Bidirectional feature map for robotic arm control

Abstract

Abstract When a neural network is used as an arm position controller, it is hoped that the network realizes at least two functions: (1) implementation of a desired input‐output relationship; (2) ensuring the ability of generalization. However, most well‐known networks, such as Back Propagation, and Feature Map clones do not satisfy these requirements simultaneously. Based on a new architecture and the learning rules, this paper proposes a Bidirectional Feature Map model which has these functions. Then the network for an arm controller with newly proposed constraints is applied to avoid ill‐posed situations inherent to articulation mechanisms with excessive freedom.