It's still keeping the self-driving unit thought, even though there's a good chance it won't ever be successful. I don't think self-driving cars are realistic at all. I bet if you did a study and ran the numbers, it would be cheaper to put down rail on existing roads and make them rail-car only, automate those and subsidize rail cars.
I feel in the minority in thinking self driving tech will ever be viable. I've worked in computer vision startups and know how powerful the tech can be, but it still can't account with even a fraction of challenging driving conditions. I don't think anything short of general purpose AI could handle it. We are some pretty big discoveries away from general AI.
There will be a large market/human-lives-saved benefit from the tech though, even if we never reach 100% self-driving autonomy
E.g. it is pretty standard now that pretty much every car on sale has automatic emergency braking that tries to avoid running over pedestrians even if the driver does not see them, many (but not all I admit) cars have lane-keeping assist that automatically keep the car in the lane, and will automatically adjust the speed to match the car in front and so on (right down to slamming on the brakes to avoid a rear-end shunt if needed). Then there are the self-parking cars, the ones with synthetic overhead 360 degree camera viewpoints, ones that prevent "unintended acceleration" in car parks, ones that read the road signs to let you know the speed limit etc. And these are "normal" cars that a common joe can afford, not high-end six-figure Mercedes/BMW/Lexus/Tesla etc
This sort of tech was science fiction not so long ago, yet now it is literally standard-fit.
There will be a slow but steady "encroachment" of more and more automation into cars as the years go by - you can guarantee that companies like Cruise are churning out the patents at the least for this sort of thing, even if not outright directly licensing technologies. It feels like a land-grab phase right now, even if there is no real intention to reach L5.
I’ve always been extremely dubious about the “self driving to save lives” angle for one very specific reason: we already know how to save lives. Slow. The. Cars. Down. Oslo managed to get their car deaths down to 1 in 2019. 1, in a city of 600,000. Los Angeles killed 240 that year, or 36x more per capita.
And yet, we don’t actually apply any known techniques to reduce fatalities. If we really cared about saving lives, we’d implement these measures today rather than banking on a long term tech solution.
I believe that self driving is about a fantasy world in which we can ignore all the negative consequences of a low density, car centric world. Pedestrian deaths? AI magic will solve it. Long commutes? Just eat breakfast in your car! Emissions? Why well just switch over to EVs! Never mind that half of this tech is either not real, or orders of magnitude less efficient than the transit that Europe has used for almost half a century!
The truth is we can fix all of these things right now, but it would involve sacrifices for those whom the system is already benefitting. So instead we’ve made up a fantasy world where we can have our cake and eat it too.
It's a cost benefit calculation. Sure we can make everyone go 25mph, but how much does that slow down commerce? It's a tradeoff people aren't willing to make. They'd rather sacrifice a few lives than take three hours to get 75 miles.
But the self driving tech allows us to get much safer while still going at normal speeds. Or maybe allows us to be equally safe at higher speeds.
First, we only need to slow down cars to 25mph where pedestrians are present. It’s not like cities with low pedestrian fatalities are stuck in the Middle Ages. Places where pedestrians are separated from cars can and should go faster.
Secondly I find it disturbing that discussions around acceptable traffic fatalities focus on unnamed and unspecified people as the apparently acceptable fatalities in pursuit of shorter commutes. I think this let’s a lot of people push off that moral calculus as something that happens to other, lesser, people. See: the reactions once people realized that Uber’s self driving car had killed a homeless woman. A better question is “what is the acceptable level of risk for you and your family to accept in return for faster commutes?”
Finally, don’t underestimate the feedback effect. High speeds are only economically necessary, if they even are, because we’ve chosen to design cities in a way that assumes high speed car traffic. Continuing to focus on keeping traffic going just puts off very necessary and politically difficult discussions about public transit and housing density. We could setup a society where 25mph in the cities works just fine, but we’ve made the explicit choice not to, and continue to reinforce that.
