Autonomous Car Progress

[EVNow]

Years ago I'd watch Omar's videos.

But he puts out raw videos showing brake/accelerator now (because people kept accusing him of cheating).

ps : One thing that is common to (most) YouTubers is that they want a lot of clicks. They are more likely to make the videos as attractive as possible to get clicks. Thats the nature of social networks. That doesn't make one a shill. Shill is someone who doesn't care about the product at all - and just saying positive things for financial gain (like people who write positive reviews of products on Amazon in exchange for money or a "gift").

 

[EVNow]

And people on this forum know I used to be a huge FSD fan, constantly defending Elon's FSD claims, but I got burned when Elon missed one too many FSD deadlines. So I soured on FSD but I still like Tesla cars.

I know this well - that is why I think you have been gullible in the past. I'm always skeptical of these kinds of claims. By Tesla or Waymo.

BTW you completely ignored the part about "acting". Do you have any proof Omar is not a fan - but is just acting ? That is what would make him a shill.

 

[powertoold]

To simply say Omar is a shill, something is lost in translation. Omar has some degree of shill. But to say he is a Tesla shill in general is wrong. No one denies FSDb is imperfect. It's funny to see people expecting perfection on initial release, pointing out every little problem. Rather than focusing on all the inevitable problems, Omar just focuses on the big picture, and in the long run, he'll be right about the big picture.

We've gone, as a forum, from pointing out deficiencies in camera vision or placement to nitpicking about merging, unnecessary slowdowns, creeping behavior, jerkiness, etc. etc. -- what's clear now is that camera vision has won. The rest is just software.

 

[spacecoin]

What's clear now is that camera vision has won. The rest is just software.

Are you claiming that anyone thought that you could autonomously without cameras or with camera only?

Both are equally wrong at this point in time. Perhaps CV will get there, but not this decade most likely. On existing Tesla cars, I predict autonomy never. There is a subtle 2000x-4000x reliability improvement needed for autonomy.

After almost three years of little progress in reliability I think we can safely put Tesla in the "L2+" basket for good on HW3/HW4.

The rest isn't just software. It's about breaking new boundaries in computer vision and machine learning.

 

[powertoold]

Are you claiming that anyone thought that you could autonomously without cameras or with camera only?

Both are equally wrong at this point in time. Perhaps CV will get there, but not this decade most likely. On existing Tesla cars, I predict autonomy never. There is a subtle 2000x-4000x reliability improvement needed for autonomy.

After almost three years of little progress in reliability I think we can safely put Tesla in the "L2+" basket for good on HW3/HW4.

The rest isn't just software. It's about breaking new boundaries in computer vision and machine learning.

What is your background and how long have you followed AVs?

 

[spacecoin]

What is your background and how long have you followed AVs?

I am a serial entrepreneur and investor with a technology background (computer science). I have been following autonomous vehicles since the start of the Google Self-Driving Car Project and have been researching AI and ML since the DeepMind/AlphaGo splash. I drive a Tesla, as it won't drive itself

 

[powertoold]

I am a serial entrepreneur and investor with a technology background (computer science). I have been following autonomous vehicles since the start of the Google Self-Driving Car Project and have been researching AI and ML since the DeepMind/AlphaGo splash. I drive a Tesla, as it won't drive itself

Ok great, I just found many logical disconnects in what you were saying wrt to CV and AV development. I find it hard to believe you are detail oriented to be involved in large projects and post what you did.

Why don't you think CV is there yet? What is missing?

 

[Doggydogworld]

But he puts out raw videos showing brake/accelerator now xx

Yes, he does now. As I clearly stated, FSD improved enough that he no longer needs to cheat so egregiously.

(because people kept accusing him of cheating).

Because he was cheating. And his actions reveal that he's perfectly willing to cheat.

ps : One thing that is common to (most) YouTubers is that they want a lot of clicks.

Ironically, it's mistakes that garner more clicks. Conegate got something like 100x as many views as all of JJRicks' other videos combined. I'm sure Omar had tons of great FSD screwups over the years that would have garnered tons of clicks. But he cares more about the narrative. That's why unbiased people call him a shill.

 

[powertoold]

Yes, he does now. As I clearly stated, FSD improved enough that he no longer needs to cheat so egregiously.


Because he was cheating. And his actions reveal that he's perfectly willing to cheat.


Ironically, it's mistakes that garner more clicks. Conegate got something like 100x as many views as all of JJRicks' other videos combined. I'm sure Omar had tons of great FSD screwups over the years that would have garnered tons of clicks. But he cares more about the narrative. That's why unbiased people call him a shill.

Do you remember what video / videos you're referring to wrt interventions? In the past, there wasn't a consensus wrt intervention vs disengagement. I've been watching his videos since the very beginning.

As for conegate, that was a fail of epic proportions, so it makes sense for the wide publicity.

 

[diplomat33]

what's clear now is that camera vision has won. The rest is just software.

I think you are jumping the gun. We don't see any driverless with vision-only, not even close. It might win eventually, but until we actually see L4/L5 with vision-only, it has not won.

 

[powertoold]

We don't see any driverless with vision-only, not even close.

I take issue with these types of statements. What isn't close about Tesla FSDb? Please consider that FSDb is trained on road geometry and semantic data WORLDWIDE.

Do you really think that FSDb wouldn't be able to achieve similar or better performance / reliablility / safety vs Cruise or Waymo if Tesla focused their training / efforts on a smaller locale?

