Vibrissae-inspired vision-based magnetic-actuated whisker

Abstract

Tactile perception is significant for robotic operation in unstructured environments. Whisker-based sensors offer lightweight bio-inspired solutions, yet most rely on single-whisker sensing and are limited by passive interaction and constrained functionality. Here we present a circular array of eight independently actuated whiskers, each driven by a pulse-switchable permanent magnet and tracked by a camera. This design enables simultaneous multi-point sensing and coordinated actuation, supporting diverse functions. Quantitative analyses demonstrate accurate pixel-to-physical mapping, consistent pixel-to-force characterization, and long-term repeatability. In this work, we show that an vibrissae-inspired vision-based magnetic-actuated whisker array integrating distributed perception with active interaction achieves reliable physical mapping, repeatable sensing, and delicate grasping, thereby advancing tactile sensing and laying the foundation for contact-driven exploration, soft manipulation, and adaptive behavior in dynamic environments.

Introduction

Tactile sensing is essential in robotic systems, enabling robots to interact with and interpret their environments through physical contact. Objects can be characterized through contact, and properties such as shape, texture, and material composition can be inferred through tactile interactions. A wide range of tactile sensors has been developed to achieve these capabilities, including pressure, strain, vibration, and slip sensors^1,2,3. Among these, artificial whisker-based tactile sensors offer distinctive advantages due to their mechanical compliance and spatial sensing capabilities. Inspired by the vibrissae of rodents (Fig. 1a, b), whisker sensors detect external disturbances through deflections of slender structures, rather than requiring direct contact on the sensing element⁴. This indirect mode of contact enables detection at a distance from the supporting structure, minimizing interference with the outside world and reducing the risk of displacing or damaging delicate or unsecured items.

Fig. 1: Illustration of the vibrissae-inspired whisker array.

a Schematic of whisker-based tactile perception in mice. Active whisking enables tactile sensing and guides behavioral responses to external stimuli. b Mice can localize object position through small whisker movements. c Schematic of a natural whisker. Within each follicle, mechanoreceptors detect follicle rotation driven by intrinsic muscles or shaft deflection caused by external contact. These signals encode direction, velocity, duration of displacement, and torque⁵³. d Artificial whisker filament inspired by biological morphology. e Bioinspired whisker array designed for tactile sensing and object interaction. f Mechanical layout of the whisker array system. g Actuation mechanism of the array: whiskers rotate under magnetic torque generated by voltage-controlled pulses. h Perception mechanism of the array: a camera tracks the motion of each whisker via visual markers. i The vibrissae-inspired whisker array can be integrated with commercial robotic platforms for downstream tasks. j Demonstrated applications include object classification, airflow sensing, waterflow sensing, viscosity estimation, and object grasping. a Reproduced from refs. ^54,55, licensed under CC BY 4.0. Copyright 2020, the authors. b Reproduced from ref. ⁵⁶, licensed under the Creative Commons Attribution License. Copyright 2007, the authors, published by PLOS. c Adapted from ref. ⁵⁷ (licensed under CC BY 4.0; Copyright 2022, the authors, published by Elsevier) and ref. ⁵⁸ (open access under terms equivalent to CC BY; Copyright 2011, the authors, published by Frontiers).

Building on these biological principles, researchers have developed artificial whiskers using diverse transduction mechanisms, including strain gauges^5,6, piezoresistive^7,8, piezoelectric^9,10, capacitive¹¹, and triboelectric^12,13,14,15 methods. These methods can measure forces, displacements, and contact interactions, but directional perception remains challenging, as many designs detect contact presence without resolving its direction. To address this limitation, researchers have explored various strategies. An approach involves embedding multiple sensing elements within a single whisker^{16,17,18,19,20,21,22,23}, which enhances multidirectional sensitivity but introduces complex wiring and integration overhead. A second strategy arranges whiskers in geometric arrays to capture spatial contact patterns^{24,25,26,27,28}. While effective, these arrays pose challenges related to mechanical constraints, precise alignment, and calibration. A third approach embeds magnets into whiskers to detect magnetic field changes^29,30,31,32. This method enables non-contact detection and is useful in low-force applications, but it often suffers from interference between adjacent magnetic fields, which limits scalability. Although each method enhances directional sensing, they increase system complexity and limit scalability, highlighting the need for a simpler yet effective approach for multi-directional whisker sensing.

Vision-based tactile systems offer a promising alternative by enabling contact-free tracking of whisker motion across sensor arrays^33,34, while avoiding many of the integration challenges associated with embedded electronics. An early implementation by Kaneko et al.³⁵ used a charge-coupled device camera to capture full whisker motion, but its performance was highly sensitive to ambient lighting, limiting robustness in unstructured environments. With advances in miniaturized imaging technology, modern systems integrate compact cameras within the whisker mount structure, allowing whisker tracking by analyzing whisker base movement³⁶. This method improves resilience to lighting variation and enables real-time monitoring. With the high information density of imaging, modern designs further extend to multi-whisker arrays^33,34,37, providing richer spatial mapping with relatively low mechanical complexity. Nevertheless, vision-based sensing remains inherently passive. It perceives contact but lacks actuation for active exploration or object manipulation. Achieving both high-fidelity perception and interactive control requires coupling vision-based tracking with an actuation mechanism.

Currently, static whisker sensors are effective in detecting object presence, texture, and surface interactions, but they are limited in their ability to interact with and manipulate objects. To address this, active whisker systems have been developed, integrating actuation to mimic biological whisking. Rotational motors can sweep whiskers for surface scanning and object detection^7,38,39, but they increase bulk and mechanical complexity. Electromagnetic actuation, using pairs of electromagnets and permanent magnets, allows programmable movement^29,40 but suffers from high power consumption and overheating, limiting long-term operation. Tendon-driven mechanisms³⁴ and electroactive polymer (EAP) actuators^32,41 provide flexible, bio-mimetic motion but face scalability and durability challenges. Shape-memory actuation⁴² requires thermal cycling, which constrains speed and repeatability. Overall, existing active whisker approaches struggle to combine low power and continuous operation in a compact form factor.

Prior research in whisker-based sensing has demonstrated the potential of precise and compliant tactile perception, yet challenges remain in achieving directional sensitivity, maintaining low system complexity, minimizing power consumption, and enabling effective active control^4,29. Addressing these issues requires an approach that combines precise perception with low-complexity active control. To this end, we introduce a vision-based whisker sensor array that integrates compact magnetic actuation for interaction. To contextualize our contribution, a comparison between this design and representative whisker arrays reported over the past decade is summarized in Table 1, detailing the sensing principles, configurations, directional resolution, actuation methods, and manipulation capabilities.

Table 1 Comparison of representative whisker array sensors with the present study

Our system draws inspiration from the vibrissae of mammals, where whiskers emerge from follicles densely packed with mechanoreceptors (Fig. 1c). In nature, passive touch occurs as external stimuli deflect the loosely held follicle, enabling subtle contact and flow detection⁴³, while active whisking driven by muscle contractions supports environmental exploration and decision-making⁴⁴. Mimicking these biological mechanisms, our sensor incorporates eight whiskers arranged in a circular configuration. Vision-based tracking serves as an artificial mechanoreceptive system, providing real-time monitoring of whisker rotations for accurate contact-direction discrimination. Structurally, the elastomer membrane acts as an artificial skin, the whisker filament replicates the natural whisker shaft, and the iron elastomer composite functions as a synthetic follicle, together forming a bio-inspired analog of the whisker-follicle-skin system for passive sensing. Coupled with pulse-switchable permanent magnet (PSPM) actuation that mimics muscle-driven protraction, the system achieves both passive and active whisking. This explicit structural and mechanistic correspondence to natural vibrissae enables sensitive detection of subtle stimuli and controlled exploratory interactions, thereby improving both perception and manipulation performance.

