Uber wasn't to blame for robo-ride crash – or was it? Witness said car tried to 'beat the lights' A police report appears to support the claim that Uber was not to blame for a recent crash of its self-driving car in Tempe, Arizona. But the incident raises serious questions about what rules the startup's engineers are putting into the car's software. The report – made available Wednesday for anyone who is able to …
My impression was that she hit the Uber car in the side with enough force to push the Uber onto it's side. In any case, it was her responsibility to make sure the intersection was clear before entering it. Make this out to be Uber's fault? That's just P.C. in my book.
If they train experts system (AI) using reading taken from people driving in the beat the lights manner then it will try to beat the light because that is how it has been teached it should drive. Using expert systems (AI) is all well and good but you have to remember to teach it using good driving habits rather than bad driving habits.
But, wait! There's more! Get ready to test my new invention; fake traffic light projections onto self-driving vehicles with inexpensive LCD mini-projector, a smartphone, and a live crab. Patent Pending. We're still working out what the crab is for, but you'd be surprised, it does not function without one for the nonce.
The Uber car was knocked onto its side by the impact! Despite what witnesses might have stated, it's extremely likely that at the moment of impact the Uber cars speed was a lot lot less than that of the other car. It's very very hard for a slow moving car to knock a fast moving one onto its side.
It's very very hard for a slow moving car to knock a fast moving one onto its side.
Other way around ?
A stationary car can roll another. If the oncoming car turned across it, then it's (relatively) low bonnet would be first into the path of the Uber car. The Uber car could then "drive up" the bonnet like a ramp and do what in the film/stunt business as a pipe roll.
On the other hand, drive into a stationary or slow moving car "side in" and it's very hard to roll that car over.
So clearly you never saw the episode of mythbusters where they tried to use the bonnet of one car as a ramp for another. That myth was busted IIRC.
But in all seriousness it is very difficult to predict when a car will roll. It's often surprising which vehicles pass and which fail the elk test. One thing you can be pretty sure of is that the Volvo, being a Volvo, will have passed the elk test. As such it should be pretty damned hard to roll.
The photographs of the aftermath don't make it clear how the car was rolled, but visible damage to both cars does not speak of a high speed impact.
"So clearly you never saw the episode of mythbusters where they tried to use the bonnet of one car as a ramp for another. That myth was busted IIRC."
Not sure you can use Mythbusters as a reliable proof of whether something can be done. They once "proved" that robin hooding an arrow in archery (eg shooting one into the back of another) couldn't be done, yet even as an average archer I did it (I still have the two arrows stuck together on my wall), and when visiting the local archery shop with its own range (which the staff regularly used when not serving) they had a bin full of the things. At best they simply prove THEY couldn't do something after a few goes in certain circumstances.
"The Uber car could then "drive up" the bonnet like a ramp"
I had an office by an intersection that saw its fair share of this kind of crash (mostly caused by straight-through vehicles running reds and hitting turning vehicles on arrows).
This kind of scenario _never_ happened (ie: in no case did the oncoming car ride up over the nose of the turning one, or vice versa), but in one case the driver of the turning vehicle saw the crash coming and tried to avoid it. Instead of braking hard she hit the gas and ended up accelerating hard into the side of the straight-through vehicle, which ended up on its roof about 20 yards further on.
One of the things I learned from working there was that crash witnesses are highly unreliable. Most aren't paying attention and only start properly looking _after_ the noise happens, then assume the rest. Statements given to the police seldom tallied and CCTV footage (which is why the police would be in my office) usually showed something completely different to their claims.
The Uber car bounced off a traffic light / street light, /THEN/ rolled. and hit two other cars.
My guess: it rolled because it mounted the kerb.
Uber was overtaking on the inside. Stationary traffic in the middle lanes of the 6 lane road, uber in the inside lane, doing 38 in 40mph zone through a yellow light.. Uber was nudged on to the Right kerb by car making a Left Hand turn. Continued on, mounted the kerp, hit a pole, spun and flipped, hitting two other cars.
(remember, driving on Right in USA)
Observation (1). No sensible driver would do 38/40 through a yellow light inside of stationary traffic.
(2) Any sensible driver would have expected [Left Hand Turn] vehicles to be waiting for the light to change to complete their turn. They always do that.
