Firmware 9 in August will start rolling out full self-driving features!!!

[Garlan Garner]

A synonym for "right way" might be "defensive driving". So for example, if FSD sees a car in front swerving left and right a bit in the lane, it increases the TACC distance between you and that car to give it more space and maybe preps the EAB to respond faster just in case. Or, if the side rear camera sees a car in the adjacent lane behind you speeding up to pass you and is going too fast, it hugs the opposite lane marking to give the speeding car some space or if safe to do so, switches lanes to give the speeding car even more space. Or, if a car behind you is riding your bumper, FSD could speed up if it is safe to do so, or switch lanes to give the car behind you a free lane to accelerate. These are good behaviors that a FSD system should have IMO.

To me....FSD needs to first be able to flawlessly operate if there are no drivers on the road but itself.

For instance....it's driving down the road.....the road disappears and drops off to a cliff or sand or something. Lastly if the road ahead dips down under 5 feet of water. FSD currently does not account for such simple things. FSD will continue to drive over the cliff because the lack of a road is not an obstacle.

 

[diplomat33]

To me....FSD needs to first be able to flawlessly operate if there are no drivers on the road but itself.

I agree. That needs to come first. But since you were talking about idiot drivers, I was talking about how I think FSD should handle them.

 

[Garlan Garner]

I agree. That needs to come first. But since you were talking about idiot drivers, I was talking about how I think FSD should handle them.

understood.

Idiocy is just too variable to counter with programing.....IMHO

 

[ahkahn]

And in doing so, I think Karpathy saved Tesla's entire Autopilot and FSD project.

10000% agree. I pin a lot of AP2/FSD's delay and early failures on Chris Lattner.

 

[rnortman]

To me....FSD needs to first be able to flawlessly operate if there are no drivers on the road but itself.

For instance....it's driving down the road.....the road disappears and drops off to a cliff or sand or something. Lastly if the road ahead dips down under 5 feet of water. FSD currently does not account for such simple things. FSD will continue to drive over the cliff because the lack of a road is not an obstacle.

Which FSD are you talking about exactly?

 

[SlicedBr3ad]

I agree. That needs to come first. But since you were talking about idiot drivers, I was talking about how I think FSD should handle them.

Ejector seats!

 

[Garlan Garner]

Which FSD are you talking about exactly?

5.

 

[rnortman]

5.

I give up.

 

[dhanson865]

The statement could be a way to increase cash on hand by having more people pre-purchase FSD.
With the ramp up on the Model 3, this could mean a lot of cash.

Funds pulled in from unreleased features are kept separate from general cash flow per Deepak Ahuja CFO of Tesla Motors.

They aren't counted as profit, aren't usable to pay R&D or operating expenses. Basically just sit there doing no good to Tesla until they release the feature that unlocks those funds.

 

[Garlan Garner]

I give up.

Yep

 

[Skotty]

Since this thread came up, I'll say it again. Don't get your hopes up.

I still believe they should initially focus largely on making interstate highway self driving, getting it to level 5, while mostly ignoring around town driving. And there should probably be a coalition of car companies and DOT organizations to define standards for how highways must be marked and how road construction must be implemented so that self driving engineers can have better and clearer limits to what the self driving hardware and software needs to be able to handle. Why? Because then you will finally be able to achieve level 5 (at least on highways) with the inevitable software and hardware limitations (cost included). And I don't know about the rest of you, but if I could trade away one type of driving for automation, it would be those long monotonous highway drives.

Also, this could be great for truckers, who could earn their money on the endpoint local driving but getting to rest on the long haul highway stretches, delivering their goods faster while also being less exhausted.

Level 5 for all roadways will continue to disappoint people for probably decades to come.

 

[rnortman]

Since this thread came up, I'll say it again. Don't get your hopes up.

I still believe they should initially focus largely on making interstate highway self driving, getting it to level 5, while mostly ignoring around town driving.

There.
Is.
No.
Such.
Thing.
As.
L5.
That.
Only.
Works.
On.
Highways.

L4 is not "something just shy of driverless". L4 is "completely driverless on certain roads". So an L5 system that only works on the highway isn't an L5 system, it's an L4 system.

Holy cow, lots of people are wrong on the Internet and I must do something about it.

 

[diplomat33]

Since we've been discussing levels of self-driving, I thought I would share this graphic to educate all involved on what we are talking about:

For reference, EAP is currently at level 2. Waymo has level 4.

 

[diplomat33]

Based on the graphic I posted above, I think the V9 software update will start pushing EAP into level 3 self-driving with features like the auto lane change without driver input, merging traffic (that Musk alluded to in his tweet) and a full on-off ramp and highway transition features.

 

[Skotty]

There.
Is.
No.
Such.
Thing.
As.
L5.
That.
Only.
Works.
On.
Highways.

L4 is not "something just shy of driverless". L4 is "completely driverless on certain roads". So an L5 system that only works on the highway isn't an L5 system, it's an L4 system.

