SCIENCE ROBOTICS | REVIEW

SENSORS Copyright © 2022

The Authors, some
Neuro-inspired electronic skin for robots rights reserved;
exclusive licensee
American Association
Fengyuan Liu1†, Sweety Deswal1†, Adamos Christou1, Yulia Sandamirskaya2, for the Advancement
Mohsen Kaboli3,4, Ravinder Dahiya1* of Science. No claim
to original U.S.
Touch is a complex sensing modality owing to large number of receptors (mechano, thermal, pain) nonuniformly Government Work
embedded in the soft skin all over the body. These receptors can gather and encode the large tactile data, allowing
us to feel and perceive the real world. This efficient somatosensation far outperforms the touch-sensing capability of
most of the state-of-the-art robots today and suggests the need for neural-like hardware for electronic skin (e-skin).
This could be attained through either innovative schemes for developing distributed electronics or repurposing the
neuromorphic circuits developed for other sensory modalities such as vision and audio. This Review highlights
the hardware implementations of various computational building blocks for e-skin and the ways they can be inte-

Downloaded from https://www.science.org at Oklahoma State University Stillwater on September 26, 2024
grated to potentially realize human skin–like or peripheral nervous system–like functionalities. The neural-like sensing
and data processing are discussed along with various algorithms and hardware architectures. The integration of
ultrathin neuromorphic chips for local computation and the printed electronics on soft substrate used for the devel-
opment of e-skin over large areas are expected to advance robotic interaction as well as open new avenues for
research in medical instrumentation, wearables, electronics, and neuroprosthetics.

INTRODUCTION obtain tactile information, such as object properties and action com-
The sense of touch is crucial to cope with the everyday challenges re- mands, from the raw tactile data. However, these models often rely
lated to interaction with objects, to safely manipulate and explore on structured interactions and do not provide accurate information
them to understand their physical properties (1), and for perception needed for control or robust perception. The alternative is to use
and self-awareness (2). When deprived of reliable tactile information data-driven methods that learn mappings from raw sensory data, or
[e.g., through the numbness of anesthetized or cold fingers (3, 4)], lower-level features, to high-level object properties and action com-
people become clumsy, and accidents are prone to occur. Similarly, mands. In this regard, supervised, unsupervised, and reinforcement
tactile sensing or haptics also has a vital role in the development of learning can be explored. Hierarchical representations, such as neural
cognitive and intelligent robotic systems because it allows them to networks, are often used to learn multiple levels of features (19, 20).
autonomously explore their surroundings. Robotic systems of the Flexible representations allow the robot to adapt the learned model to
future thus need touch sensing to safely interact in dynamic, un- the specific task based directly on data. The analytical, data-driven, or
structured, and often uncertain environments. As a result, during the algorithmic approaches can be more effective with distributed com-
past several decades, researchers have explored numerous ways to puting in tactile skin (21). The neural-like hardware developed for
create an artificial sense of touch through various types of sensors in other sensory modalities, such as vision and audio, or the chips
bendable and stretchable form factors (5–8) [e.g., resistive (9), piezo­ developed to imitate the working of central nervous system (CNS)
resistive (10, 11), capacitive (12), optical (13), piezoelectric (14, 15), (22–24) can be repurposed for tactile data processing. However,
acoustic (16)]—either individually or as a stack (17), and yet, we are such solutions are not ideal for tactile sensing, because unlike other
far from the tactile sensing capabilities possessed by humans. Some of sensory modalities, tactile sensing is physically distributed all over
the key developments for tactile sensing or electronic skin (e-skin) in the body and requires mechanical softness to interact with other
robotics are shown in Fig. 1. objects.
For a robot to have human-level perceptual capability, it is im- Robotics has considerably advanced from using few sensors in the
portant to associate the tactile sensors with a similar data processing hands to using large numbers of sensors all over the body (Fig. 1) to
system in the way that receptors work in the peripheral nervous sys- meet the requirements of emerging tasks that exploit the large-area or
tem (PNS). This requires physically distributed computing hardware whole-body contact to manipulate objects or navigate through un-
on soft substrates, along with tactile sensors. Although tactile sensors structured or cluttered environments. With increasing numbers of
have received substantial attention in the past, the data encoding and sensors, the amount of tactile data they generate can rapidly approach
processing using dedicated hardware has not been explored as much. practical limits, such as occupying the communication bandwidth. As
The tactile information processing in robotics so far has mainly in- such, it is impractical to send all of the data to the robot’s centralized
volved analytical or data-driven approaches, using a software plat- computing hardware (an equivalent of the brain). Likewise, the power
form (18). Analytical approaches exploit physics-based models to requirements can be considerably high. This calls for e-skin to have
efficient data handling capability, and it could be achieved through
1
distributed low-power electronic hardware for computing. This no-
Bendable Electronics and Sensing Technologies (BEST) Group, James Watt School
of Engineering, University of Glasgow, G12 8QQ Glasgow, UK. 2Neuromorphic Com-
tion aligns with the way the PNS complements the CNS.
puting Intel Labs, Munich, Germany. 3Department of Research, New Technologies, Focusing on the computing hardware for e-skin, this Review com-
Innovation, BMW Group, Parkring 19, 85748 Garching bei Munchen, Germany. plements previous review articles that have presented topics such as
4
Cognitive Robotics and Tactile Intelligence Group, Donders Institute for Brain, various types of tactile sensors (25–28), techniques, and materials (for
Cognition, and Behaviour, Radboud University, Nijmegen, Netherlands.
*Corresponding author. Email: [email protected] example, using liquid metal and hydrogel) to realize sensors in soft
†These authors contributed equally to this work. and flexible form factors (29, 30), identification of object properties

Liu et al., Sci. Robot. 7, eabl7344 (2022) 8 June 2022 1 of 15

 SCIENCE ROBOTICS | REVIEW

Downloaded from https://www.science.org at Oklahoma State University Stillwater on September 26, 2024
Fig. 1. The tactile sensation and perception in human skin and the evolution of artificial tactile skin in robots. Skin is the major component of human PNS (left), which
has inspired robotic tactile skin research over the years. The tactile sensing technologies have advanced from single-touch sensor to large-area skin (middle) covering the whole body
of a robot (right). This poses challenges in data handling and energy consumption. The images (from bottom to top) in the timeline (middle) are adapted from (146, 147, 184–187) with
permissions. Distributing computing in tactile skin can further boost the interaction capabilities of robots. The block diagram below shows the key steps to attain the same
and outlines the structure of this paper. The image of the robot on the right-hand side is an adaptation from the image “iCub Humanoid platform,” created by “antoni gràcia.” Link of the
original image: https://www.behance.net/gallery/17635773/iCub-Humanoid-platform Link for the license: https://creativecommons.org/licenses/by-nc/4.0/legalcode.

and interactions (31), and distributed energy (32, 33). This article in areas such as medical instruments, wearables, and neuroprosthetics,
also complements previous reviews covering neuro-/bio-­inspired and therefore, researchers working in these diverse areas will benefit
e-skin (34–36), providing a systematic and comprehensive discussion from this article.
on the computing element in tactile sensing. In this regard, this Re-
IMAGE CREDITS: TOP TO BOTTOM FOR TIMELINE (146, 147, 184–187)/IEEE

view is distinct and timely. The discussion starts from a biological

perspective, summarizing the state-of-the-art viewpoints related to TACTILE SENSATION AND PERCEPTION IN HUMAN SKIN
tactile data encoding and processing in PNS. The discussion also Constructing a robust perceptual system for robots based on the tac-
highlights the key message for the hardware implementation of tile sensory information is a critical but challenging task. Because hu-
neural-like processing in e-skin. This is followed by the discus- mans manage this almost effortlessly, the human body is an excellent
sion on various building blocks for hardware implementations. reference to follow. This section presents some fundamental but vital
Although some of these building blocks are not developed for features in the tactile sensation and perception in human body, espe-
e-skin, they can be repurposed to develop one to address the challeng- cially at the PNS level with a view to realizing computational e-skin
es stated above. The pathways for the integration of these building blocks for the next generation of robots.
with possible algorithms, leading to the development of computation- The human body acquires the tactile sensory information through
al e-skin, are discussed afterward. We then present how the imple- thousands of mechanoreceptors distributed over the skin (Table 1),
mentation of large-area computational e-skin could gain from the as shown in Fig. 2A. This large number of receptors ensures a reliable
advances in flexible and printed electronics technologies. The intelli- sensation but poses challenges in data transmission. The biological
gent robotic skin capable of extracting low-level information from solution is to use action potentials, sometimes with adaptation, to
the abundant tactile data could also open new application avenues encode, communicate, and control the body (37). Such a manner is

Liu et al., Sci. Robot. 7, eabl7344 (2022) 8 June 2022 2 of 15

SCIENCE ROBOTICS | REVIEW

Table 1. Typical properties of human skin (78, 194–204). NA, not available.
Relative area (%) Strain (%) No. of mechanoreceptors Spatial acuity (cm)*
Location
(78, 194, 203) (196–201) (195, 202) (204)
Finger ~1.3 35–45 ~13,350 ~0.2
†
Foot ~6.1 <30 1,000–5,000 0.8–1.8
Chest ~12.8 <30† ~13,000 ~3.2
Back ~13.9 <30† 2,000–14,000 ~1.3
Shoulders ~1.9 NA ~4,000 ~3
Abdomen ~3.6 NA ~4,000 ~3.6
Thigh ~18.3 <30† ~30,000 ~2.3
Wrist ~0.7 10–40 ~1,500 ~4.2

Downloaded from https://www.science.org at Oklahoma State University Stillwater on September 26, 2024
Knee ~1.2 30–40 ~2,000 ~4.7
Elbow NA 60 NA ~4.2

*Spatial acuity denotes the two-point discrimination threshold.   †Data are anticipated according to the value from other areas of the skin.

highly beneficial in terms of reducing power consumption and data achieving edge computing at the skin level itself could greatly lower
latency. the data transmission burden. Like the audio and vision, it is also
Reverse engineering to develop the e-skin requires knowledge argued that the tactile sensory data are processed via time division:
beyond neuroscience principles. For example, from a structure The firing pattern of several neurons at the same time would trigger
viewpoint, the skin is soft and present over a large area. With a large the firing of the next-level neuron and, in turn, strengthen the afferent
number of receptors (of varying thresholds) distributed at different synapse (Fig. 2C) (47). By doing so, the tactile sensation aroused by
depths, the skin can respond to the stimuli of various frequencies an object is correlated with the firing of one (or multiple) second-­
[fast adapting (FA) and slow adapting (SA)] (38). Moreover, each order neuron(s), and the generated spiking patterns are sent for fur-
subtype of receptors shows a different size of receptive field; they ther processing (Fig. 2C). This lays the foundation for tactile signal
overlap with each other and are interconnected locally (39–41). Such computing using the spiking time–dependent plasticity (STDP) learning
an intricate nature forms the foundation of the fine spatial sensitivity rule. Nevertheless, this is not the only rule available. Other possible
of the skin. Toward this, several works have reported various sen- learning rules, whether or not they are biomimetic, can be implemented
sors that provide similar functionalities as SA and FA receptors and in the e-skin to enable the intelligent data processing. This is discussed
also suggest their stacking to mimic the mechanoreceptor arrange- in the “Neural system implementation and algorithms” section.
ment in the skin (Fig. 2A) (17, 42, 43).
Compared with tactile sensation, the understanding of the tactile
perception is far less developed. The general view is that the percep- COMPUTATIONAL BUILDING BLOCKS AND THEIR
tion starts at the cuneate (second-order tactile) neuron (in the spinal HARDWARE IMPLEMENTATIONS
cord), while the first-order neuron (sensory neuron) is solely re- To fully mimic the functionality of the biological skin and the asso-
sponsible for the tactile sensing (38). However, it is also argued that ciated PNS, the building blocks that can function as sensors, neu-
the mechanical properties such as the softness of skin may also be rons, and synapse are required (Fig. 2A). This section reviews the
related to the perception process. Some level of computing takes place hardware devices mimicking sensory neurons and cuneate neurons
in the skin itself owing to the location-specific tactile sensing char- and the synaptic devices needed between these two sets of neurons.
acteristics inside the receptive field of sensory neuron and the soft The functionalities required from the individual building blocks,
nature of the skin, which deforms during contact (Fig. 2B) (44, 45). along with their preliminary integration in a neural pathway for
In this regard, the spiking pattern from the first-order neuron could robotic applications, are discussed in the subsequent subsections.
contain some information of the edge orientation. Although more
similar studies are needed to unravel the perceptual element of tac- Artificial neuron
tile sensing in biological skin, hardware implementation of some level The neuron is the basic processing unit in the biological neural sys-
of computation in the artificial skin will be beneficial. It is because, tem, where analog incoming signals are integrated and converted
unlike biosystem where hyperconnectivity is regular, the electronic into action potentials when the spatiotemporal summation exceeds
system usually has a much lower fan-out/fan-in ratio, with one fan-in the firing threshold of the neuron. Previous studies have identified
leading to three to four fan-out on average in digital electronics approximately 20 typical spiking features in the biological system,
(46). The realization of a hyperconnected electronic system is chal- such as tonic spiking, tonic bursting, and phasic spiking (48). Among
lenging from both a design and a fabrication perspective. Another them, some features have been widely observed and play important
reason is that the flexible, stretchable nature of the artificial skin roles in somatosensation. The capability of delivering such spiking
poses a major hurdle to reliable data transmission. The large num- features, for example, frequency adaptation, should be one of the
ber of receptors confined in a limited area could markedly increase criteria for the hardware implementation of neuron block for the
the difficulty of data transmission without losing fidelity. Instead, e-skin. Here, we review the state-of-the-art strategies that have been

