Applications of learning method for dynamic control of robot manipulators

Some types of learning control were developed for a class of robot manipulators to realize a desired motion. In those control schemes, the actuator input to the robot is modified by the acceleration error signal or the velocity error signal so as to make the robot motion approach the desired one by repeating operation of the robot. However, since actual measured signals of those variables are apt to be contaminated by noise, it is desirable to correct the input by an actual position signal which is more accurately measured than the acceleration or velocity signal. Motivated by this observation, a new type of learning control is proposed, in which the input is modified by the position error signal. It is theoretically shown that by using this control method the robot motion converges to the desired one with repetition of maneuvering the robot. It is also shown that this learning method can be applied to not only position control but also force control of robot manipulators. Moreover, it is experimentally shown by using such a learning method that the real robot can be trained to polish a curved surface of an object and eventually carry the task to fulfillment.