To accomplish a particular manipulation task, a robot needs a detailed description of how to execute it. However, it is not possible to specify all potential tasks of a manipulation robot beforehand. For example, robotic assistants operating in industrial contexts are frequently faced with changes in the production process. As a consequence, novel manipulation skills become relevant on a regular basis. For this reason, there is a need for solutions that enable normal users to quickly and intuitively teach new manipulation skills to a robot.

More concretely, we consider the problem of learning generalized descriptions of object manipulation tasks from human demonstrations. We employ dynamic Bayesian networks (DBN) as a compact representation where special nodes encode geometrical constraints between the relevant objects in the scene and the hand of the demonstrator. This formulation allows the robot to learn generalized task descriptions from multiple demonstrations so that it can reproduce them also under changed conditions. Furthermore, novel constraints can easily be added during the reproduction of a task. This is, for example, relevant to allow a robot to deal with obstacles. To reproduce a task, the robot searches for the action sequence that maximizes the data likelihood in the DBN. In experiments carried out in simulation as well as on a real manipulation robot, we show that our approach enables robots to efficiently learn novel manipulation skills from human demonstrations and to robustly reproduce them in different situations.