Engineering Applications of Artificial Intelligence 126 (2023) 107175

Contents lists available at ScienceDirect

Engineering Applications of Artificial Intelligence

journal homepage: www.elsevier.com/locate/engappai

Artificial Neural Networks for inverse kinematics problem in articulated

robots
Daniel Cagigas-Muñiz
Department of Computer Architecture and Technology, Universidad de Sevilla, Seville, Spain

ARTICLE INFO ABSTRACT

Dataset link: https://github.com/dcagigas/Rob The inverse kinematics problem in articulated robots implies to obtain joint rotation angles using the robot
otics end effector position and orientation tool. Unlike the problem of direct kinematics, in inverse kinematics
Keywords: there are no systematic methods for solving the problem. Moreover, solving the inverse kinematics problem
Robot kinematics is particularly complicated for certain morphologies of articulated robots. Machine learning techniques and,
Inverse kinematics more specifically, artificial neural networks (ANNs) have been proposed in the scientific literature to solve this
Industrial robots problem. However, there are some limitations in the performance of ANNs. In this study, different techniques
Artificial Neural Networks that involve ANNs are proposed and analyzed. The results show that the proposed original bootstrap sampling
Machine learning and hybrid methods can substantially improve the performance of approaches that use only one ANN. Although
all of these improvements do not solve completely the inverse kinematics problem in articulated robots, they do
lay the foundations for the design and development of future more effective and efficient controllers. Therefore,
the source code and documentation of this research are also publicly available to practitioners interested in
adapting and improving these methods to any industrial robot or articulated robot.

1. Introduction One of the most popular methods for solving the direct kinematic
problem is to use the Denavit–Hartenberg (DH) method (Hayat et al.,
Industrial robot kinematics is dedicated to the analysis and solution 2013). The algorithm enables moving from one coordinate system of a
of problems arising from the processing of the elements of a manipula- link to another using 4 basic transformations consisting of a succession
tor (that is, a robotic arm). Articulated robots are a subset of industrial of rotations and translations (Barrientos, 2007).
robots/manipulators and are the most widely used robots in industry In contrast, there is no systematic method to solve the inverse
today. This characteristic makes them particularly interesting to study.
kinematics problem. There are usually three classical approaches to
The elements of a robot manipulator are defined by its mechanical
solving this problem: geometrical (Jiang et al., 2023), analytical (Haug,
structure and its degrees of freedom. The mechanical structure of an in-
2023), and numerical methods (Larsson and Grönlund, 2023). The
dustrial manipulator robot is made up of several metal links connected
by joints. These joints define the robot’s degrees of freedom and are the former can also be used to solve the direct kinematics problem. It is
elements that position the terminal element of the mechanical structure suitable when the kinematic chain of the manipulator robot is rela-
at a point in space (with a certain orientation). tively simple. The second approach, although theoretically possible,
Kinematics is concerned only with defining the position of the is not always feasible in practice when equations are complex. Here,
manipulator robot with respect to a coordinate system. Kinematics homogeneous transformation matrices of the direct kinematic problem
solves two characteristic problems in manipulator positioning (Liana are used. The third approach involves much more computation than
López-Pacheco, 2012): analytical approaches, provides only point solutions, and may fall into
local minima.
• Direct kinematic (DK) problem: it solves the position and ori- Inverse kinematic (IK) equations can also be obtained using geo-
entation of the robot manipulator terminal element from the
metric techniques or kinematic decoupling techniques (Nosova, 2020).
parameters defining the degrees of freedom, i.e. the angles of the
Nevertheless, kinematic decoupling techniques are only possible for
joints.
some articulated robot configurations. In Predescu and Stroe (2015),
• Inverse kinematic (IK) problem: obtains the parameters of the
Cole et al. (2007) different equations are proposed for Scorbot ER VII,
robot’s degrees of freedom (i.e. the angles of the joints) from
the position and orientation of the robot manipulator terminal which is the example articulated robot used in this study.
element.

E-mail address: [email protected].

https://doi.org/10.1016/j.engappai.2023.107175
Received 12 April 2023; Received in revised form 9 August 2023; Accepted 15 September 2023
Available online 28 September 2023
0952-1976/© 2023 Elsevier Ltd. All rights reserved.
D. Cagigas-Muñiz Engineering Applications of Artificial Intelligence 126 (2023) 107175

