Slashdot Mirror
Autonomous Cars Aren't As Smart as They're Cracked Up To Be (computerworld.com)
Gill Pratt, executive technical adviser at Toyota, offers a note of caution, even as more car companies start putting AI elements into their cars. Speaking in Tokyo at the announcement of a Silicon Valley AI research center that Toyota is to open in early 2016, Pratt pointed out the big shortcoming in an AI system as applied to automobile: Autonomous cars might look great in controlled tests or on pristine highways, "but soon fail when faced with tasks that human drivers find simple." From the article:
Drivers, for example, can pretty much get behind the wheel of a car and drive it wherever it may be, he said. Autonomous vehicles use GPS and laser imaging sensors to figure out where they are by matching data against a complex map that goes beyond simple roads and includes details down to lane markings. The cars rely on all that data to drive, so they quickly hit problems in areas that haven't been mapped in advance. ... A truly intelligent self-driving car needs artificial intelligence that can figure out where it is even if it has no map or GPS, and manage to navigate highways and follow routes even if there are diversions or changing in lane markings, he said.
I regularly drive a stretch of road that's just a few miles long, but between construction, accidents, poor marking, bicycles, and heavy traffic I'd be nervous about letting an AI system navigate. In what real-world driving scenarios would you most want humans to take over?
You don't have to have the car drive everywhere, 95% of the places you drive will probably have all of the factors needed for the car to navigate easily. Just don't have the car drive in areas where it can readily get in trouble.
You don't start teens off in ambiguous hard to drive conditions, but rather low traffic side streets or empty parking lots, etc.
We don't need self driving cars that are perfect from the start, merely good enough to drive us most places most of the time, and do not have accidents in the areas that are suitable for it to drive.
People keep saying this, but the truth is that the car is going to [be programmed to] follow the law. That means it's going to approach intersections at safe speeds, and it's going to avoid hitting pedestrians in crosswalks but will simply murderize them even if there's ten of them in your lane, and a cancer-ridden octagenarian driving a yugo in the other lane — even if the car has enough sensors to smell cancer, it's still going to run right into those pedestrians like you've gone bowling rather than deviate from the marked lane. It's going to make a good-faith best effort to stop. But remember, it's not going to go around a blind curve at a speed at which it can't stop if there's an obstacle. It's simply going to decelerate for the curve, and then accelerate again on the other side. If someone is in the road, it won't hit them, because it's not driving for fun. It's driving to minimize risk.
"You're right," Fisheye says. "I should have set it on 'whip' or 'chop.'"
Translation: Toyota is woefully behind in autonomous car development, and rather worried about it.
The FUD begins.
Well, neither are human drivers.
Post may contain irony: discontinue use if experiencing mood swings, nausea or elevated blood pressure.
The biggest problems aren't actually going to be point to point navigation or even obstacle avoidance, though those aren't trivial. Navigation when you know the destinations is a solved problem and we've got a pretty good idea how to handle obstacle avoidance and terrain following though there is progress to be made
Possibly the hardest problem to solve it you want completely autonomous cars will be navigation in the last quarter mile and for destinations where you aren't actually sure exactly where you are going. This is a human interface problem and those are always challenging. In those circumstances it is REALLY hard to instruct a computer efficiently without actually taking the controls yourself. For example how do you explain to the computer that you want the parking space 3 places over but you want to back in? Or that you don't want to block in the car so park next to it on the lawn? Sounds easy but it really isn't - not yet anyway. Humans can do it mostly competently but we don't have any computer that is anywhere close to human level processing of verbal commands. Stuff like parking lots will be surprisingly hard to automate in a way that will be pleasing to most people. There are solutions but they are going to take a long time and require a lot of infrastructure. Probably several decades away at minimum. Sort of how we had autopilot for planes many year before we had the ability to do autonomous takeoffs and landings. (and the aviation problem is arguably easier as it has fewer variables)
I think we will see semi-autonomous systems relatively soon particularly for stuff like highway driving. But I think there is going to remain driver controls for quite some time because steering into that parking space or instructing the car to back up to the front door is actually pretty hard to do well. What will happen is that you'll program in your destination, the car will take you close to where you want to go and then you'll probably drive the last little bit yourself in a lot of cases. I think this piece of navigation will be solved last if at all.
