Do You Need to Model Everything or Do You Just Model Them?

In dreamer, at least, you're using those links for reconstruction as well to part of your loss function. And so there's a reason to put a lot of extra junk in there that you really didn't need just for the rewards. I would be very interested in seeing people playing around more with these hybrid algorithms and trying to build something in between. Have you experimented with anything like this? Not yet, but I think it's something that is very interesting. How can you generalize zero-shot? How can you do a little bit of fine-tuning to work well? But I could also see policy reuse in much broader context, that if you're training a policy to