Received: 12 November 2019 Revised: 10 February 2020 Accepted: 27 March 2020

DOI: 10.1002/asjc.2356

SPECIAL ISSUE

Design of a PID optimized neural networks and PD fuzzy

logic controllers for a two-wheeled mobile robot

Chiraz Ben Jabeur | Hassene Seddik

Department of Electrical Engineering,

University of Tunis, CEREP, ENSIT,
Abstract
Tunisia In this paper, two intelligent techniques for a two-wheeled differential mobile
robot are designed and presented: A smart PID optimized neural networks
Correspondence
Chiraz Ben Jabeur, Department of based controller (SNNPIDC) and a PD fuzzy logic controller (PDFLC). Basi-
Electrical Engineering, University of cally, mobile robots are required to work and navigate under exigent circum-
Tunis, CEREP, ENSIT, Tunisia.
stances where the environment is hostile, full of disturbances such as holes
Email: chirazbenjabeur@gmail.com
and stones. The robot navigation leads to an autonomous decision making to
overcome an obstacle and/or to stop the engine to protect it. In fact, the actua-
tors that drive the robot should in no way be damaged and should stop to
change direction in case of insurmountable disturbances. In this context, two
controllers are implemented and a comparative study is carried out to demon-
strate the effectiveness of the proposed approaches. For the first one, neural
networks are used to optimize the parameters of a PID controller and for the
second a fuzzy inference system type Mamdani based controller is adopted.
The goal is to implement control algorithms for safe robot navigation while
avoiding damage to the motors. In these two control cases, the smart robot has
to quickly perform tasks and adapt to changing environment conditions while
ensuring stability and accuracy and must be autonomous with regards to deci-
sion making. Simulations results aren't done in real environments, but are
obtained with the MATLAB/SIMULINK environment in which holes and stones are
modeled by different load torques and are applied as disturbances on the
mobile robot environment. These simulation results and the robot perfor-
mances are satisfactory and are compared to a PID controller in which param-
eters are tuned by the Ziegler–Nichols tuning method. The applied methods
have proven to be highly robust.

KEYWORDS
fuzzy logic controller, hostile environment, neural network controller, PID controller, tracking
control, wheel differential drive mobile robot

© 2020 Chinese Automatic Control Society and John Wiley & Sons Australia, Ltd

Asian Journal of Control. 2021;23:23–41. wileyonlinelibrary.com/journal/asjc 23

24 BEN JABEUR AND SEDDIK

1 | INTRODUCTION rate. In [17] a fuzzy logic based control system was

implemented for mobile robot target tracking and obsta-
To this day, the control of robotic systems is still the main cle avoidance in a dynamic hostile environment in which
theme in robotics research. In fact, mobile robots, auton- particular attention is paid to integrate computer vision
omous vehicles being one example, are gaining popular- based sensing mechanisms to a robust fuzzy logic based
ity among common users; among which we find the two- navigation control method. In [18], the authors proposed
wheeled and differential mobile robots that are the most a wireless control approach based on fuzzy logic tech-
used due to their simplicity and ease of control. They nique to control a swarm of mobile robots in an
consist of two wheels propelled by two actuators and a unknown environment with static and dynamic obsta-
castor wheel that can move freely in space. By adjusting cles. In [19], a navigation method was presented and con-
the motors power, the robot can be planned to accom- sists of a static navigation strategy which drives the robot
plish different missions. Several research works have to avoid the static obstacles by using fuzzy logic control-
focused on the control of such robots in a wide range of ler. But in the case of moving obstacles, the robot should
applications such as navigation, tracking control and generate a corresponding Trajectory Prediction Table to
obstacle avoidance [1]. In the field of trajectory tracking, predict the obstacles in future trajectories. The work in
some applications use conventional methods [2–7] and [20] presents implemented control architecture for
others are based on artificial intelligence [8–14]. For the mobile robot target tracking and obstacle avoidance in a
first ones, on one hand the PID controller is a form of dynamic hostile environment.
feedback control which is the most dominant today and Given the existing body of research results in the field
it is quite reasonable to predict that certainly they will of obstacle avoidance and path planning, which is
remain in use in the future [2,3]. On the other hand, the reviewed in this context, particular attention is paid to
theory of “Robust” control systems has grown remark- integrate computer vision based sensing mechanisms to a
ably over the past ten years. Its popularity is now spread- robust fuzzy logic based navigation control method.
ing over the field of robotics control where it is an Depth and color information for both navigation and tar-
invaluable tool for analysis and design. This rapid pene- get tracking are to be captured using a Asus Xtion PRO
tration is due to two major advantages: its applied nature sensor.
and its relevance to practical problems of automation However, most of these works based on fuzzy logic
engineering [4,5,7]. In addition, model predictive control found solutions to avoid obstacles and do not resolve situ-
is a very effective advanced control technique compared ations in which there is no way to avoid these obstacles;
to standard PID controllers, and has had a significant especially the problem of actuator damage in the case of
impact on industrial process control [2,6]. an over current when the robot is moving in a hostile
Because of their success, intelligent controllers such environment full of holes and stones in the case of an
as neural networks and fuzzy logic have become one of open environment. Also in the case of industrial area,
the most favorable areas of research for controlling sys- many objects are able to represent obstacles to the robot
tems. On one hand, fuzzy logic offers the important con- navigation.
cept of computing with words and thus becomes an This work deals with intelligent tracking control
important tool for approximate reasoning where the applied to a mobile robot that is moving in a hostile envi-
problems are poorly defined or subject to imprecise ronment with the use of artificial intelligence. In fact, in
description. On the other hand, neural networks have the its navigation, the robot may encounter some obstacles
capability for learning and adaptation. that depend on the robot environment, and that can
In the last few years, many problems facing naviga- damage its actuators. In this case, the smart robot should
tion and trajectory tracking, especially in hostile environ- behave intelligently taking into account some cases stud-
ments with static and dynamic obstacles, were solved ies regarding disturbances in the form of load torques.
with the use of artificial intelligence based controllers. Thus, it should attribute correspondent actions to each
Inside this scope, neural networks [9,11,15] and fuzzy case. So that the robot should stop its navigation, deceler-
logic [16–20] took an important place in the field of sys- ate or change its trajectory according to the situation
tems control and navigation of mobile robots or for tun- change. To solve mobile robots issues, some characteris-
ing parameters of controllers. In fact, in [16], the motion tics are included in intelligent control structures. These
model of mobile robot is established by mechanism anal- elements are the learning ability of the intelligent tech-
ysis. Then, a fuzzy PID controller is designed for trajec- niques, the data processing used for learning and the
tory tracking of the mobile robot. The controller consists adjustment of the gains.
of a PID and a fuzzy inference unit to tune the parame- Thus, the main objective is the implementation of a
ters of PID controller according to the error and the error PD fuzzy logic controller (PDFLC) imposing on the robot
BEN JABEUR AND SEDDIK 25

a desired behavior while tracking a desired trajectory. Firstly, we need to define two different coordinate
The mobile robot must be autonomous in the decision frames:
making process and should be able to protect itself by
avoiding the engine damage. A comparative study will be • An inertial and global reference frame in the environ-
done with the SNNPIDC based neural networks pres- ment, in which the robot moves in. It is denoted as:

! ! !
ented in [12]. Simulation tests are done with SIMILNK R 0 O0 , i 0 , j 0 , k 0
environment, in which real obstacles are modeled by load
torques with several degrees of strength. • A local frame which is attached to the robot and is
 
This paper is composed of five sections. The first ! ! !
0 0 0 0 0
denoted as: R O , i , j ,k .
section concerns the introduction. The second
section deals with kinematic and dynamic modeling. The
third section accounts for the control strategies and goals.
In the forth section both simulations and results are pres- The origin of the local robot frame is the mid-point O0 on
ented, and the fifth section concerns the conclusion. the axis between the wheels and the center of mass C is
at a distance d from the origin O0 .
The posture of any point in the robot can be represen-
2 | MODELING OF DIFFERENTIAL ted in the local frame as: q0 = [x0 y0 θ]T and in the iner-
DRIVE ROBOT tial frame as: qI = [xI yI θ]T.
Where:x0 ,y0 : are coordinates in the local frame,
Modeling of a differential drive robot is a task of pre- xI,yI: are coordinates in the inertial frame,
senting the overall model of the two-wheeled mobile θ: is the angle between the heading direction and the x
robot based on the kinematic and dynamic models in axis.
addition to the DC motor dynamics that must be taken The two coordinates are linked by the orthogonal
into account. In fact, the control inputs of the robot rotation matrix as follows:
dynamic model are the torques delivered by the two
2 3
DC motors incorporated in the left and the right cosθ −sinθ 0
6 7
wheels [21]. RðθÞ = 4 sinθ cosθ 0 5: ð1Þ
0 0 1

2.1 | Coordinates and non-holonomic

constraints We can apply the same transformation between frames
on motions as:
The typical model of a non-holonomic wheeled mobile
q_ I = RðθÞq_ 0 : ð2Þ
robot is shown in Figure 1. The robot is operated with
two wheels in addition to a castor one ensuring its This equation is very important, because it defines rela-
stability. tions between velocities in the inertial frame and the
local frame. In fact, motions of a differential drive mobile
robot are described by two non-holonomic conditions
which are:

•1st Condition: in the local robot frame, the velocity of

the center-point O0 is zero along the lateral axis:

y_ = 0: ð3Þ

Considering the orthogonal rotation matrix, the velocity

in the inertial frame gives:
_
− xsinθ + y_ cosθ = 0: ð4Þ

•2nd Condition: each wheel maintains one contact point

S with the ground. The velocities of the contact points in
FIGURE 1 Posture of non-holonomic wheeled mobile robot the robot frame are related to the wheel velocities by:
26 BEN JABEUR AND SEDDIK

2 3
2 3 x_
0 0 66 y_ 7
7

vsr = r φ_ wr 6 −sinθ cosθ 0 76 7
6 76 _ 7
ð5Þ AT ðqÞq_ = 6 cosθ sinθ L −r 0 76 θ 7: ð11Þ
vsl = r φ_ wl 4 56 7
cosθ sinθ −L 0 6 φ_ 7
−r 4 r 5
φ_ l
where φ_ wr and φ_ wl are angular velocities of right and left
wheels.
The positions of left and right wheels in the inertial From the matrix constraints, the derivative posture
frame are expressed with: as functions of wheels angular velocities, and the for-
ward robot kinematic model is determined and
 
x wl = x −Lsinθ x wr = x + Lsinθ defined by:
and : ð6Þ
x wl = y + Lcosθ x wr = y− Lcosθ 2 3
2 3 cosθ 0
x_ 6 7
6 7 6 sinθ 0 7
6 y_ 7 6 7
Considering the rotation matrix R(θ), the velocities of 6 7 6 7 v
right and left wheels are given by:
6
q_ = 6 θ_ 7 6
7=6 0 1
7
7 ð12Þ
6 7 6 7 ω
6 φ_ wr 7 6 7
  4 5 6 1=r − L=r 7
x_ wl −r φ_ wl cosθ = 0 x_ wr −r φ_ wr cosθ = 0 4 5
and : ð7Þ _φwl
y_ wl −r φ_ wl sinθ = 0 y_ wr −r φ_ wr sinθ = 0 =r =r
1 L

and
Replacing positions equations in the velocities equations,
2 3
we obtain: rcosθ rcosθ
6 7
8 6 rsinθ rsinθ 7
: 6 7
>
>
>
zfflfflfflfflffl}|fflfflfflfflffl{ 6 7 φ_
>
> x −Lsinθ −r φ_ wl cosθ = 0 6 7 wr
> q_ = 6 r=2L − r=2L 7 : ð13Þ
>
> 6 7 φ_ wl
>
> : 6 7
>
> zfflfflfflfflffl}|fflfflfflfflffl{ 6 07
< y + Lcosθ −r φ_ wl sinθ = 0 4 1 5
: : ð8Þ
>
> zfflfflfflfflffl}|fflfflfflfflffl{ 0 1
>
>
> x + Lsinθ −r φ_ wr cosθ = 0
>
>
>
>
>
> :
: zfflfflfflfflffl}|fflfflfflfflffl{
>
y −Lcosθ − r φ_ wr sinθ = 0
2.2 | Kinematic modeling

Thus the rolling constraint equations are formulated as The main principle of kinematic modeling is to provide
follows: relationships between linear and angular robot velocities
8 and angular velocities of left and right wheels. The over-
>
> _
x_ − Lθcosθ −r φ_ wl cosθ = 0 all study is based on the mechanical system motion with-
>
>
>
< y_ −Lθsinθ
_ out taking into consideration forces.
−r φ_ wl sinθ = 0
: ð9Þ From the matrix constraints, the derivative posture as
>
> _
x + L _
θcosθ −r φ_ cosθ = 0
>
> wr functions of wheels angular velocities is given by:
>
: _
y_ + Lθsinθ −r φ_ wr sinθ = 0
The matrix of non-holonomic constraints AT(q) is 8 r
>
> x_ = ðφ_ wr + φ_ wl Þcosθ
expressed by these following equations: >
> 2
>
< r
2 3 y_ = ðφ_ wr + φ_ wl Þsinθ : ð14Þ
0 0 >
> 2
6 − sinθ cosθ 0 7 >
>
6 7 >
: θ_ = r ðφ_ wr − φ_ wl Þ
AT ðqÞ = 6 cosθ sinθ L − r 0 7: ð10Þ 2L
4 5
cosθ sinθ −L 0
−r

