Using Data Augmentation as an Alternative to Domain Randomization?

In theory, you could augment this with many models. They would all feed back into the encoder. Maybe it makes it too complicated and now the encoder gets confused. We were never able to get it working. And I think that's one thing is also just the complexity of actually, in terms of wall time, actually doing this. It becomes impractical. If you want to run something in a real robot, there's a certain frequency that you need to interact with. That was a constraint. But then what we tried to do is at least try to probe the model, I figure out which out of these five self-supervised objectives would be more effective for this task that we