TACTILE TEXTURE DISCRIMINATION IN THE ROBOT-RAT

PSIKHARPAX

Steve N’Guyen, Patrick Pirim

Institut des Syst

emes Intelligents et de Robotique, Universit

e Pierre et Marie Curie, Paris 6, France

Brain Vision Systems, Paris, France

Jean-Arcady Meyer

Institut des Syst

emes Intelligents et de Robotique, Universit

e Pierre et Marie Curie, Paris 6, France

Keywords:

Tactile perception, Whiskers, Texture discrimination, Kinetic signature, Resonance, Robot.

Abstract:

We endowed a whiskered robot with a simple algorithm allowing to discriminate textures. Its efﬁciency and

robustness have been demonstrated using both a ﬁxed head and a mobile platform. Comparatively to previous

similar approaches, this system affords greater behavioral capacities and proves to be able to complement or

supply vision in simple navigation tasks. The corresponding results suggest that the length and number of

the whiskers involved play a role in texture discrimination. They also suggest that two hypotheses that are

currently considered as mutually exclusive to explain texture recognition in rats - i.e., the “kinetic signature

hypothesis” and the “resonance hypothesis” - may be, in fact, complementary.

1 INTRODUCTION

Touch is a very important sensory modality for many

species of insects and mammals. For example, the

whiskers of a rat are often compared to human ﬁn-

gertips in terms of their tactile - or haptic - ability.

In particular, they make it possible to ﬁnely discrimi-

nate textures (Carvell and Simons, 1990; Guic-Robles

et al., 1989) or objects (Brecht et al., 1997) and

even to precisely determine an aperture width (Krupa

et al., 2001). Biologists have studied rat whiskers for

decades and know quite precisely the pathway from

an individual vibrissa to the somatosensory cortex.

One remarkable property of this haptic system is that

whiskers project somatotopically to this part of the

cortex, into a structure named “barrel cortex”. A “bar-

rel” is a discrete neural structure that receives an input

principally from a given whisker, with little inﬂuence

from neighboring whiskers (Petersen and Diamond,

2000). This relatively simple system, as compared to

vision for example, facilitates the study of the neural

coding scheme, as well as its use for perception and

higher-level cognition.

Being simple, efﬁcient and robust, whiskers

should become popular in robotics (Hartmann,

2001) although few robots have been equipped

with such devices in the past. The corresponding

implementations were calling on various sensors

ranging from the simplest binary switch to a very

accurate bi-dimensional torque sensor. Brooks

(1989), for example, used a simple sensor made of a

metal shaft ﬁxed on a push button, providing a very

robust security sensor for a walking robot. Another

implementation called upon probe whiskers made of

a stem glued to a potentiometer with return springs

and was used to evaluate the contour of an object

(Russell, 1985). Even wind sensitive sensors have

been designed (Chapman et al., 2000) allowing a

robot to navigate through a labyrinth. Basically, this

sensor was made of small springs surrounded by

electric contacts and was able to detect the direction

of the wind.

Recently, several artiﬁcial whisker systems

have been used in robotics to discriminate textures.

Whisker hairs of real rats, glued to capacitive sensors

(electret microphone), served (Fend et al., 2003;

Lungarella et al., 2002) to produce very precise

haptic sensors, with an uni-dimensional measurement

of dynamic signals. Using an active whisker array

of such sensors mounted on a mobile robot, Fend

et al. (2003) successfully discriminated a set of 11

N’Guyen S., Pirim P. and Meyer J. (2010).

TACTILE TEXTURE DISCRIMINATION IN THE ROBOT-RAT PSIKHARPAX.

