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

Technical Summary: OptiTac – Biomimetic Optical-Fiber Tactile Sensing

Problem Statement
Robotic systems require tactile sensing capabilities that approach the adaptability, resolution, and distributed nature of human skin to enable dexterous manipulation and safe interaction. While human skin utilizes a complex architecture where mechanoreceptors (e.g., Merkel cells, Meissner corpuscles) transduce mechanical stimuli and nerves transmit signals to the spinal cord and brain for processing, current robotic tactile sensors face limitations. Vision-based tactile sensors (VBTS) like the TacTip offer high resolution but require cameras to be placed directly behind the sensing skin, limiting scalability and distributed sensing in complex robotic geometries. Conversely, optical fiber array (OFA) sensors often rely on deep learning "black-box" models for inference or use reflective skins that lack the biomimetic mechanical pre-processing found in human skin. There is a need for a sensor architecture that routes tactile information away from the sensing surface while preserving high spatial resolution and enabling interpretable, analytical inference without relying solely on opaque deep learning models.

Methodology
The authors propose OptiTac, a biomimetic tactile sensor that emulates the human mechanoreceptor-to-nerve architecture. The system consists of two distinct modules separated by a coherent optical fiber array (OFA):

1. Soft Tactile Sensing Module: A soft skin containing an array of rigid pins with white markers on their tips, embedded in a transparent elastomeric gel. This module mimics the dermal papillae structure of human skin.
2. Remote Imaging Module: A camera housed in a separate casing that captures light transmitted through the OFA.
3. Optical Fiber Array (OFA): A coherent hexagonal array of 217 individual plastic optical fibers (1 mm diameter) that physically connects the two modules.

Key Design Principle: The sensor employs a one-to-one pairing between each mechanical pin on the soft skin and a single optical fiber. When the skin deforms, the pins lever, causing the white markers to move in and out of alignment with the corresponding fibers. This alignment change modulates light intensity, which is transmitted to the remote camera.

Data Processing and Inference:
Instead of using end-to-end deep learning, the authors treat the tactile data as images and apply simple analytical methods:

• Pre-processing: Raw RGB images are converted to grayscale, cropped, and segmented using watershed segmentation to identify individual fiber endfaces.
• Surface Plot Generation: The discrete intensity values from the 217 fibers are interpolated (cubic interpolation) to generate high-resolution light intensity surface plots.
• Centroid Localization: Image moments are used to calculate the contact centroid. A calibration process using k-d trees and Gaussian filtering corrects for spatial errors, particularly at the periphery of the array.
• Width Estimation: Contact width is estimated by calculating eigenvalues from central moments of the intensity distribution, assuming a Gaussian contact profile.
• Shape Classification: Hu moments (seven moment invariants) are extracted from binarized surface plots. A Gaussian Mixture Model (GMM), an unsupervised learning method, is used to cluster and classify contact shapes (edge, circle, square, triangle, ellipse).

Key Results

• Signal Transmission: The system successfully transmits structured tactile information from the compliant sensing interface to a physically separate imaging module, maintaining signal coherence.
• Localization Performance: The one-to-one pin-fiber pairing produced localized intensity changes, unlike sparse or dense patterns which generated complex, non-intuitive signals. The calibrated system achieved a coefficient of determination ($R^2$) of 0.99 for both seen and unseen datasets. The Root Mean Square Error (RMSE) for localization was 0.4 mm in both x and y axes for unseen data (interpolated between fibers), demonstrating hyperacuity (localization precision significantly finer than the 1.5 mm spacing between fiber pairs). This performance is comparable to or better than existing OFA sensors using deep learning (e.g., Lu et al., 2023 reported 0.02–0.03 mm, but Baimukashev et al., 2020 reported 1.1–1.4 mm with larger fiber spacing).
• Contact Width Estimation: The model showed a strong linear relationship ($R^2$ of 0.95–0.97) between true and measured contact diameters. RMSE values were 1.2 mm and 1.0 mm for major and minor axes, respectively, which are smaller than the fiber spacing. Accuracy was higher for indenters smaller than 10 mm.
• Shape Classification: Using a combination of the first three Hu moments ($M_1, M_2, M_3$) with a GMM, the sensor achieved an average classification accuracy of 96%. It correctly classified an edge shape with 100% true positives (F1-score of 1.00) and achieved minimum F1-scores of 0.90 for flat shapes (circle, square, triangle, ellipse).

Significance and Claims
The paper claims that OptiTac establishes a scalable, interpretable route toward distributed tactile sensing in robots. Its primary contributions are:

1. Biomimetic Architecture: It demonstrates a physical pre-processing mechanism (one-to-one pin-fiber pairing) that structures tactile signals for intuitive extraction, mimicking the separation of sensing and transmission in biological systems.
2. Interpretability: By using analytical image descriptors (image moments) rather than deep learning "black boxes," the system links inferred contact properties directly to physically meaningful deformation patterns. This supports sensor design, calibration, debugging, and transfer across geometries.
3. Hyperacuity without Deep Learning: The system achieves sub-millimeter localization and contact size estimation comparable to state-of-the-art deep learning approaches, but through a transparent analytical pipeline.
4. Proof-of-Concept for Distributed Sensing: The physical separation of the sensing skin from the imaging hardware via optical fibers offers a practical pathway for integrating tactile sensing into complex robotic end-effectors where embedding cameras is difficult.

The authors note that while the current prototype uses short fibers and controlled conditions, the design principle is generalizable. Future work will focus on integrating the sensor with robotic grippers, extending calibration to arbitrary contact distributions, and assessing robustness under longer fiber routing and environmental variations.