Will the Google Car Turn Out To Be the Apple Newton of Automobiles?

An anonymous reader writes The better question may be whether it will ever be ready for the road at all? The car has fewer capabilities than most people seem to be aware of. The notion that it will be widely available any time soon is a stretch. From the article: "Noting that the Google car might not be able to handle an unmapped traffic light might sound like a cynical game of 'gotcha.' But MIT roboticist John Leonard says it goes to the heart of why the Google car project is so daunting. 'While the probability of a single driver encountering a newly installed traffic light is very low, the probability of at least one driver encountering one on a given day is very high,' Leonard says. The list of these 'rare' events is practically endless, said Leonard, who does not expect a full self-driving car in his lifetime (he’s 49)."

Comments

my motoo is, just let go of the wheel

[Anonymous Coward]

if you just let go of the wheel and trust the car to drive itself, then that's what matters. hasn't failed me yet!

Re:my motoo is, just let go of the wheel

[kelemvor4]

My motto is "Find a fucking dictionary"

You're welcome

Re:my motoo is, just let go of the wheel

[Anonymous Coward]

Let's be realistic, he won't need that dictionary, this is slashdot after all. Perhaps if it was a dicktionary.

How hard is it to recognize a stoplight?

[Spy Handler]

didnt RTFA but seriously? Google car can't recognize a red light??

I would've thought some of the better Slashdotters could write software that recognizes a traffic light from a camera feed, let alone the geniuses at Google.

Re:How hard is it to recognize a stoplight?

[Anonymous Coward]

I'd imagine it's pretty damn hard. Harder still is figuring out whether that stop light applies to the lane you're in, and the direction you're turning. Stop lights don't always line up with lanes exactly, they don't always point straight, etc etc.

Re:How hard is it to recognize a stoplight?

[Anonymous Coward]

I've seen quite a few times when even drivers can't figure it out. What we need is better geo-mapping from cities themselves. And eventually cars may be compatible only for certain cities/states. (And maybe go into a "ok, now we need a human!" mode if you cross state lines.)

Re:How hard is it to recognize a stoplight?

[ShanghaiBill]

didnt RTFA but seriously? Google car can't recognize a red light??

Yes, the Google car can recognize a traffic light. TFA is written by a confused journalist. He found out that Google maps out roads, keeping a database of signs, lane markings, etc. He then concluded that the Google car only works on pre-mapped roads. That is not true. If the car is driving on a pre-mapped road, it will use the info from the database. But it can still drive on other roads with good accuracy. There are still come problems to be worked out, and plenty of testing to be done, before SDCs are ready for sale to the public. But TFA is a very inaccurate description of those problems.

Re:How hard is it to recognize a stoplight?

[Anonymous Coward]

Must recognize traffic light in all lighting and weather conditions, then find lines for stopping, determine direction signals if present, EG "can turn left on 'normal' green, or is there a green left signal that must go?"

It's a much more complex problem than just "is traffic light yes/no?" I appreciate Google et. al. doing this research. All these articles whining about how hard fully automated cars will be often feel like attention seekers. Yes it will take a while, that's why it's good that someone is putting a lot of money into the research it will take yeah? "It's too hard so let's never do it." Is the attitude that's caused us to still use coal and gas instead of fusion reactors.

So it's difficult, but less media whoring of "it's too hard" and more "it's awesome so let's get on with it even if it is hard" please

Re:How hard is it to recognize a stoplight?

[Hadlock]

I think the problem is that "good accuracy" is not yet to the point where the driverless car is less likely to run over a pedestrian at an intersection than a car piloted by a human. Until that threshold is met it's just a science experiment.

Re:How hard is it to recognize a stoplight?

[Anonymous Coward]

Military driverless vehicles handle that just fun.

Re:How hard is it to recognize a stoplight?

[ShanghaiBill]

I think the problem is that "good accuracy" is not yet to the point where the driverless car is less likely to run over a pedestrian at an intersection than a car piloted by a human.

I very much disagree with this assessment. Google's SDC has been tested thousands of times with a huge range of pedestrian scenarios. It may not be better than an alert and primed human, but it is almost certainly better than an average human, which is the important criteria. If I was walking across an intersection, I would trust a Google SDC far more than someone late for an appointment, driving a Chevy Tahoe with a cellphone in one hand, a Starbucks latte in the other, and two screaming kids in the back seat.

Re:How hard is it to recognize a stoplight?

[Anonymous Coward]

The issue, as I understand it, with the SDC is road conditions, not obstacles of decent size (humans, bikes, other cars, etc). Potholes or things like ice or a broken piece of wood with nails sticking out of it.

Honestly, I'd prefer if they just implemented some sort of token under each lane at a traffic light. They already put down wires for metal detection around here, how much more would something like a RFID chip cost?

Re:How hard is it to recognize a stoplight?

[Ichijo]

I'm interested in whether the car can set an appropriate speed when it comes upon a sign saying "Speed Limit 35 When Children Are Present" or "...When School In Session", and whether the car can read and obey hand signals, and whether the car knows right turns are prohibited at a particular intersection during rush hour.

Re: How hard is it to recognize a stoplight?

[Anonymous Coward]

Oooh, it would be so fun to clone those RFID chips and put them in incorrect locations then watch the cars freak out...

Re:How hard is it to recognize a stoplight?

[phantomfive]

I can write software to recognize a red light, but in this case it needs to have very high accuracy. A 90% accuracy rate isn't going to be good enough, and I'm not sure I could even do that well.

The important thing to notice is how carefully Google has controlled their communication about what their car can do. They release heartwarming stories about driving a blind person, and we hear about it when a human crashes the car, but what sorts of algorithms is it using? What kinds of situations can it handle? We don't even know. A clear example from the article of careful messaging:

"The company frequently says that its car has driven more than 700,000 miles safely, but those are the same few thousand mapped miles, driven over and over again."

And that is even before getting to the engineering problem of reliable software. When Boeing built their recent airline software package, it took 5-8 years to get something that was reliable enough for air travel. And air travel is a much simpler problem than driving a car.

Re:How hard is it to recognize a stoplight?

[ShanghaiBill]

I'm interested in whether the car can set an appropriate speed when it comes upon a sign saying "Speed Limit 35 When Children Are Present" or "...When School In Session"

It can use OCR to read signs. It also can make a decision about whether school is in session. Since it cannot reliably detect when children are present, it would most likely just default to the lower speed.

whether the car can read and obey hand signals

Hand signals are a problem. Google is working on it. But SDCs aren't going to just pop up on the road one day. Their release will be coordinated with the police. So one solution is for the police to use LED flashlight wands that the SDCs are programmed to recognize.

whether the car knows right turns are prohibited at a particular intersection during rush hour.

Sure. An SDC would know the same why a human would. By reading the sign, or by "remembering" by accessing the database on that intersection.

Re:How hard is it to recognize a stoplight?

[ls671]

> What we need is better geo-mapping from cities themselves

For driver less cars to work, the whole city should be wired so the google car doesn't have to recognize the red light, it would just get the information through some type of wireless transmission thus knowing it has to stop. It may fall back on A.I. on country roads with not much traffic but then again.

A.I. is not advanced as we sometimes think it is. Even good drivers can have trouble recognizing a red light in bright sunlight conditions.

Re: How hard is it to recognize a stoplight?

[ShanghaiBill]

Oooh, it would be so fun to clone those RFID chips and put them in incorrect locations then watch the cars freak out...

The SDC would notice the conflict between the RFID and the sensor data, and pull over and stop, while alerting the human driver/passenger.

Someone with sufficient education and expertise to clone an RFID would probably have better things to do than pull a stupid juvenile prank that could send them to prison for a long time. There are numerous ways that people can sabotage infrastructure and kill people. We can't prevent that, but it is rarely a problem, because most people are not motivated to murder random strangers. You can make human drivers "freak out" if you shoot at them with a sniper rifle. How often does that happen?

Re:How hard is it to recognize a stoplight?

[cheater512]

Doing it is easy. Doing it with a very very high degree of confidence? Hard.

You just have to miss that one stop light for something bad to happen.

Re:How hard is it to recognize a stoplight?

[cheater512]

If the cars that fall back to AI then communicate their observations and decisions back to Google then to other cars then the next car wouldn't need AI and could improve knowledge of the area, plus any particularly bad problem spots can be highlighted for further investigation at Google HQ.

Normal drivers don't have LIDAR. I assume it is a massive assistance for some aspects of Google's work.

Re:How hard is it to recognize a stoplight?

[McGruber]

I imagine that the hard part would be recognizing every traffic signal when the signals are dark (power is out).

Re:How hard is it to recognize a stoplight?

[jklovanc]

Sorry but this article has different information.

The key advantage is that the car isn’t just seeing and figuring out the world as it drives along. It’s basing its actions on vast amounts of data the Google Self-Driving Car Project has already compiled about every road it travels. Before the car drives itself into new territory, the project team collects detailed information on permanent features: lane markers, the precise location of the curbs, the height of traffic lights, local speed limits, and so forth.

“We require digital maps in order for our cars to be able to drive,” Andrew Chatham, who leads mapping on the project, said at a press event Tuesday. That data “makes the job of building the self-driving car software much simpler.”

The car has a good idea of what to expect from any stretch of road, freeing up the software to deal with cars, pedestrians, cyclists, construction, and any other new obstacles in real time.

That’s the “magic of maps,” Software Lead Dmitri Dolgov said. But that “magic” inherently limits the range of the self-driving car to areas Google has the data for. As Chatham pointed out, “If we have not already built our own maps in an area, the car cannot drive there.” He noted that as the car’s sensors get better, they will rely less on perfect accuracy, but Chris Urmson, the project director, emphasized the key role these maps play.

Regular Google maps do not have enough accuracy.

Re:How hard is it to recognize a stoplight?

[TheGavster]

I think the real goal would be to have all vehicles self-drive; then they can be coordinated to interlace at intersections, removing the need for stop lights and saving a ton of fuel!

Re:How hard is it to recognize a stoplight?

[NicBenjamin]

That should be easier then identifying a stoplight, because the signs are very standard, are in easily predictable places, and don't look like anything else. You simply program the car to recognize that a sign with this format means this speed limit under these circumstances.

Stoplights can be multiple colors (most are yellow, but some are black), look like quite a few things you drive by every day (ie: a police car with lights on, or a street sign warning there's a light ahead), may be strung across the road or on a poll above the road, etc.

Re:How hard is it to recognize a stoplight?

[mjwx]

If I was walking across an intersection, I would trust a Google SDC far more than someone late for an appointment, driving a Chevy Tahoe with a cellphone in one hand, a Starbucks latte in the other, and two screaming kids in the back seat.

The problem you have is, someone like that wont let the car drive itself because a self driving car will stick to speed limits and slow down at pedestrian crossings because it will be programmed to anticipate stopping at a pedestrian crossing (like a defensive driver is trained to do). Nope, someone that self adsorbed and with such poor time management skills will be taking manual control with the pedal pressed to the floor whilst screaming into their phone. You simply cant overcome selfishness with a new technology.

But actual autonomous cars are years away from practical use. Decades away from the way you're thinking. The first autonomous cars will be normal cars with an autonomous mode that only works on specially upgraded roads which you can guarantee will be limited access roads (freeways and motorways) with no traffic lights.

Re: How hard is it to recognize a stoplight?

[Anonymous Coward]

Yes, your argument certainly convinced me that no one would ever vandalize street signs or reprogram construction zone flashing message boards to read "WARNING ZOMBIES AHEAD".

Re:How hard is it to recognize a stoplight?

[Vellmont]

  If I was walking across an intersection, I would trust a Google SDC far more than someone late for an appointment, driving a Chevy Tahoe with a cellphone in one hand, a Starbucks latte in the other, and two screaming kids in the back seat.

If you think that's supposed to instill confidence, you might want to re-think that. Your're compairing a computer to a severely distrtacted human. A human, I might add that's breaking the law. Distracted driving is illegal.

You need to compare the SDC to a fully aware human being, not a fully distracted one. You sound like someone that might have inside knowledge. So listen carefully. EVERYONE thinks they're an above average drive that's fully aware. THAT'S your standard, not a distracted latte sipping soccer mom with kids yelling in the back seat. If this think is ever going to succeed it has to be better than an actual good driver, since everyone thinks they're that.

Re:How hard is it to recognize a stoplight?

[phantomfive]

Yes, the Google car can recognize a traffic light. TFA is written by a confused journalist.

If you're going to quibble about credentials, you ought to at least read the summary, where the MIT roboticist says he doesn't expect to see fully self-driving cars in his lifetime. Do you think he is confused too?

Re:How hard is it to recognize a stoplight?

[Charliemopps]

Describe for me, programmatically, the difference between a stoplight and a taillight.
and a police light
and a neon sign
and every other red light on earth...

and also, please include all the many shapes and sizes of the various stoplights all over the country.

Stop signs have a very specific shape, and text printed on them. They do not very from place to place. They're piratically a damned bar code as far software is concerned. It's almost like they were designed for the task.

Re:How hard is it to recognize a stoplight?

[phantomfive]

Google's SDC has been tested thousands of times with a huge range of pedestrian scenarios. It may not be better than an alert and primed human, but it is almost certainly better than an average human,

I'd really be interested if you have a reference for this. Even if your reference is just a Google PR person, that's still better than nothing.

Re:How hard is it to recognize a stoplight?

[phantomfive]

Incidentally, the author may be confused, but Google themselves have confirmed that the car has problems in a number of scenarios, including:

The Google car doesn’t know much about parking: It can’t currently find a space in a supermarket lot or multilevel garage.
It can't consistently handle coned-off road construction sites, and
its video cameras can sometimes be blinded by the sun when trying to detect the color of a traffic signal.
it can't tell the difference between a big rock and a crumbled-up piece of newspaper

Re: How hard is it to recognize a stoplight?

[ShanghaiBill]

Yes, your argument certainly convinced me that no one would ever vandalize street signs

The question isn't whether someone would "ever" do it, but whether people would do it often enough for it to matter. Snipers occasionally shoot into traffic. But we don't consider that when we design roads and cars. Nonetheless, the algorithm to handle a sabotaged location sensor (RFID, Magenetic, or whatever) is brain dead obvious: pull over and stop the car. Stolen stop signs cause a few deaths every year. Those deaths would be eliminated with SDCs, because they could access their database and know the stop sign was supposed to be there. SDCs would likely be more resistant than human drivers to most forms of sabotage.

Re:How hard is it to recognize a stoplight?

[JasonGoatcher]

Yeah, you don't want the military vehicle running over people, then the soldiers wouldn't have the fun of shooting them.

Re:How hard is it to recognize a stoplight?

[gumbi west]

I walk my children to school past a "When Children Are Present" sing. I've never seen a single driver obey the lower speed limit. So if that's appropriate action, then yes, I'll bet it can do that. I'd take just obeying the posted normal limit.

Re:How hard is it to recognize a stoplight?

[ShanghaiBill]

the MIT roboticist says he doesn't expect to see fully self-driving cars in his lifetime. Do you think he is confused too?

If he actually believes that, then yes, I think he is confused. He is 49. If even 10% of life extension research pans out, he should easily live to 90. That is 41 years. Computers will be thousands, if not millions, of times more capable. Even if we don't yet have general AI by then, huge progress will be made in algorithms for computer vision, pervasive sensors, etc. Even ten years ago, the car that won the DARPA Grand Challenge was able to navigate dirt roads through the desert. We can do far better than that today. Even if every road has to be pre-mapped, so what? Mapping a road is 0.1% the cost of building the road. So the lack of maps is not going to cause SDCs to not happen. Maybe we won't yet have SDCs in five years, but unlikely that we won't have them in ten. To think that we won't have them in 40 years is just silly.

Re:How hard is it to recognize a stoplight?

[roger10-4]

The field of computer vision (which this would fall into) is extraordinarily challenging. Just creating an algorithm that can simply *recognize* an object is hard let alone the ability to take that information and make a decision based on it. As a bit of a contrived example, imagine a person was holding a sign on a street corner that looked like a stop light (same colors, shape, size, etc.). As humans, we know it's not a stop light because we know people generally don't hold stop lights, we know a stop light is 3-dimensional, the lights glow, etc. Getting a computer to understand this type of contextual information is not easy.

Re:How hard is it to recognize a stoplight?

[Anonymous Coward]

Care to test that notion yourself? I can give someone a Google a call to make it happen.

Attitudes and answers change once proof is required. Folks have done it in the past (think chemists, so succeeded some failed with death).

