What self-driving cars still cannot do

While I am enthusiastic about self-driving cars and its promise of increasing the mobility of those unable to drive as well as the possibility that such cars may be better drivers than humans, via Machines Like Us I came across this article by Lee Gomes says that we should not overestimate what they can currently do, because the road tests that they have done so far that produced 700,000 miles of accident-free driving have been under very limited conditions. There is still a long way to go and some major technological hurdles to overcome.

He says that the cars cannot be driven on 99% of the current roadways, cannot park themselves, cannot be used in heavy rain or snow, and cannot detect potholes. So Cleveland is definitely out for the near future.

Many of the remaining problems are because of the map=based technology these cars use.

Google often leaves the impression that, as a Google executive once wrote, the cars can “drive anywhere a car can legally drive.” However, that’s true only if intricate preparations have been made beforehand, with the car’s exact route, including driveways, extensively mapped. Data from multiple passes by a special sensor vehicle must later be pored over, meter by meter, by both computers and humans. It’s vastly more effort than what’s needed for Google Maps.

Google’s cars are better at handling some mapping omissions than others. If a new stop light appeared overnight, for example, the car wouldn’t know to obey it. However the car would slow down or stop if its on-board sensors detected any traffic or obstacles in its path.

Maps have so far been prepared for only a few thousand miles of roadway, but achieving Google’s vision will require maintaining a constantly updating map of the nation’s millions of miles of roads and driveways.

Among other unsolved problems, Google has yet to drive in snow, and [director of the Google car team Chris] Urmson says safety concerns preclude testing during heavy rains. Nor has it tackled big, open parking lots or multilevel garages.

Pedestrians are detected simply as moving, column-shaped blurs of pixels—meaning, Urmson agrees, that the car wouldn’t be able to spot a police officer at the side of the road frantically waving for traffic to stop.

The car’s sensors can’t tell if a road obstacle is a rock or a crumpled piece of paper, so the car will try to drive around either. Urmson also says the car can’t detect potholes or spot an uncovered manhole if it isn’t coned off.

Sigh. It looks like I may not be able to get a self-driving car by the time I am incapable of driving a car myself.


  1. AsqJames says

    Google’s vision will require maintaining a constantly updating map

    I hadn’t thought about this previously, but that strikes me as an almost insurmountable technical challenge from a security point of view. When you have safety critical information being transmitted on an ongoing basis the possibility exists for someone to hi-jack the signal or find a way to smuggle bad information into the genuine signal. However slight that possibility can be made, if you have millions of vehicles depending on it not happening the potential impact of a security breach are so high it’s still a massive problem.

    I’m no expert though, so maybe it’s really not that simple.

  2. Crimson Clupeidae says

    I think by the time this is ready to be widespread, the cars will rely more on sensors and processors (visual, short range radar, wifi network with nearby vehicles, etc.) and will be not very reliant on a built in database of roads.

    The technology is getting there, but it’s hard (for me, anyway) to say when it might be viable. As one of the few people who actually enjoys driving, and pays attention (my cars are all standard transmissions, and I mostly commute on my bicycle and motorcycle), I think most drivers will eventually be glad to not have to focus on driving, since they don’t now….

  3. moarscienceplz says

    This strikes me as a bit of Ludditism. Even if all a computer-contrilled car can do is drive in an interstate in fair weather, that is a big boon for truckers and others who have to drive long hauls. That is a big improvement! Self-driving cars may be a gradual thing, but so what? Even if I have to hold on to the wheel and merely have a robotic back-seat driver, that is better than what we have now, and will probably save thousands of lives!

  4. DsylexicHippo says

    It is not Ludditism to question their ability. The current runs are carefully staged.

    The reality is that there’s still a long, long way to go before self-driving cars would merit serious consideration as viable alternatives for human driven vehicles. For me the technological benchmark would be when the self-driven car would be able to handle unplanned events, specifically how it would handle a police officer directing traffic in a way contrary to normal traffic rules. Remember, the police officer is the law and her on-the-spot instructions override all written traffic laws. This would require a level of AI sophistication that I don’t think current generation technology could handle.