The linear velocity of each driving wheel in the robot

This matrix of constraints and the posture vector verify: frame is:
BEN JABEUR AND SEDDIK 27


vwr = r φ_ wr € + Cðq, q_ Þq_ + Fðq_ Þ + GðqÞ + τq = BðqÞτ + AT ðqÞλ
MðqÞq
: ð15Þ
vwl = r φ_ wl ð21Þ

Therefore, the linear velocity of the robot is the average where:

of the linear velocities of the two wheels: M ðqÞ∈Rnm : is a symmetric positive definite inertia
matrix,
1 r Cðq, q_ Þ ∈Rnm : is the centripetal and coriolis matrix,
v = ðvwr + vwl Þ = ðφ_ wr + φ_ wl Þ ð16Þ
2 2 F ðq_ Þ∈Rn1 : is the surface friction matrix,
GðqÞ∈Rn1 : is the gravitational vector,
and the angular velocity of the robot is expressed with τp ∈Rðn − mÞ1 : is the bounded unknown disturbance,
the two wheels angular velocities as: BðqÞ∈Rnðn − mÞ : is the input matrix,
τ∈Rðn − mÞ1 : is the control input vector provided by
r
ω = θ_ = ðφ_ − φ_ wl Þ: ð17Þ actuators,
2L wr AT(q): is the matrix associated with the kinematic con-
straints and verify AT ðqÞq_ = 0, and λ is the Lagrange mul-
On the contrary, we can obtain the two wheels angular tipliers vector.
velocities as functions of linear and angular robot veloci- The approach used for such modeling is the Lagrange
ties as: dynamic approach [22,23]. The principle of the method is
analytically based on the kinetic and potential energies of
1 L 1 L the system.
φ_ wr = v − ω and φ_ wl = v + ω ð18Þ
r r r r The following equation describes the Lagrange
equation:
Finally, the forward robot kinematic model is given by:
   T
2 3 d ∂L T ∂L
2 3 − = BðqÞτ + AT ðqÞλ ð22Þ
x_
cosθ 0 dt ∂ q_ ∂q
6 7
6 7 6 sinθ 0 7
6 y_7 6 7
6 7 6 7 v
6
q_ = 6 θ_7 6
7=6 0
7
7 : ð19Þ where:
6 7 6 1 7 ω
6 φ_ wr 7 6 7
4 5 6 1=r − L=r 7
4 5 Lðq, q_ Þ = U ðq, q_ Þ− T ðqÞ ð23Þ
φ_ wl 1 L =r =r
is the Lagrangian function, U, is the kinetic energy,
This model can also be considered: T is the potential energy, q are the generalized coordi-
nates, B(q) is a matrix that links the actuators torques
2 3
2 3 rcosθ rcosθ and the generalized forces. Since the robot is moving in
x_ 6 7 the inertial frame, its potential energy is considered to
6 7 6 rsinθ rsinθ 7
6 y_ 7 6 7
6 7 6 7 φ_ be zero.
6
q_ = 6 θ_ 7 6
7 = 6 r=2L − r=2L
7 wr
7 : ð20Þ The generalized coordinates are selected as:
6 7 6 7 φ_ wl
6 φ_ wr 7 6 7
4 5 6 1 07
T
4 5 q_ = x_ y_ θ_ φ_ wr φ_ wl : ð24Þ
φ_ wl
0 1

2.3 | Dynamic modeling The global kinetic energies of the robot consist of the
sum of the kinetic energy of the robot platform with the
The main purpose is to represent the robot dynamic kinetic energies of the wheels and actuators. This relation
model which is crucial for the simulation study of various is given by:
motion control-algorithm designs. The robot dynamic
model should take into account different forces which U ðq, q_ Þ = U c ðq, q_ Þ + U wr ðq, q_ Þ + U wl ðq, q_ Þ ð25Þ
affect its motion.
Generally, a non-holonomic differential robot with where:
m constraints, n generalized coordinates and n-m inputs U c ðq, q_ Þ : is the kinetic energy of the robot platform
is illustrated by the motion equation above: defined as:
28 BEN JABEUR AND SEDDIK

1 1 2 Applying the Lagrangian derivative equation, the

U c ðq, q_ Þ = mc v2c + I c θ_ ð26Þ
2 2 dynamic equation is given by:
2 3
0
U wr ðq, q_ Þ : is the kinetic energy of the right wheel given 0
6 −mc dsinθ 7
6 0 7
by: 6 m 07
6 −m dcosθ 0 7
6 m c 7
6 0 7
1 1 1 6 −m dsinθ −m dcosθ 07
U wr ðq, q_ Þ = mw v2wr + I m θ_ + I w φ_ 2wr
2
ð27Þ 6 I 0 7q
7€ ð33Þ
c c
2 2 2 6
6 7
6 0 0 7
6 0 Ir 07
6 7
U wl ðq, q_ Þ : is the kinetic energy of the left wheel given by: 6 0 7
4 0
0 5
0
1 1 1 Ir
U wl ðq, q_ Þ = mw v2wl + I m θ_ + I w φ_ 2wl
2
ð28Þ
2 2 2 2 3
0
0
6 −mc dsinθ 7 2 3
6 0 7 0
where, mc is the mass of the robot platform, mw is the 60 7 0
6 −mc dcosθ 0 07 6 7
mass of each driving wheel, Ic is the moment of inertia of 60 0 7 60 07
6 7 6 7
the robot through the center of mass, Im is the moment of 60 07 6 07 τ
−6
6
0 0 0 7q_ = 6 0
7 6
7 r q_
7 τ
inertia of each driving wheel with a motor about the 60 7 61 7 l
wheel axis, and Iw is the moment of inertia of each driv- 6 0 7 6 07
6 0 0 7 4 5
60 07
ing wheel with a motor about the wheel diameter. 6 7 0
4 0
0 5 1
With consideration to the center of mass coordinates
0
and the two wheels coordinates in the local frame, these 0
equations depend on the generalized coordinates x and y 2 3
as: 2 3 λ1
0 0 6λ 7
6 −sinθ cosθ 0 76 27

1

2  1 2 6 76 λ 7
2
+ 6 cosθ sinθ L −r 0 76 37
_
U c ðq, q_ Þ = mc x_ −dθsinθ _
+ y_ + dθcosθ + I c θ_ 4 6
56 7 7
2 2 cosθ sinθ −L 0 4 λ4 5
ð29Þ −r
λ5
1

