As the tentative launch date for Tesla's Robotaxi service approaches, more details are emerging, including the presence of a human safety driver and other key aspects.
That’s because Tesla self-driving takes a different, and imo way worse, approach.
Waymo relies on mapping, the entire city is basically 3D modelled and loaded into the car memory. It’s more or less ‘on rails’.
It also uses LIDAR for live data alongside imaging cameras, again building a 3D model of its environment combined with image recognition.
Tesla decided that, for some reason, they want their cars to drive ‘like humans’, only relying on vision and deployable anywhere, without pre-mapping.
Demanding a computer to behave like humans, instead of using a computer’s strengths, seems like a very poorly thought out move to me.
Arguably mapping out cities to this degree across the globe is a ginormous effort, on an order of magnitude more so than what Google Maps etc. currently provide. Thus I don’t think it’s entirely unreasonable to try designing something that operates purely in terms of sensory input (and of course map data where available, those approaches don’t have to be mutually exclusive).
drive ‘like humans’, only relying on vision […] without pre-mapping.
Demanding a computer to behave like humans
So basically their taxis will go into the job of driving a taxi without any prior knowledge of the city? Like a human? Only relying on road signs? Will it also stop to ask for directions? Like wtf? What kind of stupid idea is this from Tesla. Sounds absolutely moronic.
A human taxi driver doesn’t work like this. They are people who know the city very well going in, or at least used to before GPS navigation in vehicles came to aid.
He’s famously said that humans are good enough with just our two eyes. So he went the cheap route of not including lidar and relying on stereoscopic cameras.
He’s an idiot. Because when I want stuff automated, I want it to be better than what a human can do.
This requires the car to have an equally sophisticated visual cortex as well, which we cannot achieve right now. Cars need those better sensors to equal our abilities.
God, hearing his stupid arguments shows us he really knows nothing about anything.
Because he seems to base his ideas purely on how cool he thinks they are rather than practicality and efficiency (such as the hidden tesla door handles).
GPS is knowledge of the city. Most cities publish detailed street map data that is immediately imported into gps maps. The car relies on the gps not just for navigating, but even deciding things like what lane. And of course gps’s have much more complete traffic data than your human taxi driver ever did.
Does the taxi driver remember every sign in the city, every road and parking spot? No. They are humans - they remember the streets, some important spots that are confusing, maybe a couple of shortcuts. There is a huge difference between having a 3d map of everything in the city in the memory, and setting a GPS to an address, reading the signs as you go by and adhering to them. Also if self driving tech is to expand, you don’t go putting the entire world into memory - that’s not scaleable.
Does the taxi driver remember every sign in the city, every road and parking spot? No.
having a 3d map of everything in the city in the memory
you don’t go putting the entire world into memory
Cars don’t do this either, do they? Surely this type of data is streamed as needed. Just like video games do. This type of optimization has been around for decades… We need not worry about that in cars either.
I’m just saying that GPS and LIDAR is needed in addition to just camera input.
No argument on gps and lidar from me. Streaming doesn’t work. You are probably thinking about Microsoft flight sim, which completely fails (and is the first “completely streamed” map). Out in the “real world”, you don’t have a fiber connection to stream gigabytes every hour.
Streaming doesn’t work. You are probably thinking about Microsoft flight sim, which completely fails (and is the first “completely streamed” map).
Streaming might be the wrong word. I’m talking about loading just enough data to do what’s relevant right now. And I’m not talking about full 3D geometry of the world, that’s not helpful to a vehicle. It needs to know in 2D where it can and cannot drive, as well as real non-static/dynamic obstacles (what the cameras and LIDAR are for).
You don’t need gigabytes of data to load a 2D geometry of a small area like a part of a city, surely. You can also cache it on disk. Your phone can even do this. In fact, you can ask it to cache however wide of an area as you’d like. You might hit several hundred megabytes but that’s like a whole midsize city probably.
While I don’t want to fanboi too much …… as long as nobody is able to do self driving, any approach has potential. Teslas approach has the huge advantage of starting with millions of potential vehicles and they will soon be able to crank out a quarter million robotaxis every year, whereas Waymo is not ready to scale up. They’re going all in with a potential approach and if it succeeds are in position to disrupt the industry.
