Supplementary Materials For: Omniskins: Robotic Skins That Turn Inanimate Objects Into Multifunctional Robots

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 16

robotics.sciencemag.

org/cgi/content/full/3/22/eaat1853/DC1

Supplementary Materials for

OmniSkins: Robotic skins that turn inanimate objects into multifunctional robots

Joran W. Booth, Dylan Shah, Jennifer C. Case, Edward L. White, Michelle C. Yuen, Olivier Cyr-Choiniere,
Rebecca Kramer-Bottiglio*

*Corresponding author. Email: rebecca.kramer@yale.edu

Published 19 September 2018, Sci. Robot. 3, eaat1853 (2018)


DOI: 10.1126/scirobotics.aat1853

The PDF file includes:

Text S1. Actuators in robotic skin prototypes.


Text S2. Capacitive sensors and signal conditioning.
Text S3. Robotic skin construction.
Text S4. Predicting deformation of a known object.
Text S5. Predicting deformation of an unknown object.
Text S6. Control algorithms.
Text S7. Locomotion gaits using robotic skins.
Text S8. Rapid prototyping with robotic skins.
Fig. S1. McKibben actuator schematic.
Fig. S2. Capacitive sensor construction and characterization.
Fig. S3. Three robotic skin prototypes that vary in components, configuration, and integration.
Fig. S4. Force versus displacement curves for SMA/McKibben actuators and soft cylindrical
bodies.
Fig. S5. A geometric model of a robotic skin deflecting a soft cylinder, with the neutral axis
located along the surface of the cylinder.
Fig. S6. A geometric model of a robotic skin deflecting a soft cylinder, with the neutral axis
located along the surface of the cylinder.
Fig. S7. Control algorithms.
Fig. S8. Possible locomotion gaits using robotic skins with and without soft cylindrical bodies.
Fig. S9. Examples of rapid prototyping with robotic skins.
References (37–39)

Other Supplementary Material for this manuscript includes the following:

(available at robotics.sciencemag.org/cgi/content/full/3/22/eaat1853/DC1)

Movie S1 (.mp4 format). Transferability: continuum robots to locomotion robots.


Movie S2 (.mp4 format). Inchworm locomotion: teleoperated with camera.
Movie S3 (.mp4 format). Inchworm locomotion: controlled by light.
Movie S4 (.mp4 format). Continuum manipulation.
Movie S5 (.mp4 format). Wearable posture shirt.
Movie S6 (.mp4 format). Tensegrity ball rolling.
Text S1. Actuators in r obotic s kin p rototypes
Pneumatic Actuators. The pneumatic actuators used for our robotic skin implementations
were McKibben actuators constructed with braided mesh (McMaster-Carr, 9284K2), latex bal-
loons, and tubing connectors (McMaster-Carr, 2974K331), following the same method de-
scribed in Park et al. (8) (Figure S1). Our actuators were built using 105 mm lengths of 6.35 mm
(0.25 in) diameter mesh wrapped around 110 mm lengths of latex balloons. The mesh had a
rest angle of θ0 = 29◦ and a rest diameter of D0 = 9 mm. A 35 mm length of Tygon tubing
(McMaster-Carr, 6516T62) was inserted into the balloon and secured with three 2 mm zip-ties
to ensure a good pneumatic seal. The tubing was then cut flush to the end of the mesh and the
tied end of the balloon was secured to the mesh on the other end. The zip ties were placed so
that there was a 100 mm length between the two inner most ties. Force-versus-displacement
measurements of a single actuator were obtained at 140 kPa in a materials testing machine (In-
stron 3345) and were found to be similar to predictions obtained using the theory presented in
Daerden & Leferber (37) (Figure S4A). The actuator force, FA , is predicted as:

π D02 P
 
3 2 1
FA = (1 − ) −
4 tan2 (θ0 ) sin2 (θ0 )
(1)

where D0 is the rest diameter of the mesh, θ0 is the rest angle of the mesh fibers, P is the applied
pressure, and  is the strain.

Fig. S1. McKibben actuator schematic.

