Tesla vs Mobileye: who will win the Full Self Driving battle?
Vložit
- čas přidán 19. 02. 2021
- How does Mobileye compare to Tesla in terms of full self driving (or autonomous driving) capability? How does the hardware compare? How about the software? How do the two companies differ in their philosophy and their execution? Do they both use deep neural networks, machine learning, Artificial intelligence, big data, and their fleet of millions of vehicles to train their models? Watch and find out!
Tesla vs Mobileye: who will win the Full Self Driving battle?
Check out our Patreon site here: / drknowitallknows . Thanks for your support!
Want some awesome Dr. Know-it-all merch? Check out our awesome new Merch store: drknowitall.itemorder.com/sale
For a limited time, use the code "Knows2021" to get 20% off your entire order!
You can help support this channel with one click! We have an Amazon Affiliate link in several countries. If you click the link for your country, anything you buy from Amazon in the next several hours gives us a small commission, and costs you nothing. Thank you very much for helping support the channel by clicking our link and shopping on Amazon! (If you live in a country not listed here, please let us know and we will add it!)
* USA: amzn.to/39n5mPH
* Germany: amzn.to/2XbdxJi
* United Kingdom: amzn.to/3hGlzTR
* France: amzn.to/2KRAwXh
* Spain: amzn.to/3hJYYFV
**What do we use to shoot our videos? Check this out:
-Sony alpha a7 III: amzn.to/3czV2XJ
--and lens: amzn.to/3aujOqE
-Rode Wireless Go II Lavalier microphones: amzn.to/3eC9jUZ
-Rode NTK Studio Microphone: amzn.to/3vjMomW
-Focusrite Scarlette 2i2 audio interface: amzn.to/3l8vqDu
-Sony MDR-7506 Professional Headphones: amzn.to/2OoDdBd
-Apple iMac 27" computer (with Final Cut Pro and Logic Pro software): amzn.to/38ybTHv
--RAM for Apple iMac: amzn.to/2PRIAJv
-DJI Mavic Mini Drone: amzn.to/2OnHCEw
-GoPro Hero 9 Black action camera: amzn.to/3vgVMrH
Here are a few products we've found really fun and/or useful:
* Our new camera!! The Sony alpha a7 III: amzn.to/3czV2XJ
* and lens: amzn.to/3aujOqE
* The Humane Interface, by Jef Raskin: amzn.to/3ajM7YT
* AI Superpowers, by Kai-Fu Lee: amzn.to/3dgRSs3
* A camping mattress for your Tesla model S/3/X/Y: amzn.to/3chSRrQ
* A great book on Tensorflow: amzn.to/3acUHaw
The Apple Pencil, for use with iPads: amzn.to/3tsuIVz
And of course the amazing "Elon Musk: Tesla, SpaceX and the Quest for a Fantastic Future": amzn.to/36hqdmS
JOWUA:
If you'd like a Jowua foldout mount, check out this link: bit.ly/3tRMSQR
If you use the code DRKNOW5, you get $5 off the price of each mount you order!
Or if you would like to purchase a more powerful wireless iPhone/Android charger for your Tesla Model 3 or Model Y, check this link: bit.ly/3b5fHSW
If you use the coupon code DRKNOW on your order, it will automatically take off the price of the 64GB MicroSD card, which is very nice!
If you are looking to purchase a new Tesla and would like 1,000 free Supercharger miles, just click this link and use our referral link: www.tesla.com/referral/john11286. Thank you!
Music by Zenlee. Check out his amazing music on instagram -@zenlee_music
or CZcams - / @zenlee_music
If you enjoy more outdoorsy, adventure type videos--like climbing the Matterhorn or trekking to Mount Everest Base camp--and a few product reviews too, please check out our sister channel, Whole Nuts and Donuts, right here: / @wholenutsanddonuts5741
Please ask questions in the comments below, or at: DrKnowItAllKnows@gmail.com. Thank you!
Twitter: @DrKnowItAll16
Instagram: @drknowitallknows
This is exactly why competition is a great thing.
I've been hoping for a video like this (Tesla vs Mobileye) for a long time. Thks!!
One very important point that I didn't hear you mention is that all the current Mobileye data is from a single center mount camera.
EyeQ4 is the first chip able to handle multiple surround cameras, so all those billions of miles are monocular vision Driver Assistance, and thus not comparable to the full suite of sensors that Tesla has been running from the beginning.
So I don't see the billions of miles of data that they have being nearly as valuable as Tesla's data, and not really useful for an entirely different system with a bunch of cameras and lidar, radar, etc.
ME has been focused on being a driver assist system all along, and now they have to make the jump to a FSD system with an entirely new processor, entirely new sensor suite, and entirely new focus, and I think it's going to be just as hard for them to make that jump as it will be for Waymo etc to make the jump from riding on the HD mapped "rails" to being able to operate in the wild as truly FSD wherever a person wants to go.
That's the big difference between Tesla and everyone else, they have always been working on the end goal of free range FSD where the car can drive with the same inputs as a human, where everyone else is giving the car a crutch of a preprogrammed route for it to follow, and no ability to truly navigate fully on its own.
ME has been focused on the Driver Assistance, and did a pretty good job, but that's not directly transferable to FSD, as they have made clear with an entirely new sensor suite for FSD, including parallel redundant vision and lidar systems, showing that they realize that their existing system is not adequate.
Tesla in the mean time is working with the full dataset from the full sensor suite, with full focus on giving it free range FSD capabilities, and doing very well with their iterative process, despite the competition throwing inaccurate shade at them.
Stuff like realizing that they need to process 360 video rather than just individual images is stuff you have to learn by doing.
Another benefit Tesla has is full access to their systems.
ME only gets data from some of the manufacturers who use it, and it's limited based on the individual contracts, and they don't have the ability to upload frequent updates to try different things, not that there is much to try since it's only a single camera, and a limited processor.
