That’s great but it’s 100 years a day of completely unusual inputs. It’s great they can navigate fairly empty streets but what happens when the flood returns?
This article is about cars driving in simulation, not on the real streets. But I do wonder how they can introduce truly "unpredictable" events in the simulations.
I don't know anything about self driving cars, but it seems like solving self driving cars isn't far off having general artificial intelligence. How well does a car react when there is a construction flagger telling it to turn around, or when there is no signs and the traffic lights go down due to a power failure?
As an aside, my recent experience ordering a burrito from a virtual assistant over the phone did not leave me too dazzled with the state of the AI industry. Of course, Chipotle's phone robots are probably not training on 100 years of simulated burritos...
> How well does a car react when there is a construction flagger telling it to turn around, or when there is no signs and the traffic lights go down due to a power failure?
It tells the driver to take over when it can't interpret the situation.
Total self driving is far in the future. Realistically in a few years we can get to the point where the car can drive itself in usual situations, but the human is needed for anything unusual.
The problem in the meantime is that you have to keep the driver alert the whole time. Deaths involving Teslas almost always are because the driver had Autopilot on and was ignoring the “take control” warning.
> This article is about cars driving in simulation, not on the real streets. But I do wonder how they can introduce truly "unpredictable" events in the simulations.
Like normal SW:
Requirement: navigate this road with those signs.
Test: Put the same signs as in requirement. Check that car navigates correctly.
> I don't know anything about self driving cars, but it seems like solving self driving cars isn't far off having general artificial intelligence. How well does a car react when there is a construction flagger telling it to turn around, or when there is no signs and the traffic lights go down due to a power failure?
It either continues or stops. When the traffic lights go off it uses the computed right of way.
> Also, my recent experience trying to order a chipotle burrito from an AI assistant over the phone did not leave me too dazzled with the state of the AI industry.
It's the same for self driving cars. It is like every SW designed after a set of requirements: you feed it with something not specified in requirements and whatch it crash or spitting an error message.
A widely-held view is that self-driving cars must be able to function in an environment that doesn't cater to them. The reasoning is that world isn't willing to bear the cost of rebuilding roads around SDCs, so SDCs must adapt to roads as-is.
I think there's a lot of wisdom in this, but I think the endgame might be to deviate from it a little bit. Maybe SDCs don't need to make 100% of the changes. Maybe if they can handle 99% of situations, they will have proven their usefulness, and the world will be willing to close the gap by making some changes to the road network to make the remaining 1% of situations easier on SDCs.
So suppose SDCs are in regular use and make up 20% of cars on the road, but they can't handle the construction flagger. Maybe by that point the highway department would be willing to change rules about how construction crews must manage cars going through a construction site.
And/or, maybe SDCs also try to avoid routes that go through construction sites. If it occasionally takes you 20 minutes longer to get to your destination because the SDC has to go the long way around, that might not be bad enough to make you give up your SDC.
It's a shame that Waymo (as well as Google Streetview?) isn't operating on the road at this time, since there's probably no better chance to do mapping as when there are far fewer cars and people to obstruct the view.
16 to 18 years in most countries. Prediction and correlation of cause and effect, spartial awaraness and other cognitive functions take a while to train.
>how many years of experience is needed to drive on public roads without supervision
If, as I strongly suspect, full self-driving requires artificial general intelligence, Waymo's algorithms will not get there no matter how long they run simulations or even how many real road tests they do.
This headline (and article) is IMHO promoting a unit-less metric.
There's no mention of the yield in quality improvement per hour of simulation. I.e., how much better does the vehicle drive using the learnings from these 876,000 sim hours?
Once a sim pipeline is set up, it's "easy" to scale up the number of hours it runs (throw it more compute resources, throw it more scenarios). But that doesn't mean the analysis scales up, or the quality of the sim results, or the application of analysis to bug fixes or feature development.
As an driver, I think too much emphasis is placed on those global mean crash/fatality rate statistics. That sort of analysis is too reductive. By not being an alcoholic and choosing to always wear a seatbelt, not driving in bad weather, choosing to obey the speed limit, etc.. individuals can deliberately boost their chances well above the national mean. Insurance companies charge different people different rates because they aren't so reductive as to only consider national averages.
The response I usually get when I talk about is "everybody thinks they're better than average, so changes are equally good you're not", but that too is too reductive. Loads of people drive drunk, while I know that I don't. That alone significantly stacks the odds in my favor for the likelihood that I'm correct in thinking I'm above the average.
