TacWhiskers: Biomimetic optical tactile whiskered robots
Nathan F. Lepora, Martin Pearson, Luke Cramphorn
Abstract— Here we propose and investigate a novel vibrissal
tactile sensor - the TacWhisker array - based on modifying
a 3D-printed optical cutaneous (ﬁngertip) tactile sensor - the
TacTip. Two versions are considered: a static TacWhisker array
analogous to immotile tactile vibrissae (e.g. rodent microvib-
rissae) and a dynamic TacWhisker array analogous to motile
tactile vibrissae (e.g. rodent macrovibrissae). Performance is as-
sessed on an active object localization task. The whisking motion
of the dynamic TacWhisker leads to millimetre-scale location
perception, whereas perception with the static TacWhisker
array is relatively poor when making dabbing contacts. The
dynamic sensor output is dominated by a self-generated motion
signal, which can be compensated by comparing to a reference
signal. Overall, the TacWhisker arrays give a new class of tactile
whiskered robots that beneﬁt from being relatively inexpensive
and customizable. Furthermore, the biomimetic basis for the
TacWhiskers ﬁts well with building an embodied model of the
rodent sensory system for investigating animal perception.
I. INTRODUCTION
Animal sensory organs have inspired diverse artiﬁcial
sensing systems, from chemosensing based on olfaction to
whiskered robots based on the vibrissal sense of rodents.
We are intimately familiar with our own human sense of
touch, which underlies our abilities to manipulate the world
and interact with other human beings. Yet for rodents and
some other mammals, their primary tactile sense is from
vibrissae, or tactile hair, which functions as a proximity
sense for navigation and to catch prey. Whiskered robots
are of interest to deploy this proximity sense in applications
where vision is compromised, such as disaster recovery, and
as biomimetic instantiations to investigate the physiological
and neuroscientiﬁc principles underlying natural sensing [1].
Accordingly, whiskered robots are an active area of re-
search, with state-of-the-art devices including the SHREW-
bot and BELLAbot robots [2], [3]. While these whiskered
robots are excellent biomimetic counterparts of biological
vibrissal systems, they are highly complex, one-off robots
that are expensive and labour-intensive to make; for example,
the SHREWbot nose cone has 24 whisker modules, each
having an actuated base with a Hall effect sensor to measure
whisker deﬂection. Meanwhile, within the related area of
artiﬁcial cutaneous (ﬁngertip) touch, there has been progress
towards 3D-printed, optical tactile sensors such as the TacTip
family [4], [5], which are inexpensive and simple to fabricate.
Here we propose a new class of 3D-printed, optical tactile
whisker sensors that we call TacWhiskers, based on the Tac-
NL was supported in part by a grant from the Leverhulme Trust on ‘A
biomimetic forebrain for robot touch’ (RL-2016-39).
All authors are with the Department of Engineering Mathematics and
Bristol Robotics Laboratory, University of Bristol, Bristol, UK.
Email: n.lepora@bristol.ac.uk
Fig. 1.
Side (left image) and front (right image) views of the dynamic
TacWhisker array mounted on an ABB robot arm. The actuation module,
sensor body and tip are visible, with the tendon that protracts the whiskers.
Tip design. We consider two versions: a static TacWhisker
array analogous to immotile tactile vibrissae (e.g. rodent
microvibrissae) and a dynamic TacWhisker array analogous
to motile tactile vibrissae (e.g. rodent macrovibrissae). The
dynamic TacWhisker uses a single motor to protract and
retract its whiskers in a rhythmic whisking motion, using
a tendon passing between two rows of whiskers (Fig. 1).
To assess the TacWhisker performance, we consider a
localization task where the position of a rod is classiﬁed
from whisker deﬂections, analogously to an experimental
protocol for rodent perception [6]. We ﬁnd the accuracy of
TacWhisker perception depends strongly on the whisker mo-
tion: the static TacWhisker has poor location perception with
a forward/back dabbing motion and the dynamic TacWhisker
has good location perception when whisking. These motions
are then applied to an active perception task, in which the
rod is both localized and centred within the whisker array.
