Hindawi

Mobile Information Systems

Volume 2022, Article ID 8715036, 9 pages
https://doi.org/10.1155/2022/8715036

Research Article
Design and Simulation Analysis of Trajectory Planning
Algorithm for 6-DoF Manipulator Based on Deep Learning

Min Zhuang ,1 Ge Li,2 and Kexin Ding1

1
Hangzhou Polytechnic, Hangzhou 311402, China
2
Zhejiang Sci-Tech University, Hangzhou 310018, China

Correspondence should be addressed to Min Zhuang; [email protected]

Received 1 April 2022; Revised 4 May 2022; Accepted 6 May 2022; Published 14 June 2022

Academic Editor: Liping Zhang

Copyright © 2022 Min Zhuang et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
At present, the robot developed only has partial intelligence, and it cannot be well controlled independently for some jobs, so it
should optimize its path and trajectory. However, as an industrial robotic arm robot necessary for advanced industrial
manufacturing enterprises, it has been replacing humans to perform simple and repetitive actions, and it is doing better and better.
For the six-degree-of-freedom (6-DoF) manipulator, this paper aims to optimize the trajectory planning of the manipulator based
on deep learning technology. Therefore, this paper presents a fuzzy control algorithm for the trajectory planning of the ma-
nipulator based on the neural network algorithm. The experimental results show that for the established three-joint dynamic
model, the moment received by joint 1 is the largest, followed by joint 2, and joint 3 is the smallest, which can better simulate the
trajectory of the robotic arm.

1. Introduction The dynamic parameters of the robotic arm itself are not
easy to obtain, and there are interference and other un-
For a 6-DoF manipulator, the end of the manipulator avoidable factors in the outside world. It is a complex,
generally performs tasks in Cartesian space. No matter strongly coupled nonlinear system. Its trajectory planning
what kind of trajectory, it is necessary to perform inter- and tracking control is also a complex process. In order to
polation in Cartesian space to obtain a sequence of in- describe the motion state of the manipulator, it is nec-
terpolation points and then map it to joint space to obtain essary to carry out kinematic analysis and dynamic
joint angles through inverse kinematics solution. To analysis. These factors cause changes in the tracking ac-
complete an ideal trajectory planning, it is necessary to curacy of the robotic arm. At the same time, the dynamic
ensure that the curves in both Cartesian space and joint model of the manipulator is difficult to establish accu-
space are smooth. It can be engaged in welding, spraying, rately. How to choose a suitable control algorithm that
handling, blanking, assembly, packaging, palletizing, and does not require an accurate dynamic model is a difficult
other work. On the one hand, it reduces the workload of problem to achieve high-precision trajectory tracking.
human beings and liberates human hands. On the other Therefore, under this situation, the development of Chi-
hand, it can improve the production efficiency of enter- nese robots faces challenges to varying degrees. This also
prises and reduce labor costs through automated pro- requires that research and technology in related fields can
duction of machinery. After planning the trajectory of the be quickly followed up, so that the development of robots
manipulator and obtaining the desired trajectory, the can keep up with the pace of social development. This
manipulator must accurately track the established tra- article is to conduct research in related fields, and expect to
jectory. First, the joint action of the robotic arm requires gain something in this field.
the servo motor to apply a torque to the joint, and the This paper takes the 6-DoF manipulator as the re-
magnitude of this torque needs to be solved dynamically. search object and studies its trajectory planning and
2 Mobile Information Systems

