Positioning a robot arm: an adaptive neural approach

Abstract

The problem of positioning a robot arm in three-dimensional space has been studied for a long time. However, most solutions developed until now, despite the fact of providing great reliability and accuracy, lack the necessary flexibility to permit the arm to move in an open environment. Most problems to be solved by a robot arm in uncontrolled environments are mostly like the ones we solve on a daily basis, such as pick and place tasks. Those tasks don't necessarily need the accuracy provided by the known methods to position an arm, but they do need the degree of adaptivity and flexibility that humans possess. We present a neural adaptive approach to solve the problem of positioning a robot arm in the space. This method works by incorporating the state of the system into the network. The network input is the current state of the system (the current arm position and orientation) and the outputs are the changes in the state variables (the joint values) in order to approximate the current state to the desired one. This is a closed loop neural control scheme and it is done in real time without needing any previous training phase.