3852 IEEE SENSORS JOURNAL, VOL. 18, NO.

9, MAY 1, 2018

Design, Modeling, and Validation of a Soft

Magnetic 3-D Force Sensor
Anany Dwivedi , Anand Ramakrishnan, Aniketh Reddy, Kunal Patel, Selim Ozel, and Cagdas D. Onal

Abstract— Recent advances in robotics promise a future where This article adresses each challenge by presenting a magnetic
robots co-exist and collaborate with humans in unstructured sensing modality and neural network based modeling of a
environments, which will require frequent physical interactions small tri-axial soft sensing element.
where accurate tactile information will be crucial for perfor-
mance and safety. This article describes the design, fabrication, In terms of sensing medium, tactile sensing literature may
modeling, and experimental validation of a soft-bodied tactile be classified into: 1) stimuli responsive and/or composite mate-
sensor that accurately measures the complete 3-D force vector rials that mainly employ resistive or capacitive measurements,
for both normal and shear loading conditions. Our research and 2) using embedded discrete electronic components or other
considers the detection of changes in the magnetic field vector physical quantities such as optical or magnetic signals within
due to the motion of a miniature magnet in a soft substrate
to measure normal and shear forces with high accuracy and the sensor body. Resistive sensing has been a popular method,
bandwidth. The proposed sensor is a pyramid-shaped tactile although it may suffer from dynamic artifacts [5]–[8]. In [9],
unit with a tri-axis Hall element and a magnet embedded in Vogt et al. uses a conductive fluid (eGaIn) placed in channels
a silicone rubber substrate. The non-linear mapping between the created on a soft matrix to measure applied forces. These are
3-D force vector and the Hall effect voltages is characterized by multi-axis force sensors and can measure forces in normal
training a neural network. We validate the proposed soft force
sensor over static and dynamic loading experiments and obtain a and shear directions. But fabrication using eGaIn involves a
mean absolute error below 11.7 mN or 2.2% of the force range. number of challenges which are discussed in [10]. Alterna-
These results were obtained for a soft force sensor prototype and tively, [11] measures pressure (or force) in the normal direction
loading conditions not included in the training process, indicating by the change in capacitance between two PDMS layers filled
strong generalization of the model. To demonstrate its utility, with carbon nanotubes. A similar measurement idea is realized
the proposed sensor is used in a force-controlled pick-and-place
experiment as a proof-of-concept case study. via conductive textiles in [12]. In [13] Yang and Chen present
pressure and position sensors made of conductive elastomers
Index Terms— Robot sensing systems, force feedback, neural co-printed into a soft actuator in a single process without
networks.
assembly. These sensors are capable of providing feedback
I. I NTRODUCTION because of their innovative design and the piezoresistive effect

D IRECT force and tactile sensing remains an open

technical problem in robotics [1]–[3]. To address this
problem, new sensing mechanisms and modeling approaches
of conductive elastomers.
On the other hand, in [14], a commercially available baro-
metric IC is embedded in a soft elastomer, to perform as a 1-D
need to be developed to achieve compliant, safe, and aware tactile sensor to measure normal forces. Yi et al. [15] presents
interactions between robots, human users, and the environ- a tactile sensor using optical fiber Bragg grating based on
ment. We observe three fundamental challenges that impede phase modulation of the optical source to determine the forces
progress [4]. First, it is difficult to obtain a reliable sensing applied on the sensor. In [8], Sohgawa et al. propose a resistive
modality between external forces and a measurable change in a tactile sensor using piezo-resistive cantilevers embedded in a
physical medium, without detrimental nonlinearities or other soft matrix. It is important to note that these existing solutions
artifacts such as hysteresis or time delay. Second, modeling suffer from various inherent challenges such as a lack of 3-D
to obtain a reliable map between the external force and force sensing capability, relatively complex fabrication and
the measured physical change is challenging, especially for signal processing circuitry, or hysteresis and time delay.
multi-dimensional measurements. Finally, scalability tends to Regardless of the sensing medium, modeling and obtaining
be a challenge, to obtain useful spatial resolution over a a reliable mapping between the force and the measured physi-
surface, while avoiding crosstalk between sensing elements. cal change is crucial and challenging, especially for multi-axis
force sensing. The work in [16] employs multiphysical finite
Manuscript received January 18, 2018; revised February 18, 2018; accepted
February 20, 2018. Date of publication March 12, 2018; date of current element analysis (FEA) to describe the behavior of a dome-
version April 9, 2018. This work was supported by the National Science shaped magnetic tactile sensor. Using FEA models are useful
Foundation under Grant CNS-1544636. The associate editor coordinating but in a multiphysics environment small errors and initial set-
the review of this paper and approving it for publication was Prof. Subhas
C. Mukhopadhyay. (Anany Dwivedi and Anand Ramakrishnan contributed tings reflect heavily during the actual experimental testing of
equally to this work.) (Corresponding author: Selim Ozel.) the sensors and since elastomers (e.g. Ecoflex 0030) are known
The authors are with the WPI Soft Robotics Laboratory, Mechanical to be highly nonlinear, FEA modeling is not straightforward.
Engineering Department and Robotics Engineering Program, Worcester Poly-
technic Institute, Worcester, MA 01609 USA (e-mail: cdonal@wpi.edu). In addition, such models are usually not applicable for real-
Digital Object Identifier 10.1109/JSEN.2018.2814839 time computation.
1558-1748 © 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
DWIVEDI et al.: DESIGN, MODELING, AND VALIDATION OF A SOFT MAGNETIC 3-D FORCE SENSOR 3853

