Review

Advances in Mechanical Engineering

2015, Vol. 7(12) 1–14
Ó The Author(s) 2015
Mobile robot nonlinear feedback DOI: 10.1177/1687814015620328
aime.sagepub.com
control based on Elman neural
network observer

Khaled Al-Mutib2, Fodil Abdessemed1, Ramdane Hedjar2,

Mansour Alsulaiman2, Mohamed Bencherif2, Mohammed Faisal2,
Mohammed Algabri2 and Mohamed Mekhtiche2

Abstract
This article presents a new approach to control a wheeled mobile robot without velocity measurement. The controller
developed is based on kinematic model as well as dynamics model to take into account parameters of dynamics. These
parameters related to dynamic equations are identified using a proposed methodology. Input–output feedback lineariza-
tion is considered with a slight modification in the mathematical expressions to implement the dynamic controller and
analyze the nonlinear internal behavior. The developed controllers require sensors to obtain the states needed for the
closed-loop system. However, some states may not be available due to the absence of the sensors because of the cost,
the weight limitation, reliability, induction of errors, failure, and so on. Particularly, for the velocity measurements, the
required accuracy may not be achieved in practical applications due to the existence of significant errors induced by sto-
chastic or cyclical noise. In this article, Elman neural network is proposed to work as an observer to estimate the velo-
city needed to complete the full state required for the closed-loop control and account for all the disturbances and
model parameter uncertainties. Different simulations are carried out to demonstrate the feasibility of the approach in
tracking different reference trajectories in comparison with other paradigms.

Keywords
Nonholonomic mobile robot, input–output feedback linearization, Elman neural network, observer

Date received: 8 September 2015; accepted: 9 November 2015

Academic Editor: Mario L Ferrari

Introduction models. Many works have investigated the trajectory

tracking problem using the kinematic model of the
Many research articles have dealt with the design of mobile robot, and many control laws have been devel-
control laws for nonholonomic mobile robots. This oped including intelligent techniques of the fuzzy logic
interest comes from the continuous need for mobile
robots in industry, field, and service robots. Mobile
robots are no more restricted to factories and some spe- 1
Department of Electronics, University of Batna, Batna, Algeria
cific places but are becoming closer to human beings 2
College of Computer and Information Sciences, King Saud University,
when they moved to our homes. More particularly, Riyadh, Saudi Arabia
robots are used to assist people at home or assist dis-
Corresponding author:
abled individuals if their powered chairs are made intel- Fodil Abdessemed, Department of Electronics, University of Batna, Batna
ligent.1,2 In the early works, the controllers developed 05000, Algeria.
for wheeled mobile robots are based on kinematic Email: [email protected]

Creative Commons CC-BY: This article is distributed under the terms of the Creative Commons Attribution 3.0 License
(http://www.creativecommons.org/licenses/by/3.0/) which permits any use, reproduction and distribution of the work without
further permission provided the original work is attributed as specified on the SAGE and Open Access pages (https://us.sagepub.com/en-us/nam/
open-access-at-sage).
2 Advances in Mechanical Engineering

and neural network. This interest has been extended to components of the linear velocity vector. The proposed
dynamic models in order to take into account para- controller is based on dynamic model. But it seems that
meters of dynamics such as mass, moment of inertia, PowerBot user interface components do not offer the
and so on, which might give unforeseen behavior in developer the possibility to access the low-level soft-
case they are neglected or badly determined. Based on ware but Arcos. Thus, to develop a control software
the dynamic model, many articles focused on tracking application, one needs to talk to Arcos interface.
and stabilization issues using different types of control- Further work in this direction is under progress, and
lers.3,4 The controllers developed for this purpose only simulation results are presented in this article.
require some accurate knowledge on the parameters to This article is organized as follows. In section
give good performances. In the presence of uncertain- ‘‘PowerBot parameters determination,’’ we elaborate the
ties, robust methodologies must be developed in order parameter determination of PowerBot, taking into
to meet the imposed performances.5–7 Moreover, the account the physical and mobility parameters that help in
developed controllers require sensors to obtain the determining the inertia tensors and the position of gravity
states needed for the closed-loop system. However, center. The kinematic and dynamic modeling of the
some states may not be available due to the absence of mobile robot are presented in section ‘‘Kinematic and
the sensors because of the cost, the weight limitation, dynamic modelization.’’ Section ‘‘Mobile robot control’’
reliability, induction of errors, failure, and so on. presents the controller’s design for path following. In sec-
Particularly, for the velocity measurements, the tion ‘‘ENN observer,’’ the proposed procedure of the
required accuracy may not be achieved in practical ENN-based observer is performed for estimating the miss-
applications due to the existence of significant errors ing state in the presence of some disturbances. Simulation
induced by stochastic or cyclical noise. For these rea- results are given in section ‘‘Simulation results’’ and sec-
sons, software sensors, also called observers, are devel- tion ‘‘Conclusion’’ concludes this article.
oped to overcome all the above-cited problems.
Tracking control using observers when velocity mea-
surement is not possible has been proposed in some
PowerBot parameters determination
articles.8–10 In fact, observers constitute the backbone The PowerBot development platform for this project is
of the system that is to be controlled by estimating the one of the most critical components when it comes to
uncertain parameters and compensate for modeling successfully achieving the project goals. The project
errors and noise.11 The success and superiority of requires us to determine the unknown parameters that
dynamic neural network over static neural network in serve in elaborating the dynamics model. For this pur-
system identification and control of dynamical sys- pose, we propose to give the physical parameters of
tems12–17 motivated us to use it as an observer. In this PowerBot and go through all the steps that lead to the
article, Elman neural network (ENN) is proposed for determination of the position of the center of gravity
that purpose to estimate the velocity needed to com- and the different tensors.
plete the full state required for the closed-loop control.
Unlike the static neural network, the ENN has certain
dynamical advantages, such as its capability to estimate PowerBot characteristics
the velocity in a fraction of the fundamental cycle, PowerBot, among many research robots developed by
avoiding the delay often introduced by static neural Adept Corporation, is a research indoor platform
networks. When the dynamic model is to be integrated mobile robot.20 Below, we report the characteristics of
with the nonholonomic kinematics constraints, the PowerBot in terms of its physical and mobility para-
problem of trajectory tracking becomes more interest- meters necessary to derive the dynamics model.
ing and more realistic. Motivated by the work of
Sarkar et al.18 and Yun and Yamamoto,19 our goal is
to develop a control law approach based on input–
Inertia tensors of PowerBot
output linearization while estimating the nonmeasured PowerBot is constituted of elements with different
system states using ENN trained with the error back- masses and dimensions (Figure 1). Referring to
propagation (EBP) algorithm. Thus, the principal con- Figure 1, we note that the batteries are fixed in the
tribution of this article is twofold. First, we provide the extreme left robot’s chassis. Therefore, the inertial ten-
PowerBot parameter determinations which are neces- sor is not in the geographic center of the robot. For this
sary in finding the PowerBot dynamic model. Second, reason, we propose to split the robot into two rectangu-
the nonmeasured mobile robot velocity is estimated lar parts as depicted in Figure 2. The first part contains
using a proposed ENN observer based on an input– the batteries and the second part the different electronic
output linearization controller. The proposed observer cards and components.
is inserted to estimate the right and left wheel angular The PowerBot characteristics and mobility para-
velocities necessary to compute the estimated meters are defined as follows:
Abdessemed et al. 3

