Fiber Bragg Grating Whiskers for Bioinspired Hydrodynamic Perception on Underwater Robots

Hao Li; Tianyu Tu; Siyue Yao; Ziyang Chang; Juhyun Jung; Xiaochi Xie; Long Yin Chung; Tian-Ao Ren; Genliang Chen; Mark Cutkosky

Source-linked AI summary

Fiber Bragg Grating Whiskers for Bioinspired Hydrodynamic Perception on Underwater Robots

Hao Li, Tianyu Tu, Siyue Yao, Ziyang Chang, Juhyun Jung, Xiaochi Xie, Long Yin Chung, Tian-Ao Ren, Genliang Chen, Mark Cutkosky

arXiv:2608.24724v1Aug 25, 2026cs.RO

TL;DR

Underwater robots need passive sensing when vision and acoustics are degraded or undesirable, while translating whisker sensing into robot behavior remains challenging. This paper develops an FBG whisker system and demonstrates flow sensing, delayed wake detection, and robot branch selection.

Problem
Underwater robots require sensing methods that remain useful when conventional optical and acoustic sensors are degraded, power-intensive, ambiguous, or undesirable.
Method
The paper develops a robot-mountable, two-axis FBG whisker with a superelastic NiTi core and urethane shell, then evaluates controlled flow, foil-generated wakes, and branch selection.
Results
17 correct decisions in 20 runs were achieved online, while held-out windows reached 77.4% accuracy, 0.863 AUROC, and 0.776 average precision.
Takeaways & Limitations
The study connects bioinspired hydrodynamic sensing to robot behavior by showing that whisker signals can initiate action on a mobile underwater platform.
Takeaways & Limitations
The demonstrated progression is evaluated within a single robot-compatible FBG whisker system.

Abstract

from arXiv · show

Harbor seals track hydrodynamic trails with their vibrissae, enabling passive perception of moving targets in dark or turbid water. Inspired by this capability, we present compact fiber Bragg grating (FBG) whiskers for underwater robots. Like seal whiskers, they have a non-uniform taper and elliptical cross-section. Controlled towing experiments show a monotonic relative-flow response from 0.1 to 0.6 m/s, a strong reduction of self-induced oscillation relative to a cylindrical baseline, and a pronounced dependence on angle of attack. Experiments with a pitching foil show that the whiskers can detect the characteristic vortices shed by a stationary or moving source, detectable several seconds after the source has passed. Using this information, a single front-mounted whisker enabled a small underwater robot to distinguish between continuing straight and executing a turn, selecting the correct branch in 17 of 20 trials (85.0%) from whisker signals alone. These results connect bioinspired hydrodynamic sensing to robot action and suggest the utility of whiskers for tracking underwater objects.

I. INTRODUCTION

The paper develops an FBG whisker system inspired by seals’ ability to sense delayed hydrodynamic trails in visually degraded water. It links whisker design and flow sensing to robot branch selection and field wake detection.

Harbor seals can follow hydrodynamic trails after delays of ten seconds or more, even when vision is blocked.This motivates passive whisker sensing for underwater robots operating in dark or turbid environments.
The study presents a robot-mountable FBG whisker with a superelastic NiTi core and urethane shell that recovers after collision-like bending.The design provides two-axis strain readout and supports connecting multiple whiskers to a single optical fiber.
Controlled experiments evaluate relative flow-speed response, angle-of-attack dependence, and vortex-induced-vibration reduction against a cylindrical baseline.These measurements establish basic flow-sensing properties needed for interpreting whisker signals on a robot.
Pitching-foil experiments characterize distinct wake patterns from stationary and moving sources, including delayed signatures relevant to trail following.The moving-source regime is used to connect wake sensing with a robot decision task.
A small ROV used whisker signals to choose whether to proceed straight or turn, while an ocean deployment detected vortices shed by a human diver’s fin.These demonstrations connect bioinspired hydrodynamic sensing to robot behavior and field wake detectability.

2) Whisker-inspired flow sensors:

The paper situates its whisker within bioinspired flow-sensing research and implements seal-like morphology in a compact optical sensor. The design combines an undulating elliptical shaft, embedded FBGs, and a recoverable NiTi–urethane structure.

