---
title: 'SandWorm: Biomimetic Subsurface Navigator'
url: https://www.emergentmind.com/topics/sandworm
type: topic
---

# SandWorm: Biomimetic Subsurface Navigator

SandWorm is a biomimetic robotic system designed for navigation and tactile perception in granular media, integrating a screw-actuated peristaltic locomotion mechanism and the SWTac visuotactile sensor. The platform fuses mechanical innovation, event-based sensing, active vibration, and real-time algorithmic filtering for robust operation in environments characterized by unpredictable particle behaviors. Its pipeline includes state-of-the-art tactile imaging, contact mask estimation with deep learning, and feedback-driven locomotion for subsurface exploration and pipeline inspection in complex, field-realistic settings [2601.14128].

## 1. Mechanical Architecture and Locomotion

SandWorm’s locomotion system leverages a rigid spiral shell described as an “Archimedean screw” with a pitch $p \approx 4$ mm, outer shell diameter 32 mm, and length 80 mm. A brushless DC motor rotates the shell at 60–100 RPM, translating angular displacement $\theta$ into axial motion:

$$
l = \frac{p}{2\pi}\,\theta
$$

Locomotion is enhanced by an internal pushrod applying alternating force $F_p$, yielding two phases:
- **Extension**: $F_{\rm extension} = F_{\rm propel} + F_p - F_{\rm friction}$
- **Retraction**: $F_{\rm retraction} = F_{\rm propel} - F_p - F_{\rm friction}$

Here, $F_{\rm propel} = mg \sin\alpha$ incorporates gravity effects on inclines. The combined screw–peristalsis action delivers a measured maximum locomotion speed of 12.5 mm/s in a 200 mm-ID pipe—a 62% improvement over screw-only drives. The pushrod stroke is approximately 30 mm at 1 Hz, with $F_{\rm friction}$ accounting for all resistive forces from the medium and boundaries [2601.14128].

## 2. SWTac Event-Based Visuotactile Sensor

The SWTac sensor integrates an actively vibrated elastomer (PDMS, Sylgard 184, 17:1 mix, Shore 20 A, 1.5 mm thick) with a decoupled event camera, ensuring high-fidelity dynamic and static tactile imaging.

### Vibration Isolation

An array of eight lateral springs (stiffness $k_x$) and two flexible-shaft couplers ($k_z$) constitute a second-order isolation system for the camera:

$$
m_{\rm cam}\,\ddot x + c\,\dot x + k\,x = F_{\rm vib}(t)
$$

Transmissibility is defined as

$$
T(\omega)\approx\frac{(\omega_n/\omega)^2}{\sqrt{[1-(\omega_n/\omega)^2]^2+(2\zeta\,\omega_n/\omega)^2}}
$$

with measured 83% vibration isolation at 50 Hz.

### Elastomer Vibration

Dual actuation is applied: vertical (electromagnetic valve, $f_v=50$ Hz, $A_v\approx200$ µm) and horizontal (offset-mass motors, $f_h=100$ Hz, $A_h\approx50$–100 µm). Optimal sensor signal-to-noise (MSNR) is observed at these vibration parameters and mid-level event thresholds [2601.14128].

## 3. Event-Based Imaging, MSNR, and Temporal Filtering

### Grayscale Event Reconstruction

Event streams $e_i=(x_i,y_i,p_i,t_i)$ are integrated across 1 ms windows ($\Delta T$), discarding polarity $p_i$, to generate sharp 1 kHz frames:

$$
G_k(x,y) = \sum_{\substack{e_i\in\mathcal E_{\rm filt}\\(x_i,y_i)=(x,y)\\t_i\in[T_k,T_k+\Delta T)}} C
$$

### Masked SNR (MSNR)

MSNR evaluates foreground image quality:

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

### IMU-Guided Temporal Filtering

Sensor output quality $IQ(t)$ fluctuates with vibration phase and is modeled as a function of vertical displacement $f(t)=A\sin(\omega t+b)$ and measured acceleration:

$$
\mathrm{IQ}(t)\approx u|A\omega\cos(\omega t+b)| \\
\mathrm{IMU}(t)=A\omega^2\sin(\omega t+\pi)+\epsilon
$$

Peak-aligned, bandpass-filtered IMU data is fitted to predict high-quality intervals; only event slices above threshold are retained, resulting in up to 24% MSNR improvement, 46% reduction in MSNR standard deviation, and 1 ms processing latency [2601.14128].

## 4. Contact Surface Estimation by Deep Learning

Finite-element simulations indicate indenting the elastomer yields asymmetric edge responses (sharp inside, blurred outside). A U-Net architecture processes 256 × 256 event frames ($G_k$) to produce binary contact masks, capitalizing on these edge features.

### Network and Training

- Four-level encoder/decoder: 3 × 3 convolution + BN + ReLU, 2 × 2 max-pooling, up-convolution, and skip connections.
- Final 1 × 1 convolution with sigmoid activation for pixelwise mask probabilities.
- Training dataset: 300 hand-annotated images of 12 textures, augmented to 3,000 samples, cross-category hold-out.
- Loss: $L_1(M, M_{gt}) = \sum|M - M_{gt}|$.

IMU-filtered inference achieves SSIM ≈ 0.969, IoU ≈ 0.81, and RMSE ≈ 0.069 [2601.14128].

## 5. Tactile and Locomotive Performance

SandWorm’s integrated system demonstrates proficiency on granular and mixed-media tasks.

### Tactile Sensing Outcomes

- **Texture Resolution**: 0.2 mm, enabling recovery of fine board patterns.
- **Material Classification**: Five stone classes (grit, gravel, pebble, cobble, eggstone) with 98% accuracy using fine-tuned ResNet-18 at 500 Hz.
- **Shear Force Estimation**: Tip displacement $(x, y, r)$ mapped to force via Random Forest; MAE = 0.15 N ($R^2 > 0.95$) [2601.14128].

### Locomotion and Task Benchmarks

- **Pipeline Inspection**: 200 mm-ID, 600 mm in 48 s (12.5 mm/s), with navigational triggers from shear force sensing.
- **Obstacle and Bend Navigation**: Reliable steering in 15° bends, wall intersections, and 90° elbows (150 mm ID).
- **Dredging**: Removal of gravel/cobble/eggstone blocks with 90% success in blocked pipeline trials (600 mm in 84–90 s).

### Subsurface and Field Performance

- **Granular Drilling**: 40 trials in beach sand, TPE ($\rho\approx700$ kg/m³), EPP (100 kg/m³), and EPE (10 kg/m³): 36/40 buried objects found (90% success, ≤120 s per trial).
- **Field Operations**: Effective on grass, bushes, cement, autonomous dredging in mud/leaves/gravel, and recovery of diverse objects (fossils, bottle caps) from natural soil [2601.14128].

## 6. Significance and Technological Implications

SandWorm exemplifies hardware–software co-design and bio-inspiration (screw plus peristaltic actuation), yielding robust subsurface locomotion and sub-millimeter-level tactile imaging in granular environments previously considered intractable for robotic agents. Co-optimization of vibratory actuation, event-based tactile sensing, IMU-guided event selection, and deep-learning-driven contact reconstruction enables precise, real-time (1 kHz) perception and control at the tip, with field demonstrations validating efficacy across a spectrum of real-world scenarios. A plausible implication is that this architectural fusion could generalize to future bio-inspired robots facing similarly challenging, dynamic contact conditions [2601.14128].

Source: https://www.emergentmind.com/topics/sandworm