The self driving uber car killed someone.
Here is the vid
www.bbc.co.uk/news/world-us-canada-43497364
The question is, would you or i have done any better?
|
Not with the range of those headlights that's for sure. You couldn't see them up ahead in that video. The bike was fairly colourful too... Too many cyclists around here often in all black and hard to see.
The 'human monitor' wasn't much use if they weren't looking ahead either.
But these cars have LIDAR on the roof so you'd have thought that would detect the pedestrian sooner. Although weather does affect LIDAR for obvious reasons.
Last edited by: rtj70 on Thu 22 Mar 18 at 18:36
|
>>But these cars have LIDAR on the roof so you'd have thought that would detect
>> the pedestrian sooner. Although weather does affect LIDAR for obvious reasons.
>>
Yes, I can't work out why the LIDAR didn't see the pedestrian. The weather was fine (if dark). It seems a pretty simple scenario, but the car couldn't handle it. Back to the drawing board!
|
Hard to say isn't it, but given that the victim set off from the other side of the road, I hope so. Not having our eyes closed might have given us an edge.
The dynamic range of human sight is usually better than can be displayed on a screen.
But I don't think driving and avoiding accidents is all about reactions. Whilst it's nice to have a pilot or driver with lightning reactions and outstanding control skills to deal with trouble, I'd rather have one that avoided the trouble in the first place. That is to me a credible model for driving, but I don't think it works for 'minding'. By the time the minder has spotted the problem, it's usually going to be too late to intervene.
|
>> The dynamic range of human sight is usually better than can be displayed on a screen.
We are judging what could be seen from a dash cam - it might be poor in low light conditions. But you'd have thought Uber had that covered in these test vehicles.
Forward lighting seemed poor to me from the video. And although Uber had permission to test these vehicles and the human monitor might have been unable to do anything.... should they have not been looking ahead at all times.... They are there to intervene surely. I find that partly liable for this but the law might see it differently.
From a legal perspective Uber had a camera recording both the road ahead and the human monitor.
|
I think the conventional lighting we see is more about being seen than seeing. The LIDAR being used in the car should have a massively higher dynamic range than our eyes under any lighting conditions. There will be a lot of amplification noise and artifacts at work I guess.
Its an unfair question really I guess. as what we can see on screen does not fully represent what could be seen in real life
|
>> >> The dynamic range of human sight is usually better than can be displayed on
>> a screen.
>>
>> We are judging what could be seen from a dash cam
The best cameras, in given conditions, have a similar dynamic range/contrast ratio to human sight, but until we can have the camera data plugged into our brains we have to see the images on a screen.
The reason I mentioned it is that the driver would have been able to see more than we could see on the video, via a screen.
It's not really relevant to what the car could "see". If the car is processing visible wavelength and/or infra red images then potentially it has human-sight-levels of resolution, anyway, but I suspect two reasons they use radar and lidar is that using optical images involves huge amounts of data processing and doesn't provide accurate ranging.
The machines are not trying to mimic the way humans drive which is more based around pattern recognition that happens at a subconcious level - we can tell, for instance, when somebody is about to change lanes, walk into the road, or make a turn without being able consciously to analyse the process or explain how we knew; we can judge relative speed without knowing how fast we or the other vehicle are going. Given we weren't evolved for it, the fact that even idiots can do it passably most of the time is astonishing.
|
Optically, the uber car will outstrip us every time.
However, as you say its intelegence is far below ours when it comes to risk assessment, experience and imaginative variety of mitigation and avoidance.
I bet it hasn't even been programmed for the "we are going to crash, whats the least painful option"
|
There are cars that would have been able to avoid that 6-7 years ago, Audi A8 and S class with expensive options of Night Vision Assist with Pedestrian Detection, AEB and Pre-sense Plus.
