SWTac: Sliding Window Tactile Sensor

• SWTac is an event-based visuotactile sensor that leverages active vibrations and mechanical isolation to enable precise tactile sensing in granular media.
• It employs IMU-guided temporal filtering to boost imaging quality by up to 24% while reducing temporal variance.
• The integrated U-Net algorithm accurately reconstructs contact surface masks, achieving superior metrics in SSIM, IoU, and force estimation.

SWTac (Sliding Window Tactile) is an actively vibrated, event-based visuotactile sensor designed to enable high-fidelity contact perception in unpredictable granular environments. Developed as part of the SandWorm biomimetic robot, SWTac integrates a mechanically isolated event camera with a vibrated elastomer and implements a suite of hardware and signal-processing innovations for robust tactile imaging, texture classification, and force estimation within dynamic particulate media (Li et al., 20 Jan 2026).

1. Mechanical Architecture and Isolation

SWTac comprises two primary subassemblies: a vibration-actuated tip and a mechanically isolated event camera. The vibrational assembly includes an electromagnetic actuator inducing vertical oscillation at 50 Hz and two eccentric-mass motors producing horizontal vibration at 100 Hz. These actuators drive a 1.5 mm PDMS elastomer tip (17:1 base:curing agent, Shore 20 A), optimized for abrasion resistance and compliance.

The event camera (DVXplorer Mini) is mechanically decoupled via eight radial steel springs (horizontal isolation) and two flexible shaft couplers (vertical isolation), forming a two-axis mass–spring system. This configuration attenuates ≈83% of vibration energy experienced at the actuator, verified through FFT analysis of camera-IMU data, while maintaining an unobstructed field of view. A bright-field illumination system with 14 WS2812B LEDs and a diffusor provides uniform lighting, and the diffusor's central mask aids feature extraction.

2. Active Vibration Sensing and Imaging Model

SWTac's elastomeric tip is oscillated according to

$f(t) = A \sin(\omega t+b)$

with amplitude $A\approx 200\ \mu$m and frequency $\omega$ (50 Hz vertical, 100 Hz horizontal). Imaging quality is governed by instantaneous velocity $|A\omega\cos(\omega t + b)|$, which empirical evidence suggests yields strong correspondence with edge-rich, high-contrast event frames.

The camera’s integrated IMU is leveraged to quantify phase and vibrational state:

$$IMU(t) = -A\omega^2\sin(\omega t + b) + \epsilon(t)$$

with attenuation ensuring residual vibrations at the sensor are ≤20% of the tip’s motion. A custom metric, masked signal-to-noise ratio (MSNR), is defined as

$$MSNR = 10 \log_{10} \left( \frac{\sum_{i\in\Omega} I(i)^2}{\sum_{i\in\Omega} [I(i)-\mu_\Omega]^2} \cdot \frac{N_\Omega}{N_{image}} \right)$$

where $\Omega$ is the foreground mask, $I(i)$ is intensity, and $N_\Omega,N_{image}$ are pixel counts.

3. IMU-Guided Temporal Filtering

Temporal consistency and spatial resolution are enhanced by IMU-guided event stream filtering. Calibration aligns IMU phase with image-quality oscillations by collecting event and IMU data under set vibration frequencies and reconstructing frames at high rates (500–1000 Hz). IQ time series, computed via MSNR, are linearly regressed against IMU absolute value, yielding a gating rule:

• Accept event windows when $|IMU_{live}(t-\Delta\hat{t})| \geq IMU_{th}$
• Drop intervals falling below this threshold

This filtering regime boosts average MSNR by up to 24% and reduces temporal variance by 40–45% at operational frequencies, improving detection of textural and edge features during both dynamic contact and stalling conditions.