Simulators can be pretty good. The FAA certifies some pretty low-end simulator setups for commercial/private pilots to get IFR training in, for example. And of course, airline pilots almost exclusively train in simulators. Very few pilots have experienced an engine fire in their 787, but I'd bet that about 100% of them would handle the situation correctly because of simulator training (and good documentation).
I don't know how good driving simulators are, but at least self-driving cars are modular enough to test components. Even if you can't test the machine vision, you can still invent data coming out of the vision to test the rest of the system. (You can probably fuzz test it, too. Generate a bunch of random scenarios, fail if an obstacle hitbox and vehicle hitbox touch. Tweak that, and then set up the scenario in the field when COVID-19 is over.)
Certainly the vision part of the equation is one of the greatest engineering challenges. That is why they are testing cars in the field. But just having a working vision system doesn't get you a working self-driving car, so quite a bit of effort has to be invested in the rest of the system. With noone allowed to go to work, now is a great time to invest in that.
I feel like HN is too skeptical of self driving. Having driven a Tesla for a road trip, I would feel safe letting it do all the work on interstates without the feedback it currently requires. That's a big step forward. Every day use cases, even if it's not every moment you're in your car, are certainly within reach.
> I feel like HN is too skeptical of self driving. Having driven a Tesla for a road trip, I would feel safe letting it do all the work on interstates without the feedback it currently requires.
How many times does a Tesla have to actively accelerate and careen into a solid concrete barrier before HN's modicum of skepticism for self driving, particularly Tesla's implementation, is justified?
1/2 as many times as a human does. I've driven about 250,000 miles in my life and had 5 accidents, none with injury. But I think a self driving car could do a lot better.
My skepticism of self driving tech is all weather based. Here in LA, I’m confident that a self driving car could do fine; slow traffic (normally), good roads, low precipitation and clear lines make for easy driving.
But back when I lived in Colorado I had multiple times when I had to drive on a highway with completely obscured lines, poor weather, and extremely dangerous consequences for failure. I recall one drive in a blizzard where all the cars driving had to negotiate their own space without any feedback of where the lines were supposed to be. If I recall correctly we all settled on 1 fewer lanes than the road “should” have. I do not believe that self driving tech is capable of doing that yet, especially on mixed human/robot roads.
The guys who can cover California with AVs will already become extremely wealthy. Maybe the north gets them ten years later, twenty years later, a hundred years later. The economic benefits to local public transit in California alone would be huge. Add in commercial intra-state trucking and you've got a nutty advantage already. Maybe it'll turbo up California's massive economy even more.
Maybe lane marker tracking isn’t the way to go? I’ve always wondered if it would be possible (in the US; through a massive initiative) to just paint lines (or embed something in or on the road) in the center of the lanes and have the cars just follow the lines. How much would that cost?
No paint will solve the problem of snow covering said paint. So you’ll have to embed them. The hardware alone would cost billions of dollars to purchase, and installation would cost either trillions of dollars to rework all the roads now, or decades to install as our road network ages out.
Sure, I’m sure there were complaints about the costs, but the value of a massive road network has been known for millennia.
Spending a trillion ripping up a perfectly good road network in order to try out a new technology that’ll generate profits for a small number of companies is an insane plan, and genuinely might not work.
We don’t necessarily need to rip up the existing roads. We could just put some radio based markers on the road. Then when they do construction, they could be easily moved, too.
It’s true that markers on the pavement are cheaper to implement than markers in the pavement, I agree.
I’m also not convinced it solves anything. In the scenario above the problem wasn’t the fact that the lines weren’t visible, the problem was other cars. This is particularly problematic when many of those other cars are human beings who will never have access to modified road information. In this scenario a smart car that knows where the lane should be is still forced to dynamically negotiate for space with other vehicles, sans communication protocols. That is a very hard problem, one I do not believe is currently solved.
Sure it is. Volvo developed a system where magnetized nails are driven into the pavement to mark the center of lanes. It's an additional hint, for when the visual, LIDAR, and GPS systems could use a little help. As in heavy snow. Volvo also points out that it's useful for snowplow guidance.
AGVs have been using wire guidance for years. You cut a notch in the floor, put in a wire, and fill in the notch. That's the technology GM used for the Firebird demo cars in the 1960s.
It would cost a lot more than just mapping the lines & state of the infrastructure like lanes & traffic lights in a database, which is what Waymo already does. Adding additional paint on the road(s) doesn't help the car know what to do when there are weird edge cases. Following lines on the road was already solved 10yrs ago and isn't really the problem.
