Content uploaded by Bastian Wibranek
Author content
All content in this area was uploaded by Bastian Wibranek on Nov 05, 2019
Content may be subject to copyright.
Building a Library of Tactile Skills Based on FingerVision
Boris Belousov*, Alymbek Sadybakasov*, Bastian Wibranek, Filipe Veiga, Oliver Tessmann, and Jan Peters
Abstract— Camera-based tactile sensors are emerging as a
promising inexpensive solution for tactile-enhanced manipula-
tion tasks. A recently introduced FingerVision sensor was shown
capable of generating reliable signals for force estimation,
object pose estimation, and slip detection. In this paper, we
build upon the FingerVision design, improving already existing
control algorithms, and, more importantly, expanding its range
of applicability to more challenging tasks by utilizing raw skin
deformation data for control. In contrast to previous approaches
that rely on the average deformation of the whole sensor
surface, we directly employ local deviations of each spherical
marker immersed in the silicone body of the sensor for feedback
control and as input to learning tasks. We show that with
such input, substances of varying texture and viscosity can be
distinguished on the basis of tactile sensations evoked while
stirring them. As another application, we learn a mapping
between skin deformation and force applied to an object. To
demonstrate the full range of capabilities of the proposed
controllers, we deploy them in a challenging architectural
assembly task that involves inserting a load-bearing element
underneath a bendable plate at the point of maximum load.
I. INTRODUCTION
Endowing robots with a sense of touch is a long-standing
problem in robotics [1]. A variety of sensor designs were
proposed in the past [2]. An early camera-based solution [3]
measured shear force by tracking reflective cones protruding
from an opaque surface of a polymer. To get a full 3D
vector of force, the GelForce sensor [4] used a camera
to track two layers of colored dots immersed in a piece
of elastomeric gel with an opaque coating. The GelSight
sensor [5] has a similar design to GelForce but in addition it
can estimate 3D topography of a contact surface and it has a
much higher resolution. Unlike the previous sensors, which
had an opaque cover and illuminated the sensor skin from
within, the FingerVision sensor [6] is transparent and relies
on external lighting for illumination. Such design allows
capturing object properties such as shape and color even
before contact. Moreover, slip detection turns into a well-
studied computer vision problem of motion estimation.
In order to efficiently use a tactile sensor for robot control,
sensor data needs to be passed to a control system in an
appropriate form. A preprocessing pipeline for FingerVision
together with a number of tactile feedback controllers was
proposed in [7]. Various behaviors, such as gentle grasp,
*Authors contributed equally
B.B., A.S., and J.P. are with Intelligent Autonomous Systems Lab, TU
Darmstadt, Germany surname@ias.tu-darmstadt.de
J.P is in addition with MPI for Intelligent Systems, Tbingen
B.W. and O.T. are with Digital Design Unit, Technische Universitt
Darmstadt, Germany surname@dg.tu-darmstadt.de
F.V. is with CSAIL, MIT, USA fveiga@csail.mit.edu
This project has received funding from the European Unions Horizon
2020 research and innovation programme under grant agreement No 640554.
handover,in-hand manipulation, and object tracking, were
demonstrated. However, in all cases, force-based controllers
relied on averaged force values over all markers [8]: either
the average force itself (object tracking), or the average of
the absolute force values (handover), or some score based
on the absolute force values (gentle grasp).
In this paper, we extend the range of capabilities of the
FingerVision sensor by developing a library of controllers
based on marker deviations and proximity vision, which
subsumes already existing tactile skills and adds a number of
novel ones. Our contributions are summarized in Table I. We
show that feedback controllers utilizing local force informa-
tion can be designed to enable in-hand object rotation and
whole-arm manipulation. Furthermore, we demonstrate the
usefulness of rich tactile feedback in two learning scenarios:
learning to apply a specific force to an object and inferring
texture and viscosity properties of a substance by stirring
it. Along with developing these novel skills, we also make
a number of improvements to already existing capabilities.
