A Machine Learning Approach to Contact Localization in Variable Density Three-Dimensional Tactile Artificial Skin

Carson Kohlbrenner † Mitchell Murray † Yutong Zhang † Caleb Escobedo Thomas Dunnington Nolan Stevenson Nikolaus Correll Alessandro Roncone · 2024

Read A Machine Learning Approach to Contact Localization in Variable Density Three-Dimensional Tactile Artificial Skin by Carson Kohlbrenner † Mitchell Murray † Yutong Zhang † Caleb Escobedo Thomas Dunnington Nolan Stevenson Nikolaus Correll Alessandro Roncone (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-00689.txt.

Introduction

Carson Kohlbrenner and seven co-authors present a 2024 arXiv study on teaching artificial skin to pinpoint touch. Real skin localizes contact effortlessly, but robot skins are usually flat sheets with evenly spaced sensors, unlike the curved, irregular surfaces robots actually wear. The team studied a semi-conical skin studded with capacitive touch sensors in an uneven, not fully known arrangement.

A neural network learned the mapping from sensor patterns to touch locations directly. Online, it localized contacts with an average error of about six millimeters, give or take three. Because the network infers the layout from data, the method tolerates unknown shapes and sensor placements that break geometry-based approaches. The authors offer it as a versatile tool for tactile robots, prosthetics, and any device whose skin cannot be a neat grid.

Catalog record

CREATORCarson Kohlbrenner † Mitchell Murray † Yutong Zhang † Caleb Escobedo Thomas Dunnington Nolan Stevenson Nikolaus Correll Alessandro Roncone
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE13 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-00689

