This tells me that we’re still a long way from full level 4 (and certainly level 5) autonomy in a busy city like San Francisco. The edge cases requiring immediate human attention are still too frequent for the human safety driver to be remote, as is the case in Phoenix.

Also, just a reminder that Waymo in Phoenix is nowhere close to being level 5, since it is still heavily geofences and requires those remote safety monitors. I still think that true level 5 (i.e. ability to drive autonomously everywhere with zero human oversight with a safety record equivalent to the median human driver) requires AGI. Would love to be proven wrong!

Did you expect something different? I can't really see a boardroom writing a roadmap that goes straight from

  test rides (no passengers) with a backup driver onboard

  actual rides with passengers - no backup driver onboard

Back in 1999 the Chinese government announced that airline execs would be airborne at the changeover as reassurance that aircraft were safe from Y2K. Like or hate them, the incentive logic was sound.

https://www.wired.com/1999/01/y2k-in-china-caught-in-midair/

https://www.cbc.ca/news/science/chinese-airlines-won-t-be-bi...

Historically, people packing parachutes could be randomly selected to jump with the parachute they had packed.

How do you know board members aren't already using Waymo product heavily? It still doesn't mean they can go straight from board members using Waymo without backup drivers to arbitrary customers using Waymo without backup drivers.

I'm guessing this is the old joke about Eric Schmidt?

The median modern bicycle has 0 computers and sensors.

[1] https://www.ceinetwork.com/cei-blog/auto-computers-means-com...

Of course the book showed some failure modes for that, but I wonder if network coverage and latency, as well as "backup driver response time" could be considered good enough, perhaps this sort of model could have an acceptable risk trade-off.

[1] https://en.wikipedia.org/wiki/Halting_State

Having customers paying for the R&D will help make it sustainable.

basically all of Italy

I don't believe people are using their full AGI when driving (and the full "AGI" may as well happen to be a set of basic pattern matching capabilities which we haven't discovered yet). After decades of driving the behavior is pretty automatic, and when presented with complex situation following a simple rule, like just brake, is frequently the best, or close to it, response.

As things stand emergency vehicles have to cope with a reasonable minority of people who completely panic and actually impede their progress.

Or hacked maliciously.

An incomplete list includes:

1) Integrating all the camera views into one 3-D vector space before training the neural network(s).

2) A large in-house group (~1000 people) doing manually labeling of objects in that vector space, not on each camera.

3) Training neural networks for labeling objects.

4) Finding edge cases where the autocar failed (example is when it loses track of a vehicle in front of it when the autocar's view is obscured by a flurry of snow knocked off the roof of the car in front of it), and then querying the large fleet of cars on the road to get back thousands of similar situations to help training.

5) Overlaying multiple views of the world from many cars to get a better vector space mapping of intersections, parking lots, etc

6) New custom build hardware for high speed training of neural nets.

7) Simulations to train rarely encountered situations, like you describe, or very difficult to label situations (like a plaza with 100 people in it or a road in an Indian city).

8) Matching 3-D simulations to what the cars cameras would see using many software techniques.

Yeah, a Tesla couldn't possibly drive into a stationary, clearly visible fire engine or concrete barrier, on a dry day, in direct sunlight.

I dislike saying anything in defense of tesla's self-driving research, but let's be accurate.

At least, I never would...

A million people are killed globally each year by motor vehicles. Staggering amounts of pain and injuries. Massive amounts of property damage. Tesla's cars are not supposed to be left to drive themselves. The chance to save so much carnage seems worth letting some people driving Tesla's, that fail to pay attention to the road, suffer the consequences of poor decisions.

Plus these problems are likely too be mostly fixed due to the fact that they happened.

Start-up religion doesn't really work when there are lives on the line. That's fine for your social media platform du jour but please don't bring that attitude to anything that has 'mission critical' in the description. That includes medicine, finance, machine control, traffic automation, utilities and so on.

Not to say that we shouldn't be cautious here, but over-caution kills people too

These things are excellent - undeniably better than humans at the boring stuff, highway driving, even major roads. They can rightfully claim massive mileage with high safety levels in those circumstances... but throw them into nastier conditions where you have to understand what objects actually are and things quickly seem to fall apart.

Waymo drove 29,944.69 miles between "disengagements" last year. That is an average California driver needing to touch the wheel once every 2.3 years.

Tesla by comparison is classed as a SAE Level 2 driver assist system and isn't even required to report metrics to the state. While they sell it to consumers as self-driving, they tell the state it is basically fancy cruise control.

What you want is "interventions". Additionally, look at where those miles were driven. Most of them are some of the most simplistic road driving scenarios possible.

From my experience of California driving, that doesn't sound too bad. Compared to the entire Eastern seaboard, y'all have great roads and better drivers.

You know better. If most of those miles were in sunny Mountain View suburbs, they don't count.

As I understand it, if you pay for FSD, you don't actually get anything like self-driving, you just get lane-changes on the highway in addition to the lane-keeping. Effectively, you get lane-keeping, which you have if you don't pay too.

All the videos of "FSD driving" are from a small number of beta-testers, and there's no way to opt into the beta.

Because of that, my assumption would be very few people on here have experienced tesla's self-driving. It's only open to a small number of beta testers, whether you have purchased it or not.

On the other hand, waymo is available for the general public to use, though only in specific geographic areas.

https://www.cnn.com/2021/08/16/business/tesla-autopilot-fede...

While we can (and should) hold autonomous vehicle developers to a much, much higher standard than we hold human drivers, it is precisely because of excellence.

It is also possible that if it develops enough it will succeed in ways that a human cannot, such as extremely long monotonous cross-country driving (think 8 hour highway driving) punctuated by a sudden need to intervene within seconds or even milliseconds. Humans are not good at this but technology is. Autonomous cars don't get tired or fatigued. Code doesn't get angry or make otherwise arbitrary and capricious decisions. Autonomous cars can react in milliseconds, whereas humans are much worse.

There will undoubtedly be more accidents if the technology is allowed to develop (and I take no position on this).

An autopilot is a system used to control the path of an aircraft, marine craft or spacecraft without requiring constant manual control by a human operator. Autopilots do not replace human operators. Instead, the autopilot assists the operator's control of the vehicle, allowing the operator to focus on broader aspects of operations (for example, monitoring the trajectory, weather and on-board systems).

3000 people die every day in automobile accidents, 10% of which are from people who are sleepy. Even standard autopilot is better than a tired driver

It's not an easy problem to fix at high speed without false positives, and they seem to really hate false positives.

Every year, I'm amazed at how quickly our personal "veneer of civilization" collapses in the snow.

The prior owners of our home would just keep their kids home from school, and work from home an average of "about a week every winter."

We're a little more aggressive with snow removal, but there are still mornings every winter where I'm getting up at 5 to spend a couple hours plowing and blowing out drifts on our driveway (after typically doing the same thing the night before) just in order for my wife to make it down to our county road which might still have a foot or so of snow covering it.

Similarly, in windy snow-covered conditions, there are a couple spots between us and town where the snow regularly drifts back over the road in a matter of hours, causing a "well, I know the road goes about here, I think I can make it through these drifts if I floor it so here it goes" situation.

Even when the roads are well plowed and clear, there are plenty of situations where it's difficult for me, a human, to distinguish between the plowed-but-still-white-road and the white snow all around it in some lighting conditions.

And let's take snow out of it. Our fastest route into town involves gravel roads. And our paved route is chip-sealed every couple years, and typically doesn't get a divider-line drawn back on it for 6-months or so after.

Which is all to say, I think it's going to be quite a while before I have a car that can autonomously drive me into town in the summer, and global warming aside, I'll probably never have one that can get me there reliably in the winter.

Sure it will, at least eventually. However, I suspect the humans at the time won’t like the answer: that it’s not always safe to drive in these conditions, and then the car refusing to drive autonomously, even if it is technically capable of navigating the route. It may deem the risk of getting stuck, etc. to be too high. Or you may need to accept a report to your insurance company that you’ve opted to override the safety warnings, etc.

Lastly, not the OP, but winter is my favorite season for the most part, and I love being around lots of snow!

Level 0 is no automation, level 1 is just a dumb cruise control, level 2 is radar adaptive cruise control plus lane keeping (which is where most production systems like Tesla Autopilot and GM Supercruise are currently at). Level 2 still requires full human supervision, if you engaged it on the road above it would either fail to engage or you'd crash and it would be your fault. Level 3 is the same plus an ability to handle some common driving tasks, like changing lanes to pass a slower vehicle.

Level 4 is where it gets really interesting, because it's supposed to handle everything involved in navigating from Point A to Point B. It's supposed to stop itself in the event of encountering something it can't handle, so you could theoretically take a nap while it drove.

However, an important limitation is that Level 4 autonomy is geofenced, it's only allowed in certain areas on certain roads. Also, it can disable itself in certain conditions like construction or weather that inhibit visibility. Waymo vehicles like these are ostensibly level 4, if you tell them to drive through a back road in the snow they'll simply refuse to do so. It's only useful in reasonably good conditions in a few big cities.

