With Tesla the complaint is that the statistics are almost all highway miles so it doesn’t represent the most challenging conditions which is driving in the city. Cruise then exclusively drives in a city and yet this isn’t good enough either. The AV-sceptics are really hard to please…
You’ll always be able to find individual incidents where these systems fail. They’re never going to be foolproof and the more of them that are out there the more news like this you’re going to see. If we reported about human-caused crashes with the same enthusiasm that would be all the news you’re hearing from then on and letting humans drive would seem like the most scandalous thing imaginable.
I do not care about situations that they work in, I care about what situations they will fail at. That’s what matters and that’s what no company will tell you. As said, we learn about the capabilities of self driving cars one crash at a time, and that’s just unacceptable when you could figure all of that out years ago in simulation.
So far none of the self-driving incidences I have seen were some kind of unforeseen freak situation, it was always some rare, but standard thing, fog, pedestrian crossing the road, road blocked by previous crash, etc.
Humans get into accidents all the time. Is that not unacceptable for you?
I feel like people apply standards to self driving cars that they don’t to human driven ones. It’s unreasonable to expect a self driving system never to fail. It’s unreasonable to imagine you can just let it practice in simulation untill it’s perfect. This is what happens when you just narrowly focus on one aspect of self driving cars (individual accidents) - you miss the big picture.
I feel like people apply standards to self driving cars that they don’t to human driven ones.
Human drivers need to pass driving test, self-driving cars do not. Human drivers also have a baseline of common sense that self-driving cars do not have, so they really would need more testing than humans, not less.
It’s unreasonable to expect a self driving system never to fail.
I don’t expect them to never fail, I just want to know when they fail and how badly.
It’s unreasonable to imagine you can just let it practice in simulation untill it’s perfect.
What’s unreasonable about that?
individual accidents
They are only “individual” because there aren’t very many self-driving cars and because not every fail ends up deadly.
Tesla on FSD could easily pass the driving test that’s required for humans. That’s a nonsensical standard. Most people with fresh license are horribly incompetent drivers.
I don’t expect them to never fail, I just want to know when they fail and how badly.
“Over 21 months in Arizona, Waymo’s vehicles were involved in 47 collisions and near-misses, none of which resulted in injuries”
“These crashes included rear-enders, vehicle swipes, and even one incident when a Waymo vehicle was T-boned at an intersection by another car at nearly 40 mph. The company said that no one was seriously injured and “nearly all” of the collisions were the fault of the other driver.”
According to insurance companies, human driven cars have 1.24 injuries per million miles travelled. So, if Waymo was “as good as a typical human driver” then there would have been several injuries. They had zero serious injuries.
The data (at least from reputable companies like Waymo) is absolutely available and in excruciating detail. Go look it up.
As already said, I want to know where they fail, preferably in the simulator, not on actual roads. Having vehicles drive in circles on carefully selected roads and making a lot of miles is no big accomplishment and not comparable with humans that have to drive on all the roads under all the conditions.
As a software developer, that’s not how testing works. QA is always trying to come up with weird edge cases to test, but once it’s out in the wild with thousands (or more) of real-world users, there’s always going to be something nobody ever tried to test.
For example, there was a crash where an unmarked truck with exactly the same color as the sky was 90° sideways on the highway. This is just something you wouldn’t think of in lab conditions.
there’s always going to be something nobody ever tried to test.
That’s not what is happening. We don’t see weird edge cases, we see self driving cars blocking emergency vehicles and driving through barriers.
For example, there was a crash where an unmarked truck with exactly the same color as the sky was 90° sideways on the highway.
The sky is blue and the truck was white. Testing the dynamic range of the camera system is absolutely something you do in in lab situation. And a thing blocking the road isn’t exactly unforeseen either.
I don’t expect self driving cars to be perfect and handle everything, but I expect the manufacturers to be transparent about their abilities and they aren’t. Furthermore I expect the self driving system to have a way to react to unforeseen situations, crashing in fog is not acceptable when the fact that there was fog was plainly obvious.
And a thing blocking the road isn’t exactly unforeseen either.
Tesla’s system intentionally assumes “a thing blocking the road” is a sensor error.
They have said if they don’t do that, about every hour or so you’d drive past a building and it would slam on the brakes and stop in the middle of the road for no reason (and then, probably, a car would crash into you from behind).
The good sensors used by companies like Waymo don’t have that problem. They are very accurate.
