Author Topic: EEVblog #1066 - Uber Autonomous Car Fatality - How?  (Read 60005 times)

0 Members and 1 Guest are viewing this topic.

Offline Fungus

  • Super Contributor
  • ***
  • Posts: 16628
  • Country: 00
Re: EEVblog #1066 - Uber Autonomous Car Fatality - How?
« Reply #50 on: March 22, 2018, 01:14:38 pm »
Google are nowhere near ready to deploy these cars on the street.

Not even Google says they are. Google is currently aiming at "2020".

no software QC, no dynamic testing with actual moving targets that need to be tracked.

Really?  :palm:
 

Offline Monkeh

  • Super Contributor
  • ***
  • Posts: 7992
  • Country: gb
Re: EEVblog #1066 - Uber Autonomous Car Fatality - How?
« Reply #51 on: March 22, 2018, 01:26:18 pm »
I am not sure that you can compare our roads and traffic to your little country

Uh, why not? They're roads, with supposedly educated adult drivers using them. Ours kill fewer people per capita and per car.. and per mile, too. Why can't we compare them, because we're too 'little'?

Quote
and I am not sure about your point

I was disagreeing with your opinion that there are few fatalities. There are many more than ours and ours are higher than they need to be. You seem to have taken this personally.
 

Online coppice

  • Super Contributor
  • ***
  • Posts: 8632
  • Country: gb
Re: EEVblog #1066 - Uber Autonomous Car Fatality - How?
« Reply #52 on: March 22, 2018, 01:40:41 pm »
Google are nowhere near ready to deploy these cars on the street.
Not even Google says they are. Google is currently aiming at "2020".
So you consider 21 months to be long term?
 

Offline Fungus

  • Super Contributor
  • ***
  • Posts: 16628
  • Country: 00
Re: EEVblog #1066 - Uber Autonomous Car Fatality - How?
« Reply #53 on: March 22, 2018, 01:45:22 pm »
In general: Not so much. I bet you can find examples where this exact set of inputs is perfectly Ok (right up until it's too late).
This is where I would ask for examples - but I cannot see you succeeding.

That's because you're only thinking of ordinary examples of everyday traffic. You need to think of really mad things, just like what happened here (seriously, what was going through her head?)

My thoughts on what happened are that somewhere in the software there has to be a variable:

float probabilityOfPedestrianWalkingAcrossTheRoadWithoutLooking = 0.5;

In the city you set it to quite a high value.

On a highway/freeway/motorway? A lower value - you don't dozens of cars to be emergency braking from 55mph every time a shrub grows towards the road.

Bottom line:

This sort of thing was bound to happen. This is very complex software, it was 100% likely that there would be a death (or three) through software bugs.

Let's not kid ourselves though: The road death rate is going to drop sharply in a few years time and it will be thanks to these cars.

(no, it won't ever be 0%, but that's no reason not to do it).

« Last Edit: March 22, 2018, 02:06:26 pm by Fungus »
 

Online coppice

  • Super Contributor
  • ***
  • Posts: 8632
  • Country: gb
Re: EEVblog #1066 - Uber Autonomous Car Fatality - How?
« Reply #54 on: March 22, 2018, 01:46:19 pm »
In the video the cyclist appears from the darkness at the very last moment, as though the dipped headlights were dipped way too far. If a human were driving on that dark road they would have had their high beams on, and it looks like they would have seen the cyclist reasonably early, when it was still realistic to stop. I don't think the "human driver would have hit this cyclist anyway" argument, which I have seen, holds water.
 
The following users thanked this post: drussell

Offline Fungus

  • Super Contributor
  • ***
  • Posts: 16628
  • Country: 00
Re: EEVblog #1066 - Uber Autonomous Car Fatality - How?
« Reply #55 on: March 22, 2018, 01:51:07 pm »
Google are nowhere near ready to deploy these cars on the street.
Not even Google says they are. Google is currently aiming at "2020".
So you consider 21 months to be long term?

a) Nobody said it was a fixed date.

b) The testing is increasing exponentially:


c) The simulated testing is increasing at an even faster rate than that! Google drove 2.7 billion miles in simulators last year.

(Simulators are where people dream up scenarios like this one and feed the input to virtual cars)

d) Maybe they meant December 2020, that's nearly 33 months!!
« Last Edit: March 22, 2018, 01:53:26 pm by Fungus »
 

Offline mdijkens

  • Regular Contributor
  • *
  • Posts: 146
  • Country: nl
Re: EEVblog #1066 - Uber Autonomous Car Fatality - How?
« Reply #56 on: March 22, 2018, 01:52:28 pm »
The fact that this DASHCAM video shows such short headlight-beams is proof that this camera has auto-adjusted the exposure for that.
The headlight-beams of an XC90 are much further but not visible here because of that auto-exposure.

