6 ms·

> The system didn't know what this was, so it smashed at full speed into the trailer. I feel like this will be a never-ending battle with a camera-based system

> The system didn't know what this was, so it smashed at full speed into the trailer. I feel like this will be a never-ending battle with a camera-based system. There will always be some visual situation not seen before. Shouldn't the default behavior be to slow/stop or throw up its metaphorical hands and release control back to the driver?

k_bx 2y ago
Adding a LIDAR will add a problem of "the system couldn't decide between believing a lidar or camera, decided to go with LIDAR which was wrong, and smashed into a perfectly visible wall".
- fredgrott 2y ago
  yes but given Tesla had lidar once and did not have these problems where does that leave your narrative? If I remember right it was lidar overriding the camera as the set up when Tesla was using lidar...
  - vel0city 2y ago
    Tesla never had lidar on production cars. They had radar.
  - k_bx 2y ago
    It doesn't look like Tesla ever was in a state that it didn't have problems. It's a very new area where all solutions are filled with unsolved problems.
- xvinci 2y ago
  This is just repeating what musk says though without giving it much thought. There are many ways to tackle this problem with different tradeoffs.
  - k_bx 2y ago
    Same answer can be given to the original question, couldn't it?
- VHRanger 2y ago
  All other successful self driving cars use LiDAR.
  - daghamm 2y ago
    Removing lidar was 100% a cost saving that was ordered by Elon. I am sure there are internal investigations showing all kind of problems when lidar is removed, and Elon choose to ignore them all to save a few bucks. Look at what he did at twitter. As soon as he was put in charge he removed all safety measures to save money and now twitter is full of bots and scammers. Why do people assume he runs Tesla any differently??
  - k_bx 2y ago
    By successful you mean none? No company is "successful", not even Tesla.
    - tsimionescu 2y ago
      Tesla is the least successful of all these companies. Waymo is successfully running a robotaxi business in a few select locations. Mercedes is successfully offering a full self driving solution in certain very limited but very well-defined parameters. Tesla is selling a driver assisted system that can't reliably turn the car or park it without manual intervention a decent percent of the time.
    - Zigurd 2y ago
      Waymo announced a $5B investment to expand service. They also announced a new sensor suite. The new vehicle will have no steering wheel. The vehicle and sensor hardware are real and being demo'ed. They charge real money and have 10,000+ paying customers per week. That's small compared to Uber, but the expansion signals that Waymo can scale up without hiring a lot more vehicle monitoring support people.
- chmod775 2y ago
  Or it just stop if either sees a wall. Plus with LIDAR being an active system you're more likely to see nothing or garbage where the ground and a road is supposed to be. Good idea to stop if there's a hole in reality! With camera+AI it's the opposite. For too many things AI won't really know what they are - it's mainly looking for things that it does recognize. With unclassified pixels everywhere, of course the default is to ignore what it doesn't understand, otherwise the system would never work. You can see this with the Tesla happily driving through some unrecognized pixels. LIDAR-based systems can operate without having to unsafely ignore things that they can't account for or that don't make sense.
- viraptor 2y ago
  Teslas have 3 front facing cameras with a lot of overlap, so that same problem has already been decided on. The issue is not new, it's just one more view to merge.
  - k_bx 2y ago
    Lol, it's like claiming that a human could have an ear instead of an eye and it would be "just as good", since it's just another sensor. Merging a camera with a camera, and a camera with a completely different kind of a sensor cannot be "the same problem". It's absurd.
    - lawn 2y ago
      What's absurd is saying that adding a better sensor to the mix would somehow make the system worse.
    - viraptor 2y ago
      I'm not saying anything about the quality change. I'm saying that adding lidar is not a change that requires new types of decision. They already have answers for "what to do if one camera identifies something but another doesn't". Same class of problem with the lidar, even if it needs to be solved differently.
- almatabata 2y ago
  That argument never made sense to me. If we hit a scenario where the camera makes a mistake, we will do the wrong thing 100% of the time. Adding more sensors would have made more sense to me but that would have made the cars more expensive and the existing fleet would need a retrofit.
onion2k 2y ago
I genuinely believe that more people would be put off buying a Tesla if the headlines were "Teslas randomly slow down" compared to "Teslas randomly kill people". People wouldn't buy a car that slows down occasionally because they'd accept it would happen to they and they'd find it annoying. They'd happily buy a car that randomly kills people because they'd tell themselves it's only happen to other people.
- BoredPositron 2y ago
  Most of the FSD following of Tesla is a cult so if it doesn't happen to them personally it doesn't happen at all.
- KingOfCoders 2y ago
  Spot on.
- daghamm 2y ago
  According to NYT Tesla has been involved in 200+ court cases. At some point new buyers attitude will change from "this will never happen to _me_" to "shit, this happens all the time".
- bostik 2y ago
  In other words: "my time is more valuable than your life".
whoitwas 2y ago
I agree. It's technically impossible to solve and that's why Waymo has had success and Tesla never will (with the current hardware).
- KingOfCoders 2y ago
  I wonder if, contrary to all Musk said, the Robotaxi will have Lidar and what Tesla buyers feel then. Because each Robotaxi is 23/7 on the road and will see much more usage compared to a car, so there should be more incidents per car, with an - assumed - increase in Robotaxis we'll see an significant increase in incidents. Also Robotaxi customers might complain more - or sue - compared to Tesla owners.
  - whoitwas 2y ago
    Why should anyone believe Musk? He promised a solar powered Tesla coup in 2008, and the stories have only become wilder since. I think robotaxi is the same sort of fantasy.
  - candiddevmike 2y ago
    This post assumes Robotaxi is a real thing and not another "high concept" lie.
- mattmaroon 2y ago
  It seems like Elon conflates what he wants to be true and what is actually true. It’s much better for the business model if cars don’t need expensive LIDAR to get full self driving. With just cameras, the cost to install them in every car and thus have the ability to sell FSD to every customer is low. With LIDAR you have to make cars with it and without it, and then you can’t sell it like an app to them later. (I suppose you could maybe do something like pre-wire the car for LIDAR and make it easily installable but you’re still probably going to need the user to take the car to the dealer which may be enough friction to reduce sales significantly.)
  - JohnFen 2y ago
    > It seems like Elon conflates what he wants to be true and what is actually true. I don't think he conflates those things in his own mind, only in his rhetoric. In other words, I think that Musk lies when it suits his purposes.
tommiegannert 2y ago
The only thing it needs to do is to build a 3D model and ensure the car's speed isn't faster than anything in front of it. It doesn't need to know whether it's a trailer or a bird or Superman. Just that if the relative speed is negative, then adapt the car's speed. Maybe ignore small things like insects or birds, if you're that inclined. There is Gaussian splatting now. That seems fast enough for real-time, and would provide all that you need for this. If enough Gaussians are about to hit the windshield, slow the f* down.
diggernet 2y ago
It's not a problem with a camera based system. It's a problem with an overly complex AI+camera based system. If the car has to recognize what the camera sees before it will react, it's going to hit things. But you can have a camera based system with good binocular vision to build a depth map, no AI stuff. It doesn't matter if it's something seen before. If something is there, you don't need to know what it is, just don't hit it. The AI stuff is helpful for a lot of self driving decision making, but detecting objects for collision avoidance needs to be simple and reliable.