Namely, we improve the marker tracking algorithm by in-
tegrating a Kalman filter into the estimation procedure, we
extend the handover skill with a leaky integrator to allow for
seamless gripper opening and closure, we enhance the force
tracking skill with a speed control feature to enable smoother
hand-guiding behavior, and we add a visual scan skill that
utilizes proximity vision to locate an object boundary in
case of objects that do not fit inside the gripper. Finally,
we showcase the capabilities of the developed controllers in
an integrated architectural assembly task shown in Fig. 1.
Fig. 1: The architectural assembly task is comprised of a
sequence of manipulations. After being located and scanned
with the FingerVision camera, a load-bearing element is
grasped and rotated. The robot then goes down until a contact
with the ground surface is detected. Lastly, the point of the
highest load is identified via continuous force measurement
by FingerVision and the part is placed there.
arXiv:1909.09669v1 [cs.RO] 20 Sep 2019
II. BUILDING CUSTOM FINGERVISION SENSOR
The original FingerVision design, manufacturing instruc-
tions, and the data processing code were presented in [6]. We
make several modifications to accommodate for a different
gripper and available materials (see Fig. 2). In particular, we
use different silicone, markers, marker pattern, and housing.
Silicone Replisil 19 N glasklar was selected among al-
ternatives for its transparency. However, it was too hard.
To increase the softness of the silicone, tactile mutator
Slacker was added, as suggested in [9]. Furthermore, to
extend the time available for infusing the markers into the
silicone, additive Slo-Jo was employed, which granted up to
30 minutes dripping time. Resulting material showed good
durability: after several months of intensive experiments, the
silicone was still transparent and in good condition. Direct
comparison to the silicone used in the original FingerVision
sensor was not carried out due to its unavailability at the time
of our experiments; however, we were able to reproduce the
results reported on FingerVision with our materials.
Using microbeads as markers, as envisaged by the original
design [6], turned out to be problematic: the microbeads did
not stick well and were shifting inside the silicone. Bad
adhesion might be due to the difference in silicone used.
As an alternative solution, iron oxide was mixed with the
silicone to color it black, and then the mix was injected
through an acrylic template to form spherical blobs inside the
transparent silicone layer. Such approach allowed for precise
control of the positioning and shape of the injected markers.
Thanks to the fisheye lens of the FingerVision camera, the
density of tracked markers could be increased by giving the
sensor body a round shape. Fig. 2 shows a schematic design
that we used and its implementation. Note the precision of
marker placement and the radial symmetry of the pattern.
In our experiments, an under-actuated parallel gripper was
employed. Due to its specific morphology, an additional
elastic degree of freedom between the sensor body and the
gripper phalange was added to enable grasping of objects
of different sizes (see Fig. 2). Our manufacturing process is
described in more detail in [10].
Two types of information were extracted from the Fin-
gerVision camera [6]: marker displacements and proximity
vision. Marker displacements, shown by red arrows in Fig. 2,
provide local force estimates at marker locations. Proximity
vision delivers object pose and motion estimation informa-
tion via histogram-based background subtraction [7].
Fig. 2: Our FingerVision sensor has a round shape allowing
for more markers in the active area. Two major data modali-
ties are shown: marker displacements (force estimation) and
proximity vision (object pose and movement estimation).
TABLE I: Tactile skills based on FingerVision. Most skills
require 2 modalities, therefore they are arranged in a table
with 4 basic modalities in rows and columns. Our new skills
are shown in bold; italic highlights learning-based skills.
