Slashdot Mirror
Self-Driving Cars Aren't Going To Be So Great Until We Make Our Maps Better (theverge.com)
Uber is debuting its self-driving cars in Pittsburgh this month, a move that has many taxi drivers upset. The Verge's Nilay Patel argues that this move should change the way we think about maps and addresses. He adds that Uber is currently unable to pinpoint his home, and often ends up at the door of a "widely different address." Citing the CEO of a "large ridesharing company", Patel writes that this issue is known as the "egress problem" -- "the way we locate buildings on a map doesn't really describe how people move in and out of those buildings." Though there are workarounds and inventive ways to pinpoint your exact address, Patel argues that when we grow reliant on self-driving cars, things will get far more complicated and futile if we don't make our maps and navigation services better. He writes: Driverless cars are one of the ultimate signifiers of the future -- the real Jetsons stuff. And we're so close to making them happen: tons of cars have sophisticated adaptive cruise control that can basically keep you going on the highway, prototypes of true self-driving cars from Google and others are already on the road, and the momentum is only increasing. But maybe we shouldn't hand control of how we get somewhere to the machines until we're entirely sure the robots know where we're going.
Its not that hard to give out your coordinates if needed. That's a small issue for self driving cars, they have much harder challenges.
Sort of. That's only half of the problem. The other half is:
* Bad data
* Incomplete or missing data
In California and Washington I've seen incorrect map data both with Google Maps and Apple Maps. Sure a better driver _could_ (and should) be able to work around that but are they forced to in the first place??
The road data should both be:
[ ] complete
[ ] accurate
Obviously construction will cause some of that but it shouldn't take _years_ for a non-moving road to be added to the system.
--
GIMP v2.8 still sucks compared to Photoshop. Where are any of the Layer Effects???
There will always be something not on the map. The AI on the cars will need to be good enough to figure out what to do in many cases or else allow the passenger to manually maneuver the car. That is why many of the plans from google are about driving from a known location to another known location, as it may be decades or more before they'll be able to figure out how to get into and out of a condo garage, negotiate mountain roads, deal with temporary obstacles (the dog is in the middle of the driveway so there's no way to recalculate a route), etc. But pickup from a street corner and get dropped off at a street corner, that's much more doable.
No, the article is about human drivers of a cab company who can't get him to his front door. He then launches into a worthless rant about how self-driving cars need better maps.
If a human driver, using the same map and their set of eyes, can't get the guy to his front door, what makes him think a car programmed by humans will be any better, especially by humans who have never seen the place you're going to?
We will bankrupt ourselves in the vain search for absolute security. -- Dwight D. Eisenhower
I was helping a startup in the next town over, who had a package that FedEx couldn't deliver, so I agreed to drive over to the local FedEx office to get the package.
While I was there, the person at the counter pointed out that the delivery person couldn't find the address, and I explained how to go into the parking lot, down and around the building, to the front door of the startup.
It was indeed a weird situation where you can't see the front door from the road, and you had to know beforehand where to go to make deliveries.
The point here is, the FedEx person at the counter typed in my instructions in the "notes" section of their database and then assured me that further deliveries should go through OK.
Will it *really* be that hard to do something similar with self-driving cars? By which I mean, report an error to the company along with the correct data, or manually direct the car to the correct destination and note the error, and similar work around.
I'm not sure this is a terribly important issue. I mean, it sure *seems* like there's a simple solution and the problem will quickly be self-correcting.
If it runs through my yard on the way to your location, then there's going to be a problem, a BIG problem.
You see, all maps have errors - printed maps, Garmin maps, Google maps, ALL maps. Garmin shows streets that aren't there and misses streets that are there. Google maps shows my house two doors down from my house. I've been using maps on my job for the past forty years. I've seen it all.
