Optimal Design Oriented Adaptive Robust Control

for A Planar 2-DOF Redundantly Actuated Parallel

Robot
Jiang Han
Hefei University of Technology
Peng Wang
Hefei University of Technology School of Mechanical Engineering https://orcid.org/0000-0001-6692-
8520
Fangfang Dong (  [email protected] )
Hefei University of Technology
Xiaomin Zhao
Hefei University of Technology
Shan Chen
Hefei University of Technology

Research Article

Keywords: Dynamic system, Adaptive robust control, Optimal design, System uncertainty, Udwadia-
Kalaba equation

Posted Date: March 9th, 2021

DOI: https://doi.org/10.21203/rs.3.rs-245881/v1

License:   This work is licensed under a Creative Commons Attribution 4.0 International License.
Read Full License

Version of Record: A version of this preprint was published at Nonlinear Dynamics on July 29th, 2021.
See the published version at https://doi.org/10.1007/s11071-021-06739-y.
Noname manuscript No.
(will be inserted by the editor)

Optimal Design Oriented Adaptive Robust Control

for A Planar 2-DOF Redundantly Actuated Parallel
Robot

Jiang Han1,2 · Peng Wang1,2 · Fangfang

Dong1,2 · Xiaomin Zhao3 · Shan Chen1,2

Received: date / Accepted: date

Abstract An adaptive robust control combined with a multi-objective pa-

rameter optimization method for the parallel robot with unknown uncertainty
is proposed. In the active joint space, the accurate dynamic model of the par-
allel robot can be obtained by combining the closed-chain constraint force and
the open-chain system’s dynamic equation. According to the Udwadia-Kalaba
theory, the closed-chain constraint force imposed by the end effector can be
calculated in a simple way. The proposed adaptive robust control could guar-
antee deterministic robust performances of the system (the uniform bounded-
ness and uniform ultimate boundedness). To seek suitable weighting factors
for the proposed control, a system performance function, which includes the
transient performance portion, the steady state performance portion, and the
control cost, is introduced. By applying D-operation, the performance function
is transformed into a multi-objective function with weighting factors. Mean-
while, the problem of choosing the optimal gain is equivalent to the problem
of finding the minimum value of the system performance index function. An il-
lustrative example illustrates the superiority of the proposed modeling method
and the proposed adaptive robust control.

Keywords Dynamic system · Adaptive robust control · Optimal design ·

System uncertainty · Udwadia-Kalaba equation.

J. Han · P. Wang · F. Dong · S. Chen

1. School of Mechanical Engineering, Hefei University of Technology, Hefei, Anhui, 230009,
China.
2. Anhui Engineering Laboratory of Intelligent CNC Technology and Equipment
E-mail: Corresponding author, [email protected]
X.Zhao
3. School of Automotive and Transportation Engineering, Hefei University of Technology,
Hefei, Anhui, 230009, China.
2 Jiang Han1,2 et al.

1 Introduction

Compared with the traditional serial robot, the parallel robot is a closed-loop
motion structure with multiple kinematic chains, which makes the kinematic
and dynamic analysis of the parallel robot tough. But it has superiorities that
traditional serial robots do not have, such as high rigidity, low inertia, and so
on. Because of the above high performances, it has been widely used in many
fields of science and engineering [1–3]. In order to realize its superiorities, the
dynamic modeling and design of a proper controller are essential. However,
the coordinated operation of the end effector by multiple kinematic chains
poses a challenge to the design of motion control. Generally, current controls
of the parallel robot are mainly composed of two branches, one is the kinematic
controller and the other is the dynamic controller.
For the kinematic controller of the parallel robot, there are many classic
control methods such as PD controller [4] and nonlinear PD (NPD) controller
[5]. PD controller does not need a dynamic solution. Its structure and the
calculation are simple. NPD controller is designed to improve system control
performance in the nonlinear system. In the high-speed condition, the nonlin-
ear dynamic characteristics of the parallel robots are more prominent, while
the control effect of the kinematic controller is often less than satisfactory.
Unlike the kinematic controller, the design of the dynamic controller is
based on the actual dynamic equation of the parallel robot. Fichter and Merlet
[6] are the first scholars to explore the dynamic modeling of parallel robots and
established the dynamic model of the Stewart robot. Some classic approaches,
such as the Lagrangian approach [7] and Newton-Euler approach [8], have
been widely applied. Others like d’Alembert’s principle, Kane equation, and
so forth also solved many dynamics problems. Reference [9] established the
dynamic model of the redundant actuation parallel robot in generalized space
by the application of the Udwadia-Kalaba method. Solving dynamic problems
is to provide the design foundation for the controller design.
For dynamic controllers, the system often has a good control effect especial-
ly in high-speed and high-precision situations, because it takes the dynamic
characteristics of parallel robots into consideration. Reference [10] analyzed
the dynamics of the simplified rigid body model of parallel robot. Then, the
expression of inertia moment was obtained, and the computed torque method
was proposed to improve the augmented PID controller. This method can re-
duce the difficulty of decoupling the control algorithm. Constraint-following
control [11] shows good control ability in the dynamic control strategy. The
LQR controller was introduced in [12] which presented a comparison of the
time specification performance between the LQR controller and PD controller
for a double inverted pendulum system. The sliding mode control [13], due
to its variable structure characteristics, is robust to nonlinear systems with
uncertain parameters, especially for parallel robots. From the perspective of
improving the robustness of the mechanism, reference [14] designed a variable
structure sliding mode controller to solve the instability problem.
Title Suppressed Due to Excessive Length 3

In a high-speed motion environment, the impact of external uncertainty

on system performance cannot be ignored. Hence, a large number of control
methods based on dynamic model emerge for the uncertain mechanical sys-
tem. When the bound of uncertainty has been obtained, many controllers can
effectively solve the uncertainty problem such as H∞ control [15], robust con-
trol [16] and so on. Adaptive robust control [17] has good control performance
under the unknown uncertainty. The uncertainty of parallel robot was handled
by estimating the uncertain parameters online [18–20] and the stochastic prop-
erties of uncertain terms were explored at the same time [21]. A fuzzy approach
is proposed in [22–24] to describe the uncertainty by the fuzzy set theoretic.
The uncertainty’s information tends to be unknown but its boundary is within
a fuzzy threshold.
The main contributions of this work are threefold. Firstly, the planar 2-
DOF parallel robot is we divided into several subsystems and the open-chain
dynamic equation is obtained. In order to quickly and effectively obtain the
complete dynamic model, a novel approach to calculate the constraint force
imposed by the end effector is introduced. This novel analytical mechanics ap-
proach can calculate the constraint force (holonomic constraints and/or non-
holonomic constraints) in an analytical form without any auxiliary variables
such as Lagrange’s multiplier. Secondly, in the high-speed motion environment,
the impact of external uncertainty on system performance cannot be ignored.
By using a simple function to describe the bound of uncertainty and a given
adaptive law to estimate the unknown parameter, an adaptive robust control
with weighting factors is proposed. Thirdly, by system performance analysis,
the system has the uniform ultimate boundedness property and uniform ul-
timate boundedness property. However, as the system’s control performance
is enhanced, the control costs increase accordingly. In order to keep a proper
balance between the system performance and the control costs, we need to
seek the optimal gains. Then, the optimal gain design problem is described
by a system performance index which includes a transient portion, a steady
state portion and a control cost part. The optimal gain design problem could
be solved by finding the global minimum of the system performance index.

2 Dynamic modeling of the 2-DOF parallel robot

2.1 Constraints of mechanical system

The dynamic equation of a general mechanical system without constraints is

given by [25]:
M (ϑ, t)ϑ̈ = F (ϑ̇, ϑ, t), (1)
here, ϑ ∈ Rn×1 is an n-dimensional generalized coordinate. In general, it
means the joint angles. The velocity vector, ϑ̇ ∈ Rn×1 , is the first deriva-
tive of ϑ with respect to time t, and ϑ̈ ∈ Rn×1 represents the acceleration
vector. The M (ϑ, t) ∈ Rn×n is the inertia matrix which is positive definite
and M (ϑ, t)=M T (ϑ, t). The F (ϑ̇, ϑ, t) ∈ Rn×1 is the given force that includes
4 Jiang Han1,2 et al.

Coriolis/Centrifugal force, gravitational force, and other control input forces of

the unconstrained system. The functions M (ϑ, t) and F (ϑ̇, ϑ, t) are considered
continuous.
In practice, unconstrained system is often affected by constraint force. Sup-
pose there are m(m < n) constraints in the following form:

n
∑
Ali (ϑ, t)ϑ̇ = cl (ϑ, t) l = 1, 2, 3, ..., m. (2)
i=1

Eq.(2) is the constraint in the first-order form (Pfaffian representation). The

functions Ali (·) and cl (·) are both C 1 . These constraints may in holonomic
form or nonholonomic form, and may be not integrable in general. The m
means the number of constraints attached to the system. The matrix form of
those constraints is given by

A(ϑ, t)ϑ̇ = c(ϑ, t), (3)

where A(ϑ, t)= [Ali ]m×n ∈ Rm×n , c = [c1 , c2 , ..., cm ]T ∈ Rn×1 . The expression
of constraints likes Eq.(3) is more universal. Generally, the constraint often
includes two forms, namely active constraint and passive constraint. The active
constraint is the control input which ensures the desired output. The control
input of system is the main problem for engineers. The passive constraint is
from the structure or the environment which could automatically generate the
required constraint force.

Remark 1 The constraints of the first-order form (or zero-order form when
A(ϑ, t) = 0) are widely applied to Gibbs and Appell equation, Lagrange’s
motion equation, Maggi equation, and so on. However, the existing analytical
methods are based on some auxiliary variables (such as Lagrange multipli-
ers, projections, or any quasi or auxiliary variables) rather than analytical
expressions.

