Why do you keep saying “AI”? Control systems for vehicles are far from needing AI level design, they just need to be adequately adaptive, responsive and redundant. Redundancy – multiple independent control monitoring processes that can identify failure more reliably that solitary controllers – is key to making autonomous vehicles usable. You simply cannot design a driverless car on the cheap.
Again, I think you’re contradicting yourself here, but it may be that we’re using too loose of a vocabulary. Adaptive driving that can handle any situation better than a backup human pretty much boils down to a set of driving software that is at least a weak AI much more advanced than what I’m aware of being generally available, possibly a strong AI.
“You can’t design a driverless car on the cheap” is a statement I would wholly agree with, but as you try to squash the last software or design flaw, the costs reach infinity. I would guess that a human backup and the controls they can use to override the system will probably remain more cost effective for at least the next 15 years. I could be wrong. If it happens in 5, I’d be mighty surprised.
No, there is absolutely no need for any form of “AI” in a self-driving car. Just because an enumeration of exception cases is beyond what you can personally for-see AI programing is in no way required for a self driving car. Almost all of the issues can be dealt with with traditional programing and or fail safe actions when those are not met.
As an example. A self driving car does not need to know if an obstruction is a child, trash can, or a yeti. Nor does it need to take action based on the fact that an obstruction is a child, trash can or a yeti. It only needs to know that there is an obstruction and deal with that situation.
As the self-driving car takes off adding in automation friendly indicators throughout the infrastructure will also greatly reduce the complexity required in the current ad-hoc system.
scab,
Self driving cars are not driverless. The initial goals were stated very well by Cheesesteak #147:
*"There are four things self driving cars can do to reduce congestion, that human drivers cannot.
First, they can safely drive closer to other cars at equivalent speeds, increasing throughput.
Second, they can maintain that distance more exactly, preventing traffic waves and rubbernecking delays/accidents.
Third, they can communicate over short distances to further reduce the need for space between cars, each car will know what is happening based on signals instead of visual cues.
Fourth, they can connect to a central repository of traffic data, potentially anonymized snapshots of traffic flow provided by member automobiles."*
This is not Ai. It is standard communication technology in a PID loop. The addition of electronic steering and visual driving allow the car to be autonomous. That does not mean the car will run down to Pizza Hut and return with dinner. It means the car can become autonomous when necessary:
The car becomes autonomous when the driver falls asleep.
The car becomes autonomous when the driver is disabled or distracted.
The car becomes autonomous when the car systems are not fully functional.
The car becomes autonomous in any situation whose resolution exceeds the ability of the driver.
There is a gray area between autonomy and some self driving. Road signs will soon be electronic so that warnings and limits are communicated directly to the car. The car will then respond by traveling at the received speed limit or slowing in anticipation of a hazard or providing a warning message to the driver.
The early adapters of these systems will be commercial fleets. Large trucks, taxis, limos, delivery vans etc. Their motivation is transit of urban areas at optimum speed, with maximum safety and minimum exposure to infractions.
Two motivating forces are just coming into play - computer controlled tolling and infraction billing.
All feeder roads to urban centers will soon be toll roads, but with a twist. The toll is only charged at peak traffic hours and the toll rate is a function of traffic density. Direct billing is made possible by license plate readers. You will be informed by a message from a roadside transmitter that you are entering a toll area and given the toll rates. If your vehicle is automated you may move to the center lane for fastest speed and lowest rate. If your vehicle is not automated you must remain in the slow lane.
Toll fees will be collected by a credit card inserted in the dash or billed to the owner of the vehicle.
Infractions will be billed directly to the owner of the car.
A legal determination made during the IVHS program is that the owner of the vehicle is totally responsible for it’s use. A friend of mine in Miami loaned her car to her son. In two weeks he had run up $875 in unpaid tolls and traffic infractions (mostly speeding). She had to pay.
This link explains it (redundant post):
Crane
Has anyone mentioned the autonomous tractor? Driverless tractor - Wikipedia
and
It's working where there are no roads.Missed the edit window
I like this video of the self driving tractor SO much better.
No, there isn’t any need for AI in a self driving car. However, once the car is autonomous, without any manual control, you’ll need AI at some point.
And I’m not the one arguing that self-driving cars are going to be driverless. Uzi has proposed that they are inevitably going to become that, and I’m arguing against it.
I think this aspect needs to be reiterated. Cars will transition from personal parlors to peripheral mass transit. Some folks will own private vehicles, but even those people will have a subscription to the local auto pool, with which they can order up a car, go somewhere, and not be arsed to find a parking space, they will just call up one for getting home.
