The issue with self-driving is (1) how it generalises across novel environments without "highly-available route data" and provider-chosen routes; (2) how failures are correlated across machines.
In safe driving failures are uncorrelated and safety procedures generalise. We do not yet know if, say, using self-driving very widely will lead to conditions in which "in a few incidents" more people are killed in those incidents than were ever hypothetically saved.
Here, without any confidence intervals, we're told we've saved ~70 airbag incidents in 20 mil miles. A bad update to the fleet will easily eclipse that impact.
> The issue with self-driving is (1) how it generalises across novel environments
That's also an issue with humans though. I'd argue that traffic usually appears to flow because most of the drivers have taken a specific route daily for ages - i.e., they are not in a novel environment.
When someone drives a route for the first time, they'll be confused, do last-minute lane changes, slow down to try to make a turn, slow down more than others because because they're not 100% clear where they're supposed to go, might line up for and almost do illegal turns, might try to park in impossible places, etc.
Even when someone has driven a route a handful of times they won't know and be ready for the problem spots and where people might surprise they, they'll just know the overall direction.
(And when it is finally carved in their bones to the point where they're placing themselves perfectly in traffic according to the traffic flow and anticipating all the usual choke points and hazards, they'll get lenient.)
You've a very narrow definition of novel, which is based soley on incidental features of the environment.
For animals, a novel situation is one in which their learnt skills to adapt to the environment fail, and have to acquire new skills. In this sense, drivers are rarely in novel environments.
For statistical systems, novelty can be much more narrowly defined as simply the case where sensory data fails a similar-distribution test with historical data --- this is vastly more common, since the "statistical profile of historical cases, as measured, in data" is narrow.. whilst the "situations skills apply to" is wide.
An example definition of narrow/wide, here: the amount of situations needed to acquire safety in the class of similar environments is exponential for narrow systems, and sublinear for wide ones. ie., A person can adapt a skill in a single scenario, whereas a statistical system will require exponentially more data in the measures of that class of novel scenarios.
I travel and drive in a lot of new places and even the novelty of novelty wears off.
At some point you’ll see a car careen into the side of the curb across three lanes due to slick and you’ll be like ehhh I’ll just cut through with this route and move on about your day.
After driving for 20 years, about the only time I got scared in a novel situation was when I was far from cell service next to a cliff and sliding a mountain fast in deep mud running street tires due to unexpected downpour in southern Utah. I didn’t necessarily know what to do but I could reason it out.
I don’t really find “using a new route” difficult at all. If I miss my exit, I’m just going to keep driving and find a U-turn — no point to stress over it.
Remember that what matters is the general driving populace, and there will always be people who drive better and who drive worse.
Also, a very significant portion of drivers overestimate their driving skills, in particular older drivers. Having only been scared once in 20 years would likely make someone lenient and dull their senses as nothing requiring notable effort or attention ever seems to happen to them.
Generalizing across novel environments is optimal, but I'm not sure the bar needs to be that high to unlock a huge amount of value.
We're probably well past the point where removing all human-driven vehicles (besides bikes) from city streets and replacing them with self-driving vehicles would be a net benefit for safety, congestion, vehicle utilization, road space, and hours saved commuting, such that we could probably rip up a bunch of streets and turn them into parks or housing and still have everyone get to their destinations faster and safer.
The future's here, even if it still has room for improvement.
I'd think congestion would go up as AVs become more popular, with average occupancy rates per vehicle going down. Since some of the time the vehicle will be driving without any passengers inside. Especially with personally owned AVs. Think of sending a no-human-passenger car to pick up the dog at the vets office. Or a car circling the neighborhood when it is inconvenient to park (parking lot full, expensive, whatever).
Up to 30% of cars on city streets at any given time are looking for parking [1].
Cars are also the least utilized asset class, being parked 95% of the time [2].
AVs, by virtue of being able to coordinate fleet-wide and ability to park anywhere rather than only one's home or destination, would be able to gain incredible efficiencies relative to status quo.
Atop those efficiencies, removing both the constraint of having a driver and the constraint of excessive safety systems to make up for human inattentiveness means AVs can get drastically smaller as vehicles, further improving road utilization (imagine lots of 1- and 2-seaters zipping by). And roads themselves can become narrower because there is less room for error with AVs instead of humans.
