This paper presents an Iterative Learning Control algorithm for direct-drive robots. The learning algorithm assumes linear dynamics, which is created using a nonlinear model-based compensator. The convergence criterion of the learning controller is derived in the frequency domain. Rules for designing the filters, used in the update law, are explained. The effectiveness of the algorithm is demonstrated in experiments on a spatial direct-drive robot. The root-mean-square values of the tracking errors in a demanding writing task are over 10 times smaller after just eight iterations of the learning algorithm, compared with the errors before learning.