MIT Creates Car Co-Pilot That Only Interferes If You're About To Crash

MrSeb writes "Mechanical engineers and roboticists working at MIT have developed an intelligent automobile co-pilot that sits in the background and only interferes if you're about to have an accident. If you fall asleep, for example, the co-pilot activates and keeps you on the road until you wake up again. Like other autonomous and semi-autonomous solutions, the MIT co-pilot uses an on-board camera and laser rangefinder to identify obstacles. These obstacles are then combined with various data points — such as the driver's performance, and the car's speed, stability, and physical characteristics — to create constraints. The co-pilot stays completely silent unless you come close to breaking one of these constraints — which might be as simple as a car in front braking quickly, or as complex as taking a corner too quickly. When this happens, a ton of robotics under the hood take over, only passing back control to the driver when the car is safe. This intelligent co-pilot is starkly contrasted with Google's self-driving cars, which are completely computer-controlled unless you lean forward, put your hands on the wheel, and take over. Which method is better? A computer backup, or a human backup? I'm not sure."

Comments

2001

[headhot]

I'm sorry David, I cannot allow you to pass that car.

Re:2001

[girlintraining]

I'm sorry David, I cannot allow you to pass that car.

Okay, that's funny, but this is more serious: "I'm sorry officer, I didn't cause that accident -- the car's controls stopped responding." Robotics are fine and dandy, we already have airplanes that can fly themselves, but those controls can be disabled or overridden in an emergency, because a computer's decisions are only as good as its programming.

Maybe the car in front of you cuts you off... and so it applies the brakes hard to avoid the collision. It doesn't see your friend's car that you're towing with just a tow rope. Crunch. Or how about this one: It's 3 am tuesday morning. You're the designated driver for a friend that's had way too much to drink because his wife just left him. You're a mile from a hospital when your passenger starts showing signs of cardiac arrest caused by alcohol poisoning. The speed limit is only 30 MPH. You try and floor it, but the car's engine keeps cutting out, saying you're travelling at an unsafe speed. Four minutes later, you make it to the ER lot, but not before you have to crawl there at 15 MPH because of the posted signs along the hospital's main entrance. It takes another three minutes before paramedics can get to your friend and start recussitation efforts. After 8 minutes without oxygen to the brain, your friend is permanently brain damaged. You could have made it to the hospital in half that time, saving your friend from a life of permanent disability. But the computer wouldn't let you. Or here's another one: You're in a bad neighborhood, driving a fancy state of the art semi-autonomous vehicle. It looks like a million bucks. Two guys in the car in front of you think the same thing and decide to pull out guns and try to car jack you. You're in traffic and there's no option to reverse as cars are behind you. Your only option is to break to the right lane and accelerate out of the kill zone... but the car detects an obstacle (the gunman) in your path, and applies the brakes. He takes aim. You keep pumping the accelerator, but nothing happens. Seconds later, rounds come through the windshield. You're dead.

Technologies like this are fine for everyday use, but in an emergency, they need to be able to be fully disabled. If they can't be, then the manufacturer needs to accept full legal liability for situations like those mentioned above.

Re:2001

[norpy]

It must be interesting living in your fantasy world.

Re:2001

[Shihar]

Oh! I have one! You are driving peacefully down the road when a driver blows through a red light, crashes into your car, and kills your entire family and leaves you a paralyzed from the neck down. It turns out that the drive that plowed into was drunk. He decided to drive home anyways.

I'm going to leave it to you to decide which is more likely, terrorist assaulting your car and only your awesome human driver skillz saving your life, or random drunk crashes and kills you.

Re:2001

[HungryHobo]

It comes down to likely scenarios vs unlikely.

The chances of you killing someone because you've had a late night and your attention drifts while you're on your way to work are quite high in comparison to the chances of a team of assasins coming after you while you're sitting in traffic.

Hell the chances of you having a psychotic break and merely *believing* that's happening are probably higher than the chances of it actually happening.

The chances of someone you're with going into cardiac arrest or similar while you're not so far away from a hospital that it makes not difference anyway but not so close that you can get there on time anyway are stunningly small and even if that happened don't forget that you very well might kill a few bystanders while playing at being a rally driver on your way.

  and it would be surprising if such a system enforced speed limits.

yes we can come up with insane scenarios, "if ninjas appeared right then" or "what if an attack chopper came after me just as I was about to..."

but they're not likely. most deaths on the road are due to mundane causes like people not paying attention or daydreaming about terrorist attacks and how they'd escape.

Re:2001

[TemporalBeing]

If you live in the inner city - places like Paterson, NJ, downtown Los Angeles, CA, Washington, D.C, - then the chances of a carjacking as described (especially for white folks in a black neighborhood) are actually pretty high, and can very well be as described by the GP.

Another way to look at it - a police officer has one of these cars, and gets called in to duty in an emergency but can't use the police issued vehicle for some reason (e..g left it at the station, what-have-you). He's got to take the car but it refuses to do what he wants it to, and makes it to the scene too late to be of any help - whether someone had a heart attack, was a victim of domestic abuse, attempted burglary, etc.

Google's solution is far better from the point that the human overrides the computer (same with Boeing) so in an emergency it is always clear what the vehicle will do. The MIT solution (much like Airbus) makes the computer override the human so you never really know what the vehicle is going to do until it does it, and you might have to spend some precious time telling the computer to take a hike so you can do what you need to do to resolve the problem. In an emergency every moment counts.

Re:2001

[ViperOrel]

Hell the chances of you having a psychotic break and merely *believing* that's happening are probably higher than the chances of it actually happening.

Best comment ever! I wish I had the karma to give to you...

This is probably a better start

[ranton]

While fully autonomous cars may be the more desirable future, computer backup systems like this are a more likely first step. Once people start getting used to cars making good decisions on the road, they will be more willing to give the computers even more control.

Re:This is probably a better start

[purpledinoz]

I can't wait for fully automated cars. It would relieve us of another monotonous task. But I guarantee this will meet great resistance from those who want the freedom to cut people off and talk on their mobile phones while driving.

Re:This is probably a better start

[WillDraven]

I would like a combination of both approaches. Full auto for when I want to turn my seat around backwards and play poker with my friends in the back, manual control for when I want to zip though some fun curvy roads, with emergency computer takeover when I forget that I'm not in a formula one car and start to do something stupid.

Re:This is probably a better start

[__aaltlg1547]

As much as I don't like the idea of turning control of my car over to a computer, I like Google's method better. I think it's much safer. MIT's system is more likely to get implemented sooner and I think that's pretty scary. If people start trusting *something else* to get you out of dangerous situations, the immediately respond by relying too much on that trust and putting themselves in more dangerous situations. People with these things in their cars will drive like complete idiots and the computers won't be able to avoid ALL of the dangerous situations. Probably in the end about the same number of crashes per driver-year will occur, but in the meantime, people will have forgotten *or never learned* how to drive a vehicle safely without computer assistance. Also, just because your auto-driver takes over and protects YOU from the consequences of your stupidity, it doesn't mean that you didn't cause another driver to react to your being out of position or going too fast and have an accident.

Something would have to be done to reduce this effect. First, you should have to drive unassisted to qualify for a license, or in an auto-driver assist vehicle with the machine making a log of the times it "came awake" because you were doing something foolish or illegal. Once you're a licensed driver, the car could emit radio message each time you do something stupid and the police could be alerted. This would act as a disincentive to be a complete shit while driving, especially if there's a threshold above which you will be fined for dangerous driving. And such a system would have to be tamper resistant so the typical driver can't turn it off.

Re:This is probably a better start

[forkazoo]

While fully autonomous cars may be the more desirable future, computer backup systems like this are a more likely first step. Once people start getting used to cars making good decisions on the road, they will be more willing to give the computers even more control.

Yeah, I think it's obvious that in the long term, the way forward is to ban the option of human control on most roads. Human driving should be relegated to a novelty passtime on specially designated roads, scenic routes, and racing tracks. Any transportation infrastructure being build today should really be built with the assumption that it will eventually be used only by robotic vehicles at very high speeds. Given how far we have come in the past ten years with the technology, I fully expect that we could do a 100% switchover within twenty years if we gave a damn. But, as a society, we aren't really that forward looking, so nobody is currently talking about that in a serious way at this point. Consequently, 'transitional' technologies like the robotic crash preventer will be the only thing widely deployed in the near-medium term, and we'll still be talking about fully automated transportation as a futuristic idea twenty years from now.

Re:This is probably a better start

[Yvanhoe]

No. Fuck that. I am tired of baby steps. I won't buy a new car just for this kind of equipment. Give me a full automated car, you'll have a selling point, but a copilot that only works in case of crash is not a selling point for me, and I suspect I am not alone.

The thing is, the technical challenges are all solved since 2009. And even since the 1990, we knew how to build a car (or more interestingly : trucks) that can follow autonomously a human-piloted car/truck. This technology never caught up because the law makers never made the appropriate changes.

So fuck baby steps : make something which usefulness is obvious and that would force politicians to understand that now is the time to allow these things spread and improve.

Re:This is probably a better start

[Jane Q. Public]

"While fully autonomous cars may be the more desirable future, computer backup systems like this are a more likely first step. "

And once put into production, they will be recalled and shelved for 10 years due to suspicion (and legal accusations) that they actually CAUSED some serious or fatal accidents.

