An Artificial Intelligence‐Motivated Skin‐Like Optical Fiber Tactile Sensor

Soft and stretchable tactile sensors have received extensive attention for their potential applications in wearables, human–robot interaction, and intelligent robots. Herein, inspired by the functions of skin somatosensory signal generation and processing, an artificial intelligence‐motivated skin‐like optical fiber tactile (SOFT) sensor is proposed. It features multifunctional touch interaction capabilities including tactile amplitude and position and tensile strain. Four fiber Bragg gratings (FBGs) are embedded in a skin‐like three‐layer laminate structure of the SOFT sensor, forming a flexible tactile sensing array with a stretchability larger than 20%. Fusing the two‐level cascaded neural network, the position and magnitude of the contact force can be distinguished simultaneously. The recognition accuracy for contact position is up to 92.41% and the error is less than 4.2% within the force range of 0–3.5 N. Several SOFT sensor‐based interactive applications including pressure password interface and music playback are achieved by combining the artificial intelligence spatiotemporal dynamic logic analysis. Furthermore, the sensor is also capable of complex scenes involving tension and tactile sensing, such as dexterous hand perception and human–robot interaction control. This work provides novel insights into artificial intelligence‐based integrated skin that shows broad promise in intelligent prosthetics and bionic robotic.

DOI: 10.1002/aisy.202200460 Soft and stretchable tactile sensors have received extensive attention for their potential applications in wearables, human-robot interaction, and intelligent robots. Herein, inspired by the functions of skin somatosensory signal generation and processing, an artificial intelligence-motivated skin-like optical fiber tactile (SOFT) sensor is proposed. It features multifunctional touch interaction capabilities including tactile amplitude and position and tensile strain. Four fiber Bragg gratings (FBGs) are embedded in a skin-like three-layer laminate structure of the SOFT sensor, forming a flexible tactile sensing array with a stretchability larger than 20%. Fusing the two-level cascaded neural network, the position and magnitude of the contact force can be distinguished simultaneously. The recognition accuracy for contact position is up to 92.41% and the error is less than 4.2% within the force range of 0-3.5 N. Several SOFT sensor-based interactive applications including pressure password interface and music playback are achieved by combining the artificial intelligence spatiotemporal dynamic logic analysis. Furthermore, the sensor is also capable of complex scenes involving tension and tactile sensing, such as dexterous hand perception and humanrobot interaction control. This work provides novel insights into artificial intelligence-based integrated skin that shows broad promise in intelligent prosthetics and bionic robotic.
inherent tensile properties of silica fibers (tensile strain <1% [29] ), they are difficult to achieve skin-like soft and stretchable compliant structures to cope with mechanical surface deformations such as robot joint bending. Many attempts have been conducted to combine micro-/nano-optical fibers, [30] flexible optical waveguides, [31] etc. with polymer materials to fabricate flexible tactile sensors to achieve greater flexibility and stretchability. However, such intensity modulation sensors are difficult to achieve distributed monitoring, which leads to challenges in array integration.
Herein, a skin-like optical fiber tactile (SOFT) sensor with the FBG-sensing unit has been proposed for intelligent tactile cognition. Imitating the skin Ruffini corpuscle, four FBGs are embedded between two flexible substrates in parallel to form a skin-like three-layer laminate structure. These embedded FBGs have been divided into two sensing areas, which use the arrangement of the "double-x" pattern and a straight line to improve stretchability and achieve tactile sensing, respectively. Such a delicated design leads to a dense, soft, and highly stretchable optical fiber sensor with hybrid-sensing capability of tactile and strain (tensile strain is up to 20%). A mapping model between the mechanical stimulation and the wavelength shift of the SOFT sensor has been trained by the two-layer cascaded neural network, enabling the force and contact position simultaneous measurement with high accuracy. Additionally, benefiting from the tactile and strain hybrid-sensing capability of the stretchable SOFT sensor, it has unique merits that can measure the tactile information at stretchable joints in complex scenes such as dexterous hands and fingers.
