Autonomous Car Progress

[daniel]

I thought the goal was L5?

Also what do you mean by "generalized L4"? L4 is autonomous driving but limited in when or where it can be used, such as geofencing, weather restricted etc... So do you mean generalized in the sense that it is vision-only with no HD maps but L4 because it will be limited? Are you saying Tesla will limit when or where we can use FSD because that is what L4 implies?

I cannot speak for powertoold, but to me, "generalized L4" would mean L4 with very broad ODD. Perhaps limited to well-maintained roads and excluding extreme weather. "Generalized L4" would suggest to me an autonomous car that people all around the country could buy, and most of the time it would operate autonomously with only very rare calls for the driver to wake up and take over. As distinct from a robotaxi that only operates in a very limited area.

Even if Tesla has a reliable L5 robotaxi, I don't think it's a good idea to enable *all* routes / locations for the public. Just as people do stupid things with AP, they'll do stupid things with a L5 robotaxi. You can use your imagination here.

L5 by definition does not need to have any means for a human to operate it. It is by definition capable of driving anywhere (presumably anywhere a human might drive) so I cannot think of anything "stupid" a person might do with it that only involves getting from one place to another. The L5 car should be able to safely transport a person between any two points connected by adequate roads. Unless you mean it should avoid "undesirable" parts of town?

ETA: JHCCAZ beat me to it.

 

[FloridaJohn]

The progenitors of the field of AI thought they could solve it over the summer in 1956. I don't think there's any field that produces a higher ratio of optimism to results.

Cold Fusion.

 

[smogne41]

Cold Fusion.

Not fair, that is dividing by zero!

 

[Microterf]

 

[jimm01]

I'm interesting in knowing how safe and reliable L5 Autonomy can be achieve without centralized monitoring and flow control (smart highway), a peer-to-peer network of ad hoc nodes, or some mixture of both is a final requirement. Knowledge of the belief-desire-intention of a vehicle, its maneuverability and performance, and human interaction (i.e., situational awareness, driver performance, improper or atypical response, etc.) need to be accounted for. Will any vehicle be capable of reliable and safe L5 Autonomy without a corroborative approach?

 

[Matias]

Keep in mind he also apparently didn't realize BTC used a lot of electricity until he'd already bought 1.5 billion dollars of it.

He does seem to approach virtually every problem he hasn't actually worked on as "Well that looks easy!"

Then, it turns out it's not.

Some have worked out really well in the end...(reusable rockets, mass producing EVs....) Some haven't turned out as well (battery swap stations... those freaky snake superchargers... over-reliance on automation in the factory...radar as a primary sensor )

The end result on FSD isn't known yet, but it's certainly more behind schedule than the other stuff that's turned out well so far.

Yes. On the other hand he’s extremely successful on difficult tasks, on the other hand he sometimes has wild hubris about the things he knows nothing. E.g. he was sure that you can make digging tunnels faster by factor 10 (or was it more), before he had even seen tunnel boring machine.

​

 

[J1mbo]

Using 1.2 MP webcams, with 2D radars from 2011/2014 that's only meant to be used for ACC due to its very very very very very low resolution and failure modes isn't going cut it.

Saying your vision system is the best and has gotten so good that it surpassed Radar and has rendered Radar obsolete when you were using radars from 2011.
Is like saying your internal smartphone is the best and has gotten so good that it surpassed the IPhone and has rendered IPhone completely obsolete when you were using the IPhone 3GS.

Its none-sense.



Get it through your head. NN and the advancements and breakthroughs of NN has nothing to do with Camera or Lidar.
The same NN are used on Lidar data. 100% of the Lidar stack is NN. Again get it through your head. Its not NN vs Lidar.




Yes pivot from a sensor that has gotten orders of magnitude better in quality, performance and price. Because a struggling auto company who in an attempt to cut costs put $5-$10 1.2 MP webcams and $60 2011 radar into their cars.

Let compare the quality, reliability and resolution of lidar that Google used in 2012 versus what's available today in the mass market and is going into dozens of cars in 2021 and 2022. Spoiler: its night and day.