Liu et al., Sci. Robot. 7, eabl7344 (2022) 8 June 2022 3 of 15

 SCIENCE ROBOTICS | REVIEW

excitable) and event-driven sensing. How-

ever, the shape of the spiking signal they
generate is different. Because the first
strategy uses an AND gate, it generates
square-shaped spikes (Fig. 3A); in con-
trast, the conventional neuron circuits
using Si complementary metal-oxide semi-
conductor (CMOS) can deliver biologi-
cally plausible spiking patterns including
the hyperpolarization and depolarization
stages, refractory period (Fig. 3B). Be-
cause the shape of the spike governs the
plasticity-based learning of artificial syn-
apses, the synaptic devices need to be

Downloaded from https://www.science.org at Oklahoma State University Stillwater on September 26, 2024
identified and designed carefully along
with the associated neurons. From the
engineering point of view, the first strategy
relies on the digital circuit and can be
more robust to disturbance. In this re-
gard, some parts have been successfully
demonstrated on flexible substrates using
printing techniques. However, consider-
ing the whole neural system, the sim-
plified spiking signal generated by the
oscillating circuits cannot offer some of
the vital spiking features observed in
biology. To realize a system with rich
neural dynamics, the second strategy may
be required.
Considering the open and unpredict-
able environment the skin typically ex-
periences, the sensors need to exhibit
excellent specificity/selectivity, particularly
because the e-skin requires multimodal
Fig. 2. The neural pathways for tactile data processing in human skin and their simple implementation schemes sensing to detect various parameters such
using basic devices and circuits. (A) Neural pathway for tactile data processing. The sensing data are collected in as stress, pressure, and temperature. This
the sensory neuron (first order), passed to the cuneate neuron (second order), and finally sent to the higher levels of can be challenging because the electronic
the nervous system. The key components (mechanoreceptors, neurons, and synapses) of human body involved in
systems themselves are temperature sen-
tactile sensing and data encoding are shown with the illustrations of their simple implementations through electronic
devices and circuits. The SA and FA mechanoreceptors are usually mimicked by a stack of capacitive and piezoelectric
sitive. Nevertheless, there are many
sensors (17) (i). The RC circuit (ii) emulates the action potential of the spiking neuron (63). Memristors or similar devices encouraging works showing excellent
(iii) can be used for the implementation of various synaptic functions such as time- and rate-based learning rules. The specificity regardless of ambient varia-
rate-based learning rule refers to BCM (Bienenstock, Cooper, and Munro) learning rule. (B) Possible tactile data pro- tions (55–59). Alternatively, one could
cessing at the first-order neuron (44, 188). Reprinted with permission from (188). (C) Possible tactile data processing use low-specificity sensors and calibrate
at the second-order neuron. Reprint from (47) with permission. later, for example, at the second-order
neuron stage.
Cuneate neuron
This subsection discusses the strategies
IMAGE CREDITS: (B) (188)/IEEE (C) (47)/SPRINGER NATURE

developed for sensory neuron (that is, neuromorphic sensors) and for the hardware implementation of spiking neuron and their eval-
then the neuron circuits in general, which can be used for the reali- uation in terms of the biological plausibility, that is, the number of
zation of the second-order tactile neuron. the biologically observed spiking features, and the implementation
Sensory neuron (neuromorphic sensor) cost measured by the active devices required in the design. The neu-
The sensory neuron is the first-order neuron in the tactile neural ron circuit discussed here can be used to realize the second-order
pathway. The sensors designed for the computational e-skin should tactile neuron for the computational e-skin. From the model point
output the spiking signal, and to this end, the two hardware imple- of view, several mathematical equations have been proposed to de-
mentation approaches have been explored so far. One is to integrate scribe the neuron spiking process. These include leaky integrate-and-fire
the sensor with oscillating and edge-detection circuits, as illustrated model (60), Hodgkin-Huxley (61), and Izhikevich model (62). These
in Fig. 3A (49–52), and the second is to interface the sensor with the neuron models can be implemented using one or two capacitors
neuron circuits, as shown in Fig. 3B (53, 54). Both approaches could with leaky resistors to emulate the important features observed in
provide biological features such as spiking rate dependency (class 1 biological neurons (63) (Fig. 2A), such as spatiotemporal integration

Liu et al., Sci. Robot. 7, eabl7344 (2022) 8 June 2022 4 of 15

SCIENCE ROBOTICS | REVIEW

Downloaded from https://www.science.org at Oklahoma State University Stillwater on September 26, 2024
Fig. 3. The hardware implementation of an artificial neuron. (A) and (B) show the strategies to develop a neuromorphic sensor. (A) An oscillating circuit–based neu-
romorphic sensor. This strategy uses digital circuits and thus is easier for the hardware implementation on soft substrates as demonstrated in (49–52). (B) A neural circuit–
based neuromorphic sensor (53). Such a strategy is able to provide a more biomimetic spiking signal. However, it is still a challenge for the hardware implementation in
soft platforms, especially over large areas. (C) and (D) show the strategies to develop spiking neuron circuit in general. (C) The phase-change device–based spiking neuron
(64). (D) The volatile resistive memory–based spiking neuron (69). (E) The strategy to implement the nonspiking neuron using a multigate transistor. Adapted from (78).
(F) The performance comparison of various spiking neurons, with data extracted from (64–75, 189, 190). The criteria of the biological plausibility are based on the spiking
patterns discussed in (48). As can be seen from the figure, the neuron circuits based on novel materials/devices require a smaller number of active devices compared with
the CMOS-based neurons.

and all-or-none rule. In this regard, the components showing a realized with only two leaky phase-change devices: Each controls a
leaky resistance are the key to the hardware implementation of neu- capacitor, thus mimicking the neuron’s opening and closing of the
ron circuits. sodium and potassium channels gated by voltage. Such a design can
One way to emulate the leaky component is to use phase-change emulate almost all the spiking features found in biological neurons—
devices (Fig. 3C). The resistance of such a device remains high under including tonic and phasic spiking, tonic and phasic bursting, and
a low voltage bias, owing to an insulating phase, and shows an abrupt frequency adaptation (Fig. 3C) (64)—and thus can be used to real-
transition to a low resistance state (metallic phase) when sufficiently ize a neural system with rich neural dynamics. It is also possible to
high bias voltage is available. Such a transition is reversible depend- implement a neuron circuit with only one phase-change device (65–67),
ing on the voltage bias. A highly biomimetic neuron circuit can be although in this case the spiking signal generated cannot fully mimic the

Liu et al., Sci. Robot. 7, eabl7344 (2022) 8 June 2022 5 of 15

SCIENCE ROBOTICS | REVIEW

biological spiking features. Overall, the phase-change device–based been explored to develop resistive random access memory (RRAM)
design offers the lowest hardware complexity (Fig. 3F). (84, 85), phase-change memory (PCM) (86–88), magnetic random
The leaky resistive device can also be realized with volatile resistive access memory (MRAM) (89), and ferroelectric random access memory
memristors (68, 69), whose resistance switching is governed by the (FeRAM) (90), leading to two- and three-terminal structures. Such
voltage bias across the device (Fig. 3D). From a fabrication view- devices exploit the change in physical properties of the material in
point, it seems relatively easier to fabricate the resistive memristors, response to external electrical stimuli to mimic the general synaptic
and thus, they may hold more promise for the large-area implementa- behavior (Fig. 4A). The memristive crossbar arrays (91) are a con-
tions on soft substrates. cise and attractive route for implementation of synapses (Fig. 4B).
A more common strategy is to use CMOS circuits (70–75). In For ANN, the operation of synapse is usually achieved by using a
this case, the subthreshold behavior of metal-oxide semiconductor single pulse signal, with the aim to achieve a long retention time
field-effect transistor (MOSFET) is used to emulate the resistance change. with multilevel weight tuning (Fig. 4C). Contrary to ANN, the SNN
Unlike phase-change and memristive devices, the MOSFETs are three-­ is more biologically plausible because it operates the synaptic devices
terminal devices whose channel resistance is controlled by the gate using a pair of pulses (pre- and post-) after various biological learn-
voltage. The CMOS neuron requires a complicated layout to con- ing rules, for example, STDP (Fig. 2A) (92–96). Figure 4D illustrates

Downloaded from https://www.science.org at Oklahoma State University Stillwater on September 26, 2024
trol the resistance of the MOSFETs channel, to mimic the leaky be- some examples of hardware-implemented large-scale artificial syn-
havior. The comparison between the hardware complexity and the apses: They are in the rigid form factor. The future development of
biological plausibility for each strategy has been illustrated in Fig. 3F. large-scale synapses for e-skin should be carried out on soft substrates.
The spiking-based computing system consumes substantially less Although crossbar arrays provide a concise architecture, they
power. For example, the energy required for the neuron circuit based could lead to several potential problems as well. For example, the
on phase-change device can be as low as 100 fJ per spike (76). This sneak path issue (current leakage through unselected cells) is a well-
value can be even lower for CMOS neurons (~4 fJ) (71). Assuming a known bottleneck for the large-area implementation of memristive
moderate firing rate of 100 Hz, the energy required per neuron per crossbar arrays. In addition, the requirement of electrical forming
second is in the order of picojoule. For a direct comparison, we use process for each memristor pixel could enhance the hardware com-
the power consumption of Intel Core i7-920 microprocessor as a ref- plexity, which may make the filamentary memristive devices less
erence. This microprocessor consumes a power of 85 W with 731 million suitable for the large-area implementation. Selector devices such as
transistors. This corresponds to ~100-nJ energy consumption per transistor (97) or self-selective (passive) memristive crossbar (98)
second for a single transistor or several hundreds of nanojoules for a have been investigated to overcome the sneak path issues. Recently,
single gate—this is substantially higher than the power consumption a passive crossbar array was directly implemented with CMOS circuit
of a neuron circuit. allowing online learning and vector matrix multiplications, thus
Nonspiking neuron providing operational neuromorphic computing hardware (99). The
The nonspiking neuron is another alternative where a continuous hardware implementation of high-performance, high-yield, and
function represents the firing rate and thus presents a simplified uniform one transistor–one memristor (1T1M) crossbar arrays for
activation function. Both digital and analog circuits have been ex- convolutional neural networks is another example of energy-efficient
plored for implementation of nonspiking functionalities, although large-scale networks (100). Using transistors as selector devices
they usually have a complex layout and consume high power. Attempts leads to 1T1M structure (101), which may be used for synaptic
have been made to seek alternatives such as on VO2-based Mott functionality in e-skin. However, their large form factor needs to be
devices (77) and nanowire-based neural FETs (78) to achieve recti- considered as well. To this end, one solution is to merge the selector
fied linear unit and threshold function, respectively (Fig. 3E). The device with RRAM (102, 103), which allows a higher lateral and ver-
benefits of using a nonspiking neuron include easy and highly accu- tical integration than the conventional 1T1M configuration.
rate training of the associated neural network. However, their rela- Despite several reports on RRAM-based artificial synapses, their
tively higher power requirements compared with the spiking neuron hardware realization is limited possibly because of challenges such
are a major drawback. Thus, the selection of spiking or nonspiking as poor resistance tuning, spatial and temporal variability, device yield,
neurons as the building block for the computational e-skin may and nonlinearity/asymmetry. These issues negatively affect the perform­
vary with the application. ance of neuromorphic computing (104). On the other hand, the
floating gate transistors offering better uniformity are actively being
Artificial synapse explored for synaptic functionalities (105–108), which are considered
As the core building block of a neural network, the synapse has two as a more mature technology due to their compatibility with CMOS
main functionalities: the synaptic efficacy (the ability to pass the sig- fabrication process. Emerging synaptic transistors with various working
nal) and the synaptic plasticity (the ability to adapt the weight accord- mechanisms like electrochemical (109), charge trapping/detrapping
ing to various learning rules). To replicate such functionalities in (110), and light assistive reaction (111, 112) have also been reported
hardware, CMOS circuits and novel electronic devices (79–83) have along with their application in neuromorphic tactile sensing and pro-
been explored, for both artificial neural network (ANN) and spik- cessing system (113, 114). Some hardware implementations of these
ing neural network (SNN). devices are worthwhile; however, efforts are needed to achieve the
For the proposed computational e-skin, the hardware imple- response uniformity over a large area, with desired retention and
mentation over large areas with flexible/stretchable substrates is desired, endurance (115). As mentioned above, for robotic skin, the flexibility/
and the incompatibility of CMOS devices with soft and conformable stretchability of artificial synapse is critical, and in this regard, the soft
materials raises some challenges. Alternatives such as memristive synaptic transistors (116–118) and memristors (119, 120) are relevant.
devices on flexible substrates have been explored to realize compact Flexible high-­performance synaptic transistors have been reported
artificial synapse. A variety of organic and inorganic materials have to mimic the native biological synapses. Likewise, on the basis of