Hall element. This measured change has a non-trivial relation

to the force applied on the sensor module.
We introduce two main novel contributions with this work.
We use a neural network to calibrate the force sensor in 3-D
and demonstrate its generalization ability to other materials
and loading conditions. We show that the network is able
to learn to respond to a large range of force directions
Fig. 1. 1) Design of the proposed tri-axial soft force sensor utilizes a
whereas only small number of training forces are applied
miniature magnet (A), Hall Effect IC (B), and a circuit layer on an acrylic at pre-determined directions. We also show that the same
sheet (C), embedded in a pyramid shaped silicone rubber substrate. 2) The network can be used with sensor prototypes that were not
sensor prototype is 8 mm tall with an 12 mm wide square base.
in the training set. This means that the network is able to
overcome minor manufacturing differences and it can be used
to scale up the sensing resolution without scaling up the
In [9], Vogt et al. approximate the sensor’s resistance calibration efforts. Recently the usage of machine learning
change response as linear and create a calibration matrix using has seen great interest for capturing complex relationships
least mean squares method. A similar approach was taken between the input space (forces) and the output space (the
in [17], where a linear regression model was employed to find sensor measurement) [25]. In [22], fully connected neural
the relationship between force applied and the corresponding networks (FCNs) are explored for characterization. However,
change in voltage. The linearity approximation makes these the results show that the network is not able to generalize
models simple enough to calculate the force mapping in real- beyond the training conditions. Here, we present FCNs based
time, but they are highly simplistic and thus, prone to errors. on the Net2Net initialization technique [26], which provides
Response reliability is also important. In [11], Lipomi et al. better regularization of the network. We show that our network
mention carbon nanotubes not being aligned after repetitive is able to generalize very well in Section III.
loading. The resulting hysteresis can be reduced by depositing The second novelty is the pyramid shape. One of the main
carbon nanotubes at a pre-stretched state, which may increase factors that affect the dependability of a tactile magnetic
complexity and repeatability issues. These limitations moti- sensing element is its shape [27]. The shape of the contact
vated us to consider approaches that use machine learning as surface that tapers to a point helps in restricting and channeling
a method for estimating the complex relationship between the the movement of the magnet inside the soft matrix. In [16]
applied force and the measured change in the magnetic field we see a dome-shaped magnetic soft sensor with the magnet
vector in the proposed sensor to offer both accurate modeling immediately below the dome which helps in obtaining a highly
and real-time computation for the nonlinear sensory mapping. accurate model of the sensor. In [22] the shape of the soft
Using an embedded miniature magnet and flexible elec- matrix is a cuboid and the magnet is placed over the Hall
tronics in a soft substrate, we have previously shown that element. A considerable problem with these designs are that
sensing local changes in magnetic field is suitable for accurate the applied force may cause rotation of the magnet about
and high-speed measurements on the curvature of a flexible its own axis in unmodeled ways, thus requiring additional
bending body [18]. In previous research, our objective was calibration data and also reducing measurement dependability.
to obtain curvature measurement from soft silicone rubber In this work, we utilize the shape of a pyramid with the magnet
segments for a specific type of bending soft actuator utilized embedded within the pyramid (as close to the centroid as
in a snake robot [19]. This article employs similar design practically possible). The advantage of this shape is that off-
principles to force/contact measurements on a soft deformable center forces acting on the sensor would tend to act about the
substrate and in a small form factor. centroid of the sensor and thus the tendency of rotation of
Hall effect based sensors have gained prominence the magnet about its own axis is reduced, greatly improving
recently [16], [20]–[24]. In [22], Tomo et al. presents a soft reliability.
sensor which utilizes a magnet for detecting forces in multiple This article is organized as follows: In Section II we discuss
axes. 16 Hall-effect ICs are used in total for one sensor the design and manufacturing methods for the proposed soft
module, hence adding to the complexity of the design. In this 3-D force sensor. We include detailed information on electron-
work, we present accurate force measurement results using a ics and fabrication process. In Section III we discuss in detail
single IC with a magnet on top. We also do not need any how data is collected for characterization and how it is used
explicit noise filter thus enabling our sensor to retain a faster to map the function space of the sensor by a Neural Network.
response. In Section IV, we discuss the results for the dynamic response
Our sensor design utilizes a 3-axis Hall effect sensor IC test on the sensor and also present a use case application of
and a small magnet placed over it at a defined location, the force sensor by performing force control on the fingers of
embedded inside a soft elastomer substrate. This gives the a Jaco arm. Finally, we conclude the paper and discuss our
composite mechatronic structure the compliance required for future plans in Section V.
force sensing. A 3-D model is shown in Figure 1. Any force
applied on the soft matrix produces a deformation on it, which II. S ENSOR D ESIGN AND FABRICATION
changes the position of the embedded magnet and this in The proposed 3-D force sensor utilizes tri-axial measure-
turn causes changes in the magnetic flux values around the ment of the magnetic field vector created by a miniature
3854 IEEE SENSORS JOURNAL, VOL. 18, NO. 9, MAY 1, 2018

Fig. 2. Manufacturing process of the proposed tri-axial soft force sensor. (a) 3D printed molds required for fabricating sensor are shown in yellow (top).
A custom flexible PCB is bonded to an acrylic carrier (bottom). (b) The molds are clamped together with the acrylic circuit carrier and liquid silicone rubber
precursor is introduced. (c) Cured silicone rubber sensor body is demolded with a cavity for the magnet. (d) Magnet is placed in the cavity and liquid silicone
rubber is injected to seal it in place. (e) Cured and finished force sensor module CAD model. (f) Actual force sensor prototype.