Bad that a supposed car of the future which is ultimately supposed to 'go it alone' can't.
www.youtube.com/results?search_query=night+vision+assist++pedestrian
Last edited by: Shiny on Thu 22 Mar 18 at 23:34
|
>> >> >
>> Given
>> we weren't evolved for it, the fact that even idiots can do it passably most
>> of the time is astonishing.
>>
But surely we are perfectly evolved for it? That's precisely what our hunter ancestors would have done - watching animals, anticipating likely movements, interpreting tiny body movements of the animal being stalked? Even watching the movements and reactions of other animals not directly the target - birds disturbed, rubber-necking troupes of monkeys, etc.
We have a sub-conscious hunch about what a pedestrian might possibly be about to do just as we'd have sensed whether the mammoth was going to veer off suddenly, stop, or turn and charge.
It's a jungle out there :)
|
>> It's a jungle out there :)
I never thought of the M25 as the natural evolution from the East African rift valley, but now you mention it.
|
Is LIDAR same as what many cars call as milimeter wave radar, used for adaptive cruise control or autonomous emergency braking?
|
Not at all the same thing. And there are variations in LIDARs as well because of the channel numbers employed.
It uses lasers to map out the surrounding in 3D. LIDAR sensors are expensive. a 64 channel one will cost tens of thousands at the moment. Costs will probably drop.
The units need to be mounted high up - e.g. on the roof.
velodynelidar.com/hdl-64e.html
Last edited by: rtj70 on Fri 23 Mar 18 at 18:49
|
|
Currently cars are using forward facing camera (some using FLIR), radar sensors, and ultrasound, for autonomous assist functions, LIDAR is not far away, GM is preparing to add it to to their top of the range systems shortly, most car makers are gearing up for it. Mounting it at the top pf the windscreen is fine, does not need to be on the roof.
Last edited by: Zero on Fri 23 Mar 18 at 19:23
|
Like I said LIDAR should have spotted the women. As I thought the Uber car had a LIDAR system on the roof. It turns out it was one from the company Velodyne I mentioned:
www.bbc.co.uk/news/technology-43523286
Cars available now like the Tesla at the moment are not as far as I know using LIDAR. They are trying to use cameras.
A 16 channel Velodyne LIDAR system is now only $4000. A 64 channel much more. And at the moment they go on the roof to give the 360 degree view.
Last edited by: rtj70 on Fri 23 Mar 18 at 23:34
|
Update on the fatal collision. The car detected the woman pushing a bicycle, but decided it was a false positive.
It's obviously how these things have to work, but I doubt whether people generally will understand that they will have to be tuned for an acceptable compromise between evasive action for every bit of stray cardboard or pigeon and running down an acceptably small number of people, logical though it is.
www.theguardian.com/technology/2018/may/08/ubers-self-driving-car-saw-the-pedestrian-but-didnt-swerve-report
|
>> I doubt whether people generally will understand that they will have to be tuned for an acceptable compromise ....
Or that their performance should be compared to that of the real world of human piloted machines and not some imaginary nirvana of 0 accidents/injuries.
|
>> Or that their performance should be compared to that of the real world of human
>> piloted machines and not some imaginary nirvana of 0 accidents/injuries.
>>
or that there are as many philosophical questions as there are technical. What is an acceptable number of deaths / injuries per year worldwide.
|
>> or that there are as many philosophical questions as there are technical. What is an
>> acceptable number of deaths / injuries per year worldwide.
The number we have now, otherwise cars would not be allowed on the roads.
Last edited by: VxFan on Wed 9 May 18 at 09:25
|
>> or that there are as many philosophical questions as there are technical. What is an
>> acceptable number of deaths / injuries per year worldwide.
>>
>The number we have now, otherwise cars would not be allowed on the roads.
Not really, there is quite a difference.
We might quite understand that a human driver swerved to avoid a child and killed an old person. We might even think that was preferable. But we do not have to dictate that this is what will happen.
With a computer you will have to dictate that the computer will kill the old person to save the child.