Liu et al., Sci. Robot. 7, eabl7344 (2022) 8 June 2022 6 of 15

 SCIENCE ROBOTICS | REVIEW

Downloaded from https://www.science.org at Oklahoma State University Stillwater on September 26, 2024
Fig. 4. Hardware Implementation of artificial synapse. (A) Various two- and three-terminal devices as artificial synapses including RRAM, PCM, floating gate transistor,
IMAGE CREDITS: (A) (191)/SPRINGER NATURE (D) (I) (99); (II) (95)/SPRINGER NATURE

and Fe-FET. The schemas of the RRAM, PCM, and floating-gate transistor are redrawn based on (191). (B) Matrix vector multiplication in crossbar arrays with both active
and passive memristive devices. Potentiation and depression characteristics of analog memristive devices. (C) Synaptic behavior as excitatory and inhibitory with long-
and short-term plasticity. (D) Hardware implementation of some memristor-based artificial synapse in (i) ANN [reprint from (99) with permission. Copyright 2019, Springer
Nature] with integration of memristor crossbars and CMOS chips on wire-bonded pin-grid array package (inset shows the testing setup) and (ii) SNN [reprint from (95)
with permission] using PCB with passive memristor switching matrix (inset shows the scanning electron image of memristor crossbar). (E) Localized neural pathway for
robots. (i) Distributed sensorimotor merging that leads to the learning of the maze exploration of the robot. Adapted from (127) with permission. (ii) Distributed sensorim-
otor correlation that leads to the acquisition of the pain reflex. Reprint from (128) with permission.

flexible memristors, a three-dimensional (3D) artificial synaptic on flexible (121) and stretchable (118, 123) artificial synapses are
network has been reported to enable direct correlated learning and encouraging because these devices are suitable for large-area
trainable memory capability (121). A new class of flexible mem- robotic skin.
ristors based on 2D materials (122) is also very attractive, with The performance of an artificial synaptic device is typically
compatibility of solution processing and printing technology. The assessed in terms of multilevel states, linearity, retention, endurance,
thickness of 2D materials in subnanometer range further facilitates dynamic range, variability, device area, device yield, and energy
low operating voltages and switching speed (115), but it is still lack- consumption. However, not all of these device metrics are critical to
ing in terms of large-area scalability and uniformity. Recent reports every application, and they can depend on the training algorithm

Liu et al., Sci. Robot. 7, eabl7344 (2022) 8 June 2022 7 of 15

SCIENCE ROBOTICS | REVIEW

and specific task. A comparison of some of the hardware-implemented NEURAL SYSTEM IMPLEMENTATION AND ALGORITHMS
synaptic devices is demonstrated in Table 2. The technology for The neural network structure
such novel neuromorphic devices is at an infancy, and despite their This section presents the possible neural system structure for the
advantages, now, they are unable to replace CMOS technology at the computational e-skin along with its training algorithms. This includes
commercial hardware level. A hybrid design, for example, memristor-­ one layer of sensory neurons and one layer of cuneate neurons, con-
based analog computing along with CMOS-based digital computing, nected via synapses (Fig. 5A). As discussed in the “Tactile sensation
maybe the solution in the near future. and perception in human skin” section, the subtypes of mechanore-
ceptors have different sizes of receptive fields, which overlap and
Peripheral neural pathway for robots are interconnected to form the basis for the spatial sensitivity of the
Robots require sensory-motor fusion and adaptive interaction with skin (40, 129). We present a similar concept for the proposed neural
the environment. Specifically, skin-like tactile sensing has been used network (Fig. 5A). The stimulation of one neuron could influence
in robotics for tactile-based environmental exploration, physical other neighboring neurons, and thus, it is possible to map the in-
human-robot interaction and collaboration, objects’ physical prop- coming tactile signal to a higher-dimensional space in the first layer;
erties recognition, tool manipulation, and locomotion (18, 124, 125). How- the output is fed into the second layer for further processing. The

Downloaded from https://www.science.org at Oklahoma State University Stillwater on September 26, 2024
ever, most of these demonstrations are achieved by software approaches proposed neural network is similar to the reservoir computing. How-
in a centralized computing unit, that is, the equivalent of brain. In ever, owing to soft tissues, the native skin is more complex: The re-
addition, these platforms usually use digital circuitries, which add ceptive fields of the neighboring mechanoreceptors are modulated
to the latency of the process. Instead, in organisms, localized sensory-­ by the external tactile stimuli (Fig. 2B) (130, 131). This is one of the
motor coordination and integration has been widely observed (126). unique aspects of tactile sensing (44), and mimicking the same could
Such a decentralized, analog processing notably reduces the data be a future direction of e-skin research. The proposed neural net-
latency for robots. For this, the building blocks discussed above work is capable of carrying out various tasks required by robots, such
can be used to construct the localized computing platform needed as local tactile feature recognition and contact/slippage detection.
for tactile sensing, opening avenues for next-generation robots Nevertheless, it is also necessary to consider what kinds of tasks are
(Fig. 4E). suitable for localized processing at the skin level. One possible di-
For example, the agent-environment interaction is required for rection is that a higher-level arrangement can be made in a similar
the exploration and learning of robots. As shown in Fig. 4E (i), rein- manner to organisms in nature. For example, as discussed in earlier,
forcement learning could be done in a localized manner using dis- differentiating the edge orientation of a tactile stimulus is a task
tributed synaptic circuits to learn the maze exploration by the robots (127). processed in the skin level for humans (45), and similar tactile per-
Similarly, robots need to work in unknown environments with the ception could be realized with the computational e-skin as well.
potential to sense the impending hazard. This can be achieved by the
correlation between the sensory and motor signal, allowing the robot The learning algorithms
to identify the pain signal and respond to it (128). Owing to the use Implementing the proposed neural network using the hardware build-
of the synaptic device [Fig. 4E (ii)], the withdrawal reflex behavior ing blocks is another important aspect to consider. Various learning
could be mimicked in a concise manner. The above examples show strategies including supervised, unsupervised, and reinforcement
how decentralized computing can be constructed in a simple neural learning are available. However, developing an all-hardware–based
configuration, benefiting the next-generation robots. In the next supervised learning system can be costly in terms of devices or cir-
section, possible neural network structures are discussed, which cuits needed (97, 132, 133) and will thus not be discussed for the
could potentially be used for carrying out more complicated tasks e-skin. On the other hand, the plasticity-based learning rule, STDP,
for robots. is the fundamental learning rule in SNNs and leads to unsupervised

Table 2. Hardware-integrated artificial synapse devices.

Memristive device Application of
Reference
configuration Cell dimension ( m2) Size of crossbar array Energy consumption
hardware system
Recognition of printed
(205) 1M 10 × 10 15 × 6 0.31 J
digits
MNIST (National
Institute of Standards
(100) 1T1M 0.5 × 0.5 128 × 16 371.89 pJ
and Technology)
image recognition
Pattern matching and
(206) 1M <0.5 × 0.5 32 × 32 719.0 J natural image
processing
Demonstration of
(99) 1M 0.5 × 0.5 54 × 108 1.12 pJ different computing
models
(95) 1M – 20 × 20 – Coincidence detection

Liu et al., Sci. Robot. 7, eabl7344 (2022) 8 June 2022 8 of 15

 SCIENCE ROBOTICS | REVIEW

Downloaded from https://www.science.org at Oklahoma State University Stillwater on September 26, 2024
Fig. 5. The neural network structure and learning algorithms for the proposed computational e-skin. (A) The neural network structure and the potential tasks it
should address. A network similar to reservoir computing has been proposed. However, for e-skin, it should be more complicated because the weights in the first layer
along with the receptive fields of the sensory neurons are dependent on the mechanical stimuli. (B) Plasticity-based learning rule, STDP, for unsupervised learning. Various
forms of STDP have been identified. Typical examples include asymmetry Hebbian learning, asymmetry anti-Hebbian learning, and symmetric Hebbian learning. Reprint
from (135). (C) Waveform engineering is one popular strategy for achieving hardware realization of the STDP learning rule using novel devices such as memristors and
synaptic transistors shown in [D (i)] and [D (ii)]. For this, the principle is to take advantage of the net bias across the device, Vpre-Vpost, under various timing. When the net
bias is larger than the threshold of the device needed to trigger the synaptic behavior (potentiation or depression), the weight (conductance) of the device would be
changed. (D) The possible circuit layout for realizing STDP learning for two-terminal memristor and three-terminal synaptic transistor. Adapted from (136).

learning (Fig. 5B) (134, 135). The implementation of such a learning memristor-based synapses, the pre- and post-neuron signals are fed
IMAGE CREDIT: (D) (136)/ELSEVIER

rule is hardware-friendly, especially with the previously mentioned into the two ends of the memristor devices. The net bias across the
synaptic devices. The change of the synaptic weight is only subjected memristor device is therefore the temporal subtraction of the pre-
to the time correlation between paired spikes from the pre- and post-­ and post-neuron spikes (Fig. 5D) (136). Thus, the difference in the
neurons, under various pairing schemes such as “nearest neighbor spike timing between the pre- and post-neurons would lead to a net
takes all,” “nearest neighbor takes more,” or “all spike pairs count bias of reverse polarity and varying amplitude (Fig. 5C). Depending
equally” (Fig. 5B) (135). This could be potentially promising for im- on the relationship between the potentiation/depression threshold
plementation on the soft computational e-skin. For example, for and the net bias across the device, the synaptic weight is modified

Liu et al., Sci. Robot. 7, eabl7344 (2022) 8 June 2022 9 of 15

 SCIENCE ROBOTICS | REVIEW

Possible examples may also exist in the

biological nervous system: The synaptic
plasticity could be influenced by a global
neuromodulator signal. For example,
dopamine can lead to facilitation, and
acetylcholine can lead to depression.
Several computational studies have shown
that they could lead to a higher learning
accuracy for various tasks (138–141). In
this regard, efforts are needed in the
hardware demonstration of the “three-­
factor STDP behavior” using the artifi-
cial synaptic devices, and the devices
offering multiple approaches to control

Downloaded from https://www.science.org at Oklahoma State University Stillwater on September 26, 2024
the synaptic behavior [chemical and light;
(142)] could be beneficial.

