Low-cost wearable Multichannel Surface EMG
Acquisition for Prosthetic Hand Control
Davide Brunelli and Andualem Maereg Tadesse
Department of Industrial Engineering
University of Trento
Trento, Italy
{name.surname}@unitn.it
Abstract—Prosthetic hand control based on the acquisition
and processing of surface electromyography signals (sEMG) is a
well-established method that makes use of the electric potentials
evoked by the physiological contraction processes of one or more
muscles. Furthermore intelligent mobile medical devices are on
the brink of introducing safe and highly sophisticated systems to
help a broad patient community to regain a considerable amount
of life quality. The major challenges which are inherent in such
integrated system’s design are mainly to be found in obtaining a
compact system with a long mobile autonomy, capable of
delivering the required signal requirements for EMG based
prosthetic control with up to 32 simultaneous acquisition
channels and – with an eye on a possible future exploitation as a
medical device – a proper perspective on a low priced system.
Therefore, according to these requirements we present a wireless,
mobile platform for acquisition and communication of sEMG
signals embedded into a complete mobile control system
structure. This environment further includes a portable device
such as a laptop providing the necessary computational power
for the control and a commercially available robotic handprosthesis. Means of communication among those devices are
based on the Bluetooth standard. We show, that the developed
low cost mobile device can be used for proper prosthesis control
and that the device can rely on a continuous operation for the
usual daily life usage of a patient.
Keywords— EMG, wireless communication, Bluetooth.
I.
INTRODUCTION
In recent years, prosthetic technology for lower arm limb or
hand amputated patients has revealed impressive advances by
evolving from simple and fixed anthropomorphic extensions to
real-time actuated electronic prosthetics. The preferred signals
used to control hand gesture are derived from non-invasive, so
called surface electromyography (sEMG) sensors, providing
broad signal information about the corresponding muscle
activity [1].
These sensors incorporate capturing and amplification of
evoked muscle action potentials by one or more related
muscles in the patient’s remaining stump. The signals are
processed by detection algorithms to control a prosthetic device
in accordance with the patient’s intended movements such as
grasping, wrist rotation or flexions. This approach tries to re-
978-1-4799-8981-2/15/$31.00 ©2015 IEEE
Bernhard Vodermayer, Markus Nowak and
Claudio Castellini
German Aerospace Research Center - DLR
Germany
{name.surname}@dlr.de
enable the natural way of controlling the lost hand by
contracting the remaining corresponding muscles, thus offering
a more intuitive approach in contrary to the EMG derivation of
other muscle groups, e.g. in the pectoral region where patients
have to train for intensive and longer sessions to regain their
formerly lived quality of life [2],[3].
The promising opportunities from this technology are
widely recognized by the scientific community which has
started to investigate sEMG also for control commands in
many other applications such as gesture-based interfaces for
virtual reality games, virtual keyboards, and human computer
interface etc. [4],[5].
Depending on the number of sEMG sensors and the types
or groups of muscles used to feed the control algorithms, it is
possible to perform different classes of hand gestures: from
coarse hand grasping, wrist and hand movements, to individual
finger flexions.
To achieve smooth and articulated prosthesis movements a
higher number of sEMG sensors along with multi-channel
signal acquisition is beneficial [6].
In this paper we address the possibility to acquire sEMG
signals up to 32 channels simultaneously, providing adequate
bandwidth and signal resolution for complex gesture
recognition and prosthetic control. A wireless, mobile low-cost
system will be introduced, that makes use of commercially
available, medical device grade components of the shelf COTS,
like the Otto-Bock 13E200 electrodes for sEMG signal reading
or Touch Bionics’ i-LIMB Ultra Revolution robotic hand
prosthesis. The interconnection of these components will be
presented to be completely wireless and providing an
independent autonomy that renders the system useable for
patients’ daily life actions helping them to regain their formerly
lived quality of life. Corresponding system performance will be
discussed.
The paper is organized as follows: in Section II we present
the state of the art of wireless prosthetic hand controlled by
sEMG signals. The modeling and the design of the wireless
acquisition system is presented in Section III with emphasis on
the design choices for power optimization. Experimental
94
measurements and discussion about performance and energy
autonomy is discussed in Section IV. Finally Section V
concludes the paper.
II.
