Meet a non-deterministic problem when two colliable kinematic rigid bodies come into intersection

[tianxing113]

As shown in the figure, I have two kinematic rigid bodies, A and B, moving in the same direction, where B is moving faster than A. They are on the same object layer, but I mistakenly set this layer to be able to collide with itself.

I ran this demo on two machines simultaneously. At the beginning, the results of each frame on both machines were completely identical. After a while, when B caught up with A, B and A intersected. At this point, the results calculated by the two machines were no longer the same. And after a while, when B completely overtook A, the calculation results returned to being consistent.

I extracted a BodyPair segment from a Contact, as follows:

machine A:
mContactNormal = 1.22139943e-06, -5.34267519e-06, 1.00000000
mPosition2 = -0.00106811523, -1.10757065, -2.31998825

machine B:
mContactNormal = 1.22139943e-06, -5.34267519e-06, -1.00000000
mPosition2 = -0.00106811523, -1.10757065, 2.31998825

The z-values of mContactNormal and mPosition2 are opposite on the two machines.

I suspect that when two kinematic objects intersect, the collision calculation may produce indeterminate results. Should kinematic rigid bodies be prevented from intersecting?

Thanks.

[jrouwe]

It only makes sense for kinematic objects to intersect with each other if there's some game mechanic that you want to trigger. The physics simulation does nothing with collisions between kinematic objects.

I assume you have set BodyCreationSettings::mCollideKinematicVsNonDynamic = true (or called Body::SetCollideKinematicVsNonDynamic) to allow these collisions to be detected?

I'm assuming you've defined JPH_CROSS_PLATFORM_DETERMINISTIC?

What CPUs/OS-es did these machines have?

There are too many unknown variables at this point for me to investigate. Can you please create a minimal reproduction case? A popular approach is to hack SimpleTest in the Samples app. If you can create something that e.g. traces the contact points and that produces different results when run multiple times / when run on different machines, then I could try to reproduce. If you're not on Windows, you could also hack the HelloWorld app to do the same.