Tesla is hiring for a DSP engineer for electronic scanning radar, so the rumor is this is temporary (camera only safety systems) until they’re ready to deploy their own, more advanced radar sensor (trained off of the visual neural net) versus the crude Continental radar sensor they’ve deprecated.
Even the best implementations dazzle other drivers.
Auto high beams are only realistic when used where normal high beams are. That is, mostly on single lane roads with no dividers, little to no street lights, where opposing traffic is not appearing often, etc.
-
It seems like such a small thing compared to driving a whole car... but this is one of the few vision features that directly affects other drivers on a regular basis.
The comments explain how for example, vehicles with tall ride heights like commercial trucks get dazzled due to their tailights sitting much lower than their mirrors
It's just not realistic to leave your high beams "potentially on at all times" unless you're willing to dazzle other drivers.
https://www.autoweek.com/news/technology/a34417809/nhtsa-sti...
https://electrek.co/2021/05/26/lucid-reveals-ux-intuitive-us...
Edit: the reason here is that road paint (and the marking on a sign) is typically retroreflective, while roads (and regular painted sheet metal) are not. Retroreflective materials are very easy to detect with LIDAR. Even if you are on older roads that don't have retroreflective paint, white objects and black objects also tend to show up differently.
Edit2: there's also no reason you can't use detection algorithms to characterize pointclouds. You can also fuse the returned objects from the lidar sensor(s) with the objects from the camera(s). You have the latter problem even with just multiple cameras. Inevitably, whatever you are using for segmentation/classification will return different results on different sensors, for sensors that share the same FOV - you do need a strategy for handling this.
Tesla has not added LIDAR because of cost (and because of Elon's ego...) - they are one of the few companies going directly for consumer sales - more parts that cost $$$ and require calibration aren't good for that.
Single cameras (even stereo cameras) require a lot more assumptions about contrast and environment lighting to do image segmentation and depth maps. They're rarely as accurate as LIDAR and often far worse.
Depth maps and image segmentation are the building blocks of object identification. You can try to AI your way into those with a single camera but it'll be less accurate and more error prone.
It will mean Tesla is done with trying to avoid adding it. They test with lidar so they might as well start using it:
https://insideevs.com/news/508669/tesla-model-y-luminar-lida...
If it wants auto high beam, it can just enable the setting itself, and restore your own preference if you disable autopilot (and even better: it can control the beams when autopilot is enabled, irrespective of any user settings, after all if autopilot can steer why couldn't it also change the lights)?
It’s not an irrelevant bit of context.
A human driver will see the headlights coming and turn their high-beams down preemptively.
Its doubly shitty because cars outfitted with auto high-beams typically have those high intensity lamps too. Ugh.
They are a safety hazard, and unless they consistently operate headlamps in accord with rhe applicable local law [0], using them already is illegal.
[0] e.g., https://leginfo.legislature.ca.gov/faces/codes_displaySectio...
it's funny to me because you'd expect music to be lower stakes and such, but it just highlights that actually driving a car is a much more well defined problem than picking the music i like