Tesla has the FSD system working in shadow mode constantly, and so is able to actually test out the system constantly on all the vehicles, and also has the ability to constantly upload new software to try out improvements in real time.
Once ME starts shipping surround camera cars, then we can get an idea of where they are in relation to Tesla, but I don't think that single camera assist has prepared them for FSD.
@@tobygraham6280 no one cares
@@axlcayden4339 Scammers.
@@skylarkesselring6075 you cared enough to comment
Thanks for the comment. I want to know more about this. Is it possible to connect directly? Thanks
Good points. From what was my impression of the CES presentation of Mobileye, they will have too much data to deal with with their multiple redundancy approach (vision. Lidar, radar). Furthermore as Elon Musk pointed out, if vision contradicts radar, which system will you trust? Better to make one system 99.999... % reliable. I found the presentation from ME very interesting but could not get rid of the feeling that it is overloaded and too complicated.
Excellent info. Had never heard of Mobileye.
Thank you. Very useful information.
Is the instruction set standardized between the two processors? If not, is it worth comparing operations per second?
My money is on Tesla, Tesla uses 360 view, where ME uses front view only.
No, their system uses 12 cameras I think
I still like Tesla’s approach better
@@stevesimeonidis5488 actually 8 cameras with 360° degree view, 12 ultrasonic sensors, 1 front facing radar
Great info on self driving technology details.
Lidar makes a necessary redundancy for an added safety. Mobileye makes high definition maps with their fleet and that continually updated and expanded. Who's going to win fsd race?
The company that will prove to the authorities to be safe.
*BINGO!* Not sure that the "mapping" is much of an operational advantage, but Tesla might already be doing this.
Thank you!
Great video,
Mobleye left Tesla because it wanted full control over the data Tesla was generating and wanted Tesla to stop its in-house development of Autopilot.
Musk was very clear that he wanted this capability to be brought in-house, but they were not ready when Mobleye pulled the plug over excuse of an accident.
No auto supplier leaves a big customer over an accident, to suggest otherwise would be naieve.
The EyeQ3 chip was phenomenal for its time but Musk integrated it into AP1 with cavalier disregard towards user safety. For L2 driver attentiveness monitoring, he installed the cheapest & least foolproof method, leaving literally 5 minutes between checks for torque from hand on steering. This choice (opposed by many engineers on the AP team, who preferred IR gaze detection) directly catered to extreme vigilance decrement [plus a nifty sales boost for the cash-strapped company] amongst tech-enthused Tesla owners such as Joshua Brown, whose cognitive dissonance (having paid upfront for the expensive future feature) pushed them to believe they already had a L3 autonomous vehicle and could safely watch Harry Potter movies at 80mph on the interstate. In addition, the radar sensor was never specced to detect stationary obstacles when moving at that speed, so when the system failed the results were spectacular. Decapitations, etc., inevitably ensued. Regardless of any data-sharing arrangement or conflicting future development plans it was therefore no surprise that serious people like Shashua would rapidly distance themselves from these disastrous cowboy methods which risked bringing the whole SDC industry into disrepute. As the wrongful death lawsuits and adverse NTSB reports have continued to accumulate, Musk has progressively reduced the HoW nag time, to now ~15s, indicating that he belatedly recognises at least the risk of liability being attached to himself, if still not actually caring more than a casual toss for his customers' bodily integrity.
@@foobarrel9046 Where do you get this nonsense from.
@@Lee-fc3yf It is common knowledge gleaned by wide reading. Is there any specific point you wish to dispute?
The reason Mobileye stop the contract with Tesla was because they were promising things that werent possible.
@@randomguy-ys1mu Don't think anyone knows the full story but its game on now, Elon wants his direction Mobileye I'm guessing will follow in the end. A car that cannot function in any scenario anywhere is not FSD. If MobileEye can achieve this then all good but bets are on Tesla in terms of approach....Humans dont need HD maps to drive. We just have a ton of edge cases where we make decisions...if AI can do this all good.
Very informative
I thought you would give an opinion at the end. What percentage chance do you think each company has at being the market leader in 5 years?
Note: as Tesla mentioned , since they have access to the cars hardware, they give more weight to the data when the driver takes over. This distinguishes them from the competitions.
Yeah I mean this pretty much applies to everyone that competes with Tesla, any argument against Tesla at the moment is pretty much default invalid.
Additionally, as was mentioned in the Autonomy Day presentation, Tesla avoid data overload by not training on the full set of data from every single car on the road. They look for edge cases that cause problems and then simply 'ask' the fleet for specific training examples.
Another point that wasn't made is that Tesla operates its hardware in 'shadow mode' where the software operates in the background (even when AutoPilot is turned off) in order to compare the car's decisions with the human driver's and any discrepancies (where the human driver's response led to a better outcome) are fed back into the network.
And a further point, this time aimed at MobilEye: not all the cars that use their systems have the same capabilities since they offer a tiered range of products (presumably tailored to a specific manufacturer's requirements and available sensor-suite). So they may well not be collecting data from every car that is as detailed as that which Tesla can collect.
For some reason I have a feeling Mobileye will win.
Well there’s probably a Chance but when Tesla dojo comes in they could take lead
It would be interesting to know how many platforms that host the MobilEye technology actually have the connectivity to return data to MobilEye to be processed and can support regular updates to the platforms so that they can gain improvements that can be used in the real world. I doubt that all of the vehicles that are claimed to be generating learning data for MobilEye actually have real time reasonable bandwidth connectivity. This connectivity is not required for automatic lane keeping and other ‘super’ cruise control purposes...
I estimate MobilEye have installed somewhere around 3..5 million EyeQ4 on cars from a half dozen OEMs in the past 2 years, which they have used to map public roads of EU & US in exquisite detail. They are all kitted with SIMcard for comms. How often they are updated will depend heavily on the OEM, but is surely possible. How much data they return for learning purposes is also subject to agreement but is likely considerable. Thus there is no reason to think Tesla has an advantage over ME in quantity, range or detail of such data.