It makes me feel much more confident that a person would have seen much more, much sooner than this video shows...
 
The following users thanked this post: SilverSolder, drussell, TheDane

Offline Fungus

  • Super Contributor
  • ***
  • Posts: 16628
  • Country: 00
Re: EEVblog #1066 - Uber Autonomous Car Fatality - How?
« Reply #57 on: March 22, 2018, 01:55:05 pm »
It makes me feel much more confident that a person would have seen much more, much sooner than this video shows...

...assuming they weren't looking down at their phone or something.
 
The following users thanked this post: TheDane

Offline AF6LJ

  • Supporter
  • ****
  • Posts: 2902
  • Country: us
Re: EEVblog #1066 - Uber Autonomous Car Fatality - How?
« Reply #58 on: March 22, 2018, 02:10:54 pm »
In the video the cyclist appears from the darkness at the very last moment, as though the dipped headlights were dipped way too far. If a human were driving on that dark road they would have had their high beams on, and it looks like they would have seen the cyclist reasonably early, when it was still realistic to stop. I don't think the "human driver would have hit this cyclist anyway" argument, which I have seen, holds water.

Agreed...
Sue AF6LJ
 
The following users thanked this post: drussell

Offline Monkeh

  • Super Contributor
  • ***
  • Posts: 7992
  • Country: gb
Re: EEVblog #1066 - Uber Autonomous Car Fatality - How?
« Reply #59 on: March 22, 2018, 02:20:47 pm »
The video is deceiving.

It's a lit road in a semi-built area immediately approaching a 4-way light-controlled junction. Are you sure you'd be using full beam?
 

Online coppice

  • Super Contributor
  • ***
  • Posts: 8632
  • Country: gb
Re: EEVblog #1066 - Uber Autonomous Car Fatality - How?
« Reply #60 on: March 22, 2018, 02:32:28 pm »
The video is deceiving.

It's a lit road in a semi-built area immediately approaching a 4-way light-controlled junction. Are you sure you'd be using full beam?
Someone else said the dynamic range of the camera might be distorting the situation, and making a well lit spot look dark. Whether lit by the street lighting or lit by the car's headlamps, a human driver should have been able to see the cyclist long before we see them in the video. This is a perfectly normal night driving with careless cross traffic scenario, and we rarely have a problem with it.
 
The following users thanked this post: drussell

Offline AF6LJ

  • Supporter
  • ****
  • Posts: 2902
  • Country: us
Re: EEVblog #1066 - Uber Autonomous Car Fatality - How?
« Reply #61 on: March 22, 2018, 02:34:50 pm »
The video is deceiving.

It's a lit road in a semi-built area immediately approaching a 4-way light-controlled junction. Are you sure you'd be using full beam?
Yes.
Sue AF6LJ
 
The following users thanked this post: drussell

Offline CNe7532294

  • Regular Contributor
  • *
  • Posts: 108
  • Country: us
Re: EEVblog #1066 - Uber Autonomous Car Fatality - How?
« Reply #62 on: March 22, 2018, 02:59:32 pm »
Overall you misunderstood my post. I'm not for shutting down the tech. I laugh at anyone thinking this is the end of automated driving. However, I'm against shutting down the extra layer of protection called human action. A quick search on piloting will turn up pilots work in tandem with the autopilot. Tech is not there to make people lazy. As I said, only there to make things easier. There is a difference between those things. Also if you notice, planes try to have 2 or more things for a reason (ie. two engines, two pilots along with the autopilot, two sets of gauges, 2 radio sets, 2-3 hydraulic systems, 2-3 fuel lines with crossfeed valves, etc.). Its called being redundant. It works. :palm:

I have no proof in saying this exact part here but there is no doubt in my mind yet uber reps told the driver everything is fine. This car drives itself. Causes her to lower her guard. Pays attention to the phone. We all know the rest. :palm: If you're by yourself and need to sleep, drunk, or pay more attention to your phone, take a real taxi or bus please. Basically saying err on the side of caution. This is yet another case to build on this wise saying. Certainly won't be the last. :rant:

To put it simply, your implication is impractical.  Put anyone in a car that drives itself and after a while, EVERYbody is going to drop their attention at times.  AS IT IS, we can't even be assured that people who have full driving responsibilities will pay attention.  Just look at the issues arising from mobile phones.

Anyway, that's not the topic of this thread.