Power consumption
Because there are a limited number of
studies on the power consumption of
e-skins, we try to extract data from other
scenarios for comparison. For example,
the neuromorphic chip “SpiNNaker” con-
tains ~250,000 neurons and 82 million
synapses (143). It consumes 36 W in
total or 20 nJ per synaptic event, at an
average firing rate of 22 Hz. The chip
“TrueNorth” contains ~1 million neu-
rons and 128 million active synapses,
consuming an average of 26 pJ per syn-
aptic event at a firing rate of ~20 Hz (22).
The e-skin targeting human-level tactile
performance would require a similar num-
ber of sensory neurons in the first layer
as in the case of human skin (~100,000;
see Table 1), but with a much less recur-
rency as compared with SpiNNaker or
Fig. 6. The technological advances to implement large-area computational e-skin on flexible substrates. (A) TrueNorth. Assuming a smaller portion
The energy autonomous e-skin with graphene-based transparent touch-sensing layer on robotic hand and on flexi- (~1/10) of cuneate neurons in the second
ble solar cells. The back side of the solar cells shows rigid off-the-shelf chips. Images adapted from (192). (B) The layer, the power consumption of the
neuromorphic chip, TrueNorth, with 1 million spiking neurons implemented using digital electronics. Adapted from proposed e-skin should be in the order
(22). (C) UTC technology could be used to obtain a flexible version of neuromorphic chips and hence the computa- of several watts at a moderate firing rate
tional e-skin with greater flexibility. Adapted from (148, 193). (D) The roll-to-roll printing of various computing building (~20 Hz).
blocks for the future e-skin. The feasibility of such e-skin is evident from the examples of printed electronic layers and
In addition to the power consump-
devices: (i and ii) transfer printed Si nanoribbons and transistor made from them. Reprint from (159). (iii) Contact
tion, there are several other problems to
printed nanowire-based devices. Reprint from (174).
consider for developing a computational
e-skin. For example, the way to address
accordingly. From the functional viewpoint, the implementation of and send the sensing signal is an aspect for study, because the active
STDP learning rule could lead to the correlation between two neu- or passive matrix manner cannot support the high spatial and tempo-
rons and further possible sensorimotor correlation for robots. ral resolution at the same time. Another aspect is the method to
Although unsupervised learning has several benefits, such a learning guarantee the system reliability, because the continuous interaction
strategy alone may be insufficient. A method to strengthen the de- between the skin and the external environment could possibly lead
IMAGE CREDIT: (C) (148)/WILEY-VCH

sired behavior and weaken the undesired behavior, based on the to localized damage (144, 145).
environmental feedback, for example, using reinforcement learning,
may also be required. In terms of the hardware, reinforcement learning
has been demonstrated in the example shown in Fig. 4E and other TOWARD COMPUTATIONAL E-SKIN IMPLEMENTATION
works using memristor crossbar (or synaptic transistors) and digital Flexible PCB and chip thinning technology
circuitry (137). However, with respect to the more biologically plausible E-skin needs to be fabricated on soft substrates to mimic the me-
scenario, SNN, the hardware implementation has not been realized. chanical properties of biological skin. Initial progress has been made

Liu et al., Sci. Robot. 7, eabl7344 (2022) 8 June 2022 10 of 15

SCIENCE ROBOTICS | REVIEW

by integrating off-the-shelf sensors and electronic components onto However, it should be noted that the printed electronics is still
flexible printed circuit boards (PCBs), ranging from hand-based mainly developed within laboratories with limited device metrics
manipulation to the whole-body area, as shown in Fig. 1. Similar such as device density, uniformity, and mobility. Taking the fabri-
approaches could be adopted for computational e-skin by interfacing cation of FETs as an example, the largest number of printed FETs
sensors with neuromorphic chips (Fig. 6, A and B). The neuro- on the flexible or stretchable substrates is in the order of a few thousand
morphic chips offer a platform of dense, interconnected neural network (182). Using standard microfabrication techniques, this number can
that can carry out various computing tasks required by robots be in the order of tens of thousands (183). At present, printing tech-
(146, 147). Such an arrangement will enable “in-hardware” computing nology could possibly be used to produce those parts of the e-skin
capability in the e-skin using deep neural networks. This is different that require lower device density, which corresponds to the part of
from the strategy we have discussed so far to develop the “skin-type” the skin with a lower density of mechanoreceptors or spatial acuity.
tactile functionality, because the achievement of the human-level
tactile sensation and perception relies on the intricate interplay be-
tween the softness of the skin and the spiking response of the neurons, CONCLUSION
using a “shallow” network of two layers. Next-generation robots are expected to be highly intelligent and

Downloaded from https://www.science.org at Oklahoma State University Stillwater on September 26, 2024
The integration of soft and rigid materials limits the bendability. autonomous, and the sense of touch is critical for them to safely
Furthermore, the mismatch in their mechanical properties increases interact in dynamic, unstructured, and often uncertain environments.
the chances of failure during extended use. As a result, it is challeng- For intelligent systems to have human levels of performance, it is
ing to use hybrid devices on sensitive body parts such as fingertips vital to develop a sensitive tactile sensory system that provides at
of a robot, where high-density tactile feedback is required for inter- least similar information. To attain this, there is a need to expand
action and manipulation. This can be potentially mitigated through the e-skin research toward perception and learning. Now, most of
wafer thinning technology using ultrathin chip (UTC) (148–151), as the e-skin research still focuses on the tactile sensation and their in-
shown in Fig. 6C. The thin chips bonded onto flexible PCBs could tegration on substrates, which can conform to curvy surfaces of ro-
offer greater flexibility. With this strategy, one critical challenge is botic body. We address this need in this review article by focusing
the lack of suitable bonding techniques to gain access to the circuitry on the computational aspect of the skin and the associated PNS to
on thin chips from a soft platform. The conventional chip bonding efficiently process the tactile data. The e-skin that can mimic the bio-
methods are not suitable because they are likely to introduce cracks logical tactile neural pathway could offer the desired preliminary
in the chips. These issues could be resolved by methods such as bond- perception capability, markedly decreasing the cognitive load on their
ing by printing (152–156). The challenges related to reliable inter- central control units. This is analogous to the PNS complementing
connects could be addressed by mechanically flexible conductive the functionality of the CNS in humans. We discussed the possible
materials such as liquid metals (157). Overall, despite these challenges building blocks of the tactile neural pathways and the integration
and limitations, using UTCs with flexible PCB is possibly the quickest that could imitate their functionality. It is highlighted that the
route toward the realization of the computational e-skin. mechanical properties of the skin and the neurological behaviors are
correlated in both tactile sensation and perception. Furthermore,
Printed electronics on soft substrates we have also discussed how the e-skin development could benefit
The proposed e-skin can also be fabricated directly on soft platforms, from advances in areas such as printed and flexible electronics. By
using printing technology with both inorganic (158, 159) and organic revisiting the discoveries from diverse disciplines and reviewing the
(158, 160) materials (Fig. 6D). This will also be an attractive direc- state-of-the-art in tactile sensing and neuromorphic computing hard-
tion for future electronics, because resource-efficient manufacturing ware, it is hoped that this article will inspire future advances for the e-skin
could help to reduce electronic waste and contribute to sustainability. research, rendering the robots with human-­like responsiveness.
Further, it is easier to process biodegradable materials using printing
technology so that the printing technology also holds promise for REFERENCES AND NOTES
future transient electronics (161–163). Several printing techniques 1. M. Kaboli, G. Cheng, Robust tactile descriptors for discriminating objects from textural
properties via artificial robotic skin. IEEE Trans. Robot. 34, 985–1003 (2018).
that are available today include inkjet printing (118, 164–166), trans-
2. R. Dahiya, N. Yogeswaran, F. Liu, L. Manjakkal, E. Burdet, V. Hayward, H. Jorntell,
fer printing (159, 167–170), contact printing (171–174), and screen Large-area soft e-skin: The challenges beyond sensor designs. Proc. IEEE 107, 2016–2033
printing (175–179). Specifically, the transfer and contact printing (2019).
techniques are promising for high-performance electronics and also 3. R. S. Johansson, G. Westling, Roles of glabrous skin receptors and sensorimotor memory
compatible with roll-to-roll manufacturing (180). Figure 6D shows in automatic control of precision grip when lifting rougher or more slippery objects.
Exp. Brain Res. 56, 550–564 (1984).
an example of roll-to-roll printed multigate devices (Fig. 3E) on the 4. R. Dahiya, M. Valle, Robotic Tactile Sensing: Technologies and System (Springer Science +
flexible substrate along with various other building blocks required Business Media, 2013).
for the computational e-skin mentioned in the “Computational building 5. N. Yogeswaran, W. Dang, W. T. Navaraj, D. Shakthivel, S. Khan, E. O. Polat, S. Gupta,
blocks and their hardware implementation” section. Specifically, the H. Heidari, M. Kaboli, L. Lorenzelli, G. Cheng, R. Dahiya, New materials and advances
in making electronic skin for interactive robots. Adv. Robot. 29, 1359–1373
synaptic device can be made from printed metal-oxide nanowires
(2015).
[Fig. 6D (iii)] (174), and the channel of the multigate devices (as 6. R. A. Nawrocki, N. Matsuhisa, T. Yokota, T. Someya, 300-nm imperceptible, ultraflexible,
nonspiking neuron) can be made of Si nanoribbons (181) shown in and biocompatible e-skin fit with tactile sensors and organic transistors. Adv. Electron.
Fig. 6D (i and ii) (inset). The multigate device has the inputs fed from Mater. 2, 1500452 (2016).
the readout circuit from various sensors, showing a simple tactile 7. O. Ozioko, R. Dahiya, Smart tactile gloves for haptic interaction, communication
and rehabilitation. Adv. Intell. Syst. 4, 2100091 (2022).
sensing scheme capable of delivering linear summation on the flexible 8. O. Ozioko, P. Karipoth, M. Hersh, R. Dahiya, Wearable assistive tactile communication
substrate. The low-temperature processing from the printed electronics interface based on integrated touch sensors and actuators. IEEE Trans. Neural Syst.
is naturally compatible with the thermally sensitive soft substrates. Rehabil. Eng. 28, 1344–1352 (2020).

Liu et al., Sci. Robot. 7, eabl7344 (2022) 8 June 2022 11 of 15

SCIENCE ROBOTICS | REVIEW

9. M. Kaltenbrunner, T. Sekitani, J. Reeder, T. Yokota, K. Kuribara, T. Tokuhara, M. Drack, 37. E. R. Kandel, J. H. Schwartz, T. M. Jessell,S. Siegelbaum, A. J. Hudspeth, S. Mack, Principles
R. Schwödiauer, I. Graz, S. Bauer-Gogonea, S. Bauer, T. Someya, An ultra-lightweight of Neural Science (McGraw-Hill, 2000), vol. 4.
design for imperceptible plastic electronics. Nature 499, 458–463 (2013). 38. V. E. Abraira, D. D. Ginty, The sensory neurons of touch. Neuron 79, 618–639 (2013).
10. P. Karipoth, A. Christou, A. Pullanchiyodan, R. Dahiya, Bioinspired inchworm 39. M. W. Matlin, H. J. Foley, Sensation and Perception (Allyn & Bacon, ed. 3, 1992).
and earthworm like soft robots with intrinsic strain sensing. Adv. Intell. Syst. 4, 2100092 40. D. R. Lesniak, K. L. Marshall, S. A. Wellnitz, B. A. Jenkins, Y. Baba, M. N. Rasband,
(2022). G. J. Gerling, E. A. Lumpkin, Computation identifies structural features that govern
11. O. Ozioko, P. Karipoth, P. Escobedo, M. Ntagios, A. Pullanchiyodan, R. Dahiya, SensAct: neuronal firing properties in slowly adapting touch receptors. eLife 3, e01488 (2014).
The soft and squishy tactile sensor with integrated flexible actuator. Adv. Intell. Syst. 3, 41. C. W. Eurich, H. Schwegler, Coarse coding: Calculation of the resolution achieved by
1900145 (2021). a population of large receptive field neurons. Biol. Cybern. 76, 357–363 (1997).
12. S. Somlor, R. S. Hartanto, A. Schmitz, S. Sugano, A novel tri-axial capacitive-type skin 42. S. Chun, W. Son, H. Kim, S. K. Lim, C. Pang, C. Choi, Self-powered pressure-and
sensor. Adv. Robot. 29, 1375–1391 (2015). vibration-sensitive tactile sensors for learning technique-based neural finger skin.
13. B. Ward-Cherrier, N. Pestell, L. Cramphorn, B. Winstone, M. E. Giannaccini, J. Rossiter, Nano Lett. 19, 3305–3312 (2019).
N. F. Lepora, The tactip family: Soft optical tactile sensors with 3d-printed biomimetic 43. K. Y. Chun, Y. J. Son, E. S. Jeon, S. Lee, C. S. Han, A self-powered sensor mimicking
morphologies. Soft Robot. 5, 216–227 (2018). slow- and fast-adapting cutaneous mechanoreceptors. Adv. Mater. 30, 1706299 (2018).
14. R. S. Dahiya, A. Adami, L. Pinna, C. Collini, M. Valle, L. Lorenzelli, Tactile sensing chips 44. V. Hayward, Is there a ‘plenhaptic’ function? Philos. Trans. R. Soc. Lond. B Biol. Sci. 366,
with POSFET array and integrated interface electronics. IEEE Sens. J. 14, 3448–3457 (2014). 3115–3122 (2011).
15. N. Yogeswaran, E. S. Hosseini, R. Dahiya, Graphene based low voltage field effect 45. J. A. Pruszynski, R. S. Johansson, Edge-orientation processing in first-order tactile

