Contour-Net

A Model for Tactile Contour-tracing and Shape-recognition

André Frank Krause

, Thierry Hoinville

1,2

, Nalin Harischandra

and Volker Dürr

1,2

Department of Biological Cybernetics, Bielefeld University, Bielefeld, Germany

Cognitive Interaction Technology - Centre of Excellence (CITEC), Bielefeld University, Bielefeld, Germany

Keywords:

Tactile Sensor, Contour Tracing, Shape Recognition, Artiﬁcial Neural Network.

Abstract:

We propose Contour-Net as a bio-inspired model for rhythmic movement control of a pair of insectoid feelers,

able to successively sample the contour of arbitrarily shaped objects. Initial object contact initiates a smooth

transition from a large-amplitude, low-frequency searching behaviour to a local, small-amplitude and high-

frequency sampling behaviour. Both behavioural states are deﬁned by the parameters of a Hopf Oscillator.

Subsequent contact signals trigger a 180

◦

phase-forwarding of the oscillator, resulting in repeated sampling of

the object. The local sampling behaviour effectively serves as a contour-tracing method with high robustness,

even for complicated shapes. Collected contour data points can be directly fed into an artiﬁcial neural network

to classify the shape of an object. Given a sufﬁciently large training dataset, tactile shape recognition can be

achieved in a position-, orientation- and size-invariant manner. Only minimal pre-processing (normalisation)

of contour data points is required.

1 INTRODUCTION

The tactile sense enables humans and animals to ac-

tively perceive their immediate surrounding through

direct physical contacts with an object (Lee and

Nicholls, 1999). In contrast to vision, direct tactile

sampling of an object allows to ’feel’ object proper-

ties like surface texture, chemical properties, temper-

ature, compliance and humidity, that are hard to ob-

tain otherwise (Lederman and Klatzky, 2009). The

sense of touch is independent of light conditions, and

works equally well night and day. Moreover, the di-

rect contact with an external object yields reliable dis-

tance information (Patanè et al., 2012). Therefore, the

tactile sense plays an important role throughout the

animal kingdom, especially in nocturnal species.

Several insect species, for example the honey bee

(Apis mellifera), the American cockroach (Periplan-

eta americana) and the Indian stick insect (Carau-

sius morosus) have become important model organ-

isms for the study of the sense of touch. Insects carry

a pair of antennae that are densely covered with sen-

sory hairs of different modalities (Staudacher et al.,

2005). Honeybees, for example, show a high con-

centration of tactile hairs at their antennal tip (Esslen

and Kaissling, 1976). Active tactile scanning of sur-

faces allows them to discriminate the micro texture of

ﬂowers (Kevan and Lane, 1985) or artiﬁcial gratings

(Erber et al., 1998).

An important constraint of the insect tactile sys-

tem is that antennae essentially are one-dimensional

structures that are incapable of providing a two-

dimensional image “at a glance”. Instead, antennae

need to be moved actively in order to sample infor-

mation from different locations within their working-

range. Active tactile sensing is of particular relevance

in near-range exploration. Many insects actively use

their antennae for obstacle localization, orientation

behaviour, pattern recognition, and even for commu-

nication (Staudacher et al., 2005). Similarly, mam-

mals like cats or rats use active whisker movements

to detect and scan objects in the vicinity of the body.

Insect antennae and mammal whiskers have in-

spired robotic research in the area of tactile sensors.

Early work by Kaneko et al. (1998) describes an arti-

ﬁcial antenna using a ﬂexible beam capable of detect-

ing 3D contact locations and surface properties. Rus-

sell and Wijaya (2003) applied an array of whiskers

that passively scan over an object to recognize its

shape using advanced pre-processing contact-points

and decision trees. In Solomon and Hartmann (2006),

robotic whisker arrays were used to generate 3D spa-

tial representations of the environment and extract ob-

ject shapes. Related work done by Kim and Möller

Frank Krause A., Hoinville T., Harischandra N. and Dürr V..

Contour-Net - A Model for Tactile Contour-tracing and Shape-recognition.

