Digitizing Human Motion via Bending Sensors toward Humanoid Robot

A humanoid robot is an advanced representation of robot techniques. It can move like a human and possesses the ability to use human tools. Sensing and intelligent control technologies have increased the intelligence of humanoid robots. However, the dexterity of a humanoid robot is still far from that of a human. Herein, a bending sensor is introduced to detect human motion, which can thus be digitized to enable a humanoid robot to learn to be more dexterous. This bending sensor is fabricated of carbon fiber with a cascade structure, resulting in excellent characteristics, including a quantitative detection capability, high gauge factor, low hysteresis, long‐term durability, and high‐frequency response. Thus, the bending sensor can recognize both the large‐scale motion of joint bending and subtle motion of muscle contraction. Fourteen bending sensors corresponding to the 14 knuckles of the human hand can accurately recognize and digitize hand gestures. These multichannel sensors are combined with nine additional bending sensors to comprehensively digitize the grasping process of the upper limb and recognize the grasping of objects with different sizes and weights. These successful applications demonstrate that the bending sensor can open a path toward learning human behavior for a humanoid robot.

DOI: 10.1002/aisy.202200337 A humanoid robot is an advanced representation of robot techniques. It can move like a human and possesses the ability to use human tools. Sensing and intelligent control technologies have increased the intelligence of humanoid robots. However, the dexterity of a humanoid robot is still far from that of a human. Herein, a bending sensor is introduced to detect human motion, which can thus be digitized to enable a humanoid robot to learn to be more dexterous. This bending sensor is fabricated of carbon fiber with a cascade structure, resulting in excellent characteristics, including a quantitative detection capability, high gauge factor, low hysteresis, long-term durability, and high-frequency response. Thus, the bending sensor can recognize both the large-scale motion of joint bending and subtle motion of muscle contraction. Fourteen bending sensors corresponding to the 14 knuckles of the human hand can accurately recognize and digitize hand gestures. These multichannel sensors are combined with nine additional bending sensors to comprehensively digitize the grasping process of the upper limb and recognize the grasping of objects with different sizes and weights. These successful applications demonstrate that the bending sensor can open a path toward learning human behavior for a humanoid robot.
provide information on the movement postures of the human body. It cannot provide in-depth information about the motion mechanism, which is important for a humanoid robot to learn human behavior. [30] An inertial motion capture system is based on microelectromechanical system (MEMS) inertial sensors, which can obtain the motion characteristics by measuring the accelerated velocity and angular velocity data of the human body along the X, Y, and Z axes in a spatial coordinate system. [31] A MEMS inertial sensor possesses excellent characteristics, including a small size, low power consumption, and high signal-to-noise ratio, and has received widespread attention for use in motion capture. [32] However, a MEMS inertial sensor also has poor compatibility with human skin because of its inherent rigidity, thus restricting the freedom of motion of the human body. Various kinds of flexible strain sensors have been developed for motion capture to increase the compatibility with human skin. [33][34][35][36] Lee et al. introduced augmented tactile-perception and haptic-feedback rings and a soft modular glove for motion capture and VR applications. The ring consisted of triboelectric nanogenerator (TENG) tactile sensors, flexible pyroelectric sensors, eccentric rotating mass vibrators, and nichrome (NiCr) metal wires. [37,38] Sundaram et al. used a scalable tactile glove sensor array and deep convolutional neural networks to learn the signatures of human grasp. [39] Ko et al. proposed an ultrasensitive skin-like sensor coupled with a deep neural network to elucidate human motions. [40] The results of these studies demonstrated that flexible sensors are robust candidates for capturing human motion. Among the various kinds of flexible sensors, strain sensors can respond to skin stretching and joint bending and thus detect human motions such as gesturing, grasping, walking, and running. [41][42][43][44] However, such strain sensors depend strongly on the stretchability of the sensing materials and lack sensitivity to subtle strain motions, such as muscle contraction. It is still a challenge to develop a high-sensitivity flexible sensor that can capture and digitize human motion to enable a humanoid robot to learn human behavior (as shown in Figure 1b,c).
