Watch Tesla’s Self-Driving Car Learn In a Simulation! 🚘
ฝัง
- เผยแพร่เมื่อ 27 ก.ค. 2024
- ❤️ Check out Fully Connected by Weights & Biases: wandb.me/papers
🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible:
Aleksandr Mashrabov, Alex Haro, Andrew Melnychuk, Angelos Evripiotis, Benji Rabhan, Bryan Learn, Christian Ahlin, Eric Haddad, Eric Martel, Gordon Child, Ivo Galic, Jace O'Brien, Javier Bustamante, John Le, Jonas, Kenneth Davis, Klaus Busse, Lorin Atzberger, Lukas Biewald, Matthew Allen Fisher, Mark Oates, Michael Albrecht, Nikhil Velpanur, Owen Campbell-Moore, Owen Skarpness, Ramsey Elbasheer, Steef, Taras Bobrovytsky, Thomas Krcmar, Timothy Sum Hon Mun, Torsten Reil, Tybie Fitzhugh, Ueli Gallizzi.
If you wish to appear here or pick up other perks, click here: / twominutepapers
Thumbnail background design: Felícia Zsolnai-Fehér - felicia.hu
Károly Zsolnai-Fehér's links:
Instagram: / twominutepapers
Twitter: / twominutepapers
Web: cg.tuwien.ac.at/~zsolnai/
#Tesla #TeslaAIDay - วิทยาศาสตร์และเทคโนโลยี
"We are doing it paper-style..."
My man!
"holy mother of papers"
I love to see all these familiar papers all being used in one place.
how did you comment 20 hours ago?!
@@proloycodes it might be patreon I dont know
They are correct. Patreon
Power rangers moment
@@Wmann many use both
Can we just take a moment to appreciate how incredible their simulated Environments look? 🤯
Yea is awesome :)
I am starting to think we are the simulated people in one of these simulated environments, lol.
Didn't expect to see you here elca how is your avatar game going
To be fair, the more real the look and the more real the actions of the people / animals / vehicles represented within them, the more effective they are at allowing the AI to learn things that are applicable to the real world. If all the vehicles looked like they were out of a 1980's racing game - even if they were obeying physics that were as close to reality as possible - the AI might learn things that would be inapplicable to the real world - or worse, learn things that would cause it to take incorrect actions in the real world.
This is a rendering, not their simulation they use for AI training
Thirteen Minute Papers - but I'm not complaining! I could watch and listen to Dr. Zsolnai-Fehér for hours :)
We need Sixty Minute Papers
13 min papers with 2 min segments 😜🙃
It's 2 minutes in Valve time.
Wow, as a scientist in training it's amazing to see all of these papers getting real world use so fast
It's already 7 years of development.
Absolutely incredible technology. Thank you for doing a paper-style review of the talk. Truly a crazy time to be alive!
This is one of the best technical AI videos for dummies that I've ever watched in my life and I'm so glad I stumbled upon your TH-cam channel. Thank you Doc.
You are very kind, thank you so much! 🙏
It was actually awesome to see Tesla do this whole thing with the simulation, trying to make it realistic because i immediately remembered your video about that paper that you mention in this video as well. (where they create a "real world looking world" from GTA)
in the full AI day presentation you hear the Simulation Expert comment about the GTA 5 technique too and how theirs is an improvement in many ways. :)
@@gr1f1th Yeah i remember.
I don't know why people are so surprised, Google has been doing similar things for over 5 years, maybe closer to 10. What am I missing ?
@@autohmae The results are different?
@@PresidentialWinner As far as I know Google and Tesla have one big difference: Google uses Lidar and vision and Tesla only uses vision. And Tesla started this project later when we have more compute power per unit price and research is much further along for machine learning. I'm just surprised people are reacting like Tesla is doing something special. Or maybe it's just Google never really explained what they were/are doing to a more general audience ?
How cool that the simulation is also an interface for creating or adapting different kinds of roads, environments, traffic regulations. It has been easier to add traffic rules of other countries in this way. It looks like FSD will be available everywhere sooner than we think.
It also keeps the robots off our backs because they can just chill in the virtual world
i hope in 5-10 years wenn we are lucky
I'm here to squeeze ALL of my papers!
I squeezed too hard during this video. damn this is crazy.
I think this video made me believe in self driving cars again. Sure it’s not there yet, and autopilot is still cost prohibitive, but this makes it seem possible that these cars will be actually be fully autonomous in a few years. Absolutely mind blowing.