Tools like this need a LOT of proving before they will be generally accepted.

Re:This is probably a better start

[Shihar]

I don't think that there would be much to lose if computers took over which shit was going down. When shit hits the fan you have a second or two to respond. Your response is going to be pure instinct. It is probably going to be wrong.

I live in the north east USA. It snows a lot. People should be good at handling a skidding emergency, yet I watched a 30 car pile up a few years ago because a dozen drivers did the exact same thing when they hit ice on the highway and realize that the brakes were doing literally nothing. I watched a dozen cars blink their brake lights a few times, realize it wasn't working, realize that they were going to hit the cars that were stopped or already crashing in front of them, and then proceed to do literally the stupidest thing you can do. They braked and turned on ice. This is a surefire way to put your car into a spin, and spin they did. I got through a massive accident because I was one of the few people that could get a hold of myself long enough to realize that as much as I wanted to both stop and turn, only turning was going to work, and I gently steered my way through the accident without touching the breaks.

A computer would have handled it vastly better than most people. A computer would have realize that breaking was stupid, and then tried to coordinate gently moving out of the way. At worst, the computer would at least realize that if all else fails, a straight rear end impact as vastly safer than going into a spin.

People can't be trusted. As soon as computers can do it better (which shouldn't be all that hard), we should surrender control and let the computers do it. If that means that every time you hit ice and break the computer steals control, so be it.

Re:This is probably a better start

[mykro76]

The trouble is, people are lazy. If given a car like this I would just be, or pretend to be, asleep/drunk at every opportunity, letting the car do all the driving.

So you end up with a "fully autonomous" car anyway, just not a very good one.

Better to aim for perfection right from the start, Google Car style. (How very at odds with their usual ship-the-beta approach!)

Re:This is probably a better start

[adolf]

But having fun on twisty backroads is probably the exact same sort of activity that something like this is likely to interdict: It must be conservative in its approach, or it won't perform properly during the normal circumstances for which it was designed.

I'm not sure that it's possible to properly have fun in a car with a robotic safety committee watching your every move.

And never successfully navigate in snow or mud: The system would go into "ZOMG!" mode, decide that what you are otherwise-successfully doing is impossible, and slow down until the car stops...probably while encouraging you to call AAA for a towtruck, when all you really want to do is get up the fucking hill.

Re:This is probably a better start

[WillDraven]

Unfortunately, what we will probably get is paranoid full auto, boring slow semi auto, and your insurance company has just been notified electronically and your policy cancelled full manual control.

Re:This is probably a better start

[Gr8Apes]

And never successfully navigate in snow or mud: The system would go into "ZOMG!" mode, decide that what you are otherwise-successfully doing is impossible, and slow down until the car stops...probably while encouraging you to call AAA for a towtruck, when all you really want to do is get up the fucking hill.

I hadn't thought of this, but my favorite tactic when we get sheet ice that's slicker than oil on our hills was to move over to the unpaved shoulder and just smoothly drive up the hill. It's pretty funny to see the looks of people trying to push each other up the hill on the ice. It also helps to have 4 wheel/all wheel drive.

Re:This is probably a better start

[WillDraven]

You could say they are difficult or maybe even impossible with our current technology, but obviously not impossible in general. Haven't you ever heard a driver say "I almost lost it in that turn." If a human driver can realize he's about to screw up and correct for it then a computer can too.

Re:This is probably a better start

[SCHecklerX]

Subaru is working on some of this stuff. I'm actually looking forward to adaptive cruise control, and wish I had it right now. Having to conatantly turn it on, off, increase it, decrease it defeats the purpose. Set and forget and let the car automatically slow down because of the idiot in front of me will make long trips much more pleasant.

Re:Much better than Google's approach

[headhot]

I'm not certain but I'm pretty sure computers are landing airplanes with the pilots overseeing the process.

I also find it hard to believe that a computer cannot get better at driving a car the most people. Sure there are emergency situations the require extreme skill and judgement calls, but how many people are good in those situations? I have seen many drivers who react 100% wrong in dangerous situations. They don't understand the dynamics of the car and get confused in a panic. Computers don't have this problem.

Volvo XC60

[marcroelofs]

Don't all Volvo XC60's have this (as an option). Only difference is it wakes me up when it thinks I'm sleeping instead of it taking over 'until I wake up' LOL.

Re:Volvo XC60

[agw]

Many VWs have this standard or available as an option, too.
E.g for the Passat you get the tiredness detection, lane assistance (keeping you in the lane) and front assistance (keeping distance or doing emergency braking at city speeds).
It's probably more basic than the MIT stuff, but always expect next years models to have more and more advanced versions of those.
Commercial systems evading obstacles should be available in 2-3 years (e.g. from Continental).

Re:Volvo XC60

[GuldKalle]

the "detect sleeping" is not the interesting part here...

Fast Lane

[headhot]

I would be all for this if the computer would take over once it determines you are driving too slow in the fast lane and blocking traffic. Maybe there can be 2 modes, emergency take over, and 'Nag' mode for when the computer determines your acting like a selfish asshole.

Re:Fast Lane

[AchilleTalon]

All automated cars are the best choice to solve this problem. No more emotionnal decisions will taint driving decisions. And fast lanes will lost completely their significance since the flow of trafic can be accurately controlled and trafic jams/slinky effect completely avoided.

Re:Fast Lane

[Ichijo]

I hope this computer resurrects the lost art of using the turn signal.

Re:Fast Lane

[Anonymous Coward]

No more emotionnal decisions will taint driving decisions.

On an individual level, sure. We'll still have groups appealing to emotion and lobbying representatives to slow us all down no matter how safe the system is at speed.

Re:Fast Lane

[theripper]

Why would I want to warn my enemies of my intentions?

Re:Fast Lane

[abigsmurf]

Personally, I'd rather the computer be used to not unlock the doors to anyone using the term 'fast lane'.

If someone's overtaking, they have the right to be there. People treating the passing lane as a fast lane, doing 30mph over the limit and making it hard for people to safely pull out to overtake cause far more tailbacks thank people 'only' doing 5mph faster overtaking. Don't get me started on the twats who instantly tailgate anyone who has the audacity to be infront of them.

Re:Fast Lane

[hawk]

whether it is the "passing lane" or the "fast lane" is a matter of state law, and varies widely. In most of the West, it is most definitely not a "passing lane."

either way, someone in it below the posted speed limit in good conditions is usually committing an offense of the same level as speeding.

hawk

Re:Much better than Google's approach

[purpledinoz]

I disagree. Human drivers are always a disaster waiting to happen. Computers don't get drunk. Computers don't get angry. Computers don't get sleepy. Computers aren't trying to impress a woman. (At least not yet...) Sure, computers fail, but humans fail too, but much more often. My concern is with the cases where a malfunction occurs in the system, maybe a broken sensor. How does a computer driver respond to these scenarios, which are guaranteed to happen in the real world?

a thought experiment.......

[Anonymous Coward]

Flip it around for a moment...
In the today world one of the most difficult tasks for pilots is to assume control from the auto -pilot when systems problems cause the auto pilot to lose the ability to control. Transitioning from "passive monitor" mode to active control is extremely difficult - and not unusually fatal.
BUT
the robotic backup does not suffer from distraction or an inability to correctly process valid inputs. My choice - robots get my back

Re:a thought experiment.......

[yesteraeon]

Absolutely correct. Investigations of aviation accidents involving autopilot errors, and the study of human factors more broadly, have shown that humans are really bad at passively monitoring while automated systems perform all tasks. This should come as no surprise, I suppose, as doing so is incredibly boring making it very difficult to maintain focus and extremely easy to become distracted.

Re:a thought experiment.......

[jbwolfe]

You may be referring to AF447. An important distinction in that case was that the automation lost the data required to control the aircraft, so as designed, it disengaged and informed the crew that they were now in control- of an aircraft in severely degraded operational state (direct law) and also in instrument meteorological conditions. Those conditions were what made it so challenging, but just transitioning from passive monitor to active control is not inherently difficult.

Re:a thought experiment.......

[jbwolfe]

Brings to mind Northwest Flight 188 from San Diego to Minneapolis. The degree of distraction here was extreme to say the least. Though one peice of automation that was not installed would have saved them from this mistake: most new transport style aircraft models include a system that warns the pilot that they have not interacted with the aircraft in a specific amount of time. Like the following:
PILOT RESPONSE: Advisory message after 30 min in cruise, 8 min in descent.
PILOT RESPONSE: Caution message & beeper after 35 min in cruise, 9 min in descent.
PILOT RESPONSE: Warning massage & siren after 40 min in cruise, 10 min in descent.
Since the automation is so capable it should include features like this to keep the humans honest...

Re:a thought experiment.......

