Forums

When will AP2 be on par with AP1?

When will AP2 be on par with AP1?

Still waiting on my new 100D and wondering about this as I love AP1 on long trips in my 70D. Any ideas?

cbmeeker2 | 8 mars 2017

@Dofpic

Same question, same to cars. Now that I have had AP1, any less functionality in EAP would be totally unacceptable.

ACH | 8 mars 2017

Am I the first one on this thread that gets to say "soon"?....:)

On a more serious note...based on what Elon has indicated on tweets and earnings call, end of March. But I think if you poll this forum I'd say the over/under on end of March is 80/20.

tesla | 8 mars 2017

I'm getting the impression that building and training the NN on the AP2 platform is much more complex than originally thought. My guess is you are looking until at least June-August until you will get full highway speed AP2 on par with AP1

The flying hippie | 8 mars 2017

I'm curious to know if the drive px computer in the new cars actually learns individually by watching how it's owner drives and comparing it to what it would have done or if it only learns over time through software updates consisting of what Tesla tells it to do in a given situation.

BigD0g | 8 mars 2017

@The flying hippie, I seriously doubt it. I pass the same overhead sign everyday coming home from work, and everyday I get a mild breaking event, and yesterday I got a major braking event from the same sign. so it's cleary not learning. I suspect what happens, is the car reports the false positives back to the mothership, and then we get a software update to make it recognize more. Think of it like software, it's just running some libraries that tell it the rules to drive under that it follows, but it doesn't make up it's own rules, only follows the rules it's been given. Then, when drivers interdict to change a behavior this is fed back to the fleet, so in the future algorithms it will understand better how to handle a given situation.

tesla | 8 mars 2017

@The flying hippie --

The Neural Net is running in the background in 'shadow mode' and it records the difference between the decision it would have made, and the decision actually made by the driver. If you aggregate enough of this data over thousands of drivers and thousands of situations you can determine the "correct" course of action for any given situation.

That information is sent back to Tesla, and using that 'error data' they train the Neural Net on a server farm that probably contains hundreds, or maybe thousands, of Nvidia GPUs. Given that the EAP/FSD neural net probably contains tens-of-millions of neurons, the single Drive PX computer in the car would never be able to perform the training in a timely fashion.

Once trained and verified, Tesla delivers the updated neural net to the cars via OTA software update, and the cycle repeats.

The flying hippie | 8 mars 2017

I just googled it and it seems deep learning in the px2 is capable of supervised and unsupervised learning. Although capable it seems for now Tesla is only allowing supervised learning. This probably is why you still have issues passing the same sign every day. If unsupervised learning was enabled it surely would have corrected its actions over time. My AP2 car barely has 60 miles on it so I haven't tried out the AP on it yet but will do some road tests soon and compare to AP1 under similar conditions.

BigD0g | 8 mars 2017

I don't get to use AP much at all, these are all TACC events i'm experiencing. I suspect for unsupervised learning, that it requires 2 px boards, we only have 1, either way, i'd love to read and learn more, do you have a link ?

tesla | 8 mars 2017

Tesla is not performing training of these Neural Nets in the vehicle, at least not for the live copy that executes when using TACC, AutoSteer, etc. That would produce the very dangerous side-effect of different copies of the NN executing differently on every single vehicle. This would be incredible dangerous, because it would produce a randomly trained Neural Net that was continually diverging.

Rather, they use the collected data to train a single 'production' neural-net that gets sent to every car.

The flying hippie | 8 mars 2017

No link, its just from Wikipedia and the nvidia site. I googles key words like deep learning and nvidia px2. Pretty much the first few hits. If it's your TACC doing that do you think maybe the radar unit is aimed a little high? Did it display anything in your instrument display? Maybe it can be adjusted by service?

BigD0g | 8 mars 2017

I got a full collision warning event from my instrument cluster, no idea regarding what's causing it, but it appears to be a fairly common problem with AP2 based on these boards and tesla motors club. I've reported the events to service and autopilotfeedback@tesla.com, I haven't been asked yet to bring the car in to the service center for this, but who knows if that will change in the future.

Sam_S | 8 mars 2017

Perhaps the programming manager,at,Tesla underestimated what mobileye accomplished with AP1.

As other software engineers can attest, it wouldn't be the first time a software project took twice as much work as originally estimated.

The flying hippie | 8 mars 2017

Thanks for the heads up, I'll be on the lookout for that issue.

jbrowning1 | 8 mars 2017

I recently traded in my P 90 D for the new P 100 D I think I might have made a mistake. I never drove any of my other cars when my Tesla operated correctly now that the hardware does not function I am driving other cars this is not the way it supposed to be. I hope they can make the new hardware work like the old hardware in my other car it was really fun to drive with everything working. Waiting till june/july is going to be a bummer..

drklain | 8 mars 2017

So you have a new Tesla where the AP functionality isn't working as well as on your old Tesla and your solution is to choose to drive a car that has NO AP capability and give up the other benefits of driving a Teska? How does that make any sense?