Only the dynamic TacWhisker is able to perform well at this
task, with trajectories quickly centring on the stimulus and
localization performance near perfect after a few contacts.
II. BACKGROUND AND RELATED WORK
A. Biological sensing with tactile hair
Almost all mammals except Homo Sapiens have a spe-
cialised form of tactile hair called vibrissae or whiskers [8].
Whiskers differ from conventional (pelagic) hair by be-
ing [8]: (i) much longer; (ii) primarily facial (although they
can also occur on the body); (iii) sited on a large, highly-
arXiv:1810.00460v2  [cs.RO]  7 Dec 2018
Fig. 2.
Physiology of tactile vibrissae - the follicle-sinus complex from
which the whisker shaft emanates, and within which are the vibrissal sensory
mechanoreceptors. (Diagram modiﬁed from [7] under a Creative Commons
Attribution-ShareAlike 3.0 License.)
innervated follicle-sinus complex (Fig. 2); and (iv) speciﬁ-
cally represented in the sensory cortex of the brain.
Tactile whiskers may be motile or immotile, depending on
the animal and where on the body they are located [8], [9].
In mice and rats, the long facial whiskers (macrovibrissae)
around the snout move bilaterally back and forth in an
active sensing motion known as ‘whisking’; meanwhile, the
short facial whiskers (microvibrissae) underneath the nostrils
are ﬁxed. Immotile whiskers are also found on other body
regions; for example, the carpal vibrissae just above cat paws.
Sensory mechanoreceptors within the whisker follicle
transduce motion of the whisker shaft into contact infor-
mation about the environment [7]. Merkel cells in a collar
around the follicle opening activate slowly-adapting (SA)
neurons that ﬁre during sustained whisker deformation.
Deeper receptors such as Lanceloate, Rufﬁni and Paccinian
endings within the whisker follicle act in concert with the
Merkel cells, signalling information about vibration, motion
and whisker deﬂection, to comprise the vibrissal tactile sense.
B. Biomimetic tactile whiskered robots
Over the last decade there have been a succession of
biomimetic tactile whiskered robots developed from a col-
laboration between Shefﬁeld Robotics and Bristol Robotics
Laboratory [1]. The initial Whiskerbot mobile platform had
6 glass-ﬁbre moulded whiskers mounted on strain gauges
to measure 2D deﬂections of the whisker shaft [10]. An
improved SCRATCHbot platform had 18 actuated 3D-printed
whiskers with Hall effect sensors to measure deﬂections
while actively whisking [11]. These single-actuated whiskers
were modularized as part of the BIOTACT project, leading
(a) TacTip transduction
(b) TacWhisker transduction
Fig. 3.
Common transduction principle of the TacTip and TacWhiskers.
(a) For the TacTip, surface strain separates the inner pins. (b) For the
TacWhisker, whisker shaft deformation deﬂects the pins. In both cases, the
movement of markers on the pin tips is tracked by a camera.
to another mobile whiskered platform called Shrewbot [2]
and a stand-alone whisker array for mounting on a robot
arm [12], both with 24 individually-actuated whiskers from
∼5-15 cm long arranged in a conical 6-by-4 array.
Other technologies have been used for whiskered robots.
An early mobile robot from the aMouse project attached real
rodent whiskers to microphones for detecting high-frequency
vibrations [13]. Soon after, a whisker array was built with
both slowly adapting (deﬂection) and rapidly adapting (ve-
locity) components of the whisker signal and used to recog-
nize shape [14]. A more recent robot, BELLAbot, combined
the BIOTACT technology with advances in electroactive
polymeric (EAP) actuation to orient, whisk and sense over
an array of 20 distinct EAP whisker modules [3]. Simpler
biomimetic whiskers have also been proposed, such as using
strain gauges in the follicle to measure the bending moments
proposed to underlie localization with rodent vibrissae [15].
C. Biomimetic optical tactile sensors
This paper investigates a novel vibrissal tactile sensor
based on modifying a 3D-printed cutaneous (ﬁngertip) tactile
sensor called the TacTip (see [5] for a recent overview).