@@foobarrel9046 Wow. Can someone please post a video of a vehicle controlled by the mobile eye system? It is impressive being able to sell 3.5M cars with self-driving capabilities and I have not seen a single (excluding marketing vehicles) of these vehicles driving themselves. Can someone please post a video I would really like to see these cars that are so far ahead and drive so much better? I should just do better job of looking it seems.
@@jacobuserasmus
Well (my understanding) Mobile Eye and TESLA do different things, TESLA trains AI mobile eye make detailed maps...
In regards to data gathering (probably different data but lets go with miles) Mobile Eye do not have any cars they need to have agreements with the manufacturer the systems are in, not a problem TESLA have.
TESLA gathered 1 billion miles in 6 months in 2020 and that is of course just getting faster with more cars, in april 2020 they had 3 Billion miles so should be at 4-5 billion now.
Mobile eye claims to collect 3.7 million miles/day lets say that is true, still TESLA gather around double that amount.
THE IMPORTANT THING THOUGH is what they do, Mobile Eye create detailed maps, TESLA trains an AI TOTALLY DIFFERENT approaches.
Would also say I find it strange that when talking about data gathering nobody have mentioned Mobile Eye if they gather this much data, then they are clear No 2.
arstechnica.com/cars/2020/01/intels-mobileye-has-a-plan-to-dominate-self-driving-and-it-might-work/
www.forbes.com/sites/greatspeculations/2020/07/03/tesla-king-of-self-driving-cars/?sh=200a1962366b
electrek.co/2020/04/22/tesla-autopilot-data-3-billion-miles/
@@jacobuserasmus
_"Wow. Can someone please post a video of a vehicle controlled by the mobile eye system? It is impressive being able to sell 3.5M cars with self-driving capabilities and I have not seen a single (excluding marketing vehicles) of these vehicles driving themselves."_
Well it is pretty simple.
1. They do not sell cars and even according to them self they had the system in around 1 million cars late 2020
2. They do not have self driving, they have cruise control look at link below, not at all like TESLA.
arstechnica.com/cars/2018/02/the-cadillac-ct6-review-super-cruise-is-a-game-changer/
@Kevin Smith
1. MobilEye sell something like 20M systems p.a. to various automakers, so it's no great stretch to think ~20% of those are EyeQ4 in 2021, namely 4M units.
2. Also where do you get the notion they collect nothing from the vehicles? How do you imagine they have mapped out practically all EU & US public roads on their REM system otherwise?
Thanks
The large stationary object not being detected is not a vision problem, but a radar problem.
13:31 where’s above?? there’s no link
Great video! My two takeaways,
I think the actual draw for Tesla hardware is closer to 80 watts but lets go with 100. So at around 350 watts per mile three and a half hours of driving would cost you 1 mile of range from FSD computer draw. So, is that even a thing?
A second radar would also seem to give redundancy to the large still object on road ahead, without requiring the integration of a whole new kind of sensor suite. But I don't know cost of radar.
so - correct me if I am wrong - it sounds like Mobileye solution doesn't scale anywhere close to the level of Tesla..... which means Tesla is still far ahead in the FSD race
They have already scaled well past Tesla, so no.
@@foobarrel9046 no they haven’t 😂
@@JesusisKing2000 There are more MobileEye L2 ADAS systems on the road than there are Tesla's with AP2. ME supplies systems to heaps of OEMs, so this is no surprise.
@@foobarrel9046 L2 is old news. Mobile eye has 0 beta cars in the hands of customers. Tesla has hundreds at the moment. Tesla will solve FSD then flip a switch to instantly make every Tesla made after 2018 FSD capable.
@@JesusisKing2000 MobilEye has not negligently contributed to the deaths/maiming of dozens of its customers/3rd parties through a reckless business model of putting half-baked & fundamentally unsafe products on the market.
"Solving FSD" cannot mean making it to the grocery store/home from work without intervention, while in the driver's seat watching the road like a hawk during a nerve-wracking trip. It means >=L4, i.e. packing your kids in the car and waving goodbye, with no driver aboard.
Tesla's FSD system will require a new sensor suite of hi-res cameras, 4D radar and HW4 to ever make it out of Beta, i.e. achieve regulatory approval for L4 applications, which IMHO seems impossible before 2025.
Unless all that new hardware is retrofitted for free by the company for those who prepaid for FSD, their HW3 Teslas will remain stuck at L2 forever.
Great work as usual
The thing with tesla is, that they aren't thinking in terms of "will we solve full self driving", but rather as "how can we provide fsd extremely cheap to a broad amount of the population".
Solving FSD is just a matter of time, it's not a question of "if" but rather a "when". The big difference is, that every tesla sold right now, already has everything it needs to use FSD in the future, while the competition has special test cars with funny looking sensors on the top usually.
Well, this is at least one of the more informed videos on the subject, but it still misses out some key points.
Mobileye's system as seen on it's entirely hands-free trips in the hectic streets of Tel Aviv and more recently Munich, is purely camera-based, nothing else. This is the Supervision system they will be releasing later this year and it is way beyond Tesla FSD in the complex driving scenarios it can navigate. That's the big difference between a hands free system and a system like FSD Beta that needs constant hands on and monitoring because it may do "the worst thing at the worst possible time".
Separately, Mobileye are developing with Intel improved radar and lidar technologies which will be unobtrusive and much more affordable. This second system will work separately and alongside their camera-based system, one assisting the other.
Whatever Elon has said, Lidars CAN and DO offer improvement beyond cameras alone. Cameras are degraded in their ability during low light conditions and direct sunlight. Lidar works fine in these scenarios, plus it can often "see" what can't be seen. For instance, a child or vehicles coming out from behind another object. Any benefit, to whatever degree, is still a benefit and in the case of Mobileye's future system, should one system inadvertently fail, the other can immediately take over.
