Self Driving Car thread

[Chuckstar]

Boneman[/url]":30imdc6k]
Have you never had a long highway commute? Are you just not getting it? How can you not see the utility of it? IT DOESN'T HAVE TO TAKE CARE OF EVERYTHING FOR YOU 100%. It just has to safely hand off control. "Alerting human to the issue and ensuring control is safely handed to human" is handling it. And they said they required the person to take control in those situations, but I don't recall it saying its because the system can't handle it. I'm sure they've got a lot of testing they want to do before they let loose their cars in high-risk situations like residential areas and on snow. Doesn't mean the car can't do it, it's just policy not to let it.

The article posted in the other thread said the guy's hands were on the wheel for only 14 minutes of his one hour commute. If I only had to be actively engaged for 25% of my two hour commute it would be a fucking godsend. Ok, sure I couldn't got into deep REM sleep in the current Google car iteration, but so long as the car is programmed to safely hand over control when it encounters a situation it can't handle, a little nap wouldn't hurt. hell even just being able to fart around on my phone would be pretty awesome. Which most people do on the road already.

I'm sure we're a long way from a car with a "layout in backseat and start counting sheep" mode, but for many people with a lot of highway driving on their commutes, the current version of Google Chauffeur would be simply amazing.

I have an hour-and-a-half each way mostly highway commute. No one, absolutely no one, wishes for a car that could take over the on-highway portion of the commute more than I do. I get it. But unless I can actually do something else (even if it's not napping), then it's a complete waste for me. I'm not interested in a situation where the car is driving but I still have to keep my eyes on the road looking for rare events.

And here's what you're ignoring: Right now these cars are only being driven by very responsible drivers who are (generally) staying alert to do things like watch for construction zones, take over in the rare emergency, and take over for the "hard" parts of the drive (i.e. surface streets, complicated interchange, not driving on snow, etc.) The question is, can you trust Joe Sixpack to be as responsible? If your answer is "no" (and mine certainly is), then they still need the car to be way smarter than it currently is before selling it to Joe Sixpack. Even Google admits as much. (It needs to be at least smart enough to know when it needs to stop and refuse to keep going unless the human takes over.)

And the way AI works, every bit smarter you need the machine to be is exponentially harder to implement than the last bit of smarts. (Maybe "exponentially" is not formally correct, but you get the point.) "Stay on this side of the yellow line" is way easier to implement than "read the orange construction signs and figure out where you're supposed to go". Hell, winding your way through a construction detour can be confusing even for humans, and it's not the kind of problem that computers do better than humans on. Even "if you see orange construction signs, make the human take over", is a difficult problem for computers, when you factor in the practical concept that it's useless if there's too many false positives.

 

[Chuckstar]

redleader[/url]":3g60k8s7]

Aatos[/url]":3g60k8s7]Wonder what the rain/snow limitations for microwave sensors are.

Rain/snow attenuate both optical and high frequency microwave beams through the same means, scattering. There are differences in how water absorbs microwave and optical frequencies, but they actually don't matter in this regime because the percentage of the air that is actually water during even a heavy rain storm is not that high. Thus absorption tends to be much less of a problem and can generally be neglected.

The types of scattering are actually somewhat different. Lidar is Mie scattered or even geometrically scattered because the rain drops are so large relative to the wavelength that they operate as miniature lenses/diffractive elements leading to forward peaked scattering. 60GHz is a 5 mm wavelength, which puts it in the rayleigh scattering range unless the rain droplets are very large. This will lead to more isotropic scattering.

In practice it doesn't really matter for coherent imaging systems like lidar/radar because they image based on time of flight. Any scattering, isotropic or forward peaked, will prevent accurate ranging if the scattering path length is much less than the imaging range. This is actually one of the advantages of time of flight based ranging as compared to other techniques.

Are there problems with the drops that hit the actual sensor? What I'm thinking of is that the raindrops on your glasses are much more of a problem for sight than the raindrops that simply fall between you and the object you're viewing. Of course, there are solutions for keep drops from forming on surfaces (wipers, RainX, etc.), but I'm as much curious as anything.

 

[Chuckstar]

LordFrith[/url]":3ig3zibp]

Peldor[/url]":3ig3zibp]

If an s-d car violates the law, then that's likely the fault of the manufacturer

Then you do see the problem very clearly. Currently the car manufactures have a very low chance of being found legally at fault in a car crash. It's almost always a driver error. When the car is the driver, now it's a design flaw.

But they seem to have been okay with adaptive cruise control, which easily could rear-end or kill someone pretty easily. My car, for example, allows cruise control down to stop-and-go traffic on city streets, meaning that someone could easily step in front of a seemingly stopped car, only to have it start up and hit them. So, manufacturers already cope with this potential liability.

No. Because with cruise control you are still in control of the car. Your hands are on the steering wheel and your foot is inches from the pedals and you're (supposed to be) paying just as much attention as if there was no cruise control. Not the case with a true self-driving car (at least not a self-driving car that's actually useful).

 

[Chuckstar]

LordFrith[/url]":2nvek8uj]Why would my foot be inches from the pedal? I keep it flat on the floor as I lean back and let the cruise handle everything.

And what is the difference if the manufacturer simply says, "This is only an aid. You were expected to be in control."

Your floor is more than a few inches from the pedals? What kind of car do you have?