Level 5 is considered to be Point A to Point B, for any two navigable points, in any conditions that the vehicle can traverse. You could build a Level 5 vehicle without a driver's seat, much less an alert driver. I kind of think this will require something much closer to artificial general intelligence; level 4 is just really difficult conventional programming.

And if you have lane-keeping but not cruise control, that's also level 1.

The difference between 1 and 2 is weird.

Strangely I've never seen Waymo vehicles practicing on that. They're all over Mountain View, but I have never once seen them in the mid-peninsula hills.

I’d gladly buy a self-driving car that require some additional input on such a road and had additional aids to spot oncoming traffic I can’t see behind the tractor that’s a few hundred meters forward of the spot linked to. It would still be safer.

To really make things work, we need cars to be able to negotiate the way humans do on the right of way, etc. There is a lot of non-verbal (and when that fails, very verbal) communication while driving. Currently, cars can’t communicate with each other and the pedestrians, which limits possibilities a lot.

My grandpa who drives on those roads primarily, sweats bullets in the city.

Maybe you’ll have different driving models to load in different scenarios …

Spinning out on a deserted highway and hitting a snowbank and getting trapped in your car kills a large number of people every year. Even with smartphones, calls for help can't always be responded to in time, resulting in death. (Have an emergency kit in your car if you live above the snow line!)

Driving in city traffic can be quite harrowing, but hitting another car at 20-30 mph isn't usually fatal. (Wear your seatbelts!)

The point that GP post was trying to make is that humans have different preferences, and what seems dangerous to one doesn't (and possibly isn't) dangerous to another. Humans are also notoriously bad at judging danger, eg some people feel threatened by the idea of wearing of papers masks.

Secondly, its dumbest errors have to be better than what the average human thinks their dumbest errors would be. If there is an accident and every one thinks they would never have made this error, it will crush the acceptance.

Looking at the general accidents stats and saying to people that, on average there are less deaths on the road but they might die in a stupid accident they would never have been into, had they been driving themselves, will be a very hard pill to swallow. Most people prefer to have the illusion of control even if statistically it means worse expectations.

Add to all this one thing and we're further than I think most people realize: Weather. Show me an FSD doing better than a human in the snow or we're not really anywhere yet.

That said, I agree with others that this is the natural progression of testing rollout and doesn't tell us anything about the pace at which the rollout will occur, in particular whether it will be faster or slower than Phoenix.

[1] https://www.dmv.ca.gov/portal/file/adopted-regulatory-text-p...

[2] https://www.cpuc.ca.gov/regulatory-services/licensing/transp...

By the time L5 arrives people will have been happily riding around in vehicles with no steering wheels for decades. L4 cars that phone home less and less every year.

Eventually someone will notice that no L4 car has phoned home for a whole year and almost nobody will care. Just a footnote to an era that already feels taken for granted.

How many of those cars have caused a traffic jam at a given moment because they’ve encountered a traffic cone? How long does each issue take to resolve? It seems like, in addition to the technical hurdles, there are many more logistical hurdles before this can be rolled out at scale.

This might be true. Most of the time (95%) I am on complete human brain autopilot when I’m driving but those other 5% need my full focus and attention. I shut of the radio and tell other passengers to be quite (if I have the time for it).

Put another way, the 5% of times I need to focus are usually the times where I am somewhere new and don’t necessarily understand the road layout - which something like Waymo may avoid through mapping for instance.

It might be true, but plenty of problems that have been thought to require true AGI have later been found to not require it after sufficient research - for example it’s not long ago that we thought good image recognition was entirely out of reach.

Because they are not even trying to be level 5. They've made it very clear that will only ever be a level 4 company and level 5 is not feasible.

L4 is enough to be viable and safe, and is all that is needed.

In fact this level crap is bullshit. It's the speak of MBAs at Bain and McKinsey at who think they understand tech, not engineers.

Real engineers don't stare at their debugging screens going "check out this data, is it L3 or L4?"

Instead engineers look at things like safety-critical interventions per kilometer, non-critical interventions per kilometer, accidents per kilometer, etc.

So how do you explain, succinctly, the difference between a car that can hold 55 mph (but do nothing else automatically) without driver intervention and one that can change from 25 mph to 55 mph and change lanes without driver intervention.

The differences between levels are drastic in terms of implications on overall utility of the technology. There's a reason the terms are used.

> L4 is enough to be viable and safe, and is all that is needed.

Needed for what exactly? Each level has its own benefits. That's exactly the reason the levels were established. I want to be able to get into a car, put in an address and fall asleep for the whole trip. When do I get that? L4? L5? If I can't do other tasks while the car is driving, then the point of full autonomous is essentially pointless.

> not engineers.

Precisely. The levels have nothing to do with engineers. It's about understand the benefits for the population. For example - can I fall asleep in the back of the Tesla autonomous car? No, because it's not L5. That's the point.

It doesn't really matter so long as there are fallback drivers in all situations. An lower level autonomy car would remotely connect to the backup driver (in a third world call center) several times per hour in order to behave like a higher level one.

A higher level one would connect more rarely. The only difference I notice as a customer is I pay more out of the box for the higher level car and less as a subscription , while the lower level car costs less out of the box and has a higher subscription fee because the human costs are higher.

https://en.wikipedia.org/wiki/Self-driving_car#Classificatio...

So when people talk about a particular company "achieving" level 4 or level 5, I don't know what they mean. Maybe they mean achieving it "safely" which is murky, since any system can crash. Maybe they mean achieving it legally on public roads, in which case, it's a legal achievement (although depending on what regulatory hoops they had to go through, maybe they had to make technical achievements as well).

[1] : https://web.archive.org/web/20161120142825/http://www.sae.or...

Not really. L5 is impossible, period.

What I think will happen is L4 with 99.999% cases covered and have it come to a safe stop for the 0.0001%, assuming there was a way to safely stop.

L5 which means 100.000% covered, will not happen, but the PR people will continue to use the term.

Agreed.

> L5 which means 100.000% covered, will not happen, but the PR people will continue to use the term.

Which is precisely why so many people are critical of the term "fully autonomous".

> 99.999% cases covered

What cases? The point is that edge cases are the issue with autonomous driving. I can fall asleep on a train or a plane because I know there is human conductor who can handle the edge cases. This doesn't exist with L4. Everything else that doesn't let me fall asleep (read a book, look at my phone, etc.) is only marginally better.

> assuming there was a way to safely stop.

That's a pretty damn strong assumption.

No communicating with a server to download maps, no perfect performance, just a car that knows the state's traffic laws driving a brand new road in any reasonable weather, and getting into less crashes than a human would.

Really? Because the L5 claims come more out of the Ubers and Teslas than the "MBAs".

I don't think any company wants to sign off on the notion that their software will handle all classes of problem, even ones they have no data for at all.

That point is taken into account under J3016_202104 § 8.8:

“There are technical and practical considerations that mitigate the literal meaning of the stipulation that a Level 5 ADS must be capable of ‘operating the vehicle on-road anywhere that a typically skilled human driver can reasonably operate a conventional vehicle,’ which might otherwise be impossible to achieve. For example, an ADS-equipped vehicle that is capable of operating a vehicle on all roads throughout the US, but, for legal or business reasons, cannot operate the vehicle across the borders in Canada or Mexico can still be considered Level 5, even if geo-fenced to operate only within the U.S.”.

It's not a remote human safety driver, it's more like a remote human safety coach.

The difference is giving high level directions vs directly driving the car. They don't remotely drive the car because that would obviously be super dangerous w/r/t connection stability/latency.

In case anyone else was wondering what AGI means, its Artificial General Intelligence. [1]

1: https://en.wikipedia.org/wiki/Artificial_general_intelligenc...

Or, on seeing an anomalous <children's toy> enter the road you can probably guess that a child might follow shortly after.

I'm not suggesting that the problem cannot be solved without AGI, but you can see why some people might think that though, right?

My personal feeling is that we shouldn't be setting the bar at making a car that can handle any situation anywhere way better than any human at any time, but that we should also try to make roads that are more suitable for self driving vehicles. I'd rather we move to driving agents that don't get bored, frustrated, or angry.

And your example isn't moving goalposts, its just another legitimate example of a situation thats gotta get figured out. If you think that things like understanding that some kid learning to skateboard nearby could fall a surprisingly far distance and thus you should exercise caution, or being aware of factors that imply fast biking children (say, an adult and a child implies the potential for another fast moving child on the same trajectory), that this sort of situational and contextual awareness is critical for proper driving.. then yeah, that would be a reasonable sounding argument to support "I think self driving cars will require some level of progress in AGI".

That's all I'm long-windedly getting at.

It would be great to see the dawn of AGI, but I don't think it will ever happen with classical computation. GPT-3 spits out nonsense with the input of the largest and easiest to parse portion of reality, and I have not seen any ML approach replicate the abilities of something as simple as bacteria. ML requires constant validation from human operators, so the same is going to hold true for ML powered vehicle navigation.

Cell data connections aren't reliable enough, and having the car emergency stop (and potentially get rear-ended) when it loses signal wouldn't be acceptable.

I think this statement is off the mark. Comparing to a human is hard. Not many accidents happen because people are bad at driving. Driving is honestly pretty easy. They happen because people are distracted, tired, drunk, or perhaps just an asshole driving recklessly for thrills or for speed.

