[Question] What actions are expected for a previously terminated sub-env in a vector (next env.step() should return reset obs so the actions shouldn't matter)?

[sven1977]

Question

Quick question:
In a vector env with say 2 sub-envs (0 and 1).
When I do env.step([action0, action1]) and now sub-env 0 is terminated (returns the final observation), what action would the env expect next for sub-env 0? I would assume that it wouldn't matter, b/c the next env.step([action0, action1]) should return the reset obs for sub-env 0 anyways, correct? So can I pass in any random action?

[pseudo-rnd-thoughts]

Yes, correct the action is ignored so can be any valid action.
We generally assume that the terminated observation is still valid therefore can we treat it like any other observation in terms of action selection (requiring no change) but a random action is a valid alternative approach

[sven1977]

Ok, thanks for clarifying @pseudo-rnd-thoughts !