As a proof of concept, the SOFT sensor is shown can not only recognize mechanical stimulation including tactile and tensile strain, but also project it into human-machine interaction control commands for multifunctional interaction. Combining with artificial intelligence spatiotemporal dynamic logic analysis, several SOFT sensor-based tactile interactive applications have been achieved, including music playback and pressure password interface. Additionally, benefiting from the unique advantages of tactile and strain hybrid sensing of the SOFT sensor, its application capabilities in complex scenes have been further explored, such as strain and tactile sensing at and near the stretchable joints of dexterous fingers and robot multi-degree-of-freedom motion control.

Design Inspiration and Fabrication of the SOFT Sensor
Biologically, external mechanical stimulation received by the human body is captured by a variety of tactile receptors (Ruffini corpuscle) in the skin and converted into receptor potentials. [32] Nerve fibers and synapses encode incoming signals from receptors into patterns of information that the brain can receive and understand. [33] The encoded postsynaptic potential is transmitted to the cerebral cortex through the spinal cord. Then the human brain performs logical analysis and judgment by interpreting the encoded information to identify the stimulation situation and finally makes stress reactions and decisions (Figure 1a, top). [34] Inspired by the functions of somatosensory signal generation and processing, an artificial intelligence-motivated SOFT sensor is demonstrated. In addition, an intelligent cognitive system integrating perception, signal transmission and processing, logic analysis, and interactive control is constructed. The SOFT sensor adopts a skin-like three-layer laminate structure. Four seriesconnected FBGs (Wuxi Ruike Huatai Electronic Technology Co., Ltd., Type: SMF-28Eþ, Central wavelength: 1540, 1545, 1550 and 1555 nm, Reflectivity ≥ 90%, the length of the grating area is 40 mm,) mimicking Ruffini corpuscle are embedded in the middle layer of a flexible substrate (Ecoflex 0020, Smooth-On Inc.) to form a sensing array (Figure 1a, bottom). Each FBG can be regarded as a narrow-band mirror that can generate spectral signals carrying the unique central wavelength. Different types of mechanical stimulation will be identified by the shift law of the FBG spectral signals ( Figure S1, Supporting Information). According to the fiber coupled mode theory, the central wavelength of the FBG can be expressed as where Λ is the natural period of the FBG, n eff is the effective refractive index of the core. Both lateral stress and axial strain can cause changes in Λ and n eff , which induce the wavelength shift of FBG. [35] Therefore, the wavelength shift of the fiber grating can be expressed as: The pressure will induce lateral stress and axial strain of the FBG, so the wavelength shift can be further decomposed as where the change of the parameter Δn eff is caused by the lateral force under the FBG plane stress; the axial strain of the FBG is represented by ε. Since the change of the bending radius of the FBG will also cause its wavelength shift, [36] the SOFT sensor is accordingly divided into a tactile unit and a tensile unit to exhibit multidimensional perception functions like the human skin ( Figure 1a, bottom). In our design, the embedded FBGs have been divided into two sensing areas: the tensile unit adopts a unique "double-x" configuration to improve the stretchability of the SOFT sensor, while the tactile unit is arranged in a straight line for easy array partitioning. As can be seen from Figure S2, Supporting Information, when the sensor is tensioned, the double x-shaped FBG driven by the Ecoflex elastomer gradually becomes straight as the tensile strain increases. Such a method of fiber shape control makes the sensor feature good tensile properties. In addition, the stretchability of the SOFT sensor can be regulated by the bending angle of the double-x-shaped fiber, which can increase with the increase of the bending angle of the double-x-shaped fiber ( Figure S3, Supporting Information).
