3416 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 14, NO.

8, AUGUST 2018

New Discretization-Formula-Based Zeroing

Dynamics for Real-Time Tracking Control of
Serial and Parallel Manipulators
Jian Li , Yunong Zhang , Member, IEEE, Shuai Li , Senior Member, IEEE, and Mingzhi Mao

Abstract—Improvement of the real-time performance of ra Actual end-effector’s position vector of serial ma-
tracking control is increasingly desirable. It is a routine nipulator.
for most conventional algorithms that the control input at rd Desired path for serial manipulator.
current time instant is to track the current desired output.
However, lagging errors resulting from computational time J(θ(t)) Jacobian matrix of serial manipulator.
and the fluctuation of the desired output exist for the track- l Length vector of the six legs of Stewart platform.
ing control. Different from conventional algorithms, a look- pa Actual effector’s position vector of Stewart plat-
ahead scheme of zeroing dynamics (ZD) is established in form.
this paper to achieve the real-time tracking control of both pd Desired path for Stewart platform.
serial and parallel manipulators. With the exploitation of
data at current time and that in history, the control inputs
generated by the proposed ZD algorithms never lead to lag- I. INTRODUCTION
ging errors with the source from the inevitable computa-
EAL-TIME tracking has been a fundamental research
tional time. To tackle prediction errors for ZD algorithms, a
new high-precision discretization formula, as an essential
part of ZD algorithms, is presented to confine the predic-
R topic in the control field [1], [2]. Various methods have
been developed as found in literature [3]–[5]. Tracking control
tion error in an ignorable range in comparison with lagging of robots has been widely applied in industry [6], [7]. In [8], the
errors. state feedback linearization was applied to the control of non-
Index Terms—Discretization formula, parallel manipula- holonomic wheeled mobile robots. Furthermore, in [9], a mod-
tors, real-time tracking control, serial manipulators, zeroing ified input–output linearization method using the generalized
dynamics (ZD). inverse of decoupling matrices was proposed for the tracking
control of two-wheeled mobile robots. In [10], a robust motion
NOMENCLATURE controller based on neural networks and backstepping was pro-
posed for a two degree-of-freedom (DOF) low-quality mobile
f (x(t), t) Nonlinear time-varying function vector. robot, which guarantees that the mobile robot follows the desired
H(x(t), t) Jacobian matrix of f (x(t), t). path. Sliding mode control is also a classical method for tracking
ft (x(t), t) Partial derivative of f (x(t), t) with respect to t. control problems [11]–[13]. In [13], practical tracking control
θ Joint angle of serial manipulator. design of robot manipulators with continuous fractional order
g(θ) Forward-kinematics mapping of serial manipu nonsingular terminal sliding mode based on time-delay estima-
lator. tion was studied, which requires no detailed information about
the robot dynamics. Many other advanced methods exist, e.g.,
Manuscript received August 7, 2017; revised November 16, 2017; ac- adaptive learning control [14] and neural networks [15], [16].
cepted December 3, 2017. Date of publication December 7, 2017; date
of current version August 1, 2018. This work was supported in part by Conventional methods established in continuous time usually
the National Natural Science Foundation of China under Grant 61473323 degrade when implemented in discrete time due to the one-step
and Grant 61401385; in part by the Laboratory Open Fund of Sun Yat- delay. It is a common practice that the control input at current
sen University under Grant 20160209; in part by Hong Kong Research
Grants Council Early Career Scheme under Grant 25214015; and in part time instant is to track the current desired output. It is known
by the Departmental General Research Fund of Hong Kong Polytech- that computation takes time while the desired output varies with
nic University under Grant G.61.37.UA7L. Paper no. TII-17-1780. (All time, resulting the existence of lagging errors for the tracking
authors of this paper are jointly of the first authorship and jointly of the
corresponding authorship.) (Corresponding author: Mingzhi Mao.) control. If conventional methods are utilized to the real-time
J. Li, Y. Zhang, and M. Mao are with the School of Information Science tracking problem, the obtained control input at current time
and Technology, Sun Yat-sen University, Guangzhou 510006, China instant assumes an one-step delay. Due to this fact, the actual
(e-mail: [email protected]; [email protected]; zhynong@
mail.sysu.edu.cn). precision of conventional methods is usually in proportion to
S. Li is with the Department of Computing, the Hong Kong Polytechnic the sampling gap [17], [18]. Zeroing dynamics (ZD) or Zhang
University, Hong Kong (e-mail: [email protected]). dynamics named after the inventor is a powerful method for
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. time-varying problems solving in real time [19]–[23], which
Digital Object Identifier 10.1109/TII.2017.2780892 is in part driven by the evolution of recurrent neural networks

1551-3203 © 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.

Authorized licensed use limited to: University of Central Florida. Downloaded on March 05,2022 at 01:22:46 UTC from IEEE Xplore. Restrictions apply.
LI et al.: NEW DISCRETIZATION-FORMULA-BASED ZEROING DYNAMICS FOR REAL-TIME TRACKING CONTROL 3417

[24]–[26]. In practice, time-varying problems may be solved by 3) Based on a new discretization formula, two stable and
digital computation, and the discrete-time version of ZD is de- high-precision ZD algorithms are proposed, and they are
veloped in [17]. The 1-step-ahead discretization-formula-based compared with algorithms using other discretization for-
discrete-time ZD has the ability to predict the future solution at mulas.
the next time instant. The ZD method is presented in the ensu-
ing section. Note that 1-step-ahead discrete-time ZD methods II. ZEROING DYNAMICS
have been widely applied to various time-varying problems
In this section, ZD (specifically, discrete-time ZD) is intro-
solving [23], [27], [28]. For example, in [27], a three-step
duced for further investigation. Note that ZD has solved vari-
discrete-time algorithm based on ZD method was developed to
ous discrete time-varying problems effectively [17], [18], [27],
solve time-varying nonlinear systems. In [28], the time-varying
and particularly, the problem of discrete time-varying nonlinear
quadratic programming was solved by utilizing ZD method.
equations is taken as an example [27]. Specifically, the follow-
Considering the superiority of ZD method to predict future
ing problem is considered, which is to be solved at each time
solutions, this work applies ZD to the real-time tracking control
interval [tk , tk +1 ) ⊆ [0, +∞):
of both serial and parallel manipulators. The corresponding ZD
algorithms are proposed to predict the future control input at arg f (xk +1 , tk +1 ) = 0 ∈ Rn (1)
next time instant by using current and past information. As a x k + 1 ∈Rn

