Neuro-Fuzzy Grasp Control for a Teleoperated Five Finger Anthropomorphic Robotic Hand

Abstract

Semi-autonomous manipulation can expand robot assistants and intelligent prosthesis use to efficiently a more comprehensive range of manipulation tasks. Nevertheless, several aspects of teleoperation, such as misaligned joint control, still restrict robotic manipulation in unstructured environments. This paper presents a novel neuro-fuzzy intelligent control system for human-like grasp behaviours. First, the remote operation of the human-like robotic hand used data collected from a sensing glove. Then, we used grasp synergies to allow each finger to move independently. The data collected enables the evaluation of the proposed Adaptive Neuro-Fuzzy Interface System (ANFIS) to grasp objects with high levels of accuracy. Several aspects were considered to increase robustness: no post-processing of collected data, variations of objects in the tests, and the use of multiple grasp attempts for trained data. The results show evidence that these aspects provided a robust model for remapping a human hand to the robotic system behaviour. The experimental study presented in this paper addresses the efficiency of the ANFIS system to grasp with a high level of accuracy. Merging grasp synergies with the ANFIS machine learning model, developments have been made to further define dexterous machine learning for teleoperation and how they can be used for human-like grasping operations. The results show, for an instance, a normalized root mean square error (NRMSE) of 2.13% of the index finger, which is strong evidence that these aspects provided a robust model for remapping a human hand to the robotic system behaviour.