Neurocomputing 458 (2021) 70–86

Contents lists available at ScienceDirect

Neurocomputing
journal homepage: www.elsevier.com/locate/neucom

Adaptive neural backstepping control for flexible-joint robot

manipulator with bounded torque inputs
Xin Cheng a, Yajun Zhang a, Huashan Liu a,b,⇑, Dirk Wollherr c, Martin Buss c
a
College of Information Science and Technology, Donghua University, Shanghai 201620, PR China
b
Engineering Research Center of Digitized Textile & Apparel Technology, Ministry of Education, Shanghai 201620, China
c
Institute of Automatic Control Engineering (LSR), Technische Universität München, München 80333, Germany

a r t i c l e i n f o a b s t r a c t

Article history: Aiming at tracking control with bounded torque inputs of the flexible-joint robot manipulators, we pro-
Received 26 September 2020 pose a generalized saturated adaptive controller based on backstepping control, singular perturbation
Revised 17 May 2021 decoupling and neural networks. First, by using the singular perturbation theory, the full-order rigid-
Accepted 6 June 2021
flexible dynamics of the robot manipulator is decoupled into a slow subsystem and a fast subsystem.
Available online 8 June 2021
Communicated by Zidong Wang
Second, saturated sub-controller by backstepping method is proposed for the slow subsystem, where
the projection-type parameter adaptation and a class of saturation functions are applied to make the tor-
que inputs bounded, and a saturated neural network approximator is involved to simplify the control law
Keywords:
Flexible-joint robot
and to compensate for the uncertain nonlinearity. Third, for fast subsystem, a new filtered tracking error
Bounded input of the elastic torque is used in the fast control law to make the boundary layer subside quickly. In addi-
Singular perturbation tion, explicit but strict stability analysis is given for the system. Finally, comparisons indicate that the
Backstepping control proposed controller results in a more satisfactory tracking performance with keeping the control inputs
Saturated neural network approximator bounded within the given range all the time and superior anti-disturbance capability.
Ó 2021 Elsevier B.V. All rights reserved.

1. Introduction For rigid robot manipulators with exact dynamics, a saturated

output feedback tracking controller is proposed in [2], where a
In recent years, singularly perturbed modeling and control of hyperbolic tangent function in the controller accounts for the sat-
flexible-joint robot manipulators have been extensively studied uration of the torque inputs to the joint actuators. To accommodate
since it was first proved to be feasible in [1]. The control system for the electric dynamics of the joint motors, a saturated propor-
designed via such theoretical strategy can be easily achieved under tional and derivative (PD) controller plus linear feedback of electric
ideal conditions with exactly known system parameters and current that suffices to ensure global asymptotic stability of the
unconstrained torque inputs, as well as accurate measurements rigid robots is proposed in [3]. In [4,10,11], several schemes of sat-
of all the state feedback needed. But the practical situations are urated proportional, integral and derivative (PID) type are pre-
not always so ideal that there usually exist unknown or uncertain sented, where the controllers in [4,10] are proved to be globally
nonlinearities (like modeling error, internal/external disturbances, asymptotically stable, the one in [11] investigates the global con-
etc.) and power output limitations of the joint actuators. tinuous control for the finite-time or (local) exponential stabiliza-
Motors with small capacity are always preferred as the joint tion of mechanical systems with bounded inputs involving the
actuators to make the mechanical structure compact enough in desired conservative-force compensation. An improved saturated
the design of the robot manipulators, thus this leads to a fact that controller with error-gain-contained arctangent function is
the joint actuators cannot offer arbitrary large torque required by designed in [5], and a pseudo signal is generated from a linear filter
the unbounded control law proposed in most previous controllers. to substitute the actual velocity error, to eliminate the velocity
In the last two decades, some works have taken into account of this measurements of joint actuators. In addition, an upgraded satu-
saturation problem [2–13]. rated controller with anti-external disturbance capability is pro-
posed in [6] via fuzzy self-tuning PD gains. From a different
perspective, an auxiliary system is designed in [13] to handle the
actuator saturation problem. Note that, the saturated controllers
⇑ Corresponding author. usually involve dynamics parameters in the compensation term,
E-mail address: [email protected] (H. Liu).

https://doi.org/10.1016/j.neucom.2021.06.013
0925-2312/Ó 2021 Elsevier B.V. All rights reserved.
X. Cheng, Y. Zhang, H. Liu et al. Neurocomputing 458 (2021) 70–86

so the implementation of them becomes difficult when there exist function helps change the approaching behavior to the satu-
any parametric uncertainties. In such situations, adaptive strategy ration bound and can be used as fine-tuning parameter of
is preferentially considered in the design of the saturated schemes the controller.
[13–15] for rigid robots as well as the non-saturated ones [16,7] for 2) To resist the adverse effects of uncertain nonlinearities, a
flexible-joint robots. class of saturated neural network approximators (SNNAs)
Although such saturation problem has been well solved for the is designed in the slow control system. It provides a unified
rigid robot manipulators in the past years as mentioned above, few framework for adaptive backstepping control based on satu-
of existing works considers flexible-joint manipulators, due to that, rated neural networks, while simplifying the control law and
taking the actual factors of both the actuator saturation and the facilitating the implementation.
joint flexibility into consideration, the dynamics of a flexible- 3) A new filtered tracking error of the elastic torque is involved
joint robot manipulator becomes much more complex, and greatly in the fast control law, which does a great favor to make the
increases the difficulty of the controller implementation. fast subsystem subside quickly and to enhance the dynamic
For flexible-joint robot manipulators, a practical method that performance of the whole system.
considers actuator saturation is developed in [8] based on a fuzzy
error governor, which helps adjust the loop gains during critical In addition, overall stability analysis is presented to show that
occasions, to make the bandwidth of the fast controller decreased. the tracking errors of flexible-joint robot manipulators by the pro-
But this approach results in nonsmooth torque inputs to the joint posed controller are uniformly ultimately bounded (UUB) via Lya-
actuators, which would generate flexible impact on the robotic punov’s direct method.
mechanism to deteriorate the control performance. Motivated by The rest of this paper is organized as follows: In Section 2 the
[9], two novel saturated PD set-point controllers are presented in dynamics of flexible-joint robot manipulator and the decoupled
[17,18], where a Gibbs parameterization is used in the control subsystems based on singular perturbation theorem are presented.
law in [17] to take advantage of its higher relative proportional Section 3 deals with the design of a class of saturation functions
control effort further away from the desired set point compared and control law for the slow and the fast subsystems. In Section 4
to the use of Euler angles. Such schemes focused on the set-point stabilities of the subsystems and the whole system are analyzed by
control, without mentioning the tracking control problem. Consid- Lyapunov’s theorem. Performance comparisons and evaluations
ering the model uncertainties and actuator saturation, an are made in Section 5. Finally some conclusions are given in
improved predictive controller for a typical flexible-joint robot is Section 6.
proposed in [19] using the present and the previous states, where Notations: Throughout this paper, km ðÞ and kM ðÞ stand for the
the torque control input is only limited by a simple inequality smallest and largest eigenvalue of the matrix ðÞ, respectively. xM
without considering its smoothness. A robust Lyapunov-based con- and xm denote the maximum and minimum values of variable x,
troller is proposed in [20] for flexible-joint electrically driven respectively.
robots, where a hard (nonsmooth) saturation function composed
by a linear function and a dead-zone function is applied to limit 2. Dynamics for flexible-joint manipulator
the inputs. In [21], an actuator saturation prevention method that
can have variable saturation limits is introduced in a gain- 2.1. Basic dynamics
scheduled controller without violating the passive properties of
the controller. Upon using the spring model for the flexible joint, see Fig. 1, the
In addition, except the methods [22–24] to handle the input dynamic equations of an n degree-of-freedom flexible-joint robot
constraints problem for flexible cable-driven parallel manipulators, manipulator [7] with unknown internal nonlinearity are given as
stochastic nonlinear systems or general multiple-input multiple-
output nonlinear systems (like flexible-joint robot manipulators), MðqÞq € þ Cðq; qÞ
_ q_ þ GðqÞ þ D ¼ Kðh  qÞ ð1Þ
it is worth noting that, some novel works [25,26] that aiming to J h€ þ Kðh  qÞ ¼ u ð2Þ
solve the output constraints problem for single-input single-
output nonlinear systems or uncertain nonaffine systems also pro- where q 2 Rn and h 2 Rn represent the angular displacements of
vide different perspectives for reference to design saturated links and motor shafts, respectively. MðqÞ 2 Rnn represents the
controllers. _ 2 Rnn is the centripetal-Coriolis matrix,
inertia matrix, Cðq; qÞ
Although such achievements have been made, it is far from GðqÞ 2 R is the gravitational vector, D 2 Rn is the unknown internal
n

enough that the saturated tracking control problem for flexible- nonlinearity assumed to be bounded, K 2 Rnn is a constant diago-
joint robot manipulators with bounded and smooth torque input nal matrix representing the stiffness of joints, the motor inertia
remains open, and requires a general formulation with accurate matrix J 2 Rnn is a constant diagonal matrix. u 2 Rn is the torque
analysis by decomposing the rigid-flexible coupling dynamics, control input vector of joint motors.
which is worthwhile to be investigated to facilitate the exploration Remark 1: It is common that there usually exist unknown
of the dynamic characteristics and the design of saturated control internal nonlinearities D in the control of flexible-joint robot
schemes in the fields of rigid-flexible coupled control systems manipulators such as modeling error, transmission error by fric-
including but not limited to the flexible-joint robot manipulators. tion that acts indirectly or directly on the joint motor, and other
In this work we develop a generalized saturated tracking con- internal disturbances. These nonlinearities seriously affect the
troller based on backstepping control, singularly perturbed decou- control performance and even invalidate the control strategy.
pling and neural network approximation for flexible-joint robot
manipulators with bounded torque inputs. The main contributions
are summarized as follows.

