Intelligent Robotic Approach for After-stroke Hand Rehabilitation

Nirvana Popescu

, Decebal Popescu

and Mircea Ivănescu

Department of Computer Science, University Politehnica of Bucharest, Bucharest, Romania

Department of Mechatronics, University of Craiova, Craiova, Romania

Keywords: Post-Stoke Rehabilitation, Robotic Glove, Intelligent Control, Speech Recognition.

Abstract: This paper presents the design of an intelligent haptic robotic glove (IHRG) model for the rehabilitation of

the patients that have been diagnosed with a cerebrovascular accident (CVA). Total loss or loss of range of

motion, decreased reaction times and disordered movement organization create deficits in motor control,

which affect the patient’s independent living. The control system for a rehabilitation hand exoskeleton is

discussed. One contribution is given by using a velocity observer and a force observer for performance

evaluation. The disturbance effects are eliminated by a cascade closed loop control with velocity and force

observers. The performance of the control system is demonstrated by the simulation. The second proposed

control implementation version has a great advantage - the possibility to specify some vocal commands, which

will help the patient to make a lot of medical exercises by themselves.

1 INTRODUCTION

The scientific community has become increasingly

interested in so-called Rehabilitation Robotics, a

branch of the areas of Robotics and Mechatronics that

addresses to the study of complex robotic systems

aiming to restore the human functions for those

people who suffer major trauma as a result of strokes

and cerebrovascular accident (CVA). A CVA occurs

when a blood vessel (an artery) that supplies blood to

an area of the brain bursts or is clogged by a blood

clot. Within minutes, the nerve cells in that area are

affected and they can die in a few hours. As a result,

the part of the body that is controlled by the affected

area of the brain cannot function properly

(Grebenstein, 2010).

Total loss or loss of range of motion, decreased

reaction times and disordered movement organization

create deficits in motor control, which affect the

patient’s independent living. Recent studies have

shown that intensive and repetitive training may be

necessary to modify neural organization and recover

functional motor skills. Several researchers

(Reynaerts, 1995; Rodriguez-Cheu and Casals, 2006;

Zhao et.al., 2006; Lucas et.al. 2004) have reported

significant improvement in patients’ daily activities

due to higher training intensities, even in the CVA

chronic phase. However, there were several opinions

in the literature claiming that the repetitive training is

not very efficient (French et al., 2010) and other

papers proved that there are some reduced

improvements (Barbay et al., 2013). On the other

hand there are authors that certified the improvement

arm movement ability (Housman et al., 2009).

The current health care system provides stroke

rehabilitation in the intensive care hospital setting, in

the rehabilitation setting and in the outpatient setting.

In the last decade, the literature presented numerous

concepts and techniques that allow evaluation of the

hand physiological properties, structure,

characteristics and especially the functional anatomy.

There is a great need to develop new approaches for

rehabilitation of the upper limb after stroke. Robotic

therapy is a promising form of neurorehabilitation

that can be delivered in more intensive regimens than

conventional therapy (Kitago et al. 2015). Numerous

studies (Carrozze et.al. 2003; Birglen and Gosselin,

2003, 2004; Biagiotti et.al., 2009, Lotti and Vassura,

2005, Brokaw et.al. 2011, Jiting et.al., 2011) have

allowed the development of kinematic structures to

reproduce as much as possible the human hand

kinematics.

This paper presents the design of a low-cost

Intelligent Haptic Robot-Glove (IHRG) for the

rehabilitation of the patients that have been diagnosed

with a cerebrovascular accident (CVA). The IHRG is

an exoskeleton that supports the human hand and

hand activities by using a control architecture for

dexterous grasping and manipulation. IHRG is a

Popescu, N., Popescu, D. and Iv

anescu, M.

Intelligent Robotic Approach for After-stroke Hand Rehabilitation.

DOI: 10.5220/0005662400490057

In Proceedings of the 9th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2016) - Volume 5: HEALTHINF, pages 49-57

ISBN: 978-989-758-170-0

medical device that acts in parallel to a hand in order

to compensate some lost function.

