Driverless shuttle in Las Vegas gets in fender bender within an hour
techcrunch.com
techcrunch.com
I think there are too many possible risks of programming the vehicle to do anything except stop dead in an emergency situation - you rapidly get into very complex programming with all kinds of failure modes - e.g. what if the problem was a faulty sensor? The vehicle might try to avoid a non-existent threat by crashing into something else. Stopping is the safe thing to do. If you want to react, you rapidly have to make lots of moral decisions like (The Trolley Problem).
Briefly sounding a horn in an unexpected emergency stop situation is probably a good idea, particularly if it's a white noise type thing, rather than a siren, so that other road users can localise it's source quickly.
Aside: emergency vehicles in the UK have stopped using the broad-band noise sirens AFAICT, I don't know why but the reason may be pertinent to any attempts to use it here? Actually, they often ride without the siren at all, which is completely ridiculous to me; they put it on at the junction meaning there's no time to get out of the way, if they used it continually you can hear them coming.
When they're putting it on at a junction and you're already in it, they're not expecting you to magically disappear out of their way. They're expecting you to look at where they're going and stay out of their path. This might involve proceeding as you were going to in order to clear it for them. What they really want is to stop anybody else from entering it until they've got through.
When the sirens are used it's great because you can tell what road they're coming along (after a while, it was much better with the "squelch" sirens) and anticipate in good time.
Agreed, waking everyone at 2am is not needed generally, but howabout alerting people at 3pm.
You appear to have inside knowledge here, presumably there's been some change in policy on this in the last - I don't know - 5 years, where it's been chosen during normal daytime callouts not to use sirens if at all possible?
Self-driving cars don't have to be better than humans at everything in order to be objectively better drivers than humans. In 2010, 32,999 people died in car crashes in the US [0]. If self-driving cars are on-balance 1% better than human beings at driving, then 329 lives are saved.
[0]: https://en.wikipedia.org/wiki/List_of_motor_vehicle_deaths_i...
Seriously speaking from perspective of a truck driver.
The reasons truck driver backed into thing are: He couldn’t see it. It wasn’t there when he started.
Horn and flash of lights would stop a human driver. When I back up a truck I use all my senses, since I can’t see much. Easiest fix for the robot thingy - add horn and lights. (And wait until all residents start complaining)
All you have to do to cement a driverless-car hegemony is to stop issuing new driver's licenses. 50 years later: no more drivers.
Some people love driving, and some people love riding horses, but you don’t see many horses on city streets today.
Perhaps that limitation should disqualify AI from driving cars.
So Navya, the French company who has deployed several of these slow moving autonomous shuttle busses does not appear to have such sophisticated behavioral algorithms, in spite of entertaining ambitions of building proper robotaxis.
To name just two potential issues that the "bus should have moved" people are ignoring:
1) What if there were pedestrians around?
2) What if moving the bus created an incursion into a stream of fast-moving traffic without warning?
Both of those could result in consequences far worse than the (apparently minor) fender-bender that actually occurred.
I agree that sounding the horn would be a good idea.
"What Makes for a Street Legal Vehicle?"
Horn – It may not seem the most important piece of safety equipment, and many big cities even limit how it can be used, but to be street legal every vehicle must have a horn that is audible for at least 200 feet. The horn can generally be any note or sound (even ones that play musical tunes are usually permitted), so long as the minimum volume requirements are met.
If a requirement for being street legal is having a horn then the assumption is that the driver has the ability to press the horn. (like having a condom sitting on a shelf, it's not effective if not used).
Not programming the car to use the horn is basically the same as not making a button in the cabin which can activate the horn for a human driver. In that case it would be not road legal.
Isn't it more akin to having a working horn but a driver who never presses it? Would a human driver ever be prosecuted for not using their horn?
If a self-driving car isn't programmed to ever use the horn, though, then that's a different situation. It's easy to prove that it can't ever use the horn because it isn't programmed to do so. That's more akin to a normal car having a horn in it that isn't wired up.
