Expected Performance of GR00T N1 in IsaacLab

[shalin-jain]

Hi, thank you for the great reference integration of GR00T with IsaacLab!

After experimenting with the base N1 model, I am struggling to see any measurable success for basic pick and place manipulation tasks (i.e. "pick up the blue cube"). From reading the published technical report, it seems like pick and place tasks would be fairly in-distribution for the model, especially given the described object generalization in the real world experiments.

I was curious if you had any insights on expected performance of GR00T N1 in IsaacLab without fine tuning? Do I need to use GR00T N1.5 to see better performance within IsaacLab, or is IsaacLab itself too out of distribution for the model?

Thanks!