Full automation of anything when lives are on the line is impractical in and of itself. Actually let me rephrase this. Having the computer bare all the work load is impractical. I also say again, having the best tools, like a computer, doesn't excuse one from using their brain. Dropping attention is negligence at its finest. :palm:

Automation will continue to exist just not in full form. At least not until we develop a system that separates people from vehicles. Actually thats exactly why we have trains. We also have ABS. But these are besides the point. Onto the my main point. Again. |O

As for topic, this is highly related plus Dave clearly titled it "How?" not specifically "How did those fancy sensors fail?". |O It would be very unwise not to look at something that already has experience before. Its disappointing that you would ignore this. You might as well just ignore what the NTSB, BEA, AAIB, ATSB and other transportation boards has to say. As the saying goes, "ignore history, history repeats". Again, the airline industry already has the answers working with flight computers.

I highly encourage you to talk to a Quantas pilot, flight engineer, or mechanic at the very least. I also would like to see Dave follow thru on this topic by having a discussion with Quantas. If you can't meet up with one I suggest you at least look and compare air disaster cases. I'll give you a start. Air France Flight 447. This is just one of many where either the sensor or flight computer failed while the pilot put in full trust into a faulty auto system. End result, many people died.

As for solutions and recommendations, I have one. Make not only a detection system for the computer but an independent detection system for the driver. Have it vibrate (like a stick shaker during a stall) or alarm (like TCAS). Put it on the phone even. Have it activate when the cross-sectional heat of a poodle comes across a separate FLIR camera feed. At least make the driver aware. It wouldn't stop the accident but she could have swerved to the left or slowed down. Could have left the victim live with injuries instead of I assume flying several yards along with her bike. Redundancy works.
 

Offline Decoman

  • Regular Contributor
  • *
  • Posts: 161
  • Country: no
Re: EEVblog #1066 - Uber Autonomous Car Fatality - How?
« Reply #63 on: March 22, 2018, 03:09:54 pm »
The driver of the car (heh) is OBVIOUSLY to blame for hitting this pedestrian walking across the road. It really does look like the car had NOT long range lights enabled (unsure what this is called in English).

Here in Norway, it is required that you expertly keep switching between short range lights and long range lights to actually get to see where you are driving at night (because you must not end up blinding other cars passing you in the other lane, opposite of your direction).

The type of driving seen in the footage, seem to be as reckless as the hilarious car accidents in Hollywood movies where cars collide into things at speed, or failing to lower speed when spotting a potential obstacle ahead.

This death is imo something that would have been prevented by a moderately skilled driver.

When learning to drive a car, it is taught that one must always adjust the speed to suit the environment. Obviously, no driver should ever be allowed to drive at speed across a dark road:
1) When being unable to see well ahead of the car
2) When being unable to come to a stop within a reasonable range relative to the conditions of the environment

This is imo obviously the driver's fault. This kind of collision is precisely what is taught as what will happen if you are at speed in a car, while having limited view ahead. If you can't see far ahead (and also the entire width of the road), you must drive more slowly to suit the conditions of the environment. Anything else is just crazy, if you are not willing to maintain the minimum amount of control over a vehicle that is otherwise required to safely and responsibly operate a motorized vehicle.
« Last Edit: March 22, 2018, 03:19:18 pm by Decoman »
 

Offline Kalvin

  • Super Contributor
  • ***
  • Posts: 2145
  • Country: fi
  • Embedded SW/HW.
Re: EEVblog #1066 - Uber Autonomous Car Fatality - How?
« Reply #64 on: March 22, 2018, 03:19:00 pm »
It would be most interesting to see the sensor data from the car. What did the car detect? Hopefully they will release the data at some point, not just this video footage. Without sensor data this video is quite useless.
 

Offline Decoman

  • Regular Contributor
  • *
  • Posts: 161
  • Country: no
Re: EEVblog #1066 - Uber Autonomous Car Fatality - How?
« Reply #65 on: March 22, 2018, 03:20:29 pm »
Without sensor data this video is quite useless.

The video shows the car to be driving on a road having no visibility ahead, and with no long range lights enabled, can we agree on this?
I also wonder if you drive a car yourself. :| Bonus question. I can drive most large vehicles myself.
« Last Edit: March 22, 2018, 03:24:13 pm by Decoman »
 

Offline mdijkens

  • Regular Contributor
  • *
  • Posts: 146
  • Country: nl
Re: EEVblog #1066 - Uber Autonomous Car Fatality - How?
« Reply #66 on: March 22, 2018, 03:22:50 pm »
The video shows the car to be driving on a road having no visibility ahead, and with no long range lights enabled, can we agree on this?
I also wonder if you drive a car yourself. :| Bonus question.