> we only need to slow down cars to 25mph where pedestrians are present.
I don't have the numbers on hand, but I would wager to guess that a good chunk of road injuries and deaths happen not just between drivers and pedestrians, but between drivers and other drivers.
Also, most areas in the US where pedestrians can be present are already limited to 25mph/35mph.
60% of the 2019 road fatalities in Los Angeles were non-drivers. So drivers represent a good chunk, but not even a majority. Of course this statistic is going to very wildly based on which city we’re talking about.
To be fair to my original point, we can also reduce driver fatalities by slowing down highways and offering non-car transit options. The lowest hanging fruit in this area is probably shifting cargo transit from road to rail as much as possible, which might actually save the public money, since interstate highway companies receive billions of dollars in subsidy via subsidized road maintenance.
Secondarily would be high speed rail between major metropolitan areas, which would potentially save both road miles (and deaths) but drastically shorten the commute for some citizens.
It kills me that we give such huge hidden subsidies to the trucking industry. Every time this discussion comes up, at least on Reddit, someone always comes out of the woodwork claiming that truckers pay a bunch of fuel tax and registration fees and that proposals to increase taxes on trucking is going to drive up the costs of all goods. For anyone who isn't already aware, the "billions of dollars in subsidy" is in reference to the substantial amount of damage large trucks do to the roads. My Honda Civic going down the road does basically no damage to it at all and it doesn't appreciably damage the road. A large truck going down the road at 80,000 lbs is going to cause far more damage to the road than the fuel tax and registration fees that they paid. Road wear scales to the fourth power of axle weight, so aside from road age itself the vast majority of wear is solely from trucking.
In this case the statistics specifically refer to pedestrians and cyclists, I just shortened it to “non driver” here, which wasn’t an accurate way to summarize the stats.
Where I am, Los Angeles, that is simply not true. 60% of the people killed here by a car in 2019 weren’t in a car. And that’s not even accounting for drivers who were driving reasonably and were killed by someone who was not.
I think it is especially relevant in the time of coronavirus to also consider the idea that maybe you just don't need to go into work all the time (or even at regular times) for a large percentage of workers. Just have less cars on the road by running society a little differently.
Because the costs of those measures are not worth the gain.
If you have a million commuters on the streets, each taking a total of an hour a day for their commute, 20 days per month, increasing the average commute time by just 1% effectively costs >300 lives per year.
The basic rule of any change you want to make to the system is that speeds need to go up, not down. Better AI assist systems help speeds go up, therefore they reduce the cost of commuting in human lives.
A car going at 5mph is still going to kill or main a pedestrian if it hits them.
I agree that speed and general "cars forever!" approach is not great, but there is always room for improvement ... e.g. could tech have saved that 1 person who died in Oslo? Why just stop at "the Oslo model"? Why not do that and autonomous tech? It would be worth it if it saved just a handful of lives a year.
Obviously cars and pedestrians should be separated, but you are dramatically understating the impact speed has on pedestrian survival rates.
At 5mph your chance of surviving a collision is extremely high. Heck, you have a 93% chance of surviving a 20mph crash, depending on your age. Survival rates drop off quickly with speed however, with only 80% surviving at 30mph, and 45mph being a 50/50 chance for most of the population.
The effect is that slowing cars down from 35mph to 20mph in areas where pedestrians are close to cars could potentially double the chance of a pedestrian surviving a crash. And that’s not even accounting for how much less likely a collision is to happen at 20mph than 35mph.
Edit: of course, we can and should add more tech to make cars safer. That’s why backup cameras and seat belts are mandatory, and that’s a good thing. But as an argument for self driving cars, I think that they really come up short when there are tons of effective strategies available right now that we aren’t applying. And that’s my point.
that makes sense given kinetic energy is mass * velocity^2. That squaring of velocity means every mph as they get higher is even more significant than the ones before.
