Do You Want to Do Domain Randomization in Robotics?

The idea of using a wild model to almost be like a learned simulator offline to online is becoming a little bit more accepted in the community. But no one's really figured out the fact that you could potentially do domain minimization in the learned simulator, which is also done into sample robotics all the time. Generated models seem like a promising approach. You can imagine instead of having a GAN that generates images, you know how something that generates entire transition functions that are plausible, but not exactly the mode of your offline aid set.