Figure 1. Side and top views of PowerBot.

Ge: center of mass of the electronic cards and com-

ponents part;
Gb: center of mass of the batteries part;
G: center of mass of the robot;
L: length of the robot: 83.2 cm;
2b: width of the Robot: 62.6 cm;
Le: length of the electronic part: 43.2 cm;
Lb: length of the Batteries part: 40 cm;
Lc: distance from the driving wheel axis to a look
Figure 2. Center of gravity placements (side view).
ahead reference point Pr: 43 cm;
H: height of the mobile robot: 48 cm;
r: drive wheel radius: 27 cm;
The distance between G and Ge is
rc: caster wheel radius: 14 cm;
me: mass of the electronic part: 62 kg; me
mb: mass of the batteries part: 50 kg; GGe = L
2mt
mt: total mass of the electronic and batteries part:
112 kg; We obtain the inertia tensors for the electronic part,
mw: mass of each drive wheel: ffi6 kg; the batteries part, and the platform as described in the
mc: mass of each caster wheel: ffi3 kg; coordinate system of Figure 2, with the origin at the
ul : angular position of the front left wheel; center of mass by applying the parallel axis theorem.
ur : angular position of the front right wheel; Using equation (1), one can determine the abscissa of
f: heading angle of the mobile robot; the center of gravity which is
Ie: inertia tensor of the electronic part;
Ib: inertia tensor of the batteries part; OG = 49:1 cm
If: inertia tensor of each front driving wheel and the
motor rotor; Note that this distance could be slightly different if
Ir: inertia tensor of each rear wheel (caster wheel); we take into account the masses of each of the SICK
Ip: inertia tensor of the platform; laser sensor, which is equal to 4.5 kg and Bumblebee
N: transmission ratio = 22.3:1; stereo vision camera, which weights 3 kg. With these
Jr: rotor inertia of one motor = 0.04 kg m2. values, the center of mass moves backward, and the dis-
tance OG would be
We calculate the abscissa of the center of mass G as
OG = 47:37 cm


me L2e + mb Le + Lb
2
OG = ð1Þ Now, if the Bumblebee is excluded from calculation,
mt the distance OG becomes
The distance between G and Gb is
OG = 48:03 cm
mb
GGb = L The inertia tensors are explained in the following.
2mt
4 Advances in Mechanical Engineering

Inertia tensor of the platform: Ip.

2 mb 3
ð4b2 + H 2 Þ 0 0
6 12   7
6
mb L 2 7
6 2 2 mb 7
Ib = 6 0 Lb + H + mb 0 7 ð2Þ
6 12 2mt 7
6   7
4
mb mb L 2 5
0 0 L2b + 4b2 + mb
12 2mt
2 me 3
ð4b2 + H 2 Þ 0 0
6 12   7
6
me me L 2 7
6 7
Ie = 6 0 L2e +H + 2
me 0 7 ð3Þ
6 12 2mt 7
6 
 7
4 2


2 me me L 2 5
0 0 Le + 4b + me
12 2mt
2 mt 3
ð4b2 + H 2 Þ 0 0
6 12 7
6 mt 2 mb L2b + me L2e L2
7
6 7
Ip = 6 0 H + + 2
m3b + m3e 0 7 ð4Þ
6 12 12 4mt 7
6 7
4 mt 2 mb L2b + me L2e L2

