Towards Artificial Nerves: Biomimetic Optical-Fiber Tactile Sensing for Robots

Abstract: Robotic systems increasingly demand tactile sensing that approaches the adaptability and resolution of human skin to enable dexterous manipulation and safe interaction. OptiTac is a biomimetic tactile sensor that emulates the mechanoreceptor-to-nerve architecture of human touch by pairing each mechanical pin on a soft skin with an optical fiber acting as an artificial nerve. This design demonstrates an architectural principle for routing tactile information away from the sensing surface while preserving high spatial resolution, establishing a practical route toward distributed tactile sensing in future robotic systems. By treating tactile signals as images, simple analytical methods, rather than opaque deep-learning models, are used to infer contact location, size, and shape, providing interpretable and scalable tactile intelligence. This work demonstrates how evolutionary principles from biology can guide the development of artificial nerve systems for robots, offering a pathway toward human-like tactile perception in next-generation robotic platforms. More broadly, OptiTac establishes an artificial nerve-inspired sensing framework for interpretable robotic touch and a scalable route toward future distributed tactile systems.

• The paper introduces OptiTac, a 217-fiber optical tactile sensor that pairs each mechanical pin with one optical fiber and uses image moments instead of deep learning to estimate contact properties.
• OptiTac achieves 0.4 mm unseen-data localization RMSE, demonstrating hyperacuity below its 1.5 mm fiber spacing, while contact-width estimates reach R² values of 0.95–0.97.
• The system classifies five contact shapes with 96% accuracy using the first three Hu moments and a Gaussian Mixture Model, but requires validation with arbitrary contacts, longer fibers, and real robotic tasks.

Overview and design principle

OptiTac is a biomimetic optical tactile sensor (OTS) that combines the pin-based mechanical transduction of the TacTip family with a coherent optical-fiber array (OFA) that transports tactile information away from the sensing surface to a remote camera. The central architectural claim is a one-to-one pairing between each mechanical pin on the soft skin and an individual optical fiber, emulating the mechanoreceptor-to-nerve pathway of human glabrous skin, where Merkel cells and Meissner corpuscles transduce deformation and route signals through nerves terminating in spatially organized regions of the spinal cord and somatosensory cortex. This separation of the compliant sensing interface from the imaging hardware is positioned as a practical route to distributed tactile sensing over large skin areas, circumventing the focal-distance and lens constraints of conventional vision-based tactile sensors (VBTSs).

The sensor comprises two modules: a soft tactile sensing module filled with transparent elastomeric gel, containing 217 plastic optical fibers (1 mm diameter) arranged in a coherent hexagonal array, and a remote camera module. Skin deformations lever the pins, displacing white markers relative to fiber endfaces; marker-fiber alignment modulates transmitted intensity, which the camera captures as a structured tactile image. Cubic interpolation of the 217 taxels onto a 100×100 mesh yields high-resolution intensity surface plots from which contact parameters are extracted analytically.

Pin-pattern selection as physical pre-processing

A key empirical result is that the one-to-one aligned pin-OF pattern functions as physical pre-processing that renders downstream inference tractable with simple analytical tools. Three skins were fabricated — sparse (fewer pins than fibers), dense (2 mm inter-pin spacing versus 1.5 mm inter-fiber spacing), and aligned — and indented with a 5 mm ball to 3 mm depth. The sparse pattern produced surface-wide intensity changes; the dense pattern produced localized but complex, non-intuitive changes spanning an area much larger than the indenter; only the aligned pattern produced a localized negative intensity change commensurate with the contact. This justifies the biomimetic pairing choice: the signal structure directly supports image-moment-based inference without learned feature extraction.

Contact localization and hyperacuity

Contact centroid localization uses first raw image moments on thresholded, inverted surface plots. A calibration protocol was developed in which indentations were made directly above each of the 217 fibers with a 5 mm cylinder; nearest-neighbor error surfaces (1000×1000 meshes via k-d tree) were smoothed with a Gaussian filter ($\sigma = 10$) to prevent overfitting.

