autonomous cars - the future of motoring is driverless
- Thread starter roryer
- Start date
NoXion
Craicy the Squirrel
Come on, it's not as if buses have ever looked especially stylish, except when seen through the eyes of nostalgia.
mauvais
on reddit or something
The 'driver' is there to provide oversight and intervene if necessary. It's slightly different here because they are there in a paid capacity, but the same exists in other autonomous systems like Tesla Autopilot. I've said this many times: it's total bullshit. Give someone a task where they're mostly redundant and you will inherently get disengagement. You can either have full autonomy or you make the driver perform some continuous activity, no middle ground.
This is basic stuff too if you understand how attention works psychologically. If a task doesn’t fully engage our attention, the excess capacity latches onto something else. If that engages us, our limited capacity system diverts to focus on that instead. Overcoming this process is very hard and the subject of its own field of endeavour (ISTR it’s called something like psycho-ergonomics but a search doesn’t reveal anything so I must have that wrong.). Putting people in a car with nothing to do except in an emergency breaks every rule in that book.
mauvais
on reddit or something
I saw the video earlier and it appears to be a failure of the car that should have been avoidable. It's a linear, perpendicular intrusion into the path of the vehicle, the classic cross-path scenario it should be designed to handle. The dashcam video presents a bit of a skewed view in that its dynamic range is very poor - a human would be able to see more - but the car systems shouldn't be confined to visible light anyway (instead: infrared, RADAR/LIDAR, etc).
It leaves me wondering various things, not least how much of this accident will be lab-reproducible and how much it's an edge case. It's hard to explain this but in certain systems designs, written under safety-critical development doctrine, someone like the NTSB would be able to identify the sort of swiss cheese condition that failed to handle a scenario, and go from there. I have a strong suspicion that they can't do this here, that it's not meaningfully safety-critical and that overall it shouldn't be being tested on the public.
It leaves me wondering various things, not least how much of this accident will be lab-reproducible and how much it's an edge case. It's hard to explain this but in certain systems designs, written under safety-critical development doctrine, someone like the NTSB would be able to identify the sort of swiss cheese condition that failed to handle a scenario, and go from there. I have a strong suspicion that they can't do this here, that it's not meaningfully safety-critical and that overall it shouldn't be being tested on the public.
Sea Star
have you ever explored your dark side?
A trans woman apparently. The Daily Mail were trying to make a big issue out of that but they had nothing. She's an ex con, but thoroughly reformed.
magneze
🎧
There's a video. Stops just before impact.
Video released of Uber self-driving crash that killed woman in Arizona
Video released of Uber self-driving crash that killed woman in Arizona
magneze
🎧
A few things spring to mind after watching.
A human wouldn't have been able to react.
The car was driving really fast, I assume legally.
Why couldn't the car detect the person in the road earlier? If they can't drive better than people what's the point?
As mentioned earlier, the "pilot" isn't there for safety. Had no eyes on the road and isn't concentrating. Even if she had doubt it'd have made a difference. Can't be blamed but I bet she feels awful.
A human wouldn't have been able to react.
The car was driving really fast, I assume legally.
Why couldn't the car detect the person in the road earlier? If they can't drive better than people what's the point?
As mentioned earlier, the "pilot" isn't there for safety. Had no eyes on the road and isn't concentrating. Even if she had doubt it'd have made a difference. Can't be blamed but I bet she feels awful.
magneze
🎧
The car also doesn't swerve. Could it have missed the pedestrian but then been involved in another accident? Was there a calculation made by the algorithms to not swerve in case the pilot or more people were hurt? I assume there is something like this built in? What are those calculations?
SpookyFrank
A cheap source of teeth for aquarium gravel
Robot cars 1; human-driven cars 3000-odd since breakfast.
mauvais
on reddit or something
Wrong.
Also, as mentioned (post 245), dashcam videos represent much worse visibility than the human eye perceives in real life.
38 in a 35, apparently. The video above is a 50 limit, FWIW, although the actual speed is unknown.
The driver is meant to be there for safety, but as discussed, it's a flawed premise.
Nothing complicated - it should have been able to detect the person. It didn't. No avoidance, no braking.
