The choice to remove radar has definitely reduced the chances that I will buy a tesla. It would be one thing if they lowered the price, but the "full self driving" upgrade continues to be $10k usd.
Plus the cost of the latest FSD computer upgrade, even if Elon Musk promised your old model "has all the required hardware for Level 5 autonomy". Your fault for believing him, pay up.
* Isn't actually Full Self Driving
I've had a Model 3 for 3 years with FSD. My experience:
Autopark - trying to get it to show up is almost impossible. When it does show up, the chance of curb rash is between 95-99%.
Summon - for the most part works if it can connect. still does some sketchy stuff some times.
Smart Summon - literal piece of shit that will always do the wrong thing and then just stop in the middle of the road.
Navigate on Autopilot - kind of works, but it makes the auto-lane change feature finicky. Without NOA, auto lane change works fine, with it enabled, fails 80% of the time with the car jerking back into the original lane.
Can't wait for FSD though /s
I think stereo vision won’t help humans much at the distances we’re talking about (might be different for a car with cameras over a meter apart, but that will suffer from bad resolution compared to the human fovea).
Parallax probably is a clue we use. You want to brake gently to optimize comfort for the passengers, so let’s say you want to start braking 10 seconds away from the traffic light. Since sin(x) ≈ x, the image of the traffic light will move up about 10% every second at that moment.
On a camera, that could be less than 100 pixels, but I think that should be detectable.
⇒ you don’t have to discriminate between the moon and a tragic light, you just have to find the nearby yellowish blobs.
Is it publicly known what Tesla’s self-driving does? Given the above, I think I would try to extract a depth field from the known speed of the car and the camera images, discard or deprioritize whatever is more than 20 seconds out, and then let some ML algorithm loose on detecting points of interest such as traffic lights, cars, etc. (As opposed to “feed a ML algorithm lots of video, and hope it will learn to deprioritize stuff that’s far away)
A human disambiguates by looking around, lots of prior knowledge about how the world works, etc. That doesn't seem to be the case with the current Teslas.
So if you’re in Europe/Asia/Oceania or in the market for an S/X - you’re in luck!