Received September 28, 2020, accepted October 10, 2020, date of publication October 13, 2020, date of current

version October 27, 2020.

Digital Object Identifier 10.1109/ACCESS.2020.3030775

Recurrent Neural Network-Based Robust

Nonsingular Sliding Mode Control
With Input Saturation for a Non-Holonomic
Spherical Robot
SHU-BO CHEN1 , ALIREZA BEIGI2 , AMIN YOUSEFPOUR 2 , FARHAD RAJAEE3 ,
HADI JAHANSHAHI 4 , STELIOS BEKIROS5,6 , RAÚL ALCARAZ MARTÍNEZ 7 ,
AND YUMING CHU 8,9
1 School of Science, Hunan City University, Yiyang 413000, China
2 School of Mechanical Engineering, College of Engineering, University of Tehran, Tehran 1417466191, Iran
3 Department of Mechatronics Engineering, Faculty of New Sciences and Technologies, University of Tehran, Tehran 1417466191, Iran
4 Department of Mechanical Engineering, University of Manitoba, Winnipeg, MB R3T 5V6, Canada
5 Department of Economics, European University Institute, 50014 Florence, Italy
6 Rimini Centre for Economic Analysis (RCEA), Wilfrid Laurier University, Waterloo, ON N2L 3C5, Canada
7 Research Group in Electronic, Biomedical, and Telecommunication Engineering, University of Castilla-La Mancha (UCLM), 16071 Cuenca, Spain
8 Department of Mathematics, Huzhou University, Huzhou 313000, China
9 Hunan Provincial Key Laboratory of Mathematical Modeling and Analysis in Engineering, Changsha University of Science and Technology, Changsha 410114,

China
Corresponding author: Yuming Chu ([email protected])
This work was supported by the National Natural Science Foundation of China under Grant 11971142, Grant 11401192, Grant 61673169,
Grant 11701176, Grant 11626101, and Grant 11601485.

ABSTRACT We develop a new robust control scheme for a non-holonomic spherical robot. To this end,
the mathematical model of a pendulum driven non-holonomic spherical robot is first presented. Then,
a recurrent neural network-based robust nonsingular sliding mode control is proposed for stabilization
and tracking control of the system. The designed recurrent neural network is applied to approximate
compound disturbances, including external interferences and dynamic uncertainties. Moreover, the controller
is designed in a way that avoids the singularity problem in the system. Another advantage of the proposed
scheme is its ability for tracking control while there exists control input saturation, which is a serious concern
in robotic systems. Based on the Lyapunov theorem, the stability of the closed-loop system has also been
confirmed. Lastly, the performance of the proposed control technique for the uncertain system in the presence
of an external disturbance, unknown input saturation, and dynamic uncertainties has been investigated. Also,
the proposed controller has been compared with a Fuzzy-PID one. Simulation results show the effectiveness
and superiority of the developed control technique.

INDEX TERMS Spherical robot, sliding mode control, recurrent neural network, external disturbance,
unknown input saturation, control singularity.

I. INTRODUCTION However, this kind of robot results in a nonlinear system with

The pendulum-driven spherical robot can move by changing non-holonomic dynamics, which means that the dimensions
the position of its gravity centre. Indeed, the rotating of the of the state space model are more than the number of control
pendulum generates the relocation of the mass centre position inputs. This condition makes the tracking control of this robot
of the robot [1]–[3]. The spherical geometry is very suitable difficult in real applications [8], [9].
to use these robots for exploration in harsh environments, Robotic systems are well-known samples of trajectory con-
such as in the space, deserts, and earthquake ruins [4]–[7]. trollable mechanical ones. Nevertheless, their highly nonlin-
ear dynamics, as well as uncertainties, cause a challenging
The associate editor coordinating the review of this manuscript and control problem [10]–[12]. Hence, so far several studies have
approving it for publication was Zheng Chen . tried to solve the challenge of controlling a spherical robot.
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
VOLUME 8, 2020 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ 188441
 S.-B. Chen et al.: RNN-Based Robust Nonsingular SMC With Input Saturation for a Non-Holonomic Spherical Robot

To this respect, a three-step technique has been developed by disturbances [26]–[30]. Thus, in these cases it is very benefi-
Li and Canny for the control of both position and orientation cial to apply a controller that is robust to unmodeled dynamics
of a spherical robot [13]. A novel mechanism has also been and external disturbances [31]–[33]. To this end, previous
proposed by Azizi and Naderi for controlling a spherical works have proposed several disturbance-observers with dif-
robot [14]. Precisely, they have investigated the dynamical ferent control schemes for some systems [34], [35].
model of the system and its control. Andani et al. [14] have Neural networks have been presented as an appropriate
proposed a sliding mode control (SMC) and a fuzzy SMC tool for approximation of any unknown function [36], [37].
to control a spherical robot motion. They have demonstrated Thus, using this advantage, several research studies have
that the controlled system can track the desired path with applied neural networks for control purposes [38]–[40]. For
minimum tracking error. However, they have not considered instance, a neural network-based SMC has been designed
important issues, such as the control input limitation and by Guo et al. for an autonomous underwater vehicle [41].
the singularity problem. Kayacan et al. [15] have introduced However, few studies have shown that when there are unex-
another SMC with an online learning algorithm for spherical pected changes in the system, recurrent neural networks
rolling robots. In one recent study, Roozegar et al. [8] have (RNN) perform better than conventional feedforward neural
investigated the control and motion of a spherical robot on networks [42]–[45]. To this respect, a RNN-based distur-
an inclined plane. They have proposed a terminal sliding bance observer has been employed to fortify the robust-
mode control (TSMC) to maintain and control the robot on ness of the controller by Salgado and Chairez [46].
a variable slope. To approximate the uncertain dynamics of a MIMO system,
One prevalent problem in TSMC is the singularity, which Salgado et al. [47] have also developed a RNN-based
causes by some terms in the terminal sliding mode surface observer for an adaptive SMC. The authors have proven that
(this kind of singularity does not occur in SMC). To avoid this control scheme can approximate the unknown states of
this singularity problem, a saturation function has been intro- a given nonlinear system and lessen the convergence time,
duced in [16] for dealing with the singularity problem in the as well as the oscillations in the steady-state responses.
case of chained nonlinear systems with matched perturba- Fei and Lu [42] have presented an adaptive SMC by using a
tions. In this method, without changing the design of the double loop RNN to approximate unknown dynamics. Zhang
controller, it was proposed to limit the control signal when and Chu [48] have designed an adaptive SMC based on the
singularity occurs. However, in this method, the stability local RNN to estimate the uncertainties for trajectory tracking
of the closed-loop system where singularity occurs was not control of an autonomous underwater robot. Xu et al. [49]
proven. A nonsingular TSMC method was presented in [17], have proposed a RNN-based robust tracking control to mea-
which simply swaps the state variables in the conventional sure an online unknown nonlinear system function. The
TSMC function while retaining the finite time convergence authors have shown that an RNN-based robust tracking con-
feature. Also, in [18] switching between TSMC and a lin- trol is able to significantly improve the performance of the
ear hyper plane-based sliding mode was proposed. Another controller. Also, there have been a lot of research works
approach is to transfer the trajectory to a prescribed region in focusing on approximating time-variant functions using an
advance where no singularity occurs, which is the so-called RNN. For instance, Chow and Fang [45] have used RNN as an
two-phase control strategy [19]. It should be noted that these estimator to develop an algorithm that can approximate any
methods are adopting indirect approaches to avoid singu- trajectory tracking accurately. Feedback connections between
larity. In [20], an adaptive non-singular integral TSMC has layers of the recurrent neural network create sophisticated
been presented for trajectory tracking of autonomous under- dynamics that can deal with time-varying outputs and esti-
water vehicles with dynamic uncertainties and time-varying mate them. Li et al. [50] have shown the excellent ability of
external disturbances, which can eliminate the singularity continuous-time recurrent neural networks in the estimation
problem. In [21], a modified time-varying nonsingular TSMC of dynamical time-variant systems.
manifold has been proposed to avoid the singularity problem. On the other hand, since input saturation is a potential
In general, another type of singularity problem, which is problem in many practical dynamic systems and has played
the result of the dynamic of systems, may occur in control an important role in many branches of control applications
systems. In this kind of singularity, due to some terms that during the past decades, several valuable control schemes
there exist in the functions of the system, the singularity for uncertain nonlinear systems have been proposed up to
will happen. This issue has been solved for TSMC in some now [51]–[53]. By employing the idea of auxiliary system
studies [22]–[25]. However, there are a few studies that have design, Esfandiari et al. [53] have introduced nonsymmetric
solved this kind of singularity through SMC. Actually, most input saturation constraints for a class of uncertain non-
studies in this field have considered nonsingular approaches affine nonlinear systems with external disturbances. In [52],
for TSMC, and solving this detrimental problem through an adaptive backstepping approach has been introduced to
SMC is neglected. control a single-input uncertain nonlinear system in the pres-
In most real-world applications, it is rare to found ence of external disturbances and input saturation. In that
accurate information about the dynamics of the systems, method, a Nussbaum function is employed to solve the prob-
and moreover, they are often in the presence of various lem of the saturation nonlinearity. By making use of the

