Exploring If Nvidia Can Beat Tesla's Self-Driving Cars

Elon Musk is the world's richest man because he's close to making self-driving cars a reality at their most recent AI day Tesla showed the world just how far ahead they are in terms of having the best sensors the most data and the most advanced self-driving cars but there's just one problem Elon Musk isn't racing against any one company thanks to Nvidia he's racing against them all Hyperion 8 can achieve full self-driving with a 360 degree camera radar lidar and ultrasonic sensor Suite Hyperion a will ship in Mercedes-Benz cars starting in 2024 followed by Jaguar Land Rover in 2025 Nvidia is the world's leading AI company and they just announced some insane leaps in technology that could make the race for full self-driving much closer than you think the amount of new technology in Thor is insane want to know who's really winning you're in the right place your time is valuable so let's get right into it let me start by saying that this isn't about Tesla or Nvidia somehow being bad and I'm not rooting for either of them to fail I've made video after video talking about the incredible engineering going on at both companies and all the different ways that they're transforming our daily lives and make no mistake self-driving cars will be one of the biggest transformations in our lifetimes think about how much time will get unlocked when every single driver can choose to relax or be productive in a car instead of always having to drive now think about being the first company to unlock that time for people while every other car still needs somebody to drive it the winner of this race just created many trillions of dollars in markets that only they will have access to that's the massive opportunity that both Tesla and Nvidia are racing for here's the thing full self-driving also comes with a lot of serious challenges challenges that can only be solved with real world artificial intelligence Tesla just held their second AI day where they showcased some of these Solutions but just 10 days later Nvidia held their GPU Tech conference or GTC where they showed off some exciting Solutions of their own that's why I want to take a look at just how close these two companies are to solving full self-driving today there are three main areas that I want to compare how each company processes and labels real world data how they train their self-driving AIS and what they do for testing and simulations let's start with data labeling data labeling is the step where the AI tries to figure out what every object that it can see actually is is it a road that the car can drive on label it purple is it a sidewalk or another flat surface that the car probably shouldn't drive on label it red is it another car label it blue and so on it's important to get these data labels right because a computer can't really understand what to do if it doesn't know what it's looking at earlier this year Tesla made a huge change by removing Radars from their cars and optimizing everything for video and I do mean everything from the software and Hardware inside the cars to their research and development teams and even their data simulators and supercomputers at their most recent AI day Tesla showed off some pretty cool techniques that they're using to help the car label almost everything it can see in real time for example Tesla can reconstruct a scene using data from multiple trips across multiple cars they can build maps from all the previous trips that a Tesla has made in any given area and see how those Teslas labeled their surroundings since all the video data is passed through the same labeling systems every Tesla should agree when they're seeing the same thing then the current Tesla should also agree on those labels as long as it knows where it actually is in this reconstructed scene and that's why what all these colored wines are in these videos pre-computed labels Based on data from other Teslas then our Tesla updates those labels for the next guy and so on we should also talk about lanes because unlike other types of objects in the scene lanes are part of the road itself when split and merge and the lines on the road don't show all the possible ways that traffic can move between these Lanes that's why just labeling each lane as purple isn't really good enough for the AI so what Tesla did was develop a separate system just for describing Lanes turns intersections and the road Network that they form now combine that with what we just talked about where Tesla uses labels from other cars when a Tesla rolls up to a busy intersection it should already know all the paths that it can take just by knowing where it is exactly even in limited visibility then as our Tesla goes through the intersection it updates those labels for each lane for every Tesla that comes after just like before that's all these teal lines in these clips the crazy thing about this is how much Tesla has been able to automate in a world where most data labeling by other companies is still being done by humans according to Tesla labeling 10 000 trips would take a human around 5 million hours to do or about 570 years it takes Tesla's Auto aboard just 12 hours to label those same 10 000 trips when Tesla shared those stats I