Hacker News Reader

Comment by jcynix | Hacker News Reader

I doubt that real self-driving cars will ever be a thing. OK, assisted driving on prepared streets (e.g. highways today or wie roads with some electronic markers in the future) but as soon as you enter smaller roads which are at most as wide as 1.5 cars the "fun" will begin. I could offer some dashcam videos of many such roads here in Europe.

Just imagine people accustomed to highly assisted driving in a metropolitan area suddenly forced to take over when two cars cannot manage to pass each other on such roads.

So I liked to watch "There's no such thing as a Self Driving Car - A MOTOMANTV RANT" by a car journalist here: https://youtu.be/8NJQCyv6BSk

View on HN

Self driving car services already exist in Phoenix and San Francisco.

Whether they can feasibly scale out to many other cities in an economic fashion is still unknown, but on a technical level they're definitely possible, because we already have them.

And yeah, cities and countries vary in difficulty, though as far as the US goes, SF is quite technically challenging in terms of urban obstacles (though the weather is of course quite mild).

Depends on one's definition of "self driving" and as a computer scientist with AI training (although that happened during last century ;-) I still see no real self driving cars, just highly assisted driving where a human always has to be prepared to take over in case something severe happens.

And when I see how most modern cars handle signs like a speed limit to e.g. 30km/h with both additional texts like "from 9-15” and additionally on "monday-friday" or limited "for 300m" I always shake my head.

> I still see no real self driving cars, just highly assisted driving where a human always has to be prepared to take over in case something severe happens.

There's no human driver in the car. At most, the cars can stop/pull over and ask for remote navigation from humans (but not remote operation).

How do you know there's no remote operation?

Because they send people out to physically recover the vehicle if it gets too stuck. They wouldn't do that if a human could just drive it out.

Both can be true. In some cases the remote driver can drive it, in other cases it requires physical presence.

Why would you believe that? How do you know human operators don't type pages of results manually every time you do a Google search?

If waymo's entire fleet was operated by real time remote drivers (extremely dangerous because of latench) we would know already

Granted, there are some niche applications where autonomous vehicles exist. And some level x assisted driving on prepared roads. But I was talking about "real self driving" cars, i.e. cars on any street in any city or even in rural areas. And I still don't see them anywhere, nor do I see them coming in the near future, expect in some tech dreams.

Cruise and waymo are running ride hailing service without any drivers at all in multiple cities including SF

So next up: will they be able to reach profitability servicing SF and Phoenix. So far, highly small scale subsidized pilot projects.

The hardware is more expensive than regular cabs. The software has cost billions, an may cost billions more.

Operationally, energy use should be marginally higher. Hourly cab-driver wages: gone. Insurance and write off — we'll see, but if the software is any good this should be considerably lower.

I guess it comes down to the big unknown: will it work at scale. Can we find enough cities with great weather, where enough people will use it? What will happen when the price goes up? What is a fair, sustainable price? Will that be higher than human-driven car?

Then ice thing about software is that you can copy it at least, even if it’s expensive. LIDAR prices are coming down at a steady rate, although Tesla is still trying to go the vision only route.

> I guess it comes down to the big unknown: will it work at scale. Can we find enough cities with great weather, where enough people will use it? What will happen when the price goes up? What is a fair, sustainable price? Will that be higher than human-driven car?

If we assume the technology won’t improve any longer, this is true. But I don’t think that is a very safe assumption. Everything seems to point to the question being when rather than if.

Another five years away it is, then ;-)

There will be progress in 5 years, will it be enough? Who knows, it’s something to play the stock market with if you are into that.

Who says they need great weather?

People said the same shit about Phoenix: "ohhhhhh, but what about the cities that don't have extremely wide streets and almost no pedestrians???" Now that they're in SF, the goalposts immediately move.

Weather will be gradually solved, just like other technical aspects.

Weird. Should tell that to the Waymo that drove me 20 minutes from my home to a restaurant downtown last night. I was the only human in the car, located in the back seat.

https://www.youtube.com/watch?v=LInDo1A0J4I

Ah yes, the famously narrow, cramped and heavily utilized by pedestrians streets of Phoenix.

Completely ignoring that they they also mentioned SF?

Aside from weather, SF is probably one of the hardest to drive in cities in the US.

From my perspective as an SF resident, it truly does not exist here.