3 5
0 0 b + + m3b + me
3 12 4m2t

Inertia tensor of each drive wheel: If. We consider the wheel the robot. All the tensors calculated above will be use-
as a homogeneous cylinder of radius r and height h. ful in the equations.
Now, considering the inertia of the motor with respect
to its revolution axis, we write the inertia tensor of each
front wheel as Kinematic model
2 3 The model of PowerBot is presented in Figure 3. The
mw r 2 mw h2w
6 4 + 0 0 7 model has two diametrically opposed drive wheels of
6 12 7
6 mw r 2 7 radius r, the distance between the wheels is 2b, the
If = 6
6 0 + N 2 Jr 0 7
7 angular speeds of the drive wheels are vL and vR, and
4 2 5
2
mw r mh2w f is the heading angle defined to be the angle between
0 0 + the x-axis attached to the robot and the x-axis of the
4 12
ð5Þ world coordinates. The coordinates of the center of
gravity PG and the origin of the robot coordinate

Inertia tensor of each rear (caster) wheel: Ir. We consider

the wheel as a homogeneous cylinder of radius rc and
height hc. The inertia tensor of each caster wheel is writ-
ten as
2 3
mc rc2 mc h2c
6 4 + 12 0 0 7
6 7
6 mc rc2 7
Ir = 6 0 0 7 ð6Þ
6 4 7
4 2 25
m c rc mc hc
0 0 +
4 12

Kinematic and dynamic modelization

In the sequel, we use the kinematic model and the
dynamic model to determine the motion equation of Figure 3. Kinematic model of PowerBot.
Abdessemed et al. 5

system PC are (xG, yG) and (xC, yC), respectively. The M ð qÞ =

distance separating these two points is d. We assume 2 3
mt + 2mw 0 2mw d sin f 0 0
that the mobile robot moves in the direction of the axis 6 7
6 0 mt + 2mw 2mw d cos f 0 0 7
of symmetry, and the driving wheels do not sleep while 6 7
6 7
moving. These assumptions can be formulated in a 6 2mw d sin f 2mw d cos f I 0 0 7
6 7
three constraint equations: one is holonomic, and the 6 7
6 0 0 0 If 22 0 7
remaining two are nonholonomic such that 4 5
0 0 0 0 If 22
y_ G cos f  x_ G sin f  d f_ = 0 ð7Þ ð14Þ
x_ G cos f + y_ G sin f + bf_  ru_ r = 0 ð8Þ The remaining terms are found to be
x_ G cos f + y_ G sin f  bf_  ru_ l = 0 ð9Þ 2 2
3
mt d f_ cos f
6 7
To ensure no singularity in the control solution, it 6 2 7
6 mt d f_ sin f 7
has been shown that the point of control Pr cannot be 6 7
Bðq, q_ Þ = 6 0 7,
located at the intersection of axis of symmetry of the 6 7
6 7
platform18,21 but in front of the mobile robot. In the 4 0 5
sequel, we will use the three constraints given by equa- 0
tions (7)–(9) and define the configuration of the mobile 2 3
0 0
robot by the vector q of the generalized coordinates 6 7
which is written as q = ½ xG yG f ur ul T . 60 07
6 7
In a more compact form, we write the above three E=6 7
6 0 0 7, and GðqÞ = ½0
constraint equations in a matrix form as 6 7
41 05
0 1
DðqÞq_ = 0 ð10Þ
I =Ip33 +2mw (b2 +d 2 )+2If 22 and t= ½ 0 0 0 tr t l T
where
such that tr and tl are the torques of the right and left
2 3 drive wheels, respectively. The vector of Lagrange mul-
 sin f cos f d 0 0
tipliers is l=½l1, l2 T . In order to eliminate the
D = 4  cos f  sin f b r 05 ð11Þ
Lagrange multipliers vector l, we use the angular
 cos f  sin f b 0 r
speeds of the left and right wheels to determine the
matrix S(q) such that
Dynamic model DðqÞS ðqÞ = ½0 ð15Þ
The dynamic study consists in determining the forces or
torques to be applied to the articulations in order to Solving the system (15), we obtain the matrix S(q),
obtain the desired configurations. The standard method such that
that uses the Euler–Lagrange equations is applied in
order to obtain the dynamic equations S ðqÞ = ½ s1 ðqÞ s2 ðqÞ  =
2 3
cðb cos f  d sin fÞ cðb cos f + d sin fÞ
d ∂L ∂L 6 7
 = Et  DT l ð12Þ 6 cðb sin f + d cos fÞ cðb sin f  d cos fÞ 7
dt ∂q_ ∂q 6 7 ð16Þ
6 c c 7
6 7
To elaborate the dynamic model, we choose the vec- 6 7
4 1 0 5
tor of the generalized coordinates as follows: Our goal 0 1
is to control the mobile robot over predefined trajec-
tories. Hence, we take xG and yG as elements of the vec- where c = r/2b; hence, the columns of S(q) are in the
tor of the generalized coordinates. However, we want null space of D, and consequently, S(q) spans N(D). As
to control the torques of each driving wheel. Therefore, q_ is in the null space of D, there exists a velocity vector
we add the angular position of the driving wheels as ele- v such that18
ments of this vector. We obtain the following form
q_ = S ðqÞv ð17Þ
M ðqÞ€q + Bðq, q_ Þ + GðqÞ = Et  DT l ð13Þ
where v = ½ u_ r u_ l T
where Multiplying equation (13) by ST(q), we obtain
6 Advances in Mechanical Engineering

 
tl System (24) can be partitioned into two subsystems
S T ðqÞEðqÞ = S T ðqÞM ðqÞ€q + S T ðqÞBðq, q_ Þ ð18Þ
tr facilitating its analysis and implementation of the con-
trol law; we write
This expression represents the inverse dynamic

model of the mobile robot. It helps in elaborating the x_ a = fa ð xÞ + ga ð xÞu
reference signals to be sent to the motors. We remark ð28Þ
x_ b = fb ð xÞ + gb ð xÞu
that the torques are written in terms of the robot accel-
eration. However, we are usually interested in control- Such that
ling the mobile robot in terms of the angular speed and 
acceleration of the wheels. Thus, we can establish a fa ð xÞ = S ðqÞv
ð29Þ
relation between the motor torques and the system fb ð xÞ = ½0
parameters. Deriving equation (17), we obtain
and
€q = S_ ðqÞv + S ðqÞv_ ð19Þ 
ga ð xÞ = ½0
ð30Þ
When substituting €q into equation (18), we obtain gb ð xÞ = I