[cellocgw]

i>You may be referring to AF447. An important distinction in that case was that the automation lost the data required to control the aircraft, so as designed, it disengaged and informed the crew that they were now in control- of an aircraft in severely degraded operational state (direct law) and also in instrument meteorological conditions. Those conditions were what made it so challenging, but just transitioning from passive monitor to active control is not inherently difficult.
I seem to recall reading several articles on plane crashes, and the overwhelming forensic consensus was that pilots who've been hands-off (autopilot functioning fine) for an extended period are in fact rather bad at evaluating the current flying conditions. This led to poor decision-making and often directly to the crash. When a computer takes over from a human, assuming the algorithms have been running "in the background" all the time, they should be able to generate a proper response to the situation. Personally, I vote for computers taking over from humans rather than the other way around.

Idiocracy in action

[wrencherd]

From TFA:

There is also the “deskilling” issue, where eventually no one knows how to drive a car (or fly a plane). This isn’t so bad if every car on the road is autonomous, and if steering wheels are removed altogether, but the in between period could be tricky.

If all cars on the road are autonomous why don't we just have trains, light rail and subways?

Re:Idiocracy in action

[Anonymous Coward]

Because none of those are point-to-point, to your home and place of work especially.

Re:Idiocracy in action

[yesteraeon]

I agree with you up to a point. However, make no mistake that there is a significant difference between a car I don't have to drive and the modes of public transit you have cited. To wit, a car (driven by me or a computer) will take me directly from A to B. No walking, no changing lines, etc. Aside from the fact that most people are incredibly lazy (I'm including myself in that number) the difference in time and convenience is significant. Yes there are cities where that difference is quite small (NYC, London, amongst a few others) but these places are very much the exception and not the rule.

Re:Idiocracy in action

[Anonymous Coward]

People movers yes. Trains, Light Rail, Subways, Buses go where they want to, not where I want to, and they waste my time at transfer points. People movers are more like automated taxicabs. They might work from station to station, but at least there is not 15-to-60 minute layover at one or more transfer points.

Re:Idiocracy in action

[ultranova]

If all cars on the road are autonomous why don't we just have trains, light rail and subways?

Rail is more expensive to build than asphalt. In fact, where I live, some roads are still just flattened dirt surfaces, and will likely stay that way in the foreseeable future. Also, light rail is not very fast, and the lighter the vehicle the greater the chances of derailment. Not to mention that it's impossible to make emergency dodges when you're on rail.

Re:Much better than Google's approach

[Joce640k]

I'm not certain but I'm pretty sure computers are landing airplanes with the pilots overseeing the process.

There's not many obstacles to avoid up in the air. On the road there's dozens of other cars all around you.

Confusion

[vlm]

Which method is better? A computer backup, or a human backup?

Both fail because both exist. Accident reports will be full of "I thought the computer was driving" and so forth.

Also any time there is none the less an accident, "its the computer's fault"

Re:Confusion

[Grygus]

Very true, but if the total number of accidents decreases, is this important?

Re:Confusion

[Anne Thwacks]

It matters to the lawyers, and the lawyers are the ones that matter.

Re:Confusion

[vlm]

Very true, but if the total number of accidents decreases, is this important?

Hmm. That was my point. It'll make driving during pre-accident conditions more difficult therefore more accidents. During stressful situations you have to add the extra dimension of "who's driving" and "I have yet another thing other than myself to blame if it goes wrong" and "This might be dumb, but if I have no computer, their computer will none the less prevent the accident"

Re:Confusion

[ShanghaiBill]

Accident reports will be full of "I thought the computer was driving" and so forth.

Simple solution: The computer should take control anytime it detects the human is not actively controlling the car. But even when the human is driving, the computer is assisting. When you drive a "normal" car, the car will drive straight ahead unless you turn the wheel. With computer assisted driving the car will stay in its lane if you do nothing, and you need to actively steer it to do otherwise.

Also any time there is none the less an accident, "its the computer's fault"

Computer controlled cars will save the data from all sensors, including cameras (external and internal), human brake pressure, human steering, etc. There will be plenty of objective evidence to determine who is a at fault.

Re:Confusion

[Grygus]

I did not miss your point. However, if the computer does actually prevent accidents, which I assume would have to be the case for people to become so complacent, the number of accidents may well decrease even with lowered human alertness, i.e., the type of accidents that are increased may be fewer in number than the type of accidents mitigated. In that case, lives are saved, damage is lessened, and it is an overall gain. Your position seems to assume that the computers do not actually work.

Re:Much better than Google's approach

[Joce640k]

Computers don't get drunk. Computers don't get angry. Computers don't get sleepy.

And computers absolutely will not stop, ever, until ...... ummm, until you arrive at your programmed destination.

Trolley problems?

[JoshuaZ]

There's a whole class of philsophical problems about when to save one life v. n lives http://en.wikipedia.org/wiki/Trolley_problem. One very awkward thing about this is that advanced emergency driving systems may need to address questions that we are fundamentally uncomfortable answering or discussing. Should a system for example protect the life of the people in a car as opposed to the life of people in a nearby car that they might crash into? Which gets higher priority. Does the number of people in each car matter? Exactly what the cars do in the few seconds leading up to a crash could alter this. Essentially this sort of thing may force us to examine difficult ethical problems.

Re:Trolley problems?

[OzPeter]

Should a system for example protect the life of the people in a car as opposed to the life of people in a nearby car that they might crash into? Which gets higher priority.

That was part of the angst of Will Smith's character in the I, Robot movie. A robot logically decided to save him rather than attempt (and probably fail) to save a little girl - a choice that deeply conflicted with his (and probably most peoples) morals.
 
While this was a functional account, I think it does a good job of showing some potential issues with life and death decisions that aren't made by humans.

Re:Trolley problems?

[martin-boundary]

All these types of ethical questions are fundamentally flawed. They simplify the world into a binary scenario where there is a choice between two hypotheticals, which are themselves far from certain.

In a realistic example, there are more than two cars on the road, and the machine is dumb, and while there is a remote possibility that the people in the first car could be saved from the uncertain possibility of some accident whose exact unfolding is unpredictable, and there is a remote possibility that the people in the second car could be saved by omission instead due to being also spared any unknowable side effects, in actual fact the machine will very likely fail unspectacularly to deal with the main event due to any combination of factors including tight physical constraints such as time and momentum, incorrect calibration of sensors and the machine learning system, and generally the inability of simple control systems to deal with unforeseen and unanticipated events.

The fact is that ethics has no role to play here, because neither of the two proposed alternatives occurs except in the mind of the philosopher.

Re:Trolley problems?

[JoshuaZ]

Once these driving machines become marginally functional, the questions will matter. If they fail as badly as you envision then aren't ready for actual implementation yet. But when they are implemented in real life, this will matter.

Re:Trolley problems?

[Anonymous Coward]

If cars ever do start comparing cargoes and weighing lives, expect to see aftermarket modifications that make your car always report that it's carrying a cartload of orphaned kittens.

This question already shows up in the real world when you are buying cars - do you want a smaller more manoeuvrable car, with well-engineered crumple zones that protect both sides in a collision, or do you want a huge block of steel that makes you more visible, and that can compete in a mass contest with whatever is unfortunate enough to crash into you?

Re:Trolley problems?

[wonkey_monkey]

While this was a functional account,

Also a fictional one :)

Re:Trolley problems?

[DaveGod]

The problem is not merely that we're uncomfortable discussing such moral quandaries.

A driver may make his own choice to put himself into a certain-death brick wall in order to avoid probable death of self + the other car's passengers. Nobody and nothing else ever has the right to force it upon him. Irrespective of how logical and regardless of any "choice" previously input into the auto-auto preferences screens.

Re:Trolley problems?

[maxdread]

Not to mention how aware will it be of its surroundings beyond other cars on the road? Last thing I want it to is steer me clear of hitting another car only to toss me off the edge of a cliff.

Re:Trolley problems?

[ultranova]

A driver may make his own choice to put himself into a certain-death brick wall in order to avoid probable death of self + the other car's passengers. Nobody and nothing else ever has the right to force it upon him. Irrespective of how logical and regardless of any "choice" previously input into the auto-auto preferences screens.

The problem is, the same applies to the other cars passengers. And Joe Driver had a choice to ride with whatever automation his car has, while the people in the other car had no choice over what automation Joe's car has. So why should they suffer just because Joe was a cheapass and chose the last years model with insufficient processing power to figure out a way to get out of this situation without harming anyone?

And this gets us to the real problem with self-driving vehicles: a vehicle that only thinks of its owner is unacceptable in public roads because it's a public hazard, and a vehicle that thinks of things besides its owner is likely to be unacceptable to a lot of people because people are selfish.

Of course, we could simply demand that the auto insurance pays a (huge) fee for each person killed in an accident involving self-driving cars and let the insurance companies sort it out. You want a car that prioritizes you over others, fine; how much extra are you willing to pay to be allowed in public roads in it?

Re:Trolley problems?

[perryizgr8]

i think will smith's character was wrong about being angry with the robots. the programmers made the decision he is angry about, not the robots. this behavior could have been fixed with a patch, if will smith had talked to the right guys at usr and made them understand.

Re:Trolley problems?

[martin-boundary]

The difference between model and reality is exactly the crux here. My point is that some models are so simplified that they become literally irrelevant. Unfortunately, philosophers often have not the scientific background needed to vet their proposed simplifications against reality. Whenever that happens, any time spent studying the model is wasted.

As Einstein said: models should be as simple as possible, but *no simpler*.

Re:Trolley problems?

