September 16, 2006
When Will Robotic Cars Become Mandatory?
Every time a piece of safety equipment becomes fairly cheap the United States government and governments of other highly industrialized nations have moved to require the use of that equipment. When automated operation of cars by embedded computers becomes safer than human driving will governments move to require installation of such artificially intelligent (AI) computer systems in all new cars? Will governments eventually go even further and some day even ban the operation of cars and trucks by humans?
The US government is going to mandate computer-controlled features to help maintain vehicle control. Electronic anti-lock brakes (ABS) and other computer-controlled means to keep tire traction on the road will become mandatory in the United States for all new cars.
The National Highway Traffic Safety Administration is set to announce a preliminary regulation requiring electronic stability control technology on all new vehicles.
I predict the US government will require all new cars be constructed to support fully automated operation within 20 to 30 years.
Auto companies think automated systems for avoiding loss of driving control are more important than air bags for reducing accidents, injuries, and deaths.
The Insurance Institute for Highway Safety estimated in a report in June that as many as 10,000 deaths a year could be prevented if all vehicles were equipped with the feature. Auto companies have said the systems are more critical in preventing deaths than air bags, which are credited with saving 1,200 lives per year.
Many computer-assisted driving technologies are hitting the market. All these technologies will become more advanced and lower in cost in coming decades. Toyota is moving from radar to optical (i.e. light-based) sensors to detect and automatically try to avoid crashes into road obstacles such as pedestrians.
Toyota Motor Corp. and Denso Corp. collaborated with NEC to implement the chip in autos. It was adopted as an image-processing unit for a pre-crash safety system for Toyota's Lexus LS460 to be introduced in the fall. Toyota already offered a pre-crash safety system that employs milliwave radar.
The Lexus road obstacle avoidance will work by automatically activating the brake.
The new Lexus LS460 with the image processors will automatically slam on the brakes a split second before hitting a pedestrian or vehicle and prevent a collision or reduce impact.
Systems that monitor driver wakefulness and attention are also starting to hit the market. Cameras aimed at drivers try to alert drivers when they are not paying attention and an obstacle looms.
And if all the coddling means your attention wanders, there is a camera checking the driver's eyes are on the road.
Lexus' Pedro Pacheco says: "In the event that there is an obstacle in front of the car and the driver is at the same time looking at the side of the road, the system will sound a warning to alert the driver.
I've also read proposals to have cars automatically detect whether a person has been drinking alcohol and is too drunk to operate a car.
Driver monitoring systems can detect a driver falling asleep behind the wheel.
Dr. Grace, a former Carnegie Mellon University professor, is CEO of Attention Technologies Inc., which just started marketing a Driver Fatigue Monitor -- a dashboard-mounted camera that measures how often a sleepy driver's eyes close at night and then sets off a warning alarm.
Attention Technologies is one of two local companies that have developed anti-drowsiness monitors.
The other is AssistWare Technology Inc., another Carnegie Mellon spinoff that is selling a forward-mounted camera that sets off an alarm when a vehicle veers out of its lane or wanders erratically within the lane.
Volvo is one of many car companies developing driver monitoring systems.
GÖTEBORG, Sweden (November 30, 2005) – Volvo Car Corporation is taking a decisive new step toward helping drivers avoid vehicle collisions with its new Volvo Driver Alert system, technology designed to monitor a vehicle's progress on the road and alert the driver if it detects signs of fatigue or distraction. The system helps drivers make the right decision, rather than taking control of the vehicle.
The company intends to patent the Driver Alert technology and plans to make the system available in Volvo vehicles within two years.
Driver fatigue is a major traffic safety problem around the world. According to the U.S. National Highway Traffic Safety Administration (NHTSA), approximately 100,000 collisions are caused every year on American highways by drivers who fall asleep. Fifteen hundred of the collisions result in fatalities and a further 71,000 lead to physical injuries. In Europe, the situation is very similar: the German Insurance Association – GDV (Gesamtverband der Deutschen Versicherungswirtschaft e.V.) – estimates that 25 per cent of all fatal collisions on German roads are caused by fatigue.
