LiDAR vs Computer Vision: Does Waymo Have A Better Strategy Than Tesla? | Random Thursday
Вставка
- Опубліковано 30 вер 2024
- At Tesla's Investor Autonomy Day, they laid out their plan for building a self-driving network of robotaxis and made the argument that their computer vision approach is better than LiDAR-based approaches.
Here I take a look at the two strategies and see which one could be the first to create the self-driving car.
Support me on Patreon!
/ answerswithjoe
Get cool nerdy t-shirts at
www.answerswith...
Interested in getting a Tesla? Use my referral link and get discounts and perks:
ts.la/joe74700
Become a channel member and get access to exclusive livestreams and content here:
/ @joescott
Follow me at all my places!
Instagram: / answerswithjoe
Snapchat: / answerswithjoe
Facebook: / answerswithjoe
Twitter: / answerswithjoe
LINKS LINKS LINKS:
CGP Grey on How Machines Learn
• How AIs, like ChatGPT,...
Drago Anguelov, Waymo presentation
• Drago Anguelov (Waymo)...
Tesla's Autonomy Day presentation
• Tesla Autonomy Day
Pseudo-LiDAR from Visual Depth Estimation: Bridging the Gap in 3D Object Detection for Autonomous Driving:
arxiv.org/pdf/...
news.cornell.ed...
"Oh it will never happen, Oh it will never happen, Oh its here." You should put that on a shirt.
I said this exact phrase to yer mum last night.
There's always some dork (haha, see what I did there?) in the comment section ready to jump right on any opportunity to make it dirty. And that's a good thing. (Thanks Martha)
Still waiting on the one who's going to make it all about American politics.
Looks like somebody missed their cue.
@@mustwereallydothis It's been 6 hours and nobody has fanboyed over Cody making a comment. So that's progress.
Was about to reply this lol. This quote is epic
Did realize this is Cody! Love your channel!
“I’ll never give up my horse. Those automobiles don’t have instincts like my Mr. Squibbles...”
I would trust any horse named such a legendary name like Mr. Squibbles
At least the horse was self-driving.
I don't know what all the fuss is about, clearly self driving cars were invented by the Knight Foundation in the 1980s...
I want my car to have Turbo Boost!
I want all self-driving cars to come equipped with crime-fighting computers. We'll put an end to all car-based crimes that happen on or near roads.
And hot roadies who show up to fix every minor problem...
Hmmm, are we related?
"A shadowy flight into the dangerous world of a man... Who does not exist."
ua-cam.com/video/oNyXYPhnUIs/v-deo.html
@@0neBadMonkey I assume you would be my alternate universe, goatee sporting evil counterpart...
On a side note, using AI to drive to *work* is a funny paradox. In the future we have intelligent machines helping us commute... to what jobs?
Esp when your job is to fix computers. Going be a weird future.
Didn't you people watch _The Jetsons_ futuristic cartoon? To what jobs? Really? George Jetson lives in a nice sky-pad, and he works a job in which he works about 3 hours a day, and 3 days a week, and apparently in the future, that is a "full time" job. What does he even do? He pushes buttons and sleeps half the time at work. The robots and computers do all the real work, and George just tweaks the computer and gives it some direction. He is a computer programmer. Who would have thought?
And once the cars drive themselves, we still need for somebody to design the self-fly flying cars of the future.
@@yosefmacgruber1920
And to add, before automobiles were invented, going to work was a hassle and slow
@@peterrobannsobrepena6029
For most people, the pace of life was slower, there wasn't as much cool stuff to buy, and taxes ere not much. People were probably in better shape and more easily could walk, since they walked most everywhere, the food was healthier, and they didn't sit on their butts watching mind-rot TV all day. There wasn't much for good jobs, well unless working the fields was your idea of a good job.
Many jobs won't need a commute.
You NEED cv either way, LiDAR cannot detect the markings on the road like lanes etc.
LiDAR is just for detecting obstacles, and their distance.
LiDAR is just RaDAR with light, right?
@@fattahrambe yep
LiDAR also has issues where purely visual systems have trouble. Snow, fog, precipitation, etc. are all opaque to LiDAR and cameras. Tesla uses Ultrasonic and RaDAR, which both don't care much about those things, whereas companies using LiDAR have...well, I just looked at Waymo, and it appears that they just have LiDAR. Additionally, they need to build 3D map data manually beforehand before their vehicles can be used. So they basically rely on Google Maps Streetview on steroids. Of course, those maps are only valid for a small bit of time before there's construction, or what have you. In point of fact, LiDAR is great for ideal conditions and environments, but anything remotely worse than a calm sunny day could cause issues. A sufficiently powerful downpour could "ground" all of their auto-taxis. Which means that these cars might not operate very economically in, say, London, or somewhere with higher-than-average rainfall.
This seems like a huge disadvantage. Strange that Joe didn't mention it and I'm wondering how the LiDAR self driving vehicles are operating. They have CV as well?
@@SirPrancelot1 I'm pretty sure the LiDAR vehicles all use cameras as well.
*sleeps in self driving car*
“Do you have any idea why I pulled you over sir?”
“Well yes but actually no”
"Your LIDAR is not working for night driving, fix your LIDAR"
I can't wait for the day that I can go on a long overnight road trip and kick-back and sleep the whole way (I may sleep with one eye open for the first few times though!) Great video as always!
It's going to be super weird at first!
Even beyond that imagine going on holiday in a driverless, electric RV. You can just go to sleep one night and wake up in a new city, if you're tired at the museum it can come and pick you up, you can visit out of town places with ease.
Funny that - I can do that any time I want. It called "trains" and "taxis" :D
@@totalermist Robo Taxi? According to Elon, maybe next year...