188442 VOLUME 8, 2020

S.-B. Chen et al.: RNN-Based Robust Nonsingular SMC With Input Saturation for a Non-Holonomic Spherical Robot

smooth nonlinear function of the control input signal, a non- II. MATHEMATICAL FORMULATION OF THE SYSTEM
affine pure-feedback stochastic nonlinear system has been A spherical robot is an active system which is led to a
investigated in [51]. More precisely, the proposed control desired position and orientation by moving the pendulum.
guaranteed convergence of the tracking error to an arbitrarily By adjusting the center of mass gravity, the motion of the
small neighborhood around the origin in the sense of mean robot could be controlled. A schematic representation of the
quartic value. system is depicted in Fig. 1, where points A, G, C, and
To the best of the author’s knowledge, very few attempts P represent the center of the shell, the mass center of the
have been made to design a neural network-based controller robot, the contact point between the plane and robot, and
for spherical robots [54], [55]. Moreover, none of these the position of the pendulum, respectively. In what follows,
works have considered control input saturation in the system. î, ĵ, and k̂, respectively, denote the unit vectors in x, y, and
However, as it is evident, because of the current limitations z directions.
in real actuators, the bounds of control input should be con-
sidered in real-world systems [56]–[58]. Similarly, although
the singularity problem can be induced to a large control
input [59]–[62], this has not been taken to account in most
previous studies on spherical robots. Moreover, the advan-
tages presented by RNNs have still not been completely
exploited for the control of this kind of robots. Hence, in the
present work, a novel controller has been designed for an
uncertain non-holonomic spherical robot in the presence of
unknown disturbances, control singularity problem, and con-
trol input saturation. Precisely, an RNN has been combined FIGURE 1. A spherical robot on an inclined surface.
with an SMC. Moreover, it has been demonstrated that the
proposed RNN-based disturbance observer can identify time- The Euler–Lagrange equations for the system are given by
varying disturbances and uncertainties when the robot is on a
d ∂L ∂L
   
variable slope inclined plane. The control input saturation has − =Q (1)
also been taken into account for evaluating the performance dt ∂ q̇ ∂q
of the robot in a practical, real-world scenario. Moreover, where q illustrate the generalized coordinates, Q indicates the
the proposed technique has been able to avoid the singularity generalized external forces and L is the Lagrangian function
problem in the spherical robot. The stability of the system has of the system and is given by
been proven by the Lyapunov stability theory and the Taylor
expansions technique, even when control input limitations L = T − U, (2)
were considered. where T indicates the total kinetic energy of the system and
Overall, the improvements reached by the proposed con- U denotes the potential energy. The total kinetic energy can
trol method, regarding other previous works on control of be expanded as
spherical robots, can be summarized as follows:
T = Tcase + Tpendulum , (3)
(1) Ability to deal with the control input saturation and
singularity problem through SMC, simultaneously. where Tcase and Tpendulum are given by
(2) Access to the estimated disturbance and uncertainty 1 1
by depicting the online assessed value of the overall Tcase = ms kVA k2 + IA kωs k2 , and (4)
2 2
disturbance. 1 2
(3) High-rate of accuracy in the disturbance estimation Tpendulum = mp Vp , (5)
2
using the outputs of each step as the inputs of the next where ms is the mass of the spherical shell, VA denotes the
one in the RNN. velocity of shell center,
IA is the spherical shell moment of
(4) Utilizing both activation function and biases in the inertia and ωs = α̇ − φ̇ k̂ represents the angular velocity of
neural network disturbance observer by taking to the spherical shell, where φ is the rotation of the spherical
account the stability constraints. shell relative to the inclined plane, and α is the angle of
The remainder of the work is organized as follows. the inclined plane. Also, mp and Vp indicate the mass of the
Section 2 details non-holonomic spherical robot formu- pendulum and velocity of the pendulum, respectively.
lation. In Section 3, the RNN-based nonsingular SMC According to the system that is shown in Fig. 1 the veloci-
(RNN-based NSMC) is designed for the uncertain spherical ties are as
robot. In Section 4, the proposed scheme is applied to control VC = x α̇ ĵ, VA/C = ρ φ̇ î → VEA = VC + VA/C
the motion of the system. Also, the performance of the pro-
posed controller is compared with a Fuzzy-PID one. Lastly, = ρ φ̇ î + x α̇ ĵ, and (6)
conclusions are presented in Section 5. VP = VA + VP/A , V P/A = r θ̇(cos (θ) î + sin (θ) ĵ), (7)
E

VOLUME 8, 2020 188443

 S.-B. Chen et al.: RNN-Based Robust Nonsingular SMC With Input Saturation for a Non-Holonomic Spherical Robot

where θ stands for the instantaneous angle of the pendulum, Eqs. (13)-(15) and assuming that the α̇ and α̈ are negligible,
relative to the inclined plane. we have
ρ is radius of the spherical shell and r denotes the radius
of the pendulum. The potential energy and external force can ẋ1 = x2 , (16)
be then expressed as 1   
ẋ2 = −   −mp ρrsin (z1 ) z22
Ic + mp ρ 2 sin2 z1
U = Mg (x sin (α)+ρ cos (α)−rG cos (α+θ)) , (8)
cos (z1 )

+ Mgρsin(α) − ρ [MgrG sin (α + z1 )]
and r
cos(z1 )
h i
τm 1 − ρ r
Q = τm , (9) + (17)
Ic + mp ρ 2 sin2 z1


where rG denotes the radius of the robot mass center and τm ż1 = z2 , (18)
is the motor torque. If L is recomputed by Eq. (3) to Eq. (7), mp ρrcos (z1 ) h  
these equations will be obtained as ż2 =   −mp ρrsin (z1 ) z21
    Ic mp r + mp ρ r sin z1
2 2 2 2 2

Ic + mp ρ 2 φ̈ + mp ρr cos (θ)θ̈ − mp ρrsin (θ) θ̇ 2 + θ̇ α̇

Ic + mp ρ 2


+ Mgρ sin (α)]) +  
− IA α̈ − M ρ (x0 + ρφ) α̇ 2 + Mgρsin (α) = τm , (10) Ic mp r 2 + m2p ρ 2 r 2 sin2 z1
mp r 2 θ̈ + mp ρr cos (θ)φ̈ + mp ρrsin (θ) α̇ φ̇ Ic + mp ρ 2 −mp ρr cos(z1 )
 
×[−MgrG sin(α+z1 )]+   τm
+ mp r (x0 +ρφ) sin (θ) α̈ + MgrG sin (α+θ) = τm (11) Ic mp r 2 +m2p ρ 2 r 2 sin2 z1
Respectively. By defining the states of the system as (19)
φ = x1 , φ̇ = x2 , θ = z1 , θ̇ = z2 , the state space equations of
the system are This model contains enormous complexity. Thereby,
reaching a favorable controller is the most challenging prob-
ẋ1 = x2 , (12) lem in this system, which motivates the rest of the present
1    study.
ẋ2 = −   −mp ρrsin (z1 ) z22 + z2 α̇
Ic + mp ρ 2 sin2 z1
III. CONTROLLER DESIGN
− IA α̈ − M ρ (x0 + ρx1 ) α̇ 2 + Mgρsin(α) In the current section, an RNN-based NSMC is designed for a
cos (z1 )   non-holonomic spherical robot, and the tracking convergence
+ρ − mp ρrsin (z1 ) α̇x2 + mp r (x0 + ρx1 ) of the closed-loop system is proven. Suppose ū = τm , we have
r
the following nonlinear system:
+ sin (z1 ) α̈ + MgrG sin (α + z1 )


h
1)
i ẋ1 = x2 , (20)
τm 1 − ρ cos(z
r
ẋ2 = fxz + 1f xz + gxz + 1gxz ū + d(t),


+ (13) (21-a)
Ic + mp ρ sin2 z1


2
1   
fxz = −   −mp ρrsin (z1 ) z22
ż1 = z2 , (14) Ic + mp ρ sin z1
2 2

mp ρrcos (z1 ) cos (z1 )