Also, street names/numbers are not permanent, they change, and so do addresses. New streets and neighborhoods are built all the time. Hell, I've been on streets where odd and even are on the same side, and where houses right across the street from each other have the same number. I've seen houses on streets where the house numbers run like - 402 next door to 1607 next door to 723. I've seen street signs with four different names for the two streets that intersect. BTW, it's not the houses that are numbered, it's the lots. Even vacant lots have numbers.
Mapmakers insert some errors on purpose. They can tell by the errors if somebody else is using their data. But even traditional paper maps have errors. I bought a map years ago (Rand McNally) that switched the names of two cities. How can you trust a map for tiny details when they can't even get the names of major cities right?
Humans can adapt to changes quicker than machines. Maybe one day AI will be good enough to totally trust. We're just not there yet despite all the wishful thinking.
How about the joy of finding a way to tell a fully autonomous car to dart into a gap at the airport arrivals/departure scrum?
There is no scrum, the traffic computer scheduled everybody's access already and instead there is a timer on the dash telling you how long the car will be stationary. The scrum is caused by two things, ignorance (of where everybody is and where they are going, etc) and indecision. Traffic computers will solve both of these.
Even in the shorter term, where there won't be an airport traffic computer, and there will be a mix of human and computer drivers, there won't be much problem because generally you will be waiting in a line for the next available spot, and the computers can probably handle that situation fairly easily. There might be edge cases where your car wouldn't take the first spot because somebody kicked a traffic cone into it, and a human driver wouldn't car about clipping the cone, or driving near it. But those won't be major frustrations to somebody reading a book or jeejah and waiting for the car to tell them that they have arrived.
Heck, the fastest computer can barely beat people at chess
No. Even a relatively slow computer, such as a typical laptop, has more power than Deep Blue had in 1996, and can easily beat a grandmaster.
A slow off-the-shelf chess computer from the '80s can beat well over 90% of the population at chess. In the chess club when I was at school, I think that there was only one person who could beat it on its hardest difficulty setting, and he was the under-13s UK chess champion. The fact that it took Deep Blue to beat the best human player in the world is irrelevant: self-driving cars don't have to be better than the best possible human driver, they just have to be better than most human drivers to be a big improvement.
Chess is also an irrelevant comparison, because the problem is very different. In chess, you have 16 pieces at the start. Once you've made a few moves and they're all free to move, each one has multiple possible moves. Let's simplify and assume that each one has only one possible move. At the end of my turn, there are 16 possible board positions. At the end of your turn there are 16 possible combinations for every one of mine, so that's only 256, but after another round we're up to 65,536 different positions. 16,777,216 after three rounds and so on. Almost all of the difficulty in chess is working out which part of this space is worth exploring. Your goal is to reach an end state dozens or hundreds of moves into the future that meets some conditions.
In contrast, when driving there the other cars have few options (speed up, slow down, turn) as do other obstacles. Your model only has to run tactically, not strategically. You don't have to worry about every step in the game, only that in the next round you become closer to your destination and you don't crash. You only have to model a few seconds into the future. As long as you are heading in the right direction and you can safely stop if one of the obstacles that you're tracking has a comes into your projected path then you win that round and you continue to the next.
I am TheRaven on Soylent News
I think you are creating generalities from your specific situation.
No, I specifically said it varies; did you miss that? You even quoted it. I also said that addressing is controlled by local governments, so places with alleyways are obviously going to be handled differently.
AFAYK. But it's not that way in real life. There is no "default". You have to know.
No, you'd don't "have to know". Enter some lat/lon coordinates into Google Maps, and it'll show you a location on the map. It doesn't ask you for your datum. That's because there IS a default.
That's funny, because I can get my location in any number of datums using GPS. Wikipedia isn't always right..
And I'm supposed to believe you over a cited article? If you think it's wrong, then go correct it. From a little bit of Googling, what I've read supports Wikipedia:
http://www.gpsinformation.org/...
http://gis.stackexchange.com/q...
The only reason anyone uses other datums is because they have old maps that are based on them, not because they're better in any way (they're not).