[Bug Report] step and info structures expected to be identical (when `record_infos=True`)

[Kallinteris-Andreas]

Hey, I am trying to make dataset of Ant-v5 and i am getting this error:

$ py create_dataset.py  # expert-v0
Traceback (most recent call last):
  File "/home/master-andreas/gym/rl/project/create_dataset.py", line 56, in <module>
    obs, rew, terminated, truncated, info = collector_env.step(action)
                                            ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/master-andreas/gym/rl/project/temp_env/lib/python3.11/site-packages/minari/data_collector/data_collector.py", line 222, in step
    self._add_step_data(self._buffer[-1], step_data)
  File "/home/master-andreas/gym/rl/project/temp_env/lib/python3.11/site-packages/minari/data_collector/data_collector.py", line 162, in _add_step_data
    raise ValueError(
ValueError: Info structure inconsistent with info structure returned by original reset.

for reference: self._reference_info is set using info from reset https://github.com/Farama-Foundation/Minari/blob/9803a700bdba1e228e335de5422511edbfb4c74e/minari/data_collector/data_collector.py#L261

which is then compared to the info from reset https://github.com/Farama-Foundation/Minari/blob/9803a700bdba1e228e335de5422511edbfb4c74e/minari/data_collector/data_collector.py#L147

Also in general, it is not in Env's specification that info s need to have the same structure each time https://gymnasium.farama.org/main/api/env/#gymnasium.Env.step

@alexdavey @younik

[younik]

Hello @Kallinteris-Andreas,

The problem is that Ant-v5 returns some info at reset, but not during step. If we collect and store inconsistent infos in HDF5 file, we may silently misplace infos during dataset reading (for example assigning the second info to the first step, instead of the second episode reset).

This is why we require infos to have the same structure always. With the current code, you can solve it by defining a StepDataCallback or simply setting record_infos=False (which at this point should be the default to avoid failing on simple envs like Gymnasium-Robotics).

We may want to pad infos automatically in the future, but I don't think it is high priority now