Option for observations as RGB array

[planetceres]

It would allow for a greater range of use cases if there was an option that could be passed to train using RGB arrays instead of only the image encodings. The readme describes a way to do this with get_obs_render, but it would be nice if it could be included as an argument.

Would you consider adding something like this?

planetceres@9c2ae3c

[maximecb]

I think that the cleanest way to do this would be to create a separate wrapper that just produces an RGB image, and maybe also drops the text string if you don't need it. I would create a separate wrapper rather than setting some boolean variable on the environment. If you create that, I would merge it.

[planetceres]

Optionally dropping the text string for baseline comparisons was another priority for me, so I'm glad you mentioned that. I had started with the wrapper approach and agree it makes more sense, but I was trying to avoid creating multiple wrappers for each variant of image type along with optional text. This may not be a concern though if the observation wrappers are more meant as a temporary solution to deal with dict observations.

[planetceres]

This makes sense to me based on the current wrapper syntax, although could use some refactoring later at some point:

planetceres@b16c3f6

[maximecb]

I don't think the wrappers are just a temporary solution. I think different people will want different inputs for their use cases. I think it's just cleaner to separate the environment from wrappers that format the output however you want it.

I would just create an RGBImageWrapper, which doesn't flatten the image and doesn't include the text (no boolean argument).