 

[Chuckstar]

dio82[/url]":2qxm3z55]

ChrisG[/url]":2qxm3z55]

LordFrith[/url]":2qxm3z55]My point is that car manufacturers already deal with automated systems and driving.

Big difference between an automated system and an autonomous car, and this seems to be the second time you've obfuscated the discussion a bit by throwing in this notion of semi-autonomous driving.

As redleader said, no-one in this thread is talking about cars with systems which reduce a driver's mental load. A self-driving car is a vehicle which has driving controls you do not touch except in an emergency, or when prompted by the car's control system. You get in. You specify a destination. The vehicle performs all of the required driving actions and delivers you to the destination with no intervention.

No. This all or nothing stance ignores the enormous sliding scale between all manual driving and fully autonomous driving.

Hell, just even adaptive cruise control with lane keeping would be an enormous boon for most drivers.

And from a liability perspective the following are very different:

1) Driver is in control of the car, but there are assistive systems

2) Driver is not in control of the car, but can take over

In the case of #1, as long as the assistive systems do not do anything that interferes with the driver's ability to control the car, then the driver is ultimately responsible for anything that happens (i.e manufacturer has liability if the adaptive cruise control fails to disengage when the driver hits the brake, but is not responsible if adaptive cruise control fails to react to a pedestrian -- as that's what the driver is there for.)

In the case of #2, the vehicle's systems are entirely responsible for what happens, at least for what happens in the short period of time before even an attentive driver might be able to take over.

Yes, the manufacturers have gotten comfortable with the lesser liability inherent in #1. But that may be a very different analysis from the liability inherent in #2, where the occupant can honestly say "I wasn't even driving".

 

[Chuckstar]

LordFrith[/url]":2vq2fqcg]If the manufacturers are able to handle that level of lawsuits, they are obviously geared up for handling the autonomous driving lawsuits.

I fail to see why that's obvious.

 

[Chuckstar]

LordFrith[/url]":3o8smql1]

Chuckstar[/url]":3o8smql1]

LordFrith[/url]":3o8smql1]If the manufacturers are able to handle that level of lawsuits, they are obviously geared up for handling the autonomous driving lawsuits.

I fail to see why that's obvious.

but is not responsible if adaptive cruise control fails to react to a pedestrian -- as that's what the driver is there for

Look at that statement. Now, if a backup system that does automatic braking if there's a pedestrian fails -- that's exactly what it was there for. It's failure would be the same as if a seatbelt failed in a crash. Or an airbag failed to deploy in a crash. You can't say, "Oh, well, no driver should have relied on that system" when you are giving that system as a safety mechanism. So, yes, there absolutely would be a lawsuit when that system fails.

That's the obvious part.

That's not obvious at all. Because if you are doing a good job backing up the car, then the computer should never have to intervene. You can't hit the pedestrian unless you and the car both fail. In which case, the law is pretty clear who is responsible.

 

[Chuckstar]

Frennzy[/url]":3eg6q1y2]

Nonsense. Product liability is a well establish area of law even if you are not fully aware of it.

What are you talking about? There are well defined partial liability laws...i.e. a continuum of liability. When it comes to "you're either driving or you aren't" you're simply wrong. What if I'm working the pedals, and someone else is steering? A cruise control can fail in interesting ways...but I'm clearly not the one in charge of the throttle...if it fails, am I liable?

It depends on the manner of failure, really. There are very limited failure modes for cruise control which would result in an accident that the driver couldn't avert if he was paying attention.

The problem with the "continuum" theory as an argument against any different liability situation for the automakers is that there is no continuum in a truly self-driving car. It is the computer's fault.

Assisted systems are fundamentally different, as they are really backups for the driver. If the blindspot warning fails, for instance, and you hit another car... well, you're too lazy to check your blind spot yourself?

I guess there is the situation where no human could possibly have stopped fast enough, but the austomatic braking system could have reacted fast enough, but for some reason failed. So that the pedestrian would be alive if the computer worked right, but still dead if that car happened to not have such a system. I'm not sure how that would come out in court... but it seems like kind of an edge case.

 

[Chuckstar]

Dmytry[/url]":8qza6h3f]The liability issues add nothing whatsoever new into the mix, as there's already dozens subsystems that can fail resulting in an accident that is car maker's fault (e.g. if brakes fail, or if the accelerator glitches due to the tin whiskers, or faulty ABS prevents car from braking, or what ever else). I have no idea why they are even being brought up at all.

edit: I propose that if you have an argument which applies to anti locking brakes just as much as to self driving cars, don't argue it.

Are you really saying "I get the last word, no one else say anything?"

Besides the fact that it's simply an unlikely request to be followed, it's particularly rude.

 

[Chuckstar]

Dmytry[/url]":8c9cnxge]

wco81[/url]":8c9cnxge]What if auto-pilot is as cautious as an old lady?

You can be one irate backseat driver.

Well, I picture that the maximally cautious autopilot would plot all physically possible trajectories of the other cars and assume the worst, in which case it technically could still make a very scary looking left turn like a crazy cab driver out of hell. (It would have to keep certain time clearances and such to avoid scaring other drivers).

I doubt they do anything as intensive as plot all possible trajectories. Simple heuristics will suffice.