Udwadia and Kalaba introduced second-order constraints (instead of first-

order constraints) into the mechanical system which is believed to be an es-
sential breakthrough in analytical mechanics. By differentiating Eq.(2) with
respect to t, we could obtain the second-order form of the constraints as fol-
lows:
n n
∑ d ∑ ϑ
( Ali (ϑ, t))ϑ̇l + Ali (ϑ, t)ϑ̈l = cl (ϑ, t), (4)
i=1
dt i=1
dt

where
n
ϑ
∑ ∂Ali (ϑ,t) ∂Ali (ϑ,t)
dt Ali (ϑ, t)= ∂ϑj + ∂t
j=1
n
d
∑ ∂cl (ϑ,t) ∂cl (ϑ,t)
dt cl (ϑ, t) = ∂ϑj + ∂t .
j=1
Title Suppressed Due to Excessive Length 5

Next, put the second-order derivative of ϑ on the left and move the rest to
the right side. Let
n
d ∑ d
bl (ϑ̇, ϑ, t) = cl (ϑ, t) − ( Ali (ϑ, t))ϑ̇l , (5)
dt i=1
dt

by combining Eq.(4) and Eq.(5), the matrix form of the second-order con-
straint is given as
A(ϑ, t)ϑ̈ = b(ϑ̇, ϑ, t), (6)
where A(ϑ, t) ∈ Rm×n , b(ϑ̇, ϑ, t) ∈ Rn×1 , rank(A) ≥ 1.
Assumption 1 The constraint in the second-order form of (6) is consistent,
i.e., there exists at least one solution ϑ̈ which satisfies the Eq.(6) for the given
A(·) and c(·).
Remark 2 The information related to constraints should not be lost for differ-
ential operations and it has been illustrated by most control problems such as
stabilization, trajectory following control, and so on.
To satisfy the constraint Eq.(6), it is necessary to provide the system Eq.(1)
with an active manner constraint force F c (ϑ̇, ϑ, t), which satisfies that the
virtual work done by the constraint force is zero (the d’Alembert’s princi-
ple). Therefore, when the mechanical system has additional constraint force
F c (ϑ̇, ϑ, t), it is called a “constrained mechanical system”, and its dynamic
equation like Eq.(1) could be rewritten as following:
M (ϑ, t)ϑ̈ = F (ϑ̇, ϑ, t) + F c (ϑ̇, ϑ, t). (7)
Consider the given A(·) and c(·), there exists at least one solution ϑ̈. All
the solutions may be called the possible acceleration. The problem of how to
calculate the constraint force is actually to find the only actual acceleration
which conforms to the d’Alembert’s principle among all possible accelerations.
In other words, the corresponding constraint force F c (ϑ̇, ϑ, t) ∈ ℜ(AT (ϑ, t)).
Udwadia and Kalaba provided the analytical form of the only actual acceler-
ation as [26]
1
q̈act = M −1 (q, t)Q(q̇, q, t) + M − /2 B + (q, t)(b(q̇, q, t)
(8)
1
− B(q, t)M − /2 (q, t)Q(q̇, q, t)),
1
where B(ϑ, t) = A(ϑ, t)M /2 (ϑ, t). The “+” is the Moore-Penrose generalized
inverse. By combining Eq.(1) and Eq.(8), we have
1
M (ϑ, t)ϑ̈ = F (ϑ̇, ϑ, t) + M /2 B + (ϑ, t)(b(ϑ̇, ϑ, t) − A(ϑ, t)M −1 (ϑ, t)F (ϑ̇, ϑ, t)).
(9)
Comparing Eq.(7) and Eq.(9), we can easily get
1
F c (ϑ̇, ϑ, t)=M /2 B + (ϑ, t)(b(ϑ̇, ϑ, t) − A(ϑ, t)M −1 (ϑ, t)F (ϑ̇, ϑ, t)). (10)
6 Jiang Han1,2 et al.

Because Udwadia and Kalaba pioneered the method, the equation was
named after them, called the Udwadia-Kalaba equation. It could be proved by
d’Alembert’s principle or extended d’Alembert’s principle and Gauss principle.
The constraint force ensures the system meets the constraint with analytical
expression.
Remark 3 Consider a constrained mechanical system subject to constraints,
the constraint force can be obtained by Udwadia-Kalaba equation in analytic
expressions, which does not need other auxiliary variables and just based on
the same generalized coordinates/velocities of the original system.

2.2 Dynamic modeling based on the active joint space

Fig. 1: 2-DOF redundantly actuated parallel robot

Consider a 2-DOF redundantly actuated parallel robot in Fig. 1, we will

formulate its precise dynamic model first. This parallel robot is composed of
three subsystems connected in the same plane. The end effector is installed
where the three subsystems are connected to each other. Each subsystem is a
two-link mechanism. The active links of the three subsystems, called the active
joints or driving joints of the parallel robot, are connected to the base. They
are driven by the servo motors installed on the base. The joint between the
active link and the passive link is not driven, which is called the passive joint.
The end effector is installed at the joint of the three passive links. The
end effector can move in the X and Y directions and is driven by three ac-
tive inputs. Therefore, the parallel robot belongs to a redundant drive type
parallel robot. The robot can be regarded as a traditional five-link mechanism
with a branch chain attached. This redundant branch chain can eliminate the
singularity of the five-link mechanism, improve the force transmission perfor-
mance, and obtain the only forward kinematics solution. At the same time,
Title Suppressed Due to Excessive Length 7

the redundant drive makes the model of the mechanism more complex. So we
introduced a novel Udwadia-Kalaba method to quickly and effectively obtain
the dynamic model of the robot. First, the parallel robot is divided into 3

Fig. 2: subsystem of the parallel robot

subsystems (as shown in Fig. 2). The mai and mbi (i=1,2,3) represent the
mass of passive link and active link. The Cai and Cbi mean the position of
the mass center. The rai and rbi mean the distance between the mass center
and the corresponding joint. The l represents the link length. For each sim-
ple subsystem, through Lagrangian mechanics, the dynamic equation of the
subsystem could be obtained easily. Particularly, since the influence of gravity
and friction are ignored in the plane, the dynamic equation of the subsystem
is given by [26]

Mi (ϑi , t)ϑ̈i + Ci (ϑ̇i , ϑi , t)ϑ̇i = τi , i = 1, 2, 3, (11)

where t ∈ R is an independent variable, ϑi = [ϑai , ϑbi ]T is the generalized

coordinates of i-th subsystem. The ϑ̇i and ϑ̈i represent the velocity vector
T
and acceleration vector respectively. The τi = [τai ,τbi ] is control input torque,
T
Mi (ϑi , t)=M i (ϑi , t) is the inertia matrix. The Ci (ϑ̇i , ϑi , t) denotes Corio-
lis/Centrifugal force matrix. The functions Mi (ϑi , t) and Ci (ϑ̇i , ϑi , t) satisfy
the corresponding condition of Eq.(1) and they are given by
[ ]
Xi Zi cos(ϑai − ϑbi )
Mi (ϑi , t) =
Zi cos(ϑai − ϑbi ) Yi
[ ] (12)
0 ϑ̇bi Zi sin(ϑai − ϑbi )
Ci (ϑ̇i , ϑi , t) = .
−ϑ̇ai Zi sin(ϑai − ϑbi ) 0

Here, Xi , Yi , Zi are constants related to the physical parameters of the sub-

systems, i.e.,
2 2

 Xi = Jai + mai rai + mbi l

2 , (13)
Yi = Jbi + mbi rbi

Zi = mbi lrbi

8 Jiang Han1,2 et al.

where the Jai and Jbi represent the moment of inertia relative to the mass
center of the corresponding link. Then combine the dynamic models of three
two-link mechanisms to get the dynamic equation of the entire open-chain
system, we get
Mo ϑ̈o + Co ϑ̇o = τo , (14)

with

0 M112 0
 
X1 0 0
 0 X2 0 0 M212 0 
0 M312 
 
 0 0 X3 0
Mo = 
M121 0

0 Y1 0 0 
 0 M221 0
 
0 Y2 0 
0 0 M321 0 0 Y3
12
 
0 0 0 C1 0 0
 0 0 0 0 C212 0  (15)
 0 0 0 0 0 C312 
 
Co = 
C121 0 0 0 0 0 

 0 C221 0 0 0 0 
 

0 0 C321 0 0 0
T
τo = [τa1 , τa2 , τa3 , τb1 , τb2 , τb3 ]
T
ϑo = [ϑa1 , ϑa2 , ϑa3 , ϑb1 , ϑb2 , ϑb3 ] .