1) A class of smooth saturation functions and projection-type

parameter adaptation law are used in the controller to keep
torque control inputs within the maximum torque that actu-
ators can provide. The saturation factor in the saturation Fig. 1. Model of the flexible joint.

71
X. Cheng, Y. Zhang, H. Liu et al. Neurocomputing 458 (2021) 70–86



Without loss of generality, in this paper, they are equivalently 0 I 0
ev_ ¼ v þ 1 € Þ:
ðu  J q ð15Þ
regarded as a bounded internal disturbance torque term on the J 1 K 1 0 J K1
motor side.
As introduced in [7,15], the model of flexible-joint robot manip- Ordering e ¼ 0, we get the unique isolated root of (15) as
ulator described by (1) and (2), has the following fundamental

I
v ¼ ½Z T eZ_ T  ¼
T
properties which can be exploited to facilitate the robot dynamics €Þ
ðus  J q ð16Þ
analysis. 0
Property 1: The inertia and centripetal-Coriolis matrices satisfy
the following skew symmetric relationship where Z _ is its first derivative
 is the desired elastic torque, and Z
respect to time t.
_
xT ½MðqÞ _
 2Cðq; qÞx ¼0 ð3Þ By substituting (16) into (11), we can obtain the expression of
_ _ þ Cðq; qÞ
MðqÞ ¼ Cðq; qÞ _ T: ð4Þ the slow subsystem as

Property 2: The inertia matrix MðqÞ is symmetric, positive def- € þ Cðq; qÞ

MðqÞq _ q_ þ GðqÞ þ D ¼ us  J q
€: ð17Þ
inite, and satisfies the following inequalities
Let
km ðMðqÞÞjjxjj2 6 xT MðqÞx 6 kM ðMðqÞÞjjxjj2 : ð5Þ
T
Property 3: For all x; y; z 2 Rn the centripetal-Coriolis matrix y ¼ ½zTf ez_ Tf  ¼ v  v
 ð18Þ
satisfies the following transformations
where zf is the tracking error of elastic torque, z_ f is its first deriva-
Cðx; yÞz ¼ Cðx; zÞy: ð6Þ tive respect to time t.
Property 4: The centripetal-Coriolis matrix, gravity terms can In the fast time scale s ¼ e1 t, after time-differentiating (18),
be bounded in the following manner substituting (15) and (16) in sequence, we get



_ 6 kC jjqjj;
jjCðq; qÞjj _ jjGðqÞjjM 6 kG ð7Þ dy 0 I 0
¼ yþ _
uf  ev ð19Þ
ds J 1 K 1 0 1
J K1
where kC and kG are positive constants.
Property 5: The joint stiffness matrix K satisfies the following and (11) can be rewritten as
inequalities
€ þ Cðq; qÞ
ðMðqÞ þ JÞq _ q_ þ GðqÞ þ D ¼ us þ ½I 0y: ð20Þ
km jjxjj2 6 xT Kx 6 kM jjxjj2 : ð8Þ
Property 6: The motor inertia matrix is bounded by Therefore, we obtain the redescribed dynamics (19) and (20) for
flexible-joint robotmanipulators in singularly perturbed form.
J m jjxjj2 6 xT Jx 6 J M jjxjj2 : ð9Þ  as constant in the fast time scale s and
According to [1], take v
y ¼ 0 in the slow time scalet. We get the slow and the fast subsys-
The properties listed above will help make the torque control inputs
tems as
bounded and will be used in the stability analysis of the system in
Section 4. ðMðqÞ þ JÞq€ þ Cðq; qÞ
_ q_ þ GðqÞ þ D ¼ us ð21Þ



dy 0 I 0
2.2. Decoupled model ¼ y þ uf : ð22Þ
ds J 1 K 1 0 J 1 K 1
In this subsection the dynamics (1) and (2) for flexible-joint
robot manipulators will be decoupled into a slow (reduced)
subsystem and a fast (boundary-layer) subsystem in the form 2.3. Control goal
of singularly perturbed systems, where the joint stiffness K is
assumed to be big enough, i.e., the robot manipulators dis- The control objective is to design a controller with bounded
cussed in this paper are expected to have slight flexibility at inputs jui j 6 uiM ði ¼ 1; 2; . . . ; nÞ, which guarantees the actual link
the joints. trajectory q converge asymptotically to the desired qd , where ui
We choose the link angle q and the elastic torque of the joints denotes the control torque input of the i-th joint, and correspond-
Zas slow and fast variables, respectively. Order K ¼ e2 K 1 (e is a ingly, uiM denotes the permitted maximum torque input of the i-th
constant small enough), then the elastic torque Z can be written as joint.
In brief, the control goal can be expressed as
Z ¼ Kðh  qÞ ¼ e2 K 1 ðh  qÞ: ð10Þ
9jui j 6 uiM ; 8e 2 Rn ; limeðtÞ ¼ 0 ð23Þ
Substitute (10) into (1) and (2), we have t!1

MðqÞq € þ Cðq; qÞ
_ q_ þ GðqÞ þ D ¼ Z ð11Þ where eðtÞ ¼ q  qd is the tracking error of the link trajectory.
In addition, we assume qd is twice differentiable, qd and its first
e2 J Z€ þ K 1 Z ¼ K 1 ðu  J q€Þ: ð12Þ
two derivatives are bounded for all t > 0 [6]
The control input u in (12) is given in a composite form as
jjqd jj 6 jjqd jjM ; jjq_ d jj 6 jjq_ d jjM ; jjq
€ d jj 6 jjq
€ d jjM : ð24Þ
u ¼ us þ u f ð13Þ
Remark 2: For almost any given trajectory, such assumptions by
where us and uf are the control inputs of the slow and the fast sub- (24) are easy to achieve through trajectory planning with splines or
systems, respectively. Such subsystems are obtained according to B-splines, more details can be found in our previous works [27].
the singular perturbation theory as follows. Note that, the composite controller (13) consists of two parts,
Here we define one is the slow control term presented by (21) and the other is
the fast control term by (22). The control law of each subsystem
v ¼ ½Z T eZ_ T  ;
T
ð14Þ
must be designed such that both the subsystems and the overall
then (12) can be rewritten as system are all stable.
72
X. Cheng, Y. Zhang, H. Liu et al. Neurocomputing 458 (2021) 70–86

3. Controller design The bounded virtual control law a1 with smooth saturation function
is proposed as
In this Section, aiming at making the outputs of the slow and
a1 ¼ c1 Satðz1 ; ke Þ þ q_ d ð29Þ
the fast controllers bounded, we will first introduce a class of sat-
uration functions in Section 3.1, and then develop control strate- where c1 and ke are constant diagonal matrix and saturation factor
gies for the decoupled slow subsystem described by (21) and fast diagonal matrix with positive elements, respectively.
subsystem described by (22), in Section 3.2 and 3.4, respectively, The differentiation of a1 with respect to time t is
where a class of saturated neural network approximators (SNNAs)
dc1 Satðz1 ; ke Þ
is designed in Section 3.3 to simplify the slow control law. a_ 1 ¼  €d :
þq ð30Þ
dt
3.1. A class of saturation functions To analyze the stability of the dynamics (25) and (26), a positive
definite Lyapunov function is defined as
To make the torque inputs bounded, saturation function is
1 T
adopted in the proposed control law. Different from the saturation L1 ¼ z z1 : ð31Þ
2 1
function presented in [28], which cannot ensure continuous differ-
entiation (only first-order differentiable) at the section point and The time derivative of L1 is
may result in nonsmooth torque outputs that would exert flexible
L_ 1 ¼ zT1 z_ 1 ¼ zT1 ðz2 þ a1  q_ d Þ: ð32Þ
impact on the joints, we apply a class of smooth saturation func-
tions to avoid such problems. Substituting (29) into (32), we get
Some properties on such saturation functions are listed in [15],
L_ 1 ¼ zT1 z_ 1 ¼ zT1 c 1 Satðz1 ; ke Þ þ zT1 z2 : ð33Þ
where Satðx; rÞ ¼ ½satðx1 ; r1 Þ; satðx2 ; r2 Þ; . . . ; satðxn ; rn ÞT 2 Rn is
T
continuously differentiable for all x ¼ ½x1 ; x2 ; . . . ; xn  2 R , and n
If z2 ¼ 0, the Eq. (33) becomes
r ¼ diag½r1 ; r2 ; . . . ; rn  2 Rnn is the saturation factor matrix with
L_ 1 ¼ zT1 c 1 Satðz1 ; ke Þ ð34Þ
positive diagonal elements, which can be used as fine-tuning
parameters of the controller to help change the approaching where Satðz1 ; ke Þ keeps the symbol as z1 , then L_ 1 is negative definite
behavior to the saturation bound. rM and rm represent the maxi- and the dynamics (25) is asymptotically stable. Hence, in the fol-
mum and the minimum values of the diagonal elements, respec- lowing step we will design the actual control law us to make z2 ! 0.
tively. Regarding the specific definition and properties of this Step 2: Design actual control law us .
type of saturation functions, we have made in-depth discussions After time-differentiating (27) and substituting (26) into it, we
in [15,43]. have