In this work, we present a vibrissae-inspired whisker sensor that combines vision-based multi-whisker tracking with compact pulse-switchable magnetic actuation. This design enables continuous operation with low power consumption and programmable control, supporting tactile sensing including contact perception, stimulus classification, and flow detection. By extending the whisker array from passive perception to active interaction, the system achieves coordinated actuation for tasks such as gently grasping and manipulating fragile objects, which are capabilities that traditional static whisker sensors cannot provide. This fully integrated architecture establishes a practical and scalable platform for soft robotics, environmental exploration, and delicate manipulation, where precise and minimally invasive interactions are essential. This work lays the groundwork for safe robotic interaction with fragile biological organisms, with potential applications in biomedical handling and ecological studies.

Results

Sensor design

The system is a vision-based whisker sensor array with magnetic actuation that merges tactile perception and object manipulation. Building on our preliminary prototype⁴⁰, the present design introduces a fully re-engineered actuation mechanism and provides a comprehensive characterization of both sensing and manipulation. To contextualize the advance, we provide a qualitative multi-dimensional comparison with representative whisker-based sensors (Table 1). The system incorporates a Raspberry Pi camera (640 × 480 pixels, 60 frames per second), eight carbon-fiber whiskers (1 mm diameter, 105 mm length), and eight pulse-switchable permanent magnets (PSPMs) coupled to iron-elastomer composite cubes (hereafter referred to as the iron composite) at the whisker roots. The whisker array’s design is inspired by the whisking motion observed in mice, where sweeps of the vibrissae enable active exploration and contact localization (Fig. 1a–c). This biologically motivated configuration supports a centralized, circular arrangement of eight actuated artificial whiskers mounted on an elastomer membrane (Fig. 1d–e).

The overall sensor layout and internal structure are illustrated in Fig. 1f. Each whisker is anchored by an iron composite positioned directly toward a Pulse-Switchable Permanent Magnet, enabling bidirectional control. A forward pulse initiates attraction by activating the PSPM, while a smaller reverse pulse deactivates the PSPM and releases the whisker (Fig. 1g). A camera positioned beneath the iron composites tracks their motion, enabling real-time inference of whisker displacements (Fig. 1h), while a supplementary light source ensures uniform illumination.

The system operates in two modes. In the active mode, whiskers are magnetically actuated to engage with objects, with their motion continuously monitored by the camera. In the passive mode, with the PSPMs deactivated, whiskers respond to external forces, functioning purely as tactile sensors. This dual-mode architecture, which combines tactile sensing and active interaction, enables more adaptive and complex robotic behaviors while maintaining a low power footprint. This system facilitates real-time tactile feedback while enabling magnetically controlled whisker movement, thereby enhancing the system’s ability to interact with and manipulate objects in complex environments.

Actuation characterization

To evaluate the actuation performance of the whisker array, three sets of experiments were conducted, which include measuring the voltage-angle relationship, quantifying the forces exerted during actuation, and assessing payload capacity during grasping.

An experiment was conducted to evaluate the relationship between the applied voltage to each PSPM and the resulting whisker rotation angle, in order to characterize the whisker movement behavior under different PSPM actuation voltages. The rotation angles of all eight whiskers across the range of actuation voltages are shown in Supplementary Fig. 1, while Fig. 2a illustrates representative images of a single whisker under different actuation voltages. A rotation angle of zero represents the whisker’s neutral position. As shown in Supplementary Fig. 1, the relationship between applied voltage and rotation angle closely follows the ascending branch of the magnetization curve for Alnico 5⁴⁵, which reflects the transition from a demagnetized state to magnetic saturation. The variation in the specific voltages at which individual whiskers reach relative saturation angles can be attributed to slight differences in the magnetic coupling efficiency between the PSPM and the iron composite. Small angular or positional misalignments, as well as minor air gaps or inconsistencies in spacing introduced during assembly, can alter the magnetic flux linkage and the overall reluctance of the magnetic circuit. These factors collectively contribute to the observed variability in the voltages required to achieve comparable rotation angles across the whiskers. The peak rotation angles also exhibit slight variations across the eight whiskers, which can be attributed to nonuniform mechanical properties of the elastomer membrane and variations in adhesion quality between each whisker and the membrane. Despite these variations, the experiment successfully validates the sensor’s actuation method, demonstrating that at 24 V, all whiskers are fully actuated, reaching the sensor’s retracted state, where the whiskers are maximally deflected toward the center.

Fig. 2: Characterization of whisker actuation and sensing performance.

a RealSense camera images showing a single whisker under different actuation voltages. b Standard deviation of pixel displacements along the x and y directions for each subgroup (one whisker, one distance, one angle) in the perception consistency experiment. c Forces exerted during actuation by the eight whiskers, measured 15 mm from the membrane (left y axis) and estimated at the tip (right y axis). Data represent mean ± 1 standard deviation (SD) across five trials. d Grasp success rates across different payloads. Robust grasp is defined as success probability > 50%, with success rate thresholds of ≥90% and ≥80% (dashed lines). e Absolute error histograms comparing fitted and original models for mapping pixel displacements to whisker deflections in the calibration experiment. f Displacement-force calibration of the eight whiskers. Measured forces (left y-axis) are compared with estimated tip forces (right y-axis). Solid lines denote mean values, shaded regions represent ± 1 SD, and dashed lines indicate linear regression fits. g–i Enlarged views of panel (j) showing whisker displacements over the first 10 cycles, middle cycles (495–505), and final 10 cycles in the repeatability test. j Displacement of a single whisker over 1000 activation-deactivation cycles in the repeatability experiment.

The forces exerted during actuation were next characterized. Because the tip forces were close to the resolution limit of the force gauge, measurements were instead performed at a position 15 mm away from the membrane surface, which was corresponding to 16.5 mm from the pivot point, with five repeated trials for each whisker. The equivalent force representation used to estimate tip forces is detailed in Supplementary Note 1 and Supplementary Fig. 2. As summarized in Fig. 2c, the whiskers produced an average measured force of 4.488 mN at the test location, corresponding to an estimated tip force of 0.775 mN, with a standard deviation (STD) of 0.597 mN for the measured force. The resulting sub-milli Newton tip force highlights the capability of the whisker array to perform gentle interactions, making it particularly well-suited for delicate grasping tasks.

Payload characterization was further performed by evaluating the whisker array grasping a container of varying weight (1.2 g upward in 0.1 g increments) under a 24 V, 2 ms activation pulse. Each weight was tested in ten grasping trials. Robust grasp payload is defined according to Dex-Net 2.0⁴⁶, where a grasp is considered robust if its success probability exceeds 50%. Based on binomial confidence intervals, a 95% confidence level requires a success rate of at least 90% over ten trials, while a 90% confidence level corresponds to a success rate of at least 80%. Under these criteria, the robust grasp payload was determined to be 1.8 g at the 95% confidence and 2.0 g at the 90% confidence (Fig. 2d). These results demonstrate that the whisker array achieves reliable grasping performance in the gram-scale payload regime. While the maximum payload is modest compared with rigid grippers, the sub-gram to few-gram payload range is sufficient for the intended application in delicate manipulation, where low contact force and compliance are critical. The observed payload also highlights the tradeoff between grasp robustness and payload capacity, reflecting the sensor’s design emphasis on gentle interaction rather than heavy lifting. Importantly, the measured payload reflects the coupled effects of friction, normal pressure, and whisker-membrane adhesion, and should therefore be interpreted as a practical indicator of grasping performance rather than a direct measure of intrinsic actuation force.

