Currently the OpenAI gym wrappers lack support for dynamically registering wrappers from third party packages. We should look into how this can be done, maybe similarly to how Environments are registered?

Are wrappers only allowed to be generic enough to be applicable to any kind of observation? What about the typical image based observation (Box(M, N, 3) or Box(M, N)) which can be downscaled before processed by the agent?

What about Tuple observations? How can we apply a wrapper which is only applicable to one specific type of observation, say Box(M, N, 3) (an image)?