A self driving car might be a lot "worse" than the average human driver but could still be a huge improvement in terms of expected safety record for driving overall.

They don't need to be better than humans, they just need to be not shit 100% of the time unlike humans.

Seems like Google and a few other tech companies could easily bootstrap a small city by planting some offices somewhere

The only thing it tells me is that regulations are more lax in Arizona than California.

I agree today's AI tech is a long way off from completely supplanting a human driver. I'm surprised the average consumer I talk to about this seems to think we're on the cusp.

But as vehicles with neural nets become more prevalent I expect we'll see the problem morph as it gets tackled from other angles as well. e.g. Self-driving corridors with road infrastructure aimed to improve AI safety (whether that be additional technology, modified marking standards, etc).

Once upon a time street signs with speed limits, curve warnings, and such didn't exist. After faster cars supplanted horse-drawn carriages, highways became a thing. Eventually when the only reason humans drive is for recreation (e.g. off-roading) the problem from the car's perspective will look somewhat different than it did during the transition.

I think its in part regulatory relating to paid rides in autonomous vehicles in CA, which is why Cruise is dodging it with passenger-carrying but unpaid rides that are fully driverless. I can't find a good summary of the rules, but I infer from the coverage I've seen that the threshold for having no safety driver when offering paid passenger rides is different from that without paid passengers.

I think you can only determine that if you know how many times the human attendant takes over.

Just having a human behind the wheel doesn't tell you much, I don't see how to get full self-driving without an intermediate step of human supervision.

Oh, but of course! I'm still surprised by people who think otherwise. The amount of corner cases that you have to solve while driving is pretty much limitless, you cannot "train" something on that, heck, some humans are not even fit for that. We are *FAR* from truly autonomous vehicles.

My guess is that they know how many miles they have to drive in order to reach that number, and it's a whole lot. Statistics and math stuff but you can probably pin it down to the month or quarter based on trends. Either that, or it's about driving every road in the city with all sorts of weather / traffic / pedestrian conditions until there's no issue. This isn't generalized AI driving (L5) but it's a much more logical approach to getting autonomous driving coverage where it's the most valuable.

My guess is that each city will involve a safety driver rollout until they have enough data to know the incident rate is zero. There might be a lot of variance between cities - maps data, weather conditions, customs, etc. Then remove the safety drivers.

I'm sure they also are experimenting with disaster/safety protocols while they do the roll out.

My prediction is that waymo will be a mainstream option within the next 5 years.

But people have hard time accepting with the notion that unmanned vehicle may be part responsible for that 1%.

can only please some of the people some of the time.

All traffic signs and signals need to be machine readable at a distance. That is, a traffic light might beam out "I am light 'SFTL783', will be green for 8 more seconds". The location and other data for SFTL783 is in a preloaded database. Same for speed limits and other signage.

An updated 3d map of all roads would also help a lot. As would car-to-car communication systems.

Once it has proven itself for a bit of time and they know how to set up their geo-fences and which streets to avoid, they can probably get rid of the person. That would be evident by that person never actually doing anything long before that.

The remote safety monitoring is a smart feature but it's not going to help much for the type of accidents people worry about most where something unexpected happens very rapidly. AIs are actually really good with dealing with those situations. Arguably better than humans for whom this is probably a leading cause of traffic fatalities. The way Waymo operates without hands on the wheel (i.e. level 3 & 4), basically means they have safety nailed already in those kind of situations.

There's no way a remote person would be quick enough to intervene. That person is there for other reasons. It's complex traffic situations that cause AIs to get stuck occasionally that require human intervention. Usually this is less of a safety concern and more of an annoyance.

The key metric of interventions per hundreds of thousands of miles of distance traveled is the key metric that both Tesla, Waymo, and other companies use for this. Both companies boast some pretty impressive statistics for that. Though of course it's hard to confirm those independently.

It's interesting the way Tesla and Waymo approach this problem in different ways. Waymo goes for level 4 but only in areas they've thoroughly vetted. It's taken them many years to start moving to new areas other than relatively safe and easy Phoenix. Tesla on the other hand offers their AI features just about anywhere they can but positions it as level 2 that is really aspiring to be a level 4 but with a requirement for hands on the wheel just in case (which means it's level 2). Level 2 is a legal tool to dodge a lot of legislation and bureaucracy. It basically means that if anything goes wrong, the driver is at fault. It will stay in place a long time for that reason. Liability, not safety, is the key concern here.

Arguably, Teslas would probably do pretty well in the areas that Waymo has vetted as well. But they are not in the ride sharing market and need to sell cars world wide and not just in specific geofences in Phoenix and San Francisco. But I wouldn't be surprised to see Tesla offer a similar ride sharing service in similarly geo-fenced areas at some point to get to level 4 & 5. I suspect that race is a lot closer than some people seem to think.

Their technology seems an order of magnitude above the rest, although factoring in its cost could make it look worse. (Can it turn a profit, when including the development costs?)

Tesla FSD is at best SAE level 3 (can need human fallback), and at worst level 2 (needs constant human monitoring).

[0]: https://blog.waymo.com/2020/10/revealing-our-approach-to-saf...

I'm curious what all the folks who claimed this was decades away are thinking?

What is the "this" you're talking about here? I don't think there were many skeptics out there saying that it'll be decades before a taxi service will be able to launch a small, manned trial in a single municipality. Given that it's manned it still isn't quite at Level 4.

The cynicism (and I'd consider myself a mild cynic I guess) was and is around the notion that the vast majority of us would be sat in self-driving cars by now. That was always wildly optimistic. We're making progress and that's great! But there were a lot of breathless predictions years back that have not come to fruition.

Each of these is different but neither's existence by itself proves unmanned in challenging locations is right around the corner. It's kind of an exercise in Baysian statistic, how much more likely this makes one think unmanned taxis in serious location is depends on what one thinks the initial probability is.

The vast majority of the contiguous United States is decidedly not "on the table" as of now.

"The contiguous United States" is still a decade (or more) away. It may literally never happen.

So the vast majority of the US has either significantly worse weather than SF, or a significantly harder ODD than SF?

Could you elaborate on that?

Regarding the weather, lots of people have mentioned snow, but much of the Midwest, Northeast, and especially the South can have sudden torrential rain. I haven't researched it but I would guess that a Florida rainstorm would be hard for both radar and visual guidance. I predict their next city will be Orlando, in partnership with Disney. Then somewhere like Boston (harder - older street pattern and more snow) or Philadelphia (easier). Each of those 3 would "unlock" new territory they can cover. I predict that NYC will be one of the last areas "unlocked".

I didn't make any statements about SF traffic or weather.

The fact that the +1 is a city as complex as SF is a really good sign as well. If they can handle SF, they can handle Charlotte, Atlanta, LA, and many other major metros just as "easily".

I'm a lot more optimistic for their rollout now.

SF does not. Also it looks like it almost never rains in Chandler.

we won't have all-weather reliable level 4/5 until an onboard computer has the equivalent power of a human brain

and he made similar claims each year after that.

You can thank him for the hype and cynicism

There's no commercial value at L4 whatsoever if you still pay a safety driver. It's usually even more expensive per hour than a normal human driver and the cars don't work under adverse weather conditions.

There are two markets:

- Level 4 driverless taxis (~$500B market)

- Level 5 driverless cars for every human being (~$10T+ market)

So, most of the commercial value for Uber, Lyft, etc. is from L4, but L5 is where the most overall economic value is.

But L4 has to be achieved in all locations, otherwise that TAM is highly variable.

Waymo has been at it for a bit over 12 years now. We'll have to see if Waymo's progress ramps up after San Francisco, but to me this is still looking like an 80/20 problem, with the added twist that each new region they expand to will have it's own unique 20% to learn that adds another 80% to the schedule. But that just validates Waymo's approach even more in my eyes.

It's basically level 5 or bust.

You have multiple groups of drivers including those who appreciate self-driving assistance on long boring highway stretches and may not even drive in cities that much and those who never want to get behind a wheel or perhaps ever own a car.

You could force the car to avoid certain intersections with lots of close calls, or give the car advance info about unexpected objects detected in the road by cars up ahead, etc. which would make the task a little easier.

Apart from regulations, level 4 doesn't require a remote operator monitoring at all times. It just requires them to be available when the car recognizes it can't handle the situation. Depending on how frequently interventions are needed, that could mean a remote operator for every car, or one for every 10 thousand.

The important part about Level 4 is that the car recognizes its own limits and doesn't act outside of them. How practical those limits are is completely undefined by SAE. Level 4 is a continuum whose limit is level 5. We will definitely be seeing practical commercial value as we approach level 5 but haven't yet gotten there.

[0] edit: that's incorrect, L4 and L5 are primarily defined by differences in geography range.

However, in terms of technology, level 5 is a big step up.

Consider for instance that a level 4 system can assume that the ground is purely flat (an assumption that Tesla’s FSD vector space makes), while a level 5 system needs 3D information to navigate some vertically-diverse terrain.

Which is exactly GC's point – they don't want a car that can drive itself in Phoenix and SF and that's it. Most people either want a car that can make their life easier for a huge portion of driving but still require them to be in the driver's seat (e.g. Tesla's Autopilot) or a car that can actually drive itself basically anywhere (e.g. Tesla's FSD). Waymo's Level 4 just isn't that appealing for a personal vehicle.