And honestly, that's why science is awesome.

Re:How hard is it to recognize a stoplight?

[phantomfive]

Computers will be thousands, if not millions, of times more capable.

Are you guessing this based on Moore's law?

Re:How hard is it to recognize a stoplight?

[edremy]

No, you compare to an average of the driving populace, not the best or worst. Why? Because that's what the insurance companies are going to do. When it becomes obvious that SDCs are better drivers than humans, you're going to start seeing a serious push to let the bots take over.

I don't think it's going to take anywhere near as long as people think. There's a *huge* market for this. My grandmother in law is 93. She basically can't drive, but wants to stay in her house. My wife's in the hospital right now and I have two kids that need to be different places at the same time. One of my old teachers is blind, etc, etc SDC taxi pools can act like super-flexible mass transit for areas that don't have any

Re:How hard is it to recognize a stoplight?

[NicBenjamin]

Air travel is very simple in terms of dealing with traffic. The sky is big, so you simply point the direction you want to go and you go.

But there's a whole host of things an aerial autopilot needs to be able to do that GoogleDrive doesn't.

For example, for a plane to fly it has to have an Airspeed above it's stall-speed. Airspeed refers to the speed of the air going over the wings. If your aircraft is x lbs, you need x lbs of air under your wings at all times or you lose lift and instead of going up you're going down. The way you get more air under the wings is go through the air faster. So your computer needs to know how fast you are moving relative to the wind. This is related to the speed the aircraft is going relative to the ground, because if you add 150 MPH to airspeed you've generally added 150 MPH to groundspeed as well, but that really depends on things like the exact geometry of your vector related to the vector of the prevailing winds.

Then there's fuel management. Commercial aircraft generally have fuel tanks in their wings because if you put the tanks in the fuselage there's less room for passengers. They also have fuel tanks bigger them most cars. Seriously. The Dreamliner package your talking about is available on an aircraft with 33k gallon fuel tanks in the small fuel tank model. That's more then 200k pounds of fuel. And if you burn all the fuel in the left fuel tank before you burn any in the right fuel tank your carefully calculated center of gravity moves to some spot on your right wing, which makes things very difficult. In routine flights this isn't an issue, because there's an engine on each wing, but if a goose runs into the right engine the computer has to know to start transferring fuel from the right tank to the left with a crossfeed pump or things get to be problematic.

Ice is another issue. Ice on a wing changes the wings shape, which means the physics of how the wing interacts with the air changes, which in turn means how the pilot uses the wing to stay in the air has to change. The ice also adds weight, which changes both fuel consumption and flight characteristics. Ice is one of the leading causes of passenger plane crashes in the US. And it's something an autopilot has to deal with.

There's a reason it's much harder to become a licensed pilot then a licensed driver.

Re: How hard is it to recognize a stoplight?

[Anonymous Coward]

And MS Office will have thousands more features, and will still be just as slow as it's always been.

Re:How hard is it to recognize a stoplight?

[phantomfive]

Is your point that building a self-driving car is easier than building autopilot on an airplane?

Re: How hard is it to recognize a stoplight?

[Anonymous Coward]

would people do it often enough. Yes. Yes they would. Around my place it happens about once a week that somebody reprograms a construction sign. And you seem to not understand how teenagers work. When they steal a stop sign, are they thinking they're hoping to kill somebody? No, no they are not. They think it's funny. They don't stop and think what the consequences of their actions are. You're problem is, you're applying logic to a fundamentally illogical situation. People and logic are separate entities most of the time.

And I'm glad your algorithm to handle sabotaged location sensors is so dead brain simple, but what about the algorithm to detect them. Remember, you have to assume somebody is trying to do it and not have it detected. That's where the lols are. Or are you proposing that a sensor bank is kept in the car? How is it updated? Kept in a cloud based server? What if there's a DDOS or wireless connection is currently unavailable? What happens if somebody forgets to enter a node into the database, remember, road workers probably aren't the most intelligent folks, and the bureaucrats will view entering that info as the road crews job.

You've oversimplified a problem in the extreme. This is a stupidly difficult problem because the world is big, complex and extremely dynamic. Seriously, if the cars simply pull over if it detects sabotage, what's the average failure rate of these chips? even a 99.99% non-failure rate means there's going to be a hell of a lot of cars stopped on the side of the road for no reason, angering everybody who's now late for work.

Re:How hard is it to recognize a stoplight?

[Joe_Dragon]

There are some lights where you have a red ball and green arrow (not just for left turns) some areas have diagonal arrow the point the way for the main flow of the road and I have seen people confused by that.

Re:How hard is it to recognize a stoplight?

[Joe_Dragon]

And then we will need to build pedestrian overpass / tunnels.

May as well build tunnels so the people put out of work by this have a place to sleep that is more open then the jail.

Re:How hard is it to recognize a stoplight?

[jcfandino]

I think the real goal would be to have all vehicles self-drive; then they can be coordinated to interlace at intersections, removing the need for stop lights and saving a ton of fuel!

However there are people that don't have cars and need to walk. Will the car overlords consider allowing us mere pedestrians to cross the street?

Re:How hard is it to recognize a stoplight?

[Anonymous Coward]

I'm not the parent, but I'd love to test that notion myself. Make that call to Google and get an internet stranger access to a driverless car.

Re:How hard is it to recognize a stoplight?

[Anonymous Coward]

Thousands of times with a huge range of scenarios...

OK dude.. most drivers are tested tens of thousands of times in their lifetime and don't run over a single pedestrian... so I don't think you can really make a claim that it's better than the typical driver based on this.

Re:How hard is it to recognize a stoplight?

[Joe_Dragon]

on local highways / limited access roads next to no one does the speed limits and the cops let you get a way with it as well. Even on the local roads you can do at least 5 MPH over

Re:How hard is it to recognize a stoplight?

[Anonymous Coward]

I think you may be confused as to what autopilot does. It follows the flight cue. That's it. It's up to whoever created the flight path to ensure that the speeds are adequate. If a goose runs into the right engine and takes it out, you can be damn sure the pilot is turning off autopilot. Sure, fuel has to be diverted. This is not a function of autopilot and in fact must be performed manually even if autopilot was never engaged. Ice is also not compensated for in the autopilot. The aircraft weight in reality will change as ice melts and forms, but the aircraft does not dynamically change the weight it is aware of. Instead, the plane notices that it it has gone off cue. It will then do whatever is needed to follow the cue (see above). The operator can override the gross weight, but it is not done manually. While it is true that ice is a leading cause of accidents, it has little to do with the change in fuel consumption and flight characteristics and almost everything to do with the pitot tube no longer working. Change all your instances of "autopilot" to "pilot" and you're almost on to something. Autopilot itself is relatively simple (much more simple than an autonomous car in its entirety would be).

Re:How hard is it to recognize a stoplight?

[Zynder]

Put out of work by what? The people building the cars? The people repairing the cars? The people installing and maintaining the off-vehicle sensor packages? The people building all the bridges and tunnels? It seems to me that this transition would INCREASE available jobs.

Re:How hard is it to recognize a stoplight?

[Zynder]

You need to compare the SDC to a fully aware human being, not a fully distracted one.

WHAT? What ever happened to "Plan for the worst and hope for the best"? That's how all the engineers around here operate. You said everyone thinks they're above average but are apparently lying to themselves so if the majority is indeed a shitty driver, why would you pretend they're anything else? And also why does someone who is a shitty driver but believes and acts as if he's a F1 driver be a worse problem than a soccer mom absolutely off in lala land? I wouldn't say either one is worse off. They're both equally dangerous and are the number one human factor to attempt to engineer out. A true Mario Andretti isn't the guy you need to worry about, it's the wanna be's and the don't cares that are going to kill you.

Re:How hard is it to recognize a stoplight?

[dgatwood]

Really, the problem is that "when children are present" is kind of ambiguous. What if there's only one child? And is the concern really all children, or just unaccompanied children? Are high school students children? Do kids in strollers count? And so on.

Most drivers would assume that the intended purpose is to increase safety around the time when kids are arriving at school or leaving school en masse. So they would interpret it to mean "Speed Limit [X] on Monday through Friday, from 7:15–8:00 and from 2:30–3:15". If the signs just said that instead of "when children are present", then automated cars could easily do the right thing every time. Also, by being more concrete, the signs would eliminate the selective blindness that causes many human drivers to ignore the lower speed limit.

Re:How hard is it to recognize a stoplight?

[Zynder]

I am fairly certain my Garmin knows where the school zones are, though it's been awhile since I've turned it on. Google ought to know too. If that "when children are present" part bothers you and/or the google car, wouldn't it be easier to remove that part and just say 35mph between 0730-0830 & 1430-1530? The cars know what time it is and where they are.

Re:How hard is it to recognize a stoplight?

[Zynder]

[Citation Needed]

Re:How hard is it to recognize a stoplight?

[dgatwood]

its video cameras can sometimes be blinded by the sun when trying to detect the color of a traffic signal.

So can people. One possible solution would be radio signals in every traffic light to indicate the light's state. No signal and can't see the light? Stop the car and tell the driver to take over. This would be useful for eliminating confusion when you have multiple lights as well, so it might be worth pursuing.

That said, the simpler fix is to use a higher quality camera with better lens coatings. I can't remember the last time I saw lens flare that blew out a picture to the point that it was truly unusable except when using old camera gear with uncoated lenses. For additional robustness, put more than one camera on the front, pointed in different directions. That way, lens flare should never be a problem, in practice. (Lens flare tends to be angle-specific, and the sun is in one spot, so if a lens at one angle is in a position to flare badly, a second lens at a different angle probably won't be, assuming your lenses aren't old, uncoated nightmares.)

it can't tell the difference between a big rock and a crumbled-up piece of newspaper

Neither can people, reliably, unless it is blowing. Whatever you see in the road, it is best to avoid it. :-)

Re:How hard is it to recognize a stoplight?

[Anonymous Coward]

but if a goose runs into the right engine the computer has to know to start transferring fuel from the right tank to the left with a crossfeed pump or things get to be problematic

Please talk about things you know stuff about. Every single autopilot on the market today will automatically disengage in the event of an engine failure.

Re:How hard is it to recognize a stoplight?

[phantomfive]

I guess you've solved all the problems. Maybe you should write to John Leonard and tell him?

Re:How hard is it to recognize a stoplight?

[dgatwood]

Describe for me, programmatically, the difference between a stoplight and a taillight.

That's easy. The stoplight is above you. Two cameras at different angle provide sufficient parallax to tell the difference between something far away on a hill and something nearby above the car. And you're done.

and a police light

Same answer.

and a neon sign

Same answer, plus the stoplight is not on the side of the road, as computed based on distance to the edge of the road when looking forward.

and also, please include all the many shapes and sizes of the various stoplights all over the country.

No need. Humans can't see the shape of the fixture when driving at night, but that limitation has never been a problem. You just need to know the color and to be able to figure out which colored light corresponds with which lane.

Re:How hard is it to recognize a stoplight?

[Anonymous Coward]

But... what about the chauffeurs? And the taxi drivers? And the DMV? And the traffic cops?

Wait a minute, nevermind, bring it on.

Re:How hard is it to recognize a stoplight?

[Anonymous Coward]

The journalist is not confused.

You need to understand how the google car deals with traffic signals. You need to distinguish between object (in this case, traffic light)
detection vs recognition. Maybe easier to describe in terms of face detection and recognition. Detection finds ANY face ANYWHERE in a given video frame.
Once a face is detected, the location can be handed off to the recognition module for deciding WHOSE face it belongs to. Detection is all about discriminating one class of object (faces, traffic lights) against the background (i.e. everything else). Recognition is about discriminating individual objects within a given class (individual faces, red or green traffic light). From 10,000 feet, both problems seem like the same type of problem. But in practice, for real-time vision, detection has to be FAST since it has to scan over the whole image frame. Recognition can afford to be more computationally intensive since it only has to process around the neighborhood of the location given the detector. So in practice, they typically use very different types of algorithms for detection (say using fast integral images) and detection (say deformable model based).

Google's car does not do ANY detection of the traffic light. Based on the currrent location and pre-mapped 3D location of the traffic light,
it can project the position of the traffic light in the image.
It simply recognizes the state of the lights around that position. This processing is cheap since it is on a small part of the image. So it is completely dependent on knowing the location of the traffic light beforehand.

One can try building a detector for traffic lights ofcourse. Ignoring the need for frame-rate processing, an ideal detector will have zero false positives and negatives. No real detector can achieve such ideal performance. In a real detector you will have to adjust the "knobs" to trade-off false negatives vs false positives. Since we are dealing with human lives here,
the knob will be adjusted so that false negatives are close to zero. Which means false positives per frame will be high. Which means your real-time performance for the recognition module is unbounded.

(PHD in Computer Vision who has worked with Sebastian Thrun)

Re:How hard is it to recognize a stoplight?

[Singularitarian2048]

If you think it should be easy to recognize a stoplight, you haven't worked on computer vision.

It's hard to recognize the most basic objects, like cats. How our brain does it so easily remains a mystery.

Re:How hard is it to recognize a stoplight?

[dgatwood]

No, I haven't solved any of the hard problems, because determining whether a colored ball or arrow is meaningful really isn't one of them. The hard problems are things like:

• recognizing and handling road signs
• dealing with potentially contradictory lane markings
• dealing with rain on the cameras
• determining which way to swerve when avoiding obstacles (like a dog running across the road), and whether to brake instead, or do both
• choosing whether it is better to hit the object in the road or swerve into the next lane (including computing the distance and speed of an oncoming vehicle correctly, even if it is a motorcycle)
• handling four-way stops when other vehicles don't follow the rules
• determining weather conditions sufficiently to compute braking distance correctly (Is it rainy or just cloudy?)
• recognizing that there are kids playing by the side of the road and you should probably slow down just in case one of them falls out into the street....

Traffic lights are relatively straightforward by comparison, so long as they are working.

Re:How hard is it to recognize a stoplight?

[Anonymous Coward]

Stop signs have a very specific shape, and text printed on them. They do not very from place to place.

Except in Quebec. In France, a stop sign says STOP. In Quebec, the French-speaking part of Canada, a stop sign says ARRET.

Oh, just checked Wikipedia. It's not just those weird French-Canadians: http://en.wikipedia.org/wiki/Stop_sign#Stop_signs_around_the_world

Re:How hard is it to recognize a stoplight?

[phantomfive]

Well good, now you're thinking more deeply.

Re:How hard is it to recognize a stoplight?

[gnasher719]

The problem you have is, someone like that wont let the car drive itself because a self driving car will stick to speed limits and slow down at pedestrian crossings because it will be programmed to anticipate stopping at a pedestrian crossing (like a defensive driver is trained to do). Nope, someone that self adsorbed and with such poor time management skills will be taking manual control with the pedal pressed to the floor whilst screaming into their phone. You simply cant overcome selfishness with a new technology.

If an accident happens, and you have a car that could have driven itself, and recorded what speed it should have gone and what speed it did go (like "crossing ahead, slow down from 50mph to 30mph" vs real speed accelerating to 60), you have an excellent witness. Jailtime for killing a pedestrian can overcome selfishness.

Re:How hard is it to recognize a stoplight?

[ThePromenader]

IMHO, cars won't be (safely) automated until roads are. Yeah, I know, catch-22.

Re:How hard is it to recognize a stoplight?

[Nikademus]

It's almost like they were designed for the task.

They were indeed designed for the task. The idea of that octogonal shape in almost every part of the world is for people to be able to recognize that sign without any confusion. It also allows drivers facing the back of the sign to identify that drivers from other lanes have a stop sign. It is also identifiable by night (because original signs were not reflective and cars lights were not particularly effective).

Re:How hard is it to recognize a stoplight?

[Anonymous Coward]

I'd read another comment against the (current state of ) Google Car which seems worth mentioning again. Apparently people are just a "blurry vertical stick figure" to them, or something to that effect as far as the car's visual recognition goes; so it can't distinguish a 'regular' person crossing the street from a cop waving you to pull over.

The removal of traffic lights and other cool things and a long ways into the future so you can't rely on that for a 1.0 or even 2.0 commercial Google Car release. Even with no pedestrians crossing the road, I imagine you'd need every car to be self-driving and able to coordinate with others for this to work. It would take one person driving their 'vintage' chevy to make one tiny mistake and there's bound to be an accident.

Also, in an era of self-driving cars, these 'vintage' cars might need to retro-fitted (or whatever the appropriate term is) to be considered street-legal by the new standards.

