PreprintPDF Available

A Method to use Nonlinear Dynamics in a Whisker Sensor for Terrain Identification by Mobile Robots

Preprints and early-stage research may not have been peer reviewed yet.

Abstract and Figures

This paper shows analytical and experimental evidence of using the vibration dynamics of a compliant whisker for accurate terrain classification during steady state motion of a mobile robot. A Hall effect sensor was used to measure whisker vibrations due to perturbations from the ground. Analytical results predict that the whisker vibrations will have a dominant frequency at the vertical perturbation frequency of the mobile robot sandwiched by two other less dominant but distinct frequency components. These frequency components may come from bifurcation of vibration frequency due to nonlinear interaction dynamics at steady state. Experimental results also exhibit distinct dominant frequency components unique to the speed of the robot and the terrain roughness. This nonlinear dynamic feature is used in a deep multi-layer perceptron neural network to classify terrains. We achieved 85.6\% prediction success rate for seven flat terrain surfaces with different textures.
Content may be subject to copyright.
A Method to use Nonlinear Dynamics in a Whisker Sensor for Terrain
Identification by Mobile Robots
Zhenhua Yu 1, S.M.Hadi Sadati2, Hasitha Wegiriya3, Peter Childs 4,Thrishantha Nanayakkara5,
Abstract This paper shows analytical and experimental
evidence of using the vibration dynamics of a compliant whisker
for accurate terrain classification during steady state motion
of a mobile robot. A Hall effect sensor was used to measure
whisker vibrations due to perturbations from the ground.
Analytical results predict that the whisker vibrations will have
one dominant frequency at the vertical perturbation frequency
of the mobile robot and one with distinct frequency components.
These frequency components may come from bifurcation of
vibration frequency due to nonlinear interaction dynamics at
steady state. Experimental results also exhibit distinct dominant
frequency components unique to the speed of the robot and the
terrain roughness. This nonlinear dynamic feature is used in a
deep multi-layer perceptron neural network to classify terrains.
We achieved 85.6% prediction success rate for seven flat terrain
surfaces with different textures.
Index Terms Robotic whiskers, Surface identification,
multi-layer perceptron, Modal analysis.
Terrain surface identification is an important function
for mobile robots moving and performing tasks in ex-
treme,unstructured environments [1]. By identifying different
terrain surfaces, mobile robots could better perceive the
surrounding environment information and decide the optimal
next move. For example, they can avoid less favorable terrain
types such as soft sand, and use the identification information
for mapping and localization objects. Although many sensors
such as cameras are used for object recognition, whiskered
tactile perception can potentially provide a robust, and eco-
nomical solution for these problems, especially in extreme
(e.g. dark, foggy) environments.
Over the years, vision [2], lidar [3], sound [4], iner-
tial measurement units [5], and tactile sensors [6] have
been studied for surface identification. Vision-based meth-
ods combined with deep learning have become a research
hotspot and significant scientific and technological devel-
opments made. However, vision-based methods can have
*This work was supported in part by the Engineering and Physical
Sciences Research Council (EPSRC) RoboPatient project under Grant
EP/T00603X/1, MOTION project under Grant EP/N03211X/2, Grant
EP/N029003/1, and supported in part by China Scholarship Council.
1Zhenhua Yu is with Dyson School of Design Engineering, Imperial
College London, SW7 2DB London, UK
2S.M.H. Sadati is with the Department of Surgical and Interventional
Engineering, King’s College London, London WC2R 2LS, U.K
3H. Wegiriya is with the Faculty of Natural and Mathematical Sciences,
King’s College London, London, WC2R 2LS, UK .
4P.Childs is with Dyson School of Design Engineering, Imperial College
London, Dyson Building, 25 Exhibition Road, London, SW7 2DB, UK.
5T. Nanayakkara is with Dyson School of Design Engineering, Imperial
College London, Dyson Building, 25 Exhibition Road, London, SW7 2DB,
Fig. 1. Proposed mobile robot equipped with a bioinspired whisker sensor
mounted on a 4-DOF servo fusion rotating platform.
a high probability of failure because of the similar visual
appearance of different terrains or objects. Moreover, visual
accuracy is severely impaired by environmental conditions
such as fog, smoke, low light levels, high-brightness and
high-temperature. Recognition and classification methods for
different objects based on lidar data have been popularly
used in autonomous cars, but its accuracy will drop in sandy
applications and dusty scenes. For an IMU installed on the
mobile robot body, a challenge is that they cannot predict
the terrain and object types in advance before the robot has
contact. For example, when the robot detects a change in
terrain from concrete to soft sand based on inertial signals,
the wheels may have already come into contact with the sand.