[martin-boundary]

I don't think so. Consider a related problem where a train is equipped with a camera to see if there is an obstruction on the track, and an AI system which can automatically decide to halt the train. Such systems certainly exist, and differ from the smart car example only in the number of dimensions available for movement (the car has two directions available, while the train has only one).

By your contention, the camera/AI system is ipso facto making an ethical choice about the life and death of a person who happens to be standing on the tracks vs the risk of accident or death of a traveller in one of the wagons who needs to go to hospital immediately (or else we do, by deciding to build it).

But that is ludicrous. The system merely solves a problem about how strongly to apply the brakes. There is no ethics invovled whatsoever, nor any choice about life and death. Merely a very simple control problem. We can certainly ask what can be done about this particular problem in general, eg how to prevent people from standing on tracks etc, but clearly the actual train/AI (and whether we should build them or not) has no ethical role at all in this.

The fact is that the statement of the problem here (a person standing on the track while a traveller may die from stopping the train) is independent of the train/AI aspect, which is just a detail. Making it *about* the train/AI is inappropriate.

Re:Trolley problems?

[JoshuaZ]

But the programmer still needs to decide what policies to give the AI. And as we can make more and more complicated driving AI, what our options will be to program them will become more broad. The ethical issues will be an issue.

Re:Trolley problems?

[martin-boundary]

Are you suggesting that going from one to two dimensions is a sufficiently complex leap to require ethical input? I'm not sure from your response if you realize what it takes to program AIs. There are no issues of ethics, merely questions of databases and geometry.

Consider the train problem. The programming task is to read the camera image, compute the two lines which represent the track, and check the distribution of colours in between the two lines on the image. If a region is detected whose colour distribution is both sufficiently different from its surroundings, and sufficiently homogeneous within itself, that indicates an obstruction. When an obstruction is detected, the projected height of it on the viewing plane can be used to infer the distance, and the extent of it implies a certain size. Once the distance is known, the current speed of the train indicates the amount of breaking required. Ethical policies are neither needed nor helpful, they are simply irrelevant.

Trucks already have this

[tomhath]

I've heard truck drivers complaining about systems like this. Apparently it has more control over the engine speed than the driver.

"keeps you on the road until you wake up again"

[kakaburra]

it doesn't wake u up??

Re:"keeps you on the road until you wake up again"

[Phrogman]

I suggest a prerecording of Bagpipe music, piped (no pun intended) through the stereo system at about 130 db, but only if the computer can detect and react to heart attacks as well :P

Re:"keeps you on the road until you wake up again"

[c4tp]

My car better not wake me up unless it has a damn good reason, such as driving by a natural wonder, a beautiful woman, or a Chick-fil-A.

Re:Much better than Google's approach

[Anonymous Coward]

I'm not certain but I'm pretty sure computers are landing airplanes with the pilots overseeing the process.

There's not many obstacles to avoid up in the air.

It's possible that you've missed the significance of the word "landing".

Re:Much better than Google's approach

[drinkypoo]

My concern is with the cases where a malfunction occurs in the system, maybe a broken sensor. How does a computer driver respond to these scenarios, which are guaranteed to happen in the real world?

The only thing that the computer can't be designed to cope with is complete hardware system failure. Are the automotive companies really prepared to put dual systems in the vehicle with backup power? And for that matter, are they going to be willing to disable the vehicle if a sensor is out of commission? They will really need to do that because drivers will become used to depending on the system.

Easier said than done

[Manip]

Anyone who has been paying attention to the "safety systems" similar to this on commercial aircraft should know that development of systems such as this always have unintended consequences. Even if they work flawlessly the flawless function could still potentially be dangerous.

Just as one example: sometimes "crashing" is the least-bad alternative available to a driver. Given the choice between hitting a person standing in the road or a row of water-filled barriers many drivers would correctly choose the barrier over the human. But this safety system will likely subvert that and take the choice away from the driver.

Re:Easier said than done

[Skinkie]

Now isn't exactly this kind of reasoning some system could be always prepared for, while the driver has need to make these kind of decisions in split seconds. Multiply this split second by not breaking, the number of choices for all parties to be safe is reduced. It would be even more interesting what would happen if two cars with this system could cooperatively "crash". Hereby saving a third party. A more complex choice would be preventing a lethal accident for multiple drivers, while in any other case all drivers would be lethally injured... now compute those chances.

Re:Easier said than done

[AchilleTalon]

You assume the safety system will make a bad decision in a particular circumstance while it can do the right choice, or even avoid completely this situation which often arise because someone failed to recognize and identify an obstacle in advance. It would be pretty easy for an all automated system to identify humans on the road from everything else well in advance and even looking on the side of the road well in advance for such obstacles to manifest. Something a human driver cannot do.

Re:Easier said than done

[aXis100]

I'm not sure if you've noticed, but humans have an attocious track record at making the correct decisions whilst driving. Most drivers simply arent trained to deal with emergency situations and have absolutely dismal respose under pressure.

Whilst it may not be ideal in some specific cases, I would bet money that a computer will on average save many, many lives.

Re:Much better than Google's approach

[Anonymous Coward]

I disagree. Human drivers are always a disaster waiting to happen.

Sounds like you'd better stay in your mommy's basement, where it's nice and safe.

Re:Much better than Google's approach

[jbwolfe]

I'm not certain but I'm pretty sure computers are landing airplanes with the pilots overseeing the process.

Correct. However, it requires a pilot to program and monitor its progress as well as very specific requirements for onboard equipment, crewmember training and triple redundancy in the event of malfunctions. I've had numerous Cat III approaches to a safe landing and it works but I wouldn't say the computers are better than the pilots. Its only used when there is not adequate visual reference for the pilot to do it. After the aircraft finishes its rollout in a straight line using ILS, the pilot still has to find his way to the gate with visibility at only a few meters.

Re:Much better than Google's approach

[jbwolfe]

Have you not seen the Terminator series? Skynet felt the same way about humans...

"Remote" - we all know what that's code for...

[Hognoxious]

Lag is the least of your worries. http://www.youtube.com/watch?v=_B7J7yjBq8Y

Re:Much better than Google's approach

[Milharis]

Auto pilot for landing exists, but it requires ground equipment that is only available in the biggest airport, and it's only installed in the biggest airliners.
The vast majority of landings are done manually by the pilots, while the autopilot is sometimes used in extreme conditions (fog especially).

Would it enforce speed limits?

[cnaumann]

Would it take over if you were attempting to drive 90MPH through a residential zone? What about doing 35MPH through a residential zone?

Which method is better?

[NEDHead]

I believe this very question distinguishes Boeing and Airbus and their autopilot philosophy. IIR, Boeing says the pilot is the senior authority, Airbus prefers the computer's judgement. Note the similarity in the sounds 'airbus' and 'skynet'.

Re:Which method is better?

[vakuona]

I quite like Airbus' philosophy. Most plane crashes are caused by pilot error, so having system in place to reduce the number of decisions they have to make in the cockpit can only be a good thing. I want pilots to do what computers cannot do, which is to reason out difficult situations.

Some pilot aids can be potentially dangerous, but only because at times pilots are not trained well enough to know their equipment. One involved an MD plane where the aircraft were fitted with automatic thrust restoration, which was a really bad idea once an engine started surging during takeoff, and you really needed to reduce thrust to save the engine.

Re:Which method is better?

[zippthorne]

I rather don't like the portion of Airbus's philosophy wherein the autopilot can pass the buck back to the pilots if some of the instruments are not working as expected, though...

I can't imagine any situation where the available instrumentation would be inferior to the pilot's sensory experience in a small compartment with tiny windows at the end of a long tube that pivots about at the other end other than failure of all of the instruments....

Re:Which method is better?

[Cassini2]

The Airbus approach is fundamentally flawed. Pilots adapt to how the plane usually works. If the plane usually works in a manner that the pilots can't make mistakes, then the pilots get used to never making mistakes.

When the automatic system quits, the pilots don't have the ability to instinctively react and fly the plane. The result is Air France Flight 447. The pilots flew a perfectly good plan into a stall, and never corrected. Had the copilots been used to flying in full manual, then they would have had the experience and instincts to react to the stall.

People make mistakes. You have to let them make mistakes and let people learn from them. Safety systems that let people repeatedly make mistakes are dangerous. Because, sooner or later, a person will make a mistake in a corner case that the automated system does not catch. When this happens, tragedy often occurs.

Re:Which method is better?

[subreality]

I agree it's flawed, but not fundamentally so. In the Airbus model where there is heavy computer assistance, the pilots are certainly more disoriented when Things Go Wrong badly enough and the plane finally admits it doesn't know what it's doing and starts progressively giving more control over to the pilots (Alternate Law 1 and 2, Direct Law). However, with all the computer assistance, Things Go Wrong less frequently.

Airbus experiences more computer-failure crashes, whereas Boeing experiences more pilot-error crashes. The net safety record for both is about the same.

I think it's not fundamentally flawed because while there's a limit to how much you can improve pilot error (the job is pretty routine and boring 99.9% of the time, and while good training usually takes care of the 0.1% case, humans are simply going to have a limit to how quickly they can reorient in an emergency), the computers can be continuously improved as flaws are found.

Re:Which method is better?