SMA Actuators. The SMA actuators were constructed using Nitinol wire that was programmed
using the same spooling and heat treatment method as described in Seok et al. (11). The SMA
actuators were formed from 0.381 mm (.015 in) diameter, d, Nitinol wire (70◦ C transition tem-
perature, Dynalloy, Inc.) coiled into springs with an inner diameter of D = 2.34 mm. The
Nitinol wire was wound around a metal rod and annealed for 2 hours at 390◦ C, followed by
quenching in water. It was reheated again for 10 minutes and quenched. This last procedure
was repeated ten times to complete the shape memory programming. Each final, un-stretched
actuator was 15 mm long with n = 39 coils. Each actuator also had a loop on each end for later
integration steps. Force-versus-displacement measurements of a single actuator were obtained
in a materials testing machine (Instron 3345) by running 0.8 A current through the SMA to
keep it in austenite phase, and having the Instron stretch the SMA from its contracted state to
80 mm. Figure S4A shows the results of this test, along with theoretical results using the two-
state thermo-mechanical model proposed by An et al. (38). For an actuated SMA coil (austenite
phase), the force is given by:

 
4 3
GA d δ  cos(αi )
FA = 3
 
8 D n cos(α )2 cos(α )2 + sin(αf )2
f f 1+ν
(2)

The left fraction in the equation is the conventional spring equation, where δ represents the
change in length (mm). The right fraction is a correction term which accounts for large defor-
mations, where αf is the current coil angle at any position and ν is the Poisson ratio. The initial
coil angle, αi , is typically assumed to be zero for a closed spring, with little error. Nitinol has a
modulus of rigidity in the austenite phase, GA , of 18.26 GPa (38).

Text S2. Capacitive sensors and signal conditioning


The capacitive sensors used in our robotic skin implementations were a modification of the
three-layer system described in White et al. (22), consisting of a ground layer, a dielectric layer
and an active layer. In our system, we added additional dielectric and ground layers such that
the active layer is surrounded by two dielectric layers and two ground layers (Figure S2). One
end of the sensor had the bottom (ground) and middle (active) composite layer attached to
electrodes, while the other end of the sensor was missing the central composite layer, therefore
having only the top and bottom “ground layers”. The ground-only end had a staple through it
to electrically connect the two sides. Because the two ground layers are connected electrically,
they act as a shield for the measurement layer, which substantially reduces environmental noise.
The conductive layers in our sensors were a silicone composite consisting of 10wt% expanded
intercalated graphite (EIG) and 90wt% DragonSkin 10 Slow silicone (Smooth-On, Inc.). The
dielectric material was DragonSkin 10 Slow.
A proxy for the capacitance of each sensor was measured using a PIC16F1825 microcon-
troller (Microchip Technologies), as described in (22), with a slight modification. Rather than
continuously passing data through an analog filter, we passed the data through a digital expo-
nential filter built into the Arduino software. Specifically, after each read request, the board
charged the sensor for a pre-specified time, then discharged and counted the number of instruc-
tion cycles that the system took to decay to ground voltage. The resulting value was transmitted
to an Arduino Uno using the I2C protocol. We note that the time required for a charge-discharge
cycle is related to the time constant of the sensor, and could be used to calculate the capacitance.
The sensor capacitance has a linear relationship with extension and the gauge factor is consistent
between sensors, as seen in Figure S2C.
Fig. S2 . Capacitive sensor construction and characterization. (A) Schematic of the five-layer sensor
construction. Two ground layers, shown in black, are connected by a staple at one end. These two layers
surround the active layer, shown in red. The gray layers indicate dielectric material. (B) A back-lit photo
of a five-layer sensor. (C) Sensor feedback versus extension. The mean, a linear least-squares fit, and two-
sided 95% confidence intervals are plotted for each sensor. The y-axis shows the number of instruction
cycles on the microprocessor, which is directly related to time required to discharge, proportional to the
microprocessor’s processing speed (i.e., a proxy for capacitance).
To minimize parasitic capacitance of the connection between the microcontroller and each
sensor, we located a signal conditioning board as close to each sensor as possible. We sol-
dered two short strips of copper to the conditioning board and sandwiched the sensor between
them. The assembly was secured by sewing the pieces together, using polystyrene sheets as a
reinforcement for the soft sensor material.

