Risk of Tesla camera-only self-driving

How does this work? Let’s say the board fails while driving. This isn’t hypothetical, this is a known point of failure. How then does a software update work around the lack of a functioning ADS computer?

The Jalopnik article suggests Tesla can beam over a temporary software fix that resolves some issues. Okay, let’s say the board goes out while while you are going around a curve at freeway speeds. Is there a software fix ready to go that can resolve the problem in time to enable a DDT fallback?

That seems incredibly unlikely. I think it is far more likely that regulators simply won’t allow current Teslas to operate at L4 on freeways, and in most other situations as well.

We’ve been off on a million tangents. Let me drag this back to where we started a bit ago (emphasis added):

Focus on the investing case.

Because of hardware limitations, I straight up don’t see how current Teslas meet the SAE/ISO L4 standards, except in very limited conditions.

This directly impacts the investment case. Future Tesla vehicles could well make tons of money as robotaxis and on autonomous driving software subscriptions. But Tesla hasn’t built any of those vehicles yet. That’s a hardware issue.

1 Like

The software work-around is, now anyway, already built into all of the vehicles, it’s not an OTA that happens after a failure, it’s coding that’s now in place in all vehicles to test and then do something other than normal if failures are detected.

You’re assuming a lot here:

  1. That current Tesla vehicles actually don’t meet the standard, which doesn’t say “redundant hardware,” but says the vehicle must do a “DDT Fallback” upon some failures. You’re guessing they can’t by your understanding of current Tesla vehicle hardware and that software isn’t already written to do so, or can’t be written to do so.

  2. You’re also guessing that regulators will come to the same conclusion as you.

  3. You’re also assuming that Tesla needs current vehicles to be L4 approved by regulators to be a good investment, and that upgrading some subset of “current vehicles” is either not practical or too costly to solve the issue in time for robotaxi deployments.

  4. But you yourself acknowledge:

If Tesla is making “tons of money,” then that’s good for the investment case, after all.

Don’t forget to account for Musk’s trillion dollar pay package. One of the 12 items Musk needs to do is to get to 10 million FSD subscriptions. 10 million FSD subscriptions at $100/month is $12 billion per year in revenue. However, Musk gets $83 billion for reaching this milestone. So that’s about 7 years of FSD revenue going straight into Musk’s pocket.

At $100/month, the margins on FSD have to be tiny. Let’s be generous and say TSLA gets 20% margin on FSD. Excluding Musk’s pay package, that’s $2.4 billion per year in profit on FSD. Accounting for Musk’s $83 billion, that means it will take 35 years for TSLA to just break even on FSD.

Another of the 12 items Musk needs to get his pay package is 1 million robotaxis. I suspect Musk will count any customer-owned car that ever does even 1 robotaxi ride in this count, so that’s going to be another $83 billion down the drain.

With Musk’s trillion dollar pay package, FSD and robotaxis are just a way to funnel money to Musk with the shareholders getting screwed.

3 Likes

..they didn’t..

Waymo has not published official numbers saying “this many calls for human assistance happen per mile/trip.” When asked about how often human help is involved, Waymo declined to share specific figures in news inquiries.

They say they staff remote operators to meet service quality standards, adjusting workforce as needed, but they don’t release a quantitative rate of interactions.

1 Like

How do you figure. The hardware is in every car so that is in the COG of the car, not the FSD. So, is the software, but even if you wanted to assume special software for FSD it would be $0, Development cost of FSD is R&D, not COGS. Seems like the margin on FSD is 100%. That makes FSD $60 billion or less than a year and five months.

Wow, looks like, after starting a whole thread and not getting much notice, you decided to quote your entire thread starter here. Well, it’s still wrong, both in how the pay package works as well as from where the money comes. I advise reading up before you post, especially before posting the same nonsense twice.

Yes, and regulators don’t ask for that information, or at least Waymo and others have interpreted the rules they operate under to not require that information on remote assistance.

And to add to that, it’s important to understand how this human help is involved. For instance, the term “remote driver” isn’t at all accurate. Waymo doesn’t have humans looking at monitors ready to hit panic buttons if vehicles do something wrong. What they have is reliance on the vehicle to mostly do the right thing, but just as importantly, ask for help when it’s not sure.

And when the vehicle asks for help, the humans still don’t “drive” the vehicle in the sense that they remotely operate the pedals or steering. What happens is that the human will do things like draw a path for the vehicle to take or tell it to pull over or tell it to wait for a thing to clear itself. In the past, Waymo would dispatch human drivers to the vehicles to sit in the driver’s seat and pilot the vehicle out of the situation, but to my knowledge Waymo hasn’t had to do that in a couple years at least.