Why should that make a difference though? Imagine e.g. a truck backed into a car when the (human) driver didn't press the horn. Is the car driver liable? Whether the driver might have pressed the horn under other circumstances doesn't seem like it has any bearing on that question. We're judged on what we did, not what we might have done - shouldn't it be the same for the software?
Replace that with some other subsystem that is practically necessary for safe operation in a commonly encountered subset of driving situation requires manual activation and we require vehicles to have.
Lights, wipers, brakes, seatbelts, defrost, etc.
See the problem?
> Lights, wipers, brakes, seatbelts, defrost, etc.
> See the problem?
No, I really don't. I think a human driver would and should obviously be prosecuted for not using lights, wipers, brakes, or defrost in a situation that called for them (seatbelts have specific laws about them). It's not at all obvious to me that this would ever happen with the horn.
The point is that code that can't use those systems is no better than them not existing which is not something we're ok with.
Society accepts that it's not unreasonable for whoever calling the shots (human or not) to misjudge a situation and not use some system when they should and that this may result in a crash.
Society does not tolerate a vehicle that does not have those systems. Code that cannot use those systems is effectively the same as not having them.
Having a horn and the ability to use it, but choosing not to, for whatever reason, is very different from not having the ability to begin with.
But if your car can only press the horn programmatically, and you never add the code to allow the horn to function, you don't actually have a horn, that's the vehicle's problem.
Driver needs to have the ability to press it. More correct analogy is driver with some physical defect that renders him incapable of pressing the horn, which would be illegal. They ask you to press horn in drivers exam .
Yes, that could happen. A driver loses control of his car on a slippery, wet roadway and doesn’t immediately honk to warn vehicles around him "Failure to worn." http://southfloridainjuryaccidentblog.com/2015/04/16/florida...
That said, it would be an unusual citation.
Someday a self driving car will me "smart" enough to honk the horn or flash the lights - or send out a panic signal to other self driving cars.
Sarcasm aside, who is deemed to be in control, who do you sue for the damage if this vehicle crashes in to you - isn't it the manufacturer that causes the crash (not in this case necessarily, I'm talking generally).
Generally, the owner of the at-fault vehicle is responsible.
Assuming normal circumstances (so not considering things like me letting a child drive, or letting you drive drunk, etc.). The person/persons in control of the vehicle are surely at fault - being the owner doesn't give you power to alter the programming and so you are powerless to avoid a crash.
The company directors, who instruct the system's programmers, are the one's able to avoid the crash by making the vehicle operate effectively.
There may be circumstance where a crash could not be avoided but there are going to be plenty when alternative actions - for which there is a strong rationale - would avoid a crash too.
Generally, the driver is at fault and liable if the driver violated a traffic law; if the accident resulted from a manufacturing defect, the manufacturer is (and potentially other participants in the stream of commerce are) liable.
The owner, as distinct from the driver, may be liable in certain cases, such as (IIRC) maintenance problems.
The absence of a driver reduces the difficulty of ascribing liability by reducing the possibilities.
Actually, a defendant then suing a third party as described by the grandparent is a thing, called “contribution” [0], and I'm fairly certain there are cases where it's been applied in exactly the manner he grandparent describes involving automakers.
[0] https://en.m.wikipedia.org/wiki/Contribution_claim_(legal)
Title should be NO_TITLE because “doofus driving truck backs into something” isn’t news.
Which leads to another question. Will driverless cars report bad drivers to police? Making tech companies police informants... will laws be pass forcing this behavior?
Here's the logs of the aggressive driving, during the last 50 seconds. These logs entail a full 3d pointcloud showing that they were not responding to other drivers' potentially erratic patterns. Other driver came within 1 foot of car on multiple occasions. License plate is ABC123 and facial scan of user is $face .
User appears to have wronged our passengers with the following laws: $list
We're talking an open and shut case here. And every car with enough of these sensors could provide "dumps".
What if driverless cars reported every jaywalker too?
How many times have you had that happen when you're driving and you see someone hesitantly about to jaywalk?
