SCIENCE ROBOTICS | REVIEW

SENSORS Copyright © 2020

The Authors, some
Electronic skins and machine learning for intelligent rights reserved;
exclusive licensee
soft robots American Association
for the Advancement
of Science. No claim
Benjamin Shih1, Dylan Shah2, Jinxing Li3, Thomas G. Thuruthel4, Yong-Lae Park5, Fumiya Iida4, to original U.S.
Zhenan Bao3, Rebecca Kramer-Bottiglio2, Michael T. Tolley1* Government Works

Soft robots have garnered interest for real-world applications because of their intrinsic safety embedded at
the material level. These robots use deformable materials capable of shape and behavioral changes and allow
conformable physical contact for manipulation. Yet, with the introduction of soft and stretchable materials to
robotic systems comes a myriad of challenges for sensor integration, including multimodal sensing capable of
stretching, embedment of high-resolution but large-area sensor arrays, and sensor fusion with an increasing
volume of data. This Review explores the emerging confluence of e-skins and machine learning, with a focus
on how roboticists can combine recent developments from the two fields to build autonomous, deployable soft

Downloaded from https://www.science.org at Northeastern University on November 15, 2024

robots, integrated with capabilities for informative touch and proprioception to stand up to the challenges of
real-world environments.

INTRODUCTION Several reviews have covered related topics on e-skins and per-
Skin plays an essential role for biological systems as a barrier between ception in soft robots, including design and fabrication of e-skins
an organism’s external environment and its internal components. (2, 3), wearable sensors (4), e-skins for interactive robots (5, 6), and
Embedded within its layers is a dense network of mechanical, chemical, future directions in sensing and perception for soft robots (7, 8).
vibrational, temperature, and pain receptors, which work in coordi- This Review examines recent developments in skin-based sensing
nation to enable somatosensation in skin. These capabilities would for soft robots, covering hardware and fabrication techniques and
also be incredibly useful for robots. Electronic skin (e-skin) research machine learning techniques that translate robot perception into
was originally motivated, in part, by a desire to understand biological action planning. To limit the scope of this Review, we consider a
sensing, but the lessons learned can help improve the design of soft robot skin to be skin sensors directly mounted on the surface
robotic systems. To sense, plan, and act, robots require a variety of [e.g., (9)] or embedded in a thin layer beneath the surface of the body
sensors embedded throughout their bodies so that they can obtain of a soft robot [e.g., (10)]. To highlight the opportunities at the
information about their environment. intersection of e-skin and soft robotics research, we cover a variety of
The field of soft robotics (1) studies the use of flexible and interdisciplinary topics including fabrication, learning, and control.
compliant materials as components for building robots, instead
of traditionally rigid components such as metals. Soft robots often
draw inspiration from nature, which has evolved organisms that can INTERDISCIPLINARY TOOLS
operate in unstructured environments. In contrast, current robotic Design and fabrication of integrated e-skins
systems are usually confined to structured laboratories or warehouse Compared with rigid robots, the high mechanical compliance of soft
environments. In addition, natural environments typically contain robots enables safer and more efficient human-robot interaction
several objects of varying material properties that further complicate (HRI) because they can seamlessly interact with the human body
tasks such as object interaction and locomotion. (11). Further advancement of soft robots requires high-performance
The overlap between e-skins, soft robotics, and machine learning electronics and sensors that can stretch continuously with their
is continually growing, and recent advances are summarized in bodies. Recent research in artificial skin has mainly focused on making
Fig. 1. Soft actuation has improved tremendously in capabilities individual sensor devices with better performance, such as sensitivity,
(Fig. 1 bottom), and soft sensors and e-skins exhibit a wide range stretchability, and reliability over many use cycles (Fig. 2). To realize
of complexities (Fig. 1 left). fully biomimetic skin for soft robotics, artificial skin should contain
Several recent advances have combined principles from each field, sensor arrays that are stretchable, cover large areas with a high
often physically manifesting in the form of sensorized fingers and spatiotemporal resolution, and have multiple functions that mimic
grippers (Fig. 1 top). Future breakthroughs in the field may come diverse receptors of the human skin (Fig. 2A). These features should
from further integration of sensors and actuators as roboticists move enable robots to use data-driven methods to extract rich information
toward designing systems that rival the abilities of biological about their environment.
organisms. Increasing sensor density and quantity normally requires a larger
number of interconnecting wires. To reduce this burden, sensor
1
Department of Mechanical and Aerospace Engineering, University of California,
arrays are normally designed in matrix form. For example, a recently
San Diego, CA, USA. 2Department of Mechanical Engineering and Materials Science, developed tactile glove comprising 548 force sensors was constructed
Yale University, CT, USA. 3Departments of Chemical Engineering and Material using readily available materials and simple fabrication tools (Fig. 2B)
Science and Engineering, Stanford University, CA, USA. 4Department of Engineering, (12). This sensor array recorded a large-scale dataset of tactile maps
University of Cambridge, UK. 5Department of Mechanical and Aerospace Engineering,
Seoul National University, South Korea. (about 135,000 frames), which was used to identify objects using
*Corresponding author. Email: [email protected] convolutional neural networks. This work highlights the ability of

Shih et al., Sci. Robot. 5, eaaz9239 (2020) 22 April 2020 1 of 11

 SCIENCE ROBOTICS | REVIEW

Used in machine learning

Natural organisms
Integration of e-skins,
Standalone sensors machine learning
and soft robotics

Actuation and sensing

1 cm Integration of e-skins with soft robotics
CREDIT: ADAPTED BY A. KITTERMAN/SCIENCE ROBOTICS FROM MATERIAL PUBLISHED BY IEEE (10, 26, 27, 32, 37, 40–42, 45, 109), SPRINGER/NATURE (19, 20, 35), WILEY (28), THE NATIONAL ACADEMY OF SCIENCE (31), AND

5 cm
Sophistication of sensing

Downloaded from https://www.science.org at Northeastern University on November 15, 2024

0.5 cm

1 cm

Actuation-focused soft robots

Sophistication of actuation
AAAS/SCIENCE (21, 29, 33, 34, 39, 43, 44, 47), LORI K. SANDERS/HARVARD UNIVERSITY (35), A. ECKERT/TUM (109)

Fig. 1. Trends in the intersections between e-skins, soft robotics, and machine learning. (Left) A range of e-skins and soft sensors that increase in complexity, from
bottom to top, by metrics including density, resolution, and fabrication (19, 21, 26–30). (Middle) Soft robots and e-skins that merge actuation and sensing (10, 37–45, 47),
from left to right and top to bottom. (Bottom) Soft robots that focused primarily on actuation (31–35), from left to right. Red boxes indicate work that has leveraged
machine learning in the processing of their sensor information.

large-scale datasets collected by a high-density sensor array to enable through detection of deformation and vibration, mimicking biological
not only a sense of touch but also the intelligent extraction of inform­ skin. Recent advancements in organic electronics by Wang et al. (18)
ation from touch. Increasing the sensor density simply by scaling down led to the creation of an intrinsically stretchable transistor array with
a passive matrix architecture will reduce the amplitude of analog sig- 347 transistors per square centimeter. Their proof-of-concept demon-
nals while increasing cross-talk between interconnects. If multiple stration illustrated that such a conformable active matrix could ac-
sensors are sampled simultaneously, each line will produce electro- curately map the force applied on each sensor. These capabilities
magnetic noise, which will corrupt the signals being carried on neigh- indicate that stretchable active matrices containing soft sensors and
boring conductive traces. Furthermore, the large number of addressing transistors are a promising step toward soft robotic skin with high
lines will be difficult to manage as the number of sensors increases resolution and high data fidelity.
substantially. These problems can be addressed with an active matrix Making multilayered sensor arrays in a three-dimensional (3D)
that pairs each sensor with a transistor to provide local signal amplifi- lattice can further increase the sensor areal density and allow greater
cation and allows sensors to take turns transmitting information (13–16). integration of different sensor modalities. Just as receptors in bio-
Active matrices with multiplexed signal transduction typically con- logical skin are embedded at various depths, engineers can embed
sume less power than passive matrices because they require fewer sensors that are sensitive to different stimuli in different spatial
sampling lines and do not need external circuitry (17). However, locations. For example, pressure, shear, and strain sensors can be
stretchable e-skins could allow better coverage of curved robot distributed in different layers of the e-skin to achieve optimized
surfaces while allowing sensing of complex texture information sensitivity. Huang et al. (19) demonstrated that stretchable electronics

Shih et al., Sci. Robot. 5, eaaz9239 (2020) 22 April 2020 2 of 11

 SCIENCE ROBOTICS | REVIEW

A Active matrix C Multimodal sensing array D Three-dimensional integration

B Passive matrix E Skin

CREDIT: ADAPTED BY A. KITTERMAN/SCIENCE ROBOTICS FROM MATERIAL PUBLISHED BY SPRINGER/NATURE (126, 127) AND SCIENCE ROBOTICS (24) AND UNPUBLISHED MATERIAL FROM THE AUTHORS OF (12)

Touch
Heat, pain, receptors
and cold

Downloaded from https://www.science.org at Northeastern University on November 15, 2024

receptor
Pressure
receptor

Merkel disk Meissner corpuscles Free nerve ending

(touch) (vibration) (temperature)
Fig. 2. Sensor arrays enable e-skins to extract information about their environment. (A) Human skin with various receptors used to sense stimuli. (B) A scalable
tactile glove containing a passive matrix of 548 force sensors for the collection of large datasets (12). (C) 2D sensor array used to generate a profile of pressure intensity
from experimental mapping of the pixel signals using an active matrix (126). The icons at the bottom represent biological analogies: Merkel disks, Meissner corpuscle,
and free nerves. (D) A 3D array of electronic sensors assembled from 2D electronics (127). (E) Multimodal sensor array that can capture both pressure and temperature
information (24).

