Hacker News Reader

  I'd be very curious if this conework was even legal

It doesn't matter if it's legal. It is not code compliant. Neither is "double parking". But it happens beyond the control of mere mortal drivers, and it must be maneuvered through safely when it happens. It's like kids dashing into the street -- it's not legal, yet you aren't empowered to simply hit them if it can be avoided.

Here in California, it happens all the time. There was a half block of EL Camino Real in Sunnyvale this afternoon that was simply coned off with no signage and no warning by construction people working on a new building. I called it in to Sunnyvale DPS, and they were totally unaware of it. There were two near-collisions in the 45 seconds I was within view. You have day laborers with no clue about traffic safety and ignorant of context given hand signs and no instructions, let alone legal standards.

Even if you place cones and signage perfectly per code, you still have to deal with signs or cones being knocked out of position or out of view.

You may want to look at this from a wider angle. Obviously the driving logic should handle this properly. But... it also has the potential of production testing road infrastructure automatically. Once it's good enough obviously.

P.S. I hope similar tech will be built for other types of road use, so cars alone don't shape our roads even more.

> I called it in to Sunnyvale DPS, and they were totally unaware of it. There were two near-collisions in the 45 seconds I was within view.

Seems like if those confused drivers had merely stopped where they were, safety would have been improved!

Again, the question isn't whether or not automation technology can be confused by bad signage, we know it can. Nor is there a question that humans can be confused by the same situation. The question is, when faced with situations like this, we can trust the automation to operate safely.

And it seems like Waymo passed the test.

It's true the externalities are different for the different technologies though. If you put down bad cones currently, you raise the likelihood of an accident. If you mark your lane badly in the era of automated vehicles, you cause a traffic jam. I know which I'd prefer.

> it seems like Waymo passed the test

It did not seem so to me. Blocking the road (when it was clearly possible to get out of the way) increased the danger to others through inaction.

Issac Asimov would tut tut.

You're not controlling your variables. The choice isn't between "block the road" and "unimpeded safe traffic", it's between "block the road" and "dangerous driving conditions". As we all know bad signage causes accidents. Here, it did not. Because the car stopped.

>The choice isn't between "block the road" and "unimpeded safe traffic", it's between "block the road" and "dangerous driving conditions".

The choice was between block the road, or pull off to the side, when it was clearly safe to do so.

> or pull off to the side

The car did pull off to the side. Those cones mark the side of the road. What you're demanding it do is to cross the cones to pull off the closed-off shoulder of the road. Which is (1) illegal and (2) only safe because of assumptions you're making about why that road was closed.

I mean, consider that they had a giant automated Waymo-branded repaving machine coming down in that lane. Would pulling off the the side have been the right choice?

Again, it's true that an AI isn't going to have the same context that you do as a functioning human driver. And so it's going to make different decisions in pursuit of safety than you would. That's not evidence that it's unsafe. This behavior simply wasn't unsafe.

>The car did pull off to the side.

That's not what I saw.

The car attempted to follow it's planned route to use the right lane, and stopped in place when it saw the next traffic cone, partially blocking the left lane.

Then it backed up and fully blocked the left lane.

I think as we come to cross a certain threshold of how widespread these autonomous vehicles are it just makes sense to shape infrastructure around the technology. For starters make it code for construction sites to place beacons that relay information to these vehicles so that there's no equivocation about what the right path for it to take is and what it needs to do.

Because making AVs work to the last step, to account for every possibility, we're not there and we're not getting there for a long long long time, so have then the bunch of AV-manufacturers come together and standardize on a secure (lol) communication protocol and give it the crutch that it apparently desperately needs.

> For starters make it code for construction sites to place beacons that relay information to these vehicles

FWIW: it really doesn't seem like it has to be that complicated. The car could see the cones just find and knew not to cross them. They just needed 4-5x more cones than they used to make it clear to the car[1] where the lane was actually going.

[1] And human drivers! Again, this conework was just plain awful, period. I'm 100% convinced that the poor Waymo device wasn't the only driver making bad decisions in that intersection that day.