 

[Chuckstar]

Control Group[/url]":12td2i51]

ChrisG[/url]":12td2i51]

Frennzy[/url]":12td2i51]Of course there is.

Cruise control would be one example.

Guess again. You're still very much in control; cruise control merely lets you take your foot off a pedal. You're still steering, required to be prepared to brake or maneuver suddenly, or you may even be required to adjust the cruise control.

Ergo you're still actually driving.

Adaptive cruise control, automatic braking, lane keeping assist, blind spot alerting, automatic parallel parking - each of them does something for the driver that the driver had to do manually before. How is this not a continuum between full manual control and fully automatic driving? Who's liable if the automatic parallel parking runs over an orphan lying on the pavement?

You are, because you still control speed and should have been paying attention.

 

[Chuckstar]

redleader[/url]":245qm1j5]

wco81[/url]":245qm1j5]Cost of LIDARs are still very high, according to a WSJ article:

I've built optical range finders. LIDAR is high resolution and is useful because the technology is ready to go for prototype purposes, but long term you wouldn't use it in a mass-produced system. Phased array radar is going to be vastly cheaper (as in orders of magnitude) because its moving into mass-production for beam forming wifi, and you can use a single sensor that is digitally scanned rather than an array of LIDAR (or equally awful, one lidar and an optical scanner).

Basically, lidar is just not a good fit for the application.

Either way, it's a little silly to worry about the cost of the prototypes they are using. LIDAR is not a mass produced item. The manufacturing isn't set up for high volume / low cost. And the R&D gets amortized across a fairly small unit volume. This is absolutely the kind of product that lends itself to orders of magnitude decline in costs when volume is ramped up to millions per year.

Your counter point of beam forming wifi is a good example. Not that long ago, someone could have written an article in the WSJ saying "hey, I don't know how they think beam forming will make it into wifi routers... current beam forming communications systems (i.e. cell tower transmitters) cost tens of thousands of dollars each".

 

[Chuckstar]

wco81[/url]":noih6jzk]But it sounds like Google and the car markers are using Lidars, not other tech.

Or have they built prototypes using something other than Lidar?

Doesn't matter what tech they are using in the prototypes. All they really need is a high enough resolution 3D map of the world in front of the car. Any sensor that could provide such a map could be plugged into the algorithms.

 

[Chuckstar]

dh87[/url]":2xdvoydr]At least for short stretches (longer if their software allowed), the Mercedes E and the Acura RLX (any maybe more) can drive themselves. I doubt they're using $10,000 worth of sensors to do it, but I can't find what they're doing immediately.

Those cars use stereo cameras in the bumper. I believe the problem with relying on stereo cameras, though, is that you really need depth perception out past their effective range. I'm not positive about the Google car, but the earlier car it's derived from (that won the DARPA challenge) had stereo cameras to build a map of the close-in 3D environment, and LIDAR for the farther away environment. I forget exactly where the cut-off was, though.

 

[Chuckstar]

Megalodon[/url]":3e7zrlpq]Could a time of flight camera be used? These have been used in mass market consumer electronics.

Current version have too low of a resolution to be useful in this application. I'm not familiar with what causes those limitations, so it's possible that's a way around the complexity of a scanning LIDAR.

 

[Chuckstar]

Zich[/url]":5gwyhorr]Also, anything that relies on blasting out infrared isn't going to work well when there's dozens of cars around all doing the same thing.

LIDAR uses a laser beam, so it's not really blasting anything out. It scans the environment basically one pixel at a time. The odds that your laser beam is hitting the same pixel of the environment at the same time as my laser beam is pretty small. And even then, you'd only end up with a single pixel that had an error, and that error would only exist until the laser passed over that pixel again. Considering these things have to able to deal with the occasional bird, bug or plastic bag passing in front of the beam, I'm pretty sure they can deal with small anomalies like a single pixel that suddenly seems in the wrong spot for the period of a single scan.

The odds that two cars' LIDAR scans interfere in a way that creates more than such occasional momentary blips are astronomically tiny. Multiply astronomically tiny by a couple hundred cars that might be nearby, and you end up with slightly less astronomically tiny. Now add the fact that any bigger anomaly would have to be something that would then make the computer do the wrong thing, and you get to even tinier odds of something going wrong. (The LIDAR scans in a wide angle in front of the car and pretty far out in front of the car. So any random anomaly is unlikely to be in a position that would require the computer to do something in reaction to it. A sudden blip two lanes over is unlikely to change the computer's decisionmaking.)

 

[Chuckstar]

Megalodon[/url]":3h30bja8]

Zich[/url]":3h30bja8]Also, anything that relies on blasting out infrared isn't going to work well when there's dozens of cars around all doing the same thing.

You could say the same of 60 ghz radar, there's ways around it. Even if everyone is in the same band different stations use different modulations.

Chuckstar[/url]":3h30bja8]LIDAR uses a laser beam, so it's not really blasting anything out.

A time of flight camera blasts, and then times the return per pixel. His response is on topic for my idea.

What simplifies it a bit is that the ranges it cares about are pretty short, like maybe dozens or perhaps hundreds of nanoseconds. Also different stations can pulse at widely different and randomly varying patterns. Like one pixel being weird wouldn't blow everything up, one frame being weird wouldn't have to either, as long as two stations don't systematically interfere with each other and the overall din of a traffic jam wouldn't prevent everyone from working.

