Design neural network-based adaptive controller

for underactuated 3-wheeled Mobile Robot

Phong Tuan Dang1, Minh Duc Vu1 , Manh Do Dung1 , Ngoc Thanh Pham1 ,Huu Nguyen Thai2 ,Hai le Xuan1*
[email protected], [email protected], [email protected],[email protected]
[email protected],
1International school, Vietnam National University
2Vinh University of Technology Education

*Corresponding author: [email protected]

Abstract 3-Wheeled mobile robots (3WMR) are used widely in various industries. They can operate in hazardous environments and
complete tasks automatically without direct human intervention. In this paper, since the 3-Wheel Mobile Robot is an insufficient actuated
system, the hierarchical sliding mode controller is designed to make the 3WMR system track to the desired trajectory. Moreover, in
reality, there are various disturbances that have negative effects on the system's operation. an RBF neural network; therefore, applied
to the closed scheme to estimate all the uncertainties and the system's nonlinearities. The stability of the closed-loop robot system is
proven rigorously by the Lyapunov function. The effectiveness of the proposed methods is verified in the simulation section by the
Matlab/Simulink platform.
Keywords: 3-Wheeled Mobile Robot, under-actuated system, Hierarchical Sliding Mode Control, RBF Neural Network
1. INTRODUCTION
Today, Wheeled Mobile Robots (WMRs) are becoming increasingly popular due to their mobility and impressive
features. Scientists and engineers worldwide are dedicating significant research and attention to these robots' autonomous
control and monitoring. WMRs are used more frequently in various industries, such as the military, space exploration, and
industry. They can operate in hazardous environments and complete tasks autonomously without human intervention
(Amirkhani et al.2020)( Matraji et al.2018). In industrial production lines, autonomous vehicles are mobile robots with wheels
used to move quickly and accurately in multiple directions (Thi Kim Duyen Ha et al.2019).WMR offers the advantage of cost
savings while ensuring flexibility and speed in design. These vehicles are equipped with two rear wheels, each driven by
an independent motor, and a freewheel in front, allowing them to move in all directions. Due to its remarkable structure, it
is suitable for various uses, such as transporting goods and materials in factories.

The three-wheeled mobile robot (3WMR) is a nonlinear mechanical model that is difficult to accurately model
kinematically and dynamically due to the many unknown parameters. This type of vehicle is also classified as an under-
actuated object class, meaning with fewer actuators than the DOF to be controlled. Furthermore, 3WMRs have non-
holonomic constraints, meaning their movement is not entirely free, and horizontal movement of the body is impossible.
As a result, designing control algorithms for 3WMRs is a challenging task, especially when deploying controllers to real
devices, as they are susceptible to mechanical uncertainties and environmental disturbances during operation. In (Do Khac
Tiep et al. 2018) the authors developed a PID controller in combination with fuzzy to determine the controller parameters.
But the method just be verified with a small number of case studies. On the other hand, some controllers were derived for
non-linear systems. Backstepping controller (Fierro, Rafael, and Frank L. Lewis (1997))., and sliding mode control (Yang, Jong-
Min, and Jong-Hwan Kim (1999)). However, these methods have not dealt with unexpected disturbances or system
uncertainties. In fact, in practical situations, the parameters of the system’s models may change over time or be difficult to
obtain accurately. Therefore, it is important to develop a new control methodology that can improve the systems’
performance adaptively and optimally.

Neural Networks are useful tools that are devised for the purpose of finding basic relationships in data sets, which act
resembling a human brain. Neural networks have relative accuracy with statistical methods such as curve graphs and
regression analysis. Neural Network contains layers that contain interconnected nodes. Each node is a perceptron with a
similar structure to the multi-linear regression function. Inside a multi-layer perceptron, they will be arranged according to
the layers that are connected to each other. The input layer will collect the input samples and the output layer will collect
the classifications or output signals that the input samples may reflect back. Neural Network is an important and
indispensable component in control engineering.

