NEURAL NETWORK-BASED LEARNING OF HUMAN SKILL IN HYDRAULIC ROBOTS

Abstract

This paper proposes an alternative direct teaching and learning method of human skillful tasks for a hydraulic robot manipulator, which realizes the compatibility between the high power performance of hydraulic actuators and the fl exibility of the direct teaching system and is useful to replace the human skill with an appropriate robot controller. The proposed skill teaching process consists of two phases; the direct teaching phase and the off-line learning phase with a neural network applied to learn the skill. In the direct teaching phase, a human operator manipulates an impedancecontrolled robot directly and executes a task to obtain the training data. In the off-line learning phase, a neural network is trained to learn a control strategy included in human tasks. Then, the operator executes the task again in cooperation with a trained neural controller and learning processes are repeated alternately. After several learning processes a neural network is re-trained and the task performance of the controller is improved. The proposed approach is experimentally applied for a control problem of an inverted pendulum by using a 2-link hydraulic manipulator to show the validity of the system. The robustness of neural network controllers trained in different ways is also investigated.