Re:How hard is it to recognize a stoplight?

[mestar]

Considering the benefits, bring on the laws that mandate that each stop light wifi/bluetooth/whatever broadcasts its state .

Re:How hard is it to recognize a stoplight?

[DrXym]

But it can still drive on other roads with good accuracy.

The lights are out at a junction. How does "good accuracy" help the car figure out when it's safe to proceed, or the order to proceed when there are buses, cars, trucks coming from all direction with an implied priority based on conditions and time people have waited?

Now a cop turns up to direct the traffic because of a fender bender. How does the car with "good accuracy" know to obey the cop's hand signals?

Now the repair crew turn up to fix the lights and put cones out so people turning have to do so from the adjacent lane. How does "good accuracy" cope with that?

Now a crazy person turns up and begins directing traffic. How does "good accuracy" tell the difference between the cop and the crazy person.

That's just a trivial demonstration of the problems a self drive vehicle would face. It's trivial to think of others - road flooding, narrow roads, diversions, vehicle break downs, animals running out, snow / leaves obstructing sensors etc. Of course in every case the simplest answer would be for the driver to override the car and manually drive it. But that naturally puts a dampener on some of the absurd expectations people have for these vehicles (e.g. that they can drive off and park themselves, self drive taxis, sleeping or drunk drivers etc.). And if the car gets confused too often or "fails to safe" for no reason then it will be infuriating.

It would be far more productive to concentrate on advanced driver assistance - cruise control, distance maintenance, lane tracking / marker detection, collision / hazard avoidance and parking assistance.

Re:How hard is it to recognize a stoplight?

[NicBenjamin]

My point is they're completely different engineering problems.

An autopilot needs to know physics, and the state of the aircraft. A self-driving car needs to be able to see and understand complex visual information.

You can't compare the two.

Re:How hard is it to recognize a stoplight?

[Anonymous Coward]

The way people drive that would be necessary anyway.

Re:How hard is it to recognize a stoplight?

[N1AK]

A human, I might add that's breaking the law. Distracted driving is illegal.

Do you have any idea how stupid that point is? I suppose the people being held by ISIS shouldn't worry because most people wouldn't behead them, only someone breaking the law and that's illegal! I can stop locking my house as well because entering and taking my stuff is a crime! That must mean I shouldn't be concerned about it because it's illegal.

It's pretty fucking obvious that illegal, distracted or poor driving causes the vast majority of accidents. Unfortunately there's a lot of illegal, distracted or poor driving. If you can suggest somewhere one can drive where there isn't then maybe people there don't need to consider it, but people everywhere else do.

Re:How hard is it to recognize a stoplight?

[N1AK]

There's a *huge* market for this. My grandmother in law is 93....

This is why we'll have it sooner rather than later. There are so many reasons why people can't drive. So many people who shouldn't be driving but are allowed to because stopping them is controversial (the elderly). So much money spent on paying people to drive (Taxis, Couriers, Hauliers).

Re:How hard is it to recognize a stoplight?

[N1AK]

Are you contradicting this based on any more than a desire to be pointlessly argumentative?

It is a more reasonable assumption that the computational power will continue to increase in line with the long term trend than that it will not, lacking evidence to the contrary. Given that the most powerful computer currently is can reach 34 Petaflops and India is planning to build a ~130 Exaflop machine before 2018 it's clear that vastly more powerful machines can be built.

Re:How hard is it to recognize a stoplight?

[StripedCow]

You don't need to describe it programmatically.
You can just train a neural network for that, then test that the miss-rate is below a certain threshold.

Re:How hard is it to recognize a stoplight?

[AmiMoJo]

Describe for me, programmatically, the difference between a stoplight and a taillight.
and a police light
and a neon sign
and every other red light on earth...

Height above ground.

Humans already have the same problem in low light conditions, and we already solved it. Still, occasionally people mistake other random red lights in the sky or on buildings for stop lights.

Re:How hard is it to recognize a stoplight?

[The Grim Reefer]

If the cars that fall back to AI then communicate their observations and decisions back to Google then to other cars then the next car wouldn't need AI and could improve knowledge of the area, plus any particularly bad problem spots can be highlighted for further investigation at Google HQ.

Personally, I think that's a brilliant idea. Some how I don't think the privacy rights crowd is going to be thrilled it though. The "Evilz Big Corporation" will be able to take over the world, control your mind and force you to buy figgy pudding. The only thing stopping them now is not knowing where you are physically located.

Re:How hard is it to recognize a stoplight?

[Anonymous Coward]

The problem you have is, someone like that wont let the car drive itself because a self driving car will stick to speed limits and slow down at pedestrian crossings because it will be programmed to anticipate stopping at a pedestrian crossing (like a defensive driver is trained to do). Nope, someone that self adsorbed and with such poor time management skills will be taking manual control with the pedal pressed to the floor whilst screaming into their phone. You simply cant overcome selfishness with a new technology.

You can, but the technology needs to be sufficiently advanced that there are no manual controls. (And sufficiently embedded that the asshat finds it difficult or uncool to get hold of a manual car.)

One thought, though: would driverless cars even need speed limits? They should be capable of maintaining a lidar map of things that *might* jump out in front of them and adjust their speed according to an algorithm that ensures they can always stop in time to avoid hitting the closest potential obstacle.

Re:How hard is it to recognize a stoplight?

[Anonymous Coward]

And to requote: "as the car's sensors get better, they will rely less on perfect accuracy."

So (a) they're collecting vast amounts of data to improve their mapping, and (b) the cars will progressively require less accurate mapping data.

So there will be a crossing point where the cars have pretty much universal mapping data and can drive about freely. It's not a finished product as of today, obviously, and it sounds like they have a roadmap to ensuring that it will be when it is released.

Re:How hard is it to recognize a stoplight?

[drinkypoo]

You need to compare the SDC to a fully aware human being, not a fully distracted one.

Why? A fully aware human being can also be aware while the car is driving. But most people drive in a state of at least partial distraction. You're setting the bar awfully high.

Re:How hard is it to recognize a stoplight?

[ShanghaiBill]

Are you guessing this based on Moore's law?

Partly. Moore's Law should continue for several more iterations. After that, who knows? But other factors will also make a huge difference. Much of the computing power in SDCs goes toward image processing. That is currently done with standard CPUs. In the future, there will be massively parallel custom hardware for vision. Even the human optic system has parallel low level wetware for things like edge detection, motion detection, etc. So far, we haven't deployed much custom hardware for vision, because it is still a niche application. But that is changing because of SDCs, general purpose robotics, security, automation, etc.

Re:How hard is it to recognize a stoplight?

[Anonymous Coward]

Considering the benefits, bring on the laws that mandate that each stop light wifi/bluetooth/whatever broadcasts its state .

Good luck with that. Most stoplights are still using incandescent bulbs and run simply on timers that don't take the actual amount of traffic at the intersection at the time of day into account.

Re:How hard is it to recognize a stoplight?

[Anonymous Coward]

I think the real goal would be to have all vehicles self-drive; then they can be coordinated to interlace at intersections, removing the need for stop lights and saving a ton of fuel!

And then the data center running that particular city could become nice, juicy target for malcontents...

Re:How hard is it to recognize a stoplight?

[Anonymous Coward]

1) Who is going to pay for their re-training and re-education?

2) You're assuming that all of these displaced workers are intelligent enough to do these new jobs.

Re:How hard is it to recognize a stoplight?

[Anonymous Coward]

I was blown away when I first saw this:

https://www.youtube.com/watch?v=csvt6JBAwBk

Re:How hard is it to recognize a stoplight?

[damnbunni]

But they're not tested much before they're given a license and allowed to drive.

And I honestly don't think I've been tested 'tens of thousands of times' by pedestrian traffic. I've had someone step out in front of me three or four times, maybe? Perhaps a hundred or so walk behind me while I'm backing up in parking lots?

Re:How hard is it to recognize a stoplight?

[Anonymous Coward]

So on a road with a busy sidewalk your car might got 3 miles per hour? I mean anyone of those people could step out into the road.

Re:How hard is it to recognize a stoplight?

[bluefoxlucid]

It's a legitimate point, but a poor example.

The car needs to also recognize hazards including pedestrians, animals, and drivers violating traffic law; if the car decides the intersection is safe to cross, and crosses illegal, it shouldn't hit anything anyway. Conversely, if the intersection is marked clear, and a driver crosses illegally, the car should recognize this. Humans often heuristically identify drivers not exhibiting a stopping behavior when approaching an intersection, projecting the driver will violate a stop sign or traffic signal; the car must do this as well.

Re:How hard is it to recognize a stoplight?

[bluefoxlucid]

Self-driving electric lorries would sharply reduce the cost of shipping by reducing fuel cost (my 12 gallon, 300-mile tank would cost $1.92 to fill if it were a Tesla electric Model S) and eliminating the wages of the truck driver for the 5000-mile, multi-day, cross-country journey.

Self-driving electric cars would eliminate ZipCar and Taxis. Like ZipCar, you could charter a vehicle for personal use, for $6/hr, insurance and fuel included; like taxis, the car is immediately returned to the service operator when you exit the vehicle. Drivers won't pay 8 hours to have the car parked 8 hours outside their job, and there won't be cabbies to demand any sort of salary at $7.25/hr (already more than the ZipCar costs in fuel, insurance, maintenance, etc.).

There would be no need for buses for mass transit, as private cars are always instantly available. Only shuttle bus, such as for mass transportation of school children, would remain.

Logically, we understand the cost of anything is the cost of labor. There is little gold, and it requires many men working many hours to fetch a small amount of gold from the earth. The same can be said for oil, and for copper, and iron. Many men work to refine steel, to transport refined girders to building sites, to rivet building frames and attach concrete forms to support structures. Labor is the cost of everything.

On this logic, automation systems must require less labor to build, transport, and upkeep than the persons they replace. A self-driving car must cost roughly as much in labor to build and maintain as any other car, with any additional costs for sensors or software being altogether less across the life of the car than the labor cost of the operator driving a modern vehicle. That is to say: one cabbie driving over the useful life of a taxi fleet vehicle must cost more at his wages than the increase in cost and maintenance for a self-driving car.

Imagine a wage worker making $20/hr produces the factory sensor part, at a rate of 100 per hour through the operation of a machine. Each part thus costs 20 cents, plus the cost of the raw materials, plus maintenance cost of the machine to build them, plus the cost of a part divided by the average number of correct parts produced per one defective part. This may all come out to under a dollar, or reach as high as a thousand dollars; but, even then, a cab service running 16 hours per day would, between two drivers, spend more than that on wages in ten days. If the part does not break within ten days, and it is the only increase in cost, then this cab is cheaper to own and operate than a cab with a salaried driver.

Automation is the art of replacing the labor of 100 low-paid workers with the labor of 10 workers each making three times as much.

Re:How hard is it to recognize a stoplight?

[Anonymous Coward]

Yep, whenever they put a new light in, they generally set it to blinking with signs letting you know about the light ahead of time. This is because people who drive that route every day will drive through the light because their brains are mapped to it not being there as opposed to drivers not from the area whom won't have the same preconceived notion.

So the first car to see the new light being put up would simply send a notice to the central server that things are changing there.

I'm pretty sure a google car might notice the change (a new light up), but not be able to identify it properly, which would require control being put back to human hands much like whenever it is going through any constructions.

Re:How hard is it to recognize a stoplight?

[ceoyoyo]

Betting on unbounded exponential growth is generally not a good idea. Those "computers" you're talking about cost billions, occupy warehouses, and are actually clusters of multiple computers. Clustering doesn't usually follow Moore's law unless the individual processors are doing so.

Processor development is not currently obeying Moore's law (http://2.bp.blogspot.com/-tcoUrJ8CNPk/UsVfZlcA4TI/AAAAAAAAKpk/PdOMedHinN4/s1600/chip2.png). It might be just a matter of waiting for new technology, as has happened in the past, or it might be something more fundamental. We are nearing some fairly hard physical boundaries. There are very different ways of building computers that might get us back on track, but it's not a sure thing.

Re:How hard is it to recognize a stoplight?

[ceoyoyo]

Sure you can. Aircraft autopilots have a fairly simple job that needs to be done superbly reliably. Most of the things you've listed are somewhat hard for a human because he has to remember to think about them, like balancing fuel. It's easy to program a computer to always make sure fuel is balanced. So easy, you don't actually need a computer for it. An aircraft autopilot doesn't "need to know physics." It needs to know that with the flaps at 15 degrees the minimum speed is X. Again, that's so simple that aircraft used to implement stall warnings with very simple electronics. Some may have even done it purely mechanically. For the hard part of piloting an aircraft, landing, the autopilot requires external help from instrument landing systems. The big challenge is making sure the human programmers don't screw up while writing the relatively simple code.

You can make a self driving car just as easily if you design the road to go with it. But if you want a self driving car that drives on regular roads you have to do computer vision, which is a lot harder.

Re:How hard is it to recognize a stoplight?

[Anonymous Coward]

If I was walking across an intersection, I would trust a Google SDC far more than someone late for an appointment, driving a Chevy Tahoe with a cellphone in one hand, a Starbucks latte in the other, and two screaming kids in the back seat.

The problem you have is, someone like that wont let the car drive itself because a self driving car will stick to speed limits and slow down at pedestrian crossings because it will be programmed to anticipate stopping at a pedestrian crossing (like a defensive driver is trained to do). Nope, someone that self adsorbed and with such poor time management skills will be taking manual control with the pedal pressed to the floor whilst screaming into their phone. You simply cant overcome selfishness with a new technology.

But actual autonomous cars are years away from practical use. Decades away from the way you're thinking. The first autonomous cars will be normal cars with an autonomous mode that only works on specially upgraded roads which you can guarantee will be limited access roads (freeways and motorways) with no traffic lights.

That's approximately 2 problems down the line.

The current problem is making a car that can be trusted to drive safely.

The next problem is getting people (lawmakers) to trust cars to drive themselves safely, so we can get the tech widely adopted.

Then we can worry about solving the "idiots who override the car and drive unsafely" problem, likely by insurance companies hacking up your deductible or denying your claim if you have an accident while driving manually, possibly by prohibiting manual driving on public roads. But that's not even something to worry about until the car is provably safer.

Re:How hard is it to recognize a stoplight?

[lister king of smeg]

Considering the benefits, bring on the laws that mandate that each stop light wifi/bluetooth/whatever broadcasts its state .

Good luck with that. Most stoplights are still using incandescent bulbs and run simply on timers that don't take the actual amount of traffic at the intersection at the time of day into account.

Based on some of the light were I live I am pretty sure some of them are using a random number generator rather than timer to decide how goes next and for how long.

Re:How hard is it to recognize a stoplight?

[0123456]

Why would anyone in their right mind want to trust a car that's relying on mapping data programmed into it rather than looking out at the real world. Do you think you could drive safely through town by ignoring the road and just turning where your GPS tells you to?

Re:How hard is it to recognize a stoplight?

[0123456]

One thought, though: would driverless cars even need speed limits?

So cities can raise revenue by changing the speed limits every day or two, and catch cars before Google has updated their maps.

Re:How hard is it to recognize a stoplight?

[0123456]

And is it a stop, or a three way stop, or a four-way stop? I think we even have a two-way stop somewhere in town, where one lane is allowed to continue through without stopping.

Re:How hard is it to recognize a stoplight?

[JimFive]

That's easy. The stoplight is above you. [...] plus the stoplight is not on the side of the road

There are places where the stop lights are on posts at the side of the road and only slightly above you, about the height of the lights on a Semi trailer. If the road is divided there may be a light on each side of the road indicating instructions for particular lanes. I saw this most often in Paris, but I've also seen it in Chicago.
--
JimFive

Re:How hard is it to recognize a stoplight?

[phantomfive]

It is a more reasonable assumption that the computational power will continue to increase in line with the long term trend than that it will not, lacking evidence to the contrary.

Then look for some evidence.

Re:How hard is it to recognize a stoplight?

[Anonymous Coward]

"Good Accuracy" isn't too far off from what human drivers are capable of. There is one traffic light that I have personally witnessed at least two people blowing through it at 55-65 MPH on a red as I pulled out into the intersection on a green. And just because the car may have accuracy issues with lights doesn't mean that it won't notice a pedestrian. Its a lot easier to recognize an obstruction in the road via LiDAR then to recognize one of hundreds of lights (street lights, lighted stop signs, brake lights, turn signals, advertisements, etc) the software is likely to be sifting through on an average drive is telling it that it should stop.