k_bx 2y ago

Adding a LIDAR will add a problem of "the system couldn't decide between believing a lidar or camera, decided to go with LIDAR which was wrong, and smashed into a perfectly visible wall".

fredgrott 2y ago
yes but given Tesla had lidar once and did not have these problems where does that leave your narrative? If I remember right it was lidar overriding the camera as the set up when Tesla was using lidar...
- vel0city 2y ago
  Tesla never had lidar on production cars. They had radar.
- k_bx 2y ago
  It doesn't look like Tesla ever was in a state that it didn't have problems. It's a very new area where all solutions are filled with unsolved problems.
xvinci 2y ago
This is just repeating what musk says though without giving it much thought. There are many ways to tackle this problem with different tradeoffs.
- k_bx 2y ago
  Same answer can be given to the original question, couldn't it?
VHRanger 2y ago
All other successful self driving cars use LiDAR.
- daghamm 2y ago
  Removing lidar was 100% a cost saving that was ordered by Elon. I am sure there are internal investigations showing all kind of problems when lidar is removed, and Elon choose to ignore them all to save a few bucks. Look at what he did at twitter. As soon as he was put in charge he removed all safety measures to save money and now twitter is full of bots and scammers. Why do people assume he runs Tesla any differently??
- k_bx 2y ago
  By successful you mean none? No company is "successful", not even Tesla.
  - tsimionescu 2y ago
    Tesla is the least successful of all these companies. Waymo is successfully running a robotaxi business in a few select locations. Mercedes is successfully offering a full self driving solution in certain very limited but very well-defined parameters. Tesla is selling a driver assisted system that can't reliably turn the car or park it without manual intervention a decent percent of the time.
  - Zigurd 2y ago
    Waymo announced a $5B investment to expand service. They also announced a new sensor suite. The new vehicle will have no steering wheel. The vehicle and sensor hardware are real and being demo'ed. They charge real money and have 10,000+ paying customers per week. That's small compared to Uber, but the expansion signals that Waymo can scale up without hiring a lot more vehicle monitoring support people.
chmod775 2y ago
Or it just stop if either sees a wall. Plus with LIDAR being an active system you're more likely to see nothing or garbage where the ground and a road is supposed to be. Good idea to stop if there's a hole in reality! With camera+AI it's the opposite. For too many things AI won't really know what they are - it's mainly looking for things that it does recognize. With unclassified pixels everywhere, of course the default is to ignore what it doesn't understand, otherwise the system would never work. You can see this with the Tesla happily driving through some unrecognized pixels. LIDAR-based systems can operate without having to unsafely ignore things that they can't account for or that don't make sense.
viraptor 2y ago
Teslas have 3 front facing cameras with a lot of overlap, so that same problem has already been decided on. The issue is not new, it's just one more view to merge.
- k_bx 2y ago
  Lol, it's like claiming that a human could have an ear instead of an eye and it would be "just as good", since it's just another sensor. Merging a camera with a camera, and a camera with a completely different kind of a sensor cannot be "the same problem". It's absurd.
  - lawn 2y ago
    What's absurd is saying that adding a better sensor to the mix would somehow make the system worse.
  - viraptor 2y ago
    I'm not saying anything about the quality change. I'm saying that adding lidar is not a change that requires new types of decision. They already have answers for "what to do if one camera identifies something but another doesn't". Same class of problem with the lidar, even if it needs to be solved differently.
almatabata 2y ago
That argument never made sense to me. If we hit a scenario where the camera makes a mistake, we will do the wrong thing 100% of the time. Adding more sensors would have made more sense to me but that would have made the cars more expensive and the existing fleet would need a retrofit.