DOI: 10.5220/0004821700920101

In Proceedings of the 6th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2014), pages 92-101

ISBN: 978-989-758-016-1

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

hopf

oscillator

+180°

sampling or

searching ?

contact

scaling

Figure 1: Block diagram of Contour-Net. The binary contact signal (red) determines the frequency and amplitude of the Hopf

Oscillator. The discrete impulse block guarantees that the oscillator phase is forwarded once by 180

◦

, resulting in a movement

away from the object surface. The output of the oscillator is then scaled and drives a simulated or real robot antenna using

velocity control of antennal joints.

(2007) used a vertical whisker array to detect the ver-

tical shape and curvature of objects.

Stick insects continuously and rhythmically move

their antennae during locomotion (Dürr et al., 2001)

and during tactile probing of external objects (Krause

et al., 2013b). Upon antennal contact with an obsta-

cle, stick insects modulate both the frequency and the

amplitude of the rhythmic antennal movement, affect-

ing both antennal joints in a context-dependent man-

ner (Schütz and Dürr, 2011; Krause and Dürr, 2012).

For example, when climbing a square obstacle, they

show a contact-induced switch in behaviour from a

broad, almost elliptical searching pattern to a local

sampling pattern with higher frequency and lower am-

plitude (Krause and Dürr, 2012). This switch to a

local sampling strategy may be interpreted as an ef-

fort to gather more detailed spatial information close

to previous touch locations, effectively leading to the

sampling of an obstacle’s contour.

A fundamental concept explaining such rhythmic

movements in vertebrates and invertebrates is the cen-

tral pattern generator (CPG, for a review see Ijspeert

(2008)). The CPG activity is often modulated by sen-

sory input, proprioceptive input and descending sig-

nals from higher level brain centres. Antennal move-

ments in stick insects are assumed to be driven by

CPGs, because rhythmic activity can be evoked phar-

macologically (Krause et al., 2013b).

Here, we present a simple but effective, bio-

inspired, CPG-based model capturing the essence of

tactile sampling in stick insects. The model is able to

trace the contour of an obstacle using an actively mov-

able tactile probe. The CPG activity is modulated by

a single, binary sensor signal: if the tactile probe is in

contact with the obstacle or not.

2 CONTOUR-NET

The contour-tracing model presented here was coined

Contour-Net, hinting at its possible integration into

existing, modular architectures for simulation and

control of hexapod walking (Walknet, Schilling

et al. (2013)) and ant-inspired navigation (Navi-Net,

Hoinville et al. (2012)). The model captures the es-

sential characteristics of antennal tactile sampling in

stick insects: rhythmic searching movements using

an antenna with two revolute joints with strong cou-

pling and a contact-triggered switch to local sam-

pling. Such rhythmic movements can be obtained as

limit cycles of nonlinear dynamical systems, typically

systems of coupled nonlinear oscillators. In our case

the optimal choice is a “Hopf Oscillator”, because the

two state variables of the oscillator exhibit a ﬁxed

phase coupling and can directly drive the two joints

of an antenna.

2.1 Hopf Oscillator

The Hopf Oscillator is a dynamical system deﬁned

in Cartesian Space by the following differential equa-

tions:

˙x = γ



− r



x − ωy

˙y = γ



− r



y + ωx

with r =

+ y

. The amplitude of the oscillator

converges to µ, with γ deﬁning the speed of conver-

gence and ω setting the frequency of the limit cycle.

Further, the phase of the oscillator can be set with

x = cos(ϕ), y = sin(ϕ).

The Hopf Oscillator has several advantages: First,

it has a stable limit cycle behavior (i.e., perturba-

tions decay quickly) with a ﬁxed, non-drifting, 90

◦

phase relationship between the x and y components

Contour-Net-AModelforTactileContour-tracingandShape-recognition

+180°

A B

Figure 2: A: Core idea of Contour-Net. Each contact induces a 180

◦

phase forwarding of a circular movement. Combined

with velocity control of a tactile probe, the resulting change in movement velocity and direction causes the probe to “bounce

off” the surface at every contact event, resulting in a successive scan along the object’s shape. B: 2D simulation of contact-

triggered contour tracing. Black: star-shaped object. Blue: trajectory of the antennal tip. Red dots: contact locations.

of the oscillator. Second, it is simple and well de-

ﬁned in terms of amplitude, phase and frequency,

which can be adjusted independently from each other.

