TMC is an independent, primarily volunteer organization that relies on ad revenue to cover its operating costs. Please consider whitelisting TMC on your ad blocker and becoming a Supporting Member. For more info: Support TMC

Speculation about EAP and FSD codelines

Discussion in 'Autonomous Vehicles' started by Matias, May 9, 2017.

Tags:
  1. Matias

    Matias Active Member

    Joined:
    Apr 2, 2014
    Messages:
    1,972
    Location:
    Finland
    How will Tesla merge EAP and FSD code lines? Or will they be merged?

    I think Tesla is currently developing two separate code lines; EAP and FSD. My belief is based to the fact, that Tesla demoed FSD already last year and is still developing EAP to reach parity with AP1 and beyond. Furthermore, Elon is speaking about the autonomous demo drive across the continent in December.

    So they must have different FSD branch at the moment, right?

    Disclaimer; I'm 100% layman.
     
  2. Derek Kessler

    Derek Kessler Member

    Joined:
    Apr 15, 2016
    Messages:
    279
    Location:
    Cincinnati
    The ongoing improvements to EAP isn't the code still being written — it's the activation of already-coded features after receiving enough active driving data to validate that it works as intended. There's no doubt some tweaking that's been happening with that data, but largely the EAP code is complete.

    FSD is built on top of that — things like lane keeping, adaptive cruise control, summon, self-parking, etc. are all part of EAP and the basis for the further functions of FSD.

    But with something as monstrously complex as EAP and FSD, I have no doubt that there are numerous development branches within Tesla for each core feature.
     
  3. lunitiks

    lunitiks Memba

    Joined:
    Nov 19, 2016
    Messages:
    1,027
    Location:
    Norway
    I think it is simpler than that. This post adequately explains the process:
     
    • Like x 1
    • Funny x 1
  4. DurandalAI

    DurandalAI Member

    Joined:
    Dec 20, 2016
    Messages:
    437
    Location:
    Little Rock, AR
    As a whole, I agree, but basic road rules must be applied on top of that as well. Essentially if you look at "big data" analytics, (even though big data is now a cliche term) the larger and more granular data sets you can provide these analytic tools, the more patterns emerge that previously were unknown. These patterns can be automatically found by the tools without even specifying what to look for. I think the same can be applied to fleet learning for driving.

    Plenty of people pooh pooh the value of this fleet data, but I think it will be amazingly important and differentiating compared to other driving AI platforms.

    But I agree with the topic as a whole, that I think there are two programs being developed, one is AP2 simply trying to mirror AP1 functionality, and the FSD program, which has little to do with AP2 beyond the hardware.
     
  5. brkaus

    brkaus Active Member

    Joined:
    Jul 8, 2014
    Messages:
    1,797
    I believe EAP will remain similar. The neural networks will continue to improve and detect more stuff. And it will do s better job if placing it in the lane.

    FSD is in development in a branch and will rely on high def map data to roughly control the car. It will know traffic signals, signs, lanes, exact radius of curves, etc. the EAP will position the car accurately within the map data.

    If you believe the lvl5 / payver sales pitch, Tesla must be pretty far along.

    I'm waiting to see the Tesla center council "game" to somehow entice people to manually drive all the road network. It works for Waze and payver is taking a pay per point/mile approach.
     
  6. verygreen

    verygreen Curious member

    Joined:
    Jan 16, 2017
    Messages:
    624
    Location:
    TN
    Sorry to burst your bubble, but there is no fleet data, whatever you mean by this.
    Also(I am no expert on this, so don't quote me here) if you don't train NN first with stuff of interest, I am not even sure it can recognize patterns, or at least it probably would quickly decide that all of Earth is covered with pavement, most of cars are black, all cars have wheels and the upper 1/4 of all pictures is blue for various values of blue as seen by onboard cameras.
    (if you refer to self learning NNs that play videogames, those don't just look at pictures, they have means of doing stuff and then seeing what happens and learn via that).

    Now, to train your NN with useful data you need a good dataset first. After that you can feed it with new data and see how it does. So just hire a bunch of interns to watch the process and correct the NN.
    Or, in case of Tesla, they have a better option.
    They can downsample the pictures from those random 10 seconds clips and feed them into mobile eye chip and see what it reports, then feed the results into their own model somehow (or compare the two models and make corrections to theirs).
    Once their model is superior - they can move to some next thing ;)

    Oh, btw, another thing. You know who has more data than Tesla from the road? Blackvue and other such companies that offer cloud storage for dashcams. And it's not in tiny 10 second chunks too.
    Also google/youtube ;)

    Yeah, right, that's why the already written code undergoes drastic transformations, I guess, because it's already finished and they are just activating new abilities after validating.
     
    • Love x 3
    • Informative x 1
    • Funny x 1
  7. AnxietyRanger

    AnxietyRanger Well-Known Member

    Joined:
    Aug 22, 2014
    Messages:
    6,166
    Location:
    EU
    These posts and the rest of the Frustrated with FSD timeline thread are still relevant reading IMO.


     
    • Love x 1
  8. AnxietyRanger

    AnxietyRanger Well-Known Member

    Joined:
    Aug 22, 2014
    Messages:
    6,166
    Location:
    EU
    Some more past posts that are relevant:

    Current HW2 Autopilot using 2 of 8 cameras * Testing Inside *

    Enhanced Autopilot 2 may never happen before Full Self Driving is ready!
     
