This Company Is Crowdsourcing Maps For Self-Driving Cars (wired.com)

mirandakatz writes: If we want self-driving cars to become mainstream, we need maps -- and not just any maps. We need ridiculously detailed and constantly updated maps of the world's roads. And there's a mad race among startups to become the definitive provider of those maps. At Backchannel, Steven Levy takes a deep look at Mapper, a startup that just came out of stealth today and that hopes to become that definitive provider by crowdsourcing the production of those maps, paying drivers to drive around with a special mapping device on their windshields. As Levy writes, "Mapper's solution is to create an army of part-time workers to gather data that will accrue to a huge "base map" for autonomous cars, and to update the map to keep it current. Think of the work as an alternative to driving for Uber and Lyft, without having to deal with customer ratings or backseat outbursts from Travis Kalanick."

Comments

AHEM

[kfh227]

Google already has the tech for this. They demoed it. Their OS will be everywhere including cars. I'm sure that this war has already been won.

Re:AHEM

[JustAnotherOldGuy]

I'll bet the business plan is "become a big enough threat to be bought out by Google for a few billion, and then laugh all the way to the bank."

Bingo.

OpenStreetMap

[Midnight Thunder]

Sounds like something OpenStreetMap already does, though there are probably accuracy and liabilities limitations? Though, based on the following sentence, I would suspect the map data they produce is not going to be open:

Mapper’s solution is to create an army of part-time workers to gather data that will accrue to a huge “base map” for autonomous cars, and to update the map to keep it current.

The other thing is whether they will reference publicly the sources of their data, otherwise there is likely to be a high risk of ripping off other sources and even including the same errors.

Re:OpenStreetMap

[jellomizer]

If self driving cars need such detailed maps, then Self Driving cars are not yet ready for the general public.

I know my GPS sometimes gets confused when going off a Ramp onto a highway, or a road parallel to a highway, and areas that fork rather rapidly. However this requires the self driving car to make the decisions not a detailed map. Because Roads change too often. Just this summer, we had some flooding that wiped out a good part of a road, for a few weeks, one lane of the road was closed causing 1 way traffic with a makeshift stoplight. Then after construction was done the road was shifted 15 feet to one side. If self driving cars are to be a reality, they will need to figure this stuff out by itself, and not with detailed maps. As this stuff can change faster (especially on low populated areas) then people can record such changes.

Re:OpenStreetMap

[Midnight Thunder]

Totally agree. Even humans get confused when trusting the map and GPS too much, instead of using reasonable judgement. There have been cases of vehicles ending up far from the expected destination, on a non-existent road or even trucks on roads that were too narrow for them. While software can compensate for some of the human errors, they will introduce others, at least at the beginning, because of sensory limitations or analysis limitations.

The other issue is GPS accuracy, whereby you can be happily going along with a 5m accuracy and then find yourself with a 100m accuracy, due to obstruction of the sky by tall buildings, cliff faces and trees. Heck I have gone for a run where there is a fairly clear path to see at least three satellites and still my recorded path jumped around, as if I was jumping from side to side of the street, even though I was running in a straight line.

Re:OpenStreetMap

[OzPeter]

If self driving cars need such detailed maps, then Self Driving cars are not yet ready for the general public.

It's way more of a problem than just maps.

How does an autonomous vehicle recognize a policeman holding up his hand to say "Stop".

Or (in the case of VA law) move over a lane when passing a police car (marked or unmarked) on the side of the road that has pulled a miscreant over.

Re:OpenStreetMap

[JohnFen]

Sounds like something OpenStreetMap already does, though there are probably accuracy and liabilities limitations?

I came here to say this.

OSM maps are very good. I have a hard time believing that you could crowdsource maps that are better.

Re:OpenStreetMap

[AvitarX]

1) they see the difference from their baseline map and exert extra processing power to that part of the image to better recognize the policemen and his gestures (which are usually opaque to me, so good luck to the autonomous car).

2) They move over for any stopped car on the side of the road, that one should be pretty easy.