The TacTip is a biomimetic tactile sensor based on the
layered structure of human glabrous skin [4]. It has an outer
biomimetic epidermis made from a rubber-like material over
an inner biomimetic dermis made from polymer gel. These
two materials interdigitate in a mesh of inner nodular pins,
based on the intermediate ridge structure of human skin that
extends under the epidermis into the dermis. The biomimetic
counterparts to sensory mechanoreceptors are markers on the
pin tips, which can be imaged through a transparent gel that
comprises the dermis. The pin movement is considered anal-
ogous to Merkel cell activity in the intermediate ridges [16].
A principal observation underlying this paper is that the
transduction mechanism in the TacTip can also be applied to
tactile whiskers (Fig 3). Information about how the TacTip
surface deforms upon contacting an object is represented in
the movement of markers on the internal pins, e.g. separating
on regions of high spatial curvature (Fig. 3a). Instead, for
internal pins attached to whiskers extending out of the
(a) CAD for the whisker
(b) CAD for the tip
(c) CAD for the base
(d) Static TacWhisker array
Fig. 4.
Design of the static TacWhisker array. (a) CAD for the 3D-printed
whisker, a tapered shaft. (b) CAD for the tip, comprising a compliant skin
with sockets for the whiskers and a rigid base. (c) CAD for the base housing
for the camera, showing also the LED ring. (d) Assembled TacWhisker array.
sensor surface, the markers represent displacement of the
whisker shafts (Fig. 3b). As the same receptor type (Merkel
cells) is implicated in both types of biological sensing, this
mechanism also gives a commonality in the biomimetics of
cutaneous and vibrissal touch (although it should be noted
that a multitude of other mechanoreceptor types are also
involved in both types of touch).
III. METHODS
A. Inspiration and design of the TacWhisker
We call the whiskered version of the TacTip a TacWhisker
array, emphasising it is based on tactile whiskers rather than
tactile (ﬁnger)tips. In this work, we consider two types of
TacWhisker array: static (immotile) and dynamic (motile).
1) Static TacWhisker array (Fig. 4): The ﬁrst design of
TacWhisker array modiﬁes just the standard TacTip tip to
house whiskers (Figs 4a,b). There is no modiﬁcation of the
3D-printed TacTip base (Fig. 4c), which contains the USB
camera (Lifecam, Microsoft) and an LED ring to illuminate
the pin markers (see [5] for details). The tip is based on
recent versions of the TacTip [5] that use multi-material 3D-
printing. The compliant surface and inner pins printed in
a rubber-like material (Tango Black+ 27) and the pin tips
and mount in hard plastic (Vero White); this outer surface is
ﬁlled with a soft clear silicone gel (Techsil RTV27905) held
in place with a clear acrylic lens cap.
For housing whiskers, the tip is modiﬁed to (Fig. 4b):
(i) reduce the number of pins to 21 (from 127) sited
near the top of the tip; (ii) space the pins further apart
(4.5 mm separation, rather than 3 mm keeping the hexago-
nal projection); (iii) enlarge and extrude the solid markers
through the compliant surface (2.2 mm dia.×3.5 mm depth
Fig. 5.
Design of the dynamic TacWhisker array. The 3D-printed tip with
mounted whiskers attaches to the base housing the camera, which attaches
to the actuation module comprising the motor and housing. A tendon runs
from the spool, through guides and across a groove in the compliant tip;
actuation of the tendon causes the tip to deform, moving the whiskers.
(a) Active protraction
(b) Passive retraction
Fig. 6.
Whisking motion of the dynamic TacWhisker array. (a) The motor
pulls on the tendon to actively protract (bring together) the two rows of
whiskers. (b) Reversing the motor releases the tendon to passively retract
(pull apart) the whiskers by elastic reformation of the tip.
pins, increased from 1.2 mm×2 mm); and (iv) include a hole
(1 mm dia.×3 mm depth) functioning as a socket for the
whiskers. These design choices were chosen to give good
pin movement upon deﬂection of the whiskers, and to site
the whiskers appropriately for contact.