The subscript o represents the open-chain system and the superscript indicates
the position (the row number and the column number) in the matrix.
Consider the open-chain equation (14), the complete dynamic equation
cannot be obtained without the closed chain constraint force. The constraint
force which puts three subsystems together turns the unconstrained open-chain
system into a complete closed-chain system, which is called “constrained me-
chanical system”. By introducing the Udwadia-Kalaba equation, the constraint
force can be easily obtained.
Suppose there are some constraints imposed by the end effector in the un-
constrained open-chain system. The constraints ensure that the end positions
of the three subsystems coincide. The detailed expressions of the constraint
equation are


 xa1 − xa2 + l(cos(ϑa1 ) + cos(ϑb1 ) − cos(ϑa2 ) − cos(ϑb2 ))
ya1 − ya2 + l(sin(ϑa1 ) + sin(ϑb1 ) − sin(ϑa2 ) − sin(ϑb2 ))

. (16)

 xa1 − xa3 + l(cos(ϑa1 ) + cos(ϑb1 ) − cos(ϑa3 ) − cos(ϑb3 ))
ya1 − ya3 + l(sin(ϑa1 ) + sin(ϑb1 ) − sin(ϑa3 ) − sin(ϑb3 ))


By differentiating the constraint Eq.(16) twice, we can convert the zero-order

form constraints into the second-order form constraints as

A(ϑo , t)ϑ̈o = b(ϑ̇o , ϑo , t), (17)

Title Suppressed Due to Excessive Length 9

where
 
− sin(ϑa1 ) sin(ϑa2 ) 0 − sin(ϑb1 ) sin(ϑb2 ) 0
 cos(ϑa1 ) − cos(ϑa2 ) 0 cos(ϑb1 ) − cos(ϑb2 ) 0 
Ao = l 
− sin(ϑa1 )

0 sin(ϑa3 ) − sin(ϑb1 ) 0 sin(ϑb3 ) 
cos(ϑa1 ) 0 − cos(ϑa3 ) cos(ϑb1 ) 0 − cos(ϑb3 )
cos(ϑa1 )ϑ̇2a1 − cos(ϑa2 )ϑ̇2a2 + cos(ϑb1 )ϑ̇2b1 − cos(ϑb2 )ϑ̇2b2
 
 sin(ϑa1 )ϑ̇2a1 − sin(ϑa2 )ϑ̇2a2 + sin(ϑb1 )ϑ̇2 − sin(ϑb2 )ϑ̇2 
bo = l  b1 b2 
cos(ϑa1 )ϑ̇2 − cos(ϑa3 )ϑ̇2 + cos(ϑb1 )ϑ̇2 − cos(ϑb3 )ϑ̇2  .
a1 a3 b1 b3
sin(ϑa1 )ϑ̇2a1 − sin(ϑa3 )ϑ̇2a3 + sin(ϑb1 )ϑ̇2b1 − sin(ϑb3 )ϑ̇2b3
(18)
Based on the Udwadia-Kalaba equation Eq.(10), the constraint force in the
analytical form is
1
Fo c (ϑ̇, ϑ, t)=Mo /2 Bo + [bo − Ao Mo −1 (τo − Co ϑ̇o )], (19)
1
where Bo = Ao (ϑo , t)Mo /2 (ϑo , t).
So far, we have obtained the dynamic equation (Eq.(14)) of the open-chain
system and the kinematic constraints (Eq.(19)) of the parallel manipulator.
Then, in order to obtain the complete dynamic model, the additional constrain-
t forces must be imposed on the open-chain system. Thus the final dynamic
model of the closed-chain system is
Mo ϑ̈o + Co ϑ̇o = τo + Fo c (ϑ̇, ϑ, t). (20)
Remark 4 Now, we have the complete dynamic model of the parallel robot
showed in Fig. 1 by Udwadia-Kalaba theory. This novel approach can calculate
the constraint force (holonomic constraints and/or non-holonomic constraints)
in an analytical form, just by its generalized coordinates. So the constraint
force could be expressed by the generalized coordinates of the system itself.
Having the dynamic equation like Eq.(20) is still not enough, because the
input torque contains passive joints. However, passive joints do not have input
torque. So, another problem is to convert Eq.(20) into the active joint space,
which only contains the input of active joints. Eq.(20) can be rewritten as
[ ] [ ] [ ]
ϑ̈a ϑ̇ τ
Mo + Co a = a + Fo c (ϑ̇, ϑ, t), (21)
ϑ̈b ϑ̇b τb
where ϑa = [ϑa1 , ϑa2 , ϑa3 ]T , ϑb = [ϑb1 , ϑb2 , ϑb3 ]T , τa = [τa1 , τa2 , τa3 ]T , τb =
[τb1 , τb2 , τb3 ]T . By kinematic analysis [27], we have
ϑ̇b = N ϑ̇a . (22)
The detailed expression of N is
  +
−n1 cos(ϑa1 ) −n1 sin(ϑa1 ) n1 cos(ϑb1 ) n1 sin(ϑb1 )
N =  −n2 cos(ϑa2 ) −n2 sin(ϑa2 )   n2 cos(ϑb2 ) n2 sin(ϑb2 ) 
−n3 cos(ϑa3 ) −n3 sin(ϑa3 ) n3 cos(ϑb3 ) n3 sin(ϑb3 ) (23)

1
ni = l sin(ϑbi −ϑai ) i = 1, 2, 3.
10 Jiang Han1,2 et al.

By differentiating Eq.(22), the acceleration relationship between the active

joints and the passive joints is
ϑ̈b = Ṅ ϑ̇a + N ϑ̈a . (24)
Combining Eq.(21), Eq.(22) and Eq.(24), we have
[ ] [ ] [ ] [ ]
E 0 E τ
Mo ϑ̈a + (Mo + Co )ϑ̇a = a + Fo c (ϑ̇, ϑ, t), (25)
N Ṅ N τb
[ ]
E
where E means the identity matrix. Let P = , then multiplying both
N
sides of Eq.(25) by P T , we get
[ ]
T τa
T T
P Mo P ϑ̈a + P (Mo Ṗ + Co P )ϑ̇a = P + P T Fo c (ϑ̇, ϑ, t). (26)
τb
[ ]
T τa
Since the passive joint has no input torque, τb = 0. Recall that P
[ ] τb
τ
= [ E N T ] a = τa , let M =P T Mo P, C=P T (Mo Ṗ + Co P ), F =
τb
− P T Fo c (ϑ̇, ϑ, t), then Eq.(26) can be simplified to
M ϑ̈a + C ϑ̇a + F = τa . (27)
Remark 5 So far, the problem of dynamic modeling is solved. This complete
dynamic model is based on the active joint space which provides a practical
basis for the design of the following controller.

3 Mechanical system with uncertainty

Although complete dynamic model based on the active joint space is obtained,
there are always many unavoidable uncertain disturbances in reality. The sys-
tem with uncertainty could be described as
M (ϑa , δ)ϑ̈a + C(ϑ̇a , ϑa , δ)ϑ̇a + F (ϑ̇a , ϑa , δ) = τa , (28)
where δ ∈ Σ is an uncertain parameter vector. The bounding set Σ is pre-
scribed and compact. The ϑa = [ϑa1 , ϑa2 , ϑa3 ]T is the active joint angles
of the system, ϑ̇a and ϑ̈a are the corresponding differential of ϑa over t,
τa = [τa1 , τa2 , τa3 ]T means the active inputs, M (ϑa , δ) is a positive definite
inertia matrix and M (ϑa , δ)=M T (ϑa , δ). The Ca (ϑ̇a , ϑa , δ) denotes Corio-
lis/Centrifugal force matrix, F (ϑ̇a , ϑa , δ) means the constraints imposed by
the end effector. The functions M (ϑa , δ), C(ϑ̇a , ϑa , δ) and F (ϑ̇a , ϑa , δ) are
continuous. The τa is the control input.
Assumption 2 The M (ϑa , δ) is considered to be uniformly positive definite
and uniformly ultimately upper bounded. That means there always exists con-
stants λM −max > λM −min > 0, such that
Title Suppressed Due to Excessive Length 11

λM −min E ≤ M (ϑa , δ) ≤ λM −max E, ∀ϑa ∈ R3 , (29)

here, E means the identity matrix. We state that M (ϑa , δ) ≤ λM −max E is not
necessarily satisfied and more extensive properties could be found in [23]. But
if there are only revolute joints in the system without slide joints or prismatic
joints, the inertia matrix M (ϑa , δ) should be uniformly positive definite and
uniformly ultimately upper bounded so that Eq.(29) is strictly satisfied.

Assumption 3 (i) Consider the initial state vector ϑa0 which consists of
ϑa0 i , i = 1, 2, 3, there always exists a fuzzy set named X0i in a universe of
discourse Ii ∈ R which is characterized by a membership function UΘi :
Θi → [0, 1]. That means

X0i = {(ϑa 0i , UΘi (ϑa 0i ))|ϑa 0i ∈ Θi } . (30)

(ii) Consider the uncertain parameter vector expressed by δi , i = 1, 2, 3, for

any δi (·), there always exists a fuzzy set named Hi in a universe of discourse
Σi ∈ R, which is characterized by a membership function UΣi : Σi → [0, 1].
That means
Hi = {(δi , UΣi (δi ))|δi ∈ Σi } . (31)
(iii) Furthermore, the function δ(·) is Lebesgue measurable.

By the fuzzy description, the system with uncertainty is called the fuzzy
dynamical system. However, it should be noted that the fuzzy model in our
work differs from the Takagi-Sugeno model or other if-then rule-based fuzzy
models.
We define the tracking error of the active joint angles as

ϑe (t) = ϑa − ϑad , (32)

where ϑad represents the ideal trajectory of the active joint.

Remark 6 Here, ϑe (t) and ϑa (t) are actually equivalent, because the ideal
trajectory ϑad (t) is known. But for the convenience of expression and the
reality of the project, ϑe (t) needs to be mentioned. Therefore, in the following
text, the function represented by ϑe (t) is actually equivalent to the function
represented by ϑa (t), just a different form of expression.

Assumption 4 (i) There exists a function named g(ϑ̇e , ϑe , δ) which repre-

sents the total uncertainty of the whole system. The detailed expression of
g(ϑ̇e , ϑe , δ) is
1
g(ϑ̇e , ϑe , δ) = C ϑ̇e − F − M ϑ̈ad − C ϑ̇ad + M S ϑ̇e + Ṁ (ϑ̇e + Sϑe ). (33)
2
According to [23], there exists a corresponding dimensional constant vector
′
ψ̄(t) ∈ Rn (maybe unknown) and a known function Π(ϑ̇(t), ϑ(t), ψ̄(t)) : Rn ×
n ′
R × (0, ∞)n → R+ , the n′ and n represent the corresponding dimensions.
12 Jiang Han1,2 et al.