From system (24), we can identify the time derivative

S T Et = S T ðqÞM ðqÞS ðqÞv_ + S T ðqÞM ðqÞS_ ðqÞv of the state variable xa and xb to be
ð20Þ
+ S T ðqÞBðq, q_ Þ 
x_ a = q_
ð31Þ
x_ b = v_
Mobile robot control
Then, the system can be expressed into its reduced
State space representation form
Considering the state vector x 
x_ a = fa ð xÞ
ð32Þ
T x_ b = gb ð xÞu
x=½q v ð21Þ

The system state equation (20) can be written as

Nonlinear feedback control
     
q_ S ðqÞv ½0 The nonlinear feedback control is used to eliminate
x_ = = + 1 t
v_ f2 ðS ðqÞM ðqÞS ðqÞÞ S T E
T nonlinearities present within the system and impose
ð22Þ desired linear dynamics. This approach was a success
of a large variety of applications. The theory of the
Such that f2 = (S T MS)1 S T (  M Sv
_  B). nonlinear feedback control has been well established in
Once the system is put in its state space form, it can Slotine and Li22 and Isidori.23 The first application in
be used for the determination of the control law. By robotics was introduced on robot manipulators in Tam
introducing a new input variable u, such that et al.24 and extended to mobile robots,25,26 where the
authors showed the usefulness of the approach and
 1
1 revealed its limitations. It is shown in Sarkar et al.26
t = (S T MS) S T E ðu  f 2 Þ ð23Þ
that input–output linearization can be achieved using a
static state feedback of a reference point whose coordi-
The state equation given by equation (22) can be fur-
nates are located on the front of the mobile robot and
ther simplified to
cannot be realized when the point is located at the mid-
 
    dle of the axis of symmetry with the driving wheel axis.
q_ S ðqÞv ½05 3 1
x_ = = + u ð24Þ For this reason, the reference point Pr located in the
v_ ½02 3 1 I2 3 1
front side and distant from Pa, by Lr, is chosen as the
In a more compact form, we can write controlling point. Based on the distance of the center
of gravity determined to be equal to 47.37 cm and given
x_ = f ð xÞ + g ð xÞu ð25Þ the length of the mobile platform L = 83.73 cm, and
  the distance from the driving wheel axis to a look
S ðqÞv ahead reference point Pr = 43 cm, one can deduce the
f ð xÞ = ð26Þ
fb distances d and Lr such that d = 7.17 cm and
  Lr = 35.83 cm.
½05 3 2 Hence, the output vector is written as
g ð xÞ = ð27Þ
I2 3 2
Abdessemed et al. 7

   
x r ðq Þ x + Lr cos f desired and measured trajectories such that e = h  y
y = hð xÞ = = G ð33Þ
y r ðq Þ yG + Lr sin f and e_ = h_  y_ , and Kp and Kv are diagonal gain
matrices whose elements are positive and well chosen to
Taking the time derivative of y yields assure stability and precision.
∂h ∂h
y_ = x_ = ðf ð xÞ + g ð xÞuÞ ð34Þ
∂x ∂x Stability analysis
y_ = Lf hð xÞ + Lg hð xÞu ð35Þ In the preceding section, we have shown that the stabi-
Letting a = c(d + Lr) and b = cb, we can verify lity and performance of the linearized system can be
that realized by a good choice of the gains Kp and Kv.
However, the dynamic behavior of the obtained system


 depends also on the internal dynamics. Since the con-
a u_ r  u_ l sin f + b u_ r + u_ l cos f
L f h ð xÞ =

trol law must take into consideration all the dynamics,
a u_ r  u_ l cos f + b u_ r + u_ l sin f
the internal behavior must be carefully considered. The
ð36Þ question is to know whether the internal states stay
and since Lgh(x) = [0], we compute once more the Lie bounded. The equations which determine internal
derivative of equation (35) in order to make the input dynamics are called zero dynamics. They are defined as
control u appear in the equation, we obtain the dynamics of the system when the state feedback law
u0 maintains the output y and its time derivatives iden-
€y = L2f hð xÞ + Lg Lf hð xÞu ð37Þ tically equal to zero, such that
 
a sin f + b cos f a sin f + b cos f L2f hð xÞ
Lg Lf hð xÞ = u0 =  ð43Þ
a cos f + b sin f a cos f + b sin f Lg Lf hð xÞ
ð38Þ
To study the behavior of the internal dynamics, con-
Deriving equation (36), one obtains sider the state transformation of the nonlinear system
" # represented by equations (25) and (33). It is trans-

