First pedestrian death by autonomous car...

[Marco]

Only thing NYT missed was the photograph showing evidence of impact about 20 meters back from where the car stopped. Still props on them for actually doing some journalism, best of the bunch.

My hypothesis, the programmers made the same mistake drivers often make. Leaving next to no margin for error for vehicles moving at right angle to themselves slowing down. A human driver will tend to be a bit more apprehensive of a squishy target though, and are aware that people slow down to get a bicycle up a curb.

[orion242]

gotta be a nightmare to sort out nonsense from something that may move in the path at any minute.

[thm_w]

According to this she walked right in front of the AVs path of travel and it never slowed down.

https://www.nytimes.com/interactive/2018/03/20/us/self-driving-uber-pedestrian-killed.html

Being equipped with both lidar and radar, it shouldn't have had much trouble spotting her.  Especially since she cross right in front!  Looking at it more, there are street lights on that portion of the road.  You can see their shadows on the road in their picture.  Given that, the headlights...even the damn person behind the wheel should have noticed her.  Again, if your used to the AV driving 99% of the time...how vigilant are you really for the 1% when the unexpected comes into the equation?

Yeah thats a bit strange, cars already have anti-collision features but I guess they don't work as well as we think.
Although they are still massively beneficial (rear end collision reduced by ~50% or something).

youtube.com/watch?v=zAeEnLr3WYk (similar event to what seems to have happened here)
youtube.com/watch?v=ooOz69CEtJ0 (test which seems to show it not activating)

[apelly]

There have been some good points raised about safety stats. I'd still like to see some apples to apples comparisons.

We know where all these AVs have driven, for how long, and at what times. Can we compare average incidents for people driving in the same places for the same amount of time at the same times of day?

[MT]

I just gently wonder if the software coders for autonomous diesel cars trucks etc have included runaway engines
particularly those who does that while transporting someone? Measuring RPM and throwing in neutral and brake
shure but turning off the engine? Some robot arm throwing in old socks into the turbo inlet?

[MT]

And what about famous Swedish moose test is that in the software as well? Well it applies to Canadians too... and Russia but they cant drive anyway..

[drussell]

Different result with purely human driver...  no.....

Really?  What are you basing that on? 

[EEVblog]

Seems they have released footage
If Chris Anderson (who is now big into autonomous cars) is saying it's not good, it's really not good.



Video link:
https://twitter.com/TempePolice/status/976585098542833664/video/1

[orion242]

Well at least they have quality people at the wheel of these "safer" AVs.  I'm sure the feedback they get on how the drive went is really useful input to the ML.

https://nypost.com/2018/03/20/ex-con-was-behind-wheel-of-self-driving-uber-that-killed-woman/

Anyone paying attention should have been able to stop.  As I said, they get complacent, they don't pay any attention because it works most of time.  That driver's eyes are anywhere but the road.  Now the fun begins....who's fault??


...trying not also point out "she" needs a serious shave in that photo...

[Halcyon]

Having now seen the video, 4 things are immediately evident to me.

1. The human driver wasn't watching the road until the time of impact.
2. The pedestrian was crossing in a poorly lit location, where she shouldn't have been, wearing dark clothing, ignoring on-coming traffic (for whatever reason).
3. The pedestrian was not visible until she was within range of the car headlights, by which time, reacting and stopping @ 60km/hr is not possible without colliding with the subject.
4. The other autonomous sensors (radar, lidar etc...) have failed to detect the subject.

Who is at fault? In my opinion 50% the pedestrian, 25% the human driver, 25% technology.

[BrianHG]

Looking at the video, yes, you cant see the pedestrian until the headlights illuminate her.
But, if the car had the latest high quality high speed 3D lidar i've seen on a BBC documentary related to the future of self driving cars, that pedestrian should have been easily picked up at 5x-10x the distance, even around that curve, that is also assuming that the AI sees the velocity and direction of that pedestrian and foresees the eminent collision.  Though, the particular lidar being demonstrated is just too expensive for now.  Human life costs less.

We will be stuck with self driving cars which have good technology, and, self driving cars with crap technology.  When deaths happen, public opinion wont care if the card used technology A vs technology B.