As the increase of bending angle of the double-x-shaped fiber will lead to the increase of the longitudinal size of the sensor, the bending angle of 120°is selected to balance the tensile property and the longitudinal size of the sensor. Particularly, it has been proved that the sensitivity of FBG to edge forces is much smaller than midforces ( Figure S4, Supporting Information). Using this law, the differential arrangement of FBGs will contribute to improving the identification accuracy of force positions. Imitating the perception process of the skin, mechanical stimulation can be sensitively detected by SOFT sensors and output as spectral signals. The spectral signals are wavelength encoded and differentiated by the signal processing system and finally transmitted to the preset neural network model in the form of wavelength shifts. The neural network model will logically analyze the wavelength shifts like a brain, and the analysis results are output to the human-machine interface as control instructions to perform interactive operations (Figure 1a, bottom) To ensure the high flexibility and stretchability of the SOFT sensor to achieve curved-conformal tactile perception on geometric surfaces, Ecoflex 0020 material with lower hardness was selected as the flexible substrate. The simple fabrication process of the SOFT sensor is depicted in Figure 1b. First, the silicone rubber (Ecoflex 0020, ratio, 1:1) was poured into molds and cured at 70°C for 10 min to fabricate the bottom substrate of the tensile unit and the tactile unit. The elastomers were taken out with wiring lines and put into corresponding positions in the assembly mold respectively. Then the distributed FBGs were filled into the elastomers along the wiring lines. Finally, the silicone rubber was injected into the assembly mold, and the fabrication of the SOFT sensor was completed after curing. The tactile unit of the SOFT sensor is divided into a 2 Â 3 sensing array and the six tactile areas are respectively recorded as buttons www.advancedsciencenews.com www.advintellsyst.com 1-6 ( Figure 1b). The fabricated SOFT sensor is highly flexible like skin and can be pressed, tensioned, and twisted ( Figure 1c), which features potential in multifunctional applications such as wearables and robot skin.

Machine Learning-Based Tactile Perception Model
As shown in Figure 2a, a calibration experimental setup was constructed to calibrate the designed SOFT sensor to provide training and test sets for the neural network. The experimental system consists of a fine-tuning frame, a sliding pressure platform, a SOFT sensor, and an ATI force sensor. The fine-tuning frame and sliding pressure platform were used to perform indentation on the surface of the sensor tactile unit, the ATI sensor was used to collect the force data generated by the contact between the pressure platform and the SOFT sensor, and the wavelength signal of the SOFT sensor was collected by an optical interrogator.
A square indenter with a size of 8 mm Â 8 mm was used to indent the sensor, which is smaller than the minimum feature size (10 mm) of objects manipulated in precise grasping tasks in daily life. Considering that SOFT sensor is mainly targeted at the gentle touch tasks that usually occur in the process of fine manipulation activities of the robot, the dynamic pressure of 0-3.5 N is applied to buttons 1-6 respectively during the sampling process. The outputs of the ATI sensor and the SOFT sensor were simultaneously acquired through a self-built LabVIEW-based user interface, enabling and facilitating data refinement.
To obtain a detailed tactile prediction model including position and magnitude, a two-layer cascade neural network was designed, as shown in Figure 2b. The force position information was predicted by the position classification model based on the BP neural network in the first layer, and the force magnitude information was predicted by the ELM neural network-based fitting model in the second layer. In more detail, the architecture of www.advancedsciencenews.com www.advintellsyst.com each neural network is a three-layer network with input, hidden, and output layers. The first layer of the neural network (NN1) has four inputs and one output, which are the wavelength shifts of four FBGs and the predicted forced button. The wavelength shifts of the four FBGs were randomly shuffled after corresponding to the button number, 70% of the data was used as the training set, and the remaining 30% of the data was used as the prediction set. Its forward transfer subprocess can be expressed as where w ij is the weight between the i-th layer and the j-th layer, b j is the bias of the j-th layer, x i is the output of i-th layer, and x j is the output of j-th layer. The performance of NN1 was evaluated by the recognition accuracy of each button as an indicator (Figure 2c). The second layer of the neural network (NN2) cascaded after NN1 with five inputs, which are the wavelength shifts of the four FBGs corresponding to the ATI-outputted force value and the button number output by NN1, and one output, which is the prediction of the force value. The relationship from the input layer to the hidden layer can be expressed as where w is the input weight, b is the hidden layer bias value, the wavelength shift Δλ i is the input of the neural network, g is the activation function, and N s is the number of samples. As shown in Figure 2d, the data enters the output layer after passing through the hidden layer, and the output of the ELM can be expressed as where β is the input weight of output layer, to obtain a higher accuracy of β; its least squares solution can be expressed as The beetle algorithm is applied to the optimization of input weight (w) and hidden layer bias value (b) in NN2, and the performance of the second-layer neural network was evaluated by the mean relative error (MRE) (Figure 2d). The entire data training and model generation were performed in the software Python.

