Multi-Object Recognition and Motion Detection Based on Flexible Pressure Sensor Array and Deep Learning

Abstract

With ongoing technological advancements, artificial tactile systems have become a prominent area of research, aiming to replicate human tactile capabilities and enabling machines and devices to interact with their environments. Achieving effective artificial tactile sensing relies on the integration of high-performance pressure sensors, precise signal acquisition, robust transmission, and rapid data processing. In this study, we developed a sensor array system based on flexible pressure sensors designed to recognize objects of varying shapes and sizes. The system comprises a multi-channel acquisition circuit and a signal transmission circuit and employs a convolutional neural network (CNN) to classify distinct signal patterns. In a test on an individual, the test results demonstrate that the system achieves a high recognition accuracy of 99.60% across two sphere sizes, three cylinder sizes, a cone, and a rectangular prism. In a group of eight people, it can achieve a recognition accuracy of 93.75%. Furthermore, we applied this sensor array system in an experimental setting involving a ball-throwing action, and it effectively recognized four distinct stages: empty hand, holding the ball, throwing, and catching. In repeated tests by other individuals, it was also able to clearly distinguish each stage. The development of artificial tactile systems allows robots to engage with their environments in a more nuanced and precise manner, enabling complex tasks such as surgical procedures, enhancing the interactive experience of wearable devices, and increasing immersion in virtual reality (VR) and augmented reality (AR). When integrated with deep learning, artificial tactile sensing shows significant potential for creating more intelligent and efficient applications.

1. Introduction

The artificial tactile system [1] is a revolutionary technology developed to replicate human tactile perception capabilities [2], enabling robots and other intelligent devices to achieve refined environmental recognition and interaction [3]. This technology integrates advanced sensors and algorithms [4], allowing the system to sense and interpret physical stimuli—such as contact force, pressure, and temperature—thereby enabling precise object recognition and manipulation [5]. Human tactile perception generally involves three main stages: sensing, transmission, and processing [6]. In the sensing stage, tactile receptors in the skin detect various types of physical stimuli; in the transmission stage, signals are transmitted through the nervous system to the brain; and finally, in the processing stage, the brain interprets these signals, enabling the recognition of object properties. The primary objective of artificial tactile systems is to emulate this process, allowing robots, prosthetics, and virtual reality systems to interact seamlessly with the physical world [7,8]. Tactile feedback is essential in this technology, enabling robots to perceive and respond to environmental information in real time and thus perform more complex tasks [9,10,11,12,13,14]. For instance, in surgical applications, robots equipped with tactile feedback can enhance both surgical precision and safety [15]. Similarly, in manufacturing, artificial tactile systems support automated production lines by ensuring that each step meets quality standards through real-time product monitoring. Tactile sensors can detect surface defects and make necessary adjustments to maintain the final product’s quality [16].

Current research is focused on enhancing the sensitivity [17,18], resolution [19,20], stretchability [21,22], and multi-modal integration [23,24] of artificial tactile systems. Researchers are developing more sensitive sensor materials capable of detecting multiple tactile cues, such as pressure, temperature, and vibration [25]. These advanced sensors more accurately simulate the sensing abilities of human skin, thereby improving system performance [26]. To meet the demands for flexibility and stretchability, significant research is devoted to creating flexible electronic materials [27], allowing sensors to conform to irregular surfaces and maintain functionality during movement. For instance, flexible sensors can be applied to robotic fingers, retaining their sensing capabilities while grasping objects of diverse shapes [28]. Furthermore, multi-modal integration has emerged as a crucial research direction [29]. By effectively combining various types of tactile information, researchers aim to replicate the human tactile experience more closely. This approach requires both advanced sensor design and robust data processing to extract meaningful information from large datasets. With advancements in Artificial Intelligence (AI), artificial tactile systems are increasingly integrated with machine learning and deep learning algorithms, allowing these systems to adapt to new tasks, enhance responsiveness to environmental changes, and support more complex decision-making and predictive capabilities [30,31,32,33]. By learning and extracting features from extensive tactile data, artificial tactile systems can more accurately identify object attributes, including material, shape, and surface texture. Convolutional neural networks (CNNs), for instance, are highly effective in processing complex tactile data and accurately recognizing various types of tactile information [16,34,35,36].

