Where's the lane? Self-driving cars confused by shabby U.S. roadways | Hacker News Reader

Where's the lane? Self-driving cars confused by shabby U.S. roadways

reuters.com

58 pointseplanit80 comments

This is one of those cases where the engineers aren't going to be able to cheat and force the world to conform to their model of reality. I await the "falsehoods programmers believe about roads." If you actually want to build autonomous cars there are a whole host of assumptions that you have to throw out. To give only one example: are all roads paved?

It would also imply that anyone with access to road marking paint can fool autonomous vehicles into misbehavior.

Such tricks could fool a significant percentage of humans, too.

And road construction crews love to play those tricks. There's a spot on the FDR highway around NYC where for years the road was redirected to swerve, but the lane markings remained straight. If you were not paying full attention and just followed the lines, you'd drive straight into the wall.

That seems less like a fun prank and more like sociopathy. I'm really surprised nobody got sued.

(I'm not a lawyer, but "We created an optical illusion directing people to crash into a wall, so those victims should have been paying attention and figuring it out" doesn't strike me as a robust defense in court.)

That seems like a ripe... opportunity for lawsuit.

Hey, in Mexico, they dig holes in the road, and don't mark them ;)

I am still trying to understand the logic of cars that follow other cars, as if you can trust them. How much deviance from the lane can the car you are following take before the autopilot just says, hell no.

the state of the roads isn't what is stopping self driving cars. considering the amount of road construction going on at any one time no amount of markings are going to help, then toss in rain, bad lighting, and snow.

No, what is holding them back is the programming and processing power that cars are expected to mount. As humans can drive with on their sense of sight cars will have the advantage of sight that see in conditions people cannot plus also have maps available at all times which should give guidance on where the roads really are.

The roads are deteriorating foil is typical election year crying. the bridges and roads are much improved since the 90s and the key fact most writers and readers of these articles don't understand is, the federal government does not have domain over many of the roads that need maintenance.

I am not aware of any vehicle with autonomous capabilities which bases it's operation on "following other cars".

A number of cars including the Infiniti Q50 have a variation of the "follow the car in front of me", I used it the other day on the highway it is more about cruise, but combined with the stay in lane feature that it also has it might quality for what you are looking for.

The Tesla Model S, X and imminent 3 use this heuristic.

The autopilot will look for lane markings or road surface edges, but seems to favour following the vehicle in front where possible. This especially helps drive better in rainy weather where road markings are obscured by reflections.

Wow...that's like literally "If all your friends jumped off a cliff..." If you had a line of self driving cars, and one mistakenly heads for a ledge in a bad rain storm, there is the potential many of the following cards could do the same as well.

Not exactly. I've followed a car in front in fog where I couldn't really see where I was going and I figured if there was a wall or cliff or something I'd see the car in front deal do odd stuff and hit my brakes.

Like Wile E Coyote painting pictures of tunnels on walls?

Well, in New Zealand we have some beaches that are legally classed as roads so I guess the answer is no.

On the subject of transport uses for beaches - there is an airport here in Scotland that uses a beach as a runway:

https://en.wikipedia.org/wiki/Barra_Airport_%28Scotland%29

It does have scheduled flights - but the schedule has to take account of the tides.

There are also plenty of ice roads https://en.wikipedia.org/wiki/Ice_road

Ninety mile beach?

Then we have to wish automatic cars have a Jeremy Clarkson mode

That's the most well known one.

Point Road, Nelson was the one that really caught me out though. It looks significantly different between Satellite and Map modes: https://www.google.co.nz/maps/place/Point+Rd,+Monaco,+Nelson...

"Shabby roadways"? If there was ever a case of "first world problems", this is it.

Frankly, US roadways are really good. I was amazed by the amount of signage present, not to mention the pavement itself.

Yes there are imperfections. But when was the last time you tore apart your car's axle due to inability to avoid the 20th gaping hole you ran across in just one way of your commute?

You want shabby roads? Go to third world countries. That's the real test for self-driving cars.

It's pointless to compare the United States with third world countries. You should be holding it to much higher standards.

