Self-Driving Uber Car Kills Arizona Pedestrian

[JoeMorgue]

Originally Posted by Elagabalus

I picture a different scene. Freeway: person1 going the speed limit in the overtaking lane, when person2 comes up behind them (not doing the speed limit, obviously) and honks or puts the brights on, but person1 refuses to move over because "they're going the speed limit!". Person2 has to overtake in the slow lane.

Sure if you live in some alternative universe where the speed of traffic on any major freeway/highway/ isn't doing 20 over the limit just as a default sure.

"The old guy driving 35 in the fast lane of the freeway and being a total menace" is one of those albino unicorns I always hear people complaining about but have seen maybe, maybe a dozen times in my entire life.

[arthwollipot]

Originally Posted by crescent

It seems as if there wasn't time. The local police have indicated that a human driver would probably not have been able to react fast enough either.

Exactly. Why is nobody talking about that?

[Myriad]

Originally Posted by CORed

I think the job of "safety driver" is damn difficult for the average human to do. 99+% of the time, you have nothing to do, so it is going to be almost impossible to stay alert, and when something does happen, you are likely to be, if not asleep, at least not paying attention. When you are actually driving, you have to pay attention.

I think it's possible to design a solution for that.

One approach: the safety driver is required to operate the vehicle's controls at all times; steering, braking, turn signaling, just as if they were actually controlling the car. Most of the time, they are not actually controlling the car, but whenever their control inputs diverge significantly from the car's automatic behavior, the car follows the safety driver's input instead. It's an automatic override.

Such events are reviewed later, so for instance if the safety driver is constantly overriding the automatic program in order to pass other cars unnecessarily, they'd come under scrutiny and if necessary be retrained or fired. But in a case of auto driver failure, the safety driver wouldn't have to decide to override the automatic program. When they see a hazard they might be expecting the program to successfully react to it, but they'll still react to it themselves which will cause an override if the program missed it.

(The rules have to be a bit more refined than that, for cases such as when the auto driver reacts to a hazard that the safety driver doesn't see. In those cases the auto driver's caution should have priority over any neutral "keep going" control input from the safety driver. This would be similar to existing devices like impending collision detection with automatic braking, that override the driver.)

So, most of the time, the human driver is essentially playing a driving sim game, where the goal is to anticipate and mimic the autonomous program. Their "scores" are monitored. This keeps them occupied. If they do nothing, the car does nothing, because the safety driver's control inputs are prioritized.

Maybe that's how it already works. It's hard to believe that in a real emergency there would be time for an idle safety driver to become alert to the hazard, grab the controls, and/or push some kind of override button, so some sort of continually active safety driving would make sense. But if that's how it works, it shouldn't be possible for a safety driver to fall asleep or disregard the road and still have the car continue operating.

[xjx388]

Originally Posted by smartcooky

Well, I guess if you were driving along at 50km/h in a 50km/h zone and had some aggressive asshat tailgating you so close that you couldn't see his number plate in the rear view mirror, who then overtook you in a dangerous manner and was more interesting in giving you the finger or shaking his fist at you than watching the oncoming traffic, putting yours, his and other road user's lives at risk, you might think that keeping up with the flow of the traffic might actually be the safer bet!!!

If it weren't for the km/h reference, I would have thought you were a native Texan! That is definitely the approach I take when I travel. My father in law is the opposite. He sets the cruise control at 5 miles under the speed limit and just lets people pass him. I wonder if there are stats on the optimal driving style and if they figure into self driving cars. All I can say is that he has been in 2 highway accidents and I haven't (fingers crossed).

[Skeptical Greg]

Originally Posted by arthwollipot

Exactly. Why is nobody talking about that?

I mentioned that a live driver would probably not have been charged in the same circumstances..

[crescent]

Originally Posted by arthwollipot

Exactly. Why is nobody talking about that?

Originally Posted by Skeptical Greg

I mentioned that a live driver would probably not have been charged in the same circumstances..

And they have not been totally clear on that yet, nor have they released the video.

Once we see the video (if they release it) it will be easier to get a feel for what happened.

[smartcooky]

Originally Posted by xjx388

If it weren't for the km/h reference, I would have thought you were a native Texan! That is definitely the approach I take when I travel. My father in law is the opposite. He sets the cruise control at 5 miles under the speed limit and just lets people pass him. I wonder if there are stats on the optimal driving style and if they figure into self driving cars. All I can say is that he has been in 2 highway accidents and I haven't (fingers crossed).