ż2 =   + Mgρsin(α) − ρ [MgrG sin (α + z1 )]
Ic mp r 2 + m2p ρ 2 r 2 sin2 z1 r
h   (21-b)
cos(z1 )
h i
× Mgρsin (α) − mp ρrsin (z1 ) z21 + z2 α̇ − IA α̈ 1−ρ r
gxz =  (21-c)
Ic + mp ρ 2 sin2 z1


Ic + mp ρ 2
i

− M ρ (x0 + ρx2 ) α̇ +  2 
Ic mp r 2 + m2p ρ 2 r 2 sin2 z1 where d (t) denotates the external disturbances. These equa-
tions include uncertainties and disturbances against whom the
× −mp ρrsin (z1 ) α̇x2 − mp r (x0 + ρx2 ) sin (z1 ) α̈

system must be robust. According to the uncertain terms 1gxz
Ic + mp ρ 2 − mp ρrcos (z1 ) and 1f xz , Eq. (21) can be rewritten as
 
− MgrG sin (α+z1 )]+   τm
Ic mp r 2 + m2p ρ 2 r 2 sin2 z1 ẋ2 = fxz + (gxz ) ū + D (t) , (22)
(15)
where D(t) is defined as
Eqs. (12)-(15) have been derived from the Lagrange equa-
tion and the total kinetic energy of the system. Combining D (t) = 1f xz + 1gxz ū + d (t) . (23)

188444 VOLUME 8, 2020

S.-B. Chen et al.: RNN-Based Robust Nonsingular SMC With Input Saturation for a Non-Holonomic Spherical Robot

A. SELECTING AN RNN-BASED CONTROLLER Adaptive laws for updating the bias and weights of the
Up to now, neural networks have been applied for enormous RNN are designed as
applications and have shown successful results in several 
c1 T ζ + b

b1 = −0S2 f˙ W
ḃ b1 , and (25)
problems. In [63], it is demonstrated that neural networks  
in conjunction with recursive least squares can be used T
c1 ζ + b
ċ = −S f˙ W
W 1 2 b1 ζ, (26)
effectively for model identification of nonlinear time-variant
processes. Actually, if we combine a feedforward neural respectively, where S2 is the sliding surface that will be deter-
networks with a recursive method, then it can estimate mined as described below. These updating rules with utilizing
time-varying functions as well. the backpropagation method update the weights and bias of
In [64], it is proven that Multilayer feedforward networks the recurrent neural network. The optimal value of compound
are universal approximators and are capable of approximating disturbance which observer can estimate is modeled as
any smooth function. However, feedforward networks need
 
Dt = 0f W1T ζ + b1 , (27)
appropriate inputs to approximate any time-varying func-
tions. Also, In [65], it is proven that RNNs are universal where the input vector ζ includes S1 , t · S 2 , X1 , X2 . The RNN
approximators even when they have only one layer. operation is shown in Fig. 2(a), whereas Fig.2 (b) shows the
An RNN is a powerful neural network that could be procedure of the proposed controller. The RNN disturbance
used for predicting complex uncertainties. In comparison observer has been combined with a SMC to compensate the
with conventional feedforward neural networks, RNNs have effects of control singularity, input saturation, and external
better performance when changes in the system are unex- disturbance.
pected [42], [66]. In addition, time-sequential can be stored Assumption 4: There exist ideal vector of weights and bias
through the recurrent weights of the network, and recurrent of the RNN such that |εn | < εm with constant εm > 0 for
neurons can then reflect time sequences. Therefore, an RNN all ζ .
could estimate disturbances better than a conventional Assumption 5: The activation function f (∗) is bounded.
feedforward techniques [48], [67]. By virtue of Assumption 4, it can be concluded that b D is also
Especially, RNN, which possesses recursive features, bounded.
is highly recommended for time-variant problems. Indeed, The manifolds of the sliding surface are designed as
due to the recurrent information which has been indirectly
S1 = e = x1 − x1d , and (28)
stored in a neural cell, RNN provides short-term dependen-
cies that create the capability of processing and learning S2 = ė + αc e = S˙1 + αc S1 , (29)
time-varying smooth functions [50], [68], [69]. Hence, RNN where e indicates the error of the system, which can be
is more applicable to the estimation of the time-evolving measured by a sensor. Moreover, αc is a positive parameter
condition [70]–[72]. that should be designed. The first-time derivative S2 is then

B. DESIGNING AN RNN-BASED NSMC Ṡ2 = ẋ2 − ẋ2d + αc (x2 − x2d ) . (30)

Consider ζ as an input vector, b
D as the estimated disturbance, Due to the limitation of the control force, the input
and 0 as a constant parameter which should be larger than the saturation function should be considered in practical, real-
value of the compound disturbance, i.e., than all uncertainties world applications. In the present study, the amount of the
and perturbations, we have controller input has been restricted by umax and umin . Hence,
  the restricted control input (ū) is considered as [74]:
D = 0f W c1 T ζ + b
b1 + εn , (24) 
umax if u > umax
b

ū = u if umax ≥ u ≥ umin (31)
where W c1 = [$1 , $2 , . . . , $n ]T is the vector of weights
umin if u < umin


of the RNN, b b1 denotes bias, εn is the estimation error, and
f (X ) is an activation function [56]–[58]. where umax and umin are the bounds for the input signal and
In the current study, we have used SoftSign as an acti- u is the control signal, which is designed as
vation function due to the fact that SoftSign is smoother gxz
u= 2 u0 , (32)
than tanh and sigmoid, and approaches its saturation regime (g xz + ε)
much slower. Consequently, compared with sigmod and tanh,
where u0 will be described below, and ε is a positive
SoftSign prevents the chattering in the estimation process
parameter. Based on Eq. (21-c) when cos (z1 ) = ρr , the
and is less likely to oscillate between minimum and maxi-
singularity problem occurs, and in this study, thanks to
mum bounds [73]. Even though tanh and softsign functions
the denominator of Eq. (32), we will rid of the singularity
are closely related, tanh converges exponentially, whereas
problem. A simple calculation yields
softsign converges polynomially. Softsign  functionsproduce
outputs in scale of [−1, +1]; hence, 0f W c1 T ζ + b
b1 will be g2xz ε
= 1− 2 . (33)
in [−0, 0]. (g xz + ε)
2 (g xz + ε)

VOLUME 8, 2020 188445

 S.-B. Chen et al.: RNN-Based Robust Nonsingular SMC With Input Saturation for a Non-Holonomic Spherical Robot

FIGURE 2. (a) Internal structure of the proposed RNN-based NSMC disturbance observer. (b) Global block diagram of the proposed RNN-based NSMC
for a non-holonomic spherical robot.

