Over several decades, robots have replaced humans for greater efficiency and manipulation of objects in various tasks in diverse environments, such as industrial automation, repetitive tasks, and social interaction/assistance. Hence, significant steps have been taken to enhance the skills of robots and bring them closer to human capabilities (Wei et al., 2023; Chuang, 2024). Robots are typically programmed with the capability to replicate intricate and multifaceted tasks. To achieve this, researchers define six important indicators, including kinematic architecture, activation, transmission, sensing, materials, and construction to enhance their functional dexterity (Controzzi et al., 2014). In line with sensing, most research efforts have focused on developing surface artificial tactile sensors that provide sufficient information for dexterous manipulation. This development includes the integration of sensors for detecting materials' characteristics, such as object edge shape (Suwanratchatamanee et al., 2007), surface properties like hardness and texture (Johnsson and Balkenius, 2008; Bok et al., 2021), and material discrimination (Lee et al., 2021). These sensors use capacitive (Tavakoli et al., 2017), piezoresistive (Fonseca et al., 2023; Kappassov et al., 2015), optical (Jiang et al., 2015), and magnetic (Kim et al., 2018) technologies, all of which are activated by mechanical stimulation. The ultimate aim is to equip robots with complex tactile perceptions similar to those of humans (Jamali and Sammut, 2011).

To assist robots in dexterous manipulation, embedded sensors must receive sufficient information from the target to effectively interact with their surroundings. For material detection, the proposed methods usually rely on sensor rubbing/friction and acoustic-based tapping on objects (Okamura et al., 1997; Spiers et al., 2016). For instance, Jamali and Sammut (2011) proposed an artificial silicon finger consisting of strain gauges and polyvinylidene fluoride (PVDF) films and embedding it into the gripper, which was able to classify eight object materials based on their texture. Friction between the sensor and different object textures induced vibrations in the silicon. Then, a classification accuracy of 95% was achieved using a Naive-Bayes tree (NBTree) classifier by extracting different frequencies. Shuaikang et al. (2023) embedded two force sensors at the fingertips (index and thumb) of the RH8D five-fingered robotic hand to measure object recognition based on hardness and texture through the force generated via rubbing the fingertips against objects. They classified seven materials via the support vector machine (SVM) algorithm using the Fourier transform (FFT) features with an accuracy of 86%. Tanaka et al. (2019) presented an artificial finger with different horizontal ridges to evaluate the effect of varying roughness on the vibration sensor output. The researchers rubbed nine objects of varying roughness with an approximate force of 0.3 N. The results indicated that various artificial fingers have different responses, reflecting differences in horizontal ridges. In 1996, Krutkov et al. introduced a method for finding materials based on their acoustic properties (Krotkov et al., 1997). To sort four materials into groups, they tapped the robot's end-effector into objects. Subsequently, other researchers presented alternative methods. Sinapov et al. (2009) used a Barrett WAM robot to create airborne sounds by performing tapping as an exploratory action on various objects. Their objective was to classify 36 different household objects using a self-organizing map (SOM), k-nearest neighbor (KNN), and support vector machines (SVM). The results showed that it is possible to identify these objects with an average accuracy of around 73%. The authors used a robotic system that implemented a latent regularized variational autoencoder (DLR-VAE) in Neumann et al. (2018). They recorded data by knocking on objects and grasping them. The encoder then mapped the input data into the latent space using a multiple-layer perceptron (MLP). Next, the features were extracted by transferring from the time domain to the frequency domain, resulting in the classification of eight materials with 76% accuracy.

However, the sensors introduced are difficult to replicate or usually bulky to achieve precise spatial resolution. On the other hand, commercial sensors, while providing good spatial resolution, only respond to stimuli that are normal to the sensor surface. To reduce wiring complexity, they use scanning techniques for data acquisition, thereby increasing the cost. Furthermore, the thickness and strength of the ridges influence mechanical stimulation. Therefore, it relies on tactile parameters that must be tailored to suit various applications (Agache et al., 1980). Furthermore, because acoustic-based material recognition necessitates the use of microphones, it has received less research than tactile-based material recognition, as soft materials don't generate sound that the microphone can record.