    • Love x 1
  9. Reciprocity

    Reciprocity Active Member

    Joined:
    Feb 27, 2017
    Messages:
    1,736
    Location:
    Chicagoland
    One hater vs one believer. We will find out by December.
     
  10. Bladerskb

    Bladerskb Like how many times do i have to be right?

    Joined:
    Oct 24, 2016
    Messages:
    601
    Location:
    Michigan
    #10 Bladerskb, Jul 29, 2017
    Last edited: Jul 29, 2017
    Hater? what.. every single thing i have said have been bulleyes while your post has several inaccuracies like Lidar not working in rain, snow plus radar being better than lidar. Lastly you can't confirm what you're doing with something that has a 30 degree FOV. 30 FOV can't back up 360 FOV of a camera.

    Me on the other hand has been spotless.

    Including The number of cameras used, The use of HD mapping and data collection for object recognition, EAP being delayed until FSD codebase is ready and actual EAP features appearing Aug-Sept and many more.

    This is with most of my predictions coming in Dec 2016. Before issues were apparent.
    That's too much of a perfect record to be a hater. Haters are typically blindly negative and because of that are mostly wrong.

    highlight

     
  11. verygreen

    verygreen Curious member

    Joined:
    Jan 16, 2017
    Messages:
    624
    Location:
    TN
    Hey, what's TSLA going to cost in August and September (of 2017 I assume), while you are still in that spotless future-seeing mode? ;)
     
    • Funny x 4
  12. croman

    croman Active Member

    Joined:
    Nov 21, 2016
    Messages:
    1,262
    Location:
    Chicago, IL
    I'd take September 2017 for EAP -- that sound downright optimistic!
     
    • Like x 1
  13. Bladerskb

    Bladerskb Like how many times do i have to be right?

    Joined:
    Oct 24, 2016
    Messages:
    601
    Location:
    Michigan
    #13 Bladerskb, Jul 30, 2017
    Last edited: Jul 30, 2017
    Next month is the calm before the storm. Theres some dark days ahead and u ppl will say alas alas that great company for in one hour thy stocks dissipated
     
    • Disagree x 1
  14. zmarty

    zmarty Member

    Joined:
    Jun 25, 2017
    Messages:
    20
    Location:
    Kirkland, WA
    As per the investors' call today, the cross country autonomous drive is still on track. Worst case scenario early next year.
     
  15. BigD0g

    BigD0g Member

    Joined:
    Jan 12, 2017
    Messages:
    222
    Location:
    VA
    Yes, and a pig will fly across the solar eclipse on Monday. Sorry, but given the current state of AP2, unless it's Elon sending 500 cars across the country in the hopes that "one" makes it, it's just not happening, and let's not even get into "soon". If we are lucky Dec 2018, maybe....

    I REALLY REALLY want to be wrong on this.
     
    • Like x 1
    • Disagree x 1
    • Love x 1
  16. lunitiks

    lunitiks Memba

    Joined:
    Nov 19, 2016
    Messages:
    1,027
    Location:
    Norway
    [​IMG]
     
    • Like x 2
  17. DurandalAI

    DurandalAI Member

    Joined:
    Dec 20, 2016
    Messages:
    437
    Location:
    Little Rock, AR
    Well, prepare to be wrong. I sincerely doubt that AP2 is running on the same code base as FSD, just in the basic behavior of current AP2 versus what was in the FSD demo. AP1 and AP2 can't even recognize relatively simple things such as red lights, green lights, or freeway exit navigation based upon your entered destination and planned route. If they were using a subset of FSD as part of AP2, you would see some of these more easily implemented features put into AP2, but you don't. It's because AP2 was developed merely as a temporary replacement as AP1 with MobilEye was lost.

    When we see these more basic features such as traffic light recognition in place, we will know they've switched over the code base to a subset of FSD.
     
  18. BigD0g

    BigD0g Member

    Joined:
    Jan 12, 2017
    Messages:
    222
    Location:
    VA
    I'm very happy to be wrong on this.... However, I suspect i'm not, we've already seen a leak of stop sign detection and indications in the debug screens for AP. I don't think @verygreen has found any "notion" of a completely separate code base.

    reference: Tesla’s new Autopilot update detected and displayed stop signs, but it didn’t act on them
     
    • Informative x 1
  19. verygreen

    verygreen Curious member

    Joined:
    Jan 16, 2017
    Messages:
    624
    Location:
    TN
    Whole idea of a separate codebase is that it is separate - i.e. no signs of it in the unrelated codebases.

    Anywaym in the ape code the is a very separate thing - the NN code and trained model - those are drop-in bits from somewhere else. Whenever it's a different codebase or there's a different team doing them and only providing periodic snapshots - I don't know, but there's this datapoint. Literally nothing else (but third-party code drops like electrobit) look the same ,everything else seems to be built from the same coedbase at the same time monolitically.
     
    • Informative x 1
    • Like x 1
  20. DurandalAI

    DurandalAI Member

    Joined:
    Dec 20, 2016
    Messages:
    437
    Location:
    Little Rock, AR
    • Like x 1

Share This Page