Int. J. Com. Dig. Sys. 2, No.

3, 109-121 (2013)

109

International Journal of Computing and Digital Systems

http://dx.doi.org/10.12785/ijcds/020302

A Practical Neuro-fuzzy Mapping and Control for a 2 DOF

Robotic Arm System
Ebrahim Mattar
Dept. of Electrical and Electronics Engineering,
College of Engineering, University of Bahrain, P.O. Box 32038, Kingdom of Bahrain.
E-mail addresses: [email protected], [email protected]
Received 23 Sep. 2012, Revised 9 Feb. 2013, Accepted 15 Mar. 2013, Published 1 Sep. 2013
Abstract: Relating an arm Cartesian space to joint space and arm dynamics, is an essential issue in arm control that

has been given a substantial attention by number of researches. Arm inverse kinematic, is a nonlinear relation, and a
closed form solution is not a straight forward, or does not even always exist. This research is presenting a practical use
of Neuro-Fuzzy system to solve inverse kinematics problem that used for a two links robotic arm. The concept here is
to learn kinematics relations for a robotic arm system. This is to learn and map its environment and remembers what it
learnt. For learning the inverse kinematics, Neuro-fuzzy needs information about coordinates, joint angles and actuator
position. Information flow needed for the training for a Neuro-fuzzy network is slow and difficult to get by measuring
the real structure. Desired Cartesian coordinates are given as input to a Neuro-fuzzy that returns actuator positions.
Hence to express them as linguistics fuzzy rules. Neuro-fuzzy system is to generalize and produce an appropriate
output. The assembled system has been equipped with C++ interface routines, as being executed from a MATLAB
environment, in addition to high-speed low-level communication with the robotic arm sensing devices.
Keywords: UOB Robotic Arm; Inverse Dynamics; Computed Torque Law;

I.

INTRODUCTION

Robotics arms are widely used and employed for

industrial and non-industrial applications. However, for
more precise and accurate motion control, dynamic
model do play important role for such applications. It is
always not an easy task to get the forward and the inverse
models for robotics structure, specifically, once
redundancy exists. Kinematics models are always
nonlinear relations, and closed form solutions are not
easy tasks to be achieved. There are a number of
approaches that have been reported in literature regarding
building kinematics models. In its boarder sense,
manipulator kinematics is the study of motion without
regard to the forces which cause it. Within kinematics, it
is possible to study position, velocity and acceleration,
and all higher order derivatives of an arm position
variables. The kinematics of manipulators involves the
study of the geometric and time based properties of the
motion, and in particular how the various links move
with respect to one another and with time. Typical robots
are serial-link manipulators comprising a set of bodies,

Neuro-fuzzy mapping.

(links), in a chain, connected by joints. Each joint has a

single Degree of Freedom (DOF), either translational or
rotational. For a manipulator with (n joints numbered
from 1 to n, there are (n+1) links, numbered from 0 to
(n). Link 0 is the base of the manipulator, generally
fixed, and link n carries the end-effector. Joint (i)
connects links (i) and i first and last links are
meaningless, but are arbitrarily chosen to be 0. Joints
may be described by two parameters. The link (o) set is
the distance from one link to the next along the axis of
the joint. The joint angle is the rotation of one link with
respect to the next about the joint axis. To facilitate
describing the location of each link we affix a coordinate
frame to it, frame (i) is attached to link (i).
Denavit and Hartenberg [1] proposed a matrix method
of systematically assigning coordinate systems to each
link of an articulated chain. Axis of revolute joint (i) is
aligned with (Zi). Parallel link and serial/parallel hybrid
structures are possible, though much less common in
industrial manipulators.

2013 UOB SPC, University of Bahrain

110

E. Mattar: A Practical Neuro-Fuzzy Mapping

A Learning ANN

A. Issues Related to Robotics Task-Space Control.

When we restrict ourselves to the control of robotic
arms, we generally faced by three dedicating issues: (i) If
a target (a grasp) position is known, usually in Cartesian and
where an arm gripper must move to, the corresponding joint
angles must be computed. This problem known as INVERSE
KINEMATICS. (ii) Secondly, a path must be generated along
which each joints must be moved from the current position.
This problem is known as PATH PLANNING. (iii) Third, the
right torques must be exerted on the joints (by giving an
actuator the accurate current). This problem is also known as
INVERSE DYNAMICS.

B. ANN Arm Control: Advanced Robotic Arm.

Control: In many studies of nonlinear system control,
Artificial Neural Networks (ANN) have been used as
effective solutions by exploiting their nonlinear mapping
properties. Learning and adaptive capabilities of ANN to
control nonlinear systems, have played an important role
in performance and have proven its promising future of
ANN as an auxiliary nonlinear controller.
Typical popular nonlinear systems, are the multi DOF
robotic manipulators. Consequently, they do need
nonlinear control methodologies. A one possibility is an
ANN based methods. Over the last few decades there has
been rapid development in both the theory and
application of Artificial Neural Networks. Taking
advantage of those characteristics of ANN, many ANN
control schemes have been proposed in the literature.
The important issue of ANN control application is to
determine an appropriate training signal for training
purposes. The known back propagation-learning
algorithm is always used to adjust the internal weights in
on-line fashion.
Since the control application of ANN requires an online tuning, hence the adaptation capability of ANN plays
more important role than that of learning capability. It
has been reported that the complete mapping from one
domain to another domain, to identify INVERSE
DYNAMICS of robot manipulators, is not easy. This
often requires an off-line training procedure, which is
time consuming.
One of the ANN applications was done by Bogdanov
and Timofeev [2]. Here ANN controller was synthesized
to compensate dynamics approximation errors in the
model of a two link robotic system. Thus providing
robust control. Obtained robustness estimates for the
developed ANN algorithms establish relation between
transients quality and parameter disturbances caused by
inaccurate approximation. The controller used to stabilize
the system could be any type because its independent of
ANN training parameters. Refer to Fig. (1), Fig. (2), and
Fig. (3) for possible robotic arm control topologies.
Among the various kinds of ANNs, great attention has
been devoted to those (called mapping networks) which,

Artificial
Neural Network

Adaptation
Mechanism

PD
Joint-Space
Controller

Nonlinear
Robotic Arm

Fig. 1. A static ANN robotic arm control. ANN is trained once.