_
2

_
2  where
U wr ðq, q_ Þ = mw x_ −Lθcosθ + y_ −Lθsinθ
2 ð30Þ τr and τl are torques control inputs of the robot.
1 1
+ I m θ_ + I w φ_ 2wr
2 Considering that the Lagrangian coefficients are
2 2 equal to zero, the robot dynamic model is given by:
8  
1

_
2

_
2 
> 2I w 1
U wl ðq, q_ Þ = mw x_ + Lθcosθ + y_ + Lθsinθ >
> m + 2 v_ −mc dω2 = ðτr + τl Þ
2 < r r
ð31Þ
1 1   : ð34Þ
+ I m θ_ + I w φ_ 2wl
2
>
> 2L2 I w L
2 2 >
: + I _
ω + m c dωv = ð τ r −τ l Þ
r2 r

After development and with considering that the where mc is the mass of the robot platform, m is the total
Lagrangian is equal to the kinetic energy: mass of the robot, Iw is the moment of inertia of each
Lðq, q_ Þ = U ðq, q_ Þ. driving wheel with the motor about the wheel diameter
The Lagrangian equation is then given by: and I is the total inertia of the robot.

1

Lðq, q_ Þ = m x_ 2 + y_ 2 + mc dðy_ cosθ − xsinθ
_ Þθ_
2 ð32Þ 2.4 | Actuators dynamic modeling
1
 1 2
+ I w φ_ 2wr + φ_ 2wl + I θ_
2 2 In general cases, the driving system of mobile robots, as
presented by Figure 2, is based on armature-controlled
where, DC motors which are considered as servo actuators.
m = mc+2mw: is the total mass of the robot, These actuators train the two wheels and drive the
and I = Ic+mcd2+2Im+2mwL2: is the total inertia of mobile robot by providing torque control inputs. The con-
the robot. trol inputs of the motors are the armatures voltages E(t).
BEN JABEUR AND SEDDIK 29

F I G U R E 2 DC motor architecture [Color figure can be viewed

at wileyonlinelibrary.com]

Electrical and mechanical parts of the motor circuit

are described by:
8
>
> diðt Þ
< E ðt Þ = Riðt Þ + L + eðt Þ with eðt Þ = K e ωm
dt : ð35Þ
>
> dω ðt Þ
: τm = J m + f ωm ðt Þ + K m iðt Þ with τ = Nτm
dt
F I G U R E 3 Hostile environment: Open environment and
As motors are mechanically coupled to the wheels by the industrial area [Color figure can be viewed at wileyonlinelibrary.
reducers, mechanical equations of motors are directly com]
linked with those of the robot:
In fact, it should be known that the couple, load tor-

ωmr = N φ_ wr and ωml = N φ_ wl que and robot behavior, must be correspondent and must
: ð36Þ
τr = Nτmr and τl = Nτml respect some rules:

• If Tlr(respectively Tll) < τr(respectively τl), that is the

As a result, the obtained equations are the dynamic load torques are less than the robot wheels torques, so
model of the two motors: the robot can overcome the obstacles (soft holes or sto-
nes), and then should evolve normally in its trajectory
8 9
> 1 > without changing its behavior;
< ðR + LpÞ ðE r −K e N φ_ wr Þ = ir and τr = NK m ir >
> =
• If Tlr(respectively Tll) > τr(respectively τl), that is the
: ð37Þ
>
> 1 >
ðE l − K e N φ_ wl Þ = il and τl = NK m il >
load torques are higher than the robot wheels torques,
: ;
ðR + LpÞ thus the robot should try to overcome the obstacles
(medium holes or stones), decelerate and then con-
tinue its trajectory normally;
3 | C O N T R O L S T R A T E G I E S AN D • If Tlr(respectively Tll) > > τr(respectively τl), that is the
G O A LS load torques are extremely higher than the robot
wheels torques, hence the obstacles are quite difficult
In this section, a trajectory tracking hindrance and speed to overcome (very deep holes or big stones), then the
control problems are presented when the mobile robot is robot should STOP to avoid motors damages, bypass
evolving in a hostile environment such as open environ- and encounter the obstacles;
ment and industrial area which are full of obstacles. • If Tlr > Tll, that is the left load torque is less than the
These obstacles present in the environments are shown right load torque (left obstacle is less important than
in Figure 3 and are modeled by different values of load the right one), then the robot should turn to the LEFT;
torques τl and τr which are applied to test the system con- • If Tlr < Tll, that is the left load torque is higher than
trol performances. the right load torque (left obstacle is more important
These figures correspond to real and industrial envi- than the right one), then the robot should turn to the
ronments in which the robot is assumed to normally RIGHT.
operate and is intended to navigate.
The major problem is that the robot must behave in an In this work, it is a task of simulating the overall system
autonomous manner. Therefore, the control objective is to with SIMULINK environment. In this simulation, a model-
design a controller which can adapt the mobile robot's ing of real obstacles is done with the introduction of load
behavior to the application of different load torques. torques as it was indicated previously.
30 BEN JABEUR AND SEDDIK

3.1 | Smart PID optimal neural networks Finally, It should also be emphasized that the PID
controller (SNNPID) gains change automatically according to the applied load
torque, that is to say that each time a load torque is intro-
3.1.1 | Neural network controller duced, the neural controller offers the gains of the
corresponding PID controller and it was a coverage on
Our case study is based on a Smart PID optimized neural the four cases of simulation since the gains of the PID
networks controller (SNNPID). Thus, the main idea is to change according to the load torque.
optimize PID controller coefficients with the use of the The mission of the neural network is to deliver the
neural networks method [24]. appropriate gains of the PID controller that automatically
The general structure of the PID controller is repre- adapt according to the applied load torque. The architecture
sented with its continuous transfer function as follows: of the overall controlled system is presented in Figure 4.
A block diagram of the neural network
Ki (NN) optimizing PID parameters is shown in Figure 4:
C ðp Þ = K p + + Kd ð38Þ
p This smart PID optimized neural networks controller
is implemented on MATLAB/SIMULINK environment and
This method is required to learn information about a data simulations are presented to demonstrate the effective-
set, in order to compute a number of weights that best ness of the proposed neural networks control strategy
allow the controller to track the given input/output data, shown by Figure 5.
and to provide PID controller with the optimal values of
its coefficients {kP, kI, kD} corresponding to the desired
mobile robot behavior. 3.1.2 | Neural network architecture
In fact, to obtain to the desired performances, the
gains of the PID controller must be optimized according The neural network used is a multi-layer network with the
to the strategy of the load torque variations. So that, a back-propagation training method. Its structure is composed
database containing the desired behaviors which is a pair of three layers: the first layer is constituted of four inputs
of applied torques and correspondent outputs will be gen- that are the two errors between angular velocities of the
erated. By learning the neural network, the appropriate wheels and their references, and the two armature voltages
gains are determined for each pair. of the two motors. The second layer is composed of eighteen
The input of the neural network are the signals input hidden neurons with “tan sigmoid” activation function and
voltages of the right and left motors Erand El, as well as the third layer has three output “purelin” neurons that are
the errors on the angular velocities of the two right and the coefficients of the PID controller.
left wheels εphi¯ l,εphi¯ r. The outputs are the gains of the The algorithm used as auto-tuning is the back-
controller, namely the proportional gain Kp, the integral propagation method, chosen for its ability to adapt to
gain Ki and the derivative gain Kd. changing environments. At the beginning, the inputs
So four models are done for four cases (which were should be applied to the network; this is propagated from
discussed before), and the data is collected and then saved the first layer to the hidden layers within, to produce an
to create groups of data. Therefore, for each case, the PID output (Kp, Ki and Kd). The output signal is compared to
adjustment is ensured by simulation and introducing the the desired output and an error signal is calculated for
appropriate load torque and the outputs are recorded. each of the outputs. The error outputs back-propagate,
The case where the right load torque is greater than starting from the output layer, to all neurons in the hidden
the left one is also studied and vice versa. layer that contribute directly to the output; however, the