Because of these properties, smooth online modula-

tion of trajectories can be achieved through chang-

ing parameters of the system at run-time. More-

over, these properties help in entrainment of the CPG

rhythm through sensory feedback, e.g., when be-

ing coupled with a mechanical system (Righetti and

Ijspeert, 2006). Hopf CPG’s have been applied suc-

cessfully to biped (Buchli et al., 2005; Righetti and

Ijspeert, 2006) and quadruped (Brambilla et al., 2006;

Ijspeert et al., 2007) locomotion.

2.2 Contact-triggered Contour Tracing

The basic idea of the contour-tracing model, as illus-

trated in ﬁgure 2A, is that each contact event triggers

a phase shift in the cyclic sampling movement of the

antenna. Ideally, this phase shift should cause the an-

tenna to “bounce off” the object after each contact.

Using a Hopf Oscillator to control the antennal posi-

tion directly would require memorizing the location

of the centre of oscillation, and shifting it along the

object surface. The need for a memory structure can

be avoided if the Hopf Oscillator output is used to

set the antennal velocity rather than position. Figure

2B shows the “scan path” (in analogy to eye-tracking

scan paths) along a star-shaped 2D-object. The veloc-

ity commands applied to the antennal joints, α and β

are given by:

∆α = s

∆β = s

where s

and s

are scaling factors, setting the max-

imum action range of the antenna. Setting these fac-

tors to distinct values leads to ellipsoid trajectories as

found in stick insect antennal movements (Krause and

Dürr, 2004).

The 2D simulation shown in ﬁgure 2B

was implemented in Matlab and contact loca-

tions are calculated using the Geom2D toolbox

(http://matgeom.sourceforge.net).The simulated an-

tenna and the Hopf Oscillator equations are iterated

using ﬁrst order, forward Euler Integration with a

ﬁxed time step ∆t.

After a contact is detected by the antenna, the

amplitude µ and the frequency ω are immediately

switched from a large amplitude, low frequency

“searching mode” to a high-frequency, low ampli-

tude “sampling mode”. Once a certain time span

sampling

has passed without encountering a further

contact event, the parameters are switched back to the

“searching mode” pattern.

A 180

◦

phase shift can be easily implemented by

negating the Hopf Oscillator state variables: x

t+1

−x

and y

t+1

= −y

. Because of the discrete-time na-

ture of numerical simulations, contacts with an obsta-

cle may last longer than a single time step. Therefore,

the 180

◦

phase shift should be applied only once for

each contact event. In the block diagram in ﬁgure 1,

this is indicated by the discrete pulse block.

2.3 Robustness Evaluation

The robustness of the contour tracing algorithm was

evaluated by scanning a rough surface. For genera-

tion of a random contour, we tested shapes with lin-

ear segments of random orientation and length. The

goal for the algorithm was to completely scan the sur-

face from the right to the left outermost side with-

out getting stuck, see ﬁgure 3A for a sample surface.

The surface was structured as a 10 units long, initially

straight line with 40 uniform segments. The locations

, y

) of the nodes connecting the segments were

ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence

0 2 4 6 8 10

100

noise level

success rate [%]

A B

Figure 3: Robustness evaluation of the contour-tracing algorithm. A: Extreme example, in which the antenna got trapped in a

cavity of the random contour. B: Success rate of contour-tracing in percentage. N=100 trials.

Table 1: Simulation parameters.

parameters 2D-Sim 3D-Sim

γ 4.0 4.0

searching

1.0 0.5

sampling

0.2 0.1

searching

1.0 1.0

sampling

2.0 2.0

∆t 0.02 0.02

= s

1.0 0.5

then randomized by x

= x

+ 0.15 ∗ rand(−r, r) and

= y

+rand(−r, r) where rand(−r, r) generates uni-

formly distributed random numbers between −r and

r. N=100 trials were performed for each noise level

r = 0.5, 1, .., 10. In almost all trials, the rough surface

could be completely scanned from start to end. Fig-

ure 3A shows one of the rare cases, where the contour

tracing algorithm got trapped in a cavity. Figure 3B

shows the overall success rate, i.e. the percentage of

trials that completely scanned the rough surface.

