Recent Advances in Smart Tactile Sensory Systems with Brain‐Inspired Neural Networks

Tactile sensory systems play a vital role in various emerging fields including robotics, prosthetics, and human–machine interfaces. However, traditional tactile sensors are typically designed to detect a single stimulus through a lock‐and‐key mechanism, which poses substantial challenges in the realization of multimodal tactile sensors. To address this issue, the convergence of tactile sensory systems with artificial neural network and machine learning (ML) platforms has been utilized to enhance the capabilities of multimodal sensors and enable signal decoupling/interpretation of mixed tactile stimuli. Herein, recent progress in multimodal sensors that can simultaneously identify various stimuli such as strain, pressure, and temperature is reviewed, providing in‐depth understanding of materials, structures, and methodologies. In addition, accurate interpretation of signals from mixed tactile stimuli under complex conditions remains challenging. This review presents a comprehensive exploration of ML algorithms that mimic human neural networks, discussing their significance in advancing smart sensory systems and improving signal interpretation in complex and dynamic environments.

cross-reactive tactile sensors which that can detect strain/ pressure, temperature/pressure, and temperature/strain stimuli are under investigation.In addition, multidimensional tactile sensors capable of detecting three or more tactile stimuli are actively researched.However, in the case of complex stimuli, decoupling of signals from each stimulus is challenging because of the intermixing or interaction between the output data.One of the feasible approaches is to assemble various tactile sensors which are not influenced by other tactile stimuli.Alternatively, machine learning (ML) and artificial neural network (ANN) can be implemented to decouple the complex stimuli based on learning and training processes.In fact, using algorithms such as convolution neural network (CNN) and spike neural network (SNN), efficient and accurate decoupling of multidimensional tactile stimuli is possible, overcoming the limitations of conventional lock-and-key type tactile sensors.In this review, various ML and ANN for multidimensional tactile sensors are discussed, providing recent research trends and future aspects.

Carbon-Based Strain Sensors
Carbon allotropes are the representative materials that have been used as the sensing layer in strain sensors.[12][13] Due to the abundance of carbon, typical carbon allotropes offer the advantage of low cost.Also, they are relatively easy to synthesize, allowing them to be utilized in diverse sensing applications.[16] Carbon-based strain sensors typically detect the change of resistance upon external mechanical stretching.In particular, the CNTs are nanometer-sized fibers that have extremely high aspect ratio (>10 6 ), high Young's modulus (≈1 TPa), tensile strength of ≈100 GPa, high carrier mobility (≈80 cm 2 V À1 s À1 ), [17] and thermal conductivity of 3500 W mK À1 . [18,19]Due to their small sizes and fibril structures, macroassemblies of CNTs can take on various organizing forms, presenting numerous opportunities for realizing different functional materials and devices based on CNTs.For example, Yamada et al. [4] demonstrated CNT filmbased strain sensors in which the number of conductive path changes due to the cracks formed inside the component during stretching, enabling the strain-dependent resistance variation.Also, graphene is a two-dimensional nanosheet which exhibits extraordinary chemical, electrical, and mechanical properties, including notable flexibility and elasticity.Its unique molecular structure allows it to adapt well to the various sensing layer of strain sensors.
Based on these advantages of carbon allotropes, strain sensors with various structures have been demonstrated.Liu et al. [20] reported a nanocellulose-based carbon aerogel with ultralight weight and superhydrophilicity using bidirectional freezing and annealing method, as shown in Figure 1a.The cellulose nanofibrils/CNT/reduced graphene oxide (rGO) composite aerogel had a regularly arranged layered porous structure that exhibits ultralow density and fast ion transport.The aerogel was used as electrodes for supercapacitors, which demonstrated significant electrochemical performance including high capacitance and outstanding cycling stability.Additionally, the aerogel exhibited good linear sensitivity (5.61 kPa À1 ) as a strain sensor and showed diverse application prospects in wearable devices, providing an effective method for constructing a multifunctional platform.Wu et al. [21] also demonstrated carbon fiber membrane (CFM)-based strain sensors which were fabricated by carbonizing highly aligned GO/polyacrylonitrile (PAN) composite fiber membrane.The CFM-based flexible strain sensors exhibited a high gauge factor (GF) of 49.8 and anisotropic electromechanical behaviors.The strain sensors also exhibited unique properties such as anisotropic strain sensing owing to their highly oriented microstructure (Figure 1b).Tan et al. [22] reported a highperformance strain sensor using graphene nanoribbons embedded in thermoplastic polyurethane matrix (Figure 1c).The sensor showed high stretchability of 160%, GF of 35.7, and long-term durability, enabling the human motion monitoring.In addition, the boron nitride nanosheets in the sensor enhanced the thermal conductivity, providing outstanding thermal stability.

Metallic Nanomaterial-Based Strain Sensors
[32] For instance, ultrathin Au/Ag NWs having ≈2 nm width and aspect ratio of >10 000 were utilized to demonstrate innovative superlattice nanomembranes and stretchable transparent strain sensors, which were inspired by the intricate structures and adaptability found in nature. [33,34]Also, as a sensitive component for strain sensors, the metallic nanomaterials are beneficial because of their unique mechanoelectrical properties.Although the strain sensing range tends to be slightly smaller or comparable to other sensing materials, NWs and NPs-based strain sensors can exhibit high GFs sufficient to compensate such drawbacks. [27,32]This is mainly due to relatively dense packing of metal nanomaterials compared to other substances, allowing to detect stimuli such as microvibrations of human skin and cracks.In addition, various structures have been investigated to enhance the strain sensitivity and widen the sensing range.For example, Kang et al. [32] presented metal NW-based strain sensors that are reversible, durable, and mechanically flexible, making them easy to attach to human skin.The sensor exhibited ultrahigh mechanosensitivity (GF = 2000), owing to the disconnection and reconnection process of the zip-like nanoscale crack junctions when subjected to strain or vibration.Furthermore, for the soft elastomer, various elastomeric materials have been suggested.For example, polydimethylsiloxane (PDMS) is highly elastic, biocompatible, and has low Young's modulus.Therefore, PDMS has been most frequently used as the soft elastomer in the strain sensors. [35,36]dditionally, various NWs and NPs-type sensing materials have been investigated to improve the stability and the sensing performance. [25]Jeon et al. reported a waterproof E-bandage for wearable strain sensors with adjustable sensitivity, utilizing CNT-based composites and fluoropolymer-coated PDMS films. [35]The E-bandage could detect the subtle movements in joints and muscles.The E-bandage showed high-sensing performance (GF = ≈100) and selective sensitivity, suitable for flexible health monitoring IoT systems.Shen et al. [37] also demonstrated a transparent and flexible strain sensor using Ag NWs/poly(3,4ethylenedioxythiophene):poly(styrenesulfonate) (PEDOT:PSS) composite film on a patterned PDMS substrate derived from a near-field electrospun PAN grid (Figure 2a).The patterned PDMS surface morphology enabled stretching-induced strain detection, making the sensor more sensitive than the flat type.Also, using the Ag NWs/PEDOT:PSS film, wider sensing range (≈100%), long-term reliability, and the detection of bending and pressing-induced deformation were possible.Based on its high-sensing performance, real-time detection of various human movements was demonstrated.Zhang et al. [25] reported highly stretchable polymer/AgNWs composite-based strain sensors for accurate recognition and rapid localization of directional vibrations (Figure 2b).The addition of AgNWs improved the electrical conductivity, while the bubble-like structure enhanced the stretchability.The sensor demonstrated a wide strain sensing range up to 500%, high linearity, low sensing limit (≈1%), and excellent durability and fatigue resistance.Also, Liu et al. [31] investigated strain sensors using geometrically designed grooves in Ag NP-coated PDMS.The sensor exhibited high GF (0-0.45%,1400; 0.46-0.65%,18 000), omnidirectional sensing, and discrimination of strain direction using an array structure.