integrated in 3D can be built with a layer-by-layer method using generation of e-skins should integrate multimodal sensor arrays to
transfer printing of predesigned stretchable circuits on elastomers capture richer sensory information than their predecessors. Recently,
with vertical interconnects. This stretchable human-machine inter- Lee et al. (24) reported a neuromimetic architecture that enabled
face had a four-layer design that offered multimodal sensing and simultaneous transmission of both tactile and thermotactile inform­
had integrated circuits for wireless data transfer. Using strain engi- ation (Fig. 2E). The pressure- and temperature-sensitive transducers
neering methods, 2D structures can also be assembled into 3D elec- can both be communicated through the pulse signatures by a single
tronic systems with sensing capabilities. Semiconductor materials can electrical conductor. As a biomimetic signaling method, this approach
play critical roles in this context, through demonstrations of com- is promising for reducing wiring and computational requirements
plex, mechanically assembled 3D systems for light-imaging ca- when a robot is covered with thousands of sensors. Multimodal sens-
pabilities that can encompass measurements of the direction, ing could also be achieved through integration of multiple stretchable
intensity, and angular divergence properties of incident light. optical fibers, which has been shown to be effective at localizing and
3D printing has also been used to directly print sensors in soft estimating force in soft actuators (25).
robots to improve both exteroceptive and interoceptive capabilities Overall, many innovations are required for realizing high-density
(20). This work highlights how a 3D integration framework enables a and multifunctional sensor arrays for soft robots. A close collaboration
higher integration density on stretchable substrates than single-layer between roboticists and materials scientists is needed to develop
approaches and allows new functionalities that would be difficult to high-performance stretchable conductors for electrodes and inter-
implement with conventional layer-by-layer designs. connections and stretchable semiconductors for building active
Processing complex tactile information from a sensor array matrices and signal amplifiers. Different sensing modalities and
requires efficient signaling and sampling methods. In human skin, integration architectures should also be explored. Lastly, hardware
stimulation of the receptors is converted into a series of voltage and algorithms for data processing should be considered during the
pulses sent to the nerves. This inspired researchers to develop arti- design of sensory systems, and their performance should be evaluated
ficial receptors and afferent nerves to convert tactile information to on a holistic range of practical robotic tasks.
digital pulses at the site of sensation (21, 22). The signal could
potentially be perceived by a user’s nerves and brain, thus directly Skin-based sensing for soft robots
linking the human brain with soft robotic prosthetics. For example, As sensors are increasingly integrated into soft robots, we can imagine
Kim et al. (23) recently developed a flexible artificial afferent nerve a conceptual plane that categorizes research based on the sophistication
that can effectively collect pressure information from arrays of pressure of actuation and sensing independently (Fig. 1). Stand-alone sensors
sensors and convert them to action potentials to activate muscles. lie on the y axis; some consist of simpler strain sensors [the bottom
Biological skin contains receptor networks that can detect various three images in Fig. 1 left (26–28)], whereas others have more sophis-
stimuli, such as vibration, humidity, and temperature. Several studies ticated sensing schemes, including distributed or multimodal sensing
on e-skin sensor arrays focused on the classification of a single type [the top four images in Fig. 1 left (19, 21, 29, 30)]. The x axis, repre-
of information, such as force, shape, or direction of motion. The next senting actuation-focused soft robots, shows examples of increasingly

Shih et al., Sci. Robot. 5, eaaz9239 (2020) 22 April 2020 3 of 11

 SCIENCE ROBOTICS | REVIEW

complex soft systems that can walk (31, 32), grow (33), swim (34), and (57) used liquid metal sensors to measure human gaits using a
operate autonomously on chemical fuel (35) (Fig. 1 bottom). Last, many sensor fabrication process first presented by Park et al. (27). The
recent works have begun exploring the intersection of the actuation resistance of these sensors increases as the embedded microchannels
and sensing (Fig. 1 middle). Several of them embedded strain sensors for inside the elastomer matrix are stretched because of the increased
state estimation or tactile sensing in a finger-like structure (10, 36–44), length and decreased area of their bulk liquid-metal channels. Others
whereas the others mounted their skins externally (45–47). As both incorporated their sensors with robots: Boutry et al. (58) paired a
areas progress, we envision further integration of increasingly sophis- shear force sensor with a robot arm to allow robotic hand control.
ticated actuation and sensing, extending into the top-right quadrant Booth et al. (47) demonstrated reconfigurable, actuatable e-skins that
of the conceptual plane. could control the motion of deformable inanimate objects from
Access to higher-resolution data about touch will increase the their surface. Zhao et al. (43) embedded optical sensors within soft
ability of soft robots to perceive the complex deformations that they pneumatic fingers, which they then integrated with a Baxter robot.
experience during tasks, including locomotion and manipulation. As skin-based sensing capabilities continue to improve, the goal is
Today’s discrete sensors, which are built with high sensitivity and to develop capabilities that match or outperform biological systems
selectivity, can be tailored to sense deformation modes in a localized (top right corner of Fig. 1).
region or known environment with high confidence (46, 48). However,
this sensing paradigm is insufficient in dynamic or unknown environ- Machine learning for soft e-skins

Downloaded from https://www.science.org at Northeastern University on November 15, 2024

ments where robots will experience substantial deformation because As e-skins increase in resolution, their signals could be processed to
robots do not yet have the level of sophistication of human skin detect higher-order deformation modes and higher-level notions about
receptors or the human brain to collect a broad range of information. the environment, such as material type. However, obtaining this
In addition, many robots are unable to process the volume of inform­ information requires algorithms that can extract useful information
ation to accurately determine the environment or the object being from large quantities of data. To handle the vast amount of data that
sensed. The transition from discrete to continuous sensing and the e-skins can provide, machine learning is emerging as a versatile tool
shift from structured to unconstrained environments both require for making sense of large quantities of data (Fig. 3). For example,
e-skins that can rapidly collect and process large amounts of inform­ Piacenza et al. (59) obtained high-resolution data from a robotic
ation. The added complexity from both transitions compounds the fingertip and used ridge regression to process this data to estimate
processing required to interpret the signals. the locations of indentations. Similarly, Larson et al. (60) used
Several designs of skin-like sensors have been used in soft robotics. convolutional neural networks to learn deformations on a sensor
Many of these sensors contain conductive and stretchable materials array that can interpret human touch in soft interfaces. At the level
to produce resistive or capacitive strain sensors (10, 26, 49, 50). of abstraction of the entire robotic system, Van Meerbeek et al. (53)
Other groups have used optical devices such as cameras and optical tested various learning algorithms to estimate the twist and bend
fibers to sense deformations within an actuator (51–53). Several of angles in sensorized foam, finding that k-nearest neighbors (kNN)
these existing sensors are well suited for measuring characteristics outperformed other common algorithms, including support vector
such as strain, pressure, and bending but do not enable the high machines (SVM) and multilayer perceptrons. In addition, researchers
sensor densities or resolutions that have been demonstrated in e-skins. have also focused on recurrent neural networks, which have been
Soft robots would benefit from integration with e-skins, such as the shown to be advantageous for learning patterns in time series data
skin-like sensor arrays that have been deployed in medical applica- (36, 39, 61, 62).
tions or directly on skin (22, 29, 54–56). Because of the complexity of the mapping between raw sensory
Currently, soft skin-like sensors have been deployed in several information and relevant functional abstractions, information theory
ways. Some groups used their sensors as wearables; Mengüc et al. and machine learning will play a large role in bringing tactile sensing

Inputs Processing Outputs Examples

Exploration strategy Direct policy learning Long short

Strain Level of abstraction ML equivalent term memory
Policy
Action planning + +
High level Reinforcement State Reward +
Shear learning +
Environment
Temperature Mid-level Supervised
learning Regression Convolutional Autoencoders
CREDIT: A. KITTERMAN/SCIENCE ROBOTICS

neural networks
Pressure Low-level Unsupervised
learning Clustering

Chemical Dimensionality reduction

Fig. 3. Machine learning techniques for processing raw sensory information, different levels of abstraction to aid in robot perception, and action planning. The
level of abstraction depends on the task, and the most effective type of learning architecture depends on the quality and structure of the sensor signals. The level of abstraction
depends on the task, and the most effective type of learning architecture depends on the quality and structure of the sensor signals. Higher-level processes can include
parallel execution of lower-level processes. End-to-end architectures [e.g., (61)] without mid-level and low-level pipelines would likely be faster and more effective but are
computationally expensive to develop.