Text S3. Robotic skin construction


We fabricated three types of robotic skin prototypes that vary in components, configuration, and
integration:

1. Rectangular pneumatic. Components: pneumatic McKibben actuators, elastomer sub-


strate, conductive composite-based capacitive sensors. Configuration: parallel compo-
nents, rectangular. Tightly integrated, shown in Figure S3A.

2. Rectangular SMA. Components: SMA actuators, fabric substrate, conductive composite-


based capacitive sensors. Configuration: parallel components, rectangular. Modular,
shown in Figure S3B.

3. Triangular pneumatic. Components: pneumatic McKibben actuators, fabric substrate,


conductive composite-based capacitive sensors. Configuration: triangulation of compo-
nents, triangular. Modular, shown in Figure S3C-D.

One of the key differences between the robotic skin designs we chose to implement is the
level of integration, shown in Figure S3. For the rectangular pneumatic skin design, sensors
and actuators were bonded along their entire length to the silicone substrate, creating a tightly-
integrated skin. In contrast, for the rectangular SMA skin design, sensors and actuators were
sewn onto the fabric substrate. The triangular skin design employs pneumatic actuators and a
fabric substrate, assembled using a modular component integration method. This difference in
integration represents a distinction between a robotic material, where all of the components are
integrated such that they cannot be separated, and a robotic material system, where individual
components are removable or interchangeable. While component-level modularity has been
advantageous in our initial exploratory work (if a component fails it can be easily repaired
or replaced), we believe tighter integration and further development of robotic materials will
enable more robust robotic skin implementations due to reduced component-count and material
interfaces.
The elastomer substrate for the rectangular pneumatic robotic skin was made using platinum-
cure DragonSkin 10 Slow rod-coated into a film approximately 2 mm in thickness. To prevent
tearing, strain limiters were fabricated from muslin cloth, impregnated with DragonSkin 10
Slow, and bonded to locations where rigid elements needed to connect to the skin. Sensors
with an active length of 90 mm were attached to the substrate using DragonSkin 10 Slow and
Fig. S3. Three robotic skin prototypes that vary in components, configuration, and integration. (A) A
rectangular robotic skin with pneumatic actuators and sensors bonded to the substrate. (B) A rectangular
robotic skin with SMA actuators and sensors sewn to the substrate. The actuators are only attached at
each end of the sensor and are not constrained in the middle. (C) A triangular robotic skin with modular
pneumatic actuators and sensors. (D) The parts of the triangular skin were made to be modular in order
to aid replacement of defective parts on a large system such as the tensegrity robot.