I myself compare the United States with our peers in Western Europe, Japan, Korea, etc. In this scope, our roadways are actually kind of sad, especially when you factor in how much we spend on them and double-especially when you factor in that we don't have solid railroads.

Our highways are generally paved really thin (compare to: German highways) and they end up cracking and breaking all the time. I've never driven between two cities on an Interstate highway without being slowed down by construction work.

Also, just from day to day observation: I live in Chicago and everywhere I walk, the streets look disastrous. They're broken apart, and some of them are half-repaired with concrete chunks sticking out (check out LaSalle and Randolph)--on some streets it's a miracle to me some cars get through without a flat tire.

We are a tremendously successful and advanced country. There is clearly a very big problem with the state of our roads.

> double-especially when you factor in that we don't have solid railroads

I think you mean 'passenger railroads'. The US freight railroad is world-class. Lots of spurious debate about railroads in the US comes from confusing the pros and cons of a freight network (which we have) with the pros and cons of a high-speed passenger network (which we don't have). The benefits of a continent-sized high-speed rail network in a continent with low population density tends to be over-emphasized.

http://business.time.com/2012/07/09/us-freight-railroads/

That is indeed what I meant.

As for benefits/needs of a high speed passenger network--all I have to say is that my current options for going Chicago - New York which I do frequently are: 1. garbage airlines that are consolidating more and more and increasing their prices for no apparent reason, 2. broken roads that are constantly being repaired, and 3. Amtrak

That said, it's clear to me that a high speed rail connection (for example Chicago - NY at ~200mph) would be a no brainer. I'd even pay more than I pay American Airlines for the route.

Between DC-NYC-Boston we have the Acela train here, which is surprisingly decent in terms of price and service. It gets up to 150mph but due to traffic and the layout of the tracks only averages 80mph or so. Still, it beats flying imho.

https://en.wikipedia.org/wiki/Acela_Express

We don't necessarily need continent-sized high-speed rail. Our population density isn't nearly as low as you think it is, or as the official numbers suggest. The problem with the official numbers is that population density is not uniform: something like 2/3 of America's population lives east of the Mississippi River. And much of that lives in the northeast corridor. When you add in all the vast, unoccupied areas in the western states like Wyoming, then the overall density seems low.

So, instead of worrying about continent-sized high-speed rail networks, maybe we should start by thinking about building a high-speed rail network that spans the Eastern Seaboard states, where so much of the population is packed, then then expand out from there (Chicago, Memphis, New Orleans, Houston, etc.). We've already tried this with Amtrak and their Acela Express, but it's kinda lame: expensive and not that fast.

If we just built a seriously fast high-speed train between Boston and DC (stopping at NYC, Philly, and Baltimore along the way), and priced it competitively with air travel, then we'd be getting somewhere. Extend it down to RTP, Atlanta, and Florida next, and then a western spur to Cleveland and Chicago.

A separate high-speed rail network along the west coast, linking San Diego, LA, SanFran, Portland, Seattle, and Vancouver would also make a lot of sense.

What probably wouldn't make so much sense is trying to link the east and west coasts with such a system. Rail (even HSR) is just too slow to compete with air travel for distances that far.

Of course, Musk's Hyperloop idea might make more sense than any of this, though it'd basically be serving the same purpose in mostly the same way, just with tubes instead of rails. But its speeds would be even higher than air travel, so it could really replace air travel for major corridors, leaving air travel for places where Hyperloop isn't built yet, or for any travel which crosses large bodies of water.

High-speed rail competes badly against air travel; what makes a huge difference is commuter rail.

> High-speed rail competes badly against air travel

For long distances, yes. For short to medium distances, high-speed rail can be much better. In the DC-Boston corridor, rail is already very competitive with air and that's with "high-speed" rail speeds averaging 80mph. True high-speed rail would be much better than air in that corridor.

> It's pointless to compare the United States with third world countries. You should be holding it to much higher standards.

Not in the context of self-driving cars. For that, US roads should be good enough. If they aren't why not just provide roads that are prepared with aids (electronic or otherwise) just so self-driving cars will work properly?

I fully agree that self-driving cars should be able to go on US roads, and of course they're "good enough." They should be designed with any type of road in mind, because you never know where you'll have to drive.

