What you omitted (without ellipses) is that my quote actually was:
And upthread I kindly answered every one of your questions, when no one else did.
Please see upthread.
1 Like
That’s correct. But if the camera can’t see the lane at all, then it can’t safely perform a DDT fallback.
As calibration and sanity check, Tesla has made robotaxi a core part of their business model since at least 2016. So it isn’t like they started working on this problem a year or two ago.
Yet, I’m unaware of any US jurisdiction who has agreed Tesla AVs can safely perform a DDT fallback with the necessary skill that they should be allowed to transport the public.
You and I can wave our hands all we want, but that’s the test. I believe there are good reasons why Tesla has not yet passed the test, and why current Teslas will not do so except under limited circumstances.
This is also why I do not believe the claim that software is the only limitation to Teslas being able to operate at L4.
1 Like
Not a problem. Musk will announce that he is working on a Time Machine, and lots of nodding heads will nod, and the stock price will go up because “the future.” (Also, ‘the past”.)
2 Likes
There is a difference between “I would like all of this to be operating correctly in order to keep driving” and “I can see well enough to pull over to the side of the road”.
1 Like
Yes, but that’s not the point. The point is if the car can’t see to the side, it can’t pull over safely in the event of system failure. If it can’t do that, it isn’t L4. No current Tesla has that capability due to hardware constraints.
This is relevant because the bull investor case for Tesla is that PE of 300 is justified in part because of all the potential robotaxi revenue that can unleashed with a software update. The reality is that no current Tesla has the necessary hardware to operate as an L4 robotaxi except possibly in restrictive environments such as geofencing, no freeway operations, good weather only, etc. And even then it will require time consuming local permitting, which Tesla has not yet achieved in any jurisdiction.
This doesn’t mean Tesla can’t achieve L4 robotaxi operations (sorry for the double negative). But the timeline of when it might be possible is pushed very far to the right. And that’s assuming vision-only is sufficient and we don’t know that yet.
2 Likes
Something I’ve been meaning to ask: what’s the big deal with adding Lidar? Heck, most iPhones have a Lidar sensor on the camera puck.
OK, I’m sure those are more rudimentary than would be required on a car, and you’d need more than one, certainly, but still…
Can it be that the addition of a few hundred bucks per vehicle is that big a deal? Apparently so, so here’s my theory:
Musk doesn’t like the look of it, and he’s convinced by his magical seeing powers that it won’t be necessary - someday, (because engineering solves everything, always) so he’s building all the cars without it and promising that it will never be necessary anyway, which nobody really knows.
It can’t be that integrating the software is “too hard”, since a half dozen other autonomy companies have already done it. And the cost of Lidar has come down rather significantly in the past few years (from what I’ve read) so it can’t really be that either.
I think it’s the “I can’t be wrong” syndrome, since he’s promised that all cars made will be able to self-drive with just a software upgrade (already proven not to be true) and it’s hard to backtrack on something so fundamental yet so vital to the elevated stock price.
3 Likes
I think it might be. I know a couple people in the auto industry, and they tell me the car companies put extreme pressure on their vendors to cut costs in every way imaginable. Literally fighting over pennies. Adding a component that costs a couple hundred bucks would cause the accounting department to have an aneurysm.
If you eliminate Musk’s foolish statements, Tesla’s approach makes sense. They have a well-rated driver assist feature and they are able to do it with a single sensor suite. Presumably, that makes it simpler and cheaper (some other manufacturers also use vision only, it should be noted).
Currently, nobody is offering L4 to the consumer market so it isn’t like Tesla is losing market share. Their FSD is good enough for now. They don’t need Lidar.
Segals’s Law:
“A man with a watch knows what time it is. A man with two watches is never sure.”
Musk has stated his rationale is simply that since humans can drive without LiDAR, so can AI.
As for cost, to cite an example from Lucid, its LiDAR driving package adds $5250 to the cost of the car. That’s certainly not all LiDAR (it also includes fancy micro-lens headlight high-beam shaping), but perhaps indicative.
Given the miles and years of data Tesla has collected, to believe Tesla knows it needs LiDAR but refuses to use it is viable only if Tesla still believes that it can solve the driving problem with the sensors it has. Which, as has been pointed out by TSLA bears repeated, Tesla has modified a few times. Even the latest HW4 vehicles now have front bumper cameras, added without increasing the hardware version number. That’s something I’ve been saying for years Tesla would need to add, even if just for parking (since USS was removed).
2 Likes
In the most recent emergency, many Waymo cars couldn’t pull over to the side safely. Are you claiming that because of that event, they aren’t L4 [anymore]?
2 Likes
I wouldn’t.
Nothing is 100%.