Markers Force Object Slip
+KF IV-A +KF IV-B
Markers ArmRot
+KF IV-C
Force ForceLearn GentleGrasp
+KF V-A
Object InHandRot ForceTrack+ ObjectTrack,
IV-E SpeedCtrl IV-B.1 VisScan VI
Slip StirLearn Handover+ InHandMan Hold
V-B LeakyInt IV-D
III. LIBRARY OF TACTILE SKILLS
We make a number of improvements upon the baseline
skills introduced in prior works [6], [7], [8]. Novel skills and
contributions are denoted in bold in Table I. Columns and
rows correspond to sensor data modalities: (i) Markers stands
for marker displacements estimated using OpenCV [11],
(ii) Force stands for the averaged force computed based
on Markers, (iii) Object stands for object information such
as shape, distance, and orientation extracted from proximity
vision, (iv) Slip stands for optical-flow-based slip estimation.
Most skills require two modalities, therefore we arrange
them in a 2D table. Whenever there is a plus sign, it
means an improvement to an existing skill. Here is a brief
summary of the improvements. First, we add a Kalman filter
to the blob tracking algorithm, which appreciably improves
both individual marker tracking performance (Markers+KF)
and force estimation (Force+KF). Second, we enhance the
Handover skill with a leaky integrator [12] to enable seam-
less gripper opening and closure during object handover
(Handover+LeakyInt). Third, we extend the force tracking
controller ForceTrack with a speed control feature that en-
ables smoother and more natural force-based interactions
with the robot (ForceTrack+SpeedCtrl).
Along with improvements to existing skills, we also in-
troduce new capabilities. First, building upon the Object
modality, we develop a visual scanning skill VisScan that
can follow the shape of an object and find its boundary
and size. We demonstrate the utility of the VisScan skill
in the assembly task (Sec. VI). Second, we introduce a
suite of skills based on raw marker displacement data:
(i) the arm rotation controller ArmRot makes use of circular
vector field estimation, which we demonstrate on rotation of
an in-finger-held asymmetric heavy object; (ii) the in-hand
rotation skill InHandRot combines Markers and Object
modalities to accomplish in-finger object rotation, which
illustrates the capacity of joint force and torque estimation;
(iii) we demonstrate that end-to-end control based on marker
displacement data is feasible by developing two learning-
based controllers—ForceLearn, capable of associating tac-
tile sensation (Markers) with applied force (Force), which we
showcase on pressing with a desired force in Newtons, and
StirLearn, capable of associating tactile sensation (Markers)
with vision-based slip detection (Slip), which we showcase
on identifying a substance through stirring.
IV. IMPROVED AND NOVEL TACTILE SKILLS
This section focuses on improved and newly introduced
tactile skills. Implementation details, evaluations, and ex-
ample use cases are presented. Subsequent sections cover
learning-based skills (Sec. V) and the showcase application
in collaborative architectural assembly (Sec. VI).
A. Blob Tracking Improved with the Kalman Filter
We start by describing the Kalman-filter-based enhance-
ment to the blob tracking algorithm of FingerVision [6].
Every marker in the silicone layer of the sensor is detected
using OpenCV, and its displacement is registered in a vector
xt=x0xty0yts0stTcontaining initial and
current (x, y)-coordinates and size sof the marker. Assuming
a first order dynamical system xt=Axt−1+εtwith linear
observations yt=Hxt+δt, we filter blob displacements
with the Kalman filter. Covariance matrices are taken to
be proportional to the identity, with system noise covari-
ance 0.01 and observation noise 0.1. Time step is taken as the
inverse of the camera frame rate, i.e., dt = 1/15. Definitions
of matrices Aand Hand further detail can be found in [10].
Fig. 3 shows a comparison of the tracking performance
between the baseline algorithm [6] and our improved version
in a static object holding task. Displacement of each blob is
estimated and an average displacement is shown separately
in x-, y-, and z-directions. The coordinate system is placed
such that the xy-plane is aligned with the image plane of the
camera, thus the z-axis is pointing from one fingertip to the
other. The baseline algorithm produces a lot of jitter due to
the robot vibration and ensuing noise in the blob detection
algorithm (blue line in Fig. 3). The Kalman filter removes
undesirable high-frequency jumps (orange line in Fig. 3).
