A Machine Learning Approach to Contact Localization in Variable Density Three-Dimensional Tactile Artificial Skin (2412.00689v1)

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 $5.7 \pm 3.0$ mm. This research contributes a versatile tool and robust solution for contact localization that is ambiguous in shape and internal sensor distribution.

• The paper demonstrates a novel ML approach using a fully connected neural network to map sensor outputs to 3D touch coordinates on a semi-conical skin.
• The study reports an average localization error of 5.7 ± 3.0 mm, validating the method’s precision against complex, non-uniform sensor layouts.
• The approach’s adaptability to irregular sensor distributions opens new possibilities for tactile applications in robotics, prosthetics, and interactive systems.

Machine Learning Approach to Contact Localization in 3D Tactile Artificial Skin

The paper presents a novel approach for contact localization employing a machine learning model on three-dimensional (3D) tactile artificial skin. This research stands out for tackling the notable challenge of contact localization on 3D surfaces with non-uniform sensor distribution. It utilizes a fully connected neural network trained to localize touch points on a semi-conical surface embedded with mutual capacitance sensors. The reported localization error of $5.7 \pm 3.0$ mm demonstrates the efficacy of the method relative to artificial skins with reported known sensor distributions.

Introduction and Methodology

Central to the paper is the implementation of artificial skin that supports the precise localization of contact points over non-uniformly distributed sensors across a curved surface, a significant departure from traditionally flat geometries. A neural network-based model is applied, leveraging capacitive sensor arrays that transmute raw sensor readings into 3D touch coordinates without requiring predefined sensor locations.

The fabrication process involved embedding a mutual capacitance sensing array onto a deformable silicone layer, conforming to a semi-conical topology. The skin hosts 64 sensors achieved through intersections of transmitter and receiver electrodes, measuring variations in capacitance that arise when touched by a grounded conductive object. Measurements are processed using mutual capacitance boards, accentuating the sensor's flexibility and integration ease.

For model training, two data collection methods, random sampling, and even spacing, were vital. Here, the authors propose the utility of 50 repeated measurements for each touch point, labeled as "point logs," to fine-tune the model's accuracy in a supervised learning setting. These logs form the foundation for mapping the sensor outputs with explicit touch coordinates using a mean square error loss, ensuring that the network is efficiently learning the intricate mappings of touch locations.

Results

Empirical results showed that the model performance improved with the augmentation of point logs; however, significant gains plateaued beyond 80 logs. The sensor's average signal-to-noise ratio (SNR) correlations with point log size affirms a direct relationship between training data size and the robustness of contact localization. Notably, the achieved localization accuracy is on par with certain areas of human tactile acuity and comparable to other artificial sensing technologies like electrical resistance tomography (ERT) and Fiber Bragg Grating (FBG) optical sensing systems.

Implications and Future Directions

This research provides an adaptable contact localization framework that implies broader applicability for artificial skins needing integration on complex, variable surfaces. Its capability to accurately localize touch points without known sensor distribution opens pathways for deployable tactile skins in robotics, prosthetics, and interactive systems, particularly in heterogeneous or deforming conditions.

The authors recognize limitations, such as lack of visual markers during data collection, possibly affecting the contact accuracy, and the current focus restricted to singular touch interactions. Prospective development may involve enhancing the robustness of data collection through structured grid layouts and extending the methodology to handle multi-touch capabilities more efficiently. Such advancements could bolster applications in nuanced robot-human interactions, enabling more intuitive communication through tactile gestures.

In summary, this paper contributes substantially to expanding the utility of artificial tactile sensing in spatially complex and mechanically challenging frameworks, leveraging machine learning to mitigate conventional structural dependencies found in tactile skins. As such, it lays significant groundwork for future exploration and refinement in the domain of intelligent tactile systems.