Re:How hard is it to recognize a stoplight?

[phantomfive]

Yes, and one of those problems is much easier than the other.

Re:How hard is it to recognize a stoplight?

[phantomfive]

That's true, good points.

Re:How hard is it to recognize a stoplight?

[Anonymous Coward]

Where I live it's a mandatory 3 day jail sentence in county lock up (and Benton County Jail is a shit place that needs to be shutdown) for speeding through a school zone. Most people obey the school zone speed limits due to this harsh deterrent.

I was in court one day for some domestic assault and battery charges and property damage charges (Wife was fucking a nigger. We were/are separated and still fuck each other from time to time. Back then, she was lying to me about her other relationships so I ended my other relationships... We are now completely transparent about our other bfs/gfs so we avoid these violent episodes completely now), and this one woman got hauled off to fucking jail right then and there for speeding through a school zone. Did not get to pass Go. Did NOT collect $200. And the shit of this is, they detained her immediately and then they take your phone.

Since the jail only turns the pay phones on for like 2 hours a day at the time that most people will be at work or otherwise at a place where you can't call collect, she had no way to inform her job and family of what happened. Fucking jack booted thugs around here.

My assault charges, etc. were completely unfounded and thrown out, of course, but this woman doing 5 over the speed limit in a school zone got hauled right the fuck off to jail. In the middle of the week in the middle of the day.

Once your wife gets the jungle fever, there's no cure. The problem is that now she is constantly being used and played by these kind of people. It drives her insane. She's been at my place every night the last two weeks because these niggers are driving her nuts. Makes it real hard to juggle my white girlfriends against all this time my wife's getting with me. Oh well... I do lover her, and she gets first dibs on my time.

They don't just smoke up all of her weed and drink all of her alcohol while they're there banging her. They actually TAKE this shit with them when they leave afterwards. They've stolen stuff behind her back. Played her into committing to a monogamous relationship while they were out fucking other apes. The list is endless.

True story.

Re:How hard is it to recognize a stoplight?

[Anonymous Coward]

I was going to link to one, but it's not self-driving...

Re:How hard is it to recognize a stoplight?

[ebno-10db]

Self-driving electric lorries would sharply reduce the cost of shipping by reducing fuel cost (my 12 gallon, 300-mile tank would cost $1.92 to fill if it were a Tesla electric Model S) and eliminating the wages of the truck driver for the 5000-mile, multi-day, cross-country journey.

Here is a simpler idea using existing technology: freight trains. You still need trucks (eh, lorries) for local delivery, but even if they have a driver they can be electric. It's also quite different than your 5000 mi example. For anything over about 200 mi, intermodal (train/truck) uses less energy and is generally cheaper. I don't know about the UK, but what the US needs is more intermodal facilities and better coordination. Even still it's more widely used in the US than many people realize. Trains can be electric of course, but even if you don't want to pay to electrify long distance lines, trains use about 1/3 the amount of diesel fuel that trucks do per ton-mile.

Re:How hard is it to recognize a stoplight?

[ebno-10db]

The issue, as I understand it, with the SDC is road conditions, not obstacles of decent size (humans, bikes, other cars, etc). Potholes or things like ice or a broken piece of wood with nails sticking out of it.

Rain and snow are also problems for SDC's. With all that testing in SV they might not have considered such things (I think everybody stays indoors when it rains), but in most of the rest of the world they're common conditions.

Re:How hard is it to recognize a stoplight?

[Charliemopps]

Really? Neon signs aren't above you? The car should stop for all red lights above 6ft tall?
What about that radio tower 2 miles away whose 50million candle power light has about the same luminosity at your location as a stoplight?

A better solution would be for google to start making the stoplight bulbs themselves, and have then pulse and a predetermined and invisible rate. It would be a cheap and easy solution.

Re:How hard is it to recognize a stoplight?

[ebno-10db]

I've had someone step out in front of me three or four times, maybe?

You obviously don't drive in Manhattan.

Re:How hard is it to recognize a stoplight?

[ebno-10db]

It can use OCR to read signs.

As long as they don't get too dirty. Think Captcha - humans do better. It's the same problem that they have with bar codes on freight cars. Not to mention other contextual cues that humans can use.

Re:How hard is it to recognize a stoplight?

[Anonymous Coward]

Interlace at Intersections?
Simples! They do that over here anyway.

(Clue: hablo español, pero no soy español)

Re:How hard is it to recognize a stoplight?

[Ichijo]

Since it cannot reliably detect when children are present, it would most likely just default to the lower speed.

And then be pulled over for obstructing traffic or otherwise being a traffic hazard.

Re:How hard is it to recognize a stoplight?

[schlachter]

Self driving cars will turn out to be faster, even if they obey the speed limit...because traffic will disappear or lessen significantly and intersections can be managed dynamically without the need for stoplights that make you sit constantly. There won't be accidents to jam up the roads, and detours will be seamless because all cars will coordinate a good distribution over alternative roads. Parents don't have to have screaming kids because they can be facing them and engaging them during the drive. And there will be no need to park when you arrive, saving time at the destination. And you'll be driving less overall as well. Need something quick at the store? Send your car out to pick it up. Or request delivery from the store. No need to physically go.

Re:How hard is it to recognize a stoplight?

[Anonymous Coward]

It's interesting to see how you all good engineers buy in on this self-driving car hype. Please do your homework on how FPs and TPs kill your FMEA analysis and how it is close to impossible to combine a low FP rate with a high TP rate, given the exposure rate to hazardous scenarios and the intrinsic FP/TP-trade-off existing in all filters and real-world algorithms, to reach a system with high availability and acceptable risk. As working in the field for a couple of years now as a systems architect, I've come to the conclusion that this problem requires no less than to solution of the AI problem due to the explosion of combinations of all the variational parameters in real-world traffic. It is, however, possible to do very good demos for situations where the variational parameters are restricted, such as pre-scanned streets or highways. These demos are enough to attract enormous amounts of financing and duping non-engineers, but I predict a "self-driving car winter".

Re:How hard is it to recognize a stoplight?

[CanHasDIY]

> What we need is better geo-mapping from cities themselves

For driver less cars to work, the whole city should be wired so the google car doesn't have to recognize the red light, it would just get the information through some type of wireless transmission thus knowing it has to stop.

Translation: Won't work without taxpayer subsidized infrastructure.

Re:How hard is it to recognize a stoplight?

[FlyHelicopters]

If I was walking across an intersection, I would trust a Google SDC far more than someone late for an appointment, driving a Chevy Tahoe with a cellphone in one hand, a Starbucks latte in the other, and two screaming kids in the back seat.

^ This, times ten, times twenty... every day of the week and twice on Sunday...

The fact is, most humans WAY over rate their own abilities and underrate computer's abilities.

If you're reading this, odds are you are NOT as good a driver as you think you are.

Re: How hard is it to recognize a stoplight?

[Anonymous Coward]

I would so love for my car to be able to do that!

Re: How hard is it to recognize a stoplight?

[Anonymous Coward]

Nonetheless, the algorithm to handle a sabotaged location sensor (RFID, Magenetic, or whatever) is brain dead obvious: pull over and stop the car.

Exactly. That's what I meant by "freak out". You're the one who presumed fatal consequences and likened it to spree killing. The failsafe mode in any autonomous vehicle MUST be to immediately stop.

This would be like being able to scatter practically invisible stop signs everywhere. From a user standpoint, having a vehicle that is proceeding blithely and suddenly starts honking warnings and pulls over to the side of the road is a very good definition of "freaking out". It's obvious the clueless users wouldn't know what happened or why their amazing wondercar has stopped and refuses to proceed.

Re:How hard is it to recognize a stoplight?

[NicBenjamin]

So it has to know the answers to thousands of physics calculations, and it has to know what to do when the answer is wrong (ie: speed is too low for flaps at 15 degrees), but that doesn't constitute "knowing physics"? That's a mighty fine distinction you're drawing there.

And, like I said, they're completely different engineering problems. With aircraft you're entering a lot of precise mathematical formulas that have to do with the current state of the aircraft. With automatic driving it's all processing visual information.

Re:How hard is it to recognize a stoplight?

[NicBenjamin]

If your point is that lots of people have done autopilots, and none of them have done self-driving cars; then yes autopilots are by definition easier then self-driving cars.

I'm curious: why do you think that's at all relevant to any other part of this thread?

Re:How hard is it to recognize a stoplight?

[phantomfive]

I'm curious: why do you think that's at all relevant to any other part of this thread?

Well, if you're curious, I believe this conversation was started with this earlier point of mine:

"And that is even before getting to the engineering problem of reliable software. When Boeing built their recent airline software package, it took 5-8 years to get something that was reliable enough for air travel."

Which is a comment on the difficulty of making reliable software. It's referring to the engineering aspect of the matter, that making software without (serious) bugs is very difficult and takes time. Since the software for self-driving cars will be much more complicated than the software for air travel, it follows naturally that making the software reliable will be even more difficult.

My point being, that even if the car could already drive itself anywhere, it would still be many years away from production.

Re:How hard is it to recognize a stoplight?

[shutdown -p now]

Tech doesn't need to overcome it, that is the territory of the law. At some point, once self-driving cars are good enough and ubiquitous, they will simply be mandatory on public roads, and taking over to control it in an area where it can drive itself would be a crime.

Re:How hard is it to recognize a stoplight?

[Lumpy]

magnets in the lanes will also solve a LOT of the problems. It would take minimal cost do retrofit and significantly help ALL cars as lane departure would be trivial, etc...

Re:How hard is it to recognize a stoplight?

[Lumpy]

"Imagine a wage worker making $20/hr produces the factory sensor part, at a rate of 100 per hour through the operation of a machine."

Will never happen. Reality is very different...

Imagine a wage worker making $8/hr produces the factory sensor part, at a rate of 100 per hour through the operation of a machine.

Executives do NOT want to pay living wages. It will be an $8 an hour job or outsourced to China where they pay $3.25 an hour.

Re:How hard is it to recognize a stoplight?

[jklovanc]

Mapping data used by the Google car goes far beyond GPS. The issue with traffic signals is that Google ca needs to know where to look for them. It can not pock out the signals from the whole video stream fast enough so it restricts it searches to specific pre-defined areas. When it finds one it can react to it.

Even then, you have a point which is why Google car is not as useful as many people think. Things change too much for pre-scanned and human edited data to ever be a valid basis for autonomous vehicles.

Re:How hard is it to recognize a stoplight?

[jklovanc]

So there will be a crossing point where the cars have pretty much universal mapping data

Roads change quite often. Every time a new road is added, traffic light changed, intersection changed, speed limit changed, etc the road may have to be re-scanned and the database updated. For example, the car can not identify the marks on a road indicating a crosswalk. That information has to be already in it's database. The issue is that the mapping data would have to be updated so often as to make that method ineffective.

Even relying on less than perfect accuracy in maps is not the complete problem. The main problem is that the AI in the vehicle can not tell the difference between a person on a sidewalk wanting to cross the street and a cardboard cutout advertisement. There are many situations where the computer can't think it's way out of a problem and a person has to take over. The issue is AI and not sensors.

Sure thee may be a point at which AI has advanced enough to drive a car. I just don't see it in the next 50 years.

Re:How hard is it to recognize a stoplight?

[Anonymous Coward]

Put out of work by what? The people building the cars? The people repairing the cars? The people installing and maintaining the off-vehicle sensor packages? The people building all the bridges and tunnels? It seems to me that this transition would INCREASE available jobs.

There is no law of economics that dictates when new technology disrupts one particular profession, an equal or greater number of jobs are created in the process.

Transportation is one of the largest industries. The number of jobs created maintaining and servicing the self driving vehicles and sensor infrastructure would not even be a drop in the bucket compared to the number of jobs that would be lost. This research has already been done.

http://hardware.slashdot.org/story/14/08/15/1512230/humans-need-not-apply-a-video-about-the-robot-revolution-and-jobs

Put out of work by what? The people building the cars?

Much of the work now is done by specially crafted robotics and even now engineers are devising ways to remove the rest of the human labor from the manufacturing process. By the time self driving cars are ready to hit the production lines people may not be involved at all.

Put out of work by what? ... The people building all the bridges and tunnels?

People are needed to build bridges and tunnels now, so I don't see how that would change much.

People will still probably be needed to repair and maintain the sensors, but nowhere near the number of drivers that are no longer needed.

Re:How hard is it to recognize a stoplight?

[Anonymous Coward]

...Google maps out roads, keeping a database of signs, lane markings, etc.....If the car is driving on a pre-mapped road, it will use the info from the database. But it can still drive on other roads with good accuracy...

People do this now. They become familiar with a road and learn what to expect and that doesn't mean they can stop paying attention to the road and its conditions. I wonder what the human rate of error is when unexpectedly encountering a new stop sign or traffic light in your daily route.

Re:How hard is it to recognize a stoplight?

[bluefoxlucid]

With high automation, shipping fuel costs become way more expensive than labor.

Re:How hard is it to recognize a stoplight?

[ceoyoyo]

An aircraft autopilot doesn't have to "know the answers to thousands of physics calculations." Aircraft autopilots aren't aircraft simulators. They're made as simple as possible so that they can be as reliable as possible. They don't simulate air flow over the wings to figure out what the stall speed of the wing is, somebody programs that number in:

stallSpeedFlaps15 = 70

That's not "knowing physics." And no, you're not "entering a lot of precise mathematical formulas." You program in some safety limits, and a set of rules to accomplish what you want (if speed = stallSpeed then throttle++). The rules are determined using complicated fluid dynamics simulations and flight testing, but none of that is programmed into the autopilot.

The two are different engineering problems, but you can certainly compare them. Aircraft autopilots perform relatively simple tasks that can be directly programmed using a relatively small set of rules. Hobbyists write them for model airplanes to run on microcontrollers. For complicated tasks like landing, they rely on external infrastructure to reduce the task to something that can be solved with simple rules.

Self driving cars could be equally simple, if they drove on highways made for the purpose. The kind Google is working on are much more difficult, involving extremely complicated rules that are too complex to program directly and so the cars have to use machine learning.

Re:How hard is it to recognize a stoplight?

[NicBenjamin]

I think you're mistaking the compiler's ability to do an awful lot of work on your behalf, for a lack of work happening.

For example, "speed = stallSpeed then throttle++" is a formula. As for "thousands," keep in mind that a commercial jet like the Dreamliner has several flaps settings, for each wing, multiple rudder settings, operates at multiple different speeds, has to keep track of three fuel tanks, altitude, etc. If there are four possible numbers entered for each of those variables into the compiler that's 4^5 potential situations the Autopilot has to deal with. As a slashdotter, you probably know that adds up (2^2)^5 or 1,024 discrete sets of numbers.

With a commercial-level autopilot they probably did a lot of the work the compiler does for you themselves, because once you reach a certain level of expertise errors the compiler adds to your program are a bigger problem then errors you would have made if you'd done it yourself.

Re:How hard is it to recognize a stoplight?

[NicBenjamin]

Again you're talking about totally different engineering challenges.

The software running a plane is very specific to that aircraft. The Dreamliner's took 5-8 years partly because it had to be perfect, but also partly because the Dreamliner's design wasn't finished, and whenever the design changed the airframe's behavior the autopilot and all the software would have to be changed to fit.

If your argument is that it will take 10-15 years minimum before we're cruising down the highway at 85 with a computer driving, then I'll agree with you.

But you're also implying that once Google has done it for a single model of car, it will take another five years to solve it for the Rice Rocket version of that car. A single spoiler would completely change the aerodynamic characteristics of an aircraft, and require the Autopilot be extensively reworked. That's just doesn't apply to Google drive. Once Google has a version of Drive that works on one rear-wheel drive car it will be roughly a week from adapting it to any other rear wheel drive vehicle. Most of that week will be figuring out where to hook up the sensors on a different vehicle's body. Much of the work that an Autopilot has to know 100% all the time could actually be learned by the car as it was driving the vehicle.

For example, it's not a tragedy if your GoogleDrive car screws up the gas consumption calculations because you forgot to tell it you're a Coal Roller; and you have to walk to the nearest station. So Google includes a simple AI to figure out your vehicle's gas consumption from experience. If you did that with a jetliner everyone would die.

Re:How hard is it to recognize a stoplight?

[phantomfive]

I'm talking about software engineering challenges. It takes a long time to write reliable software, apparently you have trouble with this topic. Maybe it's because you don't know how to write reliable software, which is fine, most people don't.

It's not about the airplane, the airplane was just an example to help people who are unaware of the engineering challenges involved.

