The problem is not that there details, the problem is constantly shifting ground. We can only rlpy on a harness to be sort of predictable but the models change all the time.