RELATED WORK
The concept of mobile medical devices including intelligent
prostheses has by now a long term history. But still big
challenges arise when physiologic processes have to be
transformed into actions of the medical device.
With respect to multi-channel acquisition of signals evoked
by the human body, Wang [7] addresses the design of a multichannel data acquisition system with 32 channels for EEG
signal based application. In this field, aggressive power
management techniques have been proposed in [8] by applying
approximate computing paradigm to the monitoring and ECG
signals.
Of course the signal requirements for EEG signals differ
from those in sEMG applications, but in this case the number
of signals and bandwidth are in the same range. Intensive
electronic filtering is utilized by a complex analog circuitry and
an on board DSP. The system is capable of acquiring up to 24
unipolar and eight bipolar electrodes. The usable signal bandwidth is 100Hz realized by complex filtering. The system does
not provide any mobility in the current state as a wired
connection to a PC, responsible for the processing of the
acquired signals.
Another application in the field by Vaca Benitez et al. [9]
presents an EMG-controlled arm orthosis in an early prototype
stage. The control system is fed with signals acquired from a
commercially available sEMG-signal recorder Brain Amp ExG
MR by the German Company Brain Products GmbH, Gilching
[10]. This type of the device is available in two versions with 8
or 16 channels and is especially prepared for the use within CT
or MRT scanners. Analog-to-digital conversion is carried out
with a resolution of 16bit with a maximum of 5 kHz per
channel. Fiber optic communication is used due to
environmental requirements, thus the device’s portability is
restricted to the cabling. The power consumption is 150mA in
active mode, yet this subsystem itself has a weight of 1.1kg.
One PowerPack for the use together with the system with a
capacity of 6500mAh, providing approximately 40h of
independent operation, has a weight of 130g [11]. Vaca
Benitez et al. further describe, that the usable signal is obtained
by filtering down to a bandwidth of 20Hz and transferring it to
a PC for further processing. Control of the orthosis will be part
of upcoming scientific work.
Lee et al. [12] also comprise a portable sEMG recording
system that can be mounted into a wireless environment using
Bluetooth connectivity. The system provides 10 parallel
channels including the complete signal chain with sensor, preamplification and initial processing, second amplification,
conversion and finally the distribution via the Bluetooth
interface to a processing station. Lee et al. report above that to
have reached a wireless maximum data rate of 723kbps while
gaining a sample rate of 2000Hz for ten simultaneous channels
acquired. The power consumption of the Bluetooth device is
described to be 25mA. However, the total power consumption
and autonomy depend also on the acquisition part.
A multichannel sEMG acquisition system comprising a
matrix based approach for the EMG-sensor array, where the
sensors are placed on two flexible PCBs within eight by four
grid each and 1 x 1 cm grid spacing is presented by Xiong [13].
As a sample rate of 200Hz is provided for the 64 channels, the
WiFi-standard is used including an access point a Laptop for
data processing to match the high demand to the data rate. It is
furthermore described that the system costs are estimated to be
at around 1000 Euro for a 32 channel version. It has to be
noted, that this setup requires additional reference electrodes to
be placed somewhere on the patient’s corresponding
extremities which has to be taken into account when dealing
with amputees.
In comparison to the state of the art, the architecture
proposed in this paper provides a good tradeoff between the
number of acquisition channels and the cost. Moreover the
energy autonomy is optimized considering the size and the
constraints need for a wearable acquisition board. The system
is also suitable to execute aggressive power management for
achieving a deep idle power and yet a substantial energy
savings during long monitoring by using compressive sampling
methods as described in [14].
III.
SYSTEM ARCHITECTURE
We started from a testbed where Analog EMG signals are
sampled and converted to digital data to ease the signal
processing, using DAQ cards with on-board AD-converter
chips, as illustrated in Fig. 1. The digital data is then sent to a
computer or an integrated microprocessor for processing of the
signal. This processing stage includes feature extraction and
classification. At this point the sampled data is classified to
certain action labels. Using these labels the signals for positionand velocity-control are computed and sent to the motors
drivers to produce the actual prosthetic hand motion based on
the control signals. This control method is capable of
recognizing movement of each fingers of the hand, making the
hand control easy so that the subject can move the hand as if
the patient would do it with a normal hand.