Inverse kinematic equations are not usually necessary to operate robots. See, for example, Srisuk et al. (2017b), Daya et al. (2010),
industrial robots. Industrial robots poses are first recorded. The robot Duka (2014), Sharkawy and Khairullah (2023). However, these are
control unit then moves it between the recorded poses in a repetitive applications that use toy examples with simulated planar robots with
process until the task is completed. In contrast, several challenges arise two or three degrees of freedom and in two dimensions. Other toys
when industrial arms are integrated into service robots. The articulated examples, this time with a 3-D articulated robot, can be found in Srisuk
robot can be commanded to reach non-recorded poses in real time. et al. (2017a), Aysal et al. (2022). In these studies a single Multi-Layer
Inverse kinematic (IK) equations or something equivalent are hence Feed Forward ANN with one hidden layer is applied.
necessary to implement the robot controller. This classical application of ANNs to solve the inverse kinematic
Machine learning methods, and more specifically Artificial Neural problem in real articulated robots has been combined with other tech-
Networks (ANNs), can be considered a fourth alternative method. They niques to improve the performance obtained. In Adar (2021) ANNs
are suitable for real-time tasks because of their low output latency. are combined with Proportional–Integral–Derivative (PID) control and
Several studies have used ANNs to solve the inverse kinematic problem
in Köker and Çakar (2016) recurrent neural networks (RNNs) with
for specific industrial robots. In the case of articulated robots, ANNs
genetic algorithms. In El-Sherbiny et al. (2018) a comparison of the
have provided limited results. The main drawback of supervised ma-
use of ANNs with genetic algorithms and fuzzy logic is made. Gener-
chine learning methods like ANN models is that they usually fail when
ative adversarial networks (GANs) were also used to solve the inverse
data input (i.e., a point in a 3D working environment) can have many
kinematic problem, but did not show better performance than previous
solutions (i.e., joint angle configurations).
solutions (Ren and Ben-Tzvi, 2020). A recent review of analytical and
However, to the best of the author’s knowledge, there have been
deep learning approaches for solving the inverse kinematic problem
no comprehensive studies of the behavior of ANNs in the case of the
of the robotic arm is described in Wagaa et al. (2023). Here, a per-
inverse kinematics problem and articulated robots. More specifically,
in the analysis of the ANNs error produced when solving the inverse formance comparative of convolutional neural networks (CNNs) and
kinematic (IK) problem. An analysis of the behavior of the errors recurrent neural networks was included. Noise was added to the input
produced by ANNs is a basic and necessary step to increase and improve dataset, and some example circular trajectories were tested. Again, only
their performance. one custom trained ANN was used in each case, and it was not clear
Here, original methods are proposed for solving the IK problem how these architectures performed (i.e. error analysis) in the whole
using data science bootstrapping and hybrid techniques. Then, the robot space area.
error prediction distribution of these methods and previous studies is In Almusawi et al. (2016) the traditional neural network design is
analyzed in detail. improved by adding the robot’s previous position to the input. In this
The main contributions of this paper can be summarized as follows: case, it is assumed that the robot starts from a position at time 𝑡 and
moves to a similar position at time 𝑡 + 1. This neural network architec-
1. An original method based on bootstrap sampling techniques to ture is hence only valid for problems where continuous trajectories are
solve the IK problem in articulated robots.
described from points in proximity to each other.
2. An analysis of the combination of bootstrap sampling techniques
In Ho and King (2022) the idea is to not only input the ANN
and other methods that combine multiple ANNs to solve the IK
architecture with a desired robot pose, but also with a ‘‘posture in-
problem in articulated robots.
dex’’. Some previous unsupervised learning is also needed to learn the
3. A detailed analysis of the error and outlier behavior of ANNs
‘‘posture indexes’’. The ANN output is then extended (post-processed)
when solving the IK problem in articulated robots.
to calculate different solutions (angle solutions for the output target
This article is organized as follows. Related works are discussed in pose). This recent approach is promising because the mean errors
Section 2. The morphology and kinematics of the Scorbot ER VII artic- are relatively small. However, no error variance or outlier analysis is
ulated robot are described in Section 3. This robot is used in this paper performed. This is because (as the authors indicate), accuracy is not
as an example for the kinematics analysis of generic articulated robots. their major concern.
Section 4 formally specifies the problem to be solved. The different ANNs implement one-to-one map functions: one input produces one
analyzed IK methods are described in Section 5. Here is included the output. This is not the case with the IK for redundant robotic arms
design of an original Multilayer Perceptron Artificial Neural Network where a pose can be reached by several joint angle values (in some
(MLP-ANN) used in the following subsections. All proposed techniques cases, virtually infinitive). To address this problem, there are studies
and solutions are tested and compared in Section 6. The experimental that collect custom or biased training datasets. Examples can be found
results are discussed in Section 7. Finally, the conclusions and future in Ghasemi et al. (2019), Almusawi et al. (2016). The main advantage
works are outlined in Section 8. of using biased datasets is that the accuracy error obtained is low.
However, the main drawback is that it is not clear that these methods
2. Related work
can infer the entire working space of the robot.
There is also another class of methods that propose to use multiple
Artificial neural networks (ANNs) have received a great deal of
ANNs in parallel. In Lu et al. (2022) the workspace of an articulated
attention from the scientific community, especially as a result of the
robot with 6 degrees of freedom is divided into several subspaces of
successes achieved by Deep Learning in applications related to image
work, each of which is assigned a different ANN. In a later step, the
recognition or the processing of natural and written language. In ad-
dition to these success stories, ANNs have continued to be successfully output of these ANNs is processed by three Deep Neural Networks
applied in other fields. Examples of these applications can be found, (DNNs) that determine the ANN output that optimally approximates
for example, in the design of time series for the prediction of energy the desired angle configuration for a given point. Finally, a last step
consumption in cities (Ghadami et al., 2021), decision support systems of refinement of this solution is applied. This class of methods (using
in financial management (Mousapour Mamoudan et al., 2023), manage- multiple ANNs) will be further developed and analyzed in detail in this
ment of water treatment plants (Pouresmaeil et al., 2022), promotion study.
of sustainable transportation systems (Zhan et al., 2023), or control The application of ANNs to solving the inverse kinematic problem
of renewable energy production in green buildings (Shahsavar et al., has not been confined to articulated robots. For example, in Jha (2014)
2021). a neural network is described to solve the inverse kinematics of a Scara-
There are also several studies in the scientific literature on the ap- type industrial robot and in Toquica et al. (2021) for a parallel robot.
plication of ANNs to solve the inverse kinematic problem in articulated This study will focus only on articulated robots.

2
D. Cagigas-Muñiz Engineering Applications of Artificial Intelligence 126 (2023) 107175

Fig. 1. Scorbot ER-VII. Figure (a) shows the real robot and Figure (b) shows the ROS rviz simulation framework used in the experiments.

Fig. 2. Scorbot ER-VII joints and links: Gazebo simulation framework used in kinematic equation experiments. Figure (a) on the left indicates the rotation angles (𝜃1 , 𝜃2 , 𝜃3 , 𝜃4 )
and figure (b) on the right indicates the links (𝑎1 , 𝑎2 , 𝑎3 , 𝑎4 , 𝑑1 , 𝑑2 ).

3. The morphology and kinematics of the Scorbot ER-VII Table 1

Denavit–Hartenberg Scorbot ER VII parameters.
Joint/Angle 𝛼 𝑎𝑖 𝑑𝑖 𝜃
It is not the objective of this research to propose a fully inverse number
kinematic controller for a specific articulated robot. But, for illustrative
1 0 0 𝑑1 𝜃1
and practical purposes, it is considered necessary to use a real artic- 2 𝜋∕2 𝑎1 𝑑2 𝜃2
ulated robot to analyze the inverse kinematics problem and propose 3 0 𝑎2 0 𝜃3
improvements. The selected articulated robot is the Scorbot ER VII. 4 0 𝑎3 0 𝜃4
5 0 𝑎4 0 0
To test and validate Scorbot ER VII kinematics, the Robot Operating
System (ROS) with Gazebo and rviz simulation frameworks were also
used (Fig. 1).
The Scorbot ER VII has four rotation angles or joints: base (𝜃1 ), • 𝑎4: distance from the joint/angle of the wrist 𝜃4 to the end
shoulder (𝜃2 ), elbow (𝜃3 ) and wrist (𝜃4 ). The gripper (end effector tool) of the gripper.
has three orientation angles: yaw, (𝛹 ) pitch (𝛷) and roll (𝜌). In this
2. Other links:
case, the angle of 𝜌 is irrelevant for positioning the gripper. Robot links
can be divided into two groups: • 𝑑1: height/distance from the base origin of the coordinates
(𝜃1 ) to the shoulder (𝜃1 ) in the X-Z plane.
1. Links in the X-Z plane through the X dimension:
• 𝑑2: displacement to the right from the base origin of the
• 𝑎1: displacement from base or join 𝜃1 to shoulder or joint coordinates (𝜃1 ) to the shoulder (𝜃2 ) in the XY plane.
𝜃2 .
• 𝑎2: distance from shoulder joint/angle 𝜃2 to elbow joint/angle The robot dimensions (angles and links) are shown in Fig. 2.
𝜃3 . Direct kinematic equations can be obtained using the Denavit–
• 𝑎3: distance from elbow joint/angle 𝜃3 to wrist joint/angle Hartenberg (DH) method mentioned in Section 1. The DH parameters
𝜃4 . for the robot are shown in Table 1.

