In comma.ai the solution is to just not run into things / go around things how a human would do and learn to peioritize what things to go around and where to be extra careful.
The important part of the thinking is at that point knowing whether a thing is a cone or not is just a technical debt in the code base, as it shouldn't be acted upon. At the same time those cones are perfect test to see if the software can learn to go around things or not, so that extra code hinders development of the final software.
It doesn’t matter what a static object is, you don’t drive into it unless you’re dang sure it fits in the class of “things that are safe to collide with.”