Performance and Characteristic of the SOFT Sensor
The SOFT sensor is divided into a tactile unit and a tensile unit to achieve a hybrid measurement of tactile and tensile strain, and their shape features are shown in Figure 3a. The 120°inferior arc ("double x")-shaped FBGs were buried into the flexible substrate of the tensile unit, leading to a highly stretchable elastomer (tensile strain is up to 20%, which is higher than the silica FBG (<1%)). The tactile unit achieves spatiotemporal tactile measurement by six array buttons. The area of each button is 10 mm Â 10 mm, which is set according to the size of the contact area between the finger and the sensor. (Figure 3a). Figure 3b shows the pressure response characteristic curve of FBG, which takes the response of FBG4 when button 1 is pressed as an example. In the range of 0-3.5 N, the sensor performs well in terms of linearity and the sensitivity is as high as 27.42 pm N À1 . In addition, the response of the SOFT sensor to cyclic mechanical stimulation of different pressure amplitudes was measured, as shown in Figure 3c. To further demonstrate the reliability of the SOFT sensor, a durability test of up to 1000 cycles with an amplitude of 1.5-2.5 N involving a hold time of 2 s was carried out (the pressing position is in the central area of the sensor). As shown in Figure 3d, the SOFT sensor maintains a stable signal output even after up to 1000 loading cycles, which proves the satisfactory durability and reliability of the sensor. Particularly, the dynamic response of the SOFT sensor to different buttons has been tested (the pressure range is 0-3.5 N), as shown in Figure 3e. Corresponding to six different buttons, the output signals of the four FBGs are combined into six different types (Figure 3e), which provides feasibility for neural network-based tactile sensing.
To accurately calibrate the strain response characteristics of the SOFT sensor, it was fastened to the fine-tuning stand and tensioned at constant intervals of 0.5 mm in the range of 0-3.5 mm. It can be seen that FBGs exhibit high sensitivity in the tensile range of 0-1.5 mm (about 0-9% tensile strain), and the maximum strain sensitivity is up to 139.3 pm mm À1 . Meanwhile, the response and recovery properties of the SOFT sensor under different tensile magnitudes inside of 0.5-1.5 mm with a constant interval of 0.5 mm (Figure 3g) and different frequencies of the mechanical stimulation ( Figure 3h) were investigated. It can be seen that when the tensile strain is applied to the SOFT sensor, the fiber is driven by Ecoflex to deform and thus output the strain signal. When the tensile strain is removed, the strain signal returns to 0, indicating that the fiber comes back to the original state. The result manifested that the SOFT sensor features excellent strain stability, which can accurately feedback tensile strain induced by skin or soft robot activity within the normal frequency range.