2.4 3D Simulation

The contour tracing algorithm also works reliably in

3D space. Virtually no change in the algorithm is re-

quired, except minor parameter tuning, see table 1.

Figure 4 shows a kinematic antennal model with or-

thogonal joint axes (similar to a cardan joint), probing

several complex objects like a Torus or a Octahedron.

Contact events were calculated using the Geom3D

Matlab package, including the contact distance along

the tactile probe.

3 TACTILE SHAPE

RECOGNITION

An application scenario for Contour-Net is tactile

shape recognition of sampled objects. Each contact

point with the surface can deliver direct information

about the contact distance, the current joint angles at

contact time and indirect values like the state of the

Hopf Oscillator. Collecting these values should al-

low a discrimination of object shapes based on pre-

vious learned examples. Here, we apply a plain but

large feed-forward neural network to solve the classi-

ﬁcation task. Previous results have shown that mul-

tilayer neural networks can classify hand-written dig-

its and temporal eye-tracking data with minimal pre-

processing, only (Krause et al., 2013a). We propose to

feed the collected and normalised contact-point data

directly as input values to the network. Normalisation

rescales the different data components to a range from

[−1, 1].

Due to the well known curse of dimensionality

problem (shrinking norm, Beyer et al. (1999)), the

training dataset should be as large as possible, requir-

ing long training times with common neural network

training algorithms. For example, standard back-

propagation has to deal with local minima (LeCun

et al., 1998) and vanishing gradients (Hochreiter et al.,

2001). Using No-Prop-fast (Krause et al., 2013a), a

special case of an Extreme Learning Machine (Huang

et al., 2006), these problems can be circumvented

and almost interactive training times can be achieved.

This allows the generation of parameter tuning curves

with many repetitions of the learning process using

personal computers.

3.1 Training Dataset

To collect a large training dataset, four different 2D-

shapes (triangle, square, star, circle) were traced us-

ing Contour-Net. The shape size was scaled between

100% and 200%, shapes were rotated randomly be-

tween 0

◦

and 360

◦

, and the initial contact location was

also randomized. Figure 6 shows a small sample from

Contour-Net-AModelforTactileContour-tracingandShape-recognition

Figure 4: 3D simulation of contact triggered contour tracing. Grey: objects. Green: insect head with left antenna. Blue:

trajectory of the antennal tip. Red dots: contact locations. A: torus, B: octahedron.

bias=1

0 „triangle“

1 „square“

0 „circle“

linear output

layer

fixed

hidden layer

0 „star“

Figure 5: General network structure used for shape recog-

nition. A plain feed forward network with a large hidden

layer is used. Data components ( ˙x, ˙y from Hopf Oscilla-

tor, contact distance d along antenna) collected during a

contour-sampling are serialized, rescaled to the range -1 to

1 and fed into the network. Shorter contour-scans can be

zero-padded. Hidden layer weight values are initialized ran-

domly and only the output layer is trained (see text). Class

labels use a “one-out-of-n” coding.

the 2D-dataset. A ﬁxed number of contact points was

selected such that all shapes in all sizes were ’encir-

cled’ at least once, hence completely traced. 400 ran-

dom samples with 35 contact points per shape were

collected, resulting in a training dataset with 1600

samples. Each contact point consisted of 3 compo-

nents: the ˙x and ˙y velocity components of the Hopf

Oscillator and the object size. Hence, the raw input to

the neural network has 105 dimensions. The network

has one hidden layer and 4 outputs. Output values use

a “one-out-of-n” coding scheme. The hidden layer

of No-Prop networks (Widrow et al., 2013) is ﬁxed

and was initialized with uniformly distributed random

numbers. Figure 5 shows the general network struc-

ture.

3.2 Classiﬁcation Performance

25 randomly initialized networks were trained on the

dataset, and performance was evaluated using 10-fold

cross validation. Figure 7B shows how the classiﬁca-

tion accuracy depends on the hidden layer size. As

shown in Krause et al. (2013a), essentially two com-

ponents inﬂuence the performance: the hidden layer

size and the hidden layer weight range. The optimal

weight range of hidden layer units does not depend

on the number of hidden layer units (ﬁgure 8) but on

the number of inputs and was estimated to be from -

0.5 to 0.5 for both the 2D and 3D training dataset (see

section 3.3). Figure 7B (blue curve) shows that an

acceptable, size- and rotation-invariant shape classiﬁ-

cation can be achieved. The discrimination rate peaks

around 600 hidden units and decreases after that. The

low spread of performance values indicates that the

algorithm has not to deal with local minima. This is

a favourable feature of the No-Prop method, see also

Widrow et al. (2013).