My standard when driving to and from work (I commute about 12 km or 9 miles each way each day) is to join the flow of traffic and do whatever speed everyone else is dong. Sometimes this can be as much as 15 km/h over the speed limit.

One thing I have to admit is that in this country, we have some absolute asshats behind the wheel. They will

- speed up to close you out at a merge lane,
- see you waiting to turn across their lane into a side road, and will speed up to block you,
- close the gap between themselves and the car in front to prevent an overtaking car from pulling in.
- speed up to 15km/h over the speed limit at a passing lane, then slow down to 15km/h under it when the passing lane ends
- pull out in front of you at the last, safe moment and then do 15km/h under the speed limit

When the traffic is heavy, and you are wanting to turn right across the opposing lane (that would be a left turn in the US) you would wait a long time before someone would stop and flash their lights to let you across, they are more likely to stop in a position to block your turn.

[Darat]

Originally Posted by smartcooky

My standard when driving to and from work (I commute about 12 km or 9 miles each way each day) is to join the flow of traffic and do whatever speed everyone else is dong. Sometimes this can be as much as 15 km/h over the speed limit.

One thing I have to admit is that in this country, we have some absolute asshats behind the wheel. They will

- speed up to close you out at a merge lane,
- see you waiting to turn across their lane into a side road, and will speed up to block you,
- close the gap between themselves and the car in front to prevent an overtaking car from pulling in.
- speed up to 15km/h over the speed limit at a passing lane, then slow down to 15km/h under it when the passing lane ends
- pull out in front of you at the last, safe moment and then do 15km/h under the speed limit

When the traffic is heavy, and you are wanting to turn right across the opposing lane (that would be a left turn in the US) you would wait a long time before someone would stop and flash their lights to let you across, they are more likely to stop in a position to block your turn.

NZ must be where the world dumps all the below average drivers as from many anti-automated cars proponents assure us all of the Top Gun level of performance of each and every driver!

[Dr. Keith]

Originally Posted by GlennB

Yep, but if you see kids playing on a lawn down beyond a parked truck you make allowances, just as you did. The kids aren't directly associated with the road and there are no direct signals they might run out, but you 'know what kids are like'. That's a hard thing to program into an AV.

Kids playing on lawn = danger

Adults milling around a bus stop = not so much danger

And yet you practically wrote the code for it in the next two lines . . .

[Disbelief]

Originally Posted by GlennB

Yep, but if you see kids playing on a lawn down beyond a parked truck you make allowances, just as you did. The kids aren't directly associated with the road and there are no direct signals they might run out, but you 'know what kids are like'. That's a hard thing to program into an AV.

Kids playing on lawn = danger

Adults milling around a bus stop = not so much danger

The AV systems know if it is a person and can anticipate potential movements, so that is already part of the programming. They even detect current movements to rate the probability of someone coming into the road.

[3point14]

Originally Posted by sadhatter

Had someone blow through a red light and hit my car with that logic.

If I had been keeping up with the flow of traffic he wouldn't have hit me avoiding to him. So it wasn't him blowing through the light it was me throwing off his timing.

Whenever I hear someone going on about the flow of traffic I translate it to "I'm a super aggressive driver and in justifying it by insinuating everyone else is just as much of a knob as I am so we should actually be following a special set of road rules made by me".

I can't see that running a red light is equivalent to breaking the speed limit to keep up with traffic.

There is massive evidence that, in a 30 limit, where all the cars are doing 30, the bloke doing 25 is more dangerous than the bloke doing 35.


When the prevailing traffic is moving at 35 MPH, the bloke doing 25 is the only one not breaking the law and also the most dangerous driver on the road.

[GlennB]

Originally Posted by Dr. Keith

And yet you practically wrote the code for it in the next two lines . . .

Hardly. Identify that they're people, estimate their age based on size, assess the nature of the location etc. What I wrote just deals with the outcome of the calculations.

[3point14]

Originally Posted by The Great Zaganza

Obviously it Uber system isn't ready for commercial use - it is sad that it took a fatality to find that out.

By that logic, a car driven by a human being isn't ready for commercial use.

[The Great Zaganza]

Originally Posted by 3point14

By that logic, a car driven by a human being isn't ready for commercial use.

Humans need driving lessons before they can be allowed to drive alone.

And as we have established that autonomous cars have to be better than humans, this system has to learn how to drive better - doing the same mistake a human would do isn't good enough.

[JoeMorgue]

Originally Posted by The Great Zaganza

Humans need driving lessons before they can be allowed to drive alone.

What do you think this was?

[Skeptical Greg]

Originally Posted by The Great Zaganza

... - doing the same mistake a human would do isn't good enough.