2

ac _
u r  u_ l cos f  bc _
u 2
 _
u 2
sin f formed to the normal form by constructing the follow-
L2f hð xÞ =
2 r l

ing diffeomorphism, z1 = h1(x), z2 = Lfh1(x),
ac u_ r  u_ l sin f + bc u_ 2r  u_ 2l cos f
z3 = h2(x), and z4 = Lfh2(x). Since the variable x is
ð39Þ seven dimensional, three more components are added.
If the matrix LgLfh(x) in equation (38) is invertible, We choose to complete the vector by adding the right
then the control vector u can be determined. This is and left wheel angles, z5 = ul and z6 = ur, as well as the
always true since its determinant is always negative: orientation angle z7 = f. However, since z7 depends on
det(LgLfh(x)) = 22ab = 22c2(d + Lr)b \ 0. z5 and z6, by the relation z7 = c(z5 2 z6), the vector z is
Hence, from equation (37), the control law is only six dimensional, z = [h1(x), Lfh1(x), h2(x), Lfh2(x),
obtained as ul, ur,]T
2 3 2 3 2 3

1   z_ 1 z2 0 0
u = Lg Lf hð xÞ €y  L2f hð xÞ ð40Þ 6 z_ 2 7 6 L2f h1 ð xÞ 7 6 Lg Lf hð xÞ 0 7
6 7 6 7 6 7 
6 z_ 3 7 6 z4 7 6 0 0 7 u1
This control law eliminates the nonlinearities and 6 7=6 2 7+6 7
6 z_ 4 7 6 Lf h2 ð xÞ 7 6 L g L f h ð xÞ 7
permits to obtain a simple input–output relationship 6 7 6 7 6 0 7 u2
4 z_ 5 5 4 u_ 5 4 0 0 5
r
which can stabilize the whole system such that z_ 6 _ 0 0
ul
€y = v ð41Þ ð44Þ

Thus, the linearized system is controlled by the outer Equations (36) can be rearranged to obtain
loop chosen as    
Lf h1 ð xÞ z2
€  Kv e_  Kp e
v=h ð42Þ = ¼
Lf h2 ð xÞ z4
 " _ #
where the vector h = (xd , yd )T represents the desired tra- a sin f + b cos f a sin f + b cos f ur
=
jectories along the x- and y-axes, respectively. The error a cos f + b sin f a cos f + b sin f u_ l
and its time derivative are the difference between the
8 Advances in Mechanical Engineering

Thus
    
u_ r b sin f  a cos f b cos f  a sin f z2
=
u_ l b sin f  a cos f b cos f  a sin f z4
ð45Þ

After applying the feedback (40), the system of the

mobile robot is written in the following form
2 3
z_ 1
6 z_ 7
6 27
6 7
6 z_ 3 7
6 7 Figure 4. Input–output control with a neural network.
6 z_ 7
6 47
6 7
4 z_ 5 5
z_ 6 has been widely used for the identification and control
2 3
0 1 0 0 0 0 of dynamical systems. In this section, we propose a
60 0 0 0 0 07
6 7 design based on this type of recurrent neural network
6 7
60 0 0 1 0 07 to estimate the time derivative of the mobile robot coor-
=6
60
7z
6 0 0 0 0 077 dinates. For our convenience, the model expressed by
6 7 equation (13) can be written in state equation form.
40 b sin f  a cos f 0 b cos f  a sin f 1 05
0 b sin f  a cos f 0 b cos f  a sin f 0 1 Choosing x1 = q and x2 = q_ as the state variables of the
2 3 system, the following state representation is obtained
0 0
61 07 
6 7 x_ 1 = x2
6 7 ð48Þ
60 07
+6 7v x_ 2 = M 1 ðx1 ÞðEt  Bðx1 , x2 Þ  DT lÞ
60 17
6 7
6 7
40 05 If we suppose that only the position state can be
0 0 measured and the inertial matrix M is approximately
known, it imposes the necessity to observe the second
ð46Þ
state and account for uncertainties. The control law
where v can be chosen as in equation (42) or by pole presented in this article is based on only estimating a
placement, that is, v = 2Kz. The internal dynamics portion of the second state vector such that
associated with the input–output linearization corre- _ ^u_ 1 , ^u_ 2 ). The assumption used is that
^x2 = (_xG , y_ G , f,
spond to the two last equations. In this case, the equa- there exist a vector W(x) and an optimal parameter
tions are split into two groups, where the second group vector u such that
corresponds to the zero dynamics and written as
x = uT W ð49Þ
    
z_ 5 b sin f  a cos f b cos f  a sin f z2
= For our purpose, the Elman recurrent neural net-
z_ 6 b sin f  a cos f b cos f  a sin f z4
work model is chosen as an observer to estimate the
ð47Þ nonmeasurable states. It is inserted in the controlled
system as shown in Figure 4. The estimated output can
The zero dynamics are defined as the dynamics of
be expressed in the following form
the system when the output y and its time derivatives
are identically equal to zero, that is, z1 = z2 = z3 =
x(2) = ^uT W ð50Þ
z4 = 0. Hence, z_ 5 = z_ 6 = 0, which means that z5 and z6
remain constant. Consequently, the zero dynamics are where ^u is the parameter vector that is computed when
stable but not asymptotically stable. Yun and the network is trained using dynamic back-propagation
Yamamoto27 have shown that the mobile robot has a algorithm. The control objective is to force the tracking
stable behavior if it moves in the forward direction but error between the desired output h_ and the estimated
an unstable behavior when it has to move backward. time derivative output ^y_ to converge to a neighborhood
of zero. Consider the ENN structure shown in Figure 5.
It is a recurrent neural network with four layers of neu-
ENN observer
rons, namely, input layer, hidden layer, context layer,
ENN is a partial neural network model that was first and output layer. The structure has an output layer
proposed for speech processing.12 Recently, the ENN weight matrix W (0) whose elements connect the hidden
Abdessemed et al. 9

2
x(1)
i = 1 ð51Þ
1 + e4(neti )
where

+5
mX
neti = Wij(0) x(0)
j i = 1, . . . , m = 7 ð52Þ
j=1

W (0) is the matrix that includes the connections

weights between the input–context layers and the hid-
den layer.

