SINERGI Vol. 25, No.

3, October 2021: 237-244

http://publikasi.mercubuana.ac.id/index.php/sinergi
http://doi.org/10.22441/sinergi.2021.3.001

SELF-LEARNING OF DELTA ROBOT USING INVERSE KINEMATICS

AND ARTIFICIAL NEURAL NETWORKS
Zendi Iklima1, Muhammad Imam Muthahhar1, Asif Khan2, Arifiansyah Zody3
1
Department of Electrical Engineering, Faculty of Engineering, Universitas Mercu Buana, Indonesia
2
School of Computer Science and Technology, Beijing Institute of Technology, China
3
Department of Information Technology, Faculty of Computer Science, Esa Unggul University, Indonesia

Abstract Keywords:
As known as Parallel-Link Robot, Delta Robot is a kind of Manipulator Artificial Neural Network;
Robot that consists of three arms mounted in parallel. Delta Robot Delta Robot;
has a central joint constructed as an end-effector represented as a Inverse Kinematics;
Motion Data;
gripper. An Analysis of Inverse Kinematic (IK) used to convert the Socket.IO;
end-effector trajectory (X, Y) into rotations of stepper motors (ZA, ZB
and ZC). The proposed method used Artificial Neural Networks Article History:
(ANNs) to simplify the process of IK solver. The IK solver generated Received: May 20, 2020
the datasets contain motion data of the Delta robot. There are 11 KB Revised: August 22, 2020
Datasets consist of 200 motion data used to be trained. The Accepted: September 18, 2020
proposed method was trained in 58.78 seconds in 5000 iterations. Published: July 2, 2021
Using a learning rate (α) 0.05 and produced the average accuracy
was 97.48%, and the average loss was 0.43%. The proposed Corresponding Author:
Zendi Iklima
method was also tested to transfer motion data over Socket.IO with Electrical Engineering
115.58B in 6.68ms. Department, Universitas Mercu
Buana, Indonesia
This is an open access article under the CC BY-NC license Email:
[email protected]

INTRODUCTION implementation of the Delta robot has become

Robots can be classified into two more widespread [2][3]. The implementation of the
topologies, namely Serial Manipulator and Delta robot is used as a type of 3D Printer,
Parallel Manipulator [1]. The parallel manipulator Sketcher Robot, Laser Cutting Robot, and Pick
has abilities such as acceleration, stiffness, low and Place robot. So far, the Delta robot has been
inertia, high precision, and faster per-cycle developed with a focus on a kinematic analysis of
operations. Serial manipulator produced slower the robot’s movement.
movement because of its connected link a chain, Kinematics is a mathematical model that
and each joint has a mass in which needs to be describes the movement of end-effectors through
calculated to achieve the dynamic manipulator. Cartesian coordinates of the active joint position
Meanwhile, The Delta Robot joints are variable or vice versa. However, mathematical
mechanically separated to achieves faster calculations of kinematics do not fully reflect
movement by using low inertia. Nevertheless, perfect results in actual application. There is a
there are still shortcomings in the Delta robot position error caused by component
that is a relatively small workspace robot. manufacturing tolerance, assembly programming,
Delta Robot consists of three arms that are and the dynamic influence of robot mechanics. On
connected to the base and effector using a the other hand, the translation of forwarding and
universal joint. The key of Delta robot design is inverse kinematics on Delta robots is a fairly
using a parallelogram mechanism to maintain the difficult problem.
orientation of the end-effector to the base. Since The Inverse Kinematics (IK) of a Delta
the Delta robot’s Original Patent expired in Robot has been derived in several methods.
December 2006 (Europe) and December 2007 Some methods are numerical algorithm,
(America), the development of research and geometrical method, genetic algorithm (GA),

