HW2 not enough for self-driving?

[LargeHamCollider]

And here I was worrying that this week we wouldn't have another new thread on this topic...

 

[Shock-On-T]

This is like asking the zoo how high the fence should be before they know what animal it's for.
The problem of self driving has not yet been solved, on any platform.
I suspect that as the software matures there will huge scope for optimisation.

 

[Tam]

...AP is used in easier scenarios...

When driverless feature is activated, not everyone will use that feature because they want to have the freedom of manual driving.

You will be able to compare the detected collisions during driverless operation and during manual driving.

When no driverless-optioned owner will do manual driving anymore, you can compare that driverless fleet against Tesla's non-driverless fleet.

 

[GregKo]

There are about 1.3 million people die from car accidents annually.

If your Tesla works as its spec of twice better than human, it would save 650,000 lives.

On the other hand, there are still the rest of 650,000 people who will die because the spec is only twice as good as human driver.

I believe this math has errors. That 1.3M number is worldwide. Many counties don't have safe or even paved roads or safe cars. Some share roads with livestock etc. A better number to compare would be first world countries where the roads and laws are similar.

"Every day, 28 people in the United States die in motor vehicle crashes that involve an alcohol-impaired driver."

Impaired Driving: Get the Facts | Motor Vehicle Safety | CDC Injury Center

In 2015 (16,484) alcohol related deaths could be wiped clean if a L5 FSD car was enabled.

Total deaths in the US (2015) was 35,092.

Fatality Facts

In order to make this number (650,000 lives saved and 650,000 still lost) statistic valid, the infrastructure in all countries would need to be equal. I don't think this comparison valid.

 

[Shock-On-T]

I believe this math has errors. That 1.3M number is worldwide. Many counties don't have safe or even paved roads or safe cars. Some share roads with livestock etc. A better number to compare would be first world countries where the roads and laws are similar.

"Every day, 28 people in the United States die in motor vehicle crashes that involve an alcohol-impaired driver."

Impaired Driving: Get the Facts | Motor Vehicle Safety | CDC Injury Center

In 2015 (16,484) alcohol related deaths could be wiped clean if a L5 FSD car was enabled.

Total deaths in the US (2015) was 35,092.

Fatality Facts

In order to make this number (650,000 lives saved and 650,000 still lost) statistic valid, the infrastructure in all countries would need to be equal. I don't think this comparison valid.

Agreed about including developing nation data.
Problem is that you might save 16,484 dui accidents, but cause 16,484 autopilot accidents (plus or minus depending on how good L5 is).
It's going to be a sliding scale. People won't accept 10,000 driverless car fatalities, even if the alternative is 16,000 dui fatalities.
I think Elon is about right with a 10:1 improvement needed for general adoption.

 

[owen.shift]

For example, an L2 system -- like Teslas AP1 or EAP [...]

I like to think of EAP as L3.

Also:

Yet, the L2 system has by definition no fallback procedure or might not be designed to watch out for flying car wheels, roadkill or man holes [...]

flying + (car wheels, roadkill or man holes).

 

[lunitiks]

I like to think of EAP as L3.

I'd like to think that too, but of course I dont

 

[owen.shift]

Why not? Because you think they will make us buy FSD for that?

 

[MXWing]

I think its impossible to answer for certain as we don't know how well the software is utilizing the potential of the hardware.

We don't even know what Tesla's most advanced build is capable of at this moment.

If we based the capabilities on when I took delivery, this fancy PX2 thingofajiggy wasn't even capable of basic cruise control. A car I drove in 1997 had more autonomy features. (RIP Nissan Maxima) Yes my MX was so fresh with AP2, the stick served no function. Now, 90 MPH autosteer on the highway - that is a big difference.

In any case is the smart money to go ahead and pay the FSDC upgrade the moment a software patch does anything with it?

 

[SoCal Tsla]

And here I was worrying that this week we wouldn't have another new thread on this topic...

oh yeah this topic will go on forever lol

 

[SoCal Tsla]

This is like asking the zoo how high the fence should be before they know what animal it's for.
The problem of self driving has not yet been solved, on any platform.
I suspect that as the software matures there will huge scope for optimisation.

honestly speaking, i don't think so. I think the software will only get more complicated as they add scenarios to consider. They may come up with new HW which can calculate some specific formulas faster but that will result in HW upgrade in the end...

 

[owen.shift]

In any case is the smart money to go ahead and pay the FSDC upgrade the moment a software patch does anything with it?

Probably. Even smarter is to let that money grow a little by way of TSLA in the meantime

 

[MXWing]

Probably. Even smarter is to let that money grow a little by way of TSLA in the meantime

Not this week, and not the way Elon has been anti-selling TSLA.

 

[owen.shift]

Yes, particularly this week. The stock has been down the past several days because people don't get what the return of FUSC means for the future of Tesla. But all this has done was to create a buying opportunity for those who do get it.

 

[MXWing]

Yes, particularly this week. The stock has been down the past several days because people don't get what the return of FUSC means for the future of Tesla. But all this has done was to create a buying opportunity for those who do get it.