F I G U R E 4 Block diagram of the

neural network (NN) optimizing PID
parameters [Color figure can be viewed
at wileyonlinelibrary.com]
BEN JABEUR AND SEDDIK 31

FIGURE 5 PID optimized neural networks control strategy [Color figure can be viewed at wileyonlinelibrary.com]

hidden layer neurons receive only a fraction of the total ωji are the weights of the hidden layer, and ωkj are the
error signal. weights of the output layer.
This process repeats iteratively, layer by layer, until η is the learning rate, δjis the error gradient at unit j,
all neurons in the network have received an error signal δkis the error at unit k and Opi and Op j are the outputs of
describing its relative contribution to the total error. The unit i and j, respectively.
back-propagation algorithm trains the neural network, The Figure 6 presents the topology of the used NN to
and compares the network's actual output Op to the tar- auto-tune the PID control gains:
get output Tp. Thus, to adjust the weights wij and the For the data required to guarantee the neural net-
biases bij reducing the energy function E: work, learn the environment and adapt its behavior,
about 50 vectors of controller parameters are used for
1
2
E= T p −Op ð39Þ learning.
2
where:


3.2 | Design of PD fuzzy logic controller
Tp = εphil Ref ,εphir Ref , ErRef , ElRef
Fuzzy logic shows some advantages over other conven-

 tional types of classical methods. Its major concept is the
Op = εphil ,εphir , Er , El :
use of linguistic variables whose values are words
expressed in natural language and also of fuzzy if-then
This difference is used to change the connection weights rules, in which antecedents and a consequence contains
between neurons in the network. This way, the weights linguistic variables.
are adjusted in a gradient descent manner, which means
the minimization of errors between the expected output
and the actual output for a particular input. In fact, syn- 3.2.1 | Fuzzy logic control strategy
aptic weights are updated as:
In our PDFLC strategy, two fuzzy controllers were used
for each of the two wheels separately to generate control
ωji ðt + 1Þ = ωji ðt Þ + ηδ j Opi ð40Þ voltages of the left and the right actuators, and are struc-
ωkj ðt + 1Þ = ωkj ðt Þ + ηδk Op j ð41Þ tured with a Mamdani-type fuzzy architecture. Each con-
troller has two inputs and one output that are logically
where: connected by nine rules.
32 BEN JABEUR AND SEDDIK

error εphi¯ l and the change rate of linear left wheel angu-
lar velocity error ε_ phil , and the output is the voltage of the
left wheel actuator El.
In addition, the inputs of the second left fuzzy con-
troller are: the linear right wheel angular velocity error
εphi¯ r and the change rate of linear right wheel angular
velocity error ε_ phir , and the output is the voltage of the
left wheel actuator Er.
The SIMULINK based fuzzy logic control strategy is
given by Figure 8.

F I G U R E 6 Topology of the used NN to auto-tune the PID a- Fuzzy controller

control gains [Color figure can be viewed at wileyonlinelibrary.
com] The block fuzzy controller is composed of three steps, as
shown in Figure 9: a fuzzification step, a fuzzy inference
As represented in the proposed PD fuzzy logic control system and a defuzzification step.
strategy shown by Figure 7, the inputs of the first right
fuzzy controller are: the linear left wheel angular velocity • Fuzzification Step

F I G U R E 7 Block diagram of the

PD fuzzy logic controller [Color figure
can be viewed at wileyonlinelibrary.
com]

FIGURE 8 PD fuzzy logic control strategy [Color figure can be viewed at wileyonlinelibrary.com]
BEN JABEUR AND SEDDIK 33

FIGURE 9 Fuzzy controller architecture

Fuzzification involves transforming, inputs from real El ∈ [0,100] and Er ∈ [0,60]; These intervals were
values to fuzzy variables. The corresponding fuzzy sets obtained from PID and PID optimized neural networks
are Gaussian membership functions for the inputs and controllers simulations.
the outputs: five linguistic variables {NB (NegativeBig), These Gaussian membership functions are shown in
NS (NegativeSmall), Z (Zero), PS (PositiveSmall), PB Figure 10 and Figure 11 as follows:
(PositiveBig)} for each of: (εphil , ε_ phil , εphir , ε_ phir , and five
linguistic variables {NB (NegativeBig), NM • Fuzzy Inference System
(NegativeMedium), NS (NegativeSmall), Z (Zero), PS
(PositiveSmall), PM (PositiceMedium), PB (Posi- Fuzzy inference system is the process of applying fuzzy
tiveBig)} for the outputs (El, Er) supplying the actuators rules to the fuzzified input values and calculating the fuzzy
with voltages. outputs. In fact, the starting point of conventional fuzzy
Gaussian membership functions are used for input control is the development of a rule base using linguistic
and output membership functions because it is shown in descriptions of control protocols. The rule base consists of
[25] that they are found to be closely performing well a set of if-then rules relating fuzzy quantities, which repre-
and better than other membership functions in control- sent process response (output) and control inputs.
ling this system. The general shapes of Gaussian mem- Table 1 represents the inference system which knowl-
bership functions are given by: edge base is constituted with eight fuzzy rules for each of
the two controllers. Each rule takes two inputs, and pro-
!
ðz −cÞ2 duces one output which is a voltage vector.
μðzÞ = exp ð42Þ
2σ
• Defuzzification Step

where the parameter c is the mean and the parameter σ In the last step, a defuzzifier converts fuzzy conclusions into
is the variance of each membership function, the parame- real values control actions. The approach opted to achieve
ter z is the crisp input μ(z) is its membership function defuzzification procedure is the center of gravity COG.
grade with numerical value in the interval (0 1). The level of the output control action is reliable with the
For the FLC, the inputs are chosen to vary within level of the input values. In fact, negative input values result
these intervals: PLEASE LOAD MATH INLINE 0065 in negative output values, zero input values in zero output
HERE, and the outputs varies within the intervals values and positive input values in positive output values.