A guy in the Netherlands got pulled over in his Tesla on autopilot, while on the phone. Was a whole fuzz about it.
Predicting the future will always be impossible... Until somebody does it. Then, everyone will be able to do it.
Whoaa, I just predicted a couple of hours ago that someone would make that exact comment!
i don't where you got that from, but something tells me you didn't come up with that. ;)
Thank you Mr. Spock
@@thecoclster8427 if you really don't get my lame joke, try listening to Joe's four minute mile story and the related ones that follow. If that doesn't work, I guess I'm just not all that funny. That's ok though. I amuse myself at least.
It begins at about the 12:30 point in this video.
@@willinwoods and an hour later, everybody was predicting it.
I have honestly wondered why Tesla and the rest use different systems to each other. Now my question is answered! Thank you Vsauce 4 very cool!
Early on in my channel I tried rising up from the bottom of the screen at the beginning of videos but it was too much like Vsauce.
@@joescott I consider you to be one of the better science channels and on par with the Vsauce channels! I'm glad that you went another way and developed your own distinct style. I really enjoy it!
This is Joe Scott my friend not Vsauce 4... but good to know
@@chapo335 Joe is your friend? Joe is everyone's friend.
I gave you a like because you had 99 likes and I couldn´t help it
Hey, Joe. You described genetic algorithms, not Machine Learning.
not only that, but machine learning is such a broad field - from linear regression to convolutional neural networks and GANs and whatever...
I didn't realize there was a difference. If you have time, would you mind explaining the differences?
The funny thing is, CGP Grey put out a follow up video to the one Joe referenced, but to be fair, the one shown here was MUCH funnier, and I loved it anyhow. :D if any of you want to learn the difference, you should check it out, it's 2 min long ua-cam.com/video/wvWpdrfoEv0/v-deo.html
@@mustwereallydothis The shortest answer - the genetic algorithms can be used as one of the tools in machine learning, but saying that GA's are the machine learning is like saying that "+" is the maths. Machine learning is largely linear algebra with flavour.
@tommy aronson There are probably two or three scientists in the world that can coherently explain what their science does to a layman. Maybe one of those three can do that while being genuinely funny. Either you only listen to three people in the entire world - and they can't entertain you 24/7, they have science to do - or you listen to other people too. And Joe is not the worst guy to listen to.
70 year old guy here: So many things about this video resonated with me. Even though I had a technology related career I have been a skeptic about the hype of new technologies and I've been wrong a lot and here it is, it looks like I was wrong again and I'm liking it. The Tesla autonomy day video that he mentioned is well worth a look. I don't know if Tesla is going to make it or not but regardless Tesla will have changed the world and that's pretty cool. And kudos to Joe on this video. It is one of his best.
Yeah, I've always been very skeptical of Tesla, but this really could be a game changer. I'm actually suspending my disbelief. Don't screw it up Elon, we are all routing for you.
what you're explaining starting at 1:50 is a kind of machine learning, but its NOT the kind (neural net + reinforcement learning) that tesla uses!
i think its more the kind that Waymo uses.
Yeah saying it "throws away" the wrong predictions is very wrong. Auto driving probably uses lstm or Gru recurrent neural networks which actually keep wrong predictions to prevent exploding or descending gradient. Essentially just cause you made a good decision doesn't mean you should just forget all your failures
@@Kaodusanya No indeed. What's interesting to me is if it can use all of that data both successful and faulure, to make decisions about other subjects, or analyze the data to determine WHY it made those decisions, and what made them right or wrong in the first place. In other words can it learn about how it learns and use that information to learn how to learn better?
@@Mrdevs96 Waymo uses Neural Networks too, genetic algorithms are by far inferior to Neural Networks with Gradient Descent based optimizations.
It was just wrong the explanation
There were several mentions of steroids in the video, Joe. Do we need to talk to your personal trainer?
I did kinda overuse that one didn’t I?
and if they work, are we in for some topless videos?
I still want to see an official drugged up olympics..
So many pros are on the juice anyway, let's make it official and see someone run the 100 meters in 5 seconds.
Then weight lift 2 tons..
followed by a 30 meter triple jump..
and a 4 story pole vault..
anything Elon does is the definition of steroids!
8:23 I mean just look at him!
Dave Webster don’t know if you have ever seen it reminds me of an old SNL skit all steroid olympics.
I want to own an autonomous car, the sooner the better. I don't think I would like the whole "not owning your car & just calling it when you need it" thing, but that could change. My bet is on Tesla, and this is largely why I have a Model Y on order.
So you want your own car, but you don’t want to be responsible for piloting it safely? If it caused an accident would you accept liability?
@@nathanchildress5596
I have a Model Y now, with Autopilot. Obviously I would accept liability if it caused a wreck today, because I am still a much better driver than it and I keep my hands on the wheel and eyes on the road at all times. But at some point in the future, if it achieves a level of safety that is, say, 10x better than me, at that point I want to be able to stop paying attention and sleep or whatever. So liability needs to switch from the person in the car to the maker of the system at that point.
@@Jer_Schmidt Well as long as you rely on yourself when it matters, then I have no gripes, sorry to sound salty. There are just so many people ready for the future to be today without realizing that the test results from current systems is terrifying. Machine learning is very different from the way we perceive the world; every scenario has to be catalogued and variables like weather, foliage, flashing lights, puddles, have all given autonomous vehicles problems. You should listen to Missy Cummings, a former fighter pilot, professor of machine interface at Duke, and now a government advisor on autonomous vehicles for real insight on what they are & are not capable of.
it might take 10 more years to reach indian traffic level
Not if it reaches complete self driving
By which it could drive on most of the indian roads
What's crazy is that autonomous driving may be here before India has and obeys traffic laws.