> I agree that speed and general "cars forever!" approach is not great, but there is always room for improvement ... e.g. could tech have saved that 1 person who died in Oslo? Why just stop at "the Oslo model"? Why not do that and autonomous tech? It would be worth it if it saved just a handful of lives a year.
The problem is twofold:
1. Cruise and co. may very well run out of money before they develop tech reliable enough to save people en masse. One proposal decades ago to solve the cars vs people problem was to physically grade separate them entirely, but as it turns out building multistory streets doesn't scale very well.
2. People and politicians, today, are using self driving cars as the magic bullet to avoid spending money on other solutions and changing bad habits. Why spend money on rail or buses if the second coming of Jesus for cars is around the corner? And to a degree, there is an opportunity cost to spending money on unproven X rather than proven Y.
I think this happens fairly often in Manhattan. There's lots of box trucks simply making turns at 5mph and pedestrians crossing obliviously with headphones on or looking at their phones.
Vehicle weight has a large impact on survivability, but not all the way down at 5mph. Plus, vehicles are easy to avoid at 5mph, the majority of the human population can out walk cars at that speed.
That being said, vehicle size is another uncomfortable subject when it comes to road deaths. If we really cared about not killing people on the road, we’d take a long, hard look at what cars we allow to operate on our roads.
Delivery trucks have the benefit of being obviously necessary; I am glad that trucks exist to deliver my food from the train depot (ideally) to my grocery store. But what’s the argument for Suburbans? SUVs not only drastically increase the risk for pedestrians, but they also make the roads more dangerous for other drivers to. For what? Status? A false perception of safety?
It is not just the energy involved from 5mph - not every accident is a cut and dry collision. A car doing 0.5mph still going to crush your skull when you go under the wheels.
People are killed frequently in London (and I am sure other places) when they are crushed by slow moving vehicles turning left or right and the driver just did not see the person there and they are either slowly run-over and pinned between the vehicle and something else and crushed.
Yes, people do get crushed under the wheels of slow moving cars. But the statistics are extremely clear about what your relative chances of survival are based on vehicle speed. To a first order approximation your chance of being killed by a 5mph car is 0%.
Should we add tech to prevent these accidents? Of course. But let’s not kid ourselves about the actual numbers of where and why cars kill people.
As a means of comparison; back up cameras are mandatory to prevent children from being backed over at slow speed. This is an obvious improvement, but numerically quite small, with estimates being that it saves about 50-70 lives nationally per year. This is far less than the number of pedestrians and cyclists struck and killed in Los Angeles per year alone.
> I don't think self-driving cars are realistic at all.
...
> I feel in the minority in thinking self driving tech will ever be viable.
Take out "cars" from that phrase and I think there might be a different answer to it, sooner or later.
Specifically I liked Otto's original proposals and it sounded almost viable. The idea of having a Sacramento -> Chicago or Houston to Dallas routes for trucking with autonomous vehicles, with the last mile routes covered by normal container trucks sounded like it would work in the near term.
The fact that these are big rig trucks ensures the unit volume cost of the hardware isn't going to be a big problem (a 25k computer in a 35k car vs a 120k one in a 800k truck), the servicing can be legally mandated (patch upgrades or cleaning cameras), the power supply management can also be handled if they want to go electric with a container + rig model.
There are fewer questions about a long haul highway only self-driving route, maybe even something which will change the economics of shipping stuff through the panama canal vs going over land in a hub-spoke model without a train depot style unloading/inventory station.
The idea of having a Sacramento -> Chicago or Houston to Dallas routes for trucking with autonomous vehicles, with the last mile routes covered by normal container trucks sounded like it would work in the near term.
Sure, we have that ... right now. Called trains. Train even carry trucks. Trains are vastly more fuel efficient than trucks.
The advantage trucks have over trains is their flexibility, their any point to any other point ability. Take that out and it can be automated but that automation isn't to be more efficient than ... 19th century technology. Ha.
