Autonomous Car Progress

[Bladerskb]

That's completely wrong. Look at how NHTSA collects crash statistics: Crash Report Sampling System | NHTSA

"CRSS obtains its data from a nationally representative probability sample selected from the estimated 5 to 6 million police-reported crashes that occur annually.

These crashes include those that result in a fatality or injury and those involving property damage. By focusing attention on police-reported crashes, CRSS concentrates on those crashes of greatest concern to the highway safety community and the general public."

By relying on police reports, only a small fraction of actual crashes are every collected by the NHTSA. Tesla's telemetry is far more comprehensive.

No Tesla FSD Beta Report is 100000000x worse because very few accidents actually lead to air-bag deployment.
While any accident of any kind can lead to police being called. I even posted accidents that had police report but there were no airbag deployed.

 

[willow_hiller]

No Tesla FSD Beta Report is 100000000x worse because very few accidents actually lead to air-bag deployment.
While any accident of any kind can lead to police being called. I even posted accidents that would have police report but there were no airbag deployed.

Maybe for typical vehicles with awful airbag systems, but from Tesla's report:

"In practice, this correlates to nearly any crash at about 12 mph (20 kph) or above, depending on the crash forces generated."

A few anecdotes aren't enough to convince me that airbags are deployed less often than police are called to crash scenes. If you have data proving otherwise, post it.

 

[diplomat33]

Time is crucial when estimating velocity? Am I missing something or is this an absurd statement?

Well, I would call it a self-evident statement. In context, it is simply making the (obvious) statement that vision NN needs to include time in order to do proper velocity estimation. For example, a few years back, Elon talked about a major rewrite that switched from image processing to video processing that he called "4D". By processing video instead of images, the Tesla vision NN would be much better at things like velocity estimation.

How significant is 0.378 m/s (0.85 mph) to autonomy software? I drive around with an mAVE considerably higher than that number.

You are correct that humans have an mAVE that is much higher. That's because humans make estimates about relative change. Humans can discern that a car is approaching too fast and we should move away to avoid a collision or that we are approaching the car in front too fast and we should brake. But we can't say the precise velocity or distance to an object. Sure it "works" most of the time but humans also get it wrong which can lead to collisions. We want AVs to be far safer than humans. So the more precision in perception, the better at making AVs safer. With cameras, radar and lidar, AVs don't need to guess like humans do but can build a very accurate and precise bird's eye view of the world which will help make AVs safer than humans.

I would say that a mAVE of 0.378 m/s is very good. That mAVE is very significant for autonomous driving. And it shows that vision NN is becoming quite capable at certain perception tasks. As the paper says, a mAVE of 0.378 m/s is almost as good as lidar. So state of the art vision is approaching lidar level accuracy when it comes to velocity estimation. But also keep in mind that the mAVE of 0.378 m/s is only from a test sample. Real-world performance in different environmental conditions like rain, snow, fog, shadow, glare, etc will likely affect that mAVE a lot. That is another big reason, L4 vehicles use sensor fusion of cameras, lidar, radar so that there is that redundancy when vision is degraded. With vision-only, when the vision degrades for whatever reason, there is no back-up. And with no back-up, when vision degrades, you either have to ask the human to take over or have the car pull over. That will happen less often if you use other sensors like lidar and radar to compliment vision.

 

[powertoold]

Tesla safety report accidents count both airbag and/or active restraint deployment.

 

[powertoold]

implying not yet as good as lidar for velocity estimation

That's not what I was saying. That's why context is important.

It's obvious cameras will never get as good as LIDAR for accuracy in certain conditions.

 

[Bladerskb]

Maybe for typical vehicles with awful airbag systems, but from Tesla's report:

Airbags in cars have to MEET safety standards.

"In practice, this correlates to nearly any crash at about 12 mph (20 kph) or above, depending on the crash forces generated."

A few anecdotes aren't enough to convince me that airbags are deployed less often than police are called to crash scenes. If you have data proving otherwise, post it.

Hitting a car/random object going 12mph will generally not get your Tesla airbag deployed.
There could be deployment for specific stationary object with specific dimension like a tree. But rear-ending/getting rear ended another car at 12mph Will not deploy your airbag. Its all about what you are hitting and what angle you are hitting it, not just your mph.

I have already posted two accidents on FSD with NO airbag deployment. The second one with police report.

Police report are generally made when someone impacts another vehicle/person and either party observes damage to their vehicle/person.

The All-time stats for airbag deployment in car accidents is 3.3 million which is a study done by the government.

There were 6.76 million police-reported crashes in 2019.

 

[diplomat33]

That's not what I was saying. That's why context is important.

Well, then please refresh my memory on the context. I seem to remember you saying that vision had won and you asked me for examples of deficiencies vision had vs lidar. I just gave you 2 deficiencies where vision has not "won" yet.