In this study, we developed a bending sensor that could digitize human motion to establish a learning platform to enable a humanoid robot to move in a more human-like manner. The operation of the bending sensor was based on the transverse pressure-sensitive behavior of carbon fiber beams (CFBs) and contact area effect between two CFBs, producing a quantitative response to a bending angle. A high gauge factor (GF) at a small strain and high reliability ensured that the sensor could detect both the large-scale motion of joint bending and subtle motion of muscle contraction. Multichannel CFB bending sensors were used for point-to-point detection for accurate gesture recognition and digitization. We also used multichannel CFB bending sensors for combined joint-bending detection and muscle contraction monitoring to digitize the grasping process of the upper limb. A hue-saturationbrightness (HSB) model was established based on muscle contraction data to recognize human motion when grasping objects of different sizes and weights. This accurate and comprehensive capability to digitize human motion using CFB bending sensors could be effectively used to improve humanoid robots. Figure 2a shows the bending sensor that is compatible with human skin that was developed in this study for use in digitizing human motion. CFBs, which are highly flexible, free-standing,  These two encapsulation layers protected the sensing layer from the exterior environment and optimized the durability of the bending sensor. Figure S1, Supporting Information, presents the detailed manufacturing process of the CFB bending sensor, which involved threading, fixing, coating with a conductive silver paste, trimming, and encapsulation. Six CFBs are threaded into the FPCB and fixed in place by coating with a conductive silver paste at the copper holes. Then, the redundant lengths of the CFBs were trimmed, and the sensor was encapsulated by the PI and PA films. In addition, the sensing layer was cascaded with three sensing units to improve both the initial resistance and variation in the resistance, enabling the practical application of the bending sensor. A single sensing unit was composed of two CFBs with a K-junction structure. Thus, there was contact between the two CFBs, and the contact area increased with the bending angle of the sensor, as shown in Figure 2b. A CFB consisted of thousands of carbon fibers (CFs), with thousands of gaps between each pair of CFs. As the bending angle of the sensor increased, the number and size of the gaps decreased and the contact area between two CFBs increased, which both increased the probability of electron quantum tunneling between pairs of CFs. Therefore, the resistance of the bending sensor decreased as the bending angle increased (or the bending curvature decreased), as expressed by (for the derivation, see Note S1, Supporting Information)

Construction and Operating Principle of CFB Bending Sensor
where R is the resistance of the sensor; φ is the bending angle of the sensor; θ 0 is the initial contact angle between two CFBs; d is the horizontal distance between two CFBs; L is the board length of the sensor; parameter A is the product of a series of initial values; parameter B is equal to V γÃb , where V is the volume of the contact region between two CFBs, γ is the radius of the CFBs, and b is the width of the contact region between two CFBs; and parameter C is the fixed resistance corresponding to the current flowing through the CFBs (excluding the contact region).

Performance Characterization of the CFB Bending Sensor
To verify the relationship between the resistance of the CFB bending sensor and the bending angle, a series of cylinders with different curvatures were fabricated using a 3D printer (see Figure 3a). The CFB bending sensor was successively mounted on each of these cylinders, where the relationship between bending curvature k and the bending angle of the sensor is given below Figure 3b shows the resistance of the sensor in response to the bending curvature or bending angle. There was a high fitting degree between the experimentally obtained data and the data derived using Equation (1), verifying the relationship between the resistance of the CFB bending sensor and the bending angle. A high fitting degree was also found for five other batches of CFB bending sensors ( Figure S2, Supporting Information), demonstrating the high sensor reproducibility. The resistance data of the sensor could be converted into the resistance change rate to give the sensor a more convenient performance (as shown in Figure S3, Supporting Information), in which ΔR R 0 ¼ R 0 ÀR R 0 . Based on the high fitting degree between the sensor measurements and data derived using Equation (1), the bending angle of the sensor could be inferred from the resistance change rate (Figure 3c). There was a low maximum error of AE6% between the calculated and measured angles (see Figure 3d), which showed that the calculated angle agreed with the measured angle. These results also demonstrated that the CFB bending sensor had the ability to provide quantitative bending angle measurements, which would be beneficial for the digitization of human motion.