I was simply responding to someone defending US roads when there's a publicly visible crisis happening with our road (among other) infrastructure that no politician seems to want to address.

anybody can just youtube for "russian road accidents" and imagine a self-driving car in those situations. Wrt. responsibility/insurance - when it comes to post-accident "discussion" would self-driving car come with a walking robot who can use baseball bat and fence using metal wheel wrench? I'd venture to guess that deluxe version which would use handgun instead of the bat would find a lot of customers among Chechens living in Moscow.

Do the third-world countries you're thinking of enforce traffic laws? Because I expect that an autonomous car could safely navigate an unmarked road if it were just paying attention to its destination and avoiding obstacles, not obeying detailed regulations.

Expect that they don't. Or if they do, not consistently.

https://xkcd.com/1232/

You can replace "exploring other planets" with poor roadways, wealth inequality, or whatever else you would like.

You want shabby roads? Go to New York City.

Or up state New York. I was driving on one limited-access highway where there were so many potholes the speed limit was 35 M.P.H. Usually I would not adhere to such a ridiculous speed on a highway, but in this case it felt prudent to actually go slower than that.

I'm an upstate (real upstate, north of Syracuse) resident. People hear "New York" and think "ooh, city!" but there are swathes of very rural, very poor areas that can barely keep up with basic maintenance. Case in point, I live in the "most fiscally stressed municipality in the state"[0]. Add to that our harsh winters and you end up with roads that are more gravel than pavement after a few years of erosion. I can't see self-driving cars ever working here as they're currently designed.

[0] http://www.northcountrypublicradio.org/news/story/31231/2016...

A self-driving car in Bombay traffic would be a monumental achievement!

"Where's the lane?"

I've often asked that myself dear robot. Welcome, you're now truly a fellow driver

Non-self-driving cars are often similarly confused. While there are situations a human can handle that a self-driving car can't as easily, many situations confound both, and necessitate cautious driving and a careful lookout in case you're not actually in your lane (or another vehicle isn't in theirs).

What?? You want to sell a self-driving car but you want it to be a good driver only in pristine, well-marked roads?? What kind of half-assed solution is that? Well, now that I think about it, the car could determine if the road wasn't good enough and tell you to take over, but then you'd have to be ready to do it and not busy reading a novel or talking with your mother on the phone, so it would defeat the purpose of doing the driving a bit. Anyway, I'll bet the only real solution will be to make the ai better. Just as with airplanes, you can't hal-ass something like this and blame it on lack of road paint, it would be too dangerous (you know, with airplanes, how they're so safe because otherwise they'd be so dangerous).

Driving in LA I often have trouble finding the lane. Either there are no markings or there are several markings of different fadedness. In rain it's even worse. You just take your best guess :-)

A fun one is when the road itself has grooves that the paint mostly follows. I naturally follow the grooves until I realize that the lane just drifted sideways on me.

There is a spot on the 405 S around Culver City that this used to happen regularly to me at.

Where in LA are you driving? It definitely isn't that bad to be "often"...

405 north in the rain is fun. Even better in the dark :-). Or just 5 north all through the LA area.

Most of LA in the rainy dark is absolutely terrible. There are no raised markers or reflectors for lane dividers, so combined with the often faded lines and reflections of headlights, it's impossible to tell where the lanes are. Most drivers (myself included) just kind of slow down and say "this seems like about the right place...".

I personally wouldn't want to be in a car that didn't work if the lane lines were faded. Apparently all terrorists will need in the future to cause mass mayhem is a can of paint.

Why wait for them to fade when you can cause much more directed mayhem by painting new lines that lead the vehicle somewhere unsafe.

How can you head a multinational automobile engineering company and not be aware of this most basic shortcoming of your product?

Who says he wasn't aware? Most likely, he wasn't aware that the specific place where they were planning the demo didn't have markings.

Uhh...many HUMAN drivers are confused by U.S. roadways...I've been in many places where the lane is more about following the car in front of me as there are no well visible markers. Kind of like when it snows, the lane is the two tire tracks, who cares where the actual lane is haha.