The whiskers (Fig. 4a) are modiﬁed versions of BIOTACT
vibrissae [2] that are 3D printed using nanocure-25. The
main change is to reduce the whisker size for the smaller
Fig. 7.
Data processing pipeline. The internal camera captures an image of the pins attached to the shafts of the whiskers. The pins are detected and
located with a blob detection algorithm. The pins are then ordered, by tracking the pins from frame-to-frame (here coloured by their row). Over multiple
frames, the processing pipeline produces a time series of pin movement, representing deﬂection of the whiskers.
scale of the TacTip (40 mm dia.) compared with the BIO-
TACT conical housing (100 mm dia.). Accordingly, we chose
whiskers 40 mm long with a 0.98 mm dia. base tapering to
0.6 mm dia. at the tip, similar in scale to real rat whiskers.
For simplicity, all whiskers had the same dimensions, but
it would be straightforward to introduce size variations like
those of rodent macrovibrissae.
2) Dynamic TacWhisker array (Fig. 5): The TacWhisker
array can join onto an actuation module that protracts and
retracts the whiskers back and forth in a whisking motion
(Fig. 6). The whiskered tip is modiﬁed to have 2 rows of
5 whiskers arranged in a bilaterally symmetric pattern. A
tendon runs through a groove between these rows and two
guides in the tip mount (Fig. 6). Forwards whisker motion
(protraction) results from tensioning the tendon to compress
the surface at the midline (Fig. 6a); backwards whisker
motion (retraction) results from releasing the tendon to elasti-
cally reform the surface (Fig. 6b). The compliant surface and
whisker mounts are shaped so that the whisker tips can meet
under modest surface compression. The dynamic TacWhisker
array can thus rhythmically protract and retract its whiskers
together and apart in a motion akin to rodent whisking.
The dynamic TacWhisker is designed to be modular and
re-use parts of the static TacWhisker. Apart from the modi-
ﬁed whiskered tip, the TacWhisker base housing the camera
and LED lighting is the same as the conventional TacTip.
The underside of the base has a bayonet ﬁtting, which is
used to connect to an actuation module for driving the tendon
(Fig. 5). This actuation module houses a Dynamixel MX 28
servomotor and spool for the tendon, with outer guides to
ensure the tendon runs smoothly from the spool, outside the
actuation and body modules, and over the TacWhisker tip.
B. Robotic platform and software infrastructure
For testing, the static or dynamic TacWhisker array is
mounted as an end-effector on a 6-DOF robotic arm (IRB
120, ABB) (e.g. Fig. 1). The arm is mounted on a table that
also contains mounting stations for the stimuli. A custom
3D-printed mount is bolted to the rotating (wrist) section of
the arm to which either sensor can be attached via a common
bayonet ﬁtting on the TacWhisker base and actuation module.
A modular software infrastructure is used in which MAT-
LAB is the primary interface for running tests and analysing
data. The ABB arm is controlled via an IronPython and
RAPID interface, and data gathered from the USB cam-
era within the TacWhisker sensor with Python OpenCV.
Similarly, a Python interface controls the dynamixel motor
of the dynamic TacWhisker array. Communication between
software modules is via TCP/IP ports and sockets.
C. Sensory transduction and data processing
1) Sensing: Following recent studies with the TacTip [5],
[17], the sensor output is a time series of pin deﬂections
extracted from the camera images. The transformation of
the camera image to marker positions requires that the
pin markers be detected, which is done via standard ‘blob
detection’ methods in Python OpenCV. Overall, the data
processing is a pipeline: camera image to pin detection to pin
identiﬁcation (nearest neighbour tracking) to give an ordered
time series of pin deﬂections measured in pixels (Fig. 7).
The resulting tactile whisker data comprises a multi-
dimensional time series of (x, y) pin deﬂections, measured in
pixels on the horizontal and vertical directions of the camera
image. For visualization, the time series plots of the x- and
y-deﬂections are labelled by colouring the tactile dimension
by its pin location (Fig. 7, right plots).