This is a massively overblown fear/discussion point, I think, that, if I’m right, derives from a possible misunderstanding of how machine learning works (unless it’s me that misunderstands how machine learning works). It’s an issue raised more by philosophers than programmers. The assumption is that the machine makes decisions in the same way that people makes decisions, which involves moral imperatives. But I don’t think the machine is capable of working through the consequences of an act more than a very short time period ahead. I don’t think it forms a possibility space of downstream consequences and chooses between them. I think it just acts to avoid a developing risk in the way that the data shows is most effective.
cupid_stunt
Chief seagull hater & farmerbarleymow's nemesis.
Looking at that video, it seems when she became visible from the shadows, she was only a couple of car lengths in front of the uber car. As the 'thinking' distance at 40 mph is 3 car lengths/12m, with braking distance another 6 car lengths, I am not convinced that a driver could have changed the outcome, and can understand the police chief saying the car is likely not at fault.
mauvais
on reddit or something
No, for all the reasons already given.
mauvais
on reddit or something
Yeah, but also: so does a human driver, if they even do that (target fixation etc). The trolley problem is something confined to philosophers and disaster movies.
cupid_stunt
Chief seagull hater & farmerbarleymow's nemesis.
Your conclusion is just based the video, from which it's hard to conclude one way or the other, but 2 seconds is spot on the time it takes a human to normally react, then you also have the braking time on top of that. The police on the ground clearly have far more to go with in order to judge the situation, so I am more inclined to accept what they have said at this point in time.
mauvais
on reddit or something
Some police spokesperson (the chief?) came out early on and said it was the victim's fault, and then the video came out, and the police department backtracked, claimed they were taken out of context.
Regardless, have you looked at many dashcam videos? They're all much worse than a human at capturing available information that you take in when driving at night. The sensor (the infeed to purely visible light visual processing) is probably worse than the human eye, and then the rendered output is worse than the sensor data.
Then, the idea of the 'car not being at fault' misses the myriad ways in which it's supposed to be able to detect things better than the human eye.
cupid_stunt
Chief seagull hater & farmerbarleymow's nemesis.
The police statement was after seeing the video...
I haven't seen anything about the police backtracking, do you have a link?
BigTom
Well-Known Member
Camera sensors are definitely worse in low light / mixed light conditions - they will auto adjust exposure to control the brightest light (from the streetlights in this case presumably), making darker bits very dark. The human eye is much better at dealing with images with mixed bright and dark areas and so to the human eye the person would have been more visible/visible earlier. Although we can't really know, I would think that a human would have been able to see the pedestrian earlier and react accordingly. HDR video/photo was developed to help deal with this problem but there's basically no chance of having this in a dashcam I shouldn't have thought.
Last line is paramount though - autonomous cars should not be reliant on light sensing when there are so many other sensor types available.
magneze
🎧
I don't see what machine learning has to do with it. There will be some basic rules about car handling. The car could have swerved but flipped over so perhaps this is a rule that it couldn't break. Maybe it should have?
The concept of machines making a calculus of life is dependent on the idea that the machine is modelling future scenarios and choosing between them. This just isn’t the case — it’s not how machine learning works.
magneze
🎧
Maybe not explicitly to do with calculus of life but there is a calculation that was made that meant the car wasn't braking or swerving. Should it have? I imagine there's a bunch of engineers looking at that now...
magneze
🎧
Police chief said Uber victim “came from the shadows”—don’t believe it
Interesting alternative video that shows the same area at the same time more lit up, somewhat supporting the above posts about dashcam quality being shit.
Interesting alternative video that shows the same area at the same time more lit up, somewhat supporting the above posts about dashcam quality being shit.
pseudonarcissus
fluttering and dancing
In the States it's almost impossible for a driver to be at fault in a pedestrian fatality, unless you run a red light (red light cameras are unconstitutional). The victim would be to blame even on a sunny day, in this instance. Even if it's where everyone crosses the road because it'd where the shops are. Zebra crossings are an amazing invention.
it's the lack of pedestrians that keeps the body count down
mauvais
on reddit or something
The New York Times has published some interesting stats:
Uber’s Self-Driving Cars Were Struggling Before Arizona Crash
Approximate miles per driver intervention:
Waymo (Google): 5,600
Cruise (GM): 1,200
Uber: < 13
Uber’s Self-Driving Cars Were Struggling Before Arizona Crash
Approximate miles per driver intervention:
Waymo (Google): 5,600
Cruise (GM): 1,200
Uber: < 13
