Robust Manipulation Primitive Learning via Domain Contraction