3
D. Cagigas-Muñiz Engineering Applications of Artificial Intelligence 126 (2023) 107175

The Scorbot ER VII direct kinematic (DK) equations obtained using hyperparameters must be defined: the number of hidden layers, the
the DH method are shown in Eqs. (1). These equations were tested and number of neurons in each hidden layer, the activation functions, the
verified in the rviz and Gazebo frameworks. learning rate, the dropout ratio, the number of epochs, and many
𝑋 = 𝑎4 𝑐1 𝑐234 + 𝑐1 𝑐23 𝑎3 + 𝑐1 𝑐2 𝑎2 + 𝑐1 𝑎1 + 𝑠1 𝑑2 more. There are no exact methods to determine the most suitable
values for each ANN and each parameter. Experience based on similar
𝑌 = 𝑎4 𝑠1 𝑐234 + 𝑠1 𝑐23 𝑎3 + 𝑠1 𝑐2 .𝑎2 + 𝑠1 𝑎1 + 𝑐1 𝑑
ANN models, heuristic algorithms, or test and proof methods is usually
𝑍 = −𝑠234 𝑎4 − 𝑠23 𝑎3 − 𝑠2 𝑎2 + 𝑑1 applied. The importance of hyperparameters in ANNs models and their
𝑌 𝑎𝑤(𝛹 ) = 𝜃1 influence on the inverse kinematic problem in articulated robots is
highlighted in Wagaa et al. (2023). Nevertheless, even in this study,
𝑃 𝑖𝑡𝑐ℎ(𝛷) = 𝜃2 + 𝜃3 + 𝜃4 (1)
only a specific, arbitrary, and close set of hyperparameters are tested.
𝑅𝑜𝑙𝑙(𝜌) = 0 (No roll is considered in the end effector tool/gripper) ANN architecture design details or hyperparameter justifications are
where 𝑠1 = 𝑠𝑖𝑛(𝜃1 ), 𝑠2 = 𝑠𝑖𝑛(𝜃2 ), 𝑐1 = 𝑐𝑜𝑠(𝜃1 ), 𝑐2 = usually omitted in the scientific literature, so it is difficult to figure
𝑐𝑜𝑠(𝜃2 ), 𝑠23 = 𝑠𝑖𝑛(𝜃2 + 𝜃3 ), 𝑐23 = 𝑐𝑜𝑠(𝜃2 + 𝜃3 ), out why those ANNs architectures are used. There are, however, some
tune tools that help to determinate optimal ANN parameters in a semi-
𝑠234 = 𝑠𝑖𝑛(𝜃2 + 𝜃3 + 𝜃4 ), 𝑐234 = 𝑐𝑜𝑠(𝜃2 + 𝜃3 + 𝜃4 ) automatic way. This tuning process is more convenient for obtaining
The Scorbot ER VII ROS simulation framework used in this work an optimal ANN architecture.
can be downloaded at Cagigas-Muñiz (2023). In this study, the Keras Tuner has been used for hyperparameter
optimization. Keras Tuner is an easy-to-use, scalable hyperparame-
4. Problem formalization and error minimization ter optimization framework that helps to solve the hyperparameter
search (O’Malley et al., 2019). It comes with built-in Bayesian Op-
The inverse kinematics problem for the case of the Scorbot ER VII timization, Hyperband, and Random Search algorithms. On the one
and other articulated robots can be defined as follows. hand, random and hyperband algorithms usually need more computa-
We define the position and orientation 𝑝 of the end of the Scorbot tion time than Bayesian optimization to achieve similar solutions. On
ER VII terminal element (gripper) as follows: the other hand, Bayesian optimization was deemed a good choice in
𝑝 = (𝑥, 𝑦, 𝑧, 𝛹 , 𝛷), where x, y, z are the Cartesian coordinates, 𝛹 is different papers (see Snoek et al. (2012), Srinivas et al. (2009)) and
the orientation 𝑌 𝑎𝑤 and 𝛷 is the orientation 𝑃 𝑖𝑡𝑐ℎ of the end of the was thus the one selected. Bayesian Optimization process determines
robot terminal element. the optimal number of neurons in each layer, activation functions, and
A possible solution of 𝑝 (angles of the robot joints) is defined as: also the number of intermediate layers. Details of the tuning and hyper-
𝑞 = (𝜃1 , 𝜃2 , 𝜃3 , 𝜃4 ), where 𝜃1 , 𝜃2 , 𝜃3 , 𝜃4 are the angles of the robot’s parameter processes can be found in the source code in Cagigas-Muñiz
base, shoulder, elbow, and wrist, respectively. (2023).
It follows that: To train and tune the ANN architecture, a data set sample of 6000
𝑝 = 𝐷𝐾(𝑞) and 𝑞 = 𝐼𝐾(𝑝), where 𝐷𝐾 is a function of direct points was randomly collected within the Scorbot ER VII working area.
kinematics and 𝐼𝐾 is a function of inverse kinematics. The idea of collecting datasets is straightforward. A random value is
The function 𝐼𝐾 is assumed to be unknown and is approximated by generated for each robot joint/angle. The DK equations (see Section 3)
another function 𝐼𝐾 ′ . It follows that calculates the position and orientation of the end effector. Thus, a single
𝑞 ′ = 𝐼𝐾 ′ (𝑝) and hence 𝑝′ = 𝐷𝐾(𝑞 ′ ), where 𝑞 ′ is the predicted data set point/pose has these 9 values: 4 joint angles (𝜃1 , 𝜃2 , 𝜃3 , 𝜃4 ), 3
solution of 𝐼𝐾 ′ for 𝑝, and 𝑝′ is the resulting position and orientation Cartesian coordinates (𝑥, 𝑦, 𝑧) and pitch and yaw orientations. Data sets
derived from 𝑞 ′ . and scripts to generate more data sets are available at Cagigas-Muñiz
The objective is to achieve a function 𝐼𝐾 ′ that minimizes the error (2023). The 70% of the points were used for the training process and
𝜀 for any 𝑝. the rest for the validation of the test. Up to 100 epochs were used for
𝜀 = |𝑝′ − 𝑝| = |𝐷𝐾(𝑞 ′ ) − 𝑝| = |𝐷𝐾(𝐼𝐾 ′ (𝑝)) − 𝑝|, where | | represents the training process. An epoch can be defined as a pass through the
the Euclidean distance between two points in Cartesian space. training set. In each epoch, the entire ANN model is updated.
In this study, the function 𝐼𝐾 ′ will be approximated by ANN-based During the optimization of an MLP-ANN, the error for the current
models and the function 𝐷𝐾 is known and implemented by Eqs. (1) in state of the model must be repeatedly estimated. This requires the
the previous Section 3. choice of an error function, conventionally called the loss function (or
just loss). To estimate the loss of the model, the weights of the neurons
5. Methodology are updated at the next evaluation (epoch).
For regression predictive modeling problems, three loss functions
This section specifies and describes the different ANN methods ana- were initially tested:
lyzed. The first method specified in Section 5.1 is based on an original
• Mean Squared Error: this is the default loss to use for regression
Multilayer Perceptron Artificial Neural Network (MLP-ANN) architec-
problems.
ture. Next, a robot work area segmentation approach is described in
• Mean Absolute Error Loss: suitable when the target variable is
Section 5.2. Then Section 5.3 presents an alternative bootstrap sam-
close to Gaussian.
pling method. This section is completed in Section 5.4 with another
• Mean squared logarithmic: more suitable for outliers.
innovative hybrid method that combines the two previous methods.
The default mean squared error provided the optimal results for
5.1. MLP-ANN architecture design 100 epochs. In Fig. 3 it can be seen the evolution of the loss function
during the training process and the tuning process for the optimal MLP-
Unlike, for example, Convolution Neural Networks (CNNs) or Recur- ANN obtained. It can be observed that at 100 epochs, the train process
rent Neural Networks (RNNs), Multilayer Perceptron Artificial Neural stabilizes. More data, details and source code on the tuning process can
Networks (MLP-ANNs) are simpler ANNs. However, they are appropri- be found in Cagigas-Muñiz (2023).
ate for solving regression problems, such as estimating the joint angles The final MLP-ANN architecture after the tuning process is shown
of a robot manipulator in an IK problem. in Fig. 4. It is composed of an input layer with the (𝑋, 𝑌 , 𝑍) coordinates
One of the critical problems with an MLP-ANN and an articulated of the position of the end effector (Scorbot ER VII gripper) and the
robot IK problem is determining the architecture of the neural net- orientation of the pitch (𝛷) and yaw (𝛹 ). As mentioned in Section 3, the
work (Karlik and Aydin, 2000). Thus, when designing ANNs, some gripper roll (𝜌) is not relevant and is thus not considered. The output