Fig. 1. Schematic block of the preliminary wired acquisition system
95
The contribution presented in this paper is to move the
acquisition and pre-processing part directly in a board the
patient can wear comfortably in the arm close to the set of the
sEMG probes. Data communication with suitable bandwidth is
provided by Bluetooth wireless protocol and low-power radios
as depicted in Fig. 2.
The capability of modern mobile devices can easily
substitute a standard desktop workstation used so far for
experimental studies. In fact, microprocessors available in
those mobile devices have more than just the adequate
computational capacity to perform feature extraction,
classification and prosthetic hand control. This allows the
design of completely novel wearable systems, capable of
controlling a prosthetic hand while performing multiple tasks
directly commanded by the patient’s nervous system in an
efficient and cost effective way.
prosthesis node. After parsing preprocessing the data into a
convenient form for the machine learning program (used for
pattern classification), the mobile platform can interact with the
prosthetic hand or directly display the analog signal by
reconstruction from the digital data. The prosthetic interface
handles the received communication data from the computer,
i.e. commands generated by the mobile device will be sent to
the prosthetic hand using this interface. With respect to the
consecutive design considerations for this system, the future
aim is a condensed, integrated circuitry providing all necessary
components. To achieve this, a detailed analysis of
requirements for each one of the components and the
possibility of merging the components in future, leading to a
considerable reduction of component size was conducted. Of
course, a further remarkable aim was to use non expensive
components of the shelf (COTS) as far as possible.
Fig. 2. Functional scheme of the wireless Prosthetic Hand Control
with Surface EMG signals
Otto bock 13E200 surface EMG electrodes are used to pick
up neuromuscular activity signals. Pattern recognition and
machine learning techniques are used to predict the
corresponding fingers’ movements.
As depicted in Fig. 3, the overall system design consists of
three interleaving subsystems: Sensor (EMG-Electrode)
interface with pre filtering and acquisition, a mobile processing
device (e.g. a small laptop) and the control interface to the
Prosthetic hand. The primary purposes of the sensor interface,
placed on the patient’s remaining forearm stump, is the
acquisition of the EMG signals from the incorporated
electrodes, the framing of data-packets and the transmission to
the mobile device with the help of a Bluetooth connection. Of
course, the acquisition system is part of a more complex
communication protocol thus it provides means of connection
establishment, security and data integrity. Mobile computation
capabilities – i.e. the mobile device - are made available by a
small Intel Core i5 520M (mobile version) based notebook
including at least two Bluetooth interfaces. It interprets and
classifies the incoming data-streams from the sEMG probes
and transmits control commands to the robotic hand prosthesis.
Thus it serves as a central Bluetooth based communication
node in between both - the sensor-acquisition node and the
Fig. 3. Schematic block of the wireless communication system
based on Bluetooth radio.
In more detail, the designed wireless EMG sensor
acquisition system depicted in Fig. 4 incorporates the following
features: 32 EMG multiplexed sensor inputs with passive
single stage pre-filtering (low pass), a 16 Bit low cost
microcontroller (MSP430 derivate with 25MHz), an onboard
Bluetooth transceiver, an optional USB port for direct PCconnection and debugging, an inertial measurement unit
(IMU), several indicator LEDs, and a rechargeable battery. In
the current version, the electronics are split, i.e. the controller
part including power management and communication can be
separated from the analog part. This offers scalability to 64 and
more EMG sensors, furthermore, the layout on the analog part
can be optimized for best analog signal integrity.
The Bluetooth 2.1 compatible transceiver, a Roving
Networks RN-41, provides a low power, bi-directional RF
communication link with a 240kbps data-rate. It is connected to
the controller by UART. Also the controller was selected
according to its low power capabilities, using less than 10mA
when operation in full performance mode. The final cost of the
wearable multichannel sEMG was required to be below 150
96
Euros, excluding the costs for the EMG probes. The prototype
developed and tested is illustrated in Fig. 4.
As the power consumption for the analog part is far less
than the one of the controller, a lightweight 45g, 2000mAh
LiPo rechargeable battery with the form factor of the board,
can keep the acquisition system alive for at least two days
running in full power mode. An optimization of the energy
management has not been done yet but could increase the
mobile capabilities by far.
Fig. 4. Prototype of the Low cost mobile Multichannel Surface
EMG Acquisition for Prosthetic Hand Control.