The individual units will be able to connect into train-like groups to improve efficiency and safety on heavily traveled routes, and getting on one of these trains outbound means you will has a personal transport unit when your trip branches off.
Weighing the cost of waiting a bit against the major long-term expense of ownership, I can imagine many people will go for the subscription, and personal cars will become a quaint, almost Luddite, notion.
If an autonomous car can handle daily driving tasks at a level of an attentive driver, it doesn’t matter about the exceptional situations. Humans are already terrible at that as well. Just raising the bar enough to handle 99% of situations will make driving immensely safer and if the other 1% is handled by attempting to get the car to come to a complete then well and good.
Btw, I’ve been in IT since the early 90’s, from front line support, back end support and design, to managing and implementing most ITIL processes (for an airline) and now back to managing telecoms and IT departments. There are crap hardware and software, and there is essentially bulletproof equivalents. Most of the robust ones are designed to do specific tasks like a control system at a plant, and they tend to do them exceedingly well for long periods of time.
I should have answered this earlier. How many commercial airliners cross the date line on autopilot? How many of them got lost because of it or ever got lost because of it?
But as I said upthread, as you start to chase down that last 10%* of situations that the (weak) AI has trouble with, your costs will increase exponentially. Barring some breakthrough in software development, the point where it won’t be cheaper to leave the steering wheel and brakes in the car as the final fail safe is at some point far in the foreseeable future.
And I’ll wager a healthy amount that almost all of those plants have a mechanical kill switch that a human can walk over to, and stop the machine. The equivalent to this in a car without today’s controls would be a relay that would shut down the power to the AI. If the car is well designed, it will drag itself to a stop in the same way a semi with air brakes would in the case of their brake system de-pressurizing.
Even this setup would allow for mischief, such as teenagers or drunks pulling the kill switch and stopping up traffic. Since that is the case, it would seem wiser to have controls that an authorized, non-malicious operator could use to bring the car to a more controlled stop than a kill switch all by itself would have. At the moment, most drivers aren’t drunk, crazed, homicidal maniacs. Allowing them a method of saving themselves in the event of colossal system failure seems like the sane option to me.
So, why did they let that bug emerge in the F-22? Re-using the old software of planes that had successfully crossed the international date line probably was not an option in this case. Even if it was an option, re-using previously successful software on new hardware (or even just in a new configuration) isn’t a panacea against introducing unexpected bugs into the system.
*I don’t think the current self driving cars can hanlde 99% of the situations a sober, alert driver can. If anyone has a reliable cite quantifying the limitations of the current state of the art, I’d appreciate it.
Scab,
The gee whiz phase has ended. Most of the need and solutions were defined in the mid 1990s. We have entered the commercialization phase. What used to be published in conference literature is now proprietary to companies.
The problem now is legal and political more than technical. I cannot find a single link with a current summary. Perhaps someone can. The following links are informative:
Summary of IVHS program:
http://www.eng.fiu.edu/mme/robotics/elib/IVHS%20Part%201%20v2.pdf
Statement of issues 1994:
http://ntl.bts.gov/lib/jpodocs/repts_pr/6144.pdf
Bosch Palo Alto positions:
I think you are reaching here. Teenagers and drunks drive recklessly all the time. If all they have the ability to do now is hit the panic button and come to a stop, then I think we’d all be willing to live with those consequences. Again, if that happens the car should inform the central computer who will query the occupants as to what is happening and dispatch the police if no satisfactory answer is received.
Because an F-22 is essentially a one off and everything is custom on it. Cars don’t have that problem. Years of research and testing and real world driving will find bugs that will be removed as found if proven to be dangerous. Most bugs don’t cause software to flip out and blow the start the computer on fire. Some of these bugs may be as innocuous as causing the windshield wiper to run faster than required rather than aim at the bus load of Nuns and accelerate.
Maybe not now, I agree. But it won’t be 20 years until they can. Think back on what the computer capabilities were 20 years ago to anticipate what it will be in another 20. I’d guess no more than 5 years, ten years tops, before you can buy a car that drives itself far better than any human can today. Same time frame before a computer controlled car wins the Indy 500 (assuming they let them play).
And, yet again, self-driving cars don’t need to be perfect, just better than humans. Suppose someone came up with a self-driving car that, if mass-implemented, would result in 1000 deaths per year. Inevitably, some of those deaths would be situations which a human driver could have dealt with. Does this mean that that car is a bad car, and shouldn’t be allowed? Quite the contrary: If such a car existed, we should be passing laws to require it, and prohibit human drivers, because 1000 deaths per year is such a vast improvement over human drivers.