No  8)

The fact that this DASHCAM video shows such short headlight-beams is proof that this camera has auto-adjusted the exposure for that.
The headlight-beams of an XC90 are much further but not visible here because of that auto-exposure.
 

Online coppice

  • Super Contributor
  • ***
  • Posts: 8632
  • Country: gb
Re: EEVblog #1066 - Uber Autonomous Car Fatality - How?
« Reply #67 on: March 22, 2018, 03:25:00 pm »
Without sensor data this video is quite useless.

The video shows the car to be driving on a road having no visibility ahead, and with no long range lights enabled, can we agree on this?
I also wonder if you drive a car yourself. :| Bonus question.
If you've read through the comments the answer is clearly no, we can't agree on that. The darkness may be entirely due to the exposure setting on the camera, while the road was well lit with street lamps.
 

Offline Decoman

  • Regular Contributor
  • *
  • Posts: 161
  • Country: no
Re: EEVblog #1066 - Uber Autonomous Car Fatality - How?
« Reply #68 on: March 22, 2018, 03:28:02 pm »
I wonder if these cars have a diagnostic system running, because if a sensor fails gracefully and stealthily, then that would be bad I think.
 

Offline mdijkens

  • Regular Contributor
  • *
  • Posts: 146
  • Country: nl
Re: EEVblog #1066 - Uber Autonomous Car Fatality - How?
« Reply #69 on: March 22, 2018, 03:29:28 pm »
It might even be that there was so much light that any human driver could have seen the victim from 100ft/300m

We can't tell based on the DASHCAM footage...
 

Offline Decoman

  • Regular Contributor
  • *
  • Posts: 161
  • Country: no
Re: EEVblog #1066 - Uber Autonomous Car Fatality - How?
« Reply #70 on: March 22, 2018, 03:30:51 pm »
Protip: When using long range lights on the road, you actually see where you are driving. To think that a camera could even auto adjust that away in a recording seems bizarre imho.
 

Offline CatalinaWOW

  • Super Contributor
  • ***
  • Posts: 5224
  • Country: us
Re: EEVblog #1066 - Uber Autonomous Car Fatality - How?
« Reply #71 on: March 22, 2018, 03:35:18 pm »
The only thing that is clear here is that a lot of technical people are willing to speculate on facts and develop firm opinions based on those speculations.  If this thread's authors are representative of those writing the self driving software it does not bode well for the results.
 

Offline Kalvin

  • Super Contributor
  • ***
  • Posts: 2145
  • Country: fi
  • Embedded SW/HW.
Re: EEVblog #1066 - Uber Autonomous Car Fatality - How?
« Reply #72 on: March 22, 2018, 03:38:40 pm »
Without sensor data this video is quite useless.

The video shows the car to be driving on a road having no visibility ahead, and with no long range lights enabled, can we agree on this?
I also wonder if you drive a car yourself. :| Bonus question.
Sure, we can agree on that, of course. Long range lights may have helped in this situation, but in the city area using long lights is not always possible due to other traffic or pedestrians. In this situation human vision might have been somewhat better than the video camera (cannot know how much better, this is just speculation), but nevertheless the self-driving cars are using sensor data to detect obstacles and other vehicles / pedestrians / just name it, so the sensor data is the key here.

Sure, I drive. We have dark fall/winter when there is no snow yet, and driving in the poorly lit areas is pretty demanding, especially when the roads are wet and absorb all light - or if someone is driving towards me with headlights reflecting from wet asphalt straight into eyes. Hitting a moose in the country side in dark roads is not uncommon either. Quite often I do hope to have some sensor fusion available.
 

Offline mdijkens

  • Regular Contributor
  • *
  • Posts: 146
  • Country: nl
Re: EEVblog #1066 - Uber Autonomous Car Fatality - How?
« Reply #73 on: March 22, 2018, 03:40:22 pm »
Compare with this:
 

Offline CNe7532294

  • Regular Contributor
  • *
  • Posts: 108
  • Country: us
Re: EEVblog #1066 - Uber Autonomous Car Fatality - How?
« Reply #74 on: March 22, 2018, 03:44:32 pm »
Here's an interesting read related to any automated vehicle. Worth a look over. Its from the Austrialian Transportation Safety Board. The incident of Quantas Flight 72.

https://www.atsb.gov.au/media/3532398/ao2008070.pdf

Start on "Executive Summary". You're welcome to read the rest of 300+ pages though.  >:D
« Last Edit: March 22, 2018, 03:46:43 pm by CNe7532294 »
 


Share me

Digg  Facebook  SlashDot  Delicious  Technorati  Twitter  Google  Yahoo
Smf