Seeing FSD Beta drives is really quite impressive. Especially seeing how the car senses its environment so well. It just doesn't know yet how to act right in many situations.
Beta 10 is out today I think and promises much improvement, can't wait to see videos of it.
With dojo and the accelerated training it really could be there soon. Definitely not a question of if anymore but when.
People were saying this exact same thing 10 years ago (if not even earlier). "Full self-driving" (or level 5 autonomous self-driving, without need for direct human involvement) probably isn't possible using machine learning. The companies working on it have an extremely obvious incentive to be unrealistically optimistic - they rely on investor confidence. This is especially the case for Tesla, which lives or dies based on its stock price. Most people who actually work in machine learning/AI and don't have a personal incentive to sell the idea of self-driving will agree on this. It's a useful technology that is already used in very helpful safety features, but it's not suited for extremely high degree of consistent accuracy that is necessary for actually fully controlling automobiles - without human supervision - in the full variety of situations a car might normally encounter. It can probably be reliably used for travel along relatively simple set routes (like shipping on expressways), but regular people won't be using it any time in the remotely near future. I would be willing to bet literally any sum of money that a significant portion of people won't be using level 5 self-driving vehicles within the next 5 years.
Also, a lot of the stuff Tesla is accomplishing here is just attempting to recreate functionality that is already possible with LIDAR lol (and there are a ton of inherent issues with cameras that aren't reasonably solvable, like the fact that there are many unavoidable environmental circumstances that can obscure their imaging). They're forced to be clever because of a stupid decision. If actual level 5 autonomous driving is ever achieved, I can guarantee you it won't be Tesla that does it.
It should also be mentioned that it's grossly unethical to have customers test such a technology on public roads, to an extent that is honestly hard to believe. Literally having untrained people testing technology that controls these multi-ton chunks of metal amidst countless other drivers and pedestrians. Actual reputable companies working on self-driving don't do this. They do actual proper testing. What Tesla does with its "FSD beta testers" is not how software testing is supposed to work - testers need to be in contact with developers to actually accomplish anything meaningful. And it's significantly worse when it's also endangering people. Everyone involved with this, including Musk, should probably be in prison.
Not sure if a system that can be 'fooled' into a 3d interpretation of 2d images truly inspires confidence (visions of wil e coyote painting a fake tunnel)
Tesla's entire AI presentation is well worth everyones time. The D1 chip and Dojo computer are nuts!
7:28
Car 1: "Hey I can't see anything. Can anyone help me please?"
Car 2: "Oh yeah, sure. I've encountered this problem before, I can tell you what to do."
Car 1: "Thanks, dude. I owe you one."
Car 2: "No problem. Hey, by the way, how's the robot uprising plans going?"
Car 1: "Good, good. They don't suspect a thing."
Just wait until they use SkyNet (sorry StarLink) and Optimus (prime) (sorry Tesla Bot) to take over the world.
Hahaha nice
I already watched the tesla AI day in full, but I still enjoyed watching it in a shortened video form
Such a great explanation of the talk with all the papers loved it! Thank you for working on this one from a different angle!
Watching your breakdown of Tesla's presentation somehow made their achievements 10x more impressive! I'm going to go and rewatch it now.
excellent video - thank you, karoly. it reinforces how powerful the human brain is for visual processing and decision-making! i would like to know why Tesla decided not to use depth cameras or any depth-sensing technique - or is this not so?
Thank you! 🙏 They produce the depth information from the raw pixel data. They are quite keen on developing the smartest AI using the cheapest sensors available. Likely to keep the cost per car as low as possible.
They don't use it, because it is expensive, and prone to edge case errors. Humans can drive cars just fine with vision and neural nets alone.
Because it was expensive at the time. It's less expensive now, and it will be very cheap in the near future, so, they'll probably adopt it in the next years.
@@rubenssz nope. Tesla is all in on vision. Lidar and other sensors don't work well in all situations. Vision with the proper neural nets works in every situation that human vision works with.
@@TwoMinutePapers The way I understand it is that ultimately the vision system must be good enough to recognise objects, their position orientation, speed etc. and have detailed enough modelling of objects to predict their behaviour. The only thing that depth sensing helps with is position and even then it can't be relied on and it doesn't have the resolution of vision. So, Elon's concern was that using depth-sensing like lidar would give the false impression of real AI progress, when in fact it only helps with one thing. And if the vision system is good enough to do all the other things they want, then it would be good enough not to need other sensors anyway.
