Footage of driverless Uber car / pedestrian hit released

[Delicieuxz]

Figured I wouldn't start a new Tech News & Discussions thread about this after the original one was closed. However, the footage has been released, and I think it shows some deficiency in the car's automation.

 

The video footage can be watched here: https://www.rt.com/usa/421978-uber-driverless-death-video-released/

 

FYI, the exterior footage cuts out the moment the car is about to make contact with the pedestrian.

 

 

Here's what I notice and think about what the footage reveals:

 

The black / light contrast in the video is really high, making the blacks darker than they would actually appear while driving.

 

The Uber vehicle made no attempt to swerve out of the way.

 

To me, it looks like the car didn't even attempt to slow down, even after the pedestrian was clearly visible in the video footage.

 

Also, the pedestrian did not jump out in front of the car, or suddenly appear in front of the car. The pedestrian was walking their bike across the road slowly, already while the car was a far distance from the pedestrian.

 

I think there should have been enough time for the car to brake, if the sensors were working and scanning an appropriate distance in front of the vehicle.

 

 

In my opinion, this accident would have been much less likely with a human driver. I think the pedestrian would have been perceived a lot earlier, and that a human driver would have recognized light reflecting from the bike's spokes and other metal. I also think that normal human braking and swerving abilities, taking into account there being no other vehicles on the road and there being plenty of available road-space to maneuver into, could have very likely prevented the vehicle's collision with the pedestrian.

[Trik'Stari]

After seeing the video, yeah the car should have been able to stop.

 

What I don't understand is how the pedestrian didn't think "maybe I should move faster and get out of the way, there's a car coming". Judging by the footage, I'd have thought she would have been able to perceive that well before the point she became visible in the footage.

 

Not that she would have thought "that car is going to hit me" but generally speaking if I am crossing a road and I know a car is coming, I move faster just because "I can't afford to take the chance that that person doesn't see me".

[Delicieuxz]

10 minutes ago, Trik'Stari said:

After seeing the video, yeah the car should have been able to stop.

 

What I don't understand is how the pedestrian didn't think "maybe I should move faster and get out of the way, there's a car coming". Judging by the footage, I'd have thought she would have been able to perceive that well before the point she became visible in the footage.

 

Not that she would have thought "that car is going to hit me" but generally speaking if I am crossing a road and I know a car is coming, I move faster just because "I can't afford to take the chance that that person doesn't see me".

Yeah, I wondered about that, too. Maybe they expected the car to slow down.

 

Regardless, though, I think that automated cars which can't slow down or swerve in this situation, and which react less dynamically than a human does, are extremely risky to have on the road, since children often play on road, run across roads, and get themselves into situations tighter than the one with this pedestrian walking across the road with their bike.

 

I think that if automated cars can't skillfully respond to the situation the video shows, then these things might well turn out to be kid-murdering machines, once there are tons of them on the roads.

[NMS]

Seems to me like that car should have been aware of the pedestrian sooner. What are all these censors for then?

Was it a software issue perhaps?

[Mr.Meerkat]

Automation does occasional fail hence why typically there is a person monitoring to ensure no accidents happen but at the same time, the person monitoring the automation can also fail so...ermmmm...

 

2 minutes ago, Trik'Stari said:

What I don't understand is how the pedestrian didn't think "maybe I should move faster and get out of the way, there's a car coming". Judging by the footage, I'd have thought she would have been able to perceive that well before the point she became visible in the footage.

This is why I think all three entities (automation, driver and pedestrian) are at fault.

The general reaction of a pedestrian when they are jaywalking (or crossing the road in general) and notice a car coming is to speed up so they can clear the road faster and cause less disruption for the driver(s). However, as the pedestrian did not react to the incoming car, if the pedestrian was distracted, they shouldn't be as they are crossing a road after all and should be paying attention to their surroundings. But, if she did notice, then she should not have been so naive in thinking the driver will react and she doesn't need to do something. 

In terms of the automation, it failed, shows that the system was not thoroughly tested and more thorough testing should have been carried out in a controlled environment. 

The driver was not doing his job and well there's not really anything else to talk about. Had he been paying attention to his job and actually carrying it out, he could have probably taken over control and swerved and/or slammed on the brakes (which would have at the very least lessened the force of impact if the driver was unable to miss the pedestrian). 

 

Anyway, although the driver isn't the only entity at fault, they probably will be the entity that gets the majority of the blame. If we have a look at aviation accidents (specifically commercial), even if a mechanical fault, weather factors or something affects the flight (that is not due to the pilot(s)) and the pilot makes a mistake somewhere during the emergency landing, generally they will end up with the biggest blame as its simply cheaper and the easiest to just put the blame on one or two person instead of a set of systems or bad design or whatever. 