Z. Iklima, et al., Self-Learning of Delta Robot Using Inverse Kinematics and Artificial … 237
 SINERGI Vol. 25, No. 3, October 2021: 237-244

reduced method, polynomial method, fuzzy

algorithm, fuzzy PID [4], Feed Forward Neural
Network, Fuzzy Neural Network, Particle Swarm
Optimization (PSO), etc. [5].
Previous work has been done to solve the
IK problem using Neural Network (NN). NN can be
used to control validations of the IK model. IK
model is created based on the robot’s geometry
and dynamic to solve the forward kinematics
problem [4][5].
Other solutions to solve the IK problem of
Delta Robot can be by using a hybrid approach
are artificial neural network and particle swarm
optimization (ANN-PSO) model in the behavior
prediction and optimization of channel connectors
embedded in normal and high-strength concrete
(HSC) [6, 7, 8].
An interesting topic also explained that IK
solution could be solved using neural networks
combined with a genetic algorithm (GA) [9, 10, 11].
NN-GA is an optimal design method for the Figure 1. IK-ANN Block Diagram
parallel robot, which maximizes the volume of the
workspace of parallel robots. The neural network Inverse Kinematics (IK) Analysis
learns the motion model of the robot. Then, the Inverse Kinematics on this robot is very
genetic algorithm uses this model to generate the necessary for the design of robot motion. This
optimal parameters of the robot [12]. function is to find the position of the entire linear
To overcome this, the proposed method to actuator or slider towards the end effector. For
solve the IK problem is using an Artificial Neural example, if the end-effector position moves to
Network (ANN). Our proposed method's loss can position X with respect to position 0, then how
be minimised by finding an optimum ANN much displacement changes (up/down) of the
architecture by constructing the layers, the three linear actuators or sliders. In this section, the
activation function, and the optimizer. On the other analysis is based on Stolfi [12].
hand, our proposed method is implemented using Delta robot consists of the top platform,
Socket.IO channels to distribute the pre-model of bottom platform, and end effector. The reference
our IK-ANN data. remains on the 𝑅(𝑂 − 𝑥𝑦𝑧) frame system located
in the middle on the upper platform (𝐴𝐵𝐶 triangle).
METHOD The 𝑍-axis is parallel to the end effector, and the
This paper created the Delta robot using the 𝑌-axis is parallel to 𝐶𝑂, as shown in Figure 2.
Inverse Kinematics method as an initializing value
for the Artificial Neural Network training data. The
results of the end-effector position forwarded to
the motor motion execution for a displacement of
the effector position. Figure 1 shows the process
of the proposed method while it trains and tests
the ANN Model.
The IK dataset was generated in small rows
(200 rows of IK data). The inputs contain any
positions of the end-effector on the base frame
𝑅′(𝑂′ − 𝑥’𝑦’𝑧′). The outputs contain any directions
on 𝑍 − 𝑎𝑥𝑖𝑠 (𝑍𝐴 , 𝑍𝐵 , 𝑎𝑛𝑑 𝑍𝐶 ). ANN Model (pre-
trained) used to train (the dataset) and to test
(random inputs) to predict the direction on 𝑍 −
𝑎𝑥𝑖𝑠 (𝑍𝐴 , 𝑍𝐵 , 𝑎𝑛𝑑 𝑍𝐶 ). Socket.IO implemented as a
communication platform in which will broadcast
any random input (𝑥, 𝑦 ) from delta robot then
tested on ANN trained model to generate a
predictive value of (𝑍𝐴 , 𝑍𝐵 , 𝑎𝑛𝑑 𝑍𝐶 ). Figure 2. Delta Robot Kinematics Frame

238 Z. Iklima, et al., Self-Learning of Delta Robot Using Inverse Kinematics and Artificial …
 p-ISSN: 1410-2331 e-ISSN: 2460-1217

