AP 2.0 on submitted orders

[Tam]

Autosteer and Summon are already features of AP1. Who knows what that advanced part is? Is that where it actually stays on the road and doesn't run into things? Sorry about your car running off the road and flipping over sir. You should have paid for Enhanced AP when you had the chance. I'll take my chances and let you know what features I'm missing in three months. If the car has truck lust, takes exits by accident, and runs itself into trailers during summoning, you must be right. Well, as long as the Enhanced systems aren't misbehaving in the same manner.

As long as it's a Tesla, it's a good choice.

1) Manual Driving
2) AutoPilot
3) Enhanced AutoPilot
4) Full Self-Driving


Just be informed! It's all good!

We know how AP1 has behaved in real life but we have no ideas about AP2's real life experience until owners will start reporting back.

In the mean time, we just have to speculate how much better it can be as demonstrated by Tesla's video.

No one has demonstrated that AP1 can avoid whacking traffic cones.

But nVidia demonstration below shows that the system learned its fault when striking a traffic cone. It then performed flawlessly in a twisting construction zone full of traffic cones. It learned the street scenarios in California and applied that in New Jersey....

You don't want your brand new car to whack a traffic cone in the process of learning real life experience. That's why there's a gap of manual driving from now until December 2016. The system needs to learn sufficient numbers of real scenarios first. I believe it is called Deep Neural Network (DNN) and branded as "Tesla Vision" which will share with all your cars of what it has learned so far in December 2016. The system will continue collect and sharing data and should be sufficient for driverless driving for Los Angeles to New York trip in December 2017.

 

[⚡️ELECTROMAN⚡️]

I'm guessing they designed the system to be upgradeable as compared to AP1. I read that the CPU is designed to be swapped out with newer faster CPU. I wouldn't be surprised if they had to make some hardware tweaks to pull off full autonomy. I would think they built that into AP2 or they wouldn't be so confident.

 

[MP3Mike]

@MP3Mike Many will probably upgrade to AP+, but even with a few thousand cars with V2 hardware and v1 functionality, Tesla now has created a specific branch of software for arguably a very small number of vehicle. It will cost them more to train and maintain and test for each new release. It is likely that they'll end up moving everyone to AP+ at some point, like they gave Nav to non Tech Package cars after a couple of years.

But there are probably more than thousands, it would be 10s of thousands. And on top of that EAP probably gracefully degrades to the AP1 level if cameras fail/get blocked instead of just turning off. So the effort is useful for every Customer going forward.

 

[Tam]

I'm guessing they designed the system to be upgradeable as compared to AP1. I read that the CPU is designed to be swapped out with newer faster CPU. I wouldn't be surprised if they had to make some hardware tweaks to pull off full autonomy. I would think they built that into AP2 or they wouldn't be so confident.

Agreed. Tesla will do a Driverless Los Angeles to New York City in December 2017. That's 1 year from now so I don't think Tesla will abandon current hardware just yet.

However, that is not a guarantee that Tesla might find something better by then.

As you might know, Tesla has continuously evolved and it just does not stop and wait for a next "year model."

I would feel more comfortable if Tesla can clarify the components of motherboard, CPU, GPU, memory on board...

Right now, we only know by patching different sources together:

Elon: It's a Titan GPU (That's a GP 102, I think because Titan is a graphic card, not a chip.)

nVidia: It's a Drive PX 2 motherboard. You need 1 GPU on that motherboard for Autocruise. 2 GPU for point-to-point travel (I would call that as Tesla Enhanced Autopilot). And you need "mulitple" boards for Driverless Driving.

 

[2virgule5]

They rushed all possible V1 cars by end of Q3 and stopped the factory for the first week of October. Even orders passed early September where built and delivered for the most part. So really there should not be more than 2 or 3 weeks worth of production for folks that ordered V1 and that will have V2.