unique feature of this proposed method, the time consumption where f (·, ·) : Rn × [0, +∞) → Rn is time-varying, differen-
for computation does not lead to lagging errors. Considering tiable, and nonlinear for all tk > 0 with updating index k =
the existence of prediction errors, a new high-precision 1-step- 0, 1, . . .. Besides, we denote tk = kτ and xk +1 = x(tk +1 ) =
ahead discretization formula is presented and analyzed to obtain x((k + 1)τ ) with τ denoting the sampling gap. During each time
much smaller prediction errors than lagging errors. Note that the interval [tk , tk +1 ), only the current and past discrete informa-
new discretization formula has a truncation error being O(τ 3 ), tion, e.g., f (xk , tk ), f (xk −1 , tk −1 ), xk , and xk −1 , is available,
and the precision of corresponding ZD algorithms is O(τ 4 ), and the future information, e.g., f (xk +1 , tk +1 ), is unknown,
which is significantly better than that of conventional methods. which means that we need to predict the future solution xk +1
Note that based on the theorems in [18], the precision of con- during each time interval [tk , tk +1 ).
ventional methods is O(τ ) for the real-time tracking control. To solve problem (1), it is first converted into a continuous-
The ZD method has been applied to robot motion control in lit- time form as
erature [17], [29]. For example, in [29], a two-link serial robot
manipulator was controlled to draw a four-leaf-clover. Note that arg f (x(t), t) = 0 ∈ Rn . (2)
x(t)∈Rn
those existing algorithms are first derived on the bases of an
equation at velocity level, and then the idea of ZD is employed To zero out f (x(t), t), the following ZD design formula is em-
to solve that equation. Because of the indirect use of ZD, the ini- ployed [19], [27], [28]:
tial errors of those algorithms must be sufficiently small, which
is less practical in real applications. Different from existing al- ḟ (x(t), t) = −λf (x(t), t)
gorithms, the proposed ZD algorithms in this paper directly where ḟ (x(t), t) is the first-order time derivative of f (x(t), t).
employ the idea of ZD to force the actual trajectory to converge Then, a continuous-time ZD model solving problem (2) is ob-
to the desired path, allowing the initial errors to be arbitrarily tained as
large for the proposed ZD algorithms. Moreover, considering
the same discretization formula, ZD algorithms in this paper ẋ(t) = −H −1 (x(t), t) (λf (x(t), t) + ft (x(t), t)) (3)
outperform existing ones in terms of precision.
where
The remainder of this paper is organized into five sections.
∂f (x(t), t)
Section II introduces the ZD method. In Sections III and IV, the H(x(t), t) = ∈ Rn ×n
ZD method is applied to the real-time tracking control of serial ∂xT
and parallel manipulators, respectively, with the corresponding ∂f (x(t), t)
theoretical analyses provided. For substantiating the effective- ft (x(t), t) = ∈ Rn
∂t
ness and superiority of proposed ZD algorithms, numerical ex-
periments for the tracking control of both serial and parallel with the superscript T denoting the transpose of a vector or
manipulators are conducted in Section V. Finally, Section VI matrix. To solve original discrete-time problem (1), a usable
concludes this paper with final remarks. The main contributions 1-step-ahead discretization formula is necessary to discretize
of this paper are shown as below. continuous-time model (3). There are only a few usable dis-
1) Different from conventional methods, the look-ahead ZD cretization formulas so far [30], including conventional Euler
method with an effective prediction ability is presented forward discretization formula [31]
for the first time to address real-time tracking control . xk +1 − xk
ẋk = (4)
problem. τ
2) The look-ahead ZD method developed in this work is fur- and recently proposed Taylor-Zhang discretization formula [32]:
ther investigated to solve serial and parallel manipulator
control, which demonstrates the outstanding performance . 2xk +1 − 3xk + 2xk −1 − xk −2
ẋk = (5)
in terms of extendibility and applicability. 2τ

Authorized licensed use limited to: University of Central Florida. Downloaded on March 05,2022 at 01:22:46 UTC from IEEE Xplore. Restrictions apply.
3418 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 14, NO. 8, AUGUST 2018

.
where “=” denotes the computational assignment operation. is real time. This problem is mathematically formulated as the
Note that Taylor–Zhang discretization formula is also termed form of (1) as
Taylor discretization formula [32], of which the former is more
arg rak +1 − rdk +1 = 0. (9)
scientific, as the formula was proposed by Zhang et al. [32] θ k + 1 ∈Rn
using Taylor expansion iteratively for many times. The trun-
cation errors of Euler forward discretization formula (4) and B. Newton Iteration and Conventional
Taylor–Zhang discretization formula (5) are O(τ ) and O(τ 2 ), Discretization-Formulas-Based ZD Algorithms
respectively. By employing the above two discretization formu-
To solve the real-time tracking control problem (9), classi-
las to continuous-time model (3), the following two algorithms
cal Newton iteration (NI) method [31] is considered, and the
can be obtained:
corresponding NI algorithm is presented as
.  
xk +1 = −H −1 (xk , tk ) (hf (xk , tk ) + τ ft (xk , tk )) + xk (6) .
θk +1 = −J + (θk ) g(θk ) − rdk + θk (10)
and
where J(θ(t)) is the Jacobian matrix of the serial manipulator
. defined as J(θ(t)) = ∂g(θ(t))/∂θ(t) ∈ Rm ×n . As mentioned
xk +1 = − H −1 (xk , tk ) (hf (xk , tk ) + τ ft (xk , tk ))
in the Introduction part, NI algorithm (10) as a conventional
3 1
+ xk − xk −1 + xk −2 (7) method has a precision in proportion to the sampling gap, i.e.,
2 2 O(τ ) [17], [18].
where h = λτ . By focusing on the above algorithms, it can Furthermore, the ZD method introduced in the above section
be found that the solution at tk +1 is predicted and obtained is also employed to solve the real-time tracking control prob-
by previous information without using f (xk +1 , tk +1 ), which lem (9). Specifically, it is first converted into a continuous-time
means that problem (1) can be solved in real time. Besides, it is form as
provable that the truncation errors of algorithms (6) and (7) are arg ra (t) − rd (t) = 0.
O(τ 2 ) and O(τ 3 ), respectively, where O(τ 2 ) denotes a vector θ (t)∈Rn
with every entries being O(τ 2 ).
Second, a vector-form error function is defined as
III. ZD FOR SERIAL MANIPULATOR TRACKING CONTROL e(t) = ra (t) − rd (t) (11)
From the above section, it can be concluded that the ZD has and we employ ZD design formula [19], [27]
the ability of prediction to solve time-varying problems, and
thus, it can solve them in real time. In this section, the tracking ė(t) = −λe(t) (12)
control of serial manipulator is considered as a time-varying to zero out e(t) with t → +∞, and by combining (8), (11), and
problem, and ZD is employed to solve this problem. Moreover, (12), the continuous-time ZD model is obtained as
a new discretization formula is presented, which has higher   
precision then Euler forward and Taylor–Zhang discretization θ̇(t) = J + (θ(t)) ṙd (t) − λ g(θ(t)) − rd (t) (13)
formulas. Based on the new formula, a new algorithm is pro-
where ṙd (t) is the desired velocity vector of the end-effector; and
posed to solve the tracking control problem in real time with
the superscript + denotes the pseudo-inversion of a matrix. Then,
higher precision.
Euler forward discretization formula (4) and Taylor–Zhang dis-
cretization formula (5) are employed to discretize continuous-
A. System Modeling and Problem Formulation time ZD model (13). The following ZD-I algorithm based on
The kinematic equation for serial manipulator is depicted as Euler forward discretization formula (4) is obtained:
follows [29], [33]: .   
θk +1 = J + (θk ) τ ṙdk − h g(θk ) − rdk + θk . (14)
g(θk ) = rak (8)
Based on recently proposed Taylor–Zhang discretization for-
n m
where g(·) : R → R is a continuous nonlinear forward- mula (5), the ZD-II algorithm is obtained as
kinematics mapping of serial manipulator; θk = θ(tk ) ∈ Rn .   
θk +1 = J + (θk ) τ ṙdk − h g(θk ) − rdk
denotes joint angle at instant tk ; and rak ∈ Rm denotes the end-
effector’s position and orientation vector at time instant tk . The 3 1
+ θk − θk −1 + θk −2 . (15)
objective of the real-time tracking control of serial manipu- 2 2
lator is to predict and obtain θk +1 during each time interval
[tk , tk +1 ) ⊆ [0, tf ] so that C. New Discretization Formula and Corresponding ZD
Algorithm
rak +1 → rdk +1
Discretization formula plays a significant role on the perfor-
where rdk +1 (the value of desired path at tk +1 ) is unknown mance of ZD algorithms, and it is always a challenging task
during [tk , tk +1 ). Note that obtaining θk +1 before tk +1 gives the to find a new discretization formula [30], [32]. First, to sat-
manipulator time to control and move, and when tk +1 comes, isfy the requirement of real-time control, a usable discretiza-
the end-effector has moved to rak +1 . Thus, the tracking control tion formula must be 1-step-ahead (i.e., approximating ẋk by