I completely agree. I don't think the only value in AI is to totally replace humans, but rather automate the boring and costly parts that don't require a ton of thinking.
In my field of AI in radiology, I see a lot of people trying to completely automate things that don't have to be, and thereby fail to innovate on the particularly hard parts. Sometimes it's better to just have a human somewhere in the loop to make the harder judgement calls.
It won't - if the idea is to keep it confined to specific overland routes, it is economically absurd vs. trains.
But that's not the idea. It is the camel's nose under the tent.
The goal is to get them on the road somewhere, and then gradually expand use as they can re-legislate at the cost of pedestrians.
There's a template for doing this - when cars were first commercialized, pedestrians had to be kicked off the roads, so they invented "jaywalkers", and launched a huge propaganda campaign to make it a thing[1].
Too much money and priming has been dumped into self-driving tech to abandon, so they'll try to create markets by pushing the squishy slow things that complain when they get smooshed out of the way.
I think we're very far away from general-purpose self-driving that can drive in all locations, conditions, and handle all scenarios. But we don't necessarily need an all-or-nothing solution - I think we could be reasonably close to a self-driving system that can handle a whitelisted set of roads and highways, especially if those roads were augmented to accommodate that self-driving in some way.
Regardless of road quality, weather conditions are unpredictable.
When the rain or snow or haboob kicks in, I guess your car will pull over to the side of the road and you'll need to drive manually until conditions improve. We'll need to keep the steering wheel around for the foreseeable future.
Yeah, I've always understood that to be the big initial play— get all those long-haul trucking miles autonomous, relegate human drivers to handling the in-city miles and cargo yards. I think Ottomotto (remember them?) might have even made a video years ago which pitched this exact scheme.
Of course, the environmentalist in me is saddened by the thought of further driving down the cost of that shipping mode which might be most compatible with being switched to rail freight.
More important than simple probability is the “expectancy value”, which is roughly spoken (and cumulatively) the probability multiplied by the effect when the probability comes to pass.
I just translated the German term Erwartungswert wrong. The English term seems to be “expected value”, and it’s basic enough that you learn it in high school.
There aren't any self driving cars on the road that don't have a safety driver, and the ones that are out there get confused all the time. The Uber one had a software dev turn off the lidar sensor to test the computer vision at night and it ended up killing a woman crossing the street.
Most self-driving is done in warm climates, in a circle. These companies are amassing thousands of empty miles on the same streets every day to impress people who don't know any better.
I lived on a branch of the main Google circuit, and often they were the only traffic for most of the day. Even saw three Google vehicles in a line, but two were more common.
What's the value of "self-driving" with no weather, traffic or route variation? PR and regulatory filing stats.
Last time cruise made a report about disengagements (where the safety driver had to take over) it was one every 5000 miles driven. That's 1 crash the car was prevented from every 5000 miles. Humans crash once per 70000 miles.
> I bet if you did a study and ran the numbers, it would be cheaper to put down rail on existing roads and make them rail-car only, automate those and subsidize rail cars.
Lane miles in the US are in the many millions. Each mile of track costs a million+ dollars. You're talking many trillions of dollars to railify every road in the country.
So no, there's basically no way that's cheaper than figuring out self-driving.
It is so strange that people don't see what is happening at Tesla. They are driving billions of miles under AI control.
It is totally fine to dislike Musk or to think that autopilot is unsafe today, but how can you look at billions of miles and say that it isn't going to work within the next few years.
There is a long long tail of edge cases that kill people.
When drivers go a few trillion miles per year in the US, any anomaly in your driving software is going to be measured in deaths per hour. There is a built-in human sympathy for accidents caused by people which will not be at all present with accidents caused by machines. You can't sure a maker for wrongful death by a faulty human, you can easily sue a car company.
It will just take one cute kid getting killed by an errant robot car to end the dream completely. Machines don't make human mistakes, they make mistakes that are alien and frightening, things you couldn't ever a person imagine doing.
The public will get scared and automatic driving will be banned.