4
D. Cagigas-Muñiz Engineering Applications of Artificial Intelligence 126 (2023) 107175

Table 2
Join workspace divisions/partitions considered for the Scorbot ER VII.
Joint Rotation Division/Partition Division/Partition Division/Partition
angle range A B C
𝜃1 (Base) (−125◦ , +125◦ ) 2 2 2
𝜃2 (Shoulder) (-85◦ , +85◦ ) 2 3 3
𝜃3 (Elbow) (−112,5◦ , +112,5◦ ) 2 3 4
𝜃4 (Wrist) (−90◦ , +90◦ ) 3 3 4
Totalsubspaces: 24 54 96

In Lu et al. (2022) the work space area of a six-degree articulated robot

is divided/segmented into several working sub-areas or sub-spaces.
Subareas are segmented or divided along the rotation range of each
joint. For example, the 4th robot joint has a rotation range between
0◦ and 180◦ . This area is divided in three sub-areas: 0◦ to 60◦ , 60◦ to
120◦ and 120◦ to 180◦ . A similar procedure is performed for the other
robot joints. Multiplying the divisions of the ranges of each of the six
robot joints gives 2 * 2 * 2 * 2 * 3 * 4 * 2 = 192 sub-working areas
in total. The decision about which partitions or divisions of the join
ranges is heuristic (based on expert experience). Each working subarea
is assigned to a single ANN. The architecture of the ANN is common to
every sub-area of the work space.
In this research, an analogous procedure is used for the Scorbot
ER VII robot. In this case, in the experiments, three different types of
partitions or divisions were chosen for each of the joints (see Section 6):
A, B, and C. The partitions are specified in Table 2.
In Lu et al. (2022) it is argued that processing the output of the 192
Fig. 3. Evolution of the loss function over 100 epochs. Train and data set loss functions ANNs (to find out which provides the optimal solution) is computation-
are shown. ally expensive. This problem is because one has to calculate with the
equations of the DK each point 𝑝𝑖 ′ obtained from the solution 𝑞 𝑖 ′ of the
ANNs (0 ≤ 𝑖 ≤ 191). Then the Euclidean distance between the target
layer has four joint angles: 𝜃 1 , (base), 𝜃 2 , (shoulder), 𝜃 3 , (elbow) and point 𝑝 and 𝑝𝑖 ′ (0 ≤ 𝑖 ≤ 191) must be calculated. The point 𝑝′ closest
𝜃 4 , (wrist). to the target point 𝑝 will provide the optimal solution. This procedure
It must be pointed out that in this original ANN architecture, each was specified in Section 4. Thus, Lu et al. (2022) includes a second step
joint angle has a subnet associated. Only one hidden layer per subnet in their algorithm. Three DNNs were trained to predict which of the
was used. A two-layer design per subnet was also tested during the 192 ANNs provided the optimal result. The problem with this solution
tuning process, but the results were not better: the mean error and is that the three DNNs may not be able to correctly predict which of
standard deviation of the test set were slightly higher. The optimal the 192 ANNs is the optimal solution or may simply give an incorrect
values for each hidden layer found by Keras Tuner were 16 neurons for solution. Moreover, training a common or unique ANN for every sub-
the base layer, 272 neurons for the shoulder layer, and 496 neurons for work area is not ideal. Each of the 192 sub-work areas is different, and
the elbow and wrist layers, respectively. The base layer uses the 𝑠𝑖𝑔𝑚𝑜𝑖𝑑 global performance is penalized. In the ideal case, 192 different ANNs
activation function and the rest of the subnets use the 𝑟𝑒𝑙𝑢 activation should be trained and tuned, which is costly.
function. The Keras Tuner also tested the activation functions 𝑡𝑎𝑛ℎ and
𝑠𝑜𝑓 𝑚𝑎𝑥 for each hidden layer. The output subnet layers use all 𝑙𝑖𝑛𝑒𝑎𝑟 5.3. Bootstrap sampling
functions.
The MLP-ANN architecture of Fig. 4 was trained by the Adam Bootstrap sampling is an important and well-known technique in
optimizer with a learning rate between 0.01 and 0.001 (optimal result data science, machine learning, and statistics. In statistics, bootstrap
for 0.01). An important characteristic of the proposed ANN architecture sampling is a method that involves drawing sample data repeatedly
is that the shoulder output (shoulder joint) is also input for the elbow with replacement from a data source to estimate a population param-
subnet input. Similarly, the elbow output (elbow joint) is the input of eter. In machine learning, bootstrap sampling is used in an ensemble
the wrist subnet. That is, in the coordinate plane 𝑋 −𝑍, the joint subnet algorithm called bootstrap aggregating or bagging (for short). An en-
𝑖 has as input the result of the joint subnet 𝑖 − 1, for 0 < 𝑖 < 𝑁, where semble algorithm averages multiple predictive models (or takes most
𝑁 is the total number of joints or degrees of freedom. Joint 0 is the votes) to be more accurate than just one model (Chen, 2021). In
rotatory base and joint 𝑁 is the final end-effector joint of the robot bagging, a certain number of equally sized subsets of a data set are
manipulator. This feedback between subnets (joints) is a key factor for extracted with replacement. Then, a machine learning algorithm is
good performance. applied to each of these subsets, and the outputs are ensembled.
All experiments in Section 6 use the MLP-ANN architecture of Fig. 4. The main problem of the classical ANNs solutions applied to artic-
The terms MLP-ANN or ANN will be used interchangeably in the ulated robot kinetics is that a single model (i.e. a MLP-ANN) does not
remainder of this paper. perform enough well. As mentioned in the Introduction Section 1, some
points in the robot work area can be reached using multiple joint/rotary
5.2. Segmentation of the work space area angle configurations. Any single MLP-ANN (no matter how tuned or
optimized the model is) cannot predict adequately in these scenarios.
Most of the extensions of ANNs applied to the IK problem in articu- However, an ensemble of MLP-ANNs can approximate a better rotary-
lated robots were cited in Section 2. Of these, the Ref. Lu et al. (2022) angle set solution. Inverse kinematics is a regression problem (not a
is the closest to the alternative methods proposed in this paper. It uses classification or decision problem). Here, not the average or most of
several ANNs models that cooperate to provide a more accurate result. votes is selected, but the optimal joint/rotary angle value set. That is,