PD is standing for Proportional Derivative Controller
synthesis.
A Learning ANN

Artificial
Neural Network
Kp

Kp

Nonlinear
Robotic Arm

Kp

Fig. 2. Another topology of ANN robotic arm control.

A Learning ANN
Artificial
Neural Network
ANN
Adaptation
Mechanism

PD
Joint-Space
Controller

Nonlinear
Robotic Arm

Fig. 3. Adaptive type ANN based arm control detailed

structure.

through simple, ordered processing structures, reproduce

the main functions of the human brain, in particular those
of learning and adaptation. These ANNs can learn a
mapping between an input and an output and synthesize
an associated memory which gives the appropriate output
when a certain input is presented to it. The output is
either one corresponding to an input which is known at
the start, or is the result of a capacity for generalization
when the input is unknown. In addition, due to their
inherent nonlinearity, these neural models are capable of
executing functional approximations better than the best
classical techniques, which are essentially based on
linearization hypotheses. Such capabilities, along with
others of less importance, are supported by the
considerable calculation speed characterizing ANNs, due
to their massively parallel architecture. Such
characteristics have made the use of ANN a valid
alternative to the better-known classical techniques in the
solution of various problems ranging from pattern

E. Mattar: A Practical Neuro-Fuzzy Mapping

recognition to process control, to telecommunications, to
robotics.
C. Feedback Error Learning Technique.
One popular non-model based robot arm control
approach is to apply a simple Proportional Derivative
Controller (PD) synthesis to a robot manipulator. The
PD controllers are continuously stable, but its tracking
performance is generally poor due to uncertainties caused
by the robot dynamics. To compensate for such
uncertainties, feedback error learning scheme has
proposed to augment the PD controller as shown in Fig.
(2). This approach requires the ANN to identify the
complete robotic arm inverse dynamics for possible
compensation, De Azevedo and Barreto [3].
D. Other ANN Control Technique.
Hisa, as in [4], has applied a PD controller system for
a robot arm. He stated that, The PD controller system is
always stable, but its tracking performance is generally
poor due to uncertainties caused by the robot dynamics.
To compensate for such uncertainties, feedback error
learning (FEL) scheme has proposed to augment the PD
controller as shown in Fig. (3). This approach requires
the NN to identify the complete robot inverse dynamics
for compensation, Hsia [4].
One of the ANN applications most explored in the
last few years is the adaptive control of robotic arms with
unknown dynamics. This kind of control is generally split
into two phases: identification of the most representative
parameters of the system dynamics to be controlled, and
design of the most suitable regulator for this control. The
presence of an identification module and the possibility
of on-line tuning of the regulator parameters allow any
external disturbance or variations in the system dynamics
itself to be compensated for, thus guaranteeing successful
control of a process in all its dynamic conditions. In the
past three decades, major progress has been made in
adaptive identification and control for linear timeinvariant plants with unknown parameters. The
mathematical theories developed for adaptive control
(both identification and design of the regulator) are now
well consolidated and are all based on linear algebra and
the theory of ordinary linear differential equations. In
other words, the choice of the identifier and controller
structures is based on well-established results in linear
systems theory.
In their research paper, Lakshmi and Mashuq [5],
have both introduced an adaptive Neuro-Fuzzy control
method. This is for a Cartesian motion control of a 4
DOF robot arm. In their paper, the foucs was the
control of Selective Compliant Assembly Robot Arm
(SCARA) type robot arm. The main controller concept
was based on the use of inverse learning Adaptive
Neuro-Fuzzy Inference System (ANFIS) model only to
train itself from certain given robot trajectories. In
reality, these trajectories should be obtained by directly

111
measuring the robot arm responses for given inputs to
capture the actual dynamics in the presence of all
uncertainties. The employed approach was used for the
design and implementation of an ANFIS controller which
has shown to work with satisfactorily performance.
A NeuroFuzzy Controller synthesis for position
control of robotic arm was also presented by Tavoosi et.
al. [6]. In their approach, they stated that, robot
manipulators have become increasingly important in the
field of flexible automation. So modeling and control of
robots in automation will be very important. But Robots,
as complex systems, must detect and isolate faults with
high probabilities while doing their tasks with humans or
other robots with high precision and they should tolerate
the fault with the controller. In this respect and
background, they introduced a Neuro-Fuzzy Controller
(NFC) for position control of robot arm. Hence, they
proposed a five layer ANN to adjust input and output
parameters of membership function in a fuzzy logic
controller. For training purposes, a hybrid learning
algorithm was also used for training of such 5-layres
ANN network. While using such a learning algorithm,
the least square estimation method is applied for the
tuning of linear output membership function parameters
and the error backpropagation method is used to tune the
nonlinear input membership function parameters. To
validate the proposed NFC algorithm, the obtained
simulation results show that Neuro fuzzy controller is
better and more robust than the PID controller for robot
trajectory control.
In [7], both Pham and Fahmy have introduced a
Neuro-Fuzzy Modelling and Control technique for
Robotics Manipulators Trajectory Tracking system. In
their research efforts, they presented a novel Neurofuzzy controller synthesis for robotic manipulators
control. First, an inductive learning technique is applied
to generate the required modelling rules from
input/output measurements recorded in the off-line
structure learning phase. Second, a fully differentiable
fuzzy neural network is developed to construct the
inverse dynamics part of the controller for the on-line
parameter learning phase. Finally, a fuzzy-PID-like
incremental controller was used and employed as
feedback servo-controller. For validation purposes, the
suggested control system was also tested using dynamic
model of a six-axis industrial robot (6-DOF) arm. The
control system showed good results compared to the
conventional-PID individual joint controller.
Furthermore, Lazarevska [8] have introduced a Neurofuzzy modeling network for the issue of inverse
kinematics problem of a 4 DOF robotic arm. In this
context, the manuscript presented a detailed structure of
Neuro-fuzzy model of the inverse kinematics of 4 DOF
robotic arm employing the relevance vector learning
algorithm. Lazarevska [8] has stated that, although the