Inspired by the aforementioned works, in this paper, an adaptive control method based on an RBF neural network and
hierarchical sliding mode technique for underactuated 3-wheeled Mobile Robots is developed. The mathematical model of
the 3WMR is separated into two subsystems. The first subsystem is utilized for controlling the vehicle's angle whereas the
remaining subsystem is used for controlling the vehicle's position. In the following step, an adaptive control method for this
system based on the neural network and hierarchical sliding mode technique is designed to achieve the desired control
quality, especially in the matter of keeping the vehicle’s position and rotation stable at the desired references. Finally, a
comprehensive analysis of the system's stability will be given during the step-by-step design procedure. That is the major
contribution of this paper.
 The remainder of this paper is organized as follows: section II presents the system description and mathematical models
of the studied mobile robot. The synthesis of the adaptive control regime is demonstrated in Section III. In Section IV,
comprehensive simulations will be undertaken to validate the effectiveness of our control regime. Finally, Section V
concludes our paper.
m n
In this paper, m rows and n columns. The notation n is an short presentation of
is a space of matrices having
1 n
. x is a column vector of n real numbers q1 , q2 ,..., qn . x 
n1 n T
is a transpose vector of x. |x| is the absolute
value of a real number x . In addition, tr ( A) is the trace operator of matrix A .

2. PROBLEM PRELIMINARY
2.1. Represent the mathematical model in the state space
The scheme of the movement of the 3-Wheels Mobile Robot (3WMR) is depicted in Figure 1, which is shown from top
to bottom. As can be seen from this figure, the constituent movement of the 3WMR is determined through 2 coordinate
systems. The Oxy coordinate system is a fixed coordinate system mounted on the floor, and it is a so-called Earth-fixed
coordinate system. The remaining one is Oxy is the coordinate system that is in the center mass of the 3WMR robotics
system. Through 2 coordinate systems, the motion of one is completely determined. That is the motion along the x-axis
x(t ) , the motion along the y-axis y(t ) , and the rotation angle of the vehicle's body  (t ) To brief and convenience, we
will be used these notations x, y ,  to replace these notations x (t ), y (t ),  (t ) in the following sectors.

Fig1.. The diagram of the 3-Wheel Mobile Robot)

By denoting u1  T1  T2 and u2  T1  T2 . The mathematical model of the 3WMR system is represented by the
below equations (Amirkhani et al.2020)
   b1u1

 x  2 ( xcos  ysin ) sin  b2u2cos (1)
 y  2 ( xcos  ycos )cos  b2u2 sin

b 1
Where b1  , b2  . By denotation 1   ,  2   , x1  x, x2  x, x3  y , x4  y , the equations (1)
rJ mr
becomes:
1   2

 2  b1u1
 x1  x2
 (2)
 x2  2 ( x cos   y sin  ) sin    b2 cos   u2
 x3  x4

 x4  2 ( x cos   y sin  ) cos    b2 sin   u2

The equations (2) are utilized to design the controller for the WMR system.
2.2. Control objectives
With the purposes of this paper, the control objectives are:
 Designing the backward-hierarchical sliding mode controller (BHSMC) for the under-actuated WMR system. The mission
of the BHSMC controller is to make the studied system track to the desired trajectory.
 Developing an RBF neural network to estimate all the system's nonlinearities and uncertainties. the convergence of the
network's weight is proven through the Lyapunov function. The controller constructed by RBF neural network and BHSMC
controller is called the ABHSMC Controller.
To achieve these above proposes, firstly, the mathematical model of the studied system (2) is separated into two
subsystems. The subsystem for controlling the vehicle body's angle and the subsystem for controlling the positions of the
vehicle. They are depicted in more detail in the next section.
3. METHODOLOGY
3.1 Vehicle Body's Angle Control
Consider the subsystem for controlling vehicle body's angle:

 1   2
 (3)
 2  b1u1

Where b1  0 is a positive constant.