Also, as a cyclist, when a cyclist doesn't ride on the road and follow the laws it drives me nuts. Are you a vehicle or a pedestrian? Please don't be both.
We'd have a high-quality signal as to where existing marked crossings were inadequate?
Human driven cars can, and their are public ad campaigns encouraging them to, and its easier for a driverless cars to do something detailed data without interfering with driving, so I'd be surprised if they didn't.
The title needs to be changed to, "Human in truck, backs up and hits vehicle (self-driving)" ?
Because that's what happened. A human caused a wreck. Doesn't seem as scare-mongery now, eh?
A persistent vision system, like 360 radar would have been fully aware of a vehicle at any direction, and would have not hit them. A human on the other hand.....
https://www.theverge.com/2017/11/8/16626224/las-vegas-self-d...
The truck driver is primarily at fault in that scenario, but I would put the autonomous bus maybe 15% at fault for not backing up or honking.
https://www.digitaltrends.com/cars/self-driving-bus-crash-ve...
This is the quote from the very article you linked to. The bus picked it up. The human driver hit it anyway. Obviously the bus' fault - it should have predicted that and stayed away from streets full with crappy drivers in trucks.
In fact, this article could start a discussion on if the risks of active accident avoidance measures (e.g. running a stop light at a clear intersection if you're about to be rear-ended) might cause more problems than they solve.
The "autonomous" cars that work are basically little trains running on digital tracks with collision avoidance systems. They are no where close to being "intelligent."
The obvious solution in this case is to get rid of the human driver of the truck that collided with the autonomous vehicle, because he was demonstrating behavior that wasn't intelligent.
Clearly this company's tech isn't up to par.
I was so surprised that I just sat on the horn for about 5 seconds even though it was all over at that point.
So, now the makers of this vehicle can code some better evasion technique and consequently all future versions will benefit, very much unlike my children, who have a high probability of reacting just like me the first time it happens to them. (Sure, maybe I will teach them, but then again maybe not.)
I wonder if sports or videogames may teach people to react in stressful situations instead of freezing, since both of those require you to apply problem solving in split second decisions.
Sorry for being off topic.
My driver training involved simulated head on collisions where an instructor would get in another car and drive at me head on to train my response. So I had reasonable training. It was just that in a parking lot I never expected someone to just throw it in reverse and back into me without looking where they were going. It simply had never occurred to me that it could happen. So when it did happen I was so busy thinking about what I could do that I didn't do any of the things I could do (honk, or put it in reverse and back out of the way.)
Since then I've avoided dozens of accidents that were new and unique in their own way, but also I've been through a few situations and have now developed good instincts about what the options are in a given situation.
It's not unusual, by the way. Pilots train emergency scenarios for just this reason.
I don't disagree with what you're saying but 3/3 times I have honked my horn as someone was backing into me and they still hit me. Another time I was a passenger when it happened.
Stopping was an evasive action, and since the result was a graze, it may have been moderately successful. While the people on the bus seem (from the article) to think there was time for more action, the perception of time under the adrenaline rush of an imminent collision may be misleading.
This may be one of the more ridiculous excuses I've seen for self driving vehicle "screwups".
Is the argument actually computers can't take action instantly compared to a human? If so, then self driving cars may as well not exist.
It's well within a computer's capability to take evasive action for an accident like this far before a human driver - assuming sufficient sensor input. This is exactly the reason I want to see self driving tech rapidly accelerate - a car that has it's situational awareness 100% at all times, and thus can react instantly in a safe manner to any situation.
I understand the software is not up to that point yet - and your response is exactly what I fear the most. I see endless DoS attacks against self driving cars once they get even a remotely substantial percentage of the market.
If all programming reverts to "failsafe stop when things may be remotely potentially unsafe" we are going to see traffic problems of a proportion we have never experienced.
>This may be one of the more ridiculous excuses I've seen for self driving vehicle "screwups".
>Is the argument actually computers can't take action instantly compared to a human? If so, then self driving cars may as well not exist.