Google used the $75k Velodyne HDL 64E, that needed constant maintenance and there were reports it broke down every month, this is in contrast to production lidars today that are specc'ed for 50,000+ hours (decades of driving) and tested for all conditions.

Velodyne HDL 64E - $75k

Key Features:​

• 64 lines
• 50m (10% reflectivity), 120m (80% reflectivity) range
• 360° Horizontal FOV
• 26.9° Vertical FOV
• 0.08° angular resolution (azimuth)
• <2cm accuracy
• ~0.4° Vertical Resolution

https://hypertech.co.il/wp-content/uploads/2015/12/HDL-64E-Data-Sheet.pdf

Luminar Iris - $500

Key Features:​

• 640 lines
• 500m max range
• 250m at <10% reflectivity
• 120° Horizontal FOV
• 30° Vertical FOV
• 0.07° horizontal resolution
• 1cm accuracy
• 0.03° Vertical Resolution
• Dust & Water Ingress, Vibration & Shock certified

Waymo shouldn't stop there. They should dump their State of the Art 4D imaging radar because Tesla deemed radar obsolete because their 2011/2014 ACC 2D radar sucks.

Right? @scottf200 @powertoold @ZeApelido @J1mbo

They should dump their HD 4k cameras aswell and yank out their camera cleaning solution aswell since Tesla doesn't have them.
Who needs those right? They only need to add the obsolete 1.2 MP cameras that Tesla has. Right guys?



In 2017 AP1 (based fully on Mobileye's EyeQ3) was still widely considered to be way better than AP. In-fact AP2 was still limited and hadn't gotten feature parity with AP1. Stop revising history.


How about Tesla actually releases a car that drives itself autonomously without any need for human supervision with safety performance that is greater than the best human driver?



Maybe because they were responsible for 99% of all Neural Network and Deep learning breakthroughs.
And your little Tesla and god Elon is responsible for 0%, nada, nil, none, zip.

You seriously don't think that Tesla (Karpathy) have been responsible for any NN/DL breakthroughs? What about OpenAI? Maybe they don't talk to each other

Anyway, we have another cool new twitter thread from @verygreen which shows the output from the depth NN, which seems pretty impressive for passive vision only:

https://twitter.com/x/status/1412597377228226562

Also cool to play with the single-frame point clouds:

https://twitter.com/x/status/1412613082224934919

He doesn't state which version these came from, so not clear if this is the same NN used by the radar-less cars and/or City Streets.

 

[Daniel in SD]

Not fair, that is dividing by zero!

Haha. True.
I should have said the difference between optimism and results, not the ratio.

 

[jimm01]

He doesn't state which version these came from, so not clear if this is the same NN used by the radar-less cars and/or City Streets.

Self-supervised may be a misnomer, maybe unsupervised? In any case, wouldn't some accurate means of measuring proximity/distance at certain intervals be required for the NN to self learn?

 

[verygreen]

He doesn't state which version these came from, so not clear if this is the same NN used by the radar-less cars and/or City Streets.

when not explicitly stated it could be assumed I am running current prod versions

 

[AlanSubie4Life]

He does seem to approach virtually every problem he hasn't actually worked on as "Well that looks easy!"

What is weird in this case is that he has been involved with this problem quite intimately since something like 2016, and he is apparently only just now realizing it is hard, in retrospect?

His optimism runneth over I guess. I guess I am too much of a negative Nancy. I guess no one would like me at a party, but for some reason I never get invited.

 

[mikes_fsd]

What is weird in this case is that he has been involved with this problem quite intimately since something like 2016, and he is apparently only just now realizing it is hard, in retrospect?

He is only now admitting it publicly, that is different from "realizing it".

 

[daniel]

The progenitors of the field of AI thought they could solve it over the summer in 1956. I don't think there's any field that produces a higher ratio of optimism to results.

Cold Fusion.

Nobody takes cold fusion seriously. Buckets of people take generalized AI seriously.

 

[MP3Mike]

What is weird in this case is that he has been involved with this problem quite intimately since something like 2016, and he is apparently only just now realizing it is hard, in retrospect?