Shih et al., Sci. Robot. 5, eaaz9239 (2020) 22 April 2020 4 of 11

SCIENCE ROBOTICS | REVIEW

to human-like performance levels. In particular, the subfield of rein­ The primary drawback of this type of approach is that when other
forcement learning (RL) will be important for developing closed-loop unplanned deformation modes are introduced, such as buckling or
control for tactile feedback. Suitable algorithms and architectures for a change of material properties through damage or natural material
analogous tasks in soft robotics can be developed by learning from biolog- aging, the models accumulate error. In addition, it is unclear how to
ical processes. For example, in computer vision and machine learning, the generalize these advances to reconfigurable soft robots (74) or
hierarchical nature of visual processing (corresponding to compositional robots that have more complex morphologies. For example, recent
functions) (63) has recently enabled deep neural networks to achieve simulations suggest that there is a wide range of soft robot morphol-
human-like performance across a variety of visual processing tasks (64). ogies that could produce useful locomotion, including quadrupedal
Processing signals from arrays of tactile sensors may benefit from shapes and various oddly shaped exteriors (75). All these classes of
similar techniques, as sets of sensor readings have information encoded robots would benefit from sensing mechanisms with fewer assump-
in spatial relationships that can be naturally represented using matrices. tions about the robot’s mechanical properties.
Tactile exploration can benefit from recent developments in The ideal shape-sensing system could stretch with the robot’s sur-
learning-based simultaneous localization and mapping algorithms. face without affecting its kinematics or dynamics, sense shape with-
Notably, Mirowski et al. (65) used an asynchronous advantage out external components, and be thin. E-skins designed for wearable
actor-critic algorithm for navigating in a complex environment and applications should accommodate the strains of about 55 to 75%
additionally solved auxiliary prediction tasks that made the RL prob- experienced by biological skin (22), and a similar range should be

Downloaded from https://www.science.org at Northeastern University on November 15, 2024

lem faster and more data efficient. Chen et al. (66) showed a direct suitable for most soft robotic applications, although different robots
policy learning algorithm with spatial memory and bootstrapped with experience different surface strains. Although a perfect solution for
human-mediated imitation learning without explicit task rewards. shape sensing of soft robots does not currently exist, recent advances
In the absence of continuous reward functions, actor-critic algorithms in the field of flexible shape-sensing e-skins (Fig. 4) have the poten-
are preferred because they require fewer samples. tial to greatly improve the capabilities of soft robots.
Similarly, tactile manipulation tasks can use insights from learning-­ In contrast to that of soft skins, most work on shape-sensing
based manipulation controllers. A general trend observed in such e-skins treats the skin as an inextensible sheet of rigid elements joined
works is the success of model-based RL (67) or learning by demon- by known axes of rotation (Fig. 4, A and B). The primary challenge
stration (68), approaches that leverage techniques from control theory is thus estimating the relative orientation between sections with known
or human knowledge, respectively. There has been a successful solu- geometries to determine the spatial locations of discrete points within
tion for the direct learning of control policies for dexterous manip- the sheet. In one early study, Hoshi and Shinoda (76) arranged 24
ulation, but it relied on the availability of an accurate simulation printed circuit board (PCB) “nodes” into a mesh and estimated
environment (69). Until robot simulators can model soft-body internode rotations using accelerometers and magnetometers (Fig. 4A).
dynamics that reliably transfer to real robot hardware, such ap- Building upon this work, Mittendorfer and Cheng (48) developed
proaches are difficult to apply to soft robots and deformable objects. rigid sensorized hexagonal PCBs that could be integrated into semi-
For specific simple tasks, it might be easier to find a direct policy flexible sheets and wrapped around robots (Fig. 4B). The nodes
than to fit a general-purpose model of the system dynamics (70). contained accelerometers similar to the work by Hoshi and Shinoda
(76) and had similar assumptions (PCBs are free to rotate but cannot
be stretched), but rotations between neighboring PCBs were calcu-
APPLICATIONS OF E-SKINS lated by obtaining at least two orientations of the skin-per-skin
Shape sensing shape and solving a constrained Procrustes problem for aligning
Whereas environmental sensing helps a robot understand its sur- matrices of data points in real time. Hermanis et al. (77) then used
roundings, having a self-model of the robot’s body is important for a grid-like arrangement of accelerometers and gravitometers on a
planning trajectories and actions within that environment. For flexible fabric sheet. The sheets were demonstrated in a dynamic state
robots primarily composed of rigid components, the geometry of estimation task where a user wore a shirt equipped with the shape-­
each segment remains the same throughout the robot’s lifetime, and sensing sheets while bending and crouching.
relative rotations or translations of links provide enough additional In contrast to the discrete sampling methods mentioned above,
information to fully specify the overall changes in shape. However, other approaches leveraged techniques from machine learning and
for soft robots, individual segments can continuously change their statistics to process various sensing signals and extract a continuous
shapes, via both intentional and unintentional deformation modes, estimate of the shape of the skins (Fig. 4, C and D). This kind of
which complicates modeling and sensing schemes. Complementing data-driven technique will be increasingly useful as the sensory
recent work on soft sensing (7), the direct sensing of surface de- spatial density increases, as discussed throughout this Review. For
formations would enhance the functionality of soft robots. instance, Rendl et al. (78) used regularized least squares to process
One approach to sensing the shape of soft robots involves pair- data from 16 piezoelectric bend sensors on a plastic sheet [polyethylene
ing a model with a relatively low number of sensors, typically on the terephthalate (PET)] to approximate the shape of the sheet as a com-
same order of magnitude as the number of controllable degrees of bination of several shape primitives. This created a flexible system that
freedom in the system. A great deal of progress has been made in could sense the bent state of the sheet with a roughly centimeter-level
modeling manipulators that can be parameterized by a curve in accuracy over an approximately A4-sized sheet. Another study
3D space (71). These models have even been coupled with sensing used relatively inextensible optical fiber Bragg gratings arranged in
mechanisms to enable closed-loop control of continuum manipulators a circle on the top and bottom of a silicone e-skin (Fig. 4C) (79). The
(72). Some approaches embedded sensors into other soft robotic relation between the strains on the fiber and the shape of the sheet
components, such as bending actuators, to achieve closed-loop was extracted from training data using a feed-forward artificial
control in a low-dimensional task space (73). neural network containing one hidden layer for computation between

Shih et al., Sci. Robot. 5, eaaz9239 (2020) 22 April 2020 5 of 11

 SCIENCE ROBOTICS | REVIEW

A Flexible at joints B
three broad categories depending on the flow of information or
energy and the primary system of concern (Fig. 5).
Manipulation
Robotic manipulation involves altering the state of an external
object to a desired set point using internal actuators. The role of
tactile sensors is mainly to obtain state information of the external
object. As energy flows to the environment, stability of the object is
of high concern. Grasp force optimization and stabilization is one
of the most basic manipulation tasks involving tactile sensors (82).
Early works were built on the estimation of normal and tangential
forces on the hand to detect slip and react accordingly (83). Recent
C Continuously flexible
works used learning-based methods for slip onset prediction with
0.17 adjustment and grasp failure detection with adjustment because of
0.16
0.16 the ability of these methods to handle complex multimodal sensory
0.15
0.15 information (84) and their generalizability (85).
Z (m)

0.14 Other manipulation studies used low-dimensional sensor space

Downloaded from https://www.science.org at Northeastern University on November 15, 2024

0.14
0.13
0.13 representations to improve performance in certain situations.
–0.02
–0.04 0.12
0.14 0.12
Van Hoof et al. (86) used autoencoders to generate a low-dimensional
X (m) representation of their complex and continuous tactile data. Con-
trol policies learned using this latent space representation required
D Ground truth: [48 , –70 ] Prediction: [48.6 , –69.3 ] fewer rollouts and were more robust to noise. Another study was on
calibration and self-modeling of a fully sensorized body for whole-
body manipulation (87). Recent work has shown that this process
can be fully automated using control signal information and other
sensor modalities, including inertial measurement units (88).
Perhaps the most complex manipulation task is in-hand manip-
ulation, which imposes strict requirements on the body, brain, and
CREDIT: ADAPTED BY A. KITTERMAN/SCIENCE ROBOTICS FROM MATERIAL PUBLISHED BY IEEE (48, 76, 79), ICS/TUM (48), AND SCIENCE ROBOTICS (53)

Fig. 4. E-skins that can sense their shape in 3D. Recent advances in shape-sensing sensors (89). Current progress in in-hand manipulation using
e-skins use several sensing modalities. (A and B) Accelerometers and/or magnetom- tactile sensors is primarily limited to rolling circular objects (90). On
eters on rigid PCBs can rotate relative to each other and reconstruct their shape the other hand, notable developments toward in-hand manipulation
at discrete points. (A) “3D capture sheet” (76). (B) Hexagonal PCBs with integrated have been achieved with external visual tracking systems (69). How-
accelerometers (48). (C and D) Continuously flexible devices can sense deformation ever, control policies trained using vision alone are scene dependent
throughout their surface and estimate their resulting shape. Data-driven methods
and require large quantities of training data, motivating further
were then used in these examples to estimate the continuous shape of the e-skin.
research into using tactile sensing during in-hand manipulation.
(C) Fiber Bragg gratings in silicone (79). (D) Plastic optical fibers in silicone foam (53).
Exploration
Tactile exploration is the process of voluntary motion of the body
the input and output layers. In a similar spirit, an array of optical based on the somatosensory feedback for identifying environmental
fibers were twisted through an elastomeric foam, and their outputs properties (91). The environmental property of interest could be low-­
were sent to several machine learning algorithms (including kNN, SVM, level features, such as surface texture (92, 93) or temperature (94), or
neural networks, and decision trees) to predict the mode of deformation midlevel tasks, such as object classification (95, 96). However, to be fully
and angle of deformation of their structure (Fig. 4D) (53). These autonomous, the higher-level process of selecting the best actions for
approaches all dealt well with a limited set of deformations and, in obtaining better sensory information, also known as active exploration,
principle, should work for a wider range of deformations when paired must be considered. This is not trivial because the concept of an ob-
with a more expressive (deeper) network. However, none of these ex- jective function and a reward function becomes difficult to define.
isting works can mechanically accommodate large in-plane strains, It is currently conjectured that human exploration is driven by a
primarily because of the inextensibility of the optical fibers used. combination of extrinsic and intrinsic reward variables (97). Extrinsic
rewards are task specific, such as classification of objects, whereas
Toward feedback control of soft robots intrinsic rewards are task independent and hence more general, such as
The intrinsic material compliance of soft robots can protect both curiosity-driven exploration. Experiments suggest that humans primarily
the robot and the environment from damage when interacting with use six types of exploratory movements when exploring objects to
unstructured environments. This property makes soft robots ap- determine their properties (98). Hence, there have been studies on
pealing in contexts such as HRI and robotic manipulation, where acquiring these specialized closed-loop policies based on intrinsic
safety around fragile objects can be important (11, 80). E-skins have rewards such as curiosity (99) or extrinsic rewards such as texture
great potential to enable soft robots to interact intelligently with discrimination ability (100). To achieve efficient exploration with soft
their environment. robots, a combination of tactile and proprioceptive feedback will
In addition, tactile information obtained through skin is vital for likely be useful for effectively implementing such reward functions.
a variety of general robotic control tasks. The type of sensor modality A first step toward an autonomous tactile exploration control
to be used, the processing algorithm, and the response from the body architecture, referred to as tactile servoing by the authors, was
all depend on the task at hand (81). These tasks can be divided into proposed by Li et al. (101). By framing the control objective as