In the future, artificial tactile systems are expected to play a critical role in a wide range of applications [37,38]. In human–robot collaboration, as robotic technology advances, artificial tactile systems will enable robots to better understand and adapt to human operational methods, facilitating smoother cooperation [39,40]. For remote operations, these systems can provide operators with realistic tactile feedback, enabling a true tactile experience even over long distances [41,42]. This capability is particularly valuable in fields such as healthcare and military operations, where real-time tactile feedback during remote surgery or other operations enhances control and precision [37,43,44]. In disaster relief, robots equipped with artificial tactile systems can carry out precise operations in complex environments, aiding rescuers in executing hazardous tasks. With tactile feedback, these robots can detect their surroundings, identify and clear obstacles, and even assist in rescuing individuals trapped during emergencies [30,45,46,47,48]. As technology advances, future tactile systems are anticipated to become more sensitive [17,18], adaptable [49,50], and capable of engaging effectively with humans in increasingly complex environments [51,52,53]. The potential applications of this technology in diverse fields promise to enhance societal intelligence and elevate the quality of life. Through ongoing research and innovation, artificial tactile systems will progress toward their ultimate objective—simulating human touch to provide precise solutions for human needs.

This paper introduces the design of an array-based, multi-channel acquisition system utilizing flexible sensors. The system incorporates high-sensitivity, flexible sensors made from PET fabric, which detect varying pressures with minimal sensitivity to frequency fluctuations. Compared with other devices described in the supplementary document (Table S1), our sensors demonstrate a significant advantage in low-pressure detection. Moreover, they feature a simple structure and are made from easily accessible materials. The sensor demonstrates rapid responses in load and unload pressure tests, with response and recovery times of 119 ms and 68 ms, respectively, while sustaining a reliable performance over 1000 testing cycles. A signal acquisition circuit was developed for the hand-shaped sensor array, and a real-time signal acquisition hardware system was constructed using the STM32F104C8T6 microcontroller. Utilizing a trained convolutional neural network (CNN) model, the system accurately identifies objects of various shapes and sizes, achieving an impressive accuracy of 99.60% on the test set. Moreover, in a test conducted by a group consisting of five men and three women, the classification accuracy on the test set reached 93.75%. Additionally, it effectively distinguishes four stages of a ball-throwing motion: empty hand, holding, throwing, and catching. Furthermore, this result is also applicable to other repetitive throws.

2. Materials and Methods

2.1. The Fabrication of Flexible Pressure Sensor Materials

The fabrication process of the flexible sensors mainly refers to reference [54], with the material and equipment information detailed in Table S2 in the supplementary document. Figure 1 illustrates the main fabrication workflow of preparing flexible pressure sensor materials. In the production of polyester (120.32 g/m², twill weave), chemical lubricants are often added to enhance the fiber’s spinnability, necessitating the washing of the polyester fibers. Initially, a 2 g/L solution of detergent 209 is prepared, followed by ultrasonic washing for 30 min at a bath ratio of 1:50, and then rinsed with deionized water. After washing, the sample is dried in an oven at 80 °C. Next, a precise amount of polydimethylsiloxane (PDMS) is dissolved in anhydrous ethanol and accurately weighed on an electronic balance. The solution undergoes 10 min of ultrasonic stirring to ensure thorough mixing. The cleaned fabric is then cut to the specified dimensions, immersed in the PDMS solution, and subjected to ultrasonic impregnation for 10 min at 30 °C. This process is repeated three times, followed by drying at 60 °C. The sample prepared using this method is designated PET-d-PDMS. Subsequently, plasma is conducted using a plasma device with an RF power source set at 13.56 MHz, employing capacitively coupled plasma (CCP) discharge. Argon gas is utilized, with adjustments made to the gas pressure (30 Pa), discharge power (100 W), and discharge time (20 s). The sample obtained from this step is labeled PET-g-PDMS. Next, 2 g of multi-walled carbon nanotubes (MWCNTs) is added to 100 mL of anhydrous ethanol, and probe ultrasonication is performed at 20 kHz for 2 h. The resulting dispersion is gently centrifuged at 2000 rpm for 20 min to remove any free MWCNTs.

The pretreated PET fabric is then immersed in the CNT dispersion for 30 min, dried in an oven at 60 °C, and subsequently cured by applying an alternating current (AC) electric field (ctp2000 k) between two electrode plates until the anhydrous ethanol completely evaporates. Finally, the fabric is cured and dried again at 60 °C. Now, the material preparation for the flexible pressure sensor is complete. The assembly and structure of the flexible pressure sensor will be introduced in Section 2.2.