It might not succeed, but no other company has succeeded yet either. The difference is they went all in, and they were willing to try something different than Internet wisdom
(Actually, as a big fan of what Tesla used to represent, this worries me about their future: Cybertruck flopped, robotaxi may not succeed for years, semi is a very conservative market they may not be able to break into or will be slow take up , and Optimus doesn’t yet have a market. They’re making some extremely risky moves at the same time their profit margins are under a lot of pressure. )
That’s because Tesla self-driving takes a different, and imo way worse, approach.
Waymo relies on mapping, the entire city is basically 3D modelled and loaded into the car memory. It’s more or less ‘on rails’. It also uses LIDAR for live data alongside imaging cameras, again building a 3D model of its environment combined with image recognition.
Tesla decided that, for some reason, they want their cars to drive ‘like humans’, only relying on vision and deployable anywhere, without pre-mapping.
Demanding a computer to behave like humans, instead of using a computer’s strengths, seems like a very poorly thought out move to me.
Arguably mapping out cities to this degree across the globe is a ginormous effort, on an order of magnitude more so than what Google Maps etc. currently provide. Thus I don’t think it’s entirely unreasonable to try designing something that operates purely in terms of sensory input (and of course map data where available, those approaches don’t have to be mutually exclusive).
So basically their taxis will go into the job of driving a taxi without any prior knowledge of the city? Like a human? Only relying on road signs? Will it also stop to ask for directions? Like wtf? What kind of stupid idea is this from Tesla. Sounds absolutely moronic.
A human taxi driver doesn’t work like this. They are people who know the city very well going in, or at least used to before GPS navigation in vehicles came to aid.
He’s famously said that humans are good enough with just our two eyes. So he went the cheap route of not including lidar and relying on stereoscopic cameras.
He’s an idiot. Because when I want stuff automated, I want it to be better than what a human can do.
This requires the car to have an equally sophisticated visual cortex as well, which we cannot achieve right now. Cars need those better sensors to equal our abilities.
God, hearing his stupid arguments shows us he really knows nothing about anything.
Because he seems to base his ideas purely on how cool he thinks they are rather than practicality and efficiency (such as the hidden tesla door handles).
Very stupid.
GPS is knowledge of the city. Most cities publish detailed street map data that is immediately imported into gps maps. The car relies on the gps not just for navigating, but even deciding things like what lane. And of course gps’s have much more complete traffic data than your human taxi driver ever did.
Exactly!
Does the taxi driver remember every sign in the city, every road and parking spot? No. They are humans - they remember the streets, some important spots that are confusing, maybe a couple of shortcuts. There is a huge difference between having a 3d map of everything in the city in the memory, and setting a GPS to an address, reading the signs as you go by and adhering to them. Also if self driving tech is to expand, you don’t go putting the entire world into memory - that’s not scaleable.
Cars don’t do this either, do they? Surely this type of data is streamed as needed. Just like video games do. This type of optimization has been around for decades… We need not worry about that in cars either.
I’m just saying that GPS and LIDAR is needed in addition to just camera input.
No argument on gps and lidar from me. Streaming doesn’t work. You are probably thinking about Microsoft flight sim, which completely fails (and is the first “completely streamed” map). Out in the “real world”, you don’t have a fiber connection to stream gigabytes every hour.
Streaming might be the wrong word. I’m talking about loading just enough data to do what’s relevant right now. And I’m not talking about full 3D geometry of the world, that’s not helpful to a vehicle. It needs to know in 2D where it can and cannot drive, as well as real non-static/dynamic obstacles (what the cameras and LIDAR are for).
You don’t need gigabytes of data to load a 2D geometry of a small area like a part of a city, surely. You can also cache it on disk. Your phone can even do this. In fact, you can ask it to cache however wide of an area as you’d like. You might hit several hundred megabytes but that’s like a whole midsize city probably.
While I don’t want to fanboi too much …… as long as nobody is able to do self driving, any approach has potential. Teslas approach has the huge advantage of starting with millions of potential vehicles and they will soon be able to crank out a quarter million robotaxis every year, whereas Waymo is not ready to scale up. They’re going all in with a potential approach and if it succeeds are in position to disrupt the industry.
It might not succeed, but no other company has succeeded yet either. The difference is they went all in, and they were willing to try something different than Internet wisdom
(Actually, as a big fan of what Tesla used to represent, this worries me about their future: Cybertruck flopped, robotaxi may not succeed for years, semi is a very conservative market they may not be able to break into or will be slow take up , and Optimus doesn’t yet have a market. They’re making some extremely risky moves at the same time their profit margins are under a lot of pressure. )