Finally, traffic lights coordinating with fleets would further reduce time to destination (hurry up and finish).
Self-driving vehicles give us the opportunity to rethink almost all of our physical infrastructure and create way more human-friendly cities.
> AVs, by virtue of being able to coordinate fleet-wide and ability to park anywhere rather than only one's home or destination, would be able to gain incredible efficiencies relative to status quo.
> Atop those efficiencies, removing both the constraint of having a driver and the constraint of excessive safety systems to make up for human inattentiveness means AVs can get drastically smaller as vehicles, further improving road utilization (imagine lots of 1- and 2-seaters zipping by). And roads themselves can become narrower because there is less room for error with AVs instead of humans.
The first part is mostly describing taxis, so the incredible efficiencies relative to the status quo can be loosely observed through them. Just subtract out wage and a slight "technological scale" bonus, and you can estimate what it would be. Then add in the expected investor returns for being a technology company and see the improvements disappear.
The second part, I wonder. Cars already average under 2 occupants, with most just being the driver. If this is what is was needed for significantly smaller cars, we would already have them. Lack of smaller cars is mostly a cultural issue, not a technical one.
> Lack of smaller cars is mostly a cultural issue, not a technical one.
And an economic/tax policy issue. Some increases in size are due to legally mandated safety features, while even more of the increased adoption of SUVs in the US is indirectly due to the CAFE standards.
¯\_(ツ)_/¯ I'd like to think there's more efficiency to be gained when the scale is city-wide rather than a small subset of demand being attended to reactively.
Ideally this would be a municipal fleet and transportation just another utility like water, electrical, and broadband. Admittedly this would require strong political power and vision, as anything that remakes physical infrastructure does.
Agree small cars are a cultural/identity issue tho usually a rather rational one as well, given safety vis-a-vis 7000lb SUVs. However, I don't think people's aversions to spend $20k+ on a city-only vehicle has any bearing on whether they would be willing to be being taken places in one when its the most convenient/safest/fastest way to get places. A city-wide transportation utility obviates most of the need/desire for individual car ownership.
To put it in tech biz terms, everything in tech is bundling or unbundling. Ownership of cars is the unbundled version of transport, and took over due to convenience and creature comforts. Now a new tech has come out that swings the pendulum toward bundled being more convenient & optimal.
Why are we using self-driving vehicles as a panacea for historical underinvestment in public transport?
Not saying that they wouldn't play a role in a functional public transport system, they'd be invaluable for the last two miles from your station to your destination.
But while our people transporting systems prioritise roads and cars, we will never have the high quality and safe public transport that high quality of life cities thrive on.
(And while I write this from NZ, with only limited experiences of LA and SF, we copied America, we went for sprawl and freeways, and it's strangling our largest city.)
I know and spend time with people who live in Berlin, Munich and Hamburg, that don't own a car, because they don't need to own a car.
They might rent one for a holiday into Italy, or they might use an app like Lime / Bird etc. to rent very short term a tiny car like a BMW i3 for a big grocery shop.
But because their cities are dense, and mix commercial with residential (e.g., ā bunch of 5 storey apartment buildings with the ground/first floor being commercial, depending on where you are), they can often buy groceries at the local market on foot on their way home from the U-Bahn, or head down to the local Getränkhandel on a bike with a basket or two to buy their beer and bottled water.
Centralising commerce away from residential, especially with big box shopping areas, is predicated on car culture, and bakes in the need for cars.
TL;DR self-driving vehicles alone are a band-aid over an unsustainable transport culture and strategy.
But they'll form a critical part of a sustainable one.
The reason we need self-driving cars is because we need cars. The reason we need cars is because of the way our cities are structured. The reason our cities are structured the way they are is due to large land availability & zoning laws, leading to massive spread.
Public transit will work for some of the people some of the time. (That is, if it can even be built - highly recommend Ezra Klein's new book Abundance on ways to get out of this).
For the people that public transit won't work for, you need to come up with a new solution if you want to see cars go away: a solution that makes going from A to B some combination of easier, cheaper, faster, more convenient than driving. Or, a solution that brings B closer to A (like changing zoning laws or building cheaper housing in metro areas).
You missed a critical (last?) step in your reasoning: The reason we have massive spread is because most humans want as much living space as they can possibly acquire and our laws and social norms reflect that.