Shih et al., Sci. Robot. 5, eaaz9239 (2020) 22 April 2020 6 of 11

 SCIENCE ROBOTICS | REVIEW

A Manipulation B Exploration C Reaction adaptation (106, 107). Otherwise, closed-loop reaction using tactile
Information processing Information processing Information processing sensing is similar to the closed-loop manipulation problem and is
often implemented in parallel with manipulation tasks as in the case
Body Sensors Body Sensors Body Sensors
of slip detection (108).
Environment Environment Environment The main challenges in whole-body sensing are the organization
and calibration of many spatially distributed multimodal tactile sensing
Flow of energy Flow of information
elements (109). Spatial calibration can be manually performed or
automated using robot kinematics and action inference techniques
(110). Data-driven methods are also promising for end-to-end models
without an explicit kinematic/dynamic calibration (111). The most
recent and comprehensive whole-body tactile sensing research was
able to self-organize and self-calibrate 1260 multimodal sensing units
and implement a hierarchical task manager composed of the fusion
of a balance controller, a self-collision avoidance system, and a skin
compliance controller (112).

Downloaded from https://www.science.org at Northeastern University on November 15, 2024

OPPORTUNITIES AND OUTLOOK
The fields of e-skins and soft robotics have both experienced rapid
progress in recent years. However, incorporating advances from both
CREDIT: ADAPTED BY A. KITTERMAN/SCIENCE ROBOTICS FROM MATERIALS PUBLISHED BY JOHAN63/ISTOCK (ELEPHANT), MIKE VEITCH/ ALAMYSTOCK PHOTO (OCTOPUS), BENJAMIN

Fig. 5. Closed-loop tasks where tactile sensing is essential. These tasks primarily
differ depending on the system that determines the objective (denoted by the fields to produce intelligent, autonomous soft robots is a challenging
TORODE/GETTY IMAGES (CAT), OSADA/SÉGUIN/DRAASM/STANFORD UNIVERSITY (DIVER), SAGE [ROBOT HAND (69)], IEEE [DIVER (128)], AND SCIENCE ROBOTICS (DRONE) (129)

shaded boxes). The middle row consists of biological demonstrations of the tasks. task that will require progress in several key areas (Fig. 6). Here, we
The bottom row contains examples (69, 128, 129) of these capabilities in current outline major open questions in this area and identify areas of
rigid robots, which we expect to further improve in parallel with the integration of research that could provide solutions.
e-skins, soft robotics, and machine learning. Note that the presented division is not
strict, and real-world tasks often involve a combination of all three elementary Design and fabrication
tasks. Yellow arrows indicate energy flow; blue arrows indicate information flow.
The primary future challenges of developing sensor arrays for soft
(A) Manipulation involves altering the state of an external object to a desired set
point using internal actuators. (B) Exploration involves motion of the body to
robots will be to design stretchable sensory arrays with wide band-
account for uncertainties in the environment based on somatosensory feedback. width and high dynamic range, resolution, and sensitivity. In addition,
(C) Reaction involves estimating and responding to environmental cues such that multimodal sensing would increase the robots’ knowledge of their
the body remains in a desired state. environment, leading to richer HRI (Fig. 6, A and B). Sensing of
pressure, shear, and vibration and even detecting the presence of
chemical and biological markers in the environment would be useful
the problem of following a trajectory in the sensor feature space, for a wide range of applications, including manipulation, disaster
various autonomous sensory exploration strategies emerged. The response, and manufacturing. Recent efforts on integrating bacteria
emergent exploration strategies included maintaining contact with cells into soft robots have made it possible to directly detect and
an object, edge tracking, and shape exploration of an unknown display chemical information on soft robots (44). Other major de-
object. Exploration has also been framed as a force and pose control sign challenges include choosing how many sensors to integrate
problem on an unknown object using tactile sensors for feedback into a skin and deciding how to place them intelligently. Resources
(102). Additional tactile information obtained during the process are limited and require careful allocation.
was then used to estimate the compliance of the object. Recent works
integrated active exploration with object discrimination (103). How- Machine learning and information processing
ever, the midlevel processes were independent from the high-level Advancing the intelligence of soft robots will also require computa-
exploration strategy, and the proposed algorithm was therefore tional models that can extract useful information from sensor arrays.
relatively inefficient and slow. The next challenge in this area is However, the details of how to develop and implement such algo-
to develop exploration strategies that run simultaneously and are rithms are unclear. For example, deciding which algorithms can most
regulated by the tactile feature extraction process. Such an algorithm efficiently accomplish tasks in classification, regression, and fault
would allow robots equipped with e-skins to efficiently process their detection; whether neural networks should be used; which architec-
sensory information to make informed decisions on how to move tures are easiest to train; and whether there are trade-offs between
within the world to gather information and achieve at least locally efficiency and reliability are all open questions that need to be
optimal exploration strategies. addressed. Answering these questions will necessitate collaboration
Reaction among computer and data scientists, materials engineers, and
Whole-body tactile skins are required for reacting to active envi- neuroscientists. The result will be robots that are more aware of
ronmental forces applied by external agents (104). Here, the control themselves, their environment, and their interactions with humans,
objective is to estimate and react to external forces such that the yielding richer and more productive experiences for human end users.
body remains stable while executing a behavior. Often, the safety Affective touch is a crucial form of nonverbal communication
of the external agent, typically a human, becomes a higher priority that humans use daily and is one application that would benefit from
than robot stability (105). Because reaction typically involves HRI, the combination of e-skins, soft robotics, and machine learning. In
additional challenges arise from safety, context prediction, and contrast, most robots currently are unable to understand gestures

Shih et al., Sci. Robot. 5, eaaz9239 (2020) 22 April 2020 7 of 11

 SCIENCE ROBOTICS | REVIEW

A Multimodal sensing B Wearable robots C Closed-loop morphing D Dexterous manipulation

Fig. 6. Potential capabilities and technologies that could be achieved with e-skins and soft robotics. (A) Multimodal sensing would be useful during manipulation
for detecting gripper states, object properties, and events such as contact and slip. (B) E-skins with an integrated human-robot interface could enable seamless assistive
wearable robots and intuitive teleoperation of anthropomorphic robots. (C) When paired with the appropriate actuators, shape sensing would enable closed-loop changes
of shape. (D) Closed-loop control algorithms would enable soft robots equipped with e-skins to succeed when performing complex tasks, including in-arm manipulation.

Downloaded from https://www.science.org at Northeastern University on November 15, 2024