Perception characterization

To evaluate the perception capability of the sensor, two experiments were conducted to quantify its ability to convert whisker movements into measurable outputs. The first experiment defined the mapping between pixel displacements and physical whisker movements. The second experiment defined the mapping between pixel displacements and externally applied forces.

The first experiment characterized the sensor’s perception by calibrating vision-based tracking of whisker movements against controlled physical displacements. This calibration ensures that raw image data can be reliably translated into tactile feedback, forming the basis for subsequent sensing and manipulation. For each whisker, at each displacement and angle, five trials were performed to record pixel displacements in the x and y directions, with axes defined in Fig. 1f. The resulting standard deviations for each set of five repeated trials under fixed displacement and angle conditions (Fig. 2b) were mostly below one pixel, with mean values of 1.03 pixels in x direction and 0.81 pixels in y direction. Inter-Class Correlation Coefficients (ICC) consistently exceeded 0.98 across all conditions (Supplementary Table 1), confirming the repeatability and stability of the tracking approach. Within this experiment, the baseline mapping model (Original Model, Supplementary Notes 2, 3) had a mean absolute error (MAE) of 1.525 mm and a root mean square error (RMSE) of 1.936 mm in linear displacement. However, as whiskers were not perfectly vertical at their neutral state (Fig. 1e), the true pivot points often deviated from the tracked roots. To address this, the Levenberg-Marquardt algorithm^47,48 was applied to fit each whisker’s pivot position and refine the mapping between pixel displacements and physical movements (Supplementary Note 3, Supplementary Table 2). The resulting Fitted Model reduced the MAE to 0.859 mm (RMSE = 1.134 mm, Supplementary Table 3) and further enabled angle estimation with an acceptable bias of 1.9% of full scale (Supplementary Table 4). Absolute error histograms (Fig. 2e) show that the fitted model substantially improves accuracy, particularly at larger displacements where pixel tracking is more reliable. Overall, this refined calibration effectively accounts for real-world deviations such as non-vertical whisker orientations, providing consistent estimates of both linear displacements and angles and offering a robust foundation for advanced tactile sensing.

The second experiment characterizes the sensor’s force perception by calibrating vision-based tracking of whisker movements against controlled force measurements. This calibration extends the pixel-to-physical displacement mapping to enable direct estimation of applied forces. A linear regression model was used, as pixel displacement scales linearly with torque and force according to the analysis in Supplementary Notes 1, 4, and 5. Across all whiskers and orientations, the model achieved a MAE of 0.271 mN and RMSE of 0.367 mN (Supplementary Table 5). This linear mapping is consistent with the theoretical framework, confirming that pixel displacement can be robustly translated into applied force. Figure 2f summarizes the displacement-force calibration results for all eight whiskers. The plots compare measured forces (left y-axis) with the corresponding estimated tip forces (right y-axis). Across whiskers, the mean trends (solid lines) exhibit clear linearity, with narrow variability captured by the shaded ±1 standard deviation (SD) bands. These results illustrate that whisker pixel displacements provide consistent and accurate estimates of milli-Newton scale forces. Histograms of absolute errors (Supplementary Fig. 4) further indicate that the regression model maintains high accuracy and repeatability across angles. Overall, this refined calibration provides a robust mapping between visual tracking signals and real-world contact forces, forming the foundation for force-sensitive tactile exploration.

Repeatability tests

To evaluate the long-term perception and manipulation stability of the whisker array, two repeatability tests were conducted. The first focused on component-level stability through repeated activation-deactivation cycles, while the second examined task-level stability through object grasping across multiple days.

In the first experiment, the whisker array underwent 1000 activation-deactivation cycles. During each cycle, the whiskers were first fully actuated by electrical pulses, causing them to rotate. Then the whiskers were deactivated and returned to their resting positions. The pixel displacement of each whisker was tracked using the onboard camera system. For each cycle, two displacement values were recorded per whisker, one at the fully actuated state and one after returning to rest. To eliminate the influence of transient oscillations following deactivation, displacement was sampled only after the whisker reached a stable state. An example trajectory of one whisker’s displacement across 1000 cycles is shown in Fig. 2j, with each cycle contributing two sequential data points. Displacement trajectories of all eight whiskers are presented in Supplementary Fig. 5, and corresponding mean and standard deviation values are summarized in Supplementary Table 6. During activation phases, mean displacements ranged from 13.319 to 18.966 pixels, with standard deviations between 0.273 and 1.578 pixels. During the deactivation phase, mean displacements ranged from 0.247 to 1.750 pixels, with standard deviations between 0.132 and 0.345 pixels. These results demonstrate high mechanical consistency across repeated cycles. While minor drift was observed in the displacement of some whiskers over time, the magnitude of such drift was small and did not noticeably affect the sensor’s perceptual accuracy or manipulation performance. No progressive deformation or mechanical degradation was observed across the 1000 trials. The consistent displacement responses also support the repeatability and reliability of perception based on whisker motion tracking. These findings confirm that the whisker array exhibits long-term repeatability and structural stability, which are essential for reliable operation in extended tactile sensing and manipulation tasks.

As a complementary task-level evaluation, object grasping was repeated on the same object (pom-pom) and setup over a ten-day period. Ten trials were conducted on days 1, 2, 5, and 10, and all trials succeeded (100% success rate). This ceiling outcome is consistent with stable end-to-end performance under the chosen task conditions, but because of the absence of failures, it does not provide sensitivity to subtle temporal degradation. Therefore, the main evidence for long-term repeatability comes from the 1000-cycle activation-deactivation experiment, while the grasping test serves as a confirmatory check that manipulation performance remained robust across days.

Airflow sensing

To evaluate the whisker array’s response to different airflow speeds, an airflow sensing experiment was conducted. A Honeywell HT-900 fan was used with three speed settings: high, medium, and low, corresponding to measured airflow speeds of 5.0 m ⋅ s⁻¹, 4.7 m ⋅ s⁻¹, and 4.3 m ⋅ s⁻¹, respectively. The average peak and valley values across the eight whiskers for each mode are presented in Table 2, and the pixel displacement of a single whisker over a 2-s interval for each airflow speed is presented in Fig. 3a. Additionally, a Fast Fourier Transform (FFT) was applied to the recorded signals to analyze the frequency characteristics of the whisker movements. The results indicate that, across all airflow conditions, the dominant frequency of whisker oscillations remained relatively consistent, averaging around 5 Hz. This consistency arises from the whisker’s cantilever-like structure, where oscillation frequency is primarily dictated by its principal natural frequency rather than airflow speed.

Fig. 3: Sensor perception of airflow and water flow stimuli.

a Airflow experiment: Pixel displacement of a single whisker over a 2-s interval under different airflow speeds. b Airflow experiment: Pixel displacement of a single whisker under different airflow speeds, including periods where the whisker made contact with an external obstacle. c Water flow experiment: Time-domain response of a single whisker over a 1-s window under different motor input frequencies. Alternating yellow and white bands denote different frequencies. d Water flow experiment: Estimated dominant whisker response frequencies computed via FFT across all whiskers and trials. The horizontal axis represents the motor’s actual oscillation frequency, while the vertical axis shows the corresponding whisker response frequency. The size and color of each scatter point represent the occurrence count of each frequency pair (i.e., larger and darker points represent more counts). e Water flow experiment: Most commonly observed response frequency among the eight whiskers for each trial. The size and color of each scatter point represent the occurrence count of each frequency pair (i.e., larger and darker points represent more counts).