112

E. Mattar: A Practical Neuro-Fuzzy Mapping

direct kinematics of the robotic arm can be modeled with

ease by the same approach, the paper focuses on the
much more interesting kinematic task, since its solution
presents a basis for robot control design. Hence, the
presented model was based on the use of a TakagiSugeno type, but its parameters and number of fuzzy
rules are automatically generated and optimized through
the adopted learning algorithm based on M. E. Tipping's
relevance vector machine. The presented model
illustrates the effectiveness of the adopted neuro-fuzzy
modeling approach.
In terms of MIMO NARX models, pham et. al. [9]
have presented Dynamic model identification of the 2Axes PAM robot arm using neural MIMO NARX model.
In their resarch, a novel inverse dynamic MIMO NARX
model is employed for modeling and identifying
simultaneously both of joints of the prototype 2-axes
PAM robot armpsilas inverse dynamic model. In reality,
the contact force variations and highly nonlinear coupling
features of both links of the 2-axes PAM robot arm are
modeled thoroughly through an inverse neural MIMO
NARX model-based identification process using
experiment input-output training data. For the first time,
the nonlinear inverse dynamic MIMO NARX model
scheme of the prototype 2-axes PAM robot arm has been
investigated. For validation, the obtained results have
shown that proposed dynamic intelligent model trained
by back propagation learning algorithm yields
outstanding performance and perfect accuracy. Kelly et.
al., [10], and in their proposed control method, they
presented and discussed a reasoning and a technique for
combining artificial neural networks (ANN) and fuzzy
logic structure. Hence, they also presented a discussion
of the problem of moving a robotic arm in the presence of
an obstacle. The approach was based on the use of
several Neuro-fuzzy controllers as are trained, using
sample data obtained from a humans control of a robotic
arm. Their performance is quantified and compared.
Kelly et. al., [10] have shown that the definition of the
fuzzy membership functions plays a significant role in
the ability of the Neuro-fuzzy controller to learn and
generalize. Possible directions for future work are
suggested.
E. Manuscript Contribution.
Having presented few literature works within this
focus, the main objective of this research is to employ a
Neuro-fuzzy architecture for a control of a home built
robotic arm system. This would include a secondly
defined objective, which is to build a robotic arm system
with an adequate sensing abilities being interfaced to a
high level computing environments (Matlab and C++).
The development of such an arm is useful for gentle
exploration of an unknown objects in un-structured
environments. Similar to a human, an object position is
sensed by the eyes and then the arm move toward it in
precise motion to get the object; here we need also
feedback sensors to be included in our system, and even

cheap feedback sensor can do the job. To accomplish

this, our strategy is to develop a good model and
controller of the system. Using an arm manipulator, by
careful modeling and parameter estimation, does the
development of the system model. Usually, a position of
an object is given in Cartesian. This dictates, to do the
inverse kinematics. After modeling the system and found
a suitable controller, a program is needed to predict the
manipulator torque required to follow a certain desired
trajectory to reach an object.
F. Manuscript Organization.
The manuscript has been divided for six Sections. In
SECTION (I), we gave an introduction to concept of
model based robotic arm control. Few literature studies
are given within this section. The UOB 2-DOF robot arm
system and associated models are also described in
SECTION (II). In specific, we present the UOB arm
system. Here both KINEMATIC and DYNAMIC models
do play important role over an arm control. Hence, this
section is focused to present the overall hardware
implementation and all the work related to
communication and interfacing with the arm. In Section
(III), we present the concept of the learning system, that
was built behind the Neuro-fuzzy system. The Neurofuzzy UOB-Arm control I s also presented in SECTION
(IV). In SECTION (V), we present few discussion
remarks, and discus the achieved results. Finally, in
SECTION (VI), we draw few conclusions remarks.
II.

UOB ROBOT ARM SYSTEM

In order to program and control a robotic arm system,

a first step towards this task, is to learn the arm dynamics
and related kinematics. Robot dynamics algorithm is a
process for calculating equations of motion of an arm
robot mechanism, and to learn the acting forces that
cause these motions. Arm dynamics do appear in two
different representations: (i) FORWARD DYNAMICS,
which calculates how the robot will move in response to
a given force, (ii) and inverse dynamics. INVERSE
DYNAMICS calculates what are the forces that are
required to make a defined motion. The former are used
in simulators and the latter in control systems.
A. UOB Robot Arm Dynamic Modelling.
Typical, robotic arms consist of a serial-link.
Manipulators are comprising a set serial-links in a chain,
as connected by joints. Each joint has a single DOF either
rotational or translational. Each joint is supported by a
torque. Joints torques are produced by DC motor. It is
important to have accurate sensing of the arm motion.
Therefore, driving motors are fitted with sensors for
control purposes. The two links robotic system consists
of two links with representations to their lengths 1 and

for (link_1) and (link_2) respectively. In addition,

links weights of m and m for link_1 and link_2. A
2

general inverse dynamic equation of a two link robotic

system is known in form of Lagranges equation. It sums

E. Mattar: A Practical Neuro-Fuzzy Mapping

113

total forces affects the two link robotic systems. In terms

of I is the moment of inertia, is the

acceleration of the rigid-body and in is torque applied.

The specific Lagranges equation for a typical n links
robotic arm is expressed by:

M N , C in

(1)

in which the used symbols designate the followings:

is arm joint trajectory and : velocity trajectory.

(rad/sec), : acceleration of the trajectory. (rad/sec2).

in 21 input torque vector. (N.m), M 22 :
inertia forces matrix. (N.m). N 21 gravity vector.
(N.m), C 21 centrifugal force vector. (N.m). The
two-links UOB manipulator with rotational joints 1 , 2
is shown in Fig. (4).