[amorsen]

They quickly ended up in a stall with approximately zero horizontal speed. This is not something you would deal with by instinct, it is so far out of normal flying conditions that only training can teach you what to do.

There were two major problems with the way the systems work on that plane. The alarms went silent when the computers were completely lost and had no idea what was going on. As soon as things approached normality, the stall alarms etc. came back. This apparently shocked the copilot enough that he kept doing the wrong thing (pulling back on the stick), which kept the alarms off. Also, the captain seems to have been unaware of the input the copilot was giving the plane. As in, the captain did not know that the copilot was pulling back hard on the stick, and the captain apparently never got the idea to ask. The pilot who is not currently flying the plane should always have an easy way to figure out what the other pilot is doing.

As far as I know neither flaw has been corrected. The first flaw may be unfixable in general, but the second one is easily fixed by using a yoke instead of a stick, and making both yokes move in unison (and in unison with what the autopilot does as well). Boeing does that, even on their fly-by-wire 777.

Re:Much better than Google's approach

[Brucelet]

Not many obstacles, but there's one really big one. And on landing, you're not tring to avoid it, but have to hit it just right.

Re: Analog is better.

[Hognoxious]

I don't even like anti-lock breaks

I don't like anti-break locks.

But then I am a burglar.

Human backup vs. computer backup

[Beryllium Sphere(tm)]

Interestingly, both approaches have been tried in aviation.

A while back, Aviation Week reported on an experimental system that could override fighter pilots when they would otherwise crash. It waited until the absolute last second, when the required maneuver was just within the structural limits of the airframe.

Using humans as backups has a long and good operational history, but it might not work as well with undertrained personnel like car drivers. Even with highly trained pilots, dropping control onto a human suddenly in a disorienting situation can be problematic, e.g. Air France 447.

Obvious problems

[Bogtha]

Firstly: How does the system detect imminent crashes? If this makes mistakes, it can wrest control away from the driver when unnecessary and cause a crash.

Secondly: How does the system react to imminent crashes? If this performs worse than what the driver was already doing, it can cause a crash.

The main problem with autonomous driving is the legal liability. The problems above still introduce the legal liability, yet without the major benefits from a broader system. I think the industry will simply skip over this straight to broader systems.

Re:Obvious problems

[Grygus]

What you're mostly saying here is that if the system doesn't work at all, then it's pointless. Okay.

The legal liability is a more interesting point, though; I think this is the highest hurdle such technology faces. Even when it is nobody's fault, our system demands that we assign blame.

Re:Obvious problems

[Jeff Carr]

The main problem with autonomous driving is the legal liability. The problems above still introduce the legal liability, yet without the major benefits from a broader system. I think the industry will simply skip over this straight to broader systems.

Liability isn't too much of a problem in my opinion. Insurance will cover any issues, and rates will change based on the performance of the autopilot. As long as the autopilot performs equally as well as drivers across the entire set of cars insured, then the insurance rates will be the same, people will pay the same rates, and insurance companies will shell out the same payments. An accident wouldn't cause rates to go up, but good driving records wouldn't bring the rates down, and everyone would pay rates based on the area they live and drive in. A horrible autopilot would be detected quickly by insurance companies, and that would be the fault of the manufacturer, a poorly modded autopilot or poorly maintained vehicle would also show up as an anomaly and would be the fault of the owner. Uninsured drivers would be responsible for themselves.

Re:Obvious problems

[Kjella]

Secondly: How does the system react to imminent crashes? If this performs worse than what the driver was already doing, it can cause a crash.

And the related question, even if the computer did everything as "right" as possible how would you prove it did? All it takes is for a person to get up on the stand and say "No sir, I did not run over and kill that man. I was going to swerve around him into the ditch but the car took over control and ran straight over him. My driving may have been reckless but it was the car that killed him." No matter if it's true or not, would be possible or not, the makers of this system would have to get up there on the stand and defend their interference and try proving they did not in any possible way made things worse.

I agree with your analysis, if the computer can handle near-crash situations well then surely it can handle the more mundane driving of making turns, following lanes and signs and traffic lights. Near crashes have all the hard parts with pedestrians, bikers and other cars, huge set of possible actions, hard to predict outcomes and is the hard part of AI driving. If you had this system you could install it on Google's car and together you'd have your fully self-driving car already, I don't see how you could create one without the other. Not to mention then the systems could actually talk to each other rather than trying to guess the thoughts of the driver.

Re:Obvious problems

[zippthorne]

In luxury cars, you can already get a speed-matching cruise control. I'm not sure if that extends to hard braking (although it would make sense to), but such a system would be a perfectly reasonable first step.

  A distance sensor for following could certainly detect the sudden acceleration of the leading car and if actually applying the brakes is unreasonable, it could certainly activate the taillights, an alarm/warning light, and disengage the throttle in anticipation of and to buy some time for the driver's decision to apply the brakes.

Re:Obvious problems

[James McGuigan]

Same way as they do in airplanes. Have a little black box recorder in a thick steel box.

When the computer kicks in, it records everything it knows to the black box. After the crash, you can look at the black box and it will tell you if and why the computer kicked in and what information it had to made the decision it did.

This absolves the driver of responsibility and gives the engineers a bug report to work on.

Re:Obvious problems

[amRadioHed]

Sorry, but that wreck was your fault. If you were allowed to skid as you wanted, your braking distance would have increased, not decreased. The ABS probably saved your from making the accident even worse, but keep on telling yourself that you could have done better with your ace skidding skills.

Re:Obvious problems

[amRadioHed]

Still doesn't make sense. Skidding reduces the coefficient of friction on pavement. If there was so little weight on the tires that the skidding wouldn't change the stopping distance, how does not skidding change anything? What about the front wheels where the weight was? Skidding would have reduced their braking power too.

On loose surfaces skidding is actually better because the tread can dig into the surface and add more friction, but that doesn't apply to your situation. And even if it did, that is a very rare situation. ABS is still better in the vast majority of situations, which means it's always better to have a car with it then without.

Re:Obvious problems

[Inda]

Yep, all your fault. You were going too fast; faster than the road conditions allowed. Or you weren't concentrating.

I don't know about your country, but both of those are illegal in my country for good reason.

The correct thing for the computer to do is slow you down. As speeds approach zero, the chance of causing a crash also approaches zero.

Re:Obvious problems

[amorsen]

Most accidents would be greatly mitigated by full-on braking (with ABS of course) that one second earlier than humans manage in most cases. The system can let the human decide which way to go around obstacles or even whether to go around at all. As long as the car prevents skidding and rolling over and keeps getting rid of as much energy as possible, that will greatly help.

Automated cars are already so good that most racing events ban most of the computer assists.

Re:Obvious problems

[amorsen]

The anti-lock brakes literally reduced my breaking power by pushing back on the brake pedal as I was trying to stop.

That is not how ABS works. Each wheel is handled individually. Even if the rear wheels were covered in oil, ABS would not take away braking power from front wheels with grip.

To beat ABS, you need a brake pedal per wheel. Good luck dealing with that. And fuck you.

Re:Much better than Google's approach

[Pikoro]

Actually, the autopilot will usually take you to the "minimums" which is usually set to several hundred feet above the deck at which point, an audible alarm is sounded "Minimums!" and the pilot is expected to take over the throttles and yoke. If that does not happen, the AP will make an attempt at landing using nothing but the ILS and glidescope, provided you are nav and gs captured (which you should be while landing).

Re:Much better than Google's approach

[Impy the Impiuos Imp]

> There's not many obstacles to avoid up in the air.
> On the road there's dozens of other cars all around you.
>
> (Score:5, Insightful) <========

Thanks for the lol, people!

(A few days ago at MIT) "Hey. Shouldn't we consider that there might be other cars on the road before we release?"

"Oh, yeah! Duh!"

What about a switch?

[Jarmihi]

What if there was a switch where the operator of the vehicle could choose between normal driving, computer-assisted driving (MIT), and human-assisted driving (Google)? I think that would be a better option than having to choose between expensive automobiles.

A servant, or a lifeguard?

[Zigurd]

Is this some kind of weird expression of the New England work ethic? Make the driver work just as hard as ever, but should he ever falter, a superior system kicks in and saves his ass?

If I have a computer that can handle emergencies more reliably than I can, surely it can handle the mundane more reliably, too.

Re:A servant, or a lifeguard?

[Lehk228]

Such an emergency system would need significantly simpler sensors and AI, fully automated driving needs total awareness of signs and location and the ability to constantly evaluate what is being sensed neither drivers nor. Police will tolerate driver AI that decides to drive on a median because it looks a bit like a lane, nor will an AI be tolerated that fails to divert where road construction requires driving in the median.

On the other hand a safety failover won't really create problems as long as it properly designates MUST NOT HIT objects, like other cars, bridges, people, jersey barriers SHOULD NOT HIT such as grass, guard rails, curbs, sidewalks, and unoccupied opposing lanes.

Re:A servant, or a lifeguard?

[maxdread]

Unless those guard rails are on the edge of a cliff in which case lets bump those up to MUST NOT HIT objects and reevaluate hitting another car.

Re:Much better than Google's approach

[Joce640k]

Not many obstacles, but there's one really big one.

