Recent reviews have surveyed advances in tactile sensing for human–robot interaction (Jassim et al., 2025) and in pushing and grasping manipulation in robotic arms (Efendi et al., 2025). In contact detection, piezoresistive sensors are widely used to detect contact onset and localize interaction points. Early studies have shown that soft sensor arrays are capable of producing real-time force maps with sub-millimeter precision, which is advantageous for localization in manipulation, establishing an example for feedback-driven grasp control (Hammond et al., 2014). Furthermore, recent work embeds piezoresistive networks directly into compliant robot hands and grippers, allowing them to maintain stable grasps when vision is unavailable, which includes fully 3D-printed wearable finger arrays for pressure-point localization and large-strain piezoresistive skins that also enable proprioceptive estimation and object classification (Pei et al., 2021; Yong et al., 2022). Moreover, in prosthetics, printable piezoresistive composites and skin-inspired tactile elements have been integrated to support safe interaction and haptic perception in unstructured environments (Lathers et al., 2017; Wu et al., 2018).

On control strategies, piezoresistive sensors enable slip detection and closed-loop grip adjustment by capturing resistance fluctuations linked to shear forces. This capability has been demonstrated in both rigid and soft grippers, allowing real-time adaptation when handling fragile objects that require minimal force. Slip detection methods can be broadly categorized into three approaches. The traditional model approach utilizes thresholds or model-based rules that focus on normal force rate, tangential components, or friction models to identify incipient slip (Stachowsky et al., 2016; Liu and Howe, 2023; Deng et al., 2017). In contrast, electrical signal detection methods directly analyze electrical readouts from piezoresistive arrays, utilizing per-taxel voltage drops or current spikes during micro-slip as indicators (Chen et al., 2024; 2021). Changes in directional force, such as an increase in specific direction, can also indicate the onset of sliding (Zhang et al., 2015). These methods can be sensitive to contact orientation, as object contact at an oblique angle can introduce normal-shear coupling, which biases tangential force estimates. This confusion may lead to incorrect detection unless pose-aware compensation or decoupling techniques are applied. Slip has also been inferred from contact acceleration or vibration, which typically requires dedicated inertial sensing elements and higher-bandwidth sampling and processing (Howe and Cutkosky, 1989). Consequently, their applicability can be constrained by the complexity of sensor integration and their sensitivity to non-slip vibrations. Lastly, the learning approach leverages time-space patterns from sensor arrays to generalize across various objects and grasp types. The limitation here mainly concerns the need for large labelled datasets to effectively cover a variety of objects, surfaces, and angles during the learning process (Zhao et al., 2025; Wang et al., 2025a).

On the application side, many studies focus on the positioning of the sensing surface parallel to the object, in which the sensing surface is placed parallel to the object by using either parallel grippers (Sui et al., 2024; Stachowsky et al., 2016) or anthropomorphic hands (Gong et al., 2021; Zhang et al., 2015; Deng et al., 2017) with fingertips with near-normal contact. However, in real-world grasping situations, each finger makes contact with the object at different oblique angles. Recently, some researchers have started using anthropomorphic hands equipped with tactile sensors in the fingertips, utilizing advanced machine learning techniques (Zhao et al., 2025; Chen et al., 2025). These developments highlight the need to create designs that can robustly adapt to pose variations and develop testing protocols that accommodate a range of contact angles.

