Port BabyAI GoToLocal Expert Trajectories To Minari

[helenlu66]

As mentioned in Source MiniGrid/BabyAI Dataset#14, once a dataset is sourced, it needs to be converted to Minari. GoToLocal expert trajectories have already been uploaded to google drive. Now it needs to be converted to Minari.

outcome:

• a repository with script and a readme on how to run the script to convert the BabyAI GoToLocal expert trajectories in .pkl format to Minari.
• uploaded HDF5 of GoToLocal expert trajectories to google drive

[helenlu66]

currently exploring the structure of the expert trajectories here

[helenlu66]

When trying to de-serialize the state observations, I get a pickle.UnpicklingError: invalid load key, '\x02'. Figuring out how to fix this and de-serialize the provided expert demonstrations on the GoToLocal task is going to require more effort than just generating expert demonstrations on the task using the BabyAI bot. So I'm setting up the steps necessary to generate expert demonstrations using the BabyAI bot for not only GoToLocal but also other levels. https://github.com/helenlu66/RLMinariDatasets/blob/master/babyai_bot_expert_data_generation.py

Keeping this issue open for now. When the alternative is done, I'll close this issue.

[helenlu66]

Currently working on it here: https://github.com/helenlu66/RLMinariDatasets/blob/master/babyai_pkl_to_minari.py.

There seems to be a major issue: when creating the data from buffer using https://minari.farama.org/api/minari_functions/#minari.create_dataset_from_buffers, Minari will try to serialize the observation space into a string. Currently for BabyAIMissionSpace, the responsible serialization and deserialization methods are not implemented

[harshsikka]

I think we should keep these issues handy as they may be useful to @jsjung00 and @snat-s going forward. We can revisit whether these issues still persist with the current setup once they have evaluate this data.