Since the temperature will cause the central wavelength shift of FBG ( Figure S5, Supporting Information), the wavelength shift of FBG can be set to zero by adjusting the initial wavelength benchmark before each use of the sensor, so as to effectively reduce the influence of ambient temperature changes. To further verify the fast response ability of the sensor to sudden mechanical stimulation, the response time of the sensor under fast pressing and fast stretching was tested ( Figure S6a,b, Supporting Information). It can be seen that the response time and recovery time of FBG1 and FBG2 to pressure are about 70 and 100 ms respectively. The response and recovery times of FBG3 and FBG4 to stress are about 60 and 110 ms, respectively. In addition, the response time and recovery time of the sensor to dynamic tensile strain are 130 and 160 ms, respectively ( Figure S6c, Supporting Information). The results show that the SOFT sensor features fast response and recovery performance, so as to realize real-time tactile/strain signal monitoring. Figure 4a shows the confusion matrix obtained from the classification of the neural network NN1 with an overall recognition www.advancedsciencenews.com www.advintellsyst.com accuracy of 92.41%. The high accuracy benefits from the remarkable multidimensional signal features obtained from the SOFT sensor. Meanwhile, the output curve fitting curves obtained from the neural network-based sensor output and the reference force have been determined (Figure 4b and S7, Supporting Information). As can be seen that the sensor output values of the six buttons are basically consistent with the reference values, their MREs are 2.77%, 4.20%, 3.09%, 3.48%, 3.99%, and 2.86%, respectively. The detailed information of the output errors has been visualized in Figure 4c. It can be found that most error points are in the range of À0.25-0.25 N, reflecting the high accuracy of the sensor output.

Multifunctional Touch Interaction
Human skin can sense external mechanical stimulation and transmit them to the brain. By processing and analyzing the different response signals from the skin, the brain can identify the form of mechanical stimulation that gives decision-making instructions. [37] Imitating the function of the biological tactile nerve, the artificial intelligence-motivated SOFT sensor enables to sense tactile information to realize human-machine interaction. As a proof of concept, the SOFT sensor-based multifunctional touch interface was constructed. Figure 5a shows the construction idea of the entire interactive system and a multifunctional touch interface has been implemented in the LabVIEW-MATLAB environment. Specifically, the first stage is the acquisition of response signals and optical-electrical conversion. The mechanical stimulation is measured by the SOFT sensor and converted into electrical signals by the FBG demodulator and input to the PC (subfigure (i) of Figure 5a). Then the sensing signals will be imported into the cascaded neural network model for real-time online prediction. The position and amplitude information of the mechanical stimulation will be recognized by the trained model and output to the interactive interface (subfigure (ii) of Figure 5a). Finally, according to the preset rules, the human-machine interface outputs control commands through logical analysis and executes actions according to the received commands as a response (subfigure (iii) of Figure 5a). www.advancedsciencenews.com www.advintellsyst.com Based on the force positioning function of the SOFT sensor, a piano interface was constructed. The six buttons of the SOFT sensor correspond to the six notes of the tonic solfa Do, Re, Mi, Fa, So, La (Figure 5b). It can be seen from Figure 5c that the output signal of the sensor changed regularly with the change of the pressed button and drove the virtual piano to produce the corresponding sound. Accordingly, by touching the SOFT sensor, the music "Happy Birthday to You" was played vividly.
Furthermore, with the assistance of the neural network model constructed in Figure 4, the SOFT sensor can not only identify the position, but also sense the force amplitude in real time when the position is correctly identified (Figure 5d). Therefore, due to the extensibility of the design principles, an intelligent pressure password interface is constructed to further explore the practical application potential of SOFT in multifunctional touch interaction. The six buttons of the SOFT sensor correspond to the numbers "1-6" on the virtual keyboard, respectively. When a finger touches the button, the SOFT sensor quickly generates a response signal and maintains it. The neural network model in the background will read the response signal in real time to predict the force button. Particularly, signal fluctuations caused by false tactile or material creep will affect practical applications. Therefore, the time threshold (>5 sampling points) and force threshold (>1 N) were set as the criteria for judging whether it is a user's subjective operation. A number will only be considered a valid number record and entered into the cryptosystem if it meets the time threshold and the force threshold. This setting can improve the robustness of the system and effectively prevent users from accidentally touching it. Two password types of "1352" and "2463" were presented to test the recognition accuracy of all buttons. It can be found that the pressed buttons were correctly identified and entered into the pressure password interface in the form of numbers. After the password was entered correctly, the "WHUT" logo at the bottom right of the interface was displayed (Figure 5e and Movie S1, Supporting Information). The sensor response during the experiment was recorded. Meanwhile, the predicted force position and force amplitude of each digit correctly identified were also collected and displayed in Figure 5f,g.