The strain coefficient is a characteristic parameter that reflects the sensitivity of both stretching and bending strain sensors, where the slope of the relative change in the electrical signal corresponding to the applied strain reflects the GF of a normal strain sensor. [45] In this contribution, the GF of the bending sensor is defined as follows www.advancedsciencenews.com www.advintellsyst.com where ε is the strain. Stretching-type strain sensor measure responses to high stretching (large strains), whereas bending sensors measure responses to small strains. [46] Because only small strains are induced in bending sensors, even at large bending angles, it is critical for bending sensors to achieve high GFs at small strain. Figure 2e and Table S1, Supporting Information, show that the strain in the proposed bending sensor was only 0.17% for a bending angle of 25°, where the elongation of the sensor was measured by an optical microscope ( Figure S4, Supporting Information). The bending sensor exhibited a high GF of 210 at a 0.07% strain and a bending angle of 15°. This GF was considerably higher than those of stretching strain sensors for a small strains ranging from 0 to 10 ( Figure 3f and Table S2, Supporting Information), showing that the bending sensor was very sensitive to small strains. As the GF of a stretching strain sensor typically increases with strain, these sensors often achieve high GFs at large strains. [47] During human motion monitoring, the bending angles of sensors mainly range from 0°to 60°i n response to joint bending and are less than 20°in response to muscle contraction. Thus, a CFB bending sensor with a high GF at a small strain could effectively be used to digitize human motion. In addition, the high bending-angle resolution of 0.5°of the CFB bending sensor ( Figure S5, Supporting Information) would ensure the recognition of subtle human motions. The use of free-standing and resilient CFBs in the CFB bending sensor resulted in low hysteresis. Figure 3g shows that the response curves of the bending and relaxation processes almost overlap when the sensor bends from 0°to 180°over four bending-relaxing cycles. Repeated bending and relaxing of the sensor was performed over tens of thousands of cycles to test the dynamic bending durability ( Figure S6, Supporting Information, shows the test equipment used to carry out the bending-relaxing cycles). Figure 3h shows the real-time evolution of the bending-relaxing cycles of the sensor. There was a negligible drift in the resistance change rate after the sensor had been bent and relaxed for 51 700 cycles, demonstrating the long-term dynamic bending durability of the CFB bending sensor. The response time reflects how quickly a sensor responds to bending, as well as its accuracy and stability. The standard response time of a sensor is usually defined in terms of a 90% time constant. [48] However, the response time is usually limited by the bending speed of the test instrument, and the response time of different sensors can only be compared using the same equipment at the same bending speed. In this study, we www.advancedsciencenews.com www.advintellsyst.com used the limit frequency response to evaluate the sensor response capability. Figure 3i shows that the CFB bending sensor completely responded to vibrations ranging from a low frequency of 10 Hz to a high frequency of 100 Hz, indicating that the sensor could respond to most human motions. Notably, the overall amplitude of the resistance change rate exhibited a declining trend as the vibration varied from a low frequency of 10 Hz to a high frequency of 100 Hz because the amplitude decreased with the increasing frequency for the same average power of the vibration generator. The combined features of low hysteresis, long-term dynamic bending durability, and a high-frequency response made the CFB bending sensor highly reliable for digitizing human motion.

