See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/224195712

Improving the performance of piezoresistive force sensors by modeling

sensor capacitance

Conference Paper · August 2010

DOI: 10.1109/ISIE.2010.5637681 · Source: IEEE Xplore

CITATIONS READS

5 120

3 authors:

Leonel Paredes-Madrid Luis Emmi

Universidad Antonio Nariño Spanish National Research Council
22 PUBLICATIONS   139 CITATIONS    19 PUBLICATIONS   453 CITATIONS   

SEE PROFILE SEE PROFILE

Pablo Gonzalez-de-Santos
Spanish National Research Council
44 PUBLICATIONS   604 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Efecto del grafeno en la resistencia mecánica y eléctrica de un material vitro-cerámico obtenido a partir de residuos industriales de escoria, ceniza volante y casco de
vidrio View project

Robot Fleets for Highly Effective Agriculture and Forestry Management (RHEA) View project

All content following this page was uploaded by Leonel Paredes-Madrid on 12 October 2018.

The user has requested enhancement of the downloaded file.

 Improving the performance of piezoresistive force
sensors by modeling sensor capacitance
L. Paredes-Madrid, L. Emmi and P. Gonzalez de Santos
Center of Automation and Robotics UPM- CSIC
Ctra. Campo Real; Km 0,2
28500 Arganda del Rey; Madrid; Spain
Abstract — Piezoresistive force sensors exhibit considerable
lower accuracy compared with load cells and force measuring
systems based in strain gauges, however a new method for
measuring forces using piezoresistive sensors is described in this
paper, leading to a considerable increase in the repeatability of
force readings. The new method consists of reading sensor’s
conductance and capacitance by applying DC and sinusoidal
waveforms, thereby allow us to determine a multivariable
estimation of force, instead of using the traditional, purely
resistive model that has been used up to now. A study of sensor
nonlinearities is also described which will allow us to determine
the optimal setup to perform capacitance readings under
sinusoidal excitation.
I. INTRODUCTION
F ORCE measurements have always been of great
importance for robots that interact with their
environment. Nevertheless, this importance has increased
along the last decade in robotic applications in which robots
and humans share a common scenario interacting each other
in closed collaboration [1], [2], [3]. This human-robot
interaction calls for continuous monitoring of the forces
exerted by robots on humans to maintain interaction forces
within both comfortable and safe limits [4]. There are,
currently, several application fields in search of solutions for
measuring interaction forces such as intelligent assist devices
[5], haptic interfaces [6], [7], human grip force analysis [8],
[9], [10], instrumented gloves [11], [12], footwear design
[13], and medical robotics, for instance. The ideal force
sensor seems to be that featuring: high repeatability,
increased robustness and reliability, low drift, independent
readings of temperature and magnetic field, small size, light
weight, and low cost.
At the very beginning of the robotics development, force was
measured by installing strain gauges on the robot links [14],
providing a reliable, small, and non-invasive method for
measuring deformation, then the forces applied to the robot’s
links could be deduced from Young’s modulus of material
and the link’s dimensions. This method has been used up
today, but a flexible element is required to achieve a link
deformation. Sometimes, that material deformation is
measured by light intensity on optical fibers [15].
Load cells –that are used in applications that require force
measurement on several axes, high repeatability, high
robustness, and low drift– are another type of force sensors.
Unlike strain gauges, load cells do not need to be installed on
978-1-4244-6392-3/10/$26.00 ©2010 IEEE
a bendy link and are available in many sizes and operation
ranges. However, they are bulky and weighty, which make
them unsuitable in many applications, such as human-grip
force research and haptic interfaces, applications that need
low invasive techniques.
Piezoresistive sensors seem to be the best solution for
those applications, since they are light-weight, thin, and
small. In addition, they can be fit inside a data glove [10] or
installed in human joints, such as knees [16] and elbows, and
they are passive and so inherently safe. However, piezo-
resistive force sensors are deficient in repeatability and exhibit
large hysteresis. These unwanted features reduce this sensor
use to applications that do not demand high accuracy [4];
however, their low cost is generating an increasing use in
industry and research interest, as well.
Two former studies on the most widely used piezoresistive
force sensors [17], [18] were considered in this paper.
Finally, we chose the FlexiForce sensor, because this sensor
manufacturer has developed many specific sensors [13], [16]
for applications in both industrial and service tasks.
II. THE INTERFACE GLOVES-ON PAYLOAD
Endpoint force control has not been an easy task to