such as a pat on the back because either they do not have the sensors highest spatial resolution (around 2 mm), as evident from the dense
necessary to measure the interaction or they are not able to make mechanoreceptor distribution at the human fingertip (119, 120).
sense of the affective contact. On the other end of the spectrum, tactile reaction likely requires the
lowest spatial resolution, as suggested by the poor spatial resolution
Shape sensing across other parts of the body. Tactile manipulation lies in between,
Despite the recent progress in shape-sensing e-skins, it is unclear how with an expected spatial resolution of 5 mm (121).
to extend these advances to the wide range of soft robots presented The type and the distribution of mechanoreceptors across the body
in literature. Soft robots experience large strains and complex de- also suggest the type of sensor technologies that would be useful for
formations; key challenges include increasing the stretchability of a particular task. Humans use distinct sensors for static and dynamic
shape-sensing skins and improving the resolution of sensors to cues. Low-bandwidth mechanoreceptors (10 to 50 Hz) can be found
detect small curvatures. mainly in the fingertip and would be essential for tactile exploration
Once the field has reliable solutions for soft robot proprioception, (122). Higher-bandwidth mechanoreceptors (50 to 400 Hz), which
it is conceivable that shape feedback would enable controlled shape respond to the vibrations induced during object slippage, are
change in robots. Current soft robots are not able to morph into distributed primarily at the palm of the hand (123). The response
specific configurations, yet even simple shape change has led to and the sensing areas of the mechanoreceptors are strongly depen-
innovative solutions for a wide range of tasks, such as obstacle avoid- dent on the skin morphology. Hence, it is vital to consider the design
ance (9), rolling locomotion (113), underwater locomotion (114), of the body and the motion capabilities for mimicking the dynamic
and camouflage (115). Larger shape changes could result in robots receptors in our body.
that switch between morphologies and corresponding locomotion Other insights can be gained by extending such an analysis to
gaits on demand (Fig. 6C). invertebrate biological organisms, such as octopuses. An octopus
has several receptors, primarily chemoreceptors, located on each
Feedback control sucker. In addition, the octopus has strain receptors associated with
Using sensorized skins to close the loop has the potential to improve its muscles and a relatively large brain for processing its receptor
the ability of soft robots to react to their environment, to locomote, information. Despite these capabilities, it has a poor proprioceptive
to explore, and to manipulate objects using their deformable bodies sense and cannot estimate the overall shape and location of external
(Fig. 6D). The use of soft tactile sensors for closed-loop control is objects that it is handling. There is local proprioceptive feedback in
still in its nascency. The few relevant studies in this area used low-­ each arm for low-level control, but the only feedback to the central
dimensional soft strain sensors for closed-loop kinematic or force nervous system comes through vision (124). Wells (125) conjectured
control (73, 116, 117). This is surprising given the wealth of litera- that in flexible animals, motor control is hierarchical and pro-
ture on soft sensing technologies and considering the intended prioceptive information must be used locally. Contrary to popular
application of these sensors (7). One reason for this discrepancy belief, the performance of an octopus in manipulation tasks is poor.
could be that soft sensors were originally developed for wearable Therefore, it might be necessary to incorporate rigid components in
devices and therefore used only for state estimation. Another reason fully soft robots, if they are to be used for tactile-based closed-loop
CREDIT: A. KITTERMAN/SCIENCE ROBOTICS

could be the demanding performance expectations placed on soft sensors. control tasks.
Although it would be useful to develop drift-free, linear sensors with
high gauge factors, biology suggests that workarounds are possible. Outlook
For example, the human tactile sensing system is hysteretic, nonlin- Researchers have developed many interesting forms of actuation
ear, time varying, and slow. Nature adapted to these drawbacks by that more closely mimic the functionality and capabilities found in
developing hyper-redundant sensing networks and intelligent data nature. The next step for the field is to develop biologically inspired
processing techniques (118). tactile sensing for soft-bodied robots that can safely interact with,
Along the same lines, various sensor design strategies can be and explore, their environments. Current work tends to concentrate
found by observing nature. Tactile exploration likely requires the on the design and fabrication of soft robots and explores how

Shih et al., Sci. Robot. 5, eaaz9239 (2020) 22 April 2020 8 of 11

SCIENCE ROBOTICS | REVIEW

machine learning can enhance soft robot perception. In the short 21. B. C.-K. Tee, A. Chortos, A. Berndt, A. K. Nguyen, A. Tom, A. McGuire, Z. C. Lin, K. Tien,
W.-G. Bae, H. Wang, P. Mei, H.-H. Chou, B. Cui, K. Deisseroth, T. N. Ng, Z. Bao,
term, the field can focus on deployable, high-resolution sensor skins,
A skin-inspired organic digital mechanoreceptor. Science 350, 313–316 (2015).
algorithms for processing the dense sensor information, and reli- 22. A. Chortos, J. Liu, Z. Bao, Pursuing prosthetic electronic skin. Nat. Mater. 15, 937–950 (2016).
able feedback control for soft robots. The longer-term goal is robots 23. Y. Kim, A. Chortos, W. Xu, Y. Liu, J. Y. Oh, D. Son, J. Kang, A. M. Foudeh, C. Zhu, Y. Lee,
that can touch and feel with the sensitivity and perception of natural S. Niu, J. Liu, R. Pfattner, Z. Bao, T.-W. Lee, A bioinspired flexible organic artificial afferent
systems. nerve. Science 360, 998–1003 (2018).
24. W. W. Lee, Y. J. Tan, H. Yao, S. Li, H. H. See, M. Hon, K. A. Ng, B. Xiong, J. S. Ho, C. Tee,
We believe that future societies will include robots tightly inte- A neuro-inspired artificial peripheral nervous system for scalable electronic skins.
grated with humanity. This includes in-home, assistive robots that Sci. Robot. 4, eaax2198 (2019).
can sense and understand gestures such as a pat on the back, collabora- 25. P. A. Xu, A. Mishra, H. Bai, C. Aubin, L. Zullo, R. F. Shepherd, Optical lace for synthetic
tive robots that work alongside humans, and exploratory robots afferent neural networks. Sci. Robot. 4, eaaw6304 (2019).
26. H. A. Sonar, M. C. Yuen, R. Kramer-Bottiglio, J. Paik, An any-resolution pressure
that can navigate the unpredictable real world.
localization scheme using a soft capacitive sensor skin, in 2018 IEEE International
Conference on Soft Robotics (RoboSoft) (IEEE, 2018), pp. 170–175.
27. Y.-L. Park, B.-R. Chen, R. J. Wood, Design and fabrication of soft artificial skin using
REFERENCES AND NOTES embedded microchannels and liquid conductors. IEEE Sensors J. 12, 2711–2718 (2012).
1. D. Rus, M. T. Tolley, Design, fabrication and control of soft robots. Nature 521, 467–475 28. T. Hellebrekers, O. Kroemer, C. Majidi, Soft magnetic skin for continuous deformation
(2015). sensing. Adv. Intell. Syst. 1, 1900025 (2019).

Downloaded from https://www.science.org at Northeastern University on November 15, 2024