You are correct, both that I wasn't thinking in terms of time-of-flight and in terms of what the solution would be.

 

[Chuckstar]

wco81[/url]":11xftyqz]Say something falls off a truck ahead of you on the freeway.

Will the self-driving car swerve to avoid running over it?

It would have to monitor cars in the lane to the left and right of it?

It would calculate the best maneuver, which might be braking or a swerve. And, yes, it constantly monitors traffic in 360 degrees. How else could it do a non-emergency lane change, after all?

 

[Chuckstar]

dh87[/url]":3u3dronf]

Megalodon[/url]":3u3dronf]Thanks for the detail on time-of-flight cameras.

Even a few khz seems wildly excessive. This would be sampling stuff every few millimeters even for obstacles at highway speed. Much lower rates, even like tens of hz, seem plausible. At highway speeds, sampling a stationary obstacle at 30 hz would be updates before the distance changes by a meter, and still probably beating attentive human reaction time by an order of magnitude.

I wonder if these calculation still underestimate the problem with interference with sensors that transmit (as opposed to cameras). In a difficult situation, for example a deer in traffic lanes with several cars in both directions, sampling rates might have to increase to match the complexity of the situation. If all cars increase their sampling rates, there could be problems. On the other hand, if there are many s-d cars, they can talk to each other as well.

You only need the sampling rates to be so high. We're talking max speeds in the hundreds mph. You still don't move very far between samples, at the kind of sample rates under discussion. Furthermore, you're still dealing with a situation where most of the time the sensor is not broadcasting. Finally, at those sample rates you also have a situation where individual anomalous frames won't matter, because you take another sample so soon after.

Ever been somewhere that a lot of people are taking flash pictures at the same time? How many pictures were ruined by those other flashes? Essentially none, right? Because the odds that someone else's flash goes off exactly when your camera's shutter is open is still tiny. Same thing here.

 

[Chuckstar]

dh87[/url]":25tcpars]

Chuckstar[/url]":25tcpars]
You only need the sampling rates to be so high. We're talking max speeds in the hundreds mph. You still don't move very far between samples, at the kind of sample rates under discussion. Furthermore, you're still dealing with a situation where most of the time the sensor is not broadcasting. Finally, at those sample rates you also have a situation where individual anomalous frames won't matter, because you take another sample so soon after.

Ever been somewhere that a lot of people are taking flash pictures at the same time? How many pictures were ruined by those other flashes? Essentially none, right? Because the odds that someone else's flash goes off exactly when your camera's shutter is open is still tiny. Same thing here.

I tend to think that the rate will be sufficient, but I wonder about one sensor collision leading to more as all sensors hurry to replace the failed reading.

I hope that the camera flash analogy isn't right, because camera shutters are open for considerably longer than flash durations.

I'm not sure why the camera flash analogy is a problem. Frankly, it just amplifies the point, since that shutter is open for so much longer and still you don't have a problem of interference from other camera's flashes.

EDIT: And I doubt you'd set up such a system to hurry to the next reading. You would just wait for the next scheduled reading. Seems like you would design the thing to be resilient to dropped frames. You wouldn't design is so that a bad frame would require speeding up the next frame. The frames will already be coming pretty quickly, after all. How much faster would you be able to hurry the next frame? (not much) How much more useful would it be to get that extra frame slightly faster? (not much)

 

[Chuckstar]

Alamout[/url]":1xpah63y]Another analogy is something like UDP streaming for video. While it will drop packets every once in a while, it's not worth trying to correct for that (like TCP would). You just move on to the next packet and ignore the error, because it's a tiny percentage of the data you're receiving.

The car is using a similar protocol for mapping out the world around it--it doesn't need to correct every bad signal, it can just smooth over it because errors are very rare.

Exactly. But I think "errors are very rare" is somewhat of a simplification. The errors just have to be rare enough. How rare is important really depends on how quickly your sensor is updating the view of the world. If you're updating the world view 100 times a second, you can probably afford to smooth over a couple large errors per second. If you're updating the world view 5 times a second, then you need errors to really be pretty rare. Off the top of my head, I don't know how many full frames per second the Google LIDAR gives them. Note that for close in things like the car in front slamming on its brakes, the Google car also uses radar (IIRC), so you're not at the mercy of the LIDAR for the really time critical things like emergency braking.

 

[Chuckstar]

Here's a problem with the cylinder view of the world: doesn't a gradual hill and a vertical wall look exactly the same in this system?

 

[Chuckstar]

Kalessin[/url]":1pfpcjga]

Chuckstar[/url]":1pfpcjga]Here's a problem with the cylinder view of the world: doesn't a gradual hill and a vertical wall look exactly the same in this system?

Doesn't your car (along with its liar) tilt eventually as it starts going up the gradual hill?

Anyway it would have be auto-leveled (like most HIDs) if it were going to work with varying vehicle loading. I'm sure the particular problem you're talking about could be solved.

But why not get some 3d info using a sheet, right?

My car only tilts once it actually gets to the hill. If that hill were instead a brick wall (which looks the same in the proposed system) then that would be a problem.

