Saving episode trajectories during training

How severe does this issue affect your experience of using Ray?

  • High: It blocks me to complete my task.

I would like to save for each episode the observation space and action space with timestep granularity.
The way I’m trying to achieve this is using the on_postprocess_trajectory callback. There, I have both postprocessed_batch and original_batches that provide access to a SampleBatch object. In that object, I have what seems to be actions, but these actions are transformed.
Is there an easy way to get the environment representation of the action?