Table 2 Whisker displacement and frequency under different airflow speeds

While the oscillation frequency remains stable, the displacement amplitude clearly varies in a nonlinear manner with airflow intensity. Whisker displacements decreased with airflow intensity, from a mean of 9.59 pixels at high airflow to 6.74 pixels at medium and 5.69 pixels at low airflow. The nonlinearity arises because the aerodynamic force exerted on each whisker is proportional to the square of the airflow velocity, rather than increasing linearly. Consequently, even modest increases in flow speed produce disproportionately larger deflection forces, which lead to a nonlinear growth in mean displacement. Additionally, higher airflow intensities introduce stronger turbulent fluctuations, further amplifying the whisker deflection. This amplitude modulation enables the sensor to distinguish airflow intensities based on the magnitude of whisker movement, demonstrating its capability for passive aerodynamic sensing and environment monitoring.

To further evaluate the sensor’s ability to detect airflow in complex scenarios, the whisker array was mounted on a UR16e industrial robot and exposed to sequential airflow and contact events (see Supplementary Movie 1). The robot moved incrementally toward an obstacle while whisker displacements were recorded. A representative whisker’s response under multiple airflow and contact conditions is shown in Fig. 3b. Based on both frequency and amplitude characteristics, whisker states were classified into three categories: Origin State (white), Dynamic State (yellow), and Contact State (purple). The classification procedure followed a combined frequency-domain and amplitude-based approach, as detailed in the Methods section under Contact and airflow feedback classification. Specifically, dynamic feedback was identified by high spectral energy in the FFT, while contact feedback was detected through large-amplitude movements or sustained non-zero displacements.

Under airflow conditions, whiskers showed clear dynamic-state behavior, with consistent oscillatory motion. As airflow speed decreased, oscillation amplitude reduced while frequency remained consistent, confirming the sensor’s ability to detect airflow intensity changes. In the absence of airflow, whiskers remained in the origin state until contact occurred. Contact events, initiated by the robot’s forward movement at approximately 10 and 15 s, caused abrupt and sustained displacement increases, accurately classified as contact feedback across all airflow settings.

Although the classification system effectively distinguishes between contact and airflow states in most cases, some misclassifications occur, particularly near transitions between contact and non-contact phases. These errors may primarily arise from transient oscillations following contact initiation or release, where sudden elastic motion leads to brief settling oscillations. If these oscillations fall within the frequency range of airflow-induced vibrations, the spectral magnitude may exceed the dynamic-state threshold of 5, leading to misclassifications as Dynamic State instead of Contact State. Conversely, under medium and low wind speed conditions, some airflow-induced oscillations were misclassified as Contact State, likely because the small-amplitude oscillations failed to generate sufficiently strong frequency components to exceed the threshold, causing the system to rely on displacement alone. In these cases, sustained but low-amplitude motion was interpreted as contact rather than airflow. Additional contributing factors include the camera’s limited frame rate (60 frames per second), which can undersample rapid transient events, and low-amplitude mechanical vibrations or sensor noise, which can occasionally generate weak periodic signals that falsely trigger dynamic state classification.

Despite these occasional errors, the combined use of frequency-domain analysis and amplitude-based detection provides a generally robust classification framework capable of distinguishing between airflow-induced oscillations, contact-induced movements, and the original state. This demonstrates the whisker sensor array’s capability to simultaneously perceive and differentiate between airflow and contact events, enabling versatile environmental interaction sensing even in mixed-contact and airflow environments. This capacity is crucial for robotic systems operating in unstructured environments, where simultaneous contact detection and airflow perception are essential for adaptive behavior.

Water flow sensing

To evaluate the whisker array’s capability for fluid flow sensing, an experiment was conducted to characterize its frequency response to oscillatory water movement. The time-domain response of a single whisker over a 1-s window under different input frequencies is shown in Fig. 3c.

The estimated frequencies across all whiskers and trials are summarized in Fig. 3d. In this figure, each point represents an estimated frequency for a single whisker at a given motor input frequency. The point size and color encode the number of occurrences, where larger, brighter points indicate more frequent observations. The dashed line represents the ideal condition y = x, where the whisker frequency perfectly matches the motor input. The results show that in the intermediate frequency range (1.84 to 4.45 Hz), the estimated whiskers’ frequencies cluster closely around the ideal line, indicating accurate and reliable tracking of external flow oscillations.

Deviations occur at both low and high ends of the frequency spectrum. At lower input frequencies (less than 1.5 Hz), whiskers often fail to synchronize with the motor due to weak flow generation, resulting in dominant self-oscillatory behavior. At higher input frequencies (larger than 5 Hz), the whisker responses are again attenuated, with many responses clustering around 3 to 4 Hz regardless of input frequency. This reflects a structural filtering behavior, likely caused by a reduced flow amplitude at high speeds due to the limited torque of the motor and the viscous resistance of the fluid.

To further analyze overall system behavior, we computed the most common response frequency among the eight whiskers for each trial and plotted the results in Fig. 3e. This plot also encodes frequency occurrence using point size and color. It confirms that in the mid-frequency range, the whisker responses align consistently with the motor input. However, in both the low- and high-frequency regions, the dominant responses cluster around 3–4 Hz, consistent with the natural structural frequencies of the whiskers.

These results demonstrate that the whisker array functions effectively as a passive hydrodynamic flow sensor within a specific operational frequency band. In the range of approximately 1.84 Hz to 4.45 Hz, the whiskers exhibit consistent and accurate tracking of periodic water flow, making the system well-suited for detecting biologically relevant fluid motion and rhythmic disturbances. Performance outside this band is limited by under-excitation at low frequencies and attenuated fluid coupling at high frequencies. Despite these boundaries, the system offers several advantages, including non-intrusive sensing, no need for direct force measurement, and scalable array configurations. These features make it a promising solution for distributed flow sensing in soft robotics, aquatic navigation, and bio-inspired environmental interaction.

Viscosity sensing

To evaluate the sensor’s capability in measuring liquid viscosity, an experiment was conducted to observe the whisker array’s response to various liquids with distinct viscosity levels. The whisker sensor was tested across a range of fluid viscosities by actuating whiskers submerged 2 cm in different liquids for five times and analyzing displacement patterns during activation and relaxation (see Supplementary Movie 2 for representative trials).

The error band plots of the whisker movements during the experiment are shown in Fig. 4a. In these plots, the mean represents the average displacement of the eight whiskers, while the standard deviation (shaded region) indicates the variation among them. The results indicate a clear viscosity-dependent trend, where liquids with higher viscosity exhibited slower whisker displacement changes. Corn sirup, having the highest viscosity, showed the slowest response, followed by maple sirup, as evident from the gradual increase during activation and the slow decay during deactivation. The pixel displacement of a single whisker is plotted in Fig. 4b, with a magnified view of the 9.8–11 s interval in Fig. 4c to highlight the whisker behavior in different fluid conditions. The results show that lower-viscosity fluids produced steeper slopes upon deactivation. In air, where frictional resistance is minimal, the whisker exhibited an under-damped response. In water and canola oil, the whisker movement was approximately critically damped due to their relatively low viscosity. In contrast, the whisker was overdamped in maple sirup and corn sirup, where high viscosity imposed significant resistance, slowing down displacement recovery.

Fig. 4: Viscosity sensing.

a Error band plots of the eight whiskers' movements across five different fluid conditions, where the mean represents the average displacement of the eight whiskers, and the shaded region indicates the standard deviation. b Pixel displacement of a single whisker over time. c Magnified view of the 9.8–11 s interval in panel (b), highlighting the whisker behavior. d Linear fit between decay rate and ${\log }_{10}$(viscosity) (R² ≈ 0.85). Each point shows the trial-averaged decay rate with error bars (±1 standard deviation across trials). Dashed line indicates the fitted regression.

