Learning sequential constraints of tasks from user demonstrations

Learning from human demonstration is likely to be one of the key features for service robots in household domains if they are to be accepted by humans. To be of most benefit possible to its user, the robot should go beyond simply imitating a user´s demonstration but try to build task knowledge that is as general and flexible as possible. One way to achieve this is equipping the robot with the ability to reason about the task knowledge he has already acquired in order to refine, generalize and complete it. A system to record and interpret manipulation task demonstrations is presented in this paper. As a representation for the sequential constraints a valid task execution must obey, task precedence graphs (TPG´s) are introduced. Means of reasoning on a task´s underlying TPG are proposed and evaluated within two tasks from the household domain