3.2. Sub-controller for the slow subsystem z_ 2 ¼ ðJ þ MÞ1 ðus  Cx2  G þ DÞ  a_ 1 : ð35Þ
A positive definite Lyapunov function is defined as
In this subsection, a saturated backstepping controller is pro-
posed through achieving system stability at each step. Backstep- 1
L2 ¼ L1 þ zT2 ðJ þ MÞz2 ; ð36Þ
ping method is an intuitive and recursive design approach for 2
constructing of state-feedback control laws, and it is associated
and its derivative with respect to t is
with Lyapunov function in each step to guarantee the stability of
the closed-loop system [28,29]. Motivated by the backstepping 1 _
L_ 2 ¼ L_ 1 þ zT2 ðJ þ MÞz_ 2 þ zT2 Mz 2: ð37Þ
method applied in [28], we propose a control law for the slow sub- 2
system considering the control goal given in Section 2.3 and invok-
Substituting (27), (30), (33), and (35) into (37) as well as using
ing the saturation function introduced in Section 3.1.
Property 1 in Section 2.1, we obtain
The design of the control law for the slow subsystem (21) is
similar to that for the rigid robot manipulators (composed by rigid L_ 2 ¼ zT1 c 1 Satðz1 ; ke Þ þ zT1 z2 þ zT2 c 1 f ðxÞ
ð38Þ
links and rigid joints), except for the saturation function invoked in þzT2 ½us  ðM þ JÞq €d  C q_ d  G þ D
each step to make the output torque bounded, see our previous
work [15].
To facilitate the expressions, we use M; C; C d and G to represent 1 ;ke Þ
T
_ Cðq; q_ d Þ and GðqÞ, respectively. where f ðxÞ ¼ ðM þ JÞ dSatðz þ CSatðz1 ; ke Þ; x ¼ ½zT1 zT2 xT1 xT2  .
MðqÞ; Cðq; qÞ; dt

Taking x1 ¼ q; x2 ¼ q_ as the state vector, the slow subsystem Invoking the saturation function introduced in Section 3.1, we
(21) can be rewritten in the state-space form as consider the following control law for (26):
€ d þ C d q_ d
us ¼ c 2 Satðz2 ; kv Þ þ ðM þ JÞq
x_ 1 ¼ x2 ð25Þ ð39Þ
þG  c 1 f ðxÞ  D
x_ 2 ¼ ðJ þ MÞ1 ðus  Cx2  G þ DÞ: ð26Þ
The controller is designed by following two steps. First, bounded
virtual control law a1 is designed for x1 to make ðx1  qd Þ ! 0; then where c 2 and kv are constant diagonal matrix and saturation factor
the actual input us is designed to make ðx2  a1 Þ ! 0. diagonal matrix with positive elements, respectively.
Step 1: Design the bounded virtual control law a1 for x1 . Substituting (39) into (38), we get
Define z1 ¼ x1  qd as the tracking error and
L_ 2 ¼ zT1 c 1 Satðz1 ; ke Þ  zT2 c 2 Satðz2 ; kv Þ
z2 ¼ x2  a1 ð27Þ ð40Þ
þzT1 z2 þ zT2 ðC d q_ d  C q_ d Þ:
where a1 is the bounded virtual control law designed for (25), then
we have
Using the Property 3 in Section 2.1 on (40), and then substituting
z_ 1 ¼ z2 þ a1  q_ d : ð28Þ
(28) and (29) into it. Subsequently, according to the Property 4 in
73
X. Cheng, Y. Zhang, H. Liu et al. Neurocomputing 458 (2021) 70–86

Section 2.1, the properties iii) and iv) on saturation functions in are combined to construct the consensus tracking control protocol
[15], we get in [30]. A two-layer neural network is used in [31] to learn the
unknown dynamics in an online manner. In addition, some other
L_ 2 6 km ðc1bke Þjjz1 jj2 þ ðkC jjq_ d jjM  km ðc2bkv ÞÞjjz2 jj2 feedforward neural networks such as wavelet based neural net-
þð1 þ kM ðc1ake ÞkC jjq_ d jjM Þjjz1 jjjjz2 jj ð41Þ work (WNN) [32], fuzzy WNN [33], etc., also play important roles
T in the adaptive backstepping control. Further, to increase the
¼ ½jjz1 jj jjz2 jjP ½jjz1 jj jjz2 jj
approximation accuracy and speed up the convergency, the feed-
back architecture is introduced subsequently into the structure of
neural networks. Tsai et al. [34] applied the recurrent neural net-
where a and b are positive constants defined in [15], the elements

 work (RNN) online learning and approximation with adaptive
p11 p12
of P ¼ are p11 ¼ km ðc1bke Þ; p22 ¼ km ðc2bkv Þ  kC jjq_ d jjM ; backstepping technique to achieve formation control of multiple
p21 p22
c k
nonlinear uncertain small-size unmanned helicopters for the first
1þkM ð 1a e ÞkC jjq_ d jjM
p12 ¼ p21 ¼  2
. time, and a new adaptive proportional-integral-derivative control
Specifically, if P is positive definite, then L_ 2 is negative definite. method using predictive control and output recurrent fuzzy
WNN is proposed in [35] for a group of nonlinear digital time-
Note that p11 ¼ km ðc1bke Þ > 0, by Sylvester’s theorem, if the elements
delay dynamic systems.
of P satisfy p11 p22  p212 > 0, i.e. Inspired by the works above, in this section, a class of saturated
c 1 ke c 2 kv c 1 ke
2 neural network approximators (SNNAs) is designed to approximate
4km ð Þðkm ð Þ  kC jjq_ d jjM Þ  ð1 þ kM ð Þk jjq_ jj Þ > 0; ð42Þ the uncertain nonlinearity term Nðf ðxÞ; DÞ.
b b a C d M
Since the outputs of the traditional neural network are not
then the positive definiteness of P is guaranteed. bounded, i.e., if it is used directly to approximate Nðf ðxÞ; DÞ in
Finally, we confirm that the proposed slow controller (39) can (43), the boundedness of the slow control law cannot be guaran-
keep the slow subsystem stable if (42) is satisfied. teed. To handle this problem, we propose a projection-type param-
It is worth noting that, the saturated sub-controller for slow eter adaptation law.
subsystem is designed by backstepping approach with considering Redefine Nðf ðxÞ; DÞ as
the saturation function in both the virtual control law a1 and the
 T /ðxNN Þ þ dðxNN Þ
Nðf ðxÞ; DÞ ¼ w
actual control law us to make the output torque strictly bounded. ð44Þ
In addition, saturation factors ke and kv can be used as fine- ^ ðxÞ; DÞ þ dðxNN Þ þ ðw
¼ Nðf T w
^ T Þ/ðxNN Þ
tuning gains to improve the tracking performance, by changing
the zero-crossing slope of the saturation function and the
approaching behavior to the saturation bound of the function. Spe- T ^ ðxÞ; DÞ ¼ w
where xNN ¼ ½zT1 zT2 xT1 xT2  and Nðf ^ T /ðxNN Þ are the input and
cially, compared with the virtual control law in the traditional the output of the SNNA, respectively. w ^ is the estimation of the
backstepping methods, e.g., a1 ¼ c1 z1 þ q_ d , in this paper,when a optimal weighting matrix  /ðxNN Þ ¼ ½/1 ðxNN Þ; /2 ðxNN Þ;    ;
w;
tracking error z1 appears, the saturation factor ke > 1 can help /m ðxNN Þ is the activation function vector, m is the number of nodes.
the virtual control law (29) approach the saturation bound imme- dðxNN Þ denotes the reconstruction error vector satisfying
diately to avoid the control performance getting worse; when the jjdðxNN Þjj 6 dN , and dN P 0 is the upper bound of jjdðxNN Þjj.
control system is stable, ke can also help improve the convergence Define the weight error w ~ as
accuracy.
To facilitate the implementation of us and enhance the practical ~ ¼w
w ^  w:
 ð45Þ
feasibility of the proposed approach, an uncertain nonlinearity Substituting (45) into (44), we get the SNNA as
term Nðf ðxÞ; DÞ is defined to replace the differential (see the first
part of f ðxÞ involved in (38)) included term c1 f ðxÞ plus the equiva- ^ ðxÞ; DÞÞ  w
Nðf ðxÞ; DÞ ¼ Nðf ~ T /ðxNN Þ þ dðxNN Þ: ð46Þ
lent internal torque disturbance term D in (39), i.e.,
^ ðxÞ; DÞÞ to substitute Nðf ðxÞ; DÞ in (43), the final saturated
Using Nðf
Nðf ðxÞ; DÞ ¼ c1 f ðxÞ þ D, and (39) becomes
adaptive backstepping controller for slow subsystem can be written
€ d þ C d q_ d
us ¼ c 2 Satðz2 ; kv Þ þ ðM þ JÞq as
ð43Þ
þG  Nðf ðxÞ; DÞ €d þ C d q_ d
us ¼ c 2 Satðz2 ; kv Þ þ ðM þ JÞq
ð47Þ
^
þG  c 1 w/ðx NN Þ:

Neural network will be applied in the following subsection to

approximate Nðf ðxÞ; DÞ so as to eliminate adverse effects caused
In order to make the parameters w^ involved in (47) bounded, we
by the lumped uncertain nonlinearity.
Remark 3: It is worth mentioning that, the neural network apply the following projection-type parameter adaptation law to
update w^
approximator designed below is expected to eliminate the differ-
ential calculations and simplify the expression of the slow control ^_ ¼ Projð/ðxNN ÞzT2 c 1 Þ
w
law us . More importantly, it avoids the differential explosion
caused by the virtual control derivation in traditional backstepping where the projection mapping is defined as
8
methods. ^i ¼ w
< 0 if w ^ M and i > 0
Projði Þ ¼ 0 if w^i ¼ w
^ m and i < 0 , w^ M and w
^ m are the maximum
:
3.3. Saturated Neural Network Approximator (SNNA) i otherwise
and minimum value of w ^ i , respectively.
Featured with good nonlinear mapping and learning ability, The parameter adaptation law is expected to have the following
neural network based adaptive strategies have recently been used properties [28]:
in backstepping control. The radial basis function neural network P1: The parameter estimates are always within the known
(RBFNN) based adaptive approach and backstepping technique bound at any time instant, wðtÞ ^ 2 Xw .

74
X. Cheng, Y. Zhang, H. Liu et al. Neurocomputing 458 (2021) 70–86

Fig. 2. Block diagram of the proposed controller.

P2: For any Cand f, there always exists quick subsiding rate will help speed up the convergence, and the
~ T ðC1 Projw^ ðCfÞ  fÞ 6 0.
w design of such controller will be discussed in the following
By Property 2, Property 4, and Property 6 in Section 2.1, prop- subsection.
erty ii) in [15], as well as (24), (47) and (48), the bound of the con-
trol input is assured with satisfying 3.4. Sub-controller for the fast subsystem

supðjusi jÞ ¼ supðc2i Þp þ supðc1i w ^ M Þ þ jjC di jjM jjq_ d jjM

ð49Þ The typical fast control law proposed in [38], which is in a brief
€ d jjM þ jGi jM
þðjjM i jjM þ jjJ i jjM Þjjq _ has been applied in many
derivative form expressed as uf ¼ kd eZ,
works recently [39]. It enables the closed-loop control without the
position measurements of joint actuators, but such fast controller
where M i ; J i ; C di and Gi denote the i-th row of M; J; C d and G; usi ; c 1i creates severe oscillation at the initial moment, especially when
and c2i denote the diagonal element in the i-th row of us ; c 1 and c2 , a non-zero tracking error exists at the beginning, since there is only
respectively. a damping term in the unbounded control law. Here we will try to
Remark 4: Compared with the traditional backstepping method handle this issue by adding a proportional term, meanwhile, with
in [36], a saturated control law us is proposed for the slow subsys- considering the problem of actuator saturation and convergence
tem, the boundedness of it is achieved by using a class of smooth rate of the fast subsystem.
saturation functions Satðx; rÞ and neural networks with Within the fast-scaled time domain, the fast subsystem (22) can
projection-type parameter adaptation law. Different from the non- be rewritten as
smooth fuzzy error governor in [8] and nonsmooth switching func-
tion in [37], the saturation function Satðx; rÞ in this paper is e€zf ¼ J 1 K 1 zf þ J 1 K 1 uf ð50Þ
continuous differentiable along the axis x, this feature makes the
where e is defined in Section 2.2.
torque outputs of joint actuators smooth enough to guarantee
In order to speed up the convergence rate of the error zf and
stable and fluent movement performance.
facilitate the fast subsystem stability analysis, a new filtered track-
Remark 5: A general expression of SNNAs is presented in this
ing error of the elastic torque r is defined as
subsection, it provides a uniform framework to design different
saturated neural networks. As special cases, when the activation r ¼ tez_ f þ zf ð51Þ
function is selected as the Gaussian function, it turns into a satu-
rated RBFNN of feedforward type. Based on this, by adding the out- where t is a positive constant.
put feedback strategy (taking the output at moment t back to the Consider the following fast control law
input as the updated input at moment ðt þ 1Þ, as presented in uf ¼ kpf Satðr; krf Þ  kvf Satðz_ f ; kdf Þ ð52Þ
[35]), the feedforward neural network (RBFNN) becomes a satu-
rated output recurrent RBFNN (named as ORRBFNN), which con- where kpf and kvf are constant diagonal matrices with positive
verges more quickly with better approximation accuracy. elements.
Meanwhile, change the activation function to wavelet function By property ii) in [15], the bound of the fast control input is
[35], and it becomes a saturated output recurrent wavelet neural assured with satisfying
network (ORWNN). supðjufi jÞ ¼ supðjkpfi j þ jkvfi jÞp ð53Þ
In this subsection, the slow control law us is designed for the
slow subsystem based on the backstepping method with a class where ufi ; kpfi , and kvfi denote the diagonal element in the i-th row of
of SNNAs, but for the whole system of the flexible-joint robot uf ; kpf , and kvf , respectively.
manipulator, the composite control law u consists of the slow con- By recalling (49), the bound of the composite controller is guar-
trol law us and the fast control law uf . Then a fast controller uf with anteed, i.e., jui j 6 uiM holds by satisfying the following inequality
75
X. Cheng, Y. Zhang, H. Liu et al. Neurocomputing 458 (2021) 70–86

uiM 6 supðjusi jÞ þ supðjufi jÞ; i ¼ 1; 2;    ; n ð54Þ Eq. (58) can be rewritten as

where uiM denotes the permitted maximum torque input of the i-th V_ 1 6 ½jjz1 jj jjz2 jjP½jjz1 jj jjz2 jjT þ kM ðc 1 Þjjz2 jjdN ð59Þ
joint actuator.
Remark 6: The essence of singular perturbation control is to where P is defined in (41), and it is positive definite if its elements
make the fast subsystem approach the manifold defined by the satisfy (42).
slow subsystem. In this paper, the control objective of the fast sub- Order e ¼ ½zT1 zT2 T , from (59) we get
system is to ensure that the actual elastic torque Z tracks the
 i.e., to make the tracking error of the elastic torque zf V_ 1 6 km ðPÞjjejj2 þ kM ðc1 ÞjjejjdN ð60Þ
desired Z,
converge to zero. As Z approaches Z,  the impact of joint flexibility due to jj
ejj2 ¼ jjz1 jj2 þ jjz2 jj2 P jjz2 jj2 .
on the system is gradually weakened. Thus the convergence perfor- From (60), it is shown that V_ 1 is negative definite for all 
e and t
mance of the fast (boundary-layer) subsystem will greatly affect as long as (55) is satisfied, and the radius of convergence of e  is
the control performance of the entire system.
kM ðc 1 ÞdN
Remark 7: The proposed fast control law uses a new filtered jjejj ¼ : ð61Þ
tracking error of the elastic torque r to make the boundary layer km ðPÞ
subside quickly and enhance the convergence performance of both
Thus, V_ 1 is negative definite outside a compact set. Eq. (61) shows
the fast subsystem and the whole system. Furthermore, different
that the control gains c1 , c2 ,ke and kv , which are contained in
from [39,40], the proposed fast saturated controller can ensure
km ðPÞ, can be selected large enough so that jj
ejj can be kept as small
the torque inputs are bounded all the time, and succeeds in reduc-
as possible. According to the standard Lyapunov theorem extension
ing the initial transient damped oscillation and suppressing the
[41,42], all signals of tracking errors are UUB.
vibration of torque even when encountering the external torque
disturbances obviously, due to the saturation function and the
4.2. Stability analysis of the fast subsystem
new filtered tracking error r applied in the control law. Such supe-
riority will be verified in Section 5.2.
Theorem 2: Consider the dynamics of the fast subsystem
described by (50) with the proposed sub-controller (52), there
3.5. Full expression of the controller
always exist proper t; kvf and kdf to make S in (64) positive definite,
and it results in exponential convergence of tracking errors in the
According to the theoretical work above, the proposed con-
sense that jjg_ jj; jjgjj ! 0, as t ! 1, with the convergence rate
troller for an n degree-of-freedom flexible-joint robot manipulator
J 1 K 1 kpf krf
by (1) and (2) is expressed as (13), (47) and (52). The block diagram 2tekm ð b
Þ.
of the controller is shown in Fig. 2. Proof: For the fast subsystem, consider the following positive
definite Lyapunov function
4. Stability analysis 1 T
V2 ¼ r r ð62Þ
2e
4.1. Stability analysis of the slow subsystem
where the vector r is defined in (51).
Theorem 1: Consider the dynamics of slow subsystem described After taking the fast-scaled time derivative of (51), and substi-
by (26) with the proposed sub-controller (47), there always exist a tuting (50) into it, we get
large enough km ðPÞ and a small enough kM ðc1 Þ if r_ ¼ z_ f  tJ 1 K 1 zf þ tJ 1 K 1 uf ð63Þ
kM ðc 1 ÞdN
km ðPÞ P ð55Þ Differentiating the Lyapunov function of V 2 with respect to the fast-
jjejj
scaled time s, substituting (51), (52), and (63) into it, and using
is satisfied, then it results that all the signals of tracking errors are property iv) in [15], we obtain
uniformly ultimately bounded (UUB).
V_ 2 6 ½jjzf jj jjz_ f jjS½jjzf jj jjz_ f jj
T
Proof: Since we add a class of SNNAs into backstepping con- ð64Þ
troller (39) to form a new control law (47) for the slow subsystem, r T tJ 1 K 1 kpf Satðr; krf Þ
then a new Lyapunov function candidate is defined as
1
~ T wÞ
V 1 ¼ L2 þ trðw ~ ð56Þ