permanent magnet embedded in a pyramid shaped silicone

rubber body as shown in Fig. 1. Magnetic fields are mea-
sured locally using a Hall element on an embedded custom
circuit board. The magnet is displaced under external forces,
which creates a corresponding change in the 3-D magnetic
field measurement. The sensor has a pyramidal shape, which
ensures that the sensor contacts the environment mostly with
its tip. In addition, the magnet is embedded deep within the
pyramidal shape. This location is chosen to limit the motion of
the magnet under off-center forces (to ensure that the magnet
translates with minimal rotation).
The sensor is fabricated using a multi-stage composite
molding process as shown in Fig. 2. For embedded electronics,
we print and etch a custom printed circuit board (PCB).
Standard circuit components and the Hall Effect IC (Melexis Fig. 3. Force versus compression data is obtained from three sensors and
MLX90363) were soldered manually. The circuit communi- compared to a solid Ecoflex 0030 segment, indicating minimal effect on the
cates with a master device using the Serial Peripheral Interface material compliance by the embedded magnet and electronics.
(SPI) protocol. Once programmed, the circuit sends 8-byte
messages of magnetic flux measurements in three axes with
14-bit resolution in each axis. We use an Arduino control board is demolded, the magnet is placed in its place, and a layer of
as the master data acquisition device and send this information silicone rubber is injected in the cavity above the magnet to
to M ATLAB where data processing is performed. seal it completely within the sensor body.
The first step of the fabrication process is bonding the A major concern in soft sensing is to ensure that the
custom PCB to an acrylic plate which has been cut into additional embedded components do not drastically modify
the shape of the PCB but with extensions on two sides. the mechanical response of the soft body. To validate this
These extensions help maintain the orientation of the PCB property for our sensor design (comprising an acrylic plate,
during molding, after which they can be snapped off. The a miniature magnet, electronic components and silicone rub-
remaining acrylic provides a rigid base to the PCB. This ber substrate), we performed compressive testing of three
custom acrylic PCB is assembled with two 3-D printed molds, prototypes. We compared the mechanical force-displacement
which also include an extruded negative of the magnet shape response of these complete prototypes to the material response
to create a cavity for the magnet. As silicone rubber (Smooth- of solid silicone rubber (Ecoflex 0030) with the same geometry
On Ecoflex 0030) is cured in this mold assembly, the sensor but without the embedded components. Our results in Fig. 3
DWIVEDI et al.: DESIGN, MODELING, AND VALIDATION OF A SOFT MAGNETIC 3-D FORCE SENSOR 3855

Fig. 4. Experimental setup for static loading shown as a CAD model (A) and the real system during operation (B). The force sensor placed below the
load cell. The bottom stage can be rotated to a desired angle to create shear forces at a defined angle. Load generation on the sensor is achieved through
lowering the load cell on the sensor by the motion stage. Thus, even though the load cell measures single axis force data, it is decoupled into normal and
shear components using the angle set at the bottom stage.

show that material properties are similar between different was designed and used to mount the sensor at desired angles
batches of the sensor. We did not observe a significant change with respect to the load cell, thus the force vector is decom-
in material response due to the composite structure. posed into normal and shear components at known combi-
nations. The load cell, the tri-axis stage, and the articulating
III. L EARNING S ENSOR M ODEL base were made up of materials which do not interfere with
U SING A N EURAL N ETWORK the magnetic flux measured by the sensor.
We collected experimental data from four different sensor
A multi-layer perceptron (MLP) also known as fully- prototypes. The data from three sensors were used to train the
connected neural (FCN) network was used to learn the func- neural network and the fourth sensor was used for validating
tion space for sensor characterization. This function space is the network. Four different loading configurations were con-
high-dimensional and nonlinear. For instance, the soft material sidered. These include pure normal force loading (at 0°), and
deformation can be described using a hyper-elastic material shear loading at angles of 30°, 45°, and 60° with respect to
model such as Mooney-Rivlin or Ogden, which require multi- the sensor normal. For pure normal loading, the sensor was
ple experimentally characterized parameters. In addition, since subjected to a maximum of 1.1 N, and for shear loading the
the relation between the magnet pose and magnetic flux is a maximum load applied was 1.5 N. These limits were chosen
nonlinear transformation, an analytical physics based model based on the saturation range of the Hall element and the
is intractable and a multi-physical finite element model may amplifier circuit and correspond to tactile contact-level forces.
prove computationally expensive to operate in real-time. Thus, Data for forces at 45° shear loading was retained only for
this work considers the use of MLPs to represent the sensor the validation dataset and was not part of the training data.
response under force loading. This was to see how well the trained mapping function is
Let f W (X i ) be the function that maps Hall Effect voltages generalized to forces at different loading conditions.
to 3-D forces. The input space is the Hall Effect voltages
measured by the sensor, where X i = [Vx , Vy , Vz ] ∈ R 3×1 and
B. Learning Approach
the force vector corresponding to these magnetic flux values is
defined as yi = [Fx , Fy , Fz ] ∈ R 3×1 . The goal of the network The technique used for learning the unknown mapping func-
is to learn a set of weights W for f W in the generic expression: tion was based on the Net2Net initialization technique [26],
yi = f W (X i ). which performs learning in a sequential manner starting with
small MLPs and then scaling the neural network up to a larger
size in width and depth by using the previous smaller network
A. Data Acquisition as a teacher to the new larger student network. This approach
The training data was obtained by applying known forces avoids the usage of a very large initial network and then re-
on the proposed soft force sensor prototypes and then mea- learning the entire network from scratch if the performance
suring the corresponding Hall Effect voltage readings. The is not suitable. In addition, we expect this approach will help
applied forces were measured using a load cell (TAL220) avoid overfitting and provide mapping functions that general-
and corresponding amplifier circuit (Sparkfun HX711). This ize well to different loading conditions and sensor prototypes.
setup measures loads up to 10 N with errors up to 5 mN. The scaling of the network is performed by initialising the
We mount the load cell on a tri-axis Cartesian stage (Newport student network with the weights of the teacher network and
9064-XYZ-PPP) as shown in Figure 4. An articulating base then widening or deepening it. Widening involves adding
3856 IEEE SENSORS JOURNAL, VOL. 18, NO. 9, MAY 1, 2018

