Extending extrapolation capabilities of probabilistic motion models learned from human demonstrations using shape-preserving virtual demonstrations.