2 S11 S12 2 ðkvf kdf ÞkM bJ m
where S ¼ , S11 ¼ tJkm ; S12 ¼ S21 ¼ et kM bþkM2bJ and
S21 S22 M m
where trðÞ is the trace of matrix .
After time-differentiating (56) and substituting (38), (46), (47), S22 ¼ et2 km ðkvfbJkdf Þkm  et.
M

and (48) in sequence, we get To assure that V_ 2 is negative definite, S should be positive def-
V_ 1 ¼ zT1 c 1 Satðz1 ; ke Þ þ zT1 z2 þ zT2 ðC d q_ d  C q_ d Þ inite. Note that S11 > 0, by Sylvester’s theorem, the following
~ T ðProjð/ðxNN ÞzT2 c 1 Þ ð57Þ inequality should be satisfied, i.e.,
zT2 c 2 Satðz2 ; kv Þ þ tr½w
2
/ðxNN ÞzT2 c 1 Þ þ zT2 c 1 d: 2 2 4km ðkvf kdf Þkm b2 J 2m 4km J 2m 2
ðt2 kM bÞ e ½ t3  t
J 2M JM
2
 2kM bðkM ðkvf kdf ÞkM  J M bÞt2 e þ ½kM ðkvf kdf ÞkM  J M b < 0 ð65Þ
Applying properties i), iii), and iv) in [15], and then substituting P2
of the projection-type parameter adaptation updating law into (57), Taking (65) as inequality equation of e, we order
we obtain
4km ðkvf kdf Þk2m b2 J 2m 2
1¼ t3  ½4kJmMJm þ 2k2M bkM ðkvf kdf Þ
c 1 ke c 2 kv J2M
ð66Þ
L_ 2 6 km ð Þjjz1 jj2 þ ðkC jjq_ d jj  km ð ÞÞjjz2 jj2
b b 2kM J M b2 t2
2
c 1 ke c ¼ 1  4ðt2 kM bÞ ½kM ðkvf kdf ÞkM  JM b2
2
ð67Þ
þð1 þ kM ð ÞkC jjq_ d jjÞjjz1 jjjjz2 jj þ kM ðc 1 Þjjz2 jjdN : ð58Þ
a
76
X. Cheng, Y. Zhang, H. Liu et al. Neurocomputing 458 (2021) 70–86

We can always find proper t; kvf and kdf to make enough and c 1 small enough. Besides, the upper bound of recon-
pffiffiffi struction error vector dN also exerts impact on the bound of jjXjj,
1 þ c > 0; c P 0; ð68Þ
so a proper neural network can make the lower bound of jjXjj smal-
then we get the solution of (65) as ler. Considering the singularly perturbed system (19) and (20) with
pffiffiffi pffiffiffi the composite controller u that consists of the slow controller us by
1 c 1þ c
maxð 2
; 0Þ < e < 2
: ð69Þ (47) and the fast controller uf by (52), if the slow control gains
2ðt2 kM bÞ 2ðt2 kM bÞ c1 ; c2 ; ke and kv are obtained according to (42) and (55), and the fast
A e that satisfies the condition (69) can be selected to make S pos- control gains kpf ; krf ; kvf and kdf are chosen according to (66)–(69),
itive definite, so from (64) we can get it can make all the error signals of the whole system UUB.
Remark 9: In the stability analysis, we can find that the tracking
V_ 2 6 rT tJ 1 K 1 kpf Satðr; krf Þ errors z1 ; z2 and zf are bounded. The assumption (24) and
J 1 K 1 kpf krf
ð70Þ x1 ¼ z1 þ qd indicate that x1 is bounded. Besides, the virtual control
6 2tekm ð ÞjjV 2 jj2 :
law a1 is bounded since that the saturation function and q_ d are
b

bounded, see (29). So it leads to the boundedness of the state x2 ,

since x2 ¼ a1 þ z2 . According to the bounded x1 ; x2 and
Finally, the fast subsystem is guaranteed to be exponentially stable projection-type parameter adaptation law (48), we get that the sat-
J 1 K 1 kpf krf
and the exponential convergence rate is 2tekm ð b
Þ. urated backstepping controller us for the slow subsystem is
bounded. Meanwhile, the fast control law uf is bounded on account
4.3. Stability analysis of the whole system of the employed saturation function, see (52). Thus we can draw a
conclusion that the control input u is bounded.
Theorem 3: Consider the dynamics of the whole system Remark 10: Although the theoretical results show that the pro-
described by (26) and (50) with the proposed sub-controller (13), posed scheme is qualified to perform trajectory tracking control for
(47) and (52), if Q and X in (74) satisfy that Q is positive definite a flexible-joint robot manipulator with bounded torque input, it
and does not mean that the permitted upper bound of the joint torque
can be arbitrarily small. First of all, the system must satisfy the
kM ðc 1 ÞdN conditions of static equilibrium, that is, the joint torque inputs
jjXjj > ; ð71Þ
km ðQ Þ should at least succeed in maintaining the body balance in any
then it results that all the error signals of the whole system are UUB posture. Second, as can be seen from (13), (47) and (49), the upper
and the control object (23) is achieved. bound of the control law is related to the desired trajectory qd . In
Proof: In the previous subsections, the stability proof of each other words, for a certain joint, tracking the same trajectory at dif-
subsystem is made. This subsection focuses on the stability analy- ferent speeds, the control input is usually not the same. Finally, the
sis of the whole system. permitted maximum torque input should not be too small, other-
By recalling the dynamics of flexible-joint robot manipulators wise it may make the actual output of link position fail to track
(19) and (20), and the composite control law (13), which consists its desired value, and further, the tracking errors will not converge
of the slow controller (47) and the fast saturated controller (52), as described in the stability analysis.
we consider the following Lyapunov function
V ¼ dV 1 þ ð1  dÞeV 2 ; 0 < d < 1: ð72Þ 5. Numerical simulations

Differentiate (72) with respect to time t, and substitute (60) and In this section, simulation comparisons are conducted to verify
(70), we get the proposed scheme in two parts. One part is to evaluate the pro-
posed controllers adopting different SNNAs, and the other is to
V_ 6 d½km ðPÞjjejj2 þ kM ðc 1 ÞjjejjdN þ jjejjjjyjj
ð73Þ make performance comparisons among several existing represen-
ð1  dÞr T tJ 1 K 1 kpf Satðr; krf Þ: tative controllers and ours.
All the tests are made on a two link direct-driven flexible-joint
manipulator, whose dynamics is given in [15] with
Considering that there always exists a positive constant u1 such J ¼ diag½0:1; 0:1 and K ¼ diag½10000; 10000. Without loss of gen-
T
that r T tJ 1 K 1 kpf Satðr; krf Þ P u1 jjyjj2 , define 
e ¼ ½zT1 zT2  and erality, the desired position trajectories with nonzero initial track-
T T ing errors for each link are given as the same in [43], external
X ¼ ½
eT y  , we obtain
torque disturbances with amplitudes of 6sinðtÞ Nm to the first
V_ 6 km ðQ ÞjjXjj2 þ kM ðc 1 ÞjjXjjdN ð74Þ joint and 4sinðtÞ Nm to the second one are added from the start

 of the 7th second to the end of the 8th second.
dkm ðPÞ  2d The most commonly used saturation function in bounded con-
where Q ¼ ; P is defined in (41).
 2d ð1  dÞu1 trol schemes in previous studies is hyperbolic tangent function
Note that dkm ðPÞ > 0, according to Sylvester’s theorem, if tanhðxi ; ri Þ [23,44,17]. In this work, the arc tangent function
2
dkm ðPÞð1  dÞu1  d4 P 0, i.e. atanðxi ; ri Þ, as a special case of the general saturation function intro-
duced in Section 3.1, is used in the proposed control law, it has a
4km ðPÞu1
d6 ; ð75Þ wider range of ðp=2; p=2Þ than ð1; 1Þ of tanhðxi ; ri Þ. In response
4km ðPÞu1 þ 1 to increasing value of ri , the zero-crossing slope of function
then Q is positive definite. atanðxi ; ri Þ increases more steeply and approaches saturation more
Hence, V_ is negative definite for all 
eand X with respect to time t rapidly than that of tanhðxi ; ri Þ. More discussions on it see [43].
as long as (71) is satisfied, i.e., all the error signals of the whole sys-
tem are UUB and the control goal (23) is achieved. 5.1. Comparisons of the proposed controllers involving different SNNAs
Remark 8: jjXjj contains tracking error information of both link
position and elastic torque for each joint. The lower bound of jjXjj The comparisons in this part are made to verify the proposed
can be achieved as small as possible by selecting km ðQ Þ large SNNAs in online learning mode, where the proposed control
77
X. Cheng, Y. Zhang, H. Liu et al. Neurocomputing 458 (2021) 70–86

