.

INTRODUCTION

It is hoped that the capabilities of modem day neural networks will solve problems that appear to be
beyond the bounds of conventional computationaldevices.

Several vivid demonstrations of the capabilities of modem day neural networks have been presented by
their advocates, including one of the authors [l].However, a major concem is they have been perceived as being
inexplicable or unprovable. Eventual users of neural technology want a clear explanation of how and what a
neural network is actually computing along with reasonable assurance of its reliability. Any really successful
applications have been able to meet these requirements.

Modem day neural networks are sophisticated extrapolationsof neural networks of the past. The user
confidence that is necessary to rapidly implement a modem day neural network application is established by
budding on the wisdom and knowledge already aquired with the most important of present day applications of
neural networks, such as high speed modems and echo cancellation.[2].

11-625

Authorized licensed use limited to: Lovely Professional University - Phagwara. Downloaded on March 18,2021 at 05:47:41 UTC from IEEE Xplore. Restrictions apply.
 The only other way of gaining the required confidence and appreciation of the technology is to solve
simpler problems. One of the objectives of this work is to present a simple application of a neural network that
can be easily validated with a large number of actual trials. Successful results from this type of apptication will
establish a degree of confidence in the results provided by a modem day neural solution.

In this study, a neural network adaptively learns the difference between a derived kinematic model of a
robot manipulator's mechanical structure and the robots actual structure. The neural network learns to improve
the accuracy of a robot (topological) transformation based only on examples of the accuracy data. Firstly, when
the examples characterize the topology the network learns that topology. If the network is presented with
additional examples of the accuracy data the network can compute the correspondingjoint angles to any unseen
correct position, that is the trained network a stationary vector field of accuracy data in a two-
dimensional planar region. In fact, very few points are required to significantly improve the accuracy of the
manipulator.

This type of computational capabii is not easily programmed at all by standard programming
techniques. A n y of the attempts of conventionally modelling the bending of a manipulator are far too complex
for the real-time control of manipulators.

Conventional Robotic Si m

Consider a jointed two degree of freedom robot arm that operates in a plane. To make the arm move,
the coordinate of the desired end point (x,y) is fed to the robot controller so that it may generate the required
joint angles (e(&) for the mechanical motors which move the arm (Fig. I).

Fig. 1Conventional robot controller.

Forward Trensfomtcrtion

The forward transformation to map j c a t coordinates (el,e2)to end effector position (x, y),
e9 --> (x, y) is:
F:(el,

where:
el = Fint angle of first arm segment
= joint angle of second arm segment
L1 = length of fust arm segment
L2 = length of second arm segment

11-626

Authorized licensed use limited to: Lovely Professional University - Phagwara. Downloaded on March 18,2021 at 05:47:41 UTC from IEEE Xplore. Restrictions apply.
 The forward transformation equations are easily calculated with conventional computing and F is a 1:l
mapping.

Inverse Tmnsftmah~on

The inverse transformation to map the end effector position to the joint angles F1:(x,y) --> (el,82)
is:

The cosine and sine of 6 2 are found (equ. (3) & (4)) and these values used in equ. (5) as arguments to
the arc tangent to get the resultant angle e2in the correct quadrant. Similarly this is done in equ. (6) and (7) to
obtain el.

The sign difference in the two forms of equations (4)and (4b) corresponds to the two possible
orientations of the elbow of the robot arm.The arm can be positioned with the elbow up or down and the end
effector stiU at the required (qy) point.

there are two possiile solutions to solving the inverse equations (equ. (4a) and (4b)), the inverse
not a 1:lmapping.

ADDliation of Neural Network to the Robotic Si-tion

The joint angles which are produced by the robot controller may not be correct to move the arm to the
desired end effector position due to changes in the physical structure of the robot.

Application of a neural network to this problem is to have the network learn the changes (differences
in the angles) ( A 8 ,A 82) needed to correct for the error between the theoretically calculated joint angles of
the kinematic model (e<&') and the actual angles (e e2)required to place the robot arm end effector at
the desired end point (Fig. 2). In effect, the network mustyearn a vector field in the 2-dimensional plane.

where: a01 = el eli

-
A82 = 82 - %'
Fig. 2 Training network with accuracy data.