This can lead to dangerous consequences. Therefore, it is
important to find robust sensing modes for the environment
to achieve terrain identification under extreme unstructured
environments and the whisker sensor is a choice worth
exploring [7].
Artificial Whiskers sensors have been demonstrated in
several studies that are a particularly high-efficiency method
for animals such as rats and sea lions which can use them to
navigate in the dark and perceive environmental information
and features without vision [8].
This has motivated researchers to design and construct
whiskers for applications such as navigation [9], obstacle
avoidance [10], object detection [11], size measurement [12],
shape recognization [13], and surface information [14]. For
example, in [15], Zurek showed that static antennae can
act as locomotory guides which can compensate for visual
methods to determine the location and distance of obstacles
during fast locomotion. By rotating the whisker against a
sequence of contact points of the object and collecting
torque information, Solomon and Hartmann put forward a
method to obtain an object’s 3D contour information [16].
In 2020, authors of [17] showed that a novel variable stiffness
controllable multi-modal whisker sensor can capture different
vibration frequencies by controlling whisking speed and the
stiffness of the follicle. In all these cases, experimental
results prove that the whisker sensors can effectively com-
pensate for the shortcomings of vision sensors in an extreme
In the work reported here, we designed and constructed a
novel whisker tactile sensor to achieve terrain identification
for a mobile robot platform (Fig. 1). Reservoir computing
[18], [19] is used in the whisker sensor to map slight
changes in the ground perturbations to distinct steady state
frequency components. This paper firstly shows that the
steady state response of nonlinear vibration dynamics can
be used to classify terrain textures even on flat terrain. The
remainder of this paper is organized as follows: Section II
introduces the design and construction of the whiskers tactile
sensor, and the experimental system including the Raspberry
robot, whiskered tactile sensor. The whisker sensor vibration
characteristics and modal analysis are studied in Section
III. Section IV reports the whisker vibration data collection
and feature extraction as well as algorithm framework based
on multi-layer perceptron. Section V presents the results
and analysis of terrain identification experiments . Finally,
Section VI concludes this paper’s contribution and discuss
the future work.
A. Bioinspired Whisker Sensor Design
The proposed whiskered tactile sensor uses SS49E linear
Hall sensors that are orthogonally mounted on one side
of a silicone rubber sensor holder base and spring beam
on the other side. The structure is shown in Figure 2. A
neodymium permanent magnet is embedded inside the spring
beam that is made of high-carbon steel, with the free length:
60 mm, wire diameter: 1 mm, outside diameter: 10 mm, and
inside diameter: 8 mm respectively. Tapered silicone rubber
is installed on the tip of the spring in order to accurately
capture the vibration of the surface . The components and
materials cost of this whiskered tactile sensor totalled less
than $5 based on one-off non-discounted prices. The low
cost construction of this novel sensor provides an advantage
to many existing surface identification sensors
B. Sensor Characterization & Working Principle
The whisker sensor is installed on the 4-DOF servo fusion
rotating platform for orientation control of the whisker,
which keeps the whisker sensor at a horizontal orientation all
the time for clarity. When the robot traverses different terrain,
external vibrations are applied to the whiskered spring shaft,
and the shaft deforms also inducing the same vibrations
on the magnet inside. This causes a continuous magnetic
flux change near the linear hall effect sensor. Consequently,
the hall effect sensor generates continuous low-frequency
Silicone Rubber
Spring beam
Neodymium magnet
Holder 𝑙
Fig. 2. The bioinspired whisker sensor.
Fig. 3. Schematic diagram of the robot whisker system flow framework
electrical voltage signals. The sensitivity of this whiskered
tactile sensor relies on the vibration of the whiskered spring
beam and the silicone rubber tip.
C. Experimental Setup Design & Procedure
The experiment is conducted by a four-wheel mobile
robot. The wheeled unit and electronic system are shown
in Fig.1. The robot is 270 mm in length, 150 mm in height
and 130 mm in width, 1.4 kg in mass, with the diameter and
width of the wheels are 70 mm and 30 mm respectively. The
robot car could traverse through the coarse ground at a speed
of up to 1.3 m/s, With a power supply of 5 V.
The robot system consists of a whiskered tactile sensor
orthogonal fixed in a 4-DOF servo fusion rotating platform,
controlled by a Raspberry Pi 4 B.The whiskered tactile
sensor has its own amplifier, and the raw voltage signals
are sent to an ADS1115 analog to digital converter.
The AD converters, servo motor driver as well as the
4-DOF rotating platform are connected to a Raspberry Pi
4B. The sensor data collection, robot’s motion and rotating
platform’s movement are synchronized controlled by the
Raspberry Pi 4B . All data are recorded in the raspberry pi
on-board flash memory, and then transferred to a computer
(1.60 GHz, 8 GB RAM) via Bluetooth for processing. Fig.3
shows the system flow framework.
