You would probably need to use sensing to improve stabilization there. The output of the approach here is just a position (I'm assuming) trajectory. They manage to do that because they have a good enough model of the beams and masses. I imagine it's harder to get a comparable model of the disturbances in a drone.
There's a pretty cool project using ML to help translate control from simulators to the real world. Video if it here, paper in the description: https://www.youtube.com/watch?v=aTDkYFZFWug
This seems like it could also help deliver something similar to what Disney has done (or possibly improve it)