That one's only dangerous if you approach it off course or at a sharp angle. Computers are pretty good at linear algebra (better than humans), getting it right isn't a massive problem (how many years have they been doing it now...?)

Guiding a car safely along an arbitrarily curved road full of unpredictable other users is much trickier than landing an aircraft.

Re:Much better than Google's approach

[GameboyRMH]

Even a suborbital hypersonic aircraft wouldn't land anywhere near those speeds. That's close to mach 1.5 at sea level. Divide by 2 and that's still close to a modern airliner's full cruising speed.

Re:Much better than Google's approach

[AchilleTalon]

Automated landing systems can rely on radioguidance devices and airstrips are usually large enough to take into account for error. That is not the case on a road where you have nothing else than road marks which may not always be visible if any. The speed of the airplane isn't really a factor provided the computers are much more faster at doing computations and evaluate sensors than a human given the overall setup is much more simple than the one required for a car. No computer vision is involved into an automated airplane landing system. And even if it were, the overall scene is a simple one compare to a car road where you can have pedestrians, bikes, cars, obstacles of all kind, curves, bumps, sidewalks, posts and so on.

Transitional

[gman003]

The difference between the two approaches is a difference of perception - in one, the *human* is considered the primary while the *computer* is the backup; in the other the *computer* is the primary and the *human* is the backup.

Now, obviously, both of those elements can fail. Humans are fallible drivers, as I well know. Computers can crash, or just fail to process events properly. No matter what, you will get accidents under any of these. Hell, we still get train crashes, and they're bound to tracks and subject to tight top-down management.

I think the most likely outcome is this: we begin using computer failovers, expanding on those we already have (antilock brakes, cruise control, lane following, automatic parallel parking). Both because this allows for gradual testing and improvement of each module, and because as a society I don't think we can handle going directly to full computer control.

Eventually, however, the computers will be good enough at driving that you'll be able to have them take full control. And eventually, doing that will be commonplace.

Re:Much better than Google's approach

[DL117]

Most airliners land around 130 to 150 knots, 140 to 160 MPH, and cruise around 400 to 500 knots, 450 to 600 mph. Don't be silly!

Re:Much better than Google's approach

[Zeromous]

But Mommy I have TO GO NOW!!!!

Dear, please hold on the car won't stop. HOW DO YOU REBOOT THIS THING?

Uh-oh Mommy I peed on the seat...

Damn Bluescreen! On-Star,help my car won't stop and nav has gone bluescreen. what does STOP 0X00C553E mean?

I pooed too....

Bad Idea

[Anonymous Coward]

As others have pointed out, have will the system make a choice between multiple bad choices? What if there's no way to avoid a crash, such as a car traveling directly towards you and the road's too tight to get out of the way. How about a tire blowing out and/or the breaks failing. How will the system handle the car not functioning as expected? Humans adapt much better than software. There's only some many situations you can test for. What if you want to use your truck to help push a broken car somewhere? The software wouldn't let you.

The other issue is humans are lazy. If you given them a car that is supposed to prevent crashes, they will speed directly into the cars ahead of them expecting the system to take over and stop for them. 'It'll break for me. Why should I have to worry about that?'

Until they're fully automatic, I don't think we should use self driving cars. The cars should only provide warnings to the driver.

One final issue, how do you disable the lock out mode? The police will use it to automatically cause cars to be pull themselves over. OnStar can already remote disable your car and does so. The plus side will greatly reduce auto theft. But dive by anyone wanted for anything and the car will see them and communicate the info back to the police. It might not be part of the first versions, but there's no reason why those features won't be added to help protect you.

On the plus side, we're going to have super detailed road maps in the future.

Google's better

[hhawk]

My 2010 Prius System 5 already stops the car if i'm about to crash (PCS) and it helps steer when I have the lane keep assist (LKA) on. LKA uses machine vision so doesn't always work if there lines in the road are missing, degraded,

While clearly the MIT system detailed as more points of constraint and while I think it's could to have PCS (Pre Crash System); that problem doesn't solve numerous problems like those who are getting older (but still need mobility), those who get fatigued, using automated car "training" to smooth out high way and perhaps local traffic so that there is more throughput, less congestion, Etc.

Clearly the MIT system could be integrated with the Google approach... and I think could and perhaps should be required for new drivers (governance of new drivers) but issues of keeping drivers and passengers safe isn't the same issue has having self driving cars.

Re:Much better than Google's approach

[nospam007]

"This is much better than the disaster-waiting-to-happen that Google is building."

How so? Google's car you put on the automatic, _then_ you go to sleep.
In this model you have to go to sleep and _then_ the automatic kicks in.

Must be a dream-car for drunks.

Just remember to bring the dog.

[protonbishop]

An old joke among pilots asks: what do you need to fly a modern airplane? A computer, a pilot and a dog.

The computer’s job is to fly the plane. The pilot’s job is to feed the dog. The dog’s job is to bite the pilot if he tries to touch anything.

Seems both harder and less valuable

[swillden]

To me it seems that the MIT approach takes on the hardest part of the problem, reacting correctly in the hard corner cases, while also adding yet another hard problem, which is determining when to take over from the driver, and being less valuable to boot. The only problem the Google approach has to handle that the MIT approach does not is navigation, and that's the easiest part.

The MIT system is still going to have to have full awareness of all of the surrounding obstacles, traffic, pedestrian and other, will have to know where the lines on the road are, etc., should probably know about traffic signs and signals, etc., road conditions (ice, sand, water, etc.), because all of that is necessary to make good decisions in an emergency situation. Actually, the MIT system needs to know more about all of that than the Google car does. The Google approach can rely on the human driver to notice and take over when things get difficult (which, BTW, is why I don't think the Google car will be a solution for people who are under the influence).

As to value, lots of people will see value in not having to drive themselves, in letting a computer do the boring, tedious part of getting from place to place. I think relatively few people will see value in an undoubtedly expensive system that only operates when they screw up, especially since most won't want to believe the system can actually do a better job than they can. Most people are willing to believe that a computer can drive a car safely in normal conditions, especially with a human to take over if things get difficult, but will find it far harder to believe that a computer can handle extreme conditions better than a human, to the degree that it makes sense for the computer to override the human driver.

Re:Much better than Google's approach

[Capt.Albatross]

I'm not certain but I'm pretty sure computers are landing airplanes with the pilots overseeing the process.

There's not many obstacles to avoid up in the air.

It's possible that you've missed the significance of the word "landing".

There is nothing of significance for joce640k to overlook. Collision avoidance in an environment full of moving road vehicles (and sometimes pedestrians) is a far harder problem than putting an airplane on the right point on the surface with the right velocity, even when you account for other air traffic around airfields.

Re:More bs from educated idiots ...

[RightwingNutjob]

You, sir/madam/AC, have made a subtle but important distinction. This is the output of one guy who's been developing and testing it in a video game simulator and with a golf-cart in an empty field. Despite the hardware, this is very much abstract and does not appear to be backed up by the level of engineering effort that aircraft autopilots were when they were introduced. So nice idea, but still quite pie-in-the-sky.

Computer backup needed ASAP

[rlwhite]

My wife has narcolepsy, which means even when medicated her 15 minute commute is a risk that she could fall asleep behind the wheel. She probably won't be allowed to drive when she has to go off of the medicine for pregnancy. This emergency autopilot would be a necessity for us if it were available.

A computer backup should be able to make it to market quite a bit faster than a computer-first human-backup driving system. The Google approach is more luxury than necessity. We should push the computer backup system first, but the nature of our economy now is that the luxury of the wealthy will likely be pushed ahead of the needs of a middle class family like mine simply because they can finance it and I can't.

Re:Computer backup needed ASAP

[aXis100]

WTF???????!!!!!!!!! Sounds like she shouldn't even be on the road now!

Stark contrast??

[robi5]

"This intelligent co-pilot is starkly contrasted with Google's self-driving cars"

There is no stark contrast except in the minds of those who dramatize without much technical understanding. The technology behind both is largely similar. As part of the technology process, it is likely that the MIT type of application will be widely commercialized first, for these reasons:

- Today's people buy into most things that increase their safety, and the market is large (elderly folks, night drivers etc.)
- It's easier to get societal acceptance to pose automatic driving as a life saving measure, unlike "taking the weels off my hand"
- This is possible at an earlier stage of maturity (after all, its main risk is that it takes control when it shouldn't)
- There is already an evolutionary trend from ABS through parking assist to lane detection to advance car brake detection to automated parking and

Vulnerable if not 100% effective

[virtigex]

No matter how many accidents that the MIT technology prevented, if this technology fails to prevent an accident the makers of this technology will get sued. The lawsuit's reasoning would go like this: Joe's standard of driving, just like everybody else's, is way above average and his super-fast reflexes were handling the traffic situation fine, but MIT's defective technology overrode his highly skilled actions and actually caused the accident. Unless the auto manufacturers and the technology's inventors could prove that the accident would also have happened with the same or higher level of damage they will be held at least partially responsible. I worked for a few auto companies and such accident preventing technology would always be put on trial when it fails to 100% prevent an accident and they would be perpetually in the courtroom defending it. This is why this kind technology rarely makes it into production.