In order to design an exoskeleton structure and to

develop techniques that allow rehabilitation of the

main anatomical features of the hand, in the context

of extremely varied range of patients with various

problems of malfunction, the architecture presented

in this paper should cover the range of issues and

anatomical structures.

Feedback is important in rehabilitation.

Rehabilitation is most effective when users get

immersive feedback that relates to the activities they

imagine or perform. For example, if people imagine

grasping an object with their left hand, then an image

of a grasping hand can help users visualize their

activity. If a stroke patient keeps trying to imagine or

perform the same movement, while receiving

feedback that helps to guide this movement, then

users might regain the ability to grasp, or at least

recover partial grasp function. In the last few years,

totally novel and promising application for motor

imagery (MI) - based Brain-Computer Interface

(BCI) has gained attention (Irimia et al, 2012, 2014).

In the case when patients can talk and the stroke

did not affect their vocal capacity of pronouncing

words, a new design of IHRG is also presented in this

paper. The implementation is made based on some

hardware platforms. One of them is Arduino Mega

2560 which is a hardware platform that determines

the movement of some small engines that will help

the patient to open or close one or more fingers. The

second one is the Raspberry Pi hardware platform

which was used in order to make all the calculations

specific to voice recognition. Using a microphone, the

patient can send vocal commands to IHRG or can

select one predefined program which will imply some

exercises for a specific period of time.

One major contribution of this paper consists of

using observers for this type of rehabilitation system.

All the analysed papers, some of them being included

in the state of the art of this work, have used velocity

sensors for performance evaluation. Instead, we used

velocity observers and force observers for doing this.

The second contribution regards the voice recognition

based implementation that offers the facility of live

interaction between the system and the patient.

The paper is structured as follows: section II

presents the design and development of the structure

of the intelligent haptic robotic glove that includes

hand biomechanics, the exoskeleton architecture and

the control architecture; section III shows a different

control approach based on voice recognition and

details the hardware implementation in this case;

section VI is concerned with conclusions.

2 DESIGN AND DEVELOPMENT

OF IHRG STRUCTURE

For a correct analysis of hand function, a researcher

must thoroughly analyse three main components: the

upper hand and wrist, fingers (four) and thumb.

Different biomechanical movements of flexion,

extension, abduction and adduction were analysed,

considering the amplitudes, movement directions,

axes and planes. The model of hand articulations is

presented in Figure 1.

Carpal osteoarticular complex is so constituted as

to permit carrying out movements of flexion-

extension, abduction – adduction and circumduction.

Because these joints are plane, each of them allows

sliding movements of small amplitude. Wrist joint

movements occur simultaneously: both in radiocarpal

and mediocarpal joints. All these movements are

accomplished by sequential displacement of region

segments: the second carpal (distal) moves on the first

carpal (proximal) and then the latter slipped on the

forearm. Functional position of the hand is the active

position, ready to grab. This means that hand makes

a dorsal flexion with 20°, the fingers are slightly

flexed and the thumb is in opposition. The muscles

are slightly tensed; the extensors of the hand and the

flexors of fingers exert a dominant action on

antagonists. This is an important aspect since in the

case of stroke occurrence and central level control

interruption, this normal anatomic position is

accentuated being accompanied by spasticity.

Using haptic interface technology allows for

reaching and grasping movement executed by hand to

be assisted by a robot that direct the movement to a

specific target. Until now, the technologies developed

for this purpose have the potential to revolutionize the

way hospitals operate, reducing the recovery stroke

cost while allowing therapists and clinicians to

manage a large number of patients in the same period

of time.

Figure 1: Hand joints model.

HEALTHINF 2016 - 9th International Conference on Health Informatics

2.1 The Exoskeleton Architecture

One of the proposed solution is to develop a

mechanical architecture consisting of a cascade of

articulated elements, whose design to cover as much

as possible the anatomic and functional finger

phalanges, providing support for the actuation

system.