This isn't just a race between automakers and tech companies. This will rapidly become a race between entire economies.
The lagging country will have to carry the drag of accidents and infrastructure inefficiencies. Those that adapt quickly could find their healthcare system going further, their road system operating more efficiently (platooning), goods and people getting around more quickly and cheaper. We fuss now about driver unemployment, but in an economy of guaranteed jobs or UBI there would be millions of eyes and hands not wasting time staring at lane markings. How quickly could we develop a corp of volunteers to look after the children, sick or elderly? How big could the creative class get?
Indian traffic seems easy. Just do not hit cows. People do not matter as much.
3 years later and there still isn't full autonomous driving. Why? humans see and process information on a sub-conscience level. We look at other people and from our lifetime of experience, we often can predict what the other person will do. Your sub-conscience sees, hears, feels, and smells a flood of information every second. Our conscience mind somehow sorts it all out and chooses what is most important this second.
There are a lot of fundamental inaccuracies in the way you explain Machine Learning and computer vision...
Agreed, missed the mark completely on that part. When he was talking about machine learning he explained a neuro evolution strategy, which is a machine learning strategy. But this is not the strategy that tesla uses to train autopilot.
As a d i disabled senior I rely on drivers to take me to hosp etc. Most of these I have noticed are
Stressed out
In a hurry...behind schedule
Working a double shift
Hungry
Getting distracting personal calls
Arguing with their dispacter and upset as a result
Tired
Inattebtive
Underpaid
Cant wait for self driving!
Will allow servuce oroviders to take more care thus do a better job eitj me and my needs
It will be a game changer. My daughter can't drive, and I guess I'll be buying a tesla for her at some point.
The nice thing about autonomous cars is the cars can share their information and learn even more quickly. That means the more ai vehicles there are the faster they will improve
This is what excites me the most and will cause the biggest reduction in traffic. It's not the learning sharing that will be the biggest deal, it's the fact that they can communicate with each other in the first place. The cars won't have egos. They won't get road rage. When it's all autonomous, they can talk to each other and coordinate. Imagine a system where two cars could leave a gap between them starting a mile before an on ramp for a car that will be merging in. No one cuts in to fill the space. The adjustment needed to make that extra gap that early barely even registers in how it affects other cars. The merging car gets in at full speed without disturbing the flow at all (obviously in places where ramps are long enough... Looking at you Eastern PA).
Imagine cars driving in a way (read: politely) that causes brakes and tires to last 2-3 times as long. Imagine cars that can automatically report road conditions to DOT.
Achieving a critical mass of autonomy is only the start; there are so many possibilities and benefits that will be seen down the road.
@@DerrickBommarito Question is, will the companies that own the autonomous systems driving these cars share information with each other? We are now in an age where data is the king of commodities, I don't see rival companies readily sharing their information. We're going to need regulatory bodies to step in to ensure that they do.
@@damionmurray8244 I expect that once critical mass is reached, if not earlier, there will be some form of autonomous cooperation standard and framework for manufacturers to work under. Like making a new WiFi spec, web standard, electrical supply procedures for the grid... There will so much to gain from making sure the cars can plan actions together I don't see how it wouldn't become a guarantee. Maybe I'm just being optimistic.
@@DerrickBommarito well not only that but, since communication would make for safer roads, it might be legally required that cars share a communication network. Even if they didnt, a network of Tesla cars would be able to relay information about the positions and velocities of individuals in a network of Company x's cars. So that communication may only be necessary within company lines.
Didn’t ieee passed a new standard for 5g, NFC, and/or Bluetooth specifically for VtoV and or vehicles to infrastructure.
I think that all new cars should broadcast their position and speed, and additional data, in next year or so. That would make autonomy roll out quickly. Cars would be like a school of fish. And magnitudes safer then current cars
I got my PhD studying human visual perception, and it is absolutely amazing that so much progress has been made on this. In the 80's, we really didn't know how you distinguished features on an object from different objects in the background. I was one of the people who laughed when MIT's first self-driving van decided that the contrast at the edge of the shadow of a tree was higher than the contrast between the edge of the road and the grass, and the van left the road and creeped slowly toward the tree. They only let it go one mile per hour for obvious reasons. I felt that we needed to figure out how we see the world before we'd have any chance of making an artificial vision system capable of driving a car. We haven't learned much about how we see since then, but artificial intelligence has made advances that are truly incredible, even to self-described experts.
You confuse machine learning with evolutionary algorithms and deep learning. Also, in Karpathy's presentation no where he mentioned that they are using evolutionary algorithms along with deep learning.
Nine8Hundred00 is correct. While evolutionary algorithms (aka genetic algorithms) can be used in conjunction with deep learning approaches (See the league process with AlphaStar), the primary method used in Tesla's approach and the method that is directly supported by their custom hardware is deep learning. In deep learning you can distribute the learning process but don't have to. To train the network the data is fed forward through and a result is computed. The difference between that result and the proper result is then fed backwards through the network and the weights are adjusted according to an algorithm called back propagation.
A more general description of machine learning is "learning from data." The actual method of learning varies, but the common factor is that the input generating the algorithm is a dataset, rather than a programmer's reasoning about how a thing should work.
Lol yeah I cringed. Machine learning has become a buzzword yet many don't know what it is
Thank you for pointing this out ! It's sad that a CGP Grey video is a bad example for a change.
Actually deep learning isn't a particularly meaningful term. Deep learning is just how standard neural networks work but with more layers. More layers allow for more complex models, but the theoretical basis is the same as a standard neural network. But yes Neural Evolution Networks/ genetic algorithms are what he was describing. Pretty much all neural networks of any use for difficult problems are using deep learning.