In Proceedings of the Third International Conference on Bio-inspired Systems and Signal Processing, pages 74-81

DOI: 10.5220/0002730200740081

 SciTePress

textures. Kim and M

oller (2004) tried both piezo

and hall-effect sensors which, mounted in orthogonal

pairs, provided a bi-dimensional measure of vibrissa

deﬂection. Like capacitive sensors, piezo sensors

cannot deliver static signals, but this can be achieved

using an extra integrator circuit. With a data process-

ing based on spectrum density, these authors were

able to discriminate a set of 7 sandpapers. Likewise,

Seth et al. (2004) performed texture discrimination

using arrays of Flex sensors, which provided an

unidimensional measure of curvature. Here, temporal

differences between pairs of vibrissæ were fed into

a barreloid system with spiking neurons. Finally,

Fox et al. (2009) used two active whiskers with

strain gage-based sensors mounted on a mobile

robot. They explored different bioinspired methods

of feature extraction and the implication of uncon-

strained whisker-texture contact on classiﬁcation

performance.

The work described herein contributes to the

Psikharpax project (Meyer et al., 2005) which aims at

designing a biomimetic artiﬁcial rat. Besides visual,

auditory and vestibular sensors, the corresponding

robotic platform is equipped with an original whisker

system described elsewhere (N’Guyen et al., 2009).

This system is intended to be used for texture discrim-

ination and object recognition and, more generally, as

an auxiliary or a substitute to vision. Its performance

in texture discrimination is the subject of the present

article.

2 SYSTEM DESCRIPTION

Insofar as the impact of the speciﬁc implementa-

tion of a rat’s whisker system on its functionalities

is currently unknown, we tried to design an artiﬁ-

cial whisker system mimicking as much as possi-

ble the natural organization. Accordingly, our sys-

Figure 1: Comparison of whisker pads in a real rat and in

our robot.

tem (N’Guyen et al., 2009) is based on a simple,

elastomer-based, artiﬁcial skin with two arrays of 33

vibrissæ and an arc/row organization (cf. Fig 1) and a

Table 1: Vibrissæ arcs, with mean measured lengths in mm

(in one adult rattus norvegicus specimen). Compared to

those of a real rat, the robot’s whiskers are approximately

6 times longer.

Arc vibrissæ rat robot

1 α, β, γ, δ,E1 41.8 250

2 A1, B1, C1, D1, E2 37.2 200

3 A2, B2, C2, D2, E3 27.6 150

4 A3, B3, C3, D3, E4 20.6 120

5 A4, B4, C4, D4, E5 12.6 100

6 C5, D5, E6 8.33 90

7 C6, D6, E7 70

8 D7, E8 55

whisker-length gradient (cf. Table 1) quite similar to

those encountered in the rat.

The deﬂection of each vibrissa is sampled in

both its anteroposterior and dorsoventral axes, pro-

viding two 8-bit measurements at 1157Hz. How-

ever, orientation information being not necessary for

texture discrimination, the two measures are normed

(

√

x2 + y2).

3 TEXTURE DISCRIMINATION

3.1 Feature Extraction

Neither the details of how a rat’s brain actually

encodes texture features, nor the exact nature of

these features, are yet known. Arabzadeh et al.

(2004) experienced different feature codings on both

artiﬁcial and natural (in vivo) whiskers. Starting from

the principle that a pure sinusoidal signal can be fully

described by its amplitude A and its frequency f ,

they stimulated a rat’s whiskers with various signals

varying in amplitude and frequency. Then, recording

the induced neural activity in the barrel cortex,

they deduced that the neural activity most probably

encodes a quantity homogeneous to the product A f .

The generalized expression of this quantity to any

natural signal is known as the “equivalent noise level

(ENL)” (for more details see: (Arabzadeh et al.,

2005)), which is usually used to measure sound

power. This quantity can also be related to the more

common ”spectral centro

ıd” (Fox et al., 2009).

