Given a desired output for a class of dynamical systems, certain kinds of iterative learning control are shown to be effective in the sense that the output converges to the desired one with repeating operations. One of those methods can be applied to a robot manipulator with high nonlinearities even though its physical parameters are not known. In this case, the velocity vector of the robot motion is regarded as the output and the input at the present operation is modified by the derivative signal of the error at the previous operation, which is the difference between the output and the desired one. Through those iterative operations, the desired motion given to the robot is obtained if some conditions for the modification of the input are satisfied. However, when the output is contaminated by noise, the derivative error is extremely different from the real derivative error. In such a case, it is difficult to realize the desired motion with high accuracy since the modified input itself becomes noisy. To overcome this difficulty, we propose an alternative learning control method for robot manipulators, which uses the error directly. To show the effectiveness of this method, fi rstly the robot dynamics is linearlized around the desired motion. Next, it is proved that by this method the robot motion converges to the prescribed motion trajectory with repeating operations. Moreover, from a practical point of view, it is discussed how the parameters concerned with this method should be determined. Finally, this learning control method is practically applied to a robot manipulator with three degrees of freedom and its effectiveness is shown experimentally.