the description of realworld data shown in the paper and the data downloaded from website seems not match

[happyflying-web]

Hi~, the description of realworld data shown in the paper and the data downloaded from https://xskill.cs.columbia.edu/data/real_kitchen_data.zip seems not match , could you help me, thank you very much!

[happyflying-web]

For example, the inference task can not be found in the dataset downloaded from https://github.com/real-stanford/xskill/issues/url

[mengdaxu]

Hi!

Thanks for your comments. Could you please clarify which aspects do not align? The training data in your screenshot appears to match the description in the paper.

Regarding the inference task, we haven't uploaded the human demonstrations for the 4-subtask combinations, but you can perform inference using the 3-subtask combinations, such as 'oven draw cloth' and 'draw cloth oven'. Meanwhile, I will try to locate and upload the 4-subtask human demonstrations. It's worth mentioning that we did not capture demonstrations involving four subtasks with the robot, as our evaluation in the paper's experiment section did not include tasks with the "same embodiment" involving four subtasks.

[happyflying-web]

I get it through your detailed reply！I am a beginner and very interested in your research. I have been studying your code carefully recently. Could you please tell me which part of the code the following two experimental parts correspond to？ Thank you！

[mengdaxu]

For dataset and instructions, please see simulation dataset and training/simulation section in readme. The repo contains both pretraining and evaluation on simulation environment. Regarding to realworld, the repo provides code to do pertaining. The realworld evaluation is largely based on diffusion policy repo: https://github.com/real-stanford/diffusion_policy.