(3) Any sensible driver would realise that the [Left Hand Turn] cars would be threading their way through and hidden by the stationary traffic, which is why you don't do full speed on the inside lane through a yellow light.
4)Uber radar and AI failed to detect [Left Hand Turn] traffic.
So two problems: driving AI took a bad risk. Detection AI failed.
Technically [Left Hand Turn] 'failed to give way'.
Practically, if uber drives like this, uber will have accidents like this.
"How can you drive so fast, while turning left at an intersection, that the freaking Volvo you hit rolls over!? Was the woman driving a tank?"
You can be doing zero, if the Volvo hits you at the right speed and angle to career off and hit something else and flip. Since the other person was turning left at an intersection, they won't have been going very fast, let's be honest. And 38 at a blind intersection is insane.
One for humans (visual), the type we are all used to.
The other (electronic) is for robot cars. All intersections have some kind of electronic signal which signals to robot cars that the traffic light is red, even though it is green. If the speed profile of the car is deemed safe, and a green is likely to be encountered at that speed profile when the car hits the junction then the light is cleared for the car to pass through. The speed profile is monitored throughout the approach and is triggered accordingly when it is assured that the overall process is under control.
I like the idea of providing non-visual information for the robot car, but I think you would be opening a new can of worms if you gave different signals to the car. In the example you give, neither the "safety driver" nor any non-automatic following vehicles, would expect the robot car to brake for a green signal.
It would perhaps be better, (and simpler from the traffic signal's point of view), to tell robot cars approaching the junction which signals were currently displaying green, and the timings of any upcoming changes. e.g. "now: E to W green, W to E green. in 29s : E to W green, E to N green ...".
That way the car would be able to decide how to handle the junction, without needing to detect the physical traffic lights.
One could combine something like a racetrack transponder (essentially an indentifying RFID) in the car with speed detecting loops under the pavement linked to the traffic controller. It should then be possible to pass information to that specific car about its chances of making through the intersection, as well as communicating the exact distance to the light to the AI.
And who should pay for all these electronic signals? Every light controlled junction in the world? That would cost billions and the people who should pay world be the manufacturers of autonomous cars, and therefore the buyers. The public purse shouldn't pay, why should the public purse subsidise this enterprise?
However these cars need to deal with the roads the way they are, the roads should not be changed to adapt to these cars.
If these things do take hold, it will be no different than any other infrastructure project. Where are the flaggers once required to walk in front of motorized vehicles? Who paid to pave the roads and build the current traffic infrastructure when the world moved away from horses and buggies? Is there a tax on the blind to equip intersections with auditory signals and warning tiles?
When all cars are fitted with a device that broadcasts locally 'here i am, this is how fast i am, this is what direction i am going in, my driver has just put the on the indicators/brakes/stomped the accelerator', something i expect to be relatively cheap, and could be used as a driver assist for ALL cars... That is the point i would expect autonomous cars to never be at fault in a RTA. When all autonomous cars feed each other this data, and they are predominant on the road, they'll be driving 2 feet apart and you wont need traffic lights...
So that's why Elon Musk is desparate to build a neural lace; it's so us cyclists can interface with the infrastructure required to make self-driving cars work properly.
And, as a thought experiment, the self-driving car has to be fully autonomous and safe in the event of a complete failure of the traffic signals, or has to be able to detect this failure and pull over until a human driver takes over or the failure is resolved.
When traffic lights turn to green, there is a delay before each successive vehicle in the queue can accelerate from rest. This is inefficient. With auto-driven cars that are in touch with one another electronically it would be interesting to investigate the possibility that vehicles "batch" their intentions together, all those doing so during the green phase of the lights that can make it through on that green phase can be bound together electronically and accelerate in unison through the intersection.
Hey, maybe we could have them follow each other in a kind of, I don't know, erm - train! Then the one at the front could have a real driver, let's call them the, erm, train driver!
If you're going to rework city infrastructure at public expense to cater for transport, do it properly and put in a decent mass transit system. This ticks not just the 'cheaper' box but also the 'safer' and 'less polluting' ones.
If Uber want something different they can lease the space to install and maintain their own equipment, city by city.
but if it is a standard 4 way intersection with traffic lights, what is the issue?