Two architectures were analysed: the 4-bar

mechanism and the tendons mechanism. In tendons

mechanism architecture, the three phalanges of the

finger are realized as a serial structure of three

rotation joints, each joint being controlled in cascade

by the previous joint through a tendon-cable that is

coupled on the pulleys system associated with each

joint. In this case, kineto-static analysis principles

were studied. Then it was considered a finger

architecture based on 4-bar mechanism where the

movement is achieved by successive deformations of

the 4-bar mechanisms associated with each phalanx

of the finger.

The proposed architecture is a control architecture

where the control is performed at the first active joint,

the other joints being passive and realizing

movements based on the associated bar system

(Popescu et.al., 2013). The global architecture of one

finger exoskeleton and of the entire structure can be

seen in Figure 2.

Figure 2: The global architecture of one finger exoskeleton

and of the whole hand.

2.2 The Control Architecture

The control architecture was developed considering

that the control is performed at the first active joint,

the other joints being passive and realizing

movements based on the associated bar system. The

proposed overall structure is shown in Figure 3.

It can be observed the distribution of forces on the

phalanges. A kinetostatic analysis was performed, the

movement of three phalanges being realized planar.

A coordinate system X

, Y

, O

and an articulated bar

system with parameters: a

, b

, c

are assigned to each

phalanx of l

length. The finger movement is achieved

by rotating each finger phalanx withe the angles θ

, θ

. The control forces on the external environment,

, F

are applied in d

, d

representing the

contact points on a human hand. It is considered the

active torque T

applied on the finger joint 1.

Figure 3: The 4-bar mechanism architecture of a phalanx.

Our approach also proposes a control scheme

dedicated to underactuated fingers with the intention

of maximizing the capabilities of the control using

position information. Position sensors are

implemented on the mechanical transmission system

and used to enhance the behaviour of the hand despite

its limited number of control signals. Several

positions of a finger in a rehabilitation exercise are

illustrated in Figure 4.

Figure 4: Finger motion for a rehabilitation exercise.

The control system needs self-tuning to adjust

dynamics and kinematics of the system. This self-

tuning has to compensate effects of internal and

external disturbances: model uncertainty, load,

friction and compliance with human finger. In some

research works, the force control for grasping is

obtained by using tactile feedback. The limited space

on the finger surface makes very difficult the

introduction of a sensor on that area. For this reason,

velocity and force observers are used to implement

the control technique.

The actuation system for each finger is based on

one dc motor moving a slider device which transmits

motion to the phalanges by several coupled „four-

bar” mechanisms. The dynamic model of the

Intelligent Robotic Approach for After-stroke Hand Rehabilitation

actuation system and load mechanism is described by

the following equations























(1)



















 



(2)

 





(3)

  



(t)

(4)

where

,,, are the state parameters: the drive

current, mechanical angular velocity, angular

position, motor torque, respectively,

,,



represent the electrical motor parameters: armature

inductance and resistance and back EMF constant,

respectively, and

,



,



,



are mechanical

parameters: nominal inertia, damping constant,

nominal torque constant and spring constant. Also





is an equivalent disturbance torque that focuses

the effect of friction, external forces determined by

the compliance with human finger and unmodeled

dynamics in the transmission mechanism.

The design of the force observer is based on

(Lucas et.al., 2004) and estimates the force from the

dynamic model of motion, avoiding the use of the

acceleration signal, measured or construct from

velocity by differentiation. By using (2), (4) (with





0, the force observer is defined by











 









(5)

where

̂ is the estimated force-torque and

represents the observer gain (in order to simplify

the notation, the variable

is omitted). We select the

observer gain as

L = α / J, α > 0 (6)

A classical cascade disturbance closed loop

control is proposed in Figure 5. The system contains

the actuation system, mechanical transmission system

and a conventional controller (for example, a PID

controller). Measuring signal is the angular position

.

Figure 5: A cascade closed loop control with observers.

