Developing robust autonomous agents require complex execution architectures operating at several levels of abstraction in order to keep the acting problem tractable. While there is a growing body of work focused on learning models at the sensory-motor level, the same cannot be said for high-level models enabling deliberative functions. In this paper, we identify the possibilities that can be offered by a learning system integrating human input for learning hierarchical operational models and present a learning algorithm that could provide a missing component for such a system.