The thing I don't understand about vision-only. It only looks forward. Those four little circles on the rear are sonar. They can "see" 2-3m at best, and at a pretty narrow field-of-view. For FSD purposes, that's fine for last-resort don't bump into things with your bumper. As a self-driving engineer, I just don't understand what they're plans for doing more complex reverse maneuvers, like backing to a crowded parking lot with a 10-20m manuever. You can sometimes get a good look at the scene with the forward-looking cameras then "hope" that the scene doesn't change as you go in reverse. My conservative nature doesn't like that.

There's also the issue of not seeing things not at bumper level at all. Say you're backing into a spot and there's an SUV with a bunch of 4x4 lumber hanging out the rear window. You're gonna put that lumber right through your rear window....no way of seeing it.

I just don't understand. On the other hand lots of people at Tesla are smarter than I am....

Edit: There is a rear-looking monocular camera. It's theoretically possible they do stereo-through-motion (instead of using two side-by-side cameras to do stereo, you use one camera, but take images in multiple positions). I'll be really impressed if they can pull this off in real time well. It's computationally brutal compared to conventional stereo.

The advantage of going vision-only is that you don't have to solve the problem of what to do with conflicting sensor data, particularly with false positives. For example if you have cameras detecting an obstacle ahead, but radar says it's clear... what do you do? Radar is supposedly more reliable in bad weather, etc. But for safety they can't just ignore that camera alert. At the same time, the car cannot run on pure radar.

So ideally a vision system is sensitive enough to avoid false negatives (missing something that should have been spotted), at which point maybe you can ditch the other sensors. However, when you tune it this way, you tend to get more false positives, which cause the behavioral issues people are mentioning like phantom braking.

This gets into precision and recall, which have been quite frequently tuned in the recent patch notes for the FSD beta.

The potential is there to have a robust vision-only system, but the devil is in the details.

also note the claimed 8m range on ultrasonics

Also one more thing that may go overlooked: the Tesla vision-only system clamps your range keeping settings down a notch to 2 through 7. The closest setting is not allowed. Which basically tells me the vision system needs more frames of data to detect if the gap is closing, and the rate. Radar has limitations but that particular aspect is *exactly* what it was designed for, and you get almost instant feedback.

It's not really the bolded. It's just the way to configure / tune vision is different if there are no other sensors in the mix.

That being said, I think their removal of the radar was premature. It smells of beancounters and should have stayed in place until the pure-vision neural network training is sufficient.