F I G U R E 1 0 Fuzzy set definition for inputs (rd/s) and output (V) variables of the right fuzzy controller [Color figure can be viewed at
wileyonlinelibrary.com]
34 BEN JABEUR AND SEDDIK

F I G U R E 1 1 Fuzzy set definition for inputs (rd/s) and output (V) variable of the fuzzy controller [Color figure can be viewed at
wileyonlinelibrary.com]

TABLE 1 Fuzzy control rule base high references velocities were applied to test the stability
of the controllers and four the two last cases, the robot
(El, Er) εphi¯ l , εphi¯ r
should turn to the right or to the left based on the applied
ε_ phil , ε_ phir NB NS Z PS PB
different load torques on the left and right wheels.
NB NB NB NM NS Z
NS NB NM NS Z PS • A linear velocity reference V_ref = 0.65 m/s;
Z NM NS Z PS PM • An angular velocity reference W_ref = 0.25 rd/s;
PS NS Z PS PM PB • A load torque reference as follows:
PB Z PS PM PB PB
When both DC motors have the same speed and the
same direction, the robot will move in a straight line. The
4 | S I MU LA T I ON R E S U L T S robot will spin at a point when both DC motors have the
same speed and the opposite direction. Running this
In this study, the training design of the SNNPID and the model will result in response curve or circular of linear
PDFLC were done with the MATLAB/SIMULINK environ- speed of the mobile robot
ment, and simulation results are carried out by consider- The following figures represent results of both train-
ing the modeled hostile environment with respect to ing control lows with different cases studies. These
various disturbances that can be holes and stones in the results are compared with those of a PID controller
form of different load torques. In addition, a comparative which parameters are tuned with the use of the Ziegler–
performances study between the two controllers is pres- Nichols tuning method [29], and are presented in
ented based on time responses and overshooting. Table 4.
In addition, some results of a PID controller are pres-
ented to demonstrate the weak point of this control strat- A 1st case study: Load torque (35 N.m) > Motor torque
egy versus the first ones.
The physical parameters for the mobile robot simula-
tions are selected in the Table 2.
Table 2 resumes the applied load torques and the B 2nd case study: Load torque (55 N.m) > > Motor
desired behavior of the mobile robot considering that the torque
motor torque maximum is 25 N.m. The form of these
applied torques are given by the Figure 12.
The control strategy consists of applying steps signals C 3rd case study: Tlr(55Nm) > Tu(40Nm)
for both linear and angular velocities references. These
values are arbitrary but should not exceed 1 m/s for both
linear and angular velocities references in the case of D 4th case study: Tlr(45N. m) < Tu(53N. m):
tracking a circular trajectory. Many works were done
with different references of linear and angular velocities
[26–28]. E Performances study
Table 3, summarize the four control strategies. In the
first case, high references velocities were chosen to test We represent by the two following figures, Figure 18 and
the robustness of the controller, in the second case, very Figure 19, the step responses of respectively linear and
BEN JABEUR AND SEDDIK 35

TABLE 2 Physical parameters for the mobile robot TABLE 3 Desired behavior versus load torques
simulations
Load
Load torques torques (N.
Situation Desired behavior (N.m) Situation Desired behavior m)
2L Distance between two wheels 0.75 m Load torque > motor Decelerate [1 40]
D Distance of point Pc from 0.3 m torque
point Po Load torque STOP moving to [40 55]
R Driving wheels radius 0.15 m > > motor torque protect the engine

mc Mass of the mobile robot 30 kg Left load torque Turn to the LEFT 55 > 40
without the driving wheels >right load torque
and DC motors Tlr > Tll

mw Mass of each driving wheel 1 kg Left load torque < Turn to the RIGHT 45 < 53
(with actuator) right load torque
Tlr < Tll
Ic Moment of inertia of the 15.625 kgm2 exposed taking into account the application of different
mobile robot about the
load torques and the effects that they have on these
vertical axis through the
responses. Evidently, in all of these cases the robot
center of mass
should behave smartly and should answer in an
Iw Moment of inertia of each 0.005 kgm2
intelligent way.
driving wheel with a motor
about the wheel axis Figure 13 represents the case of a load torque equal to
35 N.m that is bigger than the actuators torques, we have
Im Moment of inertia of each 0.0025 kgm2
driving wheel with a motor
noticed that speed responses evolves with low variations
about the wheel diameter at the time (1.5 s) of load torque application followed by
speed compensation and deceleration, in addition the
R Armature winding resistance 1.6 Ω
static error is equal to zero which makes the system quite
L Armature winding inductance 0.048 H
accurate.
Ke Back emf constant 0.19 rd/s Figure 14 presents the case of a load torque equal
Km Torque constant 0.2613 N.m/A to 55 N.m which is much more important than the
N Gear ratio 62.55 actuators torques; we remark that behavior satisfy the
specifications imposed in the table below and is able
to achieve the goal. In fact, the robot stops for fear to
angular velocities of the robot to study static and damage the motors wheels that are training the robot
dynamic performances. for both controllers. In addition, the robot is able to
follow circular trajectory. If compared with simulation
results of PID controller, Figure 15 demonstrates that
5 | DISCUSSION AND ANALYSES when applying the load torque, the system became
instable due to the high variations of the load torque
In the previous section, four cases of desired behaviors that leads to high variations on the responses. It is
were applied to the smart robot controlled by both also demonstrated by Figure 16, that if applying differ-
PDFLC and SNNPIDC, and the velocities responses were ent load torques on the left and right wheels, the sys-
tem is also instable.
Figure 17 and Figure 18 present two cases studies of
applying different load torques on the left and right
wheels, so that the robot should turn to the right or to
the left, we noticed that both controllers permit good
behavior. In fact, when applying a right wheel load tor-
que superior to the left wheel load torque, the right wheel
decelerate and the left wheel accelerate which permit the

TABLE 4 PID controller parameters

Parameters Kp Ki Kd
FIGURE 12 Load torque reference [Color figure can be
Values 0.5 40 0.1
viewed at wileyonlinelibrary.com]
36 BEN JABEUR AND SEDDIK