Downloaded from https://www.science.org at Oklahoma State University Stillwater on September 26, 2024
transistor coupled with biodegradable piezoelectric material based dynamic pressure neurons. Nat. Neurosci. 17, 1404–1409 (2014).
sensor. ACS Appl. Mater. Interfaces 12, 54035–54040 (2020). 46. S. C. Liu, T. Delbruck, G. Indiveri, A. Whatley, R. Douglas, Event-Based Neuromorphic
16. D. Hughes, N. Correll, Texture recognition and localization in amorphous robotic skin. Systems (John Wiley & Sons, 2014).
Bioinspir. Biomim. 10, 055002 (2015). 47. R. S. Johansson, J. R. Flanagan, Coding and use of tactile signals from the fingertips
17. W. Navaraj, R. Dahiya, Fingerprint-enhanced capacitive-piezoelectric flexible sensing skin in object manipulation tasks. Nat. Rev. Neurosci. 10, 345–359 (2009).
to discriminate static and dynamic tactile stimuli. Adv. Intell. Syst. 1, 1900051 (2019). 48. E. M. Izhikevich, Which model to use for cortical spiking neurons? IEEE Trans. Neural Netw.
18. Q. Li, O. Kroemer, Z. Su, F. F. Veiga, M. Kaboli, H. J. Ritter, A review of tactile information: 15, 1063–1070 (2004).
Perception and action through touch. IEEE Trans. Robot. 36, 1619–1634 (2020). 49. B. C. K. Tee, A. Chortos, A. Berndt, A. K. Nguyen, A. Tom, A. McGuire, Z. C. Lin, K. Tien,
19. G. Thuruthel Thomas, B. Shih, C. Laschi, T. Tolley Michael, Soft robot perception using W. G. Bae, H. Wang, P. Mei, H. H. Chou, B. Cui, K. Deisseroth, T. N. Ng, Z. Bao, A
embedded soft sensors and recurrent neural networks. Sci. Robot. 4, eaav1488 (2019). skin-inspired organic digital mechanoreceptor. Science 350, 313–316 (2015).
20. T. Hellebrekers, N. Chang, K. Chin, M. J. Ford, O. Kroemer, C. Majidi, Soft magnetic tactile 50. Y. Kim, A. Chortos, W. Xu, Y. Liu, J. Y. Oh, D. Son, J. Kang, A. M. Foudeh, C. Zhu, Y. Lee,
skin for continuous force and location estimation using neural networks. IEEE Robot. S. Niu, J. Liu, R. Pfattner, Z. Bao, T. W. Lee, A bioinspired flexible organic artificial afferent
Autom. Lett. 5, 3892–3898 (2020). nerve. Science 360, 998–1003 (2018).
21. M. Kaboli, D. Feng, G. Cheng, Active tactile transfer learning for object discrimination 51. Y. Wu, Y. Liu, Y. Zhou, Q. Man, C. Hu, W. Asghar, F. Li, Z. Yu, J. Shang, G. Liu, M. Liao,
in an unstructured environment using multimodal robotic skin. Int. J. Humanoid Robot. R.-W. Li, A skin-inspired tactile sensor for smart prosthetics. Sci. Robot. 3, eaat0429 (2018).
15, 1850001 (2017). 52. S. Kim, Y. Lee, H.-D. Kim, S.-J. Choi, A tactile sensor system with sensory neurons
22. P. A. Merolla, J. V. Arthur, R. Alvarez-Icaza, A. S. Cassidy, J. Sawada, F. Akopyan, and a perceptual synaptic network based on semivolatile carbon nanotube transistors.
B. L. Jackson, N. Imam, C. Guo, Y. Nakamura, B. Brezzo, I. Vo, S. K. Esser, R. Appuswamy, NPG Asia Mater. 12, 76 (2020).
B. Taba, A. Amir, M. D. Flickner, W. P. Risk, R. Manohar, D. S. Modha, A million spiking- 53. S. Caviglia, L. Pinna, M. Valle, C. Bartolozzi, Spike-based readout of POSFET tactile sensors.
neuron integrated circuit with a scalable communication network and interface. Science IEEE Trans. Circuits Syst. I Regul. Pap. 64, 1421–1431 (2017).
345, 668–673 (2014). 54. T. Birkoben, H. Winterfeld, S. Fichtner, A. Petraru, H. Kohlstedt, A spiking and adapting
23. C. S. Thakur, J. L. Molin, G. Cauwenberghs, G. Indiveri, K. Kumar, N. Qiao, J. Schemmel, tactile sensor for neuromorphic applications. Sci. Rep. 10, 17260 (2020).
R. Wang, E. Chicca, J. Olson Hasler, J. S. Seo, S. Yu, Y. Cao, A. van Schaik, R. Etienne-Cummings, 55. G. Bahuguna, V. S. Adhikary, R. K. Sharma, R. Gupta, Ultrasensitive organic humidity
Large-scale neuromorphic spiking array processors: A quest to mimic the brain. Front. sensor with high specificity for healthcare applications. Electroanalysis 32, 76–85
Neurosci. 12, 891 (2018). (2020).
24. J. Seo, B. Brezzo, Y. Liu, B. C. Parker, S.K. Esser, R. K. Montoye, B. Rajendran, J. Tierno, 56. M. Teng, Z. Zhou, Y. Qin, Y. Zhao, C. Zhao, J. Cao, A water-soluble fluorescence sensor
L. Chang, D. S. Modha, D. J. Friedman, A 45nm CMOS neuromorphic chip with a scalable with high specificity for detecting hydrazine in river water detection and A549 cell
architecture for learning in networks of spiking neurons, in 2011 IEEE Custom Integrated imaging. Sens. Actuators B Chem. 311, 127914 (2020).
Circuits Conference (CICC) (IEEE, 2011), pp. 1–4. 57. L. Rivas Yepes, E. Demir, J. Y. Lee, R. Sun, M. Smuck, I. E. Araci, Skin mountable capillaric
25. M. L. Hammock, A. Chortos, B. C. K. Tee, J. B. H. Tok, Z. Bao, 25th anniversary article: strain sensor with ultrahigh sensitivity and direction specificity. Adv. Mater. Technol. 5,
The evolution of electronic skin (E-skin): A brief history, design considerations, and recent 2000631 (2020).
progress. Adv. Mater. 25, 5997–6038 (2013). 58. S. Wang, M. Z. Hossain, K. Shinozuka, N. Shimizu, S. Kitada, T. Suzuki, R. Ichige, A. Kuwana,
26. A. Chortos, Z. Bao, Skin-inspired electronic devices. Mater. Today 17, 321–331 (2014). H. Kobayashi, Graphene field-effect transistor biosensor for detection of biotin
with ultrahigh sensitivity and specificity. Biosens. Bioelectron. 165, 112363 (2020).
27. K. Takei, W. Gao, C. Wang, A. Javey, Physical and chemical sensing with electronic skin.
59. L. Zhang, G. Wang, D. Wu, C. Xiong, L. Zheng, Y. Ding, H. Lu, G. Zhang, L. Qiu, Highly
Proc. IEEE 107, 2155–2167 (2019).
selective and sensitive sensor based on an organic electrochemical transistor
28. X. Wang, L. Dong, H. Zhang, R. Yu, C. Pan, Z. L. Wang, Recent progress in electronic skin.
for the detection of ascorbic acid. Biosens. Bioelectron. 100, 235–241 (2018).
Adv. Sci. 2, 1500169 (2015).
29. B. Shih, D. Shah, J. Li, T. G. Thuruthel, Y.-L. Park, F. Iida, Z. Bao, R. Kramer-Bottiglio, 60. W. Gerstner, W. M. Kistler, Spiking Neuron Models: Single Neurons, Populations, Plasticity
M. T. Tolley, Electronic skins and machine learning for intelligent soft robots. Sci. Robot. 5, (Cambridge Univ. Press, 2012).
eaaz9239 (2020). 61. A. L. Hodgkin, A. F. Huxley, A quantitative description of membrane current and its
30. J. C. Yang, J. Mun, S. Y. Kwon, S. Park, Z. Bao, S. Park, Electronic skin: Recent progress application to conduction and excitation in nerve. J. Physiol. 117, 500–544 (1952).
and future prospects for skin-attachable devices for health monitoring, robotics, 62. E. M. Izhikevich, Simple model of spiking neurons. IEEE Trans. Neural Netw. 14, 1569–1572
and prosthetics. Adv. Mater. 31, 1904765 (2019). (2003).
31. S. Luo, J. Bimbo, R. Dahiya, H. Liu, Robotic tactile perception of object properties: 63. N. Brunel, M. C. Van Rossum, Quantitative investigations of electrical nerve excitation
A review. Mechatronics 48, 54–67 (2017). treated as polarization. Biol. Cybern. 97, 341–349 (2007).
32. R. Mukherjee, P. Ganguly, R. Dahiya, Bioinspired distributed energy in robotics 64. W. Yi, K. K. Tsang, S. K. Lam, X. Bai, J. A. Crowell, E. A. Flores, Biological plausibility
and enabling technologies. Adv. Intell. Syst. 2021, 2100036 (2021). and stochasticity in scalable VO2 active memristor neurons. Nat. Commun. 9, 4661
33. M. Soni, R. Dahiya, Soft eSkin: Distributed touch sensing with harmonized energy (2018).
and computing. Philos. Trans. R. Soc. A 378, 20190156 (2020). 65. H. Liu, T. Wu, X. Yan, J. Wu, N. Wang, Z. du, H. Yang, B. Chen, Z. Zhang, F. Liu, W. Wu,
34. A. Chortos, J. Liu, Z. Bao, Pursuing prosthetic electronic skin. Nat. Mater. 15, 937–950 (2016). J. Guo, H. Wang, A tantalum disulfide charge-density-wave stochastic artificial neuron
35. M. Wang, Y. Luo, T. Wang, C. Wan, L. Pan, S. Pan, K. He, A. Neo, X. Chen, Artificial skin for emulating neural statistical properties. Nano Lett. 21, 3465–3472 (2021).
perception. Adv. Mater. 33, 2003014 (2021). 66. P. Stoliar, J. Tranchant, B. Corraze, E. Janod, M. P. Besland, F. Tesler, M. Rozenberg, L. Cario,
36. P. Li, H. P. Anwar Ali, W. Cheng, J. Yang, B. C. Tee, Bioinspired prosthetic interfaces. A leaky-integrate-and-fire neuron analog realized with a Mott Insulator. Adv. Funct. Mater.
Adv. Mater. Technol. 5, 1900856 (2020). 27, 1604740 (2017).