Re:OpenStreetMap

[AvitarX]

Unlike a human, the car will be able to see while reading the map, it will know it didn't take the left fork, and better be able to position itself.

The crowd sourced maps will hopefully include the extra details you mention as problems with current GPS maps (I'm less optimistic there).

Re:OpenStreetMap

[vtcodger]

Complete, accurate, current maps would certainly help. But people have gotten by without them (mostly) for over a century. And I think autonomous cars are going to have to be able to do so as well. Otherwise every street line painting project, pothole repair, jack-knifed semi-trailer or new/altered traffic signal is going to throw them for a loop. One of the requirements for a serious autonomous vehicle would seem to be to be able to distinguish a road from a parking lot from a playground and handle each appropriately. They'll also need to be able to recognize and read street signs -- although that might require some sign standardization.

My bet is that open-street maps will be more than sufficient for mature autonomous vehicle technology.

There are BTW a few other, larger, problems that need to be solved e.g. non-standard traffic control devices (flashing red left turn arrows for example), "double intersections" where the traffic lights are easily confused with those for another intersection, etc, etc, etc.

Re:OpenStreetMap

[Kjella]

Eventually yes, once you want to have cars where nobody else is licensed or capable of driving. But for the most part though I'd be very happy with a car that needs a human to take over every time the baseline changes. Because 98% of the time, 98% of the road it's exactly like last time and if SDCs become popular 98% of the changes have probably been recorded by somebody else. It's the reverse of the OTA updates to the car, "here and here and here my human had to take over", you upload the sensor data and they run it through some heuristic to either automatically or manually set a new baseline. And if they find a lot of these corrections/changes are simple, they might let the car improvise some of them on the fly.

Which is also why I think this business idea is stupid and basically a VC scam, say now they start up that taxi service in Phoenix. Pretty soon they can probably sell a consumer car that's useful to Phoenix residents, maybe they add the main highways in and out of Phoenix too since those are fairly simple. It won't work all the time, but it'll work much of the time. And all those people - or at least sufficiently many - will happily let them record the sensor data when they go out of coverage. What do they get? A car that might drive itself next time. It's a win-win, why pay someone to map out the roads when people will probably do it for free? All you need is to give them an SDC that works some of the time....

Re:OpenStreetMap

[fluffernutter]

I know someone who works in the GIS department of my city. There is no freaking way they have the resources to keep up with this.

No.

[Type44Q]

If we want self-driving cars to become mainstream, we need maps

No. If these systems require more accurate maps [than human motorists require] to keep passengers alive, then the self-driving peons still don't get it.

Agreed

[Baron_Yam]

A human can handle things like, "Upon reaching the entrance to that rural property, follow the gravel until it looks like a better idea to be on the dirt path, and maybe park on the grass next to that other car".

People are already talking about self-driving cars without any operator controls, and I just don't see that happening without a general AI, because there's so much knowledge and abstract understanding of the world required to operate a car in unusual circumstances.

Or, for those of us in snowy climates, to operate a car when the road isn't visible and you have to make judgement calls on where the pavement most likely is based on whatever cues you can find.

Or how to cooperatively route through or around construction without being so polite you sit there forever while all the pushy drivers jam in front of you.

I doubt I'll still be alive before there's a car that doesn't need a human to operate it a significant amount of the time. I would very much like to be able to put my car on 'autopilot' while on a long highway drive in good conditions and take a nap without fear of dying in my sleep, and I think that's achievable.

Re:Agreed

[fluffernutter]

Along those lines, frequently in the winter snow clearing equipment will be blocking your lane and humans are good at just slowly driving around them in the opposing lane. The loader will stop, and you go around it. This kind of situation is going to paralyze a self driving car.

Obligatory xkcd

[Allasard]

https://xkcd.com/1897/
Oddly, only from a week or so ago.

Re:Obligatory xkcd

[Midnight Thunder]

https://xkcd.com/1897/

Oddly, only from a week or so ago.

Crowdsourcing also depends on the crowd for the most part being non-malicious and honest. Just imagine 'pranking' the system and the potential outcomes.

