I'd be impressed if I see a self-driving car making a left turn in a crowded junction without a specific left-turn signal at LA. The only way to turn left in such roads is to make an illegal turn after the signal just becomes red. I wonder how the self-driving car would do in such a situation.

In California (and most places I've lived), this is actually legal. To complete the turn, you first enter the intersection while the light is green, then wait for traffic to clear. If it only clears after the light turns red, you're still in the right - because only when you fully entered the intersection (on the green light) is considered. See here: https://patch.com/california/sanbruno/ask-a-cop-should-i-pul...

Of course, this only allows one car (or maybe two, if the intersection is particularly large) per light cycle, which isn't much. People who enter the intersection on the yellow or red light (tailing the person who was in the intersection waiting to turn left) are turning illegally.

In my country there is very specific provision along the lines "it is illegal to enter intersection if an obstacle would force a stop in intersection and block traffic" which means that in a "turn left or go straight" lane you enter intersection and cannot turn left due to traffic and block traffic for cars going straight making that technically illegal. Human drivers, of course, do this all the time.

Are you sure that moving, oncoming traffic is seen as an obstacle under that definition?

The real trick against gridlock is to equate the switch to green not as "go", but as "go, once the previous wave has cleared the intersection". The "don't enter before there is room for your car on the far side" is only an optimization on top of that.

I am not a lawyer so I may not be 100% correct. As far as I remember my driving classes and other resources (cannot quickly find exact provision), anything that is not part of road infrastructure (there are intentional obstacles to guide traffic) and causes vehicle to change speed/direction is an obstacle, so moving traffic kind of is. I understand that this particular provision is specifically there to prevent gridlocks and is never enforced in this particular situation.

The way traffic works is somewhat dependent on region. I can go to another city an hour away and already feel a bit alien traffic wise there. This is relevant in discussion on autonomous vehicles: in any foreseeable future we need them to coexist with human drivers and abide by unwritten "everyone drives like that here" rules.

If you come to LA, you'll see there's no way to avoid it. The majority of lights are unprotected left turn lanes. The cops here even break the laws and no one judges. Sure, everyone could make 3 right turns instead, but I'd imagine if everyone did that the damage from wasted fuel, smog, and extra traffic would outweigh the few accidents that happen due to cheating the left turns.

In situations in which my country has that same provision, it does not treat moving traffic that crosses your path as an "obstacle", so if you're turning across traffic, you're allowed to enter the intersection and wait there. It might be worth double-checking whether your country has a similar subtlety.

You are allowed to enter the intersection on yellow in California.

21452. (a) A driver facing a steady circular yellow or yellow arrow signal is, by that signal, warned that the related green movement is ending or that a red indication will be shown immediately thereafter.

But CA law also says you can't enter an intersection without having clearance to vacate it. In other words, if the cross street itself is backed up such that its traffic backs up or into that intersection and prevents you clearing the intersection as you complete the turn, you violate the "anti gridlock" law.

That's not relevant to this situation. Waiting for oncoming to traffic to clear, then turning while the light is red is legal, and orthogonal to whether "there is sufficient space on the other side of the intersection". If there isn't sufficient space to turn, whether the light is green, red, or yellow, then you can't turn and are violating the anti gridlock law. If you entered the intersection before it turn red, and there is space to turn, you can turn on green, yellow, or after it turns red.

Most jurisdictions do not consider it blocking the box if you enter the intersection to turn and it turns red after. They consider it blocking the box if you enter with the intention of going straight while it is not clear to do so.

This is the case for NY state, see VTL 1175: http://www.safeny.ny.gov/rowa-vt.htm

CA Vehicle Code 22526(a): "Notwithstanding any official traffic control signal indication to proceed, a driver of a vehicle shall not enter an intersection or marked crosswalk unless there is sufficient space on the other side of the intersection or marked crosswalk to accommodate the vehicle driven without obstructing the through passage of vehicles from either side."

https://leginfo.legislature.ca.gov/faces/codes_displaySectio...

This does not seem to be relevant to the situation here (waiting for the oncoming traffic to clear): usually, there's plenty of space to turn into, but one has to yield to oncoming traffic before occupying that space.