Here's a tweet I sent to the City of San Carlos about 4 weeks ago. Fortunately they've since remedied it. This is Holly (Redwood Shores Parkway) looking west towards an overpass over 101.

If it's not clear, 2 lanes come over the overpass towards where I shot the photo from. An offramp from 101 North terminates where the white pickup is.

This has always been a terrible merge. 2 lanes come across the freeway, and the offramp is a new 'added' lane, but the striping/bott's dotts between the #1 and #2 lane coming down from the overpass are virtually nonexistent. There are cracks in the road (the dark marks) where the lane markings should be.

It was always a big risk because cars in the right lane coming over the overpass used to just drift over into the offramp's new lane, and half the cars coming down the offramp didn't know they had a new lane, so they'd just STOP at the end of the ramp.

THEN the City, in its infinite wisdom, decided to fix none of these problems, but instead add a striped bike lane that cuts across the offramp. Absent lane markings to follow, drivers in the right lane would just follow the dashed bike lane and force themselves into the offramp's new lane. I witnessed at least 5 near collisions in the several minutes I was there taking photos.

They've since SOMEWHAT improved the markings between #1 and #2 lanes, but the bike lane markings are much more substantial and people still tend to follow them. Also, it still tends to confuse people using the offramp.

But this is an endemic problem; I bike commute and often a bike lane switches sides of a car lane in the middle of an intersection. It's the traffic version of "Step 1: right lane. Step 2: ??? Step 3: Profit!" Or "good luck, hope you come out on the other side of the intersection alive.

https://twitter.com/reidconti/status/705866795417075712

To be clear - this is a bike lane on a road with onramps? Uh, what? There aren't even bollards (in the part without an onramp, obviously) or any semblance of protection.

It's the surface street side of the ramp there in the photo. Here's the approx location:

http://www.openstreetmap.org/?mlat=37.51639&mlon=-122.25424#...

The photo is taken looking towards the freeway, showing the offramp.

A typical bike lane in the US is made out of paint, the lack of protection isn't noteworthy.

The road is a surface street -- a 30 or 35mph divided arterial which goes over a freeway. The ramp is an offramp from a freeway. But yes, there is no protection, which is normal.

at a certain point, fast approaching, we should start optimizing road markings for electronic sensors over human sight. How costly would it be to strip the middle of every lane with a small magnet or RFID that could detected by vehicles and even convey static metadata, such as mile maker, speed limit, number of lanes, distance to next stop light etc.

I would think radio towers with GPS-like positioning information would be cheaper and better, particularly since they could include other information like road conditions.

If they are confused by US roadways, they'll melt when they come to India and see this - http://www.thehindu.com/multimedia/dynamic/00397/20080331593...

Survival of the fittest.

I saw this in Honolulu last week: http://i.imgur.com/WIdrgiG.png

I assume the extra "markings" are related to a road repair, but I couldn't imagine a worse situation for a self-driving car.

That looks like it's maybe this: https://en.m.wikipedia.org/wiki/Dowel_bar_retrofit

Thanks! Even includes a section on confusion with pavement markings.

Ha! I was waiting for this! I was on a multi-lane road recently and I couldn't see the lane delimiters. They had faded to the point they were no longer visible. It required a lot of concentration to avoid hitting cars near me and it was quite disconcerting.

The solution to this problem, at least for the foreseeable future, is to have the autopilot only active when the car is on a certified autonomous road. This would have to be a legal definition, in which road construction would de-certify that road section, and the certification continuously updated via the autonomous cars themselves. As soon as cars start seeing degraded conditions they update a central database, which alerts other cars to switch to manual mode well before entering that road section.

O.k. so lets say the roads got nice stripes. What about:

1) Under rainy conditions?

2) Heavy snow fall?

3) Incompletely plowed roads?

4) Roads under construction - i.e. partially striped and the little flags that haven't been completely moved around between road realignments?

Do you mean to tell me Alpha Go can beat a human at Go but can't be trusted not to drive a car into a ditch?