BMW is also working on warning systems to alert drivers when they are drifting out of their lanes.
Safe motoring in style means maintaining maximum attention also on long distances. And this is precisely why BMW Group engineers have developed an assistance system helping the driver to avoid situations where they begin to lose their attention and awareness. Lane Departure Warning therefore informs the driver in good time of any unattended deviation from the car’s proper course, telling the driver through clear signals to countersteer and move back in the desired direction. This serves to avoid driving errors resulting from lack of concentration on traffic conditions.
Adaptive Cruise Control is another example of an electronic assist technology that makes some decisions for the driver.
Adaptive Cruise Control is a new technology that automatically adjusts vehicle speed to maintain a driver-selected distance from the vehicle ahead in the same lane. This next generation of cruise control uses forward-looking radar, installed behind the grill of a vehicle, to detect the speed and distance of the vehicle ahead of it, and then automatically adjusts your speed accordingly.
How does it work? The radar headway sensor sends information to a digital signal processor, which in turn translates the speed and distance information for a longitudinal controller. The result? If the lead vehicle slows down, or if another object is detected, the system sends a signal to the engine or braking system to decelerate. Then, when the road is clear, the system will re-accelerate the vehicle back to the set speed.
The fancy driving automation systems are coming out first on the more expensive models.
BMW AG, DaimlerChrysler AG and Toyota Motor Corp. are among the makers of premium models that are starting to market cars that automate many parts of the driving experience: self-parking cars (with the driver inside or out); parking guidance systems (for the less-lazy driver); enhanced cruise-control systems that work in stop-and-go-traffic and maintain a safe distance between cars; and warning systems that tell you when you've strayed from your lane.
Brakes and gas pedals are not the only parts of cars which are coming under partial computer car. Automotive engineers are also targetting control of the steering wheel. Mercedes has just introduced a radar-based parking assistance technology as an option in their CL-class cars.
First, the parking guidance system automatically uses side mounted radar sensors to monitor whether the space on either the driver or passenger side is adequate to park the vehicle. If it is, the dashboard displays a "P" icon, alerting the driver to the fact a suitable space is available.
Once the driver stops the vehicle and changes to reverse gear, the instrument cluster then displays a bird's eye view of the parking situation along with guide lines showing how best to park. Red lines indicate the current steering angle, while yellow lines show the steering angle needed. As the driver turns the steering wheel and the two lines coincide, they then combine to form green lines, alerting the driver that he is now in the appropriate angle to reverse.
Finally, as the driver is reversing, an audible signal gives notice when it's time to countersteer, perfectly situating them and their new coupe in that ordinarily too-tight-to-fit-in parking space.
You can see where all this is going. As optical and radar sensors and computers drop in price and become more powerful and as software algorithms become more sophisticated computers are going to gradually take over more driving tasks from truck and car drivers. Computers will become better than drivers and computer-operated cars will become safer than human-operated cars.
Biogerontologist Aubrey de Grey foresees a human future where biotechnologies allow us to live in eternal youth and never grow old. Once we reach that future accidents, murder, and suicide will become the major causes of death. Aubrey expects political movements will successfuly bring about a ban on cars in order to reduce death from accidents. I do not think most governments and polities will go down that path. More likely computers will make car operation so incredibly safe and accidents so rare that cars will not have much impact on average longevity.
Update: Skepticism about the feasibility of computer-operated vehicles became harder to maintain when the Defense Advanced Research Project Agency's Grand Challenge periodic contest for autonomous driverless vehicles finally produced winners in October 2005.
A robotic Volkswagen called “Stanley”, developed by a team from Stanford University in Palo Alto, California, won a $2 million prize on Sunday for winning a tough desert race of driverless vehicles.