A velocity observer is used to generate the estimated

value of velocity

 using the reference value of the

torque, τ

ref

This signal is used in the force observer to

generate the estimated value of the torque

̂. The

forces between hand and exoskeleton act as the

disturbances





. The classical conventional tuning

rules are used for the control parameters in order to

compensate the disturbances and to ensure the motion

performances. The disturbances are determined by

the resistance increasing to passive finger extension.

Characterization of hypertonia can be difficult due to

the highly variable nature of the hyperactivity of the

finger flexors and the characteristics of each patient.

The observer error will be





=  ̂

(7)

From (6) yields

















(8)

We assume that the torque varies slowly relative

to the observer dynamics (Lucas et al., 2004; Chen et.

al., 2000; Xian et.al., 2004)] and it was supposed that





0

(9)

In terms of this condition, from (5), (10), the error

dynamics becomes,











 /

(10)

that proves that the observer is globally

asymptotically stable. The parameter

∝ determines

the time constants for the estimation. In order to

simplify the observer construction, a function





defined as





̂ 



(11)

By using (5), (10), (11), we obtain the new

observer equation as





 























∝



















(12)

The equation (12) allows the estimation of the

torque using the position and velocity signals,

̂ 







(13)

The velocity observer aims to estimate the

inaccessible velocity signal







using only the position

. For a reference torque-force τ

ref

, the motion

equation (2) becomes







 



  



(14)

and the observer model is selected as,

(15)

HEALTHINF 2016 - 9th International Conference on Health Informatics

(16)

where e

is the position error,

(17)





is the estimated value of the angular position

and k

, k

, and k

are constant positive observer design

parameters.

In order to demonstrate the performance of the

control system, a numerical simulation was realized.

So, we considered the dynamic model of the actuation

system and load mechanism described by (1)-(4),

where the electrical parameters of the dc motor are:

L=0.0138 H, R= 26.44Ω, k

=0.1656 Nm/A, k

0.982 Vs/rad and the mechanical

parameters of the

system are: J= 0.000254 kg m

Ω

=0.002031Nms/rad, k

=2.45 Nm/rad. The model

dynamics (1)-(4) are simulated in MATLAB

SIMULINK. First, the force and velocity observers

(12)-(13), (15)-(16), respectively, are implemented

(Popescu et.al., 2013).

Figure 6: The estimation of velocity by observer.

The velocity signal estimated by the velocity

observer is presented in Figure 6. We remark the good

quality of estimation. This signal is used as input

variable in the force observer. A PID controller was

implemented and a cascade closed loop control

system (Figure 5) was studied. The results are

illustrated in Figure 7.

Figure 7: The force control by velocity and force observers.

3 HARDWARE

IMPLEMENTATION AND

VOICE CONTROL

In this approach, the design of IHRG is made based

on some hardware platforms. One of them is Arduino

Mega 2560 which is a hardware platform that

determines the movement of small engines that help

the patient to open or close one or more fingers. The

second one is the Raspberry Pi hardware platform

which was used in order to make all the calculations

specific to voice recognition. Based on a microphone,

the patient can send vocal commands to IHRG or can

select one predefined program which will imply some

exercises for a specific period of time (Popescu et.al.

, 2014).

The glove must be very thin in order to have a

comfortable environment for exercises. Another

aspect, which was taken into consideration, was the

fact that this type of glove is like an infrastructure

suitable for a lot of wires and sensors. The wires will

help the patient to close or open the hand, for

example, as long as the sensors will calculate real

time the resistance force generated by the patient.

All these calculations are mandatory for knowing

the action time of every engine. In fact this time

implies the force which is applied for every finger,

and this force must be correlated with the reaction

obtained from the patient in order to avoid finger

fracture. Nylon threads are used for connecting the

top fingers with the small engines. There are 5 small

engines, one for every finger and they have the major

role in implementing operations like open or close. Of

course, that these operations will be made partially –

the reaction force (the feedback) from the patient is

taken in real time and, based on these values, the

operation time for every small engine is determined.

For this purpose a bending sensor is used– one sensor

for every finger.