2. S. Wagner, S. P. Lacour, J. Jones, I. H. Pai-hui, J. C. Sturm, T. Li, Z. Suo, Electronic skin: 29. D. H. Kim, N. Lu, R. Ma, Y. S. Kim, R. H. Kim, S. Wang, J. Wu, S. M. Won, H. Tao, A. Islam,
Architecture and components. Phys. E. 25, 326–334 (2004). K. J. Yu, T. I. Kim, R. Chowdhury, M. Ying, L. Xu, M. Li, H. J. Chung, H. Keum, M. McCormick,
3. J. Case, M. Yuen, M. Mohammed, R. Kramer, Sensor skins: An overview, in Stretchable P. Liu, Y. W. Zhang, F. G. Omenetto, Y. Huang, T. Coleman, J. A. Rogers, Epidermal
Bioelectronics for Medical Devices and Systems (Springer, 2016), pp. 173–191. electronics. Science 333, 838–843 (2011).
4. M. Amjadi, K.-U. Kyung, I. Park, M. Sitti, Stretchable, skin-mountable, and wearable strain 30. M. Kaltenbrunner, T. Sekitani, J. Reeder, T. Yokota, K. Kuribara, T. Tokuhara, M. Drack,
sensors and their potential applications: A review. Adv. Funct. Mater. 26, 1678–1698 (2016). R. Schwödiauer, I. Graz, S. Bauer-Gogonea, S. Bauer, T. Someya, An ultra-lightweight
design for imperceptible plastic electronics. Nature 499, 458–463 (2013).
5. N. Yogeswaran, W. Dang, W. T. Navaraj, D. Shakthivel, S. Khan, E. O. Polat, S. Gupta,
31. R. F. Shepherd, F. Ilievski, W. Choi, S. A. Morin, A. A. Stokes, A. D. Mazzeo, X. Chen,
H. Heidari, M. Kaboli, L. Lorenzelli, G. Cheng, R. Dahiya, New materials and advances
M. Wang, G. M. Whitesides, Multigait soft robot. Proc. Natl. Acad. Sci. U.S.A. 108,
in making electronic skin for interactive robots. Adv. Robot. 29, no. 21, 1359–1373 (2015).
20400–20403 (2011).
6. R. Dahiya, N. Yogeswaran, F. Liu, L. Manjakkal, E. Burdet, V. Hayward, H. Jörntell,
32. D. Drotman, S. Jadhav, M. Karimi, P. Dezonia, M. T. Tolley, 3D printed soft actuators for a
Large-area soft e-skin: The challenges beyond sensor designs. Proc. IEEE 107, 2016–2033
legged robot capable of navigating unstructured terrain, in 2017 IEEE International
(2019).
Conference on Robotics and Automation (ICRA) (IEEE, 2017), pp. 5532–5538.
7. H. Wang, M. Totaro, L. Beccai, Toward perceptive soft robots: Progress and challenges.
33. E. W. Hawkes, L. H. Blumenschein, J. D. Greer, A. M. Okamura, A soft robot that navigates
Adv. Sci. 5, 1800541 (2018).
its environment through growth. Sci. Robot. 2, eaan3028 (2017).
8. K. Chin, T. Hellebrekers, C. Majidi, Machine learning for soft robotic sensing and control.
34. R. K. Katzschmann, J. DelPreto, R. MacCurdy, D. Rus, Exploration of underwater life
Adv. Intell. Syst. 2020, 1900171 (2020).
with an acoustically controlled soft robotic fish. Sci. Robot. 3, eaar3449 (2018).
9. D. S. Shah, M. C. Yuen, L. G. Tilton, E. J. Yang, R. Kramer-Bottiglio, Morphing robots using
35. M. Wehner, R. L. Truby, D. J. Fitzgerald, B. Mosadegh, G. M. Whitesides, J. A. Lewis,
robotic skins that sculpt clay. IEEE Robot. Autom. Lett. 4, 2204–2211 (2019).
R. J. Wood, An integrated design and fabrication strategy for entirely soft, autonomous
10. B. Shih, D. Drotman, C. Christianson, Z. Huo, R. White, H. I. Christensen, M. T. Tolley,
robots. Nature 536, 451–455 (2016).
Custom soft robotic gripper sensor skins for haptic object visualization, in 2017 IEEE/RSJ IROS
36. G. Soter, A. Conn, H. Hauser, J. Rossiter, Bodily aware soft robots: Integration of
(IEEE, 2017), pp. 494–501.
proprioceptive and exteroceptive sensors, in 2018 IEEE International Conference on
11. T. Arnold, M. Scheutz, The tactile ethics of soft robotics: Designing wisely for human–robot
Robotics and Automation (ICRA) (IEEE, 2018), pp. 2448–2453.
interaction. Soft Robot. 4, 81–87 (2017).
37. L. Chin, J. Lipton, M. C. Yuen, R. Kramer-Bottiglio, D. Rus, Automated recycling separation
12. S. Sundaram, P. Kellnhofer, Y. Li, J.-Y. Zhu, A. Torralba, W. Matusik, Learning the signatures
enabled by soft robotic material classification, in 2019 2nd IEEE International Conference
of the human grasp using a scalable tactile glove. Nature 569, 698–702 (2019).
on Soft Robotics (RoboSoft) (IEEE, 2019), pp. 102–107.
13. J. Viventi, D.-H. Kim, L. Vigeland, E. S. Frechette, J. A. Blanco, Y.-S. Kim, A. E. Avrin,
38. R. Deimel, O. Brock, A novel type of compliant and underactuated robotic hand
V. R. Tiruvadi, S.-W. Hwang, A. C. Vanleer, D. F. Wulsin, K. Davis, C. E. Gelber, L. Palmer,
for dexterous grasping. Int. J. Robot. Res. 35, 161–185 (2016).
J. Van der Spiegel, J. Wu, J. Xiao, Y. Huang, D. Contreras, J. A. Rogers, B. Litt, Flexible,
39. T. G. Thuruthel, B. Shih, C. Laschi, M. T. Tolley, Soft robot perception using embedded soft
foldable, actively multiplexed, high-density electrode array for mapping brain activity
sensors and recurrent neural networks. Sci. Robot. 4, eaav1488 (2019).
in vivo. Nat. Neurosci. 14, 1599–1605 (2011).
14. D.-H. Kim, J.-H. Ahn, W. M. Choi, H.-S. Kim, T.-H. Kim, J. Song, Y. Y. Huang, Z. Liu, C. Lu, 40. L. Scimeca, P. Maiolino, D. Cardin-Catalan, A. P. del Pobil, A. Morales, and F. Iida,
J. A. Rogers, Stretchable and foldable silicon integrated circuits. Science 320, Non-destructive robotic assessment of mango ripeness via multi-point soft haptics, in
507–511 (2008). 2019 International Conference on Robotics and Automation (ICRA) (IEEE, 2019), pp. 1821–1826.
15. T. Someya, Y. Kato, T. Sekitani, S. Iba, Y. Noguchi, Y. Murase, H. Kawaguchi, T. Sakurai, 41. R. A. Bilodeau, E. L. White, R. K. Kramer, Monolithic fabrication of sensors and actuators in
Conformable, flexible, large-area networks of pressure and thermal sensors with organic a soft robotic gripper, in 2015 IEEE/RSJ International Conference on Intelligent Robots and
transistor active matrixes. Proc. Natl. Acad. Sci. U.S.A. 102, 12321–12325 (2005). Systems (IROS) (IEEE), pp. 2324–2329.
16. T. Sekitani, Y. Noguchi, K. Hata, T. Fukushima, T. Aida, T. Someya, A rubberlike stretchable 42. N. Farrow, N. Correll, A soft pneumatic actuator that can sense grasp and touch, in 2015 IEEE/
active matrix using elastic conductors. Science 321, 1468–1472 (2008). RSJ International Conference on Intelligent Robots and Systems (IROS) (IEEE), pp. 2317–2323.
17. X. Ren, K. Pei, B. Peng, Z. Zhang, Z. Wang, X. Wang, P. K. Chan, A low-operating-power 43. H. Zhao, K. OBrien, S. Li, R. F. Shepherd, Optoelectronically innervated soft prosthetic
and flexible active-matrix organic-transistor temperature-sensor array. Adv. Mater. 28, hand via stretchable optical waveguides. Sci. Robot. 1, eaai7529 (2016).
4832–4838 (2016). 44. K. B. Justus, T. Hellebrekers, D. D. Lewis, A. Wood, C. Ingham, C. Majidi, P. R. LeDuc, C. Tan,
18. S. Wang, J. Xu, W. Wang, G.-J. N. Wang, R. Rastak, F. Molina-Lopez, J. W. Chung, S. Niu, A biosensing soft robot: Autonomous parsing of chemical signals through integrated
V. R. Feig, J. Lopez, T. Lei, S.-K. Kwon, Y. Kim, A. M. Foudeh, A. Ehrlich, A. Gasperini, Y. Yun, organic and inorganic interfaces. Sci. Robot. 4, eaax0765 (2019).
B. Murmann, J. B.-H. Tok, Z. Bao, Skin electronics from scalable fabrication of an intrinsically 45. A. Alspach, J. Kim, K. Yamane, Design of a soft upper body robot for physical human-
stretchable transistor array. Nature 555, 83–88 (2018). robot interaction, in 2015 IEEE-RAS 15th International Conference on Humanoid Robots
19. Z. Huang, Y. Hao, Y. Li, H. Hu, C. Wang, A. Nomoto, T. Pan, Y. Gu, Y. Chen, T. Zhang, W. Li, (Humanoids) (IEEE, 2015), pp. 290–296.
Y. Lei, N. H. Kim, C. Wang, L. Zhang, J. W. Ward, A. Maralani, X. Li, M. F. Durstock, A. Pisano, 46. P. Mittendorfer, E. Dean, G. Cheng, 3D spatial self-organization of a modular artificial skin,
Y. Lin, S. Xu, Three-dimensional integrated stretchable electronics. Nat. Electronics, in 2014 IEEE/RSJ International Conference on Intelligent Robots and Systems, 14 to
473–480 (2018). 18 September 2014 (IEEE, 2014), pp. 3969–3974.
20. B. Shih, C. Christianson, K. Gillespie, S. Lee, J. Mayeda, Z. Huo, M. T. Tolley, Design 47. J. W. Booth, D. Shah, J. C. Case, E. L. White, M. C. Yuen, O. Cyr-Choiniere, R. Kramer-Bottiglio,
considerations for 3D printed, soft, multimaterial resistive sensors for soft robotics. Omniskins: Robotic skins that turn inanimate objects into multi- functional robots.
Front. Robot. AI 6, 30 (2019). Sci. Robot. 3, eaat1853 (2018).

Shih et al., Sci. Robot. 5, eaaz9239 (2020) 22 April 2020 9 of 11

SCIENCE ROBOTICS | REVIEW

48. P. Mittendorfer, G. Cheng, 3D surface reconstruction for robotic body parts with 74. S. Y. Kim, R. Baines, J. Booth, N. Vasios, K. Bertoldi, R. Kramer-Bottiglio, Re-configurable
artificial skins, in 2012 IEEE/RSJ International Conference on Intelligent Robots and soft body trajectories using unidirectionally stretchable composite laminae.
Systems, 7 to 12 October 2012 (IEEE, 2012), pp. 4505–4510. Nat. Commun. 10, 3464 (2019).
49. R. K. Kramer, C. Majidi, R. Sahai, R. J. Wood, Soft curvature sensors for joint angle 75. N. Cheney, R. MacCurdy, J. Clune, H. Lipson, Unshackling evolution: evolving soft robots
proprioception, in Intelligent Robots and Systems (IROS), 2011 IEEE/RSJ International with multiple materials and a powerful generative encoding, in Proceedings of the 15th
Conference on (IEEE, 2011)pp. 1919–1926. Annual Conference on Genetic and Evolutionary Computation (ACM, 2013), pp. 167–174.
50. J. T. Muth, D. M. Vogt, R. L. Truby, Y. Mengüc, D. B. Kolesky, R. J. Wood, J. A. Lewis, 76. T. Hoshi, H. Shinoda, 3D shape measuring sheet utilizing gravitational and geomagnetic
Embedded 3D printing of strain sensors within highly stretchable elastomers. Adv. Mater. fields, in 2008 SICE Annual Conference, 20 to 22 August 2008, pp. 915–920.
26, 6307–6312 (2014). 77. A. Hermanis, R. Cacurs, M. Greitans, Acceleration and magnetic sensor network for shape
51. A. Alspach, J. Kim, K. Yamane, Design and fabrication of a soft robotic hand and arm sensing. IEEE Sensors J. 16, 1271–1280 (2016).
system, in 2018 IEEE International Conference on Soft Robotics (RoboSoft), 24 to 28 April 2018 78. C. Rendl, D. Kim, S. Fanello, P. Parzer, C. Rhemann, J. Taylor, M. Zirkl, G. Scheipl,
(IEEE, 2018), pp. 369–375. T. Rothlnder, M. Haller, S. Izadi, Flexsense: A transparent self-sensing deformable surface,
52. J. Jung, M. Park, D. W. Kim, Y.-L. Park, Optically sensorized elastomer air chamber in Proceedings of the 27th Annual ACM Symposium on User Interface Software and
for proprioceptive sensing of soft pneumatic actuators. IEEE Robot. Autom. Lett. 5, Technology (UIST’14), Honolulu, Hawaii, USA (ACM, 2014), pp. 129–138.
2333–2340 (2020). 79. T. L. T. Lun, K. Wang, J. D. L. Ho, K.-H. Lee, K. Y. Sze, K.-K. Kwok, Real-time surface shape
53. I. Van Meerbeek, C. De Sa, R. Shepherd, Soft optoelectronic sensory foams sensing for soft and flexible structures using fiber Bragg gratings. IEEE Robot. Autom. Lett.
with proprioception. Sci. Robot. 3, eaau2489 (2018). 4, 1454–1461 (2019).
54. G. Schwartz, B. C.-K. Tee, J. Mei, A. L. Appleton, D. H. Kim, H. Wang, Z. Bao, Flexible 80. L. Pinto, D. Gandhi, Y. Han, Y.-L. Park, A. Gupta, The curious robot: Learning visual
polymer transistors with high pressure sensitivity for application in electronic skin representations via physical interactions, in European Conference on Computer Vision