tracking control. For trajectory planning, a kinematic 3. Robotic Arm Trajectory Planning and Deep
analysis of the object is required first. It obtains its inverse Learning Neural Network
kinematics solution and then selects an appropriate
planning algorithm to obtain the function of joint pa- 3.1. Robotic Arm Trajectory Planning. The trajectory plan-
rameters changing with time. The research of tracking ning of the robotic arm is to plan the angle, speed, and
control algorithm requires dynamic modeling of the acceleration of each degree of freedom, so that each joint can
manipulator. However, the dynamic model of the 6-DoF coordinate and complete the task continuously and stably
manipulator is very complicated, and it is difficult to [7]. The motion of the manipulator has joint friction, which
build it completely. Therefore, we propose several in the long run will lead to the wear and tear of the ma-
tracking control algorithms without detailed model in- nipulator structure, resulting in a large error of the ma-
formation and compare their tracking performance by nipulator [8]. Among them, the mechanical arm with more
simulation. applications is shown in Figure 1.
Trajectory planning algorithms have experienced a
2. Related Work process from easy to difficult, and there are many kinds. The
advantages and disadvantages of various planning algo-
The application of 6DOF manipulator in industry is very rithms [9, 10] are shown in Table 1.
extensive, and related scholars have conducted in-depth Trajectory tracking control is defined as the ability to
research on it. Huang and Lan proposed a robot-assisted and accurately track the given desired trajectory of the manip-
remotely controlled ultrasound scanning system for 3D ulator during motion after determining various manipulator
imaging. A 6-DoF (DOF) robotic arm that acts as an upper- parameters and state variables of the working end [11].
limb prosthesis remotely controlled by an operator drives an Because manipulator control is a complex system, it has
ultrasound probe to scan the skin surface. The quantitative characteristics such as uncertainty, strong coupling, and
experimental results show that the volume measurement nonlinearity, and there will be some factors in manipulator
error is less than 1.1%. This shows that the system in this modeling. As shown in Table 2, these factors are frequently
paper can accurately and flexibly control the probe scanning unavoidable. As a result, accurately establishing the ma-
to produce high-quality 3D ultrasound images [1]. Silveira nipulator’s dynamic model is difficult.
further exploited epipolar geometry and its inherent de- These uncertain factors as shown in the table may reduce
generacy. This degradation always occurs when the stability the accuracy of tracking control, or even fail to meet ex-
is close enough to equilibrium, regardless of the shape of the pectations, making the system unstable. In order to deal with
object. He then introduced three new local methods and uses these problems, the currently commonly used manipulator
both planar and nonplanar objects. He simulated and ex- trajectory tracking control methods are as follows.
perimentally evaluated their closed-loop performance using (1) PID control: the main feature is that it is simple and
6DOF manipulators at small and large displacements [2]. easy to understand, and it is the most widely used control
Donnici M designed a system. The system consists of a high- algorithm for industrial robots. However, PID control is
precision 6-DoF self-balancing arm that is able to move in difficult to ensure the dynamic performance of the manip-
passive and active modes, allowing physicians to accurately ulator, and the output torque is large when starting, which is
insert needles [3]. Zwonarz and Turnau believed that tra- easy to cause damage to the manipulator. (2) Adaptive
jectory optimization must be done for the optimization of control: when the system is running, it continuously deter-
industrial robots [4]. Smyrnaiou et al. conducted a com- mines the state of the controlled object and compares it with
parative study of classical control methods for testing the expected state. It corrects the structure and parameters of
mathematical models. This model uses a single controller to the controller in real time and adjusts the system adaptively by
control the six actuators of a 6DOF robotic arm and aims to changing the operation of the system according to the
constructively simplify the system. In more detail, he adaptive law, so that the system can reach the optimal state.
designed the mathematical model of the system, simulating Adaptive control needs to identify parameters online, and the
all mechanical components, including the mathematical real-time requirements are relatively strict. (3) Iterative
model of the five-way pneumatic reversing valve, pneumatic learning control: each iterative process is equivalent to a
actuator/piston, and controller [5]. Song et al. combined control attempt, correcting the deviation between the output
CAD technology to optimize the application of manipulator. and the expectation by the error between the two, and then
He thought that the good drawing and model fitting proceeding to the next attempt until full tracking is achieved.
function of CAD can provide good help for the path This control algorithm is more suitable for objects with re-
planning of the robot arm. In the experiment, he simulated petitive actions, and the tracking effect of iterative learning is
with three different mechanical arms, and the results showed different from that of progressive tracking. Maybe the
that the recognition rates of the three objects were all above tracking effect of the first iteration is not very good. However,
90% [6]. In the actual tracking control research, many as the number of iterations increases, it is able to achieve full
scholars use the hybrid control method. They use the tracking effect, requiring less prior knowledge [12, 13]. (4)
abovementioned control methods in combination, because Fuzzy control: using the combination of expert experience
different control methods have their own advantages in and control methods, it can be combined with other control
different aspects, and the combined use can complement methods and does not require an accurate mathematical
each other. model of the robotic arm system. However, there are many
Mobile Information Systems 3

Figure 1: Different types of 6-DoF robotic arms.

