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.

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.

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.

Either way it's a win.

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.

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.

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.

https://www.denenapoints.com/relationship-vehicle-weight-roa...

Non-driver != pedestrian. Non-drivers include car passengers.

The main person I put at risk when driving fast is me. Others too, sure, but mostly me.

So it's not at all bout killing "other lesser people". People mostly just think some risks are worth the reward.

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.

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.

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.

Source: https://www.propublica.org/article/unsafe-at-many-speeds

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.

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.

what?

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.

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.

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.

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.

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.

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.

[1] https://www.vox.com/2015/1/15/7551873/jaywalking-history

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.

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.

How do you explain actual self-driving cars on the road? I find this attitude confusing as it seems to not jive with empirical reality.

I agree that we should focus on mass transit, but Waymo drives hundreds/thousands of miles without driver intervention.

I've never heard of "expectancy value" in my life - it's definitely not a standard probability term.

Waymo cars go 11,000+ miles without driver intervention on average.

I would encourage doing actual research into this subject, rather than relying on your gut or the news articles you read.

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.

There was a driver, but I'm fairly certain that the drivers are not actively driving but are legally required to be there.

What? That's not at all what disengagements mean

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 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.

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.

I'm sure they'll get there eventually, but having a ton of data isn't a cheat code that makes rapid progress inevitable.

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.

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 have to kept clear of obstacles. If an SDC would have to stop, so would a human-driven car.

For the others, stopped cars are very common.

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.

For those costs, you're probably getting somewhere between 5 and 100 miles of rail.

Construction cost of subways outside of places like NYC that apparently have no concept of cost control is typically about $100-300 million/mile.

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.

Where within the United States is this true? (note: United States is known for high construction costs)

e: or just downvote me - but I don't know anywhere with a $100 million/mile subway system.

That being said, elevated railways have always been a very hard sell in dense areas, and ground level has its own host of problems.

Just level 3 or 4 autonomous vehicles would probably be good enough to be worth buying.

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.

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.

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.

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.