We have some interesting experiments (Cruise, Waymo) wandering around (a subset of) our streets, but they simply can't drive nearly as well as 98% of humans.

They achieve some measure of "safety" by failing safe e.g. stopping and putting hazard flashers on when confused. But they're not an effective or useful "service" for us, in the current state. Have you ever seen one try to deal with a contested left turn, or a 4 way stop with pedestrians constantly crossing?

> or a 4 way stop with pedestrians constantly crossing?

The problem is by the letter of the law, the car can simply never proceed. People drivers don’t need to adhere to that letter so much, so can get creative, auto taxis can’t.

Every waymo ride I have had has been at least as good of a driver as a mediocre Uber driver.

Haven't tried it, but I hear Cruise is considerably worse.

> Have you ever seen one try to deal with a contested left turn, or a 4 way stop with pedestrians constantly crossing?

I've done considerable mileage in Waymo and these two situations it handles pretty well. I think the software is improving rapidly - I've reported some problems that happened consistently that a week later seem to have been resolved

Austin too. I signed up to try it, though I haven't got around to it yet.

I'll ask you what I always ask folks who say that something will never be a thing - do you really think that 1000 years from now, cars won't be able to fully drive themselves under all conditions better than humans? As a reference for the amount of progress we can make in 1000 years, roughly that long ago, someone invented the compass.

If you can agree that in 1000 years, it's extremely likely that we'll have self-driving cars, then to paraphrase Churchill, now we're just haggling over timeline. 500 years from now, will we have self-driving cars? 500 years ago, gunpowder and windmills were new technology. How about 200 years? The big thing was steam power then.

And so forth and so on. Unless there is some natural or physical law that would prevent self-driving cars, it's just a matter of when. And honestly, if you watch some of these Tesla FSD drives that people post on Twitter, it's pretty clear that self-driving is production-ready under a significant number of circumstances.

The problem is, of course, that there are a lot of edge cases, and because the stakes are high, we have to solve most of those to get true full self driving. But given that cars can pretty reliably traverse the streets of SF, I struggle to see the argument that the remaining edge cases are impossible to solve.

The probability that humans are still driving around in cars, self driving or not, in 1000 years is incredibly unlikely.

Yes but also deeply besides the point the post you are replying to is trying to make.

His question is - can your mind wrap around the amount of progress that can happen in 1000 years? If yes, can you imagine that going from "now" to "fully automated cars" is a small amount of progress, relative to progress that's possible on 1000 years? If yes, great - now you can dial back that it probably won't take 1000 years.

> do you really think that 1000 years from now, cars won't be able to fully drive themselves under all conditions better than humans?

Ignoring the questions of if there will even be people -- let alone cars as we known them -- in 1,000 years, my answer to this is:

I have no idea. I don't see any special reason to suspect that such cars will exist in 1,000 years. I also see no special reason to suspect that they won't.

> If you can agree that in 1000 years

We don't agree on this.

Let me rephrase the question, as folks who are answering in the negative are mostly avoiding the main point.

At any time between now and 1000 years from now, will we have the capability to build a fully self-driving car, assuming that there are no cataclysmic, apocalyptical, or other type of events that render any question of technology irrelevant?

There will be no need for self driving cars in 10 years, because assuming we are still around, we will have car driving humanoid robots that drive our same cars. They will fit into our driver seat, be able to reach all the controls, and when they are not driving they can take care of other stuff.

This will be the best of both worlds because we’ll still be able to drive when we want or need to (on roads where it’s allowed), and free to be driven when we prefer that, and the existing fleet of cars can age safely, preserving that investment for its useful lifetime.

In 100s of years, many other things will have changed which could make self-driving cars (as we think of them today) completely irrelevant.

The question wasn't about relevance, though; it was about technical capability. Whether or not anyone actually uses one, will the technology to make one exist? Regardless of whether they're in use in 1000 years, will one ever have existed?

Driving through Mount Hamilton I realized this as well, I would never trust self driving to those roads up there with thousands of feet of elevation just over the side of the road and parts which must be shared by both directions of traffic.

I've taken about 40 rides in Waymo in SF with no safety driver. One thing it did handle surprisingly well was a narrow 1.5 car width road, with bumper to bumper on-coming traffic.

It took a bit longer than a human driver would, but it made it through. I should have videoed it because I thought for certain it would get stuck.