Another kind of sensor used in the design is the

force sensor. This sensor was added for a scenario in

which the patient will want to take some fragile

objects. These sensors are putted on the top of every

finger and will establish when the function of the

small engine will be stopped. A two-cell LiPo battery

(7.4 V) and 1000 maH battery used in design are

enough for 1 day high intensive use of the glove.

For this experiments, Arduino Mega 2560

platform was used in the implementation due to some

advantages like:

 5 ports for PWM signals which are sent to the

engines,

 10 analog ports which are used to read the

signals coming from the bending and force sensors,

 4 serial lines – one of them is used for wireless

communication between IHRG and the speech

recognition system,

 I2C busses used in order to communicate with

the LCD. On this LCD the patient can read the current

Intelligent Robotic Approach for After-stroke Hand Rehabilitation

program, chosen by himself, and some values for

some forces will be displayed here.

For the wireless communication (between the

intelligent glove and the control device) 2 xBee

modules were used. These modules work at 3.3V and

the Arduino board works at 5V. For this reason an

xBee adaptor must be included in the design. Another

conversion must also be realized – the battery offers

7.4V and the glove operates at 5V – using an

integrated circuit LM 7805.

The Raspberry Pi board is used to receive vocal

commands, to recognize these commands (speech

recognition) and to send these commands wireless to

the glove. The microphone and the soundboard are

connected to the Raspberry Pi board via USB port.

The whole design of the control system is shown in

Figure 8.

The software includes some libraries and open

source programs like: a voice recognition engine

(Julius HTK) which has a large words database,

Pexpect used for automatically speech recognition

and a Python module developed to control the

applications. The first step was to setup the

environment such that the Julius engine can start and

load at startup all the libraries and dependencies. In

the next step the library implementation was made.

This means to establish the vocabulary that will be

further used to control the intelligent glove. Then, the

vocabulary was translated to a phonetic form and

saved in a file with a specific extension (.voca). This

file contains every word, which will be used splitted

into the corresponding phonemes.

Figure 8: The control system of IHRG using Arduino.

These phonemes represent the sound unit of word.

The chosen language was English due to the fact that

the recognition level is very high. The English

language has usually 44 phonemes, and the

determination of them was made verifying 2 sources:

the lexicon file, which is available in Julius and the

CMU dictionary, which contains phonemes. The next

step involves rules definitions for the grammar, rules

that must be respected. After this step the result is a

file with .grammar extension in which is defined the

whole subset of possible sentences. The grammars are

described using Backus-Naur form and they will be

applied for getting the exact description of the

required language. After that, the mkdfa.pl script is

used to convert the results from the previous steps in

a Julian / Julius format. Julian is a special version of

Julius doing speech recognition based on acoustic

models.

The next step supposes as input the acoustic

models correlated with Julius HTK. In this case it is

necessary to define the configuration file used by

Julius. This is a very important step because in this

manner it is possible to define the properties for

samples creation. These samples will be processed by

HTK. When the Julius application will be started, the

patient can also start to use the vocabulary defined in

the previous steps. In the identification process, for

each word or sentence a score will be obtained. This

score will reflect the recognition accuracy level. Also,

a Viterbi score is obtained. This Viterbi score is got

via identification Viterbi algorithm for the same

sequence of hidden state.

HTK interaction involves some steps like the

following ones:

Step 1 – install the application and setup the

environment in order to have all the dependencies

available and also the libraries.

Step 2 – The acoustic models and the training

phase are done by the patient. Initially, a number of

samples and their exact form will be defined in a file.

After that the patient will register his voice using the

audacity samples defined in the file.

Step 3 – After all the samples were recorded, it is

necessary to define a configuration file that converts

the file with .wav extension to a file with .mfc

extension because only this extension is recognized

by HTK. This configuration file will be parsed by a

script, designed for the corresponding conversion.

For software configuration, in order to obtain

acoustic models in audacity, the following settings

were made: sample rate of 48Khz, 16 bits per sample,

mono channel and the resulting wave file was

converted to Microsoft 16-bit PCM format.

