Slashdot Mirror
Programming Safety Into Self-Driving Cars
aarondubrow writes Automakers have presented a vision of the future where the driver can check his or her email, chat with friends or even sleep while shuttling between home and the office. However, to AI experts, it's not clear that this vision is a realistic one. In many areas, including driving, we'll go through a long period where humans act as co-pilots or supervisors before the technology reaches full autonomy (if it ever does). In such a scenario, the car would need to communicate with drivers to alert them when they need to take over control. In cases where the driver is non-responsive, the car must be able to autonomously make the decision to safely move to the side of the road and stop. Researchers from the University of Massachusetts Amherst have developed 'fault-tolerant planning' algorithms that allow semi-autonomous machines to devise and enact a "Plan B."
If you're not aware of the level of performance of current self-driving cars, let me break it down for you. They can't stop for construction or understand rerouting from it or obey temporary signs. They can't see stoplight colors while the sun is setting anywhere near behind them. They can't drive on snow at all. They will slam on the brakes for a piece of newspaper blowing across the road or other low density objects. They think puddles are obstructions and will slam on the brakes.
They're basically deathtraps on wheels and they don't work at all plus they're illegal in several states.
this is a HUGE pet peeve of mine! if we deployed self-driving cars tomorrow we'd see a huge drop in overall accident rates but we're not doing it out of fear of edge cases! guess what: human beings encounter unforeseen scenarios on the road all the time & have to make reflexive decisions in real time. guess what? we f up a large % of the time! if a computer can reduce overall accidents by double-digit %s I'll be the first to say I'll accept the risk of being one of the edge cases that may (or not) have survived had a human been behind the wheel.
it's like the vaccine debate - guess what? there ARE people who have bad outcomes who would not have otherwise but the overall net gain to society is so big we (rightfully) shame people who don't participate...
So the car is travelling at 60 MPH on automatic when a situation arises that requires the car to switch to human-control ... and there might be a problem with the human not reacting correctly?
I think that the problem would be expecting the human to take control and do anything useful at that speed if the programming couldn't handle it.
No one said they should be perfect. You basically made something up and then hated it. Good job.
I'm not surprised that AI experts thinks that it is unrealistic with self driving cars. The technology they work with it way to immature/insecure to be used for a function like that.
As it looks now it would have to be solved with more traditional automation methods, something that AI experts are about as qualified to speculate in as your average linguist or any expert in any other field.
Like in California where constant earthquakes sometimes open huge gaps in the roadway and present a danger to drivers?
Or perhaps they really are moving forward with fault tolerance by brib er lobbing to make it completely the passengers fault when accidents occur? I know it worked super well for credit ratings so maybe they really are fast tracking its deployment.
Who the hell thinks a Segway vehicle is suitable for high speeds?!??!? That thing is top heavy as hell and when it needs to emergency brake actually has to acc fucking ellerate before slowing down! Basically anything over 15mph and your asking for head plant brain salad with a side of ranch dressing. On the bonus side the organ donation program wouldn't see business this good since the invention of the crotch rocket.
I'm reminded of the Allstate Insurance commercial where the Mayhem actor is playing the part of the driver's GPS.
From: legal@google.com
To: larry.page@google.com, sergey.brin@google.com
Subject: Self driving cars
And you thought building a search engine created previously unheard of legislative scenarios.
Simple example to see how complex this idea is, would be just looking at CAPTCHA, anti-spam solution by google. It's only one little static image with couple of numbers. But cracking how to teach machine to read it took years.. And even now it's not 100% accurate on every case. So imagine now that you need to decode whole dynamic reality surrounding the car, reality in which everything can have similar shapes and colors. If i see drunk person walking on sidewalk i can determine if that person will maybe walk on road or not, how can computer determine that this person is even drunk? maybe it's someone that is showing other person how to dance or playing some silly game. Reality is so complex and unpredictable, so many variables, that until we get real A.I. you can forget about self driving cars. All the hype about this now is just marketing gimmick. Simple tasks like parking are doable because there is not much of computing there, just couple of sensors that can determine from reflection the distance and algorithm to count if and how to park there.... but real driving is unreachable, even if you will pack fastest supercomputer cluster in the world, i assure you that this car will make accident or injure someone. Computer can't achieve the same computation power as our brains in certain situations. It's just unreal. Ask anyone that is really working with AI and neural networks. If we would gather all computation power that was made by humankind (from computers, phones, microcontrollers etc.) it would not be enough with current our knowledge and architecture to make fully safe self-driving car.
The problem of over dependance on automation eroding piloting skill has already been addressed in the flying biz. Read about Children of the Magenta Line.
Once people give up hands-on driving experience, expect a rapid descent into complete dependence on the AI. At which time it would be better to take the steering wheel away and admit to ourselves that everyone in the car is a passenger. Even seeing a Zipcar coming down the road is enough to strike fear into the heart of the experienced driver. Here comes someone who thinks they can keep up their skill level by borrowing a car a couple of times a month.
Have gnu, will travel.