Liu et al., Sci. Robot. 7, eabl7344 (2022) 8 June 2022 12 of 15

SCIENCE ROBOTICS | REVIEW

67. S. M. Bohaichuk, S. Kumar, G. Pitner, C. J. McClellan, J. Jeong, M. G. Samant, H. S. P. Wong, 95. M. Prezioso, M. R. Mahmoodi, F. M. Bayat, H. Nili, H. Kim, A. Vincent, D. B. Strukov,
S. S. P. Parkin, R. S. Williams, E. Pop, Fast spiking of a mott VO2–Carbon nanotube Spike-timing-dependent plasticity learning of coincidence detection with passively
composite device. Nano Lett. 19, 6751–6755 (2019). integrated memristive circuits. Nat. Commun. 9, 5311 (2018).
68. Z. Wang, S. Joshi, S. Savel’ev, W. Song, R. Midya, Y. Li, M. Rao, P. Yan, S. Asapu, Y. Zhuo, 96. Q. Duan, Z. Jing, X. Zou, Y. Wang, K. Yang, T. Zhang, S. Wu, R. Huang, Y. Yang, Spiking
H. Jiang, P. Lin, C. Li, J. H. Yoon, N. K. Upadhyay, J. Zhang, M. Hu, J. P. Strachan, M. Barnell, neurons with spatiotemporal dynamics and gain modulation for monolithically
Q. Wu, H. Wu, R. S. Williams, Q. Xia, J. J. Yang, Fully memristive neural networks for pattern integrated memristive neural networks. Nat. Commun. 11, 3399 (2020).
classification with unsupervised learning. Nat. Electron. 1, 137–145 (2018). 97. C. Li, D. Belkin, Y. Li, P. Yan, M. Hu, N. Ge, H. Jiang, E. Montgomery, P. Lin, Z. Wang,
69. Z. Wang, M. Rao, J. W. Han, J. Zhang, P. Lin, Y. Li, C. Li, W. Song, S. Asapu, R. Midya, W. Song, J. P. Strachan, M. Barnell, Q. Wu, R. S. Williams, J. J. Yang, Q. Xia, Efficient
Y. Zhuo, H. Jiang, J. H. Yoon, N. K. Upadhyay, S. Joshi, M. Hu, J. P. Strachan, M. Barnell, and self-adaptive in-situ learning in multilayer memristor neural networks. Nat. Commun.
Q. Wu, H. Wu, Q. Qiu, R. S. Williams, Q. Xia, J. J. Yang, Capacitive neural network 9, 2385 (2018).
with neuro-transistors. Nat. Commun. 9, 3208 (2018). 98. S. K. Kim, Y. J. Jeong, P. Bidenko, H. R. Lim, Y. R. Jeon, H. Kim, Y. J. Lee, D. M. Geum,
70. C. Mead, Analog VLSI and Neural Systems (Addison-Wesley, 1989). J. H. Han, C. Choi, H. J. Kim, S. H. Kim, 3D stackable synaptic transistor for 3D integrated
71. I. Sourikopoulos, S. Hedayat, C. Loyez, F. Danneville, V. Hoel, E. Mercier, A. Cappy, A 4-fJ/ artificial neural networks. ACS Appl. Mater. Interfaces 12, 7372–7380 (2020).
spike artificial neuron in 65 nm CMOS technology. Front. Neurosci. 11, 123 (2017). 99. F. Cai, J. M. Correll, S. H. Lee, Y. Lim, V. Bothra, Z. Zhang, M. P. Flynn, W. D. Lu, A fully
72. J. H. B. Wijekoon, P. Dudek, A CMOS circuit implementation of a spiking neuron with integrated reprogrammable memristor–CMOS system for efficient multiply–accumulate
bursting and adaptation on a biological timescale, in 2009 IEEE Biomedical Circuits and operations. Nat. Electron. 2, 290–299 (2019).
Systems Conference (IEEE, 2009), pp. 193–196. 100. P. Yao, H. Wu, B. Gao, J. Tang, Q. Zhang, W. Zhang, J. J. Yang, H. Qian, Fully

Downloaded from https://www.science.org at Oklahoma State University Stillwater on September 26, 2024
73. A. van Schaik, Building blocks for electronic spiking neural networks. Neural Netw. 14, hardware-implemented memristor convolutional neural network. Nature 577,
617–628 (2001). 641–646 (2020).
74. G. Indiveri, Synaptic plasticity and spike-based computation in VLSI networks 101. S. Ambrogio, S. Balatti, V. Milo, R. Carboni, Z. Q. Wang, A. Calderoni, N. Ramaswamy,
of integrate-and-fire neurons. Neural Inf. Process. Lett. Rev. 11, 135–146 (2007). D. Ielmini, Neuromorphic learning and recognition with one-transistor-one-resistor
75. J. H. B. Wijekoon, P. Dudek, Compact silicon neuron circuit with spiking and bursting synapses and bistable metal oxide RRAM. IEEE Trans. Electron Devices 63, 1508–1515
behaviour. Neural Netw. 21, 524–534 (2008). (2016).
76. M. D. Pickett, R. Stanley Williams, Sub-100 fJ and sub-nanosecond thermally driven 102. D. Zhang, C. H. Yeh, W. Cao, K. Banerjee, 0.5T0.5R—An ultracompact RRAM cell uniquely
threshold switching in niobium oxide crosspoint nanodevices. Nanotechnology 23, enabled by van der waals heterostructures. IEEE Trans. Electron Devices 68, 2033–2040
215202 (2012). (2021).
77. S. Oh, Y. Shi, J. del Valle, P. Salev, Y. Lu, Z. Huang, Y. Kalcheim, I. K. Schuller, D. Kuzum, 103. H. Abbas, Y. Abbas, G. Hassan, A. S. Sokolov, Y. R. Jeon, B. Ku, C. J. Kang, C. Choi, The
Energy-efficient Mott activation neuron for full-hardware implementation of neural coexistence of threshold and memory switching characteristics of ALD HfO2 memristor
networks. Nat. Nanotechnol. 16, 680–687 (2021). synaptic arrays for energy-efficient neuromorphic computing. Nanoscale 12,
78. W. Taube Navaraj, C. García Núñez, D. Shakthivel, V. Vinciguerra, F. Labeau, D. H. Gregory, 14120–14134 (2020).
R. Dahiya, Nanowire FET based neural element for robotic tactile sensing skin. Front. 104. S. Yu, Neuro-inspired computing with emerging nonvolatile memorys. Proc. IEEE 106,
Neurosci. 11, 501 (2017). 260–285 (2018).
79. H. Jeong, L. Shi, Memristor devices for neural networks. J. Phys. D 52, 023003 (2019). 105. S. Dai, Y. Zhao, Y. Wang, J. Zhang, L. Fang, S. Jin, Y. Shao, J. Huang, Recent advances
80. S. Choi, J. Yang, G. Wang, Emerging memristive artificial synapses and neurons in transistor-based artificial synapses. Adv. Funct. Mater. 29, 1903700 (2019).
for energy-efficient neuromorphic computing. Adv. Mater. 32, 2004659 (2020). 106. S. Wang, C. Chen, Z. Yu, Y. He, X. Chen, Q. Wan, Y. Shi, D. W. Zhang, H. Zhou, X. Wang,
81. X. Pan, T. Jin, J. Gao, C. Han, Y. Shi, W. Chen, Stimuli-enabled artificial synapses P. Zhou, A MoS2/PTCDA hybrid heterojunction synapse with efficient photoelectric dual
for neuromorphic perception: Progress and perspectives. Small 16, 2001504 (2020). modulation and versatility. Adv. Mater. 31, 1806227 (2019).
82. F. Chen, Y. Zhou, Y. Zhu, R. Zhu, P. Guan, J. Fan, L. Zhou, N. Valanoor, F. von Wegner, 107. H. Ling, D. A. Koutsouras, S. Kazemzadeh, Y. van de Burgt, F. Yan, P. Gkoupidenis,
E. Saribatir, I. Birznieks, T. Wan, D. Chu, Recent progress in artificial synaptic devices: Electrolyte-gated transistors for synaptic electronics, neuromorphic computing,
Materials, processing and applications. J. Mater. Chem. 9, 8372–8394 (2021). and adaptable biointerfacing. Appl. Phys. Rev. 7, 011307 (2020).
83. Q. Xia, J. J. Yang, Memristive crossbar arrays for brain-inspired computing. Nat. Mater. 18, 108. L. Van Tho, K.-J. Baeg, Y.-Y. Noh, Organic nano-floating-gate transistor memory
309–323 (2019). with metal nanoparticles. Nano Converg. 3, 10 (2016).
84. S. G. Kim, J. S. Han, H. Kim, S. Y. Kim, H. W. Jang, Recent advances in memristive materials 109. J. Y. Gerasimov, R. Gabrielsson, R. Forchheimer, E. Stavrinidou, D. T. Simon, M. Berggren,
for artificial synapses. Adv. Mater. Technol. 3, 1800457 (2018). S. Fabiano, An evolvable organic electrochemical transistor for neuromorphic
85. F. Zahoor, T. Z. Azni Zulkifli, F. A. Khanday, Resistive random access memory (RRAM):
applications. Adv. Sci. 6, 1801339 (2019).
An overview of materials, switching mechanism, performance, multilevel cell (mlc)
110. S. Bhattacharjee, R. Wigchering, H. G. Manning, J. J. Boland, P. K. Hurley, Emulating
storage, modeling, and applications. Nanoscale Res. Lett. 15, 90 (2020).
synaptic response in n- and p-channel MoS2 transistors by utilizing charge trapping
86. W. Zhang, R. Mazzarello, M. Wuttig, E. Ma, Designing crystallization in phase-change
dynamics. Sci. Rep. 10, 12178 (2020).
materials for universal memory and neuro-inspired computing. Nat. Rev. Mater. 4,
111. S. Dai, X. Wu, D. Liu, Y. Chu, K. Wang, B. Yang, J. Huang, Light-stimulated synaptic devices
150–168 (2019).
utilizing interfacial effect of organic field-effect transistors. ACS Appl. Mater. Interfaces 10,
87. A. Athmanathan, M. Stanisavljevic, N. Papandreou, H. Pozidis, E. Eleftheriou, Multilevel-
21472–21480 (2018).
cell phase-change memory: A viable technology. IEEE J. Emerg. Sel. Top. Circuits Syst. 6,
112. H. K. Li, T. P. Chen, P. Liu, S. G. Hu, Y. Liu, Q. Zhang, P. S. Lee, A light-stimulated synaptic
87–100 (2016).
transistor with synaptic plasticity and memory functions based on InGaZnOx–Al2O3 thin
88. S. W. Fong, C. M. Neumann, H. P. Wong, Phase-change memory—Towards a storage-class film structure. J. Appl. Phys. 119, 244505 (2016).
memory. IEEE Trans. Electron Devices 64, 4374–4385 (2017). 113. J. Yu, G. Gao, J. Huang, X. Yang, J. Han, H. Zhang, Y. Chen, C. Zhao, Q. Sun, Z. L. Wang,
89. D. Apalkov, B. Dieny, J. Slaughter, Magnetoresistive random access memory. Proc. IEEE Contact-electrification-activated artificial afferents at femtojoule energy. Nat. Commun.
104, 1796–1830 (2016). 12, 1581 (2021).
90. K. Lee, S. Jang, K. L. Kim, M. Koo, C. Park, S. Lee, J. Lee, G. Wang, C. Park, Artificially 114. Y. Zang, H. Shen, D. Huang, C.-A. Di, D. Zhu, A dual-organic-transistor-based
intelligent tactile ferroelectric skin. Adv. Sci. 7, 2001662 (2020). tactile-perception system with signal-processing functionality. Adv. Mater. 29, 1606088
91. Y. Li, K.-W. Ang, Hardware implementation of neuromorphic computing using large-scale (2017).
memristor crossbar arrays. Adv. Intell. Syst. 3, 2000137 (2021). 115. K. Berggren, Q. Xia, K. K. Likharev, D. B. Strukov, H. Jiang, T. Mikolajick, D. Querlioz,
92. S. Saïghi, C. G. Mayr, T. Serrano-Gotarredona, H. Schmidt, G. Lecerf, J. Tomas, J. Grollier, M. Salinga, J. R. Erickson, S. Pi, F. Xiong, P. Lin, C. Li, Y. Chen, S. Xiong, B. D. Hoskins,
S. Boyn, A. F. Vincent, D. Querlioz, S. L. Barbera, F. Alibart, D. Vuillaume, O. Bichler, M. W. Daniels, A. Madhavan, J. A. Liddle, J. J. McClelland, Y. Yang, J. Rupp,
C. Gamrat, B. Linares-Barranco, Plasticity in memristive devices for spiking neural S. S. Nonnenmann, K.-T. Cheng, N. Gong, M. A. Lastras-Montaño, A. A. Talin, A. Salleo,
networks. Front. Neurosci. 9, 51 (2015). B. J. Shastri, T. F. de Lima, P. Prucnal, A. N. Tait, Y. Shen, H. Meng, C. Roques-Carmes,
93. N. Du, X. Zhao, Z. Chen, B. Choubey, M. D. Ventra, I. Skorupa, D. Bürger, H. Schmidt, Z. Cheng, H. Bhaskaran, D. Jariwala, H. Wang, J. M. Shainline, K. Segall, J. J. Yang, K. Roy,
Synaptic plasticity in memristive artificial synapses and their robustness against noisy S. Datta, A. Raychowdhury, Roadmap on emerging hardware and technology for machine
inputs. Front. Neurosci. 15, 696 (2021). learning. Nanotechnology 32, 012002 (2021).
94. S. Kim, C. du, P. Sheridan, W. Ma, S. H. Choi, W. D. Lu, Experimental demonstration 116. H. Wan, Y. Cao, L. W. Lo, J. Zhao, N. Sepúlveda, C. Wang, Flexible carbon nanotube
of a second-order memristor and its ability to biorealistically implement synaptic synaptic transistor for neurological electronic skin applications. ACS Nano 14,
plasticity. Nano Lett. 15, 2203–2211 (2015). 10402–10412 (2020).