Output layer
In this application, only terms that cannot be measured
are estimated. Therefore, only a portion of the second
Figure 5. Structure of the ENN. state vector is observed, that is, x(2) = ½b
u_ 1 , b
u_ 2 T . Each
neuron of the output layer is described by an activation
linear function, and the output of the jth output is
layer to the output layer, and a hidden layer weight expressed as
matrix W (0) = ½W (0i) , W (0c)  consisted of two sub-weight
matrices W (0i) and W (0c) , whose elements connect the
x(2)
i =^x2i = G 3 ðneti Þ, i = 1, 2 ð53Þ
input layer and the context layer to the hidden layer,
respectively. where PG is a certain chosen gain and
=7 (1) (1)
neti = m j = 1 Wij xj , such that W
(1)
is the matrix that
includes the connection weights between the hidden
Input layer and output layers. We can define the measured error as
In this layer, the input variables are the measured posi- the squared error between the desired output and the
tion and orientation of the mobile robot: (xG , yG , f), and estimated output defined as
the corresponding computed controller vector: (ur , ul ).
1 2 1  d _
2
2 
E= h_  ^y_ = x_  ^y1 + y_ d  ^y_ 2 ð54Þ
2 2
Context layer
The aim is to find the appropriate weights by mini-
The context layer behaves like the input layer. The con-
mizing the error E. The output error signals are found
text node input values are simply the calculated values
by taking the ordered derivative given by the
of the hidden output nodes transmitted via recurrent
expression
connections through a unit delay, that is, x(1) (t  1).
Thus, the number of context nodes is equal to the
∂+ E ∂E ∂E ∂^y_ 1 ∂E ∂^y_ 2
hidden layer neurons. In this case, the input data of e2, i = = = +
the input layer 0 are x(0) = (xG , yG , f, ur , ul , x(1) ∂x(2)
i ∂x(2)
i
_ (2)
∂^y1 ∂xi ∂^y_ 2 ∂x(2)
i
ð55Þ
(t  1), rxc(0) (t  1)), such that x(1) (t  1) and rx(0)
c (t  1)
∂^_
y
∂ ^_
y
denote the output signals of the hidden and context e2, i =  x_ d  ^y_ 1 1
 y_ d  ^y_ 2 2
∂x(2)
i ∂x (2)
i
neurons in the previous time, respectively, and 0  r\1
is the self-connecting feedback gain. Proceeding with each error signal, we obtain


e2, 1 =  x_ d  ^y_ 1 ða sin f + b cos fÞ
Hidden layer

 y_ d  ^y_ 2 ða cos f + b sin fÞ
In this layer, each neuron is described by an activation
ð56Þ
sigmoid function. The number of hidden layer neurons e2, 2 =  x_ d  ^y_ 1 ða sin f + b cos fÞ


is equal to m. This number is chosen arbitrary; one can  y_ d  ^y_ ða cos f + b sin fÞ
2
always begin with a minimum number of nodes and
keep increasing until no more improvement could be For the hidden, input, and the context layer nodes,
seen. The output x(1)
i of the ith hidden neuron can be the error signals can be derived using the functional
expressed as decomposition rule
10 Advances in Mechanical Engineering

(l + 1)
NX
∂x(lj + 1)
el, i = el + 1, j ð57Þ
j=1 ∂x(l)
i

where N(l + 1) is the number of neurons in layer

(l + 1). One can verify that the obtained error signal
back to the hidden layer is obtained as

=5
nX
∂x(2)
j
e1, i = e2, j , i = 1, . . . , 7 ð58Þ
j=1 ∂x(1)
i

Proceeding with the calculations, one can find

=5
nX
e1, i = e2, j Wji(1) , i = 1, . . . , 7 ð59Þ
j=1

The error signal back propagated to the input layer Figure 6. Evolution of the mobile robot along a straight line
is obtained as trajectory.

=7
mX
∂x(1)
j
=7
mX
∂x(1)
j ∂ netj
e0, i = e1, j = e1, j
∂+ E ∂x(2)
j=1 ∂x(0) j=1 ∂ netj ∂x(0) ð60Þ = e2, i 3 i
= e2, i x(1)
j ð64Þ
i i
∂Wij(1) ∂Wij(1)
i = 1, . . . , 12
For the input-to-hidden layer weight adjustment:
It can be easily shown that l=0
  2 
=7
mX
∂x(1)
j
=7
1 mX ∂+ E ∂x(1) ∂x(1) ∂neti
e0, i = e1, j = e1, j 1  x(1)
j Wji(0) (0)
= e1, i 3 (0)
i i
= e1, i
3
j=1
(0)
∂xi 2 j=1 ∂Wij ∂Wij ∂neti ∂Wij(0)
  ð65Þ
1  2
i = 1, . . . , 12 = e1, i 1  x(1) x(0)
i j
ð61Þ 2