Authorized licensed use limited to: University of Central Florida. Downloaded on March 05,2022 at 01:22:46 UTC from IEEE Xplore. Restrictions apply.
LI et al.: NEW DISCRETIZATION-FORMULA-BASED ZEROING DYNAMICS FOR REAL-TIME TRACKING CONTROL 3419

xk +1 , xk , xk −1 , . . .). Second, for the effectiveness of control, expressed as

ZD algorithm must be stable while many discretization formu- .  
θk +1 = J + (θk ) (1 + h)rdk − rdk −1 − hg(θk ) + θk .
las generate unstable algorithms, which are useless in practice.
Moreover, the precision of discretization formula determines The value of ṙdk in ZD-II algorithm (15) can be approximated
the precision of corresponding ZD algorithm. by the following formula[31]:
In this section, a new 1-step-ahead discretization formula
. 3y(tk ) − 4y(tk −1 ) + y(tk −2 )
is presented, which leads to a stable ZD algorithm. Its trun- ẏ(tk ) = (19)
2τ
cation error is O(τ 3 ) while truncation errors of conventional
Euler forward discretization formula (4) and recently proposed and ZD-II algorithm with time-derivative information unknown
Taylor–Zhang discretization formula (5) are O(τ ) and O(τ 2 ), can be further expressed as
  
respectively. The new discretization formula is presented by the . + 3 1 d
following theorem. θk +1 = J (θk ) + h rk − 2rk −1 + rk −2 − hg(θk )
d d
2 2
Theorem 1 (New discretization formula): The new dis-
3 1
cretization formula is formulated as + θk − θk −1 + θk −2 .
2 2
. 4 1 1
ẋ(tk ) = x(tk +1 ) + x(tk ) − x(tk −1 ) The value of ṙdk in ZD-III algorithm (17) can be approximated
9τ 18τ 3τ
by the following formula [31]:
5 1
− x(tk −2 ) + x(tk −3 ) (16) . 11y(tk ) − 18y(tk −1 ) + 9y(tk −2 ) − 2y(tk −3 )
18τ 9τ ẏ(tk ) = (20)
6τ
of which the truncation error is O(τ 3 ).
Proof: The proof process is given in Appendix A.  and ZD-III algorithm with time-derivative information unknown
By discretizing continuous-time ZD model (13) using new can be further expressed as
discretization formula (16), ZD-III algorithm is proposed as . 9   1
θk +1 = J + (θk ) r̃dk − hg(θk ) − θk
. 9    4 8
θk +1 = J + (θk ) τ ṙdk − h g(θk ) − rdk 3 5 1
4 + θk −1 + θk −2 − θk −3
1 3 5 1 4 8 4
− θk + θk −1 + θk −2 − θk −3 . (17)
8 4 8 4 where r̃dk = (11/6 + h) rdk − 3rdk −1 + 3/2rdk −2 − 1/3rdk −3 .
Note that the truncation error of ZD-III algorithm (17) is O(τ 4 ),
D. Theoretical Analyses
as to be proved theoretically in the ensuing section. During the
control process, the control input is θ, which can be converted to Theoretical analyses are based on four fundamental results for
pulses per second (PPS) signal to control the robot manipulator. multistep methods for the stability and convergence properties
It is known that θk +1 is to track the value of desired path at instant and criteria of multistep formulas [34]–[36]. The relative results
tk +1 . Calculating θk +1 does not need the information at tk +1 are presented in Appendix B. Based on these results, the fol-
for ZD-III algorithm (17). Thus, algorithm (17) is implemented lowing theorems are provided. Specifically, about the stability
during the time interval [tk , tk +1 ) when the sampling gap τ of ZD-III algorithm, the following theorem holds.
is larger than the computing time per updating, and the end- Theorem 2 (Stability of ZD-III Algorithm): The new dis-
effector has successfully tracked the desired path at tk +1 with cretization formula (16)-based ZD-III algorithm (with time-
θk +1 obtained already when instant tk +1 arrives. The whole derivative information known or unknown) is 0-stable.
control process is thus real time. Proof: The new discretization formula (16)-based ZD-III al-
Focusing on the three ZD algorithms (14), (15), and (17), gorithm (with time-derivative information known or unknown)
it can be found that, to solve for θk +1 during time interval can be expressed as a multistep method, of which the character-
[tk , tk +1 ), the time derivative of the desired path (i.e., the value istic polynomial is
of ṙdk ), which is unknown in some applications but is needed. 1 3 5 1
Considering this situation, backward discretization formulas can P4 (η) = η 4 + η 3 − η 2 − η + .
8 4 8 4
be used for approximation. To achieve both the precision and
simplicity, the truncation error of backward discretization for- It has four roots (i.e., η1,2 = −0.7160 ± 0.5495i, η3 = 0.3069
mula approximating the value of ṙdk should be equal to that of and η4 = 1). It is evident that |η1,2,3 | < 1 and η4 = 1. Thus,
1-step-ahead discretization formula approximating the value of based on Result 1 in Appendix B, it can be proved that ZD-III
ẋk . Thus, the value of ṙdk in ZD-I algorithm (14) can be approx- algorithm is 0-stable. The proof is thus completed. 
imated by the following formula: About the convergence of ZD-III algorithm, the following
theorem holds.
. y(tk ) − y(tk −1 ) Theorem 3 (Convergence of ZD-III Algorithm): The new
ẏ(tk ) = (18)
τ discretization formula (16)-based ZD-III algorithm (with
of which the truncation error is equal to that of Euler forward time-derivative information known or unknown) is consistent
discretization formula (4) being O(τ ) [31], and then, ZD-I algo- and convergent, which converges with the truncation error
rithm with time-derivative information unknown can be further order of O(τ 4 ).