Blind cars

[Anonymous Coward]

It's pretty clear that all of the automated cars have an extremely limited view of their environment.

They have the senses, but apparently can't use them. As carbon based life forms, we've been able to manage all sorts of safe driving tasks with no local location or map knowledge. Simply following street signs, road marking, and watching curb formations. We can do this day and night, in the dry, rain and snow, in most any vehicle.

But most of the cars seem to have very limited local awareness (i.e. something is too close), and limited lane awareness. Everything else is "built it", and relies upon specific locations.

There was a video about an Audi doing 150 MPH around a racetrack, driverless. Nice and impressive, but what it was doing was just a step above what a video game does. It already "knew" the track, it didn't SEE the track, it was told what the track was, and followed internal programming that coincidentally coincided with how the track was laid out.

Impressive demo, but not really practical.

There's been chatter about making cars "aware" of each other through other mechanisms. That is, besides sight. Having cars broadcasting telemetry that local cars can pick up. "I'm a 6ft x 9ft car at lat, long going Z m/s".

Until cars can legitimately "see", see and interpret things that they are not originally aware of, they'll be little more than tech demos and not suitable for the wild.

Re: Blind cars

[Anonymous Coward]

I think it's more about biological vs mechanical and not about carbon based. Real time systems with sensors have been around a long time but driverless cars is a new challenge. It will be worked out within a decade I believe.

Re: Blind cars

[Molochi]

Just in time for fusion powerplants.

Rain and snow?

[iotaborg]

Perhaps more of a concern is the issue where the car will fail in rain/snow, both of things people in the Bay Area rarely experience.

Re:Rain and snow?

[BLToday]

That's not true. I've lived in the Bay Area, there's a lot of rain and fog. SoCal is completely different: no rain, occasional fog.

Re:Rain and snow?

[ShanghaiBill]

Perhaps more of a concern is the issue where the car will fail in rain/snow

LIDAR does not "fail" in rain/snow/fog. It just doesn't work as well. So what? Neither do human eyeballs. Sure performance will be degraded in bad weather, and the car will have to slow down to compensate. Which is exactly what humans do.

Re:Rain and snow?

[Anonymous Coward]

> slow down to compensate. Which is exactly what humans [are supposed to] do.

FTFY

Re:Rain and snow?

[unimacs]

Yes, but how badly does it degrade? Will it just stop if there is too much ice or snow on its sensors to "see" adequately. Will it just sit at a green light not knowing whether it's safe to proceed? Can it determine what the road conditions 10 to 30 yards ahead are and react accordingly?

Re:Rain and snow?

[Type44Q]

Sure performance will be degraded in bad weather, and the car will have to slow down to compensate. Which is exactly what humans do.

Considering that Texans and Okies tend to speed up when the streets are slippery and visibility's been reduced, I suspect this confirms my suspicion: hicks aren't human! :p

Re:Rain and snow?

[Guspaz]

LIDAR still works to detect surfaces, but when the road is covered in snow, the surface it's seeing isn't the road.

How can the car know where the lanes are if it can't even tell where the road starts and ends? GPS isn't accurate enough for the car to guess which lane it's in without visual cues, and if there are no visual cues, then the car effectively can't operate.

And, as it happens, Google does say that their cars don't work on snow-covered roads. That's a bit of a problem where I live, where you can have snow up to 7 months a year.

Re:Rain and snow?

[Anonymous Coward]

It will degrade until it has to stop... just like a human driver. I've been in rain that was coming down so hard that I pulled over until it was over.

Re:Rain and snow?

[Anonymous Coward]

Will it just stop if there is too much ice or snow on its sensors to "see" adequately.

Yes. And, I should add, so should YOU.

Re:Rain and snow?

[DNS-and-BIND]

Wow. Just...wow. Open bigotry and nazi-like denial of humanity of fellow humans...modded positively on Slashdot.

The mask is starting to come off now, isn't it?

Re:Rain and snow?

[Type44Q]

Your totally Aspie-ish inability to spot humor notwithstanding, I actually agree with your idealism... but it's quite clear that you've never gotten to enjoy the "pleasure" of living amongst the wonderful [and admittedly all-too-human] folks in question! ;)

Re:Rain and snow?

[schlachter]

Cars will use LIDAR, ultrasonic, and video for first party sensing.

But a lot of sensing will come from their party. Other cars reporting their position on the road. Their intent and upcoming moves. Details about the environment that they sense. Additional the infrastructure can provide info. The roads can report if cars are present. If there is ice, etc.

Anyways, the issue isn't if there is ice or snow on it's sensors. That will be easy to mitigate. The problem is the ice and snow in the environment that it needs to see through.

Re:Rain and snow?

[FlyHelicopters]

It will degrade until it has to stop... just like a human driver. I've been in rain that was coming down so hard that I pulled over until it was over.

This...

I've pulled over under a bridge before when the rain was so bad, you couldn't see the front of the car, much less the car in front of you.

Idiots on the road kept driving, watched two people have accidents that evening.

People are actually pretty crappy drivers on average. Yes, the good ones are good, but average drivers suck.

And I'm no expert, I fully believe a self-driving car is going to do a better job, overall, than I would.

Re:Rain and snow?

[unimacs]

Cars will use LIDAR, ultrasonic, and video for first party sensing.

But a lot of sensing will come from their party. Other cars reporting their position on the road. Their intent and upcoming moves.Details about the environment that they sense. Additional the infrastructure can provide info. The roads can report if cars are present. If there is ice, etc.

Anyways, the issue isn't if there is ice or snow on it's sensors. That will be easy to mitigate. The problem is the ice and snow in the environment that it needs to see through.

How much time have you spent scraping ice off of a windshield where it's gone from above freezing and raining when you parked your car to well below freezing overnight and sometimes even well below 0 F? How many times have you had to dig a car out of a snowbank? If you've had to do it over the course of a few winters you'll know that it's often not so easily mitigated.

I think people either don't know or often forget what a hostile environment winter can be in the the Northern part of our country. How densely packed are the sensors going to be in the road to tell reliably where there is ice, whether the ice is smooth or rough, whether there is snow and how deep it is and whether it is packed or fresh? How long are these sensors going to last when the roads I drive on end up filled with potholes each spring?

I'm sure it would all work great in a laboratory setting when every other car, plus all the roads and traffic control devices can talk. I'm not so convinced that it will work when things are in some state of disrepair as they often are in the real world.

Re:Rain and snow?

[Anonymous Coward]

The sooner we round all you hicks up in Jesusland and build a wire fence with guard towers around it, the better. You fucked up the country so bad that at this point, it would be genuine self defense.

Too early to tell

[Anonymous Coward]

Too early to tell. But the winner is not the one who develops a self driving car first. The winner is the first company to mass produce and sell a self driving car the consumers want.

I could see it used in specific cases....

[ArcadeNut]

If they built special lanes or only worked on places like the Freeway. It would be nice to have a self driving car for a 6 hour road trip and then manually take over in the cities or where the car had issues.

Re:I could see it used in specific cases....

[Anonymous Coward]

If they built special lanes or only worked on places like the Freeway. It would be nice to have a self driving car for a 6 hour road trip and then manually take over in the cities or where the car had issues.

A driver less car is pretty much worthless to me if I ever have to take over. Either I am focused on the task at hand (driving) or I am not. To do otherwise only maximizes distraction and invites accidents and injury. Imagine sinking into the daily Facebook updates after 2.5 hours of mind-numbing automated driving when suddenly a chime goes off indicating that the automated systems have failed and you have 1.5 seconds to make a decision critical to the safety of the vehicle and its occupants.

Re:I could see it used in specific cases....

[phantomfive]

I think his point was:

1) The car can drive for 2.5 hours on the freeway by itself, while you are not paying attention.
2) When the car arrives to an offramp, it will notify you (in advanced) that it's your turn to start driving.
3) If there is a problem along the way, the car will pull over and stop (or similar) before handing you control.

That way, you don't have to focus intently while the car is in control.

Re:I could see it used in specific cases....

[Anonymous Coward]

That is what I am interested in. But, it is the impatient drivers that can't drive that will ruin it. I don't care about fully automated driving in the city, alerts and warnings would be nice, but it is the boring long highway drives that I think they should get to work first.

The next problem is that on some scenic highways, there are no guard rails and sheer cliffs. If the car is off by a foot or two, it would be a long way down. Or if the road conditions aren't good and it has to apply the brakes, would it know not to skid? Would it know that it is better to crash than to go over the edge?

Re:I could see it used in specific cases....

[Wycliffe]

A driver less car is pretty much worthless to me if I ever have to take over. Either I am focused on the task at hand (driving) or I am not. To do otherwise only maximizes distraction and invites accidents and injury. Imagine sinking into the daily Facebook updates after 2.5 hours of mind-numbing automated driving when suddenly a chime goes off indicating that the automated systems have failed and you have 1.5 seconds to make a decision critical to the safety of the vehicle and its occupants.

I agree it's worthless at 1.5 seconds but not if the warning time was 15 minutes and if you didn't respond then it just pulled over to the side of the road.
This would be a huge benefit for travellers, business people, and truck drivers. Imagine getting in a car at 8pm in chicago, falling asleep and being woke
up at 8am saying you are about to arrive in New York. If I was google, I would take what I learned from city driving, learn from it, and switch full gear
in supporting major interstates. I would do it piecemeal and hit one interstate at a time like they are doing with google fiber. San Francisco to Los Angeles
would be a decent one to start with. There are probably lots of people in CA that would buy a car if it could drive hwy 5 from SF to LA while they slept.
I live in MO and I would gladly pay a decent price for a car that could go from Kansas City to Saint Louis unassisted even if that was the only route it could do.

Re:I could see it used in specific cases....

[Anonymous Coward]

And taking over while driving will be just as easy as un-pausing a driving game.

I.e. resulting in a crash 80% of the times.

Re:I could see it used in specific cases....

[swb]

This seems to be the way its actually being adopted if you start to consider the collision detection, adaptive cruise control and lane detection systems in some cars.

There was an article hear about a luxury car that has the ability to let you take your hands off the wheel for a certain amount of time under "cruise control" conditions and it will actually drive itself for a brief period. I think the article was about some "hack" that let you evade the built-in time limit for taking your hands off the wheel, meaning that the self-driving component was manufacturer-limited to avoid being used as a self-driving car.

I'd guess that self-driving cars will get adopted more this way than suddenly buying a car that can drive itself everywhere perfectly.

Re:I could see it used in specific cases....

[0123456]

I'd guess that self-driving cars will get adopted more this way than suddenly buying a car that can drive itself everywhere perfectly.

Bingo. Rather than the Lords Of Google coming down from the sky with the One True Driverless car that has no steering wheel at all, in the real world we'll start with increasingly driver-free cruise control on highways and progress from there. The average car will probably have hands-free cruise decades before it can drive itself around town.

But that makes the fanboys sad.

Re:I could see it used in specific cases....

[swb]

I listened to an economics podcast that discussed the role of automation in the economy.

The Google car came up and the MIT professor who was being interviewed said that contrary to what a lot of people think, the Google car is highly dependent on very detailed annotated maps and can't just detect stuff like traffic controls (lights, signals, etc).

I can see a driverless cars happening in urban areas if "self-drive" features start getting built into roads, like RFID chips embedded in lane stripes, traffic controls that have RF signalling. Even then you would probably need car-car signalling.

But at some point it seems like they aren't driverless cars as much as they are trackless trains or some kind of personalized mass transit. Decide you want a ride someplace, request a car and it follows GPS + signalling to your house from its parking place and then delivers you to a destination. Uber without the driver.

welp

[Anonymous Coward]

It will start with just freeway driving...then a partnership with builders of infrastructure to make the transportation system more uniform.

Was it ever intended to be a commercial product?

[harryjohnston]

It isn't clear to me that Google ever intended this to be a commercial product, or at least not in the short-to-medium term. Treated as a research project, it is impressive regardless of the practical limitations.

Limited Vision

[pubwvj]

"said Leonard, who does not expect a full self-driving car in his lifetime (heâ(TM)s 49)."

He is a man of limited vision. I did a lot of AI research and development, long ago, back in the dark ages of computing, and I disagree. I'm only a few years older and I do expect to see fully self-driving cars in my lifetime. Perhaps I merely will live longer...

Re:Limited Vision

[Anonymous Coward]

I'm currently doing a lot of research and development in AI and machine learning, I agree with him.

Machines are getting pretty good at doing things if the start state is well defined and the motions to complete something are very standard. There's a cool video of a robot learning how to swing a ball in to a cup based off of being shown once. But if you watch carefully, you notice that the ball must be completely stationary before the robot begins each attempt. Throw in even a little wobble, and it completely falls apart.

This is the state of robotics as of 2014. As long as nothing unexpected happens, and it only has to correct for minor variations, they do alright, anything beyond that is complete failure of the system.

Don't get me wrong, it's come a long way. But what, 40 years of research, and we have gotten to where a robot can deal with a fixed initial point with only small variations.

Re:Limited Vision

[Anonymous Coward]

And that is the crux of the problem - a self driving car has to deal with the unexpected. A policeman flagging down traffic, someone installing a fake traffic light (or even just a red light), or someone installing a bright green light at a junction, road works, another car crashing in front of it, all manner of weather conditions, working out where it can park not just how, etc.

Until a car can do all of those and cope with both real and fake data being thrown at it, as there will be those that try and subvert the system for fun or criminal gain, then it's unlikely to take off or be licensed for general use. Those are incredibly difficult problems that even the human brain sometimes struggles with.despite its sophistication and computing power.

What we will see are continuations of Tesla's system of tackling subsets of the problem like motorway / highway driving, and the systems that self park the car when the motorist selects the location. However there will be bumps in the road ahead when those systems get it wrong whilst the driver of the car is distracted or has stopped paying attention, and there may well be a backlash against them when that happens. It is currently illegal to even use cruise control in Belgium for example and that does little to remove responsibility away from the driver for the operation of the car.

Re:Limited Vision

[lorinc]

I don't agree with you. Scientific expansion in ML is of exponential growth: what took 20 years to achieve will take only 10 to be doubly improved. When I see the state of the art in computer vision 2014, I have almost no doubt the vision problems associated with automatic cars will be solve in a fewer amount of time that anyone expected.

At least ...

[mschaffer]

At least the Newton was a bit revolutionary. It could have been worse. At least it isn't the iPod HiFi of Automobiles.

Re:At least ...

[sexconker]

Beat up Martin.
Eat up Martha.
http://i.imgur.com/06fu9sE.jpg

Re:At least ...

[MightyYar]

Unless my history is rusty, Newton begat the ARM chip that we all know so well.

Re:At least ...

[phantomfive]

The ARM chip was started for the ABC Acorn and was released in 1986.

Re:At least ...

[MightyYar]

You might be disappointed if you tried to run the full-power desktop ARM from the pre-Apple era in your mobile device!

Weak criticism

[Anonymous Coward]

While there is serious obstacle to overcome for a fully autonomous car, the point the author make is weak.

A car working in a limited area, with good condition, let's say you start with a city. And sensor mapping everything all the time (including when human drive) would bring massive input of information. With increase capability of detection. Incrementally increasing the covered area. And participating city happily providing information on the traffic changes. Look at tesla already providing very limited autonomous driving. You could provide only highway at first. No need to map far off trails.

Automation of detection of light/traffic sign is a problem on which a lot of people are working, not only google, expect this problem to be "solved" before long.
All the car are connected. This would make it that much harder to miss one given all the "eyes" watching for them.

   

Re:Weak criticism

[phantomfive]

While there is serious obstacle to overcome for a fully autonomous car, the point the author make is weak.

The author made more than one point.

A car working in a limited area, with good condition, let's say you start with a city. And sensor mapping everything all the time (including when human drive) would bring massive input of information. With increase capability of detection. Incrementally increasing the covered area.

Google themselves say the problem of mapping the US in enough detail is too big. They're hoping to find a better way to do it (though of course with uninvented technology everything is possible).

Automation of detection of light/traffic sign is a problem on which a lot of people are working, not only google, expect this problem to be "solved" before long.

People have been working on this problem, probably for longer than you've been alive.

All the car are connected. This would make it that much harder to miss one given all the "eyes" watching for them.

That's another very interesting technology that still exists in the realm of theory. It's nice but it's not the Google car.

Government doesn't have records

[Anonymous Coward]

Safer (autonomous) cars could save tens of thousands of lives. Given that safety is one of the governments biggest reasons for existing, they should be doing everything they can to make autonomous cars a reality as soon as safely possible. Accordingly, if that means having up to date records for traffic signals, roads etc, then, well, they should have up to date records.