(ii) Consider all members of the parameter vector ψ̄(t), namely ψ̄(t)i , i =
1, 2, 3, there always exists a fuzzy set named Wi in a universe of discourse
Ξi ∈ R, which is characterized by a membership function UΞi : Ξi → [0, 1].
That means
{ }
Wi = (ψ̄(t)i , UΞi (ψ̄i ))|ψ̄i ∈ Ξi . (34)

The function Π(ϑ̇a (t), ϑa (t), ψ̄(t)) is a suitable upper bound for g(ϑ̇e , ϑe , δ).
Such that for all ϑe (t)

Π(ϑ̇(t), ϑ(t), ψ̄(t)) ≥ g(ϑ̇e , ϑe , δ). (35)

It should be noted that the function Π(ϑ̇a (t), ϑa (t), ψ̄(t)) is not arbitrary.
It must meet the following assumptions.

Assumption 5 (i) The function Π(ϑ̇a (t), ϑa (t), ψ̄(t)) is C 2 (means two-times
continuously differentiable) and concave. For any parameter ψ̄(t), it is a non-
decreasing function, such that for any ψ1 , ψ2

Π(ϑ̇(t), ϑ(t), ψ1 (t)) − Π(ϑ̇(t), ϑ(t), ψ2 (t)) ∂Π

≥ (ϑ̇(t), ϑ(t), ψ1 (t)). (36)
ψ1 (t) − ψ2 (t) ∂ ψ̄

(ii) The function ∂Π/∂ ψ̄ is continuous.

Although the bound of uncertainty is not know, the existence of such bound
is doubtless. Then we assume that there is a function named Π(ϑ̇a (t), ϑa (t), ψ̄(t)).
The structure of this function is known, but one of the parameters ψ̄(t) is
unknown. We estimate the value of ψ̄(t) through ψ̃(t) so that the value of
Π(ϑ̇a (t), ϑa (t), ·) will be infinitely close to the actual value. The accurate
bound of uncertainty will be obtained. The estimated value ψ̃(t) is used in
the controller design, and it satisfies a given adaptive law. That is

∂(ψ̃) −1 ∂Π
=k (ϑ̇(t), ϑ(t), ψ̃(t)) ϑ̇e (t) + Sϑe (t) − ψ̃(t), (37)
∂t ∂ ψ̄

where k is a given constant positive scalar, S = diag[Sii ]3×3 , Sii > 1, i =

1, 2, 3. The scalar k determines the “rate” of adaptation. If k is appropriately
large, the adaptive parameters will increase rapidly, so that the control input
changes correspondingly in order to compensate for the uncertainty effectively.

Remark 7 Here, g(ϑ̇e , ϑe , δ) represents the total uncertainty of the whole

system, in other words, it means the bound of uncertainty. However, it is
obvious that the form of g(ϑ̇e , ϑe , δ) is very complicated. How to get the
value of g(ϑ̇e , ϑe , δ) is a challenging problem. But using a simple function
Π(ϑ̇a (t), ϑa (t), ·) to describe the bound is a reasonable idea.
Title Suppressed Due to Excessive Length 13

4 Adaptive robust control design

We design the adaptive robust control for the system as following:

τa (t) = ατ1 (t) + βτ2 (t)

τ1 (t) = −(ϑ̇e (t) + Sϑe (t))Π 2 (ϑ̇a (t), ϑa (t), ψ̃(t)) (38)
τ2 (t) = −(ϑ̇e (t) + Sϑe (t)),

where τa (t) means the input torque which consists of τ1 (t) and τ2 (t). The τ1 (t)
stands for the robust control part whose main contribution is to compensate
for the uncertainty. The τ2 (t) stands for the PD control part, and it ensures
that the controller has the ability to track the ideal trajectory. The α and β are
positive constants which respectively represent weighting factors of adaptive
robust portion and PD portion.
Subject to the proposed control input, the system should achieve the follow-
ing goals. As t approaches to infinity, the tracking error ϑe (t) and its first-order
derivative ϑ̇e (t) approaches to zero. Meanwhile, the parameter ψ̃(t) approach-
es to ψ̄(t). Let X = [ϑe (t)T , ϑ̇e (t)T ]T , ψ = ψ̃ − ψ̄, Z = [X T , ψ T ]T . In general,
Eq.(28) could be rewritten as following for convenience:

M (ϑ, δ)ϑ̈e + C(ϑ̇, ϑ, δ)ϑ̇e + F c (ϑ̇, ϑ, δ) + M (ϑ, δ)ϑ̈d + C(ϑ̇, ϑ, δ)ϑ̇d = τa . (39)

Theorem 1 Consider the system which satisfies the Assumptions 2-4, the
control input (38) guarantees Z has the following properties:
(i) Uniform boundedness: For any constant χ > 0 with ∥Z(t0 )∥ ≤ χ, there
exists a d(χ) ∈ [0, ∞) such that if Z(·) is a solution with ∥Z(t0 )∥ ≤ χ, then
∥Z(t)∥ ≤ d(χ) for all t ∈ [t0 , ∞).
(ii) Uniform ultimate boundedness: Given any χ ∈ [0, ∞), χmin = min d(χ)
and ∥Z(t0 )∥ ≤ χ, for any χ̃ > χmin , we have ∥Z(t)∥ ≤ χ̃ for all χ̃ > χmin in
a finite time T (χ̃, χ) < ∞.

Proof Select a Lyapunov function candidate for the dynamical system Eq.(39)
as following:
V = Vϑ e + Vψ . (40)
To prove V is a legitimate Lyapunov function for the fuzzy dynamical system,
we must prove that V is positive definite and decreasing globally. According
to Assumption 2, we have
1 2
Vϑ e ≥ λM −min ϑ̇e (t) + Sϑe (t) + ϑTe (t)βSϑe (t)
2
n n
1 2
(ϑ˙e i (t) + 2ϑ˙ei (t)Sii ϑei (t) + Sii 2 ϑei 2 (t)) + β
∑ ∑
= λM −min Sii ϑei 2 (t)
2 i=1 i=1
n [ ]
1∑ ϑei (t)
= [ϑei (t), ϑ˙ei (t)]Λi ˙ ,
2 ϑei (t)
i=1
(41)
14 Jiang Han1,2 et al.

here

Vψ = 21 ψ T kψ
(42)
Vϑ e = 21 (ϑ̇e (t) + Sϑe (t))T M (ϑ, δ)(ϑ̇e (t) + Sϑe (t)) + ϑTe (t)βSϑe (t),

with
λM −min Sii 2 + 2βSii λM −min Sii
[ ]
Λi = . (43)
λM −min Sii λM −min
Based on Eq.(43), we know λmin (Λi ) = λM −min , so
n [ ]
1∑ ˙ ϑei (t)
Vϑ e ≥ [ϑei (t), ϑei (t)]Λi ˙
2 i=1 ϑei (t)
n
(44)
[ ]
1∑ ˙ ϑei (t)
≥ λmin (Λi )[ϑei (t), ϑei (t)] ˙
2 i=1 ϑei (t)
2
=λM −min ∥X∥ .

Moreover,

V = Vϑ e + Vψ
2 1
≥ λM −min ∥X∥ + ψ T kψ (45)
2
2
≥ λ1 ∥Z∥ ,

where λ1 = min{λM −min , 12 k}. So, we obtain the possible lower bound of V .
Next, we analyze the upper bound.

1 2
Vϑ e ≤ λM −max ϑ˙e (t) + Sϑe (t) + ϑTe (t)βSϑe (t)
2
n n
1 2
(ϑ˙e i (t) + 2ϑ˙ei (t)Sii ϑei (t) + Sii 2 ϑei 2 (t)) + β
∑ ∑
= λM −max Sii ϑei 2 (t)
2 i=1 i=1
n [ ]
1 ϑei (t)
[ϑei (t), ϑ˙ei (t)]Λi ˙
∑
= ,
2 ϑei (t)
i=1
(46)

where
λM −max Sii 2 + 2βSii λM −max Sii
[ ]
Λi = . (47)
λM −max Sii λM −max
Based on Eq.(43) we know λmax (Λi ) = 2βSii , So
n [ ]
1∑ ϑei (t)
Vϑ e ≤ λmax (Λi )[ϑei (t), ϑ˙ei (t)] ˙
2 i=1 ϑei (t) (48)
2
= max{βSii }∥X∥ .
Title Suppressed Due to Excessive Length 15

Moreover
V = V ϑ e + Vψ
2 1
≤ max{βSii }∥X∥ + ψ T kψ (49)
2
1 2 2
≤ k∥Z∥ =: λ2 ∥Z∥ .
2
If k is appropriately large, the adaptive parameters will increase rapidly,
so that the control input changes correspondingly in order to compensate for
the uncertainty effectively. If we choose k ≥ 2β max(λS ) and λ2 = 21 k, the
information of upper bound is obtained.
To illustrate that V is decreasing, we analyze the derivative of it. The
derivative of V is given by
V̇ = V̇ϑe + V̇ψ . (50)
We analyze the two parts of V separately. First,
T 1 T
V̇e = (ė(t) + Se(t)) M (ë(t) + S ė(t)) + (ė(t) + Se(t)) Ṁ (ė(t) + Se(t))
2
+ 2βeT (t)S ė(t)
T T 1
= (ė(t) + Se(t)) (M ë(t) + M S ė(t)) + (ė(t) + Se(t)) ( Ṁ (ė(t) + Se(t)))
2
+ 2βeT (t)S ė(t)
T 1
= (ė + Se) [τa − C ė − F c − M q̈d − C q̇d + M S ė + Ṁ (ė + Se)]
2
+ 2βeT S ė.
(51)

By combining Eq.(33) and Eq.(51), we have

V̇ϑe = (ϑ˙e + Sϑe )T (τa + g) + +2βϑe T S ϑ˙e . (52)

Recalling Eq.(35) and Eq.(38), we get

T T
V̇e ≤ (ė + Se) Π(q̇(t), q(t), ψ̄(t)) + (ė + Se) [ατ1 (t) + βτ2 (t)]
+ 2βeT S ė
T 2
(53)
= (ė + Se) Π(q̇(t), q(t), ψ̄(t)) − α∥ė + Se∥ Π 2 (q̇(t), q(t), ψ̃(t))
− ėT β ė − eT βS 2 e.
2
Let λ3 = min{Sii , 1}, Eq.(53) can be rewritten as
T
V̇e ≤ (ė + Se) Π(q̇(t), q(t), ψ̄(t))
2
− α∥ė + Se∥ Π 2 (q̇(t), q(t), ψ̃(t)) − ėT βλ3 ė − eT βλ3 e
T
(54)
=(ė + Se) Π(q̇(t), q(t), ψ̄(t))
2 2
− α∥ė + Se∥ Π 2 (q̇(t), q(t), ψ̃(t)) − βλ3 ∥X∥ .
16 Jiang Han1,2 et al.