Conductive Composite-Based Strain Sensors
The development of various conductive nanomaterials [38][39][40][41][42] and stretchable elastomeric polymers [25,39,[43][44][45][46] has led to the research of composite material-based strain sensors.In particular, due to the advances in chemical and material engineering, Figure 1.a) Illustration of the assembly of solid-state symmetric supercapacitors and their strain response from 0 to 80% strain, and the durability test under 50% strain.Reproduced with permission. [20]Copyright 2022, Wiley-VCH.b) Schematic diagram and strain sensing characteristics of a strain sensor consisting of two cross-plied CFMs.Reproduced with permission. [21]Copyright 2021, Wiley-VCH.c) Schematic diagram of the stretchable sensor and the mechanism of change in thermal conductive pathways during stretching-releasing process.Reproduced with permission. [22]Copyright 2020, Nature Publishing Group.
the fusion of various material combinations has become possible, allowing a wider selection materials and structures for strain sensors.For example, conductive polymers [28,39,47] such as PEDOT:PSS, [37] polypyrrole (PPy), [38] and poly(3-hexylthiophene-2,5-diyl) nanofibrils (P3HT-NFs), [48] as well as liquid metals such as eutectic-gallium-indium (EGaIn) [42,49,50] have been actively investigated.The composite materials with adjustable properties can offer a broad application spectrum due to their tunable electrical properties which depend on the conductive material concentration and synthesis methods.The composite material-based strain sensors are typically fabricated by dispersing conductive fillers in a stretchable polymer matrix via melt compounding or solution mixing.The melt compounding involves dry blending of conductive fillers and polymer particles at certain temperatures, then hot-pressing or extrusion is followed, offering facile production scaling due to its simple fabrication process.The solution mixing process can provide composites with good conductive filler dispersion by employing ultrasound or shearing to disperse the fillers and dissolve polymers in a solvent.By casting the homogenized solution and hot curing, conductive composites can be fabricated.Gallium-based liquid metal alloys and hydrogels containing ionic liquid (IL) conductors have gained much attention owing to their good stretchability and adaptability.In particular, the liquid metal alloys have the advantages of extremely high stretchability, adaptability to deformation, low electrical hysteresis during dynamic deformation, superior thermal and electrical conductivity, and excellent contact with heterogeneous materials.
Strain sensors using composite materials as the sensing layer typically incorporate two or more materials in an elastomer simultaneously or employ a bilayer structure.Sahlberg et al. [50] presented high-resolution patterning of liquid metal alloys using a shrinking-substrate-process (SSP) technique with an elastomer substrate, overcoming the resolution limitation of conventional printing of liquid metal alloys (Figure 3a).Specifically, using the SSP method, the patterns of liquid metal alloy can be maintained with an encapsulating layer that is cured after the substrate relaxation.With the process, a 100 μm line width can be shrunk down to 25 μm, enabling the realization of high-density sensor arrays.Xu et al. [38] also developed a natural rubber-based flexible strain sensor with highly dispersed PPy/nanofiber conductive networks.The strong interfacial interactions in sensing layer enable the formation of continuous network throughout the natural rubber matrix, resulting in significantly improved conductivity and mechanical properties at only 6% of PPy loading.With a GF of 355.3 in the strain range of over 300 %, the sensor outperforms previously reported natural rubber-based stretchable strain sensors (Figure 3b).The sensor also endures over 1000 stretching and releasing cycles under various strain ranges.As possible applications for wearable motion monitoring, sensing of swallowing, finger motion, wrist bending, elbow bending, and knee bending were demonstrated.This work offers a simple and scalable approach for constructing an effective conductive network in elastomeric strain sensors for various applications.Chen et al. [49] demonstrated a highly stretchable PU/polymethacrylate/EGaIn (PPE) fiber featuring superior electrical conductivity.The fiber consists of a core fiber, an intermediate modified layer, and an EGaIn liquid metal at the outer layer.The PPE fiber boasts exceptional electrical conductivity up to 10 3 S cm À1 , and the sensor exhibited over 500% of stretchability, GF of 1.5, and excellent thermostability with a maximum operating temperature of ≈250 °C (Figure 3c).This PPE fiber can serve as both a highly conductive stretchable fiber and a strain sensor due to its high stretchability, stable resistance variation, and outstanding performance.Li et al. [51] fabricated highly stretchable and sensitive Ag@rGO@PDMS fiber strain sensors which utilized the wrinkle-guided differential microcracking mechanism in the strain-sensing Ag@rGO bilayer shell, as shown in Figure 3d.The wrinkled rGO interlayer acts as a template, guiding controlled microcracking in the Ag NP layer along the surface wrinkles.The strain sensor demonstrated high sensitivity (0-2%, GF = 420; 110-125%, GF = 1.1 Â 10 9 ), ultra-low strain detection limit (0.01%), and fast response/recovery time (0.13 ms).Additionally, the sensors can detect various acoustic waves, human motions, and physiological signals, displaying considerable potential in wearable voice recognition and healthcare monitoring systems.

Resistance Type Pressure Sensors
Piezoresistive pressure sensors have gained considerable attention due to their simple device structures, wide sensing range, and easy-to-read signal processing.Piezoresistive sensors convert pressure inputs into changes in resistance.[56] The active material should have adequate charge transport pathways for electrical current flow and sufficient elasticity to accommodate various mechanical deformations during operation.Composites made up of elastic matrix and conductive fillers are the preferred material choice to satisfy these criteria.Commonly chosen elastic components include polymers such as PDMS, [57,58] polyimide, [59] and PU, [60] and hydrogels such as polyacrylamide and bio-derived materials (cf.fibroin and cotton).For the conductive components, carbon-based fillers (cf.carbon black (CB), CNT, MXene, graphene, and carbonized biomass), metallic nanomaterials (cf.Au/Ag NWs and NPs), and conductive polymers (cf.PPy and polyaniline (PANI)) have been explored and demonstrated satisfactory device performance.Although numerous emerging nanomaterials and nanocomposites with unique geometrical microstructures have been explored to develop resistive pressure sensors, conventional resistive-type sensors with planar structures tend to exhibit relatively low sensing performance.To address this issue, highly sensitive networks for pressure sensing have been investigated utilizing innovative geometrical microstructures such as interlocked structures, percolative networks of nanomaterials (cf.NWs, CNTs, and rGO), patterned microstructures, and porous structures.
Hybrid composite materials fabricated by various coating methods have been employed to realize highly sensitive and scalable piezoresistive type pressure sensors.Zhou et al. [61] designed ultrathin, flexible, and highly sensitive piezoresistive pressure sensors using electrospinning technology and an all-fiber network structure.The sensing layer is composed of two different fibers (conductive fibers: PEDOT:PSS; elastic fibers: polyamide-6 [PA6]) with a large specific surface area, providing abundant contact sites.Furthermore, sensors performing different sensitivity and detection ranges could be fabricated by adjusting the electrospinning time.The sensors exhibited ultrahigh sensitivity of 6,554 kPa À1 and a wide detection range of 0-60 kPa (Figure 4a).Using the PEDOT:PSS/PA6 composite fiber-based pressure sensors, applications for health monitoring such as human physiological signals and operational joint signals were demonstrated.Yang et al. [60] reported a flexible piezoresistive pressure sensor with self-healing capability, exhibiting sensitivity of 8.7 kPa À1 , low detection limit of 50 Pa, and response/recovery time of 40 ms/117 ms.The ridge-like microstructure of the sensor, which is constructed using sandpaper as a template, contributed to its enhanced sensing performance.Additionally, the use of self-healable elastomer film enabled the self-healing capability.Also, the sensor was utilized to monitor the health Figure 3. a) Optical image of liquid metal alloy-based strain sensor and the strain sensing performance for 20 repeated cycles.Reproduced with permission. [50]Copyright 2018, Wiley-VCH.b) GF and strain sensing behavior of cPPy/nanofiber composite elastomer up to 388% strain.Reproduced with permission. [38]Copyright 2021, Elsevier.c) Schematic diagram and the performance of stretchable sensor based on EGaIn composite fibers.Reproduced with permission. [49]Copyright 2020, American Chemical Society.d) Device structure and the strain sensing characteristics of Ag@rGO@PDMS fiber strain sensors.Reproduced with permission. [51]Copyright 2020, Wiley-VCH.
status of human body through the Bluetooth transmission circuit.Yu et al. [52] investigated flexible pressure sensors using MXene and cellulose-based fabrics through a simple and scalable impregnation method (Figure 4b).The crisscrossed fiber structure of the nanowoven fabric was found to be responsible for enhancing the pressure sensitivity with Ti 3 C 2 T x exhibiting high piezoresistive properties due to its high conductivity and better conductive pathways.The sensors demonstrated sensitivity of 6.31 kPa À1 , sensing range up to 150 kPa, response/recovery time of 300 ms/260 ms, and durability up to 2000 cycles, suitable for medical diagnosis, human-computer interactions, and electronic skin applications.Cao et al. [56] reported a piezoresistive sensor fabricated using a 3D porous PU sponge and a conductive rGO layer that provides conductive paths during the compression, as shown in Figure 4c.The sensor showed enhanced sensitivity and long-term durability, and accurate detection of regular human motions and identification of different weights of objects with distinct electrical signals were possible.The integrated sensor arrays with signal collection circuits were applied for flexible electronic piano manufacturing, electronic skin, and  [61] Copyright 2022, American Chemical Society.b) Sensing mechanism of MXene/cellulose fabric-based pressure sensor, cross-sectional SEM images, and the pressure-sensing characteristics.Reproduced with permission. [52]Copyright 2022, American Chemical Society.c) Pressure-sensing mechanism of the 3D sponge-based pressure sensor and the piezoresistive performance.Reproduced with permission. [56]Copyright 2023, American Chemical Society.
remote real-time control of home electronics, demonstrating a feasible option for flexible electronics in smart application scenarios.

