See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/354162491

Evaluating the Quality of Intelligent Controllers for 3-DOF Delta Robot

Control

Article · August 2021

DOI: 10.18178/ijmerr.10.10.542-552

CITATIONS READS

4 317

5 authors, including:

Minh-Thanh Le Chi-Ngon Nguyen

Vinh Long University of Technology and Education Can Tho University
9 PUBLICATIONS 25 CITATIONS 145 PUBLICATIONS 466 CITATIONS

SEE PROFILE SEE PROFILE

Tung Thanh Pham Luong Hoai Thuong

20 PUBLICATIONS 10 CITATIONS 5 PUBLICATIONS 16 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Chi-Ngon Nguyen on 27 August 2021.

The user has requested enhancement of the downloaded file.

 International Journal of Mechanical Engineering and Robotics Research Vol. 10, No. 10, October 2021

Evaluating the Quality of Intelligent Controllers

for 3-DOF Delta Robot Control
Le Minh Thanh, Luong Hoai Thuong, Pham Thanh Tung
Vinh Long University of Technical Education, Vietnam
Email: thanhlm@vlute.edu.vn, {thuonglh, tungpt}@vlute.edu.vn

Cong-Thanh Pham
Viet Nam Aviation Academy, Vietnam
Email: thanhpc@vaa.edu.vn

Chi-Ngon Nguyen *
Can Tho University, Vietnam
Email: ncngon@ctu.edu.vn

Abstract—Delta robots have been successfully researched with the fixed plate
and manufactured in many countries. In this paper, the Bi m The connection point of the upper legs
authors will research and compare many different with the lower legs
controllers to control the delta robot so that it works Di m The connection point of the lower legs
stability when changing the working speed and load. The with the moving plate
regression fuzzy neural network along with PID controller
(RFNNC-PID) is used to observe the output error Abbreviation
parameters of the robot through the identifier to update and PID Proportional Integral Derivative
adjust the optimal input parameters to control the robot, DOF Degrees of freedom
contributing error reduction of the closed-loop control RFNNC Recurrent Fuzzy Neural Network Controller
system. The advantage of this controller is that it does not RFNNI Recurrent Fuzzy Neural Network Identifier
care about the robot's mathematical model and the RFNNC-
PID controller has been successfully simulated by the
authors in MATLAB/Simulink through the robot's
I. INTRODUCTION
trajectory control. The proposed controller will be Parallel robot is a kind of closed multi-loop
compared to the single neuron PID controller and the mechanism and is widely used in industrial fields thanks
traditional PID controller in MATLAB/Simulink. The
to its high rigidity, high precision and outstanding weight
simulation results show that the proposed controller is
better than the single neuron PID controller and the and load ratio [1], [2]. Parallel Delta robot was adopted in
traditional one with obtaining response time about 3.8 ± 0.1 many complex fields, for example: microelectronics [3],
(s) and without steady-state error.  [4], medicine [5], [6], logistics intelligence [7], [8], and
3D printing [9], [10]. In those complex applications, the
Index Terms—Delta robot, single neural PID, recurrent high-precision control of Delta parallel robots has
fuzzy neural network, Identifier, trajectory tracking. become an important issue for researchers. One of these
studies is the PD and LQR controllers presented by Joao
Symbol Unit Meaning Fabian [11]. In this paper, Joao Fabian used embedded NI
1 , 2 , 3 rad/s Angles the upper leg of the robot
my RIO hardware programmed in LabView software to
1 ,  2 , 3 rad/s Passive angles determine the position compare the two PD controllers and LQR robot trajectory
of the connection points controllers. The nonlinear PID shown by the HAN [12] is
R m Radius of the fixed plate used to replace increment scheduling with a nonlinear
r m Radius of the moving plate gain function by introducing a continuous dynamic
L1 m Length of upper arm
L2 m Length of upper arm
nonlinear function to achieve better noise cancel-lation
m1 kg Mass of upper arm and better tracking, this is achieved by the synthesis of a
m2 kg Mass of lower arm function consists of a linear function close to zero error
mP kg Mass of the moving plate and a nonlinear function far from zero error. The third
O m The center of the fixed plate technique is using a PD controller with intelligent
P m The center of the moving plate compensation that is used to solve the trap trajectory
Ai m The connection point of the upper legs tracking for a parallel delta robot with three degrees of
freedom presented by Ahmed Chemori [13]. Ahmed
Manuscript received November 16, 2020; revised March 25, 2021. Chemori tested a PID controller that performs the
(*) Corresponding author: Chi-Ngon Nguyen, Can Tho University,
Email: ncngon@ctu.edu.vn

© 2021 Int. J. Mech. Eng. Rob. Res 542

doi: 10.18178/ijmerr.10.10.542-552
 International Journal of Mechanical Engineering and Robotics Research Vol. 10, No. 10, October 2021