11-627

Authorized licensed use limited to: Lovely Professional University - Phagwara. Downloaded on March 18,2021 at 05:47:41 UTC from IEEE Xplore. Restrictions apply.
 The trained neural network can be used to moddj the output angles from the robot controller as in
Fig. 3.

Fw 3 Controller supplemented with trained neural network.

The error of position for each point in the Zdimensional plane is considered to be within a regular
shaped error neighborhood N(x,y) about the desired point (qy).

For this application it was assumed that:

o the error neighborhood N(x,y) is a circle of radius less than or equal to a pre-defined tolerance

o the error field is continuous throughout the plane.

-
A network consisthg of three layers using a moditied version of the back-propagation learning rule [3]
was applied to the training set. There were 4 neurons in the input layer, 24 in the hidden layer, and 2 output
neurons. The input layer was completely COMected to the hidden layer, and the Hidden layer which was
completely connected to the output layer. There were no direct comections from the input to output layers.

The training set for the network consisted oE

IllpUtS:
1. the desired end point (x,y)
2. the uncorrected joint angles ( 8 { , 8 i )

Target Outputs:
1. the joint angle corrections (Ael,
A 62)

Two cases were considered in which the robot's physical structure was altered in some way. In one case
the base of the robot was shifted by a small amount within the working plane. In this case, for every point input

11-628

Authorized licensed use limited to: Lovely Professional University - Phagwara. Downloaded on March 18,2021 at 05:47:41 UTC from IEEE Xplore. Restrictions apply.
 to the robot controller, the end effector would be displaced from the target point by the same distance and
direction that the base was shifted. This was designated as the "ConstantError" case.

In the second case, the distal link of the robot arm was bent by a small amount at a certain position
along the link. Here, for different points input to the robot controller, the end effector would be displaced from
the target point by distances and directions that were functions of the two joint angles, the position and angle of
the bend on the distal link. This was designated as the "Bent Arm" or "Variable Error" case.

For both cases, the neural network was trained on a variable number of points (Fig.4) and then tested
by presenting sets (Fig. 5) of previously unseen points. The accuracy of end-effector positioning was calculated
for each point presented. Many of the test pohts lie outside of the regions of the training sets.

1.1-

1.s- 1.5-

1.1-
I I I I I I I I I
1.1 1.S 1
. I.. I.
¶ .
1 I.
I 1.s .
1

Fw 4 Representative sample of training point sets.

Fig.5 Test sets.

Simulation Results

The results were virtuallv identical for both est cases. After the network had be n trained on 1point,
the accuracy of positioning the end-effector to a desired point was improved by an average of 60% and for all
test sets presented, the accuracywas greater than the accuracy of the uncompensated or naked controller.

After training on 3 points, the accuracy was improved on average by a factor of 6. The improvement in
accuracy levelled off at approximately 11 times the accuracy of the naked controller after 8 points had been
learned.

11-629

Authorized licensed use limited to: Lovely Professional University - Phagwara. Downloaded on March 18,2021 at 05:47:41 UTC from IEEE Xplore. Restrictions apply.
 ..... ...............-....
.. .. .. .. .. . .............._.....
..-. . . . . . ......___.,
. ........... . . .
. .. ,...*...
.. .. .. .. .. .. . .
...,.... . . . . . .
._* .. .. ...
.. .....*....
.. .. ..
1 1 1 1 , ,
.
A ..._.....A

............ ....... ... ... -.... .......

. ....-..-
. .
L__.L.__I_._____I

.. .. .. ..
. . . . . . . . . . .
I ,., & , . I I. I . . .:___
a s - ; .....;....
. .1.... .1..;1._.__:. .. .. .. .. ____:
........-~..........-.._.~........-.-
. ., ,, , , ,, .,, . ,,. .,. \,. , .,
.. . .
..!....-.. . . . . . . . . . .. ..
................................. I

I . .

. . -.-,% . ., . . . . . . .......-. . . ............-..-.........

.).-.1. -...... . . ...-..-., ,. .
. I , . ,

I , , I I
I , . I

_..,
I

....
..?...$...!:..?.--.-
., ,. . ---
... ...-.......- .. ......
I
.. .....-...
. . . .
I I I I 1
1.-.,>

!...,
-.. ...
L.
1

. ......, . . . . . . . .........
. I

. I . ,
-
,...+..-~....?...?__._~
, I

I I I . # ,

. . . , . r .:...:.....I....L....:....:....~. ....:....:....I-.a.&:.~ ...:...I...'._:_._..:...1...:...:...L...:

,

.. .* .*.......... . . . . . . . . . . . . . . . . . . . . .
I , .