2) Perception: Tactile perception is the process of infer-
ring the properties of a stimulus from data collected by
contacting that stimulus. Here we use a likelihood model
that transforms tactile data D into a likelihood probability
P(D | Hi) for a set of perceptual hypotheses {H1, ..., HN},
which could be the labels (e.g. location xi in mm) for training
data used to construct the model. The perceptual decision is
then the hypothesis Hi = arg maxi P(D | Hi) that has the
maximum likelihood for some sensed tactile data D.
Following recent studies with the TacTip, here we use a
histogram likelihood model [18], [19], which bins the sensor
data into intervals and counts bin frequency to form sampling
distributions that are multiplied over sensor dimension and
time. While simple, this model is effective for the TacTip
and other sensors [17], [19], bears analogy with neural
processing [19] and is fairly robust and efﬁcient. That said,
(a) Static TacWhisker array with dabbing motion
(b) Dynamic TacWhisker array with whisking motion
(c) Dynamic TacWhisker array with self-motion calibration
Fig. 8.
Location data collected from the static (a) and dynamic (c,d) TacWhisker arrays. In all cases, the sensor is moved across a horizontal range from
left to right (static: 50 mm range; dynamic: 40 mm range). The plot colour denotes the identity of the whiskers (right images).
the likelihood model is not the focus of this study, and so any
model that works reasonably well would have been sufﬁcient.
All quantitative analyses of perceptual accuracy in this
paper are based on cross validation over 10 repeated runs for
data collection (representative single sets shown in Fig. 8).
Monte Carlo sampling of a randomly-chosen training set and
test data chosen from a different random set is then used.
Typically 10,000 samples are used per analysis.
3) Active
perception:
We
follow
the
approach
of
‘biomimetic active touch for ﬁngertips and whiskers’ [19]
in which active perceptual decisions are sequential over
multiple tactile contacts D(1), · · · , D(T) with actions made
between contacts to fulﬁl a goal, such as centring the sensor
on a stimulus. Bayes rule is applied recursively to the
likelihoods P(D | Hi) to integrate evidence over contacts
P(Hi|D(t)) =
P(D(t)|Hi)P(Hi|D(t −1))
P
j P(D(t)|Hj)P(Hj|D(t −1))
beginning from ﬂat priors P(Hi|D(0)) = P(Hi) = 1/N.
Here we use a simple active perception policy in which
actions move the tactile sensor towards a goal location xﬁx =
HN/2, here taken as the stimulus centre. Then the actions are
translations ∆x(t) = (xﬁx −xj(t)), with xj the jth location
class j = arg maxi P(D(t)|Hi). The decision is made when
the probability crosses a threshold P(Hi|D(t)) > θ that sets
how many contacts (on average) are needed for a decision.
IV. RESULTS
A. Inspection and comparison of TacWhisker data
Whisker contact data from two distinct experimental situa-
tions were collected (Figs 8). We chose motions in which the
TacWhisker made discrete contacts with the stimulus, with
the whiskers leaving the stimulus between contacts:
(a) Static TacWhisker array with dabbing motion. The sen-
sor was moved horizontally across a rod keeping, dabbing
vertically down (15 mm) to make a tapping contact onto the
rod stimulus, with 50 taps across 50 mm.
(b) Dynamic TacWhisker array with whisking motion. The
experiment was repeated using the dynamic array to whisk
onto the rod stimulus, with 40 whisks across 40 mm. (The
shorter range was due to to a narrower whisker ﬁeld.)
A further dataset was created from modifying set (b):
(c) Dynamic TacWhisker array with self-motion calibration.
A reference signal (taken at the centre of the location range)
was subtracting from the whisking data; this calibration
makes a whisker contact more visually apparent since the
self-motion dominates otherwise.
In all experiments, good quality data were obtained from
the TacWhisker arrays, as evident in the smoothly varying
plots in Fig. 8 with signal dominating over noise. Further-
more, the sensor data clearly covaries with contact location,
which is especially evident in the dynamic array calibrated
for self-motion (Fig. 8c), suggesting they will accurately
perceive location. However, as we will explore below, the
manner in which the whisker contacts the stimulus is key
for perception, as is evident in the signiﬁcant differences
between the data in Figs 8a,b at the same locations.