And in a stunning improvement on 2004’s Grand Challenge, when no car completed more than 5% of the course, four other vehicles also finished. The 212-kilometre race across the Nevada desert is set by the US Defense Advanced Research Projects Agency (DARPA).
It says something about the rate of advance of the underlying technologies that so many contestants could all suddenly achieve the needed capability in the same year.
On May 1, 2006 DARPA announced a more difficult next phase Grand Challenge for autonomous vehicles to navigate an urban terrain. (PDF format)
The Defense Advanced Research Projects Agency (DARPA) today announced plans to hold its third Grand Challenge competition on November 3, 2007.
The DARPA Urban Challenge will feature autonomous ground vehicles executing
simulated military supply missions safely and effectively in a mock urban area. Safe operation
in traffic is essential to U.S. military plans to use autonomous ground vehicles to conduct
DARPA will award prizes for the top three autonomous ground vehicles that compete in
a final event where they must safely complete a 60-mile urban area course in fewer than six
hours. First prize is $2 million, second prize is $500,000 and third prize is $250,000. To
succeed, vehicles must autonomously obey traffic laws while merging into moving traffic,
navigating traffic circles, negotiating busy intersections and avoiding obstacles.
Does anyone doubt this challenge will be conquered within 5 years?
"You can see where all this is going. As optical and radar sensors and computers drop in price and become more powerful and as software algorithms become more sophisticated computers are going to gradually take over more driving tasks from truck and car drivers. Computers will become better than drivers and computer-operated cars will become safer than human-operated cars."
I can see some other less desireable effects as well, though they should be ameliorated over time. Primarily there are two: software bugs and software/hardware sensor vulnerabilities.
For software bugs, one only need recall the Therac-25 disaster of the 1980s, or the more recent NASA Mars mission bugs. Multiply by the large number of vehicles on the road, and the potential for disaster becomes clear.
For sensor vulnerabilities, there are many possibilities with almost infinite variants availble: radar interference, or for video systems some visual artifact such as literal "picket fence" situations driving by sparse forests at certain times of days, or odd reflections from some uncommon road contdition, and the like. All these may, or may not cause unexpected problems. I'm just speculating.
I'm not being a Luddite, or predicting disaster here. I'm just saying that some relatively exotic but life-threatening issues are likely to reveal themselves at some time with any of these systems **after** they have been tested and production systems are on the road.
We can only hope that death and injury do not accompany the discovery of the bugs or vulnerabilities.
I think this fairly obvious fact of technological life also supports some kind of mandatory "panic button" that will return complete control of the vehicle to the driver. This might be something like the automatic disengagement of "autopilot" when an airplane pilot moves the control yoke, except that these systems seem to be designed to operate at least in part with a human driver in the loop. So the disengagement mechanism would have to be different.
Once we are talking about Aubrey's live-forever time scales, and extremely safe robot cars, (in other words, about 40 years from now, to be conservative, provided there is no global catastrophe first), then won't we all have migrated to the matrix anyway? Who will still be driving around in cars? Won't reality be horribly boring? Why live in the real world when you can live in a perpetual action/erotic adventure, like a neverending James Bond movie, but so much better?
Of course, this can be accomplished through a wired connected to the neck (as in the matrix), and that will just be the latest development in interfaces, following mail, telegram, phone, instant messaging, and now video chatting, and soon to be video chatting with stereoscopic vision. But the ultimate interface will simply be scanning and uploading the brain onto some sort of circuit. Ray Kurzweil notes that we will scan the brain, in sufficient detail, in the coming decades, and Marshall Brain says we will "discard our bodies". If there is no global catastrophe, I think they're largely right.
But they are talking about the same time scales as anti-aging therapy and extremely safe robot cars. So wouldn't it be quaint to talk about driving cars, and making them safer, in a world where people scan their brains and discard their bodies to live in the matrix?