Substituting Eq. (33) into Eq. (22), Eq. (34) can be obtained in which Actually, the uncertainties and the disturbances are
as assumed to satisfy |Dt | < 0. Considering Eq. (35), then,
Eq. (34) can be written as
ẋ2 = fxz + (gxz ) (u + 1ū) + D(t)
ε
= fxz + u0 + gxz 1ū + D(t)− u0 , (34) ẋ2 = fxz + u0 + Dt . (36)
(g2 xz + ε)

where 1ū = ū − u and its value is unknown. By considering Finally, by considering the singularity problem and control
impacts of the nonsymmetric input limitation, the compound input saturation, the RNN-based NSMC is designed as
disturbance can be expressed as
ε u0 = −αc (x1 − x1d ) − δS2 − ψsign(S 2 ) − fxz − b
D + ẋ2d
Dt = gxz 1ū + D (t) − 2 u0 , (35)
(g xz + ε) − αc (x2 − x2d ) , (37)

188446 VOLUME 8, 2020

S.-B. Chen et al.: RNN-Based Robust Nonsingular SMC With Input Saturation for a Non-Holonomic Spherical Robot

where δ and ψ are positive parameters and ψ should be a Finally, considering the updating rules described in
large constant to fulfill ψ > |εn | . Eqs. (25) and (26), we have
The stability of the closed-loop system based on the control   
laws described in Eqs. (37)-(38) is demonstrated with the V̇ = S2 −δS2 − ψsign(S 2 ) − 0f W c1 T ζ + b
b1 − εn
following theorem. 
c1 T ζ + b
 
c1 T ζ +b
 
+ 0f W b1 + 0 f˙ W b1 (W̃1T ζ + b˜1 )
Theorem 1 Under the proposed RNN-based NSMC,  
Eqs. (24) an (38), the uncertain spherical robot converges to − αc S12 − W̃1T 0S2 f˙ Wc1 T ζ + bb1 ζ
the desired trajectory in the presence of unexpected distur-  
− b˜1 0S2 f˙ Wc1 T ζ + b


bances, singularity, and input saturation. b1 = S2 −δS2 −ψsign(S 2 )−εn
Proof: Supposing a Lyapunov function candidate as
− αc S12 , (46)
1 1 1 1 2
V = αc S12 + S22 + W̃1T 0 W̃1 + b˜1 , (38) and having in mind that ψ > |εn |, the following inequality is
2 2 2 2
obtained
where W̃1 = W c1 − W1 and b̃1 = b b1 − b1 , considering
˜ḃ = ḃ
1 ˙˜
b1 , W1 = W
ċ , and the first-time derivative of V is
1
V̇ ≤ −δS22 − αc S12 . (47)

V̇ = αc S1 Ṡ1 + S2 Ṡ2 + W̃1T 0 W

ċ + b˜ ḃ
1 1 b1
At the present stage, the proof is completed, and the states
of the system converge to the commanded values, even when
T ċ
b1 , (39)
= αc S1 (S2 − αc S1 ) + S2 Ṡ2 + W̃1 0 W1 + b˜1ḃ there exist external disturbances, unknown input saturation,
substituting Eqs. (29)-(30) and the proposed control law and the singularity problem.
described in Eq. (37) into Eq. (39), yields
IV. NUMERICAL SIMULATIONS
V̇ = S2 Ṡ2 + αc S1 − αc S12 + W̃1T 0 W
ċ + b˜ ḃ


1 1 b1 Herein, the performance and effectiveness of the proposed
= S2 (αc S1 + ẋ2 − ẋ2d + αc (x2 − x2d )) RNN-based NSMC are demonstrated. The design parameters
of the controller were chosen as αc = ψ = 20, δ = 10, and
− α S 2 + W̃ T 0 W
c 1 1 1 b .
ċ + b˜ ḃ
1 1 (40)
the initial weights in the RNN were considered as uniform
Then, using Eq. (36), Eq. (40) can be written as follows random functions in the range (0,10). The system’s parame-
ters were supposed as mp = 0.639 kg, M = 1.139 kg, ρ =
V̇ = S2 (αc S1 + fxz + u0 + Dt − ẋ2d + αc (x2 − x2d )) 0.2 m, rG = 0.101 m, and Ic = 0.05 m4 . Also, to investigate
− αc S12 + W̃1T 0 W1 1 b1 .
ċ + b˜ ḃ (41) the robust performance of the suggested control, the
external
2
disturbance was assumed as d (t) = 5 cos t 2 rads (By
Considering Eqs. (41) and (37), Eq. (41) can be expressed 2
considering dmax = 5 rads ) and initial conditions for the
as
system were [X1 (0) , X2 (0) , Z1 (0) , Z2 (0)] = [0.1, 0, 0, 0].
V̇ = S2 −δS2 − ψsign(S 2 ) − b
D + Dt − αc S12 + W̃1T 0 W The sampling time for simulations were chosen to be 0.01

ċ
1
second.
+b b˜ ḃ (42) 1 1.

According Eqs. (24) and (27), we have: A. TRACKING CONTROL

   In this section, the performance of the RNN-based NSMC
V̇ = S2 −δS2 − ψsign(S 2 ) − 0f W c1 T ζ + b
b1 − εn on position tracking is illustrated. For this purpose, the slope
of the plane was varying by α (t) = π8 + 12 π π


 sin 40 t rad
+ 0f (W1T ζ + b1 ) − αc S12 + W̃1T 0 W1 1 b1 .
ċ + b˜ ḃ (43) and the control input was limited to the values umax =
6 N.m, and umin = −5 N.m. Then, Fig. 3 presents the
Using Taylor expansion of f W1T ζ + b1 about W c1 T ζ +


simulation results. It can be easily observed that the proposed
b1 , we have
b RNN-based NSMC can track the desired reference signal in
  
c1 T ζ + b
 
c1 T ζ + b
 the presence of disturbances, input saturation, and dynamic
f W1T ζ + b1 = f W b1 + f˙ W b1
uncertainties. The maximum absolute angle of the pendulum
× (W̃1T ζ + b˜1 ), (44) is less than 0.4 rad. Therefore, it can be verified that the angle
of the pendulum does not exceed π2 rad, which shows that
in which, f W1T ζ + b1 is linearized by employing the math-


these results are appropriate for a practical system.
ematical Taylor polynomial. By substituting Eq. (44) into On the other hand, Fig. 4 demonstrates that the RNN-based
Eq. (43) yields disturbance observer can identify disturbances and uncer-
 
c1 T ζ + b
 tainties precisely. In fact, Fig. 4(a) depicts the torque input,
V̇ = S2 −δS2 − ψsign(S 2 ) − 0f W b1 − εn
     which implies that the proposed RNN-based NSMC has been
+ 0f W c1 T ζ + b
b1 + 0 f˙ Wc1 T ζ +b
b1 (W̃1T ζ + b˜1 ) saturated, and Fig 4. (b) shows the estimated disturbance.
To show more extensively the capacity of the RNN-based
− αc S12 + W̃1T 0 W1 1 b1 .
ċ + b˜ ḃ (45) approximation and the proposed control scheme in tracking

VOLUME 8, 2020 188447

 S.-B. Chen et al.: RNN-Based Robust Nonsingular SMC With Input Saturation for a Non-Holonomic Spherical Robot

FIGURE 3. Time history of the closed-loop system for a step input signal.

FIGURE 4. (a) Time history of the control input when there are input saturation (b) Actual and estimated values of the disturbance d(t).

control, the designed controller has been used for another This algorithm was mainly selected by two reasons. On the
trajectory target. Thus, Figs. 5 and 6 depict the state of the one hand, many previous works have proven that neural
system, control signal, and the estimated disturbances for network-based disturbance estimators considerably improves
a ramp input signal. The numerical results conspicuously the performance of many controllers [75]. On the other hand,
demonstrate that, using the proposed RNN-based NSMC, Fuzzy control has been suggested as suitable for complex
the spherical robot can track the desired trajectories, even robots, whose models cannot be easily established from a
when there exist uncertainties and external disturbances. mathematical point of view [76]. Moreover, a Fuzzy-PID
controller has been recently proposed to control a spherical
robot with excellent performance [47].
B. COMPARISON BETWEEN THE PROPOSED METHOD Consequently, the method developed by
AND FUZZY-PID CONTROLLER Roozegar et al. [77] has been implemented and analyzed
To illustrate the benefits of the proposed control scheme, its in the present work. Although all details for this controller
performance has been compared with a Fuzzy-PID controller. can be found in [47], it should be noted that the gains for

188448 VOLUME 8, 2020

S.-B. Chen et al.: RNN-Based Robust Nonsingular SMC With Input Saturation for a Non-Holonomic Spherical Robot

FIGURE 5. Time history of the closed-loop system’s response for a ramp input signal.

FIGURE 6. (a) Control input with input saturation for a ramp input signal (b) Actual and estimated values of the disturbance d(t).