Also, the video fails to take into account the main approach behind both companies. Autopilot FSD is being developed for Tesla's own vehicles, whereas Mobileye's system will be used as the autonomous system for multiple vehicle manufacturers. Therefore, they are not necessarily in direct competition.
Personally, I think both companies systems will do well in the long term but I believe Intel Mobileye will win the race to full self driving.
Tesla will be licensing it's FSD system, so they are a competitor to Mobileye. The winner might the one who get's legislative approval first, and this will likely go to the guy with the most gee-wiz gadgetry, who impresses the law makers. I think that Tesla should deploy lidar for some obvious reasons, least of which is the fact that they must get approval to operate FSD in any jurisdiction and most often....right or wrong.... public opinion matters.
@@eugeniustheodidactus8890 Yes, Elon has mentioned the possibility of licencing out the software, but he says a lot of things. I was referring to the primary focus of the companies.
As for regulatory approval, currently favor is vastly on the side of Mobileye. They are the first company to get approval to test their autonomous vehicles in New York, undisputably the hardest driving environment in the States. Furthermore the NHTSA are launching an investigation into Tesla Autopilot due to the number of crashes into stopped emergency vehicles, another of which happened only days ago. FSD Beta isn't much better in this respect. The software has a propensity to happily drive into stationary objects. Don't believe me, check Car Autonomy channel who have extensively documented the flaws which Tesla's Autopilot team need to address: czcams.com/channels/OO7UoILYqivK7NSunIekhg.html
Lastly, as far as lidar is concerned, yes, I hope Elon will reconsider his stance. The technology is rapidly advancing, so much so that some technologies can even read signs, detect bullets, and see vehicles 1000 meters ago. Of course, Intel with their chip photonic technology will be making its own enhanced lidars and radars specifically for the self driving market.
I believe Tesla's process to get to FSD (appears) to be making the system act on "instinct". Will they succeed? I have no idea , but if they do their cars will become adaptable any place on Earth or for that manner the Moon and Mars. Watching your videos and reading articles about Tesla FSD shows me that. If they succeed all planes, boats and ships will adapt to all conditions they find themselves in. Watching Natgeo how predators adapt to conditions (dolphins). It is awesome.
The most difficult hurdles will be bad whether conditions and intelligent decision making. Just as an example, in winter often the drivable space differs a lot from the situation without snow so cars need to drive based on vision and not on maps. Regarding intelligence there will regularly occur situations where the system has to decide intelligently, as not every situation can be preprogrammed.
Somebody please explain how a car manufacturer tries to sell at scale MobileEye FSD enabled cars that work only in a few specific cities. Thanks in advance
Robotaxi services using ME. Limited to the city and to communities adjacent. Perhaps/probably main highways, allowing city to city driving. It would take a considerable amount of time and funds to map all the towns and rural areas.
But early on car manufactuers using ME could make some money. They would be competing against taxis and Uber/Lyft. It will take Tesla time to scale up. But one demand begins to be saturated Tesla should have a competitive edge.
They already have all of EU mapped and most of the US.
Great vid, thanks! This is the first time someone has seriously suggested that Tesla won't win.
Is this worth a revisit - one year later?
You have to train all your reflexes, when you need it....boom it's there.
- Bruce Lee
whoomp there it is!
-tag team
So, why can't radar see, "hard still object in the road"?
It should, especially metal objects. That is why I am puzzled why it was a problem for the current hardware to detect the semi truck.
Radar struggles to detect non metalized balloons, or cardboard. But it should work well on metal object, and objects with high water content (like humans).
Elon Musks OpenAi needs to be touched on more. Quit making it sound like they just “dump” the data in.
Note that Elon left the board of OpenAi in 2019 and doesn't seem to be very involved with it since it started reducing its openness. Since Elon still co-founded it it's not incorrect to say Musks OpenAi but I'm also not sure it's benefitting Tesla that much. That said, Tesla is capable enough on its own and contrary to what this video lets on they do skip like 99% of the data. They even have a system in place to query the fleet in real time for more data that looks like a certain edge cases where the system performs badly.
@@DanielSchlaug as far as i remember didnt musk actualy leave openAI board so there wouldnt be any conflicts of interest between his role at tesla and openAI
When considering Tesla’s massive data input into their AI algorithm, it reminds me of differentiation as in stepping half way toward a wall over and over again. You never really contact that wall, but at some point you have gotten so close as to arrive within a negligible tolerance. There is one caveat though, it seems to me that the amount of data required, whether brute or selective, is something akin to an inverse relationship with those smaller and smaller half steps. Just wondering if you agree with this concept. Either way, would you be so kind as to explain what you think are the data requirements along this path to win level 5 for Tesla.
Thanks,
Love your channel, and amazed by your “Know it all” depth of knowledge, well organized presentations, and all the while accompanied by a quiet almost negligible ego.
Who has a batter ASIL Rating?
The car in the MobilEye video is driving in Germany. How is that possible w/ regards to european regulations?
Unfortunately the license plates are blurred, so I cannot say which city that is.
I complete agree. Why excluding Lidar if it can help you in some cases and gives you real redundance, if it doesn't cost you much? And Mobileye creates and update their HD maps very rapidly, almost on the fly.
I was deciding what company to invest and you gave me the missing piece of the puzzle. Thank you!
$TSLA :)
Sometimes we can be dazzled by the hardware and software technology without consideration for whether it can ever be realistically utilised. I'm willing to bet that fully autonomous vehicles will never be the panacea they're made out to be for reasons other than technological capability.
I saw an interview with Mobileye CO where he said that their vision system utilizes two chips like Testla but Mobileye calculates route map with cameras and radar and checks the result against Lidar. Is this correct?
I think you are pretty much correct. The "maps" provide an additional accuracy point of reference, for such things as the road moving due to earth quakes.
My understanding was the EV1s were not recalled, but rather all of the EV1's were leased. When the lease ended, they took the cars back and crushed them because they would lose more money providing parts for the required 20 years (or whatever it is) then they would make from selling the 3-year old EVs.