A. Otto-Bock sEMG electrodes:
The Myobock 13E200 EMG electrodes from Ottobock are
used to acquire the sEMG signal for the control of the
prosthetic hand. These electrodes are firmly placed on the skin
to gather the information about the contraction of the muscles
underneath. The Myobock electrodes are selected because of
some special features than the other electrodes, which include:
linear and proportional signal output, highly sensitive
amplifiers, high common mode rejection ratio (CMRR) for low
frequency which provides a good rejection of electromagnetic
interference and high input impedance which can reduce the
effect due to the variability of the skin impedance. The
Myobock electrodes have frequency bandwidth of around
30Hz, and can operate with power source of 5V to give an
output voltage proportional to the level of muscle activation.
The electrodes have a 3 pin connector receptacle (Analog
output, Power input and ground).
The signal resolution, accuracy, signal range and sampling
rate define the quality of the EMG signal. The input
impedance of the analog front end system must be as high as
possible since the skin-electrode contact impedance is variable
from a few KΩ to several MΩ. The quality of the EMG signal
is also largely dependent on the resolution, accuracy and
sampling rate of the ADC. Present day ADCs used in the
EMG systems use 10 to 24bit resolution, however the
resolution of ADCs incorporated in most of today’s low cost
microcontrollers is limited to 12 or 16 bits. Considering the
number of the acquired channels and the bandwidth of the
sensors, at least ten times oversampling will be considered to
improve the output resolution of the data stream. This also
enables reduction of aliasing and phase distortion.
B. Acquisition and filter unit
The used microcontroller MSP430F5529 by Texas
Instruments, a 16-bit MCU with built in modules such as ADC,
UART and DMA greatly showed the possible reduction of the
consumed energy used for conversion and data-transfer. The 12
bit ADC, i.e. one of the core components in the data acquisition
system eliminates the need for external ADC chips. The ADC
and the UART are configured to work in parallel, thanks to the
MCU’s DMA capability, which frees the microcontroller’s
performance for interrupt handling and auxiliary computation,
such as digital inline filtering. Combined with the external,
passive low pass filters, a precise acquisition process is
possible. Due to the separation of the analog part from the
computational and communication depended part, a future
extensibility to at least 64 input channels is possible with minor
configurations. At the moment, eight out of 14 channels are
consumed. Further 32 input channels can be added by stacking
a second multiplexer module on top of the main board.
C. Bluetooth communication.
The Bluetooth communication is realized by a RN-41
Bluetooth Module from Roving Networks. It is available with a
small form factor of approximately 3 cm² including the
antenna. It supports the Serial Port Profile (SPP) up to 240
kbps, allowing its host device to appear as a virtual COM port
on standard mobile devices. This greatly simplifies the
communication protocol, as only the application layer has to be
dealt with. Moreover the module is optimized in power
consumption (20mA in full transceiver mode) and it requires
no external circuitry, further decreases additional power losses.
A tailored application layer protocol is used on the elaboration
platform to establish a Bluetooth connection with the remote
sEMG acquisition board.
Fig. 5. Picture of the used prosthetic hand.
97
D. Prosthetic Hand Interfacing
The prosthetic hand, used in this research, is the i-LIMB
hand produced by Touch Bionics (see Fig. 5). It is an
externally powered, multi-fingered hand with interesting
features about controlling and consumption. It is suitable for
experimenting on prosthetic control allowing the user to
control each one of the four fingers and the motorized
rotatable thumb. The control of the hand can be achieved by
an UART interface connected to a further Bluetooth adapter
(COTS based interface). This connection is established with
the second Bluetooth adapter of the mobile device.
E. Machine Learning Methods for sEMG-Based Hand
Movement Classification
The state of the art comprises a lot of machine learning
methods that are applied to sEMG signals with promising
results [15],[16],[17]. The main blocks of the classification
chain needed in the prosthetic hand control consist on i)
filtering and pre-processing; ii) segmentation; iii) features
extraction and iv) classification.
The signals acquired using by sEMG sensors are usually
affected by noise filtering is necessary as part of the
preprocessing step. The most critical is the one coupled by the
power line available in the environment. Even in batteryoperated systems, the noise induced by cabling in the 50-60Hz
range can be lead to inferior signal quality, since the useful
signal bandwidth correlates with the noise. However, this can
be easily addressed exploiting the potential of the on-board
microcontroller or using a suitable low power hardware
accelerator [18]. Segmentation is fundamental when it is
necessary to split streams of sensor data into fragments for
post-processing algorithms like features extraction and
classification.
one of the two sets of sEMG sensor setups, i.e. the wireless
and the wired version (see Fig. 6).