achieve in robots, mainly because a detailed model of robot’s
dynamics is needed [20] in order to ensure that the system
remains dynamically stable. The process of distinguishing
between operator-applied forces and forces caused by the
load itself may be a possible cause of instability. To
overcome such difficulty, a new type of haptic interface,
based in data gloves [21], is intended to resolve that problem
by performing non-invasive force measurement. The control
interface consists of a data glove with a 3DOF Inertial
Measurement Unit, IMU, (yaw, pitch, and roll) and a force
measurement system based in piezoresistive force sensors,
specifically in FlexiForce sensors.
Figure 1 shows a sketch of the palm and dorsal side of the
control interface. The IMU and Finger Flexion Sensors (FFS)
are installed on the dorsal side, whereas the force sensors are
installed on the palm side. As soon as the operator puts the
haptic gloves on and exerts force on the load, the
piezoresistive force sensors will measure the magnitude of
such force. Corresponding direction is given to the force
magnitudes by means of the data glove, specifically, through
the FFS which are electro-goniometers.
458
 III. PROPOSAL OF AN ELECTRICAL MODEL FOR THE SENSOR
Finger Flexion FlexiForce sensors
Sensors (FFS) We propose an RC electrical model for the FlexiForce
sensor model A201-100 (see Fig. 3b and 3c). It is possible to
determine that the RC model is a fit for the sensor through
many different ways; one of them is to study the frequency
response of the sensor and compare such results with
IMU
theoretical curves for a RC parallel circuit. The experiments
carried out on this article were done on eight sensors with the
aim of obtaining significant results.
Fig. 3a shows the circuit proposed by the manufacturer
[17] with little changes. The DC input source was replaced
by two selectable input sources: a DC voltage source and a
Force
Wireless
measurement
sinusoidal (AC) voltage source. The AC source will let us
communication
unit unit study the sensor frequency response, estimate sensor
capacitance, and also, to propose a multivariable estimation
(a) (b) of force, whereas, the DC source will be used to compare the
Fig. 1: Sketch of the haptic interface based in data gloves. (a) Dorsal and traditional conductance model with the new method
(b) palm side of the interface showing the most relevant components.
proposed. Input signals of Fig. 3a were chosen to be:
There are as many force vectors as force sensors are
Vs1 = −5V (1)
installed in the palm side of the glove. This procedure
generates a series of force vectors that are totaled to obtain a Vs 2 = AS sin ( 2π ft ) (2)
resulting force and consequently a resulting velocity, which If sensor model depicted in Fig. 3c is assumed and the
is proportional in magnitude and direction to the operator sinusoidal input is chosen (2), a linear equation for Fig. 3a
intention of motion. Also, it is necessary to translate the may be written as:
resulting force vector from the glove reference system to the Vs 2 dV V
robot reference system, so an IMU is installed just on top of + Cs s 2 = − o (3)
the glove dorsal side. Rs dt Rg
Since a high correspondence is desired between the forces where Rs and Cs are sensor’s resistance and capacitance
exerted by the operator and the consequent robot motion, respectively, Rg is the feedback resistor of the amplifier, and
every force contribution must be accurately measured in Vo is the output voltage. Solving (3) for sinusoidal input (2)
order to achieve high colinearity and a proportional response. gives us the following expression of Vo:
However, piezoresistive sensors such as FlexiForce sensors ⎛ sin ( 2π ft ) ⎞
are inaccurate, and therefore a new method for reducing Vo = - As Rg ⎜ + 2π fCs cos ( 2π ft ) ⎟ (4)
force estimation error is required. In this paper we developed ⎝ Rs ⎠
a new method to increase accuracy in force readings by Equation (4) may be written as a sine function with output
means of proposing an electrical model for the sensor, and amplitude (Ao) and phase shift ( φ ) as shown next:
developed a multivariable estimation of force.
Vo = Ao sin ( 2π ft + φ ) (5)
z Linking equations (4) and (5) give us a sensor’s
y capacitance expression which will be used later for
x
Interface Robot estimating applied forces to the sensor. An expression for
“Gloves-on
Payload”
sensor’s resistance can also be found.
A sin (φ ) Rg As
Cs = o (6) and Rs = (7)
Rg As 2π f Ao cos (φ )
On the other hand, in order to measure sensor’s
conductance it is necessary to choose the DC voltage (1) as
Fr Vr the input of circuit in Fig. 3a. Then, a linear-varying output
Load
Controller
Wireless
Receiver
Motor
Fig 2: Overview of the interface “Gloves-on Payload”. The operator exerts
force on the load, which is measured by the haptic glove and transmitted to (a) (b) (c)
the receiver. The controller indicates the motion to be performed by the Fig. 3: Piezoresistive sensor under study. (a) Sensor’s driving circuit (b)
robot according to the data received from the glove. Picture of an A201-100 FlexiForce sensor. (c) Sensor’s electrical model