B. Force Tracking vs. Object Tracking
The FingerVision sensor is unique among vision-based
tactile sensors in that it has a transparent coating and can thus
utilize the raw video stream seen through the ‘skin’ of the
sensor. But is it also useful for control? To answer this ques-
tion, we implement ForceTrack and ObjectTrack controllers
from [7] and compare their performance on the FollowMe
Fig. 3: Blob tracking with (orange) and without (blue) the
Kalman filter on a static object holding task.
Fig. 4: Robot following the motion of a human in the Fol-
lowMe task. ForceTrack and ObjTrack controllers compared.
task. The human places an object inside the gripper (e.g., as
shown in Fig. 4) and then either pulls or pushes the object;
the robot has to follow. The ForceTrack skill relies on marker
displacement estimation to move proportionally to the force
applied. The ObjectTrack skill, on the other hand, relies on
optical flow estimation to move proportionally to the velocity
of the object displacement. Experimental results in Fig. 5
show that the force-based approach is more robust and is
easier to use. The vision-based approach, however, results
in jerkier behavior and completely fails if a movement in
the z-direction is required (perpendicular to the fingertip),
for such movement cannot be detected using vision. In the
following, we elaborate on the setup, describe the details of
the controller implementations, and discuss the results.
1) ForceTrack+SpeedCtrl:This controller moves the
robot end effector in the direction of the average
force measured by FingerVision. If ft=fx
tfy
tfz
tT
is the current force estimate, then the desired Carte-
sian position at the next time step is given by the
rule xt+1 =xt+α·(ft−fmin)·1(|ft|> ε). Parameters
α, fmin, ε are set depending on the desired responsiveness
of the robot. Notation 1(b)stands for an indicator function
that converts bfrom True/False to 1/0. Parameter αcontrols
(a) pull, force (b) push, force
(c) pull, vision (d) push, vision
Fig. 5: ForceTrack+SpeedCtrl (upper row) vs. ObjectTrack
(lower row) on the FollowMe task. ForceTrack controller
follows the input (Force) better with fewer abrupt changes.
Fig. 6: ArmRot skill. An asymmetric object held in-fingers
produces torque (blue). The robot rotates the arm to reach
zero torque (orange).
the movement speed, which we call SpeedCtrl feature. We
set α= (vmax −vmin)/(fmax −fmin )to control the range
of commands xt+1 through the velocity range.
2) ObjectTrack: This controller tries to keep the object in
the center of the field of view of the FingerVision camera.
Assume the object is moved in the y-direction (see Fig. 4).
If Mtis the current object area (computed using image
moments [13] as described in [7]), and (¯xt,¯yt)are the
coordinates of the object center in the image plane, then
the desired Cartesian position is set as
xt+1 =xt+ min {x, xmax } · 1(|x|> x),
yt+1 =yt+ ∆ (1(Mt≤M)−1(Mt> M)) (1)
where ∆is the sensitivity to the displacement of the ob-
ject inside the gripper. Parameters xmax and ymax are the
maximum allowed deviations in the xand ydirections
respectively, and thresholds x,y,Mprevent false pos-
itive controller activations. For moving in the x-direction,
equations in (1) should be swapped.
3) ForceTrack+SpeedCtrl vs ObjectTrack: Fig. 5 com-
pares ForceTrack (upper) and ObjectTrack(lower) controllers
under two conditions: pulling (left) and pushing (right).
Each plot shows the input signal at the top (either force
or centroid shift) and the command generated from it at the
bottom (end-effector velocity in both cases). Note that the
force is not measured in Newtons but is proportional to the
markers displacements, as in [6]. ForceTrack controller with
SpeedCtrl was consistently better than without. As seen from
Fig. 5, the force-based controller follows the input signal
more faithfully. Additionally, it is more robust with respect to
the object used [10], in contrast to the vision-based controller.
Based on our experiments, the force-based controller should
be preferred for contact-tracking tasks, such as the FollowMe
task showcased here.