Re:Obligatory xkcd

[DerekLyons]

Crowdsourcing also depends on the crowd for the most part being non-malicious and honest. Just imagine 'pranking' the system and the potential outcomes.

Happily, filtering data to remove the outliers is a long solved problem. This makes pranking the system much more difficult.

Tesla

[EnsilZah]

And Tesla is getting this data from every car they sell, for free.

Re:Tesla

[Trailer Trash]

And Waze, which is part of Google. When people come out with big announcements like this, I have to wonder if they're either 1) stupid or 2) just trying to sucker some VCs out of cash.

Re:Tesla

[just another AC]

not sure tesla et al is gathering the lidar etc data to get the level of quality needed for autonomous progess

(although given enough investment in algorithms they can probably work with poorer source data)

Windshield blocking device?

[Anonymous Coward]

That mapping device, if that is the current iteration, is illegal in most if not all states. It is below the AS1 line and obviously blocks a large amount of view.

We need maps

[OneHundredAndTen]

Which just shows, once again, how pathetically limited AI systems are. Albeit we people do need maps, when driving we make our immediate decisions based on the information we get in real time, without having to cross-check with a map.

Re:We need maps

[ThosLives]

This was kind of my thought too - why do they need maps? I learned to drive without a map. I had to learn directions to local areas as I went. And when I do need a map, they are not "ridiculously detailed" - a road atlas has a pretty coarse scale.

Maybe part of the difficulty in achieving autonomous cars is that people seem to be trying to have the autonomous agents "know everything" at time of manufacture, rather than learn as they go.

Re:We need maps

[110010001000]

What we call "AI" is BS. There is no such thing. Computers don't "learn" and they never have. They just run programs.

Re:We need maps

[JohnFen]

I agree that "AI" as its currently being imagined is something that doesn't exist and isn't close to existing.

However, if you define "learn" as "adapting behavior according to past experience", then machines can certainly learn. They've been learning for a very long time. But learning does not require AI.

will they cover full data plans + roaming

[Joe_Dragon]

will they cover full data plans + roaming as this can really run that up. maybe even say 10GB satellite internet plans.

also just hope that for fords and others don't rip people off on map updates / hdd upgrades.

What we need

[nospam007]

is self-driving cars that collect self-driving maps data for self-driving cars.

accuracy some time you are off by one house

[Joe_Dragon]

accuracy some times you are off by one house where the numbers are not lined up right to the gps.

Payver is doing this too

[wirelessjb]

The Payver iPhone app is doing something similar. I've earned $30 already taking videos out the front of my windshield (woo hoo). Unfortunately it's iPhone only, and you have to supply your own mounting hardware, but it's an easy way to earn some lunch money. As a bonus you can tap the screen if you get in an "incident," and it will save before and after video.

Comma.ai

[DrYak]

as does comma.ai

(With the small difference that their software is opensource,
and their hardware is not vendor specific)

Down to the exact lane

[DrYak]

The point of self-driving grade maps is to be precise down to the exact lane.

like "the two left lane are for turning left. the two right lane should be for 'straight ahead', except that the second from the right is currently blocked by construction works".

(i.e.: even more precise than what's available on commercial satnavs).

OSM doesn't go that precisely into details.

For that you need to crowdsource it from video feeds :
- the first time the car goes there (either in actual autopilot mode, or in 100% pure manual mode with the camera only working as a glorified dash-cam, like comma.ai's chffr android app), do object recognition on the video feed, notice the street signage, lane arrows and construction work, and beam the info up to the mothership.
- the next time one of the car goes there in autopilot mode while connected to the same cloud, the car knows to change to the right most lane, if the satnav's path finding calls for continuing straight ahead.

Google^H Waymo, Tesla, Comma.ai, and countless others are doing this kind of stuff now.

Re:Down to the exact lane

[JohnFen]

(i.e.: even more precise than what's available on commercial satnavs).

But not more precise than Google Maps, which does exactly this.

OSM doesn't go that precisely into details.

True. I wasn't saying OSM was good enough for this task -- I was saying that it's hard to think that a new crowdsourced service would do any better.