[raptor1956]

OK, so the video puts this into a much different situation than first suggested.  Seems to me any number of sensors should have detected her and with more than enough time to react/stop.  The cameras might not have been able to see her given the light level, but radar and lidar must certainly should have.  And, I'll repeat, IR is the way to go for detecting people/animals and in my opinion IR needs to be one of the sensors that become mandatory for self driving cars.

The safety driver thing is more of a liability provision as a human driver will soon lose attention if they are not in control -- I can not imagine a job where I'm required to do nothing 99.9% of the time yet be ready to grab control and have both situational awareness and an appropriate response within a fraction of a second.  Given the time when the woman was first noticeable in the video and the point of impact was not enough time given the speed and reaction time.

What no one has said yet as best I can tell is comment on the fact that this is Uber -- one of the if not THE most douche-baggy companies on the planet.  They stole much of the tech for autonomous cars from Waymo (Google) and that's hardly the worst of there corporate sins.  Waymo is likely to win some kind of settlement against Uber but the damage done because of this death may cost everyone working on this big time.  The upside, however, and it's hard saying that given the fact someone died, is that perhaps everyone in this space takes things more seriously.  One gets the sense that many Silicon Valley tech types think they are smarter than anyone and should be allowed greater license than others -- perhaps this rests things, I have my doubts.


Brian

[Brumby]

Who is at fault? In my opinion 50% the pedestrian, 25% the human driver, 25% technology.

I might question the human occupant component.  I hesitate to call them the driver, since they were not actively engaged in the driving process and, as such, would have not been in a position to have executed any actions of a driver within any useful time frame.

My main reasons for this include:
 1. As a passive occupant of the vehicle, they would not have engaged in any normal driver activity, such as a constant state of situational awareness expected from a "hands on" driver.
 2. As well as their eyes, their hands and feet were not in contact with or positioned for use of the controls.  These conditions simply could not occur with a normal "hands on" driver situation.
 3. Previous experience of the autonomous vehicle would have trained them in the lack of the need for doing so.

There may be some others, but considering only 1. and 2. above - this is what would have needed to happen.....
 a) The "safety driver" would need to have observed the hazard before impact (which I think they may have). 
 b) They would immediately expect the AV to have responded and they would hesitate.
 c) When the AV hadn't responded, they then would have had to remember that they had the ability to take control.
 d) They then would have been challenged by the taking the decision to do so.
 e) Once they had done that, they would have had to "find" the controls - the steering wheel and brake pedal as a minimum, since their hands (certainly) and feet (possibly) were not in normal driver position.
 f) They then would apply whatever action they chose.

For a normal "hands on" driver, only steps a) and f) are involved - the others are not automatic steps and they will take a finite time to be processed.

While a "hands on" driver would have been in a far better position for noticing the pedestrian in the first place, whether they could have been more effective in lessening the severity of the outcome is questionable.

[Brumby]

You cannot be serious.

Don't ever get into the legal profession.  A first year law student would carve you up like a Sunday roast.

[Halcyon]

I might question the human occupant component.  I hesitate to call them the driver, since they were not actively engaged in the driving process and, as such, would have not been in a position to have executed any actions of a driver within any useful time frame.

I dare say the law will view that differently, however I'm admitting I don't know the laws in the area where this crash occurred. I'm guessing that because there was a human inside the car, it was required by law and the law will probably view that person sitting in the driver's seat as the "driver", regardless if they were actually physically controlling the movement of the vehicle at that particular time.

However that being said, it's clear that most of the blame can be attributed to the pedestrian. I think the only thing the driver could have reasonably done (even if he was paying full attention) would be to swerve in an attempt to avoid the collision.

Also, on another point, since this was essentially a "test vehicle", I'm curious whether or not Volvo's own safety systems (city safety etc...) were disabled and the control of the car was relying solely on third-party sensors, cameras, software etc... Volvo's technology (whilst still being a computer and not 100% fool proof) is designed to detect and avoid exactly this kind of collision. Those cars are also fitted with pedestrian air bags which lift the bonnet and deploy around the windscreen. There did not appear to be any evidence of this system activating.

[Brumby]

The law as written may have clear directions, but I will suggest it has no consideration for the "driver's" particular circumstances.  The case may go to court based on that, but I doubt proceedings will follow anything like they would for a normal driver situation.

I daresay the items b) through e) I listed would feature strongly - plus whatever instruction and/or direction Uber gave the safety driver.  I'm sure you've seen some colourful courtroom calisthenics using less than that.