Another reference is the system 𝑅′(𝑂′ −

𝑥’𝑦’𝑧′) located in the middle of the end effector
(triangle 𝑃1 , 𝑃2 , 𝑃3 ). The 𝑍-axis ’is in the direction
perpendicular to the end effector, and the 𝑌-axis’
is parallel along 𝑃3 𝑂’.
As a geometry parameter, where [4]:
𝑂𝐴 = 𝑂𝐵 = 𝑂𝐶 = 𝑅 (1)
𝑂’𝑃1 = 𝑂’𝑃2 = 𝑂’𝑃3 = 𝑟 (2)
𝐵1 𝑃1 = 𝐵2 𝑃2 = 𝐵3 𝑃3 = 𝐿 (3)
The object of inverse kinematics is to
determine the position of three linear actuators
due to changes in point 𝑂′ (middle of the end
effector) to the global reference of the 𝑅 system.
Point 𝑂′ in frame 𝑅 can be written as:
[𝑂’]𝑅 = (𝑥𝑦𝑧)𝑇 (4) Figure 3. Sketch of dots for Dataset in
The coordinates of the point 𝑃𝑖 in frame 𝑅′ can be Workspace
written as:
The coordinate frame of the Delta Robot shown in
[𝑃𝑖 ]𝑅’ = (𝑟𝑐𝑜𝑠(𝑛𝑖 ), 𝑟𝑠𝑖𝑛(𝑛𝑖 ), 0)𝑇 (5) Figure 2 can be depicted as the desired
coordinates of the end-effector. Figure 3 shows a
Where 𝑖 = 1, 2, 3 and: sketch of dots for Dataset in Workspace. This
method is easily used to see the data distribution
4𝑖−3
𝑛(𝑖) =
6
𝜋 (6) of the delta robot workspace.
Adding from two vectors at once will get the
Artificial Neural Network (ANN)
position of 𝑃𝑖 with respect to the 𝑂 − 𝑥𝑦𝑧 frame [3].
An intelligence platform inspired by the
[𝑃𝑖 ]𝑅 = (𝑟𝑐𝑜𝑠(𝑛𝑖 ) + 𝑥, 𝑟𝑠𝑖𝑛(𝑛𝑖 ) + 𝑦, 𝑧)𝑇 (7) biological neurons can conveniently learn patterns
and predict high-dimensional data distributions
Meanwhile, to determine the Bi point on frame R: [6][11]. ANN design is needed as an initial
description of how ANN works as a substitute for
[𝐵𝑖 ]𝑅 = (𝑅𝑐𝑜𝑠(𝑛𝑖 ), 𝑅𝑠𝑖𝑛(𝑛𝑖 ), 𝑥𝑖 )𝑇 (8) inverse kinematics' repetitive mathematical
Where 𝑍𝑖 Above is the 𝑍-axis position to be calculation process to determine the 𝑍-axis
searched as the position of each linear actuator position of the slider against the 𝑋 − 𝑌 coordinate
concerning the frame R reference. end-effector. ANN consists of the input layer,
As it is known: hidden layer, and output layer. Figure 4 shows the
components among the layers [13, 14, 15].
𝐿 = |[𝑃𝑖 − 𝐵𝑖 ]| (9)
Then it can be specified as:
(𝑥 − 𝑥𝑖 )2 + (𝑦 − 𝑦𝑖 )2 + (𝑧 − 𝑧𝑖 )2 = 𝐿2 (10)
Where:
𝑥𝑖 = (𝑅 − 𝑟)𝑐𝑜𝑠(𝑛𝑖 ) (11)
𝑦𝑖 = (𝑅 − 𝑟)𝑠𝑖𝑛(𝑛𝑖 ) (12)
By rearranging Equation 10: Figure 4. The architecture of ANN Delta Robot [6]

𝑧 = ±√𝐿2 − (𝑥 − 𝑥𝑖 )2 − (𝑦 − 𝑦𝑖 )2 (13) The Input layer contains activations of end-

effector coordinates defined as 𝑥 and 𝑦 in which
To find the desired position 𝑧 point on each formulated as
actuator from the Delta robot, only the positive
form of (13) is used. 𝐹 = 𝑓(𝑛𝑒𝑡) = 𝑓(𝑥, 𝑦) (14)
𝑓(𝑥, 𝑦) defines as the activation function in the
desired direction should be [6, 16, 17],
1
𝑓(𝑥, 𝑦)𝑗 = ∑𝑛 𝑤 𝑥 +𝑏
(15)
1+𝑒 𝑗=1 𝑗𝑘 𝑗 𝑗

Z. Iklima, et al., Self-Learning of Delta Robot Using Inverse Kinematics and Artificial … 239
 SINERGI Vol. 25, No. 3, October 2021: 237-244

