The aim of the project is to develop a new methodology that will enable home robots to learn new goal-directed actions in a natural way and without requiring experts to help during learning. Thus no physical models of the task will be assumed to be available because only experts can provide such models.