Model Learning for Probabilistic Movement Primitives