Hear me out, I've spent years thinking about this :)
Yes, past underinvestment is bad. And yes, initially they are a band-aid, until yes they do become critical.
My excitement for self-driving tech isn't about the short term changes, but just how powerful a technology this is in the longer term. Ultimately this tech is not about cars, it's about the ability to automate the movement of mass. This is novel and meaningful.
An obvious medium-term implication of self-driving is that cities will ban human drivers, because that way cities can ditch a bunch of high-cost infrastructure required because of human fallibility. Up until that point, self-driving would be a band-aid. After that point, the dominoes start to fall.
1. Form factors change: cars become 1-4 person pods, stripped of the unnecessary bulk of excessive safety systems and unused capacity.
2. Ownership changes: municipalities will buy fleets of cheap mass-produced pods to replace extremely capex intensive public transport.
3. What is transported changes: now you have shipping drones dropping off standardized (reusable) packages into standardized intakes. Think The Box [1] but smaller.
4. Infrastructure changes: Roads narrow, parking becomes drop-off spots, larger cafes, actual parks. Cut and cover roads multiply, leaving more space above ground for people. Cities grow 20% without getting bigger, just by obviating the need for half their roads. The blight of various parking signs and warnings to drivers disappear. People can walk about freely or ride their bikes. It's quieter. The air quality improves.
5. Housing changes: Garages transform into rooms. People ditch bulky refrigerators in favor of ordering drone-delivered fresh produce in minutes. Drones deliver upstairs not just at street level. Pods become elevators. We've seen all this in science fiction... guess what the enabling technology is?
If you extend the implications of the automated movement of mass, the logical conclusion is the physical infrastructure of the city will transform to take advantage of every gain that creates. Cities dedicate 25-40%+ of their land mass to roads. In dense urban cores, 20% of their land mass is just parking spots. We can't route people-driven cars underground unless we really really mean it and build a highway. We waste a huge amount of space on transportation. We also shape all of our buildings around the constraints imposed by car-shaped objects and all their various externalities, including noise and air pollution.
My belief is that self-driving is easily the most transformative tech to hit cities since the car, and may exceed the impact that cars have had on the built world.
> 2. Ownership changes: municipalities will buy fleets of cheap mass-produced pods to replace extremely capex intensive public transport.
Producing more small things isn't usually more efficient that fewer equivalent large things. You can't just will some "pods" into existence that are magically cheaper (per person!) than trams, trains and busses. Also, once you have a system running capex and opex aren't that different - replacing a set number of vehicles per year is pretty much the same thing as operating expenses.
> 5. Housing changes: Garages transform into rooms. People ditch bulky refrigerators in favor of ordering drone-delivered fresh produce in minutes. Drones deliver upstairs not just at street level. Pods become elevators. We've seen all this in science fiction... guess what the enabling technology is?
My prediction is that no one will ever be fine with the amount of noise a "drone" (read helicopter) makes, especially as a replacement for the very noise-free and orders of magnitude more efficient elevators we have right now.
Agree to disagree about economies of scale, but FYI busses are ~$500k and seat ~40, meaning ~$12.5k/person. There are a dozen manufacturers of electric 2 seaters today that can build for a quarter of the cost per person, or half if you assume 1 person occupancy. Yes the area per person is larger (tho not by much), but you can make up for that with increased throughput by way of point-to-point operation without stops, faster speeds, and more.
Focusing on rollout, municipal light rail almost never gets deployed in US-style cities due to huge capex, not opex. Smaller vehicles allow incremental roll-out and can use preexisting road infrastructure. Ergo, that's the form of public transit you're most likely to see grow over the next decades.
Drone here doesn't imply flying, it's about scaling down wheeled vehicles and the coexistence of a wider variety of vehicle sizes on roads that is unlocked by the automated movement of mass. Delivery to higher up floors can be done through small in-building elevators. If you think that's unrealistic, consider that it was once extremely popular to use pneumatic tubes to send mail in buildings. Built infrastructure changes based on what is possible, and mass needs to move.
I can't imagine any of that since it relies on >99% of cars being self-driving whereas currently <1% are. Even under the most optimistic estimates, how many decades would it take to get to that? 5? 10?
> Self-driving vehicles give us the opportunity to rethink almost all of our physical infrastructure and create way more human-friendly cities.