The input torque vector in 1 in 2 do represent input

T

torques to both arm links 1 , 2 . Arm inertia force

matrix:
m12
m

(3)
M 11
m
m
22
21

m11 m1 m 2 21 m 2 22 2m 2 1 2 cos 1

m12 m 2 22 m 2 1 2 cos 2

(4)
M
m 21 m 2 22 m 2 1 2 cos 2

m 22 m 2 22

The arm centrifugal forces, are also given by N , :

2
n m2 1 2 212 2 sin 2
(5)
N , 1
2

m2 1 21 sin 2
n2

The arm gravity force vector, is therefore given as C

:

c m m2 g 1 cos 1 m2 g 2 cos 1 2
(6)
C 1 1
m2 g 2 cos 1 2
c2

Further analysis of the UOB robot arm system, shows

that it can be represented as a one dynamically defined
computational block. This involves the inverse dynamics
for the two links arm system and solving for 1 , 2 .
Ignoring effect of F , and rewriting the relationship
between joint velocity and Cartesian space velocity, this
involves the followings:

(i)
z
Arm Gripper

m2

M N , C in
x J J
x J and

1
J x J

L2

Joint 2
2

m1

UOB Arm
Kinematics
Model

L1

M J 1 x J N C in
y

Join 1

Inertial Frame

in J F

M J x J N C J T F
J T 1 M J 1 x J N C F
T

(7)
(8)
(9)
(10)
(11)
(12)
(13)

(ii)

Fig. 4. (i) UOB Two-links two actuation robotic arm

system. (ii) Arm kinematics and related frames assignments
model, [11].

The above derived equation, do represent the actuator

forces in Cartesian space.

Each UOB link has a point masses m1 and m2 at

distal end of links. The dynamic equation for a two DOF
manipulator in joint space coordinates are given by the
form:
M N , C F in
(2)

B. Jacobian Based Inverse Kinematics.

In the field of robotics, we generally refer to the
Jacobian matrix. Jacobian relates joint position and
velocity to Cartesian position and velocity arm posture.
An approximation of the kinematics equation for two
links arm is obtained as shown below:

114

E. Mattar: A Practical Neuro-Fuzzy Mapping

Px cos 1 2 cos 2 3 1 cos 2

Py sin 1 2 cos 2 3 1 cos 2 (13)

P sin cos cos
1
2
2
3
1
2
z

P x ,y ,z

P x , y ,z

Px
Px


Py 1 , 2 Py
P
P
z
z
P x , y ,z

Py

Px

1
P
x
2

1
Py
2

Pz
1
Pz
2

then

(14)

K d e K p e

K d e M 1 K p e M 1 M N C f

(17)

In above last equation of Equ. (17), this makes and

behave unpredictable as the robot arm configuration and
dynamics do change once the arm is in motion.
However, when controller gains are high such that
K d 1 and K p 1 , the above equation of Equ. (17)
can be approximated by a first order differential equation
of the following:

e K

1
d

Kp

C
Adaptation
Mechanism

Kd

(J-1)T M (J-1)

Two Links
Nonlinear
Robotic Arm

J-1

(J-1)T N-M (J.J-1)

A leaning ANN

Fig. 5. Computed torque control in Cartesian space.

K e K e M x N F
*

*
f

(20)

(15)

In Equ. (15), the term e do represent the error in

Cartesian arm posture. This is due to joint space motion
demand. Substitute (2) in torque equation (1). Evidently,
the error dynamics can also be represented by the left
hand side:
Me Kd e K p e M N C f
(16)

e M

C. Cartesian Control : Non-model Based Control.

Computed torque control in Cartesian space is shown
in Fig. (5). The actual arm model system performance in
Cartesian space will be elaborated on more in SECTION
(IV.) Here the arm performance is degraded and
unpredictable. Thus the computed torque based position
control in Cartesian space (using system Jacobian) is not
robust in practice. In our situation we shall extend the
same above design approach to non-model based robot
control as will be shown in later in Fig. (7). From Fig.
(5) a defined control law is stated as:
(19)
S d K d e K p e

In our work we will not use the Jacobians due to

number of reasons. Specifically, some of them are: (i)
Do not give precise results. (ii) The inverse could not be
defined due to singularities. (iii) Specific closed form
solutions, may be out of the environment range. (iiv)
Computed Torque Controller (iv) Defining an Error
Based Control Law. As we saw, from the above, that an
arm dynamic equation is a highly nonlinear and complex
to model. In an absence of an accurate arm model,
however, control of a robotic arm motion is frequently
carried out with a standard PD controller. Such an
approach is referred to as non-model based control. The
PD based control for robot manipulators has been of great
interest because of its simplicity and stability. The
control law has the form:

Multiplying by inertia matrix inverse, i.e. M

has crucial effects on the performance of robot

manipulators.

K p e K d1 Me K d1 M N C f 0 (18)

From Equ. (18), the error dynamics are now nearly

independent of arm model. Thus the choice of PD gains

A learning mechanism is used to learn parts of the

Computed torque control parameters. A computer code
was hence written for simulation (using numerical
integration to solve the dynamics), where the controller
parameter K d and K p were estimated by standard
tuning approaches. The resulting input torque to the
system will be in the form:

(21)
K e K e N
, C
in M
d
d
p
Finally, simulating the computer torque method, will
generate a set of input-output patterns, that can be used
for learning purposes.
The learning mechanisms is based on using the
architecture of Neuro-fuzzy system. This is additional
enlightened within SECTION (III).
III.

A LEARNING NEURO-FUZZY ARCHITCTURE

Neuro-fuzzy systems combine the positive attributes

of a ANN and a Fuzzy system. ANN became largely
popular, as due to their ability to universally approximate
continuous nonlinear function using only the information
contained in a set of input/output training pairs. Learning
rules are used to adapt the interconnection weights. In
other hand, a major criticism of most ANN is their

E. Mattar: A Practical Neuro-Fuzzy Mapping

115

opaque structure where the information stored cannot be

easily interpreted by the designer. Fuzzy systems consist
of a rule base composed of vague production rules such
as: if (input is small) then (output is small). The rules
are generally linguistic representations and because the
information can be easily interpreted by the designer, it is
said to be transparent. The power of Fuzzy system lies in
the way these production rules are given a precise
mathematical meaning so that the resulting system can
generalize to produce an appropriate output for
previously un seen inputs. Fuzzy systems have a serious
drawback when applied to many applications, their rules
are often very difficult or even impossible to determine.
This has motivated the development of an adaptive
FUZZY SYSTEMS, that adjusts their rule base
parameters via heuristic training rules about which little
can be proved. This is known as the Neuro-Fuzzy shown
in Fig. (6).
Outputs
Memberships

Inputs
Memberships

Joints
Motion

N Outputs
Fourth Layer

First Layer

Second Layer

Fifth Layer

Third Layer

Fig. 6. Fuzzy implementation via cascaded multi-layers ANN

( Neuro-fuzzy System).