The output layer represents the desired Table 1. Hyperparameter of ANN

position of 𝑧 positions on each actuator described Parameters Choices
Selected
in (13) (𝑍𝐴 , 𝑍𝐵 , 𝑎𝑛𝑑 𝑍𝐶 ). To calculate the Parameters
Number of
backpropagation process defined as 𝑓′ (or called Layers
[3,4,5,6,7,8] 3
𝛿𝑗 Aka “delta”). The derivative of the activation Number of
function 𝑓(𝑥, 𝑦) and 𝛾 as a learning rate. So, the Neurons in [2, 4, 6, …] 2, 6, 12
Hidden Layers
weights will be graded when the gradient descent [‘ReLU’,
is calculated by [18, 19, 20], Activation
‘Sigmoid’, ‘TanH’, ‘Leaky ReLu’
Functions
‘Leaky ReLu’]
𝛿𝑗 = 𝑓′(𝑥, 𝑦) ∑𝑛𝑗=1 𝑤𝑗𝑘 𝛿𝑗 (16) Optimizer [‘SGD’, ‘ADAM’] ‘ADAM’
Activation [0.01. 0.02, ….
0.05
Functions Rate 𝛾 1.0]
Socket.IO
Socket.IO is used to distribute the data The hyperparameters were selected as the
among the environment by using events called parameter to build ANN architecture in which to
‘onListening’ and events ‘onEmit.’ Events find the optimum values of the number of layers,
‘onListening’ is an event to retrieve incoming data number of hidden layers, activation functions,
in which distributed through its event. Events optimizer, and the learning rate.
‘onEmit’ is an event to transmit any data into its
event [21][22]. RESULT AND DISCUSSION
The following is the prototype that was
completed as describes in Figure 2. As shown in
Figure 3, the inverse kinematics inputs taken by
the coordinate of the end-effector (𝑥 and 𝑦) to
produce the desired 𝑍 position in direction 𝐴, 𝐵 and
𝐶 (𝑍𝐴 , 𝑍𝐵 , 𝑎𝑛𝑑 𝑍𝐶 )

Figure 5. Socket.IO Events [20]

Figure 5 represents the Socket.IO events

used to distribute data among platforms. The data
was trained on the Django server and distribution
through socket.IO event onListening ‘𝑥_𝑦_𝑐𝑜𝑜𝑟’ in Figure 6. Prototype of Delta Robot
which waiting to receive end-effector data (𝑥, 𝑦)
from the robot delta. then emitting an event
′𝑧_𝑎𝑐𝑡_𝑑𝑎𝑡𝑎’, which mean the predictive z actuator The dataset from robot motion in Figure 6.
data (𝑍𝐴 , 𝑍𝐵 , 𝑎𝑛𝑑 𝑍𝐶 ) 200 rows of motion data were collected and stored
as a CSV file.
Proposed Design
Table 1 shows the hyperparameter that ANN Training
might be used to find the optimum ANN The training process was performed using
architecture to the Sole IK problem as described Intel(R) Core© i5-6300HQ [email protected], 16GB
above. RAM, and NVIDIA GeForce GTX 960M 4GB
VRAM. The training process followed the
hyperparameter shown in Table 1. The training
performance was captured in Figure 7.

240 Z. Iklima, et al., Self-Learning of Delta Robot Using Inverse Kinematics and Artificial …
 p-ISSN: 1410-2331 e-ISSN: 2460-1217

Figure 7. The ANN Loss and Accuracy in: (a) 1000 iterations, (b) 2500 iterations, and (c) 5000
iterations

Figure 8. The Testing Performance of Single Motion: (a) Motion Set 1, (b) Motion Set 2, (c) Motion Set
3, and Motion Set 4

Figure 7 shows the performance of the ANN and 5000 iterations. The loss was decreased by
Loss and Accuracy. The loss is presented as the 0.0043 in 5000 iterations, and the accuracy was
red line of the left graph, and the accuracy is increased by 0.9748 in 5000 iterations. This
presented as the blue line of the left graph. Figure iteration step was executed in 58.7826 seconds.
7(a) shows the loss decreased and the accuracy
increased in 1000 iterations. Figure 7(b) shows ANN Testing
the loss decreased and the accuracy increased in Single point of end-effector coordinated
2500 iterations. Figure 7(c) shows the loss tested as desired coordinated of 𝑥 and 𝑦 shown in
decreased and the accuracy increased in 5000 Figure 7.
iterations. The detailed values of ANN Loss and The proposed method was tested in single
accuracy were captured in this Table 2. desired coordinated as a set of motion. Figure 8
Table 2 shows the selected hyperparameter represents four motions that were tested in Figure
was performed in 1000 iterations, 2500 iterations, 8(a), Figure 8(b), Figure 8(c), and Figure 8(d). By