Yet, contour tracing trajectories were found to be

fairly variable, depending on angle of attack of the

tactile probe relative to the surface. Figure 7A shows

an example where the probe arrives at a ﬂat angle to

the surface. The subsequent, contact-induced 180

◦

phase shift of the oscillator causes the probe to leave

the surface in a suboptimal, backward direction. The

resulting pattern alternates between a ﬂat and large

arc.

ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence

Figure 6: Sample collection of the four different shapes used for the classiﬁcation task. The size and orientation of the shapes

as well as the initial contact location was randomized.

0 200 400 600 800 1000

100

hidden layer size

classification accuracy [%]

CBA

Figure 7: Classiﬁcation results improve after incorporating surface-normal information into Contour-Net. A: Sampling trajec-

tory using a ﬁxed 180

◦

phase shift of the Hopf Oscillator after contact. B: The phase of the oscillator is set such that the tactile

probe leaves the object surface in the normal direction after a contact. The resulting sampling trajectory is more regular. C:

Classiﬁcation accuracy using a ﬁxed phase shift (blue), and variable phase shift depending on the object surface normal (red).

A training dataset with four shapes as shown in ﬁgure 6 with n=400 samples per shape was used. N=25 repetitions. Solid

curve = mean value, shaded area = min and max values.

Approaching a surface perpendicular with an op-

timal, 90

◦

angle, not only avoids potential slip, see

Kaneko et al. (1995). Improving the regularity of the

scan path might also improve the shape classiﬁcation

performance. A simple solution would be to use the

estimated surface normal

for setting the outgoing

In the simulation, the surface normal was estimated by

a local, 360

◦

subsampling of the surface around the contact

position. From the resulting list of intersections with the

angle after contact. Figure 7B shows an example of

an improved sampling trajectory. The main advan-

tage for the classiﬁcator is a lower spread of contact

parameters along ﬂat regions of an object and a more

pronounced change in values at sharp edges. This is

reﬂected in more than 10% improvement of the classi-

ﬁcation accuracy. Figure 7C, red curve, shows the im-

object, an average surface normal vector could be calculated

similar to the mean resultant vector in circular statistics.

Contour-Net-AModelforTactileContour-tracingandShape-recognition

Figure 9: Sample collection of three different 3D objects: 1. row: soccer ball; 2. row: cube; 3. row: tetrahedron. The size,

position and spatial orientation of the shapes as well as the initial antennal location were randomized.

78.5

79.5

80.5

81.5

82.5

hidden layer size

500 15001000750 1250

weight range

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Figure 8: Classiﬁcation accuracy (color coded) depend-

ing on weight range and hidden layer size. Used training

dataset: 3D shapes, n=1000 samples per shape, n=20 repe-

titions. The optimal weight range value is independent from

the hidden layer size.

proved recognition rate using a dataset identical in pa-

rameters to section 3.2, but generated using Contour-

Net with variable phase shift depending on the surface

normal.

3.3 3D Shape Classiﬁcation

Subsequently, classiﬁcation performance using 3D

objects was tested. Three different objects were

used: a soccer ball, a cube and a tetrahedron, see

ﬁgure 9. The objects were scaled between 100%

and 150%; placed at a random position (x, y, z) =

[20..30, −5..5, −5..5] in front of the antenna; and ro-

tated randomly around all axes in the full range of

360

◦

. The initial antennal location around the object

was also randomized. Three datasets with 400, 1000

and 3000 random samples per shape with 35 contact

points per shape were collected. From each contact

event, three components were used: the ˙x and ˙y values

of the Hopf Oscillator and the contact distance along

the antenna. The neural network had 105 inputs and

three outputs.

