102 ANN による腱駆動ロボットの制御 : 関節剛性の学習

概要

論文の詳細を見る
A tendon-driven robotic mechanism can adjust its joint stiffness just same as a human arm using non-linear spring tensioning devices (NST). Therefore, it is expected that the robot will be able to work co-operatively with persons in a house, a hospital, an amusement park, and so on. However control of the mechanism is rather difficult compared to a usual direct-driven robot manipulator due to the large degrees of redundancy and the high non-linear characteristics of tendons. Hence, in this paper, we use an artificial neural network (ANN) controller that is composed of three local ANNs with radial basis functions (RBF) as an activation function. Since the tendon-driven robotic mechanism has large numbers of actuators so the dimension of the input vector for the network would become very large and the learning speed also would be slow. To cope with this problem, we use two local ANNs; one of them controls the joint torques and the other controls the bias forces. As described above, the tendon-driven robotic mechanism can adjust the joint stiffness semiactively using the tension bias forces. But the suitable stiffness for a desired task is not so clear. Hence the third local ANN learns it using GA. It is shown that the ANN can learn the joint stiffness values suitable for a positioning task and a ball-hitting task efficiently.
一般社団法人日本機械学会の論文
2000-09-01