Z. Iklima, et al., Self-Learning of Delta Robot Using Inverse Kinematics and Artificial … 241
 SINERGI Vol. 25, No. 3, October 2021: 237-244

given any input 𝑥 and 𝑦, will produce any desired algorithm performance to generate 50 motions in
actuators positions in Z-directions (𝑍𝐴 , 𝑍𝐵 , 𝑎𝑛𝑑 𝑍𝐶 ). the direction 𝑍𝐴 (Figure 9(a)), 𝑍𝐵 (Figure
The detailed values of ANN testing performance 9(b)), 𝑍𝐶 (Figure 9(c)), and combined (Figure 9(d)).
were captured in this Table 3.
The proposed method was implemented to
perform 𝑛 Motions and captured in actuator
direction 𝑍𝐴 , 𝑍𝐵 , 𝑎𝑛𝑑 𝑍𝐶 . Figure 9 shows the

Table 2. The ANN Loss and Accuracy in (a) 1000 iterations, (b) 2500 iterations, and (c) 5000 iterations
1000 iterations 2500 iterations 5000 iterations
Executed time (seconds) 12.5525 29.7373 58.7826
Num of Epoch 1000 2500 5000
Average Loss 0.0787 0.0131 0.0043
Average Accuracy 0.9322 0.9528 0.9748
Activation Function Leaky-ReLu Leaky-ReLu Leaky-ReLu
Optimizer Adam Adam Adam
Learning Rate (𝜶) 0.05 0.05 0.05
Layers 2 6 12 6 3 2 6 12 6 3 2 6 12 6 3

Figure 9. The Testing Performance of 50 Motions: (a) Motion Data of 𝑍𝐴 , (b) Motion Data of 𝑍𝐵 , (c)
Motion Data of 𝑍𝐶 And (d) Motion Data of 𝑍𝐴 , 𝑍𝐵 , and 𝑍𝐶

Figure 10. The Testing Performance of 100 Motions: (a) Motion Data of 𝑍𝐴 , (b) Motion Data of 𝑍𝐵 , (c)
Motion Data of 𝑍𝐶 And (d) Motion Data of 𝑍𝐴 , 𝑍𝐵 , and 𝑍𝐶

242 Z. Iklima, et al., Self-Learning of Delta Robot Using Inverse Kinematics and Artificial …
 p-ISSN: 1410-2331 e-ISSN: 2460-1217

Table 3. The Testing Performance of Single Motion

(a) Motion Set 1, (b) Motion Set 2, (c) Motion Set 3, and Motion Set 4
Motion Set 1 2 3 4

Inputs X 0.7 0.5 0.3 −0.4

(meter) Y 0.2 0.3 0.6 0.7

𝒁𝑨 0.6086 0.5482 0.4784 −0.3098

Target
𝒁𝑩 −1.0696 −0.5705 −0.1597 0.5508
(Radian)
𝒁𝑪 −0.5474 −0.5971 −1.1973 −1.5789

𝒁𝑨 0.6256 0.5405 0.4675 −0.3031

Predicted
𝒁𝑩 −1.0751 −0.5763 −0.1646 0.5553
(Radian)
𝒁𝑪 −0.5255 −0.6099 −1.2226 −1.5499