Step 1: Consider system (3), the tracking error is defined by:

e    d (4)

Where d is the desired angle of the vehicle's body, for simplicity, it is assumed d to be constant, then it is easy to
deduce that  d  0 and  d  0 . We are had a linear system which is equivalent to the subsystem (3):
.
e  Ae  Bu1 (5)

0 1 T e
Where: A  , B   0 b1  and e    .
0 0 e
Step 2: Design of the pole placement controller:
u1  R e (6)
12
Where R  is a matrix such that it makes the matrix A  BR  is stable. Thus, lim e  0 . It means that
t 

lim    d . In this paper, we are used the Roppenecker algorithm in (Nguyen Doan Phuoc(2016))to determine the matrix
t 

R . The chosen poles for the system (5) are 5, 6 yields:
R   42.25 15.49  (7)

3.2. Vehicle's Position Control

Consider the subsystem for controlling the vehicle's position:

 x1  x2

 x2  f1  x1 , x2   g1u
 (8)
 x3  x4
x  f  x , x   g u
 4 2 3 4 2

Where x1  e1  x  xd and x3  e3  y  yd these functions f ( x1 , x2 )  2 ( x cos   y sin  ).sin  ;

g1 ( )  b2 cos  ; f 2 ( x1 , x2 )  2 ( xcos  ycos ) and g 2 ( )  b2 sin  The step-by-step design procedure is
represented as the following:
Step 1: Consider subsystem of equations (8)

 x1  x2
 (9)
 x2  f1  x1 , x2   x d  g1 ( )u1
Where u1 is the control input that makes the subsystem (9) stable. Let us define the sliding surface

.
S1   x1  x2   x1  x1 ,   0 (10)

1 2
Using the candidate positive definite Lyapunov function: V1  S1 Differentiating the Lyapunov function from both
2
sides yields:
.
V1  S1S1  S1 ( x1  x1 ) (11)

The input signal u1 consists of two elements u1eq and u1sw Thus, u1  u1eq  u1sw . The element u1eq is the control signal
that brings the system stabilize in the sliding surface. u1sw is the control signal that steers the sliding surface to the origin.
Using (9), (11) becomes:

S1   x2  f1 ( x1 , x2 )  g1 ( )u1eq  g1 ( )u1sw  (12)

Therefore, if we are chosen u1eq and u1sw satisfying:

g1 ( )u1eq   x2  f1  x1 , x2   xd (13)
g1 ( )u1sw  ksignS1  S1 (14)

Substituting (13) into (14) yields:

V1  kS1signS1  S12  0 (15)

Therefore, the subsystem (9) is asymptotically stable. The control signal that makes the subsystem stable is satisfying the
following equation:

g1 ( )u1  u1eq  u1sw   x2  f1  x1 , x2   ksignS1  S1 (16)

Step 2: Similarly, for the second subsystem of (8), u2eq is also satisfied:

g 2 ( )u2 eq   x4  f 2  x3 , x4   yd (17)

By the definition of sliding surface S 2 and the Lyapunov function V2 as

S 2   x3  x4 (18)
1
V2  S22 (19)
2
Step 3: Construct the sliding manifold of the system for controlling the vehicle's position:

S   1S1   2 S 2 (20)

If choosing  1 ,  2 appropriate S  0, S1  0 , and it can be deduced that S 2  0 . Differentiating from both sides of
equation (20) yields:

S   1S1   2 S2   1   x1  x1    2  x3  x3  (21)

Noting that x3  x4 ; x3  x4  f 2  x3 , x4   g 2 ( )u . Equation (21) becomes:

S   1   x2  f1  x1 , x2   xd  g1 ( )u    2  x4  f 2  x3 , x4   yd  g 2 ( )u  (22)

