-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Tesla Radar Speculation
- Thread starter Takaa
- Start date
Personally, I think that people are making WAY too much of TESLA’s decision to remove USS. IMO, the ONLY time that matters is when the car is stopped in motion, and even then, the forward camera/s can capture from its FOV when someone/something is entering that blind spot from the LEFT or the RIGHT—or from up ahead. While the FSD TESLA is in motion, it can clearly see what is entering that blind spot from its forward FOV. As the FSD TESLA moves forward, it can see objects and people that were 20ft ahead and predict that—even as those people/objects are no longer visible—they must be in that blind spot. FSD already does this by predicting that a vehicle or a person moving behind another object (such as another large/r vehicle) is actually still there—especially with pedestrians. This has been thoroughly documented during TESLA’s AI Day 2021 (last year). How would predicting what enters into the FWD blind spot be any different? Sure, if you back into your garage, and then you later attempt to drive away from your garage not knowing that your cat, dog, or 3yr old niece might happen to be lying horizontally on the garage floor across the front fascia of your TESLA, then I guess in that rare or highly unlikely scenario, your cat, dog, or niece would be in danger.
Have you seen the Andrej Karpathy interview on Lex Fridman’s podcast????
Removal of USS:
Benefits of cameras to FSD:
Andrej not only ran TESLA’s FSD AI/DOJO team for a few years, he is also a VERY talented developer and coder and stated that he wants to “get back in the lab”, so to speak. He also indicated that more data from more input sources (cameras + radar + USS + LiDAR) is NOT conducive to long-term viability and success. He actually described all of that as noise. Even implementing higher resolution cameras, he said, was not necessarily and does not lend itself toward increased success of FSD long-term. He said the existing cameras all TESLA’s are equipped with are little better than human vision.
Remember, FSD doesn’t need to be perfect. It merely needs to become better, safer, and more reliable than human drivers, which is very achievable. Personally, I am VERY encouraged when I watch the FSD Beta videos published by several different YouTubers. It is remarkable how far the FSD system has come! One thing that FSD Beta continues to struggle with is road construction. I think it may become necessary for stricter standards to be adopted by road construction crews in the placement of cones and temporary signage. Perhaps, in the future, road construction crews will have to submit a construction plan and schedule that clearly illustrates using schematics where cones and signage will be placed. Candidly, even as a human driver, I sometimes struggle with figuring out where I am supposed to drive and not supposed to drive in construction zones—especially at night.
Last edited:
Thanks for sharing the links to these videos!It looks like an abstract painting made by a vision system
I watched both.
arbe boasts higher resolution radar and redundancy in its Lynx and Phoenix systems.
Andrej Karpathy, former head of TESLA AI/DOJO and FSD, said during his interview on Lex Fridman’s podcast that higher resolution (cameras—not radar) was not needed and regards all redundancy as noise and unnecessary. I’m sure you are aware that TESLA scrapped its radar some time ago, as well as recently scrapping USS. Andrej wasn’t merely a manager of techies. He is a brilliant coder and developer and has an excellent understanding of neural networks and machine learning AI.
The proof is in the pudding. You point me to another mobility solution that is more autonomous right now than TESLA, and let’s have a look. Please, please don’t even think about bringing up Waymo! Much easier to geo-fence much of Phoenix than it is to build a fully autonomous system capable of driving anywhere on the planet. Not even a close comparison at all!
Thank you for pointing out a general problem with Tesla. They are approaching FSD as 90% problem, when, in reality, it is 99.999% problem. Having your nephew across the floor behind the car is a very, very realistic scenario that other cars would accommodate for. BTW, almost all other cars in this price range have cross traffic monitoring.
Andrej interview is interesting. He makes some postulates that are questionable. If various sensors increase noise then why aircraft have so many? Downward pointing cameras could measure altitude, why do we need Pitot tubes and radar on top? The aircraft industry solved the problem with “noisy” input 50years ago and it works pretty well. In fact, humans not only have immensely more intelligence but also rely on multiple sensors when driving. There is a reason why it is illegal in most states to cover both ears.
Don’t get me wrong - Andrej may be right and Tesla is coming up with this amazing TV, FSD, etc. (Although, I think they are underestimating the problem). But at this point it is not proven technology. There are early adopters out there. I thank them and encourage them to use the FSDb - we need people on the cutting edge. However, majority of the customers do not care about that; for them the car is a tool and it has to be reliable. Which means behind the cutting edge, proven, albeit not very sexy, technology.
The main issue is that Tesla mixes already proven technology (the amazing EV drivetrain) with cutting edge development technology (the AI). Every seasoned product manager will tell you that not isolating them is heading for a disaster. Those are two very, very different market segments.
Andrej interview is interesting. He makes some postulates that are questionable. If various sensors increase noise then why aircraft have so many? Downward pointing cameras could measure altitude, why do we need Pitot tubes and radar on top? The aircraft industry solved the problem with “noisy” input 50years ago and it works pretty well. In fact, humans not only have immensely more intelligence but also rely on multiple sensors when driving. There is a reason why it is illegal in most states to cover both ears.
