In parameter identification of robot manipulators, measurement noise from acceleration and velocity, together with unmodelled dynamics, are the main sources of errors in the least-squares estimation and can also considerably increase the time necessary for parameter convergence. Instead of using Instrumental Variables or filtered estimates of the acceleration, we discuss the application of a repetitive controller in order to minimize estimation bias caused by measurement noise. We prove that the controller converges to a neighbourhood of the desired trajectory, taking into account measurement and actuator noise. Thus, desired values of acceleration and velocity can be used in the regression vector as estimates of their real values. The method is applied in the parameter estimation of a direct-drive arm, experimental issues are discussed and the validity of the results are experimentally verified.