Detection of Large-Scale and Subtle Motions by CFB Bending Sensor
The human motor system consists of bones, joints, and skeletal muscles. During human motion, joints are subjected to bending strain, and muscles undergo deformation from subtle stretching and contraction. [49] We used the CFB bending sensor to respond to both the large-scale motion of joint bending and subtle motion of muscle contraction to digitize human motion. First, a single CFB bending sensor was sewn into a knitted glove, which was used to steadily monitor the bending of a joint of the index finger ( Figure 4a). Figure 4b shows the resistance change rate of the sensor in response to the bending angle of the index finger.
The sensor accurately captured and digitized bending angles of 0°, 30°, 90°, and 110°, spanning the bending range of the index finger. The bending sensor also recognized when the index finger was bent from 0°to 80°in 5°increments ( Figure S7, Supporting Information), indicating that the sensor could achieve a high resolution of 5°for the finger bending angle. Muscle contraction involves a smaller deformation than the bending strain of a joint. A single CFB bending sensor was mounted on the skin of the biceps (Figure 4c), which is more active than other muscles during standing dumbbell curl exercises. Figure 4d presents the variation in resistance for the sensor response during five cycles of contraction and relaxing of the biceps. Figure S8, Supporting Information, shows the detection limits for the subtle motion of the muscle when the elbow was bent at angles of 10°, 20°, 40°, and 70°, demonstrating that the CFB bending sensor could recognize the subtle strain induced by muscle contraction. These results proved that the CFB bending sensor could detect both large-scale motion from joint bending and subtle motion from muscle contraction. The recognition of subtle motion would enable the CFB bending sensor to accurately digitize human motion.

Gesture Recognition and Digitization using Multichannel CFBs Bending Sensors
The human body contains over 300 joints, of which 78 provide the basis for human motion, such as knuckles, wrist joints, elbow  www.advancedsciencenews.com www.advintellsyst.com joints, and knee joints. [50] There are 14 knuckles in the hand (Figure 5a), which operate with the wrist joint and elbow joint to make the human hand an ideal natural tool. To completely digitize finger gestures and recognize gestural information, 14 CFB bending sensors corresponding to the 14 knuckles were sewn into a knitted glove for point-to-point detection (Figure 5b). These 14 knuckles include five metacarpophalangeal (MCP) joints, four proximal interphalangeal (PIP) joints, and four distal interphalangeal (DIP) joints. Notably, the thumb has only one interphalangeal joint in addition to the MCP joint. The real-time evolutions of the resistance change rates of these 14 CFB bending sensors corresponded to the ten different hand gestures (1-10), as shown in Figure 5c. These 14 sensors covered all the knuckles of the hand and could realistically reflect the bending of each knuckle and improve the accuracy of gesture recognition (Table S3, Supporting Information). Figure S9a, Supporting Information, shows that the bending of the pinky inevitably caused the ring finger to bend forward with a small angle, where the MCP and PIP joints of the ring finger bent but the DIP joint remained straight. Figure S9b, Supporting Information, shows that the resistance change rates of the sensors corresponding to the L-MCP, L-PIP, L-DIP, R-MCP, and R-PIP joints increased as the pinky bent, while the R-DIP joint remained unchanged, demonstrating that these sensors with point-to-point detection capability could accurately and thoroughly digitize the motion of the different joints. Similarly, the bending of the ring or middle finger caused the adjacent knuckle to bend (see Figure S10 and S11, Supporting Information). To directly demonstrate the gesture recognition ability of the multichannel CFB bending sensors, a real-time video was obtained (Movie S1, Supporting Information). This video showed that the multichannel CFB bending sensors could quickly and correctly recognize gestures 1-10, further demonstrating the capability of the bending sensor for gesture recognition and the digitization of human motion. A glove integrated with five CFB bending sensors could also recognize gestures 1-10 (as shown in Figure S12, Supporting Information) with a high accuracy of 98.8% (Table S4, Supporting Information).