thought it was game over for companies like Nvidia who still label their data by hand for ground true data we use our deepmap HD mapping human labeled data and Omniverse replicator and video will never catch Tesla if they can't label enough data to train their AI right well it turns out that they've been making serious strides in that department as well Nvidia went from Human labeled data to automatically labeling everything in 3D space and applying those labels to each camera just like Tesla but wait there's more just like Tesla Nvidia is able to take in sensor data digitally reconstruct entire scenes label everything in those scenes and then update those labels with each New Journey in the real world see what I mean the race for full self-driving is closer than we think and while Tesla uses nothing but cameras nvidia's newest system called Hyperion 9 will support up to 14 cameras 9 Radars three lidars and 20 ultrasonic sensors this is how Nvidia will provide self-driving as a platform which will allow the rest of the Auto industry to crowdsource many of the challenges that Tesla is tackling alone that's why Tesla is in a race against every single car company not just any one of them alright let me point out a few things here first it's not totally clear which sensors and video relies on to extract all these features was the scene reconstructed using only cameras or did Nvidia need lidar or radar data to achieve these amazing results if nvidia's platform relies on much more expensive sensors than Teslas they really won't be able to compete on price which will mean more Teslas on the road over time collecting more data that other companies aren't getting and sec second Tesla already has millions of cars on the road driving and collecting billions of miles of data today while Nvidia system is just starting to be rolled out to other automakers hyperion8 will ship in Mercedes-Benz cars starting in 2024 followed by Jaguar Land Rover in 2025. that really limits the amount of real world data that Nvidia can use to train their labeling algorithms and their self-driving AI That's why they're still behind Tesla when it comes to data notice how noisy and jittery these labels are even for stationary objects like curbs and parked cars that's one of the symptoms of not having enough training data alright I spent a lot of time on sensor data and labeling because the rest of full self driving really Builds on top of that for example Tesla uses that information to decide what parts of the scene to pay attention to Tesla's AI chunks its surroundings up into tiny cubes then it assigns each cube of color based on the potential movement of whatever is occupying that Cube So Tan cubes can contain buildings or mailboxes or Lamppost but what they all have in common is that those volumes are occupied with things that won't move red volumes are occupied by things that could move like a pedestrian or a parked bus and once the bus in the scene starts to move it turns blue indicating that the Tesla predicts that it will continue to move breaking the world up into cubes like this lets Tesla spend computer power on the volumes that matter more like the ones colored red or blue while not wasting resources trying to predict the motion of things in the tan cubes which won't end up moving at all then Tesla's AI predicts the paths of everything in the red or blue volumes and adjusts its own route in real time in the case of somebody running a red light that means recognizing that one potential path is them Crossing through our lane and being ready to act accordingly as soon as this car's nose turns into Arlene the Tesla applies some brakes and turns to avoid the car in a separate case somebody is parked in the left lane of this road while the person on the right is slowing down at a light because the Tesla could correctly classify the car on the left as parked and not just momentarily stopped it knew to switch lanes instead of staying behind this parked car forever it was pretty mind-blowing to see Tesla's be able to predict the motion of everything that they could see at least when they first showed this feature off but now we're starting to see the same kind of thing from Nvidia where their AI can anticipate the behavior of other vehicles and adapt accordingly by the way I'm not saying that nvidia's route planner is anywhere near as advanced as Tesla's all I'm pointing out here is that the gap between them is starting to close just like with their automatic data labeling systems but one massive advantage that Tesla does have is the insane amount of real world data that they've already collected and labeled because they have so many cars already collecting that data Tesla can quickly find all the footage where something very specific happens see all the ways that the driver intervened and then use that footage to retrain their self-driving AI to handle each specific case even if the driver doesn't have self-driving enabled the FSD computer can compare what it would have done to what the driver actually did and then send that data back to Tesla for analysis Nvidia obviously can't really do that yet since their system is still being rolled out to different brands so how can Nvidia even begin to close this massive real world data Gap the