The mean and standard deviation of the rate of displacement decay upon deactivation for each liquid are presented in Table 3. The data reveal a clear inverse correlation between viscosity and the rate of whisker displacement recovery. Fitting the trial-averaged decay rates against the logarithm of viscosity yielded a strong log-linear correlation (R² ≈ 0.85), shown in Fig. 4d. This result shows that, despite inter-trial variability, the decay rate reliably reflects fluid viscosity and can be effectively modeled using logarithmic scaling, which captures the rapid changes at low viscosities and the saturation behavior at high viscosities. Notably, the standard deviations for low-viscosity fluids are higher. This is attributed to the rapid whisker responses in low-viscosity fluids, where the camera’s frame rate was insufficient to capture the fastest transitions with high accuracy. In contrast, high-viscosity fluids showed smaller variability, suggesting more stable whisker motion constrained by the fluid’s resistance.

Table 3 Mean and standard deviation of the rate of pixel displacement decay during deactivation for each liquid

This experiment successfully demonstrates the sensor’s capability to characterize liquid viscosity without requiring additional robotic motion. By simply inserting the whisker array into a liquid, the sensor can actuate and sense viscosity variations, eliminating the need for external force application or complex robotic maneuvers. This integrated sensing and actuation approach highlights the sensor’s efficiency and potential for automated viscosity measurement in robotic applications and environmental exploration.

Object classification

To evaluate the sensor’s ability to differentiate objects based on tactile interactions, the object classification experiment was designed using ten objects of varying shapes and sizes (Fig. 5a–c). Each data sample was obtained by placing the whisker sensor at eight positions with respect to an object and recording the whisker pixel displacement values.

Fig. 5: Object classification experiment.

a Experiment setup of the object classification experiment. b The sensor contacts the object from four directions at varying distances. c The ten tested objects: foam pear, foam lemon, plastic strawberry, plastic cherry, plastic lemon slice, eraser, yuanbao, pinecone, mini tape dispenser, and light bulb. d Confusion matrix for the object classification experiment, showing 98.7% accuracy across ten classes. CHR denotes plastic cherry, ERS denotes eraser, LM denotes foam lemon, LMS denotes plastic lemon slice, LB denotes light bulb, PER denotes foam pear, PNC denotes pinecone, STB denotes plastic strawberry, TDP denotes mini tape dispenser, and YB denotes yuanbao.

The classification task was performed using a Multi-Layer Perceptron (MLP) model. The model achieved a classification accuracy of 98.7% on the test set, demonstrating the sensor’s strong capability for object identification based on tactile interactions. The confusion matrix in Fig. 5d illustrates the classification performance across different objects, with most samples correctly classified. Misclassifications were few and primarily occurred between objects with similar shapes and sizes, such as between the plastic cherry and the mini tape dispenser, which shared overlapping tactile features. Overall, the high classification accuracy indicates the sensor’s effectiveness in distinguishing objects based on a limited number of contact interactions. This result validates the potential of the whisker array for advanced tactile sensing applications, demonstrating its ability to extract meaningful object characteristics with minimal contact data.

Lightweight object grasping

This experiment evaluated the sensor’s ability to grasp lightweight objects using different numbers of whiskers. As illustrated in Fig. 6a, the sensor was mounted on a UR16e robotic arm, which provided precise vertical motion for consistent grasp attempts. For each object, the sensor performed 20 grasp attempts under one whisker configuration at a time. Three configurations were tested: grasping with all eight whiskers, four whiskers, and two whiskers. Example grasps for each configuration are shown in Fig. 6b, and representative trials are provided in Supplementary Movie 3. The success rates of these configurations for each object are summarized in Table 4. Grasping with the eight whiskers achieved an average success rate of 90%, with the lowest at 70%. This configuration consistently yielded the highest success rates across all objects, indicating stable and well-controlled grasping performance.

Fig. 6: Grasping experiments using the whisker array sensor and image processing pipeline.

a–d Lightweight object grasping. a Setup. b Grasping a pom-pom using eight, four, and two whiskers. The actuated whiskers in the four- and two-whisker configurations are highlighted. c Six tested objects: paper flower, mini pinecone, popcorn, pom-pom, cotton, and mini bow (left to right). d Mean displacement of one whisker across successful trials. Shaded regions denote ±1 SD. Four phases are marked with orange dashed lines: (1) whisker actuation, (2) robot lift, (3) robot stop, and (4) whisker deactivation. e–g Crystal grasping. e Setup. f Example crystal cluster. g Successful grasp. h–j Snail pick-and-place. h Setup. i Close-up of a snail grasping. j Snail in the Petri dish 30s after the placement, with tentacles extended, showing intact survival. k–r Image processing pipeline for whisker tracking. k Raw image. l Preprocessing. m Mask of red region and detection of contour centers. n Zoomed-in view of the yellow box in panel (m). o Tracker centers (white dots) without external forces. p Zoomed-in view of the yellow box in panel (o). q Tracker displacements (yellow arrows) under external forces. r Zoomed-in view of the yellow box in panel (q).

Table 4 Grasp success rate of the three whisker configurations for the objects (%)

The variation in grasp success rates among the tested objects can be attributed to differences in object configuration, weight distribution, and surface properties. In the eight-whisker configuration, most failures were caused by slippage between the whisker filaments and the objects. Items with higher surface friction, such as pom-poms and cotton, reached a 100% success rate because their furry exteriors provided enhanced grip. Similarly, the mini pinecone also attained a 100% success rate, aided by its overlapping scaly structure that offered additional friction. In contrast, the four-whisker configuration provided fewer contact points, which led to reduced friction and created greater difficulty for irregular or spiky surfaces. Even so, it maintained an average success rate above 60%, demonstrating its viability even in less favorable conditions. The two-whisker configuration demanded precise alignment, as the whisker forces needed to pass through the object’s center of mass while maintaining adequate friction. Due to the reduced fault tolerance, this configuration resulted in an average success rate of 36.67%. Although this is lower, it remains acceptable given the limited contact points. Objects like the paper flower, popcorn, and mini bow, which have more irregular shapes, experienced more frequent misalignment or slippage with fewer whiskers.

All grasping trials were performed without force control, and the objects remained undamaged throughout. This result highlights the whisker array’s ability to handle delicate items without causing harm, underscoring its potential benefits in applications that require gentle manipulation. Furthermore, frames recorded during the eight-whisker configuration were analyzed to assess whisker actuation and tactile sensing. Figure 6d shows the displacement of a single whisker over successive frames for each object, with lines indicating mean values and shaded regions denoting ±1 SD. The data come from successful grasp trials using all eight whiskers. Four distinct phases are visible in the displacement plot: (1) PSPM activation, (2) upward movement of the robot arm, (3) motion stop at the target height, and (4) whisker deactivation. The displacement traces show minimal variation across trials in the case of cotton, indicating stable performance and consistent actuation. By contrast, the mini bow data exhibit greater variability, highlighting the whiskers’ sensitivity to differences in object shape and orientation. This comparison highlights the sensor’s ability to perceive and respond to variations in manipulated objects, as evidenced by the displacement variations. Popcorn trials exhibited larger average displacements due to their small size, suggesting tighter whisker convergence during grasping. These findings demonstrate the whisker array’s ability to sense and adapt to different object characteristics through distributed tactile feedback.

Delicate crystal grasping

To further demonstrate the sensor’s affinity for grasping fragile objects, we conducted an experiment using delicate crystal clusters, as shown in Fig. 6e–g. Across ten trials, the sensor consistently achieved successful grasps, fully securing the crystals (see Supplementary Movie 4 for representative trials). In a few trials, some small crystal particles detached, but the recorded weight loss was negligible (below 0.001 g). After release, visual inspection confirmed that the crystals remained structurally intact in most cases, with no obvious breakage.

