about robocasa action

[baiyu858]

@xiaogangjia
Great job🎉. I'd like to ask whether the action in the Robocasa dataset you used has 12 dimensions or 7 dimensions (for the robotic arm only).
Thank you very much🤔.

[xiaogangjia]

Thanks for your interest. We use 7 dimensions and fix the last 5 as [0, 0, 0, 0, -1].

[baiyu858]

Thanks for your interest. We use 7 dimensions and fix the last 5 as [0, 0, 0, 0, -1].

Thank you very much. I have another question: For Robocasa, do you train one model (checkpoint) for each task or train one unified model (checkpoint) for all tasks?

[xiaogangjia]

In this xil paper, we train one model for all 5 tasks. You can also refer to the paper https://arxiv.org/abs/2502.12320, where we evaluate all tasks there. In that paper, we split the tasks into different groups and train one model for each group.

I am also preparing the rlds data format for robocasa and incorporating rlds dataloader into our codebase. I am currently very busy with a ddl. I will try to update the code after that.

[baiyu858]

Thank you very much for your reply!

I also want to ask how many resources were spent and what the training time was when you trained a model for the five Robocasa tasks.

[xiaogangjia]

I think I used 1 A100 and trained for 200 epochs, which took roughly 4.5 hours.

[baiyu858]

I think I used 1 A100 and trained for 200 epochs, which took roughly 4.5 hours.

Thank you very much for your reply. May I ask whether the RoboCasa dataset you used is raw data or has it undergone keyframe extraction? If keyframe extraction was applied, could you please provide the program used for keyframe extraction?

[xiaogangjia]

I just used the raw data without keyframe extraction. However, since the raw data does not provide depth, I generate the point cloud based on the raw data. Here is the code I used for generating the point cloud, you can also get any more information you want:
https://github.com/ALRhub/custom_robocasa/blob/main/custom_robocasa/robocasa/scripts/dataset_states_to_obs.py

[baiyu858]

I just used the raw data without keyframe extraction. However, since the raw data does not provide depth, I generate the point cloud based on the raw data. Here is the code I used for generating the point cloud, you can also get any more information you want: https://github.com/ALRhub/custom_robocasa/blob/main/custom_robocasa/robocasa/scripts/dataset_states_to_obs.py

Thank you very much. I noticed that the simulation verification in RoboCasa is very slow. Are there any faster parallel verification solutions available?

[xiaogangjia]

Yes, I also found that the simulation is quite slow. I am not sure if it's because of generating the point cloud or my code problem. I have the code for multiprocessing, but I don't know how much improvement it will make. You can check it here if you are interested
https://drive.google.com/file/d/1WwWQ-IYpzuK7DZAvdHKA-YU2SOaD427I/view?usp=sharing

[baiyu858]

Thank you very much for your detailed explanation! I would like to know if you are using the official dataset released by RoboCasa, or a dataset that you have collected yourself? I am looking forward to your reply with great anticipation! Thank you very much.

[xiaogangjia]

Hi, I am using the official dataset released by RoboCasa