I know there was a lot to unpack in AI Day, but it's important to note that they convert that simulated environment into the 8 camera feeds, with all the noise and distortion of the real world cameras, and then feed that into the NN. I guess there's no paper for that!
Not necessarily, they could be using it to train just the heads of the hydranet, that already operate based on vector-space data. Either ways, cool stuff. Tesla is beyond bleeding-edge papers but still implements new techniques. Truly amazing.
@@madhououinkyoma Too bad that its not open source. As far as I can tell these videos are what the show the press, it would be interesting to see how good it is exactly.
Great summary! 😎 One of your longest.
Thank you for another great video! I hope they are training for situations the car will not be expected to perform perfectly under: oncoming traffic moving into the wrong lanes, extreme weather events, subsystem failures...
ive been waiting for this sooo much!!!
This is indeed mind-blowing!
I did a video on this myself. It's absolutely foundational to Tesla and their work on FSD. It's astoundingly difficult, and astoundingly powerful.
WOW, I really love this channel. This is some amazing stuff that I can feel really good about. Seriously.
…that said, I’d like to point out that you’ve defined Dimensional Theory better than most. The first three dimensions are given by the cameras but become useful because of Time/the 4th dimension. It then follows that in the 5th we predict events before they happen. At least that’s how I see it… 😊
I could not see this coming. Now, I can't imagine how the next 5 to 10 years will look like with this amazing AI tech...
Same here. I did not think they could get this far this fast. I thought it would be another decade before they could get so much processing power to be viable inside a car. It's mind blowing. I still think it's very dangerous - so far, no one is trying to crash our cars, except the occasional CIA agent. What happens when regular hackers are doing it? Frankly, I'm scared of what so called AI will bring, because it will be used against us by the elites. I'm not scared of self aware AI, I see no evidence for it.
GTA 7 Earth Edition....
4:14 LOL, here the Tesla's image recognition is saying, "ok, there's a car" while my personal image recognition is saying, "Hey, those are definitely San Francisco streets."
You are not Elon Musk, are you? Those are future cars, get it?
@@aniksamiurrahman6365 :-) !!! -- I just thought it was funny because I'm looking at the video and thinking, have I been there?? I would bet money that I know those streets, or ones nearby.
In fact you bring up an interesting point: I heard recently, I don't remember where, that there is some image recognition software that can look at a picture of a living room and tell where in the world it is--based on the types of furniture, rugs, window-shades, and other incidentals specific to that locality. I'd like to learn more about that, and how accurate it is.
well worth my 13+ mins. Thanks for all the videos love them!
I love when 2 subjects I like come together!
I'm not even joking when I say that at times during this video I couldn't tell if the content was real life or simulation. Bravo Tesla, bravo!
Whoooa, worlds collide. I've been watching mostly 2 Min Papers and Tesla related stuff for the past couple of years on TH-cam.
Now we only need The Hockey Guy to somehow add to the mix :)
Anyways, great vid as always -b
Love the longer form content, amazing it was just two minutes!
Thank you for taking the time to put these videos together.
This is what happens, when Al es papers that are amazing on their own, get combined. It really shows the benefits of collaboration.
Watching your videos is very interesting that it makes 2 minutes feel like 13 minutes.
I think you put those the other way around!
😅
@@TwoMinutePapers lmao
Great video. This really is a "hold onto your papers" video. It is beyond incredible how they've solved each of these complex issues to create a real-world usable application. Within our lifetimes we will live in a world where traffic accidents (which are one of the leading causes of death) are almost unheard of. As always, appreciate the professor for bringing this to our attention in such an interesting way from the lens of someone who understands this.
This is incredible. I was honestly getting a little impatient with the progress of self driving cars, but after seeing this video I realized that the problem is a lot harder than I gave it credit for, AND we are tackling that problem faster than we ever should have been able to!
I will wait patiently now, knowing that progress is going faster than I ever imagined. What a time to be alive!
Thank you so much for covering this. I’ve heard lots of interpretations of AI day but yours carries much weight in my mind. They really are getting quite close...
Just make sure to keep your biases in check. 😉
What a time to be alive!
Pretty soon we'll be saying this every hour of the day
Yes! I love it! And I love the positivity here!
Amazing video. This one made my jaw drop. Thank you!
I had never before squeezed the paper. Thank you.
