PxGjkQuery::proximityInfo() produces inacurate results

[yiftahw-fsr]

Issue

Hi!,
I'm using the PhysX SDK (tag 107.3-physx-5.6.1) to perform collision calculations using the GJK API with models loaded from obj files.
during testing, a call to PxGjkQuery::proximityInfo() resulted in A, B points that are inside the actors, which causes a false positive scenario where I assume there is more distance to travel safely, but the objects are actually closer than the query reports.
self contained issue: https://github.com/yiftahw-fsr/physx-gjk-issue
the issue (and the fix) described here are already commited in the tag I'm using, and the issue still exists.
Any help or suggestion would be appreciated.

Description:

the repo contains an example with 2 actors (left and right)
left actor has 2 transformations, and right actor has 1 transformation.

• left with first transformation, call GJK, we get wrong results.
  • pictured below in blender in red, we can see the points are not on the surface of the objects.
• left's second transformation is shifted 0.1 mm in the X axis in comparison to first, but the answer differs in 0.29 mm.
  • pictured below in green, the points are on the surface of objects.

Numbers for comparison (numbers are in mm scale):

Left actor's first transformation (and right actor with its transformation)
PxGjkQuery::proximityInfo() yielded points:
a : (-104.0039980, 49.8885190, 395.7634890) b : (-103.6410830, 52.4601140, 393.8701780)
separation: 3.2139437897964096

Manual calculation:
a: (-103.9832019, 50.0363439, 395.6541406) b: (-103.5821103, 52.3519518, 393.9241805)
separation: 2.918163179333445

---------------------------------------------------------------------------------------------

Left actor's second transformation yields:
a: (-103.9186096, 49.9972229, 395.6325684), b: (-103.5878906, 52.3405457, 393.9073486)
separation: 2.928638

Manual calculation was done by checking what is the point of intersection between the vertices reported by GJK and the polygons/edges of the target polygon.

I loaded the models to Blender and rendered the objects and the vertices.

red: vertices from GJK, green: manual calculation.

we can see that if I scale down the spheres, the red spheres will be contained inside the polygons, while the greens stay on the surface.

[slava-forsight]

Hi,
We ended up fixing this by fixing proximityInfo: turns out that when gjk returns a degenerate result, proximityInfo deflates the second support by 0.1%, runs gjk again and returns that answer. That causes the small error we observed. We patched proximityInfo to return the degenerate result and it solved all of our problems.

[vreutskyy]

Hi @yiftahw-fsr, @slava-forsight
Thanks for the detailed report and investigation. I've created an internal ticket for this issue so it'd be fixed in one of the next releases.
Thanks again, regards.