To compute the latter, instead of using a Fast Fourier

Transform algorithm - of which no natural equivalent

is known - we simply called upon a time domain

“on-the-ﬂy” estimate of the quantity Xω. The corre-

sponding algorithm (cf. Fig 2) can be compared to

those used in auditory feature extraction, like ZCPA

that is used for speech recognition (Ghitza, 1994;

Kim et al., 1999; Sreenivas and Niederjohn, 1992). It

TACTILE TEXTURE DISCRIMINATION IN THE ROBOT-RAT PSIKHARPAX

Figure 2: Feature extraction algorithm. ”Peaks” are de-

tected through the monitoring of the signal’s derivative and

frequencies are estimated through the inverses of the time

intervals between successive peaks. Then, the peak ampli-

tude is multiplied by the peak frequency, averaged within a

time window.

provides a quantity homogeneous to the ENL which

we call the “Instantaneous Mean Power” or IMP

feature.

This approach relies on the strong hypothesis that

the peaks thus characterized provide enough informa-

tion to describe a texture. Such hypothesis is rein-

forced by the fact that, when Licklider and Pollack

(1948) assessed the effects of various signal distor-

tions in human speech recognition, they found that

“inﬁnite clipping” - a treatment that only kept a sig-

nal’s periodicity - did not prevent speech recognition

in humans.

Be that as it may, the corresponding algorithm

is very simple and computationally very cheap as

it necessitates only one division per peak detection

(

Peak amplitude

Peak period

) plus one addition (to compute the

peak’s period) at each time step. As for peak detec-

tion proper, it only entails one subtraction (x

−x

t−1

)

and a comparison.

There are however possible limitations to the pro-

posed algorithm. In particular, input data are drasti-

cally reduced by this procedure according to which

a pure sinus input of frequency F and a triangle in-

put of fondamental frequency F will lead to the same

feature value although they obviously don’t have the

same spectrum. Likewise, turns out that amplitude

modulations cannot be detected by a single vibrissa.

Our hypothesis is that such limitations are alleviated

by the fact that the natural ﬁltering of vibrissæ, due to

their intrinsic mechanical characteristics, decomposes

complex signals along the pad in a manner similar to

how the cochlea decomposes auditory signals.

3.2 Fixed Head Experiment

3.2.1 Experimental Appartus

At ﬁrst, we tested this haptic system according to a

relatively constrained ﬁxed head experiment that con-

sisted in sweeping a whisker pad over a set of eight

sandpapers whose grits varied from P180 to P50 (cf.

Fig 3). Sandpaper provide a complex random texture

appropriate for this task and has been used on vari-

ous experiment with real rats. Using this material, we

performed qualitative experiments with humans that

clearly indicated that the task of discriminating such

textures by tactile contact only is a very difﬁcult one,

an observation also made by Hipp and coll. (2005) .

Figure 3: The texture set used for discrimination.

A vibrissa pad was ﬁxed on the robot’s head which

could move in pan-tilt directions. The pan axis was at

a ﬁxed distance from the texture sample (cf. Fig 4)

that was presented with a small amount of variabil-

ity in position (∼ ±1cm) between each trial, with an

appropriate angle to provide contact with a maximum

number of whiskers.

For each texture, 400 experiments were made, 300

for learning and 100 for testing. The raw data (x and

y deﬂections, 8 bits resolution sampled at 1157Hz)

were normed (

√

x2 + y2). For each vibrissa, this mea-

sure was fed into the feature extraction algorithm that

output the IMP as one ﬂoat value. Finally we summed

these IMP values for each vibrissa during the sweep.

Having thus obtained an input vector of 33 ﬂoats for

each trial, we fed it into a simple multi layer percep-

tron (MLP) with 33 input neurons, 33 hidden layer

neurons and 8 output neurons, to perform supervised

learning. We used the FANN library (Nissen, 2003)

with the iRPROP training algorithm (Igel and H

uskel,

2000). The ﬁnal classiﬁcation was done by a winner-

take-all (WTA) on the 8 output neurons.

BIOSIGNALS 2010 - International Conference on Bio-inspired Systems and Signal Processing

Figure 4: Schematic of the experimental protocol. A: start

position, B: mid position, C: end position.

Figure 5: Top view of the experimental protocol.