Figure 10 shows the algorithm performance ACKNOWLEDGMENT

to generate 100 motions in the direction 𝑍𝐴 (Figure The author expresses gratitude to the
10(a)), 𝑍𝐵 (Figure 10(b)), 𝑍𝐶 (Figure 10(c)), and Electrical Engineering Department and Research
combined (Figure 10(d)). Center of Mercu Buana University, Jakarta,
The desired 50 motions` 𝑍 values are Indonesia. Under the guidance, the authors have
shown in the blue line, which is compared with the been completed this research paper.
actual values represented in the red line.
Table 4 shows the Socket.IO response REFERENCES
within a set of motion with 115.6B processed in [1] Z. Pandilov, V. Dukovski, “Comparison of The
6.68 ms, 50 sets of motion with 912.3B processed Characteristics Between Serial and Parallel
in 216.35 ms, and 100 sets of motion with 1712.1B Robots,” Acta Tehnica Corviniensis: Bulletin
processed in 335.56 ms. of Engineering, vol. 1, no. 7, pp. 143-160,
2014
[2] J. Brinker, N. Funk, P. Ingenlath, Y. Takeda,
Table 4. Socket.IO Response and B. Corves, “Comparative Study of Serial-
Inputs Parallel Delta Robot with Full Orientation
1 50 100
(motions) Capabilities,” IEEE Robotics and Automation
Letters, vol. 2, no. 2, pp. 920-926, April 2017,
Data Size
(𝑩𝒚𝒕𝒆)
115.6 912.3 1712.1 doi: 10.1109/LRA.2017.2654551
[3] Z. Iklima, A. Adriansyah, S. Hitimana, “Self-
Socket.IO Collision Avoidance of Arm Robot using
Response 6.68 216.35 335.56
(𝒎𝒔) Generative Adversarial Network and Partical
Swarm Optimization (GAN-PSO),” SINERGI,
vol. 25, no. 2, pp. 141-152, 2021, doi:
CONCLUSION 1022441/sinergi.2021.2.005
This study has developed the method to [4] H. Yuancan and H. Zongllin, “Neural Network
determine the end-effector position using inverse Based Dynamic Trajectory Tracking of Delta
kinematics and substitute by Artificial Neural Parallel Robot,” IEEE International
Network. All the equations are used to determine Conference on Mechatronics and Automation
the position of the slider that will be dataset for (ICMA), Beijing, China, Aug. 2015, pp. 1938-
ANN. The proposed method used Artificial Neural 1943, doi: 10.1109/ICMA.2015. 7237782
Networks (ANNs) to simplify the process of IK [5] M. Heidari, S.M.R Faritus, S. Shateyi, “Using
solver. The IK solver generated the datasets Artificial Neural Network for Feed Kinematic
contain motion data of the Delta robot. 11KB problem of Under-Constrained Cable
Datasets consist of 200 motion data used to be Robots,” Journal of Vibroengineering (JVE),
trained. The proposed method was trained in vol. 20, no. 1, pp. 385-400, Feb. 2018, ISN:
58.78 seconds in 5000 iterations. Using a learning 1392-8716, doi: 10.21595/jve.2017. 18633
rate (α) 0.05 produced an average accuracy of [6] M. Shariati, et al., “Application of a Hybrid
97.48%, and the average loss was 0.43%. The Artificial Neural Network-Particle Swarm
proposed method was also tested to transfer Optimization (ANN-PSO) Model in Behavior
motion data over Socket.IO with 115.58B in Prediction of Channel Shear Connectors
6.68ms. Embedded in Normal and High-Strength

Z. Iklima, et al., Self-Learning of Delta Robot Using Inverse Kinematics and Artificial … 243
 SINERGI Vol. 25, No. 3, October 2021: 237-244

