Imprecise collision detection between Kinematic and Static shape

[uzername]

I have created an application in C# WPF and Helix toolkit (as 3D renderer) to learn Bepu Physics 2 and physics of collision detection in general. Kynematic Cylinder moves all the way down to a Static Box with step of 0.1 . I observe that in some cases collision gets detected when it should not really be there, have a look at attached picture.

It's really interesting that collision depth in Narrow Phase Callback is reported to be negative.

Collision between two bodies with 2 contact(s).
... Simulation paused ...
Contact 0: Offset=<-1.8566809  -0.7241254  0.52761614>, Depth=-0.08074403
Contact 1: Offset=<1.8566809  -0.7241254  0.52761614>, Depth=-0.08074403

On next step of simulation collision - depth is positive, when collision is observed on graphical display. Here are values from next two steps

... Simulation continues ...
+++ Coordinate of cylinder: [ 0 ; 0 ; 1.2000000000000188]
Collision between two bodies with 2 contact(s).
... Simulation paused ...
Contact 0: Offset=<1.8700132  -0.7124277  0.49547675>, Depth=0.013225213
Contact 1: Offset=<-1.8700132  -0.7124277  0.49547675>, Depth=0.013225213
... Simulation continues ...
+++ Coordinate of cylinder: [ 0 ; 0 ; 1.1000000000000187]
Collision between two bodies with 2 contact(s).
... Simulation paused ...
Contact 0: Offset=<1.8826301  -0.70072997  0.46333727>, Depth=0.10719454
Contact 1: Offset=<-1.8826301  -0.70072997  0.46333727>, Depth=0.10719454

Also in this case I observed such mesages from Narrow Phase Callback, for example:

+++ Coordinate of cylinder: [ 0 ; 0 ; 1.400000000000019]
 Collision registered but no contact points reported

What is mathematical meaning of observed phenomenons:

• negative depth of collision
• collision that reported too early
• collision in narrow phase without actual points of collisions
  (I believe they may be related)

[uzername]

For above mentioned example I did calculations (using school geometry) and collision should appear when cylinder is at 1.21407400337, which is confirmed by observations in Blender

[uzername]

Well, ChatGPT answered:

This is a very insightful observation — and what you're seeing is expected behavior in BepuPhysics2 under certain conditions.

Why does this happen?

In BepuPhysics2, speculative contacts are used for kinematic bodies and CCD-like behavior. This means:

A negative penetration depth indicates that two shapes are not yet overlapping.

But Bepu predicts that they will touch or slightly penetrate by the next timestep, so it generates a contact early.

This allows BepuPhysics to:

• Avoid tunneling
• Provide smoother collisions by anticipating contact

Your cylinder is still above the box by ~0.08 units.

But the engine anticipates that it will penetrate next frame.

So it gives you early contact information (with negative depth).

This is not a bug, it’s speculative contact working correctly.

[uzername]

Next, ChatGPT suggested the following:

If you want real, actual contacts only (i.e. penetration has occurred):

In ConfigureContactManifold, filter out negative depth contacts:

if (manifold.Count > 0)
{
    for (int i = 0; i < manifold.Count; ++i)
    {
        ref var contact = ref manifold.GetContact(i);
        if (contact.Depth >= 0f)
        {
            // Real contact
        }
        else
        {
            // Speculative contact — may skip or log
        }
    }
}

Tweak Precision if Needed

You can reduce how early speculative contacts occur by lowering the speculativeMargin, e.g.:

new CollidableDescription(shapeIndex, 0.01f)

This makes speculative contacts more conservative (closer to true contact).
In my case I had also to change AllowContactGeneration from Narrow Phase Callback

public bool AllowContactGeneration(int workerIndex, CollidableReference a, CollidableReference b, ref float speculativeMargin)
{
    speculativeMargin = MathF.Max(speculativeMargin, COLLISION_PRECISION);
    return true;
}

These measures helped, so it's resolved

[RossNordby]

I can confirm that chatgpt is correct; negative depths are expected for speculative contacts. You can indeed filter them out if you want to pay attention only to touching/penetrating contacts.

And yes, you can modify the maximum speculative margin to clamp these to small or zero separation. Doing so would harm collision response stability for dynamics, but for a query-like use case it's fine.