The Elman back-propagation algorithm is used to Simulation results

update the output and hidden layers recursively at each
iteration so that E(t) is minimized. Momentum method We present in this section the computer simulation
is used to accelerate the convergence of the EBP learn- results. The mobile robot is supposed to track two dif-
ing algorithm. In this case, a fraction of the most recent ferent reference trajectories, namely, a straight line path
weight adjustment is added according to the following and a circular path (cf. Figures 6 and 7, respectively).
formula These trajectories are purposely chosen to be analogous
to those found in Coelho and Nunes.7 In the presented
∂+ E simulations, the system is linked with the observer
Wij(l) ðt + 1Þ = Wij(l) ðtÞ  a + bDW (l) ðtÞ, based on an ENN design. The training algorithm is car-
∂Wij(l) ðtÞ ð62Þ
ried out through a back-propagation algorithm. The
for l = 0, 1 gradient of the objective function with respect to neural
network parameters is calculated and used to update
where a is the learning rate coefficient of the EBP and the weight parameters, starting from the output layer
b is a user-selected positive momentum constant typi- and propagating backward. To accelerate convergence,
cally chosen between 0.1 and 0.9. The third term can be a momentum term is being added to the learning
determined as follows expression. The values of the learning rates depend on
the problem being solved since the convergence
∂+ E ∂+ E ∂x(li + 1) ∂x(li + 1) depends significantly on these values. Hence, there is a
= 3 = e(l + 1), i 3 ð63Þ
∂Wij(l) ∂x(li + 1) ∂Wij(l) ∂Wij(l) tradeoff between small values and large values. Small
values may lead to the solution but with a small con-
For the hidden-to-output layer weight adjustment: vergence rate; however, large values may overshoot the
l=1 solution. After some trials, the values retained in this
application are found to be a = 0.8 and b = 0.3. We
Abdessemed et al. 11

Figure 9. Distance error between actual and desired paths

Figure 7. Evolution of the mobile robot along a circular performed by the mobile robot in following a circular path.
trajectory.

Figure 8. Distance error between actual and desired paths Figure 10. Heading angle plot.
performed by the mobile robot in following a straight line path.

consider errors of 10% on all parameters of the system actual and the desired paths. Figures 10 and 11 depict
in order to test the robustness of the proposed scheme the evolution of the heading angles in both cases. In
with respect to the parametric and structural errors. fact, in case of a straight line path tracking, the transi-
Simulation results are performed for a specific initial ent angle lasts about 8 s before the heading angle stabi-
condition which incorporates the true nonlinear nature lizes to a fixed positive value of 45°. However, this
of the problem with the gains Kp = 3 and Kv = 7. In all angle is observed to increase linearly as the mobile
the experiments, the velocities are reconstructed from robot turns around the circular path. Note that the
the observer. In Figure 6, the initial pose of the mobile desired paths are followed with the mobile robot in the
robot is (0°, 0°, 90°), while in Figure 7, the initial pose forward motion and the measured forward measured
is (5°, 30.6°, 225°). In both cases, the mobile robot tries linear velocities are depicted in Figures 12 and 13. The
in the first step to cancel the position error by rapidly estimated and measured angular velocities of the right
approaching the reference paths. Once the paths are wheels are shown in Figures 14 and 15 for the straight
reached, the path following is achieved in a well-stable line and circular paths, respectively. Figures 16 and 17
behavior. This tracking behavior is observed in Figures depict those of the left wheels for both tracking paths.
8 and 9, which show the distance error between the The curves of the estimated angular speeds attained the
12 Advances in Mechanical Engineering

Figure 11. Heading angle plot. Figure 14. Estimated and measured right wheel speeds of the
mobile robot in following a straight line path.

Figure 12. Forward velocity of the mobile robot in following a Figure 15. Estimated and measured left wheel speeds of the
straight line path. mobile robot when tracking a circular path.

Figure 13. Forward velocity of the mobile robot in following a Figure 16. Estimated and measured right wheel speeds of the
circular path. mobile robot in following a straight line path.
Abdessemed et al. 13

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with
respect to the research, authorship, and/or publication of this
article.

Funding
The author(s) disclosed receipt of the following financial sup-
port for the research, authorship, and/or publication of this
article: This work is supported by NPST program by King
Saud University (project no. 08-ELE-300-02).

