Source-linked AI summary
Functional mimicry of Ruffini receptors with Fiber Bragg Gratings and Deep Neural Networks enables a bio-inspired large-area tactile sensitive skin
Luca Massari, Giulia Fransvea, Jessica D'Abbraccio, Mariangela Filosa, Giuseppe Terruso, Andrea Aliperta, Giacomo D'Alesio, Martina Zaltieri, Emiliano Schena, Eduardo Palermo, Edoardo Sinibaldi, Calogero Maria Oddo
TL;DR
The paper addresses the need for soft, conformable artificial skins that can detect both contact location and intensity over large, curved robot surfaces. It develops a modular FBG-based skin with deep-learning decoding, achieving 35 mN force and 3.2 mm localization median test errors. Generalization and robustness of the deep-learning strategies remain topics for future work.
Problem
Soft tactile skins still need to detect contact location and intensity over large, curved robot surfaces while adapting to collaborative interaction settings.
Method
The authors combine curved modular FBG sensorized patches with a CNN, feed-forward networks, and multigrid neuron integration to decode tactile inputs.
Results
35 mN (IQR = 56 mN) and 3.2 mm (IQR = 2.3 mm) were the median test-set errors for force and contact-source localization, respectively.
Takeaways & Limitations
The proposed modular tactile patches demonstrated simultaneous prediction of contact intensity and location for collaborative-robot applications.
Takeaways & Limitations
Future work will further investigate the generalization ability and robustness of the deep-learning strategies.
Abstract
from arXiv · showhide
Collaborative robots are expected to physically interact with humans in daily living and workplace, including industrial and healthcare settings. A related key enabling technology is tactile sensing, which currently requires addressing the outstanding scientific challenge to simultaneously detect contact location and intensity by means of soft conformable artificial skins adapting over large areas to the complex curved geometries of robot embodiments. In this work, the development of a large-area sensitive soft skin with a curved geometry is presented, allowing for robot total-body coverage through modular patches. The biomimetic skin consists of a soft polymeric matrix, resembling a human forearm, embedded with photonic Fiber Bragg Grating (FBG) transducers, which partially mimics Ruffini mechanoreceptor functionality with diffuse, overlapping receptive fields. A Convolutional Neural Network deep learning algorithm and a multigrid Neuron Integration Process were implemented to decode the FBG sensor outputs for inferring contact force magnitude and localization through the skin surface. Results achieved 35 mN (IQR = 56 mN) and 3.2 mm (IQR = 2.3 mm) median errors, for force and localization predictions, respectively. Demonstrations with an anthropomorphic arm pave the way towards AI-based integrated skins enabling safe human-robot cooperation via machine intelligence.
Introduction
The paper addresses the challenge of building soft, conformable, large-area tactile skins that support safe physical human–robot interaction by detecting contact properties across curved robot surfaces. It proposes an FBG-based, bio-inspired skin using overlapping receptive fields and deep learning to infer contact location and magnitude.
- Safe physical cooperation depends on tactile feedback that lets robots sense and respond to contact throughout their bodies.
- Large-area robot skins must combine millimetric localization, milliNewton force sensing, millisecond accuracy, softness, stretchability, light weight, and conformability.
- Conventional tactile and proximity sensors remain bulky and rigid, limiting flexibility, deformability, and adaptation to unconstrained environments.
- Covering anthropomorphic robots with soft, curved sensing components remains a major challenge, with dense sensor arrays also creating wiring problems on complex shapes.
- The proposed skin uses curved modular patches with photonic FBG sensors and one wavelength-multiplexed wiring element to cover collaborative robots.
- Overlapping FBG receptive fields and AI strategies decode both contact-force magnitude and localization across the large-area artificial skin.
Results
The soft forearm-like skin embedded FBG transducers whose spatial arrangement and cross-talk supported learned inference of force and contact location. Deep-learning models achieved millimeter-scale localization and milliNewton-scale force prediction across the curved surface, with larger errors at the edges.
- Skin structure and receptive fields: FBG spacing varied from 12.9 mm to 24.5 mm, implementing a nonuniform sensor density along the artificial forearm.
- Skin structure and receptive fields: 50.6 mN was the contact-detection threshold at 75% probability for increasing microfilament diameters.
- Skin structure and receptive fields: 15.9 mm2 was the median receptive-field hot-spot area, while neighboring FBG activation and wavelength changes reflected load location and magnitude.