reference trajectory tracking of a real delta robot with A. Kinetic Models of the 3-DOF Delta Robot
extremely high acceleration up to 100 G as shown in [14] The dynamics of the delta robot are shown in Fig. 1
Recently, several controllers are proposed to control which consists of two equal triangles (fixed plate and
the parallel robot. In [15], R. Anoop and K. Achu movable plate).
designed and verified the performance of the adaptive
PID driving a parallel Delta controller. In that work, the
controller was able to track the desired trajectory without
any discouragement. In [16], Zhang et al. studied the
problem of dynamic control for redundantly actuated
planer 2-DOF parallel manipulator. The work proposed
an augmented PD controller based on the forward
dynamic compensation control technique, which showed
better performance when compared to conventional PD
controls. In [17], Hussein Saied et al. proposed different
model-based (augmented PD and adaptive feedforward Figure 1. Rotation angle and position of delta robot.
with PD) controllers and non-model-based (PD, PID, and
nonlinear PD) controllers for a 4-DOF parallel VELOCE The common angles are 1 ,  2 ,  3 , the point P is the
robot. Experimental results indicated that the nonlinear end point location with coordinates (xP, yP, zP). To
control method can achieve a superior performance [18]. calculate the inverse kinetics of delta robot, we have to
Su et al. developed a nonlinear proportional integral know in advance the position of the end point P(xP, yP,
derivative algorithm in link space to achieve high
zP), from there we go back to the opposite angle 1 ,  2 ,
precision tracking control for a general 6-DOF parallel
manipulator.  3 . Conversely, to find the forward dynamics of delta
The PID controllers have been successfully developed robot, we have to know the angles in advance, and then
for 6-DOF robots. However, when changing robot’s we find the endpoint position P (xP, yP, zP)
parameters such as load, input coupling and friction, the 1) Reverse kinetics of delta robot
PID controllers are difficult to archive control criteria by The reverse kinetics in this paper require to obtain the
their fixed parameters. So that, the neural networks and desired angular position of the actuator given the desired
fuzzy logic are applied to improve controlling of Delta endpoint of the effect in Cartesian space. This is a
robot [19]. The RFNN controllers and identifiers have geometrical method, so weight and moment of inertia are
been developed and applied for nonlinear systems [20-24]. not considered during modeling.
During control process, the RFNN identifier can estimate The physical geometry and the real robotic delta model
the object's sensitivity, called Jacobian information. That are shown in Fig. 2
information is used for online training the RFNN
controller. Therefore, by using RFNN, the control
technique is flexible to adjust its parameters online
adapting to the control conditions.
In this paper, the group of authors makes two main
contributions. Firstly, the team set up the motion equation P (xP, yP, zP)
of the parallel delta robot (including the equations of the
forward kinetics, the reverse kinetics, the delta robot
dynamics) and successfully simulated the motion
equations of the delta robots on MATLAB / Simulink.
Secondly, the optimal RFNNC-PID is built and
The geometrical shape of the delta Realistic delta robot model
successfully simulated to control the trajectory of the robot made by authors
delta robot on an ellipse curve and eight curve in
Figure 2. The physical geometry and the real robotic delta model.
MATLAB / Simulink and make sure this trajectory is
always stable with the changes of control conditions.
In Fig. 2: R is the radius of the fixed plate, r is the
This paper is organized consists of five sections. An
radius of the moving plate, f represents the length of the
introduction is Section I; Section II presents the construct
equilateral triangle of the upper platform, e is the length
mathematical models of the 3-DOF delta robot; trajectory
of the equilateral side of the lower-moving disc, L1 the
tracking control of the 3-DOF delta robot is presented in
upper leg length, L2 length of the lower leg,
Section III; compare the simulation results of the three
controllers are presented in Section IV and Section V is P  xP , yP , zP  is the center of the low-moving disc,
the conclusion.  
D1 xD1 , yD1 , zD1 is the midpoint of the equilateral triangle

II. CONSTRUCT MATHEMATICAL MODELS OF THE 3- of the lower-moving disc, and D1 is the connection point
DOF DELTA ROBOT of the lower leg to the motion plate, 0 (0,0,0) is the center
of the top fixed plate, the point A1 is the midpoint of the
equilateral triangle of the top fixed plate, and A1 is the
connection point of the upper leg to the fixed plate. The

© 2021 Int. J. Mech. Eng. Rob. Res 543

 International Journal of Mechanical Engineering and Robotics Research Vol. 10, No. 10, October 2021

angles 1 ,  2 ,  3 , are the driving angles of the three location of the endpoint P  xP , yP , zP  , then we find the
mechanical arms, point B1 is the intersection point angle 1 as shown in Fig. 4
between the upper leg and the lower leg of the delta robot.
Due to the structure of the delta robot, A1B1 can only
rotate around the YZ axis to form a circle with the
center point A1 and radius L1. In contrast to A1B1, point D1
is seen as a composite joint, meaning that D1B1 can rotate
freely depending on point D1 forming a sphere with
center D1 and radius L2.
The intersections between a sphere and a circle are
shown in Fig. 3

Figure 4. The YZ-axis and basic dimensions for finding angle θ1.

From Fig. 4 we have:

 zB 
1  arctan  1 ? (3)
  
 y A1  yB1 
Solving equations (1) and (2) and then substituting (3)
we find the angle 1 .
Due to the algebraic simplicity is a good choice of the
reference frame: A1B1 joint moves in the YZ-axis, so we
Figure 3. The geometric intersection between a sphere and a circle.
can completely ignore the X axis. To take advantage of
In the Fig. 3 points D’1 is the projection of point D1 on this for the rest we find the angle  and  , and since the
2 3

the YZ plane. The point B1 can be found at the
intersection point of two circles C1 and C2 with the
centers are A1 and D'1 respectively, and the radius L1 and
D'1B1 respectively, if we find B1 then we can Calculates
symmetry of the delta robot is used, let's first rotate the
system coordinates in the XY plane around the Z-axis
through an angle of -120 degrees counterclockwise as
shown in Fig. 5 to get the angle  , and we continue to
2

the rotation angle of the robot and B1 is found the rotate the system coordinates in the XY plane around the
following circle formulas Z-axis through an angle of +120 degrees clockwise as
Equation of the circle C1: shown in Fig. 5 so we get the angle  3

x    z 
2 2
B1  yA1  yB1  yA1 B1  z A1 L
2
1 (1)


With points A1  0,
f
 2 3 

, 0  and B1 0, yB1 , zB1   are

found from Fig. 3, then deduced:

2
 f 
 yB1    zB1  L1
2 2

 2 3 
Equation of the circle C2:

y  z   D' B   L22   D1' D1 

2 2 2
 yD '1  zD '1
2
(2) Figure 5. Symmetrical rotation and coordinates of delta robot.
B1 B1 1 1

 e  Rotate anticlockwise at an angle -1200 then 1 →  2 ;

With points D1'  x P , y P  , z P  and D1' D1  xP , A1→ A’1 and B1→ B’1
 2 3 
Rotate clockwise at an angle +1200 then 1 →  3 ;
then deduced:
A1→ A’’1 and B1→ B’’1
2
 e  We have the following axis conversion formula:
 
2
 yB1  yP    zB1  zP  L22  xP2
 2 3 xPi  xP cos  1200   yP sin  1200 

We need to find the point B1 0, yB1 , zB1   through yPi  yP sin  1200  xP cos  1200  (4)
equations (1) and (2), and if we know in advance the z Pi  z P

© 2021 Int. J. Mech. Eng. Rob. Res 544

 International Journal of Mechanical Engineering and Robotics Research Vol. 10, No. 10, October 2021

With i = 2, 3 Bi Di  L2 , i Bi  L1
From the axis conversion formula we deduce the
formula of angle  2 as follows: where point i ∈ {1, 2, 3} refers to pins on 1, 2, 3 and Σ is
the global reference system, Σi is the reference system for
xP2  xP cos  1200   yP sin  1200  the above pins, Σeff is the reference system for the end
yP2  yP sin  1200   xP cos  1200  (5) points, Bi is the connection point between the upper and
lower legs, Di is the connection point between the lower
z  zP
2
P leg and the fixed plate,  i are the angles of the upper legs
and  i are the angles separating each upper leg and the
With i = 2, Similar to finding angle 1 , we substitute (5)
XY plane.
into (1) and (2) and solve the solution for solution, then
Lower leg length can be calculated using the following
we replace the solution we just found in (6) we find angle
vector relation:
2 :
  Di Bi  Di   Bi  (9)
zB '1
 2  arctan   (6)
 
 y A '1  yB '1  

Equation (9) can be written in terms of the geometrical
length of the robot and the position Di with due the
We have the following formula for changing the axis correlation to reference Σ, as follows:
of angle  3 :
 L1 cos i     n cos  i    xDi 
     
xP3  xP cos 1200   yP sin 1200  Di Bi  Ri  0     n sin  i     yDi  (10)
 L1 sin i      z 
yP3  yP sin 1200   xP cos 1200  (7) 0
 Di 
z  zP
3
P Inside, Δn = m - n and Ri is the rotation matrix
With i = 3 between the Σi and Σ of the reference system
We substitute (7) into (1) and (2) and solve the The Ri contained in the equation (10) is bound by the
solution for solution, then we replace the solution we just
found in (8) we find angle  3 following length:

 L22 , with i =1,2,3

2
  Di Bi (11)
zB ''1
3  arctan   (8)
 
 y A ''1  yB ''1  

Equation (11) is rewritten as follows:

x  y  z 
2 2 2
Constructing the inverse kinetic block of the delta Di  xBi Di  yBi Di  zBi  L22 (12)
robot in Simulink by giving the endpoint position
P  xP , yP , zP  , we will find the angles 1 ,  2 ,  3 of the With i = 1, 2, 3
Substituting equation (12) for (10) gives the following
delta robot through the formulas (3), (6) and (8). formula:
2) The forward kinetics of delta robot
A forward motion is used to reach the end position xBi    n  L1 cos i    cos  i 
using the upper leg angles. The representative geometry
of a robot leg is shown in Fig. 6 [25] yBi    n  L1 cos i    sin  i  (13)
zBi  L1 sin  i 

From equation (13), if we give the angles  i and  i

then we will find the position B1 xB1 , yB1 , zB1   of delta
robot. After finding the point B1 xB1 , yB1 , zB1   from
equation (13), then we replace equation (12) solve

equation finding for point D1 xD1 , yD1 , zD1 , we deduce 
the end point location P  xP , yP , zP 
Constructing the forward kinetic block of the delta
Figure 6. Geometrical parameters of a robotic delta leg. robot in Simulink by giving the corners 1 ,  2 ,  3 of the
From Fig. 7, we have a global variable like this: delta robot, we will find the endpoint position
P  xP , yP , zP  of the delta robot through the formulas (12)
 i  m , eff Di  n and (13).

© 2021 Int. J. Mech. Eng. Rob. Res 545

 International Journal of Mechanical Engineering and Robotics Research Vol. 10, No. 10, October 2021