.:...:...:
. .....:...:...:
. . .....:...:...:...A...!
..(____L___I
I
L-.f...A_-- C...).--f...A---'
, , I
2-
.
..
)
..
.
,-
-
.+
.
.A
.
+.
.
!.
-
.. I ,

..L

L. L.
m o l led tick ond mor mion rr&*
L. ...
I

ERW nclm Iw led cicks ofla 1

I.. ...I
a.$
Mla 3 lromnq poinls leaned.
I t
&.a
a.L :..:..A..:.

a
.
I

After 16 lroininq als kamed.

...
.:...:...1...:...:...1...:
I I
..e

naked conlrder. Iroininq point learned.

Fig.6 Residual positioning errors before and after neural network training.

0.07
r Errors at 0 trainiq points are off-scde
at approximately 1.00
- Circle 1

-
0.06
__c_
Circle 2
0.05

-
Circle 3
0.04
Average Error Circle 4
in Position 0.03

A1 1
0.02 _jt_

No Network
0.01

0
0 2 4 6 8 10 12 14 16

Number of Points Learned

Fig.7 Improvement in accuracy of end-effector positioning as a fundion of training.

This study indicates that it is not necessary to learn many training points in order to gain a substantial
improvement in accuracy.

There were two major thresholds in the rate of improvement as the number of points learned by the
network increased. After the network learned only 1 point, the accuracy was greater than that of the naked
controller by 60%. The second threshold was after the network had learned as many points necessary to
characterize the topology, in this case 3 points. At this point the accuracy improvement jumped by a factor of 10
over the 1 point case.

The accuracy is improved for target points both interior and exterior to the region covered by the
training set.

11-630

Authorized licensed use limited to: Lovely Professional University - Phagwara. Downloaded on March 18,2021 at 05:47:41 UTC from IEEE Xplore. Restrictions apply.
 Training times are short for few training points. This implies that a slightly damaged robot could be
made operational quickly without repairing it physically. To do this conventionally would mean solving new
potentially difficult transformations theoretically, generating new software to implement the new
transformations, and then installing them into the robot controller.

Once the network is trained, it provides required accuracy improvement information very quickly
compared to conventional schemes which are calculation intensive.

In the event that it is impossible for the robot to reach points previously within its operational domain
due to some physical change, the network could be trained to move to the nearest point still within its domain
or not to move at all.

The feedback information used for training the neural network can be provided by an external trainer,
automaticallyby some external monitoring devices, or in some cases by sensors built into the robot itself.

This study p r o d e s an example of an application that can immediately benefit from a modem day
neural network solution. Supplementing the existing robot controller with a network that learns a topology
which improves the accuracy illustrates the advantage of building on the wisdom and knowledge that we have
already obtained with conventional techniques.

It is indisputable that the application of a modem day neural networks to the three dimensional case
will work. The extension of these results to the higher dimensional case involves no new coI1cepts, thereby
providing the necessary user codidence. Further, a neural network application reduces the required
computational power, the robot calibration time, maintenance cost and engineering time when developing
controllers for new robots by Virtue of its' firmly established emergent genemlizorion,f a i t tolemnt and self-
orgcmization properties.

ACKNOWLEDGEMENTS

The authors gratefully acknowledge the National Research Council's Industrial Research Assistance
Program and the Science Council of British Columbia for their contributions.

REFERENCES

[l] Josin, G., "Neural-Space Generalization of a Topological Transformation", Biological Cybernetics, in

press.

[2] Widrow, B. & Stearns, S.D., "AdaptiveSignal Processing", Prentice-Hall, Inc., NJ., 1985

[3] Rumelhart, D. E.,McClelland, J. L. and the PDP Research Group, "Parallel Distributed Processing"
MIT/Bradford Book, 1986.

11-631

Authorized licensed use limited to: Lovely Professional University - Phagwara. Downloaded on March 18,2021 at 05:47:41 UTC from IEEE Xplore. Restrictions apply.