Fig. 6. Phase of the experimental measurements of the proposed
system.
For both setups we used a standard PC workstation of
similar performance to receive and process the acquired data,
as the focus had been on the wireless acquisition node and not
on demonstration of mobile capability. A LabVIEW
application was used to display and compare the accuracy
online, the experimental data is shown in Fig. 7.
In our system, we aim at using methods already well
experimented in the wired testbed, as depicted in [19],[20].
Even though the description of the methods are out of the
scope of this paper, for the sake of completeness, they are
based on SVM models previously built. They clearly show
that machine learning-based classification and regression
applied to surface EMG achieve good performance, in
particular in recognition rate for grasp-type classification.
Summarizing, by the combination of a high number of surface
electromyography, a significant increases in accuracy can be
achieved.
The main challenge in moving toward a low cost and low
power wireless prosthetic hand control system is introducing
software optimizations in all the steps presented, for reducing
the power consumption of the classifier. At the same time, we
aim at boosting the classification phase by reducing the delays
between EMG stimuli and actuation.
IV.
EXPERIMENTAL SETUP
We performed several experiments to assess the signal
quality of the data received by comparing the data received
with a standard wired, high quality data acquisition card by
National Instruments. The same EMG stimuli are fed to each
Fig. 7. Experimental comparison between a traditional DAQ
system and the presented sEMG wireless acquisition board.
A waveform graph is used to display a single set of samples
received from the computers serial port buffer. Lab-View
provides an easy access to the computers serial ports through
the use of the integrated VISA interface, including the
Bluetooth dongles that emulate virtual COM-ports in the serial
port profile (SPP). To visualize the EMG signals sent from the
acquisition board, we parse the data byte-wise. Matlab is then
used for quantitative analysis on the samples. The number of
98
samples acquired by the wireless acquisition module and the
National Instruments DAQ card may differ, also slight jitter is
possible due to the lacking synchronization of the both signal
sources; consequently, interpolation techniques must be used
to compare samples from the wireless acquisition system with
the National Instruments DAQ system.
As depicted by the results in Fig. 7, the mean square error is
quite below 50mV with direct differential comparison. Slight
oscillations and big overshoots are assumed to be a result of
jitters between both source signals. This result is considered
very promising considering the priority set requirements for
sEMG acquisition systems.
The current consumption of the developed wireless sEMG
acquisition board has been measured to be at 32mA, which
results in a maximum of two and a half days of continuous,
independent operation together with the proposed 2000mAh
rechargeable battery. This is well beyond the interval of 24
hours usually considered a critical threshold for pushing a
mobile device onto the market. Concerning the production
price, the developed wireless acquisition system showed –
with approximately 60 Euros – to be far less expensive than
the required target price, with the prototype multilayer board
being the most expensive part.
V.
CONCLUSION
In this paper, the development of a fully wireless, low cost
sEMG acquisition system has been presented. The EMG
signals are acquired from the electrodes attached on the
patient’s remaining forearm stump using Otto-bock 13E200
EMG electrodes. This system, based on the MSP430F5529
microcontroller, converts the analog signals into a digital data
stream that is sent to a mobile processing device by
Bluetooth. Experimental analysis shows that the output
signals of the wireless acquisition system correlates to an
industrial data acquisition board based on National
Instruments NI-DAQ, the maximum mean square error is
below 50mV. The aimed pricing for the system could well be
undercut by more than the half. The systems scalability and
the possibility to draw distributed subsystems together is a
promising strategy for further steps.
ACKNOWLEDGMENT
The work presented in this paper was supported by the
project GreenDataNet, funded by the EU 7th Framework
Programme (grant n.609000).
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
REFERENCES
[1]
[2]
Castellini, C. & van der Smagt, P. Surface EMG in advanced hand
prosthetics Biological Cybernetics, 2009, 100, 35-47
Guanglin Li: Electromyography Pattern-Recognition-Based Control of
Power Multifunctional Upper-Limb Prostheses, 2007
[20]
Kuiken, T.; Li, G.; Lock, B.; Lipschutz, R.; Miller, L.; Stubblefield, K.
