Home / General / The future of self-driving cars

The future of self-driving cars



The self-driving car appears to be sort of almost here.

A driverless future will obviously have enormous economic, social, and cultural consequences. To mention a relatively trivial one from my own piece of the pie: A large percentage of the work done by many small law firms and solo practice lawyers involves things — traffic accidents, drunk driving, etc. — that will pretty much disappear in a world without drivers.

A more consequential effect will involve the millions of people in the US alone who currently make their living by driving cars and trucks.

In short, driverless cars could be a major technological shock in all sorts of ways, both good and bad. Which raises the question (btw I hate it when people use the phrase “begs the question” to mean “raises;” also get off of my lawn) of how soon this harbinger of our robot overlord future will be upon us.

I have zero expertise or even vague lay knowledge on this subject, so I put it to you, quasi-omniscient LGM readership:

(1) When will your typical local car dealership first sell driverless cars?

(2) When will a significant percentage — say 10% — of all cars on the road be driverless?

(3) When will the majority of cars no longer have human drivers?

(4) When will the human-driven car be a freakishly rare site, causing wonder among the young, who will have lost this crucial skill set, even as their grip strength continues to decline in our increasingly machine-dominated and decadent age?

  • Facebook
  • Twitter
  • Google+
  • Linkedin
  • Pinterest
  • j_doc

    I am a pessimist.

    Let’s assume we’re talking about “take a nap in the backseat” autonomy, and not “keep both hands on the wheel at all times autonomy”.

    There are two major components to this problem. The first is interpreting and safely navigating the physical environment: streets, lane markings, road signs, etc. This is very hard, but probably tractable. If you sterilized a city or town and had nothing but an autonomous car, or a fleet of intercommunicating autonomous cars, we can probably get them working well within a decade.

    The second component, however, is that driving is a fundamentally social exercise. You have to interact with, interpret and predict the actions of, all of the human drivers, bicyclists, pedestrians, and pets around you. This is almost all non-verbal, implicit, and ambiguous – yet crucial. I am often amazed that we can all drive together as well as we do. To be truly, fully autonomous, an AI has to be able to navigate this social space better than an average human. That’s not weak AI, that’s human-equivalent strong AI. We’re not even entirely sure that is possible. If it is, it will create a revolution far beyond autonomous cars.

    I’m pessimist because I doubt that true “take a nap in the backseat” autonomy can be 100% accomplished without strong AI. I think we’ll get to 80% solutions quick enough, which will find widespread but constrained use. Interstate driving, farm/mine vehicles, military vehicles, etc. Situations where the social interaction piece is either negligible or can be mostly ignored.

  • Rob in CT

    Bring on the self driving car!! I was nearly killed this afternoon. A car crossed (drifted is too weak) over and crashed head on into a car two behind me. I had to swerve, but only a little. The car behind me had to swerve more but managed. The car behind that… never had a chance. It’s a 45 zone and I’d bet the cars involved were doing 50-55. 100 mph closure speed, give or take.

    Fuck me that was scary. The car s involved got utterly clobbered.

    • You assume that a self-driving car could never do that?

      I’ve seen autopilots do some crazy shit, but at least I have an “off” switch and a set of controls at my disposal.

      • Rob in CT

        No, I don’t assume a zero failure rate. There is no such thing.

        I do think it’s likely they can out-perform us. Holy shit that one was bad, Kong. It’s not like the driver drifted over a foot or two into the other lane. By the time of impact, he/she was fully on the wrong side of the road. I’m thinking either the driver fell asleep or had a heart attack.

        • It’s an eye-opener.

          I had a guy come at me wrong-way on an interstate on-ramp in Memphis once.

          I flashed the lights, hit the horn, finally I took the shoulder as he went right on his merry way.

          • Rob in CT


            I got off a honk, but he/she was past me by the time I actually managed it.

    • No Longer Middle Aged Man

      Route 6 about a mile from where it meets 384? Because I saw the aftermath of a really nasty one there at around 6:30 today, one of the vehicles was basically destroyed, the other one merely looked like it had been totaled.

      • Rob in CT

        Yeah, that was it. It actually happened just before 5pm. Just past Munson’s Chocolate.

        I was pretty freaked out, I gotta tell you. I avoided it, but if I was 2 cars back…

        I’m a pretty good driver, really. I don’t think I could’ve gotten out of the way. I’d have been fucked.

        I saw the impact in my rear-view mirror. It was… impressive. Explosive, even.

        • Rob in CT


          No details, but the picture gives you an idea…

        • No Longer Middle Aged Man

          Yeah,that’s the one I saw. I’m glad for you that you were able to avoid it because I have to believe very serious injuries.

          • Rob in CT

            Yeah, my first thought was “oh shit somebody just died” but I’m hoping that modern tech (crumble zones, airbags, etc) saved lives.

            And thanks.

            • Rob in CT


              Crumple. Crumble zones sound kinda yummy.

        • bender

          How horrible.

  • AdamPShort

    SAE Level 4 is here now – fully autonomous driving on fixed routes. That’s actually not that hard a problem.

    SAE Level 5 will never arrive – fully autonomous driving not limited to fixed routes. That’s a more or less impossible problem, at least with current thinking.

    Computers cannot reliably decide what to do next when they do not have a complete list of the situations they might encounter. We will have to solve that problem, which we have made very little progress on in many decades of AI research, before we can start on true self-driving cars.

