Here's an interesting glimpse into Tesla's corporate thinking when it comes to LIDAR:
Disruption Research on Twitter
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Will Tesla ever do LIDAR?
- Thread starter diplomat33
- Start date
@diplomat33
Or alternatively the VP was just parroting Musk what had planted in their head.
Look.
The problem with this ”first principles” question is that you can bend it to what you want.
I could make well founded first principles argument that autonomous driving shoud forgo the empirical modelling based on humans and instead go directly to the optimal way of sensing the world — on the level of science, not on the level of human anatomy — that, who knows, might be Lidar or whatever, not vision.
It has never, ever made sense to me someone invokes ”first principles” and then proceeds to compare the selection to human anatomy. That is not first principles, that is empirical modelling at its finest.
Or alternatively the VP was just parroting Musk what had planted in their head.
Look.
The problem with this ”first principles” question is that you can bend it to what you want.
I could make well founded first principles argument that autonomous driving shoud forgo the empirical modelling based on humans and instead go directly to the optimal way of sensing the world — on the level of science, not on the level of human anatomy — that, who knows, might be Lidar or whatever, not vision.
It has never, ever made sense to me someone invokes ”first principles” and then proceeds to compare the selection to human anatomy. That is not first principles, that is empirical modelling at its finest.
Daniel in SD
(supervised)
"First principles" say that cars should be bipedal machines instead of using wheels!
I did chuckle when I read the VP guy say that talking about using LIDAR could get you fired at Tesla. I'm thinking it might be more about being on the same page as the boss (Musk) than it is about first principles although Musk thinks it's about first principles. The truth is that Musk is super anti-LIDAR and if you want to keep your job, you better agree. I am sure the word to the newbies around Tesla HQ is "don't talk about LIDAR when Musk is around". LOL.
"First principles" say that cars should be bipedal machines instead of using wheels!
That’s the thing, right... first principles is not analogy, it is counter analogy. The wheel does not appear in nature yet in long-distance transport it is superior to human legs...
Whenever I see people talk of first principles and then jump to the analogy of humans driving cars I cringe. That is decidedly not first principles to compare to what humans need to drive.
Which is why Musk's alter ego developed his own LIDAR to solve the orbital rendezvous problem, of course.
I guess I should have been more specific. Musk is not anti-lidar for everything, he's just anti-lidar for self-driving cars.
The problem with this ”first principles” nonsense is that too often it has become a hammer to hit your opponents with, instead of a real analytical tool.
First principles thinking suggests that if you select Lidar just because everyone else does, you are not employing first principles. Similarly it suggests if you select vision only because humans also use vision, you are not employing first principles. But that’s a big IF — you don’t know why some company selected what...
Which is, curiously enough, indeed why SpaceX depeloped a Lidar based solution. The human solution of using eyes for the docking would not have been the optimal solution from the point of view of science. Even though humans indeed dock spacecrafts using only their eyes.
First principles thinking suggests that if you select Lidar just because everyone else does, you are not employing first principles. Similarly it suggests if you select vision only because humans also use vision, you are not employing first principles. But that’s a big IF — you don’t know why some company selected what...
Which is, curiously enough, indeed why SpaceX depeloped a Lidar based solution. The human solution of using eyes for the docking would not have been the optimal solution from the point of view of science. Even though humans indeed dock spacecrafts using only their eyes.
I tend to agree with the logic Tesla put forward last week for this particular case.
Lane markings, signs, brake lights, traffic lights - we designed our entire road system around eyeballs.
Because of that, you *have to* solve the image recognition/processing problem to safely navigate the roads we use today. And if you've done that adequately, you don't benefit much from adding LIDAR as well.
If we'd taken a first principles approach to designing roads for autonomous cars with no human drivers, I'd probably have a different perspective.
Last edited:
I think it is important to recognize — and judging by your post I do think you recognize it— for the sake of others: First principles thinking does not somehow magically suggest or claim Lidar is bad for autonomous driving and vision is the way to go.
First principles does suggest the avoidance of analogy (be it others in the sphere or indeed humans) and it does suggest reducing things to the irreducible. Yet, as I believe @Saghost points at the end of his post, we don’t really know what that irreducible is when it comes to autonomous driving.
It might be tempting to suggest the irreducible in this case is vision, because you will need vision for many things in autonomous driving, so you can not do without it, but that would be a simplified view. It would be analogous to the technologies we employ, not rooted in the science or the ultimate goal underneath.
First principles is a question. It is not an answer. We don’t know the answer to autonomous driving yet.