Come to Europe, we don't really have many road junctions like that because of the aforementioned batshitinsanery of them. Far too confusing and slows down traffic flows.
We have roundabouts instead.
"Curious what you think is "batshit insane" about US intersections"
Well, four-way stop is a reasonable thing, but I've seen a (very small) number of four-way yield intersections on American roads. And right turn on red is a classic, especially as it isn't universal. In New York (the state), it is permitted except where prohibited by a "No Turn On Red" sign; In New York (the city), it is prohibited except where permitted by a "Right Turn On Red" sign.
And there was the set of traffic lights I saw one time near the Capitol building in Albany, NY, with side-by-side lights.
And other gems, like intersections governed by flashing red in one direction ( = STOP sign ) and flashing yellow in the other ( = YIELD / Give Way sign ).
But they don't have anything like the Magic Roundabout in Swindon.
Where I live we've been introducing roundabouts here and there to avoid four way controlled intersections. Those of us who have been to the UK and successfully navigated them (it is harder than you think when you aren't used to them, and driving on the wrong side of the road to boot, so we apologize for slowing down unduly when approaching them on occasion) know what to do, but a lot of people are lost. Especially those from out of town.
Don't underestimate how difficult it is to deal with something you aren't used to. People who grew up where there are roundabouts understood implicitly how to drive them before they ever got behind the wheel, simply by being in the car when others did. There's a lot more thinking required to navigate them than a four way with lights. You just do what the lights tell you, easy peasy. Yeah, it isn't always the most efficient, but assuming you do like most do around here and have two seconds of all red to avoid problems from people who chase the last nanosecond of yellow, they are probably safer in the US than roundabouts. Maybe things would be different if we had roundabouts our whole lives, but there are very few places in the US where that's the case (and they still have problems with visitors)
One thing that I think has helped with the yellow light chasers is that most of the walk/don't walk signs in intersections have a countdown timer. When the clock reaches 0, the light goes yellow. So you can tell a block away whether you will be able to make the light or not. It acts as sort of a pre-yellow warning so you don't see very many cases of someone racing through while it is in the process of turning red.
They have one in Ellijay, Georgia, US and ever since I've encountered it, I've avoided the town completely when going that direction.
I cannot compare this roundabout with anything else due to poor State of Georgia signage and Google Maps losing it's fucking mind when reaching that roundabout. Just nope.
We're forgetting a major factor here: The human driver that turned left **Into oncoming traffic** without ensuring that traffic was clear. If the light was red, yellow, green, or pink, they seem to be the ones at fault, not the Uber. (weasel wording, because I haven't read the traffic report).
The Uber car in this case:
1) Was going under the speed limit.
2) Did not disobey a semaphore
3) Did not hit anything (it was the vehicle that got hit)
Also note the Bloomberg article states "the light turned yellow as the Uber vehicle entered the intersection". It did not say: "The light turned yellow when the Uber was 1/2 a block away, at which point onlookers heard the roar of the engine as the crazed robocar raced to beat the light".
Now: what if ALL cars in the intersection that day used Uber's self driving software? would the Honda have chosen to turn into oncoming traffic?
One other thing to consider -- the unreliability of eyewitnesses.
In this case, if the other cars traveling in the same direction as the Ubermobile slowed when the drivers saw the yellow light, then the witness who claimed that it "sped up to beat the light" may have simply been perceiving the relative differences in delta-V and, in the moment, attributed it to what s/he EXPECTED to see, rather than to what actually happened.
Eyewitness accounts are wildly unreliable. I'd be happier with video evidence.
FWIW attributing fault can be a complex issue with road traffic incidents, usually best left to insurers - those are the guys who will ultimately decide if Uber's cars are fit to be on the road.
Yes, the other driver 'failed to give way' and will be cited.
On the other hand, when you over-take on the inside, through a yellow light, where visibility is blocked by stationary trafic, while driving at full speed, then when you get nudged by someone trying to complete a turn, you will mount the kerb, hit a light pole, spin out and flip over, and hit two other cars.
Your choice.
(There weren't any pedestrians waiting at the corner.)
In a laned traffic light controlled intersection where traffic in the turning lane is stationary, you can expect traffic in the straight through lanes to be travelling the speed limit - no matter which country you're in.
If you suddenly decided to slow down to 20mph in such a case in London you'd have a bunch of pissed-off drivers on their horns at you.