Digitization of the Grasping Process using Multichannel CFB Bending Sensors
In addition to joints, the human body has over 600 muscles, including 50 in each upper limb. [51] All the bones, joints, and muscles of the upper limb are engaged when a human uses their hand to grasp a stainless steel cup with a diameter of 7.64 cm and weight of 0.8 kg. To digitize the grasping process, additional nine CFB bending sensors were sewn into a cuff to work in tandem with the aforementioned 14 CFB bending sensors (Figure 6a). These nine sensors were used to monitor the strain from the wrist joint, elbow joint, extensor digitorum, wrist extensor, wrist flexor, triceps (two sensors), and biceps (two sensors). The aforementioned muscles are the most active during the grasping process of the upper limb (Figure 6b), which is representative of human operational dexterity. Figure 6c,d presents the real-time evolutions of the resistance change rates of the 14 sensors in the knitted glove and nine sensors in the cuff in response to the grasping process, respectively. Each sensor exhibited a similar regular resistance variation during the grasping process over four cycles ( Figure S13, Supporting Information), indicating that all the sensors recognize the strain from the grasping process. Note  Figure S14, Supporting Information). The resistance change rate of the sensors corresponding to the muscles had lower amplitudes than those for the sensors corresponding to the joints, consistent with the strain in the muscles being smaller than those in the joints. These results showed that the multichannel CFB bending sensors could faithfully digitize human motion, as vividly demonstrated by a real-time video (Movie S2, Supporting Information). The size and weight of an object were the two key elements that influenced the amplitude of the electrical signals of the sensors during the grasping process. A paper cup with a diameter of 6.31 cm and weight of 0.0058 kg and a dumbbell with a diameter of 4.3 cm and weight of 4 kg were used in addition to the aforementioned stainless steel cup with a diameter of 7.64 cm and weight of 0.8 kg to investigate the effects of the size and weight of a graspable object ( Figure S15, Supporting Information). The real-time evolutions of the resistance change rates of the 23 CFB bending sensors in response to the hand grasping the paper cup, stainless steel cup, and dumbbell are shown in Figure S16, Supporting Information. The smallest and largest amplitudes for the resistance change rates of the sensors occurred when the hand grasped the paper cup and dumbbell, respectively. Generally, for a fixed object weight, the fingers need to bend at larger angles when grasping an object with a smaller diameter.
However, the weight of the object becomes the major factor when the object weight increases exponentially. This occurs because the hand needs to exert a larger gripping force, and the fingers need to bend more tightly to grasp an object with a larger weight. In addition, the muscles must undergo relatively large contractions or stretches because of the large weight of the object being grasped. To directly visualize this situation, the relative magnitudes of the resistance change rates of the corresponding sensors are indicated in terms of the color intensity. In Figure 7a, the darkest red and blue regions correspond to the largest magnitudes for the positive and negative ranges, respectively, whereas the lightest red and blue regions indicate magnitudes close to zero. The resistance change rate for the positive or negative range depended on whether the CFB bending sensor was initially in a relaxed or bent state, respectively. Thus, a HSB model could be established for the digitization of human motion. Figure 7b-d presents the HSB responses when the hand grasped the paper cup, stainless steel cup, and dumbbell, respectively. A single cycle of the grasping process could be decomposed into five actions, including the initial state, grabbing the object, picking up the object, putting down the object, and separation from (or restoration of ) the object. The dark blue or red color appears for the action of picking up the object (irrespective of whether the object is the paper cup, stainless steel cup, or dumbbell), whereas a nearly white color appears for both the initial state and separation. The hue associated with grabbing, picking up, and putting down an object deepens as the weight of the grasping object increases. Taking the region of triceps-1 as an example, the color of this region changes www.advancedsciencenews.com www.advintellsyst.com from white for the initial state to dark red when picking up the cup, and then returns to a light color for the separation. This result demonstrated that the triceps stretched while the cup was being picked up. Therefore, the multichannel CFB bending sensors with the capability of recognizing subtle motions from muscle contractions could feed back the motion information of the upper limb during the grasping of objects of different sizes and weights. These digital models of human motion could enable a humanoid robot to learn more "human-like" behavior. In addition to the size and weight of an object, the shape of the object had a significant influence on the electrical signals produced by the sensors during the grasping process. Three main object shapes were considered: cylindrical, spherical, and cuboid. Figure S17, Supporting Information, shows cylindrical objects consisting of a pen and an umbrella, spherical objects consisting of a tennis ball and an apple, and cuboid objects consisting of an eraser and a notebook, which were used as objects to be grasped to investigate the recognition function of the multichannel CFB bending sensors. Figure S18-S20, S22 and S23, Supporting Information, show the real-time evolutions of the resistance change rates of the sensors when the hand grasped the pen, umbrella, tennis ball, apple, eraser, and notebook, respectively. Each group of electrical signals had a distinct waveform and amplitude, demonstrating that the multichannel CFB bending sensors could both recognize and classify graspable objects.