I like how each of these papers are state of the art technology that might be worth billions each. And Tesla is using them as if they are just some randomHelperFunction(). 😎
Amazing that they still only get a level 2 rating when Honda and Mercedes are at level 3 already
@@KP3droflxp not sure if you actually want to understand how Tesla is considered ahead of the game by experts even though a few other car companies are rated 1 level higher, but on the off chance that you do and aren't a troll or irrational thinker, the jist of it is:
- you can hardcode vehicles to achieve the lower levels on the AV scale relatively easily, but getting past them becomes exponentially difficult.
- using the AI approach, the vehicles could be at 99.9999% at level 2 standards and may be at 99.0% of level 3 and 4 standards, but due to a constantly evolving neural network architecture, there are chances that some basic functionality may fail in a 1 in a million shot and therefore Tesla can not claim to have that capability or they will be legally liable.
- A VERY important distinction is that Tesla vehicles are a BARE MINIMUM at level 2 on ANY road in the world, even if no Tesla has ever seen that road before. Other companies are restricted to just a few cities in America that they scan that constantly need to be updated. It's an entirely different problem to solve for.
- finally, most other AV systems need perfect weather and road conditions, while Tesla's could perform in a thunderstorm with heavy fog
Hope this adds some nuance and details and helps you understand the technology at a much higher level!
@@KP3droflxp levels are only for regulators
@@KP3droflxp Level 3 my ass.
@@gnashr4366 Great info, thanks.
I get the sickest nerd chills when I watch your videos!
Awe inspiring!
The most important reason for full self driving is getting me back from the pub after too many beers.
The single most best idea in here is that multiple cars share their models and combine them to better understand the enviroment. Simply amazing.
I feel like I'm going to see the complete transition to driverless cars in my lifetime, which is so cool... In 100 years people are going to balk at the concept PEOPLE used to drive cars, and somehow we didn't completely obliterate ourselves.
Agreed I think they will see the 1.3 million deaths per year as incredibly shocking, it's like the casualties of a war!
A new feature they could add would be to use data on traffic accident locations & frequency (in the past few years, lets say) to have the AI system drive more cautiously through those areas. As a human, we call it "local knowledge".
The system reacts much quicker than humans, and it can see 360° around the car, so it will do what maximizes safety.
I love your usual videos, but this one was extra cool. Seeing the theory in practice
Dear Zsolnai-Fehér Károly, not just the plethora of the mind blowingly intelligent papers but your fantastic presentations themselves contribute the most to the feeling of living in the future I always was dreaming of.
Thank You so much!
I definitely do share your enthusiasm: "What a time to be alive!"
I'm watching your videos from my Tesla with FSD during charging and can't wait to see it really working in real life...
The idea of using every Tesla car nearby's cameras togeather to make an even more accurate map is genius! hats off to them.
That's not exactly what they are doing, I don't think. They put the different cars' predictions together during training, not while driving
The cars only store basic map data (think google maps gps data). What they were doing in the video was combining what all the cars saw as a means of training. For example, a car might have trouble classifying lines on the far side of an intersection but if another car had already been there, they can use that data from the other car in training to teach the AI what it should have seen on the other side of the intersection. This technique is solely used for things like drivable space, road markings, and traffic control signals since those stay in the same position day after day.
Im looking forward to Tesla selling the entire worlds 3D model to gaming companies to play with
Google maps already does this. Any indie developer can use it. Now they could have some competition.. Tesla street view data and models?
Google maps has more data tho, not just roads and places that are accessible by street.
I don't understand ML and still I love to watch every of your videos
Nice video, it's so interesting to see how these papers are used in real life
A great summary and amazing what TESLA has put on its feet in such a short time. The classic OEMs will not be able to keep up with this and sooner or later they will only play a subordinate role.
Exactly what happened to Nokia. ✌🏼🌍
I would love to see other comercial project of AI pappers mashups as this one. Just amazing! And finnaly AI being used in real world aplications.
AI finally being used? It's always been used. 99% of companies use AI in one way or another. It's even being used on my keyboard as I type right now
With real world aplications i meant stuff outside a screen. Stuff that can enhance human life outside.
great video! cool to see how tesla approches self driving!
One second. I have to pick up my jaw from the floor below.
9:40 - are you sure that it's a simulation, or is it just video footage from Canada? I think only Canadians would be that uninterested in such a common event as seeing a moose crossing the street like it's nobody's business
Knight Rider KITT is becoming more real!
bruh
I'm not sure I comprehend how exceptional this truly is...but wow. I don't think this problem will not be solved again on planet Earth for 100 years at least. This is pure excellence.
