manantomar / agent-centric-representations Goto Github PK

Eq. 2 in the paper states that the observation x is merged with random gaussian noise epsilon, but at line 123 in the infogating file https://github.com/manantomar/agent-centric-representations/blob/main/infogating.py#L123, it seems that the observation is mixed with an all zeros mask.

Combined with the random noise (which is actually dropped after 1000 steps in the UNet), I believe this is instead equivalent to ig(x) + clip(N(ig(x), 0.5), ig(x) - 0.3, ig(x) + 0.3)x.

Can you help explain the discrepancy? Which version is correct, or am I misunderstanding the code?

Thanks!

Hi,

I am wondering what is the motivation behind using InfoNCE for InfoGating as opposed to vanilla multistep inverse dynamics predictor as in ACRO. Also, did you do comparison between InfoNCE and vanilla predictor?

Thanks for the time!

Hello, in the dmc.make function, you have distracting_wrapper (as well as fb_mtenv_dmc) which is not in your codebase. Where can I get this?

Thank you for the time