B. Dynamic Model of a 3-DOF Delta Robot 𝛿𝐴 = 𝜏1 𝛿𝜃1 + 𝜏2 𝛿𝜃2 + 𝜏3 𝛿𝜃3 (16)

The best model for a 3-DOF delta robot is a system of
So we have: Q1np  1 , Q2np   2 , Q3np   3 , cases Qknp  0
rigid bodies connected by joints. The parallelogram
mechanisms that connect the driving links to the mobile with k = 4, 5, 6. Substituting kinetic expressions,
platform are modeled as homogeneous rods with potentials and equations into equation (14), we get the
universal and spherical joints at two ends. By using this motion equation of the robot as the system. The
model, the robot is seen as a multibody system with seven differential equation-algebra is as follows:
bodies: three legs each leg having two links and the 1
(𝐼𝐼𝑦 + 𝑚𝑏 𝑙12 )𝜃̈1 = 𝑔𝑙1 ( 𝑚1 + 𝑚𝑏 ) 𝑐𝑜𝑠 𝜃1 + 𝜏1
mobile platform. The calculating models are shown in Fig. 2
7 [21], [27]. 𝑠𝑖𝑛 𝜃1 (𝑅 − 𝑟) − 𝑐𝑜𝑠 𝛼1 𝑠𝑖𝑛 𝜃1 𝑥𝑝
−2𝜆1 𝑙1 ( ) (17)
− 𝑠𝑖𝑛 𝛼1 𝑠𝑖𝑛 𝜃1 𝑦𝑝 − 𝑐𝑜𝑠 𝜃1 𝑧𝑝
1
(𝐼𝐼𝑦 + 𝑚𝑏 𝑙12 )𝜃̈2 = 𝑔𝑙1 ( 𝑚1 + 𝑚𝑏 ) 𝑐𝑜𝑠 𝜃2 + 𝜏2
2
𝑠𝑖𝑛 𝜃2 (𝑅 − 𝑟) − 𝑐𝑜𝑠 𝛼2 𝑠𝑖𝑛 𝜃2 𝑥𝑝
−2𝜆1 𝑙1 ( ) (18)
− 𝑠𝑖𝑛 𝛼2 𝑠𝑖𝑛 𝜃2 𝑦𝑝 − 𝑐𝑜𝑠 𝜃2 𝑧𝑝
1
(𝐼𝐼𝑦 + 𝑚𝑏 𝑙12 )𝜃̈3 = 𝑔𝑙1 ( 𝑚1 + 𝑚𝑏 ) 𝑐𝑜𝑠 𝜃3 + 𝜏3
2
𝑠𝑖𝑛 𝜃3 (𝑅 − 𝑟) − 𝑐𝑜𝑠 𝛼3 𝑠𝑖𝑛 𝜃3 𝑥𝑝
−2𝜆3 𝑙1 ( ) (19)
− 𝑠𝑖𝑛 𝛼3 𝑠𝑖𝑛 𝜃3 𝑦𝑝 − 𝑐𝑜𝑠 𝜃3 𝑧𝑝

(𝑚𝑝 + 3𝑚𝑏 )𝑥̈ 𝑝 =

Figure 7. Calculating models of a 3-DOF delta robot [21].
Putting  i , i = 1, 2, 3 be the driving angles of the
actuated links; 1 , i = 1, 2, 3 be the passive angles that
determine the position of the connecting rods; and
P  xP , yP , zP  be the position of the center of mass of
T (𝑚𝑝 + 3𝑚𝑏 )𝑧̈𝑝 =
the mobile platform. Hence, the position of robot is
determined by the generalized coordinates:
−2𝜆1 (𝑐𝑜𝑠 𝛼1 (𝑅 − 𝑟) + 𝑙1 𝑐𝑜𝑠 𝛼1 𝑐𝑜𝑠 𝜃1 − 𝑥𝑝 )
−2𝜆2 (𝑐𝑜𝑠 𝛼2 (𝑅 − 𝑟) + 𝑙1 𝑐𝑜𝑠 𝛼2 𝑐𝑜𝑠 𝜃2 − 𝑥𝑝 ) (20)
−2𝜆3 (𝑐𝑜𝑠 𝛼3 (𝑅 − 𝑟) + 𝑙1 𝑐𝑜𝑠 𝛼3 𝑐𝑜𝑠 𝜃3 − 𝑥𝑝 )
(𝑚𝑝 + 3𝑚𝑏 )𝑦̈𝑝 =
−2𝜆1 (𝑠𝑖𝑛 𝛼1 (𝑅 − 𝑟) + 𝑙1 𝑠𝑖𝑛 𝛼1 𝑐𝑜𝑠 𝜃1 − 𝑦𝑝 )
−2𝜆2 (𝑠𝑖𝑛 𝛼2 (𝑅 − 𝑟) + 𝑙1 𝑠𝑖𝑛 𝛼2 𝑐𝑜𝑠 𝜃2 − 𝑦𝑝 ) (21)
−2𝜆3 (𝑠𝑖𝑛 𝛼3 (𝑅 − 𝑟) + 𝑙1 𝑠𝑖𝑛 𝛼3 𝑐𝑜𝑠 𝜃3 − 𝑦𝑝 )
−(3𝑚𝑏 + 𝑚𝑝 )𝑔 + 2𝜆1 (𝑧𝑝 + 𝑙1 𝑠𝑖𝑛 𝜃1 ) (22)
+2𝜆2 (𝑧𝑝 + 𝑙1 𝑠𝑖𝑛 𝜃2 ) + 2𝜆3 (𝑧𝑝 + 𝑙1 𝑠𝑖𝑛 𝜃3 )

l12   cos 1  R  r   l1 cos 1 cos 1  x p 

2
𝑞 = [𝜃1 𝜃2 𝜃3 𝑥𝑃 𝑦𝑃 𝑧𝑃 ]

  sin 1  R  r   l1 sin 1 cos 1  y p 

2
Set the differential equation of motion of the 3-DOF (23)
delta robot
  l1 sin 1  z p   0
2
The motion equation is established by using the
Lagrange formula:
l22   cos  2  R  r   l1 cos  2 cos  2  x p 
2
𝑑 𝜕𝑇 𝜕𝑇 𝜕𝑓𝑖
( )− = 𝑄𝑘 − ∑𝑟𝑖=1 𝜆𝑖 (𝑘 = 1,2, … . , 𝑚) (14)
𝑑𝑡 𝜕𝑞̇ 𝑘 𝜕𝑞𝑘 𝜕𝑞𝑘
  sin  2  R  r   l1 sin  2 cos  2  y p 
2

where qk is the extrapolation coordinates of the robot, fi is (24)

the linking equations, Qk is the extrapolation force, i is   l1 sin  2  z p   0
2

the Lagrange factor. With this model, the vector of

extrapolation coordinates   R6 and the number of
l32   cos  3  R  r   l1 cos  3 cos 3  x p 
2

associated equations is three, so m = 6, r = 3. We divide

  sin  3  R  r   l1 sin  3 cos 3  y p 
the forces acting on the robot into potential forces and 2
(25)
forces without potential energy, the extrapolation force
  l1 sin 3  z p   0
2
Qk is calculated as follow

