Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

Autonomous Car Progress

This site may earn commission on affiliate links.
Elon was asked in the earnings call whether or not he would add LIDAR to Teslas if it were free. He said: "If it were free? No..."

He is a true believer in the superiority of cameras! LIDAR is a crutch!
Cameras are obviously superior, you can't drive using only LIDAR. I guess this means they've finally solved phantom braking and collision into stationary objects?
 
I hope it is that good but we don't know yet. We have not seen what the FSD beta can do.

To me answering "no" to that question implies that they have now solved (or are very confident they will shortly solve) all the depth perception, object recognition, and environment mapping issues using their camera-only system, and have no limits on resolution of that depth perception.

You wouldn't answer no to that question posed that way if those problems had not yet been solved or there were still outstanding "difficult problems."

As an example, I'm curious how they will solve the problem of a load extending out of the back of a vehicle unsafely (and obviously they have to solve it without a flag). For example, a corner case of this vehicle with an illegal load (this one may be easy enough to detect). I don't know whether this is an easy problem for a camera system to solve, but it seems relatively easy for LIDAR. And obviously it is trivial for humans, as long as they are paying attention.

I'm actually less concerned about a rear-ending event (it probably would not get close enough). Slightly more concerned about dealing with and driving past such a vehicle when it is parked somewhere, etc.

Screen Shot 2020-10-21 at 3.45.09 PM.png
 
Last edited:
To me answering "no" to that question implies that they have now solved (or are very confident they will shortly solve) all the depth perception, object recognition, and environment mapping issues using their camera-only system, and have no limits on resolution of that depth perception.

You wouldn't answer no to that question posed that way if those problems had not yet been solved or there were still outstanding "difficult problems."

Elon has always been anti-lidar from the very start and firmly believes in his camera-only approach. I think he would answer "no" even if they were not close to solving these problems.

If Tesla has solved vision, great. I will be excited to get this FSD beta in my Model 3. But I want to see proof before I declare that Tesla has won FSD.
 
To be fair, even if it were free, it wouldn't be "free." They're so far down the no LIDAR pathway, they'd have to back up and sort of start over with a lot of the development, which would be a lot more development cost.

Yeah, the software would probably be the easiest part of the equation (though also not “free”). Whether it would improve performance enough to justify re-engineering cars in order to include it is a separate matter of course.

Elon has always been anti-lidar from the very start and firmly believes in his camera-only approach. I think he would answer "no" even if they were not close to solving these problems.

If Tesla has solved vision, great. I will be excited to get this FSD beta in my Model 3. But I want to see proof before I declare that Tesla has won FSD.

I was listening to the call, and the length of his pause before the “no” was somewhat concerning to me. I do think we read too much into his tweets or whatever he says though, so it’s not really worth extrapolating any conclusions beyond the simplest ones. He doesn’t think lidar is necessary and it’s clearly not cost-related, thus his answer (IMO).
 
Yeah, the software would probably be the easiest part of the equation (though also not “free”). Whether it would improve performance enough to justify re-engineering cars in order to include it is a separate matter of course.

Actually I would think software is the hardest part of the equation. Adding hardware is easy, utilizing to its fullest is all based on software. And rewriting code takes time and the unit is in months and years. In the free market, business time and developer time = lots of money.
 
Actually I would think software is the hardest part of the equation. Adding hardware is easy, utilizing to its fullest is all based on software. And rewriting code takes time and the unit is in months and years. In the free market, business time and developer time = lots of money.

It's certainly not trivial, but since lidar is only needed for perception and wouldn't necessarily factor into the rest of the driving logic, and they're already performing sensor fusion (cameras + radar + gps + odometry), they could add it as another input that would strengthen existing predictions for distance/orientation of objects. Hardware would at minimum require redesigning various structures of the car to add sensors (wiring, mounting points, integrating seamlessly into body panels, etc) as well as retooling production processes to support them, not to mention offsetting increased power consumption and whatever maintenance costs might end up being.

But I only say all this as someone more familiar with development of software rather than hardware and I'm likely underestimating the effort of sensor fusion with existing NNs trained on vision alone :confused:
 
  • Like
Reactions: hh83917
I was listening to the call, and the length of his pause before the “no” was somewhat concerning to me. I do think we read too much into his tweets or whatever he says though, so it’s not really worth extrapolating any conclusions beyond the simplest ones. He doesn’t think lidar is necessary and it’s clearly not cost-related, thus his answer (IMO)

"Yes" would mean that a million cars needing a hardware upgrade
 
  • Disagree
Reactions: mikes_fsd
"Yes" would mean that a million cars needing a hardware upgrade

Yeah, it was clearly a "gotcha" type question, and a "yes" (which would have been a perfectly reasonable response for capability that was completely free) would have been followed up with a lot of splashy news stories along those lines. As others have said, hard to read much into the response, since "no" was really the only possible answer for Elon.

But that being said, they might still need lidar to keep up with the competition. We'll see!
 
  • Like
Reactions: thewishmaster
It's certainly not trivial, but since lidar is only needed for perception and wouldn't necessarily factor into the rest of the driving logic, and they're already performing sensor fusion (cameras + radar + gps + odometry), they could add it as another input that would strengthen existing predictions for distance/orientation of objects. Hardware would at minimum require redesigning various structures of the car to add sensors (wiring, mounting points, integrating seamlessly into body panels, etc) as well as retooling production processes to support them, not to mention offsetting increased power consumption and whatever maintenance costs might end up being.