  5. Who Cares says

    @Crimson Clupeidae(#2):
    Those cars already have a long(er) range radar for adaptive cruise control, ladar/lidar for environment detection (effective range 100 meter), short range radar for near object detection, cameras to identify objects detected with the ladar/lidar & ocr signs (note that road markings are included as cameras are being used for lane keeping assistance), and ultra sound for very close object interaction (for example when turning on the parking assistant). More sensors will only be added if there is a clear need to track another class of object that can’t be done by the current set.
    The biggest step up on maps would be if they can ditch the ultra-ultra-ultra-high quality (1 cm details are mapped) ones they have to make now and just use GPS coordinates, but it will be a while before they can get the cars software to react flexible enough that that is all that is needed for position keeping.

    The can’t park and can’t handle potholes are spurious since the car software can be trained for those situations.
    It is weather and the need for those centimeter precision maps, as I mentioned in an earlier post, that are the big stumbling block. I would add in bad maintenance of roads so that lines and signs on it are in a bad condition but that is technically the same problem as trying to drive on a (partially) snowed in/on road.
    The problem with bad weather is that it can significantly degrade the effectiveness of the different sensors, I don’t think they’ll ever find a way to effectively push that laser through fog or heavy rain. The problem with snow, or even heavy leaf fall, is that it alters the environment so that it no longer matches the maps created.
    What DsylexicHippo(#4) says about police officers directing traffic is true as well. The car currently can detect pedestrians but can’t identify the type. So unless that officer would be using a traffic sign, like a pedestrian crossing guards stop sign, (s)he would just be classed as a pedestrian and ignored. Same goes for people being nice and giving signaling you to move while they have that right.

  6. says

    Perhaps the police will have ways to communicate securely with nearby cars anyway, to clear access for emergency vehicles and such. Probably also including a ‘slow down, park, and shut off’ mode of override. I’d be surprised if they didn’t develop some kind of IFF system to allow police to identify themselves securely.

    As to being hackable…our current traffic control systems are hackable right now. If anything, the massive increase in litigation possibilities should help carmakers focus on securing their products well.

  7. jesse says

    This strikes me as a bit of Ludditism. Even if all a computer-contrilled car can do is drive in an interstate in fair weather, that is a big boon for truckers and others who have to drive long hauls.

    You are aware that truckers must often operate in less-than-optimal weather? Probably half the trips you take will be in rain or snow. The trucking industry volume doesn’t drop because of bad weather, and the only thing that stops them is pretty serious snow or storms.

    The issue with self-driving cars is that the current technology can’t handle the very complicated business of navigating anything that isn’t prepared beforehand. So for self-driving cars to be a real alternative for anything but carefully managed routes, they have to pack enough processing power to account for all the unpredictable things that even a well-maintained road can throw at you. And if they only work in sunny southern California weather then that means they will only be good for sunny southern California from about March to October. (It does rain there, you know, but like many semi-desert climates there is a distinct rainy and dry season).

    Even sensors built in have limits, and it just shows that human senses are pretty damned good. Lidar would be pretty well useless in the fog. Radar would too, at certain wavelengths. (Like the ones weather radar uses). So you need something at many different frequencies. Then you need a system that can account for all the cars around it. Some of this is addressed by internet-based data sharing, but then you need a rather large infrastructure in place.

    On top of all that is the security issue. Any over-the-air system is vulnerable, and even someone who isn’t being malicious could accidentally kill people – pranks when you are moving at 35 miles an hour can go horribly wrong.

    All in all I am not that optimistic about the future of self-driving because even though in a purely technical sense it is possible, there are simply gigantic organizational obstacles and you end up creating more problems than you solve. Some technologies are technically cool, but they never got used because they were like that — creating problems where there were none before that made them tough to use. For instance, we could theoretically make fingerprint keys ubiquitous on apartment doors. But why aren’t they? The technology itself isn’t that expensive.

    Well, they need electric power, they are easily damaged, they would have to be changed every time a new tenant moves in and if they stored the fingerprints in a central database you’d better hope your local internet connection never has any glitches.

    Metal keys are a centuries old technology, and they work with a minimum of fuss. Problems are easy to fix on the spot. It is the same reason doors open like, well, doors on hinges, rather than the sliding doors so beloved of science fiction. Pens and pencils and paper — I could try hand-writing notes on an iPad, but if you want to just dash off a shopping list or a note to someone to call them it’s just simpler to use a post-it and a pen.

    Automated cars have the singular problem of adding steps and vulnerabilities to an already complicated process (driving and traffic control). That’s not usually a recipe for success.

Leave a Reply

Your email address will not be published. Required fields are marked *