Success Rate Is Always Either 0 or 1 During MetaWorld Training

[SCI-I]

Hi, thank you for releasing the code.

I ran the FAME code on the MetaWorld benchmark and noticed that the success rate displayed during training is always either 0 or 1, rather than a continuous value between 0 and 1.

Could you please clarify whether this is the expected behavior? Is the logged value the success indicator of a single episode instead of the average success rate over multiple episodes or environments?

If this is not expected, what might cause this issue, and which part of the evaluation or logging code should I check?

[initial-h]

Hi,

This is a common behavior in the MetaWorld benchmark. We evaluated 10 or 30 episodes at each evaluation step, I don't remember very clearly but definitely not 1 episode. The reason is that the tasks in MetaWorld don't have much randomness: when the policy begins to master the task, it succeeds almost all the time, which tends to give these 0/1 values. If you want to check some smoother curves, you could test on some medium-hard tasks, which may have some intermediate values, or check the reward curve, which is smoother.