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

A Machine Learning Approach to Contact Localization in Variable Density Three-Dimensional Tactile Artificial Skin Thanks: Department of Computer Science, University of Colorado Boulder Thanks: † represents equal contributions. Thanks: [email protected] Carson Kohlbrenner † Mitchell Murray † Yutong Zhang † Caleb Escobedo Thomas Dunnington Nolan Stevenson Nikolaus Correll Alessandro Roncone Abstract Estimating the location of contact is a primary function of artificial tactile sensing apparatuses that perceive the environment through touch. Existing contact localization methods use flat geometry and uniform sensor distributions as a simplifying assumption, limiting their ability to be used on 3D surfaces with variable density sensing arrays. This paper studies contact localization on an artificial skin embedded with mutual capacitance tactile sensors, arranged non-uniformly in an unknown distribution along a semi-conical 3D geometry. A fully connected neural network is trained to localize the touching points on the embedded tactile sensors. The studied online model achieves a localization error of 5.7 ± 3.0 5.7\pm 3.0 mm. This research contributes a versatile tool and robust solution for contact localization that is ambiguous in shape and internal sensor distribution. 1 Introduction Figure 1 : Contact localization model takes in a sensor image from any configuration of artificial tactile skin and determines the location of touch through a feedforward neural network. Artificial skin sensors are a direct way to measure external contact and have a wide range of applications in healthcare [ 2 , 11 ] , prosthetics, and robotics [ 25 , 4 ] . As robots are increasingly deployed in close proximity to humans, the ability to localize where touch occurs allows actuating bodies to react appropriately to detected contact, interactable objects, and obstacles. Artificial skin that is capable of localizing contact to a greater number of points on its sensor array than the number of sensors present is currently limited to flat surfaces [ 6 , 11 , 24 , 26 ] . The localization methods used are typically constrained by assuming either a flat plate array or a uniform density of sensors. This paper presents a method of localizing contact on 3D artificial skin with non-uniform sensor distribution. We present a curved, variable-density artificial skin equipped with mutual capacitance sensors that detect touch on the surface of the artificial skin. A neural network infers a relationship between the sensor readings and the geometry of the tactile skin. We evaluate this model by comparing known touch locations to those returned by the model to characterize the accuracy and uncertainty. Our model robustly localizes the touching region on the 3D skin with variable sensor distribution. The code is available at https://github.com/HIRO-group/TactileContactLocalization 2 Related Work Artificial skin sensors have a range of sensing methods such as acoustics [ 23 , 21 ] , computer vision [ 3 , 8 , 9 , 10 , 1 ] , capacitive sensing [ 22 , 7 , 12 , 13 , 14 , 18 , 24 , 16 ] , electrical resistance tomography (ERT) [ 20 , 19 ] , and fiber Bragg grating (FBG) optical sensing [ 17 ] . Artificial skins composed of sensor arrays have increased scalability and conformity to non-flat geometries compared to computer vision tactile sensors. Our artificial skin uses a low-cost mutual capacitance sensing array method that is highly flexible to shape and size during fabrication. Explicit contact localization methods can achieve high accuracies with capacitive sensor arrays and are standard in touchscreens and artificial skins with known sensor distributions [ 22 , 19 ] . These explicit methods utilize the known positions of sensors to interpret locations of contact. Machine learning has been used in previous artificial skins to enhance contact localization accuracy with known sensor locations due to its ability to learn complex patterns [ 17 , 15 ] . Our method treats capacitive sensor readings as images and uses a neural network to bypass the embedded spatial distribution of sensors by directly learning touch localization patterns. 3 Method 3.1 Fabrication Figure 2 : (a) CAD model for the curved geometry. (b) General wiring scheme. Sensors are located at each intersection of transmitter (TX) and receiver (RX) wires. (c) Fabricated skin sensor array where the sensing circuits are embedded within a layer of silicone. A challenge in implementing an artificial sensing skin array on a 3D object is conforming the sensing array to a curved surface. This is addressed by fabricating a two-dimensional sheet embedded with sensors, and then overlaying the artificial skin onto a semi-conical surface. As shown in Fig. 2 , the curved surface is outfitted with a total of 16 wires, divided evenly into 8 transmit and 8 receive elements, forming a network of 64 intersections. These intersections serve as the loci for capacitance measurements, effectively creating 64 sensors distributed across the surface. A mutual capacitance board (Muca) attached to the wires measures the capacitance values at each intersection [ 22 ] . A layer of 6mm thick silicone rubber is then molded on top of the wiring to provide a dielectric layer between wire intersections to ensure effective capacitive sensing. A thin copper sheet affixed to the underside of the 3D-printed structure functions as the grounding plate, mitigating electromagnetic noise from beneath the artificial skin. The result is a flush, 14.2\times 16.4\times 8.1 cm semi-conical surface capable of contact localization through mutual capacitive sensing. 3.2 Calibration The calibration process includes collecting a small dataset in a given operational environment to establish a correlation between raw capacitance readings and specific touch locations in 3D space. Each pair of raw sensor readings corresponding to a touch location are referred to as a "point log". Data is gathered by touching the artificial skin with an index finger at known locations. Two methodologies are employed for collecting point logs: random sampling and even spacing. For random sampling, the location is chosen as a point on a randomly selected edge from the CAD model of the skin. For even spacing, the surface of the CAD model is discretized into evenly spaced points depending on the selected number of point logs. When the skin sensor array is touched, 50 samples of the capacitance readings for each sensor are stored and correlated to the prompted touch position. This process is repeated until the desired number of point logs has been processed. 3.3 Sensing Mutual capacitance coupling occurs at the intersection of transmitter and receiver nodes. The intensity of the coupled capacitance can be measured by sending a known signal to the transmitter wire and measuring the signal of the receiver wire. The coupled capacitance is affected by grounded conductive objects entering the electromagnetic field near the intersection, such as a human finger. Grounded conductive objects reduce the measured capacitance at the receiver electrodes which correlates to higher contact outputs for a particular sensor. The accuracy of the sensor measurements is determined by their noise levels through a signal-to-noise ratio. Fifty sensor measurements are taken for every point log and calibration measurement. Using these measurement samples, the average sensor values for each point log ( \bar{S} ), the average initial sensor values without contact ( S_{0} ), and the standard deviation from the average sensor value without contact ( \sigma_{0} ) are used to calculate the signal-to-noise ratio (SNR) of

Read the complete text