Learning Bimanual Coordinated Tasks From Human Demonstrations

In robot programming by demonstration dealing with high dimensional data that comes from human demonstrations is often subject to embedding prior knowledge of which variables should be retained and why. This paper proposes an approach for automatizing robot learning through the detection of causalities in the set of variables recorded during demonstration. This allows us to infer a notion of coherence and coordination between multiple systems that apparently work independently. We test the approach on a bimanual scooping task, consisting of multiple phases. We detect the coordination between the two arms, between the arms and the hands and between the fingers of each hand and observe how these coordination patterns change throughout the task.