C. Torque-Driven Arm Rotation ArmRot
A torque estimation procedure for FingerVision was pro-
posed in [7]. However, it was not used for control. Here, we
demonstrate a controller that can successfully utilize such
torque information. Fig. 6 demonstrates a potential use case:
an asymmetric stick with a heavy head is grasped by the tail
and held horizontally in the beginning. The torque created
by the stick in the fingers is detected by FingerVision and is
used in a feedback loop to drive sensor readings to zero.
Fig. 7: LeakyInt skill triggers gripper closure when both
force and slippage signals are active. In general, any external
signal can be used to trigger the leaky integrator.
D. Handover with Leaky Integrator LeakyInt
Gentle opening and closing of the gripper is the most basic
skill one can expect. We implement the leaky integrator as
described in [12] for the FingerVision. The idea is to follow
an external input with some inertia. More concretely, the
position command xtfor the gripper is computed as xt=
αxt−1−(1 −α)Lwhere αcontrols how fast the gripper
reacts and Ldetermines the set point. Note that Lcan be
any external input. For example, Fig. 7 demonstrates how a
combination of slippage and force can be used to trigger L
(pink line in the top subplot). Since Lis only activated when
both Slippage (middle subplot, green points mean ‘active’,
red points mean ‘not active’) and Force (bottom subplot,
the same color convention) are active, the gripper is closing
when an object is detected between the fingers and at the
same time the object is touching the sensor (pink line).
E. In-Hand Object Rotation InHandRot
We compare two approaches to rotating an object inside a
parallel gripper: based on torque estimation and based on
slip detection. Both approaches perform well, the choice
may depend on the application. The slip-detection approach
was better for letting objects rotate under their own weight
because torque estimation was unreliable for light objects
and small torques. Fig. 8 shows the torque signal used for
rotating a pen inside the gripper. We slowly open the gripper
using the leaky integrator till the value of torque goes to zero.
V. LEARNING-BASED SKILLS
In the previous section, improvements over baseline con-
trollers and a few novel analytic controllers utilizing the
Markers modality were described (see Table I). This section
is aimed to demonstrate the feasibility of utilizing the raw
input in the form of marker displacement data and optical
flow estimation data for controlling the robot. To that end,
two example applications are presented: ForceLearn which
puts emphasis on force estimation data and StirLearn which
puts emphasis on the optical flow estimation data.
Fig. 8: InHandRot. A pen is left to rotate under its own
weight between the fingers from the position parallel to the
ground. Although zero torque was detected (upper plot),
the final pen orientation was less than 90 degrees which
means the pen got stuck. This is a typical problem of torque
estimation for light objects with FingerVision.
A. ForceLearn: Learning to Press with a Given Force
Associating marker displacements with the exerted force
is a non-trivial task. On one hand, marker displacements
should be proportional to the force, at least for small de-
formations. On the other hand, the displacements depend on
other factors, such as the material of the objects in contact
and the kinematic configuration of the problem. Learning-
based approaches may potentially be sufficiently powerful to
extract invariant information which can be used for control.
We provide a proof-of-concept demonstration that this is
indeed the case and machine learning can be effective for
encoding the mapping from sensor readings to force values.
We train a regression model to predict force from marker
displacements. Our setup consists of an electronic scale that
delivers data over ROS and the robot pushing a stick held
in-between the fingers against the scale. Fig. 9 shows the
results obtained using kernel ridge regression with radial
basis functions. The left figure shows measured vs. predicted
values; the right figure shows the output of the predictor
obtained in a real-time test while releasing a button. Using 20
pressing movements each of duration 5sec recorded at 15 Hz
(a) Regression (b) Prediction
Fig. 9: ForceLearn. Training (left) and testing (right) of a
learned predictor of the force from marker displacements.
Prediction is accurate and can be used in downstream tasks.