Conclusion
In summary, we developed a bending sensor to digitize human motion based on CFBs as the sensing material and a cascade structure. This CFBs bending sensor exhibited excellent characteristics, including the ability to quantitatively measure bending angles, a high GF of 210 at small strains, low hysteresis, long-term dynamic bending durability, and a high-frequency response. These characteristics enabled the sensor to recognize both the large-scale motion of joint bending and the subtle motion of muscle contraction. Fourteen bending sensors corresponding to the 14 knuckles in the human hand were able to recognize ten different gestures, demonstrating the capability of the bending sensor in gesture recognition and the digitization of human motion. Combining these 14 bending sensors with an additional nine sensors mounted at the wrist joint, elbow joint, and the most active muscles of the upper limb realized the comprehensive digitization of the motion of grasping a cup. Data for the subtle motion of muscle contraction provided feedback on the motion that occurred during the grasping of objects of different sizes and weights. These results experimentally proved that the CFB bending sensor is a key tool for the digitization of human motion to provide a learning platform for a humanoid robot.

Experimental Section
Materials: Carbon fiber was purchased from Conston Technology Co., Ltd.; PI film was purchased from Suzhou Mai Case Plastic Products Co., Ltd., and PA film was purchased from Shanghai Xingxia polymer products Co., Ltd.
Fabrication of CFB Bending Sensor: The bending sensor was constructed of a flexible substrate, CFBs, PI film, and PA film, in which a FPCB (PI) with a length of 3 cm, a width of 1 cm, and the thickness of 0.09 mm as the substrate, and the size of the bending sensor could also be determined according to the actual application. Two CFBs were attached to the FPCB through four copper holes to form a sensing unit, and a sensor  Cylinders with different radius sizes printed by 3D printers were used for bending sensing tests. According to the length of the designed sensor, the radius corresponding to the required angle was calculated by a mathematical formula. Here, the sensor's length of 3 cm and the corresponding radius were calculated every 5°. From 5°to 180°, a total of 36 cylinders were obtained. The sensor was attached to the surface of the cylinder, and the resistance of the sensor was recorded in real time using a SourceMeter (Keithley 2450).
The Dynamic Bending Durability Testing of CFB Bending Sensors: Mechanical cyclic deformation system for flexible electronics was purchased from Shenzhen PURI Materials Technologies Co., Ltd. The sensor with a length of 3 cm was fixed on the fixture at both ends of the bending tester. The test speed of 5 mm s À1 , the bending radius of 28.6 mm, the number of cycles over 60 000, and the electric resistance of the sensor were recorded in real time by using a SourceMeter (Keithley 2450).