There's no such thing as "Overtaking on the inside" when the passed traffic is stationary.
The driver of the turning car proceeded to move when the way was not _visibly clear_ (obstructed visibility) and as such she's 100% at fault. This kind of crash is fairly regular with squishies in control and the short answer is that if you put yourself in the path of a vehicle which has right of way, then you're at fault, no matter what colour the lights might be. Never assume they're going to stop until they actually have.
There is an interesting point developing here, given what we are told about the traffic conditions, approaching the junction at that speed was possibly optimistic (being polite) or blatantly dangerous. Was the traffic stream flowing through at 40mph?, if so then one would expect a closely following vehicle to be involved as well. The indications are that this was not the case and the Uber vehicle chose to approach a junction, with adjacent slow or stationary traffic restricting visibility. This in itself suggests that their behaviour modelling is totally inappropriate (not a surprise).
Of course what needs to happen is that in the same way a human driver is tested, so should an AI driver. This could be taken further, with modelling by independent "Driving Test/Instructors", regulated and approved by the authorities. Possible those same authorities should have the code in ESCROW so that in the event an incident (not accident, there is a difference) occurs, what is in the vehicle can be independently assured.
There is an overhead, but given most of the push for this is tech companies trying to make money, then it should just be seen as another expense. Personally I would not trust any tech company to not bury poor code in the even of an incident, as the ensuing court cases would be a lawyers wet dream.
This is all to come and once again is where the technology is ahead of the regulation and equally, due to the amount of money and influence these companies have, it is in their interests for that status quo to remain.
Before they can get to grips with the automation they need to get to grips with the human element.
In Manchester today I had to try hard to avoid running over an Uber bicycle courier (on a pizza delivery???) who appeared to be toying with road-assisted suicide. If they can't train the riders of pedal powered vehicles to obey the rules, what chance have they of making an automated car work?
This argument about whether the light was green or yellow or red is totally beside the point.
The fact is, the self-driving robot car did not detect (or anticipate) something in its path, or try to evade the collision.
Lets say for example you had a small child who got loose and ran into traffic - what good is the self driving car sensors if it cannot see something the size of a car? Presumably if it misses something big, it cannot sense anything smaller.
I fear the Uber self driving car is programmed with defeatist instructions: "collisions are inevitable; lets put this beta system out there and see what its flaws are". POOR SHOW UBER. DON'T LIKE IT. WON'T HAVE IT.
> Lets say for example you had a small child who got loose and ran into traffic - what good is the self driving car sensors if it cannot see something the size of a car? Presumably if it misses something big, it cannot sense anything smaller.
Ah - the omniscient AI, and anything less is a failure. Reminds me of the Marvin Minsky AI book where the AI was helping someone cross the road - it wouldn't let them because it was conceivable that an out-of-view/earshot formula 1 car (or some such) would be able to hit them in the time it would take to cross the road with the given road layout, therefore it was not safe.
The idea of autonomous cars should be to at least remove the "stoopid errors" made by so many drivers - whether it is aggressively hopping lanes causing a bump, driving too close and not being able to react in time, or just not getting tired/drunk/high. Perfection is in-achievable with meatbags at the wheel - anything demonstrably better is a plus. This scenario is almost the equivalent of the Minsky one - the sensors on the car couldn't see the approaching hazard, so they couldn't react. If a car is driving at 40 mph and a small child runs out, it's going to be bad news with either a meatbag or AI.
On the other hand, I can see rules along the lines of "only pass cars with a delta-v of less than X mph" being useful in at least urban scenarios, or have a different X for different scenarios, anyway.
"On the other hand, I can see rules along the lines of "only pass cars with a delta-v of less than X mph" being useful in at least urban scenarios, or have a different X for different scenarios, anyway."
Definitely. A delta-v in excess of 120 kph should be reserved for exceptional circumstances.
A delta-v of 110 kph is common on twelve lane roads under some circumstances. While these roads are inside urban centers, they don't really qualify as urban in my book.
In urban settings a delta-v of more than 80 kph should be unusual, while a delta-v of 60+ kph occurs in urban settings multiple times in a normal trip, on roads with six or more lanes.
@David Nash
"As I understood it, the self-driving robot car did not have something in its path. It was in the path of the other car, which hit it side-on."
