In certain industrial applications there is a growing interest for robots with path deviations below 0.1 mm. Usually special purpose machines are used to reach this level of accuracy. Interest for using generic robots instead has two reasons: on one hand a robot is a standard equipment with significantly lower costs than a specially designed machine, on the other hand robots allow higher flexibility and more freedom in product design by offering 6 DOF. However, the precision required cannot be achieved with current standard robot systems. This article presents an approach to reach highest accuracy for robots - based on a learning control scheme. A learning phase is required to reach this accuracy and the process has to be repeatable. However, these two prerequisites often are not a problem for mass production or for parts of highest precision. Practical tests show the potential of reducing path deviations by a factor of up to 10 (deviations down to 10–50 µm). The presented functions are already available in KeMotion robot controllers from KEBA.