The software uses a Python module called

Pexpect in order to create a pseudo terminal, which is

used to connect to Julius engine. The reason is to

obtain the generated output. After that, this output is

processed using a filter for initial lines in which a

sequence is identified and a score is obtained for it.

Then, another filter is applied based on scores and

some threshold values. Finally, the value, which

HEALTHINF 2016 - 9th International Conference on Health Informatics

satisfies all the criteria will be recorded and translated

to a specific format which, will be further processed

by IHRG.

The communication between Raspberry Pi and

IHRG is made wireless and uses a protocol. When a

command is identified from the patient, a translation

code command is generated in order to activate the

sensors of IHRG. The codification is:

- for every bending sensor a value between 0

and 9 will be transmitted, where 0 means standby and

9 represents the maximum bending.

- for the force sensor a value between 0 and 4

will be generated, where 0 means a contact and 4

means maximum fundraising amount.

This codification is sent like a string of 6 values.

The most significant value is # and the least

significant value is ~. For example, the code #55552~

represents the command GRAB.

Table 1: Implemented commands and their codes.

wave

444440

grab

555552

soft grab

333331

strong grab

999994

point 909990

open 000000

close 777770

There are a number of 7 commands implemented

and learnt by Julius as it can be seen in Table 1, along

with their codes.

The tests that have been done have proved a very

good detection for simple commands like: open,

close, grab, point or wave. Figure 9 shows the

implemented model.

The noise is a factor which can influence the

recognition process. If the noise is almost zero, the

results of speech recognition are very good; when the

level of noise is increasing, the accuracy of speech

recognition decreases a lot. Also, in case of Raspberry

Pi the tests show that the detections of commands are

not so good like in the case in which a personal

computer is used.

The precision of results can be improved by

modifying the recorded values or changing the

Viterbi score. The training phase for the acoustic

model is also another way to improve the accuracy.

In time, the data grow up and it will be reached a point

in which the accuracy remains the same. This point is

the maximum point. In order to create an acoustic

model used by more patients, it will be mandatory as

each patient to contribute with his own audio input in

the system.

Figure 9: IHRG implementation for the voice recognition

control approach.

4 CONCLUSIONS

This paper presented the design of an intelligent

haptic robotic glove (IHRG) model for the

rehabilitation of the patients that have been diagnosed

with a cerebrovascular accident (CVA). Total loss or

loss of range of motion, decreased reaction times and

disordered movement organization create deficits in

motor control, which affect the patient’s independent

living. Some studies have shown that intensive and

repetitive training may be necessary to modify neural

organization and recover functional motor skills.

The IHRG is an exoskeleton that supports the

human hand and hand activities by using a control

architecture for dexterous grasping and manipulation.

The five-fingered assistive robotic glove was

designed with mechanical compliance of human

finger. The biomechanical elements of the

exoskeleton assistive hand were designed

considering: motion in different planes, adapted to

patient’s hand, possible to train pinch and grasp,

opening/closing game. A global architecture for the

hand exoskeleton was presented. The actuation

system was also studied and designed. The velocity

and force observers are used to implement the control

technique.

The post stroke rehabilitation robotic glove was

also designed and implemented in our laboratory

considering the fact that for any patient it will be more

comfortable to make medical exercises in his own

home. The second proposed control implementation

version has a great advantage - the possibility to

specify some vocal commands, which will help the

patient to make a lot of medical exercises by

themselves. To prove the robustness of this approach,

a lot of tests have been done and a very good detection

was obtained for simple commands like: open, close,

grab, point or wave and also for complex commands

like: soft grab or strong grab. Different factors which

Intelligent Robotic Approach for After-stroke Hand Rehabilitation

can influence the recognition process were also

analysed.

One major contribution of this paper consists of

using observers for this type of rehabilitation system.

All the analysed papers, some of them being included

in the state of the art of this work, have used velocity

sensors for performance evaluation. Instead, we used

velocity observers and force observers for doing this.

The second contribution regards the voice recognition

based implementation that offers the facility of live

interaction between the system and the patient. It is

also important to mention the reduced cost of the

proposed solution.