3.2.2 Results

Table 2 gives the confusion matrices obtained on 100

test data. The mean performance is clearly above 90%

(here the chance level is

= 12.5%), which greatly

improves the human aptitude at solving the same task.

Table 2: Confusion matrix obtained for the 8 textures using

IMP.

IMP 1 2 3 4 5 6 7 8

1 100 0 0 0 0 0 0 0

2 0 99 1 0 0 0 0 0

3 0 2 95 0 3 0 0 0

4 0 0 0 96 4 0 0 0

5 0 0 1 0 99 0 0 0

6 0 0 0 0 1 93 6 0

7 2 0 6 0 3 9 80 0

8 2 0 0 0 0 1 0 97

Using the data thus acquired, we tested the inﬂu-

ence of the number of vibrissæ on the classiﬁcation

performance. Starting with data obtained with one

arc (Arc 1, 5 vibrissæ cf. Table 1), then with two

arcs (Arc 1 + Arc 2, 10 vibrissæ) etc, up to the whole

whisker pad, we assessed at each stage the quality

of the discrimination. Results are summarized on

Figure 6. The percentage of successful discrimina-

tions is quickly rising with the number of vibrissae

involved and reaches values comprised between 90

and 95% when three or more arcs (15 vibrissæ) are

concerned. This result conﬁrms previously obtained

ones in (Fend et al., 2003; Hipp et al., 2006).

Figure 6: Mean performance (% of successful discrimina-

tions) obtained with IMP, over the number of vibrissæ in-

volved.

Figure 7: Mean performance obtained with IMP for the 5

longer arcs across the 8 textures.

When analyzing the performances obtained with a

single arc (cf. Fig 7), one observes a great variability.

Indeed, it seems that each arc separately performs bet-

ter on a subset of the textures. For example, arc 2 is

very bad at recognizing texture 5, but quite good with

texture 6. This suggests that iso-length arcs comple-

ment each other and probably explains the increase in

performance with the number of arcs involved.

The relative quality of these results demonstrates

that the IMP is a suitable feature for texture recog-

nition. However, as Fox et al. (2009) pointed out,

TACTILE TEXTURE DISCRIMINATION IN THE ROBOT-RAT PSIKHARPAX

the kind of ﬁxed head task used so far is very differ-

ent from that of a robot moving in an environment,

where the distances and angles with which whiskers

touch any texture are constantly varying. Therefore,

to assess the robustness of the IMP, we also performed

such a complementary experiment.

3.3 Mobile Robot Experimentation

3.3.1 Experimental Appartus

In this experiment, a set of complex textures (cf. Fig

8) made of plexiglass were ﬁxed on the sides of two

small corridors (1m long). A different texture was

Figure 8: The four textures used in the mobile robot exper-

iment. These textures were made of relief decorated plexi-

glass.

assigned to each side of each corridor. The robot’s

task was to follow the walls in its environment, to en-

ter a corridor if encountered, to recognize the textures

on its sides, and to learn to turn left or right at the

end of the corridor, depending on the left/right com-

bination of the textures thus recognized. A main dif-

ference with the previous experiment was the “touch

strategy”. We previously swept whiskers on a tex-

ture by rotating the head, trying to maximize the num-

ber of whiskers in contact with the texture. Here, the

whole robot was moving, the head didn’t rotate and

only a subset (∼ 10 vibrissæ, the two longer arcs) of

whiskers were actually touching the textures, from a

varying distance.