TABLE 1. Parameters of the Fuzzy-PID control scheme. functions are depicted in Fig. 7 and the fuzzy rule table is
listed in Table 2.
Figure 8 depicts the time history of the system with the
proposed RNN-based control scheme as well as the imple-
mented Fuzzy-PID controller used for comparison. In this
simulation, both controllers have been applied to the robot
after 0.2 seconds. As can be seen, both algorithms converge
to the desired position. Nonetheless, it is noteworthy that
the control scheme are given in Table 1. Moreover, input when we have tried to consider the same input saturation
is divided into 7 Fuzzy logic values, including zero (ZO), for the Fuzzy PID controller, we could not obtain a proper
positive small (PS), positive medium (PM), positive big (PB), result. Hence, we inevitably applied the fuzzy PID controller
negative big (NB), negative medium (NM), and negative without control saturation, which means a great drawback for
small (NS). The degree of error and its time derivative are practical, real-world applications. As can also be observed
expressed by these linguistic variables. Then, membership in Figs. 8(a) and (b), the proposed RNN-based NSMC was

VOLUME 8, 2020 188449

 S.-B. Chen et al.: RNN-Based Robust Nonsingular SMC With Input Saturation for a Non-Holonomic Spherical Robot

FIGURE 7. Membership functions for the Fuzzy-PID controller used for

comparison.

TABLE 2. Fuzzy rule table for the fuzz-PID controller used for comparison.

FIGURE 8. Comparison of the (a) angular position and (b) angular

velocity reached by a spherical robot with the proposed controller and
the Fuzzy-PID one sued for comparison.

faster than the Fuzzy-PID controller. Moreover, the simula-

tion also confirmed that the proposed RNN-based controller
was able to overcome the external disturbance and uncertain-
ties better than the Fuzzy-PID. Furthermore, the Fuzzy-PID
produced a high-frequency oscillation in the response of the
system.
On the other hand, it should also be noted that conven-
tional SMC generates chattering for uncertain systems due
to the existence of the sign function in the control input. This FIGURE 9. Comparison of the control inputs obtained by the proposed
phenomenon then causes vibration in the system, because it RNN-based controller and the Fuzzy-PID one.

takes time to converge the sliding surface to zero. However,

Fig. 8 shows that the proposed RNN-based NSMC is able The control input signals for both control schemes are
to substantially minimize this problem. Actually, one of the depicted in Fig. 9. As can be seen, in both cases the input
major issues that causes chattering in a conventional SMC is signals dropped to zero with some oscillation. Nonetheless,
the existence of the uncertainties and disturbances. However,, it can be noticed that, when the target changed, the con-
as it is shown in Fig. 4 (b) and 6 (b), the controller proposed trol input for the Fuzzy-PID controller had a big overshoot.
in the present study is able to estimate quickly and accurately Otherwise, the control input for the proposed RNN-based
the uncertainties and disturbances. Actually, our numerical controller presented a smaller overshoot and quicker conver-
results confirm that after 3 second the estimation errors for gence time. Thus, this figure corroborates that the proposed
ramp and step input respectively are less than 9% and 2% control approach needed less control effort compared to the
in which the estimation errors (d e ) is calculated in (48), as Fuzzy-PID controller. To summarize, the proposed control
shown at the bottom of the page. method had smaller oscillations, faster response, and less con-
Therefore, the vibrations in the response of the system will trol effort. Hence, the numerical simulations have confirmed
be significantly decreased in comparison with a conventional the superiority of the proposed RNN-based NSMC over the
SMC. Fuzzy-PID controller.

|ACTUAL DISTURBANCE − ESTIMATED DISTURBANCE|

de = (48)
MAX (ACTUAL DISTURBANCE)

188450 VOLUME 8, 2020

S.-B. Chen et al.: RNN-Based Robust Nonsingular SMC With Input Saturation for a Non-Holonomic Spherical Robot