Table 1: Trajectory planning algorithms.

Trajectory planning
Advantage Shortcoming
methods
Lines and arcs Relatively simple and intuitive, small amount of calculation The applicable surface is relatively narrow
Unrestricted acceleration, shocks exist, and
Cubic polynomial Simple, small amount of computation
limitations are large
Higher-order High motion efficiency and smoothness, the higher the
Vibration prone to high frequency
polynomials polynomial degree, the better the accuracy and fit
Fully guarantee the smooth transition of the speed and There is a step in the acceleration curve, which is
S Speed curve
acceleration of the entire trajectory easy to cause mechanical wear
Large amount of calculation and high hardware
Spline Smooth trajectory, continuous speed and acceleration
requirements

Table 2: Uncertainties. robots in China can be divided into five stages, and now it
Robotic arm factor Joint factor External factors
has entered a period of rapid development. However,
China’s research on robotics technology is relatively back-
Working environment
Quality Kinetic friction ward, and some high-end core technological achievements
noise
Length Static friction Measurement error are mainly concentrated in the United States, Japan, Ger-
Moment of inertia Joint flexibility Other factors many, and other countries. Therefore, it has certain practical
significance to conduct in-depth research on all aspects of
robotics, as shown in Figure 2.
parameters to be adjusted in fuzzy control, and it is not used
much in practical engineering. (5) Sliding mode variable 3.2. Deep Learning Neural Networks. A brief description of
structure control: this is a type of nonlinear control. There is a machine learning [14, 15] is as follows: it is given a source
switching function in the system, which can transform the space x and a target space y. The goal is to learn a mapping
system structure according to the law of the switching from x to y. This makes it possible to obtain the corre-
function according to the current state of the controlled sponding image in y through f for any sample. For example,
object. If the state reaches the value of the switching function, for a Chinese to English translation task, x is the source
the system switches the structure, which makes the system language (Chinese) space, y is the target language (English)
state slide along the sliding surface and finally reaches the space, and f is the translator. Deep learning is to use the
equilibrium point of the system. The advantages of this artificial neural network to simulate the neural network of
control method are fast response, insensitivity to parameter biological brain to process data information; that is, f is an
changes and simple implementation. artificial neural network. It requires a large amount of data
In 2018, a report on the development of the robot in- for effective learning, and at the same time, it has high
dustry pointed out that China has nearly one-third of the computational complexity and requires hardware system
world’s robots, has a vast application market, and ranks support. Therefore, although it has been proposed in the last
among the best in the world. In the future, China’s industrial century, it has not been paid attention to, and it has been
robot market will increase. The development of industrial developed rapidly until the last decade.
4 Mobile Information Systems

tandem robotic arm. The serial manipulator uses multiple

Installed capacity of industrial robots

45
40 links to connect the joint axes of the manipulator in series. It
35 uses the motor to drive the movement of each joint axis and
drives the connecting rod to perform relative movement, so
(10,000 units)