TABLE I
MSE L OSS IN T RAINING D URING W IDENING O PERATION ON P YRAMIDAL S ENSOR

TABLE II
MSE L OSS IN T RAINING D URING D EEPENING O PERATION ON P YRAMIDAL S ENSOR

TABLE III
R ESULTS ON THE T EST D ATASET D URING W IDENING O PERATION ON P YRAMIDAL S ENSOR

additional neurons to a layer. Deepening involves adding a from there. Training was performed for 550 Epochs for each
new hidden layer to the network. The advantage of the method teacher network. Scaling of the network was done by widening
is that it ensures that the student network improves upon the the network first and then deepening it. Widening operation
teacher network. was done to a maximum of 18 neurons and then deepening
operations were performed up to 5 hidden layers. The final
C. Metrics for Evaluation network used for learning the sensor model contains 5-hidden
layers with 18 neurons at each layer. We implemented our
We measure the neural network performance using mean network models using the software package Keras [28].
absolute error (MAE) and standard deviation (σ ). Specifically,
as each sensor data (X i ) is passed through the network,
a prediction ( pi ) is obtained and the error (ei ) is defined as E. Training Results
the difference between this prediction and the actual force (yi ). The MSE obtained during training is shown in Table I (for
ei = pi − yi where, pi = f W (X i ). MAE and σ are calculated widening operation) and Table II (for deepening operation).
in standard form: Data from Table I indicate that the training loss converges
N and does not improve from 16 neurons to 18 neurons during
i=1 ei
M AE = , (1) the widening operation. This is the primary reason why
 N the widening operations were not pursued after 18 neurons.
N 2
i=1 ei As hidden layers are added, the training loss is almost halved
σ = . (2) for every added layer and is almost 1e−4 thus showing that the
N
network is learning the function space effectively based on the
D. Training, Testing, and Validation training data. After adding the 5th hidden layer we concluded
that further increase in depth will be prone to overfitting.
The total data points were split into training and testing However, this does not provide any insight into the capability
datasets following the 80-20 convention. Data points obtained of the network to generalize for different loading conditions
from the fourth sensor prototype and for 45° loading were and sensor prototypes. Section IV-A will demonstrate the
excluded from this dataset. response of the neural network for loading conditions and
The technique used for training all the Net2Net networks sensor prototypes that were not included in the training set.
utilized the following hyper parameters. All the models were
trained with Adam as the optimizer and mean-squared-error
(MSE) was used as the cost metric to train the network F. Testing Results
since this is a regression problem. The learning rate was Table III shows the results obtained on the test dataset
scheduled with initial value being the default 1e−3 (for Adam during widening operations. As we scale the neurons in the
optimizer). A reduction in learning rate by a factor of 10 was first layer we see that the Mean Absolute Error (MAE) and
effected whenever the loss failed to reduce in 3 consecutive the Standard Deviation (σ ) are reduced. This decrease is very
epochs. When the “Learning Rate Scheduler” function is rapid initially and then slowly converges to steady state at a
evoked, the best weights (in terms of least loss) obtained layer width of 18 neurons. Comparing the MAE results from
until the function call are loaded and training is continued the 16 Neuron and 18 Neuron architectures, we see that there is
DWIVEDI et al.: DESIGN, MODELING, AND VALIDATION OF A SOFT MAGNETIC 3-D FORCE SENSOR 3857

TABLE IV
R ESULTS ON THE T EST D ATASET DURING D EEPENING O PERATION ON P YRAMIDAL S ENSOR

Fig. 5. Validation Dataset Error Analysis.

negligible difference between them. This pattern is also seen in

the training data and this trend shows us that further increase
in the number of neurons would not provide better results.
Hence widening operation was stopped at 18 Neurons and we
move to performing deepening operations with 18 Neurons in
each hidden unit.
Table IV shows the results when deepening operation is
performed while keeping the number of neurons in each layer
constant at 18 Neurons. We see that MAE and σ are halved
everytime a new hidden layer is added. We stop at 5 hidden
layers as adding more layers would affect the time taken
for prediction during test time. At the 5th layer, we obtain
a MAE of 5.6 mN and a σ of 9.2 mN which demonstrate
strong prediction capability of the proposed network after both
widening and deepening operations of Net2Net learning.

IV. E XPERIMENTAL R ESULTS

To validate the accuracy and utility of the proposed 3-D soft
force sensor and neural network based modeling approach,
we performed three sets of experiments. First, we applied
known static and dynamic forces on the sensors and compared
the outputs from the neural network with the applied forces in
a validation dataset. Next, we performed a proof of concept Fig. 6. Forces measured from load cell are plotted on X-axis whereas
force controlled pick and place experiment using the proposed forces measured from the soft force sensor are plotted on Y-axis. The circles
represent the mean measurement value and a standard deviation for each point
soft triaxial force sensor mounted on the gripper of a commer- is also provided. The top row presents the Normal Forces case. The second,
cial robotic manipulator to demonstrate a usage scenario where third, and fourth rows represent 30, 45 and 60 degree cases, respectively.
this common manipulation task greatly benefits from real-time Since the shear forces were applied in the XZ plane, we omit forces along
Y-axis and in Normal Force case we omit forces along both X and Y axes,
3-D force measurements. which remain close to zero.

A. Static Validation Results

As previously discussed, the data obtained from the fourth layers was used to predict the forces for this sensor, and
sensor was not included in the training set and it was 1000 randomly sampled error points are shown in Figure 5.
reserved for validation. The final neural network with 5 hidden Figure 6 displays the actual forces (on the horizontal axis
3858 IEEE SENSORS JOURNAL, VOL. 18, NO. 9, MAY 1, 2018

Fig. 7. Pyramidal sensor with Dragonskin 30 as the soft substrate. The Fig. 8. Rectangular sensor with Ecoflex 0030 as the soft substrate. The
comparison of linear regression (dashed magenta lines with circle markers) comparison of linear regression (dashed magenta lines with circle markers)
and trained neural network forces (dash-dot black lines with square markers) and trained neural network forces (dash-dot black lines with square markers)
are overlaid. Forces measured from load cell are plotted on the X-axis whereas are overlaid. Forces measured from load cell are plotted on the X-axis whereas
forces measured from the soft force sensor are plotted on the Y-axis. The top forces measured from the soft force sensor are plotted on the Y-axis. The top
row presents pure normal loading. The second, third, and fourth rows represent row presents pure normal loading. The second, third, and fourth rows represent
30, 45, and 60 degree loading cases, respectively. Since the shear forces were 30, 45, and 60 degree loading cases, respectively. Since the shear forces were
applied in the XZ plane, we omit forces along Y-axis and in Normal Force applied in the XZ plane, we omit forces along Y-axis and in Normal Force
case we omit forces along both X and Y axes. case we omit forces along both X and Y axes.

