Humans?
There are people missing an eye, or not being able to use either of them. They are still able to navigate.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
how does the new end to end FSD work, need a block diagram from of data flow from the fleet to DoJO to an individual's car
- Thread starter Ridgeline
- Start date
There are people missing an eye, or not being able to use either of them. They are still able to navigate.
Ben W
Chess Grandmaster (Supervised)
For dolphins and whales in particular (and probably bats and sharks) these "specific" senses are very general-purpose for navigation and environment-sensing. But that's beside the point; none of these species could survive in their current ecological niche without these "specific" senses. Example: in your car, you are only actually pressing the physical brake about 0.2% of the time. (Especially in a Tesla with regenerative braking.) That doesn't mean you could have a viable car without it!
It's the same with radar/lidar. They may be "specific", but I predict they will still become necessary to get the car's reliability from 99% (where it is now) or 99.99% (where I think they can get with pure vision) to 99.99999% (wide-ODD L4 FSD).
Apples and oranges. I believe that Elon "hates" automotive Lidar primarily because he's hamstrung with a fleet of millions of Lidar-less cars (and years of over-promises regarding those cars) that require him to hate it. His initial valid reasons for avoiding Lidar, which were that it was crazy expensive and overkill for L2 autonomy, no longer apply, as he is now chasing L4 functionality in a cheap-Lidar environment.
Last edited:
Ben W
Chess Grandmaster (Supervised)
People missing both eyes may still be able to navigate their local environment (albeit with significantly reduced efficiency), but they can emphatically not drive, either legally or practically. I do grant, however, that FSD is currently a much better driver than the average blind person.
To the car analogy, no one is suggesting that the visual cameras be replaced by pure lidar/radar. We want the car to function as well as possible; indeed Tesla has an obligation to make the car function as well as possible, since lives are literally on the line. (See my post in the Trolley Problem thread.) Given that the cost of high-quality sensors of all types has plummeted, this to me implies they should use more (and more varied) sensors, not fewer.
Last edited:
Ben W
Chess Grandmaster (Supervised)
?? My post was comparing cameras to lidar.
Congratulations! Now drive 7,000,000 miles and see if you can still make that claim. (That's what's required for L4 FSD.)
Most of the startup companies in the Dot Com bubble also came and went, or closed down. Because obviously the Internet was completely useless! \s That argument is like saying that the NYC Subway is a terrible environment for cockroaches, because you see so many dead cockroaches there.
The truth is that most Lidar startups have struggled for two reasons. First, because most startups in ANY industry or environment always struggle. (Competition is fierce, and not all companies are competent.) Second, because the TIMELINE for AV's was overoptimistic; NOT because Lidar's usefulness for AV's was overhyped. AV's are inevitable; the Lidar startup bubble was just a bit too early. Most forecasters are predicting a 20% annual growth rate for the Lidar industry over the next decade.
Also note that, to my knowledge, Tesla is the ONLY developer of L4 autonomous vehicles or systems in the world (I won't even say "manufacturer", because they aren't manufacturing L4-capable autonomous vehicles yet) that hasn't fully committed to Lidar. To give an example, Baidu's Robotaxi costs $28,000, and has 40 separate sensors, including five Lidars.
Elon's original criticism of Lidar was that it was too expensive, and that it was a mistake to lean on it too heavily. (By e.g. mapping cities down to the millimeter and then relying on Lidar as the primary sensor for navigation.) The expensive argument is no longer valid (see Baidu). I do completely agree with Elon's second argument, which is that pre-Lidar-mapping is not the right approach for what Tesla is trying to do. And I'm not suggesting Tesla should go back to that approach. Only that by incorporating the strengths of Lidar into their existing system they would make it far more robust, a necessity for achieving L4 and eventually L5.
Ben W
Chess Grandmaster (Supervised)
This would only be applicable in the medium term (3-6 years), where the autonomous systems are not quite good enough to drive without this crutch. But note that the auto fleet will take about 20 years to substantially turn over (such that more or less "every car on the road" would have this technology). By that time, autonomy will almost certainly be completely solved at an L5 level with vastly superhuman reliability, and there will be no need for this at all.
stopcrazypp
Well-Known Member
Except there is zero evidence that lidar will provide any help to the problems FSD is currently facing. Maybe Tesla will eventually reach a point like that, but it's clear it's pretty far from that at this point. As noted very early on in this thread, Tesla's major problems are not in the perception stage, but rather the planning stage.
Ben W
Chess Grandmaster (Supervised)
Completely agreed that today's FSD makes significantly more planning mistakes than perception mistakes. But in my experience (~30,000 miles on various flavors of FSD), it still makes plenty of both. Lidar obviously doesn't help with the planning, but I expect perception to soon loom a lot larger as Tesla continues to improve the software on the planning front.
How far do you believe cameras can take FSD towards being a robotaxi? Assuming daytime and clear air, can cameras provide the reliability needed for an L5 system? From there, how far into fog at night can a camera system go? At what point would additional sensors be needed?
That's again, an absolutely made-up response.
He doesn't hate LIDAR, it's just not needed. It's just as easy to calculate the same information from video cameras. And you end up with even more information from cameras than you get with LIDAR.
Again, "seeing" hasn't really been a problem for years.
And yes, in ways LIDAR is still overkill. The accuracy isn't needed. And it is still an extra, unneeded cost.
Waymo recently had an issue and the news showed one of the cars.
There's no way that many sensors can be economical.
Until a single car is deployed driverless without a similar sensor suite, it’s just wishful thinking that Tesla pr anyone else will get to autonomy with camera-only.
Perhaps it will happen, perhaps not. Right now, we are at zero cars deployed driverless.
Sensor cost is or will be negligible at scale compared to have a human driver.
Ben W
Chess Grandmaster (Supervised)
In good lighting conditions, yes. In adverse lighting conditions (e.g. rain, fog, dirty lens, low light, sun glare), no. The cost for a current high-performance Lidar unit is ~$1000 these days, and will probably drop in half again in a couple years.
The Baidu Robotaxi has 40 sensors, including 5 Lidars, and costs $28,000. (For the whole robotaxi, not just for the sensors.) Sensor cost is not the issue. Also, with modern designs all the sensors (including Lidar) can be incorporated into the existing body shape; they don't need to stick out like this.
Ben W
Chess Grandmaster (Supervised)
Lane markings only need to be visible from a few meters away to establish lane position, and even compromised vision + the car's own lighting can handle that. Most road signs (e.g. speed limits) can be inferred well enough from maps, and again even compromised vision can still make them out at close range. It's medium-range detection where Lidar+Radar has the most value in compromised vision environments, for detecting e.g. obstacles that compromised vision cannot. Or confirming lack of obstacles, such as when vision sees a shadow that looks like a car that would otherwise trigger phantom-braking. Pure vision could be made arbitrarily safe if you're ok with a ton of phantom braking. (Though even then you might get rear-ended, so maybe not.) Adding radar+Lidar would make for a much more comfortable and enjoyable driving (or being driven) experience in marginal weather conditions.
All yall anti-FSDs are starting to get really worried. You gotta beef your game up. You got to come up with better arguments.
Vision took me on a 70 mile trek the other night. Without HD maps and on roads rarely travelled by Teslas.
At this point, the more you dig in, the funnier you are going to look. And right now it's getting to be comical.
Similar threads