The details of the other part V̇ψ are as follows

V̇ψ = ψ T k ψ̇
( )T . (55)
= ψ̃ − ψ̄ k˜˙ψ

Considering the adaptive law in Eq.(37) and Assumption 5, we have

( )T ∂Π
V̇ψ = ψ̃ − ψ̄ kk −1 (q̇(t), q(t), ψ̃(t)) ∥ė(t) + Se(t)∥
∂ ψ̄
( )T
− ψ̃ − ψ̄ k ψ̃
( ) (56)
≤ Π(q̇(t), q(t), ψ̃(t)) − Π(q̇(t), q(t), ψ̄(t)) ∥ė(t) + Se(t)∥
( )T
− ψ̃ − ψ̄ k ψ̃.

According to Eq.(54) and Eq.(56), the complete expression of V̇ is as follow

V̇ = V̇e + V̇ψ
T 2 2
≤ (ė + Se) Π(q̇(t), q(t), ψ̄(t)) − α∥ė + Se∥ Π 2 (q̇(t), q(t), ψ̃(t)) − βλ3 ∥X∥
( )T
+ [Π(q̇(t), q(t), ψ̃(t)) − Π(q̇(t), q(t), ψ̄(t))] ∥ė(t) + Se(t)∥ − ψ̃ − ψ̄ k ψ̃
T 2
= (ė + Se) Π(q̇(t), q(t), ψ̃(t)) − α∥ė + Se∥ Π 2 (q̇(t), q(t), ψ̃(t))
( )T
2
− βλ3 ∥X∥ − ψ̃ − ψ̄ k ψ̃.
(57)
1
According to the inequality −ax2 − x ≤ 4a , a ∈ R, we can get
1 2
V̇ ≤ − βλ3 ∥X∥ − ψ T k ψ̃
4α
1 2
= − βλ3 ∥X∥ − ψ T k(ψ̃ − ψ + ψ) (58)
4α
1 2 2
= − βλ3 ∥X∥ − k∥ψ∥ − ψ T k ψ̄.
4α
As being stated above, we know k ≥ 2β max(λS ). So
1 2 2
V̇ ≤ − βλ3 ∥X∥ − k∥ψ∥ − ψ T k ψ̄
4α
1 2 2
≤ − βλ3 ∥X∥ − 2βλ3 ∥ψ∥ − 2βλ3 ∥ψ∥ ψ̄
4α
1 2 2 (59)
= − βλ3 ∥Z∥ − (∥ψ∥ + 2 ∥ψ∥ ψ̄ )βλ3
4α
2 2 1
≤ −βλ3 ∥Z∥ + βλ3 ψ̄ +
4α
2
= −λ̂∥Z∥ + ĥ,
Title Suppressed Due to Excessive Length 17

2 1
where λ̂ = βλ3 , ĥ = βλ3 ψ̄ + 4α . Now, we have been proved that V is
decreasing if
ĥ2
∥Z∥ ≥ (60)
λ̂
From [24], we can obtain the uniform boundedness performance. That is
{
λ2 χ2 −λ1 Γ 2
if χ > Γ̃
T (χ̃, χ) = λ̂Γ̃ 2 −ĥ
0 if χ ≤ Γ̃ (61)
√
Γ̃ =χ̃ λ1 /λ2 ,
where √ /
Γ = ĥ λ̂. (62)

In addition, the uniform ultimate boundedness performance is as following:

√
χmin = Γ λ2/λ1 . (63)

Consider all χ̃ > χmin

{
λ2 χ2 −λ1 Γ 2
if χ > Γ̃
T (χ̃, χ) = λ̂Γ̃ 2 −ĥ
0 if χ ≤ Γ̃ (64)
√
Γ̃ =χ̃ λ1 /λ2 .

Remark 8 The value of χmin determines the radius of uniform ultimate bound-
edness ball. It is clear that χmin will decrease correspondingly as Γ decreases.
Meanwhile, Γ is determined by the weighting factors α and β. The Γ will
decrease as α and β both increase. So the uniform boundedness performance
of the system will be enhanced as α and β increase. However, the bigger α or
β does not always mean the better. Because as they increase, the control input
of the system also increases, which means greater control cost. Therefore, how
to keep a proper balance between the control cost and the control performance
is really necessary.

5 Optimal gain design based on D-operation

The system performance guaranteed by the control input has been analyzed.
Through the analysis, it can be seen that the uniform boundedness perfor-
mance will be enhanced when α and β increase. However, as they increase,
the control cost increases accordingly, which becomes another issue. In order
to keep a proper balance between the control cost and the system performance,
we need to seek the optimal α and β. The so-called balance will be described
by a system performance index function, which includes a transient portion,
a steady state portion and a control cost portion. The problem of finding the
18 Jiang Han1,2 et al.

optimal gains can be transformed into the problem of finding the minimum
value of the system performance index.
First of all, we will further explore the uncertain system. Based on the
lower bound and the upper bound of V , we have
2 2
λ1 ∥Z∥ ≤ V ≤ λ2 ∥Z∥ , (65)

and hence
V 2
− ≥ −∥Z∥ . (66)
λ2
By combining Eq.(59) and Eq.(66), we have

2 2 1
V̇ ≤ −βλ3 ∥Z∥ + βλ3 ψ̄ +
4α
V 2 1
≤ −βλ3 + βλ3 ψ̄ + (67)
λ2 4α
−2λ3 β 2 1
≤ V + βλ3 ψ̄ + .
k 4α
The analysis of this differential inequality is explained in detail in Chen
[28]. For further analysis of this differential inequality, the following is needed.

Definition 1 Consider a scalar function ω(s(t), t) with the scalars s(t) and
t in some open connected set D, we hold a function s(t) with t ∈ [t0 , t1 ] is a
solution of the differential inequality

ṡ(t) ≤ ω(s(t), t), (68)

if s(t) is continuous and its derivative satisfies Eq.(68) with any t ∈ [t0 , t1 ].

Theorem 2 Supose that ϕ(t) is continuous on an open connected set D and

there exists a unique solution of the differential equality

ϕ̇(t) = ω(ϕ(t), t), ϕ(t0 ) = ϕ0 . (69)

If s(t) and ϕ(t) are the solutions of Eq.(68) and Eq.(69) respectively and
the initial condition satisfies s(t0 ) ≤ ϕ(t0 ), then s(t) ≤ ϕ(t) for all t ∈ [t0 , t1 ].

Theorem 3 For the differential inequality Eq.(68) and the differential equa-
tion Eq.(69), for all (x1 , t), (x2 , t) ∈ D, there always exists a constant L > 0
which ensures the function ω(·) satisfies the Lipschitz condition

|ω(x1 , t) − ω(x2 , t)| ≤ L |x1 − x2 | . (70)

Then, for all t ∈ [t0 , t1 ], any solution s(t) of the differential inequality Eq.(68)
also satisfies the following inequality

s(t) ≤ ϕ(t), t ∈ [t0 , t1 ]. (71)

Title Suppressed Due to Excessive Length 19

Exploring the solutions of Eq.(68) often involves many difficulties because

the solution may be not available or non-unique. On the contrary, studying
the upper bound information of the solution may be a feasible way and it can
avoid the trouble caused by the non-unique solution. Because the analytic form
of upper bound can be solved, it gives an approach to explore the performance
index function.
Based on the above theorems and definition, a differential equation is con-
structed as
−2λ3 β 2 1
ϕ̇(t) = ϕ(t) + βλ3 ψ̄ + , (72)
k 4α
with ϕ(t0 ) = V (t0 ). Then, by solving the differential Eq.(72), we have
2 1
k(βλ3 ψ̄ + 4α )
−2λ3 β
t k 2 1
ϕ(t) = (ϕ0 − )exp k + (βλ3 ψ̄ + ). (73)
2λ3 β 2λ3 β 4α
Therefore, combining Eq.(71) and Eq.(73), we get

V (t) ≤ ϕ(t)
2 1
k(βλ3 ψ̄ + 4α )
−2λ3 β
t k 2 1
≤ (V0 − )exp k + (βλ3 ψ̄ + )
2λ3 β 2λ3 β 4α (74)
2
k ψ̄ k −2λ3 β k 2 k
≤ (V0 − − )exp k t + ψ̄ + ,
2 8λ3 αβ 2 8λ3 αβ
where V0 = Ve (t0 )+Vψ (t0 ), t0 = 0. It should be noted that the initial condition
does not necessarily to be t0 = 0. Because V (t) ≤ ϕ(t) is satisfied at t0 = 0,
so for any t ≥ t0 , the inequality V (t) ≤ ϕ(t) is also satisfied. We assume that
t0 = 0 is just for the control input starts at t = 0.
Notice that for each t ∈ [t0 , ∞), let
2
k ψ̄ k −2λ3 β
η(α, β, ψ̄, t) = (s0 − − )exp k t
2 8λ3 αβ (75)
k 2 k
η∞ (α, β, ψ̄, t) = ψ̄ + .
2 8λ3 αβ
We can find that for each α, β and ψ̄, η(α, β, ψ̄, t) approaches to 0 as
t → ∞. So η(α, β, ψ̄, t) reflects the transient performance which depends on
the initial state. On the other hand, η∞ (α, β, ψ̄, t) represents the steady state
performance.
The exact knowledge of the uncertainty may be not available. Therefore,
it may be feasible to study η(α, β, ψ̄, t) and η∞ (α, β, ψ̄, t) when analyzing the
system performance. Although η(α, β, ψ̄, t) and η∞ (α, β, ψ̄, t) are related to
the value of ψ̄ and the knowledge of exact value may not have, the value of
ψ̄ is characterized by a known membership function. In order to find a mean-
ingful reference value about the fuzzy number ψ̄ in the fuzzy set to accurately
describe the function η(α, β, ψ̄, t) and η∞ (α, β, ψ̄, t), the D-operation is intro-
duced as following.
20 Jiang Han1,2 et al.