In the above two steps, the control signal u is designed as:

u1  u1eq  u 1sw
 u2  u2 eq  u 2 sw

u  u1eq  u sw u2eq (23)

Where usw  usw1  usw 2 . Substituting $u$ calculated in (23) into (22) and using equations (17) and (13), we are had

S   1 g1u2 eq   2 g 2u1eq  ( 1 g1 ( )   2 g 2 ( ))usw (24)

1 2
Using the candidate Lyapunov function V S . Differentiating from both sides of the Lyapunov function along 24
2
we are had:

 1 g1 ( )u2eq   2 g 2 ( )u1eq   S
usw   (25)
 1 g1 ( )   2 g 2 ( )
Where   0 is an arbitrarily positive number. Then the derivation of the Lyapunov function is V    S  0 . From
2

equations (13), (17) and (23), the input signal of the BHSMC controller is designed as:

 1 f1 ( x1 , x2 )   2 f 2 ( x1 , x2 )
u  u1eq  u sw u2eq    ( x2 , x4 , S ) (26)
 1 g1 ( )   2 g 2 ( )
Where f1 ( x1 , x2 )  2 ( x cos   y sin  ) sin  ; g1 ( )  b2 cos  ; f 2 ( x1 , x2 )  2 ( xcos  ycos )cos and

 1 x2   2 x4  S   1 xd   2 yd
g 2 ( )  b2 sin  and ( x2 , x4 , S ) 
 1 g1 ( )   2 g 2 ( )

3.3. RBF Neural network design

3.3.1. Neural RBF network structure

Fig. 2. Neural RBF network structure

The control signal u in (26) only implemented in case of the known functions f1 ( x1 , x2 ) and f 2 ( x1 , x2 ) . But in reality,
these functions are difficult to determine due to the ignored factors in the modeling process. Thus, this paper proposed the
RBF neural network to estimate them. The neural network structure is shown in Figure (2) including 3 layers.
Layer 1: The input layer of the network. In this layer, the input signals are state variables and x, y ,  their derivatives x ,
y ,  These input signals are passed directly to the next layer.
Layer 2: The hidden layer. This layer consists of an array of compute units called hidden nodes. Each hidden node
N 1
(neuron) is activated by a radial basis function. The output vectors of the hidden layer h  [h1 , h2 ,..., hn ]T  are
calculated as follows:
 q  cj 2
 q  cj 2

exp ( )
b 2j
hj  , j  1, 2..., N (27)
q  ck 2
 q  ck 2
 exp (
n
k 1 )
bk2
Where N is the number of the hidden notes.

Layer 3: the output layer. At this layer, the output signal is the two components to be estimated in the model fˆ1 ( x1 , x2 ) ,
fˆ2 ( x1 , x2 ) by:
n
fˆi ( x1 , x2 )   w ijh j  W
ˆ T h, i  1, 2
i (28)
j 1

 N 1

In which Wˆ1  w11 , w12 ,..., w1N  R and Wˆ2  [w 21 , w 22 ,..., w 2N ]  R
N 1
are the estimated ight of this
network. Therefore, the exact nonlinear functions can be computed by:

fi ( x1 , x2 )  fˆi ( x1 , x2 )   i  Wˆi T h   i , i  1, 2 (29)

Where  i is the approximation error of the RBF neural network. The trunction error of network's weight is

W i  Wi  Wˆi . The next task is to determine a update rule such that lim W i  0.
t 

3.3.2. Design neural network-based adaptive controller for 3WMR

By using the RBF neural network depicted in Figure (2), the estimated inout signal u in (26) is calculated by:
 1 f1 ( x1 , x2 )   2 f 2 ( x1 , x2 )
uˆ    ( x1 , x2 , S ) (30)
 1 g1 ( )   2 g 2 ( )
 1W1T h   2W2T h
  ( x1 , x2 , S ) (31)
 1 g1 ( )   2 g 2 ( )
And then, the derivation of the sliding surface is:
T T
S   1 W 1 h   2 W 2 h   11   2 2   S (32)

