Actually, if you increase the axis samples on the cone to above 2, then this problem goes away.

sbayless said:

Actually, if you increase the axis samples on the cone to above 2, then this problem goes away.

Then the collision routine is correct - it assumes the cone to be a perfect one, not an approximated one.

The collision routine shouldn't "assume" anything… This is triangle-accurate collision, it doesn't care that it's a box or a cone, it just checks each of the mesh's triangles if they collide with another mesh's triangles.

Right, collision does not know anything about the shape other than the triangles and related data. Wild guess, but maybe there is some kind of winding issue going on here.