Humans don't need LIDAR.
Until you have something like the human brain powering the car you cannot just use "two eyes"
That means the only thing we know can work is eyes + brain.
The answer to your question is: we don't know what's needed, so people saying "sorry you need LIDAR for FSD" is much more "drank the coolaid" than my stance of "we don't know, maybe no LIDAR + lots of training data can work just fine". Or, turn the question on its head – how much better at driving do you think humans would be if we had LIDAR on top of our heads? My guess is... not very. Hence my belief that the Tesla approach (less SOTA sensing, more training data) is a reasonable one.
You've still got things a bit mixed up. My stance does not imply that a computer functions as well as a human brain. I am not claiming that better sensors won't (eventually) give you better results. My stance merely implies that the brain (digital or wet) is a very important part of the equation, and just throwing sensors at the problem won't fix the fundamental need to have a highly sophisticated ML system running the show, and it's easier to get the training data you need for such a system if you're not using LIDAR.
“Less data is better for self-driving cars” is not the take I expected to hear from the Tesla mob today.
Tesla is clearly gathering more data for their self-driving program than Waymo, due almost entirely to the fact that the Tesla sensing package does not require installing LIDAR on every car.
LIDAR is not mandatory for better driving than humans.
The fact is Tesla is not submitting test data to the California DMV, which means they are either not developing a self driving system and have no right to speak to what technology is or isn't mandatory - or they are doing illegal testing, which speaks volumes about the trust you should put in their opinion.
Nobody has proven "better than human" capabilities, LIDAR or no.
Computers don't (yet) have the hardware nor software of a human brain. So they better come with extra tricks, like lidar.
That is, if humans only need two eyes to drive, then why not just only use two cameras? Isn't using more than two cameras (and radar) admitting that cars need more "help" than humans? If using more than two cameras is okay, why not go further and use LIDAR or HD-mappings?
I believe that Musk chose this line because he knows that LIDAR (in its current state) is too expensive to sell to consumers. Same with HD-mapping. We know that LIDAR based systems have achieved "better" self-driving than camera based systems.
It's still possible that Musk is correct. But at present, it seems to me that Musk's position therefore seems like a mostly ideological stance, and not based on any real data. That's why people doubt Tesla's ability to execute on Level 4/5 self driving. The market will punish ideological purists when they are wrong.
You could also argue that because LIDAR is expensive, it's hard to get billions of miles of training from LIDAR-equipped vehicles. This is not the case with Tesla's sensing package, which is racking up the miles as we speak. So sure, expense makes it less appealing from a consumer perspective, but its also possible that the difficulty of deploying LIDAR fleets makes it less likely to achieve FSD in the short term due to less capacity to generate training data.
Long-term we'll probably want all our vehicles to have as many sensors as is feasible. But nobody trying to get to FSD today is worried about what it will look like in 100 years – they're all trying to get their first.
Er, Waymo is doing a commercial L4 service. Cruise is doing pretty good. So no, they are not all "ideological positions". Also, no one except Tesla is targeting L5 which should tell you what the realistic goal is.
> You could also argue that because LIDAR is expensive, it's hard to get billions of miles of training from LIDAR-equipped vehicles.
LIDAR is expensive was a good argument in 2015. The prices are plummeting. Waymo reduced their top of line LIDAR cost from $75,000 to under $7500 and this was in 2017. They sell their short range LIDAR (Honeycomb) for under $1000. And their 5th gen hardware costs even less now (my guess is under $5000). Which way do you think the price will go as a number of companies mass produce it?
The fact is Elon made a bad bet a few years ago and didn't anticipate how much the cost would drop. Now Tesla is stuck with his ideological stance trying make a pipe dream work.
That last sentence seems like a strange one. "Stuck" would be if you massively invested in LIDAR and it didn't help. If no-LIDAR FSD is not as feasible as Tesla hopes, they can just... use LIDAR. And they'll arguably be in a better position tahn early adopters because they weren't investing when the hardware was expensive.
Waymo is L4. L5 means it can take you to your favorite campground (anywhere) and handle any scenario. Waymo says it’s not a realistic goal. But yes, the normal levels are not very well defined.
> If no-LIDAR FSD is not as feasible as Tesla hopes, they can just... use LIDAR.
Tesla has promised the current fleet has all the necessary hardware to make it work (they couldn’t possibly know this at the time). So they can’t walk back on that without incurring huge costs for retrofitting existing cars and development cost for overhauling their tech stack.
Meanwhile, they would’ve sunk money and time pursuing a much harder problem. If they had used additional sensors from the beginning, they could’ve provided much more value to their paying customers by now.
Edit: Just saw your edit.
> From what I've seen of Tesla's FSD beta, it could very probably be T4 if Tesla geofenced certain areas and worked out any particular kinks manually.
Yes, that’s what they should have done. Make realistic promises. I mean, you’re talking about making it work on _any_ road, _any_ intersection and _any_ scenario. Without additional sensors and mapping, it’s seems pretty unlikely (if not impossible) to me. Certainly given the timeframe Elon Musk promises.
It's your own fleet that's collecting the really good detailed data. And adding LIDAR to your own fleet isn't tremendously expensive. You don't have to worry about an extra 10k per vehicle crippling your sales.
Level 4+ using only cameras is risky and aspirational.
In truth, most of Elon Musk's high level goals are risky and aspirational. So far his record is pretty good, but eventually one or more are likely to fail.
The track record of making things work is good, but it's a very different kind of problem when you're allowed to redesign.