To determine the updating law, the candidate Lyapunov function is defined as:

1 2 1  T  1  T 
Vo  S  tr  W 1 F1 W1    W 2 F2 W2  (33)
2 21   2 2  
Where 1  0 and  2  0 are two positive numbers. These matrices F1 and F2 are two symetric positive-define
matrices. tr (.) is the trace operator. Derivative (33) with respect to time yields:

1  T ˆ  1  T 
Vo  SS  1 1
tr  W 1 FW tr  W 2 F2Wˆ2  (34)
1   2  
Substituting equation (32) into equation (34) yields:
T
1  T ˆ  T
1  T 
Vo    S 2   ( S )  S 1 W 1 h  tr  W 1 FW
1 1  S  W 2 h  tr  W 2 F2Wˆ2  (35)
1  
2
2  
T
 T 
Where  ( S )  S   11   2 2  satisfying  ( S )   ,   0 . Noting that S i W i h  tr  W i  i hS  with i  1, 2
 
Equation (35) becomes:
  T   T 
1
Vo    S 2    tr W 1   1hS  FW ˆ    tr W 2   hS  1 F Wˆ   (36)
  1 1 1
  
2
2 2 2

Therefore,if the updating laws for two weight vectors W1 , W2 are:

Wˆ1  F111 1hS (37)

Wˆ2  F21 2 2hS (38)

and  is chosen large enough, the stability of the slosed-loop system is ensured:

Vo    S 2      S 2 (39)

In addition, it is easy to see that the bounded of the second-order Lyapunov function Vo  vo , vo  0 . Thus, the closed-
loop system is asymptotically stable. The scheme of the closed-loop system is shown in Figure (3).

Figure 3 The scheme of the closed-loop system

Remark 1: Due to the system's holonomic constraint, the desired value for the vehicle's angle must be determined by:

 yd 
 d  arctan   (40)
 xd 
4. NUMERICAL SIMILATION
In this section, the effectiveness of the proposed adaptive BHSMC controller is verified through two cases with the influence
of random external disturbances. The parameters of the vehicle are selected the same with paper (Dang Son Tung et al .2023)
such as m  15(kg ) , J  3.818(kgm 2 ) , r  0.07(m) and b  0.19(m) . The parameters of the adaptive BHSMC
controller are chosen as  1  10 ,  2  2.5 ,   6 , 1  1.5 and  2  0.25 . Two matrices F1  F2  I N . The initial
conditions for the 3WMR system is assumed that x (0)  0( m) , y (0)  0( m) and  (0)  0( rad ) . The random
external disturbance is created by random function in Matlab and its graph is shown in Figure 4. All the numerical results
are simulated within 25 seconds.
 Figure 4 The randomly external disturbance

To validate the performance of the vehicle's position controller with sinusoidal desired references and compare it to the
traditional BHSMC controller. The desired references in this scenario are selected as the following:

xd  t   0.25t  0.75 (41)

yd  t   0.5  0.25sin  0.2 t  (42)
And by using equation (40), the desired value for the vehicle's angle is calculated a  d (t )  arctan  0.2 cos  0.2 t  
.

Figure 5 The trajectory response of the 3WMR system

Figure (5) shows the trajectory responses with 15 kg mass. From this Figure, it is easily seen that the ABHSMC controller
by integrating the RBF neuron network in the design procedure still makes the closed-loop system to be stable at the desired
trajectory with a finite time. In spite of the non-asymptotic stability of the Robot system, the convergence of the Mobile
Robot’s states still goes to the origin even if the high-frequency external disturbances with a large amplitude.
 Figure 6 The tracking errors of the vehicle's position along the x-axis and y-axis

The tracking errors of the vehicle's positions along the x-axis and y-axis guaranteed by the ABHSMC with combining the
RBF neural network under the high-frequency disturbances are reflected in Figure 6. The special thing here is that the
positional response under the steer of BHSMC does not track the references. By contrast, for the ABHSMC controller, the
neural network’s weight has been always changed and updated for compensation for the error affected by lumped
disturbance. Consequently, The stability of the closed-loop system is still maintained in a short time interval. With the pole-
placement controller for the vehicle's angle, the rotation of the studied system is asymptotically stable.