If the data supports that Waymo (or anyone) can operate with sufficient safety, including achieving minimum risk condition, then it’s good enough.
Why is everyone so focused on anecdotes and small samples?
It’s the 10s of millions of miles of safety validation data that matters (unless the incidents are so common that you can readily reject the hypothesis of human-level safety).
Medications have rare side effects, but if the totality of the data (clinical trials) supports sufficient efficacy and safety, then we adopt the medication.
Don’t be silly. It doesn’t become you. I claimed no such thing and you know it.
I’m simply restating what the SAE/ISO standard says. I provided an example from SAE regarding this issue above.
And for the record, I also addressed this scenario just a few posts above:
In the recent “emergency” what happened? Waymo vehicles stopped in their lanes on city streets. That’s not ideal, but it also seems like a reasonable thing to do in the event of a system failure.
I’ve read your posts. My take is you are smarter than you are pretending to be. Could be wrong.
1 Like
Yeah well that’s pretty thin. Engineers are often convinced they can engineer their way out of a problem when they can’t. We’re still working on fusion energy a half century later. Urban planners can’t figure out how to relieve congestion, and we don’t have self-driving cars almost a decade after it was declared “a problem nearly solved.”
We can’t build roads (economically) that don’t break up in short order, we can’t make computers secure, and nobody is riding around on Segways. We’re working diligently on battery capacity and making progress, but there’s still a long way to go before it matches energy density of petroleum, and we don’t have a good answer to carbon sequestration (though that’s political as well as technological.) The “re-usable” Space Shuttle was a dead end, as was the Concorde.
My point is: maybe “just video cameras” is a goal that’s just out of reach, and may be forever so. Maybe not, but it seems a big gamble to rest everything on it based on “because I say so.”
It’s not good enough for “We’ll just download this into all our cars and we’ll have robotaxis everywhere”, at least not yet. That’s the vision. That’s what supports a 300 P/E.
1 Like
The point is that you can keep making up things that you think TESLA can’t do and announce that this disqualifies TESLA, but reality doesn’t work that way. For starters, if one side camera fails, the overlap is likely sufficient to pull to the side. Possibly plus the memory of there being no car in that lane when the failure happened. Plus slowing down slightly and looking for a gap with rear facing cameras. Plus being in the right lane to start with. Possibly plus such failures being rare enough not to be considered. I.e., far from being an absolute block. Not to mention that I am sure one could make up such things about Waymo … a power outage for example … did you see them pulling to the side or just stopping? Lots of things are relative to how often they happen. And whether they can be fixed.
Exactly none, zero, of your examples are analogous to Elon’s belief that if humans can drive with limited senses, cars can too.
Fusion: No human counterpart to mimic.
Urban congestion: No human counterpart to mimic.
Roads that don’t break up: No human counterpart to mimic.
Computers secure: No human counterpart to mimic.
Segways: No human counterpart to mimic.
Carbon sequestration: No human counterpart to mimic.
On wait, “re-usable Space Shuttle,” that’s something a Musk company has gotten further than anyone else. The early stages even land themselves vertically.
Keep saying that while you still can. The recent progress has convinced me driving autonomy is coming, and relatively soon. 
2 Likes
Of course there are no perfect parallels because it’s speculation about the future. My point is that there are many instances of engineers thinking they can engineer their way out of a problem and finding that they can’t. Is this one of them? We won’t know until we know.
[Waymo, Zoox, GM, Aurora, Motional, Waabi, Mobileye, Mercedes, Toyota, BMW, Latitude, Kodiak, Gatik, Baidu, and others]
There are more than a dozen companies working on self driving, only one thinks it can do it without additional sensors. Maybe they’re right and all the other ones are wrong. You don’t know, and I don’t know. Someday we will. Your “strong belief” is no more indicative of truth than my skepticism. You might contemplate that.
1 Like
Are you sure about that? Seems like there has been some growing LIDAR skepticism. China maybe?
Camera-only (or “mostly”) might work to some degree, but the benefit of the simpler sensor suite might come at a cost along other dimensions:
- more difficult to train model (larger parameter space)
- higher variance model (more sensitive to overfitting)
- sensor suite less robust to environment (dirt, glare, water, challenge to infer 3-d from 2-d, etc)
- less sensor redundancy
The above could mean longer time to obtain a sufficiently safe autonomous product, a more narrow operational domain with respect to weather and tail events, and a less safe product per mile per compute size.
I’m not saying camera only won’t work, only that it could have some downsides/costs.
But in the other direction one has the problem of resolving conflicts in what different sensor types “see”.
Nope.
This is flat wrong for reasons explained many, many times.
No one knowledgable about statistics and machine learning believes this.
No one.
I refer you to an earlier thread:
1 Like