Agreed - the driver proceeding straight through an intersection has the right-of-way over a left turning driver (in North America). In this case the left turning driver would be considered at fault as there was a greater onus on them to make sure they could turn across the other lane safely.
While a real person in the Uber car may have recognized a potentially dangerous situation and used more caution at the intersection, the Uber car failed to respond to the potentially dangerous situation and proceeded strictly bases on the fact that there were no cars in its path. While what it did was not strictly illegal, programming the extra analysis of abstract details that a person does would be a huge task.
Maybe when all cars are automated (and I am not looking forward to that) they can communicate enough to work out coordination of movement, but outside a closed system the AI just isn't there yet.
It's not different. Typically there are no pauses, when all traffic lights at an intersection are red. In the UK, this happens in order to provide a period of time for pedestrians to cross while all the traffic is stopped.
In Canada and the US, at a standard N-S-E-W intersection (read crossroads with lights), all of the lights are red for only a split-second. Never much more than that. After that split second, in most cases, the following happens:
E-to-W and W-to-E traffic will be held on red lights. Simultaneously, N-to-S and S-to-N traffic will be held on red lights too. Also simultaneously, S-to-W and N-to-E traffic will be shown green flashing arrows pointing left, allowing them to turn across the intersection. This period allows cars to make left turns across the intersection with relative ease. All pedestrians are held at this stage.
After a while, the green flashing left arrows for the cars turning will turn to solid amber left arrows then disappear. At the same time as the amber left arrows disappear, the main S and N lights will go green, allowing N-to-S and S-to-N traffic to proceed straight. Pedestrians can now also proceed to cross the road, in the same direction as the cars. So cars turning right, must give way to a stream of pedestrians crossing the road who, previously, were waiting for the cars turning on the flashing green arrows.
During this time, in most cases, people who wish to turn left must wait because there is now a stream of traffic that they must cross. So these turners must now wait for a break in the traffic to cross (if there IS a break).
When the main lights turn amber, cars streaming from S-to-N and N-to-S should stop, if safe to do so, allowing the turning cars to clear the intersection. It is at this point that S-to-N and N-to-S amber-gamblers tend to squeeze the accelerator in order to get through the intersection, just as the waiting turners think they should be able to complete their turn.
All throughout this, cars can normally turn right on red in most states and provinces. More dangerous, congestion-causing and frustrating, is that when the flashing green turn arrows go amber and disappear, pedestrians are then shown the 'green man' to cross in the same direction as the cars are now travelling.
In a matter of a few seconds, the above will be repeated for E-to-W and W-to-E traffic.
If this all sounds confusing, try doing it at night, in the rain, in a city that refuses to use reflective paint but insists on using implausibly reflective tarmac, whilst driving on the wrong side of the road. Welcome to Vancouver...
"Everyone knows that when you see a yellow light, you are supposed to slow down and stop."
Nonsense. What you are supposed to do is use your head. The yellow light is just a warning to expect a red light soon. The law is clear on this. If you can safely proceed through the intersection, you should do so. Slowing down and stopping unexpectedly is both annoying and dangerous.
My (admittedly extremely limited) understanding of US road rules puts rules on yellow/amber lights roughly the same as they are in the UK - which is to say that you're supposed to stop unless doing so is dangerous in some way. That's a judgement call that you might some day have to justify but certainly it should never ever be dangerous to what's behind you in any circumstance unless they're doing something they shouldn't be (and that's a universal reality).
Quoting from www.drivinglaws.org,
In Arizona it is not illegal to deliberately drive through a yellow light. A yellow light means only that traffic facing the light is “warned” that a red light will soon follow. As long as your vehicle entered the intersection or passed the crosswalk or limit line before the light turned red, you haven’t broken the law.
"If you can safely proceed through the intersection, you should do so. "
You aren't going to like this, but this is exactly backwards I'm afraid.
https://www.gov.uk/guidance/the-highway-code/light-signals-controlling-traffic
"AMBER means ‘Stop’ at the stop line. You may go on only if the AMBER appears after you have crossed the stop line or are so close to it that to pull up might cause an accident"
(And if you don't understand the difference between "you should go unless..." and "you should stop unless..." then I hope to God I never have to use any software written by you.)