F I G U R E 1 3 Right and left wheels velocities - linear and angular velocities of SNNPID and PDFL controllers [Color figure can be
viewed at wileyonlinelibrary.com]

robot to turn to the right and vice versa. This case study Basically, both SNNPID and PDFL controllers allow
is not satisfactory with the PID controller; the system improved responses of the velocities while varying the
remains unstable intensities of obstacles by applying some load torques

F I G U R E 1 4 Right and left wheels velocities - linear and angular velocities of SNNPID and PDFL controllers [Color figure can be
viewed at wileyonlinelibrary.com]

F I G U R E 1 5 Right and left wheels velocities - linear and angular velocities of PID controller [Color figure can be viewed at
wileyonlinelibrary.com]
BEN JABEUR AND SEDDIK 37

with different degrees. In fact, the robot can be qualified techniques that obey the desired and imposed specifica-
as a smart robot evolving in a hostile environment and tions resumed in Table 2. These results are due to the
the control strategies are to be considered as robust neural networks learning ability about a data set in order

F I G U R E 1 6 Right and left wheels velocities - linear and angular velocities of SNNPID and PDFL controllers [Color figure can be
viewed at wileyonlinelibrary.com]

F I G U R E 1 7 Right and left wheels velocities - linear and angular velocities of PID controller [Color figure can be viewed at
wileyonlinelibrary.com]

F I G U R E 1 8 Right and left wheels velocities - linear and angular velocities of SNNPID and PDFL controllers [Color figure can be
viewed at wileyonlinelibrary.com]
38 BEN JABEUR AND SEDDIK

F I G U R E 1 9 Step response of linear and angular Velocites of SNNPID and PDFL controllers [Color figure can be viewed at
wileyonlinelibrary.com]

to adjust the PID gains and concerns also the choice of As is proved in Table 5, based on PDFLC the
the appropriate membership functions of the fuzzy logic responses to high variations in the dynamic regime do
controller. not affect the overall system stability at all. In fact, the
In addition, the robot conserves its circular trajectory
in all of the control cases. Complex, Lemniscates of infin-
ity and ovoid trajectories are also taken.
The final parameters of PID controllers after training
the neural network are given by Table 4.
A comparative study between these two control strate-
gies is given in Table 3, based on the evaluation of the robot
performances such as rapidity, stability and precision.
These performances were deduced from the figures
Figure 19 and Figure 20. In fact, the circled fluctuations
at the beginning of the simulations give up the results
corresponding to the dynamic performances of the robot
by considering the linear and the angular velocities. Also F I G U R E 2 1 Complex trajectory [Color figure can be viewed
static performances are presented in the same table. at wileyonlinelibrary.com]

F I G U R E 2 0 Step response of right and left Velocitis of SNNPID and PDFL controllers [Color figure can be viewed at
wileyonlinelibrary.com]

TABLE 5 Static and dynamic robot performances

Rapidity Stability Precision

Controller Time response at 5% Rise time Phase margins Over shooting Steady-state error

tr5% (s) tm (s) Δφ( ) D% εs
Linear velocity NN controller 0.386 - - - 0
Fuzzy controller 0.12 0.033 48 19 0
Right wheel velocity NN controller 0.386 - - - 0
Fuzzy controller 0.136 0.0426 46 17 0

NN, neural network

BEN JABEUR AND SEDDIK 39

overshooting of the linear and the right wheel velocities

(19% and 17%) leads to adequate and acceptable values of
the phase margins (48 and 46 ) which are more than 45
for the two velocities. It is important to remember that
the optimal normalized values of the phase margins are
constrained between (45 and 60 ). In addition, the con-
trolled system has demonstrated its effectiveness on fol-
lowing several trajectories as is shown in Figures 21, 22,
23 and 24.

F I G U R E 2 4 Lemniscates of infinity trajectory [Color figure

6 | C ON C L U S I ON can be viewed at wileyonlinelibrary.com]

In this paper, two control schemes based and PDFL and Then, we can assure that the PDFLC responses are
SNNPID Controllers were proposed for a two-wheeled better than those of the SNNPID ones in terms of quick
and a non-holonomic robot that is evolving in a hostile responses confirmed by the values of the rise time and
environment with full of external influences. The work the time responses at 5%. In addition, the stability of the
sorts out the problem of actuators damages in the case of system is confirmed as long as the phase margin is more
an overcurrent, and proposes a solution to this problem. than 45 for the two velocities.
It was shown that both controllers were sturdy against For trajectory tracking, it is clear that PDFLC strategy
different intensive load torques. It also guarantees stabil- offers best following than the SNNPIDC one. In fact, it is
ity and precision for static error performances. Simula- clear that with the first controller the trajectory is closer
tion results for the mobile robot control strategies were to the reference which is different to that allowed by the
given to verify and ensure the effectiveness of the con- second controller.
trolled system in terms of quick and accurate responses. However, the PID controller simulations demonstrate
that this controller cannot obey different cases studies.
Finally, we can declare that we have reached the goal
of this work. In fact, when evolving in a hostile environ-
ment, the controlled robot is autonomous in decision
making: it decelerates and then continues its trajectory,
stops depending on the size of the obstacles, and as a
result avoids the damage of the actuators, or turns to the
left or to the right. It is also smart and able to protect
itself against high load torques without affecting either
its trajectory tracking or its stability.

ORCID
Chiraz Ben Jabeur https://orcid.org/0000-0002-4763-
FIGURE 22 Circular trajectory [Color figure can be viewed at
8074
wileyonlinelibrary.com]

RE FER EN CES
1. H. Yang et al., Nonlinear control for tracking and obstacle
avoidance of a wheeled Mobile robot with nonholonomic
constraint, IEEE Trans. Contr. Syst. Technol., 24 (2016),
741–746.
2. L. Pacheco and N. Luo, Testing PID and MPC performance for
Mobile robot local path-following, Int. J. Adv. Robotic Syst., 12
(11) (2015), 1-13.
3. J. Ye, Adaptive control of nonlinear PID-based analog neural
networks for a nonholonomic mobile robot, Neurocomputing, 71
(2008), 1561–1565.
4. S. Roy et al., Robust path tracking control of nonholonomic
F I G U R E 2 3 Ovoid trajectory [Color figure can be viewed at wheeled Mobile robot: Experimental validation, Int. J. Contr.
wileyonlinelibrary.com] Autom. Syst., 13 (2015), 897–905.
40 BEN JABEUR AND SEDDIK