Re:Government doesn't have records

[ewibble]

It really depends on how much "safer" autonomous cars are doesn't it. The current problem with software is that when it fails it fails usually catastrophically, what do you do when you fail, stop that could be dangerous, keep going dangerous too. People don't usually fail as completely as software they make lots of small mistakes but usually do a good enough job.

the road toll is 14.9 per 100,000 per year that is quite low considering how much people drive, you would need a lot of testing, in real life scenarios to convince me that automated car was safer. And each release would need that level of testing. Yes you may get one driver who is bad kill a few people but a bad software release could kill much more.

I am not saying automated cars aren't safer, just that just because they are automated doesn't automatically make them so.

Re:Government doesn't have records

[phantomfive]

Given that safety is one of the governments biggest reasons for existing

Wow. I'm not sure that's actually true........

Re:Government doesn't have records

[Noah Haders]

Power is the reason why govts exist. Their. Continuing mission is to consolidate more power.

Re:Government doesn't have records

[Fwipp]

Sure it is. Government exists (partly) to help secure our right to life, by protecting from threats both internal and external.

Re:Government doesn't have records

[phantomfive]

Government exists to execute the combined will of members of society.

Re:Government doesn't have records

[Anonymous Coward]

"People don't usually fail as completely as software they make lots of small mistakes but usually do a good enough job."

Stroke, falling asleep and many other things cause human drivers to fail just as spectacularly as a hung computer would.

Heh

[Falos]

> the Apple Newton of Automobiles
There's not many websites you could really drop that line on, if you think about it.

It like "this latest iteration was the SimCity2013 of coffee-makers"

Re:Heh

[Anonymous Coward]

Doesn't Nabisco make Apple Newtowns? Delicious with a cold glass of milk.

Cars will be the secondary market

[Bryan Bytehead]

The first vehicle with this technology is not going to be a personal car, or anything that resembles a personal car (like a taxi). It's going to be semi trucks with trailers.

From a conference I sat in on last week (dealing with railroads, not trucks themselves), the turnover rate for truck drivers is over 100% per year. This is considered a plus for the railroads. I say that this is a plus for autonomous trucks. They drive autonomously site to site, and then, a driver takes over to get them parked into the loading dock (most likely), the trucks manage to do this autonomously (maybe, but not the scenario I see winning out, not at the beginning), or the docks are redesigned to make it easy for the autonomous trucks to park them in loading position (what will happen once autonomous trucks are widely used).

Yes, I realize other changes will have to be made. Refueling will have to be done manually in the beginning. That may mean the truck stop hires a person or two, that then takes care of the autonomous trucks, and I'm sure the owners will gladly pay a bit of a premium to get their trucks fueled. At least until the automated fuel pumps for the trucks are in place, at existing or new truck stops.

I have zero doubt that my great grandchildren won't have to learn how to drive a vehicle. I have grandchildren, and yes, I expect that they will have to learn how to drive, the technology is moving that fast.

Yet.

Re:Cars will be the secondary market

[Anonymous Coward]

I don't remember the title, but there's an older research paper that used a neural network to train an AI to park a semi-truck. It worked extremely well, for the simulated environment. I don't remember if they put it on a real truck or kept it within a physics simulation.

We already have buyable, self-parking cars.

Re:Cars will be the secondary market

[Osgeld]

over 100% turnover per year?

think they are selling you something...

for instance there is a driver in my family, been driving for a few decades and at least 1 decade for the same company

Re:Cars will be the secondary market

[hedleyroos]

Sounds like a dream scenario for robbers. The truck will drive defensively and will stop when blocked by another car. It won't evade. So now we need a terminator on each truck.

Re:Cars will be the secondary market

[Neil Boekend]

Sounds like a nightmare scenario for robbers: any decent software will send all data to the owner (and police) in such an event. Data like images of the robbers, location, images of the cars (with license plates) and which way they went.

Most truck thefts in the Netherlands are done on truck stops. When the driver sleeps. Guess what? An SDC doesn't do that.

Re:Cars will be the secondary market

[N1AK]

I'm really not sure why you think the anecdotal evidence of one driver someone defines an entire industry. I'd also like to see something backing up the over 100% figure but I expect that driving is quite a high turnover role.

Re:Cars will be the secondary market

[Anonymous Coward]

just do it in a LTE / 4G dead zone

Re:Cars will be the secondary market

[MMC Monster]

Just remember to drop an EMP so that you fry all the digital video kept on the truck.

Re:Cars will be the secondary market

[darkmeridian]

You know, you're absolutely right. From a liability standpoint, I don't think that trucks will go fully automated without a man in the cockpit. But certainly this technology would make a lot of sense for long haul truckers. These guys have very well-plotted routes, so they can chose major thoroughfares that have been carefully mapped by Google. For most of the "routine driving" the trucks can drive themselves. The driver can take control in bad weather or other emergencies. I'm sure the computers can detect when they cannot see the lanes or have some other issue pop up and then they can tell the driver to take over. So the driver can relax a bit while the truck is taking care of things on easy driving, and the driver can be more alert when things are dodgier.

Re:Cars will be the secondary market

[0123456]

How's video going to help when the robbers are, you know, wearing masks and stuff?

The best idea

[able1234au]

Is it start with a limited environment. Have these operate in large resorts, amusement parks, wildlife reserves etc. Build a big base of realworld usage before venturing further afield.

Humans have rules for driving

[Beeftopia]

Humans have rules for driving. For example:

-> If you see a traffic light, identify what color it is, then continue, slow down, or stop based on one of those 3 colors.

So the Google Car cannot identify a traffic light? Or if it does, it cannot identify its color? If so, is that a weakness in the computing power? Like, a supercomputer could do these things, but a reasonably sized onboard computer cannot? Or a weakness in "vision" sensors?

-> Paper versus rock in the road: This, I can understand. There are a myriad things in the road. The decision here is, can the car safely pass over it? Inability to determine this is due to vision sensors or limitations in computing power?

I saw an interesting problem the other day: a piece of wood baseboard trim (for a wall) blew off a truck. It seemingly hung suspended in air then came down. I hit my brakes but kept going straight, hoping for the best. It hit the ground, bounced and lay flat. I imagine that might legitimately freak out an autonomous car.

A moron can drive safely, through city traffic, if he's highly motivated, manages to keep his attention on the road and his speed down. I guess a moron is still more capable of navigating the world than a computer.

Re:Humans have rules for driving

[Anonymous Coward]

I think the hard part isn't identifying the traffic light, but figuring out what it means. Is it for my lane? Is there a sign in English next to it indicating whether to pay attention to it (e.g. "No turn school days 8am to 4pm")? Is there a policeman directing traffic, superseding the traffic light?

dom

Another stupid viewpoint from slate that is

[Beck_Neard]

almost genius in its idiocy. If self-driving cars really start to hit the roads, cities would definitely mandate that all traffic lights show up in maps, and require that traffic lights show up in maps before being installed. This is not a problem of the driving car, it's a problem of trying to imagine future technology in a current context, which is of course always going to lead you astray.

Plus, as other commenters have said, self-driving cars can definitely recognize traffic lights. It's just that right now they aren't quite as good at doing that as humans are. The reason is that traffic lights and construction cones and stuff like that are optimized for human visibility, not robot visibility. It's quite trivial to adapt them for robot visibility as well (perhaps even incorporating stuff like specialized radio signals).

I predict that horseless carriages will never take off because without an animal like a horse with hooves on the ground, you could hit rocks and fall into ditches without knowing it.

Re:Another stupid viewpoint from slate that is

[Anonymous Coward]

traffic lights and construction cones and stuff like that are optimized for human visibility, not robot visibility.

Exactly. It's not that hard to imagine a future where traffic 'lights' are even lights at all.

Re:Another stupid viewpoint from slate that is

[jklovanc]

I predict that horseless carriages will never take off because without an animal like a horse with hooves on the ground, you could hit rocks and fall into ditches without knowing it.

A horseless carriage still has a human driver with much more intelligence than a computer. Computers do not yet have the AI to recognize most objects and figure out what to do about them.

Re:Another stupid viewpoint from slate that is

[complete loony]

It's quite trivial to adapt them for robot visibility as well (perhaps even incorporating stuff like specialized radio signals).

Or blink a bright IR diode... In the short term the cars will need to learn how humans do it. In the longer term the cars may have their own information channel to augment how we currently do it.

Re:Another stupid viewpoint from slate that is

[drinkypoo]

If self-driving cars really start to hit the roads, cities would definitely mandate that all traffic lights show up in maps,

So what? they already mandate that the maps be correct, but roads are often shown going through (or not going through) when they don't (or do.)

Re:Another stupid viewpoint from slate that is

[Anonymous Coward]

And you are assuming that every City, County, property owner, etc wants to and will outfit roads with sensors and markers and update a database, all for Google's benefit. The problem with your "insightful" solution is that if the car must rely on a third party source for its information, there is a huge risk that the third party will not have provided it at all or has provided incorrect information. I can imagine countless scenarios, mostly emergencies, where updating a traffic light database is not at the top of everyone's mind.

I know we have become obsessed with marking and tracking everything digitally but to base life and death on how well physical objects are represented in a database is crazy. Why not just LET THE PHYSICAL WORLD BE THE DATABASE. The cars need to be able to drive themselves without this data and it was incredibly short sighted of Google to base their car on scanning everything. Seems to be like they want people to talk a lot about their "working" car rather than have a car that doesn't work but at least is making progress toward a workable solution.

Re:Another stupid viewpoint from slate that is

[Beck_Neard]

> The cost is too great to do it for the small few. So fuck them. Basically that's what it amounts too. And rightfully so.

True. Electric charging stations are insanely expensive and take up space. But adding an IR or radio transmitter (for example) to an existing traffic light would cost like $2. Compared to the cost of installing a traffic light, that's nothing.

> And even if the city does change, the rural world won't. Rural towns don't have the money to paint some of the roads.

True, but then, rural towns don't have nearly as much traffic. So it seems there's far less danger of a self-driving car causing a serious accident in your average rural setting, unless you have dangerous situations like mountain passes without guard rails.

Re:Another stupid viewpoint from slate that is

[Beck_Neard]

> And you are assuming that every City, County, property owner, etc wants to and will outfit roads with sensors and markers and update a database, all for Google's benefit.

Do you know how much it costs to install a traffic light? It would not be demanding much to also type in a few numbers in a database. And it's not 'for Google's benefit', it's for the benefit of those who are riding in Google's cars.

> The problem with your "insightful" solution is that if the car must rely on a third party source for its information, there is a huge risk that the third party will not have provided it at all or has provided incorrect information.

True, but human drivers rely on third-party sources for information. When you're driving you rely on numerous signs and markings to give you information about the driving situation ahead. You also rely on other drivers to give you information about their intentions and how you can react. A database is only one part of the equation. Another part - which I mentioned in my comment - is making existing traffic markers more accessible to machines.

> Why not just LET THE PHYSICAL WORLD BE THE DATABASE.

As opposed to what? The spiritual world?

> The cars need to be able to drive themselves without this data and it was incredibly short sighted of Google to base their car on scanning everything.

But it doesn't. It uses machine vision for a huge number of things. It's not just LIDAR. That would not work.

Re:Another stupid viewpoint from slate that is

[Beck_Neard]

s/LIDAR/maps and LIDAR.

Finally, some sanity

[Anonymous Coward]

I'm shocked it's taken this long for people to start questioning the general applicability of this technology. I challenge anyone here who drives on US public roads to imagine what it would take to create a car that can operate truly autonomously, in different weather and traffic conditions, while surrounded by pedestrians plus cars driven by (often borderline insane or highly distracted) human beings.

I think the best we can hope for is very limited use of autonomous vehicles, as in special, highly instrumented auto-only highway lanes, with humans being required to take over on all other roads.

Re:Finally, some sanity

[Teresita]

Tell the car you wanna see the Gateway arch, and it takes you to EAST St. Louis just like the Grimwalds...

Re:Finally, some sanity

[CanHasDIY]

the view of the gateway arch is much better across the river at certain spots in east st. louis than it is right at the base of the monument...

I'm not sure the risk of being mugged, raped, and/or murdered is really worth the view. There is never any reason to intentionally enter East STL... unless you're in the market for a 14-year-old prostitute.

Re:WHY ALWAYS APPLE?

[Anonymous Coward]

You should check your analogies before accusing people of being retarded.

Hint: Apple Newton was a failed product.

Some issues not mentioned in the summary

[phantomfive]

Here is a list of issues mentioned in the article, but not in the summary:

The Google car doesn’t know much about parking: It can’t currently find a space in a supermarket lot or multilevel garage.
It can't consistently handle coned-off road construction sites, and
its video cameras can sometimes be blinded by the sun when trying to detect the color of a traffic signal.
Because it can't tell the difference between a big rock and a crumbled-up piece of newspaper, it will try to drive around both if it encounters either sitting in the middle of the road.

Use a little imagination and you can surely think of other issues.

Re:Some issues not mentioned in the summary

[windwalkr]

For added amusement: a rock inside a crumpled-up piece of newspaper. They might actually fair better than a human on that test.

Re:Some issues not mentioned in the summary

[Anonymous Coward]

its video cameras can sometimes be blinded by the sun when trying to detect the color of a traffic signal.

I have a very similar problem.

Re:Some issues not mentioned in the summary

[jklovanc]

Right now objects are just blobs of pixels. The way the current car tells the difference between a postbox and a person is that the route has been pre scanned and gone over by a person to identify all post boxes. What happens when a post box is installed after the scan is done? The Google car will assume it is a person.

The second issue is that Google car is very good at not running into a person that is moving but not so good a yielding to a person waiting to cross. Say you have a person standing on a corner waiting to cross. Since Google car can not identify what that person is doing it may not stop and yield as required by law.

Pre-mapped environments are a dead end

[fyngyrz]

The only way a car can be designed to safely self-drive is doing it just the way we do: by creating a local, up-to-date mapping of the surrounding area in real time and working within that representation with sufficient skill to respond to anything that might appear.

Pre-existing environmental mapping simply cannot keep up. Construction, pets crossing the road, wild animals, falling rocks, pedestrians, vandalism of road signs and traffic indicators and lane painting, washouts, drunks, heart attacks, stinging insects, oversize loads swinging around traffic lights and signs, special transports, some guy at the side of the road madly waving a hand-printed sign that says "BRIDGE IS OUT!"... the list of unpredictable effects upon the local driving environment seems almost endless -- and keep in mind these things can occur in combinations of more than one type and more than one incident. Often suddenly.

Further, if the car is smart enough to be capable of updating the environmental map in real time and deal with any combination of changes, then it's already smart enough to maintain a completely dynamic local mapping and doesn't need a pre-existing mapping for anything but gross navigational purposes (route planning) and even that can require the vehicle to adapt.

Contrariwise, if it isn't smart enough to maintain a full local environmental mapping, then it is inherently unsafe.

Someone(s) at Google didn't think this one through.

Re:Pre-mapped environments are a dead end

[mjwx]

Someone(s) at Google didn't think this one through.

I think quite a few people at Google have thought about that, came to the same conclusion as you and started working on the problem.

The thing that people dont get is that it will take years, if not decades to get fully autonomous cars onto the road. They aren't due out in 2018 and yes we know what models are coming in 2018, an updated 370z, a new NSX and a few others no-one has any interest in.

The first autonomous cars wont be by Google, in fact I doubt there will be a Google car, the first autonomous cars will be Merc's or Toyotas built using Googles technologies and the autonomous part will only work on specially outfitted roads (and they will be controlled, limited access roads at first) so you'll still be required to drive a car. In fact you probably wont see a car without a steering wheel or other controls in your lifetime.

You're quite right that roads will need to be upgraded to provide telemetry to autonomous cars, and this will happen gradually over many, many decades.

Re:Pre-mapped environments are a dead end

[fyngyrz]

You're quite right that roads will need to be upgraded to provide telemetry to autonomous cars, and this will happen gradually over many, many decades.

Didn't say that, didn't mean to imply it, and don't believe it. Other than that, no. :)

Re:Pre-mapped environments are a dead end

[FlyHelicopters]

some guy at the side of the road madly waving a hand-printed sign that says "BRIDGE IS OUT!"...

The irony is that is the sort of example that people give, but happens so rarely as to be almost not worth bringing up.