To allow the robot to navigate in its environment

using only its whiskers as sensory input, we devel-

oped a simple obstacle avoidance strategy. A distance

information was ﬁrst computed by taking into account

the iso-length arcs. One minus the mean arc deﬂec-

tion was weighted by the mean within-arc vibrissa

size. Thus, the more vibrissæ were bent, the smaller

was the output distance. Repeating this computation

for each arc, we obtained a value that decreased with

the contact distance,

D =

∑

(1 −V

) ×L

(1)

being the mean deﬂection of the i

arc and L

the

mean length of the i

arc. One can notice that the

smaller whiskers - the most frontal ones - contribute

Figure 9: Robot environment showing the 2 corridors and

direction convention used (Top). Robot inside a corridor,

with dimensions and command vector (Bottom).

less to this measure than the longer ones. This may

seem counter intuitive as, when an object touches

the small whiskers, it is probably closer than if it

only touches the longer ones. But generally in the

described task, when an object touches the smaller

whiskers, it also touches the longer ones and the

above weighting prevents an over-reaction. Moreover

this method has shown a better stability in corridors,

where a small variation of vibrissa deﬂection should

provoke a small orientation reaction in order to make

the corresponding trajectory as straight as possible.

This centering strategy was an important component

of the robot’s behavior since the corridors were 35cm

wide, while the robot’s width was 25cm (including

wheels) and the maximum whisker range was 50cm,

leaving a small error margin for steady forward move-

ment and whisker crack avoidance (cf. Fig 9). We

controlled the robot through a speed vector V applied

at the axis of the neck whose orientation component

was given by:

= (D

le f t

−D

right

) ×G (2)

with the gain G = 0.01. V

, the translation speed,

BIOSIGNALS 2010 - International Conference on Bio-inspired Systems and Signal Processing

Table 3: Confusion matrix obtained for 20 runs for each

corridor, in each direction.

Corridor-direction 1-1 1-2 2-1 2-2

1-1 75 15 10 0

1-2 0 100 0 0

2-1 15 0 85 0

2-2 0 15 0 85

was ﬁxed to 10cm/s. This simple control produced an

obstacle avoidance behavior, with a tendency to wall

following. Additionally, this control produced a rela-

tively stable corridor centering behavior - which was

its principal objective. Finally, using D

le f t

and D

right

values, we could roughly determine the corridor’s

aperture size and trigger the learning/recognition pro-

cedure only when the robot was inside a corridor as

determined by a distance threshold.

We ﬁrst ran a series of 10 experiments for each

corridor and each direction. We simply positioned

the robot approximately in front of the corridor and

recorded the IMP feature output at each time step

within the corridor. We when fed 7 data runs to a

MLP (2 ×33 neurons in the input layer, 2 ×33 neu-

rons in the hidden layer and 4 neurons in the output

layer), keeping the 3 others runs to test the learning

result. A typical data run length was ∼ 7000.

3.3.2 Results

Once the learning was completed, we ran four series

of 20 additional experiments to evaluate the capacity

of the robot to turn in the right direction at each end

of each corridor. While the robot moved in a given

corridor, in a given direction, we fed the smoothed

(low pass ﬁlter) 2 ×33 IMP output to the previously

learned MLP and computed the WTA on the output

layer. By accumulating this winner value through the

whole corridor, we obtained a mean decision vector

which served to take the ﬁnal decision (once again by

a WTA). The corresponding results are summarized in

Table 3. As expected, the trajectory stability played a

role in performances as dithering in the corridor in-

duced variations in the perceived vibrations. Most of

the errors occurred when the robot’s trajectories were

unstable (lot of dithering).

We ﬁnally conducted qualitative experiments in

the whole maze using the above described naviga-

tion rules. The maze was a round corner rectangle

of 2.20m by 4.10m made of cardboard boxes with 2

corridors (cf. Fig 9). We added a simple hand ca-

bled behavior consisting of turning left or right at the

end of a corridor, depending on the recognized tex-

tures. The robot was initially positioned near the wall

on the top of the maze with left or right orientations.

Any other starting position could have been used with

the limitation of avoiding a direct wall facing, as no

“reversal” behavior was implemented. In these condi-

tions, the robot succeeded to autonomously circulate

around the maze, following either direction indicated

by the textures on the corridor’s sides. Several tours

could be completed in a row thus demonstrating the

efﬁciency of the robot’s haptic system.

