Uber is rolling out self-driving cars in Pittsburgh, which is cool — but it means that the way we think about maps and addresses is about to get really complicated. For example, here is something that happens to me several times a week: I call an Uber to my apartment building, and the driver gets lost because they can’t find the door.
See, the address of my building is on a main street, but the actual entrances are on the sides of the building — and when I drop a pin at those locations, Uber maps them to nonexistent addresses. So the drivers either pull up to the correct building address and don’t see a door, or they pull up to a door that has a wildly different address on it.
A few months ago, the CEO of a large ridesharing company told me this 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. There are probably a lot of solutions to this problem — you could add a button to the map that lets people mark the actual doors of various addresses, you could predefine set locations at each address for riders and drivers to meet, and so on — but right now it’s a real problem that most people in cities have encountered with some frequency.
For the moment, it’s a pretty minor issue — the easiest solution is just for the drivers to call the rider, and it works itself out. But you can’t call a self-driving cars and say "oh, I live in the white building and door is around the corner." You can’t say "oh, actually the route you’re following in Waze is wrong and you should just turn right here" to a robot driver. And you certainly can’t wave your arms frantically to have a driverless car pull over and stop as they blow right past your door to the preprogrammed address on a totally different street.
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.
Comments
But also what happens with roadworks and all the general complexity and chaos of a metropolis?
Isn’t the driverless car close to a PR boondoggle? As in we’re almost certainly decades from a vehicle capable of autonomously seeing, reacting and navigating any city?
By swimtwobirds on 08.24.16 2:39pm
Same here – I simply cannot see a self driving car in any of the French city or villages.
Even in Paris, you’ll find streets so narrow or clutered with nonsensical visual clues that even humans do pause to understand what’s going on!
By FReDs72 on 08.24.16 2:54pm
Don’t think of it that way. Apart from road rules, the only clues a self-driving car needs is clearance on the sides, above, behind and ahead. Sensors will take care of that. If anything, it will tackle these types of situations far more efficiently than humans, in particular if all cars involved are self-driving and share data. Lastly, if it can handle 90% of roads in self-driving capacity, it’s already a huge leap.
By MindYourMind on 08.24.16 3:01pm
Sensors won’t do it. It has to be able to see, read temporary signs, understand the context of essential and non-essential signs, tell the difference between a dog, a trolley and a child…
It’s why I think it’s bull. A truly autonomous vehicle is just another word for unsupervised learning in unpredictable environments. That’s rock hard general purpose AI. And we’re so far from that we have no frame of reference for knowing if it’s even possible based on what we’re currently doing. All that machine learning neural net stuff is massively oversold..
By swimtwobirds on 08.24.16 4:17pm
There are still tons of roads where they can be used outside of city centres. Self driving or not, but city centres are not made for cars and cities try actively to get rid of cars there anyway. Sure there are some terrible streets where I live too (legacy of all medieval cities), but there are also big avenues where drivers are separated from cyclists and pedestrians and where the problem is not as big. For example, a driverless car IMHo should get me to work from where I live easily, because there are no narrow roads or passengers running in front of cars like mad.
As Google guy said, they won’t be able to handle 100% of cases, but they should be able to cope with many peoples’ commute or weekend getaway.
By navaicojmaniskaidsasu on 08.24.16 4:23pm
Totally – that is for sure true. Long haul freight would spring to mind. As long as they can successfully segregate the lanes I guess. It’s just every now and again you hear about the immense legislative safety hurdles, the fact that a lot of people think that, in the absence of hard AI, they might need to put stuff in the roads to ensure safety for that kind of fleet…
I just think a lot of people (cough – Google – cough) are wildly talking it up as basically a massive futurist PR exercise. I think there is tons of hot air and huge huge obstacles that never get referenced because it’s a grabby thing for a media site to do a short article on. And it gives Google, say, a Willy Wonka sheen. Makes them look a little less like a search monopoly completely reliant on advertising…
By swimtwobirds on 08.24.16 4:35pm
What the hell??!? That’s like, exactly the opposite of what Google said…
By alexander.riccio on 08.24.16 8:04pm
I’m going to agree with you here. You can do lots of cool stuff with neural networks, but there’s very little neuroscience involved, so you’re going to have trouble getting anything more than superficial intelligence out of them.
The key word is trouble. The half-assed attempts by everybody and their mother to catch up to Google’s self-driving car are really just PR.
Google figured years ago that maps cannot be used as the primary source of information. That’s why it very aggressively looks at everything around it – far more than a human could ever do – and interpret behavioral cues to actually do reasonable things.
They came up with all kinds of wacky scenarios in testing. They threw beach balls from behind rocks at the car. They hid people inside potato sacks and had them pop out while the car was watching. They had other cars break all of the rules of traffic at once. It took forever – they went to an obscene amount of trouble – but they got the car to be smart enough to do the right thing every time.
Later, when testing the car on the streets of Palo Alto, they did come across one totally kooky scenario that they didn’t anticipate: a lady in a wheelchair chasing a flock of ducks in circles with a broom. There is no sense to be made of that situation. Yet, the car slowed to a stop, waited for the duck-chasing-lady and the ducks she was chasing to get out of the road, and then resumed normal driving.
There’s a great Astro Teller presentation about this. Lemme see if I can find it.
By alexander.riccio on 08.24.16 7:47pm
That’s cool, but – and correct me if I’m wrong on this, but google’s car doesn’t work in the rain?
That seems a tad problematic.