 

[Kay Kay]

I received the same email this morning asking to upgrade but I still think it is not fair because they had already charged for the auto pilot and now they named it enhanced auto pilot and self driving separately. I think it was reasonable to charge us for extra hardware but not sepately. Any way I need a self driving car So I opted for both but just thinking down the road they might name something else and ask for more money.

 

[MorrisonHiker]

So really there should not be more than 2 or 3 weeks worth of production for folks that ordered V1 and that will have V2.

There's probably another week's worth from delayed delivery. I know a lot of people, including myself, delayed delivery while waiting for AP 2. We placed orders for two Model S three days before the refresh, knowing it was coming. Before my order confirmed, I delayed delivery because it was missing AP 2 and a couple other features. I'll finally be taking delivery in the coming weeks.

 

[rainforest]

I received an official email from Tesla this morning to upgrade to the new enhanced autopilot. My MX60 order was placed back in July with AP for $2500, but I had it pushed back to December for delivery. Now I have the option to upgrade for an additional $2500 for enhanced AP or $5500 for the whole Full Self Driving + EAP package! So glad that my solid white MX, which is now a rare breed, will be getting the new AP hardware 2.0!

But then you lose the discount. AP was $3000, and now it is $5000. Shouldn't you only be paying $2000?

 

[MorrisonHiker]

But then you lose the discount. AP was $3000, and now it is $5000. Shouldn't you only be paying $2000?

It's $5000 total for Enhanced Autopilot. Some of us with older delayed orders only paid $2500 originally so now we have to pay $2500 again to reach the $5000 total. Others who paid $3000 for more recent autopilot orders only have to pay $2000 to reach the $5000 total.

 

[krazineurons]

Is there a way to figure out if the BB8 had 4 active cameras (enhanced AP) or 8 (full self driving)?

 

[rainforest]

It's $5000 total for Enhanced Autopilot. Some of us with older delayed orders only paid $2500 originally so now we have to pay $2500 again to reach the $5000 total. Others who paid $3000 for more recent autopilot orders only have to pay $2000 to reach the $5000 total.

Yes, but technically, the difference between AP1 and AP2 is only $2000. You already have AP1 (though at a lower price), so the right thing to do would be to just charge $2000. Then again, I'm not in charge!

 

[DJ 240V]

Hey all - i'm curios to know where we all stand w.r.t to upgrading to EAP - could you please cast your votes on this thread if you are one of the few lucky ones that in the transition window? Who is upgrading to self driving?

 

[StormShadowX]

But then you lose the discount. AP was $3000, and now it is $5000. Shouldn't you only be paying $2000?

Seems like Tesla is pretty much set on $5000 for the enhanced AP regardless of any previous discount. They just charge you the difference.

 

[Tam]

Is there a way to figure out if the BB8 had 4 active cameras (enhanced AP) or 8 (full self driving)?

According to their write up:

https://arxiv.org/pdf/1604.07316v1.pdf

I am a bad reader so I just skimmed through it but it sounds that for actual self-driving they only utilized 1 front facing cameras.

But for data acquisition, you'll need at least 3 cameras or more.

Labeled by Evino:

Remember, to get the data for training, you don't need the same car. You can install multiple cameras in a different car in California then feed it to train a car in New Jersey like BB8 that has only one single front facing camera.

Ideally, for real world application, you'll need as much sensors as your budget can afford. But for the purpose of academic demonstration of what the least numbers of active sensors you can have, BB8 proves that 1 single camera (with no Radar, no Sonars, no LIDAR...) will do!

 

[krazineurons]

According to their write up:

https://arxiv.org/pdf/1604.07316v1.pdf

I am a bad reader so I just skimmed through it but it sounds that for actual self-driving they only utilized 1 front facing cameras.

But for data acquisition, you'll need at least 3 cameras or more.

Labeled by Evino:

Remember, to get the data for training, you don't need the same car. You can install multiple cameras in a different car in California then feed it to train a car in New Jersey like BB8 that has only one single front facing camera.