Don’t get me wrong - Andrej may be right and Tesla is coming up with this amazing TV, FSD, etc. (Although, I think they are underestimating the problem). But at this point it is not proven technology. There are early adopters out there. I thank them and encourage them to use the FSDb - we need people on the cutting edge. However, majority of the customers do not care about that; for them the car is a tool and it has to be reliable. Which means behind the cutting edge, proven, albeit not very sexy, technology.
The main issue is that Tesla mixes already proven technology (the amazing EV drivetrain) with cutting edge development technology (the AI). Every seasoned product manager will tell you that not isolating them is heading for a disaster. Those are two very, very different market segments.
Read my post once again. I am not disputing Tesla disruptive power. Just the opposite - it is something to celebrate because we would still be driving ICEs without them.
My concern is mixing unproven with proven tech in the same product. I bet you that the vast majority of customers buy Tesla because it is EV, not because it has AI. Would you buy an ICE Tesla because of the AI? Would you buy EV Tesla without the AI? Those are two very different things appealing to very different segments.
Let them experiment as much as they could with AI but don’t let unproven tech spoil the main reason people are buying the car. It will backfire.
BTW, there is a big question how advance their AI is. They are the most vocal but there are other “driver assistance” tech that is more reliable. Self driving that is equivalent to human driving is a super hard problem to solve (don’t give me Tesla stats - they have winner’s bias). The aircraft manufacturers are working on this for the past 70 years, it is a much simpler problem in the air, and still the aircraft are not fully autonomous. But I digress…
I am giving you that example not because I want them to stop working on that. No, let them push the boundary but keep it separate from the EV.
and that is why FSD is sold separately. It is not forced upon anybody.Appreciate your insights!
One clarification: When Andrej refers to more sensors as noise, I think he means that in the context of machine learning and AI powered by neural nets. I can see your point about sensors in aeronautics with human operators benefitting from that, but I think Andrej’s point is that the neural nets don’t benefit from sensors the way human operators do.
Yes, it is forced. By removing the radar, by using cameras for auto wipers and auto beams… Now, we expect the removal of USS. All of that is connected to the unproven and experimental Tesla Vision. Not only it is unproven tech but it is also forced on the existing fleet. Frankly, if I had known that I would not have bought the car. They reduce functionality _after_ they sold the product. Also called “bait and switch”.
Boza, sell your TESLA and buy a Polestar, for GOD‘a sake. The world will be a better place w/o retaining you as a TESLA customer. Meanwhile, the rest of us will happily keep purchasing TESLA EV’s.
I was referring to aircraft AP where the functionality is delivered by old school computers, not neural nets.
But you are bringing an interesting point. NN is modeled after nature and animals do use multiple sensors. We have either 5 or 6, depending of who you are asking
. And we do have a problem with conflicting data (e.g. vertigo). However, it seems that benefits outweigh the issues with multiple sensors.I don’t have a problem with Tesla EV; I have a problem with Tesla AI. Unfortunately, there are no good EV alternatives yet.Boza, sell your TESLA and buy a Polestar, for GOD‘a sake. The world will be a better place w/o retaining you as a TESLA customer. Meanwhile, the rest of us will happily keep purchasing TESLA EV’s.
All valid points. You commented that TESLA might be going in the wrong direction with its approach to FSD. Honest question: Do you know of another autonomous tech or another firm that is doing mobile autonomy better than TESLA right now??? If so, then who???I was referring to aircraft AP where the functionality is delivered by old school computers, not neural nets.
But you are bringing an interesting point. NN is modeled after nature and animals do use multiple sensors. We have either 5 or 6, depending of who you are asking
(I’d like to point out that DOJO is the same engine powering both FSD Beta AND their Optimus humanoid robot right out of the gate!)
While I understand the "who else right now" argument, I mean its great... sadly, Tesla will be losing so many frustrated Tesla owners in the next 2~3 years with EV cars/trucks that were designed better from ground up with better safety feautures than current Tesla TV/AP. Not BETA waivered feautures or gimmicks. Veoneer+Arbe sensor suites will be sold in most of those EV cars/trucks (Mercedes-Benz, BMW, Audi, Ford, Dodge etc etc).
I've got a Model Y LR 7 Seater. Had radar, currently still has USS... and I can afford to wait for better safety features with every dollar I'm not spending on gas with a quirky EV with questionable safety features that keep getting disabled or nerfed.
I test drove the MB EV and their “driver assistance” (more accurate description than FSD) is pretty good. At least Tesla level on the highway.
But who has better AI is a digression. Majority of customers do not care - they want something that works and it is reliable. At the moment, Tesla camera-only system is less reliable than dedicated sensors (wipers sensor, USS). When they reach parity and prove it then they release it to the masses. It is common, well established practice in systems development. By mixing beta AI with mature EV they mess up the EV experience for the majority of customers.
Yes, they cannot sell them fast enough because they ride the wave of implicit expectations - people expect the same level of reliability as the established brands have. But a few more crazy things like radar and USS and people will wise up. You can already see quite a few people debating whether to accept a non-USS car (in a way, they are cornered, which creates bad taste from the beginning of the relationship). I would not be surprised that, at least until/if Tesla gets their act together, non-USS cars will have a lower residual value than the USS cars. I already saw on a few forums people asking for the specific firmware of the car - the one before the radar was disabled.
Similar threads