answer is by simulating that data while simulated data will never be as good as the real thing almost nobody beats Nvidia when it comes to making realistic simulations that's basically their entire business Nvidia is solving the simulation challenge in multiple ways first they have Drive map which is a crowdsourced real-time digital twin of every road that they get data for kind of like how we saw with Tesla's scene reconstructions using multiple trips worth of data by 2024 Nvidia expects to map every major highway in North America Western Europe and Asia that's hundreds of thousands of miles of physical roads that will have digital twins so if Nvidia knows where a car is in real life they'll also know where it is inside this digital twin which means they'll know where all the road lines are and the traffic whites and the curbs and so on then the Hyperion sensors in the car can double check what they're seeing against that digital twin Nvidia also has Drive Sim which can take in recorded Drive data turn everything in the scene into an interactive object and harvest every object for reuse in other scenes Drive Sim also lets Nvidia remove add or change anything in the recording and rerun a bunch of simulations in that environment or they can change the environment entirely by adding different lighting effects and weather conditions and road hazards and so on Nvidia already does these things in many different markets today for example gaming nvidia's Drive map and drive SIM can also simulate what different sensors would see in a given scenario so for example they can simulate a city street and then show what a lidar would see and simulate the depth map that it would return this means that they can train their self-driving AI on lidar data even though they didn't collect it in the real world and because Drive Sim and drive map are tied to nvidia's Omniverse many different companies can work together to improve these features over time and catch up to Tesla but just because they can doesn't mean they will like I said earlier Tesla really streamlined everything by removing radar from the equation since the whole system is Vision only they don't need to simulate radar data at all when they want to generate new data that the fleet can't provide instead they use a photorealistic game engine to simulate scenes and traffic with a high degree of visual accuracy a busy City intersection might take two weeks for artists and game creators to build out and populate with realistic traffic but Tesla is able to do that in just 10 minutes using all the data and the labels and the AI models that we just talked about that's pretty crazy and on top of that they can change the vehicle and pedestrian traffic patterns to train and test the AI in different situations within the same location or they can change the surrounding trees and buildings to generate similar data sets increasing the amount of data they have for a specific scenario that they might not have seen enough enough times in real life taking it one step further they can also randomize the weather and the visibility to measure the ai's performance in low light poor visibility or when water is causing unusual Reflections in the road and of course since this is a simulation Tesla can turn on and off labels depending on whether they want to use data for training or testing ultimately Tesla can use this simulator to create an up-to-date virtual world filled with every road that they've ever mapped complete with accurate labels and now to come full circle because we generated all these tile sets from ground truth data that contain all the weird intricacies from The Real World and we can combine that with the procedural Visual and traffic variety to create Limitless targeted data for the network to learn from it's worth noting that only Tesla has access to this data since it's their cars collecting it Nvidia can't just easily copy or recreate it this is what really impressed me the most about Tesla they seem to be just as good at making these large-scale digital twins of Road networks as Nvidia a company that's been doing this kind of stuff for decades meanwhile since the Hyperion system is a self-driving platform for others to build on Nvidia can only move as fast as the car companies that end up using it want to know how bad Tesla is crushing these other car companies check out this episode next and if you feel I've earned it consider hitting those like And subscribe buttons which lets me know to make more content like this either way thanks for watching and until next time this is ticker symbol U my name is Alex reminding you that the best investment you can make is in you

Channel: Ticker Symbol: YOU

Views: 26,029

Keywords: Tesla Stock, TSLA, TSLA stock, NVDA, Nvidia stock, NVDA stock, nvidia stock analysis, tesla stock analysis, tsla stock news, nvidia gtc 2022, Elon Musk, jensen huang, tesla fsd, ai stock, best stocks to buy now, nvidia vs tesla, tesla stock tomorrow, tsla stock tomorrow, nvda stock tomorrow, nvidia stock tomorrow, tesla stock news, nvda stock analysis, artificial intelligence stocks, computing stocks, robotics stocks, nvidia omniverse, nvidia omniverse explained

Id: ko5VrUnqiQ0

Length: 14min 10sec (850 seconds)

Published: Fri Oct 28 2022