Who do you get to write this code that excludes all human mistakes?

[Elagabalus]

Originally Posted by JoeMorgue

What do you think this was?

An example of a machine learning by "doing".

[The Great Zaganza]

I don't get some of the blowback here: Uber is pulling the cars to check if there is something to be done to avoid similar accidents; not because a human could have avoided it, but because in order to gain acceptance, autonomous cars must be safe both for the occupants and everyone outside.
If the engineers believed that the code was flawless already, there would have been no need for the test drives.

[Dr. Keith]

Originally Posted by GlennB

Hardly. Identify that they're people, estimate their age based on size, assess the nature of the location etc. What I wrote just deals with the outcome of the calculations.

That is just defining variables. Pretty routine for software.

The people will already be defined. Their size and age may be easily correlated. The location is pretty damn basic and the nature of the location is about equally well known. I bet the Uber AI knows the location of more bus stops and yards in your city than you do.

[Dr. Keith]

Originally Posted by The Great Zaganza

I don't get some of the blowback here: Uber is pulling the cars to check if there is something to be done to avoid similar accidents; not because a human could have avoided it, but because in order to gain acceptance, autonomous cars must be safe both for the occupants and everyone outside.
If the engineers believed that the code was flawless already, there would have been no need for the test drives.

I think it is presumptuous to assume they pulled them because they think there is a fault. They may have pulled them just to give the police time to discern that there is not a fault. Public perception here is more important than engineering.

And I don't think they believe it is flawless. No engineer I have ever met believed his product was flawless. But they may believe that it is sufficiently robust to be safer than most human drivers. In fact, I think they wouldn't be test driving on roads if they didn't believe that. And the test driving is as much about public perception as it is about gathering data.

[Jack by the hedge]

Of course the great thing about autonomous cars is that they're like the Borg. When one of them encounters a new problem, they all learn the solution.

The requirement for autonomous cars to be better than human drivers is going to hit the problem that they each make different mistakes. There will come a point where the cars are demonstrably less lethal than people overall but there are circumstances where the humans are safe but the machines can't cope. We'll be back to the old seat belt argument "I'd rather be thrown clear" which depends on what kind of accident you intend to have.

[JoeMorgue]

And we'll get into the airplane/car paradox.

People will accept a much, much higher degree of risk if they feel they are in control of it. By any standard per mile traveled flying in a plane is monumentally safer than driving in a car but people don't like the risks associated with air travel because they have near zero control over it. This is why everyone on the road thinks they are the best driver and everyone else isn't. All those people who die in car accidents are other people, not you you're a safe driver.

You can't do that sort of rationalization when you're not in control.

And I'm legit worried that people will start thinking it's funny, clever, or somehow proving something to intentionally mess with self driving cars and turn this into a self fulfilling prophecy.

[Dr. Keith]

Originally Posted by JoeMorgue

And we'll get into the airplane/car paradox.

People will accept a much, much higher degree of risk if they feel they are in control of it. By any standard per mile traveled flying in a plane is monumentally safer than driving in a car but people don't like the risks associated with air travel because they have near zero control over it. This is why everyone on the road thinks they are the best driver and everyone else isn't. All those people who die in car accidents are other people, not you you're a safe driver.

You can't do that sort of rationalization when you're not in control.

Agreed.

Quote:

And I'm legit worried that people will start thinking it's funny, clever, or somehow proving something to intentionally mess with self driving cars and turn this into a self fulfilling prophecy.

Not to be flippant, but there's an XKCD for that.

[Darat]

Originally Posted by Dr. Keith

That is just defining variables. Pretty routine for software.

The people will already be defined. Their size and age may be easily correlated. The location is pretty damn basic and the nature of the location is about equally well known. I bet the Uber AI knows the location of more bus stops and yards in your city than you do.

Yeah but it might be a meeting of exceptionally short people and not kids so the car slowed down for no good reason whereas a human would certainly have slowed down to be a douchebag and stare at the very short people.

[Mikemcc]

I'm sure there will be some unusual set of circumstances that may reveal why the system reacted as it did. Tesla had a similar issue when the forward cameras failed to detect a white-walled semi-trailer that pulled out, the Lidar didn't pick it up either because they were aimed too low. It resulted in changes to the software and hardware to avoid future instances of the same event.

As unfortunate as the circumstances are, I'm sure there will be a learning process and changes to avoid the situation again and this will be pushed out across the fleet. It's quite right that they have suspended the trial while they investigate the issues and produce a fix.