30
25 that the end of the mechanical arm can complete various
20 tasks. For the convenience of research, it numbers each axis
15
of the manipulator in turn. From the base to the end effector,
10
there are 1–6 axes from bottom to top. As shown in Figure 3:
The robotic arm uses a motion control card as a control
5
unit, providing an open experimental platform, and con-
0
2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 trolling the robotic arm by accepting commands from the host
years computer. Its basic parameters are shown in Tables 3 and 4.
The overall hardware structure of the robotic arm mainly
Figure 2: The global installed capacity of industrial robots an-
consists of two parts: the robotic arm control and the upper
nounced by IFR in 2019.
computer control part. The upper computer realizes the
planning algorithm, the tracking control algorithm, and the
Deep learning [16, 17] uses artificial neural network as design of the human-computer interaction interface through
the architecture to perform representation learning on data, software programming. The command of the host computer is
and automatically mine and extract the features and in- transmitted to the motion controller through the 10 M Ethernet,
formation contained in the data for learning. Therefore, the and the driver of each joint controls the motor to rotate, so that
structure of deep neural network [18] is very important for the joint of the robotic arm rotates to an appropriate angle. The
the effect of deep learning. The network model is also getting system structure diagram is shown in Figure 4.
more and more complex, and the effect is getting better and The hardware part of the upper computer consists of a
better. Residual network (ResNet) in 2015 has become a computer and a network card. Data transmission between
standardized module for processing image information in devices is realized through the network card. The computer
many tasks due to its excellent performance. The network mainly does the work of controlling external devices. The
model applied to target detection is more complex and specific work is as follows: it sends control commands and
contains many functional modules. Recurrent Neural Net- query commands to the motion control card.
works (RNNs) applied to sequence learning, especially Long The robotic arm control system receives and executes the
Short-Term Memory Networks (LSTMs) and Gated Re- control commands from the host computer, receives the
current Units (GRUs), are widely used in various sequence query commands from the host computer and feeds back the
learning problems (such as natural language processing). In working conditions of the motor. It is mainly composed of
2016, the model structure of the Transformer based on the four parts, namely, DMC-21X2 motion control card, driver,
attention mechanism proposed by Google was based on the and motor. The components of the robotic arm motion
self-attention network, which greatly surpassed the model control system are shown in Figure 5. The specific functions
based on the long short-term memory network in the neural of each part are as follows:
machine translation task, and was extended to other natural
language processing tasks. In speech processing tasks, neural (1) The motion controller DMC-21x2 is an economical
networks also show their advantages. For example, in speech multiaxis independent controller. When the value of
synthesis (TTS) tasks, Google’s Tacotron series, Microsoft’s x is 1–8, the motion controller can be used to control
Transformer-based model, and Baidu’s DeepVoice series all up to 8 axes. The controller can smooth curves for
surpass traditional algorithms. It can be seen that different complex contours. In order to reduce the motion
neural network model structures need to be designed for impact and achieve smooth tracking, the controller
different tasks and different data. This process is not done provides infinite linear and arc vector feeds. That is,
overnight, but it requires deep expertise. This requires both any complex curve can be represented by straight
mastery of deep learning techniques and knowledge of the lines and arcs. (2) The driver should be matched with
field of application. This makes the use of deep learning the servo motor, and the pulse signal of the controller
technology more expensive, and the cost of designing a cannot be directly used to control the servo motor. It
suitable neural network model structure is relatively high. needs to be converted into electric current by the
Therefore, researchers have explored a method that can servo driver to drive the servo motor. (3) The motor
automatically design the neural network structure, and converts current into force that produces motion.
proposed a neural network structure search. It has also The magnitude of the current determines the mag-
become the hottest research area in automated machine nitude of the force of motion. The motor drives the
learning. joint movement. (4) The encoder is used to record
the running state of the motor. It uses its own
4. Six-Degree-of-Freedom Robotic Arm function to convert the motion state of the motor
into electrical pulses and feed it back to the con-
The robotic arm used in this article is the REBot-V-6R-650 troller, so as to correct the deviation of the motor in
6-DoF robotic arm produced by a company. It is a typical the actual operation.
Mobile Information Systems 5

PC
Trajectory Human-computer
310 mm host planning interaction
computer

1 2 3
4
DMC-21X2

220 mm
server Driver server Driver
Robotic arm

……
5
control system
servo motor servo motor

6
75 mm
Six degrees of freedom robotic arm

235 mm Figure 4: Hardware system structure diagram.