 

[Chuckstar]

redleader[/url]":jl9uf95o]

Kalessin[/url]":jl9uf95o]

redleader[/url]":jl9uf95o]

cogwheel[/url]":jl9uf95o]
Wouldn't the system need to deal with speed bumps and road debris as well? Speed bumps are usually only a few cm tall, and road debris can easily puncture a tire while being small. Also, I'd think you'd want some amount of elevation resolution to deal with stuff like swinging barriers such as those used on low traffic railroad crossings, controlled access parking lots, and toll booths.

Do you need elevation resolution for that though? Even without it you can still see the return from the road. A long linear section of more reflective road perpendicular to the direction of motion would probably be a speed bump, and you could confirm via camera that its not something more serious (like a fallen tree).

Swinging barriers, booths, etc. I think ideally the car needs to know about them. If not and it's unknown but in the cylinder-view of the car, then it needs to be avoided, as in: the car will not drive through them without some external confirmation (could be an electronic signal from the booth, for instance). This applies to anything reasonably big that can't be reasonably avoided: tires, trees, dead animals, etc.

It needs to know about them, but does it need to resolve them? Not clear to me that it gets you much. A solid object is a solid object, you don't want to hit it, no matter if it is suspended from above or sitting on the road. A single elevation pixel would be enough to detect anything, or maybe a few if you want to be really sure. I don't think high resolution like most lidar systems is needed though.

Chuckstar[/url]":jl9uf95o]Here's a problem with the cylinder view of the world: doesn't a gradual hill and a vertical wall look exactly the same in this system?

No, unless the wall is essentially transparent, since a solid wall is a single reflector, while a gradient would be a series of reflectors at different distances. Though this is an interesting point, if you need to be able to detect steep hills, have a few elevation pixels would make sense.

Transparent wall? You mean like a fence? Or a glass storefront? Or a hedgerow?

 

[Chuckstar]

redleader[/url]":1ds2vg3l]

Chuckstar[/url]":1ds2vg3l]

redleader[/url]":1ds2vg3l]

Kalessin[/url]":1ds2vg3l]

redleader[/url]":1ds2vg3l]

cogwheel[/url]":1ds2vg3l]
Wouldn't the system need to deal with speed bumps and road debris as well? Speed bumps are usually only a few cm tall, and road debris can easily puncture a tire while being small. Also, I'd think you'd want some amount of elevation resolution to deal with stuff like swinging barriers such as those used on low traffic railroad crossings, controlled access parking lots, and toll booths.

Do you need elevation resolution for that though? Even without it you can still see the return from the road. A long linear section of more reflective road perpendicular to the direction of motion would probably be a speed bump, and you could confirm via camera that its not something more serious (like a fallen tree).

Swinging barriers, booths, etc. I think ideally the car needs to know about them. If not and it's unknown but in the cylinder-view of the car, then it needs to be avoided, as in: the car will not drive through them without some external confirmation (could be an electronic signal from the booth, for instance). This applies to anything reasonably big that can't be reasonably avoided: tires, trees, dead animals, etc.

It needs to know about them, but does it need to resolve them? Not clear to me that it gets you much. A solid object is a solid object, you don't want to hit it, no matter if it is suspended from above or sitting on the road. A single elevation pixel would be enough to detect anything, or maybe a few if you want to be really sure. I don't think high resolution like most lidar systems is needed though.

Chuckstar[/url]":1ds2vg3l]Here's a problem with the cylinder view of the world: doesn't a gradual hill and a vertical wall look exactly the same in this system?

No, unless the wall is essentially transparent, since a solid wall is a single reflector, while a gradient would be a series of reflectors at different distances. Though this is an interesting point, if you need to be able to detect steep hills, have a few elevation pixels would make sense.

Transparent wall? You mean like a fence? Or a glass storefront? Or a hedgerow?

Those things won't be transparent to radar. Its actually quite hard to not backscatter RF. Lidar maybe for a shop window, but then again, people don't often place unsupported glass windows in the middle of streets.

Anyway what is the point saying that you can't see invisible things? I mean I agree with you, invisible things aren't visible, but if the object is actually invisible (however unlikely that is), then adding elevation scanning doesn't help you one bit.

Radar? Haven't we been talking about LIDAR this whole time?

And I'm disagreeing about elevation not helping, because that's only true if you assume that that any return is a obstacle, and doesn't just represent topography.

And, btw, the car doesn't have any way to know whether something like a window is anomalously in the middle of a street, or whether, for instance, a street fair has been set up that day and the first booth has glass walls.

 

[Chuckstar]

I should probably back off my position at least a little. If you are combining radar information with video (especially binocular video), then it might be enough to just have angle/distance data, without having vertical information. My point was more that there will be types of radar returns that will be inconclusive (as to whether it's an obstacle or not) without additional information.

 

[Chuckstar]

Are neural nets considered robust enough for mission critical applications? With a neural net, I mean, aren't you always just one potential edge case away from a failed result? You can't apply a rigorous approach to validate the system. All you can do is run data through and decide if you go the answer you wanted or not.

I guess the answer could be "if you safely drive enough miles with a self-driving car, then you can be confident in it's performance, especially as compared to the not-particurarly-perfect humans."

Also, I guess another potential answer is "with a complicated enough system, the rigorous approach tends to have holes with edge cases also".