The edge case thing is interesting, and the quotes of miles needed are not what many are interpreting them to be. Long ago, Tesla realized that they had a Dickensonion problem: both too much data and not enough. Far too much of just driving in a lane on a highway, not enough of weird things like a truck towing a truck towing a truck, or a truck carrying highway signs:

So, Tesla started, and still performs, culling/curation of the data they gather, as well as create synthetic scenarios of variations of what they’ve seen, or can envision. Tesla doesn’t need to train on 10Billion miles - they need to see 10 Billion miles to know they seen what needs to be seen. And this is Elon’s current guess, which as we’ve seen in the past, could be quite wrong.

Another note is that another advantage of a camera-only system is that when you want to generate synthetic data, you only have to generate images. And thanks to all the gaming engines that have been developed, there is a wealth of efficient code for doing that from 3D models. If your autonomy also depends on radar, USS, and/or LiDAR, then you also have to generate those signals as well, and they all need to be co-ordinated. Do-able, but harder.


What’s probably happening now with Tesla is:
• CyberCab design probably satisfies any regulators concern about acheiving a Minimal Risk Condition (eg DDT Fallback).
• CyberCab production line is being completed. We’ll see how quickly that ramps once completed, but Tesla has a good track record here.
• Tesla has registered 1,655 vehicles for the service with the California PUC. We don’t know how many are actually in service for the 400 square miles they cover in the SF Bay Area, and don’t what, if any, mods were made to those vehicles. I also don’t know Austin numbers.
• Tesla is using these safety driver/monitor rides in SF Bayt Area and Austin to gather the information requested by the regulators - interventions, accidents, etc. In addition to the rides, Tesla gathers data on the deadheading those vehicles take without riders, and probably doing rides just to gather data.
• We’ve seen some CyberCabs with manual controls being manually driven. We don’t know if that’s for vehicle testing and/or for FSD training and/or for FSD certification.

California has a 3-step process for robotaxi approval:

  1. AV Testing Permit with Driver (no paid fares)
  2. Driverless AV Testing Permit (no paid fares)
  3. Driverless AV Deployment Permit (paid fares allowed)

Tesla got approval for 1) back in Sept last year. They are gathering data so they can get approval to move on to 2) when ready.

The currnt paid safety-monitored Robotaxi service is permitted under a separate Transportation Charter Party permit that Tesla has.

1 Like

I’ll define terms so we can speak about the same things.

“autonomous driving with remote assist” = “L4 driving” = an AI model that can complete trips sufficiently safely but may sometimes request assistance from a human, the AI driver is capable of entering a safe state (minimal risk condition) when it encounters a situation requiring human assistance

Waymo appears to be demonstrating autonomy as defined by L4 driving above by accumulating millions of rider-only miles per week with published safety data (100+ million miles and growing).

So when I say Waymo has solved edge cases, I mean that they have solved edge cases sufficiently well in L4 driving to have achieved their published autonomy mileage and safety metrics and run a service of millions of miles per week.

To place Tesla in this context, Tesla has achieved basically zero true L4 driving miles.

It would be interesting for Waymo to publish data on frequency of human intervention. Presumably they have determined that this frequency is low enough for them to provide millions of miles of autonomous taxi trips each week. That’s about all we can conclude. (If the interventions are really high, then their taxi service wouldn’t be very viable.)

So my question remains, how did Waymo achieve this L4 driving progress with almost certainly much, much less real-world driving data than Tesla while Tesla has yet to begin accruing true autonomous miles?

Since you’re talking timing, you need to remember that Waymo and Tesla were and are going down different paths:

Waymo started with Robotaxis, but with an announced partnership with Toyota is supposedly pursuing autonomy for what they call POVs (Personally Owned Vehicles).

Tesla started with POVs, and then later (only recently) started pursuing Robotaxis.

You can’t compare items on the different timelines for different companies in a single timeline. Period.

If you want, you can argue whether Waymo is further along on Robotaxis than Tesla is on POVs, or whether Waymo is further along on POVs than Tesla is on Robotaxis. Is Tesla going to have a harder time without LiDAR (OMG, on topic for the thread title!), or is Waymo going to have a harder time pruning their sensor suite to be affordable for POVs made by Toyota?

And of course, this supposedly being an investing board, we might want to discuss who’s going to be making significant profits first. Right now it seems that when/if Tesla solves the Robotaxi software issues, it’s going to be able to scale up way faster than Waymo has - all these years later and Waymo is still in a literal handful of cities (5), and with less than 3500 vehicles. How long before Tesla scales up CyberCab production with just a couple cities to 5000 vehicles? And with us seeing exactly zero reported Waymo-equipped POVs (are they doing that in secret?), how long before that market is entered by Waymo-Toyota?

BTW, what’s Toyota’s best EV?

I see this path:

  1. Develop AV (L4, march of 9s to safety)
  2. Scale and improve AV in company fleet via taxi business.
  3. Sell autonomy in consumer-owned (and maintained/managed) vehicles/fleet.

Both Waymo and Tesla are on this path.

Step 2 has a dependency on 1.
Step 3 has a dependency on 1. and 2.
Step 3 at L4 (sleep while your own car drives, the promise) is years away.