5
D. Cagigas-Muñiz Engineering Applications of Artificial Intelligence 126 (2023) 107175

Fig. 4. MLP-ANN architecture to solve the IK problem and the Scorbot ER VII robot used in the experiments.

the MLP-ANN set selects the angle set (solution) 𝑞 ′ that minimizes implies that, for a time-constrained or a real-time environment, a par-
the distance between the end effector position 𝑝′ (provided by the allelization framework could be needed. The general response time of a
joint/rotation angles) and the target position 𝑝. For this purpose, the bootstrap solution will usually be lower and will be more complicated
direct kinematic equations (DK) obtained from Denavit–Hartenberg are to deploy than a single MLP-ANN or an analytic solution. Fortunately,
used as indicated in Section 4. In theory, the more models or ensemble modern parallel computer infrastructure (e.g., multiprocessors, or even
group size of MLP-ANNs, the greater the probability of finding a better GPUs) can help overcome most of these drawbacks. Nevertheless, a
rotatory angle solution. However, as shown in Fig. 5 there are some trade-off between bootstrap sampling and other techniques should be
practical limits. made before deploying a final inverse kinematic robot control solu-
Note that the data set is virtually infinite here: There are virtually tion. For example, mobile service robots (with articulated arms) may
infinite (𝑥, 𝑦, 𝑧) points and (𝛹 , 𝛷) orientations in an articulated robot have real-time requirements that on-board computers cannot meet.
Fortunately, cloud robotics (Siriweera and Naruse, 2021) can overcome
working area. Using a huge data training set in a single MLP-ANN can
computational constraints.
lead to overfitting problems, and worse still, to more problems with
Fig. 5 contains the results of applying bootstrap bagging to a 100
outliers (a set of points bad predicted by a model). However, a group
test set inside the Scorbot ER VII work area, and up to 100 models
of MLP-ANNs provides different solutions for a single 𝑝 = (𝑥, 𝑦, 𝑧, 𝛹 , 𝛷)
were tested. The model 𝑖th uses an ensemble of 𝑖 MLP-ANNs, where
point, so there are more probabilities to overcome the outlier effect
1 ≤ 𝑖 ≤ 100. The mean error and standard deviation of the 100 sample
and provide accurate solutions. Another advantage compared to the test dataset are calculated for each model. It can be seen that the mean
previous method in Section 5.2 is that the same ANN training is valid error decreases above 33% when 100 models are used instead of a
for the ANN ensemble. single MLP-ANN model. However, the mean error decreases quickly
Bootstrap sampling has some minor drawbacks in this context. Some after using only 20 models (mean error decrease above 26%) and
of them are common to the previous work space area segmentation there is no great improvement above 50 models (mean error reduction
method (Section 5.2). The training process grows linearly and accord- above 31%). The standard deviation behaves differently due to outliers.
ing to the number of models used (MLP-ANNs in this case). The number The standard deviation error reduction is only noticeable in the first
of models must be heuristic and experimental, determined according 10 models and then remains almost constant. This constant standard
to the robot and the application. During the robot control operation deviation is almost equal to the mean error. An analysis of the outlier
process, several MLP-ANNs must be calculated and then compared. This behavior is performed in Section 6.

6
D. Cagigas-Muñiz Engineering Applications of Artificial Intelligence 126 (2023) 107175

Fig. 5. Mean error and standard deviation for a 100 point Scorbot ER VII test data set and 100 models. Model 1 is the single MLP-ANN and model 𝑖th uses bootstrap bagging
with an ensemble of 𝑖 MLP-ANNs, where 1 ≤ 𝑖 ≤ 100.

Table 3
IK errors (in meters): mean, standard deviation, and quartiles. Four approaches were analyzed: a one single MLP-ANN model, work area segmentation, bootstrap
sampling, and hybrid methods. Two experiments were conducted using the same test data set (5000 samples). In the first experiment, the work area segmentation
and bootstrap sampling used 54 models (ANNs) and the hybrid method used 48 models (ANNs). In the second experiment, the number of models (ANNs) used
were 96 and 108.
54/48 One single Work area Bootstrap Hybrid method
models MLP-ANN segmentation sampling
Mean 0.112 0.080 0.076 0.065
Standard
0.077 0.075 0.069 0.067
deviation
Quartiles 0.056 0.096 0.148 0.026 0.057 0.112 0.023 0.058 0.106 0.018 0.043 0.089

96/108 One single Work area Bootstrap Hybrid method

models MLP-ANN segmentation sampling
Mean 0.112 0.071 0.073 0.059
Standard
0.077 0.070 0.068 0.064
deviation
Quartiles 0.021 0.096 0.148 0.022 0.050 0.099 0.021 0.055 0.103 0.015 0.036 0.079