approaches without SNNA, and with different SNNAs (RBFNN/ relatively large amplitude oscillations, while the other three
ORWNN/ORRBFNN) in the slow sub-controller, are named as controllers with SNNAs show superior tracking performances with
SB + SP-SD, SAB-RBF + SP-SD, SAB-ORW + SP-SD and SAB- much smaller errors, due to that the SNNAs produce adaptive tor-
ORRBF + SP-SD, respectively. que compensation for time-varying errors, as shown in Figs. 5–8,
All the parameters in the control are selected with considering especially at the initial moment and when the external disturbance
both the bound conditions by (49), (53), (54) on the control inputs, occurs. In addition, benefiting from the RNNs of output type
and the stability conditions by (42, (60), (66)–(69) on the control (ORNNs), controllers SAB-ORRBF + SP-SD and SAB-ORW + SP-SD
system. The parameters in the slow sub-controller us by (47) are have better dynamic response performance and anti-disturbance
chosen as c1 ¼ diag½1:5; 3; c 2 ¼ diag½25; 7; ke ¼ kv ¼ diag½3; 5. ability, than controller SAB-RBF + SP-SD with RBFNN of feedfor-
The upper and the lower bounds of the adaptive parameters of ward type.
SNNA are set as w ^ M ¼ 20 and w ^ m ¼ 20. The parameters in the fast Remark 11: As the results indicate that, the overall performance
sub-controller uf by (51) are chosen as e ¼ 0:0001; t ¼ 200; krf ¼ of the controllers (SAB-ORRBF + SP-SD and SAB-ORW + SP-SD)
diag½2; 2; kpf ¼ diag½10; 5; kdf ¼ diag½2; 2, and kvf ¼ diag½0:005; 0:009. adopting output recurrent SNNAs are superior to the controller
In addition, the internal disturbance torque is set as D ¼ 6sinðtÞ (SAB-RBF + SP-SD) adopting feedforward SNNA. More specifically,
Nm throughout all the tests. with respect to the performance presented by the controllers with
The tracking results are as shown in Figs. 3, 4 and Table 1, where output recurrent SNNAs, controller SAB-ORRBF + SP-SD is very
five evaluation criteria for the tracking errors are adopted: the root close to controller SAB-ORW + SP-SD. From the perspective of neu-
mean squared error (RMSE), integral squared error (ISE), integral ral network structure, both of the output recurrent SNNAs adopted
time-weighted squared error (ITSE), integral absolute error (IAE), in controllers SAB-ORRBF + SP-SD and SAB-ORW + SP-SD have 8
and integral time-weighted absolute error (ITAE) [34]. Without nodes in the input layer, 2 nodes in the output layer, but the
any SNNA to generate compensation to resist the external and ORRBFNN in the former has only one hidden layer including 14
internal torque disturbances, the controller SB + SP-SD presents nodes, it is much fewer than that of the ORWNN in the latter,

Fig. 3. Tracking error of the first link.

Fig. 4. Tracking error of the second link.

78
X. Cheng, Y. Zhang, H. Liu et al. Neurocomputing 458 (2021) 70–86

Table 1
Controller performance comparisons.

SB + SP-SD SAB-RBF + SP-SD SAB-ORW + SP-SD SAB-ORRBF + SP-SD

2 Link 1 2.27 1.80 1.73 1.74
RMSE (10 rad)
Link 2 2.14 1.79 1.55 1.56

ISE(103 rad) Link 1 8.22 5.21 4.80 4.80

Link 2 7.32 5.13 3.84 3.91

ITSE (104 rad) Link 1 211 5.95 4.80 4.90

Link 2 140 2.66 1.33 1.38

IAE (102 rad) Link 1 22.41 4.83 4.19 4.18

Link 2 17.39 3.55 2.18 2.15

ITAE (102 rad) Link 1 145.54 3.83 1.82 1.87

Link 2 114.55 4.03 0.50 0.42

Fig. 5. Torque control input for the first joint.

Fig. 6. Torque control input for the second joint.

where there are two hidden layers, one has 56 nodes for wavelet 5.2. Comparisons with three existing approaches
decomposition and the other has 14 nodes for reconstruction.
Considering that ORRBFNN has good performance and simple To verify the proposed scheme, simulation comparisons among
structure as mentioned above, it is preferentially used in the pro- a typical controller without model decomposition in [45], and two
posed approach to design the SNNA of controller SAB- composite controllers under singular perturbation frame in [40,39]
ORRBF + SP-SD. are made. The controller in [45] is a sliding mode controller under

79
X. Cheng, Y. Zhang, H. Liu et al. Neurocomputing 458 (2021) 70–86

Fig. 7. Working process of different SNNAs for the first joint.

Fig. 8. Working process of different SNNAs for the second joint.

Fig. 9. Tracking error of the first link.

80
X. Cheng, Y. Zhang, H. Liu et al. Neurocomputing 458 (2021) 70–86

Fig. 10. Tracking error of the second link.

Table 2
Controller performance comparisons.

SMC -FSF PD-DC + CC EPID + DC SAB-ORRBF + SP-SD

Adjusting time (sec) Link 1 2.39 4.43 1.33 0.85
Link 2 4.30 4.79 4.46 0.51
Recovery time (sec) Link 1 1.87 1.78 0.71 0.33
Link 2 3.31 3.61 3.18 0.35

Steady-state error (106 rad) Link 1 5311 2162 1528 2.58

Link 2 2863 1070 652 3.09

Maximum deviation (104 rad) Link 1 56.31 52.89 7.89 4.49

Link 2 27.43 48.94 78.88 4.45
Maximum torque (Nm) Joint 1 230.11 294.96 183.29 84.72
Joint 2 135.58 172.91 117.80 17.98

Table 3
Controller performance comparisons.

SMC-FSF PD-DC + CC EPID + DC SAB-ORRBF + SP-SD

2 Link 1 2.87 2.65 1.49 1.74
RMSE (10 rad)
Link 2 3.06 4.23 2.71 1.56

ISE (103 rad) Link 1 1.61 11.20 3.61 4.80

Link 2 1.79 28.72 11.68 3.88

ITSE (104 rad) Link 1 16.10 42.21 8.22 4.90

Link 2 3.19 120.02 26.96 1.38

IAE (102 rad) Link 1 5.61 13.67 5.76 4.18

Link 2 3.21 22.58 10.68 2.15

ITAE (102 rad) Link 1 37.50 18.09 18.05 1.87

Link 2 15.94 25.50 19.80 0.42

full-sate feedback originally aimed at rigid robots and is modified participates in the comparison as a representative of the proposed
here for flexible-joint robots by adding joint position constraints approach, where the parameters are as the same as that in
(named as SMC-FSF); the composite controller in [40] consists of Section 5.1.
a proportional-derivative sub-controller plus dynamics compensa- As shown in Figs. 9, 10 and Table 2 (where five evaluation crite-
tion for the slow subsystem, and a corrective sub-controller of ria for the control performance are employed: Adjusting time,
vibrational damping for the fast subsystem (named as PD- Recovery time, Steady-state error, Maximum deviation and Maxi-
DC + CC); the composite controller in [39] consists of an efficient mum torque, corresponding definitions can be found in
proportional-integral-derivative (PID)-type sub-controller for the [46,15,43]), the proposed controller SAB-ORRBF + SP-SD has the
slow subsystem, and a derivative-type sub-controller for the fast shortest adjusting time, recovery time and smallest steady-state
subsystem (named as EPID + DC). Controller SAB-ORRBF + SP-SD error, maximum deviation, maximum torque, which demonstrates

81
X. Cheng, Y. Zhang, H. Liu et al. Neurocomputing 458 (2021) 70–86

Fig. 11. Torque control input for the first joint.

Fig. 12. Torque control input for the second joint.

Fig. 13. Tracking error of elastic torque of the first joint.

82
X. Cheng, Y. Zhang, H. Liu et al. Neurocomputing 458 (2021) 70–86

Fig. 14. Tracking error of elastic torque of the second joint.

Fig. 15. Torque control input of the fast controller for first joint.

Fig. 16. Torque control input of the fast controller for second joint.