Liu et al., Sci. Robot. 7, eabl7344 (2022) 8 June 2022 13 of 15

SCIENCE ROBOTICS | REVIEW

117. H. Shim, K. Sim, F. Ershad, P. Yang, A. Thukral, Z. Rao, H.-J. Kim, Y. Liu, X. Wang, G. Gu, 143. E. Painkras, L. A. Plana, J. Garside, S. Temple, F. Galluppi, C. Patterson, D. R. Lester,
L. Gao, X. Wang, Y. Chai, C. Yu, Stretchable elastic synaptic transistors for neurologically A. D. Brown, S. B. Furber, SpiNNaker: A 1-W 18-core system-on-chip for massively-parallel
integrated soft engineering systems. Sci. Adv. 5, eaax4961 (2019). neural network simulation. IEEE J. Solid-State Circuits 48, 1943–1953 (2013).
118. F. Molina-Lopez, T. Z. Gao, U. Kraft, C. Zhu, T. Öhlund, R. Pfattner, V. R. Feig, Y. Kim, 144. W. W. Lee, Y. J. Tan, H. Yao, S. Li, H. H. See, M. Hon, K. A. Ng, B. Xiong, J. S. Ho, B. C. K. Tee,
S. Wang, Y. Yun, Z. Bao, Inkjet-printed stretchable and low voltage synaptic transistor A neuro-inspired artificial peripheral nervous system for scalable electronic skins. Sci. Robot.
array. Nat. Commun. 10, 2676 (2019). 4, eaax2198 (2019).
119. X. Feng, Y. Li, L. Wang, S. Chen, Z. G. Yu, W. C. Tan, N. Macadam, G. Hu, L. Huang, L. Chen, 145. C. Bader, F. Bergner, G. Cheng, A robust and efficient dynamic network protocol for a
X. Gong, D. Chi, T. Hasan, A. V. Y. Thean, Y. W. Zhang, K. W. Ang, A fully printed flexible large-scale artificial robotic skin, in 2018 IEEE/RSJ International Conference on Intelligent
MoS2 memristive artificial synapse with femtojoule switching energy. Adv. Electron. Robots and Systems (IROS) (IEEE, 2018), pp. 1600–1605.
Mater. 5, 1900740 (2019). 146. Y. Ohmura, Y. Kuniyoshi, Humanoid robot which can lift a 30kg box by whole body
120. X. Yan, Z. Zhou, J. Zhao, Q. Liu, H. Wang, G. Yuan, J. Chen, Flexible memristors contact and tactile feedback, in 2007 IEEE/RSJ International Conference on Intelligent
as electronic synapses for neuro-inspired computation based on scotch tape-exfoliated Robots and Systems (IEEE, 2007), pp. 1136–1141.
mica substrates. Nano Res. 11, 1183–1192 (2018). 147. G. Cheng, E. Dean-Leon, F. Bergner, J. Rogelio Guadarrama Olvera, Q. Leboutet,
121. C. Wu, T. W. Kim, H. Y. Choi, D. B. Strukov, J. J. Yang, Flexible three-dimensional artificial P. Mittendorfer, A comprehensive realization of robot skin: Sensors, sensing, control,
synapse networks with correlated learning and trainable memory capability. and applications. Proc. IEEE 107, 2034–2051 (2019).
Nat. Commun. 8, 752 (2017). 148. W. T. Navaraj, S. Gupta, L. Lorenzelli, R. Dahiya, Wafer scale transfer of ultrathin silicon
122. W. Huh, D. Lee, C.-H. Lee, Memristors based on 2D materials as an artificial synapse chips on flexible substrates for high performance bendable systems. Adv. Electron. Mater.

Downloaded from https://www.science.org at Oklahoma State University Stillwater on September 26, 2024
for neuromorphic electronics. Adv. Mater. 32, 2002092 (2020). 4, 1700277 (2018).
123. Y. Lee, J. Y. Oh, W. Xu, O. Kim, T. R. Kim, J. Kang, Y. Kim, D. Son, J. B.-H. Tok, M. J. Park, 149. S. Gupta, W. T. Navaraj, L. Lorenzelli, R. Dahiya, Ultra-thin chips for high-performance
Z. Bao, T.-W. Lee, Stretchable organic optoelectronic sensorimotor synapse. Sci. Adv. 4, flexible electronics. npj Flex. Electron. 2, 8 (2018).
eaat7387 (2018). 150. A. Vilouras, A. Christou, L. Manjakkal, R. Dahiya, Ultrathin ion-sensitive field-effect
124. T. Okatani, H. Takahashi, T. Takahata, I. Shimoyama, Evaluation of ground slippery transistor chips with bending-induced performance enhancement. ACS Appl. Electron.
condition during walk of bipedal robot using MEMS slip sensor, in 2017 IEEE 30th Mater. 2, 2601–2610 (2020).
International Conference on Micro Electro Mechanical Systems (MEMS) (IEEE, 2017), 151. A. S. Dahiya, D. Shakthivel, Y. Kumaresan, A. Zumeit, A. Christou, R. Dahiya, High-
pp. 1033–1035. performance printed electronics based on inorganic semiconducting nano to chip scale
125. A. G. Eguíluz, I. Rañó, S. A. Coleman, T. M. McGinnity, Reliable robotic handovers through structures. Nano Converg. 7, 33 (2020).
tactile sensing. Auton. Robot. 43, 1623–1637 (2019). 152. T. Seifert, M. Baum, F. Roscher, M. Wiemer, T. Gessner, Aerosol jet printing of nano particle
126. S. J. Huston, V. Jayaraman, Studying sensorimotor integration in insects. Curr. Opin. based electrical chip interconnects. Mater. Today Proc. 2, 4262–4271 (2015).
Neurobiol. 21, 527–534 (2011). 153. J. Wang, R. C. Y. Auyeung, H. Kim, N. A. Charipar, A. Piqué, Three-dimensional printing
127. I. Krauhausen, D. A. Koutsouras, A. Melianas, S. T. Keene, K. Lieberth, H. Ledanseur, of interconnects by laser direct-write of silver nanopastes. Adv. Mater. 22, 4462–4466
R. Sheelamanthula, A. Giovannitti, F. Torricelli, I. Mcculloch, P. W. M. Blom, A. Salleo, (2010).
Y. van de Burgt, P. Gkoupidenis, Organic neuromorphic electronics for sensorimotor 154. H.-L. Kang, S. M. Sim, Y. Lee, J. H. Yu, K. Y. Shin, S. H. Lee, J. M. Kim, Flip chip bonding using
integration and learning in robotics. Sci. Adv. 7, eabl5068 (2021). ink-jet printing technology. Microsyst. Technol. 25, 4753–4759 (2019).
128. F. Liu, S. Deswal, A. Christou, M. S. Baghini, R. Chirila, D. Shakthivel, M. Chakraborty, 155. H. Y. Zhang, X. Li, H. N. Jiang, Y. H. Mei, G. Q. Lu, Large-area substrate bonding with single-
R. Dahiya, Printed synaptic transistors based electronic skin for robots to feel and learn. printing silver paste sintering for power modules. IEEE Trans. Compon. Packag. Manuf.
Sci. Robot. 7, eabl7286 (2022). Technol. 11, 11–18 (2021).
129. A. Zylberberg, D. Fernández Slezak, P. R. Roelfsema, S. Dehaene, M. Sigman, The brain's 156. S. Ma, Y. Kumaresan, A. S. Dahiya, R. Dahiya, Ultra-thin chips with printed interconnects
router: A cortical network model of serial processing in the primate brain. PLOS Comput. on flexible foils. Adv. Electron. Mater. 8, 2101029 (2021).
Biol. 6, e1000765 (2010). 157. Y.-G. Park, H. Min, H. Kim, A. Zhexembekova, C. Y. Lee, J. U. Park, Three-dimensional,
130. H. P. Saal, B. P. Delhaye, B. C. Rayhaun, S. J. Bensmaia, Simulating tactile signals high-resolution printing of carbon nanotube/liquid metal composites with mechanical
from the whole hand with millisecond precision. Proc. Natl. Acad. Sci. U.S.A. 114, and electrical reinforcement. Nano Lett. 19, 4866–4872 (2019).
E5693–E5702 (2017). 158. A. Carlson, A. M. Bowen, Y. Huang, R. G. Nuzzo, J. A. Rogers, Transfer printing techniques
131. F. Vega-Bermudez, K. Johnson, SA1 and RA receptive fields, response variability, for materials assembly and micro/nanodevice fabrication. Adv. Mater. 24, 5284–5318
and population responses mapped with a probe array. J. Neurophysiol. 81, 2701–2710 (2012).
(1999). 159. A. Zumeit, A. Dahiya, A. Christou, D. Shakthivel, R. Dahiya, Direct roll transfer printed
132. Z. Wang, C. Li, P. Lin, M. Rao, Y. Nie, W. Song, Q. Qiu, Y. Li, P. Yan, J. P. Strachan, N. Ge, silicon nanoribbon arrays based high-performance flexible electronics. npj Flex. Electron.
N. McDonald, Q. Wu, M. Hu, H. Wu, R. S. Williams, Q. Xia, J. J. Yang, In situ training 5, 18 (2021).
of feed-forward and recurrent convolutional memristor networks. Nat. Mach. Intell. 1, 160. T. Sekitani, U. Zschieschang, H. Klauk, T. Someya, Flexible organic transistors and circuits
434–442 (2019). with extreme bending stability. Nat. Mater. 9, 1015–1022 (2010).
161. E. S. Hosseini, S. Dervin, P. Ganguly, R. Dahiya, Biodegradable materials for sustainable
133. M. Prezioso, F. Merrikh-Bayat, B. D. Hoskins, G. C. Adam, K. K. Likharev, D. B. Strukov,
health monitoring devices. ACS Appl. Bio Mater. 4, 163–194 (2021).
Training and operation of an integrated neuromorphic network based on metal-oxide
162. J. K. Chang, H. P. Chang, Q. Guo, J. Koo, C. I. Wu, J. A. Rogers, Biodegradable electronic
memristors. Nature 521, 61–64 (2015).
systems in 3D, heterogeneously integrated formats. Adv. Mater. 30, 1704955 (2018).
134. K. Buchanan, J. Mellor, The activity requirements for spike timing-dependent plasticity
in the hippocampus. Front. Synaptic Neurosci. 2, 11 (2010). 163. M. Chakraborty, J. Kettle, R. Dahiya, Electronic waste reduction through devices
135. H. Shouval, S. Wang, G. Wittenberg, Spike timing dependent plasticity: A consequence and printed circuit boards designed for circularity. IEEE J. Flex. Electron. 1, 4–23 (2022).
of more fundamental learning rules. Front. Comput. Neurosci. 4, 19 (2010). 164. J. A. Rogers, Z. Bao, Printed plastic electronics and paperlike displays. J. Polym. Sci. A
136. L. Zhao, Q. Hong, X. Wang, Novel designs of spiking neuron circuit and STDP learning Polym. Chem. 40, 3327–3334 (2002).
circuit based on memristor. Neurocomputing 314, 207–214 (2018). 165. U. Kraft, F. Molina-Lopez, D. Son, Z. Bao, B. Murmann, Ink development and printing
137. Z. Wang, C. Li, W. Song, M. Rao, D. Belkin, Y. Li, P. Yan, H. Jiang, P. Lin, M. Hu, J. P. Strachan, of conducting polymers for intrinsically stretchable interconnects and circuits. Adv.
N. Ge, M. Barnell, Q. Wu, A. G. Barto, Q. Qiu, R. S. Williams, Q. Xia, J. J. Yang, Reinforcement Electron. Mater. 6, 1900681 (2020).
learning with analogue memristor arrays. Nat. Electron. 2, 115–124 (2019). 166. C. Kotlowski, P. Aspermair, H. U. Khan, C. Reiner-Rozman, J. Breu, S. Szunerits, J. J. Kim,
138. J. N. J. Reynolds, B. I. Hyland, J. R. Wickens, A cellular mechanism of reward-related Z. Bao, C. Kleber, P. Pelosi, W. Knoll, Electronic biosensing with flexible organic transistor
learning. Nature 413, 67–70 (2001). devices. Flex. Print. Electron. 3, 034003 (2018).
139. V. Pawlak, J. N. D. Kerr, Dopamine receptor activation is required for corticostriatal 167. T.-H. Kim, A. Carlson, J. H. Ahn, S. M. Won, S. Wang, Y. Huang, J. A. Rogers, Kinetically
spike-timing-dependent plasticity. J. Neurosci. 28, 2435–2446 (2008). controlled, adhesiveless transfer printing using microstructured stamps. Appl. Phys. Lett.
140. M. Mikaitis, G. Pineda García, J. C. Knight, S. B. Furber, Neuromodulated synaptic 94, 113502 (2009).
plasticity on the SpiNNaker neuromorphic system. Front. Neurosci. 12, 105 (2018). 168. C. Linghu, S. Zhang, C. Wang, J. Song, Transfer printing techniques for flexible
141. R. S. Sutton, A. G. Barto, Introduction to Reinforcement Learning (MIT Press, ed. 1, 1998). and stretchable inorganic electronics. npj Flex. Electron 2, 26 (2018).
142. C. Shen, X. Gao, C. Chen, S. Ren, J.-L. Xu, Y.-D. Xia, S.-D. Wang, ZnO nanowire 169. M. A. Meitl, Z. T. Zhu, V. Kumar, K. J. Lee, X. Feng, Y. Y. Huang, I. Adesida, R. G. Nuzzo,
optoelectronic synapse for neuromorphic computing. Nanotechnology 33, 065205 J. A. Rogers, Transfer printing by kinetic control of adhesion to an elastomeric stamp.
(2022). Nat. Mater. 5, 33–38 (2006).