TABLE II: StirLearn: Classifying substances by stirring.
precision recall f1-score support
flour 1.00 0.94 0.97 16
sugar 1.00 1.00 1.00 16
peas 0.94 1.00 0.97 16
avg/total 0.98 0.98 0.98 48
was sufficient to predict force values with the measurement
accuracy. However, one has to stress that the learned mapping
is object-dependent due to the use of silicone as the skin in
FingerVision: different materials behave differently when in
contact with silicone. For example, the FingerVision gets
stuck on glass. It is not yet clear what a general object-
agnostic tactile sensor should look like.
B. StirLearn: Density and Texture through Stirring
As the final demonstration, we showcase the use of all
input modalities in a challenging prediction task. We consider
a problem of discerning substances that have different density
and texture, such as flour, sugar, and peas, based on tactile
interaction with them. This problem setup is inspired by [14],
where a system was trained to discriminate liquids of varying
viscosity by detecting surface changes with a depth camera.
In our case, the features are not based on visual observations
but rather on tactile sensations. As the feature vector, we
use all available information, i.e., deviation of each marker,
centroid {x, y}, object orientation θand area M.
A data set consisting of 120 trials was collected via stirring
substances with a wooden stick using a set of 8pre-defined
stirring movements and grasping settings. A training set
and a test set were created, comprised of 72 and 48 trials,
respectively. A multi-layer perceptron (MLP) with 3hidden
layers, 10 neurons per layer, and logistic activations was
used. The evaluation metrics of the trained MLP are provided
in Table II. A virtually ideal classifier could be obtained.
VI. HUMAN-ROBOT COLLABORATION IN
ARCHITECTURAL ASSEMBLY TASKS
Previous two sections introduced a variety of skills based
on tactile feedback. In this section, we want to demonstrate
the utility of developing such a repertoire of skills by
showing how they can be combined together to solve a
complex contact-rich task. As an application, we consider
collaborative architectural assembly based on digital design
models [15]. The common practice today requires an archi-
tect to precisely define in advance both the local and global
geometry of a structure to be built. However, unforeseen
changes often occur after the construction has been started.
Especially interesting are the cases where the changes are
not due to robot mistakes but rather indicate architect’s
changing design goals in an interactive fashion, such as in
the collaborative positioning task showcased below.
Fig. 10 shows the model used in our experiments. A
digital model describing a desired structure is passed to the
robot for assembly. Locations of the vertical load-bearing
elements are not pre-programmed but rather determined by
the robot through tactile sensing online as the locations of the
highest load. Subsequently, the observed state of the erected
Fig. 10: A prototype pipeline for collaborative architectural assembly. The outer feedback loop shows the data flow from
the design software to the robot and from the sensors registering the actual configuration back to the design model. The
assembly process detailed in Fig. 1 and comprised of skills from Table I allows for interaction via the Handover+LeakyInt
(Sec. IV-D) and ForceTrack+SpeedCtrl (Sec. IV-B.1) skills. See Fig. 5 and the accompanying video for details.
structure is passed back to the design software to update the
plan. Interaction with the robot is enabled through the tactile
feedback and can be used to guide the robot and reposition
the elements. See the accompanying video for details.
External lighting and bright-colored objects were used in
the assembly task. Experiments exposed high sensitivity of
the vision-based blob detection algorithms [6] to lighting
conditions and object color. Installing a light source at the
wrist could potentially solve these problems.
VII. CONCLUSION
Tactile-enabled applications have been a long-standing
vision in robotics [1]. In the recent years, the combination
of inexpensive hardware with advances in machine learning
are providing a unique opportunity to experiment with novel
designs and applications of tactile sensors. In this paper, we
attempted to build a library of useful robot behaviors by
utilizing tactile feedback (see Sec. III). To accomplish that,
we created a modified version of the FingerVision sensor [6]
that suits our robot hardware and provides a few design
enhancements (see Sec. II). We further extended the existing
software framework around the FingerVision [7] with a suite
of improved and novel tactile skills (see Sec. IV). Beyond
hand-designed controllers, we for the first time demonstrated
the feasibility of using the raw sensory data from the Finger
vision to learn skills such as pressing with a specified amount
of force (Sec. V-A) and identifying substances through
stirring (Sec. V-B). Finally, we showcased a potential fu-
ture application of tactile sensing in interactive architectural
assembly [15] based on human-robot collaboration.