Ok, that's just giving me a stroke. We already have that. It's called public transport, walkability, bikeability. These have the upside of being extremely well understood and use technology that's available today. We could start seeing benefits within a few years, not decades.
Even in your dream scenario, 50 or so years from now, cars would still have a lot of the same downsides they have today of using way too much space and causing way too much pollution per person for the utility they provide.
I live carless in NYC and primarily use a mix of public transit and cycling. That being said, infrastructure costs so excruciatingly much in the US that it would cost tens or hundreds of Waymos (company, not individual car) to replicate an NYC subway-tier system in most American metropolitan areas.
Forcing people to take public transit that is any worse than NYC subway will definitely and rightfully lose an election for that party. Building such a system at modern American construction costs will also lose an election. What is left to do but embrace autonomous driving as the first step toward retrofitting American cities to be slightly more people friendly?
Besides, this is the decadent west, we can afford for people to use more resources for more comfort. Even the well off in china have embraced cars as mobile living rooms.
> I can't imagine any of that since it relies on >99% of cars being self-driving whereas currently <1% are
Technology famously has a linear adoption curve, and convenience is famously not something that drives adoption /s
> We already have that. It's called public transport, walkability, bikeability.
Do we have that though? In the US, mostly not. So what's the path? Hoping that sprawled out cities somehow magically get the political will to build $billions in light rail? What do you think is the path of least resistance to these goal states?
> Even in your dream scenario, 50 or so years from now, cars would still have a lot of the same downsides they have today of using way too much space and causing way too much pollution per person for the utility they provide.
Read other comments, don't get stuck on the notion of 'cars' as-is.
Every city street would have 4 lanes without the need for free car storage. Maybe make one of them a bike lane or widen the sidewalks and have 3 lanes for 30% more capacity. Also, traffic engineering could be optimized to a much greater extent since you wouldn't have to worry about all the affordances which keep humans from getting confused, keep them from getting aggressive, keep them from speeding, etc. Also, most congestion is caused by drivers causing turbulence by switching lanes, stopping each other from switching lanes, getting in the wrong lane, etc. A city of only AVs would probably flow much more smoothly.
I don't agree with this novel environment argument about routes. As a human, there are a limited number of roads that I have driven on. A taxi driver drives better than me because none of the routes are considered novel: the taxi driver has likely driven on every road in a city in his/her career. The self-driving machine has most definitely driven on every single road in the city, perhaps first as testing with human backup, then testing with no passengers, and finally passenger revenue miles.
I think you underestimate how many novelties the car will encounter on existing routes and how adept these cars are at navigating novel routes.
I imagine this route data is an extra extra safeguard which allows them to quantify/measure the risk to an extent and also speed up journey's/reduce level of interventions.
I wonder if you can decrease the impact of (2) with a policy of phased rollout for updates. I.E. you never update the whole fleet simultaneously; you update a small percentage first and confirm no significant anomalies are observed before distributing the update more widely.
Ideally you'd selectively enable the updated policy on unoccupied trips on the way to pick someone up, or returning after a drop-off, such that errors (and resultant crashes) can be caught when the car is not occupied.
One measure of robustness could be something like: the ability to resist correlation of failure states under environmental/internal shift. Danger: that under relevant time horizons the integral of injury-to-things-we-care-about is low. And then "safety", a combination: that the system resists correlating failure states in order to preserve a low expected value of injury.
The problem with machines-following-rules is that they're trivially susceptible to violations of this kind of safety. No doubt there are mitigations and strategies for minimising risk, but its not avoidable.
The danger in our risk assessment of machine systems is that we test them under non-adversarial conditions, and observe safety --- because they can quickly cause more injury than they have ever helped.
This is why we worry, of course, about "fluoride in the water" (, vaccines, etc.) and other such population-wide systems... this is the same sitation. A mass public health programme has the same risk profile.
You would save more lives by harshly punishing drunk or influenced driving; however, most of the lives you save would be that of the drinker or the abuser.
You would save more lives by outlawing motorcycles; however, it would just be the motorcyclists themselves.
Another thing people don't consider is that not all seats in a vehicle are equally safe. The drivers seat is the safest. Front passenger is less safe but still often twice as safe as sitting in the backseat. If you believe picking up your elderly parents and then escorting them in your backseat is safer than them driving alone you might be wrong. This is a fatality mode you easily recognize in the FARS data. Where do most people in a robotaxi sit?