Amazing that you have been following Elon/FSD for so long and you think this is the first time he has said it is hard. Here are a couple examples I found quickly that go back as far as 2018:

https://twitter.com/x/status/1048118186518835200

https://twitter.com/x/status/1353667962213953536

https://twitter.com/x/status/1382549229726498821

So this isn't a new realization for him.

 

[daniel]

His optimism runneth over I guess.

And that optimism is why we have Tesla and SpaceX today. But it's also why he's pursuing some ridiculous ideas like Hyperloop and Mars colonization. He thinks hard problems are easy so he throws himself into them and expects them to be solved soon. Without that optimism, we would not be driving electric cars today.

 

[powertoold]

Not presuming to answer for @powertoold, but to me "generalized L4" would mean ability to operate autonomously, i.e. to plan and complete requested trips on paved roads, with a high degree of confidence, in a broad market area like the USA.

Neighborhood or short-list of selected cities L4 doesn't seem (to me) to qualify as "generalized", until it becomes most neighborhoods in most cities & counties.

If Tesla deploys a robotaxi, it will likely only service 90-95% of the public roads in the US. I don't think it's a smart business decision to allow the public to use the service without any destination geofences, as there are higher risk areas and routes. It's unlikely any company will ever deploy a true, limitless L5 service.

 

[FloridaJohn]

Nobody takes cold fusion seriously. Buckets of people take generalized AI seriously.

I would posit that no one take cold fusion seriously because after decades of false promises, everyone stopped paying attention to the constant "cold fusion in 10 years" predictions.

However, there are still a few people that take cold fusion seriously, including a small company some of us may be familiar with:

"... cold fusion research has been funded by private and small governmental scientific investment funds in the United States, Italy, Japan, and India. For example, it was reported in Nature, in May, 2019, that Google had spent approximately $10 million on cold fusion research."

Conclusion: 10 years away! Joking, actual conclusion was:

"A group of scientists at well-known research labs (e.g, MIT, Lawrence Berkeley National Lab, and others) worked for several years to establish experimental protocols and measurement techniques in an effort to re-evaluate cold fusion to a high standard of scientific rigor. Their reported conclusion: no cold fusion."

 

[AlanSubie4Life]

Amazing that you have been following Elon/FSD for so long and you think this is the first time he has said it is hard. Here are a couple examples I found quickly that go back as far as 2018:

Poor memory for Elon maybe?

“Just now realizing” does NOT imply that I think it is the first time he has said so. (Nor did I say so!!!)

I am well aware he has talked about the difficulty before.

I think it is just that optimism coming into play. He sees the problem is hard, then comments on that fact. Then he keeps thinking that maybe it isn’t hard after all, maybe? Then gets reminded by new challenges that it is a hard problem?

Hard to analyze, really.

I’m not sure what to think of boundless optimism in this context. Elon is a great salesman, so I think he plays both sides to some extent.

 

[ItsNotAboutTheMoney]

And that optimism is why we have Tesla and SpaceX today. But it's also why he's pursuing some ridiculous ideas like Hyperloop and Mars colonization. He thinks hard problems are easy so he throws himself into them and expects them to be solved soon. Without that optimism, we would not be driving electric cars today.

And that optimism is why we have Tesla and SpaceX today. But it's also why he's pursuing some ridiculous ideas like Hyperloop and Mars colonization. He thinks hard problems are easy so he throws himself into them and expects them to be solved soon. Without that optimism, we would not be driving electric cars today.

I think that's a very naive view of Musk. I view his approach as:
(1) Identifying a difficult engineering problem with a large reward
(2) Sell stories of hope to investors and customers
(3) Invest in engineering to try to solve the problem
(4) If problem not solved, go to (2)
(5) Profit

 

[Knightshade]

I think that's a very naive view of Musk. I view his approach as:
(1) Identifying a difficult engineering problem with a large reward

I view it as a solved problem. We know exactly what to do we will be there in a few years

He goes on to say a fully autonomy vehicle that is much safer than a person "is much easier than you would think"


So it sure does not sound like he IDed it as a difficult problem back then.

Quite the opposite.