Authorized licensed use limited to: University of Central Florida. Downloaded on March 05,2022 at 01:22:46 UTC from IEEE Xplore. Restrictions apply.
3420 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 14, NO. 8, AUGUST 2018

Proof: The proof process is divided into two parts, i.e., with control problem (9) is O(τ 4 ), where  · 2 denotes the two-norm
time-derivative information known and unknown. of a vector.
1) With time-derivative information known: To find the trun- Proof: Let θk∗ +1 be an exact solution of problem (9). We have
cation error of ZD-III algorithm with time-derivative informa- g(θk∗ +1 ) = rdk +1 for all k values. Thus
tion known, the truncation error of new diacretization formula
(16) is considered, which is O(τ 3 ), and thus the following equa- lim sup rak +1 − rdk +1 2 = lim sup g(θk +1 ) − g(θk∗ +1 )2 .
k →+∞ k →+∞
tion is obtained:
From Result 4 in Appendix B, Theorems 2 and 3, it can be
4 1 1
ẋ(tk ) = x(tk +1 ) + x(tk ) − x(tk −1 ) concluded that θk +1 = θk∗ +1 + O(τ 4 ) when k is large enough.
9τ 18τ 3τ Thus, with J(θk∗ +1 ) = ∂g(θk∗ +1 )/∂θk∗ +1 , it is derived that
5 1
− x(tk −2 ) + x(tk −3 ) + O(τ 3 ). (21) lim sup g(θk +1 ) − g(θk∗ +1 )2
18τ 9τ k →+∞
By using (21) to directly discretize continuous-time ZD model
= lim sup g(θk∗ +1 + O(τ 4 )) − g(θk∗ +1 )2
(13), the equation k →+∞

9 +    = lim sup g(θk∗ +1 ) + J(θk∗ +1 )O(τ 4 ) + O(τ 8 ) − g(θk∗ +1 )2

θk +1 = J (θk ) τ ṙdk − h g(θk ) − rdk k →+∞
4
1 3 5 1 = lim sup J(θk∗ +1 )O(τ 4 )2
− θk + θk −1 + θk −2 − θk −3 + O(τ 4 ) (22) k →+∞
8 4 8 4
is obtained. Thus, the truncation error of ZD-III algorithm with = O(τ 4 )
time-derivative information known is O(τ 4 ). Moreover, it is where  · F denotes the matrix Frobenius norm. Thus, the
consistent of order 4. Based on Results 2–4 in Appendix B, MSSTE of ZD-III algorithm (with time-derivative information
ZD-III algorithm with time-derivative information known con- known or unknown) to solve real-time tracking control problem
verges with truncation error order O(τ 4 ). (9) is O(τ 4 ). The proof is thus completed.
2) With time-derivative information unknown: Since the trun-
cation error of backward discretization formula (20) is O(τ 3 ), IV. ZD FOR PARALLEL MANIPULATOR TRACKING CONTROL
we have
In this section, ZD method is employed to the real-time track-
11 3 3
ẏ(tk ) = y(tk ) − y(tk −1 ) + y(tk −2 ) ing control of Stewart platform [37], which is a typical parallel
6τ τ 2τ
robot manipulator.
1
− y(tk −3 ) + O(τ 3 ).
3τ A. System Modeling and Problem Formulation
By using the above equation to directly discretize ṙdk in (22), the Stewart platform contains a mobile platform and a fixed base
equation connected together by six independent prismatic legs. The effec-
9 +   1 tor is set on center of the mobile platform. The tracking control
θk +1 = J (θk ) r̃dk − hg(θk ) + O(τ 4 ) − θk of Stewart platform is to change the lengths of the six legs, so
4 8
that the effector could track a desired path. The objective of the
3 5 1
+ θk −1 + θk −2 − θk −3 + O(τ 4 ) real-time tracking control of Stewart platform is to predict and
4 8 4 obtain lk +1 ∈ R6 during each time interval [tk , tk +1 ) ⊆ [0, tf ]
9 +  d  1 so that
= J (θk ) r̃k − hg(θk ) − θk
4 8
pak +1 → pdk +1
3 5 1
+ θk −1 + θk −2 − θk −3 + O(τ 4 ) where lk = [lk1 , lk2 , lk3 , lk4 , lk5 , lk6 ]T is a vector containing the
4 8 4
lengths of six legs at time instant tk ; pak +1 ∈ R3 is the actual
is obtained. Thus, the truncation error of ZD-III algorithm with
position of the effector; and pdk +1 ∈ R3 is the desired path to be
time-derivative information unknown is O(τ 4 ). Moreover, it
tracked, which is unknown during each time interval [tk , tk +1 ).
is consistent of order 4. Similarly, based on Results 2–4 in
This problem is also mathematically formulated as the form of
Appendix B, ZD-III algorithm with time-derivative informa-
(1) as
tion unknown converges with truncation error order O(τ 4 ). The
proof is thus completed.  arg pak +1 − pdk +1 = 0. (23)
About the tracking error of ZD-III algorithm, the following l k + 1 ∈R6

theorem holds.
Theorem 4 (Steady-State Tracking Error of ZD-III Algo- B. ZD Algorithms With Theoretical Analyses
rithm): Maximal steady-state tracking error (MSSTE) Following the design process of ZD algorithms, this problem
lim k →+∞ sup rak +1 −rdk +1 2 of the new discretization is first converted into a continuous-time form as
formula (16)-based ZD-III algorithm (with time-derivative arg pa (t) − pd (t) = 0.
information known or unknown) to solve real-time tracking l(t)∈R6
(24)

Authorized licensed use limited to: University of Central Florida. Downloaded on March 05,2022 at 01:22:46 UTC from IEEE Xplore. Restrictions apply.
LI et al.: NEW DISCRETIZATION-FORMULA-BASED ZEROING DYNAMICS FOR REAL-TIME TRACKING CONTROL 3421

An error function is defined as Corollary 2 (Convergence of ZD-VI Algorithm): The new