American guy: "I love driving my car", puts in D, and cheerfully proceeds along the endless straight-line highway. European guy: "Gotta go, see you later", puts in first, second, third gear, then second, third, fourth, keeps turning at every corner, breaks, dodges the old lady on the left, and proceed through the Italian peninsula while the Sun sets down on. Also he drinks an espresso.
While there are boring sections of the interstate you've clearly not been to America if you think this is how the roads are. There are plenty of places with scenic twists and turns, then there are all the potholes to dodge, then there are the funny road side advertisements for churches and adult fun stores (usually side by side). Then when you get in the cities when you're not dodging pedestrians who jay walk constantly you also have to contend with confusing road signs that cryptolinguists have hard times following.
@@tomstdenis True, I've never been in the States but mine was a joke, making fun of both parts. Didn't mean to be mean or anything. Indeed, I'd love to come visit your side of the world, hopefully one day I'll be able to.
american: drives down on his gas guzzling car
european: takes great public transport, saving money on insurance and car cost, not to forget the environmental benefits while reading the morning news paper bc he doesn’t need to keep his eyes on the road
ohmygoddahal great public transport? you obviously don’t live in the uk!
@@ohmygoddahal Insurance is crazy expensive and there's no affordable public transport in most places here, maybe you went to London but outside of London public transport sucks. For perspective, a day pass for my local bus service is £11 ($14), and you can go a whole 20 miles from one end of the bus route to the other for that! A train to London would cost $40 ($50), making it drastically cheaper to pay for fuel, or electricity for my Model 3 in my case!
I wish the UK was as cheap as you suggest, that'd be great. :P
Thx for video. As an AI developer CV is and will be the solution. LIDAR is cool and clever but is a patch and temporary. CV all the way.
Thx again !
Computers still have a long way to go. They can't come close to a human. It's simple, if a human had been in control of the beersheet space craft, it would be on the moon's surface. When people talking about AI seems worried about the progress, I just pull out the old Google Voice. That damn thing can't even understand basic words. I personally don't think full AI will ever happen. I think I still still see the 0 and 1 in computers. It all got started with the telegraph. It's going to take consciousness for anything to come close. Even then, look how far we are above animals.
@@Bryan-Hensley Pretty sure humans were in control, time lag to the moon is less then 2 seconds and they had real time telemetry all the way.
@@romanr1592 I figured you'd be more up to date on that mission. I love watching the computers work. In a way it's kinda basic. It's got a box in the telemetry and if it drifts near the edge the computer responds. Even though software has advanced drastically, computer are basically still just big fast calculators. Something is going to have to change at the core level of computers before true AI become anywhere near reality. Something like chemical memories. Chemical instead of electrical (Just a wild guess)
I hope I didn't get the science too wrong on this. :)
I'm drifting off the subject. Sorry guys.
Just back from a trip to San Fransisco I drove in and out of silicon valley everyday for 10 days , I kept seeing at least 1 or 2 waymos each drive and every time it had two people up front with one guy with hands on the wheel, I was also surrounded by teslas in traffic.
Kinda feel like when ever tesla go for that self driving taxi service it will be game over for competitors.
What I want is an autonomous you tube client that automatically beams joes videos into my brain.
in another 5 years Neuromesh/link made up of autonomous computer bots at the nano scale, will bring joes videos to your brain, via the Starlink Internet of Things backed up by Open-AI, That will just be one of their functions, not connected with augmenting your immune system to keep you looking and feeling 29 for the next 150-200 yeas. . And if you are on Mars, Starlink's Interplanetary Sol System Wide Quantum Communications Channel will be a mental switch away by 2030 :D
Elon Tine ⌚ is still more accurate than Valve Time ⌚ 😁
half life 420 confirmed
Half-Musk 3
So, Joe, do you like Tesla?
[that's a joke]
2023 -- Tesla still don't have full self driving cars.. 😂😂😂 #MuskSucks
10:40 is where I walk to school every day xD how random is that?
mmm, good to know
Did that on purpose...
“I always feeel like, somebody’s waatching me....”
If anyone is here now, search for Tesla FSD Beta video. They currently look less capable than Waymo, but the improvement is extremely quick and they have a general solution that will work in all US cities on day 1.
I expect 2021 to be the year nearly anyone who can afford it can go from A to B for 99% of drives with no intervention.
My brain-head thanks you for this video.
My brain feet are confused.
My mindbrain appreciates your brainhead's appreciation.
dustin fisher Hulk: BRAIN-HEAD HURTS!
I wouldn’t compare light detection and ranging with a vision system (cameras). Compare LiDAR to radar both are for mapping objects. The Camera is used for object detection, and reading signs/ stop lights.
You don't have LIDAR on your head but you can still drive just fine...
You don't have a human brain in your car either.
LIdar is like a blind man poking around his environment to find his path. Just like Lidar, he knows he's poking something in his path but not exactly sure what exactly is the object he's poking.
"...except they have 8 (eyes) that see in all directions."
COME ON, photo of a SPIDER, Joe. 😜
Waymo only works in perfect laid out roads in sunny day in closed neighborhoods.
It's why they chose Phoenix. Few rain days and no snow days. It's also within a controlled section of Phoenix so they've dramatically reduced the amount of variables you'd get including suburbs.
Speaking about generational, how will the elder lovers of cars react? It's hard to give up driving, but an autonomous driving car may bridge the age Gap?
The thing about LiDAR is that it's better than human vision. Important when you're chasing those nines.
Pretty sure camera is better than human vision too. Just don't buy those from China and you'll be fine.
One and a half year after this yt was released Tesla releases a beta version of FSD that pretty much drives around in cities without much user interaction, stopping at stop signs and red lights and going around parked cars and roundabouts and turning etc. It seems to me that they are very close to FSD.
