Thursday, December 15, 2011

Algorithmic Driving

In my post last week, my commentors took me to task on my prediction that cars will drive us in ten years. Some thought Americans would wise up and learn to love mass transit. They don't know Americans.

Others thought the hardware cost would even in ten years remain out of reach. Google did not build an autonomous car by creating the hardware but by harnessing and training good machine learning algorithms. No amount of hardware would have given you a car able to navigate the streets of San Francisco five years ago.

What hardware do you need for an autonomous car, beyond the car itself? A good camera, a GPS device, wireless Internet access, gigabytes of RAM and a fast processor. I carry all that in my pocket. Google does use other sensors including lasers and radar but as the algorithms get better, the cost and need for this hardware can be reduced. Wiring the car to drive itself won't be difficult, already the steering wheel and pedals are mostly just a user interface into a computer that is controlling the car.

I have no doubts that technologically we will have autonomous cars in ten years adding at most a couple of hundred dollars over the cost of the car itself.

Other problems could get in the way. One is legal but Nevada is already changing their laws that will allow a testbed for autonomous cars in that state. Once the cars are viewed as safe one would expect the law to expand and other states to open up as well.

The other issue is social. As with every technological change we will have the usual technological life cycle: Innovators willing to pay the big bucks to try stuff first, Early Adopters who love to jump on new technology (where I usually sit), the early and late majorities following the crowd and finally the laggards who still insist on manual transmission and pumping their own brakes.

There are other issues like patents and industries, like auto insurance companies, that will try to fight autonomous cars. Autonomous cars will be too much of a win, in terms of parking, fuel efficiency, shorter and more productive travel time and most of all safety, not to prevail.


  1. 1) Why will auto-insurance companies not like
    driverless cars- in fact if they really are safer
    then auto-ins companies may like them.

    2) The sociology - people like driving themselves-
    may take quite a while to overcome.

  2. Some of us get motion sickness when we're not in control of the vehicle. Is this brave new world going to accept us? Will motion sickness be classified as a disability? Will self-driving cars have accessibility features for those who suffer from it? (What do people with chronic motion sickness do in countries where car ownership isn't common, by the way? I've never actually thought to ask before.)

  3. Because the individuals are not driving, they do not hold liability. The car companies manufacturing those autoautomobiles are too big to need insurance. Therefore insurance companies will eventually die out. And that is a good thing.

  4. I think lack of judgment is a problem. Someone runs a red light in front of you and you have to choose between hitting them, swerving into the car next to you or running up on the sidewalk. Not sure I want a computer making that decision.

  5. Car accidents in the US cost something like 3% of GDP (even aside from all the deaths and injuries they cause). So it would be worth spending a lot of money if we could solve the problem.

  6. Excellent article shares results of modeling an ideal mass transit system vs self-driving cars

  7. A search of the USPTO Patent Application database for "automobile AND driver AND computer" finds 10,316 pending applications.

    That's applications, not patents.

    Are we destined to replace gridlock on the streets with gridlock in the courts? It's not clear (to me) that this would be an upgrade.

    By the way, the US Patent Office *is* hiring. :)

  8. Years ago people didn't believe that airplanes could fly by themselves. Yet airplanes do fly by themselves thanks to autopilot. Most airplanes fly on autopilot for most of the duration of the flight. Pilots are still here (as will be drivers) to make adjustments. Pilots are also here (as will be drivers) to respond to unusual circumstances.

    An example of an unusual circumstance is the red light situation described by 11:18 am anonymous. In such situation, the driver takes control.

    I think that realistically we will have a hybrid approach to machine/man driving, much like the one to airplane flight.

  9. Hybrid human/machine driving as Leo is suggesting won't work. Pilots are trained to keep alert despite the machine doing the driving, normal people aren't. It is an all or nothing deal.

  10. A hybrid approach probably won't work if there is no training or advanced licensing required to operate the auto-driven car.

    Simple forms of machine/human driving already work. Active cruise control, for example is available. ABS. Automatic parallel parking. These are small steps toward -- I hope -- more advanced forms of hybrid m/h driving.

    The path to reliable machine/human-driven cars can also lead to a fully machine driven vehicle. The option, however, for human intervention will always be there in my opinion.

    Back to Lance's point, I think that whatever path such automation follows it will be mainly a matter of disruptive innovation in software and evolutionary innovation in hardware.

  11. If you are right, Lance, there needs to be much more discussion of this. Airport expansions, highway expansions, rail expansions, these are all projects with 10+ year timescales that could be majorly affected by more efficient driving.

  12. why are there still cars being made with maual transmissions?

  13. @2.33PM, Dec. 17

    Automatic transmissions still add about $1000 purchase price of a vehicle. Plus automatic transmissions are often much more expensive to repair. Especially on cheap cars (under $15000) this is a non-trivial impact.

  14. Manual transmissions are more efficient on account of not having a torque converter.

  15. After knowing about the deleterious effects of auto-pilot on flights that allow pilots to sleep, be distracted, or just not know what the hell's going on (cf. the Air France debacle), it is amazing that people belive that ordinary car drivers, who are often teenagers, tired commuters, or DUIers will actually handle their automated cars responsibly.

  16. I agree it is a bit like the automatic vs manual transmission. You;ll see it showing up on high end cars. I bet it will start as an automated "chauffeur" mode for some luxury cars. However it will take many years to turn over the stock of cars - e.g., ours is 18.5 years old. Adoption will be slow despite the advantages because cars are simply too expensive to be upgraded as we do with cell phones. I don't expect it to move quickly down market because its benefits are enough that it will be introduced at the point that it is still a rather expensive premium (not like push-button starts or LED head lamps).