of these curves) that were applied on the sensor for pure of the forces on this dataset. Also, the FCN network is able
normal loading, and shear (plus normal) loading at 30°, 45°, to accurately measure the forces for shear loading at 45°,
and 60°, all about the XZ plane and with the corresponding (a loading condition for which it was not trained and on a
forces measured by the sensor using the neural network model sensor prototype which was not included in training). From
(shown on the vertical axis). Only the relevant forces which these experiments, we conclude that the network is able to
were subject to change are plotted in the figure (i.e. the shear generalize very well on new sensors and loading conditions.
measurements in Y-axis remain at zero and are not shown). The time required for predicting forces during validation for
We see that the actual to measured force curve closely follows each input vector (three Hall voltages) to the neural network
the expected diagonal line with a slope of 1, thus showing was 0.34 msec and thus the FCN does not introduce time delay
that the applied forces are measured accurately by the sensor into the sensing system during real-time operation.
using the proposed neural network model. We also see that the It is true that complex modeling schemes are usually unde-
deviation of the measurements from actual forces is very low at sirable and thus, we need to justify our choice of using a
low forces and more pronounced at higher loads. This could neural network model to calculate the 3-D forces from Hall
be due to the Hall element approaching the saturation limit voltage signals. To show neural network modeling is useful
and reducing its linearity between voltage to magnetic field for soft force sensors with complex shapes and hyper elastic
values. A MAE of 11.7 mN was obtained on the magnitude materials, we compared sensor measurements from the neural
DWIVEDI et al.: DESIGN, MODELING, AND VALIDATION OF A SOFT MAGNETIC 3-D FORCE SENSOR 3859

Fig. 9. Dynamic testing results indicate that applied and measured forces follow a 1-1 diagonal (left). Hysteresis of loading and unloading for the applied
and measured force-displacement data are in agreement (middle). The tracking of applied and measured forces during dynamic loading overlap in time axis
(7 seconds shown from the 150-second experiment) (right).

network with a simpler calibration approach, linear regression. instrument at WPI Biomedical Engineering Department. Only
The results are shown in Figure 7. Sensed forces from the FCN normal compressive forces were applied on the sensor and
outperformed linear regression results in all cases. An MAE the frequency of the applied forces was set at 0.6 Hz (limited
of 1.23 N was observed on the validation data when trained by the speed of the instrument) and the forces applied were
with linear regression. The MAE from the the neural network between 0.3 N and 1.0 N as shown in Figure 9. The sensor
on the same dataset was 0.3014 N. In the 45° case, which was pre-compressed using a force of 0.3 N to eliminate any
was not included in the training set for either approach, linear potential shifting of the sensor during the experiments. The
regression significantly underperformed as well. application of dynamic forces on silicone rubber results in
The sensor data shown in Figure 7 is obtained from a hysteresis due to viscoelastic effects (i.e. the deflection of
sensor made with Smooth-On Dragonskin 30 as opposed to the material differs for the same force during loading and
Ecoflex 0030. Since Dragonskin 30 is a stiffer material, a cor- unloading). The hysteresis plot is shown in Figure 9-Middle
responding increase in the measured force range is observed. Panel. This curve indicates that the measured forces follow
Our results indicate that the same neural network calibration the actual force closely. In other words, the proposed sensor
approach adjusts well to different material types with the same is able to map the effect of hysteresis very well and provide a
sensor shape. measure of dynamic forces accurately. Figure 9-Right Panel
In addition, a justification is needed for our choice of using shows the time response during a representative dynamic
a pyramid shape for the sensor since it is more complex loading experiment, where the sensor output tracks the applied
than a simple rectangular block design. To this end, we made dynamic forces. The accuracy of dynamic force measurements
the same set of experiments on a rectangular sensor design is better inferred from Figure 9-Left Panel, which shows
made from Ecoflex 0030 and kept the distance between measured force with respect to corresponding applied force.
the magnet and the IC surface same as in the pyramidal Here we see that the relationship between the measured and
shape sensor and following the same calibration routine. Our applied forces follows a line with a slope equal to 1.0 with
experimental results are shown in Figure 8. Just as with the minimal variation.
Dragonskin 30 pyramid design, forces obtained from neural
network are more accurate as opposed to linear regression C. Force Controlled Pick and Place Case Study
where MAE on validation dataset was 0.1110 N and 0.2827 N, The proposed force sensor can be used to grasp soft or deli-
respectively. However, small deviations in 30°, 60° and a large cate objects with manipulators that are not designed to handle
deviation in z-axis measurement in 45° suggests rectangular such objects. To demonstrate this application, we perform a
shape to be undesirable for accurate force measurements. pick and place experiment with the Jaco arm (Kinova robotics,
We attribute these deviations to free magnet rotations when Boisbriand, QC, Canada) manipulating an egg. The arm is
pushed at an angle on the rectangular surface. The pyramid a 6-degree-of-freedom (DOF) manipulator with a 3-finger
shape, on the other hand, allows us to position the magnet gripper. With the proposed soft tri-axial force sensor we are
near the centroid and reduce undesired rotations of the magnet, able to perform force control and thus the manipulator is able
enabling repeatable and accurate force measurements. to transport the egg from one location to another without
breaking it.
B. Dynamic Validation Results The arm was programmed to perform the pick and place
Dynamic loading experiments were performed using task between two known locations on a tabletop. The goal is
an Instron Electroplus-e1000 Linear-Torsion force testing for the arm to sense the forces applied on the egg in real time
3860 IEEE SENSORS JOURNAL, VOL. 18, NO. 9, MAY 1, 2018

Fig. 10. Significant time intervals from the grasp experiment are shown. At t1 robot is at its initial state and no force is measured on the sensor. Shear force
and normal force are detected as fingers get in contact with the object during t2 . The interval t3 represents the movement of the object from initial point
to the target point. The peak in this interval is due to stabilization of the egg between two fingers. Finally the object is released by detecting the change in
y-axis force in interval t4 . A coordinate frame attached to the first snapshot at t1 represents the measurement axes of the force sensor.

