Rapid learning of inverse robot kinematics based on connection assignment and topographical encoding (CATE)

Abstract

An adaptive neural structure for robot control based on homogeneous encoding in a topographical manner is developed. An intermediate representation (IRep) is adaptively generated using a novel learning scheme, CATE. The connection assignment rules of CATE keep the number of IRep-neurons as small as possible, while maintaining the desired mapping accuracy. This adaptive net (CATEnet) was successfully applied to embed the inverse kinematics of a redundant, planar robot arm (four-joint-machine) with only a few presentations of the learning set. The mapping solution incorporated local optimization of a cost function to account for a limited joint range and to avoid singularities.<>