& Englehart, K: Targeted muscle reinnervation for real-time myoelectric
control of multifunction artificial arms, JAMA., vol. 301, pp. 619-628.,
2009
K.R. Wheeler, C.C. Jorgensen, Gestures as input: neuroelectric joysticks
and keyboards, Pervasive Computing 2 (April–June (2)) (2003) 56–61.
M.A. Oskoei, H. Hu, Myoelectric control systems—a survey,
Biomedical Signal Processing and Control 2 (October (4)) (2007) 275–
294.
Xun Chen, Z. Jane Wang, Pattern recognition of number gestures based
on a wireless surface EMG system, Biomedical Signal Processing and
Control, Volume 8, Issue 2, March 2013, Pages 184-192
Wang C.S., Methodology Report - Design of a 32-channel EEG System
for Brain Control Interface Applications, Hindawi Journal of
Biomedicine and Biotechnology, Volume 2012, Article ID 274939
Bortolotti, D., Mamaghanian, H., Bartolini, A., Ashouei, M., Stuijt, J.,
Atienza, D., Vandergheynst, P., Benini, L., Approximate compressed
sensing: Ultra-low power biosignal processing via aggressive voltage
scaling on a hybrid memory multi-core processor, (2014) Proceedings of
the International Symposium on Low Power Electronics and Design, pp.
45-50.
L.M. Vaca Benitez, M. Tabie, N. Will, S. Schmidt, M. Jordan, E.A.
Kirchner, Exoskeleton Technology in Rehabilitation: Towards an EMGBased Orthosis System for Upper Limb Neuromotor Rehabilitation,
Hindawi Journal of Robotics, Volume 2013, Article ID 610589
BRAIN PRODUCTS GmbH, BrainAmp ExG mR Technical
Specifications, www.brainproducts.com, 2015
BRAIN PRODUCTS GmbH, PowerPack Technical Specifications,
www.brainproducts.com, 2015
S.S. Lee, K.-Y. Shin, J.H. Mun, Development of a Portable and Wireless
Surface EMG, Key Engineering Materials Vols 321-223 (2006), pp.
1107-1110
Q. Xiong, Design of an Advanced Portable System for high Density
Surface EMG Recording with Wireless Control of Signal Quality.
Politecnico di Torino, http://porto.polito.it/2592680, March 2015
Caione, C., Brunelli, D., Benini, L., Compressive sensing optimization
for signal ensembles in WSNs, (2014) IEEE Transactions on Industrial
Informatics, 10 (1), art. no. 6523111, pp. 382-392.
B. Karlik, M. Osman Tokhi, M. Alci, A fuzzy clustering neural network
architecture for multifunction upper-limb prosthesis, IEEE Transactions
on Biomedical Engineering 50 (11) (2003) 1255–1261.
Jun-Uk Chu, Inhyuk Moon, Mu-Seong Mun, A real-time EMG pattern
recognition system based on linear-nonlinear feature projection for a
multifunction myoelectric hand, IEEE Transactions on Biomedical
Engineering 53 (11) (2006) 2232–2239.
K.S. Kim, H.H. Choi, C.S. Moon, C.W. Mun, Comparison of k-nearest
neighbor, quadratic discriminant and linear discriminant analysis in
classification of electromyogram signals based on the wrist-motion
directions, Current Applied Physics 11 (3) (2011) 740–745.
Benatti, S.; Milosevic, B.; Casamassima, F.; Schonle, P.; Bunjaku, P.;
Fateh, S.; Huang, Q.; Benini, L., "EMG-based hand gesture recognition
with flexible analog front end," Biomedical Circuits and Systems
Conference (BioCAS), 2014 IEEE , vol., no., pp.57,60, 22-24 Oct. 2014
Gijsberts, A.; Atzori, M.; Castellini, C.; Muller, H.; Caputo, B.,
"Movement Error Rate for Evaluation of Machine Learning Methods for
sEMG-Based Hand Movement Classification," Neural Systems and
Rehabilitation Engineering, IEEE Transactions on , vol.22, no.4,
pp.735,744, July 2014
Tommasi, T.; Orabona, F.; Castellini, C.; Caputo, B., "Improving
Control of Dexterous Hand Prostheses Using Adaptive Learning,"
Robotics, IEEE Transactions on , vol.29, no.1, pp.207,219, Feb. 2013.
99