5 ms·

It doesn’t take a lot to retro-fit roads with autonomous guardrail markers when we repave. It’s going to take a federal-level effort to accomplish this (which s

by lvl102 3y ago

It doesn’t take a lot to retro-fit roads with autonomous guardrail markers when we repave. It’s going to take a federal-level effort to accomplish this (which should have been apart of pandemic infrastructure stimulus). I strongly believe this would improve safety for autonomous vehicles by 1000x.

justapassenger 3y ago
Detecting lane is mostly solved problem. Detecting AND predicting what other things will do on the road (not only cars, but people, animals, trash, etc) is not.
- pfannkuchen 3y ago
  > predicting what other things will do on the road… [like] trash At least we do have a model for this one (physics). I wonder what additional signals the car would need to accurately simulate trash movement? Wind?
  - justapassenger 3y ago
    Wind, yes. But even without it - thing get tricky, when you’re traveling at 80mph. You need to detect enough of movements to be able to predict future path. It not easy, even with very advanced sensors. And you cannot “play it safe and assume it’ll collide with us”. You get phantom breaking, which is very dangerous.
  - MattRix 3y ago
    I don’t think you could ever accurately simulate trash movement. Trash is irregularly shaped, even with the highest resolution cameras, it’d be impossible to know what the far side of the trash is shaped like. Then on top of that there’s no way to predict highly localized gusts of wind. The best way to deal with trash is to just have an ML model deal with lots of it so it can make predictions about what it is likely to do (which is basically what Tesla is already doing).
- ikekkdcjkfke 3y ago
  Have you seen the video where a tesla tries to hard left into the guard rail on a freeway because of construction/redirect lines going to the left
aedocw 3y ago
Lane keeping is not the hard part, that’s been solved for a while. It’s dealing with pedestrians and other drivers doing unpredictable things. Slow moving infrastructure projects will not help that at all.
zaroth 3y ago
I used to think that V2X road infrastructure was a major necessary component for self driving. I no longer hold that opinion. The “self-driving” infrastructure would not be magically any better than the “human-driving” infrastructure, and how to handle conflicting data from the infra in terms of sensor fusion is not at all clear. In short, it creates as many problems as it solves, and doesn’t really solve the problems that it sets out to in the real-world. Even simpler things like supplemental signaling that could be used in special circumstances like road work and emergency vehicles, if they are not used properly and consistently in all cases, it doesn’t change the fact that the system needs to correctly handle these cases internally.
- lvl102 3y ago
  I definitely agree state of the art is good enough for 97% of driving scenarios but they’re also not fail-safe and robust. In order for autonomous to be truly acceptable, they need to be 100x more safe. No, nearly flawless. In order to accomplish that you need physical and hardcoded methods as well. In addition, and equally as important, is that road infrastructure gives you improved ability to coordinate cars on the road without relying on individual and disparate compute units.
kbos87 3y ago
There’s just no need for this. Tesla can detect even the most poorly defined edges of a dirt road without a problem.
- belter 3y ago
  Maybe they should start at detecting children? Luminar Tech can... - https://youtu.be/3mnG_Gbxf_w https://youtu.be/3mnG_Gbxf_w

by lvl102 3y ago

justapassenger 3y ago

Detecting lane is mostly solved problem. Detecting AND predicting what other things will do on the road (not only cars, but people, animals, trash, etc) is not.

pfannkuchen 3y ago
> predicting what other things will do on the road… [like] trash At least we do have a model for this one (physics). I wonder what additional signals the car would need to accurately simulate trash movement? Wind?
- justapassenger 3y ago
  Wind, yes. But even without it - thing get tricky, when you’re traveling at 80mph. You need to detect enough of movements to be able to predict future path. It not easy, even with very advanced sensors. And you cannot “play it safe and assume it’ll collide with us”. You get phantom breaking, which is very dangerous.
- MattRix 3y ago
  I don’t think you could ever accurately simulate trash movement. Trash is irregularly shaped, even with the highest resolution cameras, it’d be impossible to know what the far side of the trash is shaped like. Then on top of that there’s no way to predict highly localized gusts of wind. The best way to deal with trash is to just have an ML model deal with lots of it so it can make predictions about what it is likely to do (which is basically what Tesla is already doing).
ikekkdcjkfke 3y ago
Have you seen the video where a tesla tries to hard left into the guard rail on a freeway because of construction/redirect lines going to the left

aedocw 3y ago

Lane keeping is not the hard part, that’s been solved for a while. It’s dealing with pedestrians and other drivers doing unpredictable things. Slow moving infrastructure projects will not help that at all.

zaroth 3y ago

I used to think that V2X road infrastructure was a major necessary component for self driving. I no longer hold that opinion. The “self-driving” infrastructure would not be magically any better than the “human-driving” infrastructure, and how to handle conflicting data from the infra in terms of sensor fusion is not at all clear. In short, it creates as many problems as it solves, and doesn’t really solve the problems that it sets out to in the real-world. Even simpler things like supplemental signaling that could be used in special circumstances like road work and emergency vehicles, if they are not used properly and consistently in all cases, it doesn’t change the fact that the system needs to correctly handle these cases internally.

lvl102 3y ago
I definitely agree state of the art is good enough for 97% of driving scenarios but they’re also not fail-safe and robust. In order for autonomous to be truly acceptable, they need to be 100x more safe. No, nearly flawless. In order to accomplish that you need physical and hardcoded methods as well. In addition, and equally as important, is that road infrastructure gives you improved ability to coordinate cars on the road without relying on individual and disparate compute units.

kbos87 3y ago

There’s just no need for this. Tesla can detect even the most poorly defined edges of a dirt road without a problem.

belter 3y ago
Maybe they should start at detecting children? Luminar Tech can... - https://youtu.be/3mnG_Gbxf_w https://youtu.be/3mnG_Gbxf_w