Figure 7 Vehicle's rotation of the 3MWR

Through the simulation section, the ABHSMC controller constructed by the BHSMC controller and the RBF neural network
is a useful tool for the under-actuated 3WMR system. The 3WMR nonlinear system with different environmental
operations, strong nonlinearities, and high uncertainties was implemented by neural networks.

5. CONCLUTION
The article proposed a new adaptive controller for a 3WMR based on the use of a pole-placement technique combined
with the BHSMC sliding manifold and RBF artificial neural network. The results showed that the proposed controller's
quality was significantly improved when compared to the traditional BHSMC controller. All the simulation results
confirmed that this was a highly effective control method that helped the vehicle perform well in noisy and dynamic
environments. In the future, the research team will integrate computer vision and machine learning technologies and
supplement specialized sensors to improve the autonomous vehicle's obstacle avoidance and navigation capabilities, and
adapt to other uncertainties in the real environment.
6. REFFERENCES
Amirkhani, Abdollah, Masoud Shirzadeh, Mohammad H. Shojaeefard, and Ajith Abraham (2020). "Controlling wheeled mobile robot
considering the effects of uncertainty with neuro-fuzzy cognitive map." ISA transactions 100, 454-468.
Matraji, Imad, Ahmed Al-Durra, Andri Haryono, Khaled Al-Wahedi, and Mohamed Abou-Khousa. (2018) "Trajectory tracking control
of skid-steered mobile robot based on adaptive second order sliding mode control." Control Engineering Practice 72: 167-176.
Thi Kim Duyen Ha, Manh Cuong Nguyen, Hoang Thuat Vo, Duc Dinh Nguyen, and Anh Dung Bui .(2019). "Trajectory tracking
control for four-wheeled omnidirectional mobile robot using Backstepping technique aggregated with sliding mode control." In 2019
First International Symposium on Instrumentation, Control, Artificial Intelligence, and Robotics (ICA-SYMP), pp. 131-134. IEEE.
Do Khac Tiep, Kinam Lee, Dae-Yeong Im , Bongwoo Kwak, and Young-Jae Ryoo (2018). "Design of Fuzzy-PID Controller for Path
Tracking of Mobile Robot with Differential Drive ".International Journal of Fuzzy Logic and Intelligent Systems Vol. 18, No. 3,, pp.
220-228 01
Fierro, Rafael, and Frank L. Lewis (1997). "Control of a nonholonomic mobile robot: Backstepping kinematics into dynamics." Journal
of Robotic Systems 14, no.3: 149-163
Yang, Jong-Min, and Jong-Hwan Kim (1999). "Sliding mode control for trajectory tracking of nonholonomic wheeled mobile robots."
IEEE Transactions on robotics and automation 15, no. 3: 578-587
Koubaa, Yasmine, Mohamed Boukattaya, and Tarak Dammak (2015). "Adaptive sliding-mode dynamic control for path tracking of
nonholonomic wheeled mobile robot." Journal of Automation and Systems Engineering 9, no. 2 : 119-131
Dang, Son Tung, Xuan Minh Dinh, Thai Dinh Kim, Hai Le Xuan, and Manh-Hung Ha. (2023). "Adaptive Backstepping Hierarchical
Sliding Mode Control for 3-Wheeled Mobile Robots Based on RBF Neural Networks." Electronics 12, no. 11 : 2345
Nguyen Doan Phuoc(2016), “Linear Control Theory”, Back Khoa Publishing House. (in Vietnamese)