DOOMED! They are all DOOMED! Can’t fault Elon’s logic. The only argument against him is that he has chosen the more difficult path. But he has a track record of choosing the most difficult path and winning. The more he does this it looks like his critics where simply unimaginative and short sighted. Cynical people look down their noses at us fan boys who admire Elon but there is a reason he attracts fans ... he has repeatedly proven that he is awesome.
He got rid of radar which I think makes the whole autonomy domination thing a little less likely.
I've seen Autonomous BMW driving last 10 years ( with one or two people in it ) and "STILL TESTING" 🤣👌
Suggestion video for random thursday, alpha zero chess by deep mind
Is that a new topic you came up with????
Super original !!!
Would be the 4th video of it in his channel.
LOVING the videos! I PREDICT: Within 20 years, Tesla cars will be self-driving. on MARS. (ʘ_ʘ)
I'll stay with my '65 Ford pickup. I can fix it.
It may soon cost you less than your fuel bill to own a new tesla self driving robo taxi...
Tesla will pay you $600 a week to use your vehicle in the 'fleet' plus 85% less running (fuel) costs and almost no servicing (no brakes, nofilters just 4 year coolant change)
@@gordonjohnson2497 I don't think your reply speaks to James Harback's sentiment... Simple, known technology is attractive and will always have some degree of merit.
Same as I with my '74 Camaro and '72 Corvette. I'm really glad I lived when I did. Speaking environmentally, I don't care about owning new cars (My newest is a 2003 F150). No one really considers all of the mining of materials for the metals, oils for the plastics, etc. that goes into making a new vehicle. When you add in electric, then you are talking a whole other set of materials to mine for the batteries. A process that leaves land polluted and contaminated.
I'm not knocking new car owners, electric vehicle owners, I mean I'm interested because I'm here aren't I? My point is that as in most things, moderation is probably the best for all. And as Joe says, it's probably going to be a generational thing.
@@Pantheragem
You only mine those materials ONCE
Batteries get recycled and the 2nd iteration of those minerals represents almost no carbon or pollutant foot print.
That and the fact that the new vehicles that self drive will save millions of lives means you will not have the option soon.
Teslas will be better, cleaner, safer and cheaper.
@@gordonjohnson2497 I'll probably be dead when driving yourself becomes outlawed. I live in a rural area anyway.
"as long as its in Pheonix" Definetly thought you said "As long as theres a pe**is" 😂
[Starts at 8:20]
This video is terribly outdated. Check out the company Quanergy Systems; they're providing discrete solid-state LiDARs for $250. Since they can precisely track the distances and velocities of surrounding objects, they are 100% going to a part of autonomous cars, most definitely alongside cameras.
Probably a high end upgrade...
If they are 250$ then probably they will be. But until they start producing them then it's a non starter.
1000% false, you don't know what lidar is. Lidar will never work in bad weather. Lidar in place of radar won't work since radar works in bad weather. All you are doing is adding lidar to get "more accurate" radar only in good weather. Even waymo is likely using lidar as a secondary sensor and is primarily visual just like tesla. 360 degree lidar is just a safety check if visual detection fails.
Waymo's primary use for lidar is to train the visual detection to learn how to read distance and build their simulation maps. There is a good chance waymo stops using lidar if they make a product they sell to car manufacturers. Even their use of lidar in their own vans is only needed while their visual detection isn't good enough. Waymo cannot collect data fast enough to have more stable visual detection, lidar is their shortcut, until they catch up to where tesla is today in 10 years with how much data they have. Remember, tesla used lidar on the roof of teslas to train their AI to see distance. But moved passed that when the AI got accurate enough. They also used the forward radar in the production cars. Once AI can do the job, you don't need lidar. Radar has an advantage by giving you the distance to the car in front in bad weather.
@@_PatrickO Precisely, well put.
Pretty sure Quanergy's lidars start at a few thousand usd per sensor for the products they actually have on the market, as opposed to their hype generator. Whatever their prices are on current products, they're not selling
waymo is not "with limitations"(except for lidar being big and expensive) it's doing exactly what tesla is doing as far as software .....plus .....having a scan of the area it operates at and plus lidar
Elon is saying they have a bit better software training and he is saying that the cars can drive perfectly fine with ap hardware 3.0 without scan of the environment and lidar .....and that is true .....especially because it's more cost effective ....but that doesn't mean that waymo is not using a bit more data ....does it absolutely need that data is it a massive difference over the claimed better training with teslas maybe not really
....but there must be a government body testing identically the 2 systems for looots of miles to see exactly what is the performance difference...saying that someone's training is better or that someone's data from the environment is better doesn't tell u which system is in fact better
perfect timing... just got finished watching Lex Fridman's Self driving cars video (MIT Self-Driving Cars: State of the Art)
vuiz if you haven’t, watch Tony Seba’s lecture on disruptive technologies.
@@GG-si7fw thanks for the recommendation! that's tonight's viewing material chosen
My thoughts on this are if both cameras and Lidar are successful (because I think both can be) will Lidar ever reach price parity with camera? And if not what would be the cost of transitioning lidar systems to camera based ones?
If you put 100 LiDAR cars on the road, do not they blind and confuse each other?
My point is that if ALL tries to scan one surface with lasers, I can easily imagine a situation where critical, life-threating errors can occur ...
Niespotykanie Spokojny Rowerzysta I had been wondering that same thing. It’s like a bunch of photographers using flash to photograph the same subjects at the same time. Or a lot of active sonar systems in the same swimming pool. What a mess of data that would be to sort out.
(A “fool’s errand” indeed.)
That's a good question. I hadn't thought about that.
