Modular network control for robot manipulator

Abstract

Plant dynamics of highly nonlinear dynamical systems often vary over its parameter space. This generates a natural partitioning of the parameter space based upon the operating points of the plant. As a result, it may be wise to generate local control strategies at, the operating points rather than a single global strategy. This article describes a modular neural network architecture that generates a piece-wise continuous control strategy designed to fuse local strategies together to form a single strategy. The modularity is achieved through a gating network that controls the competition and cooperation of local experts. The gating network is a high order dynamical system network, while the local expert is a multi-layer feedforward network. The capability of this technique is demonstrated by building a neurocontroller for the two-link robot manipulator with two revolute joints.