[RexinOridle]

Well, sensory input to the bran aren't always processed the right way. Sometimes people drink. May be the pedestrian had an input lag.

[Okjoek]

On one hand I thought these cars were supposed to react to this sort of thing.

 

On the other hand. MY GOD is that bicyclist stupid. Being that oblivious to the headlights and they walk out going that slow anyways? In the middle of the night too. Like friggen darwin award level stupid. I wouldn't blame an active driver for hitting them.

[dalekphalm]

Based on my rough calculations, the moment the pedestrian is visible, to the moment right before impact, is approximately 1.5 seconds. At 38 mph, that's 55.7333 feet per second. 1.5s x 55.7333 fps = 83.59995 feet (rounded to 84).

 

So that means she was 84 feet away when she became visible.

 

According to this website, average stopping distance for 40 mph is 80 feet (not counting reaction time).

http://www.government-fleet.com/content/driver-care-know-your-stopping-distance.aspx

So that means that if the SUV (which we don't know the specific stopping distance for that specific vehicle) started braking instantly, then it might have avoided the pedestrian.

 

If we assume a 100 ms (0.1s) reaction time for the autonomous system to see the input, process the data, and activate a response, we'd actually be looking at over 84 feet including stopping distance:

55.7333 fps x 0.1s = 5.57333 feet

 

5.57333 feet (reaction time) + 80 feet (stopping distance) = 85.57333 feet to come to a stop.

 

Granted, that would likely have been a non-terminal speed, when they collide at 84 feet, but still, I'm unconvinced that the system could have stopped completely. Though it's likely it could have reduced the speed - possibly to non-terminal levels, yes.

 

We could assume that a person could see her sooner (the vehicle operator certainly didn't). We could assume that the sensors (LIDAR, RADAR, etc) could see her sooner. But these are assumptions, and we don't know for sure, unless they release all sensor data.

 

Definitely, this tells me that the autonomous system needs more work - though I still firmly believe that they are - even in their present state - safer than most human drivers. Hopefully this pushes increased safety to autonomous vehicles, rather than a knee-jerk reaction trying to ban them, etc.

[Delicieuxz]

14 minutes ago, dalekphalm said:

Based on my rough calculations, the moment the pedestrian is visible, to the moment right before impact, is approximately 1.5 seconds. At 38 mph, that's 55.7333 feet per second. 1.5s x 55.7333 fps = 83.59995 feet (rounded to 84).

 

So that means she was 84 feet away when she became visible.

 

According to this website, average stopping distance for 40 mph is 80 feet (not counting reaction time).

http://www.government-fleet.com/content/driver-care-know-your-stopping-distance.aspx

So that means that if the SUV (which we don't know the specific stopping distance for that specific vehicle) started braking instantly, then it might have avoided the pedestrian.

 

If we assume a 100 ms (0.1s) reaction time for the autonomous system to see the input, process the data, and activate a response, we'd actually be looking at over 84 feet including stopping distance:

55.7333 fps x 0.1s = 5.57333 feet

 

5.57333 feet (reaction time) + 80 feet (stopping distance) = 85.57333 feet to come to a stop.

 

Granted, that would likely have been a non-terminal speed, when they collide at 84 feet, but still, I'm unconvinced that the system could have stopped completely. Though it's likely it could have reduced the speed - possibly to non-terminal levels, yes.

A human driver would have combined braking with swerving into the other open lane, though.

 

Also, like I mentioned before, I'm pretty confident that the person and their bike would be more visible than the video makes it seem, because of the video crushing all the dark colours into black. I also think that the headlights likely would reflect in some small though detectable manner off the bike's metal to indicate there was something there, even though they don't show up in the video.

[dalekphalm]

1 minute ago, Delicieuxz said:

A human driver would have combined braking with swerving into the other open lane, though.

 

Also, like I mentioned before, I'm pretty confident that the person and their bike would be more visible than the video makes it seem, because of the video crushing all the dark colours into black. I also think that the headlights likely would reflect in some small though detectable manner off the bike's metal to indicate there was something there, even though they don't show up in the video.

A human driver might have combined with swerving. you cannot say that definitively, when so many car accidents are caused by human drivers not taking evasive action.

 

As for the rest, all assumptions. Possibly true. But I will not draw any conclusions without further sensor data. The bike could have been painted with black paint to reduce reflection - and since it's going sideways, you wouldn't see the Pedal reflectors, etc.