Turning after the light turns red isn’t illegal, it is what you are supposed to do to clear the intersection. Otherwise you’d be sitting there blocking traffic.

Ideally, isn't the point to have only autonomous vehicles and these situations become obsolete?

Maybe after all cars become autonomous. Earlier there will be a hybrid period where both manned and unmanned cars drive together I think.

Even all cars become autonomous, the problem will still be hard I believe. Hard coded rules may solve the problem but relying on communication and cooperation will not be easy.

And computer-driven pedestrians, too - the car had to stop multiple times for turning left/right across a pedestrian crossing which was green at the time. (If you have watched the video, you have seen that most potentially dangerous situations did not originate from cars.)

In vancouver the general rule in dense traffic seems to be that two cars can left-turn through a stale orange light, right as the light is changing and before the opposing traffic has a chance to start rolling into the intersection from their newly green light.

Sometimes only one car, if the timing works out that way. Three is almost certainly too many. The ability of a car to judge the light timing requires near 180 degree vision and head pivoting on the part of a driver.

Equally impressive would be four self driving cars turning left (in right-hand traffic) at equal priority intersection. At least in my country the rules are muddy in this case and no car has explicit priority. The most general "priority in right" loops around and cars should be stuck there indefinitely :)

I like to imagine to avoid such situations there would some kind of control on the dash that allows a stalemate to be overridden (not the steering wheel :P). It would be great if it were a "give way" button rather than a "I'm going now" button.

Doesn't matter how the controls look - you can't have the cake of Level 5 autonomy and eat it too by having a human giving control inputs.

(GP: That's actually easy: random exponential binary backoff. It works for other domains quite well.)

In this particular instance yes, that is easy. The core of the problem is that hard rules force traffic into this deadlock. To avoid this deadlock autonomous vehicles must be taught to sometimes bend the rules. Which brings its own can of worms - how do you teach a robot under what specific circumstances rules can be bent.

Not a good example - "Yield right _in the order of arrival_" is still a hard rule, but practically deadlock-free.

The rule in California is “first to arrive”, with “yield right” as fallback for simultaneous arrival. Deadlock is unlikely (and a tertiary fallback by compass direction would solve it completely.)

Oh man I've driven through canada to blue mountain during a blizzard. That was scary. I really have to wonder how computer vision will account for that. Or similarly torrential downpours here in the Tropics.

To be fair to computers I’m not sure human vision accounts for that, we just get lucky and fall prey to survivor bias :-)

> Good thing about radar is that, unlike lidar (which is visible wavelength), it can see through rain, snow, fog and dust - Elon Musk

https://twitter.com/elonmusk/status/753843823546028040

Radar has been used on snowploughs to avoid rocks/curbs/etc buried in the snow (article from 2004):

http://www.nature.com/news/2004/040302/full/news040301-2.htm...

Ford developed a LIDAR algorithm that can "detect" snow flakes and still be effective, this video shows driving in real snow:

https://qz.com/637509/driverless-cars-have-a-new-way-to-navi...

Computers can provide near "perfect" speed/traction ratios based on active road conditions... which humans are merely estimating based on "feel", without instant direct feedback from the wheel traction sensors, combined with whatever (potentially limited) winter driving experience they have.

Computers can also "remember" any specific location of the route from hundreds of thousand of other rides in the past and of other cars on the road communicating in real time.

In 5yrs I'd expect to prefer putting my trust in any self-driving car on winter roads over a human any day...

I don't think CV ever will account for that. I always have thought the answer is going to be somewhere between great technology, and infrastructure to support it. We need "smart" roads.

Ultimately I think it'll have to be smart highways. You could embed a coded marker in the centre of each lane which the car can pick up using RF (e.g. an induction loop). This plus other constraints like GPS, visual-inertial odometry, etc should be enough to localise and remain safely within a lane boundary even if it's invisible. This would work fine through a layer of snow - some of the more expensive ski jackets have built in loops for detection in case of avalanches.

It's hard to see how we could get from where we are right now, infrastructure-wise, to there, at least here in the US. We can barely keep our roads driveable in some places, much less modernized with the latest in smart beacons and such.