Yes, software bugs are a problem. The real world is really complex. But the totally autonomous software will get built to in steps with software that generates alerts and recommendations running for years before it takes over entirely.
BTW, autopilots could land airplanes right now. But that step is still done by humans. I expect computers to land UAVs long before computers land passenger airplanes. Airplane guidance is simpler.
If virtual reality reaches the same level of realism as real reality will that end the desire to travel? Maybe.
Will we reach the ability to fully reverse aging and and become young again before or after virtual reality becomes completely convincing?
I thought a few years ago robotic driven cars would show up, and some nation would accept them. Then other nations would accept them, then over time they would become mandatory.
Now I see its going to be much simpler then that. We are not going to make the car autonomous in one fell swoop. Its going to be all these little upgrades to assist drivers, until the final step which no one really notices by that point the cars are driving themselves. And honestly this trend has been going on for a long time already, think of anti-lock brakes.
As for whether its a good thing. At a minimum as soon as possible new vehicles should come mandatorily with systems that prevent accidents. For example if someone is coming up on another vehicle too close to be safe, the vehicle will automatically keep a safe minimum distance relative to the speed.
A positive spin off from this might be people more willing to buy smaller vehicles. Right now you can buy a small car, but there is always some moron driving recklessly in a big truck and putting your life in danger. When you buy a big truck you are relatively safer.
Sooner or later all cars will be linked up online reporting to each other road conditions, accidents, traffic jams and best detours, things like cars just over a hill which can't be seen, telling traffic lights where the volume is to move seemlessly. Stolen cars will be reported and automatically deactivated.
good post. Again, I think 20 to 30 years is too far into the future. I think there will be pressure to implement and expand systems just 10-15 years from now.
I looked at the article from Skeptic and simply put, I'm baffled. It's impossible to determine whether the author thinks that
a) AI is impossible (within what philosophical framework is he working?)
b) AI lacks a unifying paradign permitting any precision in estimation of its difficulty
The latter position seems fairly reasonable to me, in the absence of molecular nanotechnology, but MNT (which Skeptic is also skeptical of, appearently because of its guilt by association with cryonics)
Regarding the achievability of AI: Computer hardware and software constantly become more powerful and sophisticated. The advances in driving assistance technologies and in autonomous vehicle operation do not provide any indications that autonomous vehicles are impossible to achieve. We have autonomous airplanes already. DARPA's contests for autonomous vehicles yield better and better vehicles. Car makers produce fancier driving assistance technologies. Do we need to argue the philosophical questions on achievability of Turing Machine AI in order to judge whether cars will drive themselves some day? I do not think so.
You wrote: "Will we reach the ability to fully reverse aging and and become young again before or after virtual reality becomes completely convincing?"
That's a great question. With respect to autonomous driving, I was assuming that they would be happening at the same time. At best, autonomous driving might happen slightly (10-20 years) before we have LE therapies and move to virtual reality. But I could be wrong. Predicting the future is extremely difficult (unless you are talking about Moore's Law and similar trends, like the playing-strength of the best chess program).
There are two important things to note:
1. My original post wanted to emphasize not just that we might be living in the matrix, but that we will be discarding our bodies and moving onto chips. If we just live in the matrix, as in the movies, then we will want to take care of our bodies and retard aging. But if we discard our bodies, there is no need to do any of that.
2. The market pressure for life extension and virtual reality technologies will be almost infinitely greater than that for safer cars. Most of us will not die in car crashes. But unless we do something (or some catastrophe happens), most of us will die because of aging and age-related diseases. And most of us would pay almost anything to live in a perpetual adventure/erotic fantasy in exotic locations.
"Do we need to argue the philosophical questions on achievability of Turing Machine AI in order to judge whether cars will drive themselves some day? I do not think so."
Especially since a car already drove itself across America a few years ago without mishap.
I work in automotive and AI fields. I see the primary bottleneck as legal not technological.