Figure 10 shows an example of the kind of trajec-

tories obtained.

Figure 10: Typical trajectory of the robot into the maze.

Top: left oriented start. Bottom: right oriented start.

4 DISCUSSION

If some research efforts have been devoted to tex-

ture discrimination in “ﬁxed head” tasks (Fend et al.,

2003; Lungarella et al., 2002; Kim, 2004; Fox et al.,

2009), very few robots have been able to navigate and

recognize tactile cues in a less constrained environ-

ment using whiskers. One such work was done using

curvature sensors with two different types of surfaces

that one may consider more as a “shapes” than as a

“textures”, as they seem to induce a mere deﬂection

sequence rather than a complex vibration (Seth et al.,

2004). This robot could be conditioned to associate an

aversive response with a given texture. A related work

concerned a smooth versus rough discrimination task

in an open arena and involved an active microphone-

based whisker sensor with a natural rat’s hair (Fend,

TACTILE TEXTURE DISCRIMINATION IN THE ROBOT-RAT PSIKHARPAX

2005). Feature extraction called upon spectral analy-

sis and lead to qualitatively good results. However, as

the author concludes, such system could not be used

to perform a more complex task without an improve-

ment of its discriminatory capability and reliability.

Finally, Fox et al. (2009) also obtained good results in

a smooth/rough discrimination task on a mobile robot

equipped with active whiskers using an “onset” fea-

ture. This “onset” feature is roughly the FFT magni-

tude within a short frequency band (2-3kHz) during

the onset period of the whisker-texture contact (the

ﬁrst 12.8ms of the contact). Moreover, this feature

is invariant to the relative position and orientation of

whiskers and textures. Experimental conditions were

slightly different from ours, as the texture position

was chosen randomly and the robot didn’t move while

touching a texture.

None of these related approaches seems suitable for

performing a more complex task than simply discrim-

inating smooth versus rough textures. In contrast, the

haptic system that has been described herein proved

to be able to use texture discrimination to afford min-

imal navigation capacities in a complex environment.

Such capacities could be used to complement vision

in daylight conditions or to replace it in the dark.

With this haptic system, texture recognition is

possible in both ﬁxed and mobile robot conditions.

This tends to indicate that, despite the underlying

simple algorithm and the various approximations on

which it relies, the IMP feature is robust.

Conversely, we already mentioned that the

whisker orientations in our system is not always well

suited. Indeed, our whiskers are oriented toward the

front (cf. Fig 5), which occasionally prevents all the

whiskers from touching a texture. Within a corridor,

for instance, about 10 whiskers only were touching

the walls. Additionally, our implementation some-

times entails brusque return jumps of some whiskers

when they are stuck on a given surface, rather than

a gentle sweeping, which makes their signals totally

unreliable. Fortunately, this problem only occurs in

corridors and with a minority of whiskers (usually

the more dorsal and ventral ones) and thus the clas-

siﬁer can see it as mere noise. Obviously, a system in

which the whisker orientation could be dynamically

controlled - such as the one used in (Fox et al., 2009)

- would be more adapted to alleviate this speciﬁc in-

convenience and would be closer to the natural active

whiskering system of rats.

Another remark concerns our feature extraction

technique. We chose to design an algorithm that ex-

tracts an estimation of the amplitude-frequency prod-

uct. This choice was based on a recent ﬁnding

about how texture signals are encoded in a rat’s brain

(Arabzadeh et al., 2004). Using such a feature, we

were able to perform ﬁne texture discrimination. This

ﬁnding is an argument in favor of the so-called “ki-

netic signature” hypothesis which stands that each

vibrissa encodes a speciﬁc signature of the touched

surface in term of magnitude and temporal pattern.