The sensor vibration characteristics are studied as a proof-
of-concept for the idea presented.
A. Cantilever Beam Vibration with Base Excitation
The sensor is modelled as an equivalent cantilever beam
with uniform mass under base excitation (Fig. 2). The steady
state response of such system for sinusoidal base excitation
yb=hbsin(ωbt)can be found based on the derivations in
[20] as a summation over the beam first five mode shapes in
Eq. 1,
y(x,t) =
Yi(x) = sin(βi·x)sinh(βi·x)...
αi=sin(Di) + sinh(Di)
cos(Di) + cosh(Di),βi=Di
qi(t) = 1
i(x)·dx,ωζi= D2
Qi(t) = Zl
where y(x,t)is the beam displacement at any length
xand time t,hband ωb=2πfare the base ex-
citation amplitude in [m] and frequency in [rad/s],
fis the base excitation frequency in [Hz], D=
[1.8751,4.6941,7.8548,10.9955,14.137]is a set of
constants related to a cantilever beam mode shapes [20],
and ζ=0.04 is the modal damping based on stainless steel
material. The equivalent beam has a length lsimilar to
the coil spring axial length lc=l=60 [mm] and cross-
section area asimilar to the spring wire cross-section area
w, where rw=0.5 [mm] is the wire radius. The
beam unit length density is ρ=Cρwwhere ρw=8050
[Kg/m3] is the wire material (stainless steel) density, and
C=lw/(np)is a correction factor to account for the coil
shape based on the spring wire pitch p=ls/n, overall length
lw=np(2πr)2+p2, and spring number of coils n=13.
We may assume that the beam bending modulus E I is
equivalent to the spring wire torsional modulus GwJw, where
Gw=Ew/3=70 [GPa] is the wire material (stainless steel)
shear modulus, and Jw=πr4
w/4 is the wire cross-section 2nd
moment of area.
B. Sensor Modal Analysis
The simulation results for the sensor lateral displacement
ys=y(xs,t), where xs=5 [mm] is the sensor distance from
the beam base, and the signal FFT (Fast Fourier Transform)
analysis in response to different excitation parameters (hb=
0.1&0.3 [mm] and fb=100&300 [Hz]) are plotted in
Fig. 4. Distinctively, different signal profiles and dominant
frequency are observed for different excitation frequencies
(i.e. due to ground texture). The signals’ first two dominant
frequencies are 100 (same as excitation) & 140 [Hz] and 120
& 300 (same as excitation) [Hz] respectively.
Fig. 5 shows the sensor maximum displacement ysmax
in [m] and dominant modal frequency fsmax in [Hz] for
different excitation frequency fband amplitude hbvalues.
Distinct regions are observed for the values of fsmax as a
function of fband independent of hb. This shows that it
0 200 400 600 800 1000
Amp. [m]
0 200 400 600 800 1000
Amp. [m]
0 200 400 600 800 1000
fs (Hz)
Amp. [m]
0 0.005 0.01 0.015
t [s]
ys [m]
hb [m]
hb = 0.0003 [m], fb = 300 [Hz]
0 0.01 0.02 0.03 0.04 0.05
ys [m]
hb [m]
hb = 0.0003 [m], fb = 100 [Hz]
0 0.01 0.02 0.03 0.04 0.05
ys [m]
hb [m]
hb = 0.0001 [m], fb = 100 [Hz]
Fig. 4. Sensor displacement y(steady-state oscillations for five base exci-
tation cycles) and FFT analysis due to various base excitation frequencies
fband amplitudes hb.
hb [m] 10-3
fb [Hz]
ys-max [m]
hb [m]
fb [Hz]
500 0
fs-max [Hz]
Fig. 5. Sensor displacement maximum value ysmax and dominant modal
frequency fsmax vs. base excitation height hband frequency fb.
is possible to successfully classify the excitation frequency
fb(i.e. the regions with different different growth patterns)
based on the sensor signal dominant frequency fsmax.ysmax
is affected by both the fband hbdespite its distinct maximum
values around fb=100 [Hz] and rapid growth for higher
values of fb. As a result, the sensor signal amplitude ysmax
is not solely enough to determine valid information about
the base excitation. However, by classifying the excitation
frequency fbbased on the sensor signal modal analysis, the
sensor signal amplitude ysmax has enough information to
determine the excitation amplitude hb(i.e. ground profile
height). Sudden changes in the slope of ysmax vs. fbfor
fb500 [Hz] show that sharper changes in the sensor signal
should be anticipated as a result of profile texture variations
for a higher frequency base excitation (e.g. rough train).