By swimtwobirds on 08.24.16 7:56pm
The rain I think is fine. Snow is a total clusterfuck. Yup that’s a problem, no questioning that. I expect them to figure that bit out, but in the mean time the Google self-driving car might be the key to my 100 year old grandmother’s continued independence. She lives in Florida (of course), where she’s never seen snow.
By alexander.riccio on 08.24.16 8:51pm
Yeah I read it there again. Rains iffy alright. Also – and this might be out of date, but don’t google’s cars require the roads to be pre mapped for pavement height, traffic light height and stuff – also that it’s informed of the local speed limit beforehand.
Does that stuff still apply? Because realistically that seems a completely insane task at scale. And also one that will require almost continuous re-mapping given complexity and entropy at scale. Again, if that still holds true, you’d be inclined to call bull on the Google car. It feels a lot like a demo setup where it’s driving around pre-mapped roads in ideal conditions in SF?
By swimtwobirds on 08.24.16 9:16pm
That’s a completely different problem; let’s not confuse sensing and decision making.
Yes, LIDAR technology has slight limitations in rain and severe limitations in snow, but you remove these issues by using HDR video cameras in combination/instead of LIDAR systems.
Technical problems like these are comparatively easy to deal with.
By Indefinite Implosion on 08.25.16 11:01am
Here it is, about 23 minutes in: https://youtu.be/DIMqIGax8Co
I’m not actually sure if this is the exact version that I remember. There are much more technical in-depth videos around, lemme look for that too.
By alexander.riccio on 08.24.16 7:58pm
Disagreed, sensors are good enough to handle 90%+ of situations presented on the road. In fact as Google’s self driving project shows us, it’s had next to no problems on any type of road/urban setting and where it was involved in a minor accident, it had been entirely human fault.
I don’t think it needs to tell the difference between a dog, trolley and a child. They are all solid obstacles that should not be present on the road and should be avoided. Breaking irrespective of the obstacle should do it. A thermal camera and motion analysis algorithm would allow the software to differentiate between living and non living things if need be, it doesn’t need to be a sophisticated AI.
You are talking about the cars self-driving ability to handle unusual/extra-ordinary situations. Such as say a bunch of debris/rocks on the road, requiring the car to drive around off road/on the grass etc. Such situations are far and few and there’s nothing wrong with the car alerting the car occupant and handing over controls once in a blue moon. As long as the cars drives themselves for 99% of the time, that’s all that matters.
By MindYourMind on 08.24.16 8:26pm
I think people are expecting too much too soon.
What they need to do is limited route autonomy and Paris (and Washington DC) are actually good cities to do it in because you have big clear routes which cover most of the city in a grid system and relatively few high buildings to interfere with GPS.
You have unmanned Mêtro, its only a small step to unmanned small bus service over a limited patterm grid them it grow as technique improve.
Nobody’s going to have autonomous cars in the small medieval Provençal village where I used to holiday but does it matter? Non.
Autonomous grape harvesting maybe…
By Dr Strange on 08.24.16 3:06pm
Totally agreed on that part.. Here in India we have a tonnes of stuff blocking our roads.. Many a times with the owners not in place to move it.. The driver’s themselves sometimes get out of their vehicles move stuff out of the way (for ex if it is a bicycle) and get along with their business..
Now what would a stupid self driving car do in this case..? Nothing.. And if it’s an uber, it probably won’t even allow the passenger to get down from the car..
By rahul.rn on 08.25.16 11:44pm
Have you really never watched any of google’s videos? These cars are already on the streets and they handle everything you mention. This tech is less than 5 years from being commercial.
By Avianzz on 08.24.16 3:17pm
One thing I (we) will surely love about self driving car is once I (we) understand how the AI will behave in some situation I (we) will take full advantage of its behave. The same way I (we) bully AI on our computer game.
If those sensors are so good which will allow the AI to avoid most of the collision by giving lane. People soon will inevitable carelessly keep on cutting the lane of self driving car?
ps Look how this smart dude trolls automatic door
By Mr.Mulderfox on 08.25.16 4:01am
So why don’t you drag the pin where it’s supposed to be and tell google to fix it? They usually approve those types of fixes very quickly.
My house is one of 3 on a single lot so I moved my pin directly on my front door, now when you google my address that’s where the pin lands.
By Avianzz on 08.24.16 2:45pm
That only works if the right address is on the same street. I have tried this dozens (hundreds?) of times.
By Nilay Patel on 08.24.16 2:53pm
I just tried an edit like that. I moved the marker, put in a more accurate address, and within minutes Google "approved" my edit. The only problem is nothing actually changed on the map.
By rahulp on 08.24.16 3:30pm
I take that back. They did change the address! But the directions are still bad
By rahulp on 08.24.16 3:34pm
I had the same thing happen. Our driveway branches off the main road a bit, but is still technically "Multiview Drive". The problem is, on Google Maps the driveway is greyed out and not a part of Multiview Drive. So, while reporting a problem with my address a few months ago, they moved the pin to the correct location, but directions are still off. I’d imagine a grayed out road is something they actually have to send a street view car out to map, photograph, and fix. So not likely to happen any time soon, if ever.
By ryan.folks on 08.25.16 12:27am
Wrong. Go to google map maker and fix it yourself. Or report the problem. If they can see it on satellite they will approve it.
By Avianzz on 08.28.16 2:32pm
The self-driving cars themselves and the sensors on them should be used to create maps.
If the user/driver permits, just send back position, terrain, 3d mapped and other imaging data to create ultra accurate, ground level, 3D maps.
Take an off-road route, map it, upload it back to Tesla, with every rock, bump and hole. Wouldn’t that be awesome.
By MindYourMind on 08.24.16 3:04pm