TABLE V
C OMPARISON OF C OMMERCIAL AND P UBLISHED F ORCE S ENSORS A LONG W ITH THE P YRAMIDAL E COFLEX 0030 S ENSOR P RESENTED IN T HIS PAPER
A LONG I MPORTANT P ROPERTIES OF S IZE , S AMPLING R ATE , H YSTERESIS , M EASURED F ORCE A XES , R ANGE AND S ENSITIVITY

to establish and maintain a gentle grip and release the egg follow a trajectory in multiple phases. The corresponding
when it touches the table top at the destination. The software forces measured in the phases are shown in Figure 10. Here,
to operate the arm was developed using Robot Operating the sensor is placed on the finger pad, where normal contact
system (ROS) [31]. The manipulator arm was configured to forces coincide with the z-axis (local normal) of the sensor and
DWIVEDI et al.: DESIGN, MODELING, AND VALIDATION OF A SOFT MAGNETIC 3-D FORCE SENSOR 3861

gravity is along the y-axis (local shear). There are no expected and shear, respectively. The bandwidth of the Melexis sensor
forces along the x-axis, which is reflected in the measurements. IC can work up to 400 Hz. Dynamic loading experiments
The gap in the gripper when fingers are fully closed was indicate that the sensor is able to accurately follow dynamic
bigger than the size of the egg. Hence, the gripper fingers were forces applied at 0.6 Hz despite the hysteresis exhibited by
padded with a silicone rubber layer in order to reduce this gap. the material.
The weight of the egg was 50 g. The arm could lift the egg We present a comparison of the pyramidal Ecoflex 0030 sen-
without any damage following a simple force-control process sor with commercially available and published force sensor
as shown as snapshots in Figure 10. The four phases during designs in Table V. Hysteresis in piezoelectric and capac-
this task are defined as follows: itive commercially available sensors seems negligible but
1) Gripper Closing Phase: Shown by the period t1 our findings suggest viscoelasticity of soft materials adds
in Figure 10, here the gripper is at the initial pose and the significant hysteresis and we expect other works that use
fingers start closing. All forces remain at 0 N throughout this similar materials to exhibit hysteresis effects. In most cases the
period, taking about 6 seconds before the fingers make contact range of the sensor can be adjusted by picking different soft
with the egg. materials or adjusting the gains of the amplifier attached to the
2) Grasping and Lifting Phase: Shown by the period t2 sensor. The pyramidal Ecoflex 0030 sensor is quite sensitive,
in Figure 10, here the gripper initiates contact with the egg as measuring forces as small as 5 mN (error range of the load
seen by the increase in FZ . The fingers continue closing until cell used for calibration) within a suitable range of 0-1 N
a predefined threshold is reached to produce a normal grasping for tactile applications. In Table V, Single Tact Sensor and
force between contact surfaces to hold on to the object due embedded microfluidic channel based sensor design stand out
to friction, with negligible deformation. The value of the safe as high sensitivity sensors within their maximum force range.
limit was determined through prior trials to be 0.15 N to 0.35 N Single Tact Sensor, however is only capable of measurements
(as measured by the force sensor) for the egg. After grasping in a single dimension. As for sensitivity the Ecoflex 0030 Pyra-
the object, the arm starts lifting up the egg at around 7 seconds. mid sensor (5 mN) is only outperformed by the commercial
In this phase, the force sensor experiences an increase in the OMD-10 sensor (2.5 mN). In terms of package size the
shear force FY due to gravity. Ecoflex 0030 Pyramid sensor has a volume of 380 mm3 .
3) Relocating Phase: Shown by the period t3 in Figure 10, Sensors having a smaller volume are Tekscan and Single
here the arm moves the gripper along with the egg over to a Tact Sensor, which only provide normal force measurements.
destination location while keeping the gripper position 10 cm This volume difference is expected since both sensors are
above the desk level. Slight disturbances are seen in this time fabricated on a thin sheet and do not have soft substrates over
period which could be attributed to a shaky movement of the them. Overall, we conclude that the pyramid sensor shape,
Jaco arm. The short pulse in this phase at 14 seconds coincides magnetic field measurements, and a soft substrate as the force
with a slight shift in the position of the egg between the measurement medium provides a good combination of size,
fingers. measured axes, and sensitivity.
4) Placing Phase: Shown by the period t4 in Figure 10, A force controlled pick and place experiment was performed
the gripper moves down towards the table. As the egg makes on an egg using the force sensor mounted on a Kinova Jaco
contact with the table top, the shear force decreases. This arm. We show that the sensor provides stable output force data
decrease in shear force allows the arm to recognise that the egg in real-time and we can use both shear and normal force data
has been placed at the destination spot. At this time, the Jaco to successfully perform fragile object manipulation.
arm opens its grip to release the egg and the arm goes back Future work will focus on the combination of multiple
to the starting position. sensor units in an array. These arrays of force sensors will
The sensing of the shear forces while placing an object help monitor accurate distributed force measurements in 3-D.
helps the manipulator to sense that the object has made contact More work is planned to further study the effect of the shape
with the ground surface and allow it to place the object of the sensor and the location of the magnet within this shape.
safely and gently at the destination spot without dropping the Usage of the sensor in tasks such as walking gait and balance
object or strongly hitting the ground surface. analysis and haptic feedback will also be explored.