B. Self-motion affects the TacWhisker data
A principal difference of the dynamic from the static
TacWhisker arrays is that the tactile data is strongly affected
by the self-motion of the dynamic array: the whiskers sweep-
ing forwards then back is the most prominent feature of the
(a) Static TacWhisker
dabbing motion
(b) Dynamic TacWhisker
(c) Dynamic TacWhisker
whisking motion
calibrated self-motion
Fig. 9.
Accuracy of location perception for the static (a) and dynamic (c,d)
TacWhisker arrays. Monte Carlo 10-fold cross validation (10000 samples)
is shown by plotting the the perceived against ground truth locations (red
markers). The variability of the location perception is shown between 25th
and 75th percentiles (gray region).
data (Fig. 8b). This self-motion effect is also a known aspect
of biological vibrissal systems (see discussion). Since the
whisking motion remains constant frequency and amplitude
in the present experimental task, its effect on the tactile data
can be removed by subtracting a reference signal, which we
choose to be at the centre of the location range. This gives
the self-motion compensated tactile data (Fig. 8c), where the
changing contact with the stimulus is clearly evident.
C. Tactile perception depends on TacWhisker motion
The accuracy of location perception in the three experi-
mental conditions was then assessed with a standard classiﬁer
of tactile data based on a histogram likelihood model (details
in Sec. III-C). Monte Carlo cross validation over 10 repeated
runs was used to generate distributions of the perceived class
label against the ground truth class label (Fig. 9).
For the static TacWhisker array (Fig. 9a), the location
perception is variable for the dabbing motion (interquartile
range IQR = 12 mm). This agrees with qualitative inspection
of the data, which on has a sparse, unstructured structure as
the rod hits or misses vibrissae (Fig. 8a).
For the dynamic TacWhisker array (Figs 9b,c), the location
perception is accurate with little variability (interquartile
range, IQR = 1.5 mm), and similar results for the uncom-
pensated and self-motion compensated tactile data. It seems
a basic compensation of self-motion does not help the per-
ception, although more sophisticated methods (e.g. adaptive
(a) Static TacWhisker
(b) Dynamic TacWhisker
dabbing motion
whisking motion
Fig. 10.
Trajectories for actively localizing a stimulus with TacWhisker
sensor for the static (a) and dynamic (b) arrays. Trajectories begin from
random locations and aim towards the central location (dashed red line). In
both cases, a posterior threshold θ = 0.5 was used to make a decision.
Dynamic TacWhisker
Location error (active vs passive)
Fig. 11.
Mean location errors for active and passive touch. Active
localization is over multiple contacts while centring on the object, with
either a posterior threshold (range 0-0.95; red histogram) or set decision
time (1-10 contacts; black histogram). Passive localization does not move
the sensor. Averages are over 1000 runs with random starting locations.
noise cancellation [20]) may be better. The compensation
does reveal a signiﬁcant covariation of the tactile data with
location (Fig. 8c), consistent with the accurate perception.
D. The dynamic TacWhisker aids active touch
The location perception was then applied to a simple task
in which the TacWhisker actively localizes a stimulus while
using intermediate estimates of the object location to centre
it within the whisker array (Sec III-C). This task is a simple
example of active touch [19] and bears analogy with rodent
behaviour when exploring stimuli.
For the static TacWhisker array (Fig. 10a), the trajectories
do not converge on the central location. It appears the quality
of the perception from a dabbing motion is not sufﬁcient to
actively localize the stimulus within the whisker ﬁeld.
The dynamic TacWhisker array (Fig. 10b) achieves suc-
cessful active localization across the range of starting loca-
tions. All trajectories converge on the central location.
