Learning models for sythesis of human motion