Dexterous Hand Perception
Benefitting from unique hybrid tactile and strain-sensing capabilities, the adaptability of the SOFT sensor to various complex scenes (such as dexterous hands and fingers, surface deformation caused by a robot or human joints, skin tension) is more conducive to its application in robot skin and wearable devices. Figure 6a,b shows that the SOFT sensor was respectively affixed to the flexor digitorum superficialis and flexor carpi radialis of the arm to monitor muscle activity caused by fist clenching and wrist swing. It can be seen that even if the strain is as tiny as less than 1%, the SOFT sensor can keenly capture human activity signals and return to the initial state when relaxed. Particularly, there are slight differences in skin strains at different positions caused by human muscle activity, [38] which can be sensitively captured by the SOFT sensor and reflected in the outputs of the four FBGs. The monitoring capability of the SOFT sensor for large strains induced by joint rotation is demonstrated in Figure 6c. As shown, when the SOFT sensor was pasted on the wrist, it can sensitively detect the large strain on the skin caused by the wrist rotation signal even if the strain reached 20% (Figure 6c). To further verify the sensing performance of the SOFT sensor to tactile and tensile strain under various complex scenes, it was wrapped on the surface of a dexterous bionic finger to simulate the skin on a robot hand. (Figure 6d). Figure 6e shows that the SOFT sensor can not only accurately perceive the force position and magnitude on the bionic finger, but also respond sensitively to the joint rotation. The successful demonstration further proves the multiparameter monitoring function of the SOFT sensor on robotic bodies even on dexterous robot hands.

Multitype Data Driven-Robot Motion Control
Furthermore, the tactile and strain hybrid-sensing capability of the SOFT sensor supports it to control the multidegree-offreedom motion of the robot arm by wrist activity detection and touch reception. As shown in Figure 7a, the SOFT sensor can distinguish and simultaneously respond to the strain induced by wrist bending and the pressure induced by finger pressing, which facilitates robotic control based on strain and tactile signals. Figure 7b shows the definition rules for controlling the up, down, left, and right movements of the robot arm through four actions: wrist up, wrist down, pressing the upper wrist, and pressing the bottom wrist. Therefore, the wrist movement-induced strain signal and tactile signal were accurately captured by the SOFT sensor and transmitted as control commands to guide the robot arm to perform precise grasping operations (Figure 7c and Movie S2, Supporting Information). Such results indicate that the application potential of the SOFT sensor is expected to be further expanded in areas such as wearable monitoring, robot control, and robot skin.  www.advancedsciencenews.com www.advintellsyst.com

Conclusion
In this article, an artificial intelligence-motivated SOFT tactile sensor is introduced. It enables to monitor mechanical stimulation including tactile and strain, as well as accurately identify the position and magnitude of the tactile. The SOFT sensor is configured as a skin-like three-layer laminate structure. Four seriesconnected FBGs have been buried in Ecoflex 0020 film, leading to a soft and highly stretchable elastomer. Meanwhile, the FBGs can be easily connected and their wavelength shifts can be distinguished by the wavelength division multiplexing technique, and the signal acquisition and transmission functions can be carried out in a single fiber. Such configurations address several key challenges such as intricate interconnects, complex structures and less stretchability of electrical tactile sensors, [39][40][41] and cumbersome array integration of intensity modulation optical sensors. [15,31] Particularly, the embedded FBGs have been deployed as the "double x" pattern and a straight line, which are employed to improve the stretchability and achieve tactile array sensing, respectively. This merit supports the monitoring of tactile and strain mechanical stimuli and the tactile sensing at or near stretchable joints that involve tensile strain change up to 20%. A two-layer cascaded neural network architecture was developed to identify the spatial contact characteristics of the SOFT sensor.