2.2. Performance Characterization

The structure of the flexible sensor is depicted in Figure 2a. The pressure sensor is composed of five layers of hydrophobic conductive fabric, each measuring 1 cm × 1 cm, with conductive copper tape utilized as the top and bottom electrodes. This size was chosen considering the size of the fingers, while the thickness was determined to achieve the best balance for wearer comfort. To mitigate the impact of contact resistance during testing, the outer sides of the electrodes are wrapped in medical-grade polyurethane (PU) film, ensuring stable electrical performance signals from the sensor. The performance of the fabricated sensor was tested using an electrochemical workstation (CHI660E, Shanghai Chenhua Instrument Co., Ltd., Shanghai, China). The sensor exhibits different signal responses under pressures of 0.1 kPa, 0.2 kPa, 0.5 kPa, 1 kPa, 2 kPa, and 5 kPa, demonstrating its capability to detect a range of pressure levels (Figure 2b). Under a constant pressure of 0.5 kPa, the sensor was tested at frequencies of 0.5 Hz, 1 Hz, 2 Hz, and 3 Hz, revealing that the different frequencies had minimal impact on the sensor’s relative current change (Figure 2c). Current–voltage curves were obtained at pressures of 0 kPa, 0.1 kPa, 0.2 kPa, 0.5 kPa, 1 kPa, 2 kPa, and 5 kPa. Within the limits of the sensor’s functionality, the resistance exhibited a linear change, which amply demonstrated that the material possessed practical piezoresistive properties (Figure 2d). Under a pressure of 0.5 kPa, loading and unloading tests were conducted, yielding response and recovery times of 119 ms and 68 ms, respectively, as shown in the inset of Figure 2e, indicating a rapid response. The sensor exhibited stability after 1000 loading and unloading cycles at a pressure of 0.5 kPa. The magnified view in the inset displays the relative current changes from the 35th to 39th and the 957th to 961st cycles, which are nearly identical, thereby verifying its long-term stability (Figure 2f).

Compared to the sensor in reference [54], the sensor fabricated in this study has a greater advantage in low-pressure detection, whereas the sensor in reference [54] is more suitable for detection within the medium to high pressure range. Additionally, the sensor in this study is also more appropriate for routine static or low-frequency dynamic pressure monitoring.

2.3. Data Acquisition

To achieve effective signal acquisition in flexible tactile systems, Bao et al. [55] developed an integrated intelligent tactile system for a humanoid robot, emphasizing the importance of real-time data collection with minimal system interference. Building on this concept, we designed a signal acquisition circuit based on a flexible sensor array, as depicted in Figure 3. This design aims to achieve real-time multi-channel data collection while preserving the flexibility of the hand. The circuit consists of five main components, as illustrated in Figure 3a: a sensor array module, data acquisition module, microcontroller (MCU) module, data transmission module, and PC data receiving module.

We use nine flexible pressure sensors to form a flexible pressure sensor array, with the sensors distributed across the five fingers of a hand, as shown in Figure 3f. Figure 3g presents a photograph of the sensor array. The nine pressure sensors are evenly distributed across the five fingers of the right hand. Compared to having just one or two sensors on a single finger, this arrangement significantly improves the accuracy and flexibility of tactile perception. The multi-point sensor design allows for finer detection of pressure differences between the fingers, providing more accurate tactile feedback, which is especially crucial for robotic grasping and object manipulation tasks. Sensors distributed across multiple fingers better simulate the tactile perception of the human hand, enhancing the system’s precision in spatial awareness, object recognition, and classification. Additionally, the distributed design of the sensors increases the system’s fault tolerance, ensuring stable tactile feedback even if one sensor fails [56,57,58]. The sensor array is responsible for converting pressure signals into electrical signals. These signals are collected by the data acquisition circuit (Figure 3c) and transmitted to the MCU module (Figure 3d). We selected the STM32F104C8T6 microcontroller (STMicroelectronics, Geneva, Switzerland) from STMicroelectronics, which utilizes its built-in analog-to-digital converter (ADC) module to convert analog signals into digital signals through programming.

Data transmission employs the asynchronous serial communication protocol (UART), facilitating the transfer of data from the microcontroller (MCU) to the PC data receiving end. UART communication utilizes two data lines, with each line connecting devices via RX (receive) and TX (transmit) pins, where the RX and TX pins are cross-connected between the two devices. Figure 3b presents a photograph of the serial-to-USB adapter based on the CH340 chip used, which is part of the STM32 development kit.