Capacitance Type Pressure Sensors
Piezocapacitive sensors convert pressure inputs into changes in capacitance.Similar to piezoresistive pressure sensors, piezocapacitive sensors are mainly composed of three components: a substrate, electrodes, and an active material which is typically a dielectric layer.In the past decades, various conductive materials and their composites including metal, [62][63][64] nanofibers, [65] CNT, [66] Ag NW, [67] and IL [68] have been investigated for flexible/stretchable electrode materials in the piezocapacitive sensors. [69]Also, various low-modulus dielectric materials have been studied as the dielectric medium including elastomer composites, [70] ionotronic films, [71,72] polymers, [72] GO, and CNT/PDMS. [73,74]In the piezocapacitive sensors, the external pressure-induced deformation of dielectrics leads to the change of capacitance.Consequently, enhancing the compressibility of the dielectric layer can directly increase the sensitivity.To achieve this, numerous dielectric layers with lower elastic modulus and diverse structures have been proposed such as micropyramids, [71,75] micropillars, [76] porous structures with air gaps, porous foam, [77] and NP-dispersed polymers. [72]Especially, in the air gap-embedded structures, the large dielectric constant change can significantly increase the sensitivity.Here, when air gap-embedded structure is pressed, the air volume decreases, resulting in a large dielectric constant change since the air has a lower dielectric constant than the elastomers.Furthermore, the viscoelastic behavior of the elastomer is reduced by the air gaps formed within its bulk, leading to shorter response and relaxation times.In addition, recently reported pressure sensors are fabricated on stretchable substrates, rather than the conventional flexible substrates. [78]Therefore, there is active research on strain-insensitive sensors for detecting only the pressure. [79]ighly sensitive capacitive-type pressure sensors mostly utilize unique sensor structures such as cylindrical or pyramidal shapes to increase the contact area of the sensing layer.Ha et al. [77] demonstrated a flexible hybrid-response pressure sensor (HRPS) which is composed of an electrically conductive porous nanocomposite (PNC) laminated with an ultrathin dielectric layer (Figure 5a).The PNC is fabricated with CNT-doped Ecoflex using a nickel foam template, exhibiting 86% porosity and electrical conductivity.Hybrid piezoresistive and piezocapacitive responses are achieved by the PNC, resulting in significantly enhanced sensitivity over wide pressure ranges (0-1 kPa, 3.13 kPa À1 ; 30-50 kPa, 0.43 kPa À1 ).Compared to the HRPS with its purely piezocapacitive counterparts, the hybrid responses are achieved from the effect of porosity or high dielectric constant.Liu et al. [80] also demonstrated a PU/IL foam-based pressure sensor with an ultrahigh sensitivity of 9,280 kPa À1 , which is benefited from its high porosity and low elastic modulus, as shown in Figure 5b.In the composite foam structure, the PU 3D foam served as an elastic structure, while the IL layer determined the electrical properties by forming interfacial electron-double-layer (EDL) capacitance.Potential applications such elastic foam-based ionotronic pressure sensors were demonstrated for wave fluctuation and vibration detection underwater, as well as the diagnosis of machine health.Lu et al. [76] presented an ionotronic pressure sensor with micropillared electrodes and an ion-gel layer to form the EDL interface.The sensor can dynamically and linearly respond to the mechanical stimuli with a high sensitivity (Figure 5c).In particular, three stages of deformation could be observed upon loading: initial contact, pillar buckling, and postbuckling, with the postbuckling stage resulting in linear and highly sensitive response (sensitivity = 33.16kPa À1 ) over a wide pressure range (12-176 kPa).The high linearity is due to the synergetic effects between the structural stiffening and contact area compensation at the electrode-gel interface, determined by the mechanical matching.The linear response and wide pressure range of the sensor-enabled applications such as stable pulse detection, planar pressure mapping, and robotic manipulation.Keum et al. [68] reported a textile-based capacitive pressure sensor utilizing a high-k ion-gel film.A highly sensitive pressure sensor was realized by optimizing the composition and thickness of the iongel film (Figure 5d).To elucidate the operation mechanism of the sensor, mechanical simulations using finite element method were conducted.The simulations showed that the capacitance of the pressure sensor varied due to changes in the contact area between the multifilament fiber electrodes and the ion-gel film with an applied pressure.Additionally, a 3 Â 3 pressure sensor array with multipoint detection capability was fabricated on a textile.A pressure monitoring circuit integrated with the pressure sensor and multicolor light-emitting diode array was also demonstrated.The results suggest that the textile-based pressure sensor has the potential for use in human-friendly smart textile applications.

Piezoelectric and Triboelectric Type Pressure Sensors
Piezoelectricity is the phenomenon in which an electric potential is generated in materials when subjected to mechanical stress. [81,82]Piezoelectric sensors are well-suited for dynamic pressure detection because of their stress variation-dependent sensing mechanisms.However, for static sensing, piezoelectric sensors may not be suitable as the output voltage generated by the sensor is impulsive.Various materials have been reported to exhibit the piezoelectric properties.Among them, polyvinylidene fluoride (PVDF) [83][84][85][86] and its copolymers such as poly(vinylidene fluoride-trifluoroethylene) [84,87,88] are promising candidates due to their flexibility, durability, lightweight nature, ease of fabrication, and chemical inertness. [89]Additionally, lead zirconate titanate (PZT) is one of the most widely used piezoelectric materials due to its high piezoelectric and electromechanical coupling coefficients.The intrinsically rigid PZT can be incorporated into flexible devices as ultrathin sheets or other nanostructured forms.Compared to the organic-based piezoelectric materials, the PZT-based sensors can offer increased sensitivity, fast response (sub-ms), low hysteresis, and high stability.Additionally, the conversion of mechanical energy to electrical energy can also occur through the triboelectric effect, in which certain materials become electrically charged when they come into frictional contact with another materials. [90,91]riboelectric nanogenerators (TENGs) have been extensively studied for low-power or self-powered pressure sensors since the first prototype was reported in 2012. [92][95][96] A full dynamic-range pressure sensor matrix with high sensitivity and resolution has been realized based on TENG and mechanoluminescent materials. [95,97]The device featured two integrated sensing components: a triboelectric sensor matrix and a mechanoluminescent sensor matrix.The sensor was capable of visualizing pressure distribution at low levels (<100 kPa).
Since most of the TENG-based pressure sensors aim to achieve low-power and self-powered operation, they possess characteristics suitable for wearable electronics.Li et al. [83] reported a flexible self-powered sensor based on poly(vinylidene fluoride-cohexafluoropropylene)/ZnO showing good stability and high sensitivity (1.9 V kPa À1 ) over a pressure range of 0.02-0.5 N.
The output performance was improved by the conjunction effect between the electric dipole inside the fiber and the asymmetric lattice of ZnO.The sensor was able to detect physical conditions for sports training to avoid injuries from overtraining.A composite nanofibrous membrane of siloxene/PVDF (S-PVDF) was proposed by Bhatta et al. [98] Using the S-PVDF as a highly negative layer for TENG, significant improvement in the electrical performance with peak power density of 13.25 W m À2 was achieved.The membrane also demonstrated good mechanical compressibility and was integrated in a capacitive pressure sensor to realize a self-powered high-pressure sensing (HPS) unit with enhanced dynamic and static pressure sensitivity.The structure and sensing mechanism are shown in Figure 6a.Copyright 2021, Wiley-VCH.b) Simulated stress distribution of structure with different porosities of 0%, 31%, 51%, 74%, and 95%, and sensing mechanism of the foam-based ionotronic pressure sensor.Reproduced with permission. [80]Copyright 2022, Springer Nature.c) Schematic illustration and cross-sectional SEM image of micropillared structure ionotronic pressure sensor and the change in capacitance as a function of pressure in the range of 0-180 kPa.Reproduced with permission. [76]Copyright 2021, Elsevier.d) Mechanical simulation results of ion-gel-based textile pressure sensors using finite element method.Three cases with different pressure conditions were simulated; without pressure and with low or high pressure.The insets show the contact area variation made between the multifilament fiber electrode and the ion-gel film under different pressures.The variation of resistance as a function of pressure and a pressure monitoring circuit on a textile.Reproduced with permission. [68]Copyright 2021, Elsevier.
Cai et al. [96] fabricated a self-powered tactile sensor based on TENG by using a wrinkled PDMS/MXene composite film.Ultraviolet-ozone treatment and the addition of MXene resulted in the formation of surface functional groups and surface wrinkles, respectively, which improved the sensitivity.The sensor exhibited a sensitivity of 0.18 and 0.06 V Pa À1 in the pressure ranges of 10-80 and 80-800 Pa, respectively.The high sensitivity in the 10-800 Pa pressure range enabled monitoring of complex human physiological signals and emulation of human touch sensation.Lee et al. [99] demonstrated a stretchable triboelectric pressure sensor (STPS) using a hierarchical superposition pattern, which maintains the pressure sensitivity regardless of strain.The STPS showed a sensitivity of 0.078 kPa À1 in the pressure range of <20 kPa, and the detection of human motions such as small and large movements was demonstrated.The proposed pattern also showed higher pressure sensitivity and larger output voltage compared to previously reported patterns.The potential applications for STPS were suggested as the robotics and smart healthcare systems (Figure 6b).

