Good post. The excerpt I took out is the only part that I somewhat differ with. I do think that more elaborate sensors could be fused more easily than was the case with the heuristic based system, but to solve the curbing I don't think it takes particularly fancy hardware.For instance, with fast precise absolute distances, including on sides of car with full-angle high res lidar or radar you could enforce policy that prevents the collisions and rim curbings in extreme situations while letting the net drive most of the time.
First, I think that ongoing training can and will solve the present situation with curbing issues, But it didn't have to happen and I also think the main original fault here is the deficit in low forward/side camera coverage.
A lot has been said about side looking cameras farther forward, and about bumper cameras for parking. I believe that both would be very helpful, and at least 3 years ago I was arguing for cameras in the headlights or elsewise at the corners. But today I'd say that the single add-on that would have the most bang for the buck (no pun intended) would be a forward-looking camera element added to the side repeaters. In simple terms, extend the existing rearward view to form a continuous panoramic view that takes in the sides, curbs and sighting along the front fender. This would do a great deal for near field, cross traffic creeping view, and perception of oncoming lanes that are obscured by oncoming left-turning traffic or stopped ehicles directly in front.
Regarding radar, maybe we will see that in the next generation (not to image curbs and near field infrastructure which it wouldn't be very good at, but for an extra layer of adversarial traffic detection). Still, I think just filling in the camera POV gaps is the highest priority and probably the most cost efficient.