Today it took me more than an hour to watch this video as my mind was getting drifted...the reason? today I testdrove an electric car for the first time...I have an electric 2 wheeler, and I am an avid follower and proponent of EVs since 2015, but the real experience of driving one is really making my mind drifty. It was particularly special because I am from India and there are only 3 proper EVs(Kona, ZS and Tata Nexon if you omit the high end EQC and I-PACE) that are being sold in India since last year, and most of last year was in lockdown. And I am living in my hometown nowadays which is bit backdated when it comes to EVs, so it was months of cajoling and promoting the local dealer to bring an EV. And guess what they not only brought it, but they are becoming an authorized dealer to sell EVs(Tata Dealer) and soon will install a fast charger at their place. All of these is first in my city..I could never imagine this happening at the beginning of 2021, I thought this was still 2 years away. But, constant nagging, phoning people works. So, yeah you might notice bit of of a bump in your average view time.
Little mistake: going from 14nm to 7nm is a 4x reduction of die size, not two 😉. If you divide each face of a square by 2, you have 4 squares of 1/4th size.
That would be true if 14nm and 7nm corresponded to actual gate sizes on the chips. Now it is more like a naming scheme for the process and have little to do with size.
I agree it's a software problem. Reminds me a day when voice recognition was a problem, eventually everything self-driving will become a commodity anyway
This is a good analogy. And with speech recognition, it's been a steady path to improvement rather than a race to the finish line. It gets more reliable in more situations every day. So it turns out there is no finish line but rather it's an ongoing race. Sometimes it's hard to tell who's ahead because the competitors are not all running the same course. Some competitors' tech will be more useful in some situations, and serve different people in different situations more effectively at different times. Tesla, MobileEye, and Waymo will probably all be successful companies for the foreseeable future.
Tesla’s voice recognition and response to its perceived input is truly awful, so I’m not sure your inference about progress in FSD is compelling!
@@Wol747 They have very little computational power dedicated to voice recognition. (I would put a small and cheap 2TOPS NPU into the console for just voice recognition and synthesis). The problem is that all A.I. engineers at Tesla work on FSD, and not the voice interface. This will hopefully change after Dojo is fully functional. It is supposed to lower the burden on the FSD team very significantly.
A super smart voice interface would be more like a gimmick in most cases, but be a nice selling feature, after FSD of course.
Tesla of course only collects data from interventions it doesn't just "dump all the data" into the neural network training, which seems like a much smarter approach to me.
Tesla does (apparently) only upload data when human and software disagree. I think Tesla may be running some FSD functions in shadow mode which would allow very large scale data collection as opposed to taking data only from their own cars and the small number of private owners who are being allowed to try the system. Drivers would be totally unaware of a portion of FSD running in shadow mod.
Tesla also sends out image requests to all (I think) of their post Oct 2016 cars. A few months ago they ask for pictures of tunnels.
@@bobwallace9753 There are over 1000 FSD beta testers right now according to Elon. That's quite a few.
@@L33TNINJA51 Thanks. That's the first number I've heard. I was thinking it might be more like 100 based on the few people posting videos.
I do wonder which features are holding back wider release. I haven't heard if FSD is performing U-turns yet. And I don't know if pothole avoidance has been released.
@@bobwallace9753 I have seen it do you turn once or twice but it was pretty bad at it and it's rare for the route to go in a u-turn fashion. Yeah Elon mentioned the 1,000 people plus number in the last earnings call, iirc. Elon tweeted that there's going to be a gap but that the next FSD beta release will be a step change and so it will be a lot better. He mentioned recently that they still aren't done transitioning all the neural networks to a stitched together video feed from all eight cameras and when that is done it will be amazingly good. So I'm guessing they're working on that right now some more. Elon also thinks it will be scary to release the FSD beta before it's totally done because if it's perfect unlike 99% of drives then people might stop paying attention to the road and end up getting into an accident sometimes. So who knows on full rollout...
@@L33TNINJA51 I'm certainly looking forward to the next major upgrade. I expect it to be good enough to defeat all but nutcase arguments that FSD is our future. And our near future. Then a year until we have permitted driverless cars? That's going to shake our world.
I've been thinking about how robotaxis are likely to impact oil consumption. We sort of operate with the assumption that one EV will replace one ICEV, cause one unit drop in petroleum use. But taxis can drive 5x to 7x more miles per year than a private car. Each robotaxi (or driven EV taxi) will have a much larger impact on oil demand than each private EV sold.
There are three main factors to be ahead in FSD - DATA, DATA and DATA.
Wait, did you get a new camera?
Edit: yup, just saw the text pop up!
$500 x 1 million cars = $500 million So maybe not that much for a single car, but we are speaking about mass production. And even $10 can make a difference.
The redundant chips on the Tesla do run in tandem, just not a 2X performance improvement. They run in a way that improves precision by averaging results aswell as adding error checking when variance is too high. Thats a way to utilize them both to improve performance while maintaining redundancy. Its not ground breaking, but it is neat.
Maybe they will take Qs from AMD with their absurdly fast interconnected chips on the upcoming GPUs. Good things ahead!
So if the FSD chip outputs agree within a certain tolerance they average the two results and perform that action. If variance is outside some limit they self-diagnose which chip failed, switch to emergency mode and perform the other output to navigate to a safe parking spot?
I remember hearing one of the chips can run newer software and run in ghost mode or something like that. So they can test new software and work out bugs before it comes out officially.
Correction:
14:00 -- MobilEye do *not* use LiDAR to build their richly detailed semantic map. It uses the EyeQ4 on-board vision cameras to triangulate distances to landmarks & objects of interest, thus continually locating the vehicle to within ~5cm CEP on the map. Details are then updated by the first car to encounter a significant change, e.g. roadworks, and broadcast this to other vehicles in the area. The whole system is called REM, Road Experience Management.
thats idiotic, thats the definition of brute force...