I had the same question. It is true that an environment with multiple cars all emitting spinning laser beams would be a real mess of laser dots moving in all directions (imagine the poor stray cats in such an environment, should the dots be visible to them!). But I surmise that each car would have no problem following it's own dot, knowing exactly where to look. So if there is some interference, it should be random, and so short as to not constitute a data set that could be interpreted in any meaningful way (in other words, just random noise).
On the other hand, I wonder about the possibility of CV being fooled by optical illusions. Remember the driver killed by a Tesla that mistook the side of a white truck for the sky. A lidar would never have made that mistake.
@@mmmikeyyy That and I bet even if it does pick up lasers from other cars it probably just sees that as light reflected off of whatever object it was already reflecting off of.
@@hondaguy9153 Probably not. It works more like radar in that it measures the time it takes for a beam to bounce back. Other lidars will have no correlation to others' bounces and could show up extremely close or extremely far point. I'm guessing they have an algorithm around that but it will get harder the more lidars there are. I suppose they're betting on lidar not being ubiquitous.
You should try to get to interview some people from Waymo.
When a car is running in traffic it is not learning/training, it is only executing the previously trained lessons. Waymo stores all this data, but Tesla can't really do that with everything their customers are doing (for obvious privacy reasons) and data management (uploading terabytes of data each time you drive isn't really realistic).
Comparing Waymo to Tesla is not comparing LIDAR to computer vision. Waymo is using both, to make the safest driver possible. Tesla is using only computer vision, to get "good enough", save cost and for esthetics.
I didn't expect it to take so long to implement self-driving. I agree with musk, however, that Lidar is a poor end game, and I never expected Musk to put out a new hardware board to achieve SD. Musk rules!
Waymo is big but Tesla is "waymo" bigger lol ;)
The problem in consumer technology is that the better solution doesn't always win. Remember Betamax vs. VHS? Sometimes it just boils down to basic economics and public perception, or who is the better self-promoter.
Doven was what you were looking for. Great video buddy
Dove : a small bird, usually soft grey in colour. Much like a pigeon, only smaller,
Dived: past tense of dive.
Hey, kids sitting in their parents' laps in the car were the ORIGINAL airbags.
Yes I can't wait for full autonomy in cars...and where is my jet pack?
Not quite a jet pack but still looks like a lot of dangerous fun ! ua-cam.com/video/GKBNTP6D1ag/v-deo.html I hope Tesla is into this . An Uber / Tesla type robo flyer would be a blast to get home with while drunk.
Homo Quantum Sapiens the joke is: about half a century ago, there was this weird obsession with jetpacks, and everyone thought we would have jetpacks by the 2000s. It’s rather common joke that will pop up when people try to talk about how they think technology will progress in the future. It basically just calls out the fact that no one really knows how this will progress and be applied in everyday life. It’s all just speculation mostly based on opinions.
A statement I have seen that sums it up pretty well is when someone said something to the affect of “I’ll laugh at all these people who think autonomies driving, Mars cities, gene editing, (just think some futuristic technology I forget what it was specifically) would never happen when we get it.” And a person replied “yeah, just like how I laughed at all the people who said it wouldn’t happen when I finally got my jetpack and cut the wheels off my car and turned it into a hover car.”
I mean all in all, it’s really just a joke. You kind of just got r/whoosh-ed.
I have a copy of the film Pretty Bird on the jet pack caper story ua-cam.com/video/3tg_eZ6ALr0/v-deo.html . Might watch it this afternoon . A Robo Jet pack might be a real crowd pleaser or air shows .
I still dig this post on air sea and road gizmo ua-cam.com/video/uGVQMdvk5Bo/v-deo.html . The theme has such a it might have been feel to it
batteries dont have enough energy density. with the amount of power required to run a few small jet engines you'd be out of juice in like 10 mins
I'm not wanting for level 3-4 self driving. Im waiting for a EV pickup truck.
I love how purposeful you use background music to indicate the end of a tangent or to indicate something is an opinion or sarcasm. Keep it up, please.
I have a PhD in statistical inference and operations research and I've spent a few years working in the self driving car industry. People are wrong. The perspective in this video is naive. And many a Tesla engineer etc are making wild extrapolations which are inapplicable. Infinite extrapolation error. So first of all 1) AGI is a *subproblem* of a fully self driving vehicle. Our journey towards AGI? We're on step 0 of 1 million steps. I asked my advisor when he thought self driving vehicles would be level 5 autonomous. My advisor owned a billion dollar remote surveillance company, was on the air force advisory board, is deeply involved at the Pentagon, was the president of the CDC society of IEEE and the chair of my dept. His answer was "never". That surprised me so I said why? He said because of the combinatorics between exception cases that couple together...
"Ai" has 0 intelligence. Curve fitting is not intelligence. If I write the least squares equation down, does that mean the equation is intelligent? Does the product of using such an equation somehow spontaneously engender/ acquire intelligence?
Elon is right that lidar is a crutch but it can make unobservable aspects of a system observable, in the statespace control theory sense. Cameras and lidar have different strengths and weaknesses.
Having lots of data is certainly helpful however good algorithms are worth more than good data. Go see what Geoffrey Hinton has to say about capsule networks.
Elon is correct concerning using the method of residuals/ design by exception cases.
People don't seem to understand that Phoenix is seen as the easiest place for a self driving car to drive at least within the US. It's cherry picking cherry picked cherries. So give me a break. L4 autonomy in Phoenix means nothing.
I didn't like your synopsis of that which machine learning is at all. You didn't miss the nail entirely but you missed 80% of it... Competitive kerning is almost an afterthought relative to everything else going on. You need to start with the notion of a line in the sand... then a multi dimensional box that can hold a set of points inside... explain a feature space... and then potentially discuss how a neural network identifies these boundaries/ separating hyperplanes/ discriminant functions...
