Machine-Learning-Enhanced Low-Cost Bionic Fingertip Sensing

Abstract

This paper presents a low-cost bionic fingertip sensing system that combines hardware design and machinelearning- based signal fusion to improve tactile perception in prosthetic hands. The system employed a compact triangular array of three force-sensitive resistors, enabling estimation of both force magnitude and contact location. To fuse signals from the sensor array, a second-order polynomial Lasso model was developed. This compensated for the inherent nonlinearity and spatial ambiguity of force-sensitive resistors by capturing both linear and non-linear interdependencies between sensor outputs. Experimental evaluation used controlled loading at seven predefined contact locations across the sensor surface, with forces ranging from 0 to 9.8 N. The mean absolute error for force magnitude prediction was below 5%, an improvement of more than 20% over the simple force summation baseline. Contact localisation prediction accuracy was 97.1%. The results demonstrate that machine learning can compensate for hardware limitations and enable clinically relevant sensing performance. At an estimated cost of approximately C18 per fingertip, this system provides a scalable and affordable solution for high-resolution tactile sensing.