@@krellin I think you mean the definition of freely crowd-sourcing vital data while being paid for millions of the ADAS product per annum. Like where do you think Musk got his initial AP ideas from? ME showed him in 2014 what they had in the pipeline for the next 10 years, and FSD is his attempt to emulate & leapfrog all their work on the cheap, recklessly neglecting user safety in the process, leading to a series of infamous fatalities/maimings on Tesla's record, such that it has been ironically nicknamed *Firetruck Super-Destruction* mode instead.
@@foobarrel9046 Are you aware of the fact that Tesla has quite a different approach to Mobileye. Calm down and stop making silly false accusations.
@@nilsfrederking62 Those are not accusations, just a neutral recounting of known facts plus a little logical deduction. Which part do you claim is false?
@@foobarrel9046 What about you delivering sources for your false claims.
Could you do an update on Mobileye vs Tesla? They are public now and they seem to make insane progress…
16:02 Well if Tesla gets FSD rght they have the whole Tesla fleet but, if Mobileye gets their full selfdriving right they have the whole automotive industry.
It’s a game of data. Whoever has the most amount of data and the most variety of data will win by a huge margin. Automakers won’t want a “beta” working system that gets updated, they want it to work now or they’ll just wait. Which doesn’t really work. All Teslas built from oct 2016 have all the hardware as newer cars so they have a lot of cars collecting data!
@@harsimranbansal5355 right but if it's a tie Mobileye will have the whole automotive market. The reason I say that is no traditional auto manufacture will want to use Tesla's FSD over Mobileye. How i came to this conclusion is that traditional automakers would already be using Tesla's charging infrastructure but they don't. They do however use Mobileye.
Tesla has FSD hardware in its cars, so a software upgrade can "potentially" makes its fleet FSd capable.
However, Mobleye's current solution, which is actually in cars, can't support FSD. If Mobleye achieves the tech, then its "automotive" partners would have to build the hardware in their new cars.
Also, Tesla is open to licensing its tech, so other automakers can also use Tesla's technology.
Not true.
There are still other players like GM's super cruise, Waymo, Zoox etc.
Plus the Europeans making their own systems.
@@neeljavia2965 Right and they may get a piece of the pie but, we were only talking about Tesla and Mobileye if they succeed hypothetically.
Surely the 'large objective in road' problem,if not properly recognised by the cameras, is picked up by the radar?
Nice of you to have faith that the Lucid Air is going to actually show up one day. ME won't get much data from that 'fleet'.
LOL
Having worked for city transportation before... if you think that dealing with multiple city halls is in any way effective... you’ve got another thing coming.
how about adding a comparison table or graph? otherwise, it's just prattling a bunch of numbers
Interesting. How does Mobileye's algorithms work with different cars with different sensor suites? How can it be a plug-and-play solution?
All sensors will be provided by mobile eye.
6:40,
or is it 300 watts? Very doubtful.
It's on the data sheet, yes.
But there is no 2.5 TerraOps chip under 100 watts.
Isn’t IQ cheap connected to single sensor while Tesla chip is connected to multiple sensors?
No mention of Teslas Dojo?
"Front lidar for security matter : long object across the road" ... does not Radar do the same ?
@Steven Gulie Radar can penetrate weather conditions that block lidar. If lidar can detect something, so can cameras. Other than very precise positioning what is the advantage of lidar? Clearly humans don't need very precise positioning at the lidar level to drive and it's clear that Teslas are doing a fine job driving without lidar.
As for radar and moving objects. Isn't the fact that the car's movement creates a change in returned radar signal? If the car is approaching an overpass the return will move higher in the "screen". A stopped truck in the road will return a signal close to the road.
@Steven Gulie "humans have hands and require years of “training” to develop an accurate 3D model of the world from visual input"
I think what a lot of people don't fully recognize is that each human has to boot up from zero and develop their accurate model but self-driving EVs (and all other intelligent machines) are able to combine their experiences and compress years of learning into a common 'brain'. And then each successive new EV springs from the womb as fully intelligent as the best of its species.
We're ratcheting knowledge. Intelligent machines are on track to be far, far smarter than people. None of us will live long enough to pack in enough experience. Unless, I guess, we wire ourselves to the great digital brain.
This evolution is likely to sneak up on most people. It's really going to disrupt us.
Tesla seems to have a huge advantage hardware wise, but I wonder how hard it would be to simply put 3 mobileye processors into a car. Then you've got a double-redundant system that's still running only 30W. And with triple processors, if one goes bad it can be easily detected because it's out of sync with the other 2.
16:05 Skynet becomes self aware
One thing that Tesla needs to do is what GM is doing in Super Cruise, watch what the driver is doing instead of requiring a hand on the steering wheel. With the inward facing camera they might be able to do this, but they also might need to add an infrared camera to see through sunglasses, or to generate a 3-D map of the driver’s face to see where his attention is. Just because the driver is touching the wheel doesn’t mean he is watching the road.
Mobileye is the one competitor to Tesla that I take seriously because they have the potential to get their system out to many different automakers around the world such as Nio and therefore will have the data that is needed to keep pace with Tesla.
I heard NIO rectory decided to go with NVidia. If true, big mistake.
You say that Tesla dumps all the data back to base for training purposes thus overloading and being inefficient. However in every presentation since Autonomous day they say they only request specific cases, such as cut ins, to avoid inefficiency. Are they lying?
Who says tesla isn't testing on specific data, they don't have to always be testing on "everything" data
I am not sure where you got the power requirements for the next FSD chip. Goinf from 14 to 5 nm will definitely reduce power requirement a lot. I think their motivation to reduce power will outweigh the need to increase TOPs.
Agree 100%. Also you can tell why Mobileye just draw that low power by looking at their motherboard. Less power processing and hardware = less power.
Lucid should put 200 sensors on their car because more=better right?
What about XPeng XPilot ?
They stole a now outdated software......
@@nilsfrederking62 somehow XPeng Navigate guided pilot and self parking is better than Tesla navigate on autopilot and self parking!