Definition 2 Consider a fuzzy set

{ ∑}
H = (δ, U∑ (δ))|δ ∈ (76)

∑
and a function ζ(δ) : → R with δ as a parameter, the D-operation of ζ(δ)
is defined as
ζ(δ)U∑ (δ)dδ
∫
∑
D[ζ(δ)] = . (77)
U∑ (δ)dδ
∫
∑

Remark 9 As mentioned before, the value of ψ̄ is characterized by a mem-

bership function. Then the D-operation is introduced to find a meaningful
reference value of the fuzzy number ψ̄ in the fuzzy set so that we can accu-
rately describe the function η(α, β, ψ̄, t) and η∞ (α, β, ψ̄, t). A special case of
D-operation while ζ(δ) = 1 is equivalent to the well-known center-of-gravity
defuzzification method.

Now we define the system performance index function named J as

∫ ∞
J = D( η 2 (α, β, ψ̄, t)dt) + µD(η 2 ∞ (α, β, ψ̄, t)) + ν(α2 + β 2 )
0 (78)
=: J1 + µJ2 + νJ3 ,

here,

∞ 2
k ψ̄ k k
∫
2
η (α, β, ψ̄, t)dt = (s0 − − )2
0 2 8λ 3 αβ 4λ 3β
2
k 2 k k 4 k2 k2 2
η 2 ∞ (α, β, ψ̄, t) = ( ψ̄ + )2 = ψ̄ + ψ̄ . +
2 8λ3 αβ 4 8λ3 αβ 64λ2 3 α2 β 2
(79)
The J1 reflects the transient performance which relates to the initial state,
J2 is related to the steady state performance. The J3 part will increase as α
and β increases. It represents the control cost.The control cost and the steady
state performance are the key points to be considered in engineering practice.
The parameters µ and ν are the weighting factors between the steady state
performance and the control cost. Next, the optimal α and β under different
weighting factors µ and ν will be further explored.
Based on the definition of D-operation, we have

J = J1 + µJ2 + νJ3
1 1 1 1 1
= ρ0 + ρ1 2 3 + ρ2 2 2 + ρ3 2 + ρ4 + ρ5 + ρ6 β + ρ7 α2 + ρ7 β 2 .
α β α β αβ αβ β
(80)
Title Suppressed Due to Excessive Length 21

Under reasonable initial conditions, ρ0 ∼ ρ7 are generally normal numbers.

The detailed expressions of parameters ρ0 ∼ ρ7 are

2
k(0.069058 − 0.043kD( ψ̄ )) µk 2 4
ρ0 = + D( ψ̄ )
4λ3 4
k3
ρ1 =
256λ3 3
µk 2
ρ2 =
64λ2 3
2
k 3 D( ψ̄ ) − 1.606k 2
ρ3 =
32λ2 3 (81)
2 2
k (2λ3 µD( ψ̄ ) − 0.043)
ρ4 =
16λ2 3
4 2
k(2.579236 + k 2 D( ψ̄ ) − 3.212kD( ψ̄ ))
ρ5 =
16λ3
0.001849k
ρ6 =
4λ3
ρ7 = ν .

Seeking the optimal parameters α and β can be transformed into a problem

of finding the minimum of J. This means

min J(α, β, V0 ) α, β > 0. (82)

α,β

To find the minimum value of J, we first calculate the first-order partial

derivative of J. For any initial condition

∂J 2ρ1 2ρ2 ρ3 ρ4
= 2ρ7 α − 3 3 − 3 2 − 2 2 − 2
∂α α β α β α β α β
1
= 3 3 (2ρ7 α4 β 3 − ρ4 αβ 2 − ρ3 αβ − 2ρ2 β − 2ρ1 )
α β
∂J 3ρ1 2ρ2 2ρ3 ρ4 ρ5
= 2ρ7 β + ρ6 − 2 4 − 2 3 − − − 2
∂β α β α β αβ 3 αβ 2 β
1
= 2 4 (2ρ7 α2 β 5 + ρ6 α2 β 4 − ρ5 α2 β 2 − ρ4 αβ 2 − 3ρ3 αβ − 2ρ2 β − 3ρ1 ).
α β
(83)

From the information of the first-order partial derivative, it is not easy

to directly find out when the function takes the minimum value. Therefore,
22 Jiang Han1,2 et al.

taking the second-order partial derivative of J, we further have

∂2J 6ρ1 6ρ2 2ρ3 2ρ4

= + 4 2 + 3 2 + 3 + 2ρ7
∂α2 4
α β 3 α β α β α β
∂2J 12ρ1 6ρ2 6ρ3 2ρ4 2ρ5
= + 2 4+ + + 3 + 2ρ7 (84)
∂β 2 2
α β 5 α β αβ 4 αβ 3 β
∂ ∂J 6ρ1 4ρ2 2ρ3 ρ4
( )= + 3 3 + 2 3 + 2 2.
∂β ∂α α3 β 4 α β α β α β

It should be declared that for any α and β, we have

∂2J ∂2J
A= 2
> 0, B = >0
∂α ∂β 2
∂2J ∂ ∂J . (85)
∂α2 ∂β ( ∂α )
C = ∂ ∂J ∂2J
>0
∂β ( ∂α ) ∂β 2

Based on Eq.(85), we know A > 0, C > 0. So for any solution αl and βl

satisfying Eq.(86), J(αl , βl , V0 ) is the local minimum.

∂J ∂J
= 0, = 0. (86)
∂α ∂β
2 2
∂ J ∂ J
Remark 10 For any α and β, ∂α 2 > 0, ∂β 2 > 0 are globally satisfied. Therefore,

J(α, β, V0 ) is a global concave function with respect to α and β so that the

local minimum of J(α, β, V0 ) is also the global minimum. Thus, the solution
that satisfies Eq.(86) is indeed the optimal solution named αopt and βopt .

Udwadia-Kalaba
equation

dynamic equation of Closed chain Convert to

three subsystems constraint force the active joint space

forward kinematics

Optimal
timal gain
ga

Adaptive robust
Bound of ontro
control
 uncertainty

Control input Dynamics model
Eq.(27)
Constant PD contr
control

 matrix

Optimal ga
gain

Velocity jacobi
matrix

the system performance Solve Function

Calculate parameters
x func
index function

Fig. 3: The main design procedure of the proposed control scheme

Title Suppressed Due to Excessive Length 23

Until now, the problems about dynamic modeling, adaptive robust control
design and optimal gain design are solved. Fig. 3 shows the main design pro-
cedure of proposed control scheme. The entire design process of the proposed
control method is summarized as follows:
Step 1: Obtain the dynamic equation of the entire open chain system like
Eq.(14).
Step 2: Calculate the closed chain constraint force by Udwadia-Kalaba
equation.
Step 3: Obtain the complete dynamic model of the parallel robot like
Eq.(20). Then convert it into the active joint space.
Step 4: Choose suitable k and bound function Π so that we can formulate
the adaptive laws by Eq.(37).
Step 5: For given S, calculate τ1 and τ2 .
Step 6: Based on the system performance analysis and D-operation, con-
struct the system performance index function J.
Step 7: Calculate parameters ρ0 ∼ ρ7 , then solve Eq.(86) to get the optimal
value αopt and βopt .

6 Illustrative example

On the one hand, we verify the accuracy of the proposed dynamic modeling
method first. The initial conditions of the base joint are A1 (0, 0), A2 (0.43, 0),
A3 (0.4269, 0.5005). We select the initial value of the revolute joints as qo (t =
0) = [1.3137; 2.9036; −1.4326; −0.8667; 1.4593; 2.972]. Consider Eq.(13), we set


 Xi = 0.0932 {
mai = 1.2525
Yi = 0.0381 l = 0.244 (i = 1, 2, 3). (87)
mbi = 1.0771
Zi = 0.0426


All initial values are based on SI units. For any input torque such as τa =
[0.1 cos(πt), 0.1 cos(πt), 0.1 cos(πt)]T , the actual output of the system is shown
in Fig. 4.
24 Jiang Han1,2 et al.