I remember, one day years ago, seeing an amber light with a good distance to go so I took my foot off the throttle and pressed the brake for a smooth controlled stop. There was the sound of squealing tyres behind me. I was being followed by a driver who applied the 'accelerate on amber' rule and thought that I would too.
From what I've heard, the 'accelerate on amber' rule is quite common in Ireland.
And in the UK, along with the "it has just turned red so I will accelerate even harder". If you look at any junction/pelican crossing, the normal behaviour now is to go through on amber and only stop after it has been red for a number of seconds. Where this falls down is at a railway crossing. At that point the same idiots then get tangled in the barrier, or worse get stuck in the crossing as well and then panic at the impending instant scrap metal event.
"If you look at any junction/pelican crossing, the normal behaviour now is to go through on amber and only stop after it has been red for a number of seconds."
Around here, if a line of vehicles approaches a stop sign, the rule is that the first of a line of cars must stop at the sign, then it and TWO cars following may proceed through the intersection as one unit (if it is safe to do so). Car number 4 must then stop, then 4, 5, and 6 may proceed, etc.
I have always just assumed that drivers had generalized that rule so that the first three vehicles at the intersection when the light turns red are allowed to go through. (I don't drive -- I'm a pedestrian and I've learned not to trust drivers!)
BTW -- "Pelican crossing"...? I've heard of "zebra crossings"; why "pelican"?
BTW -- "Pelican crossing"...? I've heard of "zebra crossings"; why "pelican"?
Zebra crossings have a singular mode. In theory pedestrians always have priority. When I was a boy, this was fine, but as traffic volumes built up a more shared system was introduced- similar to traffic lights.
The cars had a sequence of green and the pedestrians had their sequence on green. The subtle difference from traffic lights being that once all pedestrians have cleared the crossing it IS OK for cars to cross on an amber.
BTW -- "Pelican crossing"...? I've heard of "zebra crossings"; why "pelican"?
Don't forget Toucan, Puffin and Pegasus Crossings. I actually saw a Pegasus crossing for the first time a few months back but alas I forget where...
Also FWIW flashing amber always means you can go but give way to pedestrians (in the UK, to be clear).
> why "pelican"?
PEdestrian LIght-CONtrolled crossing - a traffic light away from any intersection, which only exists to stop traffic and let pedestrians cross.
Toucan - is a pelican with an adjacent cycle lane - "Two can cross"
There are a bunch of other silly bird-related names.
They seemed like a good idea at the time but statistics show that there are more, and more serious injuries at these crossings than at the Zebra crossings they replaced, so they're strongly discouraged by the transport department and it's trying to encourage their removal too.
(This is mostly because drivers speed up when they see long-duration greens and try to barge through on orange even if the crossing isn't clear - unlike the USA, cars never have right of way over pedestrians (there's no such thing as jaywalking in most of the world, That was something that the carmakers managed to get the USA government to pass as part of car culture))
Because of the safety problems, London's trying to remove as many as possible - along with "Safety fencing" - which isn't at all safe for anyone if it's hit by a vehicle, and parking restrictions (yellow lines).
The statistical perverse effect of virtually all (expensive) roadside furniture intended to "improve safety" is to make the roads more hostile to pedestrians by speeding up drivers and giving them "tunnel vision" due to a perceived increased demarcation between road and footpath that doesn't actually exist. Even painting narrower road lanes or centrelines speeds up traffic. Local authorities generally refuse to admit they got it wrong, so instead of undoing a bad decision, they then add even more stuff to try and correct the increased danger, usually with even more negative results - like the old lady who swallowed a fly.
If you're interested in this level of geekery it's worth reading the "Safer London 2020" road documents.
"Around here, if a line of vehicles approaches a stop sign, the rule is that the first of a line of cars must stop at the sign, then it and TWO cars following may proceed through the intersection as one unit (if it is safe to do so). Car number 4 must then stop, then 4, 5, and 6 may proceed, etc."
Never heard of this. Where is 'here'?
In some jurisdictions, the vehicle at a stop line associated with a stop sign, after stopping, can cross the intersection when it has precedence, based on arrival time and turning rules. What any car in front of that car may do is irrelevant.