the actuators were attached to the sensors with tin-cure Mold-Max 10 silicone (Smooth-On,
Inc.). In contrast to the platinum-cure silicone, the tin-cure silicone was not cure inhibited in
the presence of the latex balloons in the McKibben actuators. In one version, a thin layer of
medical-grade platinum-cure adhesive silicone was applied to the back of the elastomer sub-
strate to improve force transduction between the skin and the body. The skin with this backing
is shown in Figure 6A.
The fabric substrate for the rectangular SMA robotic skin was made from Spandex fabric,
with muslin cloth underneath the ends of the sensors to limit strain to the sensors’ active length.
The sensors were sewn onto the substrate. The SMA actuators were attached to 3D printed poly-
lactic acid (PLA) mounts that were sewn onto the fabric at the ends of the sensors. These mounts
also can be used to attach a zip-tie as a strain-limiting element, to improve force transduction
between the skin and deformable body. This approach was used in the SMA demonstrations.
The un-stretched sensor lengths were 80 mm for the SMA skin. For all demonstrations shown,
a 10 mm pre-strain was applied, generating an active sensor length of 90 mm, to prevent the
sensors on the active side of the skin from buckling during use.
The fabric substrate for the triangular pneumatic robotic skin was made from Spandex fab-
ric, with muslin cloth at each corner to limit the strain where the sensors and actuators connect.
The sensors and actuators were attached to the substrate using fabric snaps, allowing the assem-
bly to be more modular. This approach was used in order to improve repairs if an actuator or
sensor were to fail. The sensor length was chosen such that the sensor would have a 10% strain
at the minimum, contracted length of the actuators. Unlike the rectangular skins, the actuators
and the sensors in the triangular skins are not directly coupled. The actuators form a trian-
gle offset from the triangle formed by the sensors, which is a consequence of the component
modularity.
The remaining components for the robotic skin assemblies included wire harnessing, clothing-
style snap buttons for connecting skins together, anchor points for actuators printed from PLA
or laser cut from polystyrene sheets, and a layer of Spandex fabric on the top of each skin to
conceal the components. Off-board components included a microcontroller (Arduino Uno), a
computer for data collection, power supplies, current drivers for the SMA actuators, and valves
(X-Valve X212LF, Parker). I2C communication protocol was used for PIC-Arduino communi-
cation, while USB serial communication was used for Arduino-PC communication. The foam
cylinders used for the deformable bodies were made from polyethylene.

Text S4. Predicting d eformation of a k nown o bject


We have shown that robotic skins are able to control inert, deformable bodies from their surface.
One useful motion that can be generated using this approach is bending, which may be achieved
by wrapping robotic skins around cylindrical foam bodies with the actuators oriented along the
length of the body. This useful configuration can produce soft joints, segments in continuum
robots, robot limbs, or locomotion. Here, we describe the method followed for predicting the
maximum deformation (i.e., equilibrium) of bending systems comprised of a robotic skin and
soft cylindrical body.
First, McKibben and SMA actuators were characterized on an Instron materials testing ma-
chine. Each test began with the actuator in its fully actuated, contracted position. Then, the
materials tester pulled the actuator to extend it while maintaining actuation and recorded the
blocked force as a function of this extension. We note that the actuators were strain-cycled at
a sufficiently slow rate to assume quasi-static conditions. In the force-displacement plot shown
in Figure S4A, actuator displacement is defined as Le − Lc , where Le is the fully extended
actuator length (in our case 90 mm) and Lc is the contracted actuator length measured by the
materials tester, such that the displacement in the extended position is zero. The blocked force
approaches zero as the actuator length approaches its fully contracted position, as expected.

Fig. S4. Force versus displacement curves for SMA/Mckibben actuators and soft cylin-
drical bodies. Force vs. displacement curves for (A) SMA and pneumatic actuators, and (B) two
foam cylinders. The cylinder displacement is the change in length along one side of the cylinder
after two break-in cycles. Shaded regions around the means represent 95% confidence intervals.
Two foam cylinders were also characterized in an Instron materials testing machine to deter-
mine their bending stiffness. For each test, a string was attached along the length of the cylinder
and connected on one end to the Instron load cell. The Instron then pulled the string, which
reduced the string length between the ends of the cylinder, causing it to bend. The bending
stiffness (quantified as a resistive force to bending) was recorded as a function of string length
displacement. In the force-displacement plot shown in Figure S4B, zero displacement denotes
that the cylinder has no deformation. Resistive force increases as the magnitude of deformation
increases, as expected.
By comparing the force-displacement plots for both actuator components and body compo-
nents, we can predict the equilibrium of the system. Where an actuator curve intersects a body
curve, a system comprised of these two components can be expected to achieve its maximum
deformation at that intersection. Figures S4A-B are shown combined in Figure 3 of the main
text to show the intersections.
Actuator length, La , was converted to deflection, θ, using a geometric model. Using the
actuator length at the equilibrium point for the system yields the maximum deflection, θmax . The
model makes the following assumptions: (1) homogeneous materials; (2) constant curvature;
(3) the segment only compresses and does not stretch as it bends, which locates the neutral
axis along the outer surface of the cylinder opposite the bend. In practice, these assumptions
are accurate for an initial approximation: There is some stretching during the bending motion,
which is captured by the sensors and used in the control loops, but it is significantly less than
the compressive strain experienced on the opposing side.

