SDF Collision Response Stability

[kv-ep]

Across a wide range of scenarios when using PhysX 5, I have encountered reduced stability when using SDF colliders on dynamic objects when compared to primitive or convex collider alternatives. Here are a few of the specific issues I have encountered.

1. More likely to become deeply entombed with another object and unable to resolve the overlap.
2. Large stacks of SDF colliders appear to have reduced contact convergence, and the pile becomes more "springy".
3. More likely to randomly "pop" and disturb the resting state of a pile of objects.
4. Large stacks of SDF colliders are less able to converge to a resting state and continue to slowly slip and settle over time.

I have included a couple videos of one such scenario below as an example. Note that in this scenario a convex collider is adequate to represent the shape, but this is simply to demonstrate the effect. I also scrub back and forth at the end of the video a bit to show the extent to which the simulations settle over time (note that the effect is much more pronounced with SDF colliders).

SDF Collider: https://github.com/user-attachments/assets/9cdc868b-2f3d-4cdf-8524-0e1d4de4673a
Convex Collider: https://github.com/user-attachments/assets/158875fe-6e1c-4011-b05f-0144142f3dc7

I have increased solver position iterations (holding velocity iterations at 1), adjusted contact and even rest offsets, tried different SDF resolutions and both the PGS and TGS solvers (TGS behaves better and is what is used in the above videos). Are there any other parameters which I should use to try to improve this behavior? Is it expected that for the same set of parameters, the SDF collider would have worse stability?

[kv-ep]

https://forums.developer.nvidia.com/t/sdf-and-convex-hull-collision-bounces/288053

I've found this which seems to be relevant, however, increasing the dynamic bodies mesh resolution does not seem to improve the response behavior as that post would indicate. Possibly this no longer applies to the current version of PhysX.

[camevor]

Hi @kv-ep, and thank you for you question.

Note that SDF colliders are less suited to stacking as they generate at most one contact per triangle (of each mesh), which can lead to reduced contact convergence and shifting, especially in stacking scenarios and where large triangles are involved. Subdividing the mesh itself can help mitigate this.

You can find more details on this and general collision tuning advice that is also applicable to PhysX in this Collision Behavior Guide and this Colliders Guide.

Generally, reducing the time step is also likely to improve the contact behavior. Are your results identical across CPU and GPU simulation? Finally, a minimal, self-contained reproduction case – similar to one of the snippets – would be very helpful.

[kv-ep]

Thank you for the reply. That extra documentation was quite helpful; I was unaware that existed.

Yes, results did appear to be the same across both CPU and GPU simulations. We have temporarily gotten around this issue by implementing a convex decomposition algorithm to make use of convex colliders and will follow up on some of these other leads when we get a chance.