Recently, the similarities between Neural Networks

and Fuzzy systems have been observed, allowing the
positive attributes of both approaches to be combine.
The result is termed a Neuro-fuzzy system since it
embodies the well-established modeling and learning
capability of neural network with the transparent
knowledge representation of fuzzy systems. To
summarize the two approaches for the arm control, this
is listed below as:
A. Benefits and Weakness of FUZZY Arm Control.
i.

Fuzzy system lets us to compute precise values for

arm data points not contained in the training data set
i.e. an appropriate output for previously unseen
inputs.
ii. Fuzzy rules are generally linguistic representations
and because information can be easily interpreted by
the designer, it is said to be transparent.
iii. Weakness of fuzzy robot arm Control: Their rules
are often very difficult or even impossible to
determine.
iv. Benefits of Neural Network Robot Arm Control:

v.

The ability to universally approximate any

continuous nonlinear function using only the
information contained in a set of input/output
training pairs.
vi. Weakness of Neural Network Robot Arm Control:
The designer cannot easily interpret their opaque
structure, where the information stored.
B. The Learning Process.
As mentioned above, the learning process is
completed by finding the appropriates weight for each
neuron. The initial step toward such learning activity, is
to collect training patterns and data from the arm sensors.
These data must be collected by moving the arm in the
space, and by changing the space variables and reading
the joints angle. Since, our work didnt include the space
sensor, we collect the data using an approximation
equation (using inverse of Jacobian) by referring to arm
simulation using the appropriate arm models. However
the actual data must come from the sensor.
The below were based on a predefined trajectory of
the manipulator arm. The patterns chosen for the training
of the neural networks in this work were taken from
points in the workspace of the arm, i.e. the area that can
be reached by the end effecter of the arm. The arm
working space is obtained by considering the robot arm
geometry of Fig. 4. For this particular research, the arm
working space is a 3-D data, where further tabulation of
training patterns will be additionally analyzed at a later
stage within this manuscript.
IV.

NEURO-FUZZY ARM CONTROL

A. Arm And Motoring System.

In reference to Fig (4), and it was mentioned already
in SECTION (II), the UOB robotic arm system has 2DOF motion in 3-D space. The system is actuated via
two high torques and high resolution d.c. motors. Each
individual motor is being sensed via a high resolution
position potentiometer. This is to measure the joint
angular position and angular velocity. In the same sense,
each motor has a driving circuitry that work both with
analog and digital domains, as will be further explained
in details in the following sub-sections.
B. Arm Actuation Closed Loop System.
In an attempt to realize such class of controller and
other controllers themes practically, an attempt has been
achieved to build a robotic arm system which has been
equipped with the right set of motion sensors. This is to
be used for implementing advanced robotics control
algorithms. In this sense, the UOB arm was implemented
with the subsequent closed loop control circuit. The
entire system hardware is shown in Fig. (7). It consists of
a number of (D/A A/D) convertors, signals conditioning
circuitries, summing points, lead controller and a simple
push pull driver for driving the two DC motors.

116

E. Mattar: A Practical Neuro-Fuzzy Mapping

In addition it consist of a number of digital

interfacing circuitries. Both joints has such inartificial
circuit structure. The UOB-arm system has two degrees
of freedom actuated via two high resolution D.C. motors.
Each motor is being sensed via a high resolution position
potentiometer. Within same sense, every arm actuator is
connected through interfacing buses to PC system as this
is illustrated by the hardware shown in Fig (7).
C. Neuro-Fuzzy Training Pattern Generation.
The most familiar or known category of the Neurofuzzy is the multilayer ANN that performs fuzzy
functions. It consists of five layers, an input layer, three
hidden layers, and an output layer. The input layer is
made up of sources nodes (sensory units), the second
layer is the hidden layer (three layers) of high enough
dimensions. Finally, is the output layer that supplies the
response of the network to the activation patterns applied
to the input layer. The transformation from the input
space to the hidden unit space is non-linear. If the input
vector is presented each neuron in the hidden layer will
output a value (weight) corresponding to how close the
input vector is to each neurons center. Each hidden
neuron has an activation function.
Card No. 1
Card
No. 2

Power Amplifiers

PC interfacing
Cards
PCL-714

PC

Fig. 8. Arm Hardware and C++ low level interfacing

windows.

High Level
Commands

C++
Codes

Matlab
Codes

Power Supply

Fig. 9. Generating training patterns. Arm dynamic simulations.

Solution of Fourth order Runge-Kutta differnational equation.

Fig. 7. UOB robotic arm interfacing hardware and

associated system layout.

Fig. (8) shows only a segment of a C++ coding that

was written for interfacing the machine with arm
hardware. There are two main cards. Individually, they
are paralleled interfaced through specific hardware
circuitry. C++ coding has been used, as the ability of such
coding to communicate with low level ports and the high
level Matlab toolboxes and commands. In addition, Fig.
(9) shows few segments of the Matlab coding that was
employed for the robotic arm control. At such high level,
there are a number of defined functions and routines as
related to arm kinematics, dynamics, training patterns
generation, and the Neuro-fuzzy training itself. Matlab
has been used extensively for such robot arm control, as
due to the ability of such high level coding to perform the
Neuro-fuzzy training and the learning process itself.

This is the overall system structure that have been

used for implement the defined Neuro-fuzzy arm control.
Few typical Matlab statements and functions used in this
research are:
Arm Motion and Dynamics Simulation: The following
Matlab coding and segments were written (using
approximation) to collect the numerical data required for
learning process:
[X,XP,tr1,tr2,tr3,tr4,tr5,tr6,y1,y2,y3]=rotine3(X,XP,Qd,Qdp,Qdpp,k);
for I =1:NX;
X1(I)=X(I)+0.5*Ts*XP(I);
end;
t=t+0.5*Ts;
[X1,XP1,tr1,tr2,tr3,tr4,tr5,tr6,y1,y2,y3]=rotine3(X1,XP1,Qd,Qdp,Qdpp,
k);
for I =1:NX;
XP(I)=XP(I)+2*XP1(I);
X1(I)=X(I)+0.5*Ts*XP1(I);
end;

E. Mattar: A Practical Neuro-Fuzzy Mapping

117
where therefore gathered and tabulated.
The simulation has taken all arm dynamics
and other effects in consideration. Such
resulting movements are shown in Fig. (10).
Gathered training patterns are also listed in
Table (1). Smoothness can be seen among
the collected data, i.e. indicating how the
arm was moving in periodic sinusoidal
fashion.