Ideally, for real world application, you'll need as much sensors as your budget can afford. But for the purpose of academic demonstration of what the least numbers of active sensors you can have, BB8 proves that 1 single camera (with no Radar, no Sonars, no LIDAR...) will do!

Awesome! Doesn't this mean good news to AP 1.0 hardware owners?

 

[Tam]

Awesome! Doesn't this mean good news to AP 1.0 hardware owners?

AP1 and AP2 have a completely different platforms.

The basis of AP1 is: Driver has to be competent and alert to take over at all times!

AP1 has a single monochrome camera that has no depth perception and has a hard time to recognize different colors at time (white tractor-trailer versus brightly lit sky on the background.)

AP1 hardware is vastly inferior (old=$3,000 vs new=$8,000)

Eventually, AP1 will be able to do handfree on free way from ramp-to-ramp but driver still has to be babysitting it to make it does not get into an accident.

AP2 is so powerful that it requires a very costly supercomputer system to handle it.

Once the machine learns on its own sufficiently, you'll be out of a babysitting job for the system because it will drive itself!

 

[krazineurons]

The basis of AP1 is: Driver has to be competent and alert to take over at all times!

Sorry, I should have clarified better. I have been following all AP 1 vs 2.0 announcement and details and we'll aware of the cameras and the Nvidia platform AP 2.0 uses.

The fact that a car equipped with single camera is able to perform complex manovers, makes me feel that all AP 1.0 cars could possibly learn to do these manovers too by virtue of all learning performed by AP 2.0 hardware equipped cars. Only thing to know is if the processing hardware of AP 1.0 would be able to process the DNN model or not.

 

[DJ 240V]

Sorry, I should have clarified better. I have been following all AP 1 vs 2.0 announcement and details and we'll aware of the cameras and the Nvidia platform AP 2.0 uses.

The fact that a car equipped with single camera is able to perform complex manovers, makes me feel that all AP 1.0 cars could possibly learn to do these manovers too by virtue of all learning performed by AP 2.0 hardware equipped cars. Only thing to know is if the processing hardware of AP 1.0 would be able to process the DNN model or not.

Though all cars can learn from Nueral network ( using AI/ Deep Learning ) , they will have to rely on local hardware and NOT network connectivity to make realtime decisions. The network will make the collective fleet smarter , but the hardware needs multiple layers of redundancy.

 

[Tam]

...Only thing to know is if the processing hardware of AP 1.0 would be able to process the DNN model or not.

It is unlikely that AP1 is able to process DNN because it has not enough power and especially, it may be hand coded the old fashion way.

That means you have to write programming codes for every conceivable scenarios: that this is a white Tractor-Trailer, this is a Road Sweeper and so on....

AP2 is different. As you can see on the nVidia demo video, you do not need to tell it this is a traffic cone, this is a cement pole, this is a chain-link fence.

When it encounters the dead end with a chain-link fence in front, it does not know what it is.

But it does know that there's free space on the right, so it just turned on its right.

With AP1, after you crash into the chain-link fence, you have to write up a programming code for that scenario....

 

[Tam]

...AP 1.0 cars could possibly learn to do these manovers too...

Mobileye and Tesla

Sorry that my previous answer AP1 which is not correct.

AP1 has Mobileye EyeQ3 chip. Originally, Tesla was planning to upgrade the hardware with 5 eyeQ3 chips (from current 1.)

If you could wait till 2018, then the upgrade only needs 1 EyeQ4 chip.

What I was wrong was: AP1 does have a Deep Learning Network but it does NOT learn in real time.

MobilEye network would learn from the data and then correct the problem for your car may be in a year.

Tesla found such delay is not good for progress so it has its own Fleet Learning Network and the turnaround time is about a week.

So if you still go with MobilEye, it is not very agile!

Since now we have a competent and powerful supercomputer from nVidia, hopefully "Tesla Vision" can learn in real time.