This is why the SAE level system is poorly thought out and not particularly useful for anything except arguing over minutiae.

It isn't clear how Tesla goes from "FSD" to "full automation". They are working on the "draw the rest of the owl" step.

Meanwhile, for Waymo to jump from level 4 to level 5 requires handling edge cases which it is actually not clear can be handled by the system Waymo has built. In huge swathes of the world, weather conditions are not conducive a system that relies heavily on LIDAR / roads change too much to be effectively mapped like they've done in very stable metropolitan areas like Phoenix and SF / etc.

Said another way: our existing roads are built for entirely visual agents (humans). Getting a system that heavily relies on non-visual sensing (e.g. LIDAR/Radar) to work well on a specific subset of those roads is clearly doable, but that doesn't mean it can be generalized to our entire automobile infrastructure.

As in, "This system is known to degrade in ways that require immediate human manual intervention under risk of serious injury or death. Do not operate it without constant driver supervision and preparation for takeover."

Tangentially, I've also noticed that Waymo has picked up pace ever since the recent leadership changes. They are publishing more blog posts, offering more insights into their tech and generally seem to have increased their PR game. I wonder if that was a mandate from Alphabet leadership to show some urgency.

The real challenges when navigating city streets are the human ones – delivery vehicles blocking lanes, municipal worker fixing a manhole with a single cone to redirect traffic, pedestrians/bicyclists appearing out of nowhere, no one following traffic signs. This is the kind of stuff that tests "intelligence".

This remark makes me wonder if you've ever lived in an area that actually experiences winter.

Around here, dead of winter, there are no lines visible on the streets. Heck, after a good snow storm the lanes are basically a function of group consensus.

(noting that Waymo requires full detail maps to be able to drive an area, including all signage).

First, it's optimistic to assume that data is accurate and up-to-date.

Second, I can't emphasize my "group consensus" point enough.

Anyone who's driven in real world winter conditions has seen a day where three lane roads turn into two. Or lanes form in the shoulder. It'd be actively dangerous to insist on driving according to the underlying lane markings during those types of road conditions.

Maybe in a world where all cars are autonomous and using map data that could work. In reality it really doesn't.

Huh, what? Human drivers can already take advantage of all of those, and they still find snowstorms and torrential rain challenging.

The challenge is understanding what you see (and hear), and dealing with very noisy and limited--sometimes actively misleading--inputs.

The harder part is driving like a human and detecting that a path has been made in the middle of two lanes in heavy snow and not obeying the lines at all.

It's not a matter of "finding" lane markers. There are no lane markers visible after it snows.

We ourselves identify and confirm other urban waymarks via captcha which feeds the nav data -- bridges, signs, hills, hydrants, chimneys, lights. There is mass live verification from android auto in vehicles. There are many yearly layers of street view images and scans.

I don't think we'll see a system that can handle that in my lifetime.

The human driver detects the pedestrian, laughs at the fall, then get worried and wait for them to get up, because a human knows someone fell and didn’t magically disappear.

Source: grew up watching subarus do 360s on the freeway.

I find SF much more challenging to drive in, at least wrt the other drivers. In Boston, drivers are aggressive and take calculated, dangerous risks to meet their goal.

In SF, a lot of people on the road act like it's their first time driving in like 5 years and they're still figuring it out. There's no rational risk-taking towards a goal, but more people bumbling around unpredictably while unsure of what their goal even is.

https://www.engadget.com/2018-05-08-waymo-snow-navigation.ht...

https://www.azcentral.com/story/money/business/tech/2019/08/...

Yes and no. The weather conditions are probably close to perfect for a project like this. A city that spends months with icy roads... _that_ would be the real nightmare.

Surely Alphabet has noticed that their competitors are nipping at Waymo's heels. If they don't pick up the pace, all sorts of business books will be written about how Waymo squandered a decade-long lead in the industry.

What I wouldn't be as comfortable with is the "random sheet of ice" or "oh look, rocks" or "suddenly washboard dirt road".

(Had, as in they changed the law few years back. Not sure how it's now)

Robotaxi(s) could be quite good solution to the problem - the drivers were often pissed if you called them for a single ride when they were home or far away.

Most products, including this one, don't need to do everything to be both useful and profitable.

Not that you are wrong, just that you should be wrong because if cities actually had useful transit rural areas would be a much larger share of demand despite not having many people.

For example, I am way more likely to take Cal Train into SF if I can use a point-to-point service like Uber/Lyft/Waymo to get me the rest of the way there. Without that missing link, I'm much more likely to just give up and drive instead.

The only problem is that they are painfully slow. If muni had more dedicated bus lanes (like, a lot more) it might very well be the best bus network in the world.

Oh it will. Animals have evolved amazing camouflage. Computer Vision will easily miss a deer hidden in a dark treeline. And radar/lidar even more so because the forest is going to have a pretty irregular geometry.

Even identifying a bicycle in a regular city street is something we have not convincingly solved yet. Animals on the side of a forest road is pretty far away.

Even fairly simple autonomous tech will have better peripheral vision at near-to-mid range than one human can manage, so for all those bikers, crazy walkers, and chaotic 15mph cars you shouldn't hit, it stands a pretty good chance of being better. And when it's not, come to a stop and you're fine (barring some honks) - few are moving fast enough to hit you dangerously hard in those human-complex areas, and you don't need to stop instantly, just fast enough.

---

Honestly, I'd put SF at dramatically easier than either residential or highway roads. Residential (and adjacent) has fast-moving cars ignoring signs with obstructed vision, and inattentive humans at relatively high speed (bikers used to low traffic and swerving, kids and animals running literally through bushes adjacent to roads, general lack of care around vision-blockers like fences due to perceived low risk, etc). Who's-at-fault doesn't matter - in car vs human, humans lose, and people rightfully get upset.

Highways also seem harder, if highly specialized: accurate decisions 100+ feet in advance are absolutely critical due to the speeds involved, computer vision at that range has fairly low detail compared to humans, and lidar is practically braille for "car". Radar has trouble distinguishing stopped cars from the road because neither are moving. Ultrasonics as an ultimate backup really only work up to around 10m (and that's about the distance to stop a car at 25MPH, which you'll regularly encounter in dense city traffic).

I'll also point out that more people have died due to Tesla's autopilot on highways than Waymo, Uber, Cruise, heck all self-driving companies I'm aware of at any size combined. They're riding all the terminology lines they can to get away with it, and they may very well have an order of magnitude or two more miles, but I believe the point still stands - highways are hard.

Both are low-hanging fruit. When it can navigate a snow-covered road at night while it's still snowing, get back to me.

I basically can't do this. If it was absolutely necessary I would go out there and drive like 5 mph and be terrified the whole time, but otherwise I would just treat whatever I needed to drive to in a snowstorm at night as temporarily inaccessible. I have lived in places where it snowed in winter before.

'cause they decided not to build most of the roads: https://www.cahighways.org/maps/1955trafficways.jpg

What makes SF difficult to drive in (from my perspective of only ever being a pedestrian there) is a) extremely hilly terrain, b) the general difficulty of a dense urban environment anywhere, and only a distant third is c) traffic, which is merely an added stressor to the complex choreography that is an urban street.

How about with the freight railroad it used to have for 75 years before the state donated the ring of land to the city and paid to build the highway?

https://en.wikipedia.org/wiki/San_Francisco_Belt_Railroad

http://sanfranciscotrains.org/sbrr_history.html

Yes, I agree, but they decided it was better to go over every hill instead of through them: https://www.flickr.com/photos/walkingsf/4182283392/

I’m not sure how highways going through SF would make it easier to drive in SF (outside of the highways): wouldn’t that generally increase traffic and conflicts?

When coupled to our additional refusal to build housing, sadly, yeah. What two things do people usually commute between?

See also: Boston's Big Dig.

It would have probably been the continuation of I-980 had that bridge been built: https://en.wikipedia.org/wiki/Interstate_980#History

At one point it was also planned to be just north of SFO. Have you ever taken I-380 east instead of one of the exits to 101N/S? There's a huge multi-lane road that dwindles to basically an airport access road exit.

https://www.flickr.com/photos/walkingsf/4047626058/

https://www.flickr.com/photos/walkingsf/4047626054/

https://www.cahighways.org/ROUTE131.html

Here you can see an idea of doubling-up the Bay Bridge, plus a view of the Southern Crossing / I-980 alignment: https://www.flickr.com/photos/walkingsf/4247129432/

It is certainly true that the taste for elevated highways through cities has waned given the pollution and dust and general unsightliness that it produces. In the 1950s, when cars were all the rage, people were very excited by these things.

Personally, as an SF resident I would much prefer all the cars to be tunneled or elevated instead of idling in front of my house or blowing loudly through my block. It's a safety issue.

When there's only so much surface area where else are people supposed to build except up and down? That's why we have skyscrapers, and those don't seem to provoke the same vitriol as the roads.

Even the famously-hated Embarcadero Fwy wouldn't have been visible if the plans for the World Trade Center (lol) at Market/Embarcadero hadn't also been canceled:

https://archive.org/details/ferrybuildingcom2919sanf

https://www.sfgate.com/opinion/article/Ferry-Building-what-m...

