That visual sensors only seem like they make no sense when they can get obstructed by weather
I work on a synthetic aperture radar system that is high-resolution enough to "see" the painted stripes on roads through fog and a thin layer of snow.
It's not for automotive use and would increase the price of every car by several hundred thousand dollars but a fusion of multi-spectrum sensors should be the direction we are headed-- not a minimally-viable mono-sensor system.
In other words, failing to realize that raising minimum requirements (and therefore prices) makes it unaffordable for an increasing number of people.
Human-parity (in terms of accident rate, not failure mode) seems a reasonable minimum bar.
For proof I offer the price of the 1969 Volkswagen Beetle, the least expensive new car for sale in the US market for almost its entire sales history: $1800. That’s for one with zero options. A rolling chassis with four seats and a motor.
That’s $16k today. Coincidentally the same price of the Nissan Versa or Mitsubishi Mirage: with backup camera, air conditioning, and airbags.
People don’t WANT the cheap cars, though.
I know the average price of a new car has exploded.
That is a conscious choice by the consumer.
When production of autonomous tech scales it won’t increase costs as much as people assume.
Blanketing the frequency slice with echos seems like a non-trivial problem. Or would that help?
I’ve seen the “human-like” behaviour of FSD 12.x praised a lot by channels like this, particularly where the car is breaking the rules in a way they consider “normal”. And it’s a fair argument that predictable behaviour improves safety.
However, behaviour that is common in the US - like making a turn into a side street while a pedestrian is beginning to cross - would be considered exceptionally aggressive and reckless here in Australia. It’s a cultural difference I’ve adapted to when moving back and forth.
At the end of the day though, when I walk across a street, I don’t want to have to worry if Tesla has fine tuned their model correctly to match our local expectations of yielding. I’d rather they just followed the law as closely as possible - because that’s the most predictable behaviour of all.
The advantage of machines is not that they have better brains than humans, but that they have better senses than humans, which is an advantage that goes out the window when you obstinately refuse to use better sensors than merely cameras "because that's what humans use".
I mean, let's be blunt: the "LIDAR vs. cameras" debate has been settled at this point, and the fancy sensors lots. Of all the things one can complain about with FSD, sensor fidelity is not one. Teslas don't hit things. The problems remaining to be solved are in the planning regime: my car still misses turns at a rate somewhat higher than I do, usually because it's in the wrong lane (and that often because it thought that the backup in the turn lane was something it could go around).