Waymo self-driving cars -- progress

[Goofyhoofy]

@mostlylong

I’d be interested in specific responses to these specific counterarguments.

[1]Finally, while I enjoy to speculate, forecast and guess about the future, I am not wed to any particular view.

[2]When such information exists, I want to see data, evidence, valid logic, etc and that will guide my view.

[3]Maybe some information new to me will surface here, I don’t follow the AV world as closely as others here.

1. OK.
2. OK.
3. OK.

Anything else? Meanwhile since such data is not forthcoming, not from Waymo and certainly not from Tesla, all we can do is speculate. You think we shouldn’t do that?

[Tamhas]

@albaby1

But if they do expand to California, then we’ll have more of an answer.

Given how little we really know about what Waymo is doing in CA, let me be skeptical …

[mostlylong]

@Goofyhoofy

Anything else?

These are the examples I provided upthread, quoted again below.

The first example is a question about how to measure (define) “extra/more information,” driven by a realistic, but non-technical example.

The second example illustrates that having “extra/more information” might require more (not less) analytical processing (more or smarter or “less dumb” brain power).

I don’t know the answers, but I believe these are very legit examples that counter or maybe deepen the discussion on the claims made upthread.

@mostlylong

Two basic counterarguments.

First:
System 1 could be very high resolution camera data, high spatial and temporal resolution. System 2 could be lower resolution camera with LIDAR data. Does S1 or S2 have more “information” with respect to AVs?

I would say you and I don’t really know this, and can’t even make a statement about this (unless maybe you want to define terms more precisely).

Second:
Processing sensory inputs (including synchronizing the data in space and time, and also calculating correlated, interacting features) from multiple sensor types (camera + lidar) might require more analytical processing (somehow appropriately defined) relative to processing from a single sensor type (camera only).

One can imagine additional neurons and connections and specialized features (in artificial and/or biological intelligence) to process vision plus sound, versus either alone.

I would say you and I just don’t know.

[dividends20]

It is ALL about real world data. The debate about vision only or vision + multiple sensors is over.

Tesla’s has automated data annotation with video feeds coming in from millions of cars all over the world every day and is using for model training.

On the other hand, Waymo is paying drivers to drive around and collect data and hdmaps. You cannot generate the data synthetically. The fleet is very small and expensive even when compared with Uber. This is a dead end.

[bjurasz]

The debate is NOT over for the very simple reason that Tesla has yet to have a working robotaxi service, and yet Waymo does. This, after years and years and millions and millions of miles driven and data collected. I’m puzzled as to how this debate is over.

As per synthetic data, this is exactly what Nvidia does with their Omniverse system for training systems from self driving to autonomous robots.

Disclaimer: I work for Nvidia, but have zero connection to the Omniverse system or data generation of any kind.

[syke6]

@bjurasz

The debate is NOT over for the very simple reason that Tesla has yet to have a working robotaxi service, and yet Waymo does. This, after years and years and millions and millions of miles driven and data collected. I’m puzzled as to how this debate is over.

I was scratching my head as well. The debate is over and the company who doesn’t have a robotaxi won???

[mschmit]

@bjurasz

As per synthetic data,

I think synthetic data is required to augment real data. AI needs lots of samples of each type of situation in order to properly train the models. There are two ways this is done. First, take an existing real image/video and augment it, such as making it a mirror image, adding noise, adding adverse weather conditions changing the color, brightness and so on.
Second, completely fabricate new data samples. This is important when you want to simulate scenarios that happen infrequently but are dangerous enough to want to add to the dataset…for example cars running red lights and potentially or actually smashing into your car or a car next to you so your car can avoid or minimize damage or injuries. These types of simulations could be run in all different lighting, weather, time of day (think building shadows) and with many different speeds and arrival times of the other cars.
You just aren’t going to happen to get this type of data by collecting lots of dash cam videos.

Mike

[mostlylong]

In response to:

@albaby1

So for any given level of functionality, it will be able to perform that level of functionality with a “dumber” brain than a car that lacked the extra sensory data and map information.

I wrote:

@mostlylong

The second example illustrates that having “extra/more information” might require more (not less) analytical processing (more or smarter or “less dumb” brain power).

I don’t know the answers, but I believe these are very legit examples that counter or maybe deepen the discussion on the claims made upthread.

@mostlylong

Two basic counterarguments.

First:
(see upthread)
…
Second:
Processing sensory inputs (including synchronizing the data in space and time, and also calculating correlated, interacting features) from multiple sensor types (camera + lidar) might require more analytical processing (somehow appropriately defined) relative to processing from a single sensor type (camera only).

One can imagine additional neurons and connections and specialized features (in artificial and/or biological intelligence) to process vision plus sound, versus either alone.

I would say you and I just don’t know.

Since no one answered the above counter examples, I’ll take a quick stab.

Here’s some info to help understand example 2, from above.

Bolded parts indicate additional sensor data (or “extra information”) might sometimes require more analytical processing (as opposed to “dumber” AI) to maintain a given level of system performance.

Yeong et al Sensors 2025 , 25 (3), 856

AVs are generally equipped with a diverse set of sensor types, including cameras, Lidar, radar, ultrasonic sensor, and GPS, each with distinct operational attributes that contribute to their strengths and weaknesses. For example, radar is resilient in poor weather but offers lower spatial resolution; Lidar offers high-resolution depth information but is computationally intensive; and cameras capture rich visual detail but are sensitive to lighting and weather conditions. Nonetheless, integrating these diverse sensor types introduces significant complexity in algorithmic design and computational processing. Doing so requires sophisticated and innovative fusion algorithms that can handle differences in sensor data format, resolution, and spatial-temporal synchronization [11] while maintaining the overall AV system performance and reliability.