Holy cow, lots of people are wrong on the Internet and I must do something about it.

Yeah, have to say, I don't really care about what the levels mean, since I think it's a dumb rating system to begin with. But you get the point. Completely driverless on highways I think is a worthwhile and attainable goal.

 

[rnortman]

Since we've been discussing levels of self-driving, I thought I would share this graphic to educate all involved on what we are talking about:

I hate that graphic, because despite what the text says, it shows hands on the wheel for L4 as part of smooth progression in driverlessness from L1 to L5, with the hands getting fainter and fainter. This is the entire source of confusion in this thread. L1-L5 is not about a smooth gradient of getting more and more driverless. These are sharp divisions. L4 is completely autonomous and does not require a steering wheel in principle.

L3 isn't just a "better" L2 system. L3 means you can take your eyes off the road and ignore what is happening in the car. An L3 car is responsible for alerting you when you need to pay attention and for continuing to operate safely while you wake up and figure out what is going on (this can take like 10 seconds btw, so the car needs to be able to pull over safely even when crazy stuff happens). That is not just "better L2", that is a step change; it's qualitatively different technology with fundamentally different problems than an L2 driver assistance system where the driver must pay attention at all times.

Similarly, L4 isn't just a better L3 system. And L4 isn't a "not quite L5" system either. There is no smooth gradient here.

 

[rnortman]

Similarly, L4 isn't just a better L3 system. And L4 isn't a "not quite L5" system either. There is no smooth gradient here.

Also, this isn't an entirely semantic discussion about the definitions of these levels. I already said I don't really care for L1-L5 and most people working in the field don't reference those categories.

But setting aside those definitions, the point is that this whole process of developing autonomous vehicles is not about smooth gradients of a driver assistance system getting incrementally better until it's an L5 system. This is not to say that it's not about evolving technology -- clearly there is an evolution here and the technology gets better over time. And you can have a good driver assistance system and a great driver assistance system. But you need step changes to get from driver assistance to fully autonomous -- there are fundamentally different problems to solve.

 

[boonedocks]

Funds pulled in from unreleased features are kept separate from general cash flow per Deepak Ahuja CFO of Tesla Motors.

They aren't counted as profit, aren't usable to pay R&D or operating expenses. Basically just sit there doing no good to Tesla until they release the feature that unlocks those funds.

There should be one sh%4t load of money waiting for EAP to be released then. Can’t wait for that to happen. ( EAP I mean).

 

[diplomat33]

I hate that graphic, because despite what the text says, it shows hands on the wheel for L4 as part of smooth progression in driverlessness from L1 to L5, with the hands getting fainter and fainter. This is the entire source of confusion in this thread. L1-L5 is not about a smooth gradient of getting more and more driverless. These are sharp divisions. L4 is completely autonomous and does not require a steering wheel in principle.

L3 isn't just a "better" L2 system. L3 means you can take your eyes off the road and ignore what is happening in the car. An L3 car is responsible for alerting you when you need to pay attention and for continuing to operate safely while you wake up and figure out what is going on (this can take like 10 seconds btw, so the car needs to be able to pull over safely even when crazy stuff happens). That is not just "better L2", that is a step change; it's qualitatively different technology with fundamentally different problems than an L2 driver assistance system where the driver must pay attention at all times.

Similarly, L4 isn't just a better L3 system. And L4 isn't a "not quite L5" system either. There is no smooth gradient here.

Here is a table that I believe includes the actual definitions of each level. This might be more to your liking.

 

[Bladerskb]

Yes, Self-Driving is a huge challenge. I don't think anyone thinks it is easy. However, what impressed me about the video is that Karpathy and Tesla are tackling the most difficult part of Self-Driving first. So, once it is overcome, a big piece of Self-Driving will be done. The rest won't be easy either but it will easier relatively speaking compared to the first part.

We should keep in mind that Tesla has already done a lot of work with TACC and Auto Steer which are the basic building blocks of Self-Driving since they control speed and steering. Once Tesla does finish the huge task of labeling datasets so that the "FSD package" can identify all objects in the car's environment, they will have completed one of the hardest parts of Self-Driving and will be one step closer to L4 autonomy.

Not according to Mobileye's Amon. Sensing is the easiest part of autonomous driving. Everything Andrej talked about with sensing, data gathering and labeling is literally the basics.

Now this isn't some research work. Both of these "easy" things according to Mobileye has been in production since late 2014. They power Tesla AP1, Volvo Pilot Assist 2, GM Supercruise, Audi A8 L3, and also Audi A7 Cross Country Demo Car.

What mobileye is now working on is the had part of self driving. Word-scale HD map. Their REM map is now live and powers the driveable part of their system. So that also is solved.

So they have solved two pillars in autonomous driving, sensing (which they have been saying is easy and solved back in 2016) and mapping (which they just solved with their eye4 which also has next gen sensing capabilities 1000x of eyeq3.

What is left is planning. the driving policy which they are now working on.

Sensing > Mapping > Planning

That's the difficulty scale.