[JoeMorgue]

Originally Posted by Dr. Keith

Not to be flippant, but there's an XKCD for that.

I posted that exact same comic back on page 1.

______________________________________________

The problem is people are using Vaccine Denier logic.

"I don't care that X is many, many more times more effective in a risk/reward sense than Y, I won't accept it until Y is 100% safe" which doesn't make any kind of sense.

So many people just want to write a Luddite anti-tech narrative that lets them be the "I told you so guy" when something does go wrong.

And this is just a phase we go through with all new tech that I wish we could just skip.

[Skeptical Greg]

Video now being shown.

https://www.aol.com/article/news/201...-suv/23392088/

It's like she just pops up out of the road.

I see why it was hard to deal with..

[Elagabalus]

Originally Posted by Skeptical Greg

Video now being shown.

https://www.aol.com/article/news/201...-suv/23392088/

It's like she just pops up out of the road.

I see why it was hard to deal with..

Wow. Just. Wow.

The car didn't even deviate from its "look-at-me-I'm-a-self-driving-car" path. No swerve ... nothing.


Needs a bit of work I'd say.

[Meadmaker]

I think a human driver would have avoided that collision most of the time. The cameras didn't have the same dynamic range and exposure adjustment that your eye would have, so a human would see the shape of the bicycle and pedestrian before it was visible in the camera footage. The tennis shoes were visible a full two seconds before the crash.

There will be lots more after-action analysis done, to figure out exactly what the car identified as a moving object, and when, but to me this looks like a highly preventable accident.

The woman was clearly walking stupidly, and so the driver/car might not be found at fault, but a human driver would have missed her most of the time, assuming the driver was sober, alert, and not distracted.

[kedo1981]

Just watched the video, it was entirely the pedestrian's fault.
Crossing into traffic on a dark road, how stupid can you be?
Why would she do that, how couldn't she see the cars headlights coming

[a_unique_person]

Originally Posted by Meadmaker

I think a human driver would have avoided that collision most of the time. The cameras didn't have the same dynamic range and exposure adjustment that your eye would have, so a human would see the shape of the bicycle and pedestrian before it was visible in the camera footage. The tennis shoes were visible a full two seconds before the crash.

There will be lots more after-action analysis done, to figure out exactly what the car identified as a moving object, and when, but to me this looks like a highly preventable accident.

The woman was clearly walking stupidly, and so the driver/car might not be found at fault, but a human driver would have missed her most of the time, assuming the driver was sober, alert, and not distracted.

Not me. All I would have had time to do was jump on the brakes which is what the car appears to have done.

Suicide by car I'm guessing. She could see it from a long way away as it had it's lights on and walked right into it.

[Elagabalus]

Originally Posted by kedo1981

Just watched the video, it was entirely the pedestrian's fault.
Crossing into traffic on a dark road, how stupid can you be?
Why would she do that, how couldn't she see the cars headlights coming

And you want this tech to be viable in Europe too? You better be ready for some more of the same. I won't mention any countries by name.

[Elagabalus]

Originally Posted by a_unique_person

Not me. All I would have had time to do was jump on the brakes which is what the car appears to have done.

Suicide by car I'm guessing. She could see it from a long way away as it had it's lights on and walked right into it.

You would have just T-boned her? No steer to the right or left (50:50 chance!)? I find that very hard to believe.

[Meadmaker]

Originally Posted by a_unique_person

Not me. All I would have had time to do was jump on the brakes which is what the car appears to have done.

Suicide by car I'm guessing. She could see it from a long way away as it had it's lights on and walked right into it.

Not so. You would have had several seconds warning.

You are judging by what you see on the camera, but the camera does lie. The intensity of the headlights makes most of the vision wash out, so she is not visible until it's too late, but most people's eyes are better than that camera, and the motion of the feet, or the reflectors on the bicycle tire would have caught your attention and even if you didn't realize it was a person, you would have been primed to react when, two seconds before the collision, she came into actual view.

I've had people do that on Woodward Avenue, and very commonly on Ten Mile Road, and I've nearly hit them, but the key word is "nearly".

She was at fault, in a legal sense, but an alert human driver would not have killed her.

[smartcooky]

Originally Posted by Elagabalus

Wow. Just. Wow.

The car didn't even deviate from its "look-at-me-I'm-a-self-driving-car" path. No swerve ... nothing.


Needs a bit of work I'd say.

Very, very harsh.

There was no chance to swerve or stop. A human driver would not have even had time to react, let alone brake or swerve.

