[theontheone]

https://twitter.com/elonmusk/status/1387901003664699392

Elon: "A major part of real-world AI has to be solved to make unsupervised, generalized full self-driving work, as the entire road system is designed for biological neural nets with optical imagers"

Seems like he's inching closer to reality by the day.

[blueblisters]

Or I could have used better sensors on my car but that doesn't sound as controversial as "anyone using LIDAR is doomed" so I fooled my customers and let my competitors get ahead.

[tr1ll10nb1ll]

Saying "anyone using LiDAR is doomed" is a controversial statement sure but the self-driving problem is indeed a computer vision problem.

George Hotz puts it quite neatly, companies like Waymo have spent millions so far and someone sitting at home could not use their product highlighting their inefficiency. He further states to think about how a human would drive a car, does a human need LiDAR to drive a car? No, not really. The legitimacy in his statements come from the fact ( imo ) that comma.ai ( his self-driving startup ) outperforms with less funding than one could imagine by solving the problem absolutely as a Computer Vision problem. They treat it slightly different than Tesla but in all, it goes on to show that you don't really need to use LiDAR to solve the problem, it can potentially solve the problem, but it's an inefficient way to do it.

[blueblisters]

> He further states to think about how a human would drive a car, does a human need LiDAR to drive a car?

I hate this argument. It doesn't hold for several reasons:

- A human driver has blind spots (which btw Tesla avoids by having cameras surrounding the car)

- Human reactions times are very poor at even moderate speeds

- Human vision is affected by glare

- Humans are terrible at driving in low light conditions

- Humans are terrible at driving in inclement weather - snow storms or heavy rainfall

LIDAR doesn't solve many of these problems but to claim that self driving is solvable by vision alone is a bad take. Multiple sensors add redundancies and reduce errors. Teslas are already equipped with ultrasound sensors and radar so it's not technically solving a Computer Vision only problem.

[atatatat]

> LIDAR doesn't solve many of these problems

Radar+sonar do.

[tr1ll10nb1ll]

It's never about being just able to solve the problems, it's more about solving the problem without spending millions and millions of dollars and still not having a product everyone can use after years.

[tucosan]

Can you give an example of competitors that use Lidar and are further ahead than Tesla?

[colejohnson66]

Doesn’t Waymo use LIDAR? They’re not as widespread as Tesla is, but some suggest they’re better than Teslas.

[kosievdmerwe]

Waymo also has actual robotaxis.

[avereveard]

Waymo also hasn't killed anyone

[matt-attack]

Furthermore the entire road system is designed to work with simply two rotating cameras primarily facing forward.

[gremlinsinc]

Seems to me that maybe smart roads would be better? If there's sensors that communicate back and forth provide additional wifi boosts for cars on the grid, etc... L5 could come via cheating...expecting a car to do everything itself is going to be rough... but when a street can give you data about everything on it -- pedestrians, bikes, etc...it'd be easier to build to l4.9 automation...

Musk needs to team up with the Biden Administration and work in tandem on infrastructure - that may be the only way we see self-driving road-ready cars before 2040.