I don't really see the issue with describing this as first principles thinking. Isn't the question "how to safely navigate the world around you?" As @Saghost said, all the key information about navigating the world is visual. Other solutions may solve a small portion of the equation better ("is there an object in my desired path"), but even knowing what the path is is a visual problem of identifying lane markings, parking lot arrows, etc. How is vision not the irreducible component?
Even if vision would be the irreducible component, it does not automatically follow that passive cameras and separate headlights are irreducible, as they are already a solution above the basic science involved.
For example, what is to say infrared cameras with active infrared beacons... or indeed active lasers wouldn’t be superior to passive cameras and separate headlights?
I don’t think we know yet.
StealthP3D
Well-Known Member
Are you trying to imply that self-driving cars have unique requirements compared to a spaceship that can dock to the International Space Station? And that Elon Musk wants the sensor suite to have the best match to these two applications having different specific needs?
Who woulda thunk?
This is a case Musk already made on AP2 announcement in 2016 so nothing new of course.
That means he was dead set on the Lidar question long before Tesla really knew how to solve the actual problem. Interesting to see how the prediction pans out.
Personally I don’t think Lidar is necessary (never did, thanks to MobilEye — not Tesla) but I also don’t think it couldn’t be very useful. There may be many solutions to this.
StealthP3D
Well-Known Member
There is nothing preventing anyone else from developing the technology with whatever technologies they believe most likely to succeed. That would be a real kick in the pants to finally show ol' Elon Musk what a fool he is!
I say, go for it! But you had better hurry!
Hint: I heard ultrasonics is the "secret sauce". This makes a lot of sense because that's how the Polaroid SX-70 focussed its lens. So it must be good at determining distances. Elon just doesn't have a clue!
@StealthP3D Musk relies on ultrasonics for parking since there are no nose cameras, so he KNOWS the secret sauce already!
The opposition is DOOMED! DOOMED I say!
The opposition is DOOMED! DOOMED I say!
StealthP3D
Well-Known Member
I'm sure there is not just one solution to anything. In this case we are not interested in all possible solutions, we are interested in the one that will be developed first to the point that it is capable enough to achieve commercial demand (mass adoption). That is how we will know who chose the best path.
For the very reasons Musk has repeatedly explained to those who know a lot less than he does on this particular topic, I'm confident Tesla (and others pursuing non-LIDAR autonomous driving) are on the right path. In other words, his reasoning withstands critical thinking and it's difficult to deny that, as experts in the field of autonomous driving gain more real-world experience, more and more are coming around to his way of thinking. It's a very new field and Musk has a long history of solving complex problems successfully.
CarlK
Active Member
People who think too far ahead of their time are often misunderstood and ridiculed. When Einstein first proposed the theory of relativity it was said only three people in the world understood it. He was ridiculed by most of the rest. Even a century later there are still experiments and observations trying to prove aspects of the theory are right or wrong (yes they still stand). So were Galileo and Darwin. There are people who still think today the earth is flat and evolution is a haox. Science and technology did advance a lot over the years but clueless people still far outnumber visionaries. Social media in some way helped educate people but it also was doing a lot harm by making it much easier to spread FUD. Ones who have open mind and wanted to learn will be educated. Others can only be drawn in those noises.
Elon may not be at the level of Einstien or Darwin but he's still too far ahead of our time. He has proved his vision in rockets and EV amidst all those noises. AI and autonomy will not be exceptions. Technology advances much faster than it used to be. For self driving technology the moment of truth will soon come if not already has. Just be a little patient and we'll soon see. I'm very optimistic on this.
Last edited:
I think we're just arguing semantics somewhat, but for the fun of it... Passive cameras are still the lowest common denominator. What's less complicated than just using free light already bouncing around. Granted, I'm ignoring your point about headlights vs other active sources, but your point ignores that all the relevant info is being broadcast in human visible spectrum already. I get your point about first principles not necessarily meshing with human factors from a purely theoretical / engineering standpoint, but I think it's entirely misplaced when we're talking about solving a human problem: navigating environments developed by humans, for humans, with other humans all around you.
That is not a bad argument in itself. I’m not quite convinced it is a ”first principles” argument though because nothing suggests, from science, the best way to approach a human world necessarily is to mimick human features — that would be analogous. Just as an example, thousands of years of science has come to the conclusion that long distance travel on land is not at its most efficient by mimicking feet, even though that is how evolution did it with humans, horses, big cats etc. This is why our cars have four wheels and not four feet.
This is basically to say, while employing first principles thinking can certainly lead you to the outcome Tesla took — because, again, first principles is a question not an answer — it could also lead you to several other outcomes just as well.
Ironically enough deep learning itself is based on the biggest analogy of them all.
Sometimes analogies do work.Similar threads
- Article
- Article
- Article