The mixed-mode problem

[Capt.Albatross]

A mixture of autonomous and human-controlled vehicles presents the problem that a subset of the drivers in the mix will game the system, recklessly cutting-off any vehicle they believe is under autonomous control, if it's to their advantage (or because their assholes tell them to do it.) Drivers in the NYC metro area might well believe this would be indistinguishable from the current situation, but I imagine it could be worse.

This would be a problem with both full-time and emergency-only autonomy, but it might be easier to identify the autonomous vehicles in the former case, by observing how they are driven.

Re:The mixed-mode problem

[cerberusti]

The solution to that problem is pretty easy.

If it becomes a problem, I imagine cars will begin to detect this and email a snapshot of the license plate plus evidence to local law enforcement. The end result would probably be much care in your actions around automated cars, as they will report your reckless driving.

Re:Much better than Google's approach

[danhaas]

In important industrial applications, a set of 3 sensors is used.

If they all agree, fine.
If one of them disagrees by a certain margin, use the information of the other two and light up a warning.
If they all disagree, turn it to manual and blast the alarms.

In really important stuff, like nuclear stuff, it is used up to 5 sensors, each with a different functioning principle.

Re:Much better than Google's approach

[dhasenan]

Humans suck in general. Combine the two: don't allow the human driver to take control. Ever.

Getting a decent UI for fine-grained destination control might be hard (I see a gas station and want to pull over, sort of thing), but that's solvable.

Re:Much better than Google's approach

[alex4u2nv]

A well tuned view finder and computerized system will definitely do a better job than most drivers; since it won't succumb to distractions such as rubber necking, cell phones, or hot chick doing her makeup in an adjacent car.

It will also save many lives with all the drunk, medicated, and/or sleep deprived drivers out there.

I don't understand

[PPH]

What's the problem? http://www.youtube.com/watch?v=PhlR3vidvp0

Re:Much better than Google's approach

[kheldan]

Why mod this guy down to -1?

I don't see the advantage of living in a world where people more and more are discouraged from learning to do things themselves. As is people on average are not really very good drivers, if you take the wheel away from them how are they going to be even minimally competent in the event of an emergency that requires human input?

If I was forced to purchase a car with either of these technologies, I'd opt for a backup system for me, and not the other way around. As a sidebar, if this is the direction things are going then I guess I'll be riding motorcycles more and more. ;-)

Thanks, but no thanks

[thesaintar]

I miss the time when we could buy cars that put the entire responsibility of keeping the car on the road in the hands of drivers. If I want to do a maneuver that seems like a better solution (flipping the car's tail out while purposefully messing with the throttle to induce a controlled sideways skid on a wet road) to avoiding a dangerous situation, traction control already messes with it. I wonder what it will be like with systems like these being applied as mandatory safety features.

Re:Thanks, but no thanks

[ultranova]

I miss the time when we could buy cars that put the entire responsibility of keeping the car on the road in the hands of drivers.

Having seen the way humans drive, I don't.

Re:Thanks, but no thanks

[perryizgr8]

and this is exactly why i look forward to more automation in cars. do you really think driving like an asshole ("flipping the car's tail out while purposefully messing with the throttle to induce a controlled sideways skid on a wet road") is safer?! life is not a fast and furious movie. i bet the guys who coded up your car's traction control and abs know much more physics than you and can better predict the way your car reacts. if you want to do stunts like this, go do it in a closed-off road, race-track, etc.

That is a |10-38| Outsider blabbing about

[Joe_Dragon]

We are going fix him like we did to jimmy hoffa.

Why not both systems?

[only_human]

It's a false dichotomy. There is no need to exclude one. One system for when the driver loses ability to drive. The other system for when the driver drives badly.

Re:Much better than Google's approach

[Sir_Sri]

That's silly.

A car can't drive now with a LOT of different electrical or mechanical failures. Adding one more into the mix isn't really changing much.

The legal compliance issue (what if the car knows one of its sensors isn't working) is actually a serious and legitimate problem. In that case who is considered responsible for 'operating' the vehicle? If a device has a tendency to catch fire or otherwise fail unsafely it tends to get recalled and the manufacturer blamed, so a car may not automatically do anything other than drive to a repair shop if part of its own diagnostic system fails. Right now you can just ignore those warnings and keep driving (at your own peril naturally).

If you fall asleep, for example, the co-pilot acti

[mapkinase]

> If you fall asleep, for example, the co-pilot activates and keeps you on the road until you wake up again.

Is there a way to keep the driver unaware of that feature?

not exactly new; Android apps

[khipu]

You can already get a number of Android apps that watch the road and alert you if you pull up too close or leave your lane.

And, of course, some cars have these kinds of assistance systems as well.

Re:Much better than Google's approach

[Sir_Sri]

There are a LOT of artificially created obstacles up in the air. Not colliding with another airplane isn't so bad, it's not getting 'too close' to the aviation authority approved box around all other aircraft, in 3 dimensions, when they're all moving at high speeds is a very difficult problem, in addition to air corridors.

Where automated car systems tend to fall apart is when their sensor systems can't get the data they want (which is explicitly opposite of how aircraft are designed, automated aircraft systems are very much about working when the pilot can't get the information s/he needs), or when there is something like construction, which the car cannot interpret. Aircraft of course fail badly (see the air france crash from brazil to paris) when the sensors start getting bad data.

Automated cars would actually be much easier if we could do things like radio tag the road or something similar, so that cars followed along electronic line markers. If you could then radio tag other cars as well that would make the whole problem significantly easier. That's essentially what aircraft do, but thus far no one seems inclined to do that on regular cars.

I would think though, that the first and biggest market for self driving cars will be for the elderly, where putting constraints on (vehicle moves slowly, don't go near construction etc. ) will be acceptable tradeoffs in usability, because those sorts of constraints are still a lot better than not being able to get around on your own at all. It also means that the government or an insurance company or the like may 'own' the vehicle, and assume liability for it, and maintenance costs etc. and the operator is renting it.

Man vs. Machine

[courcoul]

Was gonna flippantly reply "if human is a healthy, reasonably young member of the species in all five senses and with sufficent experience, computer should stand back, else the old fart should RIDE in the back."

But then someone mentioned planes. Anyone up to date with the news and who read the final BEA report on the Air France crash in the South Atlantic with 200+ dead will recall that the primary cause was lack of crew preparedness. Dumb pilots who couldn't fly a plane? Yes, but not cause of their choice.The airline chose to invest more in wiz-bang automated avionic wizardry instead of proper and traditional pilot training; pilots were just for show and just expensive chauffeurs who drove the plane in and out of the landing/takeoff strip, rest of the time the thing flew itself. Until the circumstances got beyond the capabilities of the dumb computer who panicked and handed over the controls to the pilots, who had no idea how to fly a big airliner out of a high altitude stall by the seat of the pants with minimal instruments cause they had never been trained to do so.

Given that there are WAY more cars, SUVs and trucks than planes, let's make sure this situation does not repeat itself in our streets and roads. How many young drivers know how to pull a motor vehicle out of a spin in an icy road, emergency brake successfully in a pouring thunderstorm, etc.?

Deer

[bill_mcgonigle]

I know some people whose insurance companies might even underwrite the cost of such a system just for the ability to avoid deer collisions.

Re:Deer

[ultranova]

I know some people whose insurance companies might even underwrite the cost of such a system just for the ability to avoid deer collisions.

Which it can't do because deer will enter the road by surprise from the forest on the side.

Re:Much better than Google's approach

[PTBarnum]

So do you only drive cars which predate ECU computers, and for that matter automatic chokes and starter motors? It's a shame that so few people know how to properly adjust spark timing and fuel air mix as they drive, just in case the automatic systems fail.

Re:Much better than Google's approach

[bill_mcgonigle]

Are the automotive companies really prepared to put dual systems in the vehicle with backup power?

Cost: $80. Markup: $330.

Sensors

[ukemike]

What happens when some mud splatters on a sensor? Does it suddenly go into catastrophic avoidance mode? What happens when sensor fails, or when a whole bank of sensors fail? My experience with cars is that when one item goes wonky (like a coilpack on a 2000-2004 VW 1.8 turbo motor) they all eventually go wonky. Since these cars will be making life and death decisions, and interacting with other cars making life and death decisions, who decides if the programming between units is compatible? Does the government set standards for expect car behavior? If a kid, hell if my own kid, runs out in front of my car and I choose to swerve into a tree, does the computer override my decision? If it overrides me and kills my kid, is the car designer guilty of murder, manslaughter, is it merely exposed to possible lawsuit, or is it exempted from liability due to the EULA?

Here's an even trickier legal question, when someone makes a bad driving decision and the car takes over during the impending accident phase, but fails to avoid the accident, and I get pulverized, who is responsible? Does the dumb driver pay my medical and repair bills? Will he argue that Google or Ford is partially at fault and I have to sue them (good luck)?

Re:Much better than Google's approach

[nukenerd]

Headhot wrote :

I also find it hard to believe that a computer cannot get better at driving a car the most people. Sure there are emergency situations the require extreme skill and judgement calls, but how many people are good in those situations? ... I have seen many drivers who react 100% wrong in dangerous situations. They don't understand the dynamics of the car .... Computers don't have this problem.