Three comparative experiments were conducted to benchmark the performance of the whisker-based sensor against existing tactile and soft grippers (Supplementary Fig. 6). The first experiment was conducted using a WSG 32 Gripper (Weiss Robotics) equipped with tactile feedback from a GelSight Mini tactile sensor⁴⁹, a high-resolution vision-based device. In ten trials, the GelSight sensor failed to detect contact in every instance, causing the gripper to continue closing and ultimately crush the crystals. None of the trials resulted in a successful grasp using this setup. A second comparison was performed using two Fin-ray soft fingers mounted on the WSG gripper (Supplementary Fig. 6). Despite its compliance and ability to conform to object shape, the Fin-ray gripper consistently compressed and fractured the delicate crystals in all ten trials. Third, a star-shaped PneuNet gripper⁵⁰ was tested (Supplementary Fig. 6). Despite its softness and adaptiveness, it achieved only two successful grasps out of ten, and in the remaining cases, the crystals were partially destroyed.

These results highlight the whisker-based sensor’s capability in handling fragile objects. Whereas the GelSight-equipped gripper failed due to insufficient contact detection and the soft grippers lacked the ability to apply a controlled, gentle force, the whisker sensor achieved reliable non-destructive grasps, demonstrating its potential for applications requiring gentle force application and high tactile sensitivity.

As a further proof-of-concept of robotic application, we performed a pick-and-place task with a live snail (Fig. 6h–j, Supplementary Movie 5), where the whisker array enabled gentle grasping and transfer from a water tank to a Petri dish. This demonstration highlights the sensor’s potential for safe robotic interaction with fragile biological organisms and points to future applications in biomedical handling and environmental exploration.

Discussion

This work presents a vision-based whisker sensor array that integrates programmable magnetic actuation with high-resolution tactile perception. By combining Pulse-Switchable Permanent Magnets (PSPMs) with a camera-based tracking system, the sensor enables low-power, gentle interaction with the environment. It supports a wide range of tactile functions, including contact classification, airflow detection, waterflow frequency estimation, viscosity discrimination, object classification, and gentle grasping of fragile items. These capabilities extend the scope of whisker-based sensing beyond traditional passive or mechanically actuated designs.

Experimental results demonstrate accurate pixel-to-physical mapping, long-term repeatability and stability, frequency-resolved fluid interaction, and robust manipulation across diverse materials. Yet, several limitations remain. Vision-based tracking introduces latency, and PSPM-driven actuation yields limited force output compared to motorized alternatives. Each whisker currently operates along a single axis, constraining three-dimensional actuation. Moreover, waterflow sensing performance declines outside moderate excitation frequencies, indicating the need for frequency-adaptive design.

Looking ahead, enhancing the temporal resolution of perception through algorithmic optimization or event-based vision could enable faster response in dynamic settings. Multi-axis actuation and material miniaturization will support more expressive whisker motion and broader applicability, particularly in soft robotics and biomedical systems. Beyond individual sensing, the whisker array holds promise as a framework for embodied exploration. Tactile feedback could be used to guide motion planning and reactive navigation in cluttered or unknown environments. Coupling this with learning-based models may further enable complex behaviors such as slip detection, adaptive grasp regulation, and predictive decision-making. While this work focused on whisker-based sensing, an important future direction lies in integrating whisker-derived tactile information with vision to construct richer multimodal representations and enhance recognition robustness. Moreover, exploring safe interaction with fragile biological organisms represents a promising direction for future research in biomedical handling, environmental monitoring, and other scenarios requiring delicate manipulation.

Together, these advances position the whisker array as a compact, energy-efficient platform for next-generation tactile sensing and interaction. It offers a pathway toward intelligent, contact-rich exploration in unstructured environments.

Methods

Mechanical design

The mechanical design centers around a vision-based whisker array (Fig. 1f). The array comprises eight whiskers uniformly arranged in a circular layout with a 12 mm radius, all supported by a single elastomer membrane fabricated from Ecoflex 00–30 (Smooth-On, Inc.) with a thickness of 3 mm. The membrane adopts an octagonal shape, with an inscribed circle of 16 mm radius. Both the material and the membrane thickness were empirically selected to achieve the optimal balance between flexibility and structural support. To mitigate the influence of external lighting on tactile perception, the upper surface of the membrane is coated with gray ink.

The whiskers pass through the membrane and are permanently secured with adhesive. The whisker filaments are carbon fiber rods with a diameter of 1 mm and a total length of 105 mm, of which 94 mm extends above the membrane. Inspired by the structure of insect legs, the tips of the whiskers are coated with thin sand (see Fig. 6g) to increase friction by applying a thin adhesive layer and gently rolling it in fine sand particles. Controlled trials confirmed that this modification enhances the sensor’s ability to grasp light objects that it could not reliably manipulate without the coating (see Supplementary Movie 6). The coating is not structurally essential, but it is functionally important for achieving stable manipulation under low contact forces. Despite consisting of exposed rigid particles, the contact forces are small and distributed across multiple whiskers, and no visible damage was observed on delicate objects during manipulation.

Each whisker is paired with an iron-elastomer composite cube mounted at its base, measuring 4 mm on each side. The iron composites are aligned with corresponding pulse-switchable permanent magnets (PSPMs) for magnetic actuation. The bottom of the iron composite is flush with the base of the whisker filament, allowing direct visual access to the red-painted bottom of the filament. These red markers allow visual tracking via a Raspberry Pi camera (Adafruit 5390, with 160° lens). The camera captures the whisker movements and streams the images to a Raspberry Pi 4 Model B for processing. An LED stripe positioned beneath the PSPMs provides consistent illumination to support robust real-time tracking under varied ambient lighting. The entire sensor assembly, including the whisker frame, camera, PSPMs, and light source, is secured using a custom 3D-printed PLA mount.

Pulse-switchable whisker actuation

Whisker actuation arises from the interaction between the iron composites at each whisker base and a corresponding Pulse-Switchable Permanent Magnet (PSPM). The PSPM replaces the previously used coreless electromagnets, which exhibited high power consumption and circuit burnout during extended operation⁴⁰. In the configuration, a semi-hard magnetic material (AlNiCo 5) is encased in a copper coil. A 2 ms 24 V forward current pulse magnetizes the core, attracting the paired iron composite at the whisker root, while a 2 ms 6 V reverse pulse demagnetizes it. Each switching cycle for one whisker requires approximately 3.6 J, comprising 2.88 J for activation and 0.72 J for deactivation. Accordingly, a complete switching cycle involving all eight whiskers requires a total of 28.8 J.

The coil is driven using a BTS7960 H-bridge controlled by an Arduino Uno, allowing independent actuation of each whisker. Crucially, once the magnet is energized, it can remain in its activated state for an extended period under normal operating conditions without drawing additional power. To limit peak current, whiskers are triggered sequentially with a 1 ms delay between pulses. Under this scheme, a 24 V DC power supply, together with appropriate decoupling capacitors, is sufficient to handle the instantaneous load. Additional details on material selection, magnetic field modeling, and drive circuit implementation are provided in Supplementary Note 6.

Image processing and whisker tracking

A Raspberry Pi Camera (Adafruit 5390) with a 160° variable focus lens is mounted below the sensor and connected to a Raspberry Pi 4 Model B. Images are captured at 640 × 480 resolution and 60 frames per second. The red whisker markers are detected by converting frames to HSV color space, followed by binary masking, erosion and dilation to remove noise. Contours are extracted using OpenCV’s findContours function, and centroid positions are computed from image moments. The image processing procedure is shown in Fig. 6k–r.