V. C ONCLUSION AND F UTURE W ORK ACKNOWLEDGMENT

In this paper we described the design, fabrication, character- The authors would like to thank Joshua R. Gershlak from the
ization, and experimental validation of a Hall effect based 3-D Biomedical Engineering Department at WPI for his valuable
soft force sensor in a pyramid shaped soft elastomer matrix. assistance on the dynamic testing of the sensor. Any opinions,
We trained a fully-connected neural network for characteriza- findings, and conclusions or recommendations expressed in
tion and mapping of measured voltages to 3-D forces. We show this material are those of the authors and do not necessarily
that the resulting mapping generalizes well for sensors and reflect the views of the NSF.
loading conditions that are not part of the training dataset.
Our experimental results show that the proposed sensor is R EFERENCES
highly accurate and it can measure forces in normal and shear
[1] R. S. Dahiya, G. Metta, M. Valle, and G. Sandini, “Tactile sensing—from
directions within a range of 0 N to 1.1 N and ±1.5 N with humans to humanoids,” IEEE Trans. Robot., vol. 26, no. 1, pp. 1–20,
an error of 2% and 2.2% of the full scale reading in normal 2010.
3862 IEEE SENSORS JOURNAL, VOL. 18, NO. 9, MAY 1, 2018

[2] H. R. Nicholls and M. H. Lee, “A survey of robot tactile sensing [26] T. Chen, I. Goodfellow, and J. Shlens. (2015). “Net2net:
technology,” Int. J. Robot. Res., vol. 8, no. 3, pp. 3–30, 1989. Accelerating learning via knowledge transfer.” [Online]. Available:
[3] H. Yousef, M. Boukallel, and K. Althoefer, “Tactile sensing for dex- https://arxiv.org/abs/1511.05641
terous in-hand manipulation in robotics—A review,” Sens. Actuators A, [27] C. Li, P.-M. Wu, S. Lee, A. Gorton, M. J. Schulz, and C. H. Ahn,
Phys., vol. 167, no. 2, pp. 171–187, 2011. “Flexible dome and bump shape piezoelectric tactile sensors using
[4] J. J. Clark, “A magnetic field based compliance matching sensor for PVDF-TrFE copolymer,” J. Micromech. Syst., vol. 17, no. 2,
high resolution, high compliance tactile sensing,” in Proc. IEEE Int. pp. 334–341, Apr. 2008.
Conf. Robot. Autom., vol. 2, Apr. 1988, pp. 772–777. [28] F. Chollet. (2015). Keras. [Online]. Available: https://github.
[5] R. B. Katragadda and Y. Xu, “A novel intelligent textile technology com/fchollet/keras
based on silicon flexible skins,” Sens. Actuators A, Phys., vol. 143, no. 1, [29] Q. Nie and F. C. Sup, “A soft four degree-of-freedom load cell based on
pp. 169–174, 2008. the Hall effect,” IEEE Sensors J., vol. 17, pp. 7355–7363, Nov. 2017.
[6] K. Noda, K. Hoshino, K. Matsumoto, and I. Shimoyama, “A shear stress [30] Y. Liu, H. Han, T. Liu, J. Yi, Q. Li, and Y. Inoue, “A novel tactile
sensor for tactile sensing with the piezoresistive cantilever standing in sensor with electromagnetic induction and its application on stick-slip
elastic material,” Sens. Actuators A, Phys., vol. 127, no. 2, pp. 295–301, interaction detection,” Sensors, vol. 16, no. 4, p. 430, 2016.
2006. [31] M. Quigley et al., “ROS: An open-source robot operating system,” in
[7] K. Kim et al., “Polymer-based flexible tactile sensor up to 32× 32 Proc. ICRA Workshop Open Sour. Softw., 2009, vol. 3, nos. 3–2, p. 5.
arrays integrated with interconnection terminals,” Sens. Actuators A,
Phys., vol. 156, no. 2, pp. 284–291, Dec. 2009.
[8] M. Sohgawa et al., “Tactle array sensor with inclined chromium/silicon
Anany Dwivedi received the B.Tech. degree in
piezoresistive cantilevers embedded in elastomer,” in Proc. Int.
electronics and communication engineering from the
Solid-State Sens., Actuators Microsyst. Conf., Jun. 2009, pp. 284–287.
LNM Institute of Information Technology, Jaipur,
[9] D. M. Vogt, Y.-L. Park, and R. J. Wood, “Design and characterization
India, in 2015, and the M.S. degree in robotics
of a soft multi-axis force sensor using embedded microfluidic channels,”
engineering from the Worcester Polytechnic Insti-
IEEE Sensors J., vol. 13, no. 10, pp. 4056–4064, Oct. 2013.
tute (WPI), Worcester, MA, USA, in 2017. As a
[10] M. Khondoker and D. Sameoto, “Fabrication methods and applications Graduate Researcher at the Soft Robotics Lab, WPI,
of microstructured gallium based liquid metal alloys,” Smart Mater. his research focus was on design and fabrication
Struct., vol. 25, no. 9, p. 093001, 2016. of soft sensors. He is currently pursuing the Ph.D.
[11] D. J. Lipomi et al., “Skin-like pressure and strain sensors based on degree with the New Dexterity Research Group,
transparent elastic films of carbon nanotubes,” Nature Nanotechnol., University of Auckland, New Zealand. He works
vol. 6, pp. 788–792, Oct. 2011. on development of brain–machine interfaces to be used to augment the
[12] L. Viry et al., “Flexible three-axial force sensor for soft and highly intuitiveness of the control of technical devices to help individuals with motor
sensitive artificial touch,” Adv. Mater., vol. 26, no. 17, pp. 2659–2664, or sensory impairments to regain their lost dexterity.
2014.
[13] Y. Yang and Y. Chen, “Innovative design of embedded pressure and
position sensors for soft actuators,” IEEE Robot. Autom. Lett., vol. 3,
no. 2, pp. 656–663, Apr. 2018. Anand Ramakrishnan received the B.Tech. degree