Qk    Qknp (15) From the motion equations of the parallel robot (17) to
qk (25), according to the proposal [21], [26], [27] we will
construct the model of the 3-DOF Delta robot in
In it Qknp are extrapolation forces corresponding to MATLAB/Simulink.
forces that are not possible. Virtual power of
extrapolation is not so:

© 2021 Int. J. Mech. Eng. Rob. Res 546

 International Journal of Mechanical Engineering and Robotics Research Vol. 10, No. 10, October 2021

III. TRAJECTORY TRACKING CONTROL OF THE 3-DOF

DELTA ROBOT

A. Delta Robot Control Using a Single Neuron PID

The control system for a robot arm is presented in Fig.
8

Figure 10. Structure of four-layer RFNN.

Layer 2 (Fuzzy layer): This layer consists of (2x5)

Figure 8. Controller structure [21]. nodes, each node representing a related function of the
Gaussian form with mean value mij and standard
The reference signal yref is the rotation angle of the deviation σij, and defined as (28).
upper three arms of the robot is sent to three adders
 2 
  Oi  k   mij  
giving three error signals e1, e2, e3 to three PID controllers
 1

one neuron, the output of three the controller was inserted O  k   exp 
2
 , i  1, 2; j  1, 2,...,5 (28)
 
ij 2
into Tau1, Tau2, Tau3 of the 3-DOF Delta Robot. At the  

same time, three RFNN identifiers will observe three 
 ij 

outputs of three single-neuron PID controllers and three
output signals y (theta1, theta2, theta3) of delta robot, At each node on the fuzzy layer, there are 2 parameters
each RFNN identifier will train the parameter Jacobian that are automatically adjusted during the online training
returned three single-neuron PID controllers to of the RFNN identifier, that is mij and σij.
continuously update the KP, KD, and KI parameters of the Layer 3 (Rule layer): This layer provides fuzzy
three controllers to control the delta robot following inferences. Each node corresponds to a fuzzy rule. The
the reference signal yref presented by the authors in [21] link before each node represents the prerequisites of the
respective rule. The qth the fuzzy law can be described:
B. The 3-DOF Delta Robot Control Using RFNNC-PID
1) The 3-DOF delta robot RFNN identifier Oq3  k    Oiq2  k , i  1, 2,...,5; qi  1, 2,...,5 (29)
i i
The identifier is shown in Fig. 9 that consists of 4
layers, with a 2-node input layer, a 10-node fuzzy layer, a Layer 4 (Output layer): This layer includes 1 linear
25-node fuzzy rule, and a 1-node output layer [28]-[32]. neuron with the defined output as follows:
Oi4  k    wij4O3j  k , i  1; j  1, 2,..., 25 (30)
j

where w ij4 is the connecting weights from 3rd layer to 4th

layer. The output of this layer is also the output of the
RFNN:
𝑂14 (𝑘) = 𝑦𝑚 (𝑘) = 𝑓̂[𝑥1 (𝑘), 𝑥2 (𝑘)] = 𝑓̂[𝑢(𝑘), 𝑦(𝑘 − 1)] (31)
The performance RFNN identifier training is based on
a cost function in (32) [20]:
1 1
EI  k    y  k   ym  k    y  k   OI41  k 
Figure 9. Delta robot model in MATLAB/Simulink. 2 2
(32)
2 2
Layer 1 (Input layer): This layer takes the input
variables and passes the input values to the next layer. The RFNN's weights are adjusted according to (33) [20]:
Feedback lines are added in this layer to embed time
WI  k  1  WI  k   WI  k 
relations into the network. The output of layer 1 is
represented as (26):  E  k   (33)
 WI  k    I   I 
Oi1  k   xi1  k   i1Oik  k  1 , i  1, 2 (26)  WI 
where, I is the learning rate, WI=[I, mI, σI, wI] T is the
With  i1 is the connection weight at the current time k.
weight vector of RFNN updated during training the
The input of the corresponding RFNN is the current RFNN. And subscript I presents the RFNN identifier,
control signal and the past output of the response: called RFNNI.
x11  k   u  k  , x12  k   y  k  1 (27)

© 2021 Int. J. Mech. Eng. Rob. Res 547

 International Journal of Mechanical Engineering and Robotics Research Vol. 10, No. 10, October 2021

Given eI(k)=y(k)-ym(k) is the error between plant’s 1

EC  k  = urfnnc  k  - u  k  
2
output and RFNN’s output, then the gradient EI(.) with 2
respect to WI is determined as (34): (40)
1
 urfnnc  k  - OC14  k  
2

EI  k  ym  k  OI41  k  2

 eI  k   eI  k  (34)
WI WI WI where u(k) is the sth input of the plant, urfnnc (k) is the sth
The weight of each RFNNI network layer is updated as output of RFNNC. After the initialization process, a
follows [21], [32]: gradient-descent-based back-propagation algorithm was
employed to adjust the controller parameters (40):
wI ij4  k  1  wIij  k  I I eI  k  OI 3i
4 w
(35)
WC  k  1  WC  k   WC  k  
mIij  k  1  mIij  k    EC  k   (41)
WC  k   C   
2 OIij  k   mIij 
 1  (36) 
 W 

I I  eI  k  wIik OIk3 
m 4 C

 Iij 
2
k
In which,  c is the learning rate, WC=[C, mC, σC, wC] T
is the weight vector of the RFNNC updated during
 Iij  k  1   Iij  k   training. Subscript C presents the RFNNC.
Given eC(k)=urfnn(k)-u(k), the gradient EC(.) with
2 OIij  k   mIij 
1 2
I (37)
I  eI  k  wIik OIk
4 3 respect to WC is defined as (42):
 Iij 
3