- Deep-learning inference: The CNN and four MLPs used 16 FBG readouts to reconstruct force and localize contacts from randomized indentations up to 2.5 N.
- Deep-learning inference: 30 mN was the median five-fold cross-validation error for force prediction, with a 1 mN IQR.
- Deep-learning inference: 35 mN (IQR = 56 mN) and 3.2 mm (IQR = 2.3 mm) were the test-set median errors for force and position predictions, respectively.
- Performance boundaries: Prediction errors increased at the skin edges, where the limited sensor count constrained learned triangulation near the boundary.
Discussion
The discussion presents a modular, curved FBG skin that combines biomimetic receptive fields with deep learning to estimate contact location and force across a large sensorized surface. Performance was strong across the tested surface and force range, while edge effects and curvature generalization remain boundaries for deployment.
- The modular tactile patches can fit different robot architectures, including retrofitting existing robots.
- The skin reproduced Ruffini-like receptive-field behavior in most FBGs, while some sensors showed clustered or multiple responsive areas linked to polymer irregularities.
- FBG sensors offer multiplexing, dense integration, high sensitivity, and electromagnetic-interference immunity for large-area tactile skins.The latter property supports compatibility with magnetic-resonance environments.
- Deep learning decoded raw FBG wavelengths into distributed contact localization and force intensity, with NIP combining four half-pitch-shifted MLP grids.
- 35 mN (IQR = 56 mN) and 3.2 mm (IQR = 2.3 mm) were the median test-set errors for force and source localization, respectively.
- Localization error was uniform over the 120 mm by 90° tested surface and the 0 N–2.5 N force range, with lower accuracy and larger fluctuations at the edges.
- Continuous skins may reduce boundary effects, and future work must test robustness when curvature changes between training and operation.
- The demonstrated curved skin identifies both the magnitude and location of normal loads, extending earlier one-dimensional sensing toward large-area coverage.
Methods
The study develops a curved, large-area soft tactile skin by embedding 16 FBGs in an 8 mm polymeric forearm cover and decoding their outputs for force and contact localization. Experiments combine FEM design analysis, Von Frey sensitivity testing, and automated force-controlled indentations.
- Sensing principle: FBGs transduce contact through strain-induced shifts in reflected Bragg wavelength, while the polymeric substrate transfers pressure and protects the optical fiber.The Bragg wavelength depends on effective refractive index and grating period, both affected by strain.
- Skin design: The artificial skin uses a soft polymeric layer with an embedded optical fiber carrying 16 FBGs across a 150 mm forearm region covering 145°.The FBGs have 8 mm lengths, wavelengths from 1530 to 1564.5 nm, and a 2.3 nm pitch.
- Automated testing: A four-degree-of-freedom mechatronic platform collected position and intensity data from force-controlled indentations for tactile decoding experiments.The platform controlled horizontal and vertical translation plus rotation, with a six-axis load cell measuring applied force.
Author Contributions Statement
The authors divided the work across artificial-skin development, experimental setup, data collection, analysis, AI algorithms, simulations, supervision, and manuscript preparation.
- Skin development: LM, JDA, GT, MZ, EDS, and CMO designed and developed the artificial skin embedding the optical fiber.
- Experiments and analysis: EP, EMS, EDS, and CMO contributed to the experimental setup, while LM, GF, JDA, and MF conducted experiments and analyzed data with support from AA, JDA, and EP.
- Algorithms and simulations: LM, GF, EDS, and CMO conceived and developed the AI algorithms, and LM and MF performed the FEM simulations.
- Supervision: CMO planned and supervised the scientific work and protocols, while EDS co-supervised the scientific work, protocols, and FEM simulations.
- Manuscript preparation: LM, GF, JDA, MF, EDS, and CMO wrote the manuscript, with AA responsible for artworks, figures, and supplementary materials.
Competing Interests Statement
The authors report a patent filing covering the developed artificial skin and a collaborative robot arm integrating FBG transducers; the remaining authors report no competing interests.
- Disclosure: LM, JDA, GT, MZ, EP, EMS, EDS, and CMO disclose a patent filed on the artificial skin and collaborative robot arm integrating FBG transducers.The application number is IT201900003657A1.
- Disclosure: The remaining authors declare no competing interests.