Don’t trust me look it up XPeng NGP
I don't think Mobileye traditionally used radar as you stated at 4:50
“Mobil Eye have more equipped cars on the road” How?
Thanks! Excellent again. For a future video, I'd like to know the specs of their radar. Beam width, ping frequency, power, penetration, etc. And does it really work? Can it see through rain and fog? My Autopilot aborts as soon as "Bad weather detected" when I really would need it the most.
And one more thing: Why doesn't Tesla have an infrared beam and camera to quicker detect people and animals on the road in the night?
I think I've heard that Tesla is going to add infrared capability. And move to a radar with shorter wavelengths that will give more accuracy.
A regular digital camera can be converted to infrared by removing the Bayer filter, IIRC.
Tesla cameras can see better than humans at night already.
@@L33TNINJA51 Do you know of any testing that has been reported? I've seen a report or two of a Tesla FSD system detecting a person/animal in the dark before the human driver did but no objective studies.
I've no doubt that digital sensors can detect better than humans in dim light. If anyone doubts that they should look at this video.
Sony video ISO 409,600 demonstrating digital sensor ability to “see in the dark”. Best estimates are that human eyes have an ISO of about 800 when fully dark adapted. A driver's eyes would not be fully dark adapted due to light from headlights and instruments/panels.
czcams.com/video/CVLBHMbRMW4/video.html
@@bobwallace9753 Not sure what you mean by objective studies. Just people doing tests with FSD beta and saying the car saw the person before they could. czcams.com/video/YxxYyetcxXk/video.html
@@L33TNINJA51 I suspect it wouldn't be hard to set up a little test with a couple of people in the car and a couple 'targets'. Make a few runs with one person 'driving' and one watching the screen to see when the car spots a person.
Problem with MobilEye is how many of those cars are actually uploading data? Not many.
The problem with Tesla is uploading too much data. If the data comes from Italy, France, Mexico and (God forbids) India your Tesla will be an absolute danger on the road. If the data comes only from the US your Tesla will drive like a grandma which is also dangerous.
3 years later and fsd looks to be ahead
Not a founder....actually he was. This was conclusively proven in courts after extensive analysis.
FYI....Before Musk, Tesla has 3 people (original founders) for 1 year, they did not even have an office, leave aside any prototype. Musk funded Series A, and 90% of Series B & C.
The word Founder is important, and in silicon valley, you only get this if you have had a significant impact.
Why is anyone's title important?
@@dougdevine27 It's the credit and honor. Wars have been fought over it. Many humans, especially those with high achievement, consider it very important.
Tesla was founded in July 2003. Musk didn't join until Feb 2004. Musk came in as an investor, not a founder.
@@danchatka8613 the courts disagree lol
The concept was founded, but Elon was there to start the actual company.
@@danchatka8613 czcams.com/video/J9oEc0wCQDE/video.html
At the end of the day the winner will be the one that is most aceppted around the world. Mobileye may have an advantage in some places that require lidar to be implemented. I believe both will do very well but I have my money on Mobileye (Nio)
Not much content covering Mobileye. Thanks.
And I don't know why Ark invest didn't mention Mobileye very often in stead of Waymo. Waymo's methodology and data size is not a threat to Tesla but Mobileye seems to be.
I would say mobeleye would win cuz they use cameras, lidar and radar while tesla only uses cameras.
Glad to see competition. That will hopefully speed things up
while Lidar might be better in identifying there is an obstruction on the road Tesla still uses also the radar that can resolve most of the issues Lidar "solves" and knowing Alon if tomorrow Lidar is sufficiently cheep and can solve for Tesla any issue it has than it might creep into the Tesla self driving system. also as you said Mobil eye is basing it's expansion on city by city that will always leave out areas of rest of the country that will not be mapped and cities have un expected road works and obstacles
Few comments on the wattages of processors and hardware. Anything less than 250 W seems to be great for range in the EVs:
250 W system:
At 120 km/h a 250 W system will consume 2.08 Wh/km equals about 5 km range loss on a full 75 kWh battery.
At 90 km/h a 250 W system will consume 2.77 Wh/km equals about 10 km range loss on a full 75 kWh battery.
At 50 km/h a 250 W system will consume 5.00 Wh/km equals about 20 km range loss on a full 75 kWh battery.
100 W system:
At 120 km/h a 100 W system will consume 0.83 Wh/km equals about 1.5 km range loss on a full 75 kWh battery.
At 90 km/h a 100 W system will consume 1.11 Wh/km equals about 4 km range loss on a full 75 kWh battery.
At 50 km/h a 100 W system will consume 2.00 Wh/km equals about 9 km range loss on a full 75 kWh battery.
10 W system:
At 120 km/h a 10 W system will consume 0.08 Wh/km equals about 0.15 km range loss on a full 75 kWh battery.
At 90 km/h a 10 W system will consume 0.11 Wh/km equals about 0.4 km range loss on a full 75 kWh battery.
At 50 km/h a 10 W system will consume 0.20 Wh/km equals about 0.9 km range loss on a full 75 kWh battery.
So the power usage is mostly important in the city and when idling, here low power will make a great impact on cost and range. As soon as you go on the highway it will matter much less. Anything above 250 W for the hardware will start impacting the range a great deal.
Also note that both EPA and WLTP test cycles are performed at rather low speed. With poor hardware systems your range will be impacted greatly on both official test. This is part of the reason why Mustand Mach-e, Audi E-tron and Taycan scores so low in EPA/WLTP but still makes for decent highway cruisers.
I'm really looking forward to FULL autonomy, meaning, no driver controls (steering, pedals, etc.) in the vehicle at ALL, no licensed driver required. That changes safety, insurance, and removes many other driving hassles. Between now and then (perhaps 5 - 7 years worst case), the systems will improve, but will NOT be (really) Full Self Driving. My CyberTruck order should be ready in 3 years, and I'll need to analyze closely whether or not it will be able to truly achieve FULL autonomy before I make the final contract to buy.