459

Fig. 4: Typical variation of resistance and conductance, and a
trendline for an A201-100 FlexiForce Sensor (Image taken from a
Tekscan White Paper: Free Sensors for design. The image legend was
modified for better comprehension).
voltage (Vo_dc) is obtained as force increases. The following
equation describes how the output voltage changes. Note that
sensor’s conductance is multiplied by the feedback resistor
(Rg) and the input voltage (Vs1).
Vo _ dc = − ( Rg / Rs ) Vs1 (8)
Note that conductance readings may be done by applying
either a DC (1) or a sinusoidal input (2), but in order to
decouple capacitive effects from conductance readings, it is
necessary to drive the circuit of Fig. 3a with a DC input
voltage. Typical variation of conductance for a FlexiForce
sensor is shown in Fig. 4. The conductance-measuring
scheme is the method proposed by the manufacturer to
perform force readings, but as mentioned before, it lacks of
high repeatability.
With the aim of developing a multivariable estimation of
force it is necessary to study if the electrical model depicted
in Fig. 3c is a fit for the FlexiForce sensor, and also to study
how sensor’s capacitance changes as force increases. A
theoretical variation of phase shift can be found by linking
(6) and (7) which yields to:
φ = arctan(2π fCs Rs ) (9)
On the other hand, an expression that relates force
Fig. 5: Test Bench pictures. (a) Close up of the stand structure with
calibrated weights on top. (b) Complete view of the test bench. (c)
Microcontroller used for digitizing the amplitude and phase reading. (d)
Interleaved configuration to handle up to eight sensors simultaneously. (e)
Close up of the conditioning circuitry.
460
readings with sensor capacitance can be obtained only by
empiric ways just as it is for the conductance in Fig. 4.
IV. TEST BENCH DESCRIPTION
In order to validate the sensor’s model depicted in Fig. 3b.
and to study force effects on sensor’s capacitance, a test
bench was built according to Fig. 5. It could handle up to
eight sensors simultaneously by means of using an
interleaved configuration with pucks between adjacent
sensors. The puck’s weight was negligible due to the low
density of the material used. This was important, because we
wanted to load every sensor with the same force. Forces were
generated by placing calibrated weights on top of the pack of
sensors.
The test bench could apply a maximum force of 250 N to
the pack of sensors. This force is equal to half of the range of
the A201-100 FlexiForce sensor. The force step was 5 N, so
a total of 50 different forces were generated during the
experiment.
V. VALIDATION OF SENSOR’S MODEL AND
EXPERIMENTAL SET-UP
A. Validation of sensor’s model
In order to validate the RC parallel electrical model, we
chose the AC input source (2) as the input in the driving
circuit of Fig. 3a. We varied the input frequency f from
100Hz to 40Khz, which matches the highest frequency the
amplifier can deal with negligible phase shift. It is important
to keep the phase shift introduced by the amplifier to at least
as possible; otherwise, such phase shifts will add to those
introduced by sensor’s capacitance, leading to inaccurate
results.
On the other hand, it is necessary to choose a small
enough input amplitude, As, for the AC source (2) that does
not saturate the amplifier output voltage as frequency
increases. Notice that expression (4) states that as frequency
increase the same will do the output amplitude. Although, if
input amplitude is too small phase readings will be subject to
noise and data will be scattered.
Fig. 6: Phase bode plot of FlexiForce sensor for three different forces
applied of 50N, 125N and 250N. Experimental data are shown with
markers and the trendlines are superposed on each one.
 We believe that changing input amplitude in (2) as
frequency increases is not a good practice, because it is a
common source of step jumps in phase readings. So, we
chose 0.5V as the input voltage for the frequency sweep.
Figure 6 is a phase bode plot of one FlexiForce sensor for
three different forces applied of 50N, 125N and 250N.
Experimental data, shown with markers, are superposed over
the theoretical curve that we obtained using (9).
Sensor’s Resistance, Rs, and Capacitance, Cs, were
estimated at 2.5KHz and then substituted in (9) for obtaining
a trendline for the experimental data. We deliberated chose
2.5KHz as the frequency for estimating Rs and Cs because
this way the trendline fits better the experimental results for