This forces you to make a value judgement ahead of an incident, rather than merely being able to understand the judgement afterwards.
And that is emotionally quite different.
Added to that, with the human driver one person decides in the heat of the moment what will happen in a split second.
With the computer the entire population of the world gets to comment and pass opinion on what should happen from the comfort and sanity of their own chairs with all the time in the world, without having to accept any implications or fallout from that 'decision'.
|
>> This forces you to make a value judgement ahead of an incident, rather than merely
>> being able to understand the judgement afterwards.
Value judgements like that are made in all sorts of scenarios, both commercial and pubic. A life, and the loss of such does have a definable cost in all of them.
|
|
Apparently the deceased tested positive for methamphetamine and marijuana. So was the accident more her fault by crossing where it was unsafe and she was not that visible? Would a human have avoided her in time?
|
>> Apparently the deceased tested positive for methamphetamine and marijuana. So was the accident more her
>> fault by crossing where it was unsafe and she was not that visible? Would a
>> human have avoided her in time?
At what level was she positive? Traces or enough to be like a drunk?
My impression of vid is that 'driver' was not paying requisite degree of attention.
|
But the car was driving ;-)
No she wasn't paying attention considering her role. But there was 6 seconds max for the car to do anything and it decided hitting the obstacle was the thing to do. But maybe a drugged up road user in black pushing a bike across a dark bit of road was more to blame.
Maybe it's lucky it was an autonomous car because the woman could have been wiped out by anyone and there'd not be the same sensors/cameras to provide evidence.
|
|
If there had been no telematic evidence, no-one would have been bothered and put it down to the idiot crossing the road. Autonomous cars won't entirely save us from idiots
|
>> If there had been no telematic evidence, no-one would have been bothered and put it
>> down to the idiot crossing the road. Autonomous cars won't entirely save us from idiots
Absolutely right. And occasionally they will be idiots themselves, especially if they have to share a physical environment with people and dumb vehicles.
I am quite sceptical about driverless cars however. Whilst they will eventually be "perfected", I suspect they are a long way off that just now.
British roads (and most elsewhere) are less than perfect in terms of presence, consistency and condition of markings for example. And while ever driverless cars have to cope with roads also used by human drivers, the task is far more complicated.
Perhaps they can cope with 99% or maybe even 99.9% of the average journey, but the last 1% or 0.1% is very difficult. The idea that errors can be mitigated by the presence of an attendant is I think a bad one.
If the autonomous vehicle is just confused (can't resolve the data) of course it can stop, and the attendant can solve the problem. But if the error is not recognised then I think it will usually be too late, unless the attendant is monitoring at the level of a driving instructor in a dual-controlled car with a novice driver. Does any driving instructor think that is easier or safer than driving themselves, particularly at speed? I doubt it.
In practice the attendant will not monitor so closely, with feet (and hands) hovering over the controls. Had the Uber driver seen the lady with the bike, would she have done anything? Initially, she wouldn't - because she has seen hazards before, and simply waited for the car to deal with them. By the time it is apparent that the car isn't doing that, the moment to intervene has passed.
The firms developing autonomous cars have far more data available and employ better brains than mine. But a feature of massive projects is a kind of tunnel vision. The owners want it to happen, if they stop they take a massive loss, and the technocrats are very invested in telling them that their problems are solvable. Rarely are problem projects stopped at the point where realisation should have dawned; they usually soldier on until the money runs out or multiples of the original budget have been spent.
This will get done eventually but the end result might not be what we imagined and it could take a lot longer.
Last edited by: Manatee on Fri 25 May 18 at 10:17
|
>>British roads (and most elsewhere) are less than perfect in terms of presence, consistency and condition of markings for example.
On the A307 alongside our fantastic cycle highway lane markings have been removed to encourage drivers to drive more slowly.
tinyurl.com/y7b7fxo3
|
The Uber Crash - rtj70