Temperature Sensors
Body temperature is a crucial and fundamental parameter for monitoring various illnesses and diseases (cf.[107] Wearable temperature sensors, as essential sensing component for artificial electronic skins, should possess high sensitivity, accuracy, fast response, wide sensing range (25-40 °C), long-term stability, and repeatability against ambient influences. [108,109]In recent developments, flexible and stretchable temperature sensors have been fabricated in various forms, depending on the materials and structures that were employed.Notably, these sensors can be categorized into resistance, [110] thermocouple, [111] or thermistor types. [112,113]Also, for the sensing materials, metals, [114] carbonnanomaterials, [115,116] and conductive polymers [117,118] have been often utilized.The resistance-type temperature sensors, also known as resistance temperature detectors, exploit the temperaturedependent electrical resistance variation.Increasing the temperature causes an elevation in resistance due to the electron vibration at higher temperatures, which hinders the free flow of electrons in conductive materials. [119,120]Thermocouples are another type of temperature sensors which consist of two distinct thermoelements (conductors, semiconductors, or a combination of both), generating a potential difference by the Seebeck thermoelectric phenomenon.When heat energy is applied to the ends of two the different conductors, a thermal voltage (electromotive force) arises between the end points.Thermistors are particularly suitable for detecting minor temperature variations and are fabricated from materials exhibiting significant changes in electrical resistance with temperature. [121]These thermally sensitive resistors can exhibit either negative temperature coefficient (NTC) [122,123] or positive temperature coefficient (PTC).
Shin et al. [120] reported a highly sensitive flexible NTC hermistor-based artificial skin, with excellent temperature sensing capability.In particular, a novel monolithic laser-induced reductive sintering scheme and unique monolithic structures were employed.The seamless monolithic structure integrates a metal electrode and metal oxide sensing channel from the same Figure 6.a) Schematic diagram of capacitive pressure sensor based on TENG and its sensing mechanism.Reproduced with permission. [98]Copyright 2022, Wiley-VCH.b) Schemes of the STPS with superposition pattern and equivalent circuit under single-electrode mode.Simulated result of the change in the output voltage as a function of external pressure with dome, wrinkle, and superposition pattern.Output voltage change of the sensor at pressures varying from 1 to 250 kPa.Reproduced with permission. [99]Copyright 2021, American Chemical Society.
material.The sensor could detect from 25 to 70 °C with temperature coefficient of resistance (TCR) of 0.4% °CÀ1 , as shown in Figure 7a.Sang et al. [114] proposed a highly sensitive, ultrathin Au-doped silicon nanomembrane (SiNM) temperature sensor array, featuring improved sensitivity to temperature and stretchable mesh structures with TCR of À37270.72 ppm °CÀ1 .Precise temperature change measurement and respiration monitoring are enabled without physical movements or moisture effects, and accurate sensing capability is demonstrated with low hysteresis and high cell-to-cell uniformity.The negative TCR property of Au-doped SiNM temperature sensor can be utilized as negative feedback within a circuit, increasing the stability and accuracy in silicon-based integrated circuit design, as shown in Figure 7b.Xiao et al. [123] also reported a flexible temperature sensor and sensor arrays using a screen-printing method, which demonstrated high sensitivity, excellent linearity, fast response time, and good repeatability in the temperature range of 18-44 °C.The negative TCR mechanism was explained based on the tunneling effects.

Pressure/Strain Cross-Reactive Sensors
With the significant popularization of portable electronic devices, the functionality of tactile sensors is no longer limited to a single stimulus but is expected to span a wide range of physical stimuli. [124]Simultaneous classification of stimuli is also a necessary feature, along with integrated diversity among various stimuli such as strain, pressure, bending, humidity, and temperature. [125,126]Multimodal tactile sensors capable of detecting pressure and strain are becoming increasingly essential in the development of stretchable and wearable electronics. [127]These sensors can play a crucial role in accurately capturing complex tactile stimuli and offer a more comprehensive understanding of the interactions between human and their environment.Stretchable electronics, as a rapidly advancing field, focuses on creating devices that can withstand significant deformations while maintaining their functionality.This characteristic makes them ideal for applications in robotics, prosthetics, and wearable systems, where they should adapt to various shapes and conform to the human body or other complex surfaces.Currently, the main types of sensors used for this purpose are resistive types (measure strain using cracks in the sensing layer and pressure using the compression), [128][129][130] capacitance types, [131,132] resistivefor-strain/capacitive-for-pressure types, [133] and combination of two sensors with different functions into one device. [134]These devices can continuously monitor vital signs, [135] track physical activities, [136] or enhance user experiences in gaming or virtual reality environments. [137]The integration of multimodal tactile  (25-30 °C).A model hand covered with the temperature-sensitive artificial skin and enlarged illustration of skin thermos-receptors.Reproduced with permission. [120]Copyright 2020, Wiley-VCH.b) Optical images of a conformally attached Au-doped SiNM temperature sensor on a forearm and a 4 Â 4 temperature sensor array on a mesh structure with PI and metal interconnection.Color heatmap of serpentine structure of Au-doped SiNM and continuous monitoring system for applications.Reproduced with permission. [114]Copyright 2022, Wiley-VCH.
sensors, especially, the pressure/strain sensors can enable precise detection and interpretation of multiple stimuli from the user's environment, leading to more accurate and reliable data acquisition for various applications. [138]ultisensing of cross-reactive tactile stimuli is vital for the development of intelligent systems that can adapt and respond to their surroundings.The ability to simultaneously measure pressure and strain allows these sensors to detect and differentiate between various stimuli such as touch, force, or deformation. [139]This feature is particularly important for applications like soft robotics and prosthetics, where the simultaneous detection of multiple stimuli is essential for precise control and accurate feedback.However, without the assistance of ML, it is very challenging to separate these continuous and simultaneously applied stimuli. [125,127,140]It is almost impossible for humans to manually distinguish the types and intensity of stimuli in real time from the continuous data that are measured. [141,142]However, with the development of ANN technology, distinguishing the cross-reactive stimuli is becoming increasingly feasible.For example, conventional sensors that operate on a lock-and-key mechanism make it difficult to separate each stimulus because it is difficult to distinguish crosstalk or electric signals induced by a certain stimulus. [143,144]evertheless, to develop a cross-reactive tactile sensor that can detect multiple stimuli simultaneously, various sensing materials have been investigated.Wei et al. [142] used a solutionprocessable approach to fabricate polyvinyl alcohol/phytic acid/ PANI/glutaraldehyde conductive composite hydrogels, exhibiting high mechanical and sensing performances (Figure 8a).The biocompatible hydrogel sensors show high sensitivity (GF = 3.4, pressure sensitivity = 0.62 kPa À1 ), conductivity, and durability over 1500 cycles, detecting both small and large strains and pressures, and weight or pressure distribution in real-time.The hydrogel is also fast-responding and able to recover quickly, providing a platform for human physiological activity monitoring and humanmachine interactions with high sensitivity and robustness.
Cao et al. [144] demonstrated a highly stretchable crumpled structure MXene film which was fabricated by brush coating the MXene ink onto an inflated latex balloon.The crumpled MXene film exhibited a pristine sheet resistance of 23 Ω/□ and maximum power density of 2.89 μW cm À2 when used as an electrode in TENG, which is 36 times higher than the flat Figure 8. a) Optical images of cross-reactive sensor for strain/pressure detection, compressive stress-strain relation, and the resistance variation as a function of pressure.Reproduced with permission. [142]Copyright 2021, American Chemical Society.b) Schematic illustration and images of MXene-based cross-reactive sensor (inset shows the equivalent circuit).Dynamic response under repeated stretching/releasing cycles and voltage variation under pressure up to 200 kPa.Reproduced with permission. [144]Copyright 2022, Elsevier.c) Device structure of nanofiber-based strain/pressure cross-reactive sensor and the resistance variation under strain (up to 1000%) and pressure (up to 30 kPa).Reproduced with permission. [145]Copyright 2021, Wiley-VCH.d) Optical images and equivalent circuit of spider web-like flexible pressure-strain sensor.The sensor can detect strain, stretched direction, and pressure simultaneously.Reproduced with permission. [146]Copyright 2021, American Chemical Society.
MXene film.This self-powered pressure sensor showed relatively high sensitivity in the pressure range of 0.3-200 kPa and good cycling stability in strain sensing.The sensing performance of the sensor is shown in Figure 8b.The self-powered pressure/ strain sensor was integrated into a real-time wireless gesture monitoring system to detect human body movements, confirming its potential in applications such as sports, healthcare, and robotics.Lu et al. [145] also developed a bioinspired nanofiberreinforced hydrogel (SFRH) with high mechanical strength, ionic conductivity of 3.93 S m À1 , and robust sensitivity (GF of 2.67) for the detection of broad range of strain (0.5-1100%) and pressure (1-28 kPa), as shown in Figure 8c.The SFRH could be used as soft sensors for detecting various body movements, blood pulses, and handwriting, for applications such as health monitoring, wearable electronics, medical diagnostic devices, tissue engineering, anticounterfeiting, and bioactuators.Zhao et al. [146] developed a spider web-like pressure/strain cross-reactive sensor by integrating a piezoresistive elastomer and stretchable electrodes into one pixel.The sensor consists of a 3D-tubular graphene sponge and spider weblike stretchable electrode, which are pressure-sensitive and strainsensitive element, respectively (Figure 8d).The sensor can monitor the magnitude and direction of the applied force showing its potential for applications such as wearable electronics and human-machine interaction systems.