the three forces applied. Although, very similar trendlines
were obtained when Rs and Cs were estimated at frequencies
between 1Khz and 4Khz.
A series of interesting facts may be taken out from Fig. 6.
First, under low frequency operation (below 1KHz)
experimental data are more scattered than at higher
frequencies. This may happen due to the fact that, under low
frequency operation the output voltage is too low, and then
phase readings are more noise-sensitive. Second, equation (9)
is a fit for the experimental data until reaching a certain
frequency, which we called the divergent frequency. As
frequency increases beyond the divergent point, experimental
phase readings saturate and both curves separate from each
other. This behavior in sensor’s response may be understood
as a frequency nonlinearity, where the RC model is no longer
valid, and thus, the equations stated in Section III do not
match for sensor’s response. Three, Divergence frequency
seems to be non-dependant of the applied force; only slight
variation of the divergent frequency is noticed as force
changes, although, the divergent frequency changes more
noticeable from one sensor to other starting at 4.5Khz for
some sensors and up to 7Khz for others.
We have demonstrated that sensor model introduced in
Section III matches for FlexiForce sensor response in a
defined range of frequencies. We are interested in
understanding why the sensor exhibits such nonlinear
response as frequency goes beyond 4.5KHz. In order to study
that, Figure 7 and 8 show Cs and Rs respectively, as a
function of frequency for the three same forces of 50N, 125N
and 250N. Sensor’s capacitance and resistance were
Fig. 7: Sensor’s Capacitance as a function of frequency for three different
applied forces of 50N, 125N and 250N.
461
estimated by performing phase, φ , and output amplitude
readings, Ao, and then using expressions (6) and (7) to obtain
Cs and Rs.
Figure 7 yields us to similar statements as previously did
Fig. 6: First, capacitance data are scattered at small frequency
values, we think this may be due to the same reason phase
values were scattered on Fig. 6 for low frequencies. Second,
capacitance values remain with little variation, for a given
force, despite frequency changes, whereas resistance values
in Fig. 8 decrease dramatically when the divergence
frequency is reached. This behavior explains why the phase
readings for the sensor saturate beyond the divergence
frequency, but actually, we can not figure out the physical
cause of such phenomenon.
On the other hand, Fig. 7 show that capacitance values
actually change as force increases. This means that the
sensors exhibited a piezocapacitive behavior; such property
had been unknown up to now, and only the piezoresistive
property of the sensor had been used for estimating forces
[4], [6], [9], [11], [17], [19]. Figure 9 shows the variation of
Cs as a function of force, note that capacitance variation is
kind of linear, but with little step jumps, just as the
conductance one in Fig. 4.
The main contribution of this paper consists in using this
piezocapacitive property to reduce force estimation errors,
given that additional information on the applied forces that
can be found in sensor capacitance.
B. Experimental set-up
We have found out that sensor capacitance remains with
little variation as frequency increases, although, under low
frequency operation (below 1 KHz) capacitance estimation
yields to scattered values, so, operating the sensor at low
frequency is not recommendable.
On the other hand, if input frequency is beyond the
divergence frequency, we will operate the sensor in a region
where the RC model is not a fit. This is not a problem itself if
we want to estimate capacitance, because as shown before in
Fig. 7, variation of capacitance is small as frequency
increases. Also, we believe that such variation will not affect
our experiments because once we have chosen an input
frequency it will be the same for all experiments. But, in
order to operate the sensor within the RC fit-region, we
Fig. 8: Sensor Resistance as a function of frequency for three different
applied forces of 50N, 125N and 250N.
preferred to set input frequency, f, to 4KHz. Input amplitude,
As, was set to 3V, but this value is, according to (4), a
compromise between Rg and As, the higher Rg the lower As.
So the input signal (2) was set for further experiments to:
Vac = 3V sin ( 2π 4000t ) (10)
The DC input signal was kept at the manufacturer
recommended value previously stated in (1).

