Hacker News Reader

Comment by espadrine | Hacker News Reader

Waymo is currently the only SAE Level 4[0] self-driving car in use, right? The wheel must not be used. Even though it is limited to two cities.

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...

View on HN

Most of the commercial value is achieved at Level 4.

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

Did anyone claim _this_ was decades away? The claim is usually about when self-driving cars will actually be viable for anyone to use in arbitrary areas, not when experimental pilot programs are launched in individual cities.

> 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.

This is the 2nd market and the other is unmanned.

Okay... let me know when their market is the contiguous United States.

That's the point of this comment chain. They don't have to make it work everywhere in the contiguous United States to have a useful product. The (robo)taxi market is concentrated in big metro areas and that's what their focus is.

The thing is, with apps providing start and end point, they don't even need to be able to handle a full city to get a benefit. They can geofence really aggressively, and the moment they can ditch the safety drivers from a sufficient subset of journeys based on routes requested, they have an advantage.

They can cherry-pick the easiest routes. Discussing the entire "contiguous USA" is pointless.

Exactly. They can outright ignore really hard markets,. and be present in markets that have a decent percentage of routes they can handle and fall back to dispatching regular cars on routes their full automation can't handle.

Surely this is just moving the goal posts. If they can launch in Chandler and in San Francisco then a huge amount of the contiguous United States is on the table.

I completely agree. I was pretty critical of them when they were just a tiny slice of Arizona, but going from n to n+1 is really major progress.

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.

Nope, not even close. There's a very specific reason they've picked Chandler, and it's that there's relatively little weather and the roads are fairly simple (among many other things specific to Chandler).

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.

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

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?

It generally doesn’t snow in SF or Chandler. I personally like snow, so I wouldn’t describe that weather as significantly worse. But my understanding is snow makes it hard for lane keeping, and then there’s all sorts of edge cases like streets that aren’t plowed and require special driving techniques, people placing cones to reserve parking spaces, etc.

This is why I wouldn't be surprised if they choose Pittsburgh next.

I feel like San Francisco is a good progression - still generally good weather, but much more crowded, more traffic, more special cases, pedestrians and bikes, one-way roads, topography, limited visibility due to hills and no-setback buildings, construction, buses, etc. It's a significant leap in urban complexity, probably greater than 98% of the rest of the USA.

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".

No sorry, the vast majority of the US has significantly worse weather and/or more complex traffic than Chandler, AZ.

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

many places have fairly strong summer storms.

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

This market is manned and a serious challenge. The trial is Arizona is in an extremely quiet, predictable suburb as well as being small scale; unmanned but not as serious a challenge.

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.

SF seems like a pretty serious location to me. Granted, its still manned, but it makes sense that they'd start out manned in any new environment out of an abundance of caution. It seems likely that we could see a switch to unmanned in SF in the next 1-2 years if things go well. I have to think at least Waymo believes this is likely, otherwise they wouldn't be doing this.

We look right on track for my prediction of 2050

flimflam man Musk stated in 2015 that their robotaxis would be up and running by the end of 2015...

and he made similar claims each year after that.

You can thank him for the hype and cynicism

AFAIK no one recently claimed level 4 with a safety driver was decades away, and I was working on self driving until a few months ago so I was kinda "plugged" into the news. A handful of serious companies have plans and funding to deploy L4 taxi fleets or L4 features to consumer cars in 4/5 years.

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.

Don't have to pay safety drivers with L4.

When someone says 'in about 5 years' it means they have no idea when.

Most commercial value is at Level 2 for quite a while.

Level 2 is becoming a commodity. Its commercial value will fall to zero very quickly.

I disagree. Only the most simple basic highway lane keeping is commodity and that is only small part of overall driving.

Level 2 varies dramatically by car maker. On the low end you have companies like Ford and Honda doing lane keep cruise control, and on the high end you have Tesla which does automatic lane changes, takes on/off ramps, and stops for traffic lights. Tesla also charges a high premium for their software making it basically the opposite of a commodity.

I would think most commercial value comes when you eliminate the human labor from driving, which is the dominant cost of transporting when there is less than ~10 passengers. What makes you think Level 4 is the most economically impactful threshold?

How is there commercial value if you still need a driver in the car?

You don't need a driver in the car for level 4. That is level 3 and below.

