Toward Universal and Interpretable World Models for Open-ended Learning Agents