Content uploaded by Ali Yacoub
Author content
All content in this area was uploaded by Ali Yacoub on Feb 17, 2021
Content may be subject to copyright.
An Incremental Learning Approach for Physical
Human-Robot Collaboration
Achim Buerkle
Wolfson School of Engineering
Loughborough University
Loughborough, UK
a.buerkle@lboro.ac.uk
Ali Al-Yacoub
Wolfson School of Engineering
Loughborough University
Loughborough, UK
a.al-yacoub@lboro.ac.uk
Pedro Ferreira
Wolfson School of Engineering
Loughborough University
Loughborough, UK
p.ferreira@lboro.ac.uk
Abstract—Physical Human-Robot Collaboration requires
humans and robots to perform joint tasks in a shared
workspace. Since robot’s characteristic strengths are to cope
well with high payloads, they are utilized to assist human
operators during heavy pulling or pushing activities. A widely
used sensor to detect human muscle fatigue and thus, to trigger
an assistance request, is an Electromyography (EMG). Many
previous approaches to process EMG data are based on training
Machine Learning models offline or include a large degree of
manual fine tuning. However, due to recent advances in
Machine Learning such as incremental learning, there is an
opportunity to apply online learning which reduces
programming effort and also copes well with subject specific
characteristics of EMG signals. Initial results show promising
potential, yet, unveil a conflict between convergence time and
classification accuracy.
Keywords—EMG, Human-Robot Collaboration, Incremental
Learning, Machine Learning
I. INTRODUCTION
Human-Robot Collaboration (HRC) in manufacturing
aims to establish symbiotic or synergetic effects between
human operators and robots [1]. This is enabled by combining
the characteristic strengths of each party. Human strengths are
considered to be adaptability to changes, decision making, and
problem solving [1], [2]. Robot’s strengths, on the other hand,
are high precision, high operating speeds, and the capability
of coping with high payloads [3]. Thus, in a physical
collaboration, robots are able to support human operators via
force amplification to handle heavy pushing and pulling
activities [4]. In order to measure human muscle activity such
as during the lift of heavy objects, a widely used sensor is an
Electromyography (EMG) [5]. The approaches typically
include pre-processing of the data, feature extraction, and a
supervised leaning of the model [5], [6]. However, recent
advances in Machine Learning regarding incremental learning
could allow to minimize the training and programming effort
of such models [7]. Furthermore, the algorithm could optimize
its performance over time in an online system [7]. In this work,
an incremental learning approach is utilized to predict EMG
data during three different states: Participants lifting light
payloads, medium payloads, and heavy payloads (struggling).
II. RELATED WORK
In a Human-Robot collaborative scenario, humans and
robots perform joint tasks in a shared workspace [1]. In order
to communicate intentions of the human to the robot, sensors
are utilized such as EMGs [5]. The EMG signals are usually
acquired from a human upper-limb since they are mostly used
in the given tasks [5]. The acquired data can be used to
communicate movement intentions. It can also provide
insights on human muscle fatigue [6]. In this case, a robot
could assist a human operator during a heavy pull or push of
an object or adapt its behavior to create more ergonomic
working conditions for its human co-worker [6], [8]. This is
intended to prevent injuries, as well as long-term health issues
related to physical fatigue [8].
Figure 1 shows the general process used to integrate
EMGs in Human-Robot Collaboration for a supervised, non-
incremental learning approach. The first stage is EMG data
acquisition. Critical attention is required during the selection
of the acquisition device, the number of channels used, as well
as the placement of each channel [5]. The channel acquisition
device also determines the sampling rate and data
transmission [9].
During the pre-processing stage, raw EMG signals are
checked for baseline offset [5]. Typically, the signal is
corrected by subtracting the average amplitude from each
instance, however, there are also approaches based on
nonlinear error-modelling [5], [6]. Raw EMG signals are
susceptible to contain noise. Thus, Butterworth filters with a
cut off frequency from 2Hz-20Hz are utilized [5]. The
remaining features are extracted in the Feature selection and
extraction stage. This is critical during EMG data processing
since it has a high impact on the classification accuracy [9].
Three properties are considered as essential: class separability
(minimize overlap), robustness (separability in noisy
environment), and computational complexity (low complexity
of features implies lower processing times) [5], [9].
The fourth stage is continuous classification of the filtered
and extracted signals. There are mainly two types of
prediction models. One is the use of kinematic models, the
second approach is to utilize Artificial Neural Networks
(ANNs) [5]. However, [9] states that Linear Discriminant
Analyses (LDA) and Support Vector Machines (SVMs) are
also widely used for EMG data classification. According to
[5] there are few critical challenges remaining. Firstly, many
Figure 1 EMG Signal Processing for HRC (adapted) [5]
3rd UK-RAS Conference for PhD Students & Early Career Researchers, Hosted virtually by University of Lincoln, April 2020
9
https://doi.org/10.31256/It7Lm1B
offline systems obtain high classification accuracies, yet the
online performances of such systems are far from satisfactory.
Secondly, there are subject-specific characteristics of EMG
signals. This can even include variation of the EMG signals
for the same person during different recording sessions. An
opportunity to increase the performance and to lower
programming and fine-tuning effort could be incremental
learning. Incremental learning algorithms have the following
characteristics: ability of life-long learning, ability to
incrementally tune the model’s performance, and no prior
knowledge about the data and its properties is needed [7].