For Tesla and everyone else doing unattended driving, it isn't a matter of covering the last few bits left, it's a matter of increasing precision by several factors of ten, and ultimately, I'm guessing robot cars won't be a reality until their intelligence starts to reach the point where people are asking questions about sentience.
But with level 4 + remote assistance, you don't need to support all the edge cases to have a car with no steering wheel. You only need to reduce the need for remote assistance enough to make the economics work out.
We will all have been riding around happily in cars without steering wheels for decades before "true self driving" comes to pass.
It's possible that remote assistance centers will never shut down entirely, even 100 years from now, still waiting on the off-chance that a .00000000001% edge case pops up that the system still can't handle.
Can you explain this more thoroughly? So the AI cannot itself resolve a situation that requires resolution within a split second, therefore it passes control (reliable, latency free?!) to a remote operations center where humans get to... assess the situation in order to remotely tell the vehicle how to react in time? I must have completely misunderstood, because I don’t see how any part of that plan is working?
The idea is that such situations are non life threatening. The vehicle will pull over and wait for a human operator.
It's not crazy to imagine that computers could handle split second decisions better than humans but choose to bail out to the side of the road if there's something they can't handle.
Tunnels are perhaps the only "no stopping" situation where there is nowhere to pull over. Luckily, tunnels are also the absolute simplest self-driving scenario.
Tunnels have to kept clear of obstacles. If an SDC would have to stop, so would a human-driven car.
Why rails? With some of this self driving tech you could just put sensors in the road and have the vehicles follow pre-determined routes. I've always thought that's where this self driving tech should have first gone in the early stages. Instead of having a large bus come by once every 30 minutes you could have a fleet of small self driving vans coming by every 5 minutes. If you don't have to pay so many drivers that could be cost effective. Something like that could change public transport and would be very accessible to even small cities.
Agreed, I was thinking the self driving vans or mini-buses would hold maybe 10 people. They wouldn't have to be as big as a full size bus as they would come by more often, making them more convenient for the riders.
I live in a smaller city and we have full size buses that only come by every half an hour, which means if you miss a connection it could take you an hour and a half to travel a distance you could easily drive in 15-20 minutes. If the buses came by more often, more people would ride them, but that would be cost prohibitive if you had to pay all the drivers.
Speechless? Where I live we have full size public buses that are frequently only carrying one or two people. In fact if I had a nickle for every time I've seen a completely empty light rail train in the middle of the day in Denver, I'd be a rich man by now. Public transport in the vast majority of the US is a failure. In Colorado Uber and Lyft probably haul 10-100 times more people every day than ride our public transportation, that's not efficient at all. The main complaint people have is public transportation doesn't run often enough, or provide enough routes, self driving vehicles could fix that and make it economic for US cities. If you had smaller buses (sometimes called vans) you could run them more often and thus make them more appealing to the masses. Is it less efficient, yes, but if it gains more riders that's a net gain IMO.
I think the main complaint about US public transit is that it's dirty and dangerous. If it was super safe and clean in+around all stations and on all trains/buses at all times, people would definitely use it more often.
Why are you comparing your scheme to a defunct system, then? The cities in Europe I’ve lived in, and the cities in Asia I’ve visited, haven’t had that problem at all.
Because there is a market there, actual money to be made, that would be a lot more realistic for companies like Cruise and Waymo to solve than some pie in the sky self driving car. They could have been working towards a public transportation system that doesn't require millions of dollars in capital construction costs to put in place first. But that's not as sexy I guess.
It would be nice and easy just putting some sensors or special paint on the road, the problem is when those automatic vans share the same space than the rest of vehicles, you will have a crash frequently, in barcelona the light rail has crashed with several vehicles even when they don’t share the platform but on the intersections with crossing streets
Agreed, the goal isn't to replace transit in places that are well suited to it, its to provide better transportation in all the places transit isn't well suited.