    I don't see the sensor part of this as being particularly expensive. We already have cars with plenty of sensors in them - e.g. a car as low end as the Ford Focus has rear and front view parking assist sensors.

    BTW: For many years automatic transmission cars were much less fuel-efficient than manual transmission cars - automatic transmissions add significant weight, fluid torque converters were much less efficient than directly geared connection, etc. There is still a power loss in the torque conversion. However, these days, despite what the EPA says, above a certain minimum power a typical driver with a similarly powered automatic transmission will get as good or better mileage in practice than the typical driver of a manual transmission.

  17. Regarding the social acceptance:
    I guess, logistics companies would be very interested in autonomous trucks. That would not only save on salaries but also on wear and consumption of trucks. As soon as it will be legal and reliable, this type of replacement will be the first thing to happen, quite independent of the social acceptance.

  18. I'm not at all convinced insurance companies would oppose the introduction of the optional ability to have the computer drive the car.

    The insurance companies want to maximize profit, so the amount they take in minus what they have to pay out. Now if automated driving is safer than human driving it would certainly cut down on payouts but how much would it cut down on premiums?

    That's a tough question because it depends on the degree each insurance company is open to adverse selection effects, any market inefficiencies and how this relates to capitalization requirements.

    I mean on the assumption that total automobile insurance payouts are essentially the sum of many independent random variables (reasonable for a national insurance company since weather is regional) the capitalization needed to offer a given level of protection against default is extremely tiny as a fraction of total revenues. Thus on this model we might as well assume there is no capitalization required.

    In this case if we assume perfect competition the profits of an insurance company should be independent of total payout amount. Obviously this model has to break somehow but how it breaks will make the difference as to whether automated driving is favorable or unfavorable to the insurance industry.

  19. >After knowing about the deleterious effects of auto-pilot on flights that allow pilots to sleep, be distracted, or just not know what the hell's going on (cf. the Air France debacle), it is amazing that people belive that ordinary car drivers, who are often teenagers, tired commuters, or DUIers will actually handle their automated cars responsibly.

    They don't need to handle them responsibly, only more responsibly than they handle their cars.

    Pilots are a bad comparison because they are selected from the best, most responsible fliers of aircraft and aren't allowed to turn over many kinds of functionality to the autopilot. Thus their distraction creates a danger not present in automated cars (namely that they will fail to handle the parts of flight they are barred by FAA regulation or cockpit tech from ceding to the autopilot) and the benefit from the use of autopilot is limited as it doesn't remove poor fliers from the air. Worse, autopilot on airplanes is only allowed to take over a section of flight in which poor piloting is virtually non-existent as a cause of crashes since planes are so widely spaced while failure to respond to a mechanical failure (such as engine shutdown or fuel leakage) is the primary danger.

    For autos there are tons of bad drivers and it's their errors in vehicle control that are primarily responsible for auto fatalities. Removing control from them results in immediate safety benefits. Unlike planes there is no need for fast reaction to a crisis situation, if something goes wrong with the engine the automated driver need only pull over to the side of the road and wait for the human to wake up as opposed to a plane where quick intelligent reaction is necessary to prevent falling out of the sky.

  20. @Mohamed

    Currently whoever is in control of the car, i.e., in the driver seat, is legally responsible for the behavior of the car regardless of any technical failings. If you drive a car with bad brakes and you hit someone as a result your legally responsible and I see no reason that driving a car with poor automated driving systems would be treated any differently. Until such systems become so safe that we eliminate the control position entirely the individual still needs insurance.


    I suspect that automated drivers will cause much less motion sickness as they will steer nice predictable perfect paths without all the minor corrections human drivers need to do. Given human reaction time when a gust of wind comes we get blown off course and swerve back but a computer need not have such a jerky driving style.


    In a crash situation where a quick deciscion about which of several bad options to take I certainly want a computer making that choice instead of a person.

    A person who has to react immediately may not remember that there was a family standing on the sidewalk before choosing whether to ride up on the sidewalk. Nor does the person have the fatality statistics for the various kind of accidents on hand when such a choice is made. The person can't exactly compute how extreme a curve can be taken before losing control nor communicate with other cars to coordinate a response.

    A good computer driver will reference both a model of accident fatalities, a non-forgetful model of the surrounding terrain, autos and people. It will then communicate with any other automated cars in the area (potentially grabbing control from the humans in certain circumstances) to ensure that everyone acts in coordination to avoid forcing more accidents and be able to compute driving trajectories that only a stunt driver could match (such as accelerating and swerving when it knows it can just barely evade the impact given maximal accel).

    People tend to do this job quite poorly and one will frequently see someone who has swerved to avoid an impact with another car collide with a street lamp or other immovable object. A computer could calmly assess the situation realize that some impact is inevitable and choose to hit the back corner of the intruding car where it will slide in response and minimize the delta v felt by the drivers.

    Heck, I suspect even quite naive crash minimization algorithms will do better than the average driver in a panic situation. They certainly won't accidentally hit the gas rather than break or guess wrong about whether they can accelerate out of trouble.

  21. One thing to keep in mind is that autonomous driving works far better for certain tasks than others. For example, driving on an interstate is a relatively simple task. I expect to see computer assistance for the long and simple parts of driving, first, much like has already happened with airplane travel. The human will take back over for the tricky parts or the parts that require social interaction with other humans.