[Kingran]

@syke6

All indicators say that Tesla isn’t close to flipping the switch

But Trump is getting ready to flip the switch…

[Kingran]

@dividends20

This is a dead end.

Oh boy, you do suffer severely from commitment bias. Do you even entertain a remote possibility of both being successful?

[dividends20]

@Kingran

Do you even entertain a remote possibility of both being successful?

Waymo is a dead end. It is a failed business that cannot survive without big daddy Google. It will need to pivot to “vision only” model if it wants to survive.

The debate is long over.

It is ironic that Google is the one that published the transformer paper which is enabling vision (and OpenAI etc.). Tesla was also stuck at a local minima and spaghetti code for long time.

[Goofyhoofy]

Not for nuthin’ but a Boeing 747 black box monitors over 1,000 discrete inputs simultaneously. And as we have seen from the 737 Max debacle, it uses those inputs to make decisions that can override the human pilots’ directions (obviously in this case in error).

Typically an airliner has 3 processors monitoring the range of inputs, and at least 2 of 3 must agree before deciding on a course of action: from things as simple as broadcasting a warning (“Terrain, Terrain! Pull up!”) to taking over various inputs to try to recover from stalls or other perilous situations.

So clearly there are microprocessors capable of handing multitudes of simultaneous inputs and making “independent decisions” based on them.

PS: You can get one for yourself for about $225 on eBay.

[bjurasz]

@syke6

The debate is over and the company who doesn’t have a robotaxi won???

He still hasn’t addressed this.

[dividends20]

@bjurasz

He still hasn’t addressed this.

Tesla has won.

There were hundreds of startups with sensor based autonomy that have folded.

It will be apparent in coming months.

[bjurasz]

Until Tesla is making money with a working robotaxi service they have not won.

What do we know today?

1. Waymo’s solution is expensive, is difficult to scale, and may not be cheaper than human drivers
2. Tesla’s system is cheaper but has not proven to be good enough to actually work as a robotaxi. It’s not even universally that much better than SuperCruise or BlueCruise for that matter.

Just because Tesla’s system is cheaper does not mean it will actually work. It is very likely that NEITHER system will make for a viable, profitable robotaxi service. There is the chance that neither will win. And you cannot claim victory until Tesla has a lot of time and a lot of miles actually running a service.

EDIT: Forgot to add “what we know number 3”: Tesla promised us this ability 6 years ago.

[dividends20]

Elon Musk

Sawyer Merritt

[MarkR]

@bjurasz

It’s not even universally that much better than SuperCruise or BlueCruise for that matter.

I agree that Tesla isn’t yet proven as a robotaxi, but this particular statement isn’t at all true. Tesla “FSD” is not better or worse than BlueCruise or SuperCruise … that’s because it is completely different than those two things. Tesla FSD drives me from my driveway to the supermarket, to the movie theater, to the mall, to my friends house, etc. It drives there, via surface roads, stopping at stop signs and traffic lights, it turns, it adjusts speed limits as necessary, and it’ll enter the highway, drive, and exit the highway, all on its own. And sometimes when it arrives at the destination it’ll even choose a parking spot and park the car! All I have to do is enter the car, choose my destination and then press the blue button to “GO”. BlueCruise and SuperCruise doesn’t do those things and doesn’t claim to do those things.

[NewInvestorKinda]

@dividends20

Of course you know more than Elon. You have so many accomplishments.

Elons only accomplishment here being that he acquired (bought) the company and got himself paid handomsly though stock option early in their technological adoption cycle? Maybe also as a very successful salesman for them?

[mostlylong]

@syke6

The debate is over and the company who doesn’t have a robotaxi won???

Consider two engineering teams for AVs.

Team 1 has a sensor array that is camera only.
Team 2 has a sensor array that is camera and LIDAR.

Both companies invest in research and design and collect many miles of test data for many years.

After years of data collection and testing, consider 3 distinct outcomes:

1. “Camera only” is the most economic to manufacture and maintain (function per dollar, not including the investment in research and design) for a commercial AV (some defined functionality)
2. “LIDAR only” is the most economic to manufacture and maintain for a commercial AV
3. “Camera and LIDAR” is the most economic to manufacture and maintain for a commercial AV

Team 1 may win* (have best unit economics) under only outcome 1.
Team 2 may win* (have best unit economics) under any of the three outcomes.

Hence, in the admittedly simplified situation above, Team 2 has the more robust strategy with respect to risk of success of different sensor configurations. (Team 2 has the more diversified sensor solution.)

There is the possibility that Team 2 requires higher investment costs to research and design more sensor types (2 vs 1), but at today’s stage of AV development I don’t think we can necessarily assume higher costs for Team 2 (because we don’t yet know how all of this tech is going to evolve into a commercial product, amount of data and testing involved (including for various sensor configurations), regulatory regime, and probably many other factors). If a team exhausts funds before developing the defined functionality, then that team fails in all outcomes.

*note the qualifier “may,” ultimate success will depend on the long-term economic viability of the commercial AV, the above outcomes are necessary conditions for economic success (superior sensor array and associated tech unit economics), the point is to illustrate the relative scope for possible success given different sensor arrays and possible technological outcomes

[buynholdisdead]

Musk said today that the pictures of Robotaxis were actually pictures of cars that they just produced going to be parked. Tesla also confirmed that all robotaxis will be remotely monitored so if something happens they can be taken over by a human. But that means you better hope they have connectivity at all times. But how many times have you lost signal on your cell phone while driving?