[BUG] `InitTracker` key flexibility

[matteobettini]

Two questions about the following code

https://github.com/pytorch/rl/blob/0063741839a3e5e1a527947945494d54f91bc629/torchrl/envs/transforms/transforms.py#L6437-L6440

• is there a reason why we are preventing the init_key to be a nested key
• is the setting of self.reset_key done properly/ever used? it seems not

[vmoens]

For 1) IIRC yes there was, because we assume that the done is at the root and that the root batch-size is indicative of the env structure. What would it mean to have a nested "init"? Happy to see a refactoring if it is well defined.

For 2) I don't know, if the value isn't used anywhere we can remove it

[matteobettini]

The transform seems to be ok with multiple reset and init keys (the only thing it needs to do is just match an init key for any reset key)

It is just the init method that needs some refreshing I think.