No, I don't think that's what the parent comment meant. It means that the people who said that it didn't take proper action (that were there) may have been mistaken in thinking there was more time to act than there was.
A driverless car should be on par with the _BEST_ human driver, not the average driver, who as we know, gets into 6 million auto accidents per year.
To your point, improving on the average doesn't require driverless cars at all. You ca get that today with just driver assistance technology, improving safety technology on roads, etc - all far easier problems to solve.
It's easy to say the car should be smart enough to move -- but what if, as it moves in one direction, the object (like a truck trying to avoid the car) suddenly swerves in that direction too? Then does the car become responsible for the collision?
And of course, it feels like there could be a real-world version of the trolley problem [1] -- what if there are 5 occupants in the vehicle who will be killed by an oncoming truck, but in the only direction where it can move out of the way, it will have to run over a single pedestrian?
Glad I'm not the one having to make these kinds of programming decisions.
There should be an emergency exit option on cars like this though.
Also, this kind of story is new and interesting right now, so people will pay attention to the details, they'll make up their mind how to feel about later stories that they won't bother to read because it will be boring by then. It's amazing luck for the company to get in a "hit while stationary" story right at the get-go to help groove the narrative that driverless cars are safer than human drivers.
From a public perception standpoint it would be so helpful (dons tinfoil hat) that you could almost imagine them arranging some accidents of that kind. I think it would be a dumb chance to take, so I don't believe that's what happened, but I bet they're fist-pumping over their good luck.
- the knowledge of the world held by cars is partial and probabilistic. As it is with real humans, only with very different obscure failure modes (see for example the "adversarial images").
- if you get into this situation, something has already failed. Either in your safety practice or someone else's. That reduces the predictability of the situation and also the predictability of what your actions will do in that case.
Put those two together and you realise that any kind of "deliberate" (in the sense that one can impart "intent" to a computer system?) running over of a pedestrian is indefensible, and that real situations don't admit of neat counter-factuals like this.
My main concern with putting in code to handle trolley problems is what happens if (when) this code is tripped as a false positive? Some person walking along minding their own business when an autocar swerves into them because something on its radar looked like a crowd of people teleported in front of it
The trolley problem is "there's something everywhere inside my stopping distance, we will inevitably hit something"; yours is only a possible solution for it: "Don't hit objects [evaluated as negative]" with a fallback of "hit the brakes, horn, and do the robot equivalent of pray [ok, gonna hit whatever is in front instead what's front left or front right]".
TL;DR: Saying "let's have this one hardcoded solution to the trolley problem" does not make the problem disappear.
How much code has to be put in for all of these edge cases? How much data needs to be learned to possibly and nowhere near reliably deal with these edge cases? How often will these edge cases occur?
It's not worth the false positives, not by a long shot. You cannot make self driving cars perfect, that's not possible. You just need to make them better than human drivers.
You assume humans make correct decisions when faced with that situation. Most of the time humans make incorrect decisions when faced with far less problematic issues.
"Swerve" is normally more reflex than thought. The number of times that drivers "swerve" to avoid something like an animal and then crash headfirst into oncoming traffic, thereby injuring humans, is non-trivial.
If an autonomous vehicle does better than this at all, it's a win.
Nothing. Whoever is controlling that moving object is responsible. So, provided the car is not stopped on railroad track (in which case - what they heck is it doing stopping on railroad track?!) it should just stay there and not add variables to the situation by jerking around.
The Machine acted in a way that was not expected by the human driver.
This gets to a fundamental issue we will have in the transition to the utopian driverless future. Learning to identify and react to this new class of driver.
The article sneers that "surely this situation is not so rare that the shuttle’s designers did not allow for it?", but of course it is not that simple - there are a great many cases in everyday driving where the projected motion of the vehicles would lead to a collision if nothing was done to stop it, and so it is clearly infeasible to react to all such situations as if nothing else will be done. The passengers of the shuttle recognized the impending collision because they understood that the truck driver had probably not seen the shuttle. Approximating these applications of our theory of mind is probably the hardest part about automated driving.
