Journal of Physics: Conference Series

PAPER • OPEN ACCESS You may also like

- Spatio-temporal warping for myoelectric
Research on Robotic Arm Movement Grasping control: an offline, feasibility study
Milad Jabbari, Rami Khushaba and
System Based on MYO Kianoush Nazarpour

- The effect of high pressures on the

yoghurt from milk with the stabilizer
To cite this article: Ye Wei and Danping Jia 2021 J. Phys.: Conf. Ser. 1754 012173 A Reps, A Jankowska and K Winiewska

- Multi-cluster problems: resonances,

scattering and condensed states
K Kat, T Myo, Y Kikuchi et al.
View the article online for updates and enhancements.

This content was downloaded from IP address 125.209.68.42 on 09/01/2024 at 15:34

ISPECE 2020 IOP Publishing
Journal of Physics: Conference Series 1754 (2021) 012173 doi:10.1088/1742-6596/1754/1/012173

Research on Robotic Arm Movement Grasping System Based

on MYO

Ye Wei1, Danping Jia1*

1
School of Information Science and Engineering, Shenyang University of Technology,
Shenyang, Liaoning, 110870, China
*Corresponding author’s e-mail: [email protected]

Abstract. This paper studies the continuous motion capture system of a robotic arm based on
MYO. The MYO sensor can collect the electromyographic signal (sEMG) and the motion signal
(IMU) of the arm during exercise, and filter the electromyographic signal, extract the features,
and return the features. After the first transformation, the machine learning algorithm is used for
model training. In order to verify the effect of the recognition of the EMG signal on the deep
learning method, the accuracy comparison experiment with the BP neural network is also added.
In addition, a two-link arm model was built, and the pose information of the end of the arm was
obtained through Euler angle transformation, and then the end pose information of the robotic
arm was obtained through coordinate mapping, and the end pose simulation was performed using
Matlab software. The simulation results prove that the performance of the system meets
expectations, and points out the direction for subsequent research.

1.Introduction
The traditional way of human-computer interaction based on program control fetters the development
of robot intelligence and is difficult to apply to robot systems that need to be directly integrated with the
human body, such as smart prostheses, medical rehabilitation robots and wearable behavior assistance
robots. Modern human-computer interaction technology needs to develop from a way that robots
passively accept instructions to a way that robots actively understand human behavior intentions. From
this, a new type of intelligent human-computer interaction technology that recognizes human behavior
intentions through bioelectric signals is derived. Among them, surface electromyogram py (sEMG) is
favored by many researchers because of its rich information and mature acquisition technology [1].
Most of the previous researches on identifying human behavior intentions through sEMG focused on
classifying human actions. The basic process includes the feature extraction of surface EMG and the
construction of classification models using feature samples. The sEMG-based model can recognize a
variety of gestures and can achieve high accuracy. These can be applied to remote grasping and
interactive robots.
Estimating continuous human joint motion through sEMG is the key point of current theoretical and
applied research. There are two ways to realize continuous joint motion estimation based on sEMG: (1)
Establish a joint dynamics model with sEMG as input combined with muscle physiology, and then
calculate joint torque/angular acceleration, etc. However, the constructed model has a complex structure
and contains many physiological parameters that cannot be directly measured, which is difficult to apply
in practice. (2) Directly establish a regression model that correlates sEMG and joint movement. Among
them, neural network is the most common application, but the structure of neural network has a great
influence on the prediction results, but there is no standardized method to define a suitable network
Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd 1
ISPECE 2020 IOP Publishing
Journal of Physics: Conference Series 1754 (2021) 012173 doi:10.1088/1742-6596/1754/1/012173

