Here, onpolicy trainer relies on the value rew - which is the mean reward from the collector:

best_reward, best_reward_std = test_result["rew"], test_result["rew_std"]

but this value is only computed by the logger here:

collect_result["rew"] = collect_result["rews"].mean()

So it seems to me that a logger, if it is switched to another loger, which does not compute the mean - will break the whole thing.