[X1,XP1,tr1,tr2,tr3,tr4,tr5,tr6,y1,y2,y3]=rotine3(X1,XP1,Qd,Qdp,Qdpp,
k);
for I =1:NX;
XP(I)=XP(I)+2*XP1(I);
X1(I)=X(I) +Ts*XP1(I);
end;
t=t+0.5*Ts;
[X1,XP1,tr1,tr2,tr3,tr4,tr5,tr6,y1,y2,y3]=rotine3(X1,XP1,Qd,Qdp,Qdpp,
k);
for I =1:NX;
X(I)=X(I)+Ts*(XP(I)+XP1(I))/6;
end;

D. Neurofuzzy Training Cycle.

% To do train
for i=1:791
ndx1(i) = data1((i),1); ndx2(i) = data1((i),2);
ndx3(i) = data1((i),3); nx1(i) = data1((i),4);
ndx12(i) = data2((i),1); ndx22(i) = data2((i),2);
ndx23(i) = data2((i),3); nx2(i) = data2((i),4);
end
trndata =[ndx1' ndx2' ndx3'];
trnout = [nx1 ];
trndata2 =[ndx12' ndx22' ndx23'];
trnout2 = [nx2 ];
Kin_Comp_1=[ndx1' ndx2' ndx3' nx1'];
Kin_Comp_2=[ndx12' ndx22' ndx23' nx2'];

numMFs=4;
mfType='gbellmf';
epoch_n=5;
in_fismat =genfis1(Kin_Comp_1,
mfType);

numMFs,

in_fismat2=genfis1(Kin_Comp_2, numMFs, mfType)

Fig. 10. Training patterns generation. Simulating the robotic

arm. (i) Joint-space arm motion. (ii) Joint-space rate of
change.

%
End

V.

ii.

Step_3: Once the training patterns are

available, a five layers ANN that can
implement the fuzzy if then rules is used for
learning the relations between arm posture
in 3-D space and the correspondence arm
joints displacements movements.

iv.

Step_4: Adjustment of fuzzy memberships.

This gives an indication that the fuzzy
system has learned relations relating arm
motion to the corresponding P x ,y ,z

RESULTS AND DISCUSSION

Having presented a detailed description of the overall

system structure hardware and software, within this
section we shall be presenting few results. In this
context, after the extensive mathematical kinematics and
dynamic modeling of the UOB robotic arm system, the
following stages were followed for validating the
proposed control methodology.
i.

iii.

Step_1: This involves the arm kinematics

and dynamics mathematical modeling and
expression in terms of a defined Cartesian
coordinate frame located at the arm base.
Step_2: Within this step, the robotic arm
system was dynamically and kinematically
simulated using Matlab. Over this stage,
large number of training data and patterns

postures.
Table 1. Part of generated training patterns.
X

035810 -130311 -130311 -130100-

133815 -137.00 -130311 -1310.5

-0.1033
-0.1033
-0.1033
-0.1033
-0.1033
-0.1033
-0.1033
-0.1033
-0.1033
-0.1033
-0.1033

-0.0178
-0.0178
-0.0178
-0.0178
-0.0178
-0.0178
-0.0178
-0.0178
-0.0178
-0.0178
-0.0178

-0.3600
-0.3600
-0.3600
-0.3600
-0.3600
-0.3600
-0.3600
-0.3600
-0.3600
-0.2744
-0.2744

-0.2744
-0.1889
-0.1033
-0.0178
0.0678
0.1533
0.2389
0.3244
0.4100
-0.3600
-0.2744

-1.8503
-1.8503
-1.8503
-1.8503
-1.8503
-1.8503
-1.8503
-1.8503
-1.8503
-1.9309
-1.9309

-0.3600
-0.3600
-0.3600
-0.3600
-0.3600
-0.3600
-0.3600
-0.3600
-0.2744
-0.2744
-0.2744

-0.1889
-0.1033
-0.0178
0.0678
0.1533
0.2389
0.3244
0.4100
-0.3600
-0.2744
-0.1889

-0.4827
-0.2792
-0.0493
0.1859
0.4022
0.5853
0.7329
0.8496
-0.9184
-0.7844
-0.6018

-0.2744
-0.2744
-0.2744
-0.2744
-0.2744
-0.2744
-0.2744
-0.2744
-0.1889
-0.1889
-0.1889
-0.1889
-0.1889

-0.1889
-0.1033
-0.0178
0.0678
0.1533
0.2389
0.3244
0.4100
-0.3600
-0.2744
-0.1889
-0.1033
-0.0178

-1.9309
-1.9309
-1.9309
-1.9309
-1.9309
-1.9309
-1.9309
-1.9309
-2.0714
-2.0714
-2.0714
-2.0714
-2.0714

-0.0178
-0.0178
-0.0178
-0.0178
-0.0178
-0.0178
-0.0178
-0.0178
-0.0178
-0.0178
-0.0178
-0.0178
-0.0178

-0.2744
-0.2744
-0.2744
-0.2744
-0.2744
-0.2744
-0.2744
-0.1889
-0.1889
-0.1889
-0.1889
-0.1889
-0.1889

-0.1033
-0.0178
0.0678
0.1533
0.2389
0.3244
0.4100
-0.3600
-0.2744
-0.1889
-0.1033
-0.0178
0.0678

-0.3594
-0.0646
0.2416
0.5086
0.7152
0.8677
0.9800
-1.0858
-0.9659
-0.7832
-0.4987
-0.0934
0.3431

Fig. (10) shows a typical full arm simulated joint

space motion. This is based on the full arm dynamic
model already derived in SECTION (III). For such a
robotic arm, as also in other similar designs and
situations, this involved a solution of a four nonlinear
state space model for both 1 , 2 . After simulated arm
model was achieved, then training data are gathered and
collected. Normalized sample data collected for both
joints and the P x ,y ,z arm x , y , z posture, were collected
as on Table (1). The arm model parameters are as listed
here: Weights of link_1 and link_2 are: m1 0.4kg
and m2 0.3kg . Lengths of link_1 and link_2 are:

1 0.3m and 2 0.25m . The defined trajectories to

the system are, ( Am 0.1 ):

initial and Final Membership Functions1

in1mf1
1

in1mf2
in1mf2

-ve large

in1mf3
in1mf3

-ve small

in1mf4
+ve large

+ve small

0.8

Degree of membership

-0.1033
-0.1033
-0.1033
-0.1033
-0.1033
-0.1033
-0.1033
-0.1033
-0.1033
-0.1033
-0.1033
-0.1033
-0.1033

E. Mattar: A Practical Neuro-Fuzzy Mapping

0.6

0.4

0.2

-0.3

-0.2

-0.1

0
input1

0.1

0.2

0.3

initial and Final Membership Functions2

in2mf1
in2mf1
1

in2mf2
in2mf2

in2mf3
in2mf3

in2mf4
in2mf4

+ve small

-ve small

+ve large

-ve large
0.8

Degree of membership

118

0.6

0.4

0.2

-0.3

-0.2

-0.1

0
input2

0.1

0.2

0.3

0.4

Fig. 11. Initial (actual) and final (learned) fuzzy membership

functions for input 1 & 2 of our system.

d 1 Am sin0.5 t , d 2 Bm cos 0.5 t (22)

d 1 Am cos 0.5 t , d 2 Am sin0.5 t (23)

2
2
2
2

d 1 Am sin0.5 t , d 1 Am cos 0.5 t (24)
2

Collected training patterns are then used for training

the designed Neuro-fuzzy system. Initial fuzzy
memberships were used at the starting stage prior to
training. Subsequently, after training, the shape of such
adopted memberships were updated. This can be
observed evidently in Fig. (11). This gives an indication
that the used Neuro-fuzzy system has learned the defined
relations. In addition, Fig. (12) shows 3-D maps
governing the adopted Neuro-fuzzy rules, whereas Fig.
(13) shows the 3-D map of the learned Neuro-fuzzy
system, i.e. the plot of the fired (fuzzy if then rule). Using
the Cartesian data as inputs to the Neuro-fuzzy, the joints
angle will be outputted from the Neuro-fuzzy and it will
be used to solve for the required joints positions. This is
also shown in Fig. (14).

(i)

(ii)

(iii)

Fig. 12. 3-D plots of relations between various fuzzy inputs, and
fuzzy outputs. i.e. 1 , 2 .

5
4
3
2
1
0
-1
-2
-3
-4
-5
5
0

5
0
-5

-5

Fig. 13. A 3-D map of the learned Neuro-fuzzy system.

The 3-D plot is displaying how actively (fuzzy if then rule) are
fired for different values of the input training sample.

E. Mattar: A Practical Neuro-Fuzzy Mapping

119
Adaptation
Mechanism

4
Training Data
ANFIS Output
3

Forward Kinematics
Real
(X, Y, Z)
Arm posture
d
d

A Learned
Neuro-Fuzzy System

Arm Motor
Sensing devices

-1

-2

Desired
X,Y,Z Arm Posture
Operators Command

-3

-4

100

200

300

400

500

600

700

800

Inverse
Forward
Kinematics

PD
Control Law

Two Links
Nonlinear
Robotic Arm

900

Fig. 16. Learned Nuero-fuzzy mapping verification. Insertion

of the trained Neuro-fuzzy system is within the controller
forward loop.

2
Training Data
ANFIS Output
1.5

0.5

-0.5

-1

-1.5

-2

100

200

300

400

500

600

700

800

900

Fig. 14. Verification of the learned Neuro-fuzzy system.

(Top): Output for a randomly excited square wave.
(Bottom): Output for a randomly excited sinusoidal wave.

After the learning phase, the adopted and trained

Neuro-fuzzy system is ready to be used within the control
loop. In this respect, Fig. (15) shows a schematic diagram
for the verification stage. In particular to Fig. (15), the
Jacobian inverse is playing a key role with the forward
loop, whereas in Fig. (16), we show the practically
implemented controller design verification in action. This
show how the learned Neuro-fuzzy system was used
within the forward loop, hence it is replacing the
Jacobian inverse. Fig. (17) shows the robotic arm
motions while using the Neuro-fuzzy maps within the
forward controller path. The robotic arm has been under
experimental trails a number of times. During such trails,
the arm has shown accurate and precise motion in 3-D.
For real verifications, the arm physical motions were
recorded while the arm was in motion. Such snaps are
also filmed as in Fig. (18).
Xd
..
Xd

Kp

J-1

Two Links
Nonlinear
Robotic Arm

Fig. 17. Arm joint space motion control,

as in reference to the controller scheme of Fig. (16).

Xa

Kd

..
Xa

Fig. 15. Jacobian inverse controller design verification.

A schematic of a non-model based robotic arm control in
Cartesian space.

Fig. 18. Snaps of recorded UOB arm video motion.

120

E. Mattar: A Practical Neuro-Fuzzy Mapping

VI.