5.4. Hybrid method aim is that all methods (except one single model, which only uses
one MLP-ANN) use approximately the same number of models/ANNs.
Fig. 5 shows that of about 50 models/ANN, the bootstrap sam- Furthermore, the maximum number of models tested is around 100
pling method does not substantially reduce the error in the Scor- since, as can be seen in Fig. 5, the bootstrap sampling method does
bot ER VII. One possibility is to first use models/ANNs based on not provide substantial improvements above 50 models/ANNs and
workspace segmentation and second to complement with models/ANNs practically none when using 100 models/ANNs.
based on bootstrap sampling. The aim is to combine the improve- For each set of experiments and methods, the same test data set
ments/advantages of both methods with respect to the one-single model was used. This test data set consisted of 5000 points in the robot
implementation of IK resolution with ANNs. The results of this alterna- workspace randomly generated using the DK equations. Each point 𝑝
tive hybrid method are analyzed in Section 6. was associated with it, in addition to its coordinates 𝑋, 𝑌 , 𝑍, the yaw
(𝛹 ), pitch (𝛷) and the four joint angles (𝜃1 , 𝜃2 , 𝜃3 , 𝜃4 ). The results
obtained are the 𝜀 errors made by each method and each 𝑝 point of
6. Experiments and results
the test data set. The mean, standard deviation, and quartiles of these
errors for each method and each of the two groups of experiments are
To compare the performance of all four methods described in Sec- shown in Table 3. The source code for the experiments described in this
tion 5, two sets of experiments were designed and performed. section can be downloaded in Cagigas-Muñiz (2023). This source code
The first set of experiments was performed using 54 models/ANNs can also be used to adapt the resolution of the inverse kinematics of
for the workspace segmentation and bootstrapping methods. This scheme other articulated robots.
is obtained from the Scorbot ER VII work area division indicated in Ta- Several initial results can be seen in Table 3. All methods give
ble 2 and Section 5.2. In the case of the hybrid method, the workspace better results (mean errors) than the classical approach (only one
is first divided into 24 subareas (2 × 2 × 2 × 3) and these models/ANNs model or MLP-ANN). The differences between the mean errors of the
are then trained. Subsequently, 24 additional models/ANNs of the bootstrap sampling method and the work area segmentation method
bootstrapping sampling technique were also trained and added. In total, are 5% better in favor of the former in the case of the 54/48 group of
the hybrid method uses 48 models/ANNs instead of 54. experiments and less than 3% in favor of the latter in the case of the
The second set of experiments uses the same procedure as above, 96/108 group of experiments. These results do not enable us to discern
but with 96 models/ANNs for the workspace segmentation and boot- notable improvements or advantages in either method. But the hybrid
strapping methods and 108 for the hybrid method (see Table 2). The method performs better in all indices and groups of experiments. Even

7
D. Cagigas-Muñiz Engineering Applications of Artificial Intelligence 126 (2023) 107175

Fig. 6. Error density distribution of the four analyzed methods. The work space segmentation and bootstrap sampling methods use 96 models/ANNs whereas the hybrid method
uses 108 in total. The mean error is indicated by a red vertical line.

in the first set of experiments, where the number of models/ANNs used result is because there is a linear correlation between the angle of
is lower than in the bootstrap sampling and work area segmentation yaw (𝛹 ) and the base angle (𝜃1 ). Therefore, in the specific case of the
methods. The hybrid method is capable of reducing the error by 90% Scorbot ER VII, the input of the yaw parameter (𝛹 ) and the output
compared to the one single MLP-ANN method and between 17% and of the base angle (𝜃1 ) could be removed from the model. The rest of
20% compared to the other two methods. the joint angles (shoulder, elbow, and wrist) have a better behavior or
When one looks at the standard deviation and quartiles, differenti- prediction in the case of the hybrid method. However, the prediction
ated behavior can be observed. The standard deviation and error in the of the elbow joint or axis (𝜃3 , red graph) stands out negatively. This
first quartile are lower in the case of bootstrap sampling. This result is joint is also the one with the greatest range of movement (see Table 2)
a first sign that this method is more robust to the treatment of outliers. and is thus the most sensitive to outliers. This is an expected result that
Even so, high values of the standard deviation indicate the presence of highlights the problem of joints with a high degree of rotation for the
extreme values (outliers) that require further analysis. To better analyze calculation of inverse kinematics.
this behavior, the error distribution in the four methods is shown in
Fig. 6. 7. Discussion
The error distribution functions in Fig. 6 (relating to the second
set of experiments performed) help to better discern the behavior of The experiments and results described in Section 6 demonstrate that
each of the methods. The one single MLP-ANN method presents a the original proposed method based on bootstrap sampling techniques
Gaussian-type error distribution, while the hybrid method presents a to calculate IK is more robust in the treatment of outliers. Better error
Poisson-like error distribution function. The differences between the accuracy is even achieved using the hybrid approach. The results have
bootstrap sampling method and the work area segmentation method shown that the use of multiple ANN models is, in general, a more
come mainly from the treatment of the outliers. The bootstrap sampling robust and accuracy method than just using one single custom ANN.
method has outliers of at most approximately 10 cm less error (20% Regardless of how customized the ANN architecture is. The use of
less error) than the work area segmentation method. However, the multiple models/ANNs to achieve more accurate IK computation is
latter method apparently has more predicted points with less error currently not a major problem for modern computers because of their
than the mean value. The hybrid method is capable of synthesizing large parallel computing power. However, as shown in Section 5.3,
the optimal performance of each of the two previous methods. A better there are limits to the massive use of ANNs. Certain groups of outliers
visualization of the outliers is shown in Fig. 7. cannot be correctly predicted by the methods described in this study.
Fig. 7 shows a section of the Scorbot ER VII workspace where Outliers are common in machine learning methods and, more specif-
the ‘‘dark areas’’ (outliers) can be seen. The workspace segmentation ically, in ANNs. In the case of IK and the Scorbot ER VII, these outliers
method performs optimally in areas close to the boundary of the robot’s are concentrated in certain areas of the robot workspace (‘‘dark areas’’).
workspace. In contrast, the bootstrap sampling method obtains lower These areas are not in the majority, and thanks to error handling
errors in areas inside the workspace where it is assumed that the robot they can be isolated and treated separately. These ‘‘dark areas’’ occur
can have several joint configurations (solutions) for the same point. because the articulated robot can reach these areas through multiple
The hybrid method combines both improvements, but shows that there joint configurations. Previous studies did not test the behavior of the
are still ‘‘dark areas’’ that cannot be predicted with a reasonably low proposed models in the whole robot work area. This can lead to
error. Another way to analyze the errors is by individually analyzing important errors if only certain type of trajectories (in certain sub-work
the prediction of each of the robot’s axes/joints (see Fig. 8). areas) are tested.
Fig. 8 show the predictions for each of the Scorbot ER VII axes/joints A possible improvement of behavior would necessarily involve lim-
(base, shoulder, elbow, wrist) with respect to the actual values of the iting the range of movement of robot joints in those ‘‘dark areas’’
test dataset. Only the cases of the single MLP-ANN method and the indicated in the Figs. 7. This aspect is especially critical in robot joints
hybrid method are shown. The blue plots describing the Base (𝜃1 ) that have a wide range of movements, as can be seen in Fig. 8 and in
joint/angle behavior indicate a near-perfect or ideal prediction. This the specific case of the ‘‘Elbow’’ joint of the Scorbot ER VII. Another