I've had motorcycle lessons, one factor in those things is that their braking power is less than that of a car (less wheels, less stability, etc), so besides the emergency stop, you also learn how to do an emergency swerve - using the motorcycle's superior agility to avoid an accident. Just an anecdote. In this case, damage would've been avoided if the car had a horn and/or was more visible and/or had backed up and/or the truck had sensors.
What will a self-driving car do if for instance someone is driving in the opposite direction, towards it? Will it just continue driving, freeze, or get out of the way? Humans would likely get out of the way, even if that means "breaking the law" by going off-road or whatever. A self-driving car probably wouldn't do that, thus it would be the kind of "unexpected driving" that could really throw humans off, and may even cause more accidents than humans, in certain specific situations.
--EDIT-- it seems my comments are being misunderstood. I am simply stating the law as it exists, and the orthogonality between land and sea in regards to these laws. I know that Newtonian physics conflicts with the land aspect of these laws.
If a truck driver can complete a run in an hour and only infrequently have an accident by bending the rules, why would a “maximize profit” company give them two hours so they could follow the rules properly?
One example I saw the other morning: a trucker pulls halfway onto the road to make a left turn, where there is no traffic control other than stop sign against them. This is against the rules; they should wait for a full opening to take that turn. However, if they attempt to wait for a full opening, it could take several tens minutes for that opening to appear. For one turn.
> It also is in the companies best interest when it comes to costly collisions, insurance, disability payments, etc.
As with everything, there's a cost/benefit ratio. One minor fender-bender every year to deliver even 110% the cargo in that year? Totally worth it. Not to mention, it's usually the driver who absorbs the bigger relative share of the cost, in terms of points on their licenses or being fired for their inevitable slip-ups while under pressure.
Something I read on Reddit a few years ago that tore away the blinders of risk management forced upon truck drivers: If a driver had the opportunity to avoid a head-on by swerving into the ditch, they won't take it. By intentionally swerving into the ditch, the blame of the wreck falls on their shoulders, and if the cargo is damaged, they've likely lost their jobs. I've since confirmed it with a few questions to truck drivers I know.
The default rules of the are basically "big class of vehicle takes precedence" and "speed takes precedence within a class of vehicles".
People yield to golf carts yield to fork trucks yield for semis which yield for trains which yield to container ships.
Traffic on slow roads crossing fast roads waits until it's clear. Stop signs codify this pretty directly. Even if you removed them it's pretty obvious who has the right of way in many cases.
The rules of the road are specifically for reducing ambiguity and managing traffic flow within a class of traffic and managing interactions between those classes. For example, the procedure for a (rule following) pedestrian crossing a busy road will be very different where there's a cross walk than where there isn't. The cross walk provides a specific exception where the right of way defaults to the lighter traffic.
However, people don't crashing stuff and there's many levels of redundancy built into "the rules" so it all gets a bit fuzzy and situation dependent. (e.g. the pittsburg left or an agricultural vehicle crossing a not too busy road)
Accordingly, if you as a pedestrian or cyclist (or golf cart driver) just cross into a busy street not at a cross walk people will stop for you giving the appearance of you having the right of way.
This is why I've been saying it's not enough for a self-driving car to just be better than the "average human driver". It needs to be way better, maybe 10x better. Because these cars are likely not going to drive the way we expect a human to drive in many situations, which could end-up causing more accidents due to this conflict with how humans actually drive, even if the car is "technically" as safe as an average human driver.
Nope, the human driver was careless and did not look around, failing to notice the obstacle. If there was a human-driven car instead, we'd have never heard of it and the matter would have been resolved by regular insurance payments that happen every day in every city a thousand times. But since there's an automatic car involved now it is a philosophical conundrum. Nope, it's not, it's just some guy in a truck being careless.
Generally, not, though there are some exceptions. But that's not relevant here, because the bus stopped in an attempt to avoid/mitigate a collision with the truck, which was illegally backing into it's path, and the driver of the semi was cited for the violation.