Your biggest clear win would be building better pedestrian infrastructure and improving roadway lighting to reduce pedestrian deaths.
We've been improving front seat safety systems for years while not adding much in the back seat. The result is obvious in the fatalities data and many institutions have involved themselves in this problem. Here's one:
There are _so many_ bad assumptions about vehicle safety it honestly drives me nuts. Especially on Hacker News. The data is available from NHTSA in a database called FARS. I encourage everyone to go look through the data. You almost certainly believe several wrong things about driving and fatalities.
I think Elon Musk is exceptionally irresponsible for using these statistics in a flatly dishonest and misleading way. He wants to sell vehicles not truly educate you about safety. People should double check.
This is a really good link. Will definitely refer to this when making a car purchase in the future. Thanks!
It seems like Volvo's reputation as one of the safest car is still well deserved after all. I don't own a Volvo--too expensive for me, but good to know.
> In safe driving failures are uncorrelated and safety procedures generalise. We do not yet know if, say, using self-driving very widely will lead to conditions in which "in a few incidents" more people are killed in those incidents than were ever hypothetically saved.
The only way for this to cause tens of thousands of death by self-driving alone is for people to suddenly need to drive the cars themselves, and not being able to do it well. Unless I'm missing something.
I usually think about it in the other direction: every time an accident occurs, a human learns something novel (even if it be a newfound appreciation of their own mortality) that can't be directly transmitted to other humans. Our ability to take collective driving wisdom and dump it into the mind of every learner's-permit-holder is woefully inadequate.
In contrast, every time a flaw is discovered in a self-driving algorithm, the whole fleet of vehicles is one over-the-air update away from getting safer.
I can imagine whole city areas well known closed to manual drivers.
Sure I would love to read a book while car is driving me to visit family in the countryside but practically I need city transportation to work and back, to supermarkets and back where I don’t have to align to a bus schedule and have 2-3 step overs but plan my trip 30 min in advance and have direct pick up and drop off.
If that would be possible then I see value in not owning a car.
I was waiting for a Waymo in Austin during the weekend storm and the Waymo suddenly cancelled on us right after a power outage that lasted a second or two. According to local news the vehicles had stopped and were blocking traffic.
> The issue with self-driving is (1) how it generalises across novel environments without "highly-available route data" and provider-chosen routes; (2) how failures are correlated across machines.
Consider London: a series of randomly moving construction sites connected by patches of city.
Waymo, as far as I recall, relies on pretty active route mapping and data sharing -- ie., the cars arent "driving themselves" in the sense of discovering the environment as a self-driving system would.
Waymo's map data is a prior, not an authoritative reference to the world. The cars report update when the map data is wrong, and a lot of what they use it for (e.g. traffic light identification, fine localization) degrades gracefully when there's new information in the environment.
Sure, my understanding is that they collectively share data, and this is combined with central mapping.
On net, yes, they are sensitive to features of the environment and via central coordination maintain a safe map of it.
The mechanism there heavily relies on this background of sharing, mapping, and route planning (and the like) -- which impacts on the ability of these cars to operate across all driving environments.
There was a problem in SF with a Waymo not avoiding a parade. Somebody didn't enter the parade in SF MUNI's list of street closures. Nobody was hurt; it was just embarrassing.
Any time there is a detour, or a construction zone, or a traffic accident, or a road flooded, or whatever else your route data is not just "worse" it is completely wrong
it's true, but if the scenarios are rare you can treat them as outliers and have the vehicle switch to a "safety mode" or similar. Even human drivers effectively do that (or should!).
machine don't make mistake when they are get perfected in certain route, sure human drive would be better in dynamic areas but you dnt need machine to be perfect either just want (80% scenario)
The issue with self-driving is (1) how it generalises across novel environments without "highly-available route data" and provider-chosen routes; (2) how failures are correlated across machines.
In safe driving failures are uncorrelated and safety procedures generalise. We do not yet know if, say, using self-driving very widely will lead to conditions in which "in a few incidents" more people are killed in those incidents than were ever hypothetically saved.
Here, without any confidence intervals, we're told we've saved ~70 airbag incidents in 20 mil miles. A bad update to the fleet will easily eclipse that impact.