They may be confusing, but I highly doubt a human driver would pull to the middle of the road and stop for 5+ minutes as drivers honked and drove around.

Honestly, if it had just stayed on the side street, I wouldn't be so critical, but the fact that even after an assistance team was dispatched, the car pulled into the main highway and then immediately stopped would have frightened me enough to get out of the car and wait on the sidewalk. You can also see a moment where the car starts to pull out into traffic as another car is passing it and then immediately slams on the brakes.

I seriously think that if the car is about to take off by itself unexpected, then it would be a bad idea to get out of it.

If control by human is so poor, this is a huge murder vehicle. Sure the systems are improving, but programmers should see all sorts of problems with this that can't be mitigated.

> If [... then] this is a huge murder vehicle

Stop it. Waymo vehicles (Tesla autonomy too, FWIW) are extremely safe, that is just an experimental fact. The circumstance in question here has nothing to do with safety, really. The bug is that it caused a traffic obstruction while the (human) operators attempted to get the car oriented.

Hyperbolizing the circumstances is really not appropriate. This was not a safety situation, really, at any point.

You stop it. They are testing huge, fast robots in public streets. It's messed up.

There's a joke among cyclists: "If you want to get away with killing someone just be sure that you're driving a car and they're riding a bike." We could make a similar joke about these killer robot cars. "If you want to test killer robots on the public just make sure they are shaped like cars." (And if you need to, you can throw in a "safety driver" as a sort of moral crumple zone to deflect the blame when your robot kills e.g. Elaine Herzberg. https://en.wikipedia.org/wiki/Death_of_Elaine_Herzberg#After... )

- - - -

We could build self-driving golf carts (slow, light, not so capable of killing) and work up to cars, eh?

We don't have to go straight to fast heavy cars, eh?

Chill. It should just give pause for thought, not stop thought.

I get confused about construction sites that are much well laid out than this. I think that I would have panicked and slowed down or actually drive in the wrong side of the cones.

I acknowledge that it is weird that the remote assistance team can't remotely stop the vehicle and I agree that it should have behaved differently but maybe there is some kind of engineering reason or just a glitch. It seems to recalculate what to do three times.

A human would have seen what all the other cars were doing, and emulated that eventually, even if the construction was poorly marked.

It was incredible to watch how Waymo was able to plot the huge number of cars accurately around the vehicle throughout the drive. But as soon as something creative was required, it all fell apart.

I still remember a talk where someone from the original DARPA challenge was recounting an episode where two autonomous cars arrived at an intersection precisely at the same time. They deadlocked because they were waiting for each other to go. A programmer added jitter to work around that situation, but it is one of those things humans innately would figure out.

> A human would have seen what all the other cars were doing, and emulated that eventually, even if the construction was poorly marked.

I have a hunch that 90% of self-driving could be solved like that, but then there's always that 10% of situations where there's nobody in front or someone who doesn't know what they're doing.

But yes it would be a good heuristic.

Until, in this case, the car in front made a left turn.

But it would still have escaped the deadlock it was in.

In this case, if it followed the car ahead through a left turn, then that would be so, but if it did not, and it were still in the section of road restricted by the cones, it would have become stuck again, just as it did after it made the initial turn onto the restricted road.

The thing is, the strategy of 'follow other traffic' needs exit rules, and while it is relatively straightforward to begin the strategy when stuck, deciding when to abandon it is less clear-cut. Did the left-turning vehicle do so because this was the correct way to deal with the problem, or did it have a different goal?

I'm reminded of Douglas Adams:

"He had a tremendous propensity for getting lost when driving. This was largely because of his method of “Zen” navigation, which was simply to find any car that looked as if it knew where it was going and follow it. The results were more often surprising than successful, but he felt it was worth it for the sake of the few occasions when it was both."

> A programmer added jitter to work around that situation,

A random sleep to avoid a race condition.

If a bunch of randoms on the Internet can think of using the other vehicles as guides for the right way to go, why hasn't Waymo?