To estimate the whisker tip positions, we modeled each whisker as a rigid rod rotating about a central pivot on the membrane. Pixel displacements of the root markers were converted to real-world coordinates using a fixed pixel-to-length scale factor. The 3D tip positions were inferred from the tracked 2D marker movements and the known whisker geometry. A full derivation of the geometric mapping and angle estimation is provided in Supplementary Note 2.

Actuation characterization

For the voltage-angle characterization, the whisker array sensor was placed 20 cm away from an Intel RealSense D455 camera, with the camera lens oriented parallel to the whisker rotation plane (shown in Supplementary Fig. 7). This ensured that the rotation angles could be accurately captured when only the magnetic force from the PSPM was applied. The actuation voltage was incrementally increased from 6 V to 24 V in 1 V steps. Each voltage pulse was applied for 2 ms to actuate the PSPM. After each actuation pulse, an RGB image was captured using the RealSense camera, and the rotation angles of the whiskers were manually measured using ImageJ software.

For characterization of the forces exerted during actuation, a Mark-10 M5-05 Advanced Digital Force Gauge (capacity 2.5 N, resolution 0.5 mN) was used. In a preliminary test, direct measurements of tip forces fluctuated between 0 and 1 mN, close to the gauge resolution. To ensure reliable quantification, forces were therefore measured at a position 15 mm from the membrane surface. The sensing element of the gauge was aligned with the direction of whisker motion and positioned in close contact with the whisker without registering preload on the gauge. Each whisker was then actuated, and the corresponding force was recorded once the reading stabilized rather than at the peak value. The experimental setup is illustrated in Supplementary Fig. 8. Five repeated measurements were conducted for each whisker.

Payload testing was conducted using a 3D-printed ABS open-ended cylindrical container. The experimental setup is shown in Supplementary Fig. 9. To increase surface friction, a fiber layer was wrapped around the container. The container itself weighed 1.2 g, which was taken as the initial payload. For each weight, the whisker array attempted ten grasps. For each trial, the sensor was aligned above the container to ensure approximate centering. The UR16e robotic arm then lowered the sensor until the whiskers reached 1 mm below the container’s supporting surface. At this position, the whiskers were sequentially actuated with 2 ms, 24 V pulses, separated by 1 ms delays. After actuation, the arm lifted the object by 3 cm and held it for five seconds. A grasp was considered successful if the object did not drop, although minor slipping during lifting was permitted. After each set of ten trials, the payload was increased by adding 0.1 g of sand into the container. The procedure was repeated to record grasp success rates.

Perception characterization

To map pixel measurements to real-world whisker movements, a characterization platform was established, as illustrated in Supplementary Note 3, Supplementary Fig. 10. A linear stage motion actuator (RATTMMOTOR CBX1605 100 mm) was combined with a rotating servo (DYNAMIXEL XM430-W210-R) to control individual whisker displacements. This calibration was conducted using whiskers without sand coating, as it only measures geometric tracking and is unaffected by the thin coating used in other experiments. The whisker sensor array was mounted such that a single whisker was aligned with the servo’s rotational axis. A 3D-printed PLA hook was mounted on the linear stage, positioned 77.5 mm above the sensor membrane, to provide consistent forward-backward displacement. To avoid introducing vertical displacement or whisker tip deformation, the hook was not attached directly to the tip but held above it. During each trial, the linear actuator pushed the whisker tip in 2 mm increments until it either reached 30 mm or made contact with its corresponding PSPM. For each linear increment, a camera image was recorded. After each forward and backward sequence, the servo rotated 5°, enabling data collection from different angular orientations. Five repeated trials for each whisker were conducted to ensure the consistency and validity of the calibration process. Pixel displacements were recorded and later used to estimate horizontal displacements at the whisker base and tip. The calibration model and geometric derivation are provided in Supplementary Note 3.

To calibrate whisker-based force sensing, the sensor was mounted on a rotary motor to allow characterization across different angular orientations, as illustrated in Supplementary Note 5, Supplementary Fig. 11. A digital force gauge (Mark-10 M5-05, resolution 0.5 mN) was fixed on a linear stage (RATTMMOTOR CBX1605, 100 mm) horizontally, which advanced in 0.5 mm steps to apply controlled forces at the whisker midpoint. At each force increment, the force and corresponding pixel displacement were recorded. The procedure continued until either a 30 mm displacement was reached or the whisker’s iron composite contacted the sensor side wall. Each orientation was tested in five repeated trials, with the process repeated in 10° increments. The collected data were then used to establish a linear regression model mapping pixel displacements to applied forces based on the analysis in Supplementary Notes 1, 4, and 5.

Repeatability tests

For the repeatability test of 1000 cycles, the sensor was subjected to repeated actuation and deactivation to evaluate its long-term stability. In each cycle, the whiskers were activated using a 24 V, 2 ms pulse, followed by deactivation to allow passive return to their resting positions. Pixel displacements were continuously monitored using the integrated onboard camera. To prevent overheating, each cycle was interleaved with a 60 s rest period after both activation and deactivation phases. The experiment was conducted over several days, with pauses at the 150th, 390th, 625th, and 910th cycles due to safety protocols prohibiting unsupervised operation overnight. During the entire testing period, the sensor was not used for any other experiments and remained in its original setup configuration to avoid introducing external variability. No external contact or loading was applied during the test to isolate the intrinsic behavior of the whisker mechanism. In each cycle, two displacement measurements were recorded per whisker, one at the activation phase and one after deactivation. To minimize the influence of transient oscillations, the displacements of the deactivation states were measured using images captured 10 s after deactivation. Displacement was defined as the Euclidean distance between the tracked whisker marker and its resting reference position, measured in pixels on the image plane.

For the grasp repeatability test, the sensor was mounted on a UR16e robotic arm and positioned directly above the pom-pom (0.35 g), ensuring approximate centering relative to the whisker array. The arm was lowered until the whisker tips contacted the surface on which the object rested. At this position, the whiskers were sequentially actuated with 24 V, 2 ms pulses, with a 1 ms delay between activations to accommodate the power supply’s current limit. After actuation, the arm lifted the sensor by 3 cm and held it briefly. A grasp was considered successful if the object remained suspended without slipping. This protocol was repeated on days 1, 2, 5, and 10, with ten trials per day, using the same object and setup to ensure consistency across sessions.

Airflow experiment setup

For different air speeds, the fan was positioned 15 cm from the whisker array sensor. During each trial, the fan operated at a given speed for 10 s, and the resulting whisker displacements were recorded. For each airflow condition, whisker displacement peaks and valleys (i.e., local maxima and minima) were identified by comparing each data point with its immediate neighbors. The airflow speed, which served as the reference for evaluating the whisker response, was measured at the same position of the whisker array sensor by using a digital anemometer (HoldPeak 866B). This device features a measurement range of 0.3–30 m ⋅ s⁻¹ with 0.1 m ⋅ s⁻¹ resolution and an accuracy of ±5% of the reading. The measured speeds were 5.0 m ⋅ s⁻¹ for the high mode, 4.7 m ⋅ s⁻¹ for the medium mode, and 4.3 m ⋅ s⁻¹ for the low mode.

Contact and airflow feedback classification

The classification process follows the pipeline described below. First, a Fast Fourier Transform (FFT) is applied to a 30-frame sliding window (0.5 s). If the maximum spectral magnitude exceeded a threshold of 5, the signal was classified as dynamic airflow feedback. Otherwise, it was considered static. To further distinguish between contact and airflow-induced movements, the peaks and valleys of the displacement waveforms are continuously tracked. If the difference between consecutive peaks or valleys exceeds 4 pixels, the feedback is classified as contact feedback. This ensures that sudden large whisker movements caused by direct object contact are not misinterpreted as airflow-induced oscillations. Additionally, within the static feedback category, if the whisker displacements are non-zero, the feedback is classified as contact feedback. If all whisker displacements remain zero, the whiskers are classified as being in the origin state.