Cheaper? An AI doesn't need a salary, or a medallion, and it's insurance will be cheaper. If taxis were cheap, most of us wouldn't have cars.
Frakly this is BS... I drive a large portion of my day for work (not a trucker, IT guy going to clients.) I run into "diversions or chaning in lane markings" and have to stop and think about what to do at times too! Why should an AI have to understand the intentions of a road worker/civil engineer better than we do before it can be accepted as intelligent?
I'm not seeing anyone in TFA saying it'd have to be better at it than us, just that it'd have to be able to do it at all would be a good start. As things stand autonomous cars are not anywhere near of being capable of doing that on their own.
I will have no clue where I am and will have to basically start driving in one direction (which these cars can do) until I figure out where I am.
No, they can't. That's the whole point here: as long as they rely on GPS and very detailed mappings for navigation they won't be able to do that -- they need to know where they are to be able to start driving at all. The author wasn't saying the car should be able to magically instantly know where it is even when no mappings or GPS was available, just that the car should still be able to try and figure it out -- quite possibly doing the exact thing you suggested and trying to find a roadsign or two. The issue here is that these cars won't know even how to get off the god damn parking lot without GPS and mappings, let alone going out and figuring their own surroundings on their own without some very extensive AI.
decide between plowing into a crowd of people to protect the driver, and smashing into a tree to protect the crowd of people
People hold up ridiculous scenarios like this as some sort of hypothetical metric, but how well would a human do with an insane choice like this, presumably with only a split second to make the decision? Not very well, I'd imagine. Don't put AI up against ridiculous situations. Put them up against realistic obstacles, which we might actually have a chance of seeing in our lifetimes. Road construction. Temporary obstacles with police directing traffic. Blizzards. Temporarily flooded road. Parking lots or garages.
There's also this false dicotomy presented, wherein some people seem to think that unless an AI can can handle ALL situations possible, it can't possibly work. I'll tell you what will happen in many situations. The AI will come to a controlled stop and tell the human "Hey, I don't know what's happening. Take over the wheel, please." That seems perfectly reasonable for crazy scenarios that only rarely occur.
The answer to what would likely happen, by the way, is that the AI in the car would have long ago started braking, so as to avoid the problem in the first place.
Irony: Agile development has too much intertia to be abandoned now.
Have you seen iRobot where the robot choses to save Will Smith (the police officer) instead of the little girl because the robot calculates the little girl's probability of survival was lower? I thought that was an interesting take on decisions like this.
Chance favors the prepared mind.
Perfect is the enemy of good.
No, the correct decision of an car's AI is to always prioritize the life and safety of it's occupant. And you can bet that's what every vehicle will be programmed to do. People on the outside can take care of themselves.
Note that this doesn't mean speeding recklessly and then plowing into a crowd to save the driver. That only occurs because of previously made poor choices. Obstacles don't magically teleport in front of cars. It only appears that way to human drivers because we have a bad habit of not paying attention. Computers don't have that little flaw, and so will be braking the car before the human occupant even realizes there's a potential situation ahead.
Car manufacturers are not exactly strangers to litigation. The notion that any single court case will doom an industry is overstating things, I believe.
Irony: Agile development has too much intertia to be abandoned now.
As customer of a Taxi, you don't have to pay any of those things, only a miniscule portion of them.
The AI is more expensive, because all new cars are super-expensive, and you have to add Research and Development costs, "brand premium", And "coolness premium" the manufacturers will charge b/c the thing can drive itself.