It is nice to know that the way we avoid the singularity is with confused road markings :-)

Seriously? Just use all the sensors you've got to determine the edges of the road, divide the width of the road in half, and the right-half is the direction you're traveling in. Based on the size of this half, determine number of lanes. Create a virtual box so the car stays in the space of where the right-most lane should be. No need to read road markings to stay in lane.

edit Nevermind, I forgot that not all roads have an even number of lanes in both directions (even considering single turning lanes and medians). But absent a median or any markings at all, the above is a relatively safe bet. I'm sure they could build a set of heuristics to accommodate the exceptions.

Humans have spatial vision, so we should also be able to create cameras that can make out objects floating in free space (such as traffic lights), piece out their components, and compare them to a set of common criteria to determine its intent, the way humans basically do.

Companies using inadequate software and sensors blame roads for engineering failure.

I'm not sure that's the case. In the greater Seattle area, roads have deteriorated to the point where it can be difficult to spot road markings even as an experience driver, and it's not unheard of for the markings to be completely worn away. I'm not sure about the rest of the country, but roads are painfully underfunded here.

I always run into this issue in downtown Seattle. There are a few places (and one major one I can think of on 5th) in which the roads do this:

     |0|1|2|3|
    _| | | | |
       / / / /
    _ / / / /
     |1|2|3|
     |1|2|3|
     |1|2|3|
     |1|2|3|

Unfortunately, the diagonals are marked incredibly poorly, so most drivers starting in lane N will go straight and end in lane N-1. So if you drive according to the law in lane 1, you would almost certainly get sideswiped eventually by someone going straight in lane 2.

I suspect an AI driver would be astute enough to see the poorly marked lines, but most humans aren't! So the AI would frequently get sideswiped here.

> roads have deteriorated to the point where it can be difficult to spot road markings even as an experience driver

Yes, but apparently self-driving systems are doing a worse job than humans. But this only means the AI isn't quite there yet. I say AI, not sensors, because their vision is probably already better than human's.

It's a combination of sensors and AI. You'd be surprised by how low quality (wrt framerate & resolution specifically) the sensors used in production autonomous features/vehicles are. See mobile-eye's papers and products: http://www.mobileye.com/wp-content/uploads/2011/09/StereoAss...

The newest reference date in that paper is 2009 so I suspect the paper itself is a bit old. I really doubt that video would be other than state-of-the-art when the whole system is so expensive.

Edit: Also that company makes add-on collision detection products. That is a different application than auto-driving.

I worked on these cars this summer at GM. I can't say what model of camera sensors we used for obvious reasons, but the resolution and framerate were not too far off from what Mobileye has in that paper!

so that's even in a fairly rich "developed" country...hello the rest of the worlds roads aren't generally any better...

this is exactly the moment when I ponder "why the hell is this whole automatic-self-driving-car thing being pushed so hard generally?"

it seems like a rather absurd and silly effort to begin with... what's wrong with a bloody system of maglev propelled capsules being shot through tubes, or the train, or a normal car you drive with a steering wheel, accellerator pedal, and brakes?

This is why we need self-driving road paving trucks first. Kinda serious, especially for filling potholes.

I want to know how they plan for self-driving cars to work in snow.

Tesla's system uses the information about the road gathered when the snow was not present. Teslas are very popular in Norway.

Notably absent from cast of complainers: Google.

At the same time I'm pretty sure they also commented that self-driving cars might be 30 years away during an event at SXSW this year.

I really liked the Rolls Royce CEO's comment on their self-driving tech: "Oh, we invented that 111 years ago. We call it a chauffeur."

http://www.bloomberg.com/news/videos/2015-01-06/rolls-royce-...

The talk by Chris Urmson is often misconstrued like this. The 30 years statement is for when self-driving cars will be able to completely replace human drivers in basically all situations and all locations. This is very different than how long it will take for self-driving cars to replace human drivers in some situations and some locations.

Most people in the industry are still expecting self-driving cars to be able to operate on normal roads in CA and some other favorable locations within 5ish years, which makes sense given the current state of the tech. In 6 years Google has gone from no self-driving car at all to many self-driving cars logging over 1.3 million miles with a single minor fenderbender of questionable fault. Given another 5 years to develop the tech, train the software, and expand the corner cases, there's no reason to think you won't be able to hail a robocab in San Jose in 2021.

Reply on news.ycombinator.com