Figure 3e displays a schematic diagram of the data acquisition circuit, which uses two operational amplifiers to extract the signals. The first operational amplifier converts the current signal into a voltage signal. When current flows through R₁, a voltage difference is generated across R₁. Based on the virtual short principle, there is no current flowing into the input of the operational amplifier, so

$$(V_5-V_2)/R_3=V_2/R_5$$

(1)

$$\left(V_4-V_1\right)/R_2=\left(V_1-V_3\right)/R_4$$

(2)

From the virtual short principle, we can conclude that

$$V_1=V_2$$

(3)

Assuming R₂ = R₃ and R₄ = R₅, we can derive the following from the above three equations:

$$V_3=R_4\left(V_5-V_4\right)/R_2$$

(4)

Here, (V₅ − V₄) represents the voltage difference across R₁, which is the converted voltage signal. When R_X0 senses a change in pressure, a corresponding voltage change will be reflected across R₁. To ensure that the output of the operational amplifier is not distorted, we conservatively set R₂ = R₄, resulting in V₃ = (V₅ − V₄).

$$V_5-V_4=V_0{R}_1/(R_1+R_{X0})$$

(5)

This section explains the relationship between (V₅ − V₄) and R_X0. Here, V₀ is directly provided by a pin of the STM32, with a value of 3.3 V. In this case, during the component selection, we ensured that the resistance of R₁ is much smaller than that of R_X0, allowing the relationship between the two to be simplified as follows:

$$V_5-V_4=V_0{R}_1/R_{X0}$$

(6)

The second operational amplifier primarily amplifies the voltage signal. Again, based on the principles of virtual short and virtual ground, we can derive the following from the second operational amplifier:

$$V_{OUT}=V_3(R_6+R_7)/(X+R_7)$$

(7)

where X represents the resistance value of the adjustable resistor R₆, which is situated near R₇. X = 0 for R₇ ensures that the inverting terminal of the operational amplifier is not directly grounded. R_X0 represents the sensors.

The relationship between the output signal and the sensor can ultimately be expressed as follows:

$$V_{OUT}=\left[V_0{R}_1\left(R_6+R_7\right)\right]/\left[\left(X+R_7\right)R_{X0}\right]$$

(8)

3. Results and Discussion

3.1. Signal Acquisition of Object Shape

In this experiment, the designed signal acquisition circuit was employed to collect signals during grasp tests on variously shaped objects, ensuring that the bottom surface of each object remained on the table throughout the testing process. The test subjects included a cone (height 21 cm, diameter 12.5 cm), a rectangular cuboid (length 6 cm, width 6 cm, height 10 cm), a sphere, and a cylinder. Specifically, two spheres (with diameters of 6 cm and 10 cm) and three cylinders (each 10 cm in height, with diameters of 4 cm, 6 cm, and 8 cm) were tested.

In Figure 4a, the signal data for grasping objects of different shapes and sizes are summarized. Significant signal differences were observed for the various object shapes. For example, when grasping the cone (Figure 4b), channels 3 and 9 exhibited relatively weak signal intensities. In contrast, for the rectangular cuboid (Figure 4c), which requires greater fingertip strength, the primary signals were concentrated in channels 1, 2, 4, 6, and 8. When grasping a sphere with a 10 cm diameter (Figure 4e), channels 3, 4, 6, and 9 displayed varying signal strengths. Specifically, channels 3 and 9 exhibited weaker signals, while channel 6 registered approximately 0.7 V, and channel 4 demonstrated the strongest signal among the four. For a cylinder with a 6 cm diameter (Figure 4g), the signals in channels 3 and 9 showed significant differences from the other channels, with their signal amplitudes being roughly half of those in the others.

Further analysis was conducted on the signal differences across channels for different object sizes (Figure 4d–h). Comparing the signals for the two spheres (Figure 4d,e), there was a notable difference in intensity in channels 6 and 8. For the smaller sphere, channel 8 was weaker, and channel 6 was slightly stronger; conversely, for the larger sphere, channel 4 exhibited a tendency toward higher signal strength. Figure 4f–h display the signal variations for grasping cylinders of three different sizes, with channels 3 and 9 showing the most significant changes. When grasping the cylinder with a 6 cm diameter, these channels exhibited the strongest signals. For the cylinder with a 4 cm diameter, channel 3 demonstrated moderate signal strength, while it was weakest for the 8 cm diameter cylinder. Conversely, channel 9 exhibited moderate signal strength for the 8 cm cylinder and was weakest for the 4 cm cylinder.