The presented approaches represent a part of a

work in progress project. We are currently involved

in testing the systems to obtain the complete results

after working with patients. This represents the

subject of a next paper.

Regarding future work, we also analyse the

possibility to implement an EEG-based brain-

computer interface that can be used to command a

semi-autonomous robotic glove by means of motor

imagery (MI). The BCI detects the intention to move

and provides online feedback to the user. At the same

time, the feedback can be used as trigger for different

pre-programmed robotic motion tasks.

ACKNOWLEDGEMENTS

This work is supported by PCCA 150/2012 grant of

the Executive Agency for Higher Education,

Research, Development and Innovation Funding

(UEFISCDI).

REFERENCES

Barbay, S., Guggenmos, D. J., Nishibe, M., Nudo, R. J.,

2013. Motor representations in the intact hemisphere of

the rat are reduced after repetitive training of the

impaired forelimb. Neurorehabilitation and neural

repair, 27: 381-384.

Biagiotti, L., Lotti, F., Melchiorri, C., Vassura, G., 2009.

How Far Is the Human Hand? A Review on

Anthropomorphic Robotic End-effectors, DEIS -

DIEM, University of Bologna.

Birglen, L., Gosselin, C., 2003. On the Force Capabilities

of Underactuated Fingers, Proc. IEEE Intl. Conf. Rob.

Aut., Taipei, Taiwan, pp. 1139-1145.

Birglen, L., Gosselin, C., 2004. Kinetostatic Analysis of

Underactuated Fingers, IEEE Trans. Rob. Aut., 20(2),

pp. 211-221.

Birglen L., Gosselin, C., 2004. Optimal Design of 2-

Phalanx Underactuated Fingers, Proc. Intl. Conf.

Intel.Manip. Grasp., Genova, Italy, pp. 110-116.

Brokaw, E.B., Black, I., Holley, R., Lum, P., 2011. Hand

Spring Operated Movement Enhancer (HandSome): A

Portable Passive Hand Exoskeleton for Stroke

Rehabilitation, IEEE Trans on Neural Systems and

Rehabilitation Eng vol 19, No4 ,August, pp391-398.

Carrozze, M.C., Vecchi, F., Sebastiani, F., Cappiello, G.,

Roccella, S., Zecca, M., Lazzarini, R., Dario, P., 2003.

Experimental Analysis of an Innovative Prosthetic

Hand with Proprioceptive Sensors, Proc. IEEEIntl.

Conf. Rob. Aut., Taipei, Taiwan, pp. 2230-2235.

Chen, W.H., Ballance, D.J., Gawthrop, P.J., O’Reilly, J.,

2000. A Nonlinear Disturbance Observer for Robotic

Manipulators, IEEE Trans on Industrial Electronics,

vol 47, No 4, August, pp 932-938.

French, B., Thomas, L., Leathley, M., Sutton, C., McAdam,

J., Forster, A., Langhorne, P., Price, C., Walker, A.,

Watkins, C., 2010. Does repetitive task training

improve functional activity after stroke? A Cochrane

systematic review and meta-analysis. Journal of

rehabilitation medicine 42: 9-14.

Grebenstein, G., 2010. A Method for Hand Kinematic

Designers, ICABB, Venice, Italy.

Housman, S.J., Scott, K.M., Reinkensmeyer, D.J., 2009. A

randomized controlled trial of gravity-supported,

computer-enhanced arm exercise for individuals with

severe hemiparesis. Neurorehabil Neural Repair 23:

505-514, 2009.

Irimia, D.C., Popescu, C.D., Poboroniuc, M.S., Ignat, B.E.,

Bolbocean, O., 2014. Using a Motor Imagery based-

BCI system for neuroprosthesis control, The 12th

Congress of the Romanian Society of Neurology,

Bucharest, Romania, Romanian Journal of Neurology

(8), ISSN online: 2069-6094, ISSN-L 1843-8148.

