拡張フィードバック系による逆運動学モデルのオフライン学習

概要

論文の詳細を見る
Recently the use of neural networks as the inverse-kinematics model of a robot arm has been proposed in learning control of the robot arm. The forward and inverse modeling, the feedback error leaning schema and the goal directed model inversion were proposed to extend the acquisition of the inverse model for the systems with many-to-one input-output correspondence. However, these methods can be used only for on-line learning. The learning of neural networks usually requires many iterations of robot arm movement and of its position measurement. In order to reduce the number of movements of the robot arm, the hybrid system which consists of a learning element and an extended feedback controller are proposed. The learning element approximates the inverse kinematics model of the robot arm. By using the extended feedback controller, the high precision solutions of the inverse kinematics problems are obtained so that these solutions can be used for the teaching signal of the learning element. After the acquisition of these solutions, the off-line learning of the learning element is conducted. The use of forward model of the robot arm is also proposed. The numerical simulations show the good performance of the proposed system.
日本ロボット学会,工業技術院機械技術研究所,東京大学工学部の論文
1995-07-15