VI. MULTIVARIABLE SENSOR MODELING

We developed four empirical models based upon
conductance and capacitance changes. The first one is the
traditional conductance model obtained when Vs1 (1) is
selected as the input in the circuit of Fig. 3a. This method Fig. 10: Mean Squared Error of the four models proposed for the eight
will be used as a reference for further comparison. The sensors under study.
second model is a linear regression of the capacitance values,
from the Vo and Cs values. Conversely, the other three
so input Vac (10) is chosen as the input in the circuit of Fig. 7.
methods only take into account one of the two variables.
The third model consists of averaging the output forces
Averaging the predicted forces from the conductance and
predicted by the first and second models. The fourth and final
capacitance models seldom produced better results than the
model is a feedforward neural network with two inputs (Vo
neural network model (which was the case with sensor 1, but
and Cs), one hidden layer with two neurons, and one neuron
not the other sensors).
output. The same network topology was used for all sensors,
A 3D plot of the surface generated by the neural network
but training data was taken separately for each one; this was
for a given sensor is depicted in Fig. 11. An arrow
necessary because capacitance is slightly different from one
superimposed on the surface indicates the typical variation of
sensor to other. The neural network was trained offline and
Vo and Cs as force increases. When either Vo or Cs values
then tested with new data.
deviate from the ideal trajectory described by the white
As mentioned before, we want to decouple conductance
arrow, force is estimated incorrectly. However, the soft
readings from capacitance effects, so, Vo values were
surface generated by the neural network tends to mitigate this
obtained for the models: one, three and four by setting (1) as
error and improve sensor response.
the input in the circuit of Fig. 3a.
Finally, the traditional conductance model produced output
The histogram in Fig. 10 summarizes the experimental
errors that were notably different from one sensor to another.
results for each sensor in terms of the mean squared error for
These imbalances are a frequent source of problems in
the four models under study.
applications in which several sensors are used [6]. Table 1
An important set of facts can be concluded from Fig. 10.
summarizes the improvements introduced by the neural
First, the capacitance model generated lower errors than the
network, compared with the traditional conductance model
traditional conductance model for all sensors. This is an
for the eight sensors under study.
important fact, because it shows that capacitance readings are
Note that the neural network model reduces MSE
more repeatable than conductance measurements.
dispersion to a narrow range of 0.107 – 0.515, whereas the
Second, averaging the predicted forces from the
dispersion of the traditional conductance model is noticeable
conductance and capacitance models sometimes reduced the
higher 0.258- 2.01. The percentage of error reduction is
output errors; this was true for sensors 1, 2, and 3, but not
computed (PER) from:
true for sensors 4 to 8.
Third, the neural network model was the best technique for ⎛ MSE neural network model ⎞
PER = ⎜ 1 − ⎟ (11)
reducing output errors. This method worked better than the ⎝ MSE traditional conductance model ⎠
others due to the fact that a two-variable function is built

