I've long expressed disappointment in the choice of camera positions that Tesla adopted (or perhaps inherited from the MobilEye platform that was conceived for a more highway-centric autopilot).
However, wishing for an improvement is not the same as saying that the exist setup is unworkable or highly dangerous.
The way I've looked at this is that
- it's predictably difficult to equal or exceed human safety levels, when no one yet knows how to achieve human level inference in the AI.
- But, a major emphasis can be placed on areas where it's already known how to achieve superhuman performance at low cost. Driving is a very complex task, and there isn't only one set of weighted talents that can conquer it.
So it's very sensible to take maximum advantage of the available technology, within the severe cost constraints of volume production. Cameras that can see all around and that can see better in the dark, a computer that never stops paying attention, large crowdsourced databases that ideally respond to short, medium and long-term route changes.
Some of these exist and are helping, others are falling short or could have been better. Specifically, some argue that the B-pillar location is not
that much of a disadvantage compared to a human view, not worse than a vehicle with a very long hood, etc., but I think the larger point is that it
could have been placed up front to look around corners and look past traffic, and to help with near-field obstacles - all better than a human in almost any vehicle. And camera Field-of-View geometry is is just one example of the superhuman advantages that could have been had for cheap, helping to offset the lack of human-level inference.
What other things besides better camera FOV coverage? Relatively inexpensive camera auto-cleaning, easier owner access to keep the interior windshield clean over the important front cameras, inexpensive night-vision (near-IR) illumination within the existing lamp and camera housings, a couple of cheap external microphones providing all kinds of potential benefits to safety and to Robotaxi capability. The BOM cost for these suggestions isn't zero, but it's probably less than the cost of the abandoned UslSS.
So in summary, I think the point isn't just isolated arguments about which specific edge cases might benefit from specific hardware changes. The design philosophy should be to extract
all the possible benefits from inexpensive hardware, let the NN process those inputs with minimal guesswork, and let it devote more cycles to the unavoidably harder decisions.