The whisker sensor array was further mounted on a UR16e industrial robot (Universal Robots) for the experiment. The fan was initially off, with no contact. The robot then remained stationary while airflow was applied, followed by a 1.5 cm forward movement contacting two whiskers, then an additional 0.8 cm to engage four. After holding the position for 5 s, it returned to the origin. Classification results aligned with observed ground truth (see Supplementary Movie 1).

Water flow experiment setup

To assess the sensor’s response to dynamic fluid environments, a servo-driven blade was used to generate periodic water flow at controlled frequencies. The setup featured a DYNAMIXEL XM430-W210-R motor connected to an L-shaped blade, with the whisker array positioned 2 cm in front of the blade. Whisker tips were submerged approximately 2 cm below the water surface to ensure consistent interaction with the flow (Supplementary Fig. 12).

The motor was programmed to operate across a frequency range from 0.5 Hz to 10 Hz in 0.5 Hz increments; however, due to mechanical limitations, the actual output frequencies ranged from 0.49 Hz to 7.1 Hz. For each frequency, the whisker displacement data were recorded for 5 s, with three repeated trials, resulting in 60 total trials. Given the eight-whisker configuration, a total of 480 whisker response sequences were collected. Each sequence was analyzed using the FFT to extract dominant frequency components.

Viscosity sensing experiment

To evaluate the sensor’s ability to discriminate fluid viscosity, the whisker array was submerged 2 cm below the surface of the test fluid. All whiskers were actuated simultaneously and held in motion for 10 s, followed by deactivation. Data recording continued for an additional 10 s, capturing both activation and relaxation phases.

Five fluids spanning a wide viscosity range were tested: air (0.01854 mPa ⋅ s), Water (1.57 ± 0.02 mPa ⋅ s), canola oil (57.74 ± 0.13 mPa ⋅ s), maple sirup (170.52 ± 0.41 mPa ⋅ s), and corn sirup (3664.99 ± 4.06 mPa ⋅ s). The viscosity values, except for air, represent reference measurements obtained at 25 °C using a TA AR-G2 Rheometer with a bob-and-cup fixture in Flow Ramp mode (shear rate range: 0.01–100 s⁻¹). For air, the viscosity value was adopted from the CRC Handbook of Chemistry and Physics⁵¹.

Each experiment was conducted over five trials per liquid, with the displacements of all eight whiskers recorded. To quantify the rate of displacement decay upon deactivation, the slopes of the whisker displacement curves were computed. The interval for slope computation was initiated at 10 s, coinciding with the moment when the whisker array was deactivated. To mitigate the influence of signal fluctuations, a mean filter was applied to smooth the displacement curves. The filter window size was determined as the smaller value between 2 s and the time span required for the minimum pixel difference to exceed -5 pixels. The length of the slope computation interval was then set to the smaller of 10 times the filter window or the remaining duration before 20 s. The slope was computed as the minimum pixel difference between consecutive data points over the respective time interval.

To further assess the relationship between decay rate and fluid viscosity, the slopes obtained from all eight whiskers in each trial were averaged, resulting in one representative decay rate per trial. These trial-averaged values were then regressed against the logarithm of the fluid viscosity using a log-linear model. The resulting regression equation was: Decay Rate $=(-81.32{\log }_{10}(\eta )+243.72)$ pixel ⋅ s⁻¹, where η denotes the dynamic viscosity of the test fluid (mPa ⋅ s).

Object classification experiment setup

The sensor, mounted on a UR16e robot arm, approached each object horizontally from four different directions. As depicted in Fig. 5b, data were collected at distances of 3 cm and 1 cm from the object’s center, maintaining a vertical gap of about 1 mm between the whisker tips and the surface on which each object rested. During each approach, the sensor collected data from all eight whiskers, yielding eight contact data points. Since each whisker measurement includes both x and y pixel displacements, each approach produced 16 data points in total. Repeating this procedure for eight contact positions resulted in 128 data points overall in one trial. To achieve comprehensive coverage, each object was placed on a rotating servo (DYNAMIXEL XM430-W210-R) that swept from 0° to 360° in 10° increments. At each angle, five trials were performed, resulting in 180 samples per object.

Object classification algorithm

The classification task was performed using a Multi-Layer Perceptron (MLP) model with one hidden layer consisting of 32 neurons and ReLU activation functions. The model’s input was a tensor of size 128, representing the whisker pixel displacements at eight contact positions, while the output was a tensor of size 10, corresponding to the object categories. The dataset was split into training and testing subsets in a 7:3 ratio. The model was trained using the Adam optimizer with a learning rate of 0.01, and cross-entropy loss was used as the classification criterion. Training was conducted over 10 epochs. The training loss and accuracy over the ten epochs were shown in Supplementary Fig. 13.

Lightweight object grasping

Six lightweight objects, each with different shapes and surface properties, were chosen for this experiment which were paper flower, a mini pinecone, popcorn, a pom-pom, cotton, and a mini bow (shown in Fig. 6c). Their corresponding weights were 1.03 g, 0.82 g, 0.24 g, 0.35 g, 0.12 g, and 0.26 g, respectively.

For each grasp attempt, the sensor was positioned directly above the object, ensuring the object was approximately centered. The UR16e robotic arm then lowered the sensor until the whiskers reached the surface on which the object rested. At this lowest position, the whiskers were individually actuated by a 2 ms 24 V pulse, with a 1 ms delay between each activation to accommodate the power supply’s current limit. Following actuation, the robot arm lifted the object by 3 cm and briefly held it in place. A grasp was considered successful if the object remained held by the sensor without dropping.

Delicate crystal grasping

The crystals were grown using the NATIONAL GEOGRAPHIC Craft Kits - Crystal Growing Kit and allowed to develop for one day at a controlled temperature of 21 °C. Due to natural variations in crystal structures, some clusters were highly fragile, breaking into pieces when dropped from a height of just 5 cm. To ensure meaningful grasping trials, a selection process was conducted to filter out loosely formed clusters, retaining only those that remained intact after a 5 cm drop. This selection allowed us to better assess the applied force effects during grasping.

For the experiment, the crystals were placed on a platform with a professional digital scale (0.001 g accuracy) positioned underneath. The sensor was moved downward to align with the platform height and then actuated to grasp the crystals. After grasping, the robot lifted the crystal by 2 cm, and the weight change on the scale was recorded to determine the amount of the crystal cluster successfully grasped.

To benchmark the sensor’s performance, three comparative experiments were conducted. In the first, a WSG 32 Gripper (Weiss Robotics) was equipped with tactile feedback from a GelSight Mini (GelSight), a commercial gel-based tactile sensor. The WSG gripper was positioned 0.5 cm below the platform level to accommodate the trapezoidal shape of the GelSight Mini, ensuring that the tactile sensor could make contact with the crystal cluster when the gripper closed. During each trial, the WSG gripper closed slowly, and once the GelSight Mini detected contact, the gripper was programmed to stop and lift the crystal. In the second experiment, a pair of Fin-ray fingers made from Ecoflex 00–30 (Smooth-On, Inc.) was mounted on the WSG gripper. The lowest points of the Fin-ray fingers were aligned with the platform level where the crystals were placed. In each trial, the WSG gripper was closed slowly to allow the compliant Fin-ray structure to conform around and attempt to grasp the crystal. In the third experiment, a star-shaped PneuNet gripper was fabricated from Ecoflex 00–30 (Smooth-On, Inc.) with fabric serving as the strain-limiting layer. The lowest point of the actuated PneuNet was aligned 1 mm below the platform to better accommodate crystal grasping. During each trial, the PneuNet was actuated using an electric air pump to close around the crystal.