Downloaded from https://www.science.org at Northeastern University on November 15, 2024

and health monitoring. Nat. Commun. 4, 1859 (2013). (Springer, 2016) pp. 3–18.
55. J. Byun, Y. Lee, J. Yoon, B. Lee, E. Oh, S. Chung, T. Lee, K.-J. Cho, J. Kim, Y. Hong, Electronic 81. M. R. Cutkosky, R. D. Howe, W. R. Provancher, Force and tactile sensors, in Springer
skins for soft, compact, reversible assembly of wirelessly activated fully soft robots. Handbook of Robotics, B. Siciliano, O. Khatib, Eds. (Springer, 2008), pp. 455–476.
Sci. Robot. 3, eaas9020 (2018). 82. M. Li, Y. Bekiroglu, D. Kragic, A. Billard, Learning of grasp adaptation through experience
56. H. Jeong, L. Wang, T. Ha, R. Mitbander, X. Yang, Z. Dai, S. Qiao, L. Shen, N. Sun, N. Lu, and tactile sensing, in 2014 IEEE/RSJ International Conference on Intelligent Robots and
Modular and reconfigurable wireless e-tattoos for personalized sensing. Adv. Mate. Technol. Systems (IEEE, 2014), pp. 3339–3346.
4, 1900117 (2019). 83. N. Wettels, V. J. Santos, R. S. Johansson, G. E. Loeb, Biomimetic tactile sensor array.
Adv. Robot. 22, 829–849 (2008).
57. Y. Mengüc, Y.-L. Park, H. Pei, D. Vogt, P. M. Aubin, E. Winchell, L. Fluke, L. Stirling,
R. J. Wood, C. J. Walsh, Wearable soft sensing suit for human gait measurement. 84. F. Veiga, J. Peters, T. Hermans, Grip stabilization of novel objects using slip prediction.
Int. J. Robot. Res. 33, 1748–1764 (2014). IEEE Trans. Haptics 11, 531–542 (2018).
58. C. M. Boutry, M. Negre, M. Jorda, O. Vardoulis, A. Chortos, O. Khatib, Z. Bao, 85. Y. Chebotar, M. Kalakrishnan, A. Yahya, A. Li, S. Schaal, S. Levine, Path integral guided
A hierarchically patterned, bioinspired e-skin able to detect the direction of applied policy search, in 2017 IEEE International Conference on Robotics and Automation (ICRA)
pressure for robotics. Sci. Robot. 3, eaau6914 (2018). (IEEE, 2017), pp. 3381–3388.
59. P. Piacenza, S. Sherman, M. Ciocarlie, Data-driven super-resolution on a tactile dome. 86. H. Van Hoof, N. Chen, M. Karl, P. van der Smagt, J. Peters, Stable reinforcement learning
IEEE Robot. Autom. Lett. 3, 1434–1441 (2018). with autoencoders for tactile and visual data, in 2016 IEEE/RSJ International Conference on
Intelligent Robots and Systems (IROS) (IEEE, 2016), pp. 3928–3934.
60. C. Larson, J. Spjut, R. Knepper, R. Shepherd, A deformable interface for human touch
recognition using stretchable carbon nanotube dielectric elastomer sensors and deep 87. Y. Ohmura, Y. Kuniyoshi, Humanoid robot which can lift a 30kg box by whole body
neural networks. Soft Robot. 6, 611–620 (2019). contact and tactile feedback, in 2007 IEEE/RSJ International Conference on Intelligent
Robots and Systems (IEEE, 2007) pp. 1136–1141.
61. D. Kim, J. Kwon, B. Jeon, Y.-L. Park, Adaptive calibration of soft sensors using optimal
transportation transfer learning for mass production and long-term usage. Adv. Intell. Syst. 88. P. Mittendorfer, E. Yoshida, G. Cheng, Realizing whole-body tactile interactions
2020, 1900178 (2020). with a self-organizing, multi-modal artificial skin on a humanoid robot. Adv. Robot. 29,
51–67 (2015).
62. D. Kim, J. Kwon, S. Han, Y.-L. Park, S. Jo, Deep full-body motion network (dfm-net)
for a soft wearable motion sensing suit. IEEE/ASME Trans. Mechatron. 24, 89. K. Hertkorn, M. A. Roa, C. Borst, Planning in-hand object manipulation with multifingered
18453663 (2018). hands considering task constraints, in 2013 IEEE International Conference on Robotics and
Automation (IEEE, 2013), pp. 617–624.
63. M. Riesenhuber, T. Poggio, Hierarchical models of object recognition in cortex.
Nat. Neurosci. 2, 1019–1025 (1999). 90. H. Van Hoof, T. Hermans, G. Neumann, J. Peters, Learning robot in-hand manipulation
with tactile features, in 2015 IEEE-RAS 15th International Conference on Humanoid Robots
64. Y. Bengio, A. Courville, P. Vincent, Representation learning: A review and new
(Humanoids) (IEEE, 2015), pp. 121–127.
perspectives. IEEE Trans. Pattern Anal. Mach. Intell. 35, 1798–1828 (2013).
91. G. E. Loeb, G. A. Tsianos, J. A. Fishel, N. Wettels, S. Schaal, Understanding haptics by
65. P. Mirowski, R. Pascanu, F. Viola, H. Soyer, A. J. Ballard, A. Banino, M. Denil, R. Goroshin,
evolving mechatronic systems, in Progress in brain research (Elsevier, 2011), vol. 192,
L. Sifre, K. Kavukcuoglu, D. Kumaran, R. Hadsell, Learning to navigate in complex
pp. 129–144.
environments. arXiv:1611.03673 (2016).
92. N. Jamali, C. Sammut, Majority voting: Material classification by tactile sensing using
66. T. Chen, S. Gupta, A. Gupta, Learning exploration policies for navigation.
surface texture. IEEE Trans. Robot. 27, 508–521 (2011).
arXiv:1903.01959 (2019).
93. J. Sinapov, V. Sukhoy, R. Sahai, A. Stoytchev, Vibrotactile recognition and categorization
67. J. Fu, S. Levine, P. Abbeel, One-shot learning of manipulation skills with online dynamics
of surfaces by a humanoid robot. IEEE Trans. Robot. 27, 488–497 (2011).
adaptation and neural network priors, in 2016 IEEE/RSJ International Conference on
Intelligent Robots and Systems (IROS) (IEEE, 2016), pp. 4019–4026. 94. H. Yousef, M. Boukallel, K. Althoefer, Tactile sensing for dexterous in-hand
manipulation in robotics––A review. Sensors Actuators A Phys. 167, 171–187 (2011).
68. A. Rajeswaran, V. Kumar, A. Gupta, G. Vezzani, J. Schulman, E. Todorov, S. Levine, Learning
complex dexterous manipulation with deep reinforcement learning and demonstrations. 95. N. Jamali, C. Sammut, Material classification by tactile sensing using surface textures, in
arXiv:1709.10087 (2017). 2010 IEEE International Conference on Robotics and Automation (IEEE, 2010), pp. 2336–2341.
69. OpenAI, M. Andrychowicz, B. Baker, M. Chociej, R. Jozefowicz, B. M. Grew, J. Pachocki, 96. B. S. Homberg, R. K. Katzschmann, M. R. Dogar, D. Rus, Haptic identification of objects
A. Petron, M. Plappert, G. Powell, A. Ray, J. Schneider, S. Sidor, J. Tobin, P. Welinder, using a modular soft robotic gripper, in Intelligent Robots and Systems (IROS), 2015 IEEE/
L. Weng, W. Zaremba, Learning dexterous in-hand manipulation. arXiv:1808.00177 (2018). RSJ International Conference on (IEEE, 2015), pp. 1698–1705.
70. D. P. Bertsekas, Dynamic Programming and Optimal Control (Athena Scientific, 1995), vol. 1. 97. J. Gottlieb, P.-Y. Oudeyer, M. Lopes, A. Baranes, Information-seeking, curiosity, and
71. J. Burgner-Kahrs, D. C. Rucker, H. Choset, Continuum robots for medical applications: attention: Computational and neural mechanisms. Trends Cogn. Sci. 17, 585–593 (2013).
A survey. IEEE Trans. Robot. 31, 1261–1280 (2015). 98. S. J. Lederman, R. L. Klatzky, Hand movements: A window into haptic object recognition.
72. T. George Thuruthel, Y. Ansari, E. Falotico, C. Laschi, Control strategies for soft robotic Cogn. Psychol. 19, 342–368 (1987).
manipulators: A survey. Soft Robot. 5, 149–163 (2018). 99. L. Pape, C. M. Oddo, M. Controzzi, C. Cipriani, A. Förster, M. C. Carrozza, J. Schmidhuber,
73. J. Morrow, H.-S. Shin, C. Phillips-Grafflin, S. Jang, J. Torrey, R. Larkins, S. Dang, Y.-L. Park, Learning tactile skills through curious exploration. Front. Neurorobot. 6, 6 (2012).
D. Berenson, Improving soft pneumatic actuator fingers through integration of soft 100. J. A. Fishel, G. E. Loeb, Bayesian exploration for intelligent identification of textures.
sensors, position and force control, and rigid fingernails. in 2016 IEEE International Front. Neurorobot. 6, 4 (2012).
Conference on Robotics and Automation (ICRA), 16 to 21 May 2016 (IEEE, 2016), 101. Q. Li, C. Schürmann, R. Haschke, H. Ritter, A control framework for tactile servoing,
pp. 5024–5031. in Robotics: Science and Systems (2013).