But I only say all this as someone more familiar with development of software rather than hardware and I'm likely underestimating the effort of sensor fusion with existing NNs trained on vision alone :confused:

Yeah, that make sense. I'm more involved in the software development industry myself as well.
I was thinking hardware engineering for Tesla would be piece of cake, considering the speed SpaceX is cranking out rockets so quickly. I was thinking there's a lot of work on software side and maybe not crazy hard logically, but could just be tedious work. Since they've always had the same set of sensors since AP2.5, they might need to tweak the new sensor and add it into their NN decision tree. I guess it depends on how modular they wrote their stack, I hope it's not like changing oil while driving the bus type architecture. :p
 
  • Funny
Reactions: thewishmaster
Yeah, it was clearly a "gotcha" type question, and a "yes" (which would have been a perfectly reasonable response for capability that was completely free) would have been followed up with a lot of splashy news stories along those lines. As others have said, hard to read much into the response, since "no" was really the only possible answer for Elon.

But that being said, they might still need lidar to keep up with the competition. We'll see!

there could well come a regulatory requirement for redundancy in sensors that can only be viable with combined fully functional lidar and camera. that is to say, both camera and lidar need to be capable independently providing full autonomy, for a car to be allowed driverless operation.

in particular, the other camera centric developer (mobile-eye) has effectively stated the above.
 
From what I am seeing, Tesla has achieved an autonomous driving prototype. The car has perception and planning for autonomous driving and can do driving hands-free. It still requires driver supervision of course since it is a beta prototype. Tesla still has a lot of work to do to get to driverless. But this is a big step. I applaud Tesla for the FSD "feature complete". Finally.

I think this is where the FSD race gets super interesting because we have several companies with FSD now but from very different perspectives and different business models. We have Tesla and Mobileye that are really aiming for the consumer market. Mobileye has really good FSD but not deployed to consumer cars yet. Tesla has FSD that can be deployed to a lot of consumer cars but it is beta and will require driver supervision. And then we have Waymo that still has the best FSD IMO based on capabilities and reliability but Waymo is aiming for the local ride-hailing market. Waymo is aiming to deploy robotaxis in local markets.

Exciting times.
 
From what I am seeing, Tesla has achieved an autonomous driving prototype. The car has perception and planning for autonomous driving and can do driving hands-free. It still requires driver supervision of course since it is a beta prototype. Tesla still has a lot of work to do to get to driverless. But this is a big step. I applaud Tesla for the FSD "feature complete". Finally.

I think this is where the FSD race gets super interesting because we have several companies with FSD now but from very different perspectives and different business models. We have Tesla and Mobileye that are really aiming for the consumer market. Mobileye has really good FSD but not deployed to consumer cars yet. Tesla has FSD that can be deployed to a lot of consumer cars but it is beta and will require driver supervision. And then we have Waymo that still has the best FSD IMO based on capabilities and reliability but Waymo is aiming for the local ride-hailing market. Waymo is aiming to deploy robotaxis in local markets.

Exciting times.

It doesn't look anywhere near feature complete to me. For a start there is no display of pedestrians at all. Can it even see pedestrians waiting at a pelican crossing (or whatever your local equivalent is)?

For that matter can it see cyclists and overtake them when safe?

So far it appears that it can barely cope with a roundabout and has very narrow situational awareness. Certainly no sign of a wide area birds-eye view showing other road users and their predicted paths.

That puts them behind where Waymo was over a decade ago when they were running Project Chauffer.
 
It doesn't look anywhere near feature complete to me. For a start there is no display of pedestrians at all. Can it even see pedestrians waiting at a pelican crossing (or whatever your local equivalent is)?

For that matter can it see cyclists and overtake them when safe?

So far it appears that it can barely cope with a roundabout and has very narrow situational awareness. Certainly no sign of a wide area birds-eye view showing other road users and their predicted paths.

That puts them behind where Waymo was over a decade ago when they were running Project Chauffer.

It is "Elon feature complete". ;)

But that is why I said in the other thread that Tesla still has a lot of work to do. It is an autonomous driving prototype. It's a big deal for Tesla. But it is far from driverless.
 
I know disengagement rates are not a perfect science but I do hope Tesla starts releasing some disengagement data for the FSD Beta when it goes wide. Disengagement data could help us gauge the overall reliability of the system on a large sample size that can help estimate how close Tesla might be to removing driver supervision. And I know disengagements rates cannot always be compared 1-1 but it would be interesting to compare Tesla's disengagement rate in a similar environment to the disengagement rates from say Waymo and Cruise to get at least get a ballpark estimate of how they compare.
 
I know disengagement rates are not a perfect science but I do hope Tesla starts releasing some disengagement data for the FSD Beta when it goes wide. Disengagement data could help us gauge the overall reliability of the system on a large sample size that can help estimate how close Tesla might be to removing driver supervision. And I know disengagements rates cannot always be compared 1-1 but it would be interesting to compare Tesla's disengagement rate in a similar environment to the disengagement rates from say Waymo and Cruise to get at least get a ballpark estimate of how they compare.
I hope they never ever release the disengagements data... until they flip the switch and say we offer generalized L4 on all our cars.
Then they can offer disengagement data.