Prior artificial whiskers use capacitive, piezoresistive, piezoelectric, triboelectric, magnetic, optical, and fiber-optic transduction.Existing systems commonly demonstrate steady-flow estimation and directional response, while wake studies often use cylinders or fin-like paddles.
FBGs provide electrically passive, saltwater-compatible, electromagnetically immune, sensitive, multiplexable, and remotely interrogable sensing.The paper builds on prior underwater FBG whiskers that demonstrated hydrodynamic and direction-related signatures.
A towed bluff body can generate a Kármán vortex street that induces vortex-induced vibration on the body.This motivates shaping the whisker to reduce self-induced oscillation during flow sensing.
The biomimetic profile uses peak and trough elliptical semi-axes, two inclination angles, and an undulation wavelength.The reported peak and trough aspect ratios are a/b = 2.95 and k/l = 1.90.
Each approximately 100 mm whisker uses a 0.4 mm NiTi wire overmolded with urethane and a 3D-printed base housing two orthogonal FBGs.The shaft is fabricated through a printed mandrel, compliant silicone mold, and urethane overmolding workflow.

III. ANALYTICAL MODEL AND PARAMETRIC DESIGN

The whisker was modeled as a hybrid elastic mechanism and calibrated through controlled loading to connect deformation with FBG wavelength shifts. Parametric simulations and experiments guided bridge selection while exposing load-dependent performance limits.

Analytical model: A hybrid parallel-serial elastic model represents the complete whisker-plus-base structure and transforms deformation into algebraic equations.Each segment is modeled as an equivalent 6-DoF elastic mechanism, with Jacobians and stiffness matrices describing the structure.
Model validation: Calculated deflections matched observations within 0.5 mm for the H = 2 mm, L = 45 mm design.Five designs varied bridge length L ∈ {45 mm, 35 mm, 25 mm} or height H ∈ {2 mm, 4 mm, 6 mm}; simulated and experimental sensitivity results were compared.
Parametric design: The testing setup applies controlled motion with a linear stage and measures contact force using a load cell.The setup also compares captured and simulated deflections and relates successive deflection states to contact forces and wavelength shifts.
Parametric design: Sensitivity improved as H decreased and L increased, leading to selection of {H, L} = {2 mm, 45 mm}.Further reduction of H was avoided because attaching the whisker consistently to the base bridge becomes difficult.
Performance boundary: The minimum load detection limit is ≈0.5 mN under the equivalent point-load scenario.The estimate considers an FBG wavelength-fluctuation noise level of approximately 2.7 pm at ±3σ.

IV. EXPERIMENTS AND RESULTS

The study evaluates hydrodynamic cues under controlled foil-generated wake conditions and tests whether delayed wake information can support a constrained robotic decision. It combines foil experiments, mobile-ROV branch selection, and an ocean deployment as a qualitative robustness check.

Research questions: The study asks which hydrodynamic cues FBG whiskers measure under relative flow and foil-generated wakes, and whether delayed cues support robotic decision-making.The decision task concerns distinguishing between continuing straight and executing a turn.
Wake experiments: A pitching NACA0012 foil generates unsteady flows for comparison between fixed and co-translating finite-Strouhal-number conditions.The co-translating condition better approximates the delayed hydrodynamic cue left by a moving source.
Robotic evaluation: A single whisker on a mobile ROV was evaluated for binary branch selection using whisker signals alone in a delayed foil-generated trail.The study also deployed the system in the ocean to qualitatively assess wake detectability outside the controlled pool environment.

A. Flow Sensing

Controlled towing established three sensing behaviors: reduced self-induced vibration, monotonic response to relative flow speed, and angle-of-attack dependence. The flow-speed calibration applies only within the tested configuration and environment.

Flow-sensing objectives: Controlled towing characterized suppression of VIV, relative-flow-speed response, and angle-of-attack dependence before wake experiments.These tests define operating behavior under well-controlled relative flow.
Experimental setup: A three-axis gantry towed the rigidly mounted whisker along the x-axis at prescribed constant speeds to generate controlled relative flow.The towing experiments were conducted in a 4.5 m × 2.2 m × 1.2 m outdoor pool.
Geometry comparison and VIV suppression: The cylindrical whisker produced a 0.132 nm FBG wavelength shift and approximately 5.4 mm tip motion, versus 0.014 nm and 0.9 mm for the wavy tapered whisker.The wavy non-tapered whisker produced 0.020 nm and approximately 1.5 mm; both wavy geometries reduced fluctuations relative to the cylindrical baseline.
Response to relative flow speed: From 0.1 to 0.6 m/s, the wavy tapered whisker’s mean baseline-referenced wavelength shift increased monotonically in magnitude with towing speed.The grouped mean response in pm was fit with a quadratic calibration curve.
Response to relative flow speed: The calibration establishes relative flow-speed sensitivity over the tested range but is not universal across orientation, mounting, or flow environment.Its interpretation is therefore bounded by the experimental configuration.