Most tactile sensor-based slip detection methods are typically designed for parallel grippers, which assume that the contact between the fingertips and the surface of an object is nearly parallel. In contrast, anthropomorphic hands often make contact with an object at various angles across different fingers, creating a challenge for current slip detection systems. The proposed method addresses this issue by introducing a per-finger closed-loop slip controller. This controller uses force readings from a tri-axial piezoresistive force tactile sensor and infers slip from temporal changes in the resultant tangential force between consecutive readings, referenced to a runtime no-slip baseline. Therefore, the proposed controller avoids the need for explicit friction modeling, data-driven slip classifiers, inertial sensing elements, or frequency-domain processing. Here, calibration-free denotes that no additional user-performed calibration, object- or pose-specific tuning, or friction-parameter identification is required beyond the factory force output from the sensor and a short runtime baseline acquisition. This approach is designed to be robust against various uncertainties associated with the object, such as rigidity, weight, and surface textures. Additionally, the per-finger sensor configuration allows for independent slip detection for each finger. Control is only activated for the finger that detects a slip, which prevents excessive force from being applied by the other fingers on the object. This simple approach ensures robust slip detection with good object conformability while maintaining safe and stable grasp control.

This section presents a slip detection algorithm that controls the grasping of each finger using tri-axial tactile feedback. The algorithm has been implemented and validated on a real-world robot setup, as shown in Figure 1. This method involves monitoring the force in the z-direction to ensure that the object remains engaged. Changes in resultant force within the xy-plane between consecutive readings serve as indicators of potential slip. When a slip is detected, commands are sent to the corresponding motor to adjust the gripping force accordingly. Moreover, a circuit protection mechanism is integrated to prevent electrical current overload in the motor. This integrated approach allows for calibration-free, real-time slip detection and correction across the fingers without prior knowledge of the properties of the object.

The pseudocode is shown in Algorithm 1, and was applied to each finger independently. The force readings from the tactile sensors were stored in the matrix F raw with three column vectors f x raw , f y raw and f z raw , as shown in Equation 1, which at any iteration i F raw i = f x raw i f y raw i f i raw i ⊤ (1)

Three objects with different rigidity, weight, and surface textures were selected for this study, as shown in Table 1. As mentioned in Section 3.1, the orientations of the sensors and the object during the grasping process were not guaranteed to be parallel. To mitigate this issue and ensure a consistent initial contact, a fixed value for f z T H was empirically determined for all fingers before all experiments. Here, a monotonic closing motion was commanded to all fingers, in which all P motor were increased from zero to the maximum motor, while f z control was logged. The minimum value at which contact was reliably detected across trials was then determined by visual inspection using a trial-and-error method, starting from a value of 50 mN with an increment of 50 mN per trial. The results were illustrated in Table 2, which found that the same f z T H value was sufficient for all fingers to touch these objects successfully. This threshold was selected once for the platform and then kept unchanged for all fingers and objects in this study.

After determining f z T H , this value was plugged into the algorithm, and the object engagement test was carried out to test its validity. The results are shown in Figure 3. The same f z T H value was used in all subsequent experiments. Also, the maximum and final forces after stabilizing for each finger were listed in Table 3. From these results, a few observations were obtained. First, f z control slightly exceeded f z T H after touching. The difference was caused by the fixed step increment in the motor control, which resulted in stepwise motion, introducing a small overshoot in f z control . Also, it was observed that the sensor values dropped after reaching their maximum values. It was mainly caused by the restoration of the shape of the elastomeric layer in tactile sensors. Moreover, if the object was soft, such as a sponge, the fluctuations were found to be even more rigorous than those of rigid objects, which was mainly due to the shape restoration of the soft object.

The results demonstrated that all motors could be effectively controlled using feedback from corresponding sensor readings, and their positions were maintained once f z T H was reached. This showcased good conformability to different objects using the proposed method.

