MIT Technology Review has an interesting article about the current state of Google's self-driving cars. Specifically, the myriad limitations that aren't usually mentioned in articles about the technology.
Some examples from the article, which are sourced from Google itself: The self-driving car can't drive itself on about 99% of the roads in the U.S. because it requires hyper-accurate maps which haven't been created for those roads. The maps are resource intensive (both computer and human) to create. It's very sensitive to changes in the stored maps; if a stoplight were erected without being included in the map, the car wouldn't know to obey it. The car hasn't been tested in heavy rain or on snow-covered roads. And it's generally incompetent at parking. There's a few other amusing (and terrifying) examples in the article.
Chris Urmson, the director of the Google car team who is the primary source in the article, says that he wants these issues will be worked out in the next few years, and believes in many cases the problems exist merely because they haven't gotten around to addressing them yet.
(Score: 2, Informative) by Leebert on Thursday September 04 2014, @12:39AM
Argh, you only notice your grammar errors once the submission goes live. "he wants these issues to be worked out".
(Score: 0) by Anonymous Coward on Thursday September 04 2014, @01:17AM
Good read, yhanks for the article!
(Score: 2, Interesting) by boris on Thursday September 04 2014, @01:45AM
Is it me or does detecting a red light seem a whole lot easier then detecting a child running across the street which I've seen them demonstrate in promotional material. I could see things like construction zones, and inclimate weather presenting real challenges.
(Score: 3, Interesting) by rumata on Thursday September 04 2014, @03:01AM
Hey,
a running child on a road surface is easy to extract because:
- A camera sees a moving blob in a uniform background (grey/black road, plus minus shadows).
- A Lidar sees a positive bump in a uniform flat surface.
- A radar sees a small doppler-shift, separate from static clutter, oncoming traffic and leading cars.
So, you'll have good signal in multiple sensors, no worries. Note that this does not require any understanding of the scene, it's simply moving-object-on-the-driving-surface--avoid. A tumbleweed, card-board box etc. would most likely trigger the same response.
The more interesting question would be if one can come up with a method to detect a child before it enters the road (say from the gap in between two parked cars), hopefully early enough that a defensively driving control algorithm will be able to avoid the kid.
The robot has some advantage here, in that it could get good all around situational awareness by throwing enough sensor and processing power at the problem. Humans still have _vastly_ better sensors/processing, but are limited by narrower focus, fatigue, and inherently being a point-sensor (so bits and pieces of the car can get into the way). Projected a few years into the future, robots may well start to have an edge here. Right now, imo not so much.
Now, why are traffic lights hard? They require scene understanding. You have to realise that there _is_ a traffic light, find and identify the correct one if there is multiples, then focus your Mk1 eyeball on a small part of it and check the colour of the tiny patch of light. The first few steps are difficult algorithmically, the last step is difficult in terms of sensing. Because traffic lights are usually elevated the (back) lighting situation varies widely, in fact so widely that _people_ sometimes have trouble, and that Mk1 eyeball has _way_ more dynamic range and than any affordable camera.
Now, seeing that traffic lights are infrastructure (built for human usage) the obvious engineering answer is to cheat and augment the infrastructure for robotic usage, e.g. use a wireless signal instead of that silly optical business or have the light blink with some pseudo-random code to lift the optical signal way out of the noise-floor etc etc.
Cheers,
Michael
(Score: 3, Funny) by c0lo on Thursday September 04 2014, @03:39AM
What?
The most cost effective and secure solution is to provide every person with a presence sensor (of course, for a price) which will make the person visible to the self-driving cars without those gimmicks that you mention. Of course those sensor will be used for multiple purposes, the more the better.
Traffic-lights? Same: it the responsibility of local administration to put and maintain an emitter which will tell the car that there's a traffic light and what color it is.
Hackers? Why, don't we have prisons for that? Just increase the penalties for hacking.
Asking for costly equipment and long and risky research? Where, in your suggestion, is the responsibility of a corporation to its shareholders? What you suggest is simply not capitalism (aka "socializing the cost and privatizing the profits") but ... how shall I put it?... yes, of course, an opposition to progress. You simply didn't think of lost jobs, did you?
https://www.youtube.com/watch?v=aoFiw2jMy-0 https://soylentnews.org/~MichaelDavidCrawford
(Score: -1, Troll) by Anonymous Coward on Thursday September 04 2014, @09:44AM
> The most cost effective and secure solution is to provide every person with a presence sensor
WTF? You going to put those on wild animals too?
Because if your "most cost effective" solution doesn't include wild animals then the system still has to deal with them and re-purposing it to also handle people and domesticated animals is only a marginal increase in cost.
Man, geeks sure can be fucking idiots.
(Score: 2) by c0lo on Thursday September 04 2014, @02:27PM
Kill them. What's the army good for if it doesn't protect the... ummm... national interest?
They sure can.
(large grin)
https://www.youtube.com/watch?v=aoFiw2jMy-0 https://soylentnews.org/~MichaelDavidCrawford
(Score: 2) by emg on Thursday September 04 2014, @06:30PM
(Score: 2) by EvilJim on Friday September 12 2014, @04:38AM
I could see fake stop lights being real challenges... say you're driving through your local red-light district and your google car stops permanently
(Score: -1, Offtopic) by SoylentsISay on Thursday September 04 2014, @02:03AM
Slashdot??? Ok I said it, the King has no clothes!! So, you copy Slashdot, and expect... to be different somehow?