With Tesla the complaint is that the statistics are almost all highway miles so it doesn’t represent the most challenging conditions which is driving in the city. Cruise then exclusively drives in a city and yet this isn’t good enough either. The AV-sceptics are really hard to please…
You’ll always be able to find individual incidents where these systems fail. They’re never going to be foolproof and the more of them that are out there the more news like this you’re going to see. If we reported about human-caused crashes with the same enthusiasm that would be all the news you’re hearing from then on and letting humans drive would seem like the most scandalous thing imaginable.
I do not care about situations that they work in, I care about what situations they will fail at. That’s what matters and that’s what no company will tell you. As said, we learn about the capabilities of self driving cars one crash at a time, and that’s just unacceptable when you could figure all of that out years ago in simulation.
So far none of the self-driving incidences I have seen were some kind of unforeseen freak situation, it was always some rare, but standard thing, fog, pedestrian crossing the road, road blocked by previous crash, etc.
Humans get into accidents all the time. Is that not unacceptable for you?
I feel like people apply standards to self driving cars that they don’t to human driven ones. It’s unreasonable to expect a self driving system never to fail. It’s unreasonable to imagine you can just let it practice in simulation untill it’s perfect. This is what happens when you just narrowly focus on one aspect of self driving cars (individual accidents) - you miss the big picture.
Human drivers need to pass driving test, self-driving cars do not. Human drivers also have a baseline of common sense that self-driving cars do not have, so they really would need more testing than humans, not less.
I don’t expect them to never fail, I just want to know when they fail and how badly.
What’s unreasonable about that?
They are only “individual” because there aren’t very many self-driving cars and because not every fail ends up deadly.
Tesla on FSD could easily pass the driving test that’s required for humans. That’s a nonsensical standard. Most people with fresh license are horribly incompetent drivers.
So why don’t we check it? Right now we are blindly trusting the claims of companies.
What are these claims we’re blindly trusting exaclty? Do you have any direct quotes?
Have you used it? It’s not very good. It tries to run red lights, makes random swerves and inputs, and generally drives like someone on sedatives.
They’ve had to inject a ton of map data to try to make up for the horrendously low resolution cameras, but “HD MaPs ArE a CrUtCh” right?
No radar or lidar means the sun can blind it easily, and there’s a blind spot in front of the car where cameras cannot see.
Is what they’ve made impressive? Sure, but it’s nowhere near safe enough to be on public roads in customer’s cars. At all.
“Over 21 months in Arizona, Waymo’s vehicles were involved in 47 collisions and near-misses, none of which resulted in injuries”
“These crashes included rear-enders, vehicle swipes, and even one incident when a Waymo vehicle was T-boned at an intersection by another car at nearly 40 mph. The company said that no one was seriously injured and “nearly all” of the collisions were the fault of the other driver.”
According to insurance companies, human driven cars have 1.24 injuries per million miles travelled. So, if Waymo was “as good as a typical human driver” then there would have been several injuries. They had zero serious injuries.
As already said, I want to know where they fail, preferably in the simulator, not on actual roads. Having vehicles drive in circles on carefully selected roads and making a lot of miles is no big accomplishment and not comparable with humans that have to drive on all the roads under all the conditions.
As a software developer, that’s not how testing works. QA is always trying to come up with weird edge cases to test, but once it’s out in the wild with thousands (or more) of real-world users, there’s always going to be something nobody ever tried to test.
For example, there was a crash where an unmarked truck with exactly the same color as the sky was 90° sideways on the highway. This is just something you wouldn’t think of in lab conditions.
That’s not what is happening. We don’t see weird edge cases, we see self driving cars blocking emergency vehicles and driving through barriers.
The sky is blue and the truck was white. Testing the dynamic range of the camera system is absolutely something you do in in lab situation. And a thing blocking the road isn’t exactly unforeseen either.
Or how about railroad crossing, Tesla can’t even the difference between a truck and a train. Trucks blipping in out of existence, even changing direction, totally normal for Tesla too.
I don’t expect self driving cars to be perfect and handle everything, but I expect the manufacturers to be transparent about their abilities and they aren’t. Furthermore I expect the self driving system to have a way to react to unforeseen situations, crashing in fog is not acceptable when the fact that there was fog was plainly obvious.
Tesla’s system intentionally assumes “a thing blocking the road” is a sensor error.
They have said if they don’t do that, about every hour or so you’d drive past a building and it would slam on the brakes and stop in the middle of the road for no reason (and then, probably, a car would crash into you from behind).
The good sensors used by companies like Waymo don’t have that problem. They are very accurate.