83
X. Cheng, Y. Zhang, H. Liu et al. Neurocomputing 458 (2021) 70–86

overall significant advantages over the other three benchmark con- the main results to (1) bounded control problems for general
trollers. Comparisons of tracking errors among the four controllers Euler–Lagrangian Systems with uncertain nonlinearity [47] and
by more quantitative evaluation criteria are listed in Table 3. (2) time-varying estimation problems for networked systems with
As shown in Figs. 11, 12 and Table 2, All the benchmark con- unknown boundness [48–50].
trollers fail to keep the maximum torque input at the initial
moment within the given limited range (120 Nm for the first joint
and 20 Nm for the second one), while the torque control inputs of CRediT authorship contribution statement
the proposed controller SAB-ORRBF + SP-SD keep bounded all the
time and just generate a maximum torque of 84.72 Nm for the first Xin Cheng: Conceptualization, Methodology, Software, Valida-
joint and 17.98 Nm for the second one. This is for the reason that tion, Formal analysis, Investigation, Writing - original draft, Writ-
there is no anti-saturation strategy adopted in the three bench- ing - review & editing, Visualization, Project administration.
mark controllers, but in the proposed controller SAB-ORRBF + SP- Yajun Zhang: Conceptualization, Methodology, Software, Valida-
SD, smooth saturation function is applied in both the slow and tion, Formal analysis, Investigation, Writing - original draft, Writ-
the fast saturated sub-controllers to create bounded inputs. In ing - review & editing. Huashan Liu: Conceptualization,
addition, SNNA with projection-type parameter adaptation in the Methodology, Resources, Data curation, Writing - review & editing,
slow control law is also used to help keep the control inputs Supervision, Project administration, Funding acquisition. Dirk
bounded strictly. Wollherr: Data curation, Writing - review & editing. Martin Buss:
Importantly, controller SAB-ORRBF + SP-SD significantly Data curation, Writing - review & editing.
reduces the initial transient damped oscillation as shown in Figs. 11
and 12, due to a new filtered tracking error of the elastic torque r Declaration of Competing Interest
applied in the fast control law (52), which accelerates the conver-
gence for both the fast subsystem (as shown by Figs. 13 and 14) The authors declare that they have no known competing finan-
and the whole system (as shown by Figs. 9 and 10). With respect cial interests or personal relationships that could have appeared
to the anti-disturbance capability, the fast subsystem plays a vital to influence the work reported in this paper.
role in the whole system, its working process is shown in Figs. 15
and 16. With a combined application of the saturation function and
the new filtered tracking error of the elastic torque, the fast con- Acknowledgments
troller responses very quickly to weaken the influence exerted by
the external disturbances. Apart from the control gains kpf and This work was sponsored in part by the Shanghai Rising-Star
kvf in (52), the error gains krf and kdf are designed to act as a fine Program under Grant 19QA1400400, the Natural Science Founda-
adjustment for the control performance within the saturation tion of Shanghai under Grant 21ZR1401100 and the Fundamental
function in the fast control law. Research Funds for the Central Universities under Grant
2232019G-09.

6. Conclusion References

This paper discusses the tracking control of flexible-joint robot [1] M.W. Spong, Modelling and control of elastic joint robots, J. Dyn. Syst. Meas.
manipulators with bounded torque inputs. We have shown that Control 109 (1987) 310–319.
[2] A. Loría, H. Nijmeijer, Bounded output feedback tracking control of fully-
the flexible-joint robot manipulator is decoupled into a slow sub- actuated Euler-Lagrange systems, Syst. Control Lett. 33 (1998) 151–161.
system and a fast subsystem within the framework of singular per- [3] V.M. Hernández-Guzmán, V. Santibáñez, A. Zavala-Río, A saturated PD
turbation theory. The subsystems have their independent state controller for robots equipped with brushless DC-motors, Robotica 28 (2010)
405–411.
variables and are directly applicable to any control strategies. A [4] Y. Su, P.C. Mueller, C. Zheng, Global asymptotic saturated PID control for robot
general bounded composite controller is presented with applying manipulators, IEEE Trans. Control Syst. Technol. 18 (2010) 1280–1288.
a class of smooth saturation functions, where a slow sub- [5] H.S. Liu, S.Q. Zhu, A generalized trajectory tracking controller for robot
manipulators with bounded inputs, J. Zhejiang Univ.-Sci. A 10 (2009) 1500–
controller is designed by improved backstepping method and sat-
1508.
urated neural network approximator (SNNA), and a fast sub- [6] H.S. Liu, S.Q. Zhu, Z. Chen, Saturated output feedback tracking control for robot
controller involving a new filtered tracking error of the elastic tor- manipulators via fuzzy self-tuning, J. Zhejiang Univ.-Sci. C 11 (2010) 956–966.
que is created to make the fast subsystem subside quickly. Impor- [7] H.S. Liu, Y. Huang, Robust adaptive output feedback tracking control for
flexible-joint robot manipulators based on singularly perturbed decoupling,
tantly, it is proved strictly in theory that the proposed composite Robotica 36 (2018) 822–838.
controller succeeds in keeping the torque control inputs to each [8] S. Ozgoli, H.D. Taghirad, Fuzzy error governor: a practical approach to counter
joint actuator bounded, and a comprehensive stability analysis is actuator saturation on flexible joint robots, Mechatronics 19 (2009) 993–1002.
[9] Y. Su, C. Zheng, Globally asymptotic stabilization of spacecraft with simple
made for the overall system. This work clearly states the design saturated proportional-derivative control, J. Guid. Control Dynam. 34 (2011)
details of the tracking controller with considering both the actua- 1932–1935.
tor saturation and joint flexibility of the robot manipulators. [10] M. Mendoza, A. Zavala-Río, V. Santibáñez, F. Reyes, Output-feedback
proportional-integral-derivative-type control with simple tuning for the
The proposed controller for flexible-joint manipulator is tested global regulation of robot manipulators with input constraints, IET Contr.
and evaluated on a two degree-of-freedom flexible-joint robot Theory Appl. 9 (2015) 2097–2106.
manipulator. The simulation results demonstrate that, motor com- [11] G.I. Zamora-Gomez, A. Zavala-Río, D.J. Lpez-Araujo, V. Santibáñez, Further
results on the global continuous control for finite-time and exponential
mand torque of the proposed controller can effectively keep stabilisation of constrained-input mechanical systems: desired conservative-
bounded, i.e., benefiting from the proposed saturation functions force compensation and experiments, IET Contr. Theory Appl. 13 (2019) 159–
and SNNAs, torque control inputs to each joint actuator are always 170.
[12] X. Hu, X. Wei, H. Zhang, J. Han, X. Liu, Robust adaptive tracking control for a
bounded within a given limited range. Moreover, the proposed fast
class of mechanical systems with unknown disturbances under actuator
saturated sub-controller is validated to be helpful to reduce the ini- saturation, Int. J. Robust Nonlinear Control 29 (2019) 1893–1908.
tially damped transient oscillations significantly. All of these fea- [13] S. Jia, J. Shan, Finite-time trajectory tracking control of space manipulator
tures guarantee a superior dynamic performance and make the under actuator saturation, IEEE Trans. Indus. Elect. 67 (2020) 2086–2096.
[14] Q. Zhou, S. Zhao, H. Li, R. Lu, C. Wu, Adaptive neural network tracking control
proposed controller as a valuable reference for other similar for robotic manipulators with dead zone, IEEE Trans. Neural Netw. Learn. Syst.
robotic systems. Further research topics include the extension of 30 (2019) 3611–3620.