Nor does the existence of a few cases where the computer fails but a human succeeds mean that we should include manual overrides in cars, even if such overrides carried no increase to the price tag. Yes, you’d be avoiding some accidents, but you’d also be enabling others, when the human used the manual overrides when they shouldn’t, and thereby caused an accident. If the human is safer than the computer, then don’t allow the computer at all, and if the computer is safer than the human, don’t allow the human at all.
The issues near term will be legal and political. Very much like the plethora of court cases that resulted from the initial use of RADAR.
Issues like:
Red light monitoring. Very effective, but the public objects.
Owner responsibility for any infraction involving his car.
Liability for damage done by an automatic parking vehicle. That’s kind of basic to this discussion - if the car is automatic, whose liable?
Placing smart cars (trucks etc) on the highway will create an elite class that has preferential access to the public road surface.
Automatic infraction billing is considered political suicide for local office holders.
Perhaps the same is true for traffic density tolling. Maybe they can blame it on the Feds.
So, the technology is running way ahead of the political and legal community.
Crane
Crane, while we are still in the stages of the self-driving car being like the Google car, where it isn’t yet able to deal with inclement weather, parking lots or temporary traffic signs - I believe we’re still in the development stage. Proposals and job postings don’t really provide evidence otherwise.
You’re assuming that coming to an immediate stop in a straight line is the optimal solution in most situations. I don’t think that’s the case in enough of them that we’ll all be willing to deal with it. In any case, I’ll probably be tooling around in a robot body by the time it really becomes a concern.
Even with the one-off nature of aircraft, their software is currently vetted at great expense more thoroughly than the software for cars. After they take their greater level of care, they still end up with bugs that would be fatal without a pilot on board. Cars currently have a dismal bug rate in comparison. On top of that, I think software written specifically for each piece of hardware has a lesser rate of failure in the end.
I do agree that computing advances a lot in even 10 years. In 1990, a 16 color graphics card in an IBM pc was awesome, not so much in 2000. However, I think most of that advance comes from increased computing power, and software appears to be the opposite of hardware. The more complex it becomes, the harder it is to maintain, and it takes more time and money for the next advancement to come.
To quote the software lead on the Google car:
And he’s not even getting into how you might develop a way to test your extreme cases once you discover them.
If you let the car run the track after it has been able to get a very detailed map of it and let it run alone, in dry weather, with no obstacles, and you’re Jeremy Clarkson - BMW can already do this. On the other side of the argument, Google predicts they’ll have their car ready to drive in snow, rain and parking lots by 2017-2020. Today I doubt a driverless car would be able to match Mark Higgins’ “moment” at the 2012 Isle of Man TT. With or without it’s detailed map, it would either crash or come to almost a stop, making it lose the time trial to Higgins. He can already do that sort of thing in the rain, snow, mud, parking lots, snowy parking lots full of mud while it’s raining, etc. He has never won the Indy 500. I don’t think that Google agrees with your assessment, even if they might try loading the software into a race car by then.
If we ever get to the point where we get such a self-driving car, there will almost inevitably be a plateau where having a combination of them will be the statistically safest option. If the AI starts winning by a little bit, start selling cars with a control system that blasts your crotch and hands with blue dye if you use it (hell, make it take a pic and post your moment of royal smurf terror to the local billboards while we’re being futurists), that way it will be ensured to only be used in an emergency. By the time you reach less than 1000 deaths per year, you very well might have strong AI driving the car, and manual controls will be a rarity mostly confined to antiques.
I’m not unreasonable, I believe we’ll go well beyond self driving cars and have strong AI someday. I just don’t think it will be any time soon, and probably later than most optimistic futurists can fathom. A good self driving car will probably come around the same time they could feasibly implant my age-delayed brain into a small car. Only I’d rather it be a spaceship with some sort of futuristic beam weapon that they shoot off into space - and I come back to fight the AI when it inevitably turns against us.
Whoa, got a little space opera-y there, sorry. Did I mention I want a robot body someday? 
You are ignoring the fact that current, real life, driverless cars are already much safer than human drivers.
This is not a “when” topic it is a reality today. It is no longer a question of technology it is a question of cost.
You’re ignoring the fact that they’re safer in only a very limited set of circumstances.
Like the established statistics from driving on public roads for years?
Cite: Here is a fairly recent youtube video showing the Google cars dealing with difficult situations.
Google’s cars are now dealing with Construction and peds quite well