Performance figures are strong:

Metric	Value
Calibrated $R^2$, seen data	0.99
Calibrated $R^2$, unseen data	0.99
RMSE reduction (seen)	0.26 mm
RMSE, unseen data (both axes)	0.4 mm
Mean repeatability error across 5 locations	< ±0.71 mm
Mean error range per location	0.09 mm

Because these errors are more than three times smaller than the 1.5 mm pin-OF spacing, the authors claim hyperacuity — localization precision finer than the sensing-element pitch — under the tested indentation conditions. Compared against prior OFA sensors, OptiTac's 0.4 mm RMSE substantially improves on Baimukashev et al.'s CNN-based sensor (RMSE 1.1/1.4 mm at 3.9 mm fiber spacing), though it does not reach the 0.03/0.02 mm achieved by Lu et al.'s Multiscale ResNet. The substantive point is that sub-millimeter accuracy is obtained with a fully transparent analytical pipeline rather than end-to-end deep learning, at roughly double the spatial density of the CNN-based predecessor.

Contact-width estimation

Contact width is estimated from eigenvalues of normalized second-order central moments under a Gaussian-contact assumption, yielding full widths along principal axes. After axis-specific offset removal (3.43 mm major, 2.47 mm minor) and scaling (2.35 and 2.51), measured widths track true diameters linearly across 1.85–20.09 mm indenters with $R^2$ values of 0.95 (major) and 0.97 (minor), and RMSEs of 1.2 mm and 1.0 mm — both below the fiber pitch. Accuracy is size-dependent: for indenters below 10 mm, RMSE drops to 0.75/0.57 mm, rising to 1.52/1.20 mm above 10 mm. The current geometry therefore favors small contacts, a limitation relevant to grasping applications involving large-area contact.

Shape classification via Hu moments

Shape discrimination uses the seven Hu moment invariants computed on binarized surface plots, classified by an unsupervised Gaussian Mixture Model (GMM) with five components corresponding to one edge and four flat shapes (circle, square, triangle, ellipse). Systematic evaluation of eight Hu moment combinations showed that only M1–M2–M3 exceeded 90% average diagonal accuracy, reaching 96%, while combinations 4–8 lagged by at least 10 percentage points — indicating higher-order moments contribute little discriminative information for this stimulus set. Within the chosen combination, the edge class achieved perfect classification (F1 = 1.00); square and triangle reached 100% true positives with non-zero false positives; all shapes scored F1 ≥ 0.90. The authors appropriately frame this as proof-of-principle for interpretable feature discrimination rather than a benchmark for unconstrained object recognition.

Limitations and open questions

Several limitations are conceded explicitly. First, pixel averaging over each fiber endface discards spatial information mixed within fibers; deep-learning approaches on raw OFA images retain this information, which partly explains their superior precision. Second, the calibration dataset consists only of indentations directly above each fiber with a single indenter size; extension to arbitrary contact distributions remains open. Third, all results were obtained under controlled laboratory conditions with short (95 mm polished) fibers and fixed packaging; robustness under longer routing, repeated bending, packaging tolerances, and environmental variation is unquantified. Fourth, shape classification used five canonical indenters at fixed 3 mm depth, leaving generalization to naturalistic contacts untested. Finally, the LED ring required for optimal illumination constrains miniaturization, though future designs may remove it.

Conclusion

OptiTac demonstrates that a one-to-one pin-to-fiber architecture can physically structure tactile signals so that contact location, size, and shape are recoverable through interpretable image-moment methods, achieving calibrated sub-millimeter localization (0.4 mm RMSE), hyperacuity relative to element spacing, and 96% shape-classification accuracy with a GMM on the first three Hu moments. The work establishes a credible analytical alternative to black-box inference for structured tactile sensors, while its deployment-scale claims — long-fiber robustness, arbitrary contact distributions, and integration with robotic end effectors — remain to be validated.