Will their Never be Full Self driving and robotaxis when will they come then?
Level 5 autonomy is a fantasy - and even if it weren't, Tesla's current hardware would not come anywhere near reaching it.
@@trainiax Based on what evidence?
driver monitoring system creeps me tf out
Yeah but Dojo is an A.I. so there is that!
You keep saying power consumption matters, but they are not. When an EV is running, they use many kilo watts of power, so 100watt is just as irrelevant as where the wind is blowing
Always a pleasure to absorb a video from somebody who knows the field. CZcams Gem-Quality content.
On another note, in 2019 I carefully researched a new camera purchase, eventually deciding that the Sony mirrorless cameras were the way to go. Settled on the A7iii and have been consistently blown away by how the camera performs...even now two years later. Tremendous package!
DOJO...
one more
Mobileye to shake hands with TESLA hmm that be great deal
you think yeah right on me too
The first company to reach effectively full self driving will be the company that first starts relying on what human vision relies on for percieving the world. Which is inferring depth information from different points of view. And you can see that nobody is doing it at the moment simply by looking at their camera configurations.
Mobileye is effectively driving cars like trains. They build their rail network and their cars cannot leave the rails. Tesla has an overall better approach but they also have the problem of overusing neural networks to solve trivial problems that can be solved more easily and accurately with simple geometry. The main problem with overrelying on neural networks is their inherent feature to occasionaly give random meaningless outputs that can cause serious problems if there is no post-processing on the NN output. Tesla clearly doesn't do any significant such post-processing which is evident by the spontaneous outputs that their autopilot sometimes gives.
The "big stationary object on the road" problem is trivial to solve if you approach self triving as a general problem of getting from point A to point B without hitting stuff, which is a very old and solved robotics problem. You simply need correct 3D information of your surroundings which you can get either with LIDAR and cameras (LIDAR alone makes it hard to identify moving objects) or with photogrametry just with cameras. The drivng-specific stuff shouldn't be a priority, but every ADAS developing company starts with that. They first develop lane marking and car detection and only later or never they develop actual 3D perception. It is so ridiculous... If you have a system that can see a thing on the road and avoid it you've instantly solved the problem of avoiding cars, humans, trucks that have fallen over, and every other object that could be on the road all at the same time. But if you decide to train NNs to identify specific objects and only avoid those you can never make a safe car because some objects that will be on the road in the future would have never been on the road in the past so it is theoretically impossible to train a NN to identify something that doesn't exist yet.
7nm is not twice as dense as 14nm. More like 4 times as dense.
Tesla focuses very heavily on edge case data and skips the bulk of what's available to them. They have a system in place to ask the fleet to look for more data that looks like certain edge cases where the system performs badly and send only that back to Tesla. Amnon Shashua should know this (at least before he gets on stage to say the opposite) so he seems to be engaging in foul play. Since Musk has hinted that Tesla will also make FSD available to other car makers as a service Mobileye must be feeling the heat.
Where did you hear that Hardware4 would be drop in replacement for hardware 3? Did you mean retrofit for old cars or "replacement" in terms of new cars? I ask because Musk said hardware 4 would be matched to high definition cameras and rumors indicate Tesla will be using new high definition 4-d radar as well. It seems like the 3x improvement of hw4 would go to increase processing needed for new high-def sensors and not needed for HW3 cars --not to mention additional cost to replace hw3 computer. I don't think HW3 cars will get HW4 computer. They might keep it backward compatible so they can stop making HW3 for repairs but I'd think they'd prefer to redesign if they can improve.
Hi Know Could please talk about China production and political land scape for Tesla.
There’s room for several players though I think MobilEye’s solution is superior to NVidia’s, both SW and HW-wise. Soon most ICEV OEMS will realize they have no choice but to join one of the two camps. Tesla is in charge of their own destiny, the only player at the moment in the driver’s seat.
Did not watch this video or anything on this channel yet but already wanna say THANK YOU! I am tired of all the intellectually lazy Tesla Bulls talking about Tesla FSD without even mentioning Mobileye - they just have no clue about the market and keep talking about Waymo. And no, I am not a hater (I have lots of Tesla stock) and don't even know what this video will say about Mobileye.
Because mobileye is not much famous like google
@@neeljavia2965 That's not the point! They have different approaches, Waymo is not relevant. But Mobileye is VERY relevant competition to Tesla. Most Tesla Bulls don't know enough and follow random CZcams guys who also don't know enough to FULLY understand the market and all the dynamics.
@@JakobHelm Granted.
But waymo is still a decent one.
420… chill very chill
thanks for covering mobileye. i've been looking at them for ages now and i do believe that they have a very interesting approach to "human nature" in their data, something i'm not sure how is implemented within tesla.
Tesla actually has more human nature type ai than mobileye.
@@neeljavia2965 I love Tesla. 8 would like to agree with you. Mobileye was better at negotiating with pedestrians and cars in Israel back in 2019. Something that was amazing then and Tesla still cannot do a month ago, see AIDRVR clips.
@@AlexandruJalea Than why haven't mobileye expanded everywhere yet?
Doing good in one city doesn't prove anything.
Even waymo can do it.
@@neeljavia2965 please do your research.
@@neeljavia2965 czcams.com/video/A1qNdHPyHu4/video.html
Just because MobileEye has their chip in all of those cars does not mean the car will be able to do anything near actual FSD.
I think you’re off base on a few points. MobilEye’s ADAS chips are in many cars but they are not being used to collect data. Tesla already has a million strong fleet collecting real data. MobileEye is still very limited in data collecting capability and mostly limited to Tel Avi. Finally LiDAR is not being used to create detailed maps. The cameras are. However Mobile eye has the best merchant solution and my bet they will beat out NVidia but they are not a threat to Tesla in any way.
HD maps make no sense. Seems like an unsustainable short term solution. Humans don’t need HD maps to drive.