References
1. Hachem Lamiti A, Ben Khlifa MM, Gorce P, et al. The
command of a wheelchair using thoughts and gaze. In:
Figure 17. Estimated and measured left wheel speeds of the 16th IEEE Mediterranean electrotechnical conference
mobile robot in following a circular path. (MELECON), Hammamet, Tunisia, 25–28 March 2012.
New York: IEEE.
2. Lankenau A and Röfer T. Smart wheelchairs, state of
measured ones after small lapse of time, which prove the art in an emerging market. Zeitschrift KI mit Schwer-
the convergence of the ENN observer. The simulation punkt Autonome Mobile Systeme 2000; 4, http:
//www.informatik.uni-bremen.de/kogrob/papers/ki00.pdf
carried on the established model of the mobile robot
3. Kim D-H and Oh J-H. Tracking control of a two-
showed the feasibility of the approach and the success-
wheeled mobile robot using input–output linearization.
fulness of the observer in estimating the missing states Control Eng Pract 7(3): 369–373.
of a system subjected to error modeling. 4. Oriolo G, De Luca A and Vendittelli M. WMR control
via dynamic feedback linearization: design, implementa-
tion, and experimental validation. IEEE T Contr Syst T
Conclusion 2002; 10: 835–852.
5. Jiangzhou LU, Sakhavat S, Ming XIE, et al. Sliding
The theoretical work in the field of wheeled mobile mode control for nonholonomic mobile robot. In: Pro-
robot control is large and offers different solutions to ceedings of the international conference on control, auto-
solve the problems of displacement. However, the inte- mation, robotics and vision, 2000, pp.465–470, http://
gration of the control laws on an experimental requires citeseerx.ist.psu.edu/viewdoc/
us to take into account certain constraints. We pre- download?doi=10.1.1.24.9760&rep=rep1&type=pdf
sented in this article the different steps in the design, 6. Hu N and Wang C. Adaptive tracking control of an
modeling, and control of a mobile robot modular uncertain nonholonomic robots. J Control Theory Appl
wheel. We initially identify important parameters for 2012; 10: 56–63.
the control of mobile robot. This allowed us to make 7. Coelho P and Nunes U. Path-following control of mobile
robots in presence of uncertainties. IEEE T Robot 2005;
choices not only on the assumptions of behavior but
21: 252–261.
also on the formalism adopted for the kinematic and 8. Ortigoza RS, Ortigoza GS, Guzmán VMH, et al. Trajec-
dynamic modeling. In this article, feedback linearizing tory tracking in a mobile robot without using velocity
control law is used to allow tracking a specified refer- measurements for control of wheels. IEEE Latin Am
ence trajectory. It is supposed that only the position Trans 2008; 6: 598–607.
state can be measured, and some of the parameters that 9. Do KD, Jiang ZP and Pan J. Simultaneous tracking and
intervene in the modelization are approximately known. stabilization of mobile robots without velocity measure-
To solve these problems, an ENN-based observer is ments. In: Proceedings of the 42nd IEEE conference on
proposed to estimate the nonmeasured velocity vector decision and control, Maui, HI, 9–12 December 2003.
of the mobile robot. Simulations are carried out and the New York: IEEE.
results obtained show that the complete controlled sys- 10. Wang Q, Tong G and Xing X. Wheeled mobile robot
tracking without velocity measurement. Appl Mech Mater
tem presented good performances and stability under
2013; 373–375: 231–237.
uncertainties on the model parameters, which confirm 11. Shu P, Ishibashi Y, Shibata H, et al. Adaptive tracking
the feasibility of the approach. Further work in this control scheme for wheeled mobile robots without mea-
direction is under progress, and our goal is to build the surement of longitudinal velocity. In: Proceedings of the
controller on the true PowerBot and carry out the 2012 international conference on advanced mechatronic
experimental tests. systems, Tokyo, Japan, 18–21 September 2012. New
York: IEEE.
14 Advances in Mechanical Engineering

12. Rothkrantz LJM and Nollen D. Speech recognition using 19. Yun X and Yamamoto Y. On feedback linearization of
Elman neural networks. In: Matousek V, Mautner P, mobile robots. Department of Computer and Information
Ocelı́ková J, et al. (eds) Text, speech and dialogue. Vol- Science Technical Reports, University of Pennsylvania,
ume 1692 of the series lecture notes in computer science. Philadelphia, PA, 1992.
Berlin: Springer, 1999, pp.146–151. 20. http://www.mobilerobots.com/ResearchRobots/Power
13. Sawaragi T and Kudoh T. Self-reflective segmentation of Bot.aspx
human bodily motions using recurrent neural networks. 21. Zulli R, Fierro R, Conte G, et al. Motion planning and
IEEE T Ind Electron 2003; 50: 903–911. control for nonholonomic mobile robots. In: Proceedings
14. Wlas M, Krzeminski Z, Guzinski J, et al. Artificial- of IEEE international symposium on intelligent control,
neural-network-based sensorless nonlinear control of Monterey, CA, 27–29 August 1995, pp.551–557. New
induction motors. IEEE T Energy Convers 2005; 20: York: IEEE.
520–528. 22. Slotine J-J and Li W. Applied nonlinear control. Upper
15. Lin FJ and Hung YC. An Elman neural network control Saddle River, NJ: Prentice Hall, 1991.
system for linear piezoelectric ceramic motor using 23. Isidori A. Nonlinear control systems. 2nd ed.Berlin:
FPGA. In: Australasian universities power engineering Springer-Verlag.
conference (AUPEC ’08), Sydney, NSW, Australia, 14–17 24. Tam TJ, Bejczy AK, Isidori A, et al. Nonlinear feedback
December 2008, pp.1–6. New York: IEEE. in robot arm control. In: Proceedings of IEEE conference
16. Zhang ZQ, Tang Z and Vairappan C. A novel learning decision and control, Las Vegas, NV, 12–14 December
method for Elman neural network using local search. Neu 1984, pp.736–751. New York: IEEE.
Inf Pro 2007; 11: 181–188. 25. D’Andréa-Novel B, Campion G and Bastin G. Control
17. Lin FJ, Lee SY and Chou PH. Intelligent nonsingular ter- of nonholonomic wheeled mobile robots by state feed-
minal sliding-mode control using MIMO Elman neural back linearization. Int J Robot Res 1995; 14: 543–559.
network for piezo-flexural nanopositioning stage. IEEE 26. Sarkar N, Yun X and Kumar V. Control of mechanical
T Ultrason Ferr 2012; 59: 2716–2730. systems with rolling constraints application to dynamic
18. Sarkar N, Yun X and Kumar V. Dynamic path follow- control of mobile robots. Int Conf Robot Autom 1994; 13:
ing: a new control algorithm for mobile robots. In: Pro- 55–69.
ceedings of the 32nd IEEE conference on decision and 27. Yun XP and Yamamoto Y. Stability analysis of the inter-
control, San Antonio, TX, 15–17 December 1993, vol. 3, nal dynamics of a wheeled mobile robot. J Robot Syst
pp.2670–2675. New York: IEEE. 1997; 14: 697–709.