Wake UP!!
I dunno, I kinda expected a site that wasn't me-too...
can we have relevant tech news pleeez?
I'm going to be minused into the ground for this comment and you know what, I DON'T CARE BECAUSE SOYLENT NEWS IS ALMOST AS MUCH BS AS SLASHDOT.
I really wanted them to be better, but instead of moderating the posts, they let the readers moderate the posts.
And what do you get? Mediocrity.
Duh.
Big Duh.
But they won't care, because Mediocrity creates popularity.
Thank-you Soylent Fox News. :)
You may now minus me into oblivion, because. rep on this site is worthless anyway. :)
(Score: 2) by mechanicjay on Thursday September 04 2014, @02:22AM
But the community moderating itself is kind of the point...
The difference we're going for here is to be a community by the community for the community, not beholden to a corporate profit driven overlord.
If you don't like the content here, please generate more of your own! The more voices and contributions we have, the stronger the community and by extension the site will become.
My VMS box beat up your Windows box.
(Score: 1) by MostCynical on Thursday September 04 2014, @03:16AM
Your helpful suggestion (which seems to amount to "wake up") will be taken on board.
You seem to have some anger issues, which are quite mis-directed. If you want a better site, help create it.
(also, if you want a world where "reputation" really matters, I suggest you read "Down and Out in the Magic Kingdom" by Cory Doctorow http://craphound.com/down [craphound.com] On SN, mod points just help you filter comments - nothing more)
"I guess once you start doubting, there's no end to it." -Batou, Ghost in the Shell: Stand Alone Complex
(Score: 2) by mrcoolbp on Thursday September 04 2014, @04:14AM
I'm going to agree with the other commenters here. This project is completely volunteer-run. We set out to make our own path. If you don't like what you see, feel free to submit some stories or jump on IRC [soylentnews.org] and help us out.
Also, I don't see it as "copying", it's more of a fork. We're not perfect, we admit that, but we want to be great.
(Score:1^½, Radical)
(Score: 0) by Anonymous Coward on Thursday September 04 2014, @09:48AM
Dude you are the problem.
You've made zero story submissions.
Every single post you've made here has been whining that the site is not what you want it to be.
So, judging by your own actions you want soylent to be nothing more than whines about how crappy soylent is.
(Score: 2) by kaszz on Thursday September 04 2014, @02:14AM
Seems like Google-car needs decades of more research. In the meantime it can be confined to follow the next car in queues etc. Even DARPA terrain cars seems to outdo this performance.
In the meantime bet your city tax money on personal rapid transit [wikipedia.org] (PRT). It will take you to your selected destination regardless of rain, snow, and other environmental factors.
(Score: 3, Interesting) by theluggage on Thursday September 04 2014, @10:42AM
I haven't heard Google promise that they'll be in the shops for Christmas.
Except the Google Car is likely to become credible as a PRT-like system long before it is a viable private car. Like PRT, the cars would only need to operate on well-defined routes covering a limited area - greatly simplifying the mapping and 'training' process. Unlike a PRT, there would be no need to construct special tracks or guideways for every inch of every route - just a limited number of 'problem' locations.
(Score: 2, Informative) by MostCynical on Thursday September 04 2014, @03:05AM
So the car can only drive as well as the maps allow.
At the moment, maps are pretty terrible.
http://theweek.com/article/index/243813/8-drivers-who-blindly-followed-their-gps-into-disaster [theweek.com]
Radar braking, radar cruise control, infra red sensors...All these things are already working.
http://www.daimler.com/dccom/0-5-1210218-1-1210321-1-0-0-1210228-0-0-135-0-0-0-0-0-0-0-0.html [daimler.com]
The car needs to behave more like the DARPA truck: http://www.youtube.com/watch?v=1pl_Pont_Zk [youtube.com]
And less like a pick-and-fill robot http://www.packworld.com/pick-and-place-robot [packworld.com]
There is a possibility that Google will insist all roads be painted with special coductive paint, that all street signs have rfid tags, and that only self-driving cars will be allowed on the specially-marked roads.
Also, you will have to watch ads on the windows. Cars with adblock will be restricted to 10km/h.
Yay google!
"I guess once you start doubting, there's no end to it." -Batou, Ghost in the Shell: Stand Alone Complex
(Score: 3, Interesting) by emg on Thursday September 04 2014, @04:18PM
The problem is, Google is going the wrong direction, trying to magic up a car that can do everything perfectly and drive itself.
Actual auto manufacturers are starting small, with collision-avoidance radar and cameras, and lane-following adaptive cruise control. Starting simple and adding new capabilities as they're developed is far more likely to result in a real driverless car than Google's approach is... but it doesn't get fawning media attention.
(Score: 2) by JeanCroix on Friday September 05 2014, @01:41PM
All their millions of miles driven with no accidents means nothing to me until they can do it in the winter. I can't be a believer until I see one of these things make it up my steep driveway with 5 inches of fresh snow, and without spinning out and ending up stuck sideways in the lawn.