Why ACT and DP performed poorly in the Pi0 Paper VI C experiment, what are their failure patterns?

[tsrobcvai]

Hi, really impressive work! and thanks for generously open-sourcing the codebase and pretrained models.

I was wondering why ACT and DP performed poorly in the Pi0 Paper VI C experiment, as shown in the figure below. What are their failure patterns or examples? Understanding this might help me better grasp why Pi0 works better. I’d appreciate any comments!