300
200
250

200
Force (N)

150
150

100
100
50

0
800 50
600 4
400 3
2
Cs (pF) 200 1 0
Vo (V)
0 0
Fig. 11: 3D plot of the surface generated by the neural network for
predicting forces applied to FlexiForce sensors. An ideal variation of Vo
Fig. 9: Capacitance as a function of Force. and Cs as force increases is indicated by a white arrow.
462
 A high value of PER for a given sensor means that the [9] M.C.F. Castro, A. Cliquet, Jr. “A Low-Cost Instrumented Glove for
Monitoring Forces During Object Manipulation”, in IEEE Transactions
neural network model has substantially reduced estimation
on Rehabilitation Engineering, vol. 5, pp. 140-147, June 1997
error compared with the traditional conductance model. An [10] J-H. Lee, Y-S Lee, S-H Park, M-C Park, B-K Yoo, S-M In, “A Study
average of these values for the eight sensors under study was on the Human Grip Force Distribution on the Cylindrical Handle by
also computed, resulting equal to 64%. Intelligent Force Glove (I-Force Glove)”, in International Conference
on Control, Automation and Systems.(Seoul, Korea), pp. 966-969,
October 2008.
VII. CONCLUSIONS AND FUTURE WORK [11] H. Kazerooni, D. Fairbanks, A. Chen, G. Shin. “The Magic Glove”, in
An RC equivalent model was presented and validated for IEEE International Conference on Robotics and Automation, (New
Orleans, LA), pp. 757-763, April 2004
the FlexiForce sensor model A201-100. Also, we [12] L. Dipietro, A. M. Sabatini, P. Dario, “A Survey of Glove-Based
demonstrated that the sensor has a piezocapacitive property Systems and their Applications”, in IEEE Transactions on Systems,
which is useful to increase sensor’s accuracy by means of Man, and Cybernetics, vol. 38, pp. 461-482. July 2008.
performing a two-variable estimation of force. [13] J. H. Ahroni, E. J. Boyko, R. Forsberg, “Reliability of F-Scan In-Shoe
Measurements of Plantar Pressure” in Foot and Ankle International,
Future work will focus on testing the two-variable 9,10 pp. 668-673. October 1998
estimation of force in a force control application, also, future [14] R. Luo, “A Microcomputer-Based intelligent sensor for Multiaxis
work will focus on proposing better and more accurate sensor Force/Torque Measurement”, in IEEE Transactions on Industrial
models that help to reduce force estimation errors. Electronics, vol. 35 pp. 26-30. February 1998.
[15] D. Chapuis, R. Gassert, L. Sache, E. Burdet, H. Bleuler, “Design of a
simple mri/fmri compatible force/torque sensor”, in IEEE International
TABLE I Conference on Intelligent Robots and Systems, vol. 3, (Sendai, Japan),
COMPARISON OF TRADITIONAL CONDUCTANCE MODEL WITH NEURAL pp. 2593-2599, September 2004.
NETWORK MODEL IN TERMS OF MEAN SQUARED ERROR (MSE) [16] J. L. Pavlovic, Y. Takahashi, J. E. Bechtold, R. B. Gustilo, and
Sensor 1 2 3 4 5 6 7 8 R.J. Kyle, “Can The Tekscan Sensor Accurately Measure Dynamic
MSE of Pressures In The Knee Joint?”, in 17th Annual Meeting, American
traditional Society of Biomechanics, October 1991.
0.547 0.258 0.285 1.06 2.01 0.527 1.17 0.839
conductance [17] Tekscan Inc, FlexiForce User Manual, Available in:
model
http://www.tekscan.com/pdfs/FlexiforceUserManual.pdf, Sept. 2009.
MSE of neural [18] Interlink Electronics, Standard Specification Sensors, Available in:
0.330 0.107 0.156 0.118 0.374 0.145 0.515 0.249
network model http://www.interlinkelectronics.com/force_sensors/products/forcesensi
Percentage of
ngresistors/standardsensors.html?specs=1, Sept. 2009.
error reduction 39.7 58.2 44.9 88.8 81.4 72.4 55.9 70.2 [19] F. Vecchi, C. Freschi, S. Micera, A. Sabatini, and P. Dario,
(%) “Experimental evaluation of two commercial force sensors for
Average Percentage of error reduction 64% applications in biomechanics and motor control.” In International
Functional Electrical Stimulation Society (IFESS), (Aalborg,
Denmark), June 2000.
VIII. ACKNOWLEDGMENTS [20] S. Eppinger and W. Seering, “On Dynamic Models of Robot Force
Control”, MIT Internal Report, A.I. Memo No. 910, 1986.
This work has been funded by the Spanish Ministry of [21] L. Paredes and P. Gonzalez-de-Santos, “System and procedure for
Science and Innovation through grant DPI2007-65728 and controlling manipulators”, Patent Nº P200930173, Spanish Patents
AECID through grant PCI-Iberoamerica D/026706/09. Office, 2009.