This is a false choice though. It would be better to design the city in such a way that we don't need personal automobiles for most trips. Building high-speed roads (elevated or not) through a city tends to have the opposite effect. If you live in SF, just think of the parts of the city that do have high speed roads. Is it pleasant to walk along division street? People choose their mode of transit based on what feels safe and convenient to them.

In the space and money taken up by an elevated highway, we could have low-speed mixed-use streets and an entire separate highway for bikes, and it would be safer and quieter.

Source (great read if your interested in the subject): https://www.amazon.com/Traffic-Drive-What-Says-About/dp/0307...

From what I've heard repeatedly, SF sounds much more irregular and messy than Phoenix, so it should be something of a stepping stone to making them usable more widely if they can crack it.

I've been expecting it to happen for literally decades, but always been disappointed.

If we want to really test self-driving cars, introduce them to Milton Keynes 'Magic Roundabout'

https://www.google.com/search?q=milton+keynes+magic+roundabo...

I used to live round the corner from the magic roundabout. It's actually fine. The best analogy I have for it is juggling, if I concentrate too much and overthink it I drop the balls. If you over-think the magic roundabout it can seem intimidating, but when you're actually there it makes much more sense and you just go with the flow. You're not dealing with the whole system in one go, most people just take it one roundabout at a time.

One of the reasons it seems to work is that people take it easy, everyone is paying attention to what they are doing, and most people take it at a sensible speed.

Needless complexity is all I see in it.

If you can navigate a roundabout, you can navigate the magic roundabout - you just apply the same rules.

If anything, this is the sort of thing an automated system would excel over humans at - where the automation won't get confused by an uncommon application of a familiar ruleset.

A couple of years ago I took a wrong turn and got lost in the Long Beach harbor - and had neither GPS nor phone. Holy Cow man.... holy cow... (admittedly, low quality comment)

Not even close - I've lived in the DC metro, Boston, and LA and all three are certainly worse than SF, and I think that is also backed up by evidence.

My scepticism around SD cars in places like that and Bath is not even the roads themselves, but the traffic.

The are quite a lot of roads around me that are quite narrow and frequently drop to single lanes. I feel like any self-drive system will need to develop quite an advanced theory of mind to perform the weird silent negotiation that happens when I'm driving a long, with three cars behind me, and come face-to-face with a bus coming the other way, and somehow we have to either backup en-masse, or perform the squeeze dance as we edge past each other with inches to spair.

- Self driving is not possible

- Self driving is not possible anytime soon

- Self driving is possible, but requires LIDAR

- Self driving is possible, and can be done with normal cameras

I've engaged with a lot of people who presume self-driving reduces onto AGI, therefore it will not be achieved anytime soon if ever. I wonder which of their assumptions is wrong, if this ends up being successful.

Many people assume that a self driving car must be 100% successful. To me it is sufficient to beat the top 20%/bottom 80% of drivers.

So the car will crash if a seagull flies in front of the camera and there is a deer just beyond it. The human would probably crash too.

Computerized safety systems can, should, and are raising the bar, in terms of safety for the top 20% of drivers. "Self-Driving Cars" aren't competing in a static environment, similar tools are raising the safety bar that they have to compete against. This means it will get increasingly difficult to beat non-self-driving vehicles.

And what is the story on driving becoming safer?

It's interesting that you set the threshold there (I presume you meant 'beat the bottom 80% of drivers'). Rationally, we should be happy if their driving performance is above average. I fear that the public (and the courts and the insurers, etc) will require airline levels of per-mile safety and still be wary.

If a human driver kills someone / injures someone / damages property, people get satisfaction or resolution when the human is punished - insurance increases, license points, jail, etc.

But when a self-driving car - even if it's better-than-average - does something wrong, there's nobody to punish and no retribution to exact, so people will be left feeling unsatisfied. That's why the bar is going to need to be so much higher.

Pick something you pay extra attention to. Now imagine being informed you can't do that any more, you have to take what [this robot] does for you. Also remember all the terrible UI changes in software that have been done for a majority instead of your use-case.

Reluctance until it's substantially better seems reasonable to me. Somewhat unfortunate on a species level, but not for many individuals.

That is still a pretty high ask for current software.

Those two obstacles won't be overcome until AGI. At best their frequency will be brought down a bit but not by nearly enough orders of magnitude.

Also even Tesla handles the obstacle in the road situation well. Their FSD beta will happily enter the left lane to go around obstacles.

If an AI system's default response is "come to a stop safely" then it's going to be way ahead of a lot of human "unexpected situation" handling in cars.

Yes, people are bad at driving, because they don't pay attention, panic, make mistakes etc. But ML models tend to freak out at slight variations on mundane circumstances; a cyclist crossing the road at just the right angle and the wrong colour of bike, that sort of thing. The thing self driving cars need to avoid is killing people in broad daylight for no discernable reason, and that seems like the kind of thing that you'd need a mind for. It's the same issue as with adversarial image manipulation to fool image recognition; if changing 3 pixels can turn a frog into a toaster, you aren't really "seeing" the frog at all in a symbolic way, and not seeing a road symbolically seems like a recipe for disaster.

This, I think, is the thing that people miss when they say "self-driving cars don't need to be perfect, they just need to be better than human-drivers, who aren't actually all that great".

From a public confidence perspective, it doesn't matter if a self-driving car crashes one tenth, one one-hundredth as often as human drivers; as soon as you see a self-driving car kill someone in a situation that a human driver obviously would have avoided (like in the adversarial image kind of scenario), you've totally destroyed any and all confidence in this car's driving ability, because "I would never, ever have crashed there."

Sure it's probably feasible for cars to handle hand signals in the happy path, but anything outside of that will be disastrous. How will the car understand and communicate with a person who doesn't use the standard signals, aside from having some level of intelligence?

Sometimes. Other times they confusingly gesticulate or just shout out things, or even give conflicting signals. Humans can interpret these without much effort but it's a hard AI problem.

I don't expect that we will ever see the day where everyone is OK with self-driving cars on the roads regardless of the safety statistics, because there will always be edge cases of crashes and personal preferences around driving styles.

We have the advantage of retrospect when looking at these claims which now may seem more possible then before due to our better understanding of the difficulties and technologies to address them.

I think many people confuse their excitement for the promise of having self driving cars and the actual technical and political barriers that still need to be addressed to bring this to reality (ranging from robustness, infinite number of edge cases, perception, insurance, or public policies).

I think most of the confusion is due to deceptive marketing. Sales people like Elon Musk have been saying that the big dream of true self driving is just around the corner for years now, even though they were nowhere close.

Define “soon”. For me it means 50 years. That’s an infinitesimally small amount of time. If we can have level 5 autonomy in 50 years I will say that progress was rapid and we really excelled.

> The first working device to be built was a point-contact transistor invented in 1947 by American physicists John Bardeen and Walter Brattain while working under William Shockley at Bell Labs. The three shared the 1956 Nobel Prize in Physics for their achievement.

https://en.wikipedia.org/wiki/Transistor

Source: https://www.sjsu.edu/faculty/watkins/transist.htm

Given sufficient qualitative and quantitative investment self driving may be technologically feasible at some point in the future but not price competitive with human driving.

Self driving proponents seem to assume the tech is free. Just keeping sub meter scale 3d street maps up to date has a massive cost.

> Just keeping sub meter scale 3d street maps up to date has a massive cost.

True, but probably not actually necessary.

Well, for one thing, I don't want anyone in populated parts of the US to live on $10 an hour, not even a cab driver. What happens when we double that? Cost of living goes up over time, but cost of technology goes down. When do we reach the tipping point? Have we already?

If wages go up to $20 an hour, then maybe there is general inflation and robotics goes up to $200 an hour. In the world today we have cheap labor and expensive energy. That is why automation is not replacing humans. Look into the history of the British industrial revolution.

Please describe the last factory you were in.

Here is the reality that self driving will confront:

cost effective, reliable, competitive automation is hard, even for seemingly closed mundane taskes like palatising products.

Yes, it's expensive, however, as we refine algorithms, and the processing systems, it gets cheaper, especially amortized over more cars, where the per-car cost becomes affordable.

ML doesn't work well enough for offline maps generation either, and all the high quality maps require human editors for final touch.

All this work is currently done because realtime perception doesn't work well enough, and you can have a much more reliable system with the aid of the maps. Having a 3D base map of the world makes the realtime perception problem far simpler, and it makes fancy sensors less critical.

In the US, where the cost of labor is very expensive, self driving will make sense, even if it's expensive, but someplace like China or India, where a middle class person can afford a driver, it probably makes less sense, though the push for it in China is probably the strongest that I've seen anywhere in the world.

Self driving does not scale at all now. Because as you said these special maps need to a lot of human labor to make and the cars need a lot of sensors on top of the auto patform. Labor is not even the main cost driver in person transportation.

Why are efforts not focused primarily on interstate/highway travel, specifically, collaboration with DoT for mesh/distributed/coordinated long-term travel?