EC  k  u  k  O4  k 
k

 eC  k  rfnnc  eC  k  1 (42)

 Ii1  k  1   Ii1  k   WC WC WC

I  2  OIij1  k   mIij  OIij1  k  1 (38) The weight of each RFNNC network layer is updated
I  eI  k  wIik OIk
4 3
as follows [21], [30]:
k
  2Iij
 EC  k  
In addition, to estimate the output of the model ym(k), wC ij4  k  1  wCij
4
 k   C w   
 wCij 
4
(43)
the RFNNI must also estimate Jacobian information
 wC ij4  k   C u pid (k )eC  k  OC i3
wC
y (k ) , determined as follows [21], [31], [32]:
u (k )
 EC  k  
y  k   O3  2  O1  k   m   mC ij  k  1  mCij  k   C m   
4   Iij Iij  
  wIij  2
Iq
 (39)  mCij 
u  k  q  s OIqs  Iij 
2
 (44)
2 OCij
1
 k   mCij 
 mCij  k  C  eC  k  wC ik OC k
mC 4 3

 
2
2) Building controller of RFNNC-PID k
Cij
In this part, for controlling a robot arm, a RFNN
controller (RFNNC) is combined with a traditional PID  EC  k  
controller with its structure is shown in Fig. 11 [21]-[23].  Cij  k  1   Cij  k   C   
  Cij 
(45)
 k   mCij 
2
2 OCij
1

  Cij  k   C  eC  k  wCik OCk

C 4 3

 
3
k
C ij

 EC  k  
Ci1  k  1  Ci1  k   C     Ci  k  
1

 Ci1 
(46)
 2 OCij
1
 k   mCij  OCij
1
 k  1
  eC  k  w O
C 4 3

 
Cik Ck 2
k
Cij
Figure 11 Control system based on RFNNs.

In Fig. 11, the RFNNI is used to identify the model of IV. COMPARE THE SIMULATION RESULTS OF THE
the robot, based on y(k-1) and u(k). That makes the THREE CONTROLLERS
RFNN model is simple and decreases number of neurons.
Through training, the RFNNI estimates the output
A. Simulation Parameters of Three Controllers
trajectories of the delta robot by (30).
Training performance criterion is defined as (40) [20]: The specifications of the real robot that the authors
have built are shown in Table I.

© 2021 Int. J. Mech. Eng. Rob. Res 548

 International Journal of Mechanical Engineering and Robotics Research Vol. 10, No. 10, October 2021

TABLE I. DELTA ROBOT MECHANICAL SPECIFICATIONS The parameters of the Single Neural PID Algorithms,
Symbol Value Unit and Recurrent Fuzzy Neural Network Controller, and
α1 0 rad/s Recurrent Fuzzy Neural Network Identifier are randomly
α2 2π/3 rad/s initialized.
α3 4π/3 rad/s
R 138.9 mm B. Simulation Results
r 25 mm
f 481.07 mm The MATLAB/Simulink control system for a 3-DOF
e 43.3 mm delta robot with the desired trajectory as an ellipse curve
L1 250 mm (47).
L2 544 mm
x d = 0.19sin( t)+0.3
m1 0.258 kg
m2=2mb 0.088 kg y d = 0.12cos( t)+0.2 (47)
mp 0.044 kg
z d = -0.8

Figure 12. The 3-DOF Delta Robot controller in MATLAB/Simulink.

Figure 13. The angles of theta 1. Figure 15. The angles of theta 3.

Figure 14. The angles of theta 2. Figure 16. Errors of responses.

© 2021 Int. J. Mech. Eng. Rob. Res 549

 International Journal of Mechanical Engineering and Robotics Research Vol. 10, No. 10, October 2021

Fig. 13, Fig. 14, Fig. 15 and Fig. 16 comparisons
angles theta of three traditional PID, Single-Neuron PID
controller and RFNN-PID controller.
Fig. 18 including load changed. Simulation results show
that the RFNN-PID controller is better than Single neuron
PID controller, with the setting time is about 3. 8±0. 1
seconds, and the steady-state error is eliminated.
The MATLAB/Simulink control system for a 3-DOF
delta robot with the desired trajectory as a Fig. 8 is shown
(48).
x d = 0.15sin(2 t)+0.3
y d = 0.15sin(2 t)*cos(2 t)+0. 2 (48)
z d = -0.7

Figure 17. Trajectory tracking of traditional PID and Single Neuron

PID and RFNN-PID controller.

Figure 19. Trajectory tracking of the Fig. 8.

And the control criteria of PID controller, Single

Figure 18. Responses when changing load from 4.71 Kg to 7.05 Kg.
The response of the traditional PID, Single neuron PID
and the RFNN-PID controllers are presented in Fig. 13 –
Neuron PID and RFNNC-PID controller are presented in
Table II. The results show that the proposed controller
must be better than the Single Neuron PID controller and
traditional PID controller.

