Learning to Unscrew a Light Bulb from Demonstrations

In this paper we show a way of learning how to sequence predefined basic movements in order to reproduce a previously demonstrated skill. Connections between subsequently executed movements are learned and represented in a graph structure. Learning the switching behavior between connected movements is treated as classification problem. Due to the graph, the overall accuracy of the system can be improved by restricting the possible outcomes of the classification to connected movements. We show how the graph representation can be learned from the observations and evaluate Support Vector Machines as classifier. The approach is evaluated with an experiment in which a 7-DOF Barrett WAM robot learns to unscrew a light bulb.