I don't need a self-driving taxi. I would pay 20K for a car which participated in a federally-regulated framework which let me let go of the wheel when I get on a highway and let teh emergent cloud determine how best to move my car and all the others on dedicated/reserved lanes in coordinated "trains" for hours.

The wins here seem like no-brainers. Sidestep all jurisdictional nonsense; optimize commerce and personal travel; automatically handle emergency vehicles and other unusual conditions; etc ad infinitum.

All my car needs to be able to do other than existing lower-tier self-driving/driver assist basics, is join the borg.

Coordination for traffic flow management seems like an unbelievable win.

But no, all we seem to be getting is cyclist-terminating taxis which cost $250K each and are, IMO, doomed in target-rich environments like SF to not forseeably adequately the last 8% of anomalous novel cases.

Just don't get it.

Other car companies are a few years behind, but even something as simple as adaptive cruise control + lane centering is a huge help on freeways.

Nobody is saying that self-driving cars are perfectly safe. Considering how many Teslas are on the road and how many miles are driven on autopilot, it would be surprising if there weren't any deaths. As long as it's safer than unaided human drivers (which it is), it's a net win.

For 95% of people I think the value is in letting them do other stuff while driving long distances/times instead of stuck behind the wheel. This seems many orders of magnitude easier than trying to tackle inner city driving.

https://www.motortrend.com/reviews/cadillac-super-cruise-is-...

This is big news, and I encourage you to apply if you're in San Francisco. For the HN crowd, I'd also recommend last week's blog post [1] which includes some more technical material on "how" we're driving.

[1] https://blog.waymo.com/2021/08/MostExperiencedUrbanDriver.ht...

I had to google "Cruise". To save others a moment or two, Cruise is a self-driving startup unrelated to Alphabet. In the quoted sentence, "they" != "Waymo" (I wasn't sure).

HN is getting to reddit's status where people make comments and everyone just upvotes because it sounds good to them.

https://news.ycombinator.com/newsguidelines.html

I became a HN regular because I saw some insightful comments backed by quality moderation. I was escaping the dumpster fire of reddit so it seemed like paradise by comparison. Over time, my usage of HN grew, and I became more exposed to the reddit-like mentality that pops up on certain topics or at certain times(i.e. weekends).

I think the quality of HN varies depending on where the more serious users are in their development schedules. It would be an interesting thing to analyze I suppose.

https://twitter.com/MumbaiPolice/status/1223090017397960705

I'm hoping I just missed that line. But I signed up anyway, so I guess that says more.

I was a self driving skeptic 2-3 years back. Now given the advances in both hardware and NNs, I do see possible solutions in the next decade.

Anyone else got this to work? I only see a message that says "It looks like you're not in our service area..."

Safety is obviously important, but we already accept a decrease in safety over money and convenience (otherwise, we would have a 30kph speed limit everywhere).

But most people use their car for commuting, groceries, dropping the kids to sports, etc. They may make one or two road trips a year for vacations or business.

To facilitate this, a middle class family will often own two vehicles, pay tax, insurance and maintenance on said vehicles, pay for parking, etc.

If that suddenly wasn't necessary, that would be a massive saving for most people. You no longer need the capital investment in something that loses value.

Even better, you don't need to bring little Suzy to football, you can put her in a waymo instead (once she's old enough obviously).

And when you want to go drive to see grandma, you rent a car for the longer trip.

All of that depends on affordable self driving transport of course.

Financial Times (Alphaville Blog): The questionable economics of autonomous taxi fleets: https://www.ft.com/content/aa05823d-e58f-33ee-a99c-d50f2fd0c...

... quotes this Harvard/MIT study: Autonomous Taxis and Public Health: High Cost or High Opportunity Cost?: https://psyarxiv.com/6e94h

<<<Drawing on a wealth of publicly available data, Ashley Nunes and his colleague Kristen Hernandez suggest that the price for taking an autonomous taxi will be between $1.58 to $6.01 on a per-mile basis, versus the $0.72 cost of owning a car. Using San Francisco’s taxi market as its test area, the academics examined a vast array of costs such as licensing, maintenance, fuel and insurance for their calculations.>>>

I can imagine that as the world grows more urbanised, and wealth concentrates in urban areas, that people will want to pay for the convenience and safety of a robot taxi.

For the autonomous taxi calculation they also include a 250k medallion plus interest paid off over 5 years, but don't include this for regular taxis. They also include a 200K/year to supervise the auto taxi for the high end ($6/mile).

With all that said, Auto taxis are unlikely to be a cost effective replacement to owning a car for distance commuters. Even with lower costs than traditional taxis, taxis are horribly cost ineffective way to get to work and back for any significant distances.

I don't doubt there will be a shift in the overall economic picture, but it may not be as large for the car _user_ as we would hope. It will need to be better/cheaper in some dimension for it to get adopted, but the rent-seeking vultures are always circling, spoiling the utopia :-(

From a public confidence perspective, it doesn't matter if a self-driving car crashes one tenth, even one one-hundredth as often as human drivers.

If you see a self-driving car cause an accident, particularly a lethal one, in a situation that almost any human driver would have avoided, you've totally destroyed any and all confidence in this car's driving ability, because "I would never, ever have crashed there."

As we've seen, there are lots of scenarios like this. The Tesla crash from last year, where the car simply didn't see a white truck against a light background. Or imagine an adversarial image attack, where some tiny insignificant detail is placed onto a stop sign or a "do-not-enter" that turns it into nothing from the perspective of the AI driver.

These kinds of scenarios obliterate public confidence in self-driving cars, because intuitively, you immediately realize that you're "a much better driver" than this car! Even if that's untrue 99/100 times, it only takes one visceral example to drive this kind of wedge.

Self-driving cars don't just have to be better than human drivers. They have to be as close to perfect as is possible, because that's what people will expect.

A large fraction of human drivers are actually all that great. The majority of accidents/deaths are caused by a minority of terrible drivers, or good drivers who found themselves in terrible but rare circumstances. The majority of drivers drive hundreds of thousands of miles without any accidents that were their fault, or even any accidents at all.

In other words, it's probably easy to beat the mean human driver, which is greatly dragged down by a minority of terrible drivers. It's probably very difficult to beat the median human driver, and near impossible to beat the top 20% of human drivers.

In 2019 in California, there were 1.06 deaths per 100 million vehicle miles traveled. Any self-driving automobile technology that doesn't have at least 1 billion vehicle miles of data is in no position to claim that it is safer than human drivers and less likely to kill people.

Self driving cars don't make the same kinds of mistakes as human drivers do, but they make different kinds of mistakes. Some of these can be fatal.

Agreed. I should have written "relatively easy."

> Any self-driving automobile technology that doesn't have at least 1 billion vehicle miles of data is in no position to claim that it is safer than human drivers and less likely to kill people.

The circumstances under which those miles are driven (e.g. road type, location, weather, time of day, etc.) also have to be consistent with circumstances under which humans are driving. 10 billion autonomous vehicle miles driven only on highways in broad daylight is a worthless point of comparison, whereas 500 million miles driven across a variety of conditions representative of the full human driving population is worth a lot more.

This is key, there's expectation and some wiggle room that as a human driver, humans will fuck up predictably and experienced drivers know how to avoid getting into incidents when this happens (usually).

Self-driving cars are weird to drive around. They will absolutely stop in situations where no human would think to stop. I think about this as a motorcycle rider, what if I'm committed to cornering on a corner I can't see around and the software decides on a self-driving car that it should just stop in the middle of the road after the apex? A human driver could do this too but many will know that this is a dangerous place to stop and try to put the car on the shoulder or minimize the amount of time it's stuck there.

I don't know if this is something we need to tolerate a temporary increased incident rate on as people get used to them being on the road, or if we need to make the software drive more like humans (with the assumption that means potentially making the behavior act sloppier than it can handle so that increased software reaction rate doesn't cause humans with slow reaction rate to slam into them)

It is. The mean is dragged down by alcoholics who drive drunk every single day. If you never drive drunk, that gives you a significant advantage.

The mean is given by the number I posted, about 1 death per 100 million miles traveled. That number includes drunk drivers, distracted drivers trying to text, everything.

https://www.ots.ca.gov/ots-and-traffic-safety/score-card/

California Numbers:

* 3,606 traffic fatalities in 2019.

* 1,066 Alcohol-impaired driving fatalities (fatalities in crashes involving a driver or motorcycle rider with a blood alcohol concentration, or BAC, of 0.08 or higher) in 2019.

* 620 Unrestrained passenger vehicle occupant fatalities in all seating positions in 2019.

* 164 Teen motor vehicle fatalities (age 16-19) in 2019.

* 972 Pedestrian fatalities in 2019.

* 133 Bicycle fatalities in 2019.

assuming the above (alcohol-impaired, unrestrained passenger, teens, pedestrian, bicycle) are all all poor-driver related that leaves 651 traffic fatalities.

I wonder what humans will actually do better than AI in 50 years. I have a personal theory but I'm a bit off topic here

As t approaches infinity, what are the chances that self-driving cars won't take over the world?

Less drunk, distracted, angry, and/or sleepy drivers on public roads, with overall significantly less crashes? Sign me up!

I'm really beginning to get sick of seeing people use the word "freedom" when they clearly mean "no personal responsibility while living in a society among other people".