8
D. Cagigas-Muñiz Engineering Applications of Artificial Intelligence 126 (2023) 107175

Fig. 7. Error magnitudes corresponding to section 𝑋 > 0, 𝑌 > 0 and 𝑍 > 0 of the Scorbot ER VII work area. The work space segmentation and bootstrap sampling methods use
96 models/ANNs whereas the hybrid method uses 108 in total. Darker areas represent bigger IK prediction errors. White areas have tiny or nearly no errors.

strategy could imply recording robot positions in the center of each predicted by ANNs controllers should be identified, isolated, and then
‘‘dark area’’ cluster. Then refine the position to approximate the target treated using specific models.
point using optimization algorithms or analytical solutions. Finally, other interesting future work related to this study could
With all this, the results of research evidence that articulated robot be done. First, some parallel implementation performance needs to
controllers based exclusively on ANNs under some restrictions could be analyzed, especially for real-time systems and service robots with
be designed and operated in real time on parallel hardware. The constrained computational resources. Second, these solutions could be
disadvantage of these methods is that the ‘‘dark areas’’ (outliers) must extended to other industrial robots types like, for example, cylindrical,
first be isolated from the robot’s working area. This process is made Scara or parallel robots.
possible by calculating the errors made in a training set. Second, the
IK controller would have to be designed so that, depending on the CRediT authorship contribution statement
working area in which the target point is located (and the application
error tolerance), select ANNs hybrid method output or use some of Daniel Cagigas-Muñiz: Conceptualization, Methodology, Software,
the alternative strategies commented in the last paragraph. Despite the Validation, Data curation, Formal analysis, Investigation, Writing –
technical difficulty of this process, it is possible to automate it and even original draft, Visualization, Investigation, Supervision, Validation,
include it in a generic software environment that could be used for Writing – review & editing.
any articulated robot. Thus, this work is a step toward such a universal
Declaration of competing interest
solution for IK and articulated robots.

The authors declare the following financial interests/personal re-

8. Conclusions
lationships which may be considered as potential competing inter-
ests: Daniel Cagigas-Muñiz reports financial support was provided by
Inverse kinematics solutions for articulated robots found in the University of Seville.
scientific literature using artificial neural networks can be substantially
improved. The results described in this paper indicate that hybrid Data availability
techniques can help provide valid solutions for generic articulated
robots under some constraints. However, the inverse kinematic problem The source code is available at github: https://github.com/dcagigas/
in articulated robots with relatively high degrees of freedom is still an Robotics.
open problem. Recent references in this field demonstrate the interest
of the scientific community in this topic. Therefore, the bootstrap sam- Acknowledgments
pling and work area segmentation techniques described in this research
can be extended, combined, and complemented with other previous This research was supported by the CHIST-ERA H2020 grant SMALL
studies proposals. These extensions may improve the inverse kinematic (CHIST-ERA-18-ACAI-004, PCI2019-111841-2/AEI/10.13039/5011000
results for any generic articulated robot. ‘‘Dark areas’’ (outliers) bad 11033).

9
D. Cagigas-Muñiz Engineering Applications of Artificial Intelligence 126 (2023) 107175

Fig. 8. Hexbin plots that compare True vs Prediction joint values (in radians) for the one single MLP-ANN architecture (a) and hybrid method and (54+54) 108 models/MLP-ANNs
(b).

References El-Sherbiny, A., Elhosseini, M.A., Haikal, A.Y., 2018. A comparative study of soft
computing methods to solve inverse kinematics problem. Ain Shams Eng. J. 9 (4),
2535–2548. http://dx.doi.org/10.1016/j.asej.2017.08.001.
Adar, N.G., 2021. Real time control application of the robotic arm using neural
Ghadami, N., Gheibi, M., Kian, Z., Faramarz, M.G., Naghedi, R., Eftekhari, M.,
network based inverse kinematics solution. Sakarya Univ. J. Sci. 25 (3), 849–857.
Fathollahi-Fard, A.M., Dulebenets, M.A., Tian, G., 2021. Implementation of solar
http://dx.doi.org/10.16984/saufenbilder.907312.
energy in smart cities using an integration of artificial neural network, photovoltaic
Almusawi, A.R.J., Dülger, L., Kapucu, S., 2016. A new artificial neural network system and classical Delphi methods. Sustainable Cities Soc. 74, 103149. http:
approach in solving inverse kinematics of robotic arm (denso VP6242). Comput. //dx.doi.org/10.1016/j.scs.2021.103149.
Intell. Neurosci. 2016, 10. http://dx.doi.org/10.1155/2016/5720163. Ghasemi, J., Moradinezhad, R., Hosseini, M.A., 2019. Kinematic synthesis of parallel
Aysal, F., Çelik, I., Cengiz, E., Oğuz, Y., 2022. A comparison of multi-layer perceptron manipulator via neural network approach. arXiv:1904.04668.
and inverse kinematic for RRR robotic arm. J. Polytech. http://dx.doi.org/10.2339/ Haug, E.J., 2023. A cyclic differentiable manifold representation of redundant
politeknik.1092642. manipulator kinematics. J. Mech. Robotics 1–9. http://dx.doi.org/10.1115/1.
Barrientos, A., 2007. Fundamentos de Robótica. Mc Graw-Hill. 4063038.
Cagigas-Muñiz, D., 2023. Robotics software repository. URL: https://github.com/ Hayat, A., Chittawadigi, R., Udai, A., Saha, S., 2013. Identification of denavit-
dcagigas/Robotics. hartenberg parameters of an industrial robot. In: ACM International Conference
Chen, L.-P., 2021. Practical statistics for data scientists: 50+ essential concepts using R Proceeding Series. pp. 1–6. http://dx.doi.org/10.1145/2506095.2506121.
Ho, C.-K., King, C.-T., 2022. Automating the learning of inverse kinematics for robotic
and Python. Technometrics 63 (2), 272–273. http://dx.doi.org/10.1080/00401706.
arms with redundant DoFs. arXiv:2202.07869.
2021.1904738.
Jha, B.B., 2014. A neural network approach for inverse kinematic of a SCARA
Cole, L., Nagy-Sochacki, A.F., Symonds, J., 2007. Drawing using the SCORBOT-ER
manipulator. Int. J. Robot. Autom. 3 (1), 56–61. http://dx.doi.org/10.11591/ijra.
VII manipulator arm. https://www.lukecole.name/doc/reports/drawing_using_the_
v3i1.3201.
scorbot_manipulator_arm.pdf. Jiang, J., You, J., Bi, Y., 2023. Kinematic modeling and simulation of a new robot
Daya, B., Khawandi, S., Akoum, M., et al., 2010. Applying neural network architecture for wingbox internal fastening application. Machines 11 (7), http://dx.doi.org/10.
for inverse kinematics problem in robotics. J. Softw. Eng. Appl. 3 (03), 230. 3390/machines11070753.
http://dx.doi.org/10.4236/jsea.2010.33028. Karlik, B., Aydin, S., 2000. An improved approach to the solution of inverse kinematics
Duka, A.-V., 2014. Neural network based inverse kinematics solution for trajectory problems for robot manipulators. Eng. Appl. Artif. Intell. 13 (2), 159–164. http:
tracking of a robotic arm. Proc. Technol. 12, 20–27. http://dx.doi.org/10.1016/j. //dx.doi.org/10.1016/S0952-1976(99)00050-0.
protcy.2013.12.451, The 7th International Conference Interdisciplinarity in Engi- Köker, R., Çakar, T., 2016. A neuro-genetic-simulated annealing approach to the inverse
neering, INTER-ENG 2013, 10-11 October 2013, Petru Maior University of Tirgu kinematics solution of robots: A simulation based study. Eng. Comput. 32 (4),
Mures, Romania. 553–565. http://dx.doi.org/10.1007/s00366-015-0432-z.