What is of greater interest, though, is why the technology did not intervene.  That is the mystery.

[Brumby]

Also, on another point, since this was essentially a "test vehicle", I'm curious whether or not Volvo's own safety systems (city safety etc...) were disabled and the control of the car was relying solely on third-party sensors, cameras, software etc... Volvo's technology (whilst still being a computer and not 100% fool proof) is designed to detect and avoid exactly this kind of collision. Those cars are also fitted with pedestrian air bags which lift the bonnet and deploy around the windscreen. There did not appear to be any evidence of this system activating.

Now that is an interesting question.

[Halcyon]

The law as written may have clear directions, but I will suggest it has no consideration for the "driver's" particular circumstances.  The case may go to court based on that, but I doubt proceedings will follow anything like they would for a normal driver situation.

I daresay the items b) through e) I listed would feature strongly - plus whatever instruction and/or direction Uber gave the safety driver.  I'm sure you've seen some colourful courtroom calisthenics using less than that.


What is of greater interest, though, is why the technology did not intervene.  That is the mystery.

Indeed and you are absolutely correct. This isn't a typical case, but the courts have to apply "typical" law, perhaps with a dash of case law. Will there be changes to come out of this? You bet! Will the human driver be found at-fault? If I were a betting man, I'd probably say no.

(And yes, some of my court matters have been... interesting... but mostly hours of boredom.)

[station240]

What I see is the pedestrian isn't looking in the car's direction right up to the point of impact. If that wasn't the case, she would have seen the car and stopped in the middle of the road.
The car's headlights would have been visible to her, long before she was visible to them.

Here is the first view from the camera that gives a hint there is something on the road.

This is 2 second before impact.

Was it possible to stop the car given this visual information
http://www.peterblight.com/Documents/eDS%20Stopping%20Distances.pdf
40 mph (64 km/h)
thinking distance: 12M
braking distance: 24M
Stopping time for ABS car is 2 seconds, add 50% for a human driver to react (+1 second).

If this wasn't a self driving car, the headline would be "Cyclist killed after walking in front of car". There is nothing a human driver could have done to prevent this accident, only reduce the crash speed.

So why didn't the autopilot software detect her using Lidar ? I'd guess they did, but are trained to accept cyclists as other road users that obey road laws. Perhaps the concept of bike + human walking it isn't in the software. What could be done to improve the software, basic face recognition to detect when fools aren't looking before crossing the road. I'd also think if the Lidar did see something moving, it should switch the headlights to highbeam so the camera can get a closer look.

[BrianHG]

What I see is the pedestrian isn't looking in the car's direction right up to the point of impact. If that wasn't the case, she would have seen the car and stopped in the middle of the road.
The car's headlights would have been visible to her, long before she was visible to them.

Here is the first view from the camera that gives a hint there is something on the road.

This is 2 second before impact.

Was it possible to stop the car given this visual information
http://www.peterblight.com/Documents/eDS%20Stopping%20Distances.pdf
40 mph (64 km/h)
thinking distance: 12M
braking distance: 24M
Stopping time for ABS car is 2 seconds, add 50% for a human driver to react (+1 second).

If this wasn't a self driving car, the headline would be "Cyclist killed after walking in front of car". There is nothing a human driver could have done to prevent this accident, only reduce the crash speed.

So why didn't the autopilot software detect her using Lidar ? I'd guess they did, but are trained to accept cyclists as other road users that obey road laws. Perhaps the concept of bike + human walking it isn't in the software. What could be done to improve the software, basic face recognition to detect when fools aren't looking before crossing the road. I'd also think if the Lidar did see something moving, it should switch the headlights to highbeam so the camera can get a closer look.

1.  That's the dash cam, not the higher quality camera shots from the AI's overhead cameras.
2.  Even with Dave's video he just posted, turn up you monitor's gamma control, the pedestrian is visible far earlier, if the image was electronically analyzed.  The screenshot image you shown here has the black level crushed out compared to the youtube video.  Not only that, but after 2 generations of compression, the original most likely has a slightly mode detail, not to mention the onboard AI cameras are operating raw uncompressed.

Here is when we first see the pedestrian.