The above analysis indicates that the sensor signal can
provide enough information for classifying the sensor base
excitation (ground profile) if an appropriate method is em-
ployed to effectively handle the real-world uncertainties.
Such a method is discussed later in this paper.
The entire terrain surface identification and recognition
process are divided into three phases: 1) whisker-based off-
line data collection and processing; 2) off-line training based
Fig. 6. Diagram of whisker vibration data collection and procession
on machine learning; 3) whisker-based online surface clas-
sification and recognition. The whisker should be designed
properly and the classification models need to be trained off-
line accurately, in order to achieve a higher success rate for
identifying similar terrain surfaces.
A. Data Feature Extraction
The whiskered robot needs to collect enough whiskered
tactile sensor vibration information for model training by
traversing different terrain surfaces several times. The operat-
ing frequency of whiskered tactile sensor during experiments
is 200 Hz, and the tactile vibration data which is collected by
the whiskered sensor was pre-processed and then segmented
for accelerating the model training speed. Every segmented
vector corresponds to one-second of data from the whiskered
tactile sensor , so a 1 200 labeled surface vector x1200
be created:
i= [v1,v2,.....v200]Si|i=1,2....7(2)
Where Siis the different terrains surfaces and
i=1,2,3,4,5,6,7 corresponds to flat, cement, brick , carpet,
soft-grass, sand and asphalt terrain surface respectively.
The raw vibration data is subsequently converted from the
time domain to frequency domain . The first stage of data
procession is standardized, and every vibration vector unit
is normalized to a unit vector whose standard deviation is 1
and the mean value is 0.
Then, the tactile vibration vector unit x1200
iwas trans-
formed from the time-domain to the frequency domain x1200
through a Fast Fourier Transformation. The Fast Fourier
transformation has been proven efficient and capable of clas-
sifying the difference between multiple terrains in Section
III and it significantly enhances the Fourier transform speed
when compared to other Fourier transform methods.
B. Network Architecture
As shown in Figure 7, a , deep neural network with seven-
layers based on multi-layer perceptron is built to achieve
the different terrain surfaces identification. The activation
functions of the first five layers are Rectified Linear Units,
and the Soft-max function is introduced to better identify
surface types. The cross-entropy loss function activation can
be applied to evaluate the deviation matrix between the
predicted and actual value.
Fig. 7. Overview of Deep Multi-Layer Perceptron Neural Network Pipeline.
For different terrains, this project collects 5 minutes of
tactile vibration data Xraw for model training and terrain
surfaces classification. These data are pre-processed and
divided into 300 terrain feature vector unit x(i,f)based on
above data feature extraction method. All the data are labeled
and connected as a terrain feature class vector X(i,f).
During the training and classification period, 75 % data
Xtrain are randomly selected from the Xffor model training,
and the remaining 25% data Xtest are used to test the
performance of network. The input layer of this network
consists of 200 neurons which corresponds to the dimension
of the surfaces feature vector unit x1200
(i,f), and the output
layer includes 7 neurons, corresponding to the seven different
terrain surface types considered here. All layers of the
network are fully connected. After the training is completed,
this network is used to classify the remaining 25% testing
vectors data Xtest and it returns an estimation of the terrain
type. By trial and error, all the network’s training parameters
were updated, but it is impossible to make it optimal.
To verify the reliability and robustness of the afore-
mentioned whiskered tactile sensors and sensor vibration
characteristics analysis based on reservoir computing, surface
identification experiments were conducted on seven different
terrains at the same speed. Then, the influence of the mobile
robot running speed on the whiskered tactile sensor’s iden-
tification capability has also been analyzed and discussed
from nonlinear dynamic resonance perspectives based on the
theory presented in Section III.
A. Terrain Surfaces Identification
Based on the above methods, the whiskered robot was
controlled by a Raspberry Pi 4 system to move on 7 different
terrain surfaces: brick, cement, flat terrain, carpet, soft-grass,
sand and a soft-soil surface respectively. All raw collected-
data are pre-processed and the deep multi-layer perceptron
neural network is pre-trained offline properly. Part of the
sample experiment terrain surface and its corresponding
vibration data are given in Figure 8.