V. CONCLUSION [15] E. Kayacan, E. Kayacan, H. Ramon, and W. Saeys, ‘‘Adaptive neuro-fuzzy

In the present study, an RNN-based NSMC has been designed control of a spherical rolling robot using sliding-mode-control-theory-
based online learning algorithm,’’ IEEE Trans. Cybern., vol. 43, no. 1,
to control the motion of a spherical robot. Unknown uncer- pp. 170–179, Feb. 2013.
tainties, including external disturbances and the control input [16] Y. Feng, X. Yu, and F. Han, ‘‘On nonsingular terminal sliding-mode
saturation, have been quickly and accurately approximated control of nonlinear systems,’’ Automatica, vol. 49, no. 6, pp. 1715–1722,
Jun. 2013.
by the approach, thus leading to successful control. Also, [17] Y. Feng, X. Yu, and Z. Man, ‘‘Non-singular terminal sliding mode con-
the prevention of the singularity problem has been taken trol of rigid manipulators,’’ Automatica, vol. 38, no. 12, pp. 2159–2167,
into consideration for the control scheme. In fact, using the Dec. 2002.
[18] M. Zhihong and X. Huo Yu, ‘‘Terminal sliding mode control of MIMO
Lyapunov stability theorem and employing Taylor series linear systems,’’ IEEE Trans. Circuits Syst. I, Fundam. Theory Appl.,
method, stability, and robustness of the controller against vol. 44, no. 11, pp. 1065–1070, 1997.
uncertainties, disturbances, control input saturation, and sin- [19] Y. Wu, X. Yu, and Z. Man, ‘‘Terminal sliding mode control design
for uncertain dynamic systems,’’ Syst. Control Lett., vol. 34, no. 5,
gularity have been guaranteed. Finally, the proposed con- pp. 281–287, Jul. 1998.
troller has been compared with a Fuzzy-PID control scheme. [20] L. Qiao and W. Zhang, ‘‘Adaptive non-singular integral terminal sliding
The obtained simulation results have shown that the response mode tracking control for autonomous underwater vehicles,’’ IET Control
Theory Appl., vol. 11, no. 8, pp. 1293–1306, May 2017.
of the system is smoother when the proposed RNN-based [21] J. Geng, Y. Sheng, and X. Liu, ‘‘Time-varying nonsingular terminal sliding
controller was used. Nonetheless, as a future suggestion, mode control for robot manipulators,’’ Trans. Inst. Meas. Control, vol. 36,
updating control gains of the proposed control scheme in a no. 5, pp. 604–617, Jul. 2014.
[22] A. Yousefpour and H. Jahanshahi, ‘‘Fast disturbance-observer-based robust
Fuzzy environment can improve the output response of the integral terminal sliding mode control of a hyperchaotic memristor oscil-
system. Moreover, an extension of the proposed controller lator,’’ Eur. Phys. J. Special Topics, vol. 228, no. 10, pp. 2247–2268,
could also be used for fractional-order systems. Oct. 2019.
[23] A. Vahidi-Moghaddam, A. Rajaei, and M. Ayati, ‘‘Disturbance-observer-
based fuzzy terminal sliding mode control for MIMO uncertain nonlinear
REFERENCES systems,’’ Appl. Math. Model., vol. 70, pp. 109–127, Jun. 2019.
[1] M. Yue, B. Liu, C. An, and X. Sun, ‘‘Extended state observer-based [24] A. Vahidi-Moghaddam, A. Rajaei, R. Vatankhah, and M. R. Hairi-Yazdi,
adaptive hierarchical sliding mode control for longitudinal movement of a ‘‘Terminal sliding mode control with non-symmetric input saturation
spherical robot,’’ Nonlinear Dyn., vol. 78, no. 2, pp. 1233–1244, Oct. 2014. for vibration suppression of electrostatically actuated nanobeams in the
presence of Casimir force,’’ Appl. Math. Model., vol. 60, pp. 416–434,
[2] C. Camicia, F. Conticelli, and A. Bicchi, ‘‘Nonholonomic kinematics and
Aug. 2018.
dynamics of the sphericle,’’ in Proc. IEEE/RSJ Int. Conf. Intell. Robots
Syst. (IROS), vol. 1, Oct. 2000, pp. 805–810. [25] M. Chen, Q.-X. Wu, and R.-X. Cui, ‘‘Terminal sliding mode tracking
control for a class of SISO uncertain nonlinear systems,’’ ISA Trans.,
[3] Y. Cai, Q. Zhan, and X. Xi, ‘‘Path tracking control of a spherical mobile
vol. 52, no. 2, pp. 198–206, Mar. 2013.
robot,’’ Mechanism Mach. Theory, vol. 51, pp. 58–73, May 2012.
[26] H. Jahanshahi, M. Shahriari-Kahkeshi, R. Alcaraz, X. Wang, V. Singh,
[4] R. H. Armour and J. F. V. Vincent, ‘‘Rolling in nature and robotics:
and V.-T. Pham, ‘‘Entropy analysis and neural network-based adaptive
A review,’’ J. Bionic Eng., vol. 3, no. 4, pp. 195–208, Dec. 2006.
control of a non-equilibrium four-dimensional chaotic system with hidden
[5] Q. Zhan, T. Zhou, M. Chen, and S. Cai, ‘‘Dynamic trajectory planning of a attractors,’’ Entropy, vol. 21, no. 2, p. 156, Feb. 2019.
spherical mobile robot,’’ in Proc. IEEE Conf. Robot., Autom. Mechatronics,
[27] H. Jahanshahi, A. Yousefpour, Z. Wei, R. Alcaraz, and S. Bekiros,
Dec. 2006, pp. 1–6.
‘‘A financial hyperchaotic system with coexisting attractors: Dynamic
[6] Q. Zhan, Y. Cai, and Z. Liu, ‘‘Near-optimal trajectory planning of a investigation, entropy analysis, control and synchronization,’’ Chaos, Soli-
spherical mobile robot for environment exploration,’’ in Proc. IEEE Conf. tons Fractals, vol. 126, pp. 66–77, Sep. 2019.
Robot., Autom. Mechatronics, Sep. 2008, pp. 84–89. [28] H. Jahanshahi, K. Rajagopal, A. Akgul, N. N. Sari, H. Namazi, and
[7] K. Al-Widyan, X. Q. Ma, and J. Angeles, ‘‘The robust design of parallel S. Jafari, ‘‘Complete analysis and engineering applications of a megastable
spherical robots,’’ Mechanism Mach. Theory, vol. 46, no. 3, pp. 335–343, nonlinear oscillator,’’ Int. J. Non-Linear Mech., vol. 107, pp. 126–136,
Mar. 2011. Dec. 2018.
[8] M. Roozegar, M. Ayati, and M. J. Mahjoob, ‘‘Mathematical modelling and [29] H. Jahanshahi, A. Yousefpour, J. M. Munoz-Pacheco, S. Kacar,
control of a nonholonomic spherical robot on a variable-slope inclined V.-T. Pham, and F. E. Alsaadi, ‘‘A new fractional-order hyperchaotic mem-
plane using terminal sliding mode control,’’ Nonlinear Dyn., vol. 90, no. 2, ristor oscillator: Dynamic analysis, robust adaptive synchronization, and its
pp. 971–981, Oct. 2017. application to voice encryption,’’ Appl. Math. Comput., vol. 383, Oct. 2020,
[9] V. A. Joshi, R. N. Banavar, and R. Hippalgaonkar, ‘‘Design and analysis Art. no. 125310.
of a spherical mobile robot,’’ Mechanism Mach. Theory, vol. 45, no. 2, [30] A. Vahidi-Moghaddam, M. Mazouchi, and H. Modares, ‘‘Memory-
pp. 130–136, Feb. 2010. augmented system identification with finite-time convergence,’’ IEEE
[10] D.-I. Curiac and C. Volosencu, ‘‘A 2D chaotic path planning for mobile Control Syst. Lett., vol. 5, no. 2, pp. 571–576, Apr. 2021.
robots accomplishing boundary surveillance missions in adversarial con- [31] H. Jahanshahi, ‘‘Smooth control of HIV/AIDS infection using a robust
ditions,’’ Commun. Nonlinear Sci. Numer. Simul., vol. 19, no. 10, adaptive scheme with decoupled sliding mode supervision,’’ Eur. Phys.
pp. 3617–3627, Oct. 2014. J. Special Topics, vol. 227, nos. 7–9, pp. 707–718, Oct. 2018.
[11] N. Chen, F. Song, G. Li, X. Sun, and C. Ai, ‘‘An adaptive sliding [32] H. Jahanshahi, A. Yousefpour, J. M. Munoz-Pacheco, I. Moroz, Z. Wei,
mode backstepping control for the mobile manipulator with nonholonomic and O. Castillo, ‘‘A new multi-stable fractional-order four-dimensional
constraints,’’ Commun. Nonlinear Sci. Numer. Simul., vol. 18, no. 10, system with self-excited and hidden chaotic attractors: Dynamic analysis
pp. 2885–2899, Oct. 2013. and adaptive synchronization using a novel fuzzy adaptive sliding mode
[12] K. Alipour, A. B. Robat, and B. Tarvirdizadeh, ‘‘Dynamics modeling control method,’’ Appl. Soft Comput., vol. 87, Feb. 2020, Art. no. 105943.
and sliding mode control of tractor-trailer wheeled mobile robots sub- [33] A. Yousefpour, A. Vahidi-Moghaddam, A. Rajaei, and M. Ayati, ‘‘Stabi-
ject to wheels slip,’’ Mechanism Mach. Theory, vol. 138, pp. 16–37, lization of nonlinear vibrations of carbon nanotubes using observer-based
Aug. 2019. terminal sliding mode control,’’ Trans. Inst. Meas. Control, vol. 42, no. 5,
[13] Z. Li and J. Canny, ‘‘Motion of two rigid bodies with rolling constraint,’’ pp. 1047–1058, Mar. 2020.
IEEE Trans. Robot. Autom., vol. 6, no. 1, pp. 62–72, Feb. 1990. [34] S. Wang, A. Yousefpour, A. Yusuf, H. Jahanshahi, R. Alcaraz, S. He,
[14] M. T. Andani, S. Shahmiri, H. Pourgharibshahi, K. Yousefpour, and and J. M. Munoz-Pacheco, ‘‘Synchronization of a non-equilibrium four-
M. H. Imani, ‘‘Fuzzy-based sliding mode control and sliding mode control dimensional chaotic system using a Disturbance-Observer-Based adaptive
of a spherical robot,’’ in Proc. 44th Annu. Conf. IEEE Ind. Electron. Soc. terminal sliding mode control method,’’ Entropy, vol. 22, no. 3, p. 271,
(IECON), Oct. 2018, pp. 2534–2539. Feb. 2020.

VOLUME 8, 2020 188451

 S.-B. Chen et al.: RNN-Based Robust Nonsingular SMC With Input Saturation for a Non-Holonomic Spherical Robot