Figure 3: Experimental simulation of the robotic arm control unit. Kinematics analysis [19] (Kinematic Analysis) includes
forward kinematics analysis (Forward Kinematic Analysis)
Table 3: REBot-V-6R-650 technical parameters. and inverse kinematics analysis [20] (Inverse Kinematic
Analysis). Forward analysis refers to the angle through the
Mechanical structure 6 degrees of freedom in series joint and the length of each link. It calculates the end in-
Load weight 3 kg formation of the robotic arm, and the process is relatively
Positioning accuracy ±0.08 mm simple. Reverse analysis refers to calculating the angle of
Body weight 60 kg
each joint according to the mathematical method under the
Battery capacity 220 V 1.5KVA
Working radius 723 mm
condition of obtaining the information of the end of the
manipulator and the length of each link. Reverse analysis is
much more difficult than forward solution. In order to solve
Table 4: Specific parameters of joint axis. the trajectory planning problem, we must first carry out
kinematic analysis to find out the poses of the other joints
Maximum range of motion Maximum speed (s) when the end effector of the robotic arm is located at a
1 axis ±170° 180° certain point, and determine its joint angle.
2 axes +135∼ −45° 180°
3 axes +130°∼−70° 225°
4 axes ±170° 300° 5. Simulation Analysis of 6-DOF Robotic Arm
5 axes ±120° 375°
6 axes ±360° 500° The idea of the inversion design method is to decompose a
complex nonlinear system into multiple subsystems. It then
designs the Lyapunov function (V function) for the sub-
The first step, as shown in Figure 5, is to establish a system to ensure the stability of the system and introduces
connection with the robotic arm. 10M Ethernet is used for virtual control quantities. The former system must achieve a
communication between the upper computer and the ro- stable effect through the virtual control of the latter system.
botic arm. It assigns each connected device in the same If the order of the original system is n, then the system needs
network segment its own IP address and uses the TCP/IP to be decomposed into n or less than n subsystems. The
communication protocol. The pose of the robotic arm is inversion method is often combined with a Lyapunov-type
judged after the connection is successful, and the trajectory adaptive law to make the entire system meet the desired
planning information is obtained. At the robotic arm’s end, dynamic and static performance. The robotic arm system is a
it decomposes trajectory information into pose sequences. complex nonlinear system. As a universal approximator,
The inverse kinematics solution is used to obtain the desired fuzzy control is widely used in the study of nonlinear sys-
angle of the joint corresponding to the robot arm pose of the tems. It can be used to approximate the control law of the
trajectory point. After receiving the command, the motion system directly, or it can be used to approximate the dy-
card edits it and sends the parsed command to the motor’s namic characteristics of the system, and then design the
driver. At this point, the motor driver directs the robotic control law. The dynamic formula of the manipulator is
arm’s joints to move, and the encoder determines the 􏼈 M(q)q€ + C(q, q)
_ + G(q) + d � τ, y � q. (1)
motor’s position and adjusts it appropriately.
6 Mobile Information Systems

computer DMC-21x2 Motion Controller

driver

power supply

Encoder
E d motor
Figure 5: Motion control system components.

It is assumed that the parameters are unknown and The main purpose of designing the control system is to
bounded, and the system has the following characteristics: make the output qd of the control system well track the
desired trajectory qd.
(1) (q) is a positive definite matrix and bounded.
According to the backstepping method, the control
(2) The inertia matrix (q) has the following relationship system is decomposed into two subsystems.
with the centripetal force and Coriolis moment It defines x1 � q, and x2 � q, then combined with formula
_
(q, q): (1), the values of x_ 1 x_ 2 and y can be solved as follows:

_
x(M(q) _ T � 0.
− 2C(q, q))x (2)

⎪
⎧ x_1 � x2 ,
⎪
⎨
⎪ x_ � M− 1 x1 􏼁τ − M− 1 x1 􏼁C x1 , x2 􏼁x2 − M− 1 x1 􏼁G x1 􏼁 − M− 1 x1 􏼁d, (3)
⎪ 2
⎩
y � x1 .

α1 � −λ1 e1 + y_d , (7)

It assumes that y is the control objective, yd is the desired λ1 > 0.

trajectory, and yd has a second derivative. For the first subsystem in formula (1), it takes the
For the ith subsystem, the alternative Lyapunov function Lyapunov function as
Vi is chosen as
1
1 V1 � eT1 e1 . (8)
Vi � Vi−1 + e2i , (4) 2
2bi
Then,
b1 � 1 and b2 � M− 1 .
V_ 1 � −λ1 eT1 e1 + eT1 e2 . (9)
Step 1. For the first subsystem x_ 1 � x2 , it defines the error as
If e2 � 0, the first subsystem is stable.
e1 and introduces a virtual control variable a1
e 1 � y − yd Step 2. It designs the control law
􏼨 (5)
e 2 � x 2 − α1 . From formulas (1) and (4), we get