This experiment collects 5 minutes of vibration data for
each terrain surface. By randomly extracting 75 % data Xtrain
from the terrain tactile vibration database as the training
Fig. 8. Comparison of two example terrain surface(top two:soft-grass;the
bottom two:brick) and corresponding raw vibration data. The time window
of this vibration voltage data is 5 seconds
75%; TES T DATA 25% )
Parameters Value
Robot Velocity 0.2 m/s
Sampling Frequency 200 Hz
Time-window 1 sec
Fig. 9. Prediction rate of seven terrains surface
data, and the rest 25% of data Xtest were used to test the
performance of this network. The experimental parameters
are set as shown in Table I. In all, 20 random experiments
were conducted to avoid accidental errors caused by random
As can be seen from the Figure 9, the predicted classifica-
tion success rates for flat terrain, carpet, cement, soft-grass,
brick, sand and softsoil terrain surface are 87.3%, 83.6%,
Fig. 10. Confusion matrix of seven terrain prediction success rate
79.2%, 91.3%, 85.4%, 86.6%, and 85.8% respectively. The
experimental results indicate that the whiskered tactile sensor
and the deep multi-layer perception neural network have
good recognition and identification capabilities of different
terrains, with an average success rate of about 85.6%.The
results shows that the steady state response of nonlinear
vibration dynamics of whisker sensor can be used to classify
different terrains.
Based on the confusion matrix in Figure 10, this network
has high identification and classification success rate for
flat and soft grass terrain surface because they have distinct
vibration features based on their vibration signals. However,
there is a high confusion rate between the flat surface and
the brick surface which can be explained by the similar gap
between the two adjacent units. In the case of the cement
terrain, this could have a complicated surface including
features such as gaps, a smooth or rugged surface which
make the network more challenging to classify it from others.
Combined with the simulation results in Section III, the
results indicate that robot’s base excitation vibration am-
plitude (i.e. ground profile height) affects the sensitivity of
whisker sensor , and the sensor signal can provide enough
information for classifying the sensor base excitation (terrain
profile) by employing deep multi-layer perceptron neural
B. Different Speed Experiments
The result of the previous section indicates that the
whiskered tactile sensor and the proposed deep multi-layer
perceptron neural network reported in this paper has good
classification and identification capabilities on seven terrains
Siat a constant velocity 0.2 m/s. The influence of different
speeds on the sensor and model prediction accuracy has been
analyzed. Data collection was performed at speeds of 0.1
m/s, 0.15m/s, 0.2m/s, 0.25 m/s and 0.3 m/s respectively.
All parameters during this experiment are the same as in
the previous section. The identification accuracy results are
shown in Table II.
Speed S1S2S3S4S5S6S7
0.10m/s 87.6% 78.8% 82.3% 88% 84.8% 86.9% 74.8%
0.15m/s86% 85.2% 88.6% 83.2% 82% 93.2% 82%
0.20m/s87.3% 83.6% 79.2% 91.3% 85.4% 86.6% 85.8%
0.25m/s90.6% 93% 90.3% 81.3% 89.1% 79.3% 91%
0.30m/s92.5% 78.5% 82.2% 85.5% 92.5% 84.2% 78.5%
Fig. 11. Identification Accuracy of Seven Terrain Surfaces at different
speeds.The colour of blue, orange,yellow, purple and green represent the
speed of 0.1 m/s, 0.15 m/s, 0.2 m/s, 0.25 m/s and 0.3 m/s respectively.
Looking at Figure 11 and Figure 12, it can be seen that this
system has good identification accuracy for seven terrains
surface at different speeds. The average prediction accuracy
of the seven terrains at different speeds are: 83.31%, 85,74%,
85.60%, 87.8%, and 84.84%. Therefore, speed has no con-
Fig. 12. Average prediction accuracy of seven terrains at different speeds
sistent effect on identification success rate. However, there
are still some details here that can be used as a reference
to improve the identification accuracy. Firstly, the average
identification accuracy at a speed of v = 0.25 m/s is higher
than the other four operating states reaching 87.8%, which
indicates that the speed of mobile robot has influence on
identification accuracy. This may be caused by the dynamic
resonance of the nonlinear experiment system. The spring
beam and the robots car achieve a better resonance effect
at 0.25m/s, which makes the spring vibration more obvious.
Since the sensitivity of the sensor depends on the vibration of
the spring, it has a better classification effect at this speed.
This also echoes the simulation results of the Section III,
indicating that robot’s base excitation vibration frequency
affects the whisker tactile sensor sensitivity. Moreover, a
nonlinear system has multiple resonance frequencies, this can
also explain why good results are also achieved at the speed
0.15m/s but relatively poor at 0.2m/s.
