Is RL Really Hard to Measure?

I love tingering with all these more details. How do you actually measure consistently which algorithm is better in a setting? It seems that even within a single lab environment, if you train two models, it's very hard sometimes to say which model is better. I think it's a huge problem in robotics, at least like robot learning research, that it's so hard to reproduce other papers. And I don't think it's because people are careful about describing all the details; I think it’s just sohard to reproduce,. Yeah, it's so sensitive to very slight distributional shifts. Exactly, exactly. Well, it doesn't generalize. Oh man, what are