Abstract: Learning of e.g. a dexterous motor skill, which is common in medicaltraining, sports or playing a musical instrument, is time-costly andneeds supervision from a teacher, to avoid injury and to progress. Importantly, learning amotor skill requires active practice, and adjusting one's own individualembodiment to perform the target movement. We introduce a meta unit for co-constructing a computational scaffold for learning, that targets to integrate the teacher, the learner and a computational optimization approach, represented by a Dyna-Q $\epsilon$-greedy reinforcement learning agent. In this preliminary work, the meta unit is implemented as a Bernoulli distribution, which is parameterized by average reward gained by the simulated learner. Sampling from this distribution regulates the interventions of the simulated teacher. On the one hand, the resulting procedure regulates the interventions of the teacher during the learning process dynamically, if the learner is failing to progress. On the other hand, the learner employs a model of their talents, provided within the Dyna-Q agent, and schedules their practice with a trade-off between exploration and self-regulated practice. Under the assumption of difference in embodiments between the teacher and the learner, this approach aims to enable the learner to discover their own optimal learning policy, while being guided by the teacher on-demand, with the ultimate goal to improve beyond the capacity of the teacher.