It's borderline impossible that they will "legislate away people's freedom to drive" anytime soon.

Even if a startup appeared tomorrow that could unequivocally show that they have a perfect self-driving car that runs on fairy dust and cleans up cities as it goes, people would still demand their right to drive ICEs.

Of course the more it starts taking off, there will always be a vocal subset of the population that is strongly opposed to it, just like there are vocal anti-vaxxer groups and there were anti-seatbelt protests back in the 80's. But I can't imagine the naysayers having a very big impact on the progression of the technology, the upsides are just too enormous.

What technology is perfect? What code is perfect and doesn't have bugs in it? And yet we adopt automated systems anyway. Yes, sometimes it is painful and an entire airline grounds all planes for half a day... But that doesn't stop the unending march towards efficiency and technological progress.

Though I disagree because people expect a lot of things and alternative outcomes happen when there are incentives (Current cars, airplanes, and elevators come to mind). Waymo seems to be aware of this from this recent video. [1]

1. https://youtu.be/yjztvddhZmI

We have seen this before with autopilots in aircraft and ship, elevator operators and so on.

All it needs is time and adoption, not perfection, and as adoption increases, the roads get safer and safer, bringing us closer to the ideal, and as time increases, the closer we get to adoption.

Watching the Waymo video just changed my viewpoint.

They have a "Pull Over" lever in the back seat.

Would I trust a self-driving vechicle without Lidar--probally no.

Would I use use a self driving vechicle commuting, and around the city--yes. Two driving chores I hate. I hate them to the point that the philosophical argument over dying by my own hands, or a computer, is put in the trunk.

On a personal note, the thought of a computer driving me off a cliff while driving to Stinson beach is not something I would chance. Even if they are statistically better drivers than most humans.

I can't imagine tumbling down a cliff, and thinking if only I drove today.

I still foresee most trucking jobs, and most driving jobs, completely gone in a few years.

Yes--it's time for a Basic Income.

I feel like there is a lot of disconnect between the different media pieces and opinions I come across specifically when it comes to Waymo's self-driving - from "we're already there" to "it won't happen in the next 10 years".

They seem to have taken over a previously public parking lot at Pacific and Sansome in the city. There are often 20/30 of these cars parked up there.

I've never seen one without a driver, but occasionally when I look at the steering wheel it is turning independently and just being monitored by the driver.

I live on Dolores park. They drive by multiple times an hour

My car was broken into over the weekend and I couldn't help but think about whether a Waymo captured the vandalism... Not like it would help, but like all mass data collection campaigns, the potential implications are strange.

(Disclosure: I work for Google, speaking only for myself.)

Hotels have many people who work there and cleaners who enter the room daily (no Hendrix hotel a la Altered Carbon yet). Rental cars are secured with your credit card so there's a massive disincentive to trash the vehicle.

Perhaps pervasive recording of the vehicle interior would suffice.

Yeah, I was thinking along those lines. Trash the vehicle, get a cleaning fee charged to your card.

I imagine though that it'll just be something you can report when it comes along, then it gets sent off for cleaning, they pull up its recent rides, and send you a different car.

System locks all doors and calls police?

Also, I don't particularly want to ride in a car with a bunch of sensors and cameras monitoring my every move.

The current situation with taxis is that you enter before confirming your identity via payment, and the problem is put onto the taxi driver (an individual) when something happens.

This is distinctly not the self-drive corporation issue: you trash the car, the issue is forwarded to corporate debt recovery, who then work a 9 to 5 slowly pushing the issue though the relevant channels. Meanwhile, the car is returned to base, maintenance rips out the absorbent materials and power washes the interior.

Trashing it becomes a line-item cost to a very large organization, not a problem which "isn't worth it" for an individual operator.

Many (most?) taxis already have both dashcams and cabin cams. This has already been normalized.

At least the current fleets of AVs (Waymo, Cruise, etc.) are "obviously" potentially autonomous. I'm honestly more cautious now as a pedestrian when I see a Tesla coming after the FSD videos. I wish I could know "Is that Tesla owner using the FSD mode?"...

I wish I could know a lot more about the car and the driver in it too, such as if the car has pedestrian airbags[0] or if the driver is having a heated argument with their spouse. It's all risk management and ultimately the person behind the wheel is responsible for the exoskeleton on wheels they pilot, as the existence of cars at all is a net negative for pedestrian safety.

0: https://www-esv.nhtsa.dot.gov/Proceedings/23/files/23ESV-000...

I count three distinct hedges there :)

It's cool tho, congrats on the expansion!

Edit: can’t reply because “rate limiting”

That video is almost half a year old. Here’s a video from the same channel that’s 6 days old rather than 6 months old. Tesla pushes updates very frequently…

https://youtu.be/y85oGY02gtg

Here’s a video where the title says it all.

https://youtu.be/tdHlKhKKOgQ

I read one of the latest waymo blog posts and was surprised to see how directly you guys reference Tesla. It felt like the whole thing was written for people at tesla… You differ in choice of hardware but could you tell me more about how you differ on software? Is there any meaningful difference?

[1] https://m.youtube.com/watch?v=antLneVlxcs&feature=youtu.be

He speculates that something very different will happen. He thinks that self-driving cars that follow the laws and unerringly yield to pedestrians will transfer ownership of the streets from cars to pedestrians. Perhaps travel through cities in self-driving cars will be tedious and slow because pedestrians will jaywalk fearlessly and the cars will always yield. It's an outcome I never really considered before.

[1]: https://www.pushkin.fm/episode/i-love-you-waymo/

To "good morning dear, do you think you could help me load my bag in the trunk"

All the way up to completely degenerate scenarios like "my friend was drunk and tried to take over command of the car to see if it would work" or "I had to call customer support because when my ride arrived to pick me up, there was a homeless guy hogging the backseat and cursing at me".

Surely there's more to commercializing the tech than just not killing pedestrians. The absence of a driver should create some new weird dynamics.

IIRC Uber et al spend a considerable amount on customer support. I wonder if Waymo can break free from Google's bad reputation on that front.

1. How much should I tip?

2. The driver tried to rape me.

3. The driver seems drunk/high, and is driving horribly.

4. The driver didn't pick me up because, pick your reason: I'm black, had a bunch of kids, etc...

5. My friend was drunk and tried to pick a fight with the driver.

If there's no driver, then there's nobody to ask to help with bags.

For various degenerate situations I'm not sure being driverless really changes the situation all that much. A drunk passenger can grab the wheel of a manned vehicle as well. With a self-driving car you might even be able to avoid that situation entirely by preventing the passengers from using any controls without authorization. They can grab the physical steering wheel but can't turn it until they punch in an authorization code (which they don't have). Likewise you prevent unauthorized passengers from entering the car by keeping the doors locked until the authorized passenger gets there. And if someone does manage to sneak in you just shut down the vehicle and call the cops (just like if someone jumps in a cab without permission).

But humans are great at flummoxing countermeasures like that so it will be an interesting thing to watch for sure.

To add on, maybe even Waymo tells/asks you if you'd like to change pickup locations to save X minutes off your trip.

Feels somewhat in line with how Google Maps will ask you if you want to save X minutes by taking a new route that was previously not as good as when you started using directions.

I'm also guessing that they want an electric car. At the time they made a deal to purchase those cars (was quite a while ago) Jaguar i-Pace was one of few premium electric cars available (if you exclude Teslas, which obviously Waymo wouldn't want as it's a competitor).

At this stage cost is not a big problem.

- if that's the car their devs are testing on, it's going to be the fleet car, at least at first

- Alphabet has $135B cash on hand[0]

0: https://abc.xyz/investor/static/pdf/2021Q2_alphabet_earnings...

Most taxis in Denmark and Germany are Mercedes-Benz so it's not uncommon for the taxi industry to use luxury brands.

I'm sure I got all the science wrong but do we really want cars with souls? Do they become non-human humans? Do they get the right to vote? Are they considered slaves?

Even if its just horse-level smarts, that's still enough to develop consciousness, maybe. Let's just say sure. But then, what kinda relationship do you have with your car then?

Another thing is that, now that everyone's working from home, there's a lot less incentive to buy cars. Who are the going to sell these cars too? Maybe they'll all be self-driving taxis. And that's a whole can of proverbial worms on its own. Politically, self-driving taxis are dead ducks.

that would be more like staircase of enlightenment. or up the down escalator to enlightenment

If ML could predict the future, there would be some very rich stock brokers right now.

Autonomous driving efforts need a major pivot to start working with concepts, causality, and the integrated space time model of the world we humans use.

Dear god, no. I don't know what you think AV companies are doing past the perception layer, but they would be pivoting away from controls and planning -- you know, stuff that actually works outside of neurips/iclr/&c fantasy land.

The aspect I dislike is the move towards hiring cars on demand instead of owning them, though.

I always find it interesting that the further down the road of self-driving we go the more it seems like its a techy version of a known but morre low tech paradigm. This if taken without the self driving part just sounds like a bus lane with bus stops. In fact the driving of a car would be a less efficient mode of transportation considering that they usually only carry one person at a time.