 

Swerving at high speed is dangerous in itself, especially with a top heavy SUV - by doing so, it could easily roll the car, which might make the situation much worse.

 

I think we both agree that with better technology, the SUV likely could have avoided this accident though. The SUV should have a combination of LIDAR and RADAR sensors, aside from the optical cameras - if so, it should have seen the pedestrian even before it became visible out of the shadows. That means the autonomous AI subsystem failed to interpret the data properly, or otherwise failed to act on the data.

 

On the other hand, if this particular autonomous vehicle system relies purely on optical cameras? Then yeah that's simply a design limitation and would need to be upgraded with better sensors.

[79wjd]

 

4 minutes ago, dalekphalm said:

As for the rest, all assumptions. Possibly true. But I will not draw any conclusions without further sensor data. The bike could have been painted with black paint to reduce reflection - and since it's going sideways, you wouldn't see the Pedal reflectors, etc.

 

I think we both agree that with better technology, the SUV likely could have avoided this accident though. The SUV should have a combination of LIDAR and RADAR sensors, aside from the optical cameras - if so, it should have seen the pedestrian even before it became visible out of the shadows. That means the autonomous AI subsystem failed to interpret the data properly, or otherwise failed to act on the data.

 

On the other hand, if this particular autonomous vehicle system relies purely on optical cameras? Then yeah that's simply a design limitation and would need to be upgraded with better sensors.

Two things, firstly, there are other videos that people made that seem to have the lighting that one would expect -- so it seems likely that the Uber video was darkener than reality (due to poor settings, or unlikely -- ulterior motives). Secondly, Uber's car supposedly do have LIDAR. 

 

https://arstechnica.com/cars/2018/03/police-chief-said-uber-victim-came-from-the-shadows-dont-believe-it/

[EPENEX]

I think the thing everyone fails to remember is that there was a human behind the wheel that should have been paying attention. Just because the car is automated gives the driver no reason to not pay attention. It is entirely the drivers fault.

[Delicieuxz]

32 minutes ago, EPENEX said:

I think the thing everyone fails to remember is that there was a human behind the wheel that should have been paying attention. Just because the car is automated gives the driver no reason to not pay attention. It is entirely the drivers fault.

But pretending that automated cars are for anything other than not paying attention to the driving is just that. If a person is to focus on the driving just the same, then they might as well be driving and there's no point for the automation.

[dalekphalm]

41 minutes ago, Delicieuxz said:

But pretending that automated cars are for anything other than not paying attention to the driving is just that. If a person is to focus on the driving just the same, then they might as well be driving and there's no point for the automation.

While that's correct in theory, you've got to understand that these vehicles are basically still prototypes. While once the AI systems are worked out and thoroughly tested and regulated for general use, yes, the operator can be expected to pay no attention at all - since at that point they are just another passenger.

 

But that isn't the case here. These are test systems. The operator is there for a few reasons:

1. To take control should the autonomous system fail

2. To take control in an emergency should the autonomous system miss something (like what happened here)

3. To monitor the system, and report back in case improvements or useful observations can be made

 

Now, in this case, I still doubt the operator would have had time to see the pedestrian, process what they were seeing, take control of the vehicle, and stop/swerve safely in time to save the pedestrian. They literally only had a few seconds to react, and typical human reaction time in an automotive emergency is around 1-2 seconds.

 

Typical reaction time at 40 mph (58.6667 feet per second - rounded up from 38) is 59 feet. This works out to 1.006 seconds before the operator even moves their foot. That's already half the distance covered between the vehicle and the pedestrian.

 

So that leaves around one second to stop or otherwise avoid the pedestrian. It's technically possible, but I highly doubt most drivers would be able to do so.

 

So, I don't think the driver is really liable in this specific case, but at the same time, the operator should be alert and paying attention at all times (as much as humanly possible) - they should have tried to stop, even if it was impossible.

[Monkey Dust]

2 hours ago, Delicieuxz said:

But pretending that automated cars are for anything other than not paying attention to the driving is just that. If a person is to focus on the driving just the same, then they might as well be driving and there's no point for the automation.

This is the problem with partial automation, when it gets in to trouble it relies on a human whose concentration has almost certainly wandered.

[mr moose]

16 minutes ago, Monkey Dust said:

This is the problem with partial automation, when it gets in to trouble it relies on a human whose concentration has almost certainly wandered.

I guarantee it has wandered.    Humans are really really bad at staying focused on tasks when there is no immediate requirement to do so (sometimes even when there is we daydream while driving).  Baring the autistic and a few rare people no one is immune.