It makes the value of e2 approach zero by choosing a1 . e_2 � −M− 1 Cx2 − M− 1 G − M− 1 d − M− 1 τ − α_1 . (10)
Then,
It takes the control law as
e_1 � x_1 − y_d � x2 − y_d � e2 + α1 − y_d . (6)
τ � −λ2 e2 − e1 − φ. (11)
It defines the virtual control quantity as
Mobile Information Systems 7

Fuzzy parameter adaptive law

φ1
system status x1
Fuzzy Recognizer 1

x1
- e1
yd K1 Fuzzy Controller 1

Fuzzy parameter adaptive law

φ2
system status x1, x2
Fuzzy Recognizer 2
-

e2 u y
x2 K2 Fuzzy Controller 2 controlled object

Figure 6: Adaptive fuzzy inversion control.

For the second subsystem in formula (4), it takes the ξ � 􏼂ξ 1 (x)ξ 2 (x) . . . ξ N (x)􏼃 ,
T
Lyapunov function as
1 􏽑nj�1 uij 􏼐xj 􏼑 (17)
V2 � V1 + eT1 Me2 . (12) ξ i (x) �
2 􏽐N n i T
i�1 􏽑j�1 uj 􏼐xj 􏼑, θ � 􏼂θ1 θ2 . . . θN 􏼃 .
Then,
For the fuzzy approximation of the f function, it adopts
V_ 2 � −λ2 eT1 e1 − λ2 eT2 e2 + eT1 (f − φ) − eT2 d. (13) the form of approximating f(1) and f(2), respectively, and the
corresponding fuzzy system is designed as
In i
􏽐N n
i�1 θ1i 􏽑j�1 uj 􏼐xj 􏼑
f � −Cα1 − Mα_1 − G, (14) φ1 (x) � � ξ T1 (x)θ1 ,
􏽐N
i�1 􏽑nj�1 uij 􏼐xj 􏼑
the function f contains the inertia matrix C the Coriolis force (18)
matrix M and the gravity matrix G of the robotic arm and 􏽑nj�1 uij 􏼐xj 􏼑
φ2 (x) � 􏽐N
i�1 θ2i i T
contains the modeling information of the system. The de- 􏽐N n
i�1 􏽑j�1 uj 􏼐xj 􏼑 � ξ 2 (x)θ2 .
rivative of the unknown function and virtual control term in
the system needs to be approximated by a fuzzy system. It defines
Assuming that φ is a fuzzy system used to approximate f, it
needs to be defuzzified by single-valued fuzzification, ξ T1 0 θ1
Φ � 􏼂φ1 φ2 􏼃 � ⎢⎣⎡ ⎥⎦⎤􏼢
T T
􏼣 � ξ (x)θ. (19)
product inference engine, and centroid-averaged 0 ξ T2 θ2
defuzzification.
It assumes that the fuzzy system consists of N fuzzy rules, It defines the optimal approximation constant θ∗ , for a
where the f fuzzy rule is expressed as given arbitrarily small constant ε (ε > 0), the following in-
equality holds as ‖f − Φ∗ ‖ ≤ ε.
Ri : IF xi is ui1 and . . . xn is uin , then y is Bi (i � 1, 2, . . . , N), It makes θ􏽥 � θ∗ − θ, and the adaptive control law is
designed as
(15)
T
uij
where is the membership function of xi (j � 1, 2, ..., N). θ_ � r􏼐eT2 ξ T (x)􏼑 − 2kθ. (20)
Then, the output of the fuzzy system is defined as
The adaptive fuzzy inversion control process is shown in
Figure 6.
􏽑nj�1 uij 􏼐xj 􏼑 For the variation range of the input value of the fuzzy
y � 􏽐N
i�1 θi i T (16)
􏽐N n
i�1 􏽑j�1 uj 􏼐xj 􏼑 � ξ (x)θ.
system, it defines the fuzzy set: negative N, zero Z, and
positive P and takes the fuzzy membership function. The
In function design is shown in Figure 7.
8 Mobile Information Systems