Temperature/Pressure and Temperature/Strain Cross-Reactive Sensors
For wearable devices designed for biosignal monitoring, detecting the temperature of human body is crucial. [147,148]herefore, sensors capable of simultaneously detecting temperature with pressure or strain are currently under development.Temperature/pressure cross-reactive sensors play a vital role in health monitoring and clinical diagnosis.Consequently, rather than aiming for a broad sensing range, researchers focused on developing sensors with optimized sensing ranges for human body temperature and biometric signal detection. [149]For this purpose, meticulously designed metal nanomaterials, [150,151] polymer composites, [152] and ion-gels [153] were investigated.The temperature sensors which employ devices such as thermocouples and thermistors often involve structures connecting the upper and lower plates.This configuration significantly simplifies the fabrication of the sensor arrays. [154,155]In the case of pressure sensors, local detection of pressure is mostly pursued rather than the overall deformation, unlike the strain sensors.This led to demonstration of array-type pressure sensors which facilitate mapping of the pressure stimuli. [111,156]Temperature/strain cross-reactive sensors, in contrast, have been developed for applications such as robotics or artificial joints that can maximize the human performance or substitute for human capacities. [157]ence, sensors operating over a wide sensing range have attracted more attention.[163] Furthermore, for the implementation of these sensors in wearable devices, fabric-type substrates can be also used.Materials such as graphene, [164] MXene, [165] or Ag NW [166] have been utilized as the sensing layer to provide mechanical flexibility.The fabric-type sensors can exhibit greater flexibility compared to rigid substrates, allowing good conformal properties. [167]ai et al. [151] fabricated a multifunctional e-skin utilizing patterned metal films (PMFs) for tactile sensing of pressure and temperature changes.The e-skin exhibited broad sensing range and good linearity for pressure (80 kPa) and temperature (60 °C, TCR of 0.083% °CÀ1 ), as shown in Figure 9a.The decoupling of pressure and temperature signals was achieved using PMFs integrated on a microprotrusion-featured soft substrate.Characterized by its good mechanical flexibility, ease of processing, and simple configuration, the e-skin can be a robust tool for monitoring physiological signals and robotic perception.Wu et al. [168] also demonstrated an integrated electronic textile (E-textile) for temperature/pressure cross-reactive sensors with high spatial precision of 2D temperature/pressure detection (Figure 9b).The temperature sensors exhibited sensitivity of 1.23% °CÀ1 with fast response and robust stability.Also, the 2D pressure sensor showed sensitivity of 0.136 kPa À1 with high spatial accuracy.Liang et al. [169] fabricated pressure/temperature sensors using a direct stamping method.The sensor showed high pressure sensitivity of 0.67 kPa À1 and temperature sensitivity of 1.41% °CÀ1 , along with good stability over 5000 cycles.Chen et al. [153] introduced a fabrication method for realizing ultra-tough and ultra-sensitive ion-gel sensors.The ion-gels exhibited fracture strength of 21 MPa, Young's modulus of 325 MPa, and toughness of 102 MJ m À3 .In addition, the pressure sensors using a micropyramid structure combined with soft and hard hybrid layers showed high sensitivity of 33.8 kPa À1 , which overwhelms the dynamic sensitivity limit of conventional piezoresistive type ion-gel pressure sensors.Pang et al. [162] used a dual-network ionic conductive hydrogel for wearable strain/ temperature sensors (Figure 9c).The hydrogel, which exhibited stretchability of 720%, response time of 265 ms, and conductivity of 0.79 S m À1 , was able to monitor various human body motions with high sensitivity and stability.Also, the hydrogel-based temperature sensor with a phase transition temperature around 37 °C and a unique "V-shaped" conductivity-temperature relation was developed, showing promise for real-time environmental temperature detection and abnormal hyperthermia identification induced by fever or infection.Xiao et al. [170] introduced a hydrophobic and conductive nanofiber composite, demonstrating high-strain sensitivity (GF up to 154.8) and stability over 1,000 cycles.The composite also showed potential as a highperformance temperature sensor due to its NTC effect.

Multidimensional Cross-Reactive Tactile Sensors
The technology of isolating and detecting multidimensional (≥3) tactile signals is significantly challenging compared to the twodimensional cross-reactive sensors.Up to now, various attempts have been reported to realize the multidimensional tactile sensors. [46,171,172]In fact, the ability to perceive and respond to multiple tactile stimuli simultaneously not only mirrors the complexity of human touch, but also expands the applicability of these sensors in a wide range of fields from robotics and prosthetics to healthcare and environmental monitoring. [7,112,173,174]he primary advantage of multidimensional tactile sensors is their capability to provide comprehensive information from a single point of contact. [175]This enriched data can enhance the functionality and flexibility of the system they are integrated into, thereby providing a more nuanced response.In robotics, for instance, it would allow for a more sophisticated manipulation of objects, considering their shape, size, texture, and temperature.In healthcare, also, these sensors can be used to monitor patient vital signs more comprehensively, tracking pressure points, skin temperature, and other parameters simultaneously.Constructing sensors with materials that respond to all stimuli and structures free of crosstalk is a challenge.Nonetheless, active research is ongoing for sensors capable of detecting individual stimuli without the ML.The performance of these sensors is mostly dependent on the materials and structures used in their design.Commonly used materials include piezoelectric materials, [176,177] polymers, [178][179][180] and composites, [181,182] selected for their sensitivity to changes in pressure, temperature, and strain.
The structures can vary with microstructured surfaces, and flexible, stretchable, [183] and sensor array [184] configurations can be adapted to enhance the sensitivity and ensure compatibility with various applications.Research is also underway on sensors that intuitively and visibly detect tactile stimuli using electrochromic materials.These sensors are designed to change their colors when stimuli are applied, offering a straightforward and easily recognizable indication of sensing activity (Figure 10a). [185]In recent years, the integration of ML with multidimensional tactile sensors has emerged as a promising area of research. [186]ML algorithms can assist in extracting more useful information from the data generated by the sensors, improving the accuracy of predictions and decision-making processes.This integration can particularly enhance the application of the sensors in complex environments, where the simultaneous interpretation of multiple sensory inputs is required.However, due to the Figure 9. a) Schematic structure of pressure/temperature cross-reactive sensors with a serpentine structure electrode.Biosignal monitoring was demonstrated by attaching the sensor to wrist and forearm.Reproduced with permission. [151]Copyright 2021, American Association for Advancement of Science.b) Schematic device structure and the temperature-and pressure-sensing capabilities of the electronic textile sensors.Reproduced with permission. [168]Copyright 2019, Wiley-VCH.c) Strain/temperature cross-reactive sensors with ionic conductive hydrogel layer for the detection of human motion and abnormal hyperthermia in the body.Reproduced with permission. [162]Copyright 2022, American Chemical Society.
interdisciplinary nature of this field, there is currently a lack of research for comprehensive understanding of both device solutions and artificial intelligence.The fusion of these distinct research areas continues to present many challenges.Ongoing research and development are expected to address these issues and reveal the full potential of multidimensional tactile sensors in the future.
While real-time analysis and classification of tactile stimuli remain challenging without ML, recent research has been increasingly reporting on the ability to detect multiple types of stimuli using a single sensor.Liu et al. [175] developed a novel paper-based flexible multimodal sensor capable of detecting strain, humidity, temperature, and pressure, simultaneously.Facilitated by a CB and rGO mixture, a hierarchical sensing layer was formed on a paper, enabling the detection of various stimuli.The sensor demonstrated an acceptable sensitivity comparable to existing paper-based single-modal sensors and exhibited the capacity for degradation and reusability (Figure 10b).Wang et al. [7] presented a highly flexible tactile sensor featuring an interlocked truncated sawtooth structure.This unique structure facilitated the transformation of normal pressure into tensile strains of graphene patterns, thus significantly enhancing the sensitivity.The sensor utilized graphene nanoplatelets/silicone rubber and silver nanofibers/silicone rubber composites, exhibiting a high GF of 13 962 and conductivity of 200 S cm À1 .The sensitivity was calibrated to be high at 0.45 V N À1 for a range from 0.05-1.5 N and slightly lower at 0.12 V N À1 for 1.5-2.0N. The sensor displayed impressive repeatability, stability, and dynamic response over hundreds of cycles, showing its potential for applications in intelligent robotics, smart prosthetics, and wearable healthcare monitoring devices.Sun et al. [180] synthesized supramolecular-polymer dual-network eutectogels.They demonstrated robust adhesion and good conductivity.Notably, the temperature sensor fabricated from the eutectogel showed superior sensitivity (8.1-65.0°CÀ1 ), a wide range for temperature detection (30-120 °C), and excellent signal repeatability.The presence of nonconductive supramolecular gel network enabled the development of a highly sensitive wearable strain and pressure sensor suitable for both large joint movements and fine, complex motion monitoring.This design strategy offers a broad platform for future wearable flexible materials with multifunctionality.Kwon et al. [186] investigated a multifunctional, stretchable Bi 2 Te 3 thermoelectric fabric sensor using an in-situ reduction process, demonstrating potential for wearable electronics.This fabric, characterized by durable Bi 2 Te 3 NP networks within cotton fabric, demonstrated high stretchability up to 300% strain and outstanding electrical reliability under 10 000 cycles of mechanical deformation.Uniquely, the resistance of the fabric decreased under lateral strain, attributed to a decrease in the band gap of Bi 2 Te 3 NPs.The fabric exhibited a power factor of 25.77 μW m À1 K À2 , electrical conductivity of 36.7 S cm À1 , and a Seebeck coefficient of À83.79 μV K À1 at room temperature.It can also simultaneously detect temperature gradients and pressure owing to the ability to generate different output voltages for various temperature gradients based on the Seebeck effect.This stretchable Bi 2 Te 3 thermoelectric fabric-based device can offer promising potential for wearable electronics.