Fig. S5. A geometric model of a robotic skin deflecting a soft cylinder, with the neutral axis located
along the surface of the cylinder. (A) An undeformed cylinder of length Lseg . (B) The cylinder deflects
with constant curvature.

With these assumptions and referring to Figure S5, we additionally define Lseg as the neutral
axis, R as the radius of the cylinder, and Lr as the length from the center of curvature to the
near edge of the cylinder. The neutral axis, Lseg , is related to θ by:

Lseg = (2R + Lr )θ (3)


and the length of the actuator, La , is related to θ using the Law of Cosines:

L2a = L2r + L2r − 2Lr Lr cos(θ) (4)


L2a
− 1 = −cos(θ) (5)
2L2r

Solving (3) for Lr and substituting (3) into (5), we obtain:

L2a
 2 + cos(θ) − 1 = 0 (6)
Lseg
2 θ
− 2R

Equation (6) may then be solved for θmax with a nonlinear equation solver, such as f solve in
MATLAB.

Text S5. Predicting d eformation of an u nknown o bject


Because robotic skins may be applied arbitrarily to passive systems, the material properties of a
host body may not be known beforehand. Therefore, it is valuable to be able to retrieve that data
from the skin itself in order to predict system performance in situ. In this section, we continue
the analytic treatment of the case of a robotic skin on a soft cylindrical body to show how
sensor data from the skins can be used to derive properties of the underlying body. Note that
the model listed below has been presented previously with full experimental validation in (39),
but is repeated here for clarity and completeness.
In order to derive the properties of the underlying body material, we first require knowledge
of system state to understand how the body reacts to surface actuation. To estimate system state,
we assume that the sensors in the robotic skin follow the curvature of the bending cylinder (i.e.,
perfect contact between the exterior skin and internal body), which requires us to relax the
previous assumption that neglects stretching and use the observed stretching in our model. We
therefore shift the neutral axis from the cylinder surface, as seen in Figure S5A-B, to the center
of the cylinder, as seen in Figure S6A-B.
We define our constant curvature segment in two ways: (1) using the length of the segment
(Lseg ), curvature (κ), and angular offset of curvature (φ); and (2) using the length of the segment
(Lseg ), curvature along the x-axis (κx ), andp curvature along the y-axis (κy ). We can convert
between these two definitions using κ = κ2x + κ2y and φ = tan−1 (κy /κx ). Our robotic
skin prototype includes four evenly spaced sensors, which, when placed around the cylinder,
can be used to define a coordinate frame such that the x- and y-axes align with the sensors
(Figure S6C).
With reliable data from all four sensors, we derive the system curvatures as:
Ls,4 − Ls,2 Ls,1 − Ls,3
κx = , κy = (7)
2Lseg R 2Lseg R
Fig. S6. A geometric model of a robotic skin deflecting a soft cylinder, with the neutral axis located
along the center of the cylinder. (A) An undeformed cylinder of length Lseg . (B) The cylinder deflects
with constant curvature. (C) Schematic of sensor placement around the cylinder to define a coordinate
frame (39).

where Ls,i refers to the length of the ith sensor and R is the radius of the cylindrical segment.
Eqn. 7 assumes that all sensors are giving accurate information. However, unless the cylinder
is being simultaneously strained and bent, it is likely that only two or three sensors are giving
reliable data while the other sensors are buckled or slack. Using only two or three sensors, we
we derive the system curvatures as:
Ls,4 − Lseg Lseg − Ls,2 Ls,1 − Lseg Lseg − Ls,3
κx = = , κy = = (8)
Lseg R Lseg R Lseg R Lseg R
where the appropriate equations are selected for determining κx and κy , depending of the axis
of curvature. Note that if there are three accurate sensors, a combination of equations can be
selected from Eqns. 7 and 8 as needed.
The state estimation model may be extended to also estimate the bending and elastic modu-
lus of the underlying material. For the stiffness estimation, we require at least one actuator that
is placed parallel to the central axis of the body. Contracting this actuator will cause an equilib-
rium between the actuator force and the passive cylinder to be reached, which is dependent on
the bending stiffness of the cylinder. We relate the bending stiffness (Kb ) to the actuator force
(F ) as:
RF
Kb = . (9)
κ
where κ and the corresponding actuator length are found from the state estimation. For a well-
characterized actuator, the actuator length relates to a given input (for example, pressure for
pneumatic actuators or voltage for SMA), which provides the actuator force. Noting that Kb =
EI if we assume linear material properties, we can also determine the elastic modulus (E) of
the material as:
RF
E= . (10)