Experimental results show that the force measurement error and positioning accuracy are less than 4.2% and up to 92.41%, respectively. These designs yield multiple breakthroughs with respect to previous FBG-based tactile sensors, [26][27][28]42] such as significantly ameliorating the stretchability of the sensor and effectively addressing the problem of signal crosstalk in FBG arrays. The application potential of the SOFT sensor in constructing various human-machine interaction systems and tactile interaction under 1D plane and 3D surface has been further explored. Integrating artificial intelligence spatiotemporal dynamic logic analysis, multifunctional touch interactive interfaces including pressure password and music playback have been designed to develop more complex interactive functions for intelligent robots. In addition, the ability of SOFT sensor to handle various complex scenes has been verified. It enables to achieve precise tactile perception and tensile strain measurement at joints on dexterous hands or fingers, which provides an effective approach for multimodal perception in robot dexterous manipulation. Meanwhile, the multi-degree-of-freedom remote manipulator control driven by SOFT sensor strain signal and tactile signal was achieved. Such a control strategy can be expanded to be employed as a novel and highly safe way for doctors to remotely care for patients with infectious diseases and intelligent prosthetics for disabled people. These successful demonstrations prove that as a novel skin-like sensor solution, the SOFT sensor features huge application potential in rehabilitation medicine, human-machine interaction, and robotics. In the future, we will expand the prototype device of the SOFT sensor to a wider range of requirements, such as restoring the tactile perception of the disabled [33,43] and interactive tactile feedback of multiple parts of the robot. [19,40,41] In addition, the next work is expected to learn the change laws of the four FBG signals through the neural network to recognize the tensile strain in different directions. To achieve high-resolution tactile recognition similar to skin, more FBGs with shorter length will be embedded in series and distributed in the flexible substrate to better simulate human skin receptors, so as to prepare a high-spatial-resolution FBG tactile sensor array. Meanwhile, integrated skins with more ideal precision, stretchability, and scalability will be further developed, which will facilitate bionic robots to mimic human sensory performance in a more intelligent posture.

Experimental Section
Preparation of the Curing Mold: There were three curing molds for the fabrication of the SOFT sensor, which were denoted as the primary curing mold 1, primary curing mold 2, and the assembly mold, respectively. The three curing molds were made of ABS materials and printed by a 3D printer (Dreamer NX, Zhejiang Flash Casting 3D Technology Co., Ltd.,).
Characterization and Measurement: The sensor signal was recorded by the FBG interrogator (Gaussian optics Co. Ltd., China. OPM-T1620. The fine-tuning frame (LY60RM and LZ60-2, Zhejiang Runjia Pneumatic Technology Co., Ltd.) was used to fix the sensor and control the displacement of the sensor. The ATI sensor (Nano 17; ATI Industrial Automation, Apex, NC.) was used to collect pressure data. The dexterous bionic finger was created by a light-curing 3D printer (Form 2, FORMLABS INC.). The human-machine interactive grasp experiment was implemented by UR5 robot arm and soft touch robot. The characterization and application experiments of the sensor were carried out at room temperature (about 25°C).
Signal Processing and Interactive Interface: The signal acquisition program and human-machine interaction interface were constructed in software LabVIEW (2016 version). Neural network training and prediction was implemented in the software Python. The entire data processing and construction of the two-layer cascade neural network model were performed in the software MATLAB.

Supporting Information
Supporting Information is available from the Wiley Online Library or from the author.