Figure 10 shows the classiﬁcation accuracy for the

three differently sized datasets. The increased com-

plexity of the task required a very large dataset to

achieve a classiﬁcation rate above 90%. Due to the

random rotation around all axes, the supervised learn-

ing algorithm has to be trained on a sufﬁciently large

number of contour projections of the objects. Re-

stricting the task to a ﬁxed object size and to a random

rotation around a single axis, only, would certainly re-

duce the necessary size of the training dataset. Figure

10B shows that the contact distance information only

slightly improved classiﬁcation performance by 5%.

Hence, the contact distance along the antenna is not

essential and does not need to be very precise or may

be omitted completely.

3.4 Performance Evaluation

The No-Prop-fast method was compared with two

ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence

0 500 1000 1500 2000 2500 3000 3500 4000

100

hidden layer size

classification accuracy [%]

distance,

dx, dy

1200 samples

n = 25

3000 samples

n=25

3000 samples

n = 25

9000 samples

n = 10

A B

Figure 10: 3D shape classiﬁcation performance strongly depends on the sample size. A: Performance for three different

sample sizes. The neural network requires a fairly large sample size to achieve high performance. It needs to “see” several

contour projections of the randomly oriented objects. B: Inﬂuence of the input components on performance. Performance is

still good with the velocity components of the Hopf Oscillator, only.

50 60 70 80 90 100

No-Prop-fast: 91.6%

backprop- : 91.9%scg

support vector machine - rbf: 98.0%

cpu time [s]

3D dataset, 9000 samples, n=25

3.0 ± 0.1

146.8 ± 0.7

19.8 ± 1.0

Figure 11: Classiﬁcation accuracy and training times

of three different algorithms. First row: No-Prop-fast

hidden

= 3000), second row: scaled conjugate gradients

backpropagation (n

hidden

= 64), third row: gaussian kernel

support vector machine. Bars show median, whiskers show

min and max values, CPU times are given as mean±SD.

other algorithms, the default pattern classiﬁcation al-

gorithm of the Matlab 2012a neural network toolbox,

scaled conjugate gradient backpropagation (BP

scg

)

and an implementation of a gaussian kernel support

vector machine (SVM

rb f

) found in the Matlab 2012a

statistics toolbox. Figure 11 shows recognition rates

and CPU times of these algorithms, applied to the

3D shape dataset with 9000 patterns. CPU time mea-

surements were performed on a Dell Precision T3500

Workstation (Intel Xeon W3530 CPU at 2.8 GHz with

8 Mbyte second level cache) running Windows 7. The

three different algorithms were evaluated using ten-

fold cross-validation and 25 repetitions with random-

ized samples and randomly initialized networks.

A separate SVM

rb f

was trained for each shape

(one-against-all classiﬁcation), using the matlab

function ’svmtrain’ with sigma_rb f = 4.2 and

boxconstraint = 10. All other parameters were kept

at their default values. The performance of BP

scg

was

tested using a feedforward network with a single hid-

den layer (64 units), created with the Matlab function

’patternnet’ and trained with all parameters at their

default values. SVM

rb f

achieved the highest median

recognition rate (98%). Recognition rates of BP

scg

were lower (92%) with high spread, but individual

networks achieved up to 96%. No-Prop-fast follows

with 91%.

While the recognition rates of all three algorithms

are above 90%, training times were found to differ

signiﬁcantly. Values given in ﬁgure 11 show the aver-

age duration of individual training runs (n=250). The

Matlab implementation of SVM

rb f

required almost

147 seconds to learn the dataset, while No-Prop-fast

required only three seconds. Hence, for the shape

recognition task, No-Prop-fast is approximately 50

times faster than SVM

rb f

and still six times faster than

the Matlab implementation of one of the fastest BP-

algorithms.

4 CONCLUSIONS

Contour-Net is a robust, bio-inspired method for tac-

tile contour-tracing. In its simplest instantiation, it

Contour-Net-AModelforTactileContour-tracingandShape-recognition

requires only a single, binary signal: if the antenna

or tactile sensor is in contact with an object or not.

Yet, the model captures essential behaviours observed

in its natural paragon, the stick insect antenna (Dürr

et al., 2001). For example, it implements the contact-

induced switch from a large amplitude searching be-

haviour to a local, higher frequency sampling of an

object (Krause et al., 2013b). Contour-Net can be eas-

ily extended to use two or more antennae including

mutual coordination and attentive visual target track-

ing, as will be shown in a follow-up paper.