Classification Model
The classification is a fundamental task in ML, aiming to categorize given data into specific classes.Classification models are a fundamental part of ML, helping to categorize data into distinct classes based on learned patterns from the training data.In supervised learning, [187,188] classification models are trained using input data and associated labels to predict new data.During this process, the model learns patterns from the data and performs classification tasks based on these patterns.Performance indicators for these trained models include confusion matrix and accuracy. [189,190]A confusion matrix is a table that compares the classification model's predictions to actual outputs, showing true positives, true negatives, false positives, and false negatives.Similarly, the accuracy represents the probability of predictions matching actual results.These can be used to evaluate the performance metrics.In response to the increasing demand for big data, deep learning approaches such as neural networks are gathering considerable attention. [191,192]The neural networks employ a layered structure that mimics human neurons, and they perform tasks by representing the probability of each class at the final output layer.Neurons in each layer automatically extract features from the input data, and through the forward process, the model derives predictions.The associated error is applied to a loss function such as cross-entropy function, [193] and model optimization is achieved through the backpropagation process. [194]Tactile sensor research has primarily focused on detecting characteristic changes in response to stimuli such as strain, [195] pressure, [196] and temperature. [197]owever, recent studies have increasingly explored the integration of classification models with tactile sensors, utilizing the sensor's characteristic changes as input data for classification tasks.The first reason for this is the ease of selecting training input dataset.Researchers can train the data according to the sensing range of the sensors and specific data intervals which have been set, and then verify the results.The second reason is the ability to classify cross-reactive stimuli.When a mapping dataset containing complex and mixed stimuli is trained through a classification model, it can reveal and separate patterns that are indistinguishable to human.Thus, by integrating tactile sensors with classification models, [196,198] they have been widely applied in wearable healthcare, robotic tactile perception, [195] intermixed stimuli, and image processing and recognition. [199]In addition, advancements in signal processing can offer expanded prospects for tactile sensors in these areas.
One advanced research combining tactile sensors and classification models, there is the field of wearable biosensors.Biosensors are commonly attached to the skin areas such as hand or head, and they are used to detect characteristics that change based on the biometric signals or movements. [199]These changes can be analyzed accurately through the classification models.Tan et al. [196] have reported a wearable wristband that combined a human-computer interaction terminals with a neural network (Figure 11a).As data used in neural networks, TENG and piezoelectric nanogenerator signals were combined.Using eight active sensors enhanced by the ML, the wristband achieved a high accuracy of 92.6% with alphabetical keyboard and gesture recognition.The wristband combined with these classification models can translate sign language and prove valuable in environments with limited language communication.Classification models that can predict external stimuli based on the changes in data obtained in response to external stimuli are also frequently used.Niu et al. [200] recently developed a full-skin-bionic electronic skin applied to tactile recognition and advanced intelligent tactile recognition (Figure 11b).Utilizing ANN for real-time material identification, this system demonstrated significant potential in material recognition that exceeds human capabilities.Future work has demonstrated that it will recognize more material types in this context and focus on improving the practical value of ML.Mahmood et al. [199] also reported a complete portability, wireless, and ergonomic skin-like hybrid scalp electronic system for realtime monitoring of steady-state visual induction potential (SSVEP) of the scalp for brain-machine interface.The compact and highly portable design of low-power systems can minimize interference and provide reasonable signals for flexion and movement.The system can perform frequency classification of SSVEP in real-time using a deep learning CNN algorithm, which extracts low-level features during image processing through shared weights and provides highly efficient information transfer rate with only two channels, enabling precise control of a variety of devices.It demonstrated that classification models cannot only be used for data classification but also have applications in image processing tasks.Classification models that can predict external stimuli based on the changes in data obtained in response to external stimuli are also frequently used.Wei et al. [201] reported a flexible, self-operating tactile sensor system that uses three sensor arrays and deep learning to accurately identify different materials, as shown in Figure 12a.The system utilized the contact signals obtained from the sensors in the CNN model, achieving a 96.62% accuracy in identifying nine common materials.Lee et al. [202] presented integration of ML with a highly stretchable cross-reactive sensor matrix for detection, classification, and discrimination of various intermixed tactile and thermal stimuli.The sensor matrix consists of 10 Â 10 stretchable electrodes and multimodal hybrid sensing elements, exhibiting high sensitivity and rapid responses to different stimuli such as strain, pressure, flexion, and temperature.The design strategy of the cross-reactive sensor matrix allowed the construction of unique areal pattern data, forming the basis for the Bag of Words-based ML process to identify each stimulus as shown in Figure 12b.Consequently, by utilizing this strategy, discrimination of specific or even unknown stimulus was possible by separating the intermingled tactile and thermal stimuli.