Once we have AGI we'll have self driving vehicles... until then we'll have self driving vehicles that do the right thing 99.9% of the time but when 3 or more exceptional circumstances coincide very, very bad things will happen...
This is better:
ua-cam.com/video/aircAruvnKk/v-deo.html
When, if you are lucky enough, you meet George Jetson...he will be flying a Tesla.
Just like me.
Tesla may start using LiDAR when it’s available solid state.
I will give up my car in a second if I can get a car to pick me up via an APP.
@Joe Scott
Hi Joe, machine learning graduate here. Major detail: What you described as Machine learning is called "Evolutionary methods" and actually makes up a very small part of machine learning. Almost all major applications, including the autopilots of waymo and tesla, use so called "Backpropagation" which is a lot closer to how we humans learn, actually adjusting the way we think by tiny amounts based on wether we were right or not.
Please let me know if that was helpful.
Also the other competitors also use Computer Vison, they just use Lidar as an additional input. Object recognition/ classification is actually way easyer using images as we have tons of examples of what a car looks like etc.
I can already see how people from the political opposition have more tragic accidents.
I have all Trump voters in my AI library 😂
Lidar isn't so expensive anymore.
Modern Lidar doesn't spin. It's a small solid state system now.
Lidar would add an additional layer of security. Not a replacement for cameras.
The "we don't shoot lasers out of our eyes" argument is dumb. We also have no Radar and supersonic sensors, but Tesla build them in.
The main reason is: Retrofitting Lidar to existing Tesla cars would be too expensive.
That's about it.
"Modern LIDAR doesn't spin" -> Literally in the video, two of the top autonomous driving companies using spinning LIDAR domes. Yeah ok.
Waymo cars get better at driving in the snow every time it snows in Phoenix.
Like it even rains there...
LiDAR can't see ice.
That’s why they are testing in Michigan as well 👍🏻
@@mitchellmurphy9556 They can't rely on lidar in inclement weather. So... computer vision. I wonder who has the recipe for success there.
I guess Tesla's approach is WAYMOre sophisticated
Another great informative and entertaining video Joe, keep up the fantastic work!
Thanks!
Both the technology will work but we won't be able to afford lidar one
If Elon Stark says liader not viable you better believe him.
It will all depend on whether Tesla can get their collision detection to the point where their cars aren't colliding into semi-trailers at full speed and decapitating the driver while they watch Harry Potter.
Noooooo!!! You cut off the extra DOOMED!!!😂😂😂😂
WHY?!? That was the best thing about the quote.
What if ?
What if i did what the roadrunner did to the coyote
And put a giant banner on the road with a picture of a tunnel painted with Vantablack material making the inner tunnel to absorbing almost 100% light
Would self driving car stop or keep going?
Stops - Tesla has a nose radar and 12 ultrasonic sensors :)
Joe Scott: cars are a symbol of stature........
Tesla owners: Eat that fellas, I’m rich and I care about saving the earth from fossil fuels.
You don't have to be rich to buy a Tesla, they aren't that expensive, and unlike high-performance gas cars they are very cheap to operate.
@@andrasbiro3007 they are something like 30-100 thousand, that is very expensive
András Bíró bruh people are getting richer I agree, it’s just I ain’t there yet 😭😭.
@@briandiehl9257
Expensive is relative. The average US new car price is $35K which is the price of a base Model 3.
I don't consider myself rich, not even by my relatively poor country's standards, yet I drive a Model S.
@@andrasbiro3007 Yes, but that is just the base price of a tesla. Also "The average US *new* car", it's important to remember that the average american drives a car that is over 11 years old
Lidar is much more expensive than cameras, so Elon needs to rationalize why they don't use it.
Elbrasch They built it and use it at spacex if it was needed to win they would get it done
Nobody seems to say this but anyway, good afternoon Joe!
Also, FIRST! (Possibly second xD)
Blitzmulthe hahahaha I think I was first lol
Plastic bag: I’m going to block your path.
Car: Oh no.
Tesla uses backpropagation though, but great vid
Who is doing the tagging on all of Tesla's data?
Hope they are not using windows operating system for the cars .. The blue screen of death has a whole new meaning lol
Linux rules 😁
Sublunary World FreeBSD is much better than Linux.
QNX all the way. I don't need my self-driving algorithm distracted by pop-ups, anti-virus alerts and Mr Clippy making inane suggestions. Who the hell needs a gui anyway?
@@jpe1
Had is on a system a year or 2 back .. it's very stable ... I have been a Debian fan for years but not sure how it will go now that it's creator is gone .. Time will tell.. I will have to give bsd another look ..see where they are now ..
I expect them to be using some sort of real time operating system for all of the computer vision and self driving aspects.
In a Tesla the center console does crash and it doesn't boost confidence in the system. At all.
First and foremost CV is being used by ALL parties, LIDAR or NOT, its practically required, they are BOTH, CV.
What is different is Tesla uses CAMERAS
Waymo isnt doing LiDAR .. its doing both CAMERAS AND LIDAR...since it covers different tradeoffs.
Also the algorithms and simulations and tools that Google is working with are far more sophisticated than what is available at the Tesla front. The dataset at Google while "smaller" is actually far more comprehensive and "feature-filled".
And one of the techniques they are using in their simulation which Tesla isnt, and this is also very important, are Adversarial Neural Networking. Using ANNs for parts of the process makes the driving algorithm stronger since it fights against itself in being slowed-down/disrupted/damaged/fooled and it learns from itself.
The type of Machine Learning you described is just one of the many types of ML techniques, and while that GNN (Generative Neural Networks) are definitely being used, Google is applying ANNs to make it far more breakproof
I would amend your statement to say that Waymo uses cameras, lidar, AND high definition maps. Lidar without HD maps is not very useful. With both it makes FSD easier to do, like a person familiar with driving in a certain locale. The problem is that when you drive out of the mapped area, performance takes a big hit. I'm sure that's why you'll never see a Waymo system in a car sold to the general public, and why Tesla has avoided using lidar and HD maps altogether. Their use will never produce a general autonomous driving solution.