Raw, minimally pre-processed (normalisation)

data collected from contact events was sufﬁcient to

achieve rotation-, size- and position- invariant shape

recognition rates of over 90% using a plain feed-

forward neural network. A drawback for robotic ap-

plications is that the training dataset needs to be fairly

large. Instead of using raw data as the network input,

extracting higher level features from collected con-

tour points should signiﬁcantly reduce the required

amount of training samples. Possible features might

be the average Euclidean distance and the spread of

Hopf Oscillator phase differences between successive

contact points. An unsupervised learning algorithm

(SOM; Boltzmann Machine) will be closer to nat-

ural, continuous learning (automatic clustering with

delayed categorization).

We have further shown that incorporating surface

normal information will improve not only the regu-

larity of the sampling trajectory, but also improve the

recognition performance of a neural network. Hit-

ting a surface perpendicular can potentially reduce

slip of an insect antenna inspired robotic tactile sensor

(Kaneko et al., 1995). The reliable and fast detection

of the surface normal using a tactile sensor will be a

challenging and interesting task.

ACKNOWLEDGEMENTS

This work was supported by EU grant EMICAB

(FP7-ICT, grant no. 270182) to Prof. Volker Dürr.

We thank Prof. Holk Cruse for valuable comments on

earlier versions of the manuscript.

REFERENCES

Beyer, K., Goldstein, J., Ramakrishnan, R., and Shaft, U.

(1999). When is "nearest neighbor" meaningful? In

Beeri, C. and Buneman, P., editors, Database Theory -

ICDT 99, volume 1540 of Lecture Notes in Computer

Science, pages 217–235. Springer Berlin Heidelberg.

Brambilla, G., Buchli, J., and Ijspeert, A. (2006). Adap-

tive four legged locomotion control based on nonlin-

ear dynamical systems. In From Animals to Animats

9. Proceedings of the Nineth International Conference

on the Simulation of Adaptive Behavior (SAB 06).

Buchli, J., Righetti, L., and Ijspeert, A. J. (2005). A dy-

namical systems approach to learning: a frequency-

adaptive hopper robot. In Proceedings of the VIIIth

European Conference on Artiﬁcial Life (ECAL 2005),

Lecture Notes in Artiﬁcial Intelligence, pages 210–

212.

Dürr, V., König, Y., and Kittmann, R. (2001). The anten-

nal motor system of the stick insect Carausius mo-

rosus: Anatomy and antennal movement pattern dur-

ing walking. Journal of Comparative Physiology A,

187(2):131–144.

Erber, J., Kierzek, S., Sander, E., and Grandy, K. (1998).

Tactile learning in the honeybee. Journal of Compar-

ative Physiology A, 183:737–744.

Esslen, J. and Kaissling, K.-E. (1976). Zahl und

verteilung antennaler sensillen bei der honigbiene

(Apis mellifera l.). Zoomorphology, 83:227–251.

10.1007/BF00993511.

Hochreiter, S., Bengio, Y., Frasconi, P., and Schmidhuber, J.

(2001). Gradient ﬂow in recurrent nets: the difﬁculty

of learning long-term dependencies. In S. C. Kremer,

J. F. K., editor, A Field Guide to Dynamical Recurrent

Neural Networks. IEEE Press.

Hoinville, T., Wehner, R., and Cruse, H. (2012). Learn-

ing and retrieval of memory elements in a navigation

task. In Prescott, T., Lepora, N., Mura, A., and Ver-

schure, P., editors, Biomimetic and Biohybrid Systems,

volume 7375 of Lecture Notes in Computer Science,

pages 120–131. Springer Berlin Heidelberg.

Huang, G.-B., Zhu, Q.-Y., and Siew, C.-K. (2006). Extreme

learning machine: Theory and applications. Neuro-

computing, 70(1-3):489 – 501.

Ijspeert, A. J. (2008). Central pattern generators for loco-

motion control in animals and robots: A review. Neu-

ral Networks, 21(4):642 – 653.

Ijspeert, A. J., Crespi, A., Ryczko, D., and Cabelguen, J.-

M. (2007). From swimming to walking with a sala-

mander robot driven by a spinal cord model. Science,

315(5817):1416–1420.