where I is the second moment of inertia. With body properties and dimensions known, system
deformation can be predicted:
RF Lseg
θ= (11)
EI
The model presented here focuses specifically on a continuum segment. However, we be-
lieve that this model could be expanded to an array of skins applied to arbitrary morphologies.
Robotic skins that can derive the stiffness, and potentially morphology, of their host body can
populate predictive models to control system performance, which is a more generalized ap-
proach than formally characterizing candidate passive bodies.

Text S6. Control algorithms


The core element of the control algorithms used for the closed-loop demonstration are three-
state bang-bang logic controllers, as shown in Figure S7. When the input to a bang-bang con-
troller goes outside of an acceptable ‘deadband’, the controller sends a command to attempt to
return to the set-point. Upon attaining the set point, the controller sends a neutral signal until
the system again deviates outside the deadband.
In the current SMA skin implementation, a simple bang-bang logic controller sends a sig-
nal to either contract, hold, or release an actuator to manipulate the skin, and therefore the
underlying body, to achieve the desired length. Here the SMA’s slow response time (on the
order of seconds) allows the system to have relatively little overshoot or undershoot, and min-
imal dynamics to deal with. The pneumatic skin, with its faster response time (on the order of
milliseconds), required the addition of a proprtional-integral controller loop to prevent system
dynamics, such as overshoot, from interfering with the logic controller.
Fig. S7. Control algorithms. The control algorithms for (A) the pneumatic-elastomer robotic skin
and (B) the SMA-fabric robotic skin. These are the algorithms used to achieve closed-loop control.

Text S7 Locomotion g aits using r obotic s kins


Different gaits are possible using a single skin to achieve locomotion. Figure S8 shows several
possibilities, some of which are demonstrated in the Supplemental Videos.
Fig. S8. Possible locomotion gaits achievable using robotic skins with and without soft cylindrical
bodies. Actuation sequencing is shown in the cross-section schematics.
Text S8. Rapid p rototyping with r obotic s kins
Robotic skins may be a useful tool for rapid prototyping of soft robot configurations and
concepts. Because the robotic skins may be easily removed-from, transferred-between, and
reconfigured-on host bodies, they enable quick on-the-fly design exploration and modifications,
which may help designers gain an intuition about a system’s performance prior to investing
significant development costs. As an example, Figure S9A shows a gripper design constructed
from a robotic skin, cardboard, and tape. Two cardboard fingers were cut out and connected
together with additional cardboard. The robotic skin was positioned between the fingers, which
when contracted, pinched the fingers together to grasp an object. This gripping mechanism was
designed and constructed on-the-fly with readily available materials, demonstrating the robotic
skins’ utility for rapid and inexpensive prototyping. This prototype served as the basis for the
gripper in Figure 6B, which was comprised of a single robotic skin attached to a laser-cut piece
of polystyrene sheeting bent into a ’U’ shape. As another example, Figure S9B shows four
robotic skins being made to turn an inanimate plush toy into a walking robot.

Fig. S9. Examples of rapid prototyping with robotics skins. A) Robotic skins can be used in
combination with prototyping materials such as cardboard and tape for rapid, on-the-fly prototyping
of a gripper design. B) Robotic skins can be used to animate existing soft bodies for prototyping
or playful exploration.

You might also like