It's hard to find out how much has been invested in self driving cars, but I think you're massively underestimating the cost of building rail in the US. A billion dollars a mile is a realistic figure for rail in cities. NYC spent more than double that recently.
For those costs, you're probably getting somewhere between 5 and 100 miles of rail.
A billion dollars per mile of rail is an excessively overpriced cost of construction that ought to have severe political pressure to explain why so much money is being wasted.
Construction cost of subways outside of places like NYC that apparently have no concept of cost control is typically about $100-300 million/mile.
Lol. The reason costs are so high is because there is severe political pressure to keep them that way from contractors, unions, etc, and we're rich enough to afford it.
For context, we have 4 million miles of road in the US. So even with $100m/mile, we're talking trillions of dollars before we even get to the actual feasibility of putting rail everywhere.
Rail is great, but there's no world where it does the same thing self-driving vehicles would.
This isn't quite a fair comparison; many of those roads are running through areas with low land values. Comparing ground-level rail through cornfields in Oklahoma is apples and oranges to a busy, high-capacity subway underneath Manhattan.
Underground stations is the majority of the cost. The tunnel boring machine is relatively cheap compared to securing enough property to build stations, and it could be said that New York's recent subways have massively overengineered their stations.
That being said, elevated railways have always been a very hard sell in dense areas, and ground level has its own host of problems.
> I bet if you did a study and ran the numbers, it would be cheaper to put down rail on existing roads and make them rail-car only, automate those and subsidize rail cars
Guess what, these people have run the numbers
> solve any real transport issues compared to traditional mass transit
I think that's thinking small. Why can't self-driving be applied to mass transit.
Let's take out all the cars and replace them with 15 passenger self-driving vans. And set up many stops that are dynamically routed by UberPool. You can get convenience of a car, the cost of a bus and far less pollution.
In my opinion self driving cars far exceed the capabilities of current ML algorithms. I mean I guess you can spend couple of billions and implement some sort of expert system on top of that but it won’t be as simple as just training the ML algorithm.
Much better approach would be using ML as a safety sensor helper and let people drive the cars instead. For example, ML algorithm can detect if a pedestrian is on the road even on low vision conditions and might warn the driver, or check if there a car is acting weirdly on the road and warn the driver to be aware.
I think there is a catch 22 situation going on with self-driving cars. By the time we get to self driving cars at human level, we will already have cybernetic beings at human intelligence level and that will change the entire industries and the society itself, which might actually decrease the profits expected by self-driving car companies.
>I don't think self-driving cars are realistic at all
They said the same thing about a Computer beating a Human at Chess or Go, or Radiology, or countless other examples. I don't see why a random person on hackernews would be any more correct than any of the previous naysayers.
This might be more like turning lead into gold, which after (at least) several hundred years of effort we can do now, with merely completely absurdly uneconomic and unpractical levels of effort and energy.
It's not getting publicized very much but they did lay off people from the core AV teams (planning, controls, perception, mapping, etc.) in addition to the otherwise cited groups.
There’s at least two success criteria for SDC tech:
1) How much safer than humans does SDC need to be before we can ethically use it?
2) When will it be legalized?
The ethics is a hard question. My preference is safety standards should be way higher than the vehicular carnage we accept today. I doubt most people think that way though.
Legalization (and commercialization, especially if corporate liability is waived), could come at any moment. Given the US government accepts 3000 daily COVID19 deaths as a reasonable baseline, it seems there would be little moral qualms. If SDCs are about as good (or maybe even no more than a little worse) than human drivers approval could be granted, ethics be damned.
I feel in the minority in thinking self driving tech will ever be viable. I've worked in computer vision startups and know how powerful the tech can be, but it still can't account with even a fraction of challenging driving conditions. I don't think anything short of general purpose AI could handle it. We are some pretty big discoveries away from general AI.
I also don't think they solve any real transport issues compared to traditional mass transit: https://battlepenguin.com/tech/self-driving-cars-will-not-so...