I used to be very pro self-driving cars, but tthte moree I explored urbanism and transportation the more it became apparent that, self-driving is over optimization of an already bad paradigm that is car-centric development. You could cut out the middle man of needing a self driving car if you didn't live far away from the things you needed or if the public transit was efficient and reliable, that seems way easier than trying to figure out how to make cars drive themselves!

That would be a pretty exciting development even in, eg, European cities with decent public transit.

Cars aren't efficient. The zoning should be such that things are close by enough. Then with proper cycling infrastructure you can just go by bike. As a result you'll need way less road. As cycling infrastructure is cheap and road maintenance is costly, it'll reduce expenses for e.g. cities.

> European cities with decent public transit.

I'd rather have the mix of public transport, bicycles and roads. Combined with proper zoning. Loads of electric cars just increases demand for a highly inefficient transportation method.

https://www.autoblog.com/2021/08/22/waymo-is-99-of-the-way-t...

My guess is Chandler is purely a testing ground for them to learn to operate a robotaxi service. Things like remote assistance, customer support, emergency protocols, fleet maintenance etc.

0: https://i.redd.it/4rsg9pui55531.jpg

Do you have to have a Google account?

This is likely to be more open, but again, requiring NDA.

That makes me worried this was rushed

"Jaguar, for its part, gets a large chunk of guaranteed sales and a chance to look like it’s at the forefront of this emerging technology."

- Waymo Expands Its Robo-Fleet with Electric Jaguar SUVs, https://www.wired.com/story/waymo-buys-jaguar-suvs/

Also, companies pushing their own self-driving or ADAS solutions may be hesitant to partner with Waymo, excluding some of the large players.

Jaguar doesn't have a high-profile competing autonomy effort of their own, that would create a conflict of interest and/or invite spying.

The i-pace is electric, which makes integration easier and won't hurt for PR purposes.

And it's luxurious enough that no matter who you're showing the tech off to, it won't be a big step down from their personal car.

While the US has a lot of Urban Sprawl, dense Asian cities have the option to invest and reap the benefits of public transportation instead. Eg im incredibly excited about all the metro lines in India, HSR in China etc.

If you drive a Tesla w/ FSD, is it chic to refer to yourself as an Autonomous Specialist?

How much are Waymo vehicle operational costs vs rider revenue? How long do Waymo vehicles last?

Even with this announcement, we do not know if Waymo has a clear path to profitability.

It's time for Elon to pay the piper and refund all users who were hoodwinked into purchasing FSD. Especially people who's cars are nearing end of life.

> If Google can't do it with its massive resources

The companies behind the SLS had and have an enormous amount of resources. Yet, they have failed to even launch it.

Arianespace has a huge amount of resources and experience too. Yet, when asked, they said landing a rocket couldn’t be done. Then, back-pedalling, that it was pointless and would never make any economic sense.

Yet, SpaceX has developed three generations of rockets (Falcon 1, 9 and heavy), builds them, launches them, iterates on them, lands them, reuses them, and is now working on a fourth rocket (SuperHeavy).

They have also created 4 different rocket engines and are working on two others.

Honestly, where are Boeing and Arianespace?

Another example? Microsoft, their incredible wealth of resources and superior position falling flat on their face with smartphones.

Funnily enough, Google getting the greatest share of the global smartphone OS market, despite having had no real experience with operating systems, consumer hardware, or licensing operating systems before.

Kodak, which had some of the first patent on digital camera sensors, going pretty much extinct despite their massive resources and market position.

Xerox, failing to ever enter the personal computer market despite designing what everyone else would copy for the next 10 or 20 years.

Sure, the amounts of resources and experience seem to help, a lot, but it also certainly doesn’t seem to help reliably predict any outcome.

Because almost no company has much more serious experience with self-driving cars, let alone FSD.

Astra, maybe?

Difficult analogy because the barrier to entry's a lot lower for car mods vs orbital rocketry kits :)

What does SpaceX have to do with Tesla? They're owned by a billionaire with his attention split, but otherwise are structured differently, are in different industries, and most likely have different cultures.

The cult of worshipping a single person has got to stop.

> While I do get your point and have my own doubts […]

There are many issues with Tesla's autopilot that are completely unrelated to the amount of data they have, and they will not be fixed with more data, and having more data will not make it easier to fix it. At this point, I would argue that the discussion about who owns more millions of miles of data is completely irrelevant.

Tesla has built out amazing infrastructure to capture extensive amounts of "hard" examples from their fleet, turn them around into labeled data for training very efficiently and then utilizing simulations to further broaden the distribution of such quirky long-tail events in their training-set. In the absence of AGI, this is a very effective "brute-force" approach and they have a huge upper hand over every other player in this space.

I say all of this even though I am very skeptical that anyone will achieve L5 self-driving with where the state of things are today. But Karpathy and team are very pragmatic and making lots of good decisions coupled with excellent engineering and infrastructure development.

Good thread on ground Tesla would need to make up to catch AV leaders here. https://twitter.com/Christiano92/status/1428671634131628033

Simulation is important for sure, and Tesla even talked about that as much. But you can only simulate the scenario's that you've thought of, it can't replace real world data, which again Tesla has several orders of magnitude more of than Waymo.

As for the comments about Dojo I think they're unfounded. Tesla has already proven that they can create their own custom chip and have shipped it to hundreds of thousands of cars. I don't see any reason to think they can't get the system running in the next year or so. But even if they can't do that in time they still have NVIDIA to fall back on. Elon even said in the talk that maybe it might not work out, and if that's the case they can always buy a solution.

this raises the question of: what happens when those environments change?

It absolutely is implausible.

People who bought a 3-year or 5-year lease from 2016 with the $5k full-self driving package probably should feel pissed for wasting money.

Fools and their money. Like most every expensive car, Teslas are luxury vehicles and fashion statements. People buy them because they are cool. That product was delivered. As for future features, nobody should believe Telsa's advertising any more than we do Microsoft's advertising about how the next version of Windows will solve all our computing problems. The hamburger never looks as good as it does in the commercial. Demanding such things post-purchase, after the test drive demonstrates the deficiencies of the real product, is simply buyers remorse.

They deceptively claimed that for additional money the customers' cars would be enhanced in the future with FSD. They then made that claim again each time they failed to deliver. They continue to make this claim and people are trusting it.

A lie hides the truth.

But they lied a lot though, right? "Cross county automous summon in 2017" "Coast to coast autonomous drive in 2018" "Tesla driverless taxis in 2019" are a few examples. One may hold the view that Elon musk can make grandiose statements about solving AGI and nuclear fusion in 1 hour and make futuristic statements about that, but when does it go from projection to lying?

https://youtu.be/SuZYACWhYSI

10 years ago this might have read:

"If General Motors can't do it with its massive resources and sensors don't ever expect a Tesla to perform better than this."

FSD uses a camera, and works on all roads - and is dramatically better then supercruise.

source - test drove both, bought the tesla because of how impressed I was with supercruise.

Tesla’s unique strength, when it comes to autonomy as opposed to battery-electric vehicles, is selling hype.

And even if they don't reach FSD, they could still generate a lot of software revenue from Autopilot features.

And beyond that Tesla still does more then final assembly of EV and selling them to dealerships.

EDIT: In case my point isn't clear, if your position is predicated on the idea that anything less than full self driving is worthless garbage, you should know that is a pretty hot take. If you are instead having an emotional reaction and feel deceived I don't know what to tell you, you might have a very valid point but that doesn't appear to figure into evaluating whether a stock is under or over priced.

I see this as the future of public transport, where you'd have a combo of smaller and bigger vehicles (including self-driving minibuses and buses) running autonomously, with options to pay more in order to walk less, but only running where people actually need them . Similar to Uber and Lyft Pool before the pandemic, where you could pay less to be picked up on a main street intersection instead of your front door.

I think Waymo will make money off of their system sooner.

Tesla, on the other hand, is trying to make their cars work anywhere. It's a larger problem space, but the end goal is a far more robust product, with much lower long-term infrastructure spending.

I personally like the robust approach, but I think both approaches are long-term viable, albeit for different use-cases.

You on the other hand, have no explanation whatsoever as to why you think a non-HD map system is robust everywhere compared to a HD map system. If you claim that HD maps are less robust by definition, you clearly do not understand deep learning or computer vision.

Consider how in the early 2000s when people would mess around with AIM's smarterchild chat bot and gave it outlandish scenarios just to see how it would react. I have to imagine you might see something similar here with these cars but you're right that having a human onboard would probably curtail these interactions.

Messing with the cars means they have a full 3D video of you doing it.

  * the first accident between two fully automated AVs.
  * the first such accident within AVs from the leading company.
  * the first fatality in fully automated AVs
  * when uber or lyft or any similar company starts to use AVs more than human drivers in a metropolitan.
  * when a city gets half the driving hours in AVs.
  * when human drivers no longer get insurance coverage by a major insurance company.

And then there's the fact that there's a shot of a driver in the front seat shown for like 400 milliseconds. What's up with that? Why is a token driver required to be in the front seat? I thought the foundation of the service was automated self-driving that obviates that...? If for whatever reason it is required, alright, then own up to it and properly integrate it into the video so it doesn't feel so awkward.

Overall this presentation feels like it's hiding something, simply because it's so poorly put together.