discretization formula (16)-based ZD-VI algorithm (with
(t) = p (t) − p (t)
a d
(25)
time-derivative information known or unknown) is consistent
and the ZD design formula and convergent, which converges with the truncation error
order of O(τ 4 ).
(t)
˙ = −λ(t)
Proof: The proof process is similar to that of Theorem 3, and
makes error function (25) tend to zero with t → +∞. Thus, the thus is omitted here.
equation About the tracking error of ZD-VI algorithm, the following
corollary holds.
ṗa (t) − ṗd (t) = −λ(pa (t) − pd (t)) (26)
Corollary 3 (Steady-State Tracking Error of ZD-VI
is obtained. Based on the robot kinematic modeling of Stewart Algorithm): MSSTE limk →+∞ sup pak +1 −pdk +1 2 of
platform, the kinematic equation of the robot system at velocity the new discretization formula (16)-based ZD-VI algorithm
level is depicted as [37] (with time-derivative information known or unknown) to solve
real-time tracking control problem (23) is O(τ 4 ).
l̇(t) = C(l(t), D(t))ṗa (t) (27)
Proof: The proof process is similar to that of Theorem 4, and
where C(l(t), D(t)) ∈ R6×3 is a matrix related to the six legs’ thus is omitted here.
lengths (i.e., l(t)) and the position matrix in global coordi-
nate (i.e., D(t)), and D(t) = [d1 (t), d2 (t), d3 (t), d4 (t), d5 (t), V. NUMERICAL EXPERIMENTS
d6 (t)] ∈ R3×6 , where dj (t) ∈ R3 (j = 1, 2, . . . , 6) is the po- In this section, two examples are provided to show the ef-
sition of jth leg. Besides, l̇(t) is velocity vector of six legs, fectiveness and superiority of the proposed ZD algorithms. The
and ṗa (t) is the actual velocity of the effector on the mobile first example is to show the tracking control of serial manipula-
platform. Substituting (26) into (27) yields continuous-time ZD tor, and the other one is to show the tracking control of parallel
model for solving continuous-time problem (24) as manipulator. They substantiate the theoretical results presented
l̇(t) = C(l(t), D(t))(ṗd (t) − λ(pa (t) − pd (t))). in above sections for real-time tracking control of serial and
parallel manipulators.
Furthermore, by using conventional Euler forward discretiza-
tion formula (4) for the discretization, the following ZD-IV A. Example 1: 5-Link Serial Manipulator Control
algorithm is obatined:
. A 5-link planar serial robot manipulator is considered, of
lk +1 = C(lk , Dk )(τ ṗdk − h(pak − pdk )) + lk . (28) which the geometry is shown in Fig. 1(a). For simplicity and
By using recently proposed Taylor–Zhang discretization for- illustration, the length of each links is 1 m, and then, the forward-
mula (5), the ZD-V algorithm is obatined as kinematics mapping g(θk ) ∈ R2 of the serial robot manipulator
. can be obtained, where θk = [θk1 , θk2 , θk3 , θk4 , θk5 ]T ∈ R5 . The se-
lk +1 = C(lk , Dk )(τ ṗdk − h(pak − pdk )) rial robot manipulator is investigated to track the desired path
3 1 at tk +1 via the information at tk expressed as below during the
+ lk − lk −1 + lk −2 . (29) time interval [tk , tk +1 )
2 2
 
By using new discretization formula (16), the ZD-VI algorithm 0.5 sin3 (πtk /5) + 1.5
is obatined as rdk = rd (tk ) = .
0.5 cos3 (πtk /5) + 2.5
. 9
lk +1 = C(lk , Dk )(τ ṗdk − h(pak − pdk )) Note that the expression of desired path for each instant
4
tk is given just for the convenience of comparison, and fu-
1 3 5 1 ture information rdk +1 cannot be used during the time inter-
− lk + lk −1 + lk −2 − lk −3 . (30)
8 4 8 4 val [tk , tk +1 ). The proposed ZD-III algorithm is employed
Similar to that of serial manipulator, ZD-IV/V/VI algorithms to solve this problem. The initial joint state is arbitrarily
also need the information of time derivative (i.e., the value of set as θ0 = [3π/4, −π/2, −π/4, π/6, π/3]T rad; task duration
ṗdk ). Thus, in this situation, the backward discretization formulas tf = 20 s; parameters of ZD-III algorithm h = 0.1.
(18)–(20) can be employed to discretize the value of ṗdk . With τ = 0.1 s and the known time-derivative information
Corresponding theoretical analyses are presented by the fol- being considered, numerical results are presented in Fig. 1(b)
lowing corollaries. About the stability of ZD-VI algorithm, the and (c). Specifically, Fig. 1(b) shows the trajectory of the serial
following corollary holds. robot manipulator, and Fig. 1(c) shows the actual trajectory of
Corollary 1 (Stability of ZD-VI Algorithm): The new dis- the end-effector corresponding to the blue dotted line, which
cretization formula (16)-based ZD-VI algorithm (with time- almost overlaps with the desired path corresponding to the red
derivative information known or unknown) is 0-stable. dash line in a short time. Thus, the effectiveness of the proposed
Proof: The proof process is similar to that of Theorem 2, and ZD-III algorithm is substantiated to solve the real-time tracking
thus is omitted here. control problem of serial manipulators.
About the convergence of ZD-VI algorithm, the following To substantiate the precision of ZD-III algorithm being
corollary holds. O(τ 4 ), which is proved theoretically in Theorem 4, numerous

Authorized licensed use limited to: University of Central Florida. Downloaded on March 05,2022 at 01:22:46 UTC from IEEE Xplore. Restrictions apply.
3422 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 14, NO. 8, AUGUST 2018

Fig. 1. Geometry of 5-link planar serial robot manipulator, and trajectories of robot and end-effector for tracking desired path in Example 1 with
τ = 0.1 s. (a) Geometry of robot manipulator. (b) Trajectories of robot. (c) Trajectories of end-effector, where the blue dotted line corresponds the
actual trajectory and the red dash line corresponds the desired path.

TABLE I
MSSTE AND ACTPU DATA WHEN USING NI AND ZD-I/-II/-III ALGORITHMS WITH DIFFERENT τ VALUES IN THE SITUATION WITH TIME-DERIVATIVE
INFORMATION KNOWN

τ (ms) NI ZD-I ZD-II ZD-III

MSSTE (m) ACTPU (ms) MSSTE (m) ACTPU (ms) MSSTE (m) ACTPU (ms) MSSTE (m) ACTPU (ms)

100 4.725 × 10−2 0.1090 1.768 × 10−2 0.1285 1.349 × 10−3 0.1157 1.869 × 10−4 0.1238
50 2.359 × 10−2 0.1039 5.960 × 10−3 0.1163 2.490 × 10−4 0.1075 1.846 × 10−5 0.1106
10 4.714 × 10−3 0.1017 2.924 × 10−4 0.1047 2.670 × 10−6 0.1044 4.328 × 10−8 0.1204
5 2.356 × 10−3 0.1043 7.379 × 10−5 0.1066 3.383 × 10−7 0.1175 2.763 × 10−9 0.1163
1 4.712 × 10−4 0.1027 2.960 × 10−6 0.1058 2.718 × 10−9 0.1057 4.454 × 10−12 0.1044
0.5 2.356 × 10−4 0.1030 7.402 × 10−7 0.1042 3.398 × 10−10 0.1055 2.824 × 10−13 0.1053

TABLE II
MSSTE AND ACTPU DATA WHEN USING NI AND ZD-I/-II/-III ALGORITHMS WITH DIFFERENT τ VALUES IN THE SITUATION WITH TIME-DERIVATIVE
INFORMATION UNKNOWN

τ (ms) NI ZD-I ZD-II ZD-III

MSSTE (m) ACTPU (ms) MSSTE (m) ACTPU (ms) MSSTE (m) ACTPU (ms) MSSTE (m) ACTPU (ms)