Likewise, the fact that our results suggest that the

texture discrimination capacities depend both on the

length and number of the involved whiskers, seems to

back up the “resonance hypothesis” (Moore and An-

dermann, 2005; Neimark et al., 2003) which stands

that the self resonance property of a vibrissa plays

a crucial role in vibration transduction and, in some

way, helps to enhance texture perception. The exact

manner in which this resonance property is exploited

in rats is still unclear, but it seems quite reasonable to

think that a kind of signal ﬁltering is involved. Addi-

tional experiments with the current system might help

clarify this issue.

Be that as it may, already acquired results strongly

suggest that two hypotheses that are currently consid-

ered as mutually exclusive to explain texture recog-

nition in rats - i.e., the “kinetic signature hypothe-

sis” and the “resonance hypothesis” - may be, in fact,

complementary.

5 CONCLUSIONS

We endowed a whiskered robot with a simple algo-

rithm allowing to discriminate textures. Its efﬁciency

has been demonstrated using both a ﬁxed head and a

mobile robot. Comparatively to previous similar ap-

proaches, this system affords greater behavioral ca-

pacities and may complement or supply vision in sim-

ple navigation tasks. Future work will be devoted

to demonstrating its ability to perform shape recog-

nition. On a fundamental level, it will also be used to

investigate the inﬂuence of whiskers resonance prop-

erties on texture transduction.

REFERENCES

Arabzadeh, E., Panzeri, S., and Diamond, M. E. (2004).

Whisker Vibration Information Carried by Rat Barrel

Cortex Neurons. J. Neurosci., 24(26):6011–6020.

Arabzadeh, E., Zorzin, E., and Diamond, M. E. (2005).

Neuronal encoding of texture in the whisker sensory

pathway. PLoS Biol, 3(1):e17.

Brecht, M., Preilowski, B., and Merzenich, M. (1997).

Functional architecture of the mystacial vibrissae. Be-

havioural Brain Research, 84(1-2):81–97.

BIOSIGNALS 2010 - International Conference on Bio-inspired Systems and Signal Processing

Brooks, R. A. (1989). A robot that walks: Emergent be-

haviors from a carefully evolved network. Technical

Report AI MEMO 1091, MIT.

Carvell, G. and Simons, D. (1990). Biometric analyses of

vibrissal tactile discimination in the rat. Journal of

Neuroscience, 10(8):2638–2648.

Chapman, T., Hayes, A., and Tilden, T. (2000). Reactive

maze solving with a biologically-inspired wind sen-

sor. In J. Meyer, A. Berthoz, D. F., Roitblat, H., and

Wilson, S., editors, From Animals to Animats 6. Proc.

of the 6th Int. Conf. on Simulation of Adaptive Be-

havior, pages 81–87. MA: MIT PRESS. A Bradford

Book.

Fend, M. (2005). Whisker-based texture discrimination on

a mobile robot. Advances in Artiﬁcial Life - Proceed-

ings of the 8th European Conference on Artiﬁcial Life

(ECAL), pages 302–312.

Fend, M., Bovet, S., Yokoi, H., and Pfeifer, R. (2003). An

active artiﬁcial whisker array for texture discrimina-

tion. In Proceedings of the IEEE/RSJ International

Conference on Intelligent Robots and Systems (IROS),

volume II, pages 1044–1049.

Fox, C. W., Mitchinson, B., Pearson, M. J., Pipe, A. G.,

and Prescott, T. J. (2009). Contact type dependency

of texture classiﬁcation in a whiskered mobile robot.

Autonomous Robots. In press.

Ghitza, O. (1994). Auditory models and human perfor-

mance in tasks related to speech coding and speech

recognition. IEEE Transactions on Speech and Audio

Processing, 2(1):115–132.

Guic-Robles, E., Valdivieso, C., and Guarjardo, G. (1989).

Rats can learn a roughness discrimination using only

their vibrissal system. Behavioural Brain Research,

31(3):285–289.

Hartmann, M. J. (2001). Active sensing capabilities of the

rat whisker system. Autonomous Robots, 11:249–254.