5. S. I. Roberto et al., Robust recursive linear quadratic 21. J. Cerkala and A. Jadlovska, Nonholonomic mobile robot with
regulator for wheeled mobile robots based on differential chassis mathematical modelling and implementation
opticalmotion capturecameras, Asian J. Contr., 21 (2019), in simulink with friction in, Acta Electrotech. Inform., 15 (3)
1605–1618. (2015), 3–8.
6. J. R. Pitanga et al., Stable model-based predictive control 22. K. Ito and K. Kunisch, Lagrange Multiplier Approach to Varia-
for wheeled Mobile robots using linear matrix Inequalitie, tional Problems and Applications, Advances in Design and Con-
IFAC Paper Line Conf. Paper Arch., 48 (19) (2015), 33–38. trol, 2008, Edition Society for Industrial and Applied
7. S. Kazuya, Y. Masahiro, and T. Kazuhiro, Robust adaptive tra- Mathematics, Philadelphia.
jectory control ofNonholonomic Mobile robot, J Syst Des Dynam, 23. R. Dhaouadi and A. Abu Hatab, Dynamic Modelling of
6 (3) (2012), 273–286. Differential-Drive Mobile Robots using Lagrange and Newton-
8. F. G. Rossomando, C. Soria, and R. Carelli, Sliding mode neuro Euler Methodologies: A Unified Framework, Advances in
adaptive control in trajectory tracking for Mobile robots, J. Intell. Robot and Automation, ISSN: 2168-9695 2013. 2(2),
Robotic Syst., 74 (2014), 931–944. 24. C. Jakub and J. Anna, Application of neural models as control-
9. L. F. Michel et al., Neural control for a differential drive wheeled lers in mobile robot velocity control loop, J. Electr. Eng., 68 (1)
Mobile robot integrating stereo vision feedback, Computacion Y (2017), 39–46.
Sistemas, 19 (3) (2015), 429–443. 25. N. M. A. Davood and H. K. Mohammad, Design of optimal
10. T. Das and I. N. Kar, Design and implementation of an adaptive Mamdani-type fuzzy controller for nonholonomic wheeled mobile
fuzzy logic-based controller for wheeled mobile robots, IEEE robots, J. King Saud Univ. Eng. Sci., 27 (2015), 92–100.
Trans. Contr. Syst. Technol., 14 (2006), 501–510. 26. M. Zhiqiang and W. Yaonan, Adaptive control for simultaneous
11. K. Khaled, S. Chiraz, and S. Hassene, Smart Navigation of stabilization and tracking of unicycle mobile robot, Asian
Mobile Robot Using Neural Network Controller, 2018 Interna- J. Contr., 17 (6) (2015), 2277–2288.
tional Conference on Smart Communications in Network 27. Z. Cao, Adaptive trajectory tracking control for a nonholonomic
Technologies (SaCoNet), Algeria (2018). Mobile robot, Chinese J. Mech. Eng., 24 (3) (2011), 546-558.
12. B. J. Chiraz and S. Hassene, Implementation of SNNPID opti- 28. S. Kazuya, Y. Masahiro, and T. Kazuhiro, Robust adaptive tra-
mized Neural Networks Controller for a two-wheeled mobile jectory control of nonholonomic Mobile robot with compensation
robot, 3rd International Conference on Control, Automation of input uncertainty, J Syst Des Dynam, 6 (3) (2012), 273-286.
and Diagnosis (ICCAD'19), July 2–4 2019, Grenoble-France 29. K. J. Astrom and T. Hagglund, Revisiting the Ziegler–Nichols
(2019). step response method for PID control, J Process Contrl, 14
13. B. Habiba, B. J. Chiraz, and S. Hassene, Speed Control of (2004), 635–650.
Non–Holonomic Unicycle Robot Using Fuzzy Logic Tech-
nique, 3rd International Conference on Control, Automation
and Diagnosis (ICCAD'19), July 2–4 2019, Grenoble-France AUTHOR BIOGRAPHIES
(2019).
14. B. Habiba, B. J. Chiraz, and S. Hassene, Mobile Robot Obstacle
Chiraz BEN JABEUR holds a Mas-
Avoidance in labyrinth Environment Using Fuzzy Logic
Approach, 3rd International Conference on Control, Automa- ter Degree (2002) and a PhD Degree
tion and Diagnosis (ICCAD'19), July 2–4 2019, Grenoble- in Electrical Engineering (2007).
France (2019). Now she is an Assistant Professor at
15. K. Khaled S. Chiraz, and S. Hassene, 3D Simulator for Naviga- the higher institute of informatics
tion of a Mobile Robot Using Simscape-SIMULINK, 3rd Inter- (ISI) in the Department of Electrical
national Conference on Control, Automation and Diagnosis Engineering and Computer Science.
(ICCAD'19), Grenoble-France (2019).
She is a member of the research laboratory: Products
16. Q. Xu et al., Fuzzy PID based trajectory tracking control of
Mobile robot and its simulation in Simulink, Int. J. Contr.
Research Center (CEREP) at the ENSIT University of
Autom., 7 (8) (2014), 233–244. Tunis. Her current research deals with systems con-
17. F. Tharindu G. Harchala, and W. Chamilla, Fuzzy logic based trol and artificial intelligence such as neural net-
mobile robot target tracking in dynamic hostile environment, works, fuzzy logic, genetic algorithms related mobile
IEEE International Conference on Computational Intelligence robots' domain.
and Virtual Environments for Measurment Systems and Appli-
cations (2015). Hassene SEDDIK is an associate
18. M. Faisal et al., Fuzzy logic navigation and obstacle avoidance professor in the ENSIT School. He
by a Mobile robot in an unknown dynamic environment, Int. has obtained the electromechanical
J. Adv. Robotic Syst., 10 (2013), 1–7. engineering degree in 1995 and
19. A. G. Andurkar and R. Tankar, Fuzzy logic based path naviga- followed by the master degree in
tion for robot using Matlab, Int. Res, J. Eng. Technol. (IRJET), 4
“signal processing: speaker recogni-
(2017), 3165–3170.
tion” and the thesis degree in data
20. F. Tharindu, G. Harchala, and W. Chamilla, Fuzzy logic based
framework for Mobile robot navigation with target tracking, Int. security and owner right privacy “watermarking using
J. Comput. Sci. Netw., 4 (5) (2015), 761–770. non-conventional transformations”. He obtained the
BEN JABEUR AND SEDDIK 41

HDR degree equivalent to the ability to supervise How to cite this article: Ben Jabeur C, Seddik H.
researches leading to obtain the full professor degree Design of a PID optimized neural networks and PD
in the field of intelligent data filtering and securing.. fuzzy logic controllers for a two-wheeled mobile
His domain of interest is: data security Audio-image robot. Asian Journal of Control. 2021;23:23–41.
and video processing applied in intelligent filtering, https://doi.org/10.1002/asjc.2356
encryption and watermarking.