III. EXPERIMENTAL SETUP
The experimental setup aims to collect EMG data during
three different stages: light payload, medium payload, and
high payload, during which a participant is slightly struggling.
The acquired data will be fed into the classifier unlabeled.
However, in order to validate the prediction results, predicted
classes and the actual classification will be compared.
IV. RESULTS AND DISCUSSION
The collected data was used to train an Online Random
Forest (ORF) model, that aims to classify the EMG signals
into low payload, medium payload, and heavy payload. In any
incremental learning approach, the most crucial property apart
from accuracy is the convergence time. Since the model aims
to minimize the prediction error live and immediately. In
Human-Robot Collaboration, this is exceptionally important
as humans and robots are physically interacting. Hence, in this
validation experiment convergence time of the ORF model
was measured with a different number of trees, which is
illustrated in Figure 3.
As, expected, Figure 3 shows that the convergence time is
directly proportional to the number of trees in the ORF model.
The correspondent accuracy of the models in Figure 3 is
shown in Table 1. The collected data for this experiment is
~6000 data points of EMG signals and the associated labels.
Based on Figure 3 and Table 1, it can be noticed that the model
must achieve a trade-off between accuracy and convergence
time. The ORF model with 20 trees seems to be the most
suitable model since it can converge in less than 2 seconds,
and it achieves the highest detection accuracy.
Table 1 Number of Trees vs Prediction Accuracy
V. CONCLUSION AND FUTURE WORK
A novel incremental learning approach was introduced to
determine physical workload from EMG data in Human-
Robot Collaboration. During the online training, a conflict
became clear between processing speed and accuracy. Lesser
trees in the model meant faster convergence, however, it also
resulted in the aforementioned lower accuracy. Overall, the
accuracy could reach 89% in only two seconds. Thus, in a
Human-Robot Collaborative Scenario this would allow the
system to recognize a human operator struggling with the
payload. The collaborative robot could then support the
operator and subsequently, create a more ergonomic
environment. However, prior to this technology being ready
to be used in a practical application, further testing is essential.
This includes the need for a larger sample size in participants
and a richer variety in lifting tasks. The current setup allows
to detect muscle contraction in participant’s forearms and
biceps. Yet, more EMG sensors placed on other muscle
groups such as triceps and shoulders are expected to provide
better results for predicting pushing activities. Furthermore,
the system could be trained to not only detect temporary high
payloads but also to recognize muscular fatigue during
endurance tasks. This could help to improve human operator’s
posture and subsequently prevent negative long-term health
effects.
Nevertheless, early results of this incremental learning
approach demonstrate a reduced manual fine-tuning effort and
it coping well with subject specific characteristics in the data.
This offers the potential to be applied for additional human
sensor technologies and subsequent data classifications.
Ultimately, this could help to make Human-Robot
Collaboration safer and more efficient.
Number of
Trees
5
10
15
20
25
30
Mean
Square Error
0.26
0.19
0.250
0.14
0.15
0.17
Accuracy
[%]
82.3
84.6
85.2
89.7
86.7
86.7
Figure 2 Experimental Setup
Figure 3 ORF convergence time
3rd UK-RAS Conference for PhD Students & Early Career Researchers, Hosted virtually by University of Lincoln, April 2020
10
VI. REFERENCES
[1] L. Wang et al., “Symbiotic human-robot
collaborative assembly,” CIRP Ann., 2019.
[2] V. Villani, F. Pini, F. Leali, and C. Secchi, “Survey
on human-robot collaboration in industrial settings:
Safety, intuitive interfaces and applications,”
Mechatronics, no. June 2017, pp. 1–19, 2018.
[3] J. Krüger, T. K. Lien, and A. Verl, “Cooperation of
human and machines in assembly lines,” CIRP Ann.
- Manuf. Technol., vol. 58, no. 2, pp. 628–646, 2009.
[4] J. Schmidtler and K. Bengler, “Fast or Accurate? –
Performance Measurements for Physical Human-
robot Collaborations,” Procedia Manuf., vol. 3, no.
Ahfe, pp. 1387–1394, 2015.
[5] L. Bi, A. Feleke, and C. Guan, “A review on EMG-
based motor intention prediction of continuous
human upper limb motion for human-robot
collaboration,” Biomed. Signal Process. Control, vol.
51, pp. 113–127, 2019.
[6] L. Peternel, C. Fang, N. Tsagarakis, and A. Ajoudani,
“A selective muscle fatigue management approach to
ergonomic human-robot co-manipulation,” Robot.
Comput. Integr. Manuf., vol. 58, no. January, pp. 69–
79, 2019.
[7] A. Bouchachia, B. Gabrys, and Z. Sahel, “Overview
of some incremental learning algorithms,” IEEE Int.
Conf. Fuzzy Syst., no. August, 2007.
[8] L. Peternel, N. Tsagarakis, D. Caldwell, and A.
Ajoudani, “Robot adaptation to human physical
fatigue in human–robot co-manipulation,” Auton.
Robots, vol. 42, no. 5, pp. 1011–1021, 2018.
[9] M. Hakonen, H. Piitulainen, and A. Visala, “Current
state of digital signal processing in myoelectric
interfaces and related applications,” Biomed. Signal
Process. Control, vol. 18, pp. 334–359, 2015.
3rd UK-RAS Conference for PhD Students & Early Career Researchers, Hosted virtually by University of Lincoln, April 2020
11