The cerebellum plays a crucial role in motor learning and it acts as a predictive controller. A biological inspired cerebellar model with distributed plasticity has been embedded into a real-time controller of a neurorobot. A cerebellum-driven task has been designed: the Vestibulo-Ocular Reflex (VOR), which produces eye movements stabilizing images on the retina during head movement. The cerebellar controller drives eye compensation, by providing joint torque based on network output activity. We compared a cerebellar controller with only the cortical plasticity and a cerebellar controller with also the plasticity mechanisms at deep nuclei, in VOR multiple sessions. The results were interpreted using a two state multi-rate model integrating two learning processes with different sensitivities to error and different retention strengths. The cerebellar model showed effective learning along task repetitions, allowing a fine timing and gain adaptation based on the head stimulus. The multisite plasticity proved superior to single-site plasticity in generating human-like VOR during acquisition, extinction and consolidation.