1) I think it will be an extremely long time until this is adapted unless tort laws change. A tort lawyer in the US will pursue and likely win any case involving an injury with these systems. If an injury occurs, in a situation where it was avoidable, the manufacturer will be sued. If an injury is induced, say by a high speed avoidance manuver causing another human driver to panic and triggering a fatality, a court case will follow. The potential awards can be be devastating to a manufacturer. This is a self reinforcing problem, avoidance of losses in courts will lead to more delay in introducing these features, which will delay the learning needed to refine these systems. While these systems will certainly statistically reduce the probability of injury and death while driving, people will still be hurt and die.
2) Early systems will have 'mental blind spots' that can only be eliminated through experience (an accidents) and redesign based on lessons learned. Any control system consists of two parts, one part senses the environment and builds a state model of that environment. Another part controls the actions taken in that environment. Sometimes these two pieces are combined. However, the state space associated with a driving environment is infinite, in other words to consider the state a car is in requires an infinite amount of information. Further, many of the states are hidden from observation and must be estimated or guessed. For example, one piece of information is about the driver in the rusty blue pickup next to me having a blood alcohol content sufficient to pickle eggs. Fortunately, most of the information is completely irrelevant to what is required to make a 'good' decision to avoid a collision. Also, some of the hidden information can estimated based on scenarios.
Worse yet, a controller will not be able to consider all possibilities, therefore, it will generalize to make decisions.
The net result, is that a controller will be designed and trained to ignore data, make assumptions about the environment and generalize. This almost guarantees that the bulk of the decisions will be good, but some bad decisions will occur!
Therefore, mistakes will be made by a controller and people will die and lawyers will live off the fat of the innovators.
3) Mature systems will make bad decisions. Eventually, after correcting all of the mental blind spots, there will occur situations where chance dominates. At a high level a collision avoidance system must select a course of action to steer right or steer left to avoid an accident. Because some state information is hidden from the controller, it can only make the choice that have the best expected outcome subject to some risk considerations. Sometimes this will be a bad decision. In a courtroom, a reconstruction of the accident could show that the accident could have been avoided, therefore the manufacturer is at fault, collect 30% contingency + fees and run for congress!
So while I am very optimistic that technologically, we can reduce the fatalities and injuries on the road using these technologies, we can not eliminate all preventable injuries and deaths. Because we can not achieve perfection, I am very pessimistic of our willingness to change the legal systems to sufficiently encourage and protect this type of innovation.
The world contains many legal systems. Some pieces of the puzzle will first get fielded in other countries. Other pieces will first get fielded by the US military and other US government entities. They can deploy systems and render their suppliers immune to lawsuits from non-government entities.
Also, the US government can force the deployment so that car companies can either deploy systems or not sell cars. Plus, the US Congress can pass laws creating arbitration mechanisms and amounts awardable for death and injury.
Widgmo's bullet item (2) is why robot driving would require strong AI to even drive at a normal human level of safety.
The alternative is limited robotic driving in carefully controlled environments. But driving on the city streets among normal traffic, pedestrians, etc. -- that requires real machine intelligence, a prospect which is always "just around the corner" but never arrives.
Your submit script is broken, it generates spurious errors which then leads users to double-post.
Strong AI and complicated environments: Highways are a lot simpler than city streets. No side streets, rarely parked cars, rarely pedestrians. So I expect machine intelligence to take over highway driving before it takes over city driving.
I picture an intermediate stage where truck drivers will drive trucks to highway on-ramps and then hop out to get a ride back to their dispatch point. Then truck computers will operate the trucks down the highways. Refueling stations built for autonomous trucks won't put humans at risk because automated equipment will open gas caps and bring pump hoses to receptacles and monitor the refueling process.
The reason that strong AI has seemed just around the corner in the past was that early boosters greatly underestimated the computing power needed to achieve strong AI. But I think that level of expectancy died out in the 1980s and since then I have not come across serious researchers who believed strong AI was just around the corner.