TABLE II. SYSTEM CONTROL QUALITY STANDARDS

Respons PID Single Neuron PID RFNN-PID
e
Rise Overshoot Settling Rise Overshoot Settling Rise Overshoot Settling
time (s) (%) time (s) time (s) (%) time (s) time (s) (%) time (s)
Theta 1 3.134 1.873 5.6±0.1 3.119 1.799 4.2±0.1 2.942 0.299 3.9±0.1
Theta 2 2.838 1.882 6.9±0.1 2.889 1.992 3.6±0.1 3.346 1.953 3.8±0.1
Theta 3 2.698 1.859 4.2±0.1 2.705 1.973 3.9±0.1 2.447 0.409 3.6±0.1

V. CONCLUSION AUTHOR CONTRIBUTIONS

In this paper, the RFNN-PID controller is proposed to
control the 3-DOF delta robot. This controller guarantees
the real trajectory converges the reference trajectory with
finite time. The simulation results show that the proposed
controller is better than the single neuron PID controller
and the traditional PID controller. The proposed RFNNC-
PID algorithm is stable while changing the control
conditions such as the speed and load of the delta robot
increasing. In further work, the proposed controller will
be experimenting with the real robot model.
CONFLICT OF INTEREST
The authors declare no conflict of interest.
Mr. Le Minh Thanh, first author, is a PhD student
under supervising of Assoc. Prof. Dr. Chi-Ngon Nguyen
(last and corresponding author), who has prepared the
manuscript. Mr. Luong Hoai Thuong and Mr. Pham
Thanh Tung, 2nd and 3rd authors have contributed on
model simulation. Dr. Cong-Thanh Pham, 4th author, has
contributed on writing correction. Assoc. Prof. Dr. Chi-
Ngon Nguyen is chief of research group, who has
supervised for this study and finalized this paper.
REFERENCES
[1] G. Gao, M. Ye, and M. Zhang, "Synchronous robust sliding mode
control of a parallel robot for automobile electro-coating
conveying," IEEE Access, vol. 7, pp. 85838-85847, 2019.