100 4.725 × 10−2 0.1090 3.486 × 10−2 0.1303 2.909 × 10−3 0.1204 3.659 × 10−4 0.1239
50 2.359 × 10−2 0.1039 1.186 × 10−2 0.1213 5.342 × 10−4 0.1189 3.577 × 10−5 0.1262
10 4.714 × 10−3 0.1017 5.850 × 10−4 0.1132 5.631 × 10−6 0.1273 8.077 × 10−8 0.1332
5 2.357 × 10−3 0.1043 1.476 × 10−4 0.1224 7.128 × 10−7 0.1216 5.137 × 10−9 0.1253
1 4.713 × 10−4 0.1027 5.921 × 10−6 0.1145 5.726 × 10−9 0.1211 8.283 × 10−12 0.1371
0.5 2.356 × 10−4 0.1030 1.480 × 10−6 0.1141 7.158 × 10−10 0.1183 5.199 × 10−13 0.1212

numerical experiments are conducted by employing ZD-III the control process is real time even when sampling gap τ is up
algorithm with different values of τ . For comparison, NI, ZD-I, to 0.5 ms.
and ZD-II algorithms are also employed to solve the above
tracking control problem. Numerical results are displayed in
Tables I and II. B. Example 2: 6 DOF Parallel Manipulator Control
Specifically, from Table I, with τ = 0.1 s and time-derivative In this example, the Stewart platform, which is a 6 DOF par-
known, MSSTEs of NI, ZD-I, ZD-II, and ZD-III algorithms allel manipulator, is considered, of which the geometry is shown
are of order 10−2 , 10−2 , 10−3 , and 10−4 , respectively. With τ in Fig. 2(a). We aim at predicting and obtaining lk +1 ∈ R6 dur-
reduced by a factor of 10 to be 0.01 s, their MSSTEs decrease ing each time interval [tk , tk +1 ) ⊆ [0, tf ] so that the effector
by a factor of 10, 102 , 103 , and 104 , respectively. Similarly, with tracks the desired path, which is expressed under cylindrical
τ reduced by a factor of 100 to be 0.001 s, MSSTEs decrease by coordinate as
a factor of 102 , 104 , 106 , and 108 , respectively. ZD-I, ZD-II, and
ZD-III algorithms with time-derivative unknown also perform ⎧
⎪
⎨ ε = 0.6tk
well as observed from Table II. Moreover, from the average
computing time per updating (ACTPU) data in Tables I and II, r = 0.2 − (0.2 sin(2.5ε))2
⎪
⎩ z = (r sin(2.5ε))2 + 1.
they are around 0.1 ms in different situations, which means that

Authorized licensed use limited to: University of Central Florida. Downloaded on March 05,2022 at 01:22:46 UTC from IEEE Xplore. Restrictions apply.
LI et al.: NEW DISCRETIZATION-FORMULA-BASED ZEROING DYNAMICS FOR REAL-TIME TRACKING CONTROL 3423

Fig. 2. Geometry of Stewart platform, and trajectories of effector for tracking desired path in Example 2 with τ = 0.01 s. (a) Geometry of Stewart
platform. (b) Trajectories of effector, where the blue dotted line corresponds the actual trajectory and the red dash line corresponds the desired path.

TABLE III VI. CONCLUSION

MSSTES WHEN USING ZD-IV/-V/-VI ALGORITHMS WITH DIFFERENT
VALUES OF τ IN THE SITUATION WITH TIME-DERIVATIVE In this paper, the ZD scheme has been established for the
INFORMATION KNOWN real-time tracking control of serial and parallel manipulators.
For the control of serial manipulators, ZD-I/-II/-III algorithms
τ (ms) ZD-IV ZD-V ZD-VI have been developed based on three discretization techniques,
100 6.0000 × 10−3 1.8000 × 10−3 1.6493 × 10−4 namely the conventional Euler forward discretization formula,
50 2.3000 × 10−3 3.7057 × 10−4 1.9823 × 10−5 the recently proposed Taylor–Zhang discretization formula and
10 1.4751 × 10−4 5.1949 × 10−6 7.1282 × 10−8 the new discretization formula, respectively. Both theoretical
5 3.8206 × 10−5 6.8359 × 10−7 4.8289 × 10−9 analyses and numerical experiments have shown that the ZD-
1 1.5476 × 10−6 5.5814 × 10−9 7.9693 × 10−12
0.5 3.8705 × 10−7 6.9817 × 10−10 4.9917 × 10−13
I/-II/-III algorithms have the precision of O(τ 2 ), O(τ 3 ), and
O(τ 4 ), respectively. For the control of parallel manipulator, ZD-
IV/-V/-VI algorithms, which also bear the precision of O(τ 2 ),
O(τ 3 ), and O(τ 4 ), respectively, are also established in this pa-
TABLE IV per. It is worth pointing out that experimental studies with mea-
MSSTES WHEN USING ZD-IV/-V/-VI ALGORITHMS WITH DIFFERENT
VALUES OF τ IN THE SITUATION WITH TIME-DERIVATIVE surement noises could be an interesting, open, and challenging
INFORMATION UNKNOWN future direction of this work.
APPENDIX A
τ (ms) ZD-IV ZD-V ZD-VI
Based on Taylor expansion [31], the following equations are
100 1.2000 × 10−2 2.8000 × 10−3 3.7284 × 10−4 obtained:
50 4.6000 × 10−3 5.5584 × 10−4 4.5393 × 10−5
10 2.9501 × 10−4 7.7925 × 10−6 1.6294 × 10−7 ẍ(tk ) 2
5 7.6410 × 10−5 1.0254 × 10−6 1.1038 × 10−8 x(tk +1 ) = x(tk ) + ẋ(tk )τ + τ
2!
1 3.0952 × 10−6 8.3721 × 10−9 1.8217 × 10−11
0.5 7.7410 × 10−7 1.0473 × 10−9 1.1429 × 10−12 x(3) (tk ) 3 x(4) (c1 ) 4
+ τ + τ (31)
3! 4!
ẍ(tk ) 2
x(tk −1 ) = x(tk ) − ẋ(tk )τ + τ
Similar to that in Example 1, future information pdk +1 is un- 2!
known and cannot be used during the time interval [tk , tk +1 ). x(3) (tk ) x(4) (c2 ) 4
+ (−τ )3 + τ (32)
Expression of desired path for each time instant tk is given 3! 4!
just for the convenience of comparison. ZD-VI algorithm is ẍ(tk )
employed to solve this problem with task duration tf = 15 s, x(tk −2 ) = x(tk ) − 2ẋ(tk )τ + (2τ )2
2!
h = 0.1, and τ = 0.01 s. Fig. 2(b) shows the actual trajectory
generated by ZD-VI algorithm corresponding to the blue dot- x(3) (tk ) x(4) (c3 )
+ (−2τ )3 + (2τ )4 (33)
ted line, which tracks the desired path corresponding to the red 3! 4!
dash line successfully in a short time. To substantiate theoret- and
ical results of Theorem 7 and show the superiority of ZD-VI ẍ(tk )
algorithm, different τ values are employed for ZD-IV/-V/-VI x(tk −3 ) = x(tk ) − 3ẋ(tk )τ + (3τ )2
2!
algorithms and corresponding numerical results are shown in
Tables III and IV. It is evident that they verify the theoretical x(3) (tk ) x(4) (c4 )
+ (−3τ )3 + (3τ )4 (34)
results. 3! 4!

