Self-supervised bootstrapping of a movement primitive library from complex trajectories

Recent approaches have been advocated to learn a movement primitive library from demonstrations by using predefined motion features to identify and extract new movement primitives. In this paper, a new bootstrapping cycle is proposed which builds a suitable movement primitive library by (i) perceiving known primitives in complex trajectories, (ii) learn either new primitives or refines old ones, and (iii) consolidate the library by deleting unused primitives. The main contribution is that the movement primitives are in the center of the bootstrapping cycle and are learned self-supervised based on the notion of co-articulation. We evaluate the learning behavior of this bootstrapping cycle in a toy example and with complex handwriting trajectories. Finally, we demonstrate the bootstrapping cycle of movement primitives in a full skill learning example, where a human tutor teaches the humanoid robot iCub how to perform "fishing" motions with a fishing rod.