10
D. Cagigas-Muñiz Engineering Applications of Artificial Intelligence 126 (2023) 107175

Larsson, A., Grönlund, O., 2023. Comparative Analysis of the Inverse Kinematics of a Sharkawy, A.-N., Khairullah, S., 2023. Forward and inverse kinematics solution of a 3-
6-DOF Manipulator : A Comparative Study of Inverse Kinematics for the 6-DOF Saab DOF articulated robotic manipulator using artificial neural network. Int. J. Robotics
Seaeye eM1-7 Manipulator with Non-Conventional Wrist Configuration. Linköping Control Syst. 3 (2), 330–353. http://dx.doi.org/10.31763/ijrcs.v3i2.1017.
University, Automatic Control, p. 58. Siriweera, A., Naruse, K., 2021. Survey on cloud robotics architecture and model-
Liana López-Pacheco, O.A.D.-R., 2012. Scorbot ER-VII virtual: modelado matemático y driven reference architecture for decentralized multicloud heterogeneous-robotics
control de movimiento. Editorial academica española. platform. IEEE Access PP, 1. http://dx.doi.org/10.1109/ACCESS.2021.3064192.
Lu, J., Zou, T., Jiang, X., 2022. A neural network based approach to inverse kinematics Snoek, J., Larochelle, H., Adams, R.P., 2012. Practical bayesian optimization of machine
problem for general six-axis robots. Sensors 22 (22), http://dx.doi.org/10.3390/ learning algorithms. In: Advances in Neural Information Processing Systems. pp.
s22228909. 2951–2959.
Mousapour Mamoudan, M., Ostadi, A., Pourkhodabakhsh, N., Fathollahi-Fard, A.M., Srinivas, N., Krause, A., Kakade, S.M., Seeger, M.W., 2009. Gaussian process bandits
Soleimani, F., 2023. Hybrid neural network-based metaheuristics for prediction of without regret: An experimental design approach, CoRR abs/0912.3995. URL:
financial markets: a case study on global gold market. J. Comput. Des. Eng. 10 http://arxiv.org/abs/0912.3995.
(3), 1110–1125. http://dx.doi.org/10.1093/jcde/qwad039. Srisuk, P., Sento, A., Kitjaidure, Y., 2017a. Forward kinematic-like neural network for
Nosova, N.Y., 2020. A review of the parallel structure mechanisms with kinematic solving the 3D reaching inverse kinematics problems. In: 2017 14th International
decoupling. In: Misyurin, S.Y., Arakelian, V., Avetisyan, A.I. (Eds.), Advanced Conference on Electrical Engineering/Electronics, Computer, Telecommunications
Technologies in Robotics and Intelligent Systems. Springer International Publishing, and Information Technology. ECTI-CON, pp. 214–217. http://dx.doi.org/10.1109/
Cham, pp. 247–255. http://dx.doi.org/10.1007/978-3-030-33491-8_30. ECTICon.2017.8096211.
O’Malley, T., Bursztein, E., Long, J., Chollet, F., Jin, H., Invernizzi, L., et al., 2019. Srisuk, P., Sento, A., Kitjaidure, Y., 2017b. Inverse kinematics solution using neural
KerasTuner. https://github.com/keras-team/keras-tuner. networks from forward kinematics equations. In: 2017 9th International Conference
Pouresmaeil, H., Faramarz, M.G., ZamaniKherad, M., Gheibi, M., Fathollahi-Fard, A.M., on Knowledge and Smart Technology. KST, pp. 61–65. http://dx.doi.org/10.1109/
Behzadian, K., Tian, G., 2022. A decision support system for coagulation and KST.2017.7886084.
flocculation processes using the adaptive neuro-fuzzy inference system. Int. J. Toquica, J.S., Oliveira, P.S., Souza, W.S., Motta, J.M.S., Borges, D.L., 2021. An
Environ. Sci. Technol. 19, http://dx.doi.org/10.1007/s13762-021-03848-4. analytical and a deep learning model for solving the inverse kinematic problem
Predescu, L., Stroe, I., 2015. On the kinematics of the scorbot er-VII robot. In: UPB of an industrial parallel robot. Comput. Ind. Eng. 151, 106682. http://dx.doi.org/
Scientific Bulletin, Series D: Mechanical Engineering, Vol. 77. pp. 35–46. 10.1016/j.cie.2020.106682.
Ren, H., Ben-Tzvi, P., 2020. Learning inverse kinematics and dynamics of a robotic Wagaa, N., Kallel, H., Mellouli, N., 2023. Analytical and deep learning approaches for
manipulator using generative adversarial networks. Robot. Auton. Syst. 124, solving the inverse kinematic problem of a high degrees of freedom robotic arm.
103386. http://dx.doi.org/10.1016/j.robot.2019.103386. Eng. Appl. Artif. Intell. 123, 106301. http://dx.doi.org/10.1016/j.engappai.2023.
Shahsavar, M.M., Akrami, M., Gheibi, M., Kavianpour, B., Fathollahi-Fard, A.M., 106301.
Behzadian, K., 2021. Constructing a smart framework for supplying the biogas Zhan, C., Zhang, X., Yuan, J., Chen, X., Fathollahi-Fard, A., Wang, C., Wu, J., Tian, G.,
energy in green buildings using an integration of response surface methodology, 2023. A hybrid approach for low-carbon transportation system analysis: integrating
artificial intelligence and petri net modelling. Energy Convers. Manage. 248, CRITIC-DEMATEL and deep learning features. Int. J. Environ. Sci. Technol.: IJEST
114794. http://dx.doi.org/10.1016/j.enconman.2021.114794. 1–14. http://dx.doi.org/10.1007/s13762-023-04995-6.

11