[RLlib] MARWIL OfflinePreLearner stuck at 0.00 row/s when reading Parquet dataset with Dict action space

What happened + What you expected to happen

I'm running an offline MARWIL experiment using RLlib and a custom environment with a Dict action space and a Dict observation space generated using gymnasium and a parquet offline dataset generated with Ray Data.
After fixing a previous TypeError related to MultiDiscrete.from_jsonable, MARWIL now starts training — but it gets stuck indefinitely with messages like:

Running Dataset: dataset_3_0. Active & requested resources: 1/8 CPU, 256.0MB/196.9MB object store
0.00 row [00:13, ? row/s]
(MapWorker(MapBatches(OfflinePreLearner)) pid=27308) array([...])

The dataset progress stays at 0.00 row/s, suggesting that MARWIL / OfflinePreLearner is not reading any rows from the dataset

Environment: custom with Dict action and observation space

Dataset format: Parquet written with Ray Data

Repo with full reproduction:
https://github.com/IdairaRodYanez/RLlib-experiments/

Forum discussion (full context):
https://discuss.ray.io/t/marwil-with-gymnasium-dict-as-action-space/23236

Versions / Dependencies

• Ray version: 2.48.0
• Python version: 3.11.4
• OS: Windows 11
• Environment: Gymnasium 1.0.0

Reproduction script

https://github.com/[IdairaRodYanez/RLlib-experiments](https://github.com/IdairaRodYanez/RLlib-experiments/)/

Issue Severity

High: It blocks me from completing my task.

Start by running the linked RLlib-experiments reproduction with Ray 2.48.0, Python 3.11.4, the Dict action and observation spaces, and the Parquet dataset. Trace the RLlib OfflinePreLearner dataset path while it reports 0.00 row/s; done means the MARWIL run reads rows and progresses instead of remaining stuck.