[14] Y. Tenzer, L. P. Jentoft, and R. D. Howe, “The feel of mems barometers: in mechanical engineering from SRM University,
Inexpensive and easily customized tactile array sensors,” IEEE Robot. Kattankulathur, India, in 2015, and the M.Sc. degree
Autom. Mag., vol. 21, no. 3, pp. 89–95, Sep. 2014. in robotics engineering from the Worcester Poly-
[15] J. Yi, X. Zhu, L. Shen, B. Sun, and L. Jiang, An Orthogonal Curvature technic Institute (WPI), Worcester, MA, USA, in
Fiber Bragg Grating Sensor Array for Shape Reconstruction. Berlin, 2017, where he is currently pursuing the Ph.D.
Germany: Springer, 2010, pp. 25–31. degree with the Computer Science Department. As a
[16] S. Youssefian, N. Rahbar, and E. Torres-Jara, “Contact behavior of soft Graduate Researcher at the Soft Robotics Lab, WPI,
spherical tactile sensors,” IEEE Sensors J., vol. 14, no. 5, pp. 1435–1442, his research focus was on designing and calibrating
May 2014. of soft sensors. He works on building end to end
[17] C. Cho and Y. Ryuh, “Fabrication of flexible tactile force sensor using deep learning systems to analyze teaching videos
conductive ink and silicon elastomer,” Sens. Actuators A, Phys., vol. 237, and predict the quality of student–teacher interactions in them.
pp. 72–80, 2016.
[18] S. Ozel, N. A. Keskin, D. Khea, and C. D. Onal, “A precise embedded
curvature sensor module for soft-bodied robots,” Sens. Actuators A,
Phys., vol. 236, pp. 349–356, Dec. 2015. Aniketh Reddy received the bachelor’s degree in
electronics and communication engineering from
[19] M. Luo et al., “Slithering towards autonomy: A self-contained soft
Jawaharlal Nehru Technological University, Hyder-
robotic snake platform with integrated curvature sensing,” Bioinspiration
abad, India, in 2017. He is currently pursuing the
Biomimetics, vol. 10, no. 5, p. 055001, 2015.
master’s degree in robotics engineering from the
[20] C. Ledermann, S. Wirges, D. Oertel, M. Mende, and H. Woern, “Tactile Worcester Polytechnic Institute (WPI), Worcester,
sensor on a magnetic basis using novel 3D Hall sensor—First prototypes MA, USA. As a Graduate Researcher at the Soft
and results,” in Proc. IEEE 17th Int. Conf. Intell. Eng. Syst. (INES), Robotics Lab, WPI, his research focus is on design
Jun. 2013, pp. 55–60. and fabrication of sensors.
[21] L. Jamone, G. Metta, F. Nori, and G. Sandini, “James: A humanoid
robot acting over an unstructured world,” in Proc. 6th IEEE-RAS Int.
Conf. Humanoid Robots, Dec. 2006, pp. 143–150.
[22] T. P. Tomo et al., “Design and characterization of a three-axis Hall
effect-based soft skin sensor,” MDPI Sens., vol. 16, no. 4, p. 491, Kunal Patel received the B.Tech. degree in electron-
2016. ics and telecommunication from the College of Engi-
[23] E. Torres-Jara, I. Vasilescu, and R. Coral, “A soft touch: Compliant neering Pune, India, and the M.Sc. degree in robotics
tactile sensors for sensitive manipulation,” Massachusetts Inst. Technol. engineering from the Worcester Polytechnic Insti-
Comput. Sci. Artif. Intell. Lab., Cambridge, MA USA, Tech. Rep. MIT- tute, Worcester MA, USA. He pursued a career in
CSAIL-TR-2006-014, 2006. embedded systems for three years at Cypress Semi-
[24] T. P. Tomo et al., “SNR modeling and material dependency test of a conductors Inc. after which he decided to explore
low-cost and simple to fabricate 3D force sensor for soft robotics,” in his passion for the field of robotics. He is currently
Proc. IEEE/SICE Int. Symp. Syst. Integr. (SII), Dec. 2016, pp. 428–433. a Robotics Engineer at Near Earth Autonomy Inc.,
[25] J. Gafford, F. Doshi-Velez, R. Wood, and C. Walsh, “Machine learning Pittsburgh, PA, USA. His current work focuses on
approaches to environmental disturbance rejection in multi-axis opto- development of GPS and IMU-based inertial navi-
electronic force sensors,” Sens. Actuators A, Phys., vol. 248, pp. 78–87, gation systems. His research interests include soft robotics, SLAM, sensor
Sep. 2016. fusion, and state estimation.
DWIVEDI et al.: DESIGN, MODELING, AND VALIDATION OF A SOFT MAGNETIC 3-D FORCE SENSOR 3863

Selim Ozel received the B.Sc. and M.Sc. degrees in Cagdas D. Onal received the B.Sc. and M.Sc.
mechatronics engineering from Sabanci University degrees from the Mechatronics Engineering Pro-
in 2011 and 2013, respectively. He is currently pur- gram, Sabanci University, Turkey, in 2003 and 2005,
suing the Ph.D. degree in robotics engineering with respectively, and the Ph.D. degree in mechanical
the Worcester Polytechnic Institute. His research is engineering from Carnegie Mellon University in
focused on the sensorization of soft bodied materials, 2009. He is currently an Assistant Professor of
soft robotics, and control of compliant mechanisms. Mechanical Engineering and Robotics Engineering
at the Worcester Polytechnic Institute (WPI), where
he is the Director and Founder of the WPI Soft
Robotics Laboratory and Co-Founder of the Human
Augmentation Laboratory. He was a Post-Doctoral
Associate with the Computer Science and Artificial Intelligence Laboratory,
MIT. His current research interests include soft robotics, origami-inspired
printable robotics, alternative actuation/sensing mechanisms, bio-inspiration,
dynamic modeling, and control theory.