Kaneko, M., Kanayama, N., and Tsuji, T. (1995). 3-d active

antenna for contact sensing. In IEEE International

Conference on Robotics and Automation, volume 1,

pages 1113–1119. IEEE.

Kaneko, M., Kanayma, N., and Tsuji, T. (1998). Active

antenna for contact sensing. IEEE Transactions on

Robotics and Automation, 14(2):278–291.

Kevan, P. G. and Lane, M. A. (1985). Flower petal micro-

texture is a tactile cue for bees. Proceedings of the

National Academy of Sciences, 82(14):4750–4752.

Kim, D. and Möller, R. (2007). Biomimetic whiskers for

shape recognition. Robotics and Autonomous Systems,

55(3):229–243.

Krause, A. F. and Dürr, V. (2004). Tactile efﬁciency of in-

sect antennae with two hinge joints. Biological Cy-

bernetics, 91(3):168–181.

Krause, A. F. and Dürr, V. (2012). Active tactile sampling

ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence

100

by an insect in a step-climbing paradigm. Frontiers in

Behavioural Neuroscience, 6(30):1–17.

Krause, A. F., Essig, K., Piefke, M., and Schack, T. (2013a).

No-prop-fast - a high-speed multilayer neural net-

work learning algorithm: Mnist benchmark and eye-

tracking data classiﬁcation. In Engineering Applica-

tions of Neural Networks (EANN 2013), pages 446–

455. Springer.

Krause, A. F., Winkler, A., and Dürr, V. (2013b). Cen-

tral drive and proprioceptive control of antennal

movements in the walking stick insect. Journal of

Physiology-Paris, 107(1-2):116 – 129.

LeCun, Y., Bottou, L., Orr, G. B., and Müller, K.-R. (1998).

Efﬁcient backprop. In Orr, G. B. and Müller, K.-R.,

editors, Neural Networks: Tricks of the Trade, volume

1524 of Lecture Notes in Computer Science, pages 9–

50. Springer Berlin Heidelberg.

Lederman, S. and Klatzky, R. (2009). Haptic perception:

A tutorial. Attention, Perception, & Psychophysics,

71(7):1439–1459. 10.3758/APP.71.7.1439.

Lee, M. and Nicholls, H. (1999). Review article tactile sens-

ing for mechatronics - a state of the art survey. Mecha-

tronics, 9(1):1 – 31.

Patanè, L., Hellbach, S., Krause, A. F., Arena, P., and Dürr,

V. (2012). An insect-inspired bionic sensor for tac-

tile localization and material classiﬁcation with state-

dependent modulation. Frontiers in Neurorobotics,

6(8):1–18.

Righetti, L. and Ijspeert, A. J. (2006). Programmable cen-

tral pattern generators: an application to biped loco-

motion control. In IEEE International Conference on

Robotics and Automation (ICRA), pages 1585–1590.

IEEE.

Russell, R. A. and Wijaya, J. A. (2003). Object location and

recognition using whisker sensors. In Australasian

Conference on Robotics and Automation, pages 761–

768.

Schilling, M., Hoinville, T., Schmitz, J., and Cruse, H.

(2013). Walknet, a bio-inspired controller for hexapod

walking. Biological Cybernetics, 107(4):397–419.

Schütz, C. and Dürr, V. (2011). Active tactile exploration

for adaptive locomotion in the stick insect. Philosoph-

ical Transactions of the Royal Society B: Biological

Sciences, 366(1581):2996–3005.

Solomon, J. H. and Hartmann, M. J. (2006). Biomechan-

ics: Robotic whiskers used to sense features. Nature,

443(7111):525.

Staudacher, E., Gebhardt, M. J., and Dürr, V. (2005). Anten-

nal movements and mechanoreception: neurobiology

of active tactile sensors. Advances in Insect Physiol-

ogy, 32:49 – 205.

Widrow, B., Greenblatt, A., Kim, Y., and Park, D. (2013).

The No-Prop algorithm: A new learning algorithm

for multilayer neural networks. Neural Networks,

37:182–188.

Contour-Net-AModelforTactileContour-tracingandShape-recognition

101