One, two lane road, flat, rain moderate to heavy, traffic in both directions, 55mph speed limit. I relied on the car to do most of the driving as oncoming lights from cars could make it hard for me to see, either had their brights on or misaligned. I could use the cars display out of the corner of my eye to see it knew the road lines
second, in fog, it certainly sees better than I do, more readily picking out the road and even detecting a vehicle in front of me whose lights were basically out.
still as I always say, with self driving I am a backseat driver sitting in the driver's seat.
Hackernews takes pride in being a data-driven, scientific community, except when it comes to Tesla self-driving, in which case everyone leans on personal anecdotes.
Not sure what the stats are on this, but I'm pretty sure people don't generally die on the freeway by steering directly into concrete barriers while accelerating.
>But I think a self driving car could do a lot better.
1 accident every 50,000 miles is not a good record. I don't want 'self-driving' cars on the road that are only marginally better than the worst drivers out there.
Please don't editorialize titles. This is in the site guidelines: "Please use the original title, unless it is misleading or linkbait; don't editorialize." (https://news.ycombinator.com/newsguidelines.html) Cherry-picking the detail you think is most important from an article is editorializing.
If it's necessary to change a title, please do that by picking the most neutral and accurate phrase you can find on the actual page which represents what the article is about. More on that here: https://news.ycombinator.com/item?id=22932244
(Submitted title was "Waymo's simulators are doing 100 years of driving per day during WFH")
I'm unnerved by the anthropomorphism of "driver". What would we think of a human saying "I've got 100+ years sim experience but out on the street I need safety watch." It obviously means their "driver's" sum of experience doesn't account for much.
The closest you could compare human actions to what their machine is doing is "dreaming". When we dream, we improve our reactions without having to go through the actual situation. But again, picture this human saying "I'm a safe driver because I've dreamt about driving a lot!"
Now you could say that theirs is just a product name. Surely they'd agree machines are different. And maybe I'm pedantic when I say they shouldn't use words that liken their system to a human. Still I think it's important to acknowledge that machines will refuse to drive like we do. Our driving is too probabilistic on multiple levels.
I think we wouldn't be stuck in uncanny valley if it were acknowledged that first we have to equip highways and participants on them with transponders. Then we can let the machines take over on boring, predictable highways.
49 comments
[ 3.2 ms ] story [ 98.2 ms ] threadI don't know anything about self driving cars, but it seems like solving self driving cars isn't far off having general artificial intelligence. How well does a car react when there is a construction flagger telling it to turn around, or when there is no signs and the traffic lights go down due to a power failure?
As an aside, my recent experience ordering a burrito from a virtual assistant over the phone did not leave me too dazzled with the state of the AI industry. Of course, Chipotle's phone robots are probably not training on 100 years of simulated burritos...
It tells the driver to take over when it can't interpret the situation.
Total self driving is far in the future. Realistically in a few years we can get to the point where the car can drive itself in usual situations, but the human is needed for anything unusual.
Like normal SW: Requirement: navigate this road with those signs. Test: Put the same signs as in requirement. Check that car navigates correctly.
> I don't know anything about self driving cars, but it seems like solving self driving cars isn't far off having general artificial intelligence. How well does a car react when there is a construction flagger telling it to turn around, or when there is no signs and the traffic lights go down due to a power failure?
It either continues or stops. When the traffic lights go off it uses the computed right of way.
> Also, my recent experience trying to order a chipotle burrito from an AI assistant over the phone did not leave me too dazzled with the state of the AI industry.
It's the same for self driving cars. It is like every SW designed after a set of requirements: you feed it with something not specified in requirements and whatch it crash or spitting an error message.
They could make it a MMORPG and get people to drive or walk around?
I think there's a lot of wisdom in this, but I think the endgame might be to deviate from it a little bit. Maybe SDCs don't need to make 100% of the changes. Maybe if they can handle 99% of situations, they will have proven their usefulness, and the world will be willing to close the gap by making some changes to the road network to make the remaining 1% of situations easier on SDCs.
So suppose SDCs are in regular use and make up 20% of cars on the road, but they can't handle the construction flagger. Maybe by that point the highway department would be willing to change rules about how construction crews must manage cars going through a construction site.
And/or, maybe SDCs also try to avoid routes that go through construction sites. If it occasionally takes you 20 minutes longer to get to your destination because the SDC has to go the long way around, that might not be bad enough to make you give up your SDC.
If, as I strongly suspect, full self-driving requires artificial general intelligence, Waymo's algorithms will not get there no matter how long they run simulations or even how many real road tests they do.
There's no mention of the yield in quality improvement per hour of simulation. I.e., how much better does the vehicle drive using the learnings from these 876,000 sim hours?
