Five fluorescent whiskers. One camera.
Result
Five fluorescent whiskers and one camera estimate where an underwater thruster is, using the flow it creates.
Question
In dark water, observation often requires us to add light or sound.Those signals can change animal behavior and the observations we make.
Solution
The source creates the flow. Fluorescence makes the sensor's response visible to a camera.
01 / Flow sensing
Harbor seals use their whiskers to follow hydrodynamic trails. Inspired by this ability, our array senses the flow-induced bending of five whiskers. A camera observes the array, and a model estimates the source position from their deformation.
02 / Local optical readout
Fluorescent pigment gives the whiskers strong contrast against the dark water. The camera reads their motion with illumination concentrated around the array.
The whiskers become visible under low-power UV excitation (approximately 100 mW).
Each whisker contains a 0.4 mm superelastic nitinol core inside a urethane shell with 5 wt% green fluorescent pigment. Two-stage molding forms the undulating, tapered profile.
System
The sensing platform combines the whisker array, local UV excitation, and a visual tracking pipeline.
The platform is a 1080p, 60 Hz camera, two 365 nm UV lights on either side of it, and five whiskers mounted at different pitch and yaw angles. A single view cannot recover the full 3D bending of one whisker, but because the whiskers point in different directions, a deformation that is hard to see on one is pronounced on another.
Computer vision pipeline
Each frame is enhanced to isolate the green fluorescence, the five whiskers are segmented, and a point at a fixed arc length along each spline centerline is tracked. Two seconds of these displacements (5 whiskers × 2 axes × 50 samples) go into a lightweight CNN that outputs the source position (r, θ).
Performance
A thruster on a three-axis gantry is moved over a fan-shaped region 350–950 mm from the array and ±30° around its axis, always facing the array at the same power. Training uses 16 grid positions (8,640 two-second samples over two days); the test set adds 108 off-grid positions the model has never seen.
Across 16 on-grid and 108 off-grid test positions.
| Test points | Δr MAE (mm) | Δθ MAE (°) | Distance MAE (mm) |
|---|---|---|---|
| 16 on-grid | 45.2 | 0.47 | 47.2 |
| 108 off-grid | 77.7 | 2.78 | 94.2 |
| All | 73.5 | 2.48 | 88.1 |
Bearing is learned better than range: angular estimates are generally more concentrated than range estimates, while off-grid range estimates are pulled toward the nearest training radius. The mapping still interpolates, with off-grid predictions clustering around their true locations.
Does the estimate follow different paths?
The model is trained only on stationary sources. The arc-motion trials above test it at radii of 350, 450, and 550 mm, with the largest errors on off-grid paths.
In a separate radial-motion experiment, the thruster moves from 450 to 750 mm at 1.6 mm/s. The estimate follows it with 51.7 mm range error and 0.63° angular error; six repeated runs show the same trend. The full results below include both radial and arc trajectories.
The localization experiments use a thruster in a pool, at the array's height, facing the array and operating at a fixed power. The results establish planar source localization under these conditions. Observing marine animals in the deep sea, and measuring how this sensing approach affects their behavior, remain future work.
arXiv:2610.07275 ↗
· October 5, 2026 · v1
Read the paper PDF ↗
@misc{xie2026fluorescenceenhancedwhiskerarrayvisionbased,
title={Fluorescence-enhanced Whisker Array with Vision-based Deformation Analysis for Underwater Source Localization},
author={Xiaochi Xie and Hao Li and Shixuan Zhao and Siyue Yao and Shuran Song and Mark R. Cutkosky},
year={2026},
eprint={2610.07275},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2610.07275},
}Deep-water biological observation is essential for understanding marine organisms and their interactions with the environment. However, conventional optical and acoustic approaches can introduce stimuli that alter animal behavior and bias biological observations. This paper proposes a fluorescence-enhanced whisker array sensing system that pinpoints underwater hydrodynamic sources through local optical readout rather than direct source imaging. Five spatially oriented whiskers, fabricated with nitinol cores and fluorescent urethane shells, are integrated with ultraviolet excitation and a monocular camera. Image enhancement and segmentation are applied to track the whisker deformation. Further, a lightweight convolutional neural network captures temporal and cross-whisker features from 2 s sequences to estimate source localization. Pool experiments achieve a mean spatial localization error of 88 mm, with 73.5 mm in radius and 2.5° in angle, across a test region of 600 mm with ±30°. Real-time localization of a moving thruster further demonstrates the capability of the proposed method in dynamic scenarios, highlighting its potential for integration into underwater robots for hydrodynamic source detection, localization, and tracking in low-light environments.
Questions about the paper or the whiskers: Xiaochi Xie.
The complete walkthrough, from sensing hardware to localization experiments.