My biggest complaint with Waymo's driving was it sometimes didn't plan ahead to be in the correct lane for a turn coming up within a few blocks and then couldn't get over due to aggro drivers. They seem to have addressed this though

The most impressive feat was when it went down a narrow road to find it was blocked. It managed a 3-point turn in a difficult situation and found another route.

The biggest "bug" I witnessed was a situation with only one road going to the destination and that road was closed with a police road block. It turned around and then kept coming back to the road block, basically looping forever. Had to have an engineer intervene.

If Waymo can handle SF roads, I think it can handle anything in the US (at least without deep snow on the roads). Dunno about other places where you might have to get "creative" with your driving.

> It turned around and then kept coming back to the road block, basically looping forever. Had to have an engineer intervene.

How does one tell the robot car to pull over, I'm done and getting out here? Can you open the door and jump out if need be?

there's a "Pull Over" button on the front seat and back seat screens

You hit the end ride/pull over button. It's generally not advisable to attempt to leave a moving vehicle.

Two cars in close quarters seems like the ideal application for driving assistants. Robots are much better at the actual mechanics of moving the car.

Does that include knowing which wheels (the driving ones) would be better to keep off the verge whilst turning?

> Robots are much better at the actual mechanics of moving the car.

[citation needed]

This is a very dubious statement and it really sounds like you have never driven a car or anything larger.

I mean, it seems self evident to me.

Human perception + reaction time is eons compared to how quickly a computer can move the wheels and pedals.

Whether the machine makes the _right_ decision is much more difficult, but a computer can actuate with speed and precision humans simply cannot.

> Whether the machine makes the _right_ decision is much more difficult,

This is why I said what I said; the OP wording implies they meant just that.

True. I saw a self-driving car (not a Tesla) approach a parked car from behind at around 15 mph, then at the last second, swerve around it, leaving what looked like a 6" gap.

I was frightened, but impressed. (Good thing nobody opened the door.)

Yeah, I agree. If everyone was in a self-driving car and predictable it would be fine. As it stands now, the car is basically trying navigate chaos. Humans already do a terrible job, yet we are expecting machines with less experience and less information to do better. It’s asinine.

Self driving is a cost problem and a software problem. Over time it will become solved fully and cheap. I can't believe that a human can outperform a computer at paying attention.

re: 1.5 car width lanes or trails or whatever

Have you seen humans drive in other countries? Italy, India and Boston (lol - many more) have driving situations where lane lines are not obeyed and cars pack pretty tightly on a regular basis. I cannot imagine a world where that is not solved.

> I can't believe that a human can outperform a computer at paying attention.

The issue are the millions of edge cases in which humans remain far better at analyzing, figuring out a solution, and compromising in.

That is the issue, but that means that to say that we'll never have self-driving cars is to say that we're either at or close to the limit on advancement of computing technology. That seems implausible to me, and if you don't accept it as a premise, then it's just a question of how long it is before computers surpass people at all of those things.

You are making a lot of sense, you are setting up logical arguments but some folks go like "uhhh I just don't think so"

>> [...] is to say that we're either at or close to the limit on advancement of computing technology.

Well, we are approaching physical limits of minimum transistor size, so that might not be as ridiculous as you think. What if the best we can do is to triple the current processing power, but that's not enough?

That would do it, and I genuinely thank you for being the first person to come up with a counterargument to any of my comments that didn't just involve trying to rules lawyer my language.

How often do these narrow-lane scenarios require one of the cars to back up? And do any self-driving cars perform this back-up negotiation maneuver?

Here's an example of a non-autonomous car backing up for a FSD Tesla: (at 22:20) https://www.youtube.com/watch?v=2l6KLC3S5gc&t=22m14s

We’ll see how this comment ages in 5 years.

Smaller roads you say?

FSD Beta handles this now, including the negotiation of which car passes first

https://www.youtube.com/watch?v=2l6KLC3S5gc

OK, if two "self driving" cars meet they can negotiate somehow, nice. But they will have to cope with all those older cars and often older drivers in rural areas. Or cows, sheep, or snow drifts or landslides.

I soon will be on holiday in rural France again and you might not believe who and what is driving on those small roads. Which is totally OK with me, but I highly doubt tech will handle that in the near future.

Personally, I stopped saying things like that when AlphaGo showed up. Thrashing a 9-dan master also wasn't supposed to be "something that tech could do."

Humans aren't really all that special.

Reply on news.ycombinator.com