© 2021 Int. J. Mech. Eng. Rob. Res 550

 International Journal of Mechanical Engineering and Robotics Research Vol. 10, No. 10, October 2021
[2] S. Qian, B. Zi, D. Wang, and Y. Li, “Development of modular
cable-drivenn parallel robotic systems,” IEEE Access, vol. 7, pp.
5541–5553, 2019.
[3] J. E. Correa, J. Toombs, N. Toombs, and P. M. Ferreira,
“Laminated micromachine: Design and fabrication of a flexure-
based Delta Robot,” J. Manuf. Processes, vol. 24, pp. 370–375,
Oct. 2016.
[4] X. J. Liu, J. I. Jeong, and J. Kim, “A three translational DOFs
parallel cube manipulator,” Robotica, vol. 21, no. 6, pp. 645–653,
Dec. 2003.
[5] K. C. Olds, “Global indices for kinematic and force transmission
performance in parallel robots,” IEEE Trans. Robot., vol. 31, no. 2,
pp. 494–500, Apr. 2015.
[6] G. Yedukondalu, A. Srinath, and J. S. Kumar, “Mechanical chest
compression with a medical parallel manipulator for
cardiopulmonary resuscitation,” Int. J. Med. Robot. Comput. Assist.
Surgery, vol. 11, no. 4, pp. 448–457, Oct. 2014.
[7] G. Borchert, M. Battistelli, G. Runge, and A. Raatz, “Analysis of
the mass distribution of a functionally extended delta robot,”
Robot. Comput-Integr. Manuf., vol. 31, pp. 111–120, Feb. 2015.
[8] R. Kelaiaia, “Improving the pose accuracy of the Delta robot in
machining operations,” Int. J. Adv. Manuf. Technol, vol. 91, no. 5–
8, pp. 2205–2215, July 2017.
[9] E. Rodriguez, C. Riaño, A. Alvares, and R. Bonnard, “Design and
dimensional synthesis of a linear delta robot with single legs for
additive manufacturing,” J. Brazilian Soc. Mech. Sci. Eng., vol. 41,
no. 11, p. 536, Nov. 2019.
[10] K. He, Z. Yang, Y. Bai, J. Long, and C. Li, “Intelligent fault
diagnosis of Delta 3D printers using attitude sensors based on
support vector machines,” Sensors, vol. 18, no. 4, p. 1298, Apr.
2018.
[11] J. Fabian, C. Monterrey and R. Canahuire, “Trajectory tracking
control of a 3 DOF delta robot: a PD and LQR comparison,” in
Proc. 2016 IEEE XXIII. Inter. Congress on Electronics, Electrical
Engineering and Computing (INTERCON), Piura, pp. 1-5, 2016.
[12] J. Han, “From PID to active disturbance rejection control,” IEEE
Transactions on Industrial Electronics, vol. 56, no. 3, pp. 900-906,
2009.
[13] J. M. E. Hernández, H. Aguilar-Sierra, O. Aguilar-Mejí a, A.
Chemori, J. Arroyo-Núñez, “An intelligent compensation through
B-spline neural network for a delta parallel robot,” in Proc. 6th
Inter. Conf. on Control, Decision and Information Technologies
(CoDIT), Paris, France, pp. 361-366, 2019.
[14] A. Chemori, G. S. Natal, F. Pierrot, “Control of parallel robots:
Towards very high accelerations. SSD,” Systems, Signals and
Devices, Mar. 2013, Hammamet, Tunisia. pp. 8, ffirmm-00809514
2013.
[15] R. Anoop and K. Achu, “Control technique for parallel
manipulator using PID,” Inter. J. of Engineering Research &
Technology, vol. 5, no. 7, pp. 56–59, 2016.
[16] Y. X. Zhang, S. Cong, W. W. Shang, Z. X. Li, and S. L. Jiang,
“Modeling, identification and control of a redundant planar 2-DOF
parallel manipulator,” Inter. J. of Control, Automation, and
Systems, vol. 5, no. 5, pp. 559–569, 2007.
[17] H. Saied, A. Chemori, M. El Rafei, C. Francis, and F. Pierrot,
“From non-model-based to model-based control of PKMS: a
comparative study,” in Proc. 1st Inter. Congress for the Advan. of
Mechanism, Machine, Robotics and Mechatronics Sciences,
pp.50–64, Beirut Lebanon, 2017.
[18] Y. X. Su, B. Y. Duan, and C. H. Zheng, “Nonlinear PID control of
a six-DOF parallel manipulator,” in IEEE Proc. Control Theory
and Applications, vol. 151, no. 1, pp. 95–102, 2004.
[19] W. Widhiada, T. G. T. Nindhia, and N. Budiarsa, “Robust control
for the motion five fingered robot gripper,” Inter. J. Mecha. Eng.
and Robotics Research, vol. 4, no. 3, pp. 226-232, 2015.
[20] J. K. Liu, “Radial Basis Function (RBF) neural network control
for mechanical systems,” Design, Analysis and Matlab Simulation.
Springer-Verlag Berlin Heidelberg, pp. 55-69, 2013.
[21] L. M. Thanh, L. H. Thuong, P. Th. Loc, C. N. Nguyen, “Delta
robot control using single neuron PID algorithms based on
recurrent fuzzy neural network Identifiers,” Inter. J. of Mechanical
Eng. and Robotics Research, vol. 9, no. 10, 2020.
[22] S. Slama, A. Errachdi, and M. Benrejeb, “Adaptive PID controller
based on neural networks for MIMO nonlinear systems,” J. of
Theoretical and Applied Information Tech., vol. 97, no. 2, pp.
361–371, 2019.
© 2021 Int. J. Mech. Eng. Rob. Res 551
[23] W. Sun, Y. N. Wang, “A recurrent fuzzy neural network based
adaptive control and its application on robotic tracking control,”
Neural Information Processing-Letters and Reviews, vol. 5, no. 1,
2004.
[24] H. Hasanpour, M. H. Beni, and M. Askari, “Adaptive PID control
based on RBF NN for quadrotor,” Inter. Research J. of Applied
and Basic Sciences, vol. 11, no. 2, pp. 177–186, 2017.
[25] J. Fabian, C. Monterrey, and R. Canahuire, "Trajectory tracking
control of a 3 DOF delta robot: A PD and LQR comparison," in
Proc. 2016 IEEE XXIII International Congress on Electronics,
Electrical Engineering and Computing (INTERCON), Piura, 2016,
pp. 1-5.
[26] J. Merlet, Parallel Robots, the Netherlands: Kluwer Academic
Publishers (2000).
[27] N. D. Dung, “Reverse dynamics of parallel delta space robots,”
PhD dissertation on Mechanical Engineering,” Vietnam National
Library, 2018.
[28] H. Hasanpour, M. H. Beni, and M. Askari: “Adaptive PID control
based on RBF NN for quadrotor,” Inter. Research J. of Applied
and Basic Sciences, vol. 11, no. 2, pp. 177–186, 2017.
[29] S. Slama, A. Errachdi, and M. Benrejeb, “Neural adaptive PID and
neural indirect adaptive control switch controller for nonlinear
MIMO systems,” Mathematical Problems in Engineering, vol.
2019.
[30] C. H. Lee and C. C. Teng, “Identification and control of dynamic
systems using recurrent fuzzy neural networks,” IEEE Transaction
on Fuzzy Systems, vol. 8, no.4, pp. 349-366, 2000.
[31] S. Wei, Z. Lujin, Z. Jinhai, and M. Siyi, “Adaptive control based
on neural network,” Adaptive Control, Kwanho You (Ed.), InTech
2009.
Copyright © 2021 by the authors. This is an open access article
distributed under the Creative Commons Attribution License (CC BY-
NC-ND 4.0), which permits use, distribution and reproduction in any
medium, provided that the article is properly cited, the use is non-
commercial and no modifications or adaptations are made.
Le Minh Thanh received a Bachelor of
Engineering degree in Electrical and
Electronic Engineering at Mekong University
in 2006, a Master's degree in Automation at
the University of Transport in Ho Chi Minh
City in 2011. He is now a lecturer of the
Faculty Electrical – Electronics, Vinh Long
University of Technical Education.
Luong Hoai Thuong received a Bachelor of
Engineering in Control Engineering from Can
Tho University in 2009, a Master's degree in
Electronic Engineering at Ho Chi Minh City
University of Technical Education in 2015. He
is currently a lecturer in the Faculty of
Electrical - Electronics, Vinh Long University
of Technical Education.
Cong-Thanh Pham received the Bachelor &
M.Sc. degree in automation and control
engineering from the Ho Chi Minh City
University of Technology, Vietnam, 2002 &
2006, the Ph.D. received degree at
Department of Control Science and
Engineering, Huazhong University of Science
and Technology (HUST), Wuhan 430074,
China, 2014. He is currently with the Faculty
of Automatiom Control, Viet Nam Aviation Academy, Ho Chi Minh
City. His current research interests include AC motor, Pes
 International Journal of Mechanical Engineering and Robotics Research Vol. 10, No. 10, October 2021

Thanh Tung Pham received degree in Chi-Ngon Nguyen received his B.S. and M.S.
Electrical and Electronic Engineering at degree in Electrical Engineering from Can
Mekong University in 2004, a Master's degree Tho University and Ho Chi Minh City
in Automation at Ho Chi Minh City University of Technology, in 1996 and 2001,
University of Transport in 2010. The degree respectively. The degree of Ph.D. was award
of Ph.D. was award by the Ho Chi Minh City by the University of Rostock, Germany, in
University of Transport, Vietnam, in 2019. 2007.
Nowadays, he has worked at Vinh Long Since 1996, he has worked at the Can Tho
University of Technical Education. University. Currently, he is an associate
professor in automation of the Department of
Automation Technology. He is working as a
position of Dean of the College of Engineering Technology at the
Can Tho University. His research interests are intelligent control,
medical control, pattern recognition, classifications, speech
recognition and computer vision.

© 2021 Int. J. Mech. Eng. Rob. Res 552

View publication stats