After finding f z T H , the slip-detection algorithm was implemented to evaluate its responsiveness and robustness. Two parameters were required to be determined for the deployment of this algorithm. The first was Δ f x y T H , which was used to determine whether a slip event occurred or not. Here, a slipping test was implemented to evaluate the value. After gently touching the object, the hand was moved upward at a speed of 100 mm/s to evaluate Δ f x y of the tested objects under slipping conditions. The results are shown in Figure 4, which clearly demonstrates that there was a rigorous fluctuation in Δ f x y when slipping occurred. This finding provided strong evidence in support of the proposed slip detection method. This parameter required a balance on sensitivity. If Δ f x y T H was set too high, slip events went undetected, leading to false-negative outcomes. Conversely, if Δ f x y T H was set too low, the detector was vulnerable to mechanical vibrations and sensor noise, resulting in false positive results. From the results, it was shown that a single value for Δ f x y T H was efficient for all tested objects, as depicted as blue dashed lines in Figure 4. The second was P motor T H . As the grasping force increased with finger bending, this value directly influenced the maximum gripping force exerted on the object. For rigid or semi-rigid objects, a larger threshold was deemed acceptable since these objects could sustain a relatively greater gripping force without damage. Conversely, for soft objects, such as the sponge considered in this study, the threshold was restricted to prevent excessive deformation or damage of the object while ensuring grasp stability. The values for Δ f x y T H , P motor T H , and other parameters required for the algorithm are listed in Table 2. Unless mentioned otherwise, the values in the table were used throughout the study.

Then the slip detection test was carried out, and the results are shown in Figure 5. Also, the force variations and response time were shown in Table 4. From the results, it was shown that slippage could be successfully detected through monitoring Δ f x y . When there was an abrupt change in Δ f x y , the motor responded quickly by increasing P motor to increase the gripping force, which was evidenced by the sudden increase in f z control and the elimination of Δ f x y . As a result, the gripping force was increased significantly to stop the slippage from continuing. Additionally, it was observed that all motors were halted promptly once the slippage was eliminated. On the other hand, there were occasional decreases in P motor , which resulted from the circuit protection mechanisms designed to prevent overcurrent and protect the circuit. Furthermore, the motor response time shows that the proposed algorithm responds quickly once a slip event is identified.

Therefore, slippage was successfully stopped on all tested objects by controlling P motor based on sensor readings. This quick response ensured that no excessive force was applied to the objects, thereby minimizing the risk of damage and highlighting the effectiveness of the detection method used.

The detection algorithm was also evaluated against different lifting speeds. In this study, an aluminium rod was lifted at three vertical speeds: 100 mm/s, 300 mm/s, and 500 mm/s. The results are illustrated in Figure 6. Also, the force variations and response time were shown in Table 5. It was found that for vertical speeds of 100 mm/s and 300 mm/s, the f z T H defined in Section 3.4 was sufficient to detect slippage. However, when the vertical speed increased to 500 mm/s, f z T H needed to be increased slightly to 300 to stop slipping successfully. It was found that as the vertical speed increased, a lower f z T H could still allow for slip detection, but the response time was not fast enough for the hand to make the necessary corrections. This delay was attributed to the sampling speed of the sensors and the response of the motors during each iteration of the slip-detection phase. Therefore, it was necessary to increase f z T H to compensate for this issue. This change reflects a control setpoint adjustment for higher speed operation, and it does not involve any additional sensor calibration.

The results showed that, despite varying lifting speeds, P motor remained consistent. The robot hand applied a similar gripping force on the aluminium rod in each case. This confirms the effectiveness of closed-loop force control for slip detection, which proved reliable regardless of the lifting speed.

The robustness of the algorithm to sudden external disturbances was also evaluated in this study. After completing the slip-detection and establishing a stable grasp, the object was held in a stationary position. A manual pull was then applied vertically to simulate a sudden external disturbance. The results, shown in Figure 7, illustrated the response following the vertical lifting.

It was observed that this sudden pull induced a rapid fluctuation in Δ f x y . In response to these abrupt changes, P motor of all fingers were quickly adjusted to accommodate them. This adaptation resulted in an immediate increase in gripping force, which was indicated by the sudden increase in f z control . Furthermore, after the external disturbance was removed, some motors showed a decrease in P motor as a result of the circuit protection implemented within the algorithm while maintaining the gripping force.

Overall, the test results verified the robustness of the proposed algorithm against sudden external disturbances, as well as its ability to maintain a safe and stable grasp after the disturbance was removed.