It's not clear to me that this is purely a software problem; it might be a "the car doesn't even have the sensors necessary to detect this kind of situation" problem.
It could also be a "the software can't anticipate that this would be happening the way a human can because it doesn't have object permanence" problem. That's probably at least part of how a human would do it. But I have a suspicion that we're very, very far away from being able to get computers to do that in any sort of reliable way. So I'm guessing we're probably looking at specialized sensors if we want to handle this kind of situation appropriately.
Basically some sort of self-driving car equivalent of a human driver getting out of the car to check and make sure everything's OK.
I would be surprised if it turns out that those self-driving cars don't have object permanence. Making things move around without a physical model looks like a really nasty infinite bug-breeder.
For the longest time, at least the visual representation of trains at a level crossing strongly implied challenges around object permanence for Tesla.
I will say that I recognize that the visual representation may not be a 1:1 (in fact quite certainly isn't) mapping of the computer's "world view".
But it was quite ridiculous and definitely disconcerting seeing a level crossing with a train going through it and seeing the Tesla I was in acting like there was a traffic light that was erratically going red, while dozens of semi trucks flickered in and out of existence across the screen in front of me.
> strongly implied challenges around object permanence for Tesla
Anyone who has ever ridden in a Tesla knows they have no real perception of the world around them. Even just driving down the road they show icons of the things around them, which will randomly appear and disappear or change from busses to pedestrians to cars.
And yet a common thread in post-mortems on so many self-driving car collisions boils down to, "the car doesn't have object permanence."
It may have some ability to track specific things in specific ways in a manner that's hard-coded by engineers. But if that's the same thing as an actual mental model of the world then that implies that we've had a significant piece of AGI in the form of many video games for about half a century now and just not known it.
yeah, another poster mentioned they would have heard the screaming.
audio is difficult because a human can differentiate a woman screaming from the sounds of construction, etc. And also can tell that it's someone in shock vs someone in pain.
I'm unsure if audio processing is good enough for that yet, but imagine if someone were to try and you could get these vehicles to stop by screaming near them.
That is the beauty of technology: once we identify a problem we can do something about it and apply the fix to everything. While this is the type of case you can argue nobody would have expected before, once it happens you know and can setup lab situations where similar things happen and then figure out how best to respond over months of effort.
This is absolutely not a case nobody would have expected before? People end up under cars in accidents regularly. Imo any company that "found out" about this scenario through the cruise incident is guilty of gross negligence.
I'm sure all the bigger ones had looked at the possibility before, but it's only natural to revisit it when something horrible like this happens to a competitor.
I am sure that you don't generally start with this scenario - but I have to say again that this is a common thing. People end up under cars all the time (in accidents). There is just no way anyone should excuse any company for "not considering" this scenario if their cars are on roads. I'm sure it's hard! Other industries recognize corner cases where they need to design to avoid really ugly long tail outcomes and do so with ethical rigor. Most planes do not crash into the ocean and yet you are educated on how to escape in a water landing at the start of every flight.
Note that it's not just a pedestrian ending up under a car, but the whole situation of "other car knocks pedestrian into your path, which then ends up under your car".
The problem is that the world is full of near infinite permutations of situations that will continue to prove difficult for these technologies. A group of pigeons, to most human drivers, is identifiable from any approachable angle, in various weather and lighting conditions. To a computer, it has to be trained on all of that and consider the countless permutations - is that two pigeons in the rain in the dark? Is it 50 in broad daylight? Is it 12 cardboard cutouts? Our brains are incredibly good at parsing the world around us, and I’m not sure we’re even remotely close to that level of accuracy with self-driving cars, and that should worry people in cities where these are being rolled out.
Or you can not code your autonomous vehicle to try to move after getting into a collision, because if someone is caught under the car, you shouldn't try to move the car anyway.
It's hilarious that the fucking things park themselves if you put a traffic cone on the hood, but if they get into a crash, especially one with a pedestrian? BEEP BOOP MUST MOVE!
That's not hard for me to imagine. Now that I think about it, it should probably be standard for autonomous cars -- if there's an accident or something unexpected it should pull over, unless it can't reach the side of the road without hitting something or someone is trapped under the car.
> it should pull over, unless it can't reach the side of the road without hitting something or someone is trapped under the car.
Or…insert how many other edge cases? There are plenty of scenarios where it shouldn’t pull over, aren’t there? Training the myriad unique situations and proper responses just feels untenable, at least with current technology.
Code that identifies if anything is stuck under the car would be useful for a lot of cases but especially if an accident already has happended.
Besides that, hard edge cases will always exist and can only be solved by a human (or an AGI?) in the loop. But they are't too common, so just detecting that the situation is not solvable by the algorithm, doing nothing and waiting for a remote operator should scale well enough.
It could also be a "the software can't anticipate that this would be happening the way a human can because it doesn't have object permanence" problem. That's probably at least part of how a human would do it. But I have a suspicion that we're very, very far away from being able to get computers to do that in any sort of reliable way. So I'm guessing we're probably looking at specialized sensors if we want to handle this kind of situation appropriately.
Basically some sort of self-driving car equivalent of a human driver getting out of the car to check and make sure everything's OK.