Secondly, the accuracy of distinguishing between the brick
and flat terrain surfaces increases when the robot speed
increases. Combined with modal analysis in Section III, these
results indicate that terrain surface profile amplitude and
robots vibration frequency affects the sensor sensitivity in
the same way. But, the surface amplitude is the dominant
factor in affecting the sensor displacement for low frequency
Whiskers, vestibular system, and the cochlear are three
examples of compliant mechanical systems in biological
counterparts that solve realtime perception problems using
nonlinear vibration dynamics. Work done on such reservoir
computing systems show that even periodic sinusoidal exter-
nal perturbations can lead to complex steady state dynamics
in the compliant mechanical system including period bifur-
cation and frequency separation at local sites. This paper
shows for the first time that the steady state response of
nonlinear vibration dynamics can be used to classify textures
even on flat terrain. We also show that a mobile robot can
use speed control to move the perturbation frequencies to
elicit unique frequency domain responses in a whisker to help
terrain classification. Experiment results show that the novel
whiskered sensor and the deep multi-layer perceptron neural
network have good recognition and identification capabilities
of different terrain surfaces for mobile robots.
In the future, it will be interesting to investigate how
realtime stiffness control of the whisker can be used as a
control parameter to elicit steady state vibration frequency
components unique to the texture and low frequency geo-
metric features of a given terrain.
[1] R. Siegwart, I. R. Nourbakhsh, and D. Scaramuzza, Introduction to
autonomous mobile robots. MIT press, 2011.
[2] A. Howard and H. Seraji, “Vision-based terrain characterization and
traversability assessment,Journal of Robotic Systems, vol. 18, no. 10,
pp. 577–587, 2001.
[3] B. Suger, B. Steder, and W. Burgard, “Traversability analysis for
mobile robots in outdoor environments: A semi-supervised learning
approach based on 3d-lidar data,” in 2015 IEEE International Confer-
ence on Robotics and Automation (ICRA). IEEE, 2015, pp. 3941–
[4] A. Valada and W. Burgard, “Deep spatiotemporal models for robust
proprioceptive terrain classification,The International Journal of
Robotics Research, vol. 36, no. 13-14, pp. 1521–1539, 2017.
[5] F. G. Oliveira, E. R. Santos, A. A. Neto, M. F. Campos, and
D. G. Macharet, “Speed-invariant terrain roughness classification and
control based on inertial sensors,” in 2017 Latin American Robotics
Symposium (LARS) and 2017 Brazilian Symposium on Robotics (SBR).
IEEE, 2017, pp. 1–6.
[6] M. Fend, S. Bovet, H. Yokoi, and R. Pfeifer, “An active artifi-
cial whisker array for texture discrimination,” in Proceedings 2003
IEEE/RSJ International Conference on Intelligent Robots and Systems
(IROS 2003)(Cat. No. 03CH37453), vol. 2. IEEE, 2003, pp. 1044–
[7] L. A. Huet, J. W. Rudnicki, and M. J. Hartmann, “Tactile sensing
with whiskers of various shapes: determining the three-dimensional
location of object contact based on mechanical signals at the whisker
base,” Soft robotics, vol. 4, no. 2, pp. 88–102, 2017.
[8] M. Salman and M. J. Pearson, “Advancing whisker based navigation
through the implementation of bio-inspired whisking strategies,” in
2016 IEEE International Conference on Robotics and Biomimetics
(ROBIO). IEEE, 2016, pp. 767–773.
[9] T. J. Prescott, M. J. Pearson, B. Mitchinson, J. C. W. Sullivan, and
A. G. Pipe, “Whisking with robots,” IEEE robotics & automation
magazine, vol. 16, no. 3, pp. 42–50, 2009.
[10] M. Fend, S. Bovet, and V. Hafner, “The artificial mouse-a robot
with whiskers and vision,” in Proceedings of the 35th International
Symposium on Robotics (ISR 2004). Citeseer, 2004.
[11] M. Salman and M. J. Pearson, “Whisker-ratslam applied to 6d object
identification and spatial localisation,” in Conference on Biomimetic
and Biohybrid Systems. Springer, 2018, pp. 403–414.
[12] M. Brecht, B. Preilowski, and M. M. Merzenich, “Functional architec-
ture of the mystacial vibrissae,” Behavioural brain research, vol. 84,
no. 1-2, pp. 81–97, 1997.
[13] M. J. Pearson and M. Salman, “Active whisker placement and explo-
ration for rapid object recognition,” in 2019 IEEE/RSJ International
Conference on Intelligent Robots and Systems (IROS). IEEE, 2019,
pp. 672–677.
[14] M. Kaneko, N. Kanayama, and T. Tsuji, “Active antenna for contact
sensing,” IEEE Transactions on robotics and automation, vol. 14,
no. 2, pp. 278–291, 1998.
[15] D. B. Zurek and C. Gilbert, “Static antennae act as locomotory
guides that compensate for visual motion blur in a diurnal, keen-eyed
predator,Proceedings of the Royal Society B: Biological Sciences,
vol. 281, no. 1779, p. 20133072, 2014.