In summary, the signal intensity across different sensor channels varied significantly when grasping different objects. These differences not only provide a basis for distinguishing objects of various shapes and sizes but also lay a foundation for the intelligent development of artificial tactile systems. Combined with deep learning algorithms, these signal differences can be utilized for more accurate object recognition and classification, enabling robots to perform complex tasks with greater flexibility and precision. In scenarios such as surgery and rescue, tactile feedback allows for safer and more efficient operations. Furthermore, leveraging these signal differences can enhance tactile experiences in virtual (VR) and augmented reality (AR), making human–computer interactions more natural and realistic, as demonstrated by Wen et al. [59], who explored the use of conductive triboelectric textiles for gesture recognition in VR/AR applications.

3.2. Signal Recognition Based on Deep Learning

We first selected a 27-year-old male participant who is right-handed and conducted the test on him individually. We constructed a dataset based on the seven signal characteristics mentioned above. The sample data for each dataset were collected by the sensor array, capturing data from all nine sensors simultaneously during a single test as one sample. Each type of grasping action data includes 500 samples, with each sample comprising data from five tests, resulting in a total of 3500 samples for the dataset. Each class of data is stored in separate folders. The dataset is divided into training, validation, and test sets, containing 2500, 500, and 500 samples, respectively.

To further validate the potential of CNN-based deep learning in this application, we implemented a CNN algorithm using the PyTorch library (Python version: 3.10.4; PyTorch version: 2.5.1+cpu). The CNN architecture, illustrated in Figure 5a, is designed to recognize grasping signals from various objects. This network consists of an input layer, two convolutional layers, two pooling layers, a fully connected layer, and an output layer. The input layer has nine channels with dimensions of 1 × 9 × 150. Here, the first “1” indicates that the data are one-dimensional; the “9” represents the nine input channels; and the “150” denotes the specified length of the data in each sample. For data that are shorter than this length, zero-padding is applied before inputting them into the CNN to reach the specified length. For data longer than this length, the excess part is truncated. The choice of data length should take into account the actual range of data lengths to ensure that the specified length can accurately reflect the data information.

In the first convolutional layer, a convolutional kernel of size 3 is employed, utilizing 32 kernels, a stride of 2, and a padding of 1. Padding is applied to prevent the length of the feature map from decreasing due to edge effects. The convolution is mathematically expressed as

$$O\left[i\right]={\displaystyle \sum _{j=1}^{k}X\left[i+j-1\right]}\cdot W\left[j\right]+b$$

(9)

where O[i] denotes the value of the output feature map at position i after convolution. In this context, X is the input, W represents the kernel weight, b is the bias term, k denotes the kernel size, and i and j are indices associated with the output and kernel, respectively. Following each convolutional layer, a ReLU activation function is applied, defined as

$$ReLU\left(x\right)=\max \left(0,x\right)$$

(10)

where x represents the input. This non-linear activation function enhances the network’s expressiveness, allowing it to model complex patterns more effectively. The first pooling layer has a window size of 2, which reduces the feature map size by selecting the maximum value within each region. This pooling process is mathematically expressed as

$$O_{pool}\left[i\right]=\max \left(X\left[i:i+k\right]\right)$$

(11)

where $X\left[i:i+k\right]$ is the value at window i, and k is the pooling window size. This operation reduces computational complexity while preserving essential features. The second convolutional layer uses a kernel size of 3, with 64 kernels, a stride of 2, and a padding of 1. The second pooling layer also has a window size of 2. Following two layers of convolution and pooling, the features are flattened into a one-dimensional vector and fed into the fully connected layer, which contains 576 neurons, mathematically expressed as

$$O_{fc}=XW+b$$

(12)

where X denotes the input feature vector, W represents the weight matrix, and b is the bias term. The output of the fully connected layer is subsequently mapped to seven categories. To assess the model’s performance, we utilized the cross-entropy loss function, defined as

$$Loss=-{\displaystyle \sum _{i=1}^C{y}_i\log \left(p_i\right)}$$

(13)

where C is the number of categories, $y_i$ is the one-hot-encoded true label, and $p_i$ is the predicted probability. A dropout layer with a dropout rate of 0.5 was added during training to prevent overfitting, and the dropout operation is given by

$$O_{dropout}\left[i\right]=\left\{\begin{array}{cc}0& with probability p\\ X[i]/(1-p)& with probability 1-p\end{array}\right.$$

(14)

where p is the dropout rate, $X\left[i\right]$ is the i-th element of the input features, and O_dropout is the output after dropout, which reduces the model’s dependency on specific neurons.

