Question about Target Drawer in Task 4 and Task 5

[quark404]

Hi, thank you for releasing RoboMemArena.
Using Task 4 and Task 5 as examples, both the training data and the BDDL task definitions indicate that the target drawer is fixed rather than randomized. In Task 4, the non-empty drawer is always the top drawer; in Task 5, the empty target drawer is always the middle drawer.
For a benchmark intended to evaluate memory-based drawer selection, the target drawer should not be fixed across all episodes.

[Songwxuan]

Thank you for your question. This is a very valuable point. This omission effectively removes the need for memory in these two tasks. We will start fixing this issue immediately and expect to complete the fix within one week. At that time, I will update the repo and explain the revised setting to you.