Hipp, J., Arabzadeh, E., Zorzin, E., Conradt, J., Kayser,

C., Diamond, M. E., and Konig, P. (2006). Tex-

ture Signals in Whisker Vibrations. J Neurophysiol,

95(3):1792–1799.

Igel, C. and H

uskel, M. (2000). Improving the rprop learn-

ing algorithm. In Proceedings of the Second Interna-

tional Symposium on Neural Computation, NC2000,

pages 115–121.

Kim, D.-S., Lee, S.-Y., and Kil, R. M. (1999). Auditory

processing of speech signals for robust speech recog-

nition in real-world noisy environments. IEEE Trans-

actions on Speech and Audio Processing, 7(1):55–69.

Kim, D. and Moller, R. (2004). A biomimetic whisker for

texture discrimination and distance estimation. From

Animals to Animats 8, pages 140–149.

Krupa, D. J., Matell, M. S., Brisben, A. J., Oliviera, L. M.,

and Nicolelis, M. A. L. (2001). Behavioural prop-

erties of the trigeminal somatosensory system in rats

performing whisker-dependent tactile discrimination.

J. Neurosci., (21):5752–5763.

Licklider, J. C. R. and Pollack, I. (1948). Effect of differ-

entiation, integration, and inﬁnite peak clipping upon

the intelligibility of speech. Journal of the acoustical

society of america, 20(1):42–52.

Lungarella, M., Hafner, V., Pfeifer, R., and Yokoi, H.

(2002). Artiﬁcial whisker sensors in robotics. Intel-

ligent Robots and System, 2002. IEEE/RSJ Interna-

tional Conference on, 3:2931– 2936.

Meyer, J.-A., Guillot, A., Girard, B., Khamassi, M., Pirim,

P., and Berthoz, A. (2005). The psikharpax project:

Towards building an artiﬁcial rat. Robotics and Au-

tonomous Systems, 50(4):211–223.

Moore, C. I. and Andermann, M. L. (2005). The Vibrissa

Resonance Hypothesis, chapter 2, pages 21–60. CRC

Press.

Neimark, M. A., Andermann, M. L., Hopﬁeld, J. J., and

Moore, C. I. (2003). Vibrissa resonance as a transduc-

tion mechanism for tactile encoding. The Journal of

Neuroscience.

N’Guyen, S., Pirim, P., and Meyer, J.-A. (2009). Elastomer-

based tactile sensor array for the artiﬁcial rat

psikharpax. In ISEF 2009 - XIV International Sym-

posium on Electromagnetic Fields in Mechatronics,

Electrical and Electronic Engineering. In press.

Nissen, S. (2003). Implementation of a Fast Artiﬁcial Neu-

ral Network Library (fann). Report, Department of

Computer Science University of Copenhagen (DIKU),

31.

Petersen, R. S. and Diamond, M. E. (2000). Spatial-

Temporal Distribution of Whisker-Evoked Activity in

Rat Somatosensory Cortex and the Coding of Stimu-

lus Location. J. Neurosci., 20(16):6135–6143.

Russell, R. A. (1985). Object recognition using articu-

lated whisker probes. In Proc. 15th Int. Symp. Industr.

Robots., pages 605–612.

Seth, A. K., McKinstry, J. L., Edelman, G. M., and Krich-

mar, J. L. (2004). Spatiotemporal processing of

whisker input supports texture discrimination by a

brain-based device. In Schall, S., Ijspeert, A., Bil-

lard, A., Vijayakumar, S., Hallam, J., and Meyer, J.,

editors, From Animals to Animats 8. Proc. of the 8th

Int. Conf. on Simulation of Adaptive Behavior. MA:

MIT PRESS.

Sreenivas, T. V. and Niederjohn, R. J. (1992). Spectral anal-

ysis for formant frequency estimation in noise. IEEE

Transactions on Signal Processing, 40(2):282–293.

TACTILE TEXTURE DISCRIMINATION IN THE ROBOT-RAT PSIKHARPAX