About the ethical rules that should govern decisions like saving one baby who's lying on the railway track to the left vs 5 grannies toddling across the track on the right, when you're at the controls of the track-switch.
Now someone gets to actually program these rules into a car.
Cool!
Where are we going and why are we in a handbasket?
I can see it now, the inevitable Nationwide commercial, as the inevitable lawsuits against self-driving cars occur, when they run over kids and pets who "shouldn't have been there".
Was it a hacker? That excuse won't fly.
-- Tigger warning: This post may contain tiggers! --
but certainly things will improve over time. to be honest I'm surprised at the negativity in the comments on this article.
sure AI can't handle all scenarios, and it is worse when only some cars are AI controlled, but I would expect there to be 'partial' implementations first.
what about motorways that are designed or updated specifically for AI vehicles? for example the AI does "see" the traffic light, it receives a signal from a traffic light controller or some other system.
what about roads where it is limited to only AI vehicles (then they can all talk to each other and you don't have the human drivers behaving unpredictably)
The state could embed chips into car license plates, drivers licenses, and perhaps the road and that would go a long way towards making AI controlled cars safer.
There are a lot of issues with unconstrained AI algorithms. Their effectiveness is usually evaluated by how many false positives and false negatives they incur. Little thought is given to the nature of those false positives/false negatives. Some of them could be extremely costly (deciding to run over a child) and the algorithm could still get a very high score when evaluated against common metrics.
The biggest problem is that these algorithms don't "see" anything. They simply process information and respond deterministically based on how they were programmed. The algorithms don't know anything and therefore are far worse off than a stoned or drunk driver. Even a drunk driver will recognize a dangerous and novel situation. An algorithm will not be able to handle a dangerous and novel situation unless it was specifically part of the training dataset (and there weren't other confounding observations in the dataset). These algorithms are not magic and they're not aware.
Constrained AI is the best way forward because the damage that can be done will be contained.
what criminal liabilities are you willing to risk your freedom to a shirty AI and a eula that let's them off the hook and makes you got to court on your own and you can't get the source code as well.
Wanna know how I know you've never been drunk?
You just need to have the AI auto manufacturer like google also sell the insurance to you. What could possibly go wrong?
My main concern with self driving cars isn't their ability to navigate the road, as they're already not bad at this and will only improve with time.
My main fears with self driving cars are their social and economic impacts. As this video states (https://www.youtube.com/watch?v=7Pq-S557XQU), quite a few people get paid to drive cars and trucks, and I don't see a new industry on the horizon that will provide comparable jobs for all these people once self driving cars/trucks get to the point they start replacing paid drivers. Sure, fewer car accidents, likely better traffic, and the ability to do things during a car commute besides focus on the road may be good for society as a whole, but those things are of little comfort (especially in the short term) to the people who will be put out of a livelihood by vehicles that don't need professional drivers.
That's not to say development of said vehicles should be artificially held back, but it's only prudent to think of possible negative impacts of a new technology just as much as the possible positive ones. This is something I feel gets forgotten in most discussions of automating jobs humans typically do. It'd be nice to live in a sci-fi world where machines do all the stuff humans don't explicitly enjoy doing and people just spend their time developing new technology and enjoying their lives, but I firmly believe human social and economic structures are nowhere near ready to transition to something like that peacefully.
allow semi-autonomous machines to devise and enact a "Plan B."
Actually Haskell would make this otherwise almost impossible plan easy with the "Either a or b" datatype! Thanks functional programming for making AI easy!
"Software Engineering" would have to be raised to the level of any other engineering discipline - requiring a license to practice and the responsibility that comes with all that. I damn sure don't want the level of incompetence driving my car as there is running most of the software on the internet today.
They should look at the drone vendors--same problem, with even worst failure (it falls from the sky).
you hope.
But this makes for a good movie.
Say some buys a auto drive car and it end ups killing some due to a software bug but due to Eula the owner / passenger is at fault and they go to prison and while doing some hard time they thing about getting revenge and when they get out and find the only jobs at mc jobs and they say I was better off in the join they to go Google (or some other place) and hunt down the coders / phb's and they (kill) or maybe just beat the shit out of them to get back in to the prison life.
I would totally watch that. It's already happening in a way like the guy they just released after three years due to the sudden acceleration defect he claimed was there from day 1 and no one believed him.
Especially when 2-4 drivers arrive at around the same time which is usually the case during rush hour. Two pattern emerge. Either they all look at each other and no one goes or if there are some more decisive people among them, more than one might decide to go at the same time. Very funny. 4 way stops. Worse traffic invention ever.
"Failure in brakes.dll."
- Zav - Imagine a Beowulf cluster of insensitive clods...
Rules like: stay on the road; don't sweat flying trashbags; stop at red stop lights; don't run into other vehicles; don't run off the road.
Computers are rule-based systems. They are really good at following rules. It is the only thing they can do. If this, do that. Do that X many times or until Y. So beef up the rule sets, improve the sensors, and voila, you have a safe driving system.
Now, I don't know if sensor technology is up to the task yet, or even if we have enough computing power. But the act of driving is something computers could potentially do very well, because it is a rules-based act.