Shih et al., Sci. Robot. 5, eaaz9239 (2020) 22 April 2020 10 of 11

SCIENCE ROBOTICS | REVIEW

102. Z. Su, J. A. Fishel, T. Yamamoto, G. E. Loeb, Use of tactile feedback to control exploratory 121. F. Mancini, A. Bauleo, J. Cole, F. Lui, C. A. Porro, P. Haggard, G. D. Iannetti, Whole-body
movements to characterize object compliance. Front. Neurorobot. 6, 7 (2012). mapping of spatial acuity for pain and touch. Ann. Neurol. 75, 917–924 (2014).
103. M. Kaboli, K. Yao, D. Feng, G. Cheng, Tactile-based active object discrimination and target 122. A. B. Vallbo, R. S. Johansson, Properties of cutaneous mechanoreceptors in the human
object search in an unknown workspace. Auton. Robot. 43, 123–152 (2019). hand related to touch sensation. Hum. Neurobiol. 3, 3–14 (1984).
104. H. Iwata, S. Sugano, Whole-body covering tactile interface for human robot 123. R. S. Johansson, J. R. Flanagan, Coding and use of tactile signals from the fingertips
coordination, in Proceedings 2002 IEEE International Conference on Robotics and in object manipulation tasks. Nat. Rev. Neurosci. 10, 345–359 (2009).
Automation (Cat. No. 02CH37292) (IEEE, 2002), vol. 4, pp. 3818–3824. 124. M. J. Wells, Octopus: Physiology and Behaviour of an Advanced Invertebrate (Springer
105. M. Frigola, A. Casals, J. Amat, Human-robot interaction based on a sensitive bumper skin, Science & Business Media, 2013).
in 2006 IEEE/RSJ International Conference on Intelligent Robots and Systems (IEEE, 2006), 125. M. Wells, Tactile discrimination of surface curvature and shape by the octopus. J. Exp. Biol.
pp. 283–287. 41, 433–445 (1964).
106. S. Haddadin, E. Croft, Physical human–robot interaction, in Springer Handbook of robotics 126. K. Takei, T. Takahashi, J. C. Ho, H. Ko, A. G. Gillies, P. W. Leu, R. S. Fearing, A. Javey,
(Springer, 2016), pp. 1835–1874. Nanowire active-matrix circuitry for low-voltage macroscale artificial skin. Nat. Mater. 9,
107. D. Silvera-Tawil, D. Rye, M. Velonaki, Artificial skin and tactile sensing for socially 821–826 (2010).
interactive robots: A review. Robot. Auton. Syst. 63, 230–243 (2015). 127. W. Lee, Y. Liu, Y. Lee, B. K. Sharma, S. M. Shinde, S. D. Kim, K. Nan, Z. Yan, M. Han, Y. Huang,
108. G. Canepa, R. Petrigliano, M. Campanella, D. De Rossi, Detection of incipient object Y. Zhang, J.-H. Ahn, J. A. Rogers, Two-dimensional materials in functional three-
slippage by skin-like sensing and neural network processing. IEEE Trans. Syst. Man dimensional architectures with applications in photodetection and imaging.
Cybern. B Cybern. 28, 348–356 (1998). Nat. Commun. 9, 1417 (2018).
109. G. Cheng, E. Dean-Leon, F. Bergner, J. R. G. Olvera, Q. Leboutet, P. Mittendorfer,
128. O. Khatib, X. Yeh, G. Brantner, B. Soe, B. Kim, S. Ganguly, H. S. Stuart, S. Wang,
A comprehensive realization of robot skin: Sensors, sensing, control, and applications.
M. Cutkosky, A. Edsinger, P. Mullins, M. Barham, C. R. Voolstra, K. N. Salama, M. L’Hour,

Downloaded from https://www.science.org at Northeastern University on November 15, 2024

Proc. IEEE 107, 2034–2051 (2019).
V. Creuze, Ocean one: A robotic avatar for oceanic discovery. IEEE Roboti. Autom. Mag. 23,
110. A. Del Prete, S. Denei, L. Natale, F. Mastrogiovanni, F. Nori, G. Cannata, G. Metta, Skin 20–29 (2016).
spatial calibration using force/torque measurements, in 2011 IEEE/RSJ International
129. D. Falanga, K. Kleber, D. Scaramuzza, Dynamic obstacle avoidance for quadrotors with
Conference on Intelligent Robots and Systems (IEEE, 2011), pp. 3694–3700.
event cameras. Sci. Robot. 5, eaaz9712 (2020).
111. R. Calandra, S. Ivaldi, M. P. Deisenroth, E. Rueckert, J. Peters, Learning inverse dynamics
models with contacts, in 2015 IEEE International Conference on Robotics and Automation
(ICRA) (IEEE, 2015), pp. 3186–3191. Acknowledgement: We also thank members of the University of California San Diego
112. E. Dean-Leon, J. R. Guadarrama-Olvera, F. Bergner, G. Cheng, Whole-body active Bioinspired Robotics and Design Lab and the Yale Faboratory for helpful comments. Funding:
compliance control for humanoid robots with robot skin, in 2019 International Conference B.S. and M.T.T. acknowledge support from the Office of Naval Research (grant nos.
on Robotics and Automation (ICRA) (IEEE, 2019), pp. 5404–5410. N00014-17-1-2062 and N00014-18-1-2277). D.S. was supported by a NASA Space Technology
113. E. Steltz, A. Mozeika, N. Rodenberg, E. Brown, H. M. Jaeger, JSEL: Jamming skin enabled Research Fellowship (grant 80NSSC17K0164). T.G.T. was supported by the SHERO project, a
locomotion, in 2009 IEEE/RSJ International Conference on Intelligent Robots and Systems Future and Emerging Technologies (FET) program of the European Commission (grant
(2009), pp. 5672–5677. agreement ID 828818). Z.B. acknowledges support of related research by the Air Force Office
114. M. Ishida, D. Drotman, B. Shih, M. Hermes, M. Luhar, M. T. Tolley, Morphing structure of Scientific Research Materials Chemistry Program (grant no. FA9550-18-1-0143). R.K.-B.
for changing hydrodynamic characteristics of a soft underwater walking robot. acknowledges support of related research by the NSF Emerging Frontiers and Multidisciplinary
IEEE Robot. Autom. Lett. 4, 4163–4169 (2019). Activities Program (award no. 1830870). Author contributions: B.S., D.S., J.L., and T.G.T
115. J. H. Pikul, S. Li, H. Bai, R. T. Hanlon, I. Cohen, R. F. Shepherd, Stretchable surfaces conceptualized the review and wrote the manuscript. Y.-L.P., F.I., Z.B., R.K.-B, and M.T.T guided
with programmable 3D texture morphing for synthetic camouflaging skins. Science 358, the topics of discussion and structure of the review. All authors provided feedback and
210–214 (2017). revisions. Competing interests: The authors declare that they have no competing financial
116. H. A. Sonar, A. P. Gerratt, S. P. Lacour, J. Paik, Closed-loop haptic feedback control using interests. Data and materials availability: Correspondence and requests for materials should
a self-sensing soft pneumatic actuator skin. Soft Robot. 7, 22–29 (2019). be addressed to M.T.T. (email: [email protected]).
117. J. Wirekoh, L. Valle, N. Pol, Y.-L. Park, Sensorized, flat, pneumatic artificial muscle
embedded with biomimetic microfluidic sensors for proprioceptive feedback. Soft Robot. Submitted 22 October 2019
6, 768–777 (2019). Accepted 24 March 2020
118. R. S. Dahiya, G. Metta, M. Valle, G. Sandini, Tactile sensing—From humans to humanoids. Published 22 April 2020
IEEE Trans. Robot. 26, 1–20 (2009). 10.1126/scirobotics.aaz9239
119. J. C. Stevens, Aging and spatial acuity of touch. J. Gerontol. 47, P35–P40 (1992).
120. R. S. Johansson, A. B. Vallbo, Tactile sensibility in the human hand: Relative and absolute Citation: B. Shih, D. Shah, J. Li, T. G. Thuruthel, Y.-L. Park, F. Iida, Z. Bao, R. Kramer-Bottiglio,
densities of four types of mechanoreceptive units in glabrous skin. J. Physiol. 286, M. T. Tolley, Electronic skins and machine learning for intelligent soft robots. Sci. Robot. 5,
283–300 (1979). eaaz9239 (2020).

Shih et al., Sci. Robot. 5, eaaz9239 (2020) 22 April 2020 11 of 11