Previous research on motor learning has examined the effects of various training schedules on learning performance. If more practice improves learning, the practice must be tailored to the learner's skill level. One approach is to design interactive learning support systems that adapt to the learner. However, when it comes to the design of such systems, there seems to be a lack of theoretical results regarding the mechanisms of adaptation. Here, we investigate the effects of real-time personalization of practice through a Multi-Armed Bandit (MAB) algorithm. We conducted a controlled laboratory study with a simple motor task involving pointing movements where the wrist drives a cursor in a channel. We show that the MAB algorithm outperforms standard algorithms by effectively reducing movement variability and that this adaptation homogenises learners' skills. These results have theoretical implications that allow us to inform the design of interactive learning support systems.CCS Concepts: • Human-centered computing → Human computer interaction (HCI); • Computing methodologies → Reinforcement learning; • Applied computing → Computer-assisted instruction.