    The good news is that to realize most of the safety benefits, we really don’t need cars to be fully autonomous. We just need them to refuse to do really dumb things that drivers ask them to do. Whether consumers will accept that is another matter.

    So, to take the questions as presented:

    1) 2019
    2) 2040?
    3) Never
    4) Never, ever

  • j_doc

    I also think Atrios is probably right that anything short of “100% remove the human controls” is not going to be world-changing. And it’s binary – 99.9% won’t do. Increasingly nice and helpful as we get closer, but dishwasher-level effects, not internet-level effects.

    • Rob in CT

      I’ll take useful upgrades that aren’t world-changing, personally.

      If the car that nearly nailed me today had some feature that said “huh, the driver is allowing me to drift over the center line into opposing traffic, let’s not do that” it would be pretty useful!

      • Ghostship

        Aren’t they here already? Lane departure warning and lane keep assist should also work for crossing the centerline into on-coming traffic and forward collision warning should also cover this.

        • Rob in CT

          Yes, they’re being rolled out. My understanding is that they are passive at this time, at least in most cases. I mean something more active – as in no, you will not do that as opposed to some sort of chime telling you not to do that.

          Also, while these things will be trickling down to cheaper cars year after year, but many don’t have them yet.

          • Ken_L

            A control that stopped you crossing the center line just because a kid ran out in front of you or a drunk was lying on the road would not be a wholly benign innovation. I suspect it would cause more problems than it solved.

            A prevention of your crash situation today is probably not achievable, but wide nature strips between opposing streams of traffic minimises the risk.

          • Ghostship

            Lane keep assist nudges you away from the line but ultimately the driver can override it. The more advanced forward collision warning systems (think BMW) preps the brakes so when you press down on the pedal they come on at full power immediately. But an add-on device for lane departure warning and forward collision warning featuring a camera, controller and buzzer/warning light shouldn’t be too expensive if produced by the million in China. You could also use it as a dashcam which might make it attractive to insurance companies to lower their costs.
            Forming the centerlines with a textured surface that causes the wheels to make a rumbling noise would also help unless the driver was suicidal.

  • tomstickler

    When will it become sport to try to make an autonomous car “flinch”?

  • Randall Smith

    One has to assume this is going to be a traditional S-curve, with a very sharp turnover — I’d be surprised at anything over 5 years, and 3 years would be my guess from “approved by US gov’t” to “functionally required by law.” At which point, yes, it’s wholesale change time, although it’s worth noting that truck drivers don’t only transfer the goods — they load and unload, frequently being the only person qualified to do the unloading at the dropoff point. Uber is toast, though, except as the lowest-cost provider of applications software coding.

  • Ken_L

    It may be somewhere deep in the thread, but the comments I’ve read ignore one very important point: driving is fun for a lot of people. They enjoy it.

    I’d hate to be told I can’t drive anymore, even if there was no inconvenience or cost involved. It would be like a keen gardener being told to sit down and take it easy, we have robots to do all that for you now. Or a fishing enthusiast being presented with a boat that would find the fish, with tackle that would decide what lure to use and how to cast and retrieve it, leaving the angler to do nothing but put the catch on ice and drink beer.

    I bet lots of people will rebel and say “Thanks all the same, but we like to drive our cars.”

  • econoclast

    Many people on this thread are working from outdated information. There has been a revolution in “AI” over the last five years that has completely changed the picture for applications like self-driving cars. (I put “AI” in quotes because it’s not really AI in the way usually think of it.) This revolution has been barely covered in the media, and when it has been covered it has been covered badly.

    Computer scientists have developed a practical algorithm that allows computers to learn from data alone, with no human intervention. (It has the pretentious name “deep learning”.) Computer are still pretty dumb, so they require lots and lots of data, and they need a clear goal (like “don’t crash” or “win at Go”), but this has radically changed the picture for self-driving cars. Self-driving cars has gone from being an open research question to being a question of engineering.

    I don’t know how long it would take, and the history of predictions more than 5 years out is dismal, but it’s happening.

    • Computer scientists have developed a practical algorithm that allows computers to learn from data alone, with no human intervention.

      Unsupervised learning? That’s pretty old hat.

      (It has the pretentious name “deep learning”.)

      That generally refers to the number of layers or cascading steps. And it can be supervised (or involve “supervision” at various steps. And even deep learning is pretty old hat though has had a resurgence in popularity.

      But you’re way overestimating what it can do (and oversimplifying by saying that its an “it”). Consider voice recognition of various sorts. That’s a MUCH simpler problem (since you are dealing with a simple stream of audio data) and it’s not like deep learning has made auto captioning of YouTube videos as good as human captioning.

    • econoclast

      It’s supervised learning (which doesn’t imply that it requires human supervision, just that that the training data has to include a measure of success or failure). If by “has a resurgence in popularity” you mean “we didn’t know how to actually do it before, but now we do”, then yes, it has had a resurgence in popularity.

      And you radically underestimate how successful deep learning has proved. In 5 years, error rates on automatic captioning of images (ImageNet) dropped from 28% to 5%, which seems to be comparable to human performance.

  • Give me a functional and not ducked up autocorrect on my phone first. Then maybe I’ll trust a self-driving car.

    [Edit: see what I mean above? Autocorrect did that.]

It is main inner container footer text