Is critic-only input normalization intended for the released policy?

[gouruiyu]

Thanks for releasing this work : )
I am looking at how observation input normalization is (not) applied on the actor and would like to confirm the intent:

From the code:

• exp/manager/universal_token/all_modes/sonic_release.yaml has the actor running_mean_std: false while the critic gets true.
• sonic_release/last.pt also confirms that policy_state_dict has no running_mean_std.* buffers so the g1_dyn decoder consumes raw observations.

Is actor without input normalization a deliberate design choice? Did you get better performance or sim-to-real without normalization?

Thanks in advance!