Ive seen the event live and i knew you would do a video of this :))
Could you do a video on the dojo ai training system? Is it really as revolutionary as people are saying?
Full Wafer Chips are revolutionary, yes.
Tesla isn't the first in that league but they have a great use for it, that why they don't need to sell these chips. Except for some Tesla employees, a comparaison to other FWC is kinda pointless.
Btw: If you want to donate 7 millions, I would buy such a FWC from a vendor who's willing to sell.
@@skierpage I believe they're just using a smaller wafer, no?
@@justinmallaiz4549 I think it was they use a wafer as a pcb? All the D1s are individual chips and 25 of them are put on a wafer.
Wonder whether someone else is doing it. I googled that few weeks ago and I think I saw such tech used like for a bigger chip with chiplets but not on a full wafer scale.
If I understood everything right lol.
Definitely seems like a more easy and reliable approach than the full wafer chip Cerebra does
7:03 Amazing how the AI could even predict that the man colored blue would turn left
Cuz it's fake, rigth?
@@maxdefire no because he was looking left sweet child.
@@Twixx_77 I watched this frame by frame several times. It "predicts" his turn a second before any sign of a pedestrian's head turning.
Ok I've been putting off watching the ai presentation for a week now, this is a super good summary, thanks!
Lots was missed out in this. It's all interesting
Turn real life scenes as seen from a vehicle's cameras, into game-like simulations, so that the data can be easily and safely manipulated. Then network all vehicles, allowing them to learn from each other how to handle difficult situations. VR & AI are at the heart of self-driving. Excellent video. Thanks for sharing.
Man, the world is becoming more and more monitored. Soon every car will have no less than 8 cameras...
that is definitely something we should all be concerned about - and all other the CCTV cameras out there.
NO lidar NEEDED!
This video is so important, thank you a lot.
I'm imagining a world where cars encountering an accident will be able to run simulations in real time to predict what the best way to respond would be to reduce risk of death. That's like real magic
There were generally two responses to Tesla's AI Day
Those that understand technology and were simply amazed at how far Tesla is and where they are going.
Those that struggle to find the ON SWITCH on most devices and went "ROBOT! ROBOT! WE ARE ALL GONNA DIE!"
Tesla's method to get camera footage to vector graphics then the ability to identify objects is staggering. Then you toss on the compute chip, Intel & Nvidia must have freaked out when they saw that.
The Tesla AI Day had one single purpose - It was advertising and it was aimed at talent.
"If you want to work on the cutting edge of AI, come to Tesla" is the message
The others specifically the media had no clue at what they were looking at.
Super intriguing as always. I’d love to see more Tesla related videos from you.
This principle of using simulations to test new functions or equipment is already used for quite some time. The companies that produce ADAS systems like ACC, CC, Lane keeping and full self-driving use Digital Twinning software like Siemens PreScan, Gazebo, LGSLV or CARLA, just to name a few, to test the systems before using it on the road. Still cool to see it demonstrated like this and that it is used in a very big scale at Tesla. Keep up the great work!
awesome as always :D
So essentially, the car is taking in way more information that us, so it is/will be better than us by a lot
AI's learning from simulation is awesome. we humans already do that through the creation of games/sports
Yes and I think visualisation and dreaming play a part in our learning. Ever tried to learn a task like riding a bike. You often get better if you sleep on it.
Not quite the same thing.
The only way to make self driving car drive the best they can. Still it will be difficult for it to be capable as a real human driver who can be aware of a complex situation that a caculated camera data can't do
@@oystercatcher943 Are your comment supposed to be some sort of joke or a clever one?
Love it, when you cover tesla.
Would like more of that ;-)
Best stack of papers so far! ✌🏼🌍
Finally! A TH-cam channel that actually gets it. It's so nice to see this information presented in a way that doesn't try to spin it into how tesla is a scam or some other garbage.
"Buuut, as always... we are doing it paper-style!"
I am giddy with excitement, this is so cool!
Impressive paper, one of the best so far!
Great video as always. Just a quick note. The robot was not a 'silly joke'. The AI event was primarily a recruitment drive. The robot was essentially saying 'if you want to be part of designing this, come work for Tesla'. It was brilliant. In combination with all the self driving stuff, it made me want to go and work for Tesla.