[35] Y. Singh and M. Santhakumar, ‘‘Inverse dynamics and robust sliding mode [56] G. V. Lakhekar, L. M. Waghmare, P. G. Jadhav, and R. G. Roy,
control of a planar parallel (2-PRP and 1-PPR) robot augmented with ‘‘Robust diving motion control of an autonomous underwater vehicle
a nonlinear disturbance observer,’’ Mechanism Mach. Theory, vol. 92, using adaptive neuro-fuzzy sliding mode technique,’’ IEEE Access, vol. 8,
pp. 29–50, Oct. 2015. pp. 109891–109904, 2020.
[36] F. G. Kong, Q. Li, and W. J. Zhang, ‘‘An artificial neural network approach [57] L. Colombo, M. L. Corradini, G. Ippoliti, and G. Orlando, ‘‘Pitch angle
to mechanism kinematic chain isomorphism identification,’’ Mechanism control of a wind turbine operating above the rated wind speed: A sliding
Mach. Theory, vol. 34, no. 2, pp. 271–283, Feb. 1999. mode control approach,’’ ISA Trans., vol. 96, pp. 95–102, Jan. 2020.
[37] Z. Yao, J. Yao, and W. Sun, ‘‘Adaptive RISE control of hydraulic systems [58] H. Dong, X. Lin, S. Gao, B. Cai, and B. Ning, ‘‘Neural networks-based
with multilayer neural-networks,’’ IEEE Trans. Ind. Electron., vol. 66, sliding mode fault-tolerant control for high-speed trains with bounded
no. 11, pp. 8638–8647, Nov. 2019. parameters and actuator faults,’’ IEEE Trans. Veh. Technol., vol. 69, no. 2,
[38] B. Ren, Y. Wang, and J. Chen, ‘‘Trajectory-Tracking-Based adaptive neural pp. 1353–1362, Feb. 2020.
network sliding mode controller for robot manipulators,’’ J. Comput. Inf. [59] S. Zeinali and M. Shahrokhi, ‘‘Observer-based singularity free nonlinear
Sci. Eng., vol. 20, no. 3, Jun. 2020. controller for uncertain systems subject to input saturation,’’ Eur. J. Con-
[39] G. Liu, G. Li, H. Song, and Z. Peng, ‘‘Neural network sliding mode control trol, vol. 52, pp. 49–58, Mar. 2020.
for pneumatic servo system based on particle swarm optimization,’’ in [60] Y. Yang, ‘‘Singularity-free model predictive spacecraft attitude regulation
Proc. 11th Int. Conf. Modelling, Identificat. Control (ICMIC), Tianjin, using a variable-speed control moment gyroscope model,’’ IEEE Trans.
China, Jul. 2019, pp. 1239–1248. Aerosp. Electron. Syst., vol. 54, no. 3, pp. 1511–1518, Jun. 2018.
[40] P. Zhang, Z. Wu, H. Dong, M. Tan, and J. Yu, ‘‘Reaction-wheel-based [61] A. Nikou, C. Verginis, S. Heshmati-alamdari, and D. V. Dimarogonas,
roll stabilization for a robotic fish using neural network sliding mode ‘‘A nonlinear model predictive control scheme for cooperative manipula-
control,’’ IEEE/ASME Trans. Mechatronics, vol. 25, no. 4, pp. 1904–1911, tion with singularity and collision avoidance,’’ in Proc. 25th Medit. Conf.
Aug. 2020. Control Autom. (MED), Jul. 2017, pp. 707–712.
[41] X. Guo, W. Yan, and R. Cui, ‘‘Neural network-based nonlinear sliding- [62] D.-D. Zheng, Y. Pan, K. Guo, and H. Yu, ‘‘Identification and control of
mode control for an AUV without velocity measurements,’’ Int. J. Control, nonlinear systems using neural networks: A singularity-free approach,’’
vol. 92, no. 3, pp. 677–692, Mar. 2019. IEEE Trans. Neural Netw. Learn. Syst., vol. 30, no. 9, pp. 2696–2706,
[42] J. Fei and C. Lu, ‘‘Adaptive sliding mode control of dynamic systems using Sep. 2019.
double loop recurrent neural network structure,’’ IEEE Trans. Neural Netw. [63] M. Nikravesh, A. E. Farell, and T. G. Stanford, ‘‘Model identification of
Learn. Syst., vol. 29, no. 4, pp. 1275–1286, Apr. 2018. nonlinear time variant processes via artificial neural network,’’ Comput.
Chem. Eng., vol. 20, no. 11, pp. 1277–1290, Jan. 1996.
[43] C.-H. Lee and C.-C. Teng, ‘‘Identification and control of dynamic systems
[64] K. Hornik, M. Stinchcombe, and H. White, ‘‘Multilayer feedforward
using recurrent fuzzy neural networks,’’ IEEE Trans. Fuzzy Syst., vol. 8,
networks are universal approximators,’’ Neural Netw., vol. 2, no. 5,
no. 4, pp. 349–366, Aug. 2000.
pp. 359–366, Jan. 1989.
[44] C.-M. Lin and C.-F. Hsu, ‘‘Supervisory recurrent fuzzy neural network
[65] A. M. Schäfer and H. G. Zimmermann, ‘‘Recurrent neural networks are
control of wing rock for slender delta wings,’’ IEEE Trans. Fuzzy Syst.,
universal approximators,’’ in Proc. Int. Conf. Artif. Neural Netw. Springer,
vol. 12, no. 5, pp. 733–742, Oct. 2004.
2006, pp. 632–640.
[45] T. W. S. Chow and Y. Fang, ‘‘A recurrent neural-network-based real-time
[66] H. Ding, Y. F. Li, and S. K. Tso, ‘‘Dynamic optimization of redundant
learning control strategy applying to nonlinear systems with unknown
manipulators in worst case using recurrent neural networks,’’ Mechanism
dynamics,’’ IEEE Trans. Ind. Electron., vol. 45, no. 1, pp. 151–161,
Mach. Theory, vol. 35, no. 1, pp. 55–70, Jan. 2000.
Feb. 1998.
[67] A. Farzad, H. Mashayekhi, and H. Hassanpour, ‘‘A comparative perfor-
[46] I. Salgado and I. Chairez, ‘‘Discrete time recurrent neural net- mance analysis of different activation functions in LSTM networks for
work observer,’’ in Proc. Int. Joint Conf. Neural Netw., Jun. 2009, classification,’’ Neural Comput. Appl., vol. 31, no. 7, pp. 2507–2521,
pp. 2764–2770. Jul. 2019.
[47] I. Salgado, H. Ahmed, O. Camacho, and I. Chairez, ‘‘Adaptive sliding- [68] T. Mikolov, A. Joulin, S. Chopra, M. Mathieu, and M. Ranzato, ‘‘Learning
mode observer for second order discrete-time MIMO nonlinear systems longer memory in recurrent neural networks,’’ 2014, arXiv:1412.7753.
based on recurrent neural-networks,’’ Int. J. Mach. Learn. Cybern., vol. 10, [Online]. Available: http://arxiv.org/abs/1412.7753
pp. 2851–2866, Oct. 2019. [69] B. Šter, ‘‘Selective recurrent neural network,’’ Neural Process. Lett.,
[48] M.-J. Zhang and Z.-Z. Chu, ‘‘Adaptive sliding mode control based on local vol. 38, no. 1, pp. 1–15, Aug. 2013.
recurrent neural networks for underwater robot,’’ Ocean Eng., vol. 45, [70] L. Jin, J. Yan, X. Du, X. Xiao, and D. Fu, ‘‘RNN for solving time-
pp. 56–62, May 2012. variant generalized Sylvester equation with applications to robots and
[49] Z. Xu, Q. Song, and D. Wang, ‘‘Recurrent neural tracking control based on acoustic source localization,’’ IEEE Trans. Ind. Informat., vol. 16, no. 10,
multivariable robust adaptive gradient-descent training algorithm,’’ Neural pp. 6359–6369, Oct. 2020.
Comput. Appl., vol. 21, no. 7, pp. 1745–1755, Oct. 2012. [71] L. Cheng, Z.-G. Hou, Y. Lin, M. Tan, W. C. Zhang, and F.-X. Wu,
[50] X.-D. Li, J. K. L. Ho, and T. W. S. Chow, ‘‘Approximation of dynami- ‘‘Recurrent neural network for non-smooth convex optimization problems
cal time-variant systems by continuous-time recurrent neural networks,’’ with application to the identification of genetic regulatory networks,’’ IEEE
IEEE Trans. Circuits Syst. II, Exp. Briefs, vol. 52, no. 10, pp. 656–660, Trans. Neural Netw., vol. 22, no. 5, pp. 714–726, May 2011.
Oct. 2005. [72] L. Jin, S. Li, and B. Hu, ‘‘RNN models for dynamic matrix inversion:
[51] H. Wang, B. Chen, X. Liu, K. Liu, and C. Lin, ‘‘Robust adaptive fuzzy A control-theoretical perspective,’’ IEEE Trans. Ind. Informat., vol. 14,
tracking control for pure-feedback stochastic nonlinear systems with no. 1, pp. 189–199, Jan. 2018.
input constraints,’’ IEEE Trans. Cybern., vol. 43, no. 6, pp. 2093–2104, [73] X. Glorot and Y. Bengio, ‘‘Understanding the difficulty of training deep
Dec. 2013. feedforward neural networks,’’ in Proc. 13th Int. Conf. Artif. Intell. Statist.,
[52] C. Wen, J. Zhou, Z. Liu, and H. Su, ‘‘Robust adaptive control of uncer- 2010, pp. 249–256.
tain nonlinear systems in the presence of input saturation and external [74] A. Yousefpour, A. H. Hosseinloo, M. R. H. Yazdi, and A. Bahrami,
disturbance,’’ IEEE Trans. Autom. Control, vol. 56, no. 7, pp. 1672–1678, ‘‘Disturbance observer-based terminal sliding mode control for effective
Jul. 2011. performance of a nonlinear vibration energy harvester,’’ J. Intell. Mater.
[53] K. Esfandiari, F. Abdollahi, and H. A. Talebi, ‘‘Adaptive control of Syst. Struct., vol. 31, no. 12, 2020, Art. no. 1045389X20922903.
uncertain nonaffine nonlinear systems with input saturation using neu- [75] A. Yousefpour, H. Jahanshahi, J. M. Munoz-Pacheco, S. Bekiros, and
ral networks,’’ IEEE Trans. Neural Netw. Learn. Syst., vol. 26, no. 10, Z. Wei, ‘‘A fractional-order hyper-chaotic economic system with transient
pp. 2311–2322, Oct. 2015. chaos,’’ Chaos, Solitons Fractals, vol. 130, Jan. 2020, Art. no. 109400.
[54] C. Xia, C. Guo, and T. Shi, ‘‘A neural-network-identifier and fuzzy- [76] J. Song, ‘‘Design and comparison of AFS controllers with PID, fuzzy-
controller-based algorithm for dynamic decoupling control of permanent- logic, and sliding-mode controllers,’’ Adv. Mech. Eng., vol. 5, Jan. 2013,
magnet spherical motor,’’ IEEE Trans. Ind. Electron., vol. 57, no. 8, Art. no. 401548.
pp. 2868–2878, Aug. 2010. [77] M. Roozegar and M. J. Mahjoob, ‘‘Modelling and control of a non-
[55] Y. Cai, Q. Zhan, and X. Xi, ‘‘Neural network control for the linear motion holonomic pendulum-driven spherical robot moving on an inclined plane:
of a spherical mobile robot,’’ Int. J. Adv. Robot. Syst., vol. 8, no. 4, p. 32, Simulation and experimental results,’’ IET Control Theory Appl., vol. 11,
Sep. 2011. no. 4, pp. 541–549, Feb. 2017.