0.3

0.295

End effector y-axis coordinates/m

0.29

0.285

0.28 subsystem 1
subsystem 2
subsystem 3

0.275
0.21 0.212 0.214 0.216 0.218 0.22 0.222 0.224
End effector x-axis coordinates/m

Fig. 4: System’s output under random input

Under random input τa , Fig. 4 shows the trajectory of the three subsystems’
ends on the plane. Due to the constraint force imposed by the end effector,
the phenomenon of the end trajectory overlap between the three subsystems
shown by the system output is consistent with the actual situation. Therefore,
the proposed dynamic modeling method is proved to be effective.
On the other hand, to illustrate that the proposed control has the ability to
deal with uncertain interference, we assume that the mass of the links is uncer-
tain and there exists external disturbing force. Let mai = m̄ai + ∆mai , mbi =
m̄bi +∆mbi . The m̄ai and m̄bi means the known values, ∆mai and ∆mbi means
unknown values related to uncertainty. We define the uncertainties in mai and
mbi are the same, that is ∆mai = ∆mbi = ∆m. ∆m is “close to 0.25” with a
membership function (sine) as
Ω∆m (υ) = sin(2πυ), 0 ≤ υ ≤ 0.5. (88)
It needs to be stated that the sine function is not the only suitable member-
ship function for the system. Before introducing the fuzzy set to describe the
uncertainty of the system, the uncertainty should be studied first. Then the
correspondingly suitable membership function for the system could be chosen
reasonably.
Next, the bounding function Π is given as
Π = ψ̃ T [ϑa T ϑa ; ϑ̇Ta ϑ̇a ], (89)
here, ψ̃ satisfies the following adaptive law

˙ ˙ ˙ T
ψ̃ = [ψ̃ 1 ,ψ̃ 2 ]
 
2
1 ϑ̇ e + Sϑ e ∥ϑ a ∥ [
ψ̃1
] (90)
=  2  − 2
k ϑ̇e + Sϑe ϑ̇a ψ̃2 .
Title Suppressed Due to Excessive Length 25

According to step 5 and step 6, we choose k = 12, S = 2, λ3 = 1. Then, we

will calculate the optimal α and β under different weighting factors µ and ν.
Based on the fuzzy description, we define the estimated value ψ̃ is also “close
to 0.25” with the membership function (sine) as
Ωψ̃ (υ) = sin(2πυ) 0 ≤ υ ≤ 0.5. (91)
Based on the D-operation, we can get
ρ0 = -0.02298 + 1.24611µ
ρ1 = 6.75
ρ2 = 2.25µ
ρ3 = 0.80166
(92)
ρ4 = 9 ∗ (0.29735µ − 0.043)
ρ5 = 1.37474
ρ6 = 0.005547
ρ7 = ν .
By the membership function Eq.(91) and D-operation in Definition 2, we can
obtain the optimal gain αopt and βopt by Eq.(86) under different weighting
factors µ and ν. The optimal solutions are summarised in Table 1. From this
table, we notice that the optimal gain (αopt , βopt ) will increase as the weighting
ratio (µ/ν) increases.

Table 1: Optimal gain under different weighting factors µ and ν

(µ,ν) µ/ν αopt βopt Jmin
(1,1) 1 1.3597 1.6620 9.1329
(5,1) 5 1.7839 1.9337 18.9766
(10,1) 10 2.0655 2.1671 29.3006
(1,10) 0.1 0.9419 1.1195 34.1584
(1,5) 0.2 1.0503 1.2535 22.3282

Another point that should be noted is that different weight ratio µ/ν rep-
resents different needs in actual situations. If the engineers pay more attention
to control performance in practice, they can choose a larger weight ratio. On
the contrary, if more attention is paid to the control cost, a lower weight ratio
should be chosen.
Until now, all parameters have been obtained. Next, we choose the desired
trajectory of the output like
{
xE (t) = b(a cos(ωt) + R cos(ωt) cos(nωt))
(93)
yE (t) = b(a sin(ωt) + R sin(ωt) cos(nωt)),
here, ω = 1, n = 5, R = 0.5, a = 3, b = 0.06/3.5. Meanwhile, we choose ∆m =
0.25 sin(πt) to describe the uncertainty caused by random mass changes and
26 Jiang Han1,2 et al.

a random τ̃ to describe the uncertainty caused by disturbance torque. All the

simulation results are as follows.

∂J ∂J
Fig. 5: The unique solution of ∂α
= ∂β
=0

 " $ $ "  $
 #  $ $

 

!$

$
$




$

$



!$

 $

  $ $   
$ $ $

Fig. 6: The system performance index function J with respect to α, β

∂J
From Fig. 5, it can be concluded that the solution of ∂α = ∂J
∂β = 0 is unique.
In other words, the optimal gain is also unique. Fig. 6 shows the relationship
between the system performance index function J and (α, β). Combining with
Fig. 6 and Eq.(85), we know that J(α, β, V0 ) is a global concave function with
∂J
respect to (α, β). Therefore, the solution satisfying ∂α = ∂J
∂β = 0 is the unique
optimal gain (αopt , βopt ). The minimum value of the system performance index
Title Suppressed Due to Excessive Length 27

function does not have multiple solutions. In other words, the result of the
proposed optimal design is unique.

0.42

Random
0.4 disturbDQFH
0.378
End effector y-axis coordinates/m

0.38 0.376

0.36 0.374

0.372
0.34
0.166 0.17 0.174

0.32

0.302
0.3 0.3
0.298
0.296 Desired trajectory
0.294 Initial
0.28 3roposed control
0.292 error
0.22 0.225 0.23 PD control
0.26
0.1 0.12 0.14 0.16 0.18 0.2 0.22 0.24 0.26 0.28 0.3
End effector x-axis coordinates/m

Fig. 7: Actual tracking effect under different control

In general, the proposed method has a better control effect through ran-
dom disturbance, initial position error, and mass change. Comparing with
traditional PD control, Fig. 7 shows the actual tracking performance of the
proposed method. For the same initial position error, the proposed method
has a faster response and smaller steady-state error when tracking the desired
trajectory. In addition, a random disturbance torque is given when t = 4s and
the detailed results are shown from Fig. 8 to Fig. 12. It can be seen that the
proposed method is more robust in dealing with random disturbance. Finally,
we choose ∆m = 0.25 sin(πt) to describe the uncertainty caused by random
mass changes in the whole process. The tracking performance of the proposed
adaptive robust method is more accurate than that of traditional PD control.
All the simulation results show that the proposed adaptive robust method has
better tracking accuracy, more rapid response, and stronger robustness than
traditional PD control.
28 Jiang Han1,2 et al.

10-3
3

Tracking error of X-axis /m

0

-1
10 -6
6
-2

Random 5 / =1
-3 / =5
disturbance
/ =10
4
/ =0.1
-4 / =0.2
3 PD
-5
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Time /s

Fig. 8: Comparison of tracking error of X-axis under different control

10-3
5
10-6
4 / =1
-4 / =5
/ =10
3 -6
Random / =0.1
Tracking error of Y-axis /m

disturbance -8 / =0.2
2 PD
-10
1 -12

-1

-2

-3

-4
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Time /s

Fig. 9: Comparison of tracking error of Y-axis under different control

Title Suppressed Due to Excessive Length 29

0.035
10 -5
0.03 / =1
/ =5
-2
0.025 / =10
/ =0.1
0.02 -3 / =0.2
Random PD

/rad
disturbance
0.015 -4
a1
0.01
Error of

0.005

-0.005

-0.01

-0.015
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Time /s

Fig. 10: Comparison of tracking error of ϑa1 under different control

0.04

0.02

0
/rad a2

-0.02
Error of

10 -5
6
-0.04 Random
5 / =1
disturbance
/ =5
4
/ =10
-0.06 3 / =0.1
/ =0.2
2
PD
-0.08
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Time /s

Fig. 11: Comparison of tracking error of ϑa2 under different control

30 Jiang Han1,2 et al.

0.015

0.01

0.005

/rad a3 -0.005
Error of

-0.01 10 -5
8
Random
-0.015 6
disturbance / =1
4 / =5
-0.02 / =10
2 / =0.1
-0.025 / =0.2
0 PD
-0.03
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Time /s

Fig. 12: Comparison of tracking error of ϑa3 under different control

From Fig.8 to Fig.12, the tracking error of the proposed method is smaller
than traditional PD control, which indicates better control performance. On
the one hand, under the interference of mass changes, the tracking error under
the proposed control converges into a neighborhood of zero rapidly in a finite
time and then remained within a narrow zone thereafter. This phenomenon
proves that the proposed adaptive robust control could guarantee determinis-
tic robust performances of the system (the uniform boundedness and uniform
ultimate boundedness). On the contrary, the tracking error under the classic
PD control does not converge into a narrow region around zero. The phe-
nomenon of error convergence shows the superiority of the proposed control
in tracking performance.

On the other hand, when the system is affected by random disturbance,

the proposed control method can react quickly and maintain stable tracking
performance. On the contrary, the classical PD control is not strong enough
in dealing with random disturbance, and cannot quickly converge to a stable
state. From the perspective of tracking error, regardless of the random distur-
bance, the proposed control method has better tracking accuracy. In terms of
the ability to deal with random disturbance, the proposed control method has
stronger robustness and stability.

In addition, from the performance index function of the system, we know

that a larger weight ratio µ/ν means that in practice we pay more attention
to the steady state performance of the system. At the same time, when the
system is in stable state, the tracking error of the system has a certain trend,
that is, it decreases as the weight ratio µ/ν increases. So we can conclude
that the tracking error in stable state will decrease as the weight ratio µ/ν
increases when t → ∞. This regularity also proves the validity of the system
performance index function.
Title Suppressed Due to Excessive Length 31

Accumulative trajectory error of X-axis /m*s

10-3

0
PD wtih disturbance
/ =1
/ =5 / =10
/ =0.1 without disturbance
/ =0.2

Fig. 13: Accumulative trajectory error of X-axis under different control

Accumulative trajectory error of Y-axis /m*s

10-3

0
PD wtih disturbance
/ =1
/ =5
/ =10 without disturbance
/ =0.1
/ =0.2

Fig. 14: Accumulative trajectory error of Y-axis under different control

/rad*s a1

0.06
Accumulative error of

0.04

0.02

0
PD wtih disturbance
/ =1
/ =5
/ =10 without disturbance
/ =0.1
/ =0.2

Fig. 15: Accumulative error of ϑa1 under different control

32 Jiang Han1,2 et al.

/rad*s a2
0.04

Accumulative error of
0.02

0
PD
/ =1 wtih disturbance
/ =5
/ =10
/ =0.1 without disturbance
/ =0.2

Fig. 16: Accumulative error of ϑa2 under different control

/rad*s a3

0.02
Accumulative error of

0.01

0
PD
/ =1
/ =5 wtih disturbance
/ =10
/ =0.1 without disturbance
/ =0.2

Fig. 17: Accumulative error of ϑa3 under different control

From Fig. 13 to Fig. 17, the most obvious difference is that the accumula-
tive errors of the proposed control method are much smaller than the classic
PD control. Regardless of random disturbance, the proposed adaptive robust
control is superior to the classical PD control in terms of accumulative errors.
This means that the tracking accuracy of the proposed control method is high-
er and the ability to deal with external uncertainty is stronger in the whole
process. Comparing the system with disturbance, the proposed control method
has a smaller accumulative error when there is no disturbance. However, the
classic PD control has no such regularity, so the proposed adaptive robust con-
trol is more stable. Generally speaking, the proposed adaptive robust control
has better control performance under different performance indexes.