Once a sim pipeline is set up, it's "easy" to scale up the number of hours it runs (throw it more compute resources, throw it more scenarios). But that doesn't mean the analysis scales up, or the quality of the sim results, or the application of analysis to bug fixes or feature development.
The response I usually get when I talk about is "everybody thinks they're better than average, so changes are equally good you're not", but that too is too reductive. Loads of people drive drunk, while I know that I don't. That alone significantly stacks the odds in my favor for the likelihood that I'm correct in thinking I'm above the average.
I don't know how good driving simulators are, but at least self-driving cars are modular enough to test components. Even if you can't test the machine vision, you can still invent data coming out of the vision to test the rest of the system. (You can probably fuzz test it, too. Generate a bunch of random scenarios, fail if an obstacle hitbox and vehicle hitbox touch. Tweak that, and then set up the scenario in the field when COVID-19 is over.)
Certainly the vision part of the equation is one of the greatest engineering challenges. That is why they are testing cars in the field. But just having a working vision system doesn't get you a working self-driving car, so quite a bit of effort has to be invested in the rest of the system. With noone allowed to go to work, now is a great time to invest in that.
How many times does a Tesla have to actively accelerate and careen into a solid concrete barrier before HN's modicum of skepticism for self driving, particularly Tesla's implementation, is justified?
But back when I lived in Colorado I had multiple times when I had to drive on a highway with completely obscured lines, poor weather, and extremely dangerous consequences for failure. I recall one drive in a blizzard where all the cars driving had to negotiate their own space without any feedback of where the lines were supposed to be. If I recall correctly we all settled on 1 fewer lanes than the road “should” have. I do not believe that self driving tech is capable of doing that yet, especially on mixed human/robot roads.
Long story short, it’s not feasible.
Sure, I’m sure there were complaints about the costs, but the value of a massive road network has been known for millennia.
Spending a trillion ripping up a perfectly good road network in order to try out a new technology that’ll generate profits for a small number of companies is an insane plan, and genuinely might not work.
I’m also not convinced it solves anything. In the scenario above the problem wasn’t the fact that the lines weren’t visible, the problem was other cars. This is particularly problematic when many of those other cars are human beings who will never have access to modified road information. In this scenario a smart car that knows where the lane should be is still forced to dynamically negotiate for space with other vehicles, sans communication protocols. That is a very hard problem, one I do not believe is currently solved.
AGVs have been using wire guidance for years. You cut a notch in the floor, put in a wire, and fill in the notch. That's the technology GM used for the Firebird demo cars in the 1960s.
RFID tags are dead cheap, and would get cheaper if 100M were need to embed in road surfaces.
Personal experience with my model 3, two cases.
One, two lane road, flat, rain moderate to heavy, traffic in both directions, 55mph speed limit. I relied on the car to do most of the driving as oncoming lights from cars could make it hard for me to see, either had their brights on or misaligned. I could use the cars display out of the corner of my eye to see it knew the road lines
second, in fog, it certainly sees better than I do, more readily picking out the road and even detecting a vehicle in front of me whose lights were basically out.
still as I always say, with self driving I am a backseat driver sitting in the driver's seat.
Hackernews takes pride in being a data-driven, scientific community, except when it comes to Tesla self-driving, in which case everyone leans on personal anecdotes.
1 accident every 50,000 miles is not a good record. I don't want 'self-driving' cars on the road that are only marginally better than the worst drivers out there.
If you want to say what you think is important about an article, please post it as a comment to the thread. Then your view will be on a level playing field with everyone else's: https://hn.algolia.com/?dateRange=all&page=0&prefix=false&qu...
If it's necessary to change a title, please do that by picking the most neutral and accurate phrase you can find on the actual page which represents what the article is about. More on that here: https://news.ycombinator.com/item?id=22932244
(Submitted title was "Waymo's simulators are doing 100 years of driving per day during WFH")
The closest you could compare human actions to what their machine is doing is "dreaming". When we dream, we improve our reactions without having to go through the actual situation. But again, picture this human saying "I'm a safe driver because I've dreamt about driving a lot!"
Now you could say that theirs is just a product name. Surely they'd agree machines are different. And maybe I'm pedantic when I say they shouldn't use words that liken their system to a human. Still I think it's important to acknowledge that machines will refuse to drive like we do. Our driving is too probabilistic on multiple levels.
I think we wouldn't be stuck in uncanny valley if it were acknowledged that first we have to equip highways and participants on them with transponders. Then we can let the machines take over on boring, predictable highways.