Thanks for the link. In Arizona (and AFAIK everywhere in the US), it is perfectly OK to enter an intersection while the light is yellow. What you must not do is enter the intersection after the light has turned red. If you followed the practice cited in your link, a great deal of confusion and possibly havoc might ensue. No need for the snark, by the way.
Those suggesting extra signaling to self-driving cars aren't thinking clearly about the problem.
Self-driving cars don't need advanced warning, or an external system to calculate if they will make it through the intersection. They are far better than humans at continuously monitoring the signal, detecting the exact point in time that it changes yellow, and calculating the likelihood of stopping before the line vs clearing the intersection. All while monitoring the surrounding traffic, including close-following traffic that might be driven by a distracted human.
Most approaches are also quite good at maintaining their target speed, without the human tendency to slow down when the green "feels stale" only to immediately speed up when it actually turns yellow. (This is one reason why a completely "machine learning" approach to self-driving is flawed -- it would undoubtedly amplify the worst behavior of human drivers.)
Most traffic lights are timed generously for yellows. If they weren't there would be an obvious pattern of vehicles stopping partially into the intersection.
Thanks. I was starting to worry that nobody had posted this.
Starman: Okay?
Jenny Hayden: Okay? Are you crazy? You almost got us killed! You said you watched me, you said you knew the rules!
Starman: I do know the rules.
Jenny Hayden: Oh, for your information pal, that was a *yellow* light back there!
Starman: I watched you very carefully. Red light stop, green light go, yellow light go very fast.
It gives two copletely oposite things the car WILL do if it can without specifiying what it will do if both options are possible.
Also there is a big inconsistency in the story: statements say the light changed to amber as it entered but it seems the other lanes had already stopped - why? unless the light had changed much earlier. Congestion is one posibility I suppose but in that case how was the robo car doing nearly 40mph.
I presume the other lanes had stopped othewise the other car would have not tunred in front of them or would have hit them instead.
Accidents usually happen when multiple factors come together but driving through a junction (ok intersection but I'm British) with stationary cars and restricted visibility from those cars at 38 mph is not good driving. I woudl go further and say it is dangerous driving certain to lead to an accident at some point even if legal. No doubt the other driver could have done better and technically the uber car may have had the right of way but if she checked the road she could see was clear saw the lights were orange and then turned thatis not unreasonable driving.
Dangerous driving and speeding are not the same thing (one could contribute to the other of course).
I'd be curious on the legality side in the US around dangerous driving in this case, if this happened in the UK, and speed was considered a factor in causing the accident, even if the driver was still within the posted speed limit, the driver could still be prosecuted for dangerous driving.
"Everyone knows that when you see a yellow light, you are supposed to slow down and stop. Everyone also knows that there is a grey zone when a light turns yellow as you arrive at an intersection."
In Austria the green light blinks four times before turning yellow and finally red, so you get a pretty good idea of whether you're going to make it through or should start slowing down well before the intersection (reducing the incidence of rear-end collisions).
That and the gay and lesbian themed pedestrian lights make Austria one of the coolest nations in the World, traffic-signage-wise.
I drive a new BMW with speed limit recognition; the car has a camera in the rear view mirror housing which can identify some speed limit signs, including some of the variable ones on overhead gantries. Sadly it is not 100% reliable, more like 75% by my unreliable reckoning, and even though it does display the posted speed limit in the glorious head up display (providing you're not wearing polarised sunglasses) it doesn't use it as a suggested upper speed if you push the limit button.
BMWs with 'Speed Limit Assist' are using the Mobileye's camera system.
The technology partnership was announced almost a decade ago (I found a press release from 2008), but took a few years to make it into production vehicles.
The Mobileye system has a fixed set of signs that it identifies, and that set is region specific. What it's doing is closer to pattern matching rather than general image recognition.
For fun, print out a page-sized 5 MPH speed limit sign, tape it to rear of another vehicle, and follow that vehicle around. Now image an ADAS system enforcing all traffic control devices.
Let's face it, AI and sensor input is not up to it. Christ if AI was sufficiently sophisticated to make the myriad subtle decisions required to drive a car safely in all conditions, smartphone personal assistants would be flawless. Aeroplanes wouldn't need pilots (given the technical infrastructure around aviation, a much easier but to cracking surely than driving a car). Based on the tech of today, these things will inevitably cause crashes.
I'll wait for the downvotes.....
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