It wouldn't necessarily need to be in-road, but there are numerous places where you could adapt. For instance, cats-eyes in the UK (retro-reflective markers) are placed on virtually all motorways at lane boundaries and the car could look for a signature from them (I imagine at night, autonomous cars already use them as a feature for lane boundary detection). Nobody bats an eyelid about re-setting cats-eyes every time a bit of road is replaced. You could also put boundary markers on either side of the road, in the central reservation or something similar.

The idea would to put the smart tech in the car. What's going at the roadside could be made very cheaply and integrated into existing features without needing to rip up the entire road.

Other countries have toll highways which are well maintained, that's another option for installation.

I don't see this happening in rural/country lanes, the cost would be obscene (but again, we do liberally use reflectors in the UK). I also don't see this as a panacea, it would simply be a backup in horrendous weather.

I've seen those cat eyes before, but never in locations where plows have to travel. They are also relatively easy to place, just requiring some tar and a manual "pick & place". Making something plow proof would require embedding it in the road, either with a drill or a "rumble strip" like divot.

Yes. Smooth, flat, pothole free tarmac would be nice.

Volvo has proposed driving magnetized nails into pavements in Sweden as lane markers. Not only are these useful for self-driving cars, they're good for snowplow guidance.

How often would the magnets need to be replaced?

As often as the asphalt topping.

Uh-oh. In other words, there's a smart road ahead, you can 100% trust it. Except when it's not there, because it hasn't been maintained for decades, and it's unclear where the missing pieces are. Can't see what could possibly go wrong.

Plus there's also no incentive for anyone to waylay the smart road off the physical road, because there are no evil people (i.e. with motives that are incompatible with the vehicle's survival).

It's a hint, like lane lines. You're going to be using GPS and LIDAR as well. Yes, some of the lamer self-driving systems put too much faith in lane lines.

Indeed. Which means that we're back to square one, "it's where most people are, which is exactly the spot where it's least needed; whereas in places where more conventional navigation fails, there's minimal incentive to also build and maintain this."

TL;DR: good roads will be getting better, bad roads won't.

Seems like self-driving cars will turn into trains and trams by another name.

If it can take you 90-95% of the way, and seamlessly pass over control for the non-beaconed start and end stretches, that's still a serious convenience boost over a train/tram.

I would love this! I love trains, we should have never stopped building them.

There's some interesting recent work at (IIRC) MIT using ground-penetrating radar techniques to image subsurface features of the roadbed in real time, as the car drives over them. Rocks, rebar, and such are almost completely weather-independent (short of torrential rains that wash out the road itself), and they don't change significantly over time.

Indeed!

Since off-road use for smart cars is comparatively small, we should optimize the roads to be smart-car friendly.

Perhaps we could install systems which would enable smart cars to stay in lanes, perhaps with some hardware preventing the smart cars from going off-road.

And while we're there, we could also have systems providing charge built right into the road! Wouldn't that be neat?

The lane-keeping and power systems could allow regular cars to operate on the same roads, for backwards compatibility reasons.

Sounds too expensive to build? Not until you realize that you've re-invented the infrastructure for trams and trolleys put in place a 100 years ago.

Perhaps better. LIDAR can distinguish between clear, solid and semi-clear (snow and slush).

Perhaps self driving cars might never support highways in places with inclement conditions to begin with, which is totally fine. I agree with the other commenter about "smart" highways with embedded markers to let the vehicle know its position despite rain/snow/mud.

It's only "fine" if you don't live in places with inclement conditions, and everyone's talking about how we're gonna ban cars that aren't self-driving. ;)

And if there is self-driving zones, and non-self-driving zones, you better believe I still expect my normal car to work if I drive it from Chicago down to San Francisco.

Accurate GPS and a well mapped area can give the car a precise road lane when snow is present.

That's easy if you're not, say, in a winding, narrow, deep canyon, such as you might find in a mountain pass. Which is, for related reasons, the same place you'll find the heaviest snow cover. Also for related reasons, "well mapped areas" are around where people live, as opposed to - again - places where you could expect uncleared snow.

"Works great near my house, might accidentally kill you if you try to ride it outside the happy path" is somewhat underwhelming.

That's assuming a very accurate GPS in very good conditions. You also need to be sure that there aren't any road works that would temporarily change the lines. Seems a bit risky to commit 100% to the GPS for that.