onion2k 2y ago

I genuinely believe that more people would be put off buying a Tesla if the headlines were "Teslas randomly slow down" compared to "Teslas randomly kill people". People wouldn't buy a car that slows down occasionally because they'd accept it would happen to they and they'd find it annoying. They'd happily buy a car that randomly kills people because they'd tell themselves it's only happen to other people.

BoredPositron 2y ago
Most of the FSD following of Tesla is a cult so if it doesn't happen to them personally it doesn't happen at all.
KingOfCoders 2y ago
Spot on.
daghamm 2y ago
According to NYT Tesla has been involved in 200+ court cases. At some point new buyers attitude will change from "this will never happen to _me_" to "shit, this happens all the time".
bostik 2y ago
In other words: "my time is more valuable than your life".

whoitwas 2y ago

I agree. It's technically impossible to solve and that's why Waymo has had success and Tesla never will (with the current hardware).

KingOfCoders 2y ago
I wonder if, contrary to all Musk said, the Robotaxi will have Lidar and what Tesla buyers feel then. Because each Robotaxi is 23/7 on the road and will see much more usage compared to a car, so there should be more incidents per car, with an - assumed - increase in Robotaxis we'll see an significant increase in incidents. Also Robotaxi customers might complain more - or sue - compared to Tesla owners.
- whoitwas 2y ago
  Why should anyone believe Musk? He promised a solar powered Tesla coup in 2008, and the stories have only become wilder since. I think robotaxi is the same sort of fantasy.
- candiddevmike 2y ago
  This post assumes Robotaxi is a real thing and not another "high concept" lie.
mattmaroon 2y ago
It seems like Elon conflates what he wants to be true and what is actually true. It’s much better for the business model if cars don’t need expensive LIDAR to get full self driving. With just cameras, the cost to install them in every car and thus have the ability to sell FSD to every customer is low. With LIDAR you have to make cars with it and without it, and then you can’t sell it like an app to them later. (I suppose you could maybe do something like pre-wire the car for LIDAR and make it easily installable but you’re still probably going to need the user to take the car to the dealer which may be enough friction to reduce sales significantly.)
- JohnFen 2y ago
  > It seems like Elon conflates what he wants to be true and what is actually true. I don't think he conflates those things in his own mind, only in his rhetoric. In other words, I think that Musk lies when it suits his purposes.

tommiegannert 2y ago

The only thing it needs to do is to build a 3D model and ensure the car's speed isn't faster than anything in front of it. It doesn't need to know whether it's a trailer or a bird or Superman. Just that if the relative speed is negative, then adapt the car's speed. Maybe ignore small things like insects or birds, if you're that inclined. There is Gaussian splatting now. That seems fast enough for real-time, and would provide all that you need for this. If enough Gaussians are about to hit the windshield, slow the f* down.

diggernet 2y ago

It's not a problem with a camera based system. It's a problem with an overly complex AI+camera based system. If the car has to recognize what the camera sees before it will react, it's going to hit things. But you can have a camera based system with good binocular vision to build a depth map, no AI stuff. It doesn't matter if it's something seen before. If something is there, you don't need to know what it is, just don't hit it. The AI stuff is helpful for a lot of self driving decision making, but detecting objects for collision avoidance needs to be simple and reliable.