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 or making a Paypal contribution here: paypal.me/SupportTMC

Mobileye and Tesla

Discussion in 'Model S: User Interface' started by Cosmacelf, Nov 3, 2015.

  1. Cosmacelf

    Cosmacelf Active Member

    Joined:
    Mar 6, 2013
    Messages:
    3,399
    Location:
    San Diego
    Just listened to the Mobileye 3Q conference call (Nov 3). Some notes WRT autonomous driving and Tesla:

    Current technology is good only for lane keeping assist (what Telsa is delivering now). Mobileye noted that Tesla uniquely is willing to push the system faster and to its limits more than any other automaker.

    True Autopilot will be rolled out starting next year and will require an 8 camera suite along with integration with radar and sonar sensors. This will require 5 eyeQ3 chips. Autopilot software will take many years to fully roll out.

    In the future, a single eyeQ4 chip will handle 8 cameras.

    Current Mobileye technology builds on 16 years of standard programming techniques, but adds a deep learning network (neural net) for free space identification and path prediction. Mobileye software also currently handles reducing speed on curves and smooth speed modulation when cars cut in front, etc. The deep learning network is a traditional neural net setup in that it does NOT learn in real time. It is a supervised learning system that requires full sensor suite recording and processing by a big computer array before creating a model that can be put into their chips.

    Mobileye has 2 production agreements for semi-autonomous driving for 2016 (presumably one is Tesla), another manufacturer coming on line in 2017 and one more in 2018.

    Tesla's current self learning capability is Tesla specific. Mobileye is not involved in this at all. Mobileye gets its training data when OEMs do validation testing. I guess the big automakers do a lot of validation testing and Mobileye thinks this is sufficient for their purposes (they pretty much dismissed the Tesla fleet info and said that the validation data set was huge).

    My comments based on this and Elon's Autopilot conference call:

    It is clear that Tesla isn't just taking what Mobileye is giving them. According to Elon, they have augmented the base Mobileye system with high resolution GPS maps, driver assisted learning, fleet learning, and rapid updates. The high resolution GPS maps, along with driver input, allows the car to behave correctly in ambiguous situations. If Mobileye technology is confused about what is going on in a particular area, it will always be confused in that area (barring very slowly changing model updates which may come out once a year (my estimate)). So Tesla tracks what the human does in that area, averaged over the fleet. If the Mobileye model says to follow the right lane marker at a certain GPS co-ordinate and exit the freeway, while the driver grabs the wheel and keeps the car on the freeway, then the Tesla software learns that behavior for next time and additionally uploads that info for fleet learning. According to Elon, fleet learning updates could occur weekly (presumably more often in the heavier traveled areas).

    In my opinion, this is a huge advantage for Tesla. Mobileye touts their millions of miles driven for their deep learning model, yet Tesla accumulates a million miles of driving per DAY across the fleet.

    According to Elon, Tesla has about 50 software people working on autopilot, and about 100 working on hardware. I don't know if this was a mis-statement, but that strikes me as odd. What are all those hardware people doing, exactly? Hmmm.

    From the forum comments, it seems that Tesla's Autopilot is pretty good already. It is a real step up in functionality (it is a God send for people stuck in stop and go freeway traffic everywhere), that will get better over time. So, another major win for Tesla.

    It is also clear that another hardware refresh is going to occur sometime, probably next year, to at least add more cameras.
     
  2. mknox

    mknox Well-Known Member

    Joined:
    Aug 7, 2012
    Messages:
    8,556
    Location:
    Toronto, ON
    Thanks for that report. I was just thinking about whether AP hardware refreshes would be coming, and I concur with your thoughts. I wonder how Tesla will handle "classic AP" cars once (or if) the new AP cars start coming out?
     
  3. apacheguy

    apacheguy Sig 255, VIN 320

    Joined:
    Oct 21, 2012
    Messages:
    3,709
    Location:
    So Cal
    Why spread out the workload among 5 Q3 chips when just a single Q4 chip could be used?
     
  4. mspisars

    mspisars Member

    Joined:
    May 23, 2014
    Messages:
    493
    Location:
    Charlotte, NC
    This just tells me that the integration of Q4 chip is not ready for prime time...
     
  5. Max*

    Max* Autopilot != Autonomous

    Joined:
    Apr 8, 2015
    Messages:
    4,866
    Location:
    NoVa
    Probably the way they handle classic Model S's. "haha, neeneer neeneer, go buy a new one if you want to upgrade" :p

    - - - Updated - - -

    I think they said they're releasing the Q4 chip soon, but it's only going to hit mass production in a couple years due to...?
     
  6. Cosmacelf

    Cosmacelf Active Member

    Joined:
    Mar 6, 2013
    Messages:
    3,399
    Location:
    San Diego
    Q4 chip is only in engineering samples soon. It is possible that the deep learning model isn't actually upgradable in these chips but instead hard wired in. That would account for a long time delay between engineering samples and production chip.
     
  7. Gizmotoy

    Gizmotoy Active Member

    Joined:
    Sep 16, 2013
    Messages:
    3,137
    Location:
    Bay Area, CA
    It's an interesting number as well. I wonder what kind of matrix they're connected in to get to a 5 chip requirement. Clearly there's overlap with some cameras going to multiple chips.
     
  8. scottm

    scottm Active Member

    Joined:
    Jun 13, 2014
    Messages:
    1,276
    Location:
    Canada
    'splains why V7 is so ... erm... great!
     
  9. Cyberax

    Cyberax Member

    Joined:
    Jul 28, 2015
    Messages:
    270
    Location:
    Seattle, WA
    Possible, but unlikely. There were many attempts to utilize hardware neural network accelerators, but they don't work that well. Reasonably priced FPGAs can simulate only a small number of nodes and ASICs are too expensive to produce.
     
  10. Cosmacelf

    Cosmacelf Active Member

    Joined:
    Mar 6, 2013
    Messages:
    3,399
    Location:
    San Diego
    Not quite what I was thinking. I assume the model is encoded in some large memory store. Rom memory is going to be a lot faster to access than flash memory...
     
  11. Cyberax

    Cyberax