High-Frequency Response Testing of Bending Sensors: To test the response time of the bending sensor, the response time of the sensor was usually limited by the speed of the bending test instrument, so we built a high-frequency (0-150 Hz) test device, the physical map as shown in Figure S24 (Supporting Information). The frequency of the horn was used as the response frequency, and the bending sensor was fixed on the horn. The frequency and amplitude of the horn were adjustable through the self-made circuit board, and the frequency value was displayed through the oscilloscope. First, the bending sensor was prebent to ensure that some bending space was reserved for the sensor, and then the frequency debugging was carried out through the circuit board, and the resistance of the sensor was recorded in real time by a SourceMeter (Keithley 2450). The schematic diagram of the high-frequency test device is shown in Figure S25 (Supporting Information). The debugging started at 10 Hz and ended at every interval from 10 to 150 Hz, and recorded the changes of sensor electrical signals in real time. The changes in sensor electrical signals corresponding to 130 and 150 Hz responses are shown in Figure S26 (Supporting Information), and the amplitude variation trend of the whole process was recorded through slow-motion shooting. It was worth noting that under the same average power, the amplitude decreased slowly as the frequency increased. Due to the forced vibration of the horn, which was not ideal, the amplitude centers of different amplitudes might not be in the equilibrium position. The response frequency of the sensor was up to 150 Hz, and the error between the response frequency and the actual frequency was obtained (see Table S5 (Supporting Information)), which ensures the reliability of the response time of the sensor.
GF Measurement Method for Bending Sensors: The bending sensor and the soft rubber body were fixed together, and the two ends of the sensor were marked with beautiful paper tape. The rubber was placed on the surface of the cylinder with different curvature radii. The stretching length of the sensor was photographed and measured under a Leica microscope (LEICA DVM6) ( Figure S27 (Supporting Information)), and the resistance was recorded by a SourceMeter (Keithley 2450).
Multichannel Signal Acquisition Method: The multichannel signal acquisition system used AD7606 with a resolution of 16 bits and a maximum rate of 200 kHz. The main control system used STM32F1, the acquisition end and the host end used WiFi for wireless data transmission, the transmission protocol used TCP, and the WiFi chip used E103-W02 of Yibaite Electronic Technology Co., Ltd. When collecting, the analog signal was first input to AD7606, and then converted to get 16-bit digital signal output. Then, the digital signal was sent to MCU for processing and transformation to get the digital size of the voltage signal, and the voltage signal was transmitted to the WiFi module through the USART protocol. Then, the signal was sent in the form of an electromagnetic wave through the 2.4 G frequency band to the upper computer for display. The equivalent circuit diagram of ADC, the equivalent circuit diagram of the multichannel acquisition system, and the signal transmission process of the bending sensor are shown in Figure S28 (Supporting Information). In this article, bending sensors are divided into 14 channels for acquisition, among which there are 3 sensors in channels 1, 2, 3, and 4, respectively, which are switched and collected by simulation switch. Channel 5 had two bending sensors and the other channels had one sensor, so there are a total of 23 bending sensors.
Algorithms for Gesture Recognition: As shown in Figure S29 (Supporting Information), first, the data that were transmitted from the equivalent circuit were further processed and classified. The state of each joint of the finger could be judged by threshold value algorithms, in which the joint was defined as bending when the resistance change rate was greater than the threshold value, otherwise, as unbending. The thumb had two joints (M-MCP and M-PIP), and the other fingers had three joints (MCP, PIP, and DIP). Therefore, the state of the finger should be judged by whether the finger had over two joints that were defined as bending. If there were over two joints of the finger bending at the same time, this finger would be defined as bending (assign the character of "0"), otherwise, as unbending (assign the character of "1"). Second, combined these characters of the five fingers to match the character strings that represented the number from 1 to 10, such as "01000" corresponding to the number of 1, "01100" corresponding to the number of 2, "11 100" corresponding to the number of 3, "01111" corresponding to the number of 4, "11 111" corresponding to the number of 5, "01110" corresponding to the number of 6, "01101" corresponding to the number of 7, "01011" corresponding to the number of 8, "00111" corresponding to the number of 9, and "00000" corresponding to the number of 10. The experiments involved human subjects had been performed with the full, informed consent of the volunteers, who were also co-authors (first, fourth, fifth, and eighth authors) of the manuscript.

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