What deficiencies in vision-only do you see vs LIDAR, please give specific categories / examples of failures or deficiencies.

 

[EVNow]

Do you really think that FSDb wouldn't be able to achieve similar or better performance / reliablility / safety vs Cruise or Waymo if Tesla focused their training / efforts on a smaller locale?

I think is the big thing Waymo apologizers ignore.

 

[diplomat33]

Good article about the progress so far but also remaining challenges for Waymo and Cruise:

The death of self-driving cars is greatly exaggerated

GM’s Cruise aims to turn self-driving into a billion-dollar business.

www.understandingai.org

 

[powertoold]

What people ignore about Elon's missed timelines is his one-in-a-generation ability and track record to envision what technology can and can't do.

Elon has missed many timelines, but Elon rarely, if ever, misses on his grand visions, and his vision for camera-only full autonomy is his *most* confident prediction. When Elon isn't sure about the success of something, he will say it.

Considering Elon is intimate with many of his company's projects, it's difficult to believe that Elon doesn't know what he's saying about wrt to the camera-only approach and its limitations. And we are all armchair visionaries who know better than he does about something he's so confident about.

That isn't to say we can't question Elon's timelines, but can anyone question that LIDAR is becoming more and more obsolete every day?

 

[spacecoin]

Ok great, I just found many logical disconnects in what you were saying wrt to CV and AV development. I find it hard to believe you are detail oriented to be involved in large projects and post what you did.

Can you give an example of what was so illogical to you? Also, name one company that is deploying vision only, or planning to deploy vision only autonomous systems in the comping 2-3 years. Hint: there are none.

Why don't you think CV is there yet? What is missing?

All the Tesla bubble folks fail to understand this basic word: Reliability - which is everything in safety- and time critical applications. No one knows how to get existing CV solutions there yet.

We don't even trust unsupervised use of solely computer vision in radiology, so not likely either in a time critical 2000kg robot scenario at 130 km/h anytime soon.

Waymo does 10000 rides per day without a driver. Tesla does doesn't do a single per year. Progress should be fast now and slow later. It's slow now - or flat even @ 5-10 mils per DE. See this: The End of Starsky Robotics

Also, watch this: Reinforcement Learning for Self Driving Cars, including the Q&A.

Maybe vision only will pan out, but if you're serious of deploying this decade, you aren't betting on it. It's easy to remove maps, lidar and scale down the sensor suite later. You also won't be selling it with a six year old hw-suite for a fixed cost if you truly believe it will happen.

 

[spacecoin]

Please consider that FSDb is trained on road geometry and semantic data WORLDWIDE.

Really? Source please? For left hand driving in UK/AUS? EU traffic lights? Street signs? India? Japan? China? Or is your definition of worldwide just North America?

Do you really think that FSDb wouldn't be able to achieve similar or better performance / reliablility / safety vs Cruise or Waymo if Tesla focused their training / efforts on a smaller locale?

Anyone with basic understanding about this field understand that more and better sensing across multiple modalities improves reliability. Maps improve reliability (you see the "into the future" and occluded objects). Tesla can't even make smart summon to work reliably in a parking area. That's a limited ODD if any.

If you train two systems separately to drive and then combine the output, you can in theory multiply the failure rate if there are no overlaps.

So if you have a vision system at failure rate 0.00001 % combined with a lidar system at a FR 0.0001% you get a theoretical failure rate of 0.000000001% when you combine them. In practice perhaps just half, but it's just an example to explain my point.

Most robotaxi companies provide complete trajectory guesses in real time for all agents in the scene even if there are hundreds of them. Tesla's compute can't cope with that.

What deficiencies in vision-only do you see vs LIDAR, please give specific categories / examples of failures or deficiencies.

Let's see. Either Lidar or radar see through sun glare, heavy fog, smoke, snow/slush on lenses (you need self cleaning)... Cameras do not handle these scenarios well... Tesla has only two cameras (hw4) with cleaning.

Also Tesla doesn't have any redundancy left in its system re compute/sensing which most people argue is mandatory for autonomous/unsupervised deployment.

 

[aataskin]

what's clear now is that camera vision has won. The rest is just software.

What's clear is that one day in the future, hopefully, camera vision only vehicles should be able to be as good as human drivers. Which neural net architectures will deliver this, and what kind of hardware will be required for them to run effectively, we don't yet know. It should also be clear that camera vision only cars cannot possibly, even in theory, be as good as radar equipped vehicles at detecting imminent collisions in the dark or in the rain (or if the cameras are covered by dust or mud), or as good as USS equipped vehicles at parking. I am surprized they are still not nearly as good as rain sensor equipped cars at detecting rain, five years after that sensor was ditched.

And so the question remains: why not keep all this useful and relatively cheap technology until you have a comparable vison only solution? Notice, most news from Musk are in the future tense.

 

[EVNow]

Anyone with basic understanding about this field understand that more and better sensing across multiple modalities improves reliability

Funny story.... The Japanese moon lander last month failed because of sensor integration issues.

 

[spacecoin]

Funny story.... The Japanese moon lander last month failed because of sensor integration issues.

How is this even remotely applicable to anything of what I said? Bugs exist in software. And?

 

[EVNow]

How is this even remotely applicable to anything of what I said? Bugs exist in software. And?

It's difficult to merge sensors.

Search - this has been discussed hundreds of times.