There doesn’t necessarily need to be a repository of virtual worlds to train them on; I don’t know exactly how this works, it could be generating the data as something other than a 3D environment, then running it through algorithms to convert it to a 3D environment after the fact. We already have software that can convert static images to 3D models, so this isn’t even far fetched.
I agree with you that the progress will likely be slow, but just the fact that it’s being actively done in any capacity is encouraging - attention being paid to it is the first step in improving it!