structure, and the determination of network parameters is excessively dependent for training data, when
there is a difference between the test data and the training data, it is easy to cause the network output to
deviate from the true value.
By collecting the motion IMU signal of the upper and lower bones of the arm, through the
acceleration and Euler angle information, the motion information of the arm can be obtained. Secondly,
a coordinate system is established for the human arm, and a two-link model is created and the posture
information of the arm is mapped to the end of the manipulator through the principle of robot kinematics,
so that the end of the manipulator and the hand keep the same movement.
Therefore, a method is proposed to recognize gestures and initialize the manipulator through sEMG,
and control the movement of the manipulator through IMU signals. This method requires: (1) Select a
highly accurate and reliable gesture recognition algorithm to predict the grasping posture;(2) Processing
the IMU signal to enable it to accurately express the arm motion information, but the accuracy of the
traditional differentiation based on the acceleration information in the IMU is often related to the
accuracy of the IMU equipment. The greater the accuracy, the more expensive the equipment, the higher
the cost. It is difficult to bear the cost in production
This topic selects MYO as the acquisition equipment of sEMG and IMU signals. The sampling
frequency of sEMG is 200hz and the sampling frequency of IMU signal is 50hz. Gestures are recognized
through sEMG signals, and IMU controls the motion of the robotic arm. An experimental platform was
built to verify the feasibility and effectiveness of the system

Figure 1. MYO EMG acquisition equipment

2. Method and Process

In this paper, operations such as filtering, feature extraction, and feature normalization were performed
on the sEMG signal, and the recognition accuracy of several commonly used machine learning methods
were compared. The flow chart is as follows [2]:
Offline training model

sEMG training Feature Feature Classification

Pretreatment
data extraction normalization model

Online real‐time control

sEMG real‐time Feature Feature Model

Pretreatment Control decision
data extraction normalization discrimination

Figure 2. Myoelectric hand control system based on traditional methods

2.1 EMG signal processing method

Because sEMG is susceptible to interference, the data collected through MYO cannot be directly sent
to the classification model for use. The IIR digital filter is selected to filter out 50Hz power frequency
interference. The system function of the IIR digital filter [3] is show in (1):

2
ISPECE 2020 IOP Publishing
Journal of Physics: Conference Series 1754 (2021) 012173 doi:10.1088/1742-6596/1754/1/012173

H(z)= (1)
𝜔 2𝜋𝑓 /𝑓 (2) is filter digital frequency, 𝑓 is filter frequency, 𝑓 is sampling frequency
The surface EMG signal needs to extract its hidden deep features before it can be used to recognize
the action intention of the human body. In traditional methods [4], the features of the sEMG signal are
generally extracted manually, including time domain features, frequency domain features, and time-
frequency domain features. Generally, the EMG features are extracted through the combination of time
window and increment window, as shown in the figure:

Figure 3. Time window and incremental window for extracting sEMG feature
Define N as the number of sampling points in the time window, xi represents the value of the i-th
sample point. Select Mean absolute value (MAV) and Root mean square (RMS) as features, the
expression as (2), is the most commonly used feature and can reflect the average
𝑀𝐴𝑉 ∑ |𝑥 | (2)

𝑅𝑀𝑆 ∑ 𝑥 (3)
level of the amplitude of the EMG signal. It is often used to detect the state of action and rest. The
expression as RMS (3), is related to muscle contraction strength.
Feature normalization is to ensure that these multi-dimensional features have similar scales and
eliminate the dimensional influence between data features. This will help the gradient descent algorithm
run faster and require fewer iterations for model convergence. More importantly, normalizing features
can improve the accuracy of action classification. Linear function normalization (Min-Max Scaling) will
linearly transform the original data to map the result to the range of [0,1] to achieve proportional scaling
of the original data. Its mathematical expression (4)
𝑥 (4)
𝑥 represents the characteristics after standardization, 𝑥 , 𝑥 represent the minimum and
maximum values of the data respectively
The data source is the EMG data of five adult males, and the EMG data of the four gestures of making
a fist, rest, ok the five fingers, and bringing the five fingers together. On the classification model, Linear
Discriminant Analysis(LDA), Random forest(RF), Decision tree(DT), Support Vector Machines(SVM),
k-nearest neighbor classification(KNN) are selected, and the BP neural network [5] is specially
processed for the purpose of comparing the applicability of the system between machine learning and
neural network.

