Why does primitive Box doesn't use special support-function but instead falls back to discrete vertices ?

[manuel-koch]

I'm comparing distance3d performance with respect to using meshes ( vertices approximating a box, sphere, cylinder ) vs using primitives like colliders.Box().
When doing many collision tests, I noticed that the former ( meshes ) appears to be faster than the latter ( primitives ).
Looking into colliders.Box() I noticed that it doesn't seem to use the generic support function support_function_box but instead falls back to plain vertices, whereas the primitive colliders.Sphere uses the builtin function support_function_sphere.

I.e. colliders.ConvexHullVertices colliding with colliders.ConvexHullVertices seems to be faster
than colliders.Box|colliders.Sphere|colliders.Cylinder colliding with colliders.ConvexHullVertices.

I even tried to patch colliders.Box()using the following implementation, but the results are the same ( ConvexHullVerticesis faster than the other primitive types ):

class Box(ConvexHullVertices):
   .....
   def support_function(self, search_direction):
        return support_function_box(search_direction, self.box2origin, self.size)

[AlexanderFabisch]

Box uses the vertex representation because it is faster AND correct. In the case of spheres, capsules, cylinders, etc. the vertex representation might be faster, but it is not entirely correct. We would need an infinitesimal resolution if we want to make it correct.

Having said that, I think the real support function should be faster if we had a properly compiled support function. Unfortunately, numba is not good at handling classes. Caching and inheritance are not supported. I am still interested in making the code faster, but at some point it is not possible anymore unfortunately...

[manuel-koch]

I'm currently testing other collision detection libs too and those that support primitives seem to perform better when using primitives instead of approximated mesh vertices:

• fcl
• pybullet

[AlexanderFabisch]

Of course, these are C++ implementations and the code can be much more efficient.

For this library I wanted to keep everything in pure Python to make it easily editable and to integrate in Python code without too many dependencies. We are currently exploring the limits of this approach with numba.

[AlexanderFabisch]

For comparison, standard FCL uses libccd as a backend. The ported libccd implementations (of GJK and MPR) are also part of this library.

[AlexanderFabisch]

pybullet internally also uses a fork of libccd.