Extremely small scale for some objects

[patrickeala]

Hi @arsalan-mousavian,

I'm generating binary grasp quality labels (0 or 1) in simulation. However, my results are far from the flex labels. I am getting many failed grasps because some of the object scales provided in the grasp data are extremely small. In the image below the scale is 0.0472991972 resulting in a very small bottle. Can you clarify how exactly the scale is used in simulation? Thanks.

[clemense]

The scale refers to the isotropic scale of the mesh. Since they were sampled such that at least one dimension is smaller than the maximum gripper aperture, some of the meshes can be small.

[patrickeala]

Hi @clemense thanks for the quick reply. My problem is that are only some cases where the scale works appropriately. For example, in the image below, the scale is extremely small at 0.0096489713 but the bottle is scaled reasonably. It seems that the values work properly on some objects while others turn out to be unreasonably small. Could you clarify on how this scale is computed? What might be causing the inconsistencies with the effective size of the objects?

[clemense]

I don't see any inconsistency. The scale value is not an absolute indication of size, it's just a factor that every vertex in the mesh gets multiplied by. If there's a tiny bottle in the raw mesh data it will end up being smaller than a huge bottle even if their scale values indicate that one gets shrunken more than the other (as in your example).