Figure 5b illustrates the loss and accuracy curves for both the training and validation sets. After 80 epochs, the loss decreased, and the accuracy increased to satisfactory levels. Figure 5c presents the confusion matrix for the training and test sets of the dataset comprising various grasped objects, where the horizontal axis denotes predicted labels, and the vertical axis represents true labels. Following 80 epochs, the accuracy of the training set reached 99.88%, while the accuracy of the test set reached 99.60%, indicating a high level of recognition accuracy for the model. This is consistent with recent findings on the effectiveness of machine learning-assisted sensors for motion and gait recognition [60] and is also in line with the theoretical basis related to tactile recognition [61].

Further, we expanded the range of participants. The selected participants’ information is listed in Table S3 of the supplementary document, consisting of five males and three females, for a total of eight people. The participants’ ages ranged from 20 to 30. Each participant tested each task 100 times with their right hand. The data obtained were used to build a dataset for different populations. Given the increased complexity of the samples, we appropriately increased the complexity of the CNN. The number of neurons in each layer was increased to 128, the stride of each convolutional layer was adjusted to 1, the kernel size of the pooling layer was adjusted to 2, and the number of training iterations was increased to 150. Finally, as can be seen in Figure S1, the classification accuracy of the training set reached 95.30%, and that of the test set reached 93.75%. It can be observed that after increasing the complexity of the dataset, the accuracy rate decreased slightly.

3.3. Throwing a Ball

To demonstrate the system’s real-time capability, we collected data during a ball-tossing motion. Figure 6a illustrates the four stages of this action: initially, the hand is empty and in a preparatory phase (Figure 6a(i)); next, the ball is placed in the hand (Figure 6a(ii)); then, the ball is tossed upward, leaving the palm (Figure 6a(iii)); and finally, the ball is caught after being tossed (Figure 6a(iv)). Figure 6b displays the data collected throughout these four stages. In stage i, the signal strength across all channels is low, generally remaining below 0.3 V. In stage ii, with the ball in hand, the signal strength in channels 2, 5, and 7 significantly increases, reaching approximately 0.5 V, 1.3 V, and 0.7 V, respectively. In stage iii, as the ball is tossed, the upward acceleration causes the signal strength in channel 5 to rise again; however, once the ball fully leaves the hand, the signal will return to its initial level. In stage iv, at the moment of catching the ball, the downward acceleration causes the sensor that first comes into contact with the ball to produce a signal amplitude greater than that when normally holding the ball in the hand. At this point, the sensor corresponding to channel 2 in Figure 6b is the first to come into contact with the falling ball, which is exactly reflected in the data. Once the ball stabilizes in the palm, the signal strength will stabilize again. The changes in the signal can clearly distinguish each stage. Due to slight variations in ball-gripping positions each time, minor differences in channel signal strength are to be expected. Moreover, when testing other individuals, similar results can be obtained, which is reflected in Figure S2a,b in the supplementary document. Additionally, in Figure S2c, during consecutive throws, each throw can be distinguished from the changes in the signal. This action sequence demonstrates that the system possesses excellent real-time responsiveness for capturing dynamic signals, making it well suited for measuring signal changes during dynamic processes.

4. Conclusions

In this study, we developed a flexible sensor array system capable of perceiving and identifying objects of various shapes and sizes. Utilizing a custom-designed multi-channel acquisition and signal transmission circuit, the system enables real-time signal transmission and has successfully collected data on two different-sized spheres, three different-sized cylinders, a cone, and a rectangular block. Leveraging these datasets, a convolutional neural network (CNN) was employed to train a deep learning model, achieving an impressive overall recognition rate of 99.60% on the test set. When testing groups composed of different genders, the model achieved a classification accuracy rate of 93.75% on the test set. Furthermore, during dynamic tests—such as the ball-tossing motion—the system accurately distinguished different action phases, which also applied to others repeating the same action. This research significantly contributes to the field of artificial tactile technology, providing valuable insights for the future development of more intelligent robotics.