[16] J. H. Solomon and M. J. Hartmann, “Extracting object contours
with the sweep of a robotic whisker using torque information,” The
International Journal of Robotics Research, vol. 29, no. 9, pp. 1233–
1245, 2010.
[17] H. Wegiriya, N. Herzig, S.-A. Abad, S. H. Sadati, and T. Nanayakkara,
“A stiffness controllable multimodal whisker sensor follicle for texture
comparison,” IEEE Sensors Journal, vol. 20, no. 5, pp. 2320–2328,
[18] K. Nakajima, “Physical reservoir computing—an introductory perspec-
tive,Japanese Journal of Applied Physics, vol. 59, no. 6, p. 060501,
[19] K. Nakajima, H. Hauser, T. Li, and R. Pfeifer, “Information processing
via physical soft body,Scientific reports, vol. 5, no. 1, pp. 1–11, 2015.
[20] J. Tao and H. Cheng, “Dynamic stress response and fatigue
life of cantilever beam under non-Gaussian base excitation,
2014, iSSN: 1392-8716 Issue: 3 Number: 3 Pages: 1284-1296
Publisher: JVE International Ltd. Volume: 16. [Online]. Available:
ResearchGate has not been able to resolve any citations for this publication.
Full-text available
Mammals like rats, who live in dark burrows, heavily depend on tactile perception obtained through the vibrissal system to move through gaps and to discriminate textures. The organization of a mammalian whisker follicle contains multiple sensory receptors and glands strategically organized to capture tactile sensory stimuli of different frequencies. In this paper, we used a controllable stiffness soft robotic follicle to test the hypothesis that the multimodal sensory receptors together with the controllable stiffness tissues in the whisker follicle form a physical structure to maximize tactile information. In our design, the ring sinus and ringwulst of a biological follicle are represented by a linear actuator connected to a stiffness controllable mechanism in-between two different frequency-dependent data capturing modules. In this paper, we show for the first time the effect of the interplay between the stiffness and the speed of whisking on maximizing a difference metric for texture classification.
Conference Paper
Full-text available
Nowadays, it is notorious the increasing use of autonomous vehicles in different outdoor field applications such as agriculture, the mining industry, environmental monitoring, among others. In this context, the determination of the traversability level of a terrain is a fundamental task for safe and efficient navigation of Autonomous Ground Vehicles(AGVs) in unstructured unknown outdoor environments. Information like roughness of the ground is important when velocity control and other dynamic issues are concerned. However, most of the techniques in the literature use camera or lidar sensors to evaluate the ground around the robot, leading to complex and high cost systems. More simpler methods use only inertial sensors to estimate the roughness properties of the terrain, however they can be very sensitive to the robot’s speed. In this paper, we propose a novel classifier capable of cluster different terrains based only on acceleration data provided by an Inertial Measurement Unit (IMU). We demonstrate with real-world experiments that, for different forward velocities, the mean accuracy of the classification exceeds 80%. Our method also incorporates a controller to regulate the speed according to the terrain identified by the robot in order to avoid abrupt movements over terrain changes. (PDF) Speed-invariant terrain roughness classification and control based on inertial sensors. Available from: [accessed Feb 12 2020].
Full-text available
Terrain classification is a critical component of any autonomous mobile robot system operating in unknown real-world environments. Over the years, several proprioceptive terrain classification techniques have been introduced to increase robustness or act as a fallback for traditional vision based approaches. However, they lack widespread adaptation due to various factors that include inadequate accuracy, robustness and slow run-times. In this paper, we use vehicle-terrain interaction sounds as a proprioceptive modality and propose a deep long-short term memory based recurrent model that captures both the spatial and temporal dynamics of such a problem, thereby overcoming these past limitations. Our model consists of a new convolution neural network architecture that learns deep spatial features, complemented with long-short term memory units that learn complex temporal dynamics. Experiments on two extensive datasets collected with different microphones on various indoor and outdoor terrains demonstrate state-of-the-art performance compared to existing techniques. We additionally evaluate the performance in adverse acoustic conditions with high-ambient noise and propose a noise-aware training scheme that enables learning of more generalizable models that are essential for robust real-world deployments.
Full-text available
Soft machines have recently gained prominence due to their inherent softness and the resulting safety and resilience in applications. However, these machines also have disadvantages, as they respond with complex body dynamics when stimulated. These dynamics exhibit a variety of properties, including nonlinearity, memory, and potentially infinitely many degrees of freedom, which are often difficult to control. Here, we demonstrate that these seemingly undesirable properties can in fact be assets that can be exploited for real-time computation. Using body dynamics generated from a soft silicone arm, we show that they can be employed to emulate desired nonlinear dynamical systems. First, by using benchmark tasks, we demonstrate that the nonlinearity and memory within the body dynamics can increase the computational performance. Second, we characterize our system's computational capability by comparing its task performance with a standard machine learning technique and identify its range of validity and limitation. Our results suggest that soft bodies are not only impressive in their deformability and flexibility but can also be potentially used as computational resources on top and for free.