1 1000

control input 1
0
membership function

-1000
0.5
-2000
0 2 4
Time (s)
1000

control input 2
0 0
-5 -4 -3 -2 -1 0 1 2 3 4 5
X -1000

Figure 7: Design of membership function of fuzzy system. -2000

0 2 4
Time (s)
Joint 1 position tracking

2 1000

control input 3
0
0
-1000

-2000
-2 0 2 4
0 2 4
Time (s)
Time (s)
Figure 9: Control input signal of three-joint robotic arm.
expect q1
actual q1
Joint 2 position tracking

2 compensate for the system’s impact from this repeated

disturbance. Although the three joints are better at tracking
the established trajectory, there are still errors. Figure 9
0 shows that for the established three-joint dynamic model,
joint 1 receives the largest moment during motion, followed
-2 by joint 2, and joint 3 receives the smallest. In addition, in
0 2 4 the presence of disturbances, different torques are required
Time (s) to control the joint movement and bring the tip to the
expect q2
desired position. Because a large torque is required to
actual q2 overcome the disturbance, the torque fluctuation is high at
the start of the movement. The moment of joint 2 and joint 3
Joint 3 position tracking

5 will, however, tend to be stable due to the repeatability of the

disturbance.
0
6. Conclusion
-5 The manipulator is an extremely complex system with non-
0 2 4
Time (s)
linearity, strong coupling, and numerous input and output
parameters. The trajectory tracking control of the 6-DoF
expect q3 manipulator must be studied step by step in order to be re-
actual q3 alized. The following is the overall research concept of this
Figure 8: Position tracking trajectory of the three-joint robotic paper. It first models the manipulator’s forward kinematics
arm. and then solves for its inverse kinematics. In order to achieve
high-precision control of the trajectory due to the presence of
disturbances in practice, it is necessary to investigate the
It adopts control law 10, adaptive law 22, and takes system’s trajectory tracking control algorithm. It plans the
λ1 � 30, λ2 � 50, k � 1.5, and c � 2. The desired trajectory of manipulator’s trajectory and simulates it using fifth-order
the three joints is yd � sin(2πt) . Its simulation result is shown polynomial interpolation and spatial linear interpolation. It
in Figure 8, and the control input signal is shown in Figure 9. obtains the changing process of each joint variable and in-
Figure 8 shows that when the system begins to move, tuitively demonstrates the trajectory planning process of the
there is a large difference between the actual and expected robotic arm. However, this paper only investigates the kine-
trajectory due to the presence of disturbance. However, over matics analysis, trajectory planning, and tracking control of
time, the fuzzy inversion control algorithm can gradually the manipulator and does not actually control the
Mobile Information Systems 9