The truth is, even the gold standard for this sort of thing (Google Maps) doesn't have the amount of accuracy required. And can't. These things change too frequently to actually rely on them for that level of detail.

For that you need to crowdsource it from video feeds

That would be an improvement, certainly. It's still hard to see how it would be effective enough in practice as a crowdsourced venture. How are they going to convince enough people to outfit their cars with dash cams?

Maybe they have some secret strategy that will make this work. I just don't see what it is at the moment.

Connectivity

[fluffernutter]

What about areas without connectivity? If self driving cars *need* these maps to work then how will they ever work if you drive away from an internet connection?

Re:Connectivity

[ItsJustAPseudonym]

How does a GPS navigation system work without internet connectivity? Answer: it has the map in memory. Likewise, the self-driving cars won't need an internet connection, they will have the map in memory.

Gathering the map is another matter. They definitely need to upload the mapping data from their app at some time, but that does not have to be done in real-time. The app could just store the data and then upload it when connectivity is available.

Re:Connectivity

[fluffernutter]

So a car is going to have enough storage for every stop sign and traffic signal anywhere in a local database, and this will be kept up to date, how?

Re:Connectivity

[fluffernutter]

But what will it do when it doesn't have accurate map data because I've been driving in another country for a month and I don't have an internet plan there, and someone has knocked over a stop sign?

Re:Connectivity

[ItsJustAPseudonym]

Lawsuits, guaranteed.

Re:Connectivity

[ItsJustAPseudonym]

Replying again with a different thought.

For something this critical, the car will probably be provided with internet connectivity via the cellular data network. Heck, I used to have a Nook reader that included 3G, and the connectivity was paid for by Barnes and Noble bookseller. They used the connectivity to provide book offers and downloads. If they could do it for a little book reader, someone could certainly engineer it into a car.

Re:Connectivity

[fluffernutter]

If that's the deal then fine. Currently us Canadians have to pay $5 per 100 Mb when we go down south.

This Company?

[Gussington]

What, you couldn't include the name of the company in the headline?
Clickbait is more important to you than conveying actual information?The dumbification of Slashdot is getting really irritating...

Lane precision, Comma.ai

[DrYak]

But not more precise than Google Maps, which does exactly this.

Oh, do they ? In my (limited) experience I've only seen the "usual lanes plan when there's no construction work" that other satnavs provide too.

But maybe it's not available yet in all cities.

The truth is, even the gold standard for this sort of thing (Google Maps) doesn't have the amount of accuracy required. And can't. These things change too frequently to actually rely on them for that level of detail.

Yup, the guys at Google have often mentioned that indeed.
Hence the whole idea of using cameras and teaching the cars to understand their environment.

That would be an improvement, certainly. It's still hard to see how it would be effective enough in practice as a crowdsourced venture. How are they going to convince enough people to outfit their cars with dash cams?

Actually, comma.ai has exactly managed that with chffr : convice enough people to install it so they can have enough data to teach their system.

Their strategy revolve around :

• general cheerfulness of the crowd being happy to collaborate on an opensource project
• chffr doubling as an actual useful dashcam with full logs of everything, including all that it can gather over the canbus from the odb port
• gamifying the whole things with people boasting about their comma points on the forum
• a few perks (top contributor could get hardware more easily)

.

If the efforts don't fall into too many different forks / split groups / etc. and if comma.ai doesn't pull an "occulus rift" fuck (e.g.: by suddenly declaring that the trained data will only exclusively be available to car companies that pay a license and to no end-user directly *), comma.ai has a *very good chance* of becoming such a "OpenStreet Map for Autopilots and Self-Driving Car"-like company.

On their blog, they even have documented preliminary experiments with generating a 3D map of the street out of the sensors - using only video and GPS, no LIDAR (chffr works with most android smartphone, it doesn't require a Caterpillar with the built-in lidar).

---

* : I think I've heard somewhere that comma.ai got approached by Tesla exactly for that, and that comma.ai refused and wanted to explicitely stay car-manufacturer-neutral.