The fact is, self-driving cars don't have to be perfect. They only have to be better than humans. Since we kill 40,000 people on the roads every year, give or take, and injure many thousands (if not millions) more...

The truth is, they are already probably "good enough".

Re:Pre-mapped environments are a dead end

[Anonymous Coward]

The first autonomous cars wont be by Google, in fact I doubt there will be a Google car, the first autonomous cars will be Merc's or Toyotas built using Googles technologies

Audi seems pretty committed to driverless. Their car's autonomous ability looks to be behind Google's, but they're gaining good real-world experience.

For our money, the real magic is in having your driverless sports car lap a race track, at speed, without piling into the barriers. That’s just what Audi has accomplished this week, sending 'Bobby', a heavily customised RS 7, around Germany’s Hockenheimring in just over two minutes. That's about as quick as a racing driver could accomplish in the same car, Audi reckons.

http://tvnz.co.nz/world-news/s...

Re:Pre-mapped environments are a dead end

[Anonymous Coward]

Much older ones have done it without telemetry http://people.idsia.ch/~juergen/robotcars.html

Re:Pre-mapped environments are a dead end

[N1AK]

This. The biggest obstacles left for autonmous cars isn't technical they're legal and ethical. Building a car that gets a new traffic light right 99% of the time is probably trivial, but would putting a car on the road that will get it wrong 1% of the time be ethical and/or a legal liability? A previous example of a google car failing was that it slowed to a virtual stop when passing people walking along the side of the road; that is seen as a problem because the normal human behaviour is to drive past within a few feet of the person at a speed that could easily kill them if they moved in the way. The car is being more cautious, in order to avoid a genuine life threatening risk, but we see that as 'failing'.

Are there technical hurdles left? Obviously, but they aren't the hardest things to resolve.

Re:Pre-mapped environments are a dead end

[bouldin]

The law still hasn't come to terms with product liability for software. That's a huge hurdle that will need to be crossed before we can trust software with life-or-death situations.

Re:Pre-mapped environments are a dead end

[ibwolf]

Building a car that gets a new traffic light right 99% of the time is probably trivial

Maybe. But considering that I go through about 15 traffic lights on my way to work (and then the same 15 again on my way home), a car that handled them correctly 99% of the time would have about a 1 in 3 of messing at least one of them up EACH DAY.

We're looking for five nines here, minimum.

Re:Pre-mapped environments are a dead end

[damnbunni]

You go through fifteen NEW, PREVIOUSLY UNKNOWN traffic lights every day?

I call bullshit. I don't think most city DOTs could manage to hang fifteen traffic lights a day.

That's what we're talking about here. NEW lights. Not existing, mapped lights. Those aren't a problem.

Re:Pre-mapped environments are a dead end

[Anonymous Coward]

Did you even read the article. These cars can't even MOVE in 99% of the USA and you say that is "good enough"? Have fun driving in circles around google headquarters.

Re:Pre-mapped environments are a dead end

[0123456]

"For our money, the real magic is in having your driverless sports car lap a race track, at speed, without piling into the barriers."

Whereas anyone who's actually thought about the problems knows that's piss-easy compared to driving down an urban street in a snow storm.

Re:Pre-mapped environments are a dead end

[Lumpy]

Programming a race car to drive a perfectly mapped track is brain dead easy for even a 2nd year CS student.

Let cattle roam the track randomly, and the car drives at racing speed avoiding cows, goats, etc randomly darting in front of the car, then I'll be impressed.

Re:Pre-mapped environments are a dead end

[Anonymous Coward]

Building a car that gets a new traffic light right 99% of the time is probably trivial

Maybe. But considering that I go through about 15 traffic lights on my way to work (and then the same 15 again on my way home), a car that handled them correctly 99% of the time would have about a 1 in 3 of messing at least one of them up EACH DAY.

We're looking for five nines here, minimum.

In the scenario you described you don't hit any NEW traffic lights on your way to and from work. So you still have a 1% chance of messing up a new traffic light when you encounter it, which is rare. Not 1 in 3 each day.

An insurance based licensing scheme

[Anonymous Coward]

Unless anyone ever gets around to reading this blogpost from 2012 that covers this exact use case :

The one piece of the puzzle that's missing is the state. Suppose that a speed limit on a particular stretch of road is changed, but the software developers aren't notified in a timely manner. In this case, the state itself has been negligent, and it's the state itself which should be fined for putting motorists at risk. In the same way that the state must adequately signpost the speed limit, so should be it's responsibility to notify the state licensed self-drive software developers.

I've used speeding as an example of unsafe vehicle behavior, but this regulatory framework extends in a natural way to all vehicle behaviors - stop signs, following distances, red light rules, yielding to buses on residential roads.

From http://missingbytes.blogspot.com/2012/12/self-drive-engage.html

Silly expectations

[NitWit005]

The expectation of this article is that Google will somehow shortly produce a car which will completely replace drivers in all circumstances. Clearly, that's the eventual goal, but that's not needed to produce something useful. Car companies are already churning out various incomplete solutions that help with highway driving or parking.

I expect their initial product to be something that works as a taxi in semi-controlled circumstances, or something that makes driving more convenient, but which requires intervention some of the time. Either of which would be a viable product.

Early cell phones were overpriced bricks, but they were still useful to some people. It took a huge investment from many companies and quite a bit of time to get to the point where people considered dropping their land lines. Replacing the old generation of technology is not usually a sudden process, but involves a lot of gradual improvement.

Re:Silly expectations

[Anonymous Coward]

I fully expect driverless cars to replace long haul drivers first. Then short haul, then taxis, then everyone else. Each one of those steps will come with a long list of restrictions. But as you move to the next step those restrictions get less and less.

Re:Silly expectations

[jklovanc]

Early cell phones were overpriced bricks.

That is a very bad comparison. The reason cell phones replaced land lines is that cell phones became smaller and had longer battery life. Cell phones do not need artificial intelligence to recognize the objects on or near a road. For example, someone standing on a corner with a crosswalk going in both directions. Are they waiting for someone? Are they waiting for you to yield? Are they waiting for other cars to yield in other direction? A human can figure that out by watching what the person is doing. Computers are not so good at that.

Re:Silly expectations

[Nikademus]

I fully expect driverless cars to replace long haul drivers first. Then short haul, then taxis, then everyone else. Each one of those steps will come with a long list of restrictions. But as you move to the next step those restrictions get less and less.

So you expect the heaviest things, those which take the most space on the road, those which can do the most damage if any error occurs to be replaced first?

Re:Silly expectations

[Anonymous Coward]

Actually, yes. The thing about trucks is that can be set up to only follow routes that are "known good" and these routes could be run periodically by a real driver to check for potential issues with the route. Sensors could be placed on the routes. Municipalities could be informed of the exact routes and agree to notify a contact person if they plan on changing anything on the route. Lots of the driving could be set up to be programmed more like current factory robots and less like what most of think of as an intelligent self diving car. In other words the job the self driving car has to do has been vastly reduced and thus the solution can be more customized rather than general.

You can start with one route and it can still make sense and be useful to the business.

Contrast this with virtually unlimited requests from people drivers. This eliminates a huge number of requirements and issues of a full self driving car.

laws

[Anonymous Coward]

how hard would it be to pass a law saying in states where driverless cars are legal, new street lights must be mapped before they are turned on?

um....

[Charliemopps]

Will the Google Car Turn Out To Be the Apple Newton of Automobiles?

seriously?

First PDA: http://en.wikipedia.org/wiki/P...

Are we going to, yet again, perpetuate the myth that Apple has ever invented anything on their own?

First Personal computer: http://en.wikipedia.org/wiki/C...
First MP3 Player: http://www.ideo.com/work/mobil...
First SmartPhone: http://en.wikipedia.org/wiki/I...

Re:um....

[unimacs]

The Newton is not noteworthy for being a pda, it's noteworthy for being a flop. Most notably the handwriting recognition. It was impressive technology at the time, but would often make glaring mistakes rendering it not so great in real world use.

Next Year ! Really !

[speedlaw]

Much like the flying car, it is possible, and coming next year.....

The power of exponentials!?!?!?

[alex_guy_CA]

RE: "who does not expect a full self-driving car in his lifetime (he’s 49)." Provided he's not suicidal, here is a man who does not understand the potential growth from exponential progress.

Re:The power of exponentials!?!?!?

[Anonymous Coward]

Overclocking a car does not make its software smarter.

Testa opensource

[Anonymous Coward]

Or will it use testa's Open Source and become the first super car?

Re:Testa opensource

[Anonymous Coward]

...Testa...

Nope

So much failure of imagination

[Anonymous Coward]

As mentinoed by others the author is imagining a future that is hard whereas things will be much easier. The new stoplight could be detected by:
1) Any other human driven vehicle that stops at it
2) Updating the database, done by:
  2a) Traffic light installer
  2b) Satellite
  2c) Anyone and everyone who notices, armed with their handheld computer
  2d) Any passing auto while light is being installed and notices the construction, marking the area for caution and further investigation by 2bcd
3) Redesign the stop light be more easily detectable by autoautos
  3a) in a way humans can't detect (wifi, infra, etc)
  3b) in a way humans can detect (ex: moving light that circles the active light)
4) "Watchout there is a new stoplight!" screamed by the passenger

Furthermore, how many distracted humans will run the new light? If the autos are working correctly, once one figures it out the rest should be updated or at least approach with caution and make their own decision, reinforcing the consensus.

Re:So much failure of imagination

[K. S. Kyosuke]

If the car has a lidar, plus a visible spectrum camera, then noticing that there's an unexpected stick with a bright colored light on top shining at you shouldn't be extraordinarily difficult.

Where is the real info on Google Car?

[aberglas]

Certainly the TFA is junk, and recognizing a traffic light would be relatively easy to do. But why is there no real information on the Google Car that is public? All I ever find is vague marketing blurbs and misleading statistics.

It would be very interesting to know what it really can and cannot do. And how the software was put together. Do they build a full scene graph or just use 2 1/2 D modelling? How do they go about the planning issues? Etc.

But nothing. Just marketing hype. So TFA is actually good if it flushes out some real info.

I wonder what the iCar would be then?

[Casandro]

Comparing the Apple Newton and the iPhone... I'd say that the iCar would be a car that's well connected and essentially controlled by Apple. It would not work on roads not approved by Apple. It would probably be controlled by a touch screen or voice. However it would not drive by itself, as that feature has been proven to be complicated. Of course it won't have a driving wheel, instead it'll have a software driving wheel on a large touch screen in front of you.

Functionality wise, the iPhone was a _huge_ step back from what the Newton could do.

Nevermind traffic lights, what about drunks?

[Anonymous Coward]

Last night after dark, I was driving past the Indian casino. Some stupid drunk wearing dark clothing crossed the road in front of me. The limit is 55, and that's about what I was doing (really, I don't speed around here). I broke HARD. There was no time to honk. I was really quite in a state of shock. I passed him at 20 mph or so, he was just trundling along holding his bottle in a paper bag like nothing had happened. His buddy was on the other side of the road. I didn't honk or anything. Now I tell people, "If you're passing the Robinson Rancheria in Nice, CA at night, drive 45 mph and use high beams".

Can Google car handle that?

Re:Nevermind traffic lights, what about drunks?

[Anonymous Coward]

That's called a road hazard and it could have been a rock or a deer. You're overdriving your lights. Use your high beams and slow down.

Re:Nevermind traffic lights, what about drunks?

[SmilingBoy]

Better than you as LIDAR works in darkness as well.

Why is this not "Ask Slashdot?"?

[Anonymous Coward]

If its not WHY THE FUCK is there a question mark there? Any article with a questionmark in the title should be thrown right in the garbage but they seem to become more and more frequent here at /..

Success in Stages

[biomass]

I for one am interested in a car that can self drive on freeways and would prefer to drive myself in town.
It seems that freeways are a simplified problem, being easier to map (and update), no stop lights, no cross streets, ...
Hopefully the freeway self driver will be a practical target for the near future.

Driving is filled with intractible problems

[DrXym]

Anyone who thinks self drive is coming to a vehicle near them soon is living in cloud cuckoo land.

Self drive cars might work on a closed track where the number of external factors are limited and can be controlled. e.g. an airport loop, or a theme park transfer. It might even work on some stretches of public road e.g. some motorways although it is more likely to be an advanced driver assist mode.

It sure as hell wouldn't work in urban settings, or for atypical conditions. It's trivial to think of scenarios that would boggle the mind of a computer and cause it to stop for no good reason, or get itself stuck, or do entirely the wrong thing. e.g. in following a traffic cop's directions. At the very least such vehicles would have to have a conscious, unimpaired driver at the wheel ready to take over at a moment's notice and chances are that self drive would suck so hard that most people leave it turned off or in some reduced mode such as hazard / collision detection, cruise control etc.

Re:Driving is filled with intractible problems

[itsdapead]

Anyone who thinks self drive is coming to a vehicle near them soon is living in cloud cuckoo land.

I wasn't aware that Google had promised to have self-driving cars in the shops for Christmas. You have to start somewhere - and since any large-scale adoption of self-driving cars is going to require the cooperation of government, that means starting the PR campaign early, not just the R&D.

I think its great that some of the billions made from the internet boom are going into blue-sky projects like self-driving cars, electric cars and space travel. Will Google be selling a viable self-driving private car before Tesla are selling a 300-mile-range compact for the price of a mid-range gas burner, or will we still be waiting when SpaceX gets to Mars? It'll be fun finding out.

Self drive cars might work on a closed track where the number of external factors are limited and can be controlled. e.g. an airport loop, or a theme park transfer.

...large business/university campuses, shopping malls, an alternative to trams/personal mass transit systems that vastly reduces the amount of "civil engineering" needed... Sounds like a business opportunity for Google to me.

Re:Driving is filled with intractible problems

[Overzeetop]

“As it turns out, what looks chaotic and random on a city street to the human eye is actually fairly predictable to a computer. As we’ve encountered thousands of different situations, we’ve built software models of what to expect, from the likely (a car stopping at a red light) to the unlikely (blowing through it).

“We still have lots of problems to solve, including teaching the car to drive more streets in Mountain View before we tackle another town, but thousands of situations on city streets that would have stumped us two years ago can now be navigated autonomously,” Urmson writes. (Chris Urmson is head of the self-driving car project at Google)

Smarter people than you have been working on these problems for years already and have made significant progress. Other locations also have research going on. Virginia Tech, for example has a self-driving/autonomous vehicle program that is also working on navigation of complex environments. Hazard collision detection and autonomous steering and pacing is already in production vehicles (and has been for a couple of years).

The good thing about computers is that they can be programmed to fail gracefully, stopping when conditions do not meet the requirements for safe continuation. Unlike humans, who can't figure out when they're too drunk, tired, old, or distracted to drive safely. Everything will come in steps - collision avoidance assistance, then highway autonomy, then known-city autonomy, then full autonomy with driver, and finally full autonomy without driver (passengers w/o driving skills). You won't get that last phase in the next couple of years, but I anticipate it will happen before I'm too old to safely navigate the roads.

Re:Driving is filled with intractible problems

[DrXym]

Smarter people than you have been working on these problems for years already and have made significant progress.

Yes at doing basic navigation. There are far too many intractable problems remaining to think self drive is remotely useful on open roads. They will fail over to the driver or stop dead so often as to be annoying. I should also note your "smarter people" can't even do basic speech recognition on a phone with sufficient accuracy to make it work well. What makes you think a far more complex problem is somehow within their grasp?

The good thing about computers is that they can be programmed to fail gracefully

Which would be great if they only fail gracefully in a critical situation such as a potential collision. If they "fail gracefully" because they're confused by the plastic bag blowing across the street, or the lights being out, or by the large puddle ahead, or the cop telling them to proceed, or leaves / ice / snow obscuring a sensor, or by any car in front of my car which decides to fail gracefully then they'll suck.

It is not sufficient to fail to safe. Cars must make good progress too for the occupant and occupants of all the other cars behind. This is why all the puffery about how many miles Google cars have driven without accidents is only a fraction of the story. Yeah it's great they haven't killed anybody. It's a significant achievement. If such vehicles just stop for no reason at all (and they will) then they will suck and they will suck hard.

Will there be unmapped traffic lights then?

[cynop]

Well, it maybe hard for a machine to visually identify a traffic light, but that's hardly the only way. In the "Internet of Things" vision, traffic lights are one of the first things to be connected to the network for traffic shaping. Hence, autonomous networked cars will be actually aware of not only the closest traffic light, but of all traffic lights in a certain radius. Assuming John Leonard lives another 30 years, I find it hard to believe, that similar functionality won't be implemented by then.
Then again, one could argue that networked traffic lights won't span the globe, and thus autonomous cars will be bounded in certain geographical areas. That was true for a lot of modes of transportation originally though, and eventually it will be minimized.

