Question on Training Data

[bingxin-xu]

Hi Author,

Thanks for the great work on building the RoboMemArena benchmark. We have several questions on the training data.

1. We find some subtasks do not have order in the name. How to decide its ordering in this case?
   For example, under multi-occlusion 4_drawer_butter_dataset, "put_butter_into_drawer_have_object_seed100_task4.hdf5" does not have a subtask order in its naming. Which order does it belong to?

2. We find subtask orders is not unique for different subtask. Could you help to update the subtask ordering?
   For example, under multi-occlusion 4_drawer_butter_dataset, "close_top_drawer_final_4_seed100_task4.hdf5" and "open_bottom_drawer_4_seed100_task4.hdf5" share the same subtask order.

3. Do you train on all provided data or any filterings on failed demonstrations before training? If later case, would you mind sharing the data filtering script?
   We find some failed subtasks in the demonstration, for example, under multi-occlusion 5_butter_middle_drawer_dataset, "open_bottom_drawer_4_seed101_task5" doesn't successfully open the drawer.

4. We find some mismatches in the subtask description and robot demonstration. Would you have any data cleaning scripts for this scenario?
   For example, under multi-occlusion 5_butter_middle_drawer_dataset, "open_middle_drawer_again_5_seed101_task5" supposes to open the middle drawer, but it actually opens the bottom drawer.

Thank you so much for your effort! Look forward to hearing back!

[wpyaoxueros]

1. We mistakenly included the data of the long task and have deleted and modified it. The corrected data has been uploaded. Please pull the corrected data.
2. In this task, we mistakenly wrote the sequence numbers of two sub-tasks as one. We have deleted and modified it. The corrected data has been uploaded. Please pull the corrected data.
3. The reasonable fine-tuning method is to train on all successful data. However, due to the problem of insufficiently strict screening, a small amount of erroneous data is mixed in the uploaded data set. The modified data has been uploaded. Please pull the corrected data.
4. After regenerating the data according to the rules in 2, the problem has been solved. The corrected data has been uploaded. Please pull the corrected data.

[Curry-ZHR]

Our error has caused some trouble regarding the issues of task4 and task5. However, a new version of the data will be updated soon. Please wait. Thank you.