Repeating over many trials, the mean active localization
errors improve with mean decision time (Fig. 11, red his-
togram), reaching near perfect accuracy after ∼5 contacts
with the threshold-crossing decision rule. A ﬁxed-time rule
also improves with decision time but is not as accurate for
longer decision times (black histogram). Conversely, mean
errors for passive perception do not improve because the
robot cannot move to gather new data (white histogram).
Overall, the dynamic TacWhisker array with active lo-
calization using a posterior threshold-crossing decision rule
gives the best localization performance.
V. DISCUSSION
In this paper, we introduced two novel whiskered robots:
the static and dynamic TacWhisker arrays (Figs 4,5). These
robots are biomimetic in being based on the sensory trans-
duction principles of biological vibrissae, bearing analogy
with mechanoreception by Merkel cells in the whisker
follicle. We based the designs on a 3D-printed cutaneous
(ﬁngertip) optical tactile sensor called the TacTip [4], [5],
which is based on sensory transduction in skin via Merkel
cell mechanoreceptors. The static TacWhisker array modiﬁes
the TacTip skin to house 21 whiskers arranged around its
tip. The dynamic TacWhisker array is actuated to move its
whiskers back and forth in a whisking motion, with the
whiskers arranged bilaterally in 2 rows of 5 whiskers.
The performance of the TacWhisker sensors was examined
by perceiving the location of a rod, motivated by similar
experiment quantifying rodent perception [6]. The quality
of the perception depended strongly on the whisker motion.
For the static TacWhisker array, a forward dabbing motion
was inaccurate and variable (IQR∼10 mm), consistent with
sparse, unstructured contacts. For the dynamic TacWhisker
array, the whisking motion resulted in accurate and reliable
perception (IQR∼1.5 mm).
In consequence, only the dynamic TacWhisker array could
perform an active localization task in which stimulus location
is estimated while centring the object in the whisker array.
Performance improved from a mean error of ∼1 mm after 1
contact to perfect performance after 5 contacts, in accordance
with previous studies of active touch [19]. Other active local-
ization strategies could work better for the static TacWhisker
(e.g. avoiding locations where the sensing is bad); this would
however require a more complex policy for moving the
sensor that would need learning. It is also possible that other
exploratory motions for the static TacWhisker would improve
performance; however, active localization would be far more
complex if the contacts were not discrete and independent,
as provided by a dabbing or whisking motion.
A potential issue is that the dynamic TacWhisker signal is
dominated by the self-generated deﬂection of the whiskers.
This can be partially compensated by subtracting a reference
signal (here taken from the central contact); although this did
not affect perception, it did improve the visible interpretation
of the data. This effect and compensation mechanism are
known from animal investigations and have also been con-
sidered with the SCRATCHbot whiskered robot [20]. Note
that the TacWhisker has greater self-generated signal than
both animal and SCRATCHbot whiskers, since the dynamic
TacWhisker is directly affected by its actuation (and not just
the inertia of its whiskers).
Overall, the TacWhisker arrays give a new class of optical
tactile whiskered robots, related to optical tactile sensors that
have progressed in recent years [5]. They have the beneﬁt of
being relatively inexpensive, readily customizable and easy
fabrication using multi-material 3D-printing. We expect they
can be applied to some of the application domains of tactile
whiskers, such as proximity and ﬂow sensing. Furthermore,
the biomimetic basis for the TacWhiskers lends them to
neurorobotic investigations of rodent tactile vibrissal sensing
as an embodied model of animal perception.
Acknowledgements: We thank Niels Burnus, Yilin Tao, and
members of the Tactile Robotics group: Ben Ward-Cherrier,
Nick Pestell, Kirsty Aquilina, Jasper James and John Lloyd.
REFERENCES
[1] T. Prescott, M. Pearson, B. Mitchinson, J. Sullivan, and A. Pipe.
Whisking with robots. IEEE Robotics & Automation Magazine, 16:42–
50, 2009.
[2] M. Pearson, B. Mitchinson, J. Sullivan, A. Pipe, and T. Prescott.
Biomimetic vibrissal sensing for robots. Philosophical Transactions
of the Royal Society B: Biological Sciences, 366(1581):3085–3096,
2011.