It has been a strange week. No way it breaks 309 support, those deep pockets at TMC is going to buy in and prop the price up for sure! 306,305,304 (uh oh). Anyway - I have some calls in for 305 this week and 310 for JUN9th. I am hoping I can stay solvent longer than the market can stay irrational.

So maybe just plopping down for FSDC would actually have been the better buy.

 

[owen.shift]

So maybe just plopping down for FSDC would actually have been the better buy.

That's what I did, FSD at delivery. But you were asking about 'smart' money.

 

[lunitiks]

Why not? Because you think they will make us buy FSD for that?

No, because EAP does not fulfill SAEs definition of L3.

Listen, SAEs levels of automation is not a question of "counting" cameras, Lidars, radars, TFLOPs etc. Nor is it a question of whether your car has a software "feature" like Autosteer+ or Smart Summon. SAEs definitions is an abstraction (an important one) where the only question is how "present" the human is required to be. L5 requiring no human presence at all.

Now the difference between L2 and L3 has to do with the specific task of monitoring the driving environment, and the question of what's expected from the human with regard to fallback-readiness:

L2 expects the human to monitor the driving environment at all times, and to take over immediately without being prompted by the system. In other words, L2 expects the human to detect, recognize, classify, prepare and respond to all and any events all the time. So all L2 does, is to automate the lateral and longitudal vehicle motion control -- whithin its designed driving mode(s).

L3 on the other hand, is by definition designed to perform the monitoring task as well. So the human is not required to monitor the surroundings: She can in principle read the news paper or work while behind the wheel -- again, whithin the systems designed driving mode(s), for example on highways. L3 also requires the human to be fallback-ready (to intervene), but only when the system requests it. For example by sound, vibration, light or otherwise.

The [human] need not supervise a level 3 [system] while it is engaged but is expected to be prepared to resume the [driving] when the [system] issues a request to intervene

You see the difference?

EAP is not an L3 system, it is merely a more capable L2 system than AP1 (well, not yet but it's supposed to become that). In other words, EAP is supposed to handle more driving modes, like for example "fully automated" lane changing.

Even Tesla consideres EAP as DAS:

That said, Enhanced Autopilot should still be considered a driver's assistance feature

(Ok, admittedly you shouldn't compare that marketing description to SAEs definitions, though, as it would be an L1 system -- but you get the point.)

I hope this clarifies things a little bit

 

[Shock-On-T]

honestly speaking, i don't think so. I think the software will only get more complicated as they add scenarios to consider. They may come up with new HW which can calculate some specific formulas faster but that will result in HW upgrade in the end...

Scenarios are computationally cheap. Every car in a arcade racing game is sticking to the track, avoiding other cars, waiting for the starting gun, etc, all on a 20 year old CPU.

The hard part is computer vision, converting the camera feeds into 'arcade game' in real-time.

Computer vision is an area of current research and is improving all the time, but most of the research has focused on specificity (is it a rose or a tulip). Good for things like image classification and search, with plenty of computing power available, and no time pressure.

Computer vision for autonomous cars is relatively new, and with limited computing power and a strict requirement for timeliness it will drive research into optimization. Just quickly finding the flowers, not wasting a cycle to see if they are roses.

 

[owen.shift]

I hope this clarifies things a little bit

So L3 requires FSD. That's what I want, and that's why I paid for FSD before delivery. And happily so.

 

[lunitiks]

So L3 requires FSD.

*sigh* You're really giving me a hard time here

Please understand that "FSD" is a marketing term, born and raised in Palo Alto, California, USA.

"FSD" is not in any way recognized by SAE as being "better", "worse" or "equal" to levels 2, 3, 4 or 5. On the contrary -- here's SAEs official take on self driving:

7. DEPRECATED TERMS
[T]his section identifies certain deprecated terms that are not used in this [document] either because they are functionally imprecise (and therefore misleading) and/or because they are frequently misused by application to lower levels of driving automation [...]

7.1.2 Self-driving
The meaning of this term can vary based on unstated assumptions about the meaning of driving and driver. It is variously used to refer to situations in which no driver is present, to situations in which no user is performing the [driving], and to situations in which a driving automation system is performing any part of the [driving].

You see, SAE has worked out precise definitions of the various levels of automation. So if you want to make a claim that Tesla has promised you a certain (SAE) level of automation, you have to think long and hard about what "FSD" gives you -- the consumer -- the impression of, with respect to SAEs definitions.

Ok, I'll challenge you: When you opted for "FSD", did you expect that the system would execute every part of the driving task under all circumstances? Or did you realize/expect that the system was limited -- "somehow"? Well if the latter is true: Exactly what/which limitation/s did you consider and accept before you made your puchase? Remember you have to justify your answer to that question: Why that/those limitation/s, and not others?

(Yeah I'm provoking you a bit here, but I think these are valid questions that must be answered when dealing with "FSD" vs. SAELOA.)