Full-text available
High visual acuity allows parallel processing of distant environmental features, but only when photons are abundant enough. Diurnal tiger beetles (Carabidae: Cicindelinae) have acute vision for insects and visually pursue prey in open, flat habitats. Their fast running speed causes motion blur that degrades visual contrast, forces stop-and-go pursuit and potentially impairs obstacle detection. We demonstrate here that vision is insufficient for obstacle detection during running, and show instead that antennal touch is both necessary and sufficient for obstacle detection. While running, tiger beetle vision appears to be photon-limited in a way reminiscent of animals in low-light habitats. Such animals often acquire wide-field spatial information through mechanosensation mediated by longer, more mobile appendages. We show that a nocturnal tiger beetle species waves its antennae in elliptical patterns typical of poorly sighted insects. While antennae of diurnal species are also used for mechanosensation, they are rigidly held forward with the tips close to the substrate. This enables timely detection of path obstructions followed by an increase in body pitch to avoid collision. Our results demonstrate adaptive mechanosensory augmentation of blurred visual information during fast locomotion, and suggest that future studies may reveal non-visual sensory compensation in other fast-moving animals.
Understanding the fundamental relationships between physics and its information-processing capability has been an active research topic for many years. Physical reservoir computing is a recently introduced framework that allows one to exploit the complex dynamics of physical systems as information-processing devices. This framework is particularly suited for edge computing devices, in which information processing is incorporated at the edge (e.g. into sensors) in a decentralized manner to reduce the adaptation delay caused by data transmission overhead. This paper aims to illustrate the potentials of the framework using examples from soft robotics and to provide a concise overview focusing on the basic motivations for introducing it, which stem from a number of fields, including machine learning, nonlinear dynamical systems, biological science, materials science, and physics.
Almost all mammals use their mystacial vibrissae (whiskers) as important tactile sensors. There are no sensors along the length of a whisker: All sensing is performed by mechanoreceptors at the whisker base. To use artificial whiskers as a sensing tool in robotics, it is essential to be able to determine the three-dimensional (3D) location at which a whisker has made contact with an object. With the assumption of quasistatic, frictionless, single-point contact, previous work demonstrated that the 3D contact point can be uniquely determined if all six components of force and moment are measured at the whisker base, but these measurements require a six-axis load cell. Here, we perform simulations to investigate the extent to which each of the 20 possible "triplet" combinations of the six mechanical signals at the whisker base uniquely determine 3D contact point location. We perform this analysis for four different whisker profiles (shapes): Tapered with and without intrinsic curvature, and cylindrical with and without intrinsic curvature. We show that whisker profile has a strong influence on the particular triplet(s) of signals that uniquely map to the 3D contact point. The triplet of bending moment, bending moment direction, and axial force produces unique mappings for tapered whiskers. Four different mappings are unique for a cylindrical whisker without intrinsic curvature, but only when large deflections are excluded. These results inform the neuroscience of vibrissotactile sensing and represent an important step toward the development of artificial whiskers for robotic applications.
The stress response of cantilever beam to non-Gaussian random base excitation isinvestigated based on Monte-Carlo simulation. First, the statistical properties and spectralcharacteristics of non-Gaussian random vibrations are analyzed qualitatively; and the conclusionis that spectral method based on power spectrum density (PSD) is not applicable for non-Gaussianrandom vibrations. Second, the stress response formula of cantilever beam under non-Gaussianrandom base excitations is established in the time-domain, and the factors influencing the outputkurtosis are subsequently determined. Two numerical examples representing different practicalsituations are analyzed in detail. The discrepancies of the stress responses to Gaussian, steadynon-Gaussian and burst non-Gaussian base excitations are analyzed in terms of root mean square(RMS), kurtosis and fatigue damage. The transmissibility of RMS and high-kurtosis of steadynon-Gaussian random base excitation is different from that of burst non-Gaussian case. Finally,the fatigue life corresponding to every base excitation is calculated using the rainflow method inconjunction with the Palmgren-Miner rule. Finite element analysis is also carried out forvalidation. The predicted fatigue lives corresponding to Gaussian, steady non-Gaussian and burstnon-Gaussian base excitations are compared quantitatively. Finally, in the fatigue damage pointof view, the discrepancies among the three kinds of random base excitations are summarized.