4. Systematic Parameter Optimization

Comprehensive parameter sweeps identify optimal operating regimes for imaging quality:

• Amplitude: $A \gtrsim 200\ \mu$m required for sufficient event generation; saturation and noise increase above this threshold.
• Frequency: MSNR peaks at $\sim$50 Hz; higher frequencies trigger viscoelastic losses in the PDMS leading to degraded signal.
• Elastomer hardness: 20 A Shore PDMS offers highest band-limited MSNR via Kelvin–Voigt modeling.
• Actuation direction: Superimposed 100 Hz horizontal vibration with vertical oscillation increases MSNR by $\sim$10%, improving edge completeness.
• Camera sensitivity: Intermediate contrast threshold balances event density and noise, maximizing MSNR.

Quantitative relationships confirm that imaging quality is proportional to instantaneous tip velocity, and viscoelastic characteristics constrain useful vibration bandwidth.

5. Contact Surface Estimation and Algorithmic Pipeline

Edge-rich event frames reveal only high-strain discontinuities, leading to asymmetric edge maps that may misrepresent the true contact footprint. To resolve this, SWTac employs a 4-level U-Net (Conv-BN-ReLU blocks, skip connections, and transposed convolutions) mapping $256\times256$ event-generated edges to binary contact masks. Training uses an $\ell_1$ pixel loss on 3,000 annotated samples (12 object categories) with category-holdout cross-validation.

Performance metrics show significant gains with IMU-based filtering:

• SSIM: $0.9688 \pm 0.0004$ (filtered), higher than raw events
• IoU: $0.8104 \pm 0.0052$ vs $0.6965 \pm 0.0077$
• RMSE: $0.0693 \pm 0.0011$ vs $0.0945 \pm 0.0016$

This network recovers the full-area stress pattern of contacts, correcting the under- or over-representation of event-derived edges.

6. Experimental Performance and Application Outcomes

SWTac demonstrates robust results across a suite of tactile sensing and robotic locomotion tasks:

• Texture resolution: 0.2 mm, assessed via calibrated boards at 1000 Hz frame rate.
• Material classification: ResNet-18, fine-tuned on five stone types under varying granular conditions, achieves 98% accuracy, with per-class precision/recall exceeding 95%.
• Force estimation: Random forest regression (input: edge-derived tip centroid/radius; output: $(F_x, F_y)$) yields $R^2 > 0.95$, MAE = 0.15 N (36 calibration trials).
• Granular penetration: Active vibration reduces penetration resistance by ~20% at 30 N load; sensor drills $\gt$200 mm into medium at 100 RPM while maintaining clear event signals.
• Pipeline tasks: Locomotion speed up to 12.5 mm/s (with peristalsis); clearing blockages in 600 mm pipes (81–90 s) with 90% success in subsurface object finding.
• Field operation: Consistent performance in outdoor soil, cement, grass, with real-time imaging at 200 Hz via smartphone.

7. Design Rationales for Granular Media Sensing

SWTac’s architecture—combining active vibrational excitation, mechanical isolation, and IMU-aware event processing—specifically targets the challenges of tactile sensing in granular, high-vibration, and noisy environments:

• Active vibration transforms static contacts into dynamic, high-SNR event data, overcoming the static-scene limitation of event cameras and mitigating motion blur at kilohertz rates.
• Spring–mass isolation shields the event camera from destructive actuator vibrations (attenuation $\gt$80%), preserving longevity and image stability.
• IMU gating enables real-time rejection of low-quality intervals, supporting reproducible tactile imaging in dynamically changing or high-noise regimes.
• Parameter tuning establishes an operating “sweet spot” ($\sim$200 μm at 50 Hz vertical + 100 Hz horizontal, Shore 20 A, mid camera threshold) for edge fidelity.
• Algorithmic estimation via deep learning accommodates asymmetric edge patterns, reconstructing continuous stress footprints not directly observable in event frames.

These findings enable downstream tasks such as high-resolution texture recognition, robust material classification, vectorial force estimation, and complex robotic navigation within natural granular environments previously inaccessible to traditional tactile apparatus (Li et al., 20 Jan 2026).