Robot task teaching has during the past years received significant attention and it has been recognized that more natural teaching methods are necessary so to allow ordinary users to teach robots new tasks by simply demonstrating them. From the viewpoint of task learning in humans it is known that such a strategy where a teacher's demonstration is used as a starting point of learning significantly speeds up the process and reduces the amount of trial-and-error steps. In robotics, such an approach to learning has been considered in frameworks of Learning by Imitation or Programming by Demonstration (PbD).

An important issue to consider is that the initial task setting will change between the demonstration and execution time. A robot that has to set-up a dinner table may have to plan the order of handling plates, cutlery and glasses in a different way that previously demonstrated by a human teacher. Hence, it is not sufficient to just replicate the human movements but the robot i)~must have the ability to recognize what parts of the whole task can be segmented and considered as subtasks so to ii)~perform online planning for task execution given the current state of the environment. The important problem here is how to instruct or teach the robot the essential order of the subtasks for which the execution order may or may not be crucial. As an example, the main dish plate should always be under the appetizer or a soup plate and the order in which these are placed on the table is important. One way of addressing this problem is to demonstrate a task to the robot multiple times and let the robot learn which order of the subtasks is essential.

Movies:

Related Publications

Back to research