Liu et al., Sci. Robot. 7, eabl7344 (2022) 8 June 2022 14 of 15

SCIENCE ROBOTICS | REVIEW

170. R. S. Cok, J. W. Hamer, C. A. Bower, E. Menard, S. Bonafede, AMOLED displays 191. K. Roy, A. Jaiswal, P. Panda, Towards spike-based machine intelligence
with transfer-printed integrated circuits. J. Soc. Inf. Disp. 19, 335–341 (2011). with neuromorphic computing. Nature 575, 607–617 (2019).
171. Z. Fan, J. C. Ho, Z. A. Jacobson, R. Yerushalmi, R. L. Alley, H. Razavi, A. Javey, Wafer-scale 192. C. G. Núñez, W. T. Navaraj, E. O. Polat, R. Dahiya, Energy-autonomous, flexible,
assembly of highly ordered semiconductor nanowire arrays by contact printing. Nano and transparent tactile skin. Adv. Funct. Mater. 27, 1606287 (2017).
Lett. 8, 20–25 (2008). 193. A. Vilouras, H. Heidari, S. Gupta, R. Dahiya, Modeling of CMOS devices and circuits
172. T. Takahashi, K. Takei, J. C. Ho, Y. L. Chueh, Z. Fan, A. Javey, Monolayer resist on flexible ultrathin chips. IEEE Trans. Electron Devices 64, 2038–2046 (2017).
for patterned contact printing of aligned nanowire arrays. J. Am. Chem. Soc. 131, 194. M. Boniol, J.-P. Verriest, R. Pedeux, J.-F. Doré, Proportion of skin surface area of children
2102–2103 (2009). and young adults from 2 to 18 years old. J. Investig. Dermatol. 128, 461–464 (2008).
173. D. Roßkopf, S. Strehle, Surface-controlled contact printing for nanowire device 195. R. S. Johansson, A. B. Vallbo, Tactile sensibility in the human hand: Relative and absolute
fabrication on a large scale. Nanotechnology 27, 185301 (2016). densities of four types of mechanoreceptive units in glabrous skin. J. Physiol. 286,
174. A. Christou, F. Liu, R. Dahiya, Development of a highly controlled system for large-area, 283–300 (1979).
directional printing of quasi-1D nanomaterials. Microsyst. Nanoeng. 7, 82 (2021). 196. C. Yan, J. Wang, W. Kang, M. Cui, X. Wang, C. Y. Foo, K. J. Chee, P. S. Lee, Highly stretchable
175. J. Suikkola, T. Björninen, M. Mosallaei, T. Kankkunen, P. Iso-Ketola, L. Ukkonen, J. Vanhala, piezoresistive graphene–Nnanocellulose nanopaper for strain sensors. Adv. Mater. 26,
M. Mäntysalo, Screen-printing fabrication and characterization of stretchable electronics. 2022–2027 (2014).
Sci. Rep. 6, 25784 (2016). 197. S. Luo, T. Liu, SWCNT/graphite nanoplatelet hybrid thin films for self-temperature-
176. W. J. Hyun, E. B. Secor, M. C. Hersam, C. D. Frisbie, L. F. Francis, High-resolution patterning compensated, highly sensitive, and extensible piezoresistive sensors. Adv. Mater. 25,
of graphene by screen printing with a silicon stencil for highly flexible printed electronics. 5650–5657 (2013).

Downloaded from https://www.science.org at Oklahoma State University Stillwater on September 26, 2024
Adv. Mater. 27, 109–115 (2015). 198. S. Yao, Y. Zhu, Wearable multifunctional sensors using printed stretchable conductors
177. P. He, J. Cao, H. Ding, C. Liu, J. Neilson, Z. Li, I. A. Kinloch, B. Derby, Screen-printing made of silver nanowires. Nanoscale 6, 2345–2352 (2014).
of a highly conductive graphene ink for flexible printed electronics. ACS Appl. Mater. 199. N. Lu, C. Lu, S. Yang, J. Rogers, Highly sensitive skin-mountable strain gauges based
Interfaces 11, 32225–32234 (2019). entirely on elastomers. Adv. Funct. Mater. 22, 4044–4050 (2012).
178. X. Cao, H. Chen, X. Gu, B. Liu, W. Wang, Y. Cao, F. Wu, C. Zhou, Screen printing 200. M. Amjadi, Y. J. Yoon, I. Park, Ultra-stretchable and skin-mountable strain sensors using
as a scalable and low-cost approach for rigid and flexible thin-film transistors using carbon nanotubes–Ecoflex nanocomposites. Nanotechnology 26, 375501 (2015).
separated carbon nanotubes. ACS Nano 8, 12769–12776 (2014). 201. J. Choi, K. Hong, 3D skin length deformation of lower body during knee joint flexion
179. R. Parashkov, E. Becker, T. Riedl, H. Johannes, W. Kowalsky, Large area electronics using for the practical application of functional sportswear. Appl. Ergon. 48, 186–201 (2015).
printing methods. Proc. IEEE 93, 1321–1329 (2005). 202. G. Corniani, H. P. Saal, Tactile innervation densities across the whole body.
180. R. Yerushalmi, Z. A. Jacobson, J. C. Ho, Z. Fan, A. Javey, Large scale, highly ordered J. Neurophysiol. 124, 1229–1240 (2020).
assembly of nanowire parallel arrays by differential roll printing. Appl. Phys. Lett. 91, 203. B. Ibrahim, J. McMurray, R. Jafari, A wrist-worn strap with an array of electrodes for robust
203104 (2007). physiological sensing, in 2018 40th Annual International Conference of the IEEE Engineering
181. A. Zumeit, W. T. Navaraj, D. Shakthivel, R. Dahiya, Nanoribbon-based flexible in Medicine and Biology Society (EMBC) (IEEE, 2018), pp. 4313–4317.
high-performance transistors fabricated at room temperature. Adv. Electron. Mater. 6, 204. F. Mancin, Whole-body mapping of spatial acuity for pain and touch. Ann. Neurol. 75,
1901023 (2020). 917–924 (2014).
182. S. Wang, J. Xu, W. Wang, G. J. N. Wang, R. Rastak, F. Molina-Lopez, J. W. Chung, S. Niu, 205. Y. Jiang, P. Huang, D. Zhu, Z. Zhou, R. Han, L. Liu, X. Liu, J. Kang, Design and hardware
V. R. Feig, J. Lopez, T. Lei, S. K. Kwon, Y. Kim, A. M. Foudeh, A. Ehrlich, A. Gasperini, Y. Yun, implementation of neuromorphic systems with RRAM synapses and threshold-controlled
B. Murmann, J. B. H. Tok, Z. Bao, Skin electronics from scalable fabrication neurons for pattern recognition. IEEE Trans. Circuits Syst. I Regul. Pap. 65, 2726–2738
of an intrinsically stretchable transistor array. Nature 555, 83–88 (2018). (2018).
183. J. Biggs, J. Myers, J. Kufel, E. Ozer, S. Craske, A. Sou, C. Ramsdale, K. Williamson, R. Price, 206. P. M. Sheridan, F. Cai, C. du, W. Ma, Z. Zhang, W. D. Lu, Sparse coding with memristor
S. White, A natively flexible 32-bit Arm microprocessor. Nature 595, 532–536 (2021). networks. Nat. Nanotechnol. 12, 784–789 (2017).
184. M. Raibert, An all digital VLSI tactile array sensor, in Proceedings of the 1984 IEEE
International Conference on Robotics and Automation (IEEE, 1984), vol. 1, pp. 314–319.
185. V. J. Lumelsky, M. S. Shur, S. Wagner, Sensitive skin. IEEE Sens. J. 1, 41–51 (2001).
186. A. Schmitz, P. Maiolino, M. Maggiali, L. Natale, G. Cannata, G. Metta, Methods Funding: This work was supported by Engineering and Physical Sciences Research Council
and technologies for the implementation of large-scale robot tactile sensors. IEEE Trans. (EPSRC) through Engineering Fellowship for Growth - neuPRINTSKIN (EP/R029644/1) and
Robot. 27, 389–400 (2011). Hetero-print Programme Grant (EP/R03480X/1). Author contributions: R.D., Y.S., and M.K.
187. P. Escobedo, M. Ntagios, D. Shakthivel, W. T. Navaraj, R. Dahiya, Energy generating conceptualized the review. F.L., S.D., A.C., and R.D. wrote the manuscript. R.D., Y.S., and M.K.
electronic skin with intrinsic tactile sensing without touch sensors. IEEE Trans. Robot. 37, guided the topics of discussion and structure of the review. All authors provided feedback and
683–690 (2021). revisions. R.D. provided overall supervision of the work. Competing interests: The authors
188. R. S. Dahiya, G. Metta, M. Valle, G. Sandini, Tactile sensing—From humans to humanoids. declare that they have no competing interests.
IEEE Trans. Robot. 26, 1–20 (2010).
189. M. D. Pickett, G. Medeiros-Ribeiro, R. S. Williams, A scalable neuristor built with Mott
memristors. Nat. Mater. 12, 114–117 (2013). Submitted 2 August 2021
190. M. E. Beck, A. Shylendra, V. K. Sangwan, S. Guo, W. A. Gaviria Rojas, H. Yoo, H. Bergeron, Accepted 13 May 2022
K. Su, A. R. Trivedi, M. C. Hersam, Spiking neurons from tunable Gaussian heterojunction Published 8 June 2022
transistors. Nat. Commun. 11, 1565 (2020). 10.1126/scirobotics.abl7344

Liu et al., Sci. Robot. 7, eabl7344 (2022) 8 June 2022 15 of 15