7 Conclusion

In this work, a novel dynamic modeling method and a systematic control

design approach for a planar 2-DOF redundantly actuated parallel robot are
proposed. Based on Udwadia-Kalaba equation, the constraint force imposed
by the end effector can be calculated more effectively. This novel analytical
mechanics approach can calculate the constraint force (holonomic constraints
Title Suppressed Due to Excessive Length 33

and/or nonholonomic constraints) in an analytical form without any auxiliary

variables such as Lagrange’s multiplier. Combining the closed-chain constraint
force and the dynamic equation of the open-chain system, we can obtain the
complete dynamic model in the active joint space. In order to ensure the
performance of the system under external uncertainty, an adaptive robust
control is proposed to guarantee the uniform boundedness and the uniform
ultimate boundedness. The unknown external uncertainty is described by the
fuzzy set theory and the bound of uncertainty is described by an estimated
value that satisfies a given adaptive law. The proposed adaptive robust control
consists of an adaptive robust control part and a PD control part. In order
to seek suitable weighting factors for these two parts, a system performance
index function is adopted, which includes the transient portion, the steady
state portion, and the control cost. The problem of finding the optimal gain
can be transformed into the problem of finding the minimum value of the
system performance index function. By analysis, there is a unique solution for
the optimal gain. We conclude that the proposed adaptive robust method has
better tracking accuracy, more rapid response, and stronger robustness than
the traditional control method. The simulation results prove the superiority
of the proposed adaptive robust control. Meanwhile, the regularity reflected
in the simulation results also proves the validity of the system performance
index function.

8 Acknowledgement

This project is supported by National Natural Science Foundation of China

under Grant No. 51905140, National Key R&D Program of China under Grant
No. 2018YFB1308400, Natural Science Foundation of Anhui Province under
Grant No. 1908085QF267, University Synergy Innovation Program of Anhui
Province under Grand No.GXXT-2019-031, Fundamental Research Funds for
the Central Universities of China under Grant No. PA2020GDSK0090.

9 Conflict of Interest

The authors declare that they have no conflict of interest.

References

1. Refaat, S., Herv, J., Nahavandi, S., Trinh, H.: Two-mode overconstrained three-DOFs
rotational-translational linear-motor-based parallel-kinematics mechanism for machine
tool applications. Robotica 25, 461–466 (2007)
2. Jamwal, P., Xie, S., Hussain, S., Parsons, J.: An adaptive wearable parallel robot for
the treatment of ankle injuries. Mechatronics, IEEE/ASME Transactions on 19, 64–75
(2014)
3. Correa, J., Toombs, J., Toombs, N., Ferreira, P.: Laminated micro-machine: Design
and fabrication of a flexure-based delta robot. Journal of Manufacturing Processes 24,
370–375 (2016)
34 Jiang Han1,2 et al.

4. Wu, F.X., Zhang, W.J., Li, Q., Ouyang, P.R.: Integrated design and pd control of high-
speed closed-loop mechanisms. Journal of Dynamic Systems Measurement and Control
124(4), 522–528 (2002)
5. Shang, W., Cong, S.: Nonlinear adaptive task space control for a 2-DOF redundantly
actuated parallel manipulator. Nonlinear Dynamics 59(1-2), 61–72 (2010)
6. Fichter, E.: A stewart platform-based manipulator: General theory and practical con-
struction. The International Journal of Robotics Research 5(2), 157–182 (1986)
7. Pang, H., Shahinpoor, M.: Inverse dynamics of a parallel manipulator. Journal of Robot-
ic Systems 11(8), 693–702 (1994)
8. Dasgupta, B., Mruthyunjaya, T.: A Newton-Euler formulation for the inverse dynamics
of the Stewart platform manipulator. Mechanism & Machine Theory 33(8), 1135–1152
(1998)
9. Hui, J., Pan, M., Zhao, R., Luo, L., Wu, L.: The closed-form motion equation of redun-
dant actuation parallel robot with joint friction: an application of the udwadia-kalaba
approach. Nonlinear Dynamics 93, 689–703 (2018)
10. Alain, Codourey: Dynamic modeling of parallel robots for computed-torque control
implementation. International Journal of Robotics Research 17(12), 1325–1336 (1998)
11. Zhao, R., Li, M., Niu, Q., Chen, Y.H.: Udwadia-kalaba constraint-based tracking control
for artificial swarm mechanical systems: dynamic approach. Nonlinear Dynamics 100(4),
2381–2399 (2020)
12. Singh, N., Yadav, S.K.: Comparison of LQR and PD controller for stabilizing double
inverted pendulum system. International Journal of Engineering Research and Devel-
opment 1(12), 69–74 (2012)
13. Choi, H.B., Konno, A., Uchiyama, M.: Design, implementation, and performance eval-
uation of a 4-DOF parallel robot. Robotica 28(1), 107–118 (2010)
14. Liu, Q., Liu, D., Meng, W., Zhou, Z., Ai, Q.: Fuzzy sliding mode control of a Multi-
DOF parallel robot in rehabilitation environment. International Journal of Humanoid
Robotics 11(1), 1223–1251 (2014)
15. Meryam, Rachedi, Boualem, Hemici, Mohamed, Bouri: Design of an H∞ controller for
the delta robot: experimental results. Advanced Robotics 29(18), 1–17 (2015)
16. Sun, Q., Yang, G., Wang, X., Chen, Y.H.: Cooperative game-oriented optimal design in
constraint-following control of mechanical systems. Nonlinear Dynamics 101(7), 977995
(2020)
17. Sun, Q., Wang, X., Chen, Y.H.: Adaptive robust control for dual avoidancearrival per-
formance for uncertain mechanical systems. Nonlinear Dynamics 94, 759–774 (2018)
18. Ramrezeria, M., Siraamez, H., Luvianourez, A., Rodrguez0geles, A.: Active disturbance
rejection control applied to a delta parallel robot in trajectory tracking tasks. Asian
Journal of Control 17(2), 636–647 (2015)
19. Natal, G.S., Chemori, A., Pierrot, F.: Nonlinear control of parallel manipulators for very
high accelerations without velocity measurement: stability analysis and experiments on
par2 parallel manipulator. Robotica 34(01), 43–70 (2016)
20. Castaneda, L.A., Luviano-Juarez, A., Chairez, I.: Robust trajectory tracking of a delta
robot through adaptive active disturbance rejection control. IEEE Transactions on
Control Systems Technology 23(4), 1387–1398 (2015)
21. Angel, L., Viola, J.: Fractional order pid for tracking control of a parallel robotic ma-
nipulator type delta. ISA Transactions 79, 172–188 (2018)
22. Chen, Y.H.: A new approach to the control design of fuzzy dynamical systems. Journal
of Dynamic Systems Measurement and Control 133(6), 061019 (2011)
23. Dong, F., Zhao, X., Han, J., Chen, Y.H.: An optimal fuzzy adaptive control for uncertain
flexible joint manipulator based on D-Operation. IET Control Theory & Applications
12(9), 1286–1298 (2018)
24. Dong, F., Han, J., Chen, Y.H., Xia, L.: A novel robust constraint force servo control
for underactuated manipulator systems: Fuzzy and optimal. Asian Journal of Control
20(5), 1818–1838 (2018)
25. Zhao, X., Chen, Y.H., Zhao, H., Dong, F.f.: Udwadia-kalaba equation for constrained
mechanical systems: Formulation and applications. Chinese Journal of Mechanical En-
gineering 31(6), 1–14 (2018)
26. Udwadia, F., Kalaba, R.: Analytical Dynamics: A New Approach. Cambridge University
Press (2002)
Title Suppressed Due to Excessive Length 35

27. Shang, W., Cong, S., Zhang, Y., Liang, Y.: Active joint synchronization control for
a 2-DOF redundantly actuated parallel manipulator. IEEE Transactions on Control
Systems Technology 17(2), 416–423 (2009)
28. Chen, Y.H.: Performance analysis of controlled uncertain systems. Dynamics and Con-
trol 6(2), 131–142 (1996)
Figures

Figure 1

2-DOF redundantly actuated parallel robot

Figure 2

subsystem of the parallel robot

Figure 3

The main design procedure of the proposed control scheme

Figure 4

System’s output under random input

Figure 5

The unique solution of @J/@α = @J/β = 0

Figure 6

The system performance index function J with respect to α, β

Figure 7

Actual tracking effect under different control

Figure 8

Comparison of tracking error of X-axis under different control

Figure 9

Comparison of tracking error of Y-axis under different control

Figure 10

Comparison of tracking error of υa1 under different control

Figure 11

Comparison of tracking error of υa2 under different control

Figure 12

Comparison of tracking error of υa3 under different control

Figure 13

Accumulative trajectory error of X-axis under different control

Figure 14

Accumulative trajectory error of Y-axis under different control

Figure 15

Accumulative error of υa1 under different control

Figure 16
Accumulative error of υa2 under different control

Figure 17

Accumulative error of υa3 under different control