And by the way, I am not saying that vision will never win. I am just saying that vision has not won yet. I think that paper that vision is catching up to lidar in certain areas but it still lags, backs me up. When we have vision-only driverless cars, then I will happily admit that vision has won.

 

[willow_hiller]

The All-time stats for airbag deployment in car accidents is 3.3 million which is a study done by the government.

I like how this study is so unreliable, you can't even link to it.

Tell me: How did the government measure airbag deployments? Police reports? If so, it's only a measure of how often airbags deploy AND police reports are made. Not airbag deployments alone.

 

[powertoold]

When we have vision-only driverless cars, then I will happily admit that vision has won.

Yup, I don't see any deficiencies in vision only that would preclude that within the next year.

 

[diplomat33]

Yup, I don't see any deficiencies in vision only that would preclude that within the next year.

I don't see that happening within the next year. I think it will take longer than that. But we shall see.

PS: I guess it depends on the ODD. Vision-only has deficiencies in darkness and rain and snow. So I don't see vision-only driverless everywhere happening in the next year. But vision-only driverless in a geofence on a clear sunny day, within the next year, maybe.

 

[JB47394]

Well, I would call it a self-evident statement.

That is an academic way of saying that it's dumb. Why embellish a fundamental disclosure of information with erudition when the same could have been equivalently conveyed in the vernacular - or simply omitted in toto? (That's a rhetorical question, by the way)

Sure it "works" most of the time but humans also get it wrong which can lead to collisions.

Sure, through inattention or poor judgement, neither of which should apply to an autonomy system. I question whether such low mAVEs are even necessary. But I suppose that's a question for somebody pursuing their master's to determine the maximum mAVE practical for a "safe" autonomy system.

 

[diplomat33]

I question whether such low mAVEs are even necessary. But I suppose that's a question for somebody pursuing their master's to determine the maximum mAVE practical for a "safe" autonomy system.

It's a valid question. But you make it sound like it is just some academic question for a thesis but mAVE has real world consequences. There could be a driving scenario where the mAVE makes a difference between crashing or avoiding a crash. Certainly, I think it makes sense to try to achieve the best mAVE possible. If you can reduce the mAVE even more, why not? Achieving lower mAVE can only help your AV, not just in safety but also in making the driving smoother, which is also important for riders.

 

[spacecoin]

I don't see that happening within the next year. I think it will take longer than that. But we shall see.

PS: I guess it depends on the ODD. Vision-only has deficiencies in darkness and rain and snow. So I don't see vision-only driverless everywhere happening in the next year. But vision-only driverless in a geofence on a clear sunny day, within the next year, maybe.

Definately not happening. It doesn't work either if the sun is low. Then we have other forms of blinding, ML hallucinations and so on. Multi-modal will be the only thing robotaxis will be using this decade, and probably forever (for safety).

 

[kabin]

It might be time for the Connor apology...

 

[EVNow]

And then you have Tesla fans using these two sentence "Safety Report" to say that Tesla is safer than Waymo/Cruise. (moderator edit)

@Bladerskb - just because the mod edited your lie about me doesn't mean you can escape. Show me where I've said this or apologize.

@diplomat33 Cool - so you are so eager to defend @Bladerskb when he is attacking us - but when he is caught lying, you don't want to be "dragged in".

 

[EVNow]

I even posted accidents that had police report but there were no airbag deployed.

You can definitely have police reports with no airbag deployed. What nonsense is this ?

 

[spacecoin]

Oh the drama... Poor "us". Also angry "us". Everyone should just accept that whatever any company uses as marketing is marketing and not science? This is only controversial because? Yeah, definitely not a cult...

 

[willow_hiller]

Oh the drama... Poor "us". Also angry "us". Everyone should just accept that whatever any company uses as marketing is marketing and not science? This is only controversial because? Yeah, definitely not a cult...

It's really unclear whether you're talking about Waymo, Cruise, or Tesla.

 

[Bladerskb]

@Bladerskb - just because the mod edited your lie about me doesn't mean you can escape. Show me where I've said this or apologize.

The one delusional here is obviously you.

Waymo/Cruise are so unsafe no way SanFrancisco would allow them to run if they had not paid off CA regulators to prevent proper oversight by SF.

FSD + Human has better accident rate than just human. Looks like a good perf goal that is being met.

1+1 = 2

I know you Tesla fans love to say one thing in a vacuum and then say another thing in another vacuum.
But the real world doesn't work like that.

 

[willow_hiller]

1+1 = 2

I know you Tesla fans love to say one thing in a vacuum and then say another thing in another vacuum.
But the real world doesn't work like that.

The entire point of your accusation was that he was misusing data to make a direct comparison between the safety of FSD Beta and Waymo/Cruise.

Quoting two separate posts defeats your own argument.