[3] T. Assaf, E. Wilson, S. Anderson, P. Dean, J. Porrill, and M. Pearson.
Visual-tactile sensory map calibration of a biomimetic whiskered
robot. IEEE International Conference on Robotics and Automation,
pages 967–972, 2016.
[4] C. Chorley, C. Melhuish, T. Pipe, and J. Rossiter. Development of
a Tactile Sensor Based on Biologically Inspired Edge Encoding. In
International Conference on Advanced Robotics (ICAR), pages 1–6,
2009.
[5] B. Ward-Cherrier, N. Pestell, L. Cramphorn, B. Winstone, M. E.
Giannaccini, J. Rossiter, and N. F. Lepora. The TacTip Family: Soft
Optical Tactile Sensors with 3D-Printed Biomimetic Morphologies.
Soft Robotics, pages 1–13, 2018.
[6] M. Diamond, M. Von Heimendahl, P. Knutsen, D. Kleinfeld, and
E. Ahissar. ’Where’ and ’what’ in the whisker sensorimotor system.
Nature Reviews Neuroscience, 9(8):601–612, 2008.
[7] S. Ebara, T. Furuta, and K. Kumamoto. Vibrissal mechanoreceptors.
Scholarpedia, 12(3):32372, 2017.
[8] A. Ahl. The role of vibrissae in behavior: A status review. Veterinary
Research Communications, 10(1):245–268, 1986.
[9] T. Prescott, B. Mitchinson, and R. Grant.
Vibrissal behavior and
function. Scholarpedia, 6(10):6642, 2011.
[10] M. Pearson, A. Pipe, C. Melhuish, B. Mitchinson, and T. Prescott.
Whiskerbot: A robotic active touch system modeled on the rat whisker
sensory system. Adaptive Behavior, 15(3):223–240, 2007.
[11] M. Pearson, B. Mitchinson, J. Welsby, T. Pipe, and T. Prescott.
SCRATCHbot: Active Tactile Sensing in a Whiskered Mobile Robot.
In From Animals to Animats, volume 10, pages 93–103, 2010.
[12] J. Sullivan, B. Mitchinson, M. Pearson, M. Evans, N. Lepora, C. Fox,
C. Melhuish, and T. Prescott.
Tactile discrimination using active
whisker sensors. IEEE Sensors Journal, 12(2), 2012.
[13] M. Fend, S. Bovet, and V. Hafner. The artiﬁcial mouse-a robot with
whiskers and vision. Signal [V], 23, 2005.
[14] D. Kim and R. M¨oller. Biomimetic whiskers for shape recognition.
Robotics and Autonomous Systems, 55(3):229–243, 2007.
[15] H. Emnett, M. Graff, and M. Hartmann. A Novel Whisker Sensor
Used for 3D Contact Point Determination and Contour Extraction.
Proceedings of Robotics: Science and Systems, XIV, 2018.
[16] L. Cramphorn, B. Ward-Cherrier, and N. Lepora.
Addition of a
Biomimetic Fingerprint on an Artiﬁcial Fingertip Enhances Tactile
Spatial Acuity. IEEE Robotics and Automation Letters, 2(3):1336–
1343, 2017.
[17] N. Lepora and B. Ward-Cherrier.
Superresolution with an optical
tactile sensor. In IEEE/RSJ International Conference on Intelligent
Robots and Systems (IROS), pages 2686–2691, 2015.
[18] N. Lepora, J. Sullivan, B. Mitchinson, M. Pearson, K. Gurney, and
T. Prescott. Brain-inspired Bayesian perception for biomimetic robot
touch. IEEE International Conference on Robotics and Automation,
pages 5111–5116, 2012.
[19] N. Lepora. Biomimetic Active Touch with Fingertips and Whiskers.
IEEE Transactions on Haptics, 9(2), 2016.
[20] S. Anderson, M. Pearson, A. Pipe, T. Prescott, P. Dean, and J. Porrill.
Adaptive cancelation of self-generated sensory signals in a whisking
robot. IEEE Transactions on Robotics, 26(6):1065–1076, 2010.