My was a generic response to this generic comment.
> Saved a click: it stopped.
A reasonable truck driver would assume the driver of the other car would back up a little bit. But in this case, the car stopped where the truck couldn't see whether it was backing up or not, and wasn't programmed to understand the truck's movements at this angle.
So, software was responsible for: 1. Stopping too close. 2. Stopping in the blindspot of a truck. 3. Having no horn. 4. Not backing up or understanding fairly common motion by a truck.
[0] https://www.engadget.com/2017/11/09/las-vegas-self-driving-s...
In most 'no-fault' states, though, I think the shuttle would shoulder some of the responsibility.
https://www.digitaltrends.com/cars/self-driving-bus-crash-ve...
I get that it wasn't the Driverless car's fault, but this brings up an important use case that the driverless cars currently don't seem to be able to handle.
In an ALL-Human situation, what would've occurred is that the parked car (if it had a passenger inside) would honk at the car that is trying to back into it, or open the door and yell at the person trying to back up, and the accident would be avoided.
Driverless car doesn't (or didn't) honk even if it does detect something backing up into it. Hence the accident.
Self-driving cars may have a faster reaction time, but they will never reach the level of human awareness of their surroundings while driving.
Let's see a self-driving car navigate through a construction zone, watch for instructions from a police officer who is directing traffic, or stop when kids are playing baseball in a yard and the ball rolls across the street. Answering, "well they'll have that capability someday" isn't a very compelling answer. Truly self-driving cars are dependent on technology that simply hasn't been invented yet.
How nice it will be sitting behind a fleet of self-driving cars dragging their asses down the highway at exactly the speed limit, or slamming on the brakes when a leaf flies in front of the sensors.
In addition to that, you think masses of people will be silent while losing their jobs because these robot overlords are taking the wheel?
Source: I work for a self-driving car startup.
Self driving cars can simultaneously see in 360 degrees with multiple types of sensors. They already see better than humans. Now they just need to understand better, which doesn't seem far off.
Maybe you just work for a bad self-driving car startup?
It's in self driving proponents own best interest to have stringent standards, because if the public loses faith its going to be an uphill battle.
Simply demonizing human drivers and hand waving away errors is too self serving to work.
Note a subtle shift, with government now shilling for the driverless vehicles. That's the first time I've seen it; I suspect it won't be the last.
Perhaps we need some sort of placard that is legally mandated and easily visible (like the "student driver" placard) to let people around these vehicles be aware of them and expect different behaviors than "ordinary" divers.
It detected that a truck was backing into it's path, and stopped in an attempt to prevent a collision. The truck driver didn't properly check his mirrors and continue backing up which resulted in the truck 'grazing' the front bumper of the car.
What exactly is a human supposed to do differently in that situation? Attempt to continue driving straight and potentially end up being t-boned by the truck instead?
Also, this vehicle looks super weird already. If the truck driver didn't see it, a 'placard' isn't going to make it more noticeable.
Honk
Apparently, simply having the ability to stop does not make a self driving car better than a human. It should evade, or at least honk.
What kind of writing is that?
Driverless software must recognize situations like this and adapt in order to win the future. We will not go from manually operated cars to 100% driverless cars overnight. They will have to prove their worth and/or superiority alongside human drivers for a long time.
It's not the responsibility of vehicles behind that car in the intersection to fix things, human or otherwise.
You could also argue that a self driving car wouldn't be going as fast behind a car chasing a yellow light and could allow more room to handle this situation before it gets to that point.
If anything this just highlights to me how much commercial vehicles need self driving tech (I guess I, lamely, agree with the mayor).
My office used to be in an industrial district. The trucks there scare me. They absolutely do not follow the rules of the road, and it's dangerous for everybody around them.
Stuff like this: just backing up and expecting everybody to move out of their way, or taking a turn too tight and expecting everybody at the light to back up, were almost daily occurrences.
Yeah, self drivers need to account for this, but the bigger problem IMHO is getting those trucks and their drivers either off of the road or in compliance with driving laws.