Originally Posted by Meadmaker

I think a human driver would have avoided that collision most of the time. The cameras didn't have the same dynamic range and exposure adjustment that your eye would have, so a human would see the shape of the bicycle and pedestrian before it was visible in the camera footage. The tennis shoes were visible a full two seconds before the crash.

There will be lots more after-action analysis done, to figure out exactly what the car identified as a moving object, and when, but to me this looks like a highly preventable accident.

The woman was clearly walking stupidly, and so the driver/car might not be found at fault, but a human driver would have missed her most of the time, assuming the driver was sober, alert, and not distracted.

100% disagree. She might just have well have materialised right in front of the car. I doubt there would be any driver who could have avoided hitting her.

Do the math

I downloaded the video (its 25 fps) and measured from the first frame where you catch a glimpse of the tennis shoes, to the moment they stopped the video just before impact. It is 43 frames = 1.72 seconds

At 40mph, the vehicle is travelling at 59 feet per second. That means from the moment the tennis shoes are spotted, the car must come to a complete halt in 101.5 feet.

Based on the results of many studies, Perception-Reaction Time (PRT)* is about 2.5 seconds across 90 percent of all drivers when confronted with simple to moderately complex highway situations. Greater reaction time should be allowed in situations that are more complex.

At 40 mph, a PRT of 2.5 seconds means the car has travelled 147.5 feet before the driver gets his foot on the brake (that is 46 feet PAST the pedestrian). A further 80 feet is required for a small car to brake to a stop from 40 mph.

Even if we use the very best PRT for perfectly ideal conditions, say 1.5 seconds, that is still 88.5 feet before the driver gets his foot on the brake that leaves 13 feet before the car hits the pedestrian. In the 80 feet braking distance past that, the car is still doing around 38mph when it hits the pedestrian.

I can confidently state that no driver could have avoid hitting the pedestrian at an almost certainly fatal speed.


*NOTE: PRT is NOT the same as (and should not be confused with) reaction time, which for humans is typically 0.3 to 0.5 seconds. PRT incorporates the average time for a person to react to an unexpected stimulus, recognise that it represents a danger, and take physical action e.g. foot on brake, turn steering wheel etc

[bruto]

I'm not sure. I doubt a human driver would have had time to do much, if in fact her vision had been the same as the camera's, but she might have been able to hit less squarely. And since, as pointed out above, we cannot entirely know what a human eye would have seen an alert driver might have seen more. In addition, the car is said to have radar, which ought to be a little better than that.

It may be that a driver could not have done better, but it's certain she could not have done worse.

[3point14]

Originally Posted by The Great Zaganza

I don't get some of the blowback here: Uber is pulling the cars to check if there is something to be done to avoid similar accidents; not because a human could have avoided it, but because in order to gain acceptance, autonomous cars must be safe both for the occupants and everyone outside.
If the engineers believed that the code was flawless already, there would have been no need for the test drives.

The report is that the pedestrian stepped in front of the car at such a time as the laws of physics did not allow the car to stop before hitting the pedestrian.

No human driver would have stopped in time, no driver at all would have stopped in time. If your bar for releasing any driver onto the road is that they would not have hit this pedestrian, then there are no drivers you would allow on the road.

[The Great Zaganza]

Originally Posted by 3point14

The report is that the pedestrian stepped in front of the car at such a time as the laws of physics did not allow the car to stop before hitting the pedestrian.

No human driver would have stopped in time, no driver at all would have stopped in time. If your bar for releasing any driver onto the road is that they would not have hit this pedestrian, then there are no drivers you would allow on the road.

luckily, that ins't my bar, and I already stated that I believe the fault is squarely with the pedestrian.

Nevertheless, it would be wrong not to try to learn from the accident and just accept that **** happens.

[3point14]

Originally Posted by The Great Zaganza

luckily, that ins't my bar, and I already stated that I believe the fault is squarely with the pedestrian.

Nevertheless, it would be wrong not to try to learn from the accident and just accept that **** happens.

If the laws of physics do not allow for the pedestrian to be detected in time for the car to stop, what is there to learn from this?

[Dave Rogers]

Originally Posted by 3point14

If the laws of physics do not allow for the pedestrian to be detected in time for the car to stop, what is there to learn from this?

Two quite important things:

(1) How close did the car's actual response approximate to the limit imposed by the laws of physics, and
(2) Is there any way it can be improved to approach that limit more closely?

Because the next time a pedestrian steps out in front of a self-driving car, it may just barely be possible to stop, and an improved control system may just barely be able to do so, thus saving a life. From an engineering perspective I suspect extreme cases like this are of critical importance because they can help to establish the performance limits of the system.

Dave