Anecdotal data isn't

[Overzeetop]

100% turnover doesn't mean everyone quits, it means that for 1000 drivers average on the road, there are more than 1000 drivers who leave. If you have two people who only last 6 months each as a driver, that's 200% turnover.

Good / Bad Idea

[DrYak]

That's an idea which could be useful in theory.
(e.g.: Cars with drivers will still be able to display warning about red lights, speed limits, etc. based on the info broadcast by trafic signs)

But it has a few problems:

- The implementation will probably be botched. Expect the thing not being properly signed/authenticated, thus enabling malicious hackers to spoof information. (Similar to how hackers hijacked RDS-TMC and broadcast "bison crossing" in Germany a few year back on /. )

- Such system lacks a fail-safe option. A human might notice that a trafic light is off and will fall back to other driving behaviours. A robots might not realise that there is no emitting signal. (The robot can't see a missing emitter unlike a human who can notice a broken traffic light even without any light colour coming off). In some case it might be okay (missing traffic light: drivers are supposed to fall-back to priority-yield, which is probably the default behaviour of a robot when arriving at a crossing without signs), but it might be problematic in other case (a "danger ahead" sign with a broken emitter).

- Car insurance companies are going to abuse the shit out of this (cue in mandatory dongles that spy if you obey trafic signs. Of course driving dangerously and ignoring signs is bad. But violating privacy is bad too) At least european countries are a bit stricter regarding privacy.

Not really ...

[TrollstonButterbeans]

Most industries depend on just-in-time processes to minimize inventory.

Timing is critical and tens of millions of dollars can be stake easily if it doesn't work like a clock.

There isn't going to be the tolerance for stranded trucks with flat tires that can't be repaired without a human, or late trucks stuck in traffic.

And changing lanes in a truck in a traffic jam? Haha, yeah. And making turns in a truck in a tight intersection? Sure. And trucks are tall, there are bridges and all kinds of height obstacles out there requiring special routes.

Road construction? A minor inconvenience for a car can be a major dilemma for a truck --- things like roads with no shoulders, etc.

And single truck wreck would carry large costs, and would insurance companies covering the transportation companies put up with it?

Red flags

[allypally]

For safety, we should enact laws that require a robot with a red flag to walk in front of any driverless car.

There is precedence: http://www.greatachievements.o...

Re:Red flags

[CanHasDIY]

Back in that day, the state of Alabama (I think) enacted a law that required automobile drivers to stop prior to an intersection, get out, take a lantern to the intersection, and verbally announce that an automobile is about to cross the intersection before actually doing so.

^ Remembered from an old book of stupid laws I had as a kid.

Re:Red flags

[iggymanz]

That's pretty much the kind of job "women in construction" do, I see them on crews never handling the jack hammer or spreading the concrete, but waving flags and such.

Google code

[StripedCow]

Wouldn't it be cool if they (Google) wrote a simulation environment that emits the (simulated) outputs of a radar (LiDAR) and other sensors to an API. If they would open-source this environment, everybody could write car-driving code and test it in the safety of their home (it doesn't even need to be real-time). They could even use this environment in a Google Code session, for example, and see which participants obtain the best results.

self-driving will always be an issue

[linuxrunner]

Between road detours, new red lights, which red light is mine in some cities... but the biggest issues are for those of us who live off in the north.

We have add things to avoid. Deer, turkeys, and lots of pot holes. I will swerve to miss a critter, but not at the cost of my life.

As a driver I'm smart enough to avoid the slippery leaves that land on the ground in the fall. There are bridges I can't cross because my truck is too wide to do it at the same time as another oncoming car. I can glimpse the car coming through the trees so I just stop and wait.

Half our roads don't even have a double yellow line, never mind the white line. Hell, we have dirt roads. What is a car going to see then? I have areas (mountains) where even my satellite radio will go out.

Lastly winter. There are times with the new snow fall that I have no idea where the road is and have to drive in the middle of the road. Not because that's where the lines might be but staying away from the edge is safer, driving in the lane where someone else has driven is safer, etc.

Lastly, for Google to originally think they weren't going to put a steering wheel in the car just shows their stupidity. After kids put cones in the road, and paint some lines via detour and watch the cars line up down a dead end street or something like that. It won't end well.

Another stupid viewpoint from slate that is

[linuxrunner]

Maybe. Would cities start? Sure. But how long would it take and at what cost? We've had electric cars for a long time now, but no charging stations. The cost is too great to do it for the small few. So fuck them. Basically that's what it amounts too. And rightfully so.

Same with cities and streetlights. It's easy to say they'll change them. They may... over time. But those first self-driving cars will find themselves in horrible fucked up situations. Then what? And of course those bad situations will make it harder to sell the car. Less cars, less incentive to change things over.

And even if the city does change, the rural world won't. Rural towns don't have the money to paint some of the roads. (My town doesn't even have a single stop light actually). Our roads aren't really wide enough for two cars in some places so it takes smart driving skills to know when to move over and off the road, the roads are not painted (no yellow line, never mind a white one) dirt roads, pot holes, snow cover where you can't see the road, and mountain ranges that will even stop a satellite signal.

The point is, there are a lot of situations a self-driving car just won't work so not everyone is going to get one, which decreases the incentive to change things over and incur a large cost for the benefit of a few.

It doesn't need to be full speed

[Anonymous Coward]

I think I'd prefer the idea that self driving cars "at normal speeds" are not gong to happen in my lifetime. I can easily conceive of something that drives at 20mph and has radar for the immediate vicinity in front. Sounds achievable to me, and would give older / disabled people the independence to travel to places in their town. The use case I want is my 200 mile weekly commute to be done mostly overnight while I sleep on a Sunday night. I don't care if it does most of it at 20mph in a special lane on the motorway and may have to wake me if it simply can't handle something. Roads will be quieter and predictable. Whatever isn't mapped I'll teach it myself. It'll be economical and may arrange to join up with other vehicles to do some of it in convoy. It will get to the city limits of my destination before rush hour and drive very gently to a place where it can park up and inductively recharge. On designated "fully mapped" freeways it will be able to go faster. Maybe getting me from Calais to the heart of the french alps while I sleep following a Friday evening chunnel crossing from the UK. Again, there'll be enough of us to go in convoy, so it will be something like a long electric train. I also expect 1 way rentals to crawl back to base by themselves.

Never in North Carolina

[gelfling]

Maybe other states as well but I know it will never happen here. Why? Because the state has their hand in your pocket as part of the 'benefit' of the state owned reinsurance pool. This gives them the 'right' to unilaterally change your insurance coverage, drop you or raise your rates to whatever they want, apart from what the insurance company wants. In fact the state of NC can drop you on their own w/o telling the insurance company and then if you get a ticket? Oh well you're driving uninsured. So if you think that driverless cars are ever going to be allowed here where their precious power could be put at risk you're wrong. If anything, the state will simply decree that driverless cars need super mega extra through the sky insurance rates. Because that's what they can do.

Like the first product in a class in the hands of

[iamacat]

Sounds good! Newton was commercially available, has a loyal fan base and inspired successive generations of more polished and popular products, including Palm and Apple's own iPhone. True, there is no guarantee that just because you release an early adopter product, you will reap most of the benefits when technology matures. But not being on a lookout for new things guarantees slide into irrelevance, like Kodak or Borders. Besides someone got to do it.

Recognizing red lights

[Cro Magnon]

The other day, I was driving down the road, and I thought I saw a red light coming up on the left lane, but the light on the right lane was green. When I got closer, I saw that the red light was really on the cross street, but the light had been knocked askew so it was visible on my street. Could a self-driving car deal with that problem? And if the light emitted some signal to the car, what happens when (not if) that doesn't work?

eat up martha

[Anonymous Coward]

eat up martha

re: trains vs. trucks

[King_TJ]

I currently live in an old railroad town in the Northeastern U.S. Our rail system is still alive and well out in this part of the U.S. -- despite the appearance of being a dinosaur in many other parts of the country. (I even take commuter rail in to work each day ... and yes, it's a full size locomotive with multiple passenger cars, including a couple of double-decker cars.)

I don't see how employee turnover has a lot to do with trains OR trucks winning the battle of who gets used to haul freight around? The real bottom line is going to be economics and efficiency. The big advantage I see trains having right now is better efficiency, in the sense that today's locomotives are pretty energy efficient. Many of the new ones have solar panels on top of them to augment power generation, and they move a MUCH larger volume of product around than a truck can. (In the case of passenger rail alone? Look how many hundreds of vehicles are taken off the roadways each weekday just from all of us who use it to commute instead of driving ... and that's just ONE of several rail lines out here that run each day.)

They also have an advantage in the fact that they don't have to deal with traffic congestion. The established railways are generally about the fastest way from point to point, so they can generally predict down the minute how long it will take to arrive at a particular stop.

IMO, there's a lot of mismanagement with the rail system, which allowed the trucking industry to eat their lunch in many cases. But it didn't HAVE to be that way.

EG. I used to work for a steel fabrication place that had a railway running right outside their back door. Up through some time in the 1970's, they always used the rail system to ship steel beams to customers. But they started running into logistics problems where customers were only willing to buy from them if they could meet deadlines for "rush" deliveries (and would pay big premiums for this as well!). The railroads couldn't adapt to accommodate this, especially when they'd often have their own logistics battles to fight, trying to get certain cars unloaded on a train before others. (They said they'd often see the train they were waiting on to pick up a load chug right on by, once or even twice, during the day, before finally stopping for them -- all because the railroad wanted to unload something else first and potentially juggle the train cars around in a yard, before coming back for the pickup.) All of that convinced them to invest in a small fleet of trucks and do their own deliveries.

But in any case? I think autonomous trucks probably will arrive before autonomous passenger cars owned by individuals. (Commercial vehicles could absorb the initially higher cost to purchase them, for one thing.) But you'll probably see them limited to driving in a designated lane, at least initially. Doing this would make their operation a little more like what the railroads do now .... follow designated paths from place to place. I'm not sure how well that will work for them, if they STILL have to have a "short haul" truck pick up their loads at some point and take them to the loading docks at their destinations, using the regular road system?

How about a remotely driven car first!

[Wargames]

Would someone mind taking over the controls for a bit while I have me a nap?

Software failure

[Anonymous Coward]

As a fellow engineer liked to say: Software never fails. It always does exactly what it says it is doing.

Two thoughts

[soft_guy]

Why do we have to assume we can't make any changes to the nature of traffic lights to make them more easily machine readable? Since it requires a car with a lot of cameras to maintain the road information accurate enough for the self driving car, why would lots of cars with fewer cameras do an equally good job? For things like potholes, we can't self driving cars report information about how smooth a particular path along the road is so that pot holes are mapped?

They did sell Newtons

[whitroth]

Google car ain't going anywhere.

Last week, we had an IPv6 transition seminar where I work, and the keynote speaker was none other than Vint Cerf.

His title of "Internet evangelist" is right - sorta like, I dunno, Pat Robertson.

He told us that the next iteration of google car will *NOT* have a steering wheel, brake, or gas peddle. I quote, "you might be in the back, drinking, or doing crosswords, and so you won't have the context if an alarm goes off, and you'll do the wrong thing".

Now, when I drive home, one road goes from two narrow lanes, with the center line going away to no center line, cars are parked on one side (there's a park on the other), and busses use this route. Go ahead, tell me that anyone's software can handle that road... and we're in a old, big suburb of Washington, DC.

                mark "and I don't want *you* to have a flying car, either, since you'd crash into my second-floor bedroom"

Work Around

[Anonymous Coward]

Couldn't we just force the driver to manually drive the car on these rare occasions they are on an unmapped street?

Colleges Have No Faith In This

[CanHasDIY]

Every college I drive past these days seems to be starting up an OTR truck driver training facility. Seems they don't have a lot of faith in the alleged immediate inevitability of auto-cars.

Re: trains vs. trucks

[Bryan Bytehead]

I'm married to a fourth generation railroader. That is the only reason why I was at the presentation, being with my wife. ;)

I don't disagree about the economics and efficiency. I think the railroads have it in spades at this point. And yes, mismanagement has been a problem with railroads as well. John Snow was well criticized for his golden parachute when he took over the Treasury for Bush (II) here in CSX Town.

I do disagree about employee turnover. Some of those new drivers are not going to be as experienced as the drivers that they just replaced. I can't see a high turnover rate being anything but detrimental to trucking's efficiency.

Railroads have always been notoriously conservative, let alone risk averse. JIT manufacturing is not kind to railroads.

I do not see designated lanes for autonomous vehicles, commercial or personal. Our freeways are already bursting at the seams, and now when they talk about expanding lanes to handle traffic, they are moving to make those lanes toll lanes to pay for getting them built, certainly here in Florida. I don't see enough traffic from autonomous vehicles to justify that kind of expense. If anything, there is already an argument to be had to force semi trucks and trailers into their own lanes due to some of the accidents that have occurred. But no one actually expects that to happen. It truly becomes the chicken and the egg problem.

Railroads know they need to work better on their logistics. My wife works for a company that sells software to those railroads to do so. In some ways, I'm amazed that they aren't further along in their development, either the software company or the railroads themselves.

At age of three people have perfect sensors

[Anonymous Coward]

At age of three people have perfect sensors, fusion and interpretation of the world.
Still, we don't let people drive cars until they're 16 or 18 because they're note mature enough.
Can you reasonable claim that you can program a computer to have common sense
on par with a 16-18 year old? I'd sure like to see that program!

News today: UK wants driverless buses

[hazeii]

Interestingly, there's a report in the Telegraph today suggesting that driverless buses could be on the roads in the UK pretty soon.

On the one hand, this makes sense - the complexity of the problem is reduced with a vehicle following a pre-programmed route.

On the other hand, I'm deeply sceptical - taking the assumption that such vehicles would have to be super-safe to be accepted, I can see a spate of teens having fun baiting autobuses into emergency stops. Oh, and cyclists will totally rule the roads - get in front of a bus and pedal as slow as you like.

Bad configuration, construction workers

[Anonymous Coward]

Stop lights don't always line up with lanes exactly.

How about this light configuration which is confusing enough for normal drivers:

One way:
[_L_|_R_]
[red][green]
[^L^]/\[^R^]_______\ /__/========

Driver L on the Left sees a Red light for his lane on the Left and a Green light for the right lane on the right.
Driver L is coming out from the highway exit West-East and try to merge into this one-way dual street lane.

Driver R on the Right sees a Green light for his right lane and a Red light for the left lane.
Driver R is coming out of the boulevard on a side-ramp and lines up side-by-side with the highway exit, so not orthogonal to that other side-lane.

First time, you encounter this configuration, you might be like WTF?! and might be confused if the traffic light is broken and which one is correct,
until you see traffic flow from the highway exit on the LEFT lane and traffic flow stopped on the RIGHT lane or vice-versa.

Now try to figure out, what a computer would do, in this configuration.

Another typical one, the traffic light is either non-functioning, red or green and a policeman or a road construction worker is doing traffic control,
how do you deal with that one?

Programmed to fail?

[Ihlosi]

If the car needs to know in advance where each and every stoplight is, then the whole concept needs to go back to the drawing board.

The car needs to be able recognize stoplights (and traffic signs!), determine whether they apply to the car or not, and follow them, without having a map of every stoplight and traffic sign in the world.

A lot of humans will fail to recognize . . .

[Anonymous Coward]

a newly installed traffic light or stop sign or other change to a roadway. People have routes they drive routinely and that they have learned as a set route. Put a new feature like a stop sign in and a significant number of them will drive right through it the first time and maybe a few times before they learn it.

Further, Google is still working on the car, and is only one of many many organizations working on such things. TFA itself states that ten years ago no autonomous cars could even complete 8 miles of a test course. The progress being made is astounding, yet not instantaneous. I guess the thing that is most jarring is the statement that "it may never happen." Autonomous cars certainly will happen; just how quickly is the real question. One, driving is just not that hard, and two, humans are not that good at it day in and day out. There are certain tricky problems such as recognizing objects and possible hazards. But these are areas that computers have already demonstrated huge strides, such as face recognition technology which is frighteningly advanced. And also areas where humans are not really all that great either. The question, as many point out, is not whether automated cars will be nearly perfect at safety; it's whether they will be better than humans.