So where's the target now? That they should be able to improvize? I'm sure that can be added too with some work. Kurzweil famously wrote a program to produce music based on the music of classical composers as a child, I'm sure it could be done for dancing as well, especially since computer vision is making leaps, which would aid analysis. But then again, that would also just be "mechanical", as we would be able to explain the model.