The program was first publicly revealed nearly 11 years ago. People saying this project would take decades aren’t wrong.

That things are progressing about as expected. More specifically, Waymo is doing a bit better I than my general expectations, Cruise is about on par, and the rest slower.

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.

I think almost no commercial value is at 4. I'm fact, it could be negative value. Because you need the expense of a "digital driver" and the expense of a human driver.

It's basically level 5 or bust.

Level 4 (e.g. highways in good weather) is something that people buying cars will pay a significant amount for as an option in good weather. It does nothing for those who want robotaxis. But the former don't care about the latter.

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.

Level 4 doesn't necessarily mean just highways - it could also easily include suburbs with well-designed roads that have been mapped in excruciating detail.

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.

A selection of highways in good weather is just one example that seems particularly easy (relatively speaking), well-defined, and genuinely useful.

Current California regulations for remote operators are identical regardless of whether the car is considered SAE level 4 or 5. I'm not sure what constitutes "continuous monitoring" there (I could interpret that as anything between eyes on road at all times down to positive SOH messages arrived).

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.

> Most of the commercial value is achieved at Level 4.

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.

What does L5 get you that L4 does not (commercially)?

AFAIK most of the cynicism around self driving cars is not that it is not possible, but rather that it doesn’t solve the problems they claim to be solving. E.g. self driving cars won’t solve traffic (busses do), they won’t be safer (trains are), they won’t be convenient (unless you are rich enough to afford one), and at the end of the day, they are still massive things that takes tonnes of space and infrastructure and pollute a bunch. Just like human driven cars.

Those SAE levels apparently don’t specifically mention the geographic range the vehicle can operate in [0], but at some point that’s pretty important. Having “full automation” on a very tiny section of roads is hardly what I would call “orders of magnitude above the rest.”

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

The difference between SAE level 4 and 5 is explicitly whether the autonomy is geographically limited or not. https://en.wikipedia.org/wiki/Self-driving_car#SAE_Classific...

Still, Level 3 can be more "advanced" than Level 4.

How? Level 3 requires a driver who needs to take over when alerted. Level 4 is fully driverless.

No. It's fully driverless in geographically limited areas.

I'm aware. It's really between "maybe works, maybe doesn't everywhere (L3)" vs "works with no driver in a defined area (L4)". I consider the latter as more advanced as they are taking full responsibility for your safety.

It depends. If I could choose one car for person use, I'd take any modern adaptive cruise control + lane-keep assist system over an L4 that only worked in one city (even if it's a major city where I live). I'm not really sure how you determine which one is "more advanced," but I would consider the "level of automation" to be the portion of my normal driving habits that are able to be automated.

You're talking about personal driving, which is a different use case than robotaxis. For that, yes, you're better served with an ADAS system. It will take a while for L4 systems like Waymo to trickle down to passenger cars.

The point is that a "level 4 system" like Waymos is never going to trickle down to personally owned vehicles, because nobody wants to pay the hardware premium required for Waymo's autonomous system if it will be locked down to a specific city with specific routes.

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.

Any l4 autonomous system is capable of adaptive cruise control anywhere.

Am I the only one seeing the change between l4 and l5 abysmal? Also L5 description, at least in the Wikipedia article, is much more vague than the rest

In terms of convenience, it can be irrelevant.

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.

There is a full document available on the SAE website with full descriptions and technical details. Its much better than the common summaries.

Expanding the geographic range for Waymo is straightforward: do the same thing in a new place (plus new hurdles like snow, but that isn't currently the limitation to growth).

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

At some point "doing the same thing in a new place" is not going to be economically possible, unless the company is somehow able to continue getting money to burn (or unless they get much much better at bringing new places online).

I'd say the opposite. It is clear how Tesla goes from where they are to "full automation" – they just improve their models. Easier said than done, sure, but there is no conceptual leap required.

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.

FSD is still technically level 2. In Tesla's eyes and the law's eyes you are driving and in liable for what happens.

It's also (and this is important for Tesla owners to keep in mind) not-italicized-technically level 2.

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."

In the automotive industry we called this "level 2.999." I've moved on from automotive now, but in my experience nearly all of the new vehicle features in development, big or small, are developed at that level.

SAE levels are pretty garbage.

Reply on news.ycombinator.com