Concrete,” Applied Sciences, vol. 9, no. 24, March 2017, pp. 153-157, doi: 10.1109/
pp. 5534, 2019, doi:10.3390/app9245534 ISESD.2016.7886710
[7] P. Srisuk, A. Sento, Y. Kitjaidure, “Inverse [15] T. Collins, W. Shen, “PASO: An Integrated,
Kinematic Solution using Neural Network Scalable PSO-based Optimization
from Forward Kinematics Equations,” The 9th Framework for Hyper-Redundant
International Conference on Knowledge and Manipulator Path Planning and Inverse
Smart Technology (KST), Thailand, 2017, pp. Kinematic,” ISI Tech Report, January 2016
61-65, doi: 10.1109/KST.2017.7886084 [16] H. Su, C. Yang, H. Mdeihly, A. Rizzo, G.
[8] EG López, W. Yu, X. Li, “Optimal Design of a Ferrigno and E. De Momi, "Neural Network
Parallel Robot Using Neural Network and Enhanced Robot Tool Identification and
Genetic Algorithm,” The 10th International Calibration for Bilateral Teleoperation,"
Conference on Intelligent Control and in IEEE Access, vol. 7, pp. 122041-122051,
Information Processing (ICICIP), Morocco, 2019, doi: 10.1109/ACCESS.2019.2936334
February 2017, pp. 64-69, doi: 10.1109/ [17] B. A. Garro and R. A. Vazquez. “Designing
ICICIP47338.2019.9012182 Artificial Neural Networks using Particle
[9] K. K. Pavan, M. J. Murali, D. Srikanth, Swarm Optimization Algorithms,”
“Generalized solution for inverse kinematics Computational Intelligence and
problem of a robot using hybrid genetic Neuroscience, ID: 369298, June 2015, doi:
algorithms,” International Journal of 10.1155/2015/369298
Engineering & Technology (IJET), vol.7 no. [18] L. F. Carlos et al., “A soft computing
4.6, pp. 250-256, 2018, p.250-256, doi: approach for inverse kinematics of robot
10.14419/ijet.v7i4.6.20486 manipulators,” Engineering Applications of
[10] E. McCormick, W. Yanjun; L. Haoxiang, Artificial Intelligence, vol. 74, pp. 104–120,
“Optimization of a 3-RRR Delta Robot for a September 2018, doi: 10.1016/j.engappai.
Desired Workspace with Real-Time 2018.06.001
Simulation in MATLAB,” The 14th [19] S. Mahdi, et al., “Application of a Hybrid
International Conference on Computer Artificial Neural Network- Particle Swarm
Science & Education (ICCSE 2019), Canada, Optimization (ANN-PSO) Model in Behavior
August 2019, pp. 935-941, doi: Prediction of Channel Shear Connectors
10.1109/ICCSE. 2019.8845388 Embedded in Normal and High-Strength
[11] A.G.C Premachandra, et al., “Genetic Concrete,” Applied Sciences, vol. 9, no. 24,
Algorithm Pick and Place Sequence December 2019, doi: 10.3390/app9245534
Optimization for a Color and Size Sorting [20] L. Peiyuan, “Deep Neural Network Based
Delta Robot,” The 6th International Subspace Learning of Robotic Manipulator
Conference on Control, Automation and Workspace Mapping,” International
Robotics (ICCAR), Singapore, April 2020, pp. Conference on Control, Artificial Intelligence,
209-213, doi: 10.1109/ICCAR49639.2020. Robotics & Optimization (ICCAIRO), 2018,
9108045 vol.1, pp. 109-120, doi: 10.1109/ICCAIRO.
[12] F. Tolfi, A. Avanzini, B. Welker, and A. 2018.00027
Isinhue, "Design Review of Delta Robot," [21] T. Dewi, C. Anggraini, P. Risma, Y, Oktarina,
MECE E3430 Engineering Design, pp. 10-19, and M. Muslikhin, “Motion Control Analysis of
2014 Two Collaborative Arms Robots in Fruit
[13] A. A. Jaber and R. Bicker, “Industrial Robot Packaging System,” SINERGI, vol. 25, no. 2,
Backlash Fault Diagnosis Based on Discrete pp. 217-226, 2021, doi: 1022441/sinergi.
Wavelet Transform and Artificial Neural 2021.2.013
Network,” American Journal of Mechanical [22] Z. Iklima, “Design of Distributed Control
Engineering, vol. 4, no. 1, pp. 21-31, January Systems in Lighting Installation in 3-Floor
2016, doi:10.12691/ajme-4-1-4 Buildings based on IoTaaS (Internet of
[14] R.Y. Putra, et al., “Neural Network Things as a Service) using Docker
Implementation for Invers Kinematic Model of Container”, Jurnal Teknologi Elektro, vol.10,
Arm Drawing Robot,” International no.1, pp. 26-33, 2019, doi: 10.22441/
Symposium on Electronics and Smart jte.v10i1.004
Devices (ISESD), Bandung, Indonesia,

244 Z. Iklima, et al., Self-Learning of Delta Robot Using Inverse Kinematics and Artificial …