84
X. Cheng, Y. Zhang, H. Liu et al. Neurocomputing 458 (2021) 70–86

[15] H.S. Liu, Y.J. Zhang, W.X. Wu, Saturated adaptive backstepping control for [46] H.S. Liu, Y. Huang, Bounded adaptive output feedback tracking control for
robot manipulators with RBF neural network approximation, in: Proc. IEEE Int. flexible-joint robot manipulators, J. Zhejiang Univ. Sci. A 19 (2018) 557–578.
Conf. Info. Autom., 2016, pp. 1550–1555.. [47] Z. Zhang, M. Leibold, D. Wollherr, Integral sliding-mode observer-based
[16] M. Razmi, C.J.B. Macnab, Near-optimal neural-network robot control with disturbance estimation for euler-lagrangian systems, IEEE Trans. Control
adaptive gravity compensation, Neurocomputing 389 (2020) 83–92. Syst. Technol. (to be published, doi: 10.1109/TCST.2019.2945904.)..
[17] R.J. Caverly, D.E. Zlotink, L.J. Bridgeman, J.R. Forbes, Saturated proportional [48] L. Zou, Z.D. Wang, D.H. Zhou, Moving horizon estimation with non-uniform
derivative control of flexible-joint manipulators, Robot. Comput. Integr. sampling under component-based dynamic event-triggered transmission,
Manuf. 30 (2014) 658–666. Automatica (to be published, doi: 10.1016/j.automatica.2020.109154)..
[18] R.J. Caverly, D.E. Zlotink, J.R. Forbes, Saturated proportional derivative control [49] L. Zou, Z.D. Wang, Q.L. Han, D.H. Zhou, Moving horizon estimation of
of a single-link flexible-joint manipulator, Trans. Can. Soc. Mech. Eng. 38 networked nonlinear systems with random access protocol, IEEE Trans. Syst.,
(2014) 241–250. Man, Cybern. B, Cybern. (to be published, doi: 10.1109/TSMC.2019.2918002)..
[19] N.O. Ghahramani, F. Towhidkhah, Constrained incremental predictive [50] L. Zou, Z.D. Wang, J. Hu, D.H. Zhou, Moving horizon estimation with unknown
controller design for a flexible joint robot, ISA Trans. 48 (2009) 321–326. inputs under dynamic quantization effects, IEEE Trans. Autom. Control (to be
[20] A. Izadbakhsh, Robust control design for rigid-link flexible-joint electrically published, doi: 10.1109/TAC.2020.2968975)..
driven robot subjected to constraint: theory and experimental verification,
IEEE/ASME Trans. Mechatron. 85 (2016) 751–765.
[21] A. Walsh, J.R. Forbes, A very strictly passive gain-scheduled controller: theory
and experiments, IEEE/ASME Trans. Mechatronics 21 (2016) 2817–2826. Xin Cheng received the B.E. degree in Automation from
[22] R.J. Caverly, J.R. Forbes, Flexible cable-driven parallel manipulator control: Donghua University, Shanghai, China, in 2017. Since
maintaining positive cable tensions, IEEE Trans. Control Syst. Technol. 26 September, 2017, he has been with the Department of
(2018) 1874–1883. Electrical and Electronic Engineering, College of Infor-
[23] H. Wang, B. Chen, X. Liu, K. Liu, C. Lin, Robust adaptive fuzzy tracking control mation Science and Technology in Donghua University,
for pure-feedback stochastic nonlinear systems with input constraints, IEEE where he is currently a PhD candidate. His current
Trans. Cybern. 43 (2013) 2093–2104. research interests include robot control.
[24] M. Chen, S.S. Ge, B.V.E. How, Robust adaptive neural network control for a class
of uncertain MIMO nonlinear systems with input nonlinearities, IEEE Trans.
Neural Netw. 21 (2010) 796–812.
[25] K.P. Tee, S.S. Ge, E.H. Tay, Barrier Lyapunov functions for the control of output-
constrained nonlinear system, Automatica 45 (2009) 918–927.
[26] W. Meng, Q. Yang, J. Si, Y. Sun, Adaptive neural control of a class of output-
constrained nonaffine system, IEEE Trans. Cybern. 46 (2016) 85–95.
[27] H.S. Liu, X.B. Lai, W.X. Wu, Time-optimal and jerk-continuous trajectory
planning for robot manipulators with kinematic constraints, Robot. Comput.
Integr. Manuf. 29 (2013) 309–317. Yajun Zhang received the B.E. degree in Electrical
[28] Y. Hong, B. Yao, A globally stable saturated desired compensation adaptive Engineering from Shanghai Institute of Technology,
robust control for linear motor systems with comparative experiments, Shanghai, China and the M.E. degree in Control Science
Automatica 43 (2007) 1840–1848. and Technology from Donghua University, Shanghai,
[29] H. Dastres, B. Rezaie, B. Baigzadehnoe, Neural-network-based adaptive China in 2014 and 2017, respectively. Since September
backstepping control for a class of unknown nonlinear time-delay systems 2017, he has been pursuing his Ph.D. degree at South-
with unknown input saturation, Neurocomputing 398 (2020) 131–152. east University, Nanjing, China. His research interests
[30] Y. Shang, B. Chen, C. Lin, Consensus tracking control for distributed nonlinear include mechatronic system of robotics.
multiagent systems via adaptive neural backstepping approach, IEEE Trans.
Syst., Man, Cybern. Syst. PP (2018) 1–9..
[31] W.C. Meng, Q.M. Yang, J. Sarangapani, Y.X. Sun, Distributed control of
nonlinear multiagent systems with asymptotic consensus, IEEE Trans. Syst.,
Man Cybern. Syst. 47 (2017) 749–757.
[32] P.M.R. Bento, J.A.N. Pombo, M.R.A. Calado, S.J.P.S. Mariano, Optimization
of neural network with wavelet transform and improved data selection
using bat algorithm for short-term load forecasting, Neurocomputing
358 (2019) 53–71. Huashan Liu received the B.E. degree in Mechanical
[33] X. Bu, Y. Xiao, H. Lei, An adaptive critic design-based fuzzy neural controller for Engineering from Wuhan University, Wuhan, China, and
hypersonic vehicles: Predefined behavioral nonaffine control, IEEE/ASME the Ph.D. degree in Mechatronics from Zhejiang
Trans. Mechatron. 24 (2019) 1871–1881. University, Hangzhou, China in 2005 and 2010, respec-
[34] C.-W. Kuo, C.-C. Tsai, C.-T. Lee, Intelligent leader-following consensus tively. Since 2010, he has been with the College of
formation control using recurrent neural networks for small-size unmanned Information Science and Technology, Donghua Univer-
helicopters, IEEE Trans. Syst., Man, Cybern. Syst. PP (2019) 1–14..
sity, Shanghai, China, where he is currently an Associate
[35] C.-C. Tsai, C.-C. Yu, C.-T. Tsai, Adaptive orfwnn-based predictive pid control,
Professor and Head of Department of Electrical and
Int. J. Fuzzy Syst. 21 (2019) 1544–1559.
[36] S.J. Yoo, Distributed adaptive containment control of networked flexible-joint Electronic Engineering. During 2015-2016 he was a
robots using neural networks, Expert Syst. Appl. 41 (2014) 470–477. Visiting Professor with the Institute of Automatic Con-
[37] T. Zhao, Y.Y. Liu, Z.J. Li, C.Y. Su, Y. Feng, Adaptive control and optimization of trol Engineering (LSR), Technical University of Munich
mobile manipulation subject to input saturation and switching constraints, (TUM), Munich, Germany. He was selected for the
IEEE Trans. Autom. Sci. Eng. 16 (2019) 1543–1555. Shanghai Rising-Star Program in 2019. His current research interests include
[38] S.S. Ge, L.C. Woon, Adaptive neural network control of flexible joint robotics, mechatronics and motion control.
manipulators in constrained motion, Trans. Inst. Meas. Control 20 (1998)
37–46.
[39] Y.P. Pan, X. Li, H.Y. Yu, Efficient PID tracking control of robotic manipulators Dirk Wollherr is currently Professor at the Chair of
driven by compliant actuators, IEEE Trans. Control Syst. Technol. 27 (2019) Automatic Control Engineering, Faculty of Electrical
915–922. Engineering and Information Technology, Technical
[40] M.A. Khosravi, H.D. Taghirad, Dynamic modeling and control of parallel robots University of Munich, Germany. He received the
with elastic cables: singular perturbation approach, IEEE Trans. Robot. 30
diploma engineer degree in Electrical Engineering in
(2014) 694–704.
2000, the Doctor of Engineering degree in Electrical
[41] K.S. Narendra, A.M. Annaswamy, A new adaptive law for robust adaptation
Engineering in 2005 and the Habilitation degree in 2013
without persistent excitation, IEEE Trans. Autom. Control 32 (1987) 134–145.
[42] C. Kwan, F.L. Lewis, Robust backstepping control of nonlinear systems using from the Technical University of Munich, Germany.
neural networks, IEEE Trans. Syst. Man Cybern., A, Syst. Humans 30 (2000) From 2001-2004 he has been research assistant at the
753–766. Control Systems Group, Technical University of Berlin,
[43] H.S. Liu, K.R. Hao, X.B. Lai, Fuzzy saturated output feedback tracking control for Germany. In 2004 he has been granted a research fel-
robot manipulators: a singular perturbation theory based approach, Int. J. Adv. lowship from the Japanese Society for the Promotion of
Robot. Syst. 8 (2011) 43–53. Science (JSPS) at the YoshihikoNakamura-Lab, The University of Tokyo, Japan. From
[44] V. Santibáñez, K. Camarillo, J. Moreno-Valenzuela, R. Campa, A practical PID 2006-2008 he has been General Manager of the Cluster of Excellence ‘‘Cognition for
regulator with bounded torques for robot manipulators, Int. J. Control Autom. Technical Systems (CoTeSys)”. He has been TUM Carl-von-Linde Junior Fellow,
Syst. 8 (2010) 544–555. Institute of Advanced Studies from 2010 until 2013. His research interests include
[45] J.Y. Lee, P.H. Chang, M.L. Jin, Adaptive integral sliding mode control with time- automatic control, robotics, autonomous mobile robots, human-robotinteraction,
delay estimation for robot manipulators, IEEE Trans. Indus. Elect. 64 (2017) and artificial intelligence for robotics.
6796–6804.

85
X. Cheng, Y. Zhang, H. Liu et al. Neurocomputing 458 (2021) 70–86

Martin Buss received the diploma engineer degree in neering and Information Technology, Technical University of Munich, Germany. He
Electrical Engineering in 1990 from the Technical has been appointed full professor, head of the control systems group, and deputy
University Darmstadt, Germany, and the Doctor of director of the Institute of Energy and Automation Technology, Faculty IV - Elec-
Engineering degree in Electrical Engineering from the trical Engineering and Computer Science, Technical University Berlin, Germany,
University of Tokyo, Japan, in 1994. In 2000 he finished from 2000 to 2003. Since 2003 he is full professor (chair) at the Institute of Auto-
his habilitation in the Department of Electrical Engi- matic Control Engineering, Technical University of Munich, Germany. He is also
neering and Information Technology, Technical Univer- coordinator of various Research Centers as the DFG Cluster of Excellence ‘‘Cognition
sity of Munich, Munich, Germany. In 1988 he was a for Technical Systems CoTeSys” (www.cotesys.org) and the DFG Collaborative
research student at the Science University of Tokyo, Research Center ‘‘High Fidelity Telepresence and Teleaction”. Martin Buss is the
Japan, for one year. As a postdoctoral researcher he author of over 100 articles and papers in journals and conferences. His research
stayed with the Department of Systems Engineering, interests include automatic control, mechatronics, multi-modal human-system
Australian National University, Canberra, Australia, in interfaces, optimization, nonlinear, and hybrid discrete-continuous systems.
May, 1994. From 1995 to 2000 he has been senior research assistant and lecturer at
the Institute of Automatic Control Engineering, Department of Electrical Engi-

86