REFERENCES
[1] M. A. Peshkin, J. E. Colgate, W. Wannasuphoprasit, C.A. Moore, R.B.
Gillespie, P. Akella. “Cobot Architecture”, in IEEE Transactions on
Robotics and Automation, vol. 17, pp. 377-390, August 2001.
[2] M. Van Damme, F. Daerden, D. Lefeber. “A Pneumatic Manipulator
used in Direct Contact with an Operator”, in IEEE International
Conference on Robotics and Automation, (Barcelona, Spain), pp.4494-
4499, April 2005.
[3] J.E. Colgate, M.A. Peshkin, J. Santos-Munné, A. Makhlin, P.F. Decker,
S.H. Klostermeyer. “Control Handle for Intelligent Assist Devices”,
U.S. Patent No. 6,738,691, May 2004.
[4] C. Lebossé, B. Bayle, M. de Mathelin, P. Renaud. “Nonlinear modeling
of low cost force sensors”, in IEEE International Conference on
Robotics and Automation, (Pasadena, CA, USA), pp. 3437-3442, May
2008.
[5] D. McGee, P. Swanson. “Method of Controlling an Intelligent Assist
Device”, U.S. Patent No. 6,204,620, March 2001
[6] M. Monroy, M. Ferre, J. Barrio, V. Eslava, I. Galiana. “Sensorized
Thimble for Haptics Applications”, in IEEE International Conference
on Mechatronics. (Málaga, Spain), pp.1-6, April 2009.
[7] Z. Ye, G. Auner. “Haptic Interface Prototype for Feedback Control on
Robotic Integration of Smart Sensors”, in IEEE International
Conference on Control Application, pp. 995-1000, 2003.
[8] K. N. Tarchanidis, J.N. Lyngouras, “Data Glove With a Force Sensor”,
in IEEE Transactions on Instrumentation and Measurement, vol. 52,
pp. 984-989, June 2003

463

View publication stats