Security margin and weird collision (penetration depth) results

[maxwave11]

Hello!

I'm trying to understand the terms "penetration depth" and "security margin" and completely confused and stuck.

Here's my simplest case with BVH models:

Here's my collision request settings:
col_req.security_margin = -0.1

1. A with C collision

Results:

distance_lower_bound: -0.320594
Num contacts: 3
contact[0]:
        Penetration depth: -0.420594 <----- why these values so big and what they mean?
        Witness point on shape1: (-1,0.949945,0.872812),
        Witness point on shape2: (-1,0.949945,0.872812),
        Point: (-1,0.949945,0.872812),
        Normal: -1  0  0
contact[1]:
        Penetration depth: -0.420594 <----- why these values so big and what they mean?
        Witness point on shape1: (-1,1,0.872812),
        Witness point on shape2: (-1,1,0.872812),
        Point: (-1,1,0.872812),
        Normal: -1  0  0
contact[2]:
        Penetration depth: -0.413701 <----- why these values so big and what they mean?
        Witness point on shape1: (-0.872812,0.949945,1),
        Witness point on shape2: (-0.872812,0.949945,1),
        Point: (-0.872812,0.949945,1),
        Normal: -0  0  1

2. A with B collision

Results:

distance_lower_bound: 0.0499448
- no contacts (no collisions)

Despite the fact that object B penetrates much deeper, we do not get any intersections at all. On the contrary, in the case of object C, we get a very large penetration depth, even if we subtract the security margin from it.

I've drawn the contact points and their normals. All of them look correct. However, I have no idea what "penetration depth" means for each point. In the case of A and C, it can't be greater than 0.1.

Security margin

Also I'm trying to understand what visually means security margin. Despite the case A && C being quite simple, collision occurs even though they shouldn't, based on the screenshot below. Could it be that using BVH (OBRSS) models for such checks won't work?

models.zip