Since she is already in the middle of the right lane and this far back, you can steer to the right.  If you look at the car in the NYT images, she was struck on the right hand bumper of the car, from this far back in Dave's video screenshot, it may have been possible to turn left and miss her, or, not make a killing blow.

Even further back going 1 mpeg generation earlier (twitter source):

If it wasn't for the .mpg maybe even further back, except, the pedestrian is already getting fairly tiny even at this distance...
Though, that pitch black long sleeve shirt in the middle of the night isn't the smartest attire for night cycling.

[Brumby]

I'm not sure if your post was directed at me.

It certainly was.  Your argument is idealistic.  It completely ignores the practical circumstances.

You are no doubt going to stand by your opinion, no matter what I think, so I'm not going to say anything more about it.  You have just as much right to be wrong as I.

[orion242]

Dash cams are total crap in the dark like any other cheap camera.  Its going to look much darker than it really is.

The road had streetlights at regular intervals.  Hit google street view.  They are all along both sides of the road.  That area is peppered with them.  Look at the news photos of the car that night and the shadow under the car.  Sure there are red/blue lights from the police, but there sure looks like a far bit of orange glow from street lights there... The speed limit was 45mph, not sure why some of the stories say 35mph.  Took 2 min on street view to spot the sign right before the overpass...

Also focusing on just stopping distance is nonsense.  That person hit the right side of the car.  12" could have been the difference of life or death.  Any human in full control would have had an instinctual reaction to turn left and stop.   Both of which likely would have saved this women's life.  It was a two lane road and the AV was the only car there, giving plenty of room to safely avoid her without leaving the roadway. 

The concept of the "safety" driver is totally BS.  No human is going to pay attention in this role.  The AV needs to be 100% or its crap.

[orion242]

And to get an idea how far away she is in the crap video....

AZ center line striping appears to be 6" wide strip 10' long with 30' spacing.  So she's got to be around 60' away when her shoes appear in the grainy video.

https://azdot.gov/business/engineering-and-construction/traffic/signing-and-marking-standard-drawings/current-test

[raptor1956]

Dash cams are total crap in the dark like any other cheap camera.  Its going to look much darker than it really is.

The road had streetlights at regular intervals.  Hit google street view.  They are all along both sides of the road.  That area is peppered with them.  Look at the news photos of the car that night and the shadow under the car.  Sure there are red/blue lights from the police, but there sure looks like a far bit of orange glow from street lights there... The speed limit was 45mph, not sure why some of the stories say 35mph.  Took 2 min on street view to spot the sign right before the overpass...

Also focusing on just stopping distance is nonsense.  That person hit the right side of the car.  12" could have been the difference of life or death.  Any human in full control would have had an instinctual reaction to turn left and stop.   Both of which likely would have saved this women's life.  It was a two lane road and the AV was the only car there, giving plenty of room to safely avoid her without leaving the roadway. 

The concept of the "safety" driver is totally BS.  No human is going to pay attention in this role.  The AV needs to be 100% or its crap.

Actually I think the data indicates that most people, if they react by turning at all will tend to turn into the direction the object is moving and increase the likelyhood of impact.  Human nature is weird that way.  Similarly, something like 70% of people being passed tend to increase there speed while being passed when they should keep the same speed or even slow a bit.  Again, human nature is weird.

That said, it is certainly true that if the driver had been paying attention and turned even a little to the left they could have avoided impact even without slowing.  I was driving my Nissan 240SX pretty fast one night on a country road down a small mountain and just after turning into a right hand turn I found a large Deer in the middle of the road with it's head to the right -- I managed, barely, to swerve left then hard right to make the turn and just missed hitting the Deer by angstroms.  I was going too fast with too little distance to stop in time so swerving was the only option -- thank goD no one was coming the other way.

Expecting a human driver to remain 100% attentive behind the wheel of a self driving cars is folly and I'd bet the percentage of cases where a human driver might react to prevent a problem and actually does is less than 30%.  The fact that the driver appears distracted will likely cause that person legal grief and sadly few people will give a damn about him.  Again, the reason there are humans in the car has more to do with liability insurance and convincing regulators that its not just tech.  In other words, a bit of legal slight of hand...


Brian

[alexanderbrevig]

Can't be bothered to read every page but try making a statistics of deaths in traffic by human drivers and shut the fuck up. That was all. Thanks for your attention.