The trend has been to develop useful incremental improvements that solve initially simple but increasingly more complicated problems that parallel with various brain subsystems can do. So, for example, image processing algorithms keep getting better and solve increasingly more difficult problems. Ditto for audio voice interpretation.
I work in robotics, and am working on the next Urban Grand Challenge.
The sensors are there to get robots driving reliably in 95% of potential scenerios.
Sun-blind, night-blind, weather-blind situations are harder.
The hard part is getting things down to a price-point that car companies can handle. My estimate is that the technology will be ready for an affordable 95% solution in 6-9 years.
Getting that last 5% is also very hard.
Convincing slow moving government agencies to have make the legal situation tenable is also hard.
Think about is like this: is the car company responsible for every accident if there are declared failure modes? If we need that last 5% to get the legal side working, you'll have to work harder.
I want a system where a camera on the car will recognize road signs, particularly speed limit signs, and alert the driver appropriately ('slow down!'). Yes, speed limit information could be obtained in other ways, but this way would not require large scale infrastructure changes. Nagging drivers for doing rolling stops, or going through red lights, or cutting off vehicles with the right of way, would also be useful.
I don't think AI is necessary. The most useful and safe early introduction of auto-automation would be on the freeway system. People could drive their cars on surface streets (or even off-road) normally, then after getting on the on-ramp their car would synch-up with an automatic system. If everyone did this, then you wouldn't need individual AI systems, just a not-too-complicated-but-big computer system. Traffic jams would be smoothed out, gas mileage would improve (less slowing down and speeding up), time would be saved, it would be safer, etc. And much easier to roll out -- you'd only need to "wire" the highway system, to start.
A necessary requirement for computer driven cars: drive by wire automobiles. There is fly by wire technology in airplanes today - a concept which would be useful in autos.
RFID devices along the road would be very cheap to install and would probably do the job more reliably than optical sign reading technology. The RFIDs wouldn't get degraded by fog or the dark of night or a tree's branches growing in front of the signs. RFIDs could contain a lot more road condition information such as info about areas where lots of fatal accidents happen.
Yes, there are certainly methods such as embedding signal emitters in roads ("intelligent highways") to guide cars along to make it much easier for cars to know what is the road.
Drive-by-wire automobiles aren't hard to develop from the standpoint of actuators. We already have automatic steering which would need to become standard equipment in order for all cars to be computer steerable. We already have computers operating the throttles and brakes. But the steering wheel is both a big technological and legal barrier to cross.
One comment, I don't think general purpose, strong AI will be required. What I do think will be required is sufficiently complex stochastic models and control laws based on those statistics. I would not expect these systems to be adaptive or learning at the level where the overall control decisions are made (e.g. turn right and accelerate). I'd expect the models in the system to be co-developed between heuristics and training data. I personally can not see letting loose a vehicle in a civilian environment that makes decisions without the designers (and liable parties) having some ability to defend how the decisions are made and the statistical properties of those decisions.
I spoke to a man working on the DARPA urban challenge (albeit on the business side of things). He said that the car would cost, now, a minimum of $1 million, and easily $1.5 million. Applying Moore's law (2 years, see Wikipedia graph and original statements), and using the average price of a new car, $27,958 (http://www.usatoday.com/money/autos/2005-11-16-car-prices-usat_x.htm), it would take 10-11 years to bring the cost down to the average price.
The real political question is whether, after the AI has been (arbitrarily) sufficiently perfected, the government will eliminate the option of 'manual' driving.
Think about the people who can not drive at all now. The blind, kids, people who've had their licenses revoked for speeding or DUI. They'll be the first people to travel in fully automated vehicles.
Also, we'll know when robotic cars have arrived when insurance companies offer lower insurance for those who buy cars that can not be driven by humans. Or when insurance companies meter your car's operation to charge you a higher rate for the human-driven miles than for the computer-driven miles.