Yes, I see using both systems instead of putting expensive LiDAR in every car put them in complex road environments like urban intersections and relay that data to cars in the area.
That "machine learning" you talked about was actually Genetic Algorithm which is only one of the many algorithms used in machine learning.
Ultimately Tesla's FSD will be the safest. Other carmakers will use Tesla's technology.
Exactly as AWS is powering half of the internet right now.
lets be honest if you had seed the original tesla presentation, and if ur allaround informed about these things by lay vids, this vid added 0 value.
See any irony in your writing this? 🌱
I think Tesla is taking the smarter path... pushing ahead but safely. Waymo will probably succeed, but it's only going to take one big accident for Phoenix citizens to pull out the pitchforks and torches on them.
(01:35) It's 'diven' - dount u no how too speke English? Morun.
I think AI driving will be here soon, but mostly as a taxi. My argument for this is that this tech will be expensive for consumers for a while. Those who want a self driving car, are those who only want a vehicle to get from A to B. They usually look at the better deal when it comes to cars and EVs like Tesla are still at the luxury price. Not saying no one will buy them, but just that it will be hard for it to become mainstream unless prices go down to at least the 20ks
you described an evolutionary algorithm, while its a subfield of ML, the Tesla autopilot and neural net does NOT work like that. It uses backpropagation to learn, its not evolution.
Even a "layman" can see Musk is right.
You can not solve self driving without solving visual interpretation and deep learning.
Hence when you solve the visual bit lidar is not needed, pretty logic.
One thing never mentioned is the advantage once a majority of cars (even better all) you really do not have to interpret that much since the cars can talk to each other and have access to more sensors than just the ones on "your own" car.
Then traffic lights or even road signs won't be needed *FOR CARS*
The reason nobody wants to realize LIDAR is the wrong way to go is there is to much money invested in it.
LIDAR is faster to develop if you go with geofenced, GPS and high resolution maps.
No traffic lights and road signs for cars would be many years into the future when all traffic is automated.
All it takes is just one car which is not automated to (possibly) break the system, which means laws need to be passed that prohibit manual driving, at least in towns/cities.
They will probably put a ban on manual driving on all public roads eventually. It's been fun driving manually (most of the time) and I probably will for quite a few years longer.
But automated driving is the future, can't deny that. There's lots of advantages to it as well.
I'm legally blind, bring on the robot cars, plz someone figure this out
_Knight Rider_ had an episode in which a blind lady was driving KITT, but she wasn't really driving but it was a ruse to trick the bad guys into thinking that she may be a suitable witness, not actually being blind.
Of course KITT was pretty much at the highest level of self-drive, at least level 5, and so much AI that KITT seemed to have a fun personality.
I think on the DVD, some commentary or something on an episode or two, said there was a guy underneath a blanket who was driving KITT, and watch closely and you can briefly see a guy's hand closing or opening the car door. Gee, Tesla can open and close its doors, without any actor nor stunt-double's hand showing.
LiDAR has come substantially down in price. I think Garmin sells a basic unit for about 2-3 figures. You have to build your own rotary unit, though.
Stephan Brun true but moving parts fail. Reminds me of TV DLP technology, the picture quality was better than LED but the spinning color wheels always failed. Now DLP TV’s are a thing of the past. The same will come of LIDAR.
Lidar is way to expensive and unnecessary
When I was 4 my parents would sit me in their lap and let me steer. I was 7 when my mom asked me to shift the Volkswagen. By the time I got my drivers license, I'd driven (as a learner), almost 10,000 miles. We'd taken several trips out of state, up and down the coast and I drove the whole time. My relatives out west had been driving their farm truck since they could see over the dash.
Someday people won't even need licenses since they'll never drive. You'll just send your Tesla to pick up the kids at school.
You wont have to think about it, it's already there. Maybe kids just hail a taxi and it takes them home without saying a word or needing to find out any info. Maybe the school itself is at home and a MOOC or something.
That day is today if you live in a country with proper funding of public transport. Cars are becoming toys rather than transport.
Waymo is already an autonomous system in the hardiest environment that is city driving and for some time already. Looks like lidar was good after all. No waymo's car have ever kill someone or ram a fire truck, barriers and so on. They must be doing something good.
I have friends deep in the autonomy industry and it's interesting to see the perception of the general public vs people on the inside, behind the curtain of secrecy and IP.
Let's just say Tesla has a lot of holes in their logic, their timelines are astonishingly optimistic, and their definition of "full self driving" is very much not what most people would define it as
I too have friends in the autonomy industry, and they 100% believe that Tesla is far more on track with this than anyone else right now.
@@thetagmarket1058 I have friends who know nothing but think they know everything.
Your machine learning description is a bit off. Evolution is a specific type, but most machine learning uses gradient descent to minimize the error of the network. Other than that, great video!
@@EdNarculus Great video nevertheless.
I spend an hour and 40 minutes commuting every workday... I would love to have the car do the driving while I relax
Douggernaut84 same here - so... I just ordered a model 3 and really hope to be able to relax a little more while driving. Driving that far per day is just exhausting!
The real question is : Should we have a home made DIY hydrogen cell production in our backyards and get a hydrogen car?
You could get an electric car and do solar panels at home? Also lower chance of accidentally blowing up your house with a diy hydrogen production, storage and transfer setups that then connect to the car.
@@jameshall2223 Where's the fun in safety though?
Can’t wait, i hope my first car will be a fully self-driving one 😄