Also, why don't the roadside assistance people or the woman on the phone not have an off button to stop their cars randomly taking off on them? Other machinery has an e-stop, why not this?

I don't know if you're new to this whole technology thing but the answer a lot of the times is: things are more complicated than the simple words that can be used to describe the thing.

Most likely because there are tradeoffs. Tesla is using other vehicles as guides, so it is possible.

Exactly. There's a famous video out there from a few years back showing Tesla autopilot dutifully following a BMW at highway speeds. Then the BMW suddenly swerves around a stopped vehicle in front of it at the last minute, and the AP doesn't have time to recognize the object and come to a full stop, so it runs right into it. (The object is a fake baloon car, but the situation was real enough.)

Tesla has since improved recognition to prevent that situation, but the point remains: blindly following other vehicles is a common mistake. Remember all the "Do Not Follow" signs on construction vehicles?

>incredible to watch how Waymo was able to plot the huge number of cars accurately around the vehicle throughout the drive

4:30 look at the cars coming from the left and turning right (down the screen). Google hallucinates two cars right there, and thats just the start of the half hour journey.

> A human would have seen what all the other cars were doing, and emulated that eventually, even if the construction was poorly marked.

Google are probably planning for 100% of the cars on the road being AI-controlled, at which point it no longer makes sense to copy other car’s behaviour, because they are all AIs.

Seeing how monopolistic Google is in other areas, it would not surprise me if they are chasing a self-driving monopoly, too.

It's perfectly clear. There are 2 cones and a sign set diagonally across the lane: https://i.imgur.com/8ViGpLP.png

You cannot drive into that lane without hitting those cones. It's hard to see from the camera angle in the backseat but a driver in the front would definitely understand that the lane is blocked.

The first cone was placed a few feet to the right of the lane divider. That and that the next cone was on the divider indicated that the right lane was closed.

I wonder if the traffic on the left lane somehow shifted the code/network into favoring that it was only the right lane that was closed. That may have gotten it moving after the first hang up.

However, once it turned into the left lane and passed the first cone (which was offset to the right of the divider), it saw the second cone placed in the center and forgot that it was the right lane that was closed.

I think it pulled over because it thought that both lanes were closed or did not know which one was closed. It’s modeling of a sequence of cones closing lanes and retention of that model seems inadequate.

I second DrScump's point, but it is also a fact that the car did recognize that the lane was a no-go area, at least in its immediate vicinity. From the video, the fragility revealed here seems to be that the system driving the car is attemptng to follow the route from the planning system very explcitly, down to which lane to use, and there seems to have been no effective way to get that plan revised in the face of the problem.

In general, I would imagine that following a route plan to the lane level of detail is important; I can see that it would be when approaching a turn, for example. Also, when the car first encounters the cones, it only recognizes that the lane is blocked for a short distance: after that, each cone encountered seems to be a separate problem. Humans easily generalize a line of cones to a blocked lane and modify their plans accordingly, but we do not have AGI yet.

> If you want to guide left-turning vehicles around the right lane you need to put something IN FRONT of the lane you closed.

The important point is that a human driver (as all the human drivers in the video) has no trouble. A lot of driving is not by the book, but humans coordinate and go with the flow. The software in these cars breaks down as soon as anything unexpected happens.

There was a cone in the middle of the blocked off (right-hand) lane at the turn where it initially got stuck. Seems maybe the spacing between the cones was more cause for the confusion?

Lovely coincidence that this is in the news feed too just now: [http://files.meetup.com/119715/WhyHeideggerianAIFailed.pdf]

I agree. In my town roadwork on a four lane road caused both lanes to be diverted into two sparsely-coned lanes on one side of the road, and some human drivers started driving in the wrong lane into what would be opposing traffic.

They did something similar by my house on a pretty lightly travelled road and it was confusing as hell. People were using both sides of the cones and realizing too late they were in the wrong side.

Direct link to relevant part of video: https://youtu.be/zdKCQKBvH-A?t=1008