ACKNOWLEDGMENT
We thank Christian Betschinske for his great help in build-
ing our FingerVision sensor and Olivier Stoos for creating the
architectural assembly model. A.S. thanks Lufthansa Industry
Solutions AS GmbH for covering the travel costs.
REFERENCES
[1] M. R. Cutkosky, R. D. Howe, and W. R. Provancher, “Force and tactile
sensors,” Springer Handbook of Robotics, pp. 455–476, 2008.
[2] R. S. Dahiya, P. Mittendorfer, M. Valle, G. Cheng, and V. J. Lumelsky,
“Directions toward effective utilization of tactile skin: A review,” IEEE
Sensors Journal, vol. 13, no. 11, pp. 4121–4138, 2013.
[3] T. Zielinska, “Shear force sensor for robots,” in Proceedings of IEEE
International Symposium on Industrial Electronics, vol. 1. IEEE,
1996, pp. 49–53.
[4] K. Kamiyama, K. Vlack, T. Mizota, H. Kajimoto, K. Kawakami,
and S. Tachi, “Vision-based sensor for real-time measuring of surface
traction fields,” IEEE Computer Graphics and Applications, vol. 25,
no. 1, pp. 68–75, 2005.
[5] W. Yuan, R. Li, M. A. Srinivasan, and E. H. Adelson, “Measurement
of shear and slip with a gelsight tactile sensor,” in 2015 IEEE
International Conference on Robotics and Automation (ICRA). IEEE,
2015, pp. 304–311.
[6] A. Yamaguchi and C. G. Atkeson, “Combining finger vision and
optical tactile sensing: Reducing and handling errors while cutting
vegetables,” in IEEE-RAS 16th International Conference on Humanoid
Robots (Humanoids). IEEE, 2016, pp. 1045–1051.
[7] ——, “Implementing tactile behaviors using fingervision,” in 2017
IEEE-RAS 17th International Conference on Humanoid Robotics
(Humanoids). IEEE, 2017, pp. 241–248.
[8] A. Yamaguchi, “Fingervision for tactile behaviors, manipulation, and
haptic feedback teleoperation,” in IEEJ 4th international workshop
on Sensing, Actuation, Motion Control, and Optimization (SAM-
CON2018), 2018.
[9] Y. Wenzhen, “Tactile measurement with a gelsight sensor,” Master’s
thesis, Massachusetts Institute of Technology, 2014.
[10] A. Sadybakasov, “Learning vision-based tactile skills for robotic
architectural assembly,” Master’s thesis, TU Darmstadt, 2019.
[11] G. Bradski, “The OpenCV Library,” Dr. Dobb’s Journal of Software
Tools, 2000.
[12] F. Veiga, B. B. Edin, and J. Peters, “In-hand object stabilization by
independent finger control,” arXiv preprint arXiv:1806.05031, 2018.
[13] M. Pawlak, Image analysis by moments: reconstruction and computa-
tional aspects. Oficyna Wydawn. Politechn., 2006.
[14] C. Elbrechter, J. Maycock, R. Haschke, and H. Ritter, “Discriminating
liquids using a robotic kitchen assistant,” in 2015 IEEE/RSJ Interna-
tional Conference on Intelligent Robots and Systems (IROS). IEEE,
2015, pp. 703–708.
[15] B. Wibranek, B. Belousov, A. Sadybakasov, and O. Tessmann, “Inter-
active assemblies: Man-machine collaboration through building com-
ponents for as-built digital models,” in Computer-Aided Architectural
Design Futures (CAAD Futures), 2019.