Authorized licensed use limited to: University of Central Florida. Downloaded on March 05,2022 at 01:22:46 UTC from IEEE Xplore. Restrictions apply.
3424 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 14, NO. 8, AUGUST 2018

where ẍ(tk ) denotes the second-order derivatives of x(t) with [7] J. Baek, M. Jin, and S. Han, “A new adaptive sliding-mode control scheme
respect to t at time instant tk ; x(3) (tk ) and x(4) (tk ) denote the for application to robot manipulators,” IEEE Trans. Ind. Electron., vol. 63,
no. 6, pp. 3628–3637, Jun. 2016.
third-order and fourth-order derivatives, respectively; symbol ! [8] B. D’Andrea-Novel, G. Campion, and G. Bastin, “Control of nonholo-
denotes the factorial operator; c1 , c2 , c3 , and c4 lie in (tk , tk +1 ), nomic wheeled mobile robots by state feedback linearization,” Int. J.
(tk −1 , tk ), (tk −2 , tk ), and (tk −3 , tk ), respectively. Let (31) mul- Robot. Res., vol. 14, no. 6, pp. 543–559, Dec. 1995.
[9] D. H. Kim and J. H. Oh, “Tracking control of a two-wheeled mobile robot
tiply 8, let (32) multiply −3, let (33) multiply −5 and let (34) using input-output linearization,” Control Eng. Practice, vol. 7, no. 3,
multiply 2. Adding together with the results yields pp. 369–373, Mar. 1999.
[10] Q. Zhang, J. Shippen, and B. Jones, “Robust backstepping and neural
4 1 1 network control of a low-quality nonholonomic mobile robot,” Int. J.
ẋ(tk ) = x(tk +1 ) + x(tk ) − x(tk −1 ) Mach. Tools Manuf., vol. 39, no. 7, pp. 1117–1134, Jul. 1999.
9τ 18τ 3τ
[11] M. L. Corradini, V. Fossi, A. Giantomassi, G. Ippoliti, S. Longhi, and G.
5 1 1 Orlando, “Minimal resource allocating networks for discrete time sliding
− x(tk −2 ) + x(tk −3 ) − τ 3 x(4) (c1 ) mode control of robotic manipulators,” IEEE Trans. Ind. Informat., vol. 8,
18τ 9τ 54 no. 4, pp. 733–745, Nov. 2012.
1 5 3 [12] Q. Zhang, L. Lapierre, and X. Xiang, “Distributed control of coordinated
+ τ 3 x(4) (c2 ) + τ 3 x(4) (c3 ) − τ 3 x(4) (c4 ). path tracking for networked nonholonomic mobile vehicles,” IEEE Trans.
72 27 8 Ind. Informat., vol. 9, no. 1, pp. 472–484, Feb. 2013.
The term −1/54τ 3 x(4) (c1 ) + 1/72τ 3 x(4) (c2 ) + 5/27τ 3 x(4) [13] Y. Wang, L. Gu, Y. Xu, and X. Cao, “Practical tracking control of robot ma-
nipulators with continuous fractional-order nonsingular terminal sliding
(c3 ) −3/8τ 3 x(4) (c4 ) can be written as O(τ 3 ), and omitting the mode,” IEEE Trans. Ind. Electron., vol. 63, no. 10, Oct. 2016.
term exactly yields new discretization formula (16). [14] J. H. Yang and S. H. Shen, “Novel approach for adaptive tracking control
of a 3-d overhead crane system,” J. Intell. Robot Syst., vol. 62, no. 1,
pp. 59–80, Apr. 2011.
APPENDIX B [15] Z. Zhang, A. Beck, and N. Magnenat-Thalmann, “Human-like behavior
generation based on head-arms model for robot tracking external targets
We have the following four results for an N -step method and body parts,” IEEE Trans. Cybern., vol. 45, no. 8, pp. 1390–1400, Aug.
[34], [35].  N
2015.
Result 1: An N -step method N i=0 αi xk +i = τ i=0 βi ψk +i
[16] B. Cai and Y. Zhang, “Different-level redundancy-resolution and its equiv-
alent relationship analysis for robot manipulators using gradient-descent
can be checked for 0-stability by determining
 the roots of its and Zhang et al’s neural-dynamic methods,” IEEE Trans. Ind. Electron.,
characteristic polynomial PN (η) = N i
i=0 αi η . If all roots de-
vol. 59, no. 8, pp. 3146–3155, Aug. 2012.
[17] L. Jin and Y. Zhang, “Discrete-time Zhang neural network of O(τ 3 )
noted by η of the polynomial PN (η) satisfy |η| ≤ 1 with |η| = 1 pattern for time-varying matrix pseudoinversion with application to ma-
being simple, then the corresponding N -step method is 0-stable nipulator motion generation,” Neurocomputing, vol. 142, pp. 165–173,
(i.e., has 0-stability). Oct. 2014.
[18] L. Jin and Y. Zhang, “Continuous and discrete Zhang dynamics for real-
Result 2: An N -step method is said to be consistent (i.e., time varying nonlinear optimization,” Numer. Algorithms, vol. 73, no. 1,
have consistency) of order p if the truncation error for the exact pp. 115–140, Sep. 2015.
solution is of order O(τ p+1 ) where p > 0. [19] Y. Zhang and C. Yi, Zhang Neural Networks and Neural-Dynamic Method.
Commack, NY, USA: Nova, 2011.
Result 3: An N -step method is convergent, i.e., x[t/τ ] → [20] L. Jin, Y. Zhang, S. Li, and Y. Zhang, “Modified ZNN for time-
x∗ (t), for all t ∈ [0, tf ], as τ → 0, if and only if the method is 0- varying quadratic programming with inherent tolerance to noises and
stable and consistent. That is, 0-stability plus consistency means its application to kinematic redundancy resolution of robot manipula-
tors,” IEEE Trans. Ind. Electron., vol. 63, no. 11, pp. 6978–6988, Nov.
convergence, which is also known as Dahlquist equivalence 2016.
theorem. [21] B. Cai and X. Jiang, “A novel artificial neural network method for biomed-
Result 4: A 0-stable consistent method converges with the ical prediction based on matrix pseudo-inversion,” J. Biomed. Informat.,
vol. 48, pp. 114–121, Apr. 2014.
order of its truncation error. [22] M. Mao, J. Li, L. Jin, S. Li and Y. Zhang, “Enhanced discrete-time Zhang
neural network for time-variant matrix inversion in the presence of bias
REFERENCES noises,” Neurocomputing, vol. 207, pp. 220–230, Sep. 2016.
[23] K. Chen, “Robustness analysis of Wang neural network for online linear
[1] H. Wang, X. Liu, and K. Liu, “Robust adaptive neural tracking control equation solving,” Electron. Lett., vol. 48, no. 22, pp. 1391–1392, Oct.
for a class of stochastic nonlinear interconnected systems,” IEEE Trans. 2012.
Neural Netw. Learn. Syst., vol. 27, no. 3, pp. 510–523, Mar. 2016. [24] S. Yang, Z. Guo, and J. Wang, “Global synchronization of multiple re-
[2] Y. J. Liu, S. Tong, and C. L. P. Chen, “Adaptive fuzzy control via observer current neural networks with time delays via impulsive interactions,”
design for uncertain nonlinear systems with unmodeled dynamics,” IEEE IEEE Trans. Neural Netw. Learn. Syst., vol. 28, no. 7, pp. 1657–1667,
Trans. Fuzzy Syst., vol. 21, no. 2, pp. 275–288, Apr. 2013. Jul. 2017.
[3] H. Wang, B. Chen, X. Liu, K. Liu, and C. Lin, “Robust adaptive fuzzy [25] Y. Zhang and S. S. Ge, “Design and analysis of a general recurrent neural
tracking control for pure-feedback stochastic nonlinear systems with input network model for time-varying matrix inversion,” IEEE Trans. Neural
constraints,” IEEE Trans. Cybern., vol. 43, no. 6, pp. 2093–2104, Dec. Netw., vol. 16, no. 6, pp. 1477–1490, Nov. 2005.
2013. [26] X. Li and S. Song, “Impulsive control for existence, uniqueness and global
[4] C. L. P. Chen, G. Wen, Y. J. Liu, and Z. Liu, “Observer-based adaptive stability of periodic solutions of recurrent neural networks with discrete
backstepping consensus tracking control for high-order nonlinear semi- and continuously distributed delays,” IEEE Trans. Neural Netw. Learn.
strict-feedback multiagent systems,” IEEE Trans. Cybern., vol. 46, no. 7, Syst., vol. 24, no. 6, pp. 868–877, Jun. 2013.
pp. 1591–1601, Jul. 2016. [27] D. Guo, Z. Nie, and L. Yan, “Theoretical analysis, numerical verification
[5] Y. J. Liu, C. L. P. Chen, G. Wen, and S. Tong, “Adaptive neural output and geometrical representation of new three-step DTZD algorithm for
feedback tracking control for a class of uncertain discrete-time nonlinear time varying nonlinear equations solving,” Neurocomputing, vol. 214,
systems,” IEEE Trans. Neural Netw. Learn. Syst., vol. 22, no. 7, pp. 1162– pp. 516–526, Nov. 2016.
1167, Jul. 2011. [28] P. Miao, Y. Shen, Y. Huang, and Y. Wang, “Solving time-varying quadratic
[6] C. Yang, Y. Jiang, Z. Li, W. He, and C. Su, “Neural control of bimanual programs based on finite-time Zhang neural networks and their application
robots with guaranteed global stability and motion precision,” IEEE Trans. to robot tracking,” Neural Comput. Appl., vol. 26, no. 3, pp. 693–703, Apr.
Ind. Informat., vol. 13, no. 3, pp. 1162–1171, Jun. 2017. 2015.