Figure 4. Four gestures for recognition

The RH56DF3 manipulator of Beijing YinShi Technology is selected as the grasping device. The
device has strong grasping force and moderate speed, which is suitable for grasping operations of robots
or prosthetic hands of amputees. The supported communication protocols are RS232, RS485 and CAN

3
ISPECE 2020 IOP Publishing
Journal of Physics: Conference Series 1754 (2021) 012173 doi:10.1088/1742-6596/1754/1/012173

bus communication, as shown in the figure

Figure 5. RH56DF3 mechanical prosthetic hand

2.2 Euler angle conversion and coordinate mapping

On the other hand, the IMU signal contains three-axis acceleration information and three-axis gyroscope
information [6]. In theory, the displacement information of the object can be obtained by quadratic
integration of the acceleration. However, due to the accuracy and interference of the equipment, the
error is found to be unacceptable in the preliminary processing, so the quaternion information of the
gyroscope in the IMU will be used in the subsequent conversion to Europe Pull angle, get the posture
information of the end of the arm. The conversion relationship as (5)

∅ ⎡ 𝑎𝑟𝑐𝑡𝑎𝑛 ⎤
⎢ 𝑞 𝑞 ⎥⎥
𝜃 ⎢arcsin 2 𝑞 𝑞 (5)
𝜑 ⎢ 𝑎𝑟𝑐𝑡𝑎𝑛 ⎥
⎣ ⎦
The results of arctan and arcsin are in the range of ， , but is not satisfy all orientations
(the value range of angle θ is already satisfied ， , so need to use atan2 instead of arctan
(6).
∅ 𝑎𝑡𝑎𝑛2 2 𝑞 𝑞 𝑞 𝑞 ,1 2 𝑞 𝑞
𝜃 asin 2 𝑞 𝑞 𝑞 𝑞 (6)
𝜑 𝑎𝑡𝑎𝑛2 2 𝑞 𝑞 𝑞 𝑞 ,1 2 𝑞 𝑞
∅, 𝜃, 𝜑 are the rotation angles around the z axis, y axis, and x axis respectively.
By establishing a two-link model of the arm, the calibration coordinate system maps the pose of the
end of the arm to the end of the manipulator to achieve the purpose of following the movement of the
hand. The coordinate system calibration method is as follows: the x-axis is parallel to the upper and
lower arm bones, the y-axis is perpendicular to the upper and lower arm bones, and the z-axis is the
vertical direction of the arm.

3 Experiment and analysis

In the previous chapter, the composition of the system and the workflow of each part were introduced,
especially the gesture recognition part was described in detail, and the data conversion of IMU was also
introduced. The experimental process and experimental results are analyzed and explained below.

3.1 Algorithm accuracy comparison

After comparing the accuracy of all gestures (Figure 6), it is found that the average accuracy of the BP
neural network and the SVM algorithm is the highest under the evaluation. Because the BP neural
network can read the characteristics of the data in multiple dimensions, the experiment found that when
the hidden layer nodes exceed 15 At times, the BP neural network will overfat, and the algorithm time-
consuming is significantly higher than other algorithms.

4
ISPECE 2020 IOP Publishing
Journal of Physics: Conference Series 1754 (2021) 012173 doi:10.1088/1742-6596/1754/1/012173

Figure 6. Comparison of different algorithm models

The following (Table 1) is a comparison of the recognition accuracy of these six algorithms for the
four gestures. It can be found that the accuracy of the OK gesture is lower in all algorithms. It may be
because the OK gesture is small and the amplitude of the EMG signal is too low, so it is easy to be Noise
is submerged; the accuracy of fist and rest are relatively high, because of the large gestures, and the
experimental data conforms to the signal law.
Table1. Accuracy comparison of different gestures
LDA DT BP SVM KNN RF
fist 85.0695 78.9985 95.9888 93.3399 84.4879 74.6666
rest 83.2965 62.2333 94.0931 92.9254 82.3654 72.6598
ok 77.7541 34.4338 89.0058 93.0000 73.5896 65.9999
five 73.1458 67.2536 93.7145 90.0001 77.6314 69.7763
Then use MYO to collect the EMG signal data of the arm, use the SVM algorithm model for gesture
recognition, and discretely control the movement of the manipulator according to the recognized results.
The final two-finger(a) and five-finger(b) grasping effects are as follows:

(a) ok (b) five

Figure 7. Fetch results

3.2 Robotic arm motion trajectory

Most of the current research on the movement method of the robot arm following the arm uses a visual
calibration method, that is, an optical sensor (in most cases a camera) is used to analyze the image of
the arm movement, and the arm's movement trajectory is calculated according to the pre-calibrated
coordinate system Finally, the trajectory points are mapped to the robotic arm. This method has high
accuracy. However, because it is a third-party perspective to observe the action and ignores the actor's
own feelings, the movement of the robotic arm is inconsistent with the movement of the human body.
The problem is that it requires multiple familiarizations to operate normally, and the use of optical
sensors is limited, which limits its usage scenarios.

5
ISPECE 2020 IOP Publishing
Journal of Physics: Conference Series 1754 (2021) 012173 doi:10.1088/1742-6596/1754/1/012173

In this experiment, two MYO bracelets are worn on the upper and lower arms of the hand respectively,
and the action is to rotate around the shoulder joint for ten seconds, and the obtained data is simulated
in Matlab (Figure 8). It can be seen that the final trajectory is roughly the same as the actual trajectory.
This control method gives full play to the user's own active trajectory correction method, similar to the
control of a mouse, the user can correct errors in real time, achieving the intention of controlling the end
of the robotic arm to reach the target point.

Figure 8. Simulation of robot arm end trajectory

4. Conclusions
This paper studies and compares several machine learning algorithms, and analyzes the pros and cons
of using EMG to recognize gestures, and finally chooses SVM as the recognition algorithm. It is further
proposed to use Euler angles in the IMU signal to obtain the pose information of the arm, and to achieve
the effect of controlling the position of the end of the manipulator following the hand by mapping it to
the end of the manipulator. The system is not restricted by the venue, and the sensor only needs a MYO
bracelet to control the robotic arm and grasp the effect. Compared with the optical calibration system,
this system is more practical. Follow-up can be further researched on control accuracy and angle.

Acknowledgments
This work is supported by Liaoning Provincial Education Department Scientific Research Fund Project
(LFGD2019001). The funders project was built for the experimental platform, and data collection,
decision to publish, or preparation of the manuscript played a huge role.

References
[1] Miguel A. Perez, Maury A. Nussbaum.(2003) Principal components analysis as an evaluation and
classification tool for lower torso sEMG data[J]. Journal of Biomechanics,36(8).
[2] Moniri Ahmad, Terracina Dan, Rodriguez-Manzano Jesus, Strutton Paul H, Georgiou Pantelis.
(2020) Real-Time Forecasting of sEMG Features for Trunk Muscle Fatigue using Machine
Learning.[J]. IEEE transactions on bio-medical engineering, PP.
[3] Hiren K. Mewada, Jitendra Chaudhari.(2019) Low computation digital down converter using
polyphase IIR filter[J]. Circuit World,45(3).
[4] Shang Li, Hongli Zhang, Lin Ye,Shen Su, Xiaoding Guo, Haining Yu,13,Binxing Fang.(2020)
Prison Term Prediction on Criminal Case Description with DeepLearning[J]. CMC:
Computers, Materials & Continua,62(3).
[5] Jing Yang, Yingpeng Hu, Kaixi Zhang, Yanghui Wu.(2020) An improved evolution algorithm using
population competition genetic algorithm and self-correction BP neural network based on
fitness landscape[J]. Soft Computing, (prepublish).
[6] Robotics;(2020) New Findings from Beijing University of Technology in the Area of Robotics
Described (Calibration of Low Cost Imu's Inertial Sensors for Improved Attitude
Estimation)[J]. Journal of Robotics & Machine Learning.