CONCLUSION

This research was focused towards two parts. The

first part was completely dedicated towards designing
and physically building a two DOF robotic arm with two,
with its associated hardware. The second part was
directed towards control synthesis, purposes and an
employing a Neuro-fuzzy learning system for controlling
such a locally designed robotic arm system. This
involves an availability of very precise robot arm
kinemtics and dynamics models, hence to verify such
models.
The arm kinemtics and dynamics models are to be
essentially available for control. This allowed to simulate
the arm motion over time, hence to generate the precise
size of the training samples. It also involves an
arrangement of the control system hardware, i.e. the
interfacing electronics and the manipulator low level
control electronics and convertors.
Results have
indicated that, the adopted Neuro-fuzzy learning
mechanism was an excellent approach to approximate the
nonlinear maps that exist between arm joint-space motion
and the arm end point displacements. Training the
employed Neuro-fuzzy was a real issue. This is due to
the complexity of such learning architectures. As the
training samples was increased, the Neuro-fuzzy needs
more time to learn the right maps. Arm control was after
then achieved using the created maps relating arm jointspace displacements to the arm end point motions.
Results have indicated high degree of accuracy of arm
point-end motion.
It is worth to mention that, the first part of this
research (i.e. building the arm physical system), was
indeed an excellent experience for the UOB control
laboratory. This is due to the involvement of a number of
design stages from the mechanical setup to the actuation
and computer interfacing. The second part was also
furthermore interesting, (i.e. model based control).
Neuro-fuzzy system for inverse kinematics has shown
good degree of accuracy, rather than inverting the UOB
arm Jacobian. The built UOB arm will furthermore be
used for even advanced control methodologies within the
Control Laboratory at UOB, especially within the extent
of intelligent control.
REFERENCES
[1] R. S. Hartenberg and Denavit, A Kinematics
Notation for Lower Pair Mechanisms Based on
Matrices,
International Journal of Applied
mechanics, vol. 77, pp. 215-221, 1955.
[2] A. A. Bogdanov, and A. V. Timofeev Robust
optimal neural control of robots. Proceedings of the
IEEE IJCNN'-99, International Joint Conference on
Neural Networks, vol. 3, 1999.
[3] F. M. De Azevedo and J. M. Barreto, IMC Scheme
Using Neural Networks for Robot Arm, Proceedings

of the CBA 6th

Congress Latinoamericano de
Controle Automtico, 2010.
[4] S. J. Hsia, On an Effective Design Approach of
Cartesian Space Neural Network Control for Robot
Manipulator, International Journal of Robotica,
vol. 15, pp. 305-312, 1997.
[5] V. Lakshmi, U. Mashuq, "An adaptive Neuro-Fuzzy
control approach for motion control of a robot arm,"
(ICIEV) International Conference on Informatics,
Electronics & Vision, vol. 1, pp.832-836, 18-19
May, 2012.
[6] J. Tavoosi, M. Alaei, and B. Jahani, NeuroFuzzy
Controller for Position Control of Robot Arm, Paper
Reference No. :(0113-795), Proceedings of the 5th
Symposium on Advances in Sciences and Technology,
Iran, May 12-17, 2011.
[7] D. T. Pham, and A. A. Fahmy. "Neuro-Fuzzy
Modelling and Control of Robot Manipulators for
Trajectory Tracking," The 16th
IFAC WORLD
CONGRESS, 2005.
[8] Lazarevska, E., "A Neuro-fuzzy Model of the Inverse
Kinematics of a 4 DOF Robotic Arm," 2012 UKSim
14th International Conference on Computer Modeling
and Simulation (UKSim), vol. 1, pp. 306-311, 28-30
March, 2012.
[9] H. Pham Huy Anh, K. Kwan Ahn, and Y. Jong Il,
"Dynamic model identification of the 2-Axes PAM
robot arm using Neural MIMO NARX model,"
Proceedings of the ICCE 2008. 2nd International
Conference on Communications and Electronics, vol.
1, pp.18-23, 4-6 June, 2008.
[10] W. Kelly, R. Challoo, R. Mclauchlan, S. Iqbal,
Neuro-fuzzy Control of a Robotic Arm,
Proceedings of the Artificial Neural Networks In
Engineering Conference, St. Louis, MO, November
10-13, 1996, pp. 837-842, 1996.
[11] E. Mattar, UOB-Two DOF Robotic Arm, College
of Engineering, University of Bahrain. UOB-Arm
Internal Report. 2003.

Ebrahim Mattar is an Associate

Professor of Intelligent Control and
Robotics at University of Bahrain.
He has received BSc. in Electrical
Engineering (from University of
Bahrain in 1986), done MSc. in
Electronics in 1989 (from University
of Southampton, UK), and in 1994
he received University of Reading
Ph.D. in Cybernetics and Robotics
Control. Dr. Mattar has interests in
Computational
Intelligence,
Robotics Control, Modeling and Control. This includes
clustering with fuzzy, Neural Networks, Evolutionary
Computation, and their real applications in Robotics and
Control. Mattar is a member of a number of professional
societies locally and internationally. Locally, he is a member of
Bahrain Society of Engineers (BSE), board member of the

E. Mattar: A Practical Neuro-Fuzzy Mapping

Academic Society in Bahrain, board member of the Technology
Transfer Society. Internationally, Mattar is a member of the
IEEE, member of IEEE Control Educational Society, IEEE
Robotics and Automation Society (RAS), an active IET
member, IET Bahrain Network Honorary Chair, X-IET
EMEA Regional Board Member, X-member of IET Knowledge
Programme Advisory Member, and member of some control
societies world-wide. Dr. Mattar has published a number of
journal and conferences articles in the area of dexterous robotic
hands, Control, Optimal robotic hand forces, Neural multi-finger
robot hand grasping and control, optimal fuzzy control, and
Neuro-fuzzy systems, intelligent control. Dr. Mattar has been a
responsible body for two major conferences in Bahrain,
including the IET ICIS-2008 (International Conference on
Intelligent Systems, December 2008), and a responsible body
for three IET symposiums in Bahrain, including (New
Directions in Automatic Control: Theories and Applications,
April 2010. Dr. Mattar is also an ABET accreditation expert, as
has been leading a team for a positive full Electrical and
Electronics Engineering Programs accreditation over the period
from 2005-2010, for College of Engineering at Bahrain
University. Elected Honorary Chair for the IEE and IET for two
Sessions, Bahrain Local Network Committee. . Organizing
Head for the International Conference on (Millennium Dawn in
Training and Continuing Education), 16-18th of April 2001. A
responsible organizer for a Power Forum : (Trends Toward
Power System Networks Enhancement), 26th February. 2008.
A responsible organizer for a Communication Forum :
Challenges and Trends in Modern Communication), 7th May
2008. A responsible organizer for an IET International
Conference on Intelligent Systems (ICIS-2008), 1-3rd December
2008. A responsible organizer for an IET Control Symposium:
(New Directions in Automatic Control: Theories and
Applications), 26th of April 2010.
Chairing Continuing
Engineering Education Program, University of Bahrain (19982002). Chairing Electrical and Electronics Engineering
Department, University of Bahrain (2004-2009), also seconded
from UOB as the Director General of the Bahrain Training
Institute (2011-2012).

121