Although the presented methods have increased the performance of robots, the noise caused by tactile movements affects the system's performance (Chen et al., 2019). On the other hand, these methods rely on applying additional mechanical energy to enhance the perception of surface roughness (Hendriks and Franklin, 2010). Furthermore, texture-based material discrimination methods are not sufficient to distinguish between different materials with similar surface roughness. Consequently, object material recognition remains an open challenge. Hence, the primary motivation is to eliminate the influence of texture and material classification with a similar surface texture via an artificial finger that can be mounted on a robot. The novelty of our research is the approach in which, rather than using tactile methods to identify materials, we use the approach of damping signals by impacting the objects with a finger to acquire the signals. Our approach is based on a cantilever beam mechanism in which a load cell and an accelerometer are embedded in a cantilever beam made of an elastic metal sheet. Furthermore, we fixed a rigid metal appendage to impact the object at its free end. This setup allowed for simultaneous recording of the damping vibration-force signals when the metal appendage impacts with objects. The collected signals were then used as input for the classification model. Additionally, we used a clustering technique to reduce the machine's input data, thereby improving system performance. The presented approach allows machine learning to make an effective association between the robot and objects.

To summarize, the main contributions of this paper are as follows: (a) Reducing noise and signal interference compared to traditional tactile methods through damped vibration signals, (b) We present a straightforward method for detecting intrinsic material properties using a cantilever beam, removing the need for surface features. This design choice makes the system less sensitive to variations in roughness and more reliable in distinguishing similar materials, (c) Reducing the amount of data processed through clustering improves the system's efficiency, lowering computational load and increasing the robot's response speed.

The following section gives an overview of the remaining work. Section 2 presents the methodology. Section 3 describes the experiments and results. Then, we detail the discussion in Section 4. The paper concludes with Section 5.

Figure 1 describes the flowchart of the material detection of the objects with the proposed sensor via the UR5e robot, including: (a) A schematic of the proposed sensor embedded in the UR5e robot. (b) Acquisition of damping vibration-force signals through an impact approach and then clustering them using the beam deflection value. (c) Material classification based on the extracted features. Each section will be detailed separately.

Robotics science aims to enhance and extend capabilities by performing tasks that may be beyond human abilities or require an exceptional level of precision and consistency without experiencing fatigue. To achieve this goal, researchers have developed Human-Computer Interaction (HCI) systems that facilitate restoring more natural behaviors. The HCI systems are designed to restore natural behaviors and movements, ultimately closing the gap between human and robotic functionality. A robotic system introduces significant noise into the system and distorts useful information, which can significantly affect the performance of pattern recognition algorithms. While previous studies concentrated on external object properties such as texture and roughness, our analysis centers on intrinsic object features to provide simple and meaningful information for classifiers. This study focuses on material detection with minimal sensor-object contact to minimize contact noise, thereby addressing the aforementioned problems in the learning process.

In this context, Table 6 compares our proposed material detection method with other studies' methods. Gandarias et al. (2017) employed two distinct AI methodologies to classify materials using pressure images obtained from high-resolution tactile sensors. The experimental outcomes revealed that the classification accuracy achieved using Speed-Up Robust Features (SURF) was 80%, whereas the classification accuracy obtained using Deep Convolutional Neural Networks (DCNN) was notably higher at 91.67%. Liu et al. (2018) introduced a tactile framework for identifying 10 different materials using an efficient feature extractor called the linear dynamic systems-based fuzzy c-means method (LDS-FCM). They then used the vector of locally aggregated descriptors (VLAD) method to derive the final features from the data. Their approach achieved an accuracy of 99%. Dai et al. (2022) created three tactile sensor designs with varying protrusions. They used a 2-axis Cartesian robot to receive signals by dragging the sensor on objects' surfaces. Ultimately, they could classify six materials based on their texture with 98.1% accuracy using the KNN algorithm and time-frequency domain features. Wang et al. (2022) utilized a tactile sensor with 16 small capacitors to capture tactile data from the robot's finger sliding across eight different types of fabric material. They were able to achieve 96% accuracy in classification by reducing the frequency domain features' dimensions using principal component analysis (PCA) and the k-nearest neighbor (KNN) algorithm. In previous studies, tactile methods have often used an array of sensors, which increases system noise and leads to more input data, necessitating more powerful and costly processors in addition to the problems mentioned earlier. Receiving signals from multiple sensors leads to signal interference and requires more complex solutions. Designing a system requires balancing the number of sensors, speed, user-friendliness, and noise levels. It's essential to find an equilibrium between model complexity and computational cost. Additionally, reducing channels can lower hardware costs and complexity, decrease controller processing time, and maintain high classification accuracy. This study introduced a novel sensor and reduced the number of sensors to one, enhancing classification accuracy by reducing the input dimensions. It's important to note that comparing different studies typically focuses on the sensor type and feature extraction algorithms, often neglecting other critical factors like object standardization, contact pressure, and exploratory measures. As a result, making valid comparisons between studies is not a straightforward task.

