Biomimetic Grasp Control of Robotic Hands Using Deep Learning

Abstract

Gripping force modulation based on pressure feedback is an essential element for intuitive and natural-like control of powered limb prostheses. This paper aims to mimic human hand-gripping control in robotic arms by processing dynamic pressure maps with state-of-the-art artificial intelligence algorithms. A pressure-sensing glove was built with integrated data acquisition to learn human grip behavior when holding various objects, and then transfer the observed control pattern to control a robotic arm. The pressure readings are processed using a recurrent convolutional neural network and were able to predict the biological gripping termination with an accuracy of 84.5% for a single type of object and 77% for mixed object types. The proposed control system has proven to be a viable approach for biomimetic handling control for an intelligent robotic arm with pressure feedback.