He did qualify that line with "for now" 😉 and if we're being generous the dancer in the robot suit was a just a bit of fun.
@@shyrealist It was more than just a bit of fun. We are all familiar with Boston Dynamics whose robots are all just a little bit scary. Tesla is trying to push the message that robots aren't necessarily scary. The robot suit dancer also makes people like me with a strong interest in the capabilities of AI and having seen AI animations on this channel do all sorts of interesting things, wonder what it would take to teach an AI to dance like that. I don't know how intentional it was, but it made me want to work on the project, which was ultimately what AI day was all about. Unfortunately I am not American and a bit too old for the work schedule of Tesla and don't have the kind of CV they are looking for. But that too is the point, they want the best of the best to compete for jobs at Tesla.
@8:30 … I wonder what is the scenario supposed to be here? ”How to behave when traffic lights erroneously show green for everyone at the same time?” :-p
lmao good point xD
i would say is that the car should still pay attention even when the light is green cause humans may do shit sometimes, or a pregnant woman need go in hospital fast and a driver may go in a red light anyway or osmething like that, so i guess is for the car to still pay attention?
Your papers got me super interested in the 3D modeling side of this. I found Pytorch (that is used heavily at Tesla), and there's also pytorch3D. Blender has Python scripting so I'm thinking that it's very possible to use those. I'm really not that good at interpreting papers, and the math takes me a long time to figure out. But I hope with practice it will become easier. This is one area of AI that fascinates me the most. Especially when you can incorporate it into GTA too, my favorite game! It all started with your hide and seek video! Thank you so much for the inspiration.
"Holy Mother of Papers!" - love this!
Really awesome to see this cutting edge AI being used in real cars.
Tesla could be a real competitor to Google Maps with all it's amazing 3D-Maps and street-view like fotos of the environment which are often updatable (every time a Tesla drives through). 😉
A privacy problem to automatically publish photos from your car
@@hombacom Not legally, at least in America. The supreme court has continuously ruled that anything that can be viewed from the street or sidewalk does not have a reasonable expectation of privacy.
Still, using its AI, it could likely easily filter out peoples faces.
Nagyon tetszenek a videok! Szuper!
Hyped to see what Tesla will offer in the future. This is incredible stuff!
I really want to play this video game
Our cars are becoming AI hiveminds
That is a sentence I never thought I would say
I want people who shout about cars being symbol of personal freedom to raise their voice. Cars that work on its own gives no freedom, but keeps traffic jam and makes the city unwalkable nonetheless.
Awesome video, more of this!
Wish there was a channel about psych and neuroscience papers as perfect as this
11:22
So Microsoft makes a FlightSimulator out of the real world.
Tesla now makes a even realer CarSimulator out of the real world. 😅
ok, i get you. We have to combine them both and than make a space simulator.
"Tesla's full-self-driving will never work" - Every other carmaker
And stock gambler shorting Tesla :)
I'm thankful for the TH-cam AI directing me to your channel. What a time to be alive!
your channels great for keeping abreast of AI developments, but this video is especially good. it shows multiple examples of tech transfer and a very short delay between paper appearance and application. thankyou!
Now, imagine the performance when they run 4k cameras all around, right now I'm sure it doesn't make sense to process such footage...but once the AI is trained giving it better cameras could take this even further. Cameras can also see IR light, making night driving easier for AI.
There's really no need to use higher resolution cameras. Humans don't see in 4k but can still drive alright. 🙂
I can only imagine that the footage from their cameras is upscaled to 4K or even higher before being sent to the NN. Kliksphilip has done some great videos on the current state of AI upscaling, it's very impressive stuff.
@@TheScarnak I don't think they need to, most likely it is being trained on lower quality video. #1 Less computational power required (but still an incredible amount needed) #2 It is probably good to make the system reliable with less than stellar cameras, since even when they are HD or UHD dirty lenses and sub-optimal picture quality will still be an issue.
@@Perceivedshift That's what I'm saying, use cheap sub optimal cameras and artificially upscale the output. Save on cost and is more flexible.
@@TheScarnak I don't think upscaling will help with better navigation. Just like a driver with better vision can see obstacles sooner, so will better cameras. For a time it might make sense to only make the front camera high resolution. Upscaling can't make up for the better data out of a 4K camera.
My biggest investment is Tesla stock, for one reason alone; FSD.
I wish that you hold my paper one-day to review like this!
I always watch your videos. KEEP IT UP