Box VS Convex is not accurate as Capsule VS convex in scene->overlap or trigger

[yananZZZZ]

Hi,
What's I am doing is I create a rather small cylinder like convex shape geometry and make it as overlap query shape. The radius is 0.45mm and 12mm for height.
As i use a box collision shape (1mm for extend) to do the overlap test, the overlap is not report anything when the box already enter the convex shape. I did the same test only change the collision shape into capsule(1mm for diameter). This time the overlap can get the result as soon as the capsule enter the volume.
So i look into the code and found out for Box vs Convex a "RelativeConvex" is used for GJK and for capsule is "LocalConvex". I just wonder is this cause the result different? I also found out for small size convex, actually the GJK algorithm is easy to become degenerated. I saw a lot hardcode number inside GJK, is there anything i can fine tune?

thank you

[vreutskyy]

Hello. It would be great if you could make a compilable code snippet reproducing this issue. Thanks

[vreutskyy]

Sorry, I can't download your archive at
https://github.com/NVIDIA-Omniverse/PhysX/files/12287001/SmallConvexShapeCannotbeTriggered.zip

Could you check the link?

[yananZZZZ]

snippettriggers.zip
How about this one?

[vreutskyy]

This one's okay, thanks 👍

[vreutskyy]

Hm. It's strangely similar to the neighboring issue here: #173

Could you try the same fix I proposed them?

Change this line:

PhysX/physx/source/geomutils/src/gjk/GuGJKSimplex.h

Line 311 in e8c8deb

if(FAllGrtrOrEq(eps, area))

from
if(FAllGrtrOrEq(eps, area))

to
if(FAllGrtrOrEq(FMul(eps, eps), area))

Thanks

[yananZZZZ]

Hi,
After modify the code as suggested, the accuracy of small convex vs box is correct. Thank you for the support.