The problem with a computer is that a situation may arise which the guy who programmed it never coded for. You get this with ordinary app coding too (think the Millenium bug), although the consequences do not matter so immediately. Humans are much better at improvising in a new situation, for instance in recognising a good spot to run off the road if an overtaker is coming at you the other way. It is not a matter of the racing-driver type skill of understanding dynamics.

I'm not certain but I'm pretty sure computers are landing airplanes with the pilots overseeing the process.

Routine landing an aircraft is a very predictable operation. Even potential complications (such as tyre burst) are few enough to be programmable. There is nothing like the infinite variation you may meet when driving a car on a public road.

Re:Much better than Google's approach

[amRadioHed]

Human drivers are a disaster-currently-happening, there are major driver malfunctions in every city on a daily basis resulting in 30,000 deaths per year in the US,and over a million worldwide. We set the bar awfully low for the computers, I don't imagine it will take long to get them driving better then us.

Re:Much better than Google's approach

[StillAnonymous]

You'll get to ride in the sidecar, while the computer drives the motorcycle for you.

Re:Much better than Google's approach

[Ztras]

They don't understand the dynamics of the car and get confused in a panic. Computers don't have this problem.

You've never seen a kernel panic?

no

[PJ6]

I would not want to be trying to manhandle the steering wheel and braking system out the control of a computer during an emergency maneuver.

Re:no

[dave420]

Why would you want to, when your reactions and awareness of the situation are worse than the computer's? Oh yeah, arrogance.

Re:no

[PJ6]

Why would you want to, when your reactions and awareness of the situation are worse than the computer's? Oh yeah, arrogance.

Do you talk to people like that in real life, or just here?


Reaction speed and awareness are not substitutes for judgement and foresight.

What if I chose to go over an embankment or deliberately hit another car for a really, really good reason, and the emergency system decides to fight me because that violates one of it's constraints.

Say, like I see a two year old that's about to run into the road after a little ball?

Re:Much better than Google's approach

[nschubach]

Where automated car systems tend to fall apart is when their sensor systems can't get the data they want (which is explicitly opposite of how aircraft are designed, automated aircraft systems are very much about working when the pilot can't get the information s/he needs), or when there is something like construction, which the car cannot interpret. Aircraft of course fail badly (see the air france crash from brazil to paris) when the sensors start getting bad data.

Makes me wonder what the car would do with a large plastic bag flying toward it and possibly getting stuck over the laser sensors.

Re:Much better than Google's approach

[Neil Boekend]

Humans in cars is better than no cars, but this is about robocars, which would be better at driving than humans.
I, for one, welcome robocars.
The classical soccermom who has dropped her cell phone (which she shouldn't be using anyways) and is searching for it should be enough of a reason to ban humans from driving eventually (10 years after robocars became better than the average driver). The humans who really wish to drive on their own can do so on racetracks IMHO.

Re:Much better than Google's approach

[adolf]

With fly-by-wire steering and throttle, we've pretty much already there for a few years now. And brakes have had parallel computer control since the advent of traction control and/or ABS. (Mercedes, for instance, has had vehicles which can predict a panic stop and preemptively apply full-on ABS braking for well over a decade, using nothing more than some fancy software, some potentiometers, and perhaps a more-substantial-than-usual ABS pump.)

Thus, the scary-in-event-of-failure parts are already in-place.

How well do they work? Dunno. Don't really care. For the foreseeable future I'll be extremely resistant to buying a vehicle with total fly-by-wire anything, so I haven't kept track of their safety record at all. (Except for the widely-publicized Toyota recall from a couple of years ago, which involved both a software and a mechanical fix for what would've been a very controllable failure-mode if the drivers involved knew how to apply the brakes properly.)

And the only reason I allow the ABS and traction control circuits to survive on my vehicles is because they're largely parallel systems, which a hydraulic failsafe.

Whatever the case: The parts are already there, tooling down the highway, along with any assumption of liability for mechanical or physical electrical failure that a manufacturer might care about.

Irrespective of whether or not that's a good thing, it is just a matter of adding software and awareness to add autonomy.

abs inferior to skidding on loose surfaces

[Chirs]

On surfaces like gravel and snow (both of which exist in plenty around where I live) ABS actually gives longer stopping distances compared to locked-wheel skids. On the other hand, it does let you continue to steer while braking.

Re:abs inferior to skidding on loose surfaces

[amRadioHed]

Yeah, I noted that in my other comment. Doesn't apply to his situation though, at least not from what he's said.

Re:Much better than Google's approach

[Gr8Apes]

And then there are places where 14 year olds can drive with a mere signature of a parent/guardian.

At least that's how it was not long ago.

Kernel Panics

[Gr8Apes]

People generally don't have those, at least not nearly as often as a computer.

Oh, wonderful

[TheSkepticalOptimist]

So instead of human instincts and experience, the computer is going to override your actions and send you flying off the cliff to your death instead of sacrificing hitting another car with minor damage. The computer is instructed to "not hit ANYTHING", but in an emergency scenario where its an easy decision for a human to hit another car vs flying off off a steep cliff, the human decision is going to win every time.

Seriously, again, this idea of car automation, in ANY shape or form, is one of the dumbest ideas ever.

Just leave my car alone. Improve systems to train drivers (such as actually putting training drivers into controlled emergent driving scenarios so they build up better driving skills and confidence), but keep the robots out of my engine.

Its not arrogance that makes me a better driver then a robot, its the fact that I want to survive a crash rather then adhere strictly to a set of programming instructions.

Security in Your AUDI

[gentryx]

I do completely agree with the parent. We've recently been to Audi to discuss a research project and they gave us a nice tour, sharing some insights into their safety systems. Current Audis (and thus also Volkswagen, Seat, Scoda etc. which all belong to the VW group) have two classes of safety systems: pre and post crash. Pre crash systems analyze the situation and might act preemtively. The more interesting ones will break before you hit an obstacle or even avoid it by steering to the side. Technologically these systems are mostly mature. The sensors aboard can track >30 objects with a latency of just a few ms form object movement to servo action. The reason why they are not found in the current fleet is simple: insurance. What do we do in the 0.01% of the cases in which the system errs? These legal problems had to be resolved first.

Contrary to their name, post crash systems may kick in before the crash, but only if it's unavoidable according to the laws of physics. These systems will try to mitigate adverse effects, e.g. your bones being crumbled. For this the computers have a physical model of your body (and the car) and can detect through sensors in the seat if a heavy or light adult, or a kid in a booster seat is sitting in the car. It can then e.g. dose the strength of the belt pretensioner. One peculiar aspect of this is that the system knows whether you will break your neck during the crash. Scary!

This is the state of the art, no SciFi. Your car already knows much more about its surroundings that you might think it does. So relax people, and welcome your new robotic overlords... ehm, drivers. They'll be better drivers than you. Or me.

To clarify...

[Antony-Kyre]

"Maximum Homerdrive" is the episode to which I am referring. And I meant "Simpsons Did It" to be a joke from "Simpsons Already Did It" from South Park. How or why I was modded down from one to zero points is beyond me. I thought someone would get the joke.

Re:Much better than Google's approach

[Wolfrider]

--Does this have anything to do with the three seashells??

/ Auto-inflate, dammit!

Re:Much better than Google's approach

[holmstar]

Aircraft of course fail badly (see the air france crash from brazil to paris) when the sensors start getting bad data.

The autopilot automatically disengaged when it decided that it couldn't trust the air speed sensors, but from what I've read, the autopilot still could have been re-engaged in that state. If the pilots had re-engaged the auto-pilot, then the computer would have corrected the high angle of attack that caused the plane to stall and ultimately crash.

Re:Much better than Google's approach

[holmstar]

(Mercedes, for instance, has had vehicles which can predict a panic stop and preemptively apply full-on ABS braking for well over a decade, using nothing more than some fancy software, some potentiometers, and perhaps a more-substantial-than-usual ABS pump.)

I owned a Mazda 3 that had this too, but it seemed a bit too sensitive. On many occasions in rush hour traffic, I found myself having to quickly pull my foot off of the brake because the car was suddenly providing too much braking force. (didn't want to get rear ended)

How to avoid criminals taking advantage of this?

[kotaro24]

If the computer takes control of the vehicle during emergencies, robbers can drive in front of the vehicle, slam on the brakes, and wait for the car to come to a full stop. There are times (like when another driver points a pistol at me (happened to me once on the California 101 back in the day)) when unsafe driving is an appropriate response to conditions.

Re:Much better than Google's approach

[Fastolfe]

Student pilots do not solo after "only a few hours training". In many places there's no minimum number of hours of supervised flying experience, but that doesn't mean competency requirements do not exist as a prerequisite for soloing.

Re:Much better than Google's approach

[Rich0]

That depends on the aircraft. On a Cat III approach with an appropriately-equipped aircraft I believe the minimums are zero. As long as the aircraft determines it is capable of autolanding it will happily do so, and usually better than a human. Aircraft capable of this check their navigational accuracy and level of redundancy/etc before OKing an autoland.

I imagine most autopilots will follow a glidescope all the way to the ground, but if it isn't designed for autoland it isn't going to flare and it will be like a carrier landing, minus the specially-designed landing gear. It also won't retard the throttles.