manipulator. The robotic arm research is extensive, and there [11] H. Zhang, X. Tong, and H. Guo, “The parameter identification
are numerous aspects that can be explored further. of the autonomous underwater vehicle based on multi-in-
novation least squares identification algorithm,” International
Journal of Advanced Robotic Systems, vol. 17, no. 2,
Data Availability pp. 17298814–22092101, 2020.
[12] W. Shang, H. Ren, M. Zhu, T. Xu, and X. Wu, “Dual rotating
The data used to support the findings of this study are in- microsphere using robotic feedforward compensation control
cluded within the article. of cooperative flexible micropipettes,” IEEE Transactions on
Automation Science and Engineering, vol. 17, no. 4,
Conflicts of Interest pp. 2004–2013, 2020.
[13] X. Zhang, Y. Xu, H. Li, L. Zhu, X. Wang, and W. Li, “Flexible
The authors have no conflicts of interest. method for accurate calibration of large-scale vision me-
trology system based on virtual 3-D targets and laser tracker,”
International Journal of Advanced Robotic Systems, vol. 16,
Acknowledgments no. 6, Article ID 172988141989351, 2019.
[14] W. Cai, M. Gao, Y. Jiang et al., “Hierarchical domain ad-
This study was supported by (1) “Zhejiang University Visit aptation projective dictionary pair learning model for EEG
Scholar Project--Research and development of intelligent classification in IoMT systems,” IEEE Transactions on Com-
production line” Digital Twin technology (grant no.: putational Social Systems, vol. 1, pp. 1–9, 2022.
FX2018140) and (2) National Natural Science Foundation of [15] L. You, H. Jiang, J. Hu et al., “GPU-accelerated Faster Mean
China: A Study on the Design method of rotary Trans- Shift with euclidean distance metrics,” 2021, https://arxiv.org/
planting Mechanism facing the Requirements of trans- abs/2112.13891.
planting Track and Posture (grant no. 51375459). [16] Z. Huang, Y. Liu, C. Zhan, C. Lin, W. Cai, and Y. Chen, “A
novel group recommendation model with two-stage deep
learning,” IEEE Transactions on Systems, Man, and Cyber-
References netics: Systems, 2021.
[17] J. Kong, H. Wang, C. Yang, X. Jin, M. Zuo, and X. Zhang, “A
[1] Q. Huang and J. Lan, “Remote control of a robotic prosthesis spatial feature-enhanced attention neural network with high-
arm with 6-DoF for ultrasonic scanning and three-dimen- order pooling representation for application in pest and
sional imaging,” Biomedical Signal Processing and Control, disease recognition,” Agriculture, vol. 12, no. 4, p. 500, 2022.
vol. 54, no. Sep, pp. 101606.1–101606.7, 2019. [18] Y. Ding, X. Zhao, Z. Zhang, W. Cai, and N. Yang, “Graph
[2] G. Silveira, “On intensity-based nonmetric visual servoing,” sample and aggregate-attention network for hyperspectral
IEEE Transactions on Robotics, vol. 30, no. 4, pp. 1019–1026, image classification,” IEEE Geoscience and Remote Sensing
2014. Letters, vol. 19, pp. 1–5, Article ID 5504205, 2022.
[3] M. Donnici, G. Lupinacci, P. Nudo, M. Perrelli, and [19] C. Fan, H. Zhao, and D. Li, “Robotic grinding of a blisk with
G. Danieli, “Using navi-robot and a CT scanner to guide two degrees of freedom contact force control,” International
biopsy needles,” International Journal of Automation Tech- Journal of Advanced Manufacturing Technology, vol. 101,
nology, vol. 11, no. 3, pp. 450–458, 2017. no. 1-4, 2019.
[4] W. Zwonarz and A. Turnau, “w dynamically modified penalty [20] C. Ahmadizadeh, L. K. Merhi, B. Pousett, S. Sangha, and
function for control of mitsubishi robotic arm in real time,” C. Menon, “Toward intuitive prosthetic control: solving
IFAC-Papers Online, vol. 51, no. 32, pp. 418–423, 2018. common issues using force myography, surface electromy-
[5] G.P Smyrnaiou, M. Papoutsidakis, and A. Xatzopoulos, ography, and pattern recognition in a pilot case study,” IEEE
“Control of SIMO systems in simulation: the challenge of the Robotics and Automation Magazine, vol. 24, no. 4, pp. 102–
multiple axes actuating pneumatic arm,” International Jour- 111, 2017.
nal of Computer Application, vol. 168, no. No.10, pp. 0975–
8887, 2017.
[6] K. T. Song, C. H. Wu, and S. Y. Jiang, “CAD-based pose
estimation design for random bin picking using a RGB-D
camera,” Journal of Intelligent and Robotic Systems, vol. 87,
no. 3-4, pp. 455–470, 2017.
[7] Q. Song, S. Li, Q. Bai et al., “Trajectory planning of robot
manipulator based on RBF neural network,” Entropy, vol. 23,
no. 9, p. 1207, 2021.
[8] M. F. Shah, Z. Kausar, and M. M. Ahmad, “PID control and
stability analysis of an ith leg of a six degree of freedom
machining bed,” Procedia Manufacturing, vol. 17, pp. 927–
934, 2018.
[9] J. Chen, C. Du, Y. Zhang, P. Han, and W. Wei, “A clustering-
based coverage path planning method for autonomous het-
erogeneous UAVs,” IEEE Transactions on Intelligent Trans-
portation Systems, vol. 1, pp. 1–11, 2021.
[10] J. Chen, Y. Zhang, L. Wu, T. You, and X. Ning, “An adaptive
clustering-based algorithm for automatic path planning of
heterogeneous UAVs,” IEEE Transactions on Intelligent
Transportation Systems, vol. 1, pp. 1–12, 2021.