Real-time generalization and integration of different movement primitives

In this paper we present a new methodology to learn and integrate different movement primitives in real-time. Our approach starts from a library of example trajectories for each primitive movement, which serves as a basis for the generation of a complete representation for the trained movement primitives by statistical generalization. To enable fast switching between different movement primitives, it is essential that on-line calculations needed to initialize and switch to a new movement primitive are done in real-time. We show that by converting the initial trajectory data into dynamic systems, we can switch to a new movement primitive within a real-time sensory feedback loop. Experimentally we also show that the accuracy of the generalized movements is sufficient to realize tasks such as feedforward grasping.