Positional embedding for envs

[supersglzc]

Hi @Stevenhunter167,

In paper Section 4.3.2, it says it uses positional encoding in RND estimation. However, I noticed that for every environment, there is a wrapper that directly embeds the env's observations. Does it mean that the modified observation is used for all training procedures, not only the RND part? Please point me out if I misunderstand something :)

Additionally, the observation is actually modified a lot depending on the environment. I am wondering how you define the observation space for each environment?

Thanks!