That is true (and I am well aware of this, as I do research on robotic grasping https://sites.google.com/view/task-oriented-grasp ). But, this would very hard to generalize to more complex manipulation where the object is not already in-hand etc. ; it's yet another piece of the puzzle, but it by itself does not get us that much closer to general purpose training of robust complex manipulation.

Well, i can see that you are a expert in this field, but i still have to say the method they demonstrate, domain randomization is a big deal because that can then be applied to anything else like picking up stuff. Domain randomization is a year old thing now, but it's a big deal - if you can build a "crappy" simulator for it, you will be able to do anything. I feel like domain randomization isnt being promoted enough in the media

sure, but they demonstrate it for a very simple use case of a single object with no background etc. ; domain randomization has yet to be demonstrated to work on more complex environments/tasks