4 ms·

I wouldn't ever bet against Musk, but Musk says a lot of things. Yes we should be able to make self driving cars with just cameras because, after all, we do it

I wouldn't ever bet against Musk, but Musk says a lot of things. Yes we should be able to make self driving cars with just cameras because, after all, we do it with just two eyes. But in what world is computer vision perfect? Tesla's Autopilot already has a fatality on its record because it couldn't differentiate the color of the sky with the color of a truck. LIDAR solves a lot of problems.

virgilp 9y ago
There's another aspect of it though. As soon as Waymo makes self-driving cars more popular, and there are less humans on the road - you just need inferior technology to be able to operate. Bootstrapping this is the hardest - after self-driving cars are a thing, normally we'll see accelerated move towards fully automated driving. In the extreme case, if Waymo cars are 80%, Tesla cars can easily be 20% just based on the fact that Waymo cars will avoid most accidents made by Tesla software errors
- galdosdi 9y ago
  You might be right, but I'll devil's advocate a bit, with a few possible predictions (none of which I believe in all THAT strongly, just speculation) - If there's a period where new self driving cars are popular but aftermarket self driving kits for older cars are unaffordable or nonexistant, then during this period, the price of non-self-driving cars will plummet dramatically. Who knows, you might be able to buy a 2 year old camry for $1000. The market for these cars will exclusively be (other than oddballs/enthusiasts) people whose wages are so low that driving yourself in a non-self-driving car is still cheaper in an overall sense than using self-driving cabs, so that will cap the possible value of such cars at a pretty low point. - As long as (possibly as a result of the above) the number of human driven cars on the road is at all significant (like, at least 1% or 0.1% or something), the benefits you mention will be hard to attain. In a given trip a car interacts with hundreds of other cars. - You forgot about pedestrians and cyclists. In many places, there are more of those than there are cars, anyway, so even if all cars are self driving, - There's going to be weather and construction and falling trees and runaway skateboards and runaway trucks on ice covered hills and wildfires and so on. This kind of software will operate at scale, so improbable events become important to handle. - Uber is probably dumb enough to hope your argument is true and think they can get away with half-assed self-driving, which is why they're going to fail miserably and go bankrupt. - There's going to be regulation. As a made up example, if the "Waymo patent tax" is "just" $5000 per car, it's going to be pretty hard for a car manufacturer to justify to regulators that they'll instead develop competing software that is somewhat inferior in safety, but "acceptable", for the sake of saving, say, half of that $5000 (it still costs the thrifty manufacturer money to develop or buy their lower quality software!). All it'll take is one fatality and people will be up in arms. If the "acceptable" technology was the first one out, it'd be different, but it'll be hard to convince anyone it's worth the risk when there's better tech already out there. BTW, Waymo management would obviously be clever about setting a good price point that captures lots of value but still is low enough to discourage this kind of competition. - My vague software and driving intuitions tell me it'd actually be easier to design a more general system anyway (massive handwaving). It's actually HUMANS (specifically those accustomed to easy suburban western driving conditions) who have a harder time driving when they can't assume everyone will follow the rules, not computers, and I think that's why so many people mistakenly (IMHO) imagine it'd be so easy to make self driving cars if only everyone followed the damned rules. The simplified laws of physics applicable to driving are simpler than the DMV manual anyway.
drcross 9y ago
>Tesla's Autopilot already has a fatality on its record because it couldn't differentiate the color of the sky with the color of a truck. Do you think in your wildest dream that this exact scenario would ever be allowed to happen again by Tesla engineers? It may have crashed causing a fatality in this instance but it doesn't mean the prospect of camera only vision systems need to be written off. What people often overlook is that these systems get stronger with more and more edge cases and objectively the Tesla team are doing outstanding work with the sensor package that they have.
dllthomas 9y ago
> But in what world is computer vision perfect? I'd guess the same world where human vision is perfect...

virgilp 9y ago

There's another aspect of it though. As soon as Waymo makes self-driving cars more popular, and there are less humans on the road - you just need inferior technology to be able to operate. Bootstrapping this is the hardest - after self-driving cars are a thing, normally we'll see accelerated move towards fully automated driving. In the extreme case, if Waymo cars are 80%, Tesla cars can easily be 20% just based on the fact that Waymo cars will avoid most accidents made by Tesla software errors

galdosdi 9y ago
You might be right, but I'll devil's advocate a bit, with a few possible predictions (none of which I believe in all THAT strongly, just speculation) - If there's a period where new self driving cars are popular but aftermarket self driving kits for older cars are unaffordable or nonexistant, then during this period, the price of non-self-driving cars will plummet dramatically. Who knows, you might be able to buy a 2 year old camry for $1000. The market for these cars will exclusively be (other than oddballs/enthusiasts) people whose wages are so low that driving yourself in a non-self-driving car is still cheaper in an overall sense than using self-driving cabs, so that will cap the possible value of such cars at a pretty low point. - As long as (possibly as a result of the above) the number of human driven cars on the road is at all significant (like, at least 1% or 0.1% or something), the benefits you mention will be hard to attain. In a given trip a car interacts with hundreds of other cars. - You forgot about pedestrians and cyclists. In many places, there are more of those than there are cars, anyway, so even if all cars are self driving, - There's going to be weather and construction and falling trees and runaway skateboards and runaway trucks on ice covered hills and wildfires and so on. This kind of software will operate at scale, so improbable events become important to handle. - Uber is probably dumb enough to hope your argument is true and think they can get away with half-assed self-driving, which is why they're going to fail miserably and go bankrupt. - There's going to be regulation. As a made up example, if the "Waymo patent tax" is "just" $5000 per car, it's going to be pretty hard for a car manufacturer to justify to regulators that they'll instead develop competing software that is somewhat inferior in safety, but "acceptable", for the sake of saving, say, half of that $5000 (it still costs the thrifty manufacturer money to develop or buy their lower quality software!). All it'll take is one fatality and people will be up in arms. If the "acceptable" technology was the first one out, it'd be different, but it'll be hard to convince anyone it's worth the risk when there's better tech already out there. BTW, Waymo management would obviously be clever about setting a good price point that captures lots of value but still is low enough to discourage this kind of competition. - My vague software and driving intuitions tell me it'd actually be easier to design a more general system anyway (massive handwaving). It's actually HUMANS (specifically those accustomed to easy suburban western driving conditions) who have a harder time driving when they can't assume everyone will follow the rules, not computers, and I think that's why so many people mistakenly (IMHO) imagine it'd be so easy to make self driving cars if only everyone followed the damned rules. The simplified laws of physics applicable to driving are simpler than the DMV manual anyway.

drcross 9y ago

>Tesla's Autopilot already has a fatality on its record because it couldn't differentiate the color of the sky with the color of a truck. Do you think in your wildest dream that this exact scenario would ever be allowed to happen again by Tesla engineers? It may have crashed causing a fatality in this instance but it doesn't mean the prospect of camera only vision systems need to be written off. What people often overlook is that these systems get stronger with more and more edge cases and objectively the Tesla team are doing outstanding work with the sensor package that they have.

dllthomas 9y ago

> But in what world is computer vision perfect? I'd guess the same world where human vision is perfect...