Regression Model
The regression model is also another category of ML.Regression models are actually used for predicting continuous outcomes  [196] Copyright 2022, Wiley-VCH.b) Process flow of neural network-assisted material cognition, encompassing data generation and acquisition, training, and actual cognition processes.The cognition interface showcases location photographs of eight materials, along with the cognition results and corresponding curves for each material.Reproduced with permission. [200]Copyright 2022, Wiley-VCH.
based on input variables.While the classification models predict discrete labels, the regression models can predict continuous labels.This characteristic allows the regression models to make predictions even for labels that are not present in the training data.Therefore, it is possible to predict the intensity of stimuli that have not trained (e.g., when 10%, 20%, 30%, and 40% of strain have been trained, the regression model can still predict when a 15% strain stimulus is applied).Regression models use various loss functions such as root mean square error, [203] mean square error, [204,205] and mean absolute error, [206] to compare the differences between predicted and actual values.These loss functions are used to optimize the model by adjusting the weights through backpropagation.To evaluate the performance of the optimized model, the loss can be plotted during the training process or the predicted and actual values can be compared to calculate the R-squared value or assess the alignment between the predicted values and the actual values.R-squared value is a statistical metric, [207] which indicates how the independent variables explain the variance in the dependent variable, and the higher R-squared value indicates better predictive performance of the model.Because of these characteristics, there have been several studies on tactile sensing tasks employing the regression models, primarily for performing tasks that are challenging to solve through classification models, such as position recognition, [208] stimulation intensity estimation, [204] and material characterization. [209]Ha et al. [210] reported that the regression Normalized voltage characteristics of various contact materials under three different contact pressures (0.05, 0.1, and 0.2 N).Improved material identification system utilizing a neural network model (Visual Geometry Group architecture).Normalized heatmaps of the tactile texture signal (TTS) corresponding to nine contact materials.Confusion matrix presenting nine groups of the dataset using the TTS, achieving a material identification accuracy of 96.62%.Reproduced with permission. [201]Copyright 2022, Wiley-VCH.b) Discrimination of intermixed stimuli using the bag-of-words algorithm.The training process involves utilizing 2D current mapping data generated by FEA for both single and intermixed stimuli.Key features are extracted from the images using the SURF algorithm.K-means clustering is then applied to remove duplicate or weak key points.During testing, new image feature points are extracted and classified using the closest feature vector in the codebook and the support vector machine based on their respective categories.Reproduced with permission. [202]Copyright 2020, Wiley-VCH.
model effectively controls the position of a soft glove using pressure information with trajectory errors within 8%. Lee et al. [204] introduced a novel approach using wireless parallel pressure cognition platform (WiPPCoP).With the regression model combined with pressure sensors, prediction of pressure intensity was possible.The WiPPCoP combined with regression model is useful in areas such as electronic skin, cognitive science, and intelligent robotics (Figure 13a).However, regression models still have their limitations.One notable issue is that most studies using combined sensors have only been able to train on a single type of stimulus.While it is possible to verify one stimulus that exhibits linear or near-linear characteristics, the issue of separating and individually distinguishing two or more input stimuli for verification remains a challenge.
Despite these limitations, recent research has been actively focused on overcoming these challenges.So et al. [209] demonstrated epidermal sensors for classification and regression of mechanical stimuli and substrate modulus.The utilization of deep learning for data processing allowed precise regression and classification, offering simplified analysis of complex datasets in preclinical and clinical studies.Specifically, the customized deep neural network demonstrated the capability to predict three variables such as sensor output, substrate hardness, pressure that exhibit a nonlinear relationship.Massari et al. [211] proposed soft sensitive skin integrating modular tactile patches that can be adapted to various robot architectures (Figure 13b).Using the regression model, it was able to predict both the location and intensity of external loads applied to the patch surface.The use of fiber Bragg grating (FBG) sensors offered advantages such as multiplexing capabilities, high sensitivity, dense integration, and immunity to electromagnetic interference.Regression networks are employed to retrieve contact .Reproduced with permission. [204]Copyright 2020, Wiley-VCH.b) Schematics showing the integration of FBG transducers in artificial skin and the working principle of FBG, highlighting the encoding of strain applied to the grating through a peak-wavelength shift in the reflected light spectrum.The structure and results of a deep learning neural network, specifically developed for detecting force intensity.The diagram of localization of contact load on the skin via a feed-forward neural network and subsequent multigrid neural image processing.Reproduced with permission. [211]Copyright 2022, Springer Nature.c) The fabrication process of highly sensitive skin sensors using laser-induced crack generation.The neural network is composed of an encoding network and a decoding network.The encoding network utilizes long short-term memory layers to analyze temporal sensor patterns and generate latent vectors.Reproduced with permission. [212]Copyright 2020, Springer Nature.localization and force intensity information from the FBG wavelengths.The results show promising accuracy in force estimation and localization.By combining hardware development, fine sensing skills, and advanced artificial intelligence approaches, the potential of multimodal approaches in artificial skins was explained.Kim et al. [212] presented a novel technique for measuring dynamic motions using a deep-learned soft sensor attached to the skin surface, as shown in Figure 13c.A synchronized deep neural network successfully decoded finger motions in real-time by regression of metric space consisting of two parameters, R (bending of a finger) and θ (angle of the moving finger).The system's concept can be extended to other body parts, offering potential for detecting various stimuli and physiological signals.This technology can be applied in the fields of health monitoring, motion tracking, and soft robots.

SNN
The SNN is a specific kind of ANN that mimics the behavior of neurons in the biological brain.Unlike traditional deep neural networks which employ continuous analog values, SNN utilizes discrete spikes to encode and process information making them more efficient for certain tasks, particularly those involving temporal data processing.In the biological brain, the stimulus received from a preneuron is converted into a spike and transmitted to a postneuron.When the accumulated spikes in the postneuron exceed a specific threshold, a spike is generated, representing the transmission of information.As a result, the SNN holds promising potential due to low power consumption and fast response time for utilization in hardware-based artificial tactile sensors.Also, the SNN represents input and output information in the form of discrete spikes, typically expressed as binary values of 0 or 1. [213,214] These spikes influence the neuron membrane potential, a dynamic variable that changes over time.The potential is adjusted by incorporating weighted inputs from other neurons in the network.When the cumulative potential exceeds a certain threshold, the neuron fires, producing an output spike and resetting its potential.To incorporate these behaviors into a tactile sensor, various devices or circuits have been employed [214][215][216][217] such as the leaky integrate-and-fire and integrate-and-fire models.These models emulate the dynamics of  [221] Copyright 2022, Wiley-VCH.b) 3D plot of spike frequency encoding as a function of temperature and circuit's variable resistance in MFSN.SNN with one hidden layer with mapping spike frequency and transmission process, featuring 400 neurons in input layer, 50 neurons in hidden layer, and 8 neurons in output layer.Recognition rate of object patterns and temperature based on data type utilized.Reproduced with permission. [214]opyright 2022, Wiley-VCH.c) Schematic of a spatiotemporal SNN utilizing EGT synapses, depicting a 3 Â 1 configuration.Three preneurons connect to a postneuron through three EGT synapses.The preneurons generate a spike sequence (V s ) with increasing spike timings (t1 < t2 < t3), denoted as 1-2-3.Schematic of a 5 Â 5 touchpad with artificial tactile sensor array for moving orientation detection.Spatiotemporal information processing in SNN using EGTs enables intelligent touch orientation recognition and output spike generation.Reproduced with permission. [216]Copyright 2020, Wiley-VCH.
neurons by integrating incoming stimuli and generating spikes when specific conditions are met.Furthermore, as the SNN utilizes spikes as input data, conversion of continuous analog raw data into discrete spike representations is necessary.This conversion process involves several methods, including threshold encoding, [218] rate encoding, [213,214] and time encoding. [219]hese encoding techniques facilitate the transformation of continuous analog information into discrete spikes, enabling subsequent processing and interpretation by the SNN.The l earning methods employed in the SNN draw inspiration from the functions of the brain.One notable approach is the spiketiming-dependent plasticity method, [220] which relies on the Hebbian learning rule.According to this method, connections between neurons are strengthened when they are simultaneously activated.Additionally, there are alternative techniques such as backpropagation and ReSuMe that utilize surrogate functions. [221]These methods play a significant role in training SNN and offer diverse approaches to model the behavior of the brain.
As mentioned above, SNN is a promising method for artificial tactile sensors due to its advantages of low power consumption and fast response speed.Recently, there have been increasing efforts to implement human-like tactile perception in SNN for classifying pressure, [216] temperature, [214] patterns, [214] and objects. [214,222]Han et al. [221] have introduced self-powered artificial mechanosensory receptor module that imitates the behavior of a biological mechanoreceptor (Figure 14a).By integrating a self-powered pressure sensor and a TENG with biristor neurons, the module could perform simultaneous pressure detection and spike encoding, making it suitable for neurotypical sensory systems.With its ability to detect low pressures around 3 kPa, a broad range of applications can be possible including robotics, hearing aids, and medical/healthcare domains.Moreover, a hardware-based respiratory monitoring system was demonstrated using the module, showing precise classification of inhalation and exhalation.These advancements could offer exciting prospects for energy-efficient and highperformance tactile systems across diverse industries.Zhu et al. [214] also demonstrated a multimode-fused spike neural (MFSN) network that combines a pressure sensor and a NbO x memristor with thermal response properties (Figure 14b).This MFSN network efficiently captured and converted multimodal sensory inputs such as pressure and temperature, into spike trains.By integrating a 20 Â 20 MFSN array with an SNN classifier, it is shown that the performance of multimode-fused method in pattern recognition tasks outperformed single-modal approaches.Li et al. [216] reported the fabrication of a 32 Â 32 oxidebased electrochemical gating transistor (EGT) array exhibiting quasi-linearity, low variation, high endurance, high speed, low readout conductance, and low switching energy density.Based on the analog switching capability of the EGT array, hardwarebased SNN was demonstrated utilizing the spatiotemporal coding, with efficient learning and recognition abilities.Furthermore, the EGT-based SNN was applied to a tactile sensory system capable of detecting moving orientations (Figure 14c).The important figure of merits for tactile sensors are summarized in Table 1, representing the performance comparison of notable crossreactive sensors operated with ML.