4) Effect of angle of attack:

Whisker response depends strongly on orientation relative to incoming flow, so angle of attack must be considered when comparing measurements. The experiments also establish the controlled setup and response behavior used to assess this dependence.

4) Effect of angle of attack:: The towing setup rotates the wavy tapered whisker from 0° to 90° in 15° increments while generating controlled relative flow.
4) Effect of angle of attack:: At 0.6 m/s, both FBG channels fluctuate weakly at 0° and 15°, but oscillations become much more pronounced from 30° onward.The trend is consistent with increasing projected cross-flow area as the whisker rotates.
4) Effect of angle of attack:: The study compares fixed and translating foil-generated unsteady-flow regimes because source kinematics determine the resulting disturbance field.
4) Effect of angle of attack:: Whisker response to foil-generated unsteady flow was measured while varying pitching amplitude and frequency at a fixed relative position.
4) Effect of angle of attack:: AoA is defined by the orientation of the whisker’s narrow side relative to the incoming flow.

3) Fixed pitching foil in quiescent water:

A stationary pitching foil in quiescent water produces a spatially varying whisker response, with stronger signals laterally and weaker responses directly downstream. Because there is no mean advection, this benchmark does not directly reproduce the delayed trail of a moving target.

3) Fixed pitching foil in quiescent water:: The fixed-foil response map shows stronger whisker signals at lateral offsets and weaker responses directly downstream.The foil pitched at f = 0.6 Hz with angular amplitude A = 0.4 rad for eight cycles.
3) Fixed pitching foil in quiescent water:: Without mean advection, the stationary foil configuration is not expected to generate the downstream-convected trail relevant to delayed branch selection.
3) Fixed pitching foil in quiescent water:: The fixed-foil case serves as a controlled unsteady-flow benchmark rather than a direct proxy for a moving target’s hydrodynamic trail.
3) Fixed pitching foil in quiescent water:: The study contrasts this fixed regime with co-translating foil-whisker experiments intended to approximate delayed sensing behind a moving source.
3) Fixed pitching foil in quiescent water:: Figure 5 maps whisker responses around a NACA0012 foil using repeated measurements over a spatial grid.

5) Connection to robot branch selection:

The study uses delayed foil-generated hydrodynamic trails to test whether whisker signals alone can drive binary branch selection by an underwater robot. A single front-mounted whisker supported online decisions between continuing straight and turning.

5) Connection to robot branch selection:: The foil generated a delayed hydrodynamic trail, after which the robot chose between straight and turning branches using whisker input alone.The robot advanced from a fixed start point to a decision region before selecting a candidate branch.
5) Connection to robot branch selection:: 77.4% held-out test accuracy was achieved by a logistic-regression classifier using engineered features from both FBG channels.The held-out test set contained 20 demonstrations; training and validation accuracies were 92.2% and 73.7%, respectively.
5) Connection to robot branch selection:: 85.0% online success was obtained across 20 run-level trials, with the robot reaching the corresponding target region in 17 cases.The classifier produced one branch decision after the robot entered the decision region.
5) Connection to robot branch selection:: The branch-selection experiment was conducted with a single whisker in a controlled pool and a binary decision rather than continuous trail centering.The authors frame it as an initial systems demonstration, not continuous biological wake tracking.

D. Field Deployment

The paper frames its contribution as a progression from seal-inspired hydrodynamic sensing to action on a mobile underwater robot. It situates the work within broader whisker-sensing research while avoiding direct ranking across differing comparison protocols.

D. Field Deployment: The work’s main contribution is an experimental progression from biological principles to robot behavior, rather than only a new whisker transducer.The progression includes sensing, wake interpretation, and action on a mobile platform.
D. Field Deployment: Seal-inspired wake interpretation motivates reducing self-induced vibration, establishing flow sensing, selecting a relevant wake regime, and initiating action on a mobile platform.The paper connects these requirements within one robot-compatible FBG whisker system.
D. Field Deployment: Prior artificial whiskers demonstrate capabilities including steady-flow estimation, directional response, wake detection, source localization, and vehicle-state estimation.The paper treats these capabilities as contextual background rather than a rank ordering.
D. Field Deployment: Comparison across studies is limited because flow-speed coverage, wake distance, sensitivity, response time, and SNR are defined and reported differently.A reported 0 m/s condition is treated as a no-flow baseline, not a demonstrated lower flow-speed limit.