Authorized licensed use limited to: University of Central Florida. Downloaded on March 05,2022 at 01:22:46 UTC from IEEE Xplore. Restrictions apply.
LI et al.: NEW DISCRETIZATION-FORMULA-BASED ZEROING DYNAMICS FOR REAL-TIME TRACKING CONTROL 3425

[29] L. Jin and Y. Zhang, “Discrete-time Zhang neural network for online Yunong Zhang (S’02–M’03) received the B.S.
time-varying nonlinear optimization with application to manipulator mo- degree in industrial electrical automation from
tion generation,” IEEE Trans. Neural Netw. Learn. Syst., vol. 26, no. 7, Huazhong University of Science and Technol-
pp. 1525–1531, Jul. 2015. ogy, Wuhan, China, in 1996, the M.S. de-
[30] Y. Zhang, Y. Chou, J. Chen, Z. Zhang, and L. Xiao, “Presentation, error gree in control theory and control engineer-
analysis and numerical experiments on a group of 1-step-ahead numerical ing from South China University of Technology,
differentiation formulas,” J. Comput. Appl. Math., vol. 239, pp. 406–414, Guangzhou, China, in 1999, and the Ph.D. de-
Feb. 2013. gree in mechanical and automation engineering
[31] J. H. Mathews and K. D. Fink, Numerical Methods Using MATLAB. from Chinese University of Hong Kong, Shatin,
Englewood Cliffs, NJ, USA: Prentice-Hall, 2005. Hong Kong, China, in 2003.
[32] Y. Zhang, L. Jin, D. Guo, Y. Yin, and Y. Chou, “Taylor-type 1-step-ahead He is currently a professor in the School of In-
numerical differentiation rule for first-order derivative approximation and formation Science and Technology, Sun Yat-sen University, Guangzhou,
ZNN discretization,” J. Comput. Appl. Math., vol. 273, pp. 29–40, Jan. China. Before joining Sun Yat-sen University in 2006, he had been with
2014. National University of Singapore, University of Strathclyde, and National
[33] S. Li, S. Chen, B. Liu, Y. Li, and Y. Liang, “Decentralized kinematic University of Ireland at Maynooth, since 2003. In addition, he is also
control of a class of collaborative redundant manipulators via recurrent currently with the SYSU-CMU Shunde International Joint Research In-
neural networks,” Neurocomputing, vol. 91, pp. 1–10, Aug. 2012. stitute, Foshan, China, for cooperative research. His main research in-
[34] E. Suli and D. F. Mayers, An Introduction to Numerical Analysis. Oxford, terests include robotics, neural networks, computation and optimization.
U.K.: Cambridge Univ. Press, 2003. His web-page is now available at http://sdcs.sysu.edu.cn/content/2477.
[35] D. F. Griffiths and D. J. Higham, Numerical Methods for Ordinary Differ-
ential Equations: Initial Value Problems. London, U.K.: Springer-Verlag, Shuai Li (M’14–SM’17) received the B.E. de-
2010. gree in precision mechanical engineering from
[36] R. J. LeVeque, Finite Difference Methods for Ordinary and Partial Dif- Hefei University of Technology, Hefei, China, in
ferential Equations: Steady-State and Time-Dependent Problems. Seattle, 2005, the M.E. degree in automatic control en-
WA, USA: SIAM, 2007. gineering from the University of Science and
[37] A. M. Mohammed and S. Li, “Dynamic neural networks for kinematic re- Technology of China, Hefei, China, in 2008, and
dundancy resolution of parallel Stewart platforms,” IEEE Trans. Cybern., the Ph.D. degree in electrical and computer en-
vol. 46, no. 7, pp. 1538–1550, Jul. 2016. gineering from Stevens Institute of Technology,
Hoboken, NJ, USA, in 2014.
After graduation, he joined the Department of
Computing, The Hong Kong Polytechnic Univer-
sity, Kowloon, Hong Kong. His current research interests include dynamic
neural networks, wireless sensor networks, robotic networks, machine
learning, and other dynamic problems defined on a graph.
Dr. Li is currently on the editorial board of Neural Computing and Ap-
plications and the International Journal of Distributed Sensor Networks.

Mingzhi Mao received the B.S. degree in com-

Jian Li received the B.E. degree in computer puter science, the M.S. degree in computational
science and technology from Hainan University, mathematics, and the Ph.D. degree in computer
Haikou, China, in 2014. He is currently working software mathematics from Sun Yat-sen Univer-
toward the Ph.D. degree in communication and sity, Guangzhou, China, in 1988, 1998 and 2008,
information systems at the School of Information respectively.
Science and Technology, Sun Yat-sen University, He is currently an associate professor at
Guangzhou, China. the School of Information Science and Technol-
His main research interests include robot, ogy, Sun Yat-sen University, Guangzhou, China.
neural networks, numerical computation, and in- His main research interests include intelligence
telligent information processing. algorithm, software engineering, and manage-
ment information system.

Authorized licensed use limited to: University of Central Florida. Downloaded on March 05,2022 at 01:22:46 UTC from IEEE Xplore. Restrictions apply.