188452 VOLUME 8, 2020

S.-B. Chen et al.: RNN-Based Robust Nonsingular SMC With Input Saturation for a Non-Holonomic Spherical Robot

SHU-BO CHEN was born in Shaoyang, Hunan, STELIOS BEKIROS received the Diploma, B.Sc.,
China, in August 1980. He received the B.S. and M.Eng. degrees in electrical and computer
and M.S. degrees from Hunan Normal University, engineering from the National Technical Univer-
Changsha, China, in 2004 and 2007, respectively, sity of Athens (NTUA), and the M.Sc. degree in
and the Ph.D. degree from Central South Uni- decision sciences (statistics, operations research,
versity, Changsha, in 2012. He has worked as an and information technology) and the Ph.D. degree
Assistant Professor and an Associate Professor in econometrics and finance from the Athens
with the Department of Mathematics, Hunan City University of Economics and Business (AUEB).
University, Yiyang, China, from 2009 to 2012 and He was a Faculty Member/Research Fellow with
2012 to 2017, respectively, where he has been a the European Institute for Statistics, Probability,
Professor and the Dean, since 2017. His current research interests include Stochastic Operations Research and its Applications (EURANDOM), Tech-
combinatorical inequalities theory and applications, functional analysis, and nical University of Eindhoven; the Center for Nonlinear Dynamics in Eco-
numerical analysis. nomics and Finance (CeNDEF), Faculty of Economics and Econometrics,
University of Amsterdam; the Department of Economics, European Uni-
versity Institute (EUI), Florence, Italy; the Hellenic Ministry of Finance;
the Centre of Planning and Economic Research (KEPE); the Department of
ALIREZA BEIGI received the B.S. degree in Accounting and Finance, AUEB; the IPAG Business School, Paris, France;
mechanical engineering from the University of and the Decision Support Systems Laboratory, National Technical Univer-
Tehran, in 2019. He is currently a Research Assis- sity of Athens. He serves as an Expert Evaluator of the Natural Sciences
tant with the Advanced Instrumentation Labora- and Engineering Research Council of Canada (NSERC) and NSF, USA,
tory, University of Tehran. His research interests an Expert Evaluator/Vice-Chair of the European Commission (EU-REA
include nonlinear control theory and its application panels, Horizon 2020), an Expert Evaluator of the Swiss National Science
to engineering problems, optimization, reinforce- Foundation (SNF), a WES Expert of the Ifo Institute (Leibniz Institute for
ment learning, and artificial intelligence. Economic Research, University of Munich), and an Expert Project Evaluator
of ‘‘la Caixa Barcelona Institute’’ Banking Foundation while he is a regular
Evaluator of Hellenic national research bodies. His work, with presently
over 160 articles, has been published in the most highly esteemed academic
journals and in books and proceedings in many scientific fields. His research
AMIN YOUSEFPOUR received the B.Sc. degree interests include among other chaos and econophysics, machine learning,
in mechanical engineering from Yasouj University, artificial intelligence, time series and spectral econometrics, forecasting,
Iran, in 2015, and the M.S. degree in mechanical quantitative finance, extreme and Bayesian statistics, complex systems,
engineering from the University of Tehran, Iran, financial engineering, macro-finance, OR, behavioral finance, big data, and
in 2018. Since 2018, he has been a Research Assis- FinTech. He is the Editor, a Co-Editor, and an Associate Editor of 17 pres-
tant with the Advanced Instrumentation Labora- tigious journals of the major publishing houses, namely Elsevier, Wiley,
tory, University of Tehran. Also, up to now, he has Springer, Taylor & Francis, and Springer-Nature.
done various projects in international collabo-
ration with prestigious universities and research
institutes. In general, his research passion is to RAÚL ALCARAZ MARTÍNEZ received the M.Sc.
solve complex real-world problems in dynamics and control by developing degree in electronic engineering from the Uni-
novel techniques using insights from numerous different disciplines of versity of Alcalá, Madrid, Spain, in 2005, and
science and engineering. His research activities began with classical control the Ph.D. degree in biomedical engineering
and vibration analysis and have transitioned towards more unconventional from the Universidad Politécnica de Valencia,
and high-tech approaches in robotics, control techniques, reinforcement Valencia, Spain, in 2008. Since 2006, he has
learning, and dynamical systems. been a Lecturer with the Department of
Electrical, Electronic, Automatic, and Commu-
nication Engineering, University of Castilla-La
Mancha, Spain. He has taught several subjects
FARHAD RAJAEE received the B.Sc. degree in related to electronic and biomedical instrumentation, analog and digital
electronic engineering from the Urmia University electronics, and biomedical signal processing; and has been the author of
of Technology, Iran, in 2015, and the M.S. degree several academic publications in these areas. His research interests include
in mechatronic engineering from the University of statistical, nonlinear, and array signal processing applied to biomedical signal
Tehran, Iran, in 2018. Since 2018, he has been and the development of new medical equipment.
a Research Assistant in advanced service robots
with the University of Tehran. His research inter-
ests include control, robotic, biomechanics, neural YUMING CHU was born in Huzhou, Zhejiang,
networks, and deep learning. China, in June 1966. He received the B.S. degree
from Hangzhou Normal University, Hangzhou,
China, in 1988, and the M.S. and Ph.D.
degrees from Hunan University, Changsha, China,
in 1991 and 1994, respectively. He has worked as
HADI JAHANSHAHI received the B.Sc. degree an Assistant Professor and an Associate Profes-
in mechanical engineering and the M.Sc. degree sor with the Department of Mathematics, Hunan
in aerospace engineering from the University of Normal University, Changsha, from 1994 to
Tehran. He has worked as a Researcher with the 1996 and 1997 to 2002, respectively. Since 2002,
University of Tehran and a Technician with the he has been a Professor and the Dean of the Department of Mathematics,
University of Manitoba. His main research inter- Huzhou University, Huzhou. His current research interests include special
ests include nonlinear control, nonlinear dynam- functions, functional analysis, numerical analysis, operator theory, ordinary
ics, robotics, fractional calculus, and optimization. differential equations, partial differential equations, inequalities theory and
He was a recipient of several Ph.D. Scholarships applications, and robust filtering and control.
from different universities.

VOLUME 8, 2020 188453