Generalization in Computer Vision

I was not super familiar with generalization or even the concept of robustness or adversarial examples. I ended up doing RL research anyway, but now I think that's what got me interested in generalization in RL. And one of the first things that I set out to do is like how can we actually quantify the generalization of the algorithms that we have right now? We took the DeepMind Control Suite, which was a common benchmark at the time and still pretty popular, which was on image based RL. It turns out they were all doing very, very poorly. But there wasn't really any solutions to that problem yet. So it doesn't seem like a scalable way to do this