APPENDIX A GEOMETRY, ROBUSTNESS, AND CROSS-STUDY CONVENTIONS

The appendix documents the whisker’s tapered geometry, durability testing, and conventions for interpreting cross-study sensing metrics. These details define the tested design and constrain comparisons.

APPENDIX A GEOMETRY, ROBUSTNESS, AND CROSS-STUDY CONVENTIONS: The whisker geometry uses an undulated cross-section with a quadratic taper applied from base to tip.The semi-axis parameterization follows Hanke et al., with t0 = 0.029 specified for the taper.
APPENDIX A GEOMETRY, ROBUSTNESS, AND CROSS-STUDY CONVENTIONS: 500 repeated indentation cycles showed no obvious loss of mechanical-to-optical response in the recorded FBG 1 waveforms.Early and late cycles were compared for the selected whisker design.
APPENDIX A GEOMETRY, ROBUSTNESS, AND CROSS-STUDY CONVENTIONS: Flow-speed comparisons list nonzero validation conditions, while 0 m/s is treated as a static baseline rather than a demonstrated lower measurement limit.This convention prevents a no-flow trace from being interpreted as positive low-speed validation.
APPENDIX A GEOMETRY, ROBUSTNESS, AND CROSS-STUDY CONVENTIONS: For this work, response time is defined as 29.74 ms from video-detected contact to the first clear FBG departure from baseline.Steady-flow SNR is defined as 20 log10(|µ_steady − µ_static|/σ_static).

A. Towing Experiments

Controlled towing and wake-map procedures characterize how whisker signals vary with geometry, speed, angle, and source motion. The analysis uses standardized preprocessing and explicitly accounts for reconstruction assumptions.

A. Towing Experiments: Geometry traces were interpolated, baseline corrected, smoothed, and evaluated in a common fixed-duration stable-velocity window without geometry-specific peak searches.Centering by subtracting the midrange changes display position but not peak-to-peak amplitude.
A. Towing Experiments: Five runs at each commanded speed from 0.1 to 0.6 m/s were reduced to grouped means and standard deviations after within-speed outlier screening.The second-order curve is a visualization fit to the grouped means.
A. Towing Experiments: Each displayed map point comes from a repeated single-location run, so the maps are not simultaneous planar flow measurements.For the fixed-foil map, the two FBG channels form a response vector whose Euclidean norm sets magnitude.
A. Towing Experiments: Moving-source runs were aligned to foil-motion onset, combined across opposite initial phases, reflected about the trajectory centerline, and transverse-sign inverted.The reconstruction assumes approximate centerline symmetry.
A. Towing Experiments: The co-translating robot condition used U = 0.46 m/s, f = 1.2 Hz, Aθ = 0.4 rad, and St = 0.35.The Strouhal number was defined as St = fA_TE/U for this operating point.

APPENDIX C MODEL IMPLEMENTATION AND BRANCH CLASSIFIER

The appendix models the whisker–bridge structure as coupled discretized elastic mechanisms and specifies a signal-processing and classification pipeline using only the two FBG channels. The frozen classifier achieved 17 correct first decisions in 20 online runs.

MODEL IMPLEMENTATION: Each whisker and bridge branch is discretized into short segments represented by equivalent six-DoF elastic mechanisms.Each segment contains three prismatic and three revolute joints, with stiffness assigned from structural properties.
CLASSIFIER DATA AND PREPROCESSING: The classifier used only FBG 1 and FBG 2, with trajectory-level splits producing 463 training, 114 validation, and 186 held-out test windows.Each decision used a trailing 2.0 s window, with resampling, padding, detrending, and mean centering applied before feature extraction.
FEATURE EXTRACTION: Each window produced 42 engineered features, including autocorrelation, Welch-PSD, dominant-frequency, RMS, mean, and successive-RMS-ratio descriptors.The features were distributed as 21 per channel across four equal temporal segments for temporal characterization.
BRANCH CLASSIFIER: The frozen logistic-regression model achieved 77.4% accuracy, 0.863 AUROC, and 0.776 average precision on held-out windows.The model used an ℓ1 penalty, C = 0.1, and a fixed 0.5 decision threshold; 16 of 42 coefficients were nonzero.
BRANCH CLASSIFIER: 17 of 20 online runs received correct first branch decisions, making the run-level result behaviorally relevant despite overlapping offline windows.The online evaluation used one first-decision window per run.