Irimia, D.C., Ortner, R., Krausz, G., Guger, C., Poboroniuc,

M., 2012. BCI Application in Robotics Control, 14th

IFAC Symposium on Information Control Problems in

Manufacturing, Bucharest, Romania, 14 (1): 1-6, ISSN:

1474-6670; ISBN: 978-3-902661-98-2, DOI:

10.3182/20120523-3-RO-2023.00432.

Kitago, T., Goldsmith, J., Harran, M., Kane, L., Berard, J.,

Huang, S., Ryan, S.L., Mazzoni, P., Krakauer, J.W.,

Huang, V.S., 2015. Robotic therapy for chronic stroke:

general recovery of impairment 5 or improved task-

specific skill?, Jurnal of Neurophisyhology,

114(3):1885-94. doi: 10.1152/jn.00336.2015.

Li, J., Wang, S., Wang, J., Zheng, R., Zhang, Y., Chen, Z.,

2011. Development of a Hand Exoskeleton System for

Index Finger Rehabilitation, Chinese Journal of

Mechanical Engineering Vol. 24,aNo. 5.

Lotti, F., Vassura, G., 2005. A Novel Approach to

Mechanical Design of Articulated Fingers for Robotic

Hands. DIEM, Mech.Eng. Dept, University of Bologna.

Lucas, L., DiCicco, M., Matsuoka, Y., 2004. An EMG-

Controlled Hand Exoskeleton for Natural, Journal of

Robotics and Mechatronics, Vol.16 No.5, pp 1-9.

Popescu, N., Popescu, D., Cozma, A., Vaduva, A.J., 2014.

Hardware Design and Implementation of an Intelligent

HEALTHINF 2016 - 9th International Conference on Health Informatics

Haptic Robotic Glove, Proc. of the the 8th

International Conference and Exposition on Electrical

and Power Engineering, EPE2014, IASI Romania,

ISSN: 978-1-4799-5848-1, 2014.

Popescu, N., Popescu, D., Ivanescu, M., Popescu, D.,

Vladu, C., Vladu, I., 2013. Force Observer-Based

Control for a Rehabilitation Hand Exoskeleton System,

Asian Control Conference (ASCC2013), pp. 1-6,

Istanbul, Turkey.

Popescu, N., Popescu, D., Poboroniuc, M., Popescu, C.D.,

2013. Intelligent Haptic Robotic Glove for Patients

Diagnosed with Cerebrovascular Accidents, 19th

International Conference on Control Systems and

Computer Science, Workshop on Medical and

Rehabilitation Engineering Applications, pp 717-721,

October, Sinaia.

Popescu, N., Popescu, D., Ivanescu, M., Popescu, D.,

Vladu, C., Berceanu, C., Poboroniuc, M., 2013.

Exoskeleton Design of an Intelligent Haptic Robotic

Glove, 19th International Conference on Control

Systems and Computer Science, vol. 1, pp. 196-201,

Bucharest.

Reynaerts, D., 1995. Control methods and actuation

technology for whole-hand dexterous manipulation, .D.

dissertation, Prod. Eng. Mach. Design Autom., Catholic

Univ. Leuven, Leuven, Belgium.

Rodriguez-Cheu L.E., Casals, A., 2006. Sensing and

control of a prosthetic hand with myoelectric feedback,

the 1st IEEE/RAS-EMBS Int. Conf. Biomed. Robot.

Biomechatron., Pisa, Italy.

Xian, C., de Queiroz, M. S., Dawson, D. M., McIntyre, M.

L., 2004. A Discontinuous Output Controller and

Velocity Observer for Nonlinear Mechanical Systems,

Automatica, 40, 4, April, pp. 695–700.

Zhao, J., Xie, Z., Jiang, L., Cai, H., Liu, H., Hirzinger, G.,

2006. A fivefingered underactuated prosthetic hand

control scheme, 1st IEEE/RAS-EMBS Int. Conf.

Biomed. Robot. Biomechatron., Pisa, Italy.

Intelligent Robotic Approach for After-stroke Hand Rehabilitation