Researchers have employed tactile and acoustic methods to enhance robots' capacity to identify object characteristics. Tactile methods are susceptible to factors such as unwanted vibrations and electrical signal interference, resulting in reduced accuracy and efficiency. Similarly, acoustic methods rely on microphones, which are not ideal for soft objects, and ambient noise can significantly degrade signal resolution. Both approaches demand substantial processing, leading to the creation of complex systems. However, literature has shown that contact control techniques can significantly increase the efficiency of robotic systems (Elguea-Aguinaco et al., 2023). Therefore, in this study, we utilized an integrated sensor employing the impact technique to minimize system noise and enable simultaneous measurement of two signals. The impact technique prevents excess noise transmission by establishing point contact, allowing the AI to focus on important information by reducing distracting signals. This approach reduces the system's processing load and enhances its detection accuracy and overall performance. By reducing the amount of processed data, the system can respond more quickly and accurately to environmental changes.

On the other hand, the feature space, defined by the type and number of features, significantly influences processor load (Asheghabadi et al., 2024). Therefore, minimizing the input data to the device streamlines the algorithm and lightens the processor load. Unlike conventional methods where all data is fed to the machine at once, in this study the data was clustered into two categories based on δ_j values before feeding into the AI. Figure 9 shows that clustering the vibration-force damping signals reduces the data input to the system. This approach enables the algorithm to utilize only relevant, high-quality data, eliminating complex feature extraction and unnecessary processing, thereby enhancing accuracy and processing speed. In particular, δ_j is a feature that is not included in the feature vector yet effectively divides the data and reduces the machine's workload. This method enhances the performance of approach A in object material classification by minimizing noise and interference, allowing the algorithm to identify more accurate patterns and improve classification accuracy. Compared to regular MFCC methods, this strategy reduced the dataset size by 50% and decreased feature extraction and classification times by 78% and 48%, respectively. This strategy significantly reduced the execution time. Ultimately, it optimizes processing time and boosts the efficiency of AI systems. Also, in previous studies, researchers classified materials by creating artificial textures on objects' surfaces, creating a predictable nature (Jamali et al., 2009; Edwards et al., 2008; Kim et al., 2005). Our work successfully classified materials based on intrinsic properties with the same natural surface. This approach also eliminated overfitting and enhanced system reproducibility. Features extracted from clustered data improved material recognition with minimal in-class variance, maximal out-of-class variance, and high repeatability. In this study, we exclusively used the δ_j value as the data clustering index (approach B). This index can be integrated with other methods, such as KNN, that employ Manhattan or Euclidean distance to create more boundaries for the data, dividing them into more categories.

The proposed method has some limitations as follows: (1) Sensitivity to the angle of impact when the sensor interacts with objects. Variations in the impact angle can affect vibration and force signal consistency, leading to inaccurate material classification. A possible solution is to use a mechanism that stabilizes the sensor's orientation or algorithms that normalize data to counteract angle variations, ensuring reliable signal acquisition. However, achieving this level of realism was beyond the scope of this study. (2) Reliance on a single-point contact for signal acquisition, while it reduces noise, might miss certain material characteristics that could be captured through a broader interaction. To address this, the sensor could be upgraded to include multiple points of contact or an array of miniaturized sensors. This would enable a more comprehensive collection of data across a larger surface area, providing a richer dataset that could improve the accuracy and robustness of the classification process without compromising the simplicity of the original setup. (3) The signals are often affected by uncertainties and inaccuracies, such as requiring preprocessing using tools based on fuzzy logic. To reduce computational load when dealing with large datasets, it is advisable to merge similar signals in a fuzzy manner (Versaci et al., 2020, 2022), thereby creating distinct fuzzy classes for each group and extracting a representative signal from each group.

It is expected that the proposed method will be used in other robot types, such as bionic hands (Moqadam et al., 2021a; Moqadam et al., 2022), where amputees face challenges interacting with the environment due to loss of touch sensation. In future work, in line with our previous studies, we intend to create sensory feedback in amputees (Moqadam et al., 2023a) with the help of myotome digits (Moqadam et al., 2023b) by embedding the proposed sensor in the hand prosthetic finger [which is controlled by electromyography signal (Bahrami Moqadam et al., 2018)] and making intelligent grasping to realize automated, robust, and accurate material detection.

This study presented an effective method for identifying objects' material by measuring damping vibration and force signals simultaneously. The direct proportionality between the measured force and the vibration of the elastic metal sheet led to the objects clustering into two groups, which reduced the input volume to the system and consequently improved classification accuracy. In this study, ten objects were classified using three different classifiers: MLP, SVM, and LDA, and the highest accuracy was 95.46% with MLP. Finally, the proposed method shows remarkable versatility, with potential applications not only in industrial automation but also in prosthetics, where the ability to recognize materials accurately can significantly enhance the interaction of robotic devices with their environment, making the research promising for future Implementations in various areas of advanced robotics.