Conclusion
Here, recent research trends of various types of smart tactile sensors are discussed, including capacitive, piezoresistive, and optical sensors, highlighting their principles of operation and advantages in different applications.Furthermore, the integration of neural networks into the tactile sensing systems is explored, emphasizing their ability to process and interpret complex tactile data for improved perception and dexterous manipulation.We highlighted key advancements in neural network architectures and learning algorithms for tactile sensing.It discussed the use of deep learning techniques, such as CNN, for feature extraction and pattern recognition from tactile signals.
Additionally, the article examined the importance of data augmentation, transfer learning, and active learning strategies in training tactile sensing systems with limited labeled data.Moreover, this article presented various applications of smart tactile sensor systems with neural networks, including robotic grasping and manipulation, prosthetics, and haptic feedback interfaces.It showcased how these systems have revolutionized these domains by enabling fine-grained tactile perception, adaptive control, and enhanced human-robot interactions.However, despite the remarkable progress in smart tactile sensor systems with neural networks, there are still challenges that need to be addressed.These include the development of robust and reliable sensors, the design of efficient neural network architectures for real-time processing, and the creation of large-scale labeled datasets for training and mimicking tactile sensing systems.In conclusion, the integration of neural networks with smart tactile sensor systems has opened up new possibilities for the field of robotics and artificial intelligence.The combination of advanced sensing technologies and powerful learning algorithms has paved the way for more sophisticated and versatile robotic systems capable of perceiving and interacting with the world in a tactile manner.As further research and advancements continue, we can expect to see even more groundbreaking applications and innovations in the field of smart tactile sensor systems with neural networks.

Figure 4 .
Figure 4. a) Schematic device structure and fabrication process of piezoresistive pressure sensor based on PEDOT:PSS/PA6 fibers, and the pressure-sensing performance.Reproduced with permission.[61]Copyright 2022, American Chemical Society.b) Sensing mechanism of MXene/cellulose fabric-based pressure sensor, cross-sectional SEM images, and the pressure-sensing characteristics.Reproduced with permission.[52]Copyright 2022, American Chemical Society.c) Pressure-sensing mechanism of the 3D sponge-based pressure sensor and the piezoresistive performance.Reproduced with permission.[56]Copyright 2023, American Chemical Society.

Figure 5 .
Figure 5. a) Schematic diagram and sensing performance of capacitive pressure sensor based on a conductive PNC composed of CNT.Reproduced with permission.[77]Copyright 2021, Wiley-VCH.b) Simulated stress distribution of structure with different porosities of 0%, 31%, 51%, 74%, and 95%, and sensing mechanism of the foam-based ionotronic pressure sensor.Reproduced with permission.[80]Copyright 2022, Springer Nature.c) Schematic illustration and cross-sectional SEM image of micropillared structure ionotronic pressure sensor and the change in capacitance as a function of pressure in the range of 0-180 kPa.Reproduced with permission.[76]Copyright 2021, Elsevier.d) Mechanical simulation results of ion-gel-based textile pressure sensors using finite element method.Three cases with different pressure conditions were simulated; without pressure and with low or high pressure.The insets show the contact area variation made between the multifilament fiber electrode and the ion-gel film under different pressures.The variation of resistance as a function of pressure and a pressure monitoring circuit on a textile.Reproduced with permission.[68]Copyright 2021, Elsevier.

Figure 7 .
Figure 7. a) Schematic representation of the monolithic laser reductive sintering process.PTC and NTC characteristics of the monolithic laser reductive sintering processed Ni electrode and Ni-NiO-Ni structure.The sensor exhibited high B-value (the material constant of thermistor) of 7350 K for entire measuring range (25-70 °C) and 8162 K for room temperature range(25-30 °C).A model hand covered with the temperature-sensitive artificial skin and enlarged illustration of skin thermos-receptors.Reproduced with permission.[120]Copyright 2020, Wiley-VCH.b) Optical images of a conformally attached Au-doped SiNM temperature sensor on a forearm and a 4 Â 4 temperature sensor array on a mesh structure with PI and metal interconnection.Color heatmap of serpentine structure of Au-doped SiNM and continuous monitoring system for applications.Reproduced with permission.[114]Copyright 2022, Wiley-VCH.

Figure 11 .
Figure 11.a) Schematic illustration of the hybrid generator in gesture recognition wristband (GRW), showing the operating principle of the piezoelectric nanogenerator (PENG) and TENG in GRW.The sensor is positioned near the tendons (represented by the dark red elongated object in the figure) beneath the skin.Voltage and current signals obtained from a single sensor during handshaking, indicating motions with minimal contact.Voltage and current signals obtained from a single sensor during fist clenches, representing motions involving greater force.Reproduced with permission.[196]Copyright 2022, Wiley-VCH.b) Process flow of neural network-assisted material cognition, encompassing data generation and acquisition, training, and actual cognition processes.The cognition interface showcases location photographs of eight materials, along with the cognition results and corresponding curves for each material.Reproduced with permission.[200]Copyright 2022, Wiley-VCH.

Figure 12 .
Figure 12. a) Variation of output voltage with ambient humidity when in contact with different materials.Each colored ball represents a distinct material.Normalized voltage characteristics of various contact materials under three different contact pressures (0.05, 0.1, and 0.2 N).Improved material identification system utilizing a neural network model (Visual Geometry Group architecture).Normalized heatmaps of the tactile texture signal (TTS) corresponding to nine contact materials.Confusion matrix presenting nine groups of the dataset using the TTS, achieving a material identification accuracy of 96.62%.Reproduced with permission.[201]Copyright 2022, Wiley-VCH.b) Discrimination of intermixed stimuli using the bag-of-words algorithm.The training process involves utilizing 2D current mapping data generated by FEA for both single and intermixed stimuli.Key features are extracted from the images using the SURF algorithm.K-means clustering is then applied to remove duplicate or weak key points.During testing, new image feature points are extracted and classified using the closest feature vector in the codebook and the support vector machine based on their respective categories.Reproduced with permission.[202]Copyright 2020, Wiley-VCH.

Figure 13 .
Figure13.a) ML-based pressure signal processing.Schematic illustration of the CNN.Comparison of the predicted pressures applied to each sensor with the actual pressure applied, after training with 100 data points.Visualization plot comparing the actual pressure levels applied to each sensor (represented by blue dots) to the predicted pressure levels after training with 10 data points (represented by black dots) and 100 data points (represented by red dots).Reproduced with permission.[204]Copyright 2020, Wiley-VCH.b) Schematics showing the integration of FBG transducers in artificial skin and the working principle of FBG, highlighting the encoding of strain applied to the grating through a peak-wavelength shift in the reflected light spectrum.The structure and results of a deep learning neural network, specifically developed for detecting force intensity.The diagram of localization of contact load on the skin via a feed-forward neural network and subsequent multigrid neural image processing.Reproduced with permission.[211]Copyright 2022, Springer Nature.c) The fabrication process of highly sensitive skin sensors using laser-induced crack generation.The neural network is composed of an encoding network and a decoding network.The encoding network utilizes long short-term memory layers to analyze temporal sensor patterns and generate latent vectors.Reproduced with permission.[212]Copyright 2020, Springer Nature.

Figure 14 .
Figure 14.a) SNN-based breath monitoring.Schematic illustration of the mechanoreceptor sensing.Comparison of the spike generated by the two mechanoreceptors at exhale and inhale motions.Plot of output current of SNN in response to exhale and inhale.Reproduced with permission.[221]Copyright 2022, Wiley-VCH.b) 3D plot of spike frequency encoding as a function of temperature and circuit's variable resistance in MFSN.SNN with one hidden layer with mapping spike frequency and transmission process, featuring 400 neurons in input layer, 50 neurons in hidden layer, and 8 neurons in output layer.Recognition rate of object patterns and temperature based on data type utilized.Reproduced with permission.[214]Copyright 2022, Wiley-VCH.c) Schematic of a spatiotemporal SNN utilizing EGT synapses, depicting a 3 Â 1 configuration.Three preneurons connect to a postneuron through three EGT synapses.The preneurons generate a spike sequence (V s ) with increasing spike timings (t1 < t2 < t3), denoted as 1-2-3.Schematic of a 5 Â 5 touchpad with artificial tactile sensor array for moving orientation detection.Spatiotemporal information processing in SNN using EGTs enables intelligent touch orientation recognition and output spike generation.Reproduced with permission.[216]Copyright 2020, Wiley-VCH.

Table 1 .
Performance comparison of multimodal tactile sensors.