So what about a splotch on the windshield, a plastic bag floating in the air, a sheet of rain, or a clump of fog and so on. You can take the Tesla route, which is to load up on cameras and neural compute in the trunk, and try to identify every object you see to attempt to divine its relevance and intent...and still fail when you see a plastic bag. Or plop on some lidars to get high accuracy, but sparser and slower data.
Fundamentally though, grenadiers are limited because they have one camera, and it's not even stereoscopic. The only way to know how far something is, is to know what it is and how large large it's supposed to be. This is a reasonably constrained problem when you talk about humans and cars and trucks. When you expand this, you very quickly end up down the Tesla rabbit hole and wish you could fit more GPUs in the trunk.