Maybe you can argue Tesla will go 1. to 3. without 2., but they haven’t completed 1. yet and they are doing/trying/claiming 2., so hard to argue they are going to 3. without at least some 2.

Waymo is at step 2.

Bulls often argue that Tesla has a manufacturing advantage.

I would argue that solving autonomy is the harder problem versus manufacturing vehicles. The world is full of vehicle manufacturers but not so full of AVs with demonstrated human-level autonomy.

We can go back and forth, but what has happened and is happening supports the above.

What evidence do you have that Waymo will have a hard time pruning their sensor suite?

Again, let’s check the available data.

Looks to me like they have already done it with their next gen (6th) sensor suite:

Karp told Business Insider that Waymo aims to launch Ojai for public riders by 2026. The company has been testing the vehicle in several cities, including San Francisco.
…
The sixth-generation Waymo Driver will have 13 cameras, four lidars, six radars, and audio receivers.
…
5th gen has 29 cameras, six radars, and five lidars.

They cut the cameras more than in half and dropped one lidar.

As far as affordable for consumer-owned L4?

Who knows? That’s not happening this year or next year at any scale. We do know hardware prices go down over time and with scale.

1 Like

Thinking about the “bad weather” issue. I wonder if it is possible that if the weather is bad enough that the 7 or 8 cameras on a vehicle can’t see sufficiently well, that nobody should be out driving at that time, it may just be too dangerous overall. I think the data shows that many more crashes occur during this kind of bad weather.

Generalizing and removing details does not make the two paths the companies are on the same. If you refuse to see that pursuing different business models in a different order with different hardware/software stacks are, in fact, different, then our conversation is over.

I thought it was obvious that I was talking pruning the suite down enough to be viable for POVs.

It’s kind of funny that "Waymo chose Ojai as a nod to the company’s roots in California and to “inspire a sense of tranquility.” Because, well the car is from a Chinese company with manufacturing in China, and “Waymo took a deliberate step in removing any Zeekr branding on the Ojai.”

Additionally, because of tariffs, Waymo has scrambled a Korean company, Hyundai for their third vehicle before the second one has even done a singe paid ride.

So much for manufacturing being easy. Waymo continues down the path of modifying existing vehicles for its fleet. Why isn’t Google putting down the Capex to churn out purpose-built Waymo vehicles at scale? What do they know that you don’t?

The New Street analysis I linked earlier shows the monetary advantage of having less expensive vehicle capital costs, and no-one’s going to beat the cost of the CyberCab.

Yes, I’ve made this point before and it’s clear the J3016 accomodates this, even for L5:

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,’

True, but I don’t think this concern deals with (or only deals with) near whiteout/rainout conditions where no one should be driving. Rather, it’s making sure that the sensor suite can drive in conditions of bad weather where humans still can drive, but the sensor suite might have some issues. One such common issue is that the sensor’s ‘field of vision’ is relatively small, so even a small clump of dirt or snow, or even just a smudge or an insect, can cause some issues for the AV that a human would just “look around. So, an AV will need more elaborate cleaning options for the small area in front of a sensor than a human driver needs for a windshield.

Certainly not an unsolvable problem - but one that might require specific hardware to solve.

2 Likes

Let me see if I get this, on our discussion board.

If I don’t agree with your point of view, then the discussion is over?

Am I interpreting this correctly?

Open question:

1 Like

My guess is that Waymo is just more intelligent than FSD. I’ve watched a lot of FSD videos on Youtube and FSD still does a lot of stupid stuff like going the opposite way as shown on the map.

Plus, Waymo is operating in a geofenced area which probably means that less data is needed to achieve L4.

1 Like

If you want to insist on a point of view that is supported by ignoring facts, then yes. What you’re doing is like saying a bumble bee and a hummingbird are the same if you stand far enough away from them, and take pictures with a low-res camera.

You can keep posting your equivalent of “when did you stop beating your wife?” but I’m not going to take the bait.

1 Like

That was V12.6 on HW3 if we’re looking at the same video.

What about recent stupid stuff like a Waymo going onto train tracks?

Waymos have recently crashed into each other, too:

2 Likes

Hey, I wouldn’t ride in either one of them. They’re both dangerous. The only reason we haven’t seen FSD cars crash into each other is because they still have human drivers in them.

Wait until Elon removes the human drivers and then we’ll see FSD doing more crazy stuff.

Also, the video I saw where FSD went the wrong way was v14.2.2.3. I can’t find it right now because it was a smaller creator. I’ll keep looking though.

The data says otherwise, but I’m not surprised you’re not looking at the data.

Here’s a video where FSD v14.2.2.3 goes the wrong way:

Check out 36:30 of the video. FSD goes the wrong way then makes a U-turn right in the middle of the road. Horrible. A human would never do that.

FSD doesn’t have any data because there is still a human in the car to take over to keep FSD from doing dangerous stuff.

1 Like