If instead of buying a $12,000 used car that meets all your needs, you spend $60,000 on a brand new car-that-can-drive itself and lasts 10 years, plus a $50 monthly service fee for the cloud maps service, then you're paying approximately $5000 extra a year for self-driving capabilities.
That would buy you 333 $15 taxi rides.
Anyways, based on that, unless you spend more than $5000 a year on the Taxi, then it just isn't a worthwhile economic proposition.
Also, the self-driving cars are probably going to be introduced at about the $120,000 price point, not the $60,000 price point.
Then you'll also spend an extra $10,000 in vehicle loan interest per year to get the self-driving feature.
The "AI" we're talking about presently and in the near future is mostly "A" and almost no "I" at all.
The kind of thing marketroids and the naive are calling AI today, including the tech that's beginning to show up in vehicles, is so one-dimensional in its "intelligence" as to be on about the same level as a toaster that "knows" not to burn my toast, or a chess playing program that can kick my ass at chess. The toaster "AI" couldn't control a robot vacuum cleaner, and the chess "AI" can't even play checkers, much less deal with anything further out of it's 1D "I" zone of competence, including not burn my toast.
In order for a vehicle to be able to "know where it is" and "know what to do about it", it will have to be more than one dimensional; it will have to be able to read signs, it will have to know destinations as things other than map references and paths other than mapped roads (parking lots, unmarked roads, etc.), it will have to make decisions based on extremely vague inputs and be able to do things such as ask for, and locate sources for, directions and understand them in pretty much whatever form they are provided. It will have to deal with the various situations that come up when the maps don't match the roads, too. Judging by my GPS, that's a lot more common than one might otherwise assume. It changes over time in random, unpredictable ways, too.
A general intelligence system designed for service (by which I mean to imply not conscious... otherwise we're talking about slavery, and we should know better than that by now) is not that close as yet. Frankly -- and I'm speaking with my AI researcher hat on now -- I think we'll get to a conscious general purpose intelligence well before we get to an unconscious one. We have a great deal of experience with imparting information to consciousnesses and we have considerable information available to us about what comprises one in our study of the human brain, whereas we have almost none about building a general purpose non-conscious intelligence, other than stacking multiple one-dimensional intelligences one upon another, which approach is approximately equivalent to solving the problem of multiplying by a million by adding one to an initial value of zero a million times. In other words, it'll eventually get the answer, but it's not in any way efficient.
As far as AI goes, all we really have right now is AI research, and various (not insignificant) benefits from the various tech insights and advances that fall out of that process. We don't have AI at all, at least not in the sense that is even slightly worthy of the term. The way AI is being used today, you'd want to be very careful telling your kid they were "intelligent", because they're likely to take away the idea that you think you just told them they're about as bright as the toaster. Not to mention the fact that when an actual AI is finally brought to light, we're not going to have anything useful left to call it. At that point, "AI" would be an insult. Not a great way to start a conversation with a new entity, IMHO.
The whole "it's AI!" meme reminds me strongly of the whole "3D TV" debacle. Again, marketroids and the ignorant built and propagated that appellation as a supposedly appropriate designation for fixed-viewpoint stereo vision, where fixed-viewpoint stereo vision is constrained, even by a relatively coarse and generous measure using whole-number degrees, to about 2 and 1/64800D or 2.000015432...D, whichever notation you prefer, leaving the viewer with something that in very few ways indeed resembles an actual 3D perception. When trying to describe actual 3D imaging, one is left with no accurate terminology. Unlike AI, we even actually have some low-performance versions of real 3D imaging now, so the linguistic problem is already on the roost, so to speak.
Sure, language evolves, that's a legitimate and real thing, but language also devolves, and that's what we're seeing in both these cases. I'm going with it, but I'm going kicking and screaming about the word-crap the marketroids are leaving on my lawn. Goddamn kids and their unleashed word-mutts. Where'd I leave my shotgun, anyway?
I've fallen off your lawn, and I can't get up.