Any AI or real-time software control guys out there know if this has been contemplated before? (It's probably been discussed. I guess I'm asking for a summary of those discussions. )

 

[Chuckstar]

Boneman[/url]":33tj0ppu]

Jim Z[/url]":33tj0ppu]what I posted in one of Dr. Gitlin's articles on the front page:

like the article mentions in the "eyes and ears" section, a lot of the stuff required for this is already in place. I drove a Lincoln MKZ for the weekend which had all the goodies. Active park assist (APA,) lane-keeping assist (LKA,) adaptive cruise with forward collision avoidance, etc. It's eerie feeling the car steer you back into the lane if you drift over too much, and it was really something else when the car started braking on its own when I had cruise set and traffic ahead stopped abruptly.

So we already have cars which can steer themselves, brake themselves, (sort of) park themselves, accelerate themselves (via Electronic Throttle Control,) and monitor and react to surroundings. I'd say we're 80% there for autonomous cars. The last 20% is software. But that 20% is probably going to be 90% of the total work of getting an autonomous car out there.

Yeah, but hasn't that software bit already been done? Google and others have cars that can be entirely autonomous, but just aren't for legal reasons right? I thought the Google car could handle things from beginning to end, but they just wouldn't let it?

Google's car is not done. It's pretty far along, but not done. They've pretty much started by tackling the problems that have the broadest application. Then they switched to a process of tackling easiest problems first, working their way up to hardest. That means progress will slow down. Last time I saw one of their blog postings (this was a year ago IIRC), they were just starting to tackle construction zones. Hell, humans sometime get confused by construction zones. It's not necessarily going to be straightforward for a computer to figure out. And they also still needed more work on snow/ice driving. A big problem with snow/ice being that even for a human it can sometimes be difficult to figure out what's a road vs not-a-road and what's a lane vs not-a-lane. I'm sure that's not the only issue with snowy conditions, but that's the one that pops into my head as especially problematic for an autonomous car.

 

[Chuckstar]

NavyGothic[/url]":39xhj7wy]Google is so far ahead of anyone else that "can't get it to work" is a very real concern. Or rather, "can't get it to work in time".

If Google and / or Google partners can put their cars in consumer hands 5 years before Toyota or GM's solution is ready, there's going to be huge pressure to give in and license for their own vehicles. Even if it's a temporary measure, it's a potentially lucrative opportunity.

As Alamout says, this is kinda unknown territory so who the hell knows how it will play out. And as advanced as Google is, there's still an awfully long way to go.

How do you know that Google is so far ahead? Google is more open about where they are in development than anyone else. I think that's all we really know.

 

[Chuckstar]

Deus Casus[/url]":2t46xcx9]

Chuckstar[/url]":2t46xcx9]

NavyGothic[/url]":2t46xcx9]Google is so far ahead of anyone else that "can't get it to work" is a very real concern. Or rather, "can't get it to work in time".

If Google and / or Google partners can put their cars in consumer hands 5 years before Toyota or GM's solution is ready, there's going to be huge pressure to give in and license for their own vehicles. Even if it's a temporary measure, it's a potentially lucrative opportunity.

As Alamout says, this is kinda unknown territory so who the hell knows how it will play out. And as advanced as Google is, there's still an awfully long way to go.

How do you know that Google is so far ahead? Google is more open about where they are in development than anyone else. I think that's all we really know.

Google has actual test cars on the road and other companies don't. I highly doubt that any company working on this wouldn't have jumped to get a license to run driverless cars on the roads if for nothing but data gathering.

Other companies don't? Or other companies self-driving cars aren't being reported heavily on Ars? There is a difference, you know.

(BMW, Mercedes, Ford, GM, Nissan, Toyota and Audi all have systems under development that they have showed off in various forums.)

 

[Chuckstar]

Alamout[/url]":2vehn6rp]While self-driving trucks are a pretty obvious win (if the last-mile stuff can be figured out), the Teamsters are definitely going to fight it tooth and nail unless they extract something juicy out of the deal.

Unless the trucks are 100% ready on day one, a strike would stop them dead. If they are ready, the union is fucked.

Something like 5% of truck drivers are unionized in the U.S. I don't think anyone is shaking in their shoes about the unions having any particular influence in the trucking industry. There are narrow markets where unions are influential, but that's not going to affect the industry as a whole.

 

[Chuckstar]

Alamout[/url]":2qfc05m4]

Something like 5% of truck drivers are unionized in the U.S.

I guess I was thinking specifically of long-haul trucking (big rigs and the like), but even there the rate is lower than I thought (around 20% for "transportation and warehousing" according to BLS data.

It's somewhere well under 5% for long-haul trucking. The unionization in transportation and warehousing includes the people who work the train system and ports, terminals and warehouses, many of which are 100% union. Note that there are ports/terminals that require union drivers on-site. The union drivers move the stuff around on-site and sometimes make short-distance runs. For long-haul, they exchange tractors somewhere near the site (often there's a lot specifically for doing that). LTL has some unionization, also. But the traditional warehouse-to-warehouse and warehouse-to-retail long-haul is essentially 0% union if I understand correctly.

 

[Chuckstar]

Alamout[/url]":3mkps2hz]

ABS constrained for most people now, no?

Sure--this is definitely an area where I don't know nearly enough about cars. My impression is that if you stomp on the brakes the car is going to stop as fast as it can--a robot couldn't do it better (the ABS is a robot that's braking for you). But ABS is designed for people, to prevent them from locking the brakes--maybe a smarter robot that really wanted to stop could do something even better.

Or at the extreme end, we could build self-driving cars that explicitly had a crash-stop ability. It could even be something semi-destructive, analogous to the airbag. We're talking about an extreme situation--it would be worth replacing some parts if it saves a life.

A computer might be able to dynamically control braking or even go to reverse (in a fixed gear EV if nothing else) to decrease stopping distance, something I don't think humans could figure out in <100 ms amounts of time.

I would think that trying to go into reverse would result in locking the wheels up/skidding. You can imagine a combination of physical braking and using the motors to slow down, but any EV could do that when you hit the brakes.

A human can beat ABS braking distance in some conditions (used to be all conditions, but the systems have gotten better). Depending on exactly which system you are talking about, it might not really be accurate to think of ABS as robot braking the car.

The simplest ABS systems wait until they sense one of the wheels start to lock, and then start pulsing the brakes. This pulsing can overcompensate, lengthening stopping distance compared to a human driver. I wouldn't generally think of this as equivalent to a robot stopping the car.

As the systems matured over the last couple decades, the pulsing got faster (less overcompensation) and the systems got smarter about moving the braking force between the wheels without requiring pulsing and even without waiting to sense the actual start of a slip. But even then, humans can stop the car faster on surfaces like snow and gravel that throw off the wheel speed sensors (wheels are always slipping a little when stopping on such surfaces, so the ABS tends to overcompensate). More importantly from the perspective of calling it a robot, many cars now can fully trigger the brakes on their own (collision avoidance). This really is like a robot stopping your car.

But where ABS wins is that in the discussion I just went through I kept using the word "can" for whether humans beat the ABS system. Even skilled humans are more likely than the ABS system to misjudge and end up in an actual skid. Skidding is a much greater danger than a few extra feet of stopping distance would be.

 

[Chuckstar]

Alamout[/url]":1i3mo5it]^ Interesting. I thought trucking was one of the remaining bastions of unionization, but I guess it died a while back.

Trucking was like 90% unionized in the early 80s. A side-effect of trucking deregulation was the unions got killed off. (I don't know the detailed history, just that basic outline.)

 

[Chuckstar]

Alamout[/url]":rt2fdfoz]

But where ABS wins is that in the discussion I just went through I kept using the word "can" for whether humans beat the ABS system. Even skilled humans are more likely than the ABS system to misjudge and end up in an actual skid. Skidding is a much greater danger than a few extra feet of stopping distance would be.

In that case it sounds like you're saying that an autonomous car could theoretically stop faster than a person using advanced ABS, depending on the surface. But it would have to recognize the surface to do that, or something. So maybe not plausible.

The autonomous car is unlikely to implement braking much differently than the advanced ABS/Collision Avoidance systems of today, except for whatever improvements such systems will gain over time. But keep in mind that all cars now have ABS. The comparison to humans is really "human braking with the ABS turned-on vs human braking with the ABS turned-off". Since no one drives their car with the ABS off, autonomous cars will stop no worse than the human driven cars on the road today.

 

[Chuckstar]

NervousEnergy[/url]":3ekgkfp5]NYC already has a very efficient above-ground people mover system... it's the enormous fleet of yellow cabs. I remember the first time I visited NYC and was totally shocked at how more than half the cars in view at any time (and sometimes nearly all of them) were cabs. Hold my hand up, and within seconds (if not instantly) a cab was pulling up right in front of me, and minutes later I was at the next destination (and $15 lighter.) Replace those with Google-style pods in a more controlled street environment and you'd change the nature of surface life in NY a tremendous amount. It'd sure be quieter at least.

One thing that gets emphasized a lot in research on self-driving cars that the public rarely seems to grasp is how a busy traffic pattern full of slower automated cars will typically get people to their destinations much faster than a faster but less consistent traffic pattern of human drivers.

Of course, if we're wishing, I'd rather ban ICE completely from NYC, dome it, and replace all the streets with variable-speed moving pedways. But then we'd probably have to replace all the cops with Sandmen...

Have you ever smelled NYC in the summer when it hasn't rained for a few weeks? You definitely do not want to put a dome over the place.

 

[Chuckstar]

Tom Dunkerton[/url]":3lthut2k]I wonder what traffic flow in the greater NYC area would be like if every car was programmed to stay in its lane. Never double park. Obey traffic signals. Drive at a slow but consistent speed. Never give the finger or toot the horn. Be aware of the situation in 360 degrees. Never rear end or side crash with other cars. I think traffic jams would become a thing of the past.

Consistent speeds and avoiding lane jockeying certainly improves traffic flow. At some point, though, there would still simply be too many cars trying to get to the same place. You cannot completely eliminate the problem by avoiding tragedy-of-the-commons-style driving. Furthermore, the double-parking that really causes traffic flow problems are generally commercial vehicles making deliveries. Automated driving doesn't fix that problem. Although if you managed to meaningfully reduce the number of cars that needed to be parked, you might free up parking on nearby side-streets. But commercial deliveries are going to be very distance-sensitive. If the nearby parking isn't very close, the guy is probably still going to double-park.

Finally, I will point out that methods to improve traffic flow tend, in general, not to fix very high traffic areas. The pent-up demand for driving through those areas is so high that any improvement in flow simply allows additional people to drive there who previously didn't bother trying. We see this a lot when they add a lane to a high-use freeway somewhere. Pretty quickly the better flow encourages additional people to travel along that route, eliminating much of the initial improvement.

But don't get me wrong, there would be a lot of situations where automated vehicles could meaningfully improve traffic flow. It just won't be a fix for every road at every time of day.

 

[Chuckstar]

dio82[/url]":vja068j2]

ZnU[/url]":vja068j2]

tie[/url]":vja068j2]
How are those nicely designed overpasses? In the first picture, pedestrians have to walk an extra football field length every time they cross the street just to get up the gratuitously wavy overpass. In the second picture, they still have to climb and descend four stories of staircases to cross the street. And there is no consideration for cyclists.

Self-driving cars can deal fine with pedestrians and cyclists. There is no need to destroy a city's walkability to integrate self-driving cars.

I was more making the point about overpasses not turing cityscapes into some dystopian-looking horror, not necessarily endorsing the functionality of those specific designs. Keeping pedestrian crossings at ground-level and building underpasses for vehicles at busy intersections is another option, and one that could incrementally lead to just burying some road segments in busy city centers entirely — dig a trench for the road, roof it over, put a park on top.

In the rest of the world where pedestrian traffic is important (i.e. pretty much everywhere except for USA), exactly this is happening in city centers. Through traffic is buried in a trench and roofed over in oder to allow for pedestrian level access. Sometimes plazas are placed on top of the roof, sometimes parks, sometimes buildings, sometimes shopping malls and very often reduced speed, reduced lane car traffic. Pedestrian overpasses are generally a pita, three stories worth of stairs and an amazingly long ramp for bicycle/wheelchair access. Underpasses are more expensive and only marginally better, and always seem to attract criminals. Especially pedestrian overpasses are generally a blight to city view.

This pretty much only happens for pass-through traffic, not for local traffic. I believe the original proposal in this thread was to grade-separate all vehicular traffic, which is just a stunningly nonsensical idea. And, frankly, it's an idea that makes even less sense in a world of automated cars, not more.

 

[Chuckstar]

ZnU[/url]":jrqa0xtq]

dio82[/url]":jrqa0xtq]
In the rest of the world where pedestrian traffic is important (i.e. pretty much everywhere except for USA), exactly this is happening in city centers.

See, this is sort of my point. Right now, this is better for pedestrians, but that's not enough incentive to actually get it to happen in many places in the US. With self-driving cars though, the fact that it lets you switch intersections to a reservation system will mean that not only is it better for pedestrians, it's also better for drivers. I suspect that's when you'll start to see some real action in the US.

Chuckstar[/url]":jrqa0xtq]
This pretty much only happens for pass-through traffic, not for local traffic.

There's no reason it couldn't happen for local traffic as well. Add stairways to access the street every couple of blocks, and some underground parking — and remember that with self-driving cars, the car doesn't have to be parked in the same location its passenger gets out. (If it has to be parked at all; there's a strong case to be made for just banning private vehicles from city centers in favor of a self-driving taxi fleet.)

Chuckstar[/url]":jrqa0xtq]
I believe the original proposal in this thread was to grade-separate all vehicular traffic, which is just a stunningly nonsensical idea. And, frankly, it's an idea that makes even less sense in a world of automated cars, not more.

There's no reason you'd have to do this for entire cities at once; you'd still see a lot of benefit doing it incrementally, starting in the locations where you'd see the largest returns.

But entirely doing away with surface roads in dense urban areas is a nice long-term goal, and over the next 30-50 years this is probably more achievable than it seems right now. If you have mature self-driving car technology, it means you've solved a lot of general problems related to having automated systems navigate the real world. Those solutions are going to be applicable to a bunch of other tasks, including, say, highly automated construction, which could totally change the economics of large infrastructure projects.

This is just a terrible, terrible proposal. It should only take secinds for you to figure out why it's a terrible idea.

 

[Chuckstar]

ZnU[/url]":2ib43v27]

Chuckstar[/url]":2ib43v27]
This is just a terrible, terrible proposal. It should only take secinds for you to figure out why it's a terrible idea.

You've now made two posts in a row insisting grade separation is somehow a crazy idea in this context, without even the slightest hint of a specific objection. It's not at all obvious to me what you're talking about. I can think of many possible objections; I can also think of solutions for addressing them.

What's your objection? Large-item delivery? Emergency or maintenance vehicles? Flooding? The solution to all of those is basically the same — maintain limited surface roads, probably at least one lane along every existing road, that are normally used as bike paths or pedestrian walkways, but can accommodate occasional vehicle traffic if required.

Cost. This would cost literally trillions for a city of any significant size. I really shouldn't have to go on past this. The few expressways that have been buried have cost so much that a proposal to go back through a city and bury every major road is absurd on it's face.

Convenience. You've separated the roadway from the entrances of buildings, parking lots, and generally where the people and businesses are. You're telling me we'd spend trillions in my city just so when I get home from the airport I have to carry my luggage upstairs out of the tunnel to the sidewalk then walk a block to my building? No thank you. Does anyone even want to live this way? I think not.

And what's the value? Traffic flows a little more smoothly? But will it really? You'll still have to have pedestrians down in these tunnels. They'll be getting in and out of cars, loading and unloading packages/luggage/belongings. You haven't even really changed anything. It's just such a monstrously terribly idea.