grid sample positions

[zywang03]

Why are the results of my grid samples mostly concentrated in the four corners, instead of capturing the edge features of the objects as shown in the paper?

[Fediory]

If you obtained the visualization results on LIBERO and ALOHA, that is indeed the case. We found the same pattern as in the paper on RoboTwin and real machines, so the results of these two were also the most improved in the end. As for why LIBERO and ALOHA can produce effects even on the four corners, I tend to believe that GridS plays a role in information compression on these two datasets, utilizing the property of differentiable pruning to compress excessive repetitive information into fewer tokens, thereby reducing the importance of position.

[Ramesii]

I have uploaded the latest improved version, which effectively resolves the previous issue of sampling points occasionally clustering outward in complex scenarios.