Dict Space Crash on to_jsonable

[rmwtoyon]

The to_jsonable function fails for Dict space types. Here is a simple example...

space = spaces.Dict({
  'test': spaces.Box(low=0, high=1, shape=(2,), dtype=np.float32)
})
sample = space.sample()
jsonable = space.to_jsonable(sample)

Here is the call stack...

  ...
  File "C:\Users\rwilkerson\AppData\Local\Programs\Python\Python38\lib\site-packages\gym\spaces\dict.py", line 71, in to_jsonable
    return {key: space.to_jsonable([sample[key] for sample in sample_n]) \
  File "C:\Users\rwilkerson\AppData\Local\Programs\Python\Python38\lib\site-packages\gym\spaces\dict.py", line 71, in <dictcomp>
    return {key: space.to_jsonable([sample[key] for sample in sample_n]) \
  File "C:\Users\rwilkerson\AppData\Local\Programs\Python\Python38\lib\site-packages\gym\spaces\dict.py", line 71, in <listcomp>
    return {key: space.to_jsonable([sample[key] for sample in sample_n]) \
TypeError: string indices must be integers

I can fix the issue by making the following change to `gym/spaces/dict.py'...

    def to_jsonable(self, sample_n):
        # serialize as dict-repr of vectors
        return {key: space.to_jsonable([sample[key] for sample in sample_n]) \
                for key, space in self.spaces.items()}

...to...

    def to_jsonable(self, sample_n):
        # serialize as dict-repr of vectors
        return {key: space.to_jsonable(sample_n[key]) for key, space in self.spaces.items()}

[rmwtoyon]

I'm using version 0.15.6 by the way. Doesn't look like this has been fixed in newer releases.

[rmwtoyon]

What is the process for fixing issues like this? Would OpenAI process a merge request to master with the proposed change above if I created one? Happy to help fix this. I'm using is_jsonable in a shared project and am trying to avoid having team members doing custom installs of gym.

Thanks in advance.

[rmwtoyon]

Pull request is here: #1849

[rmwtoyon]

Addressing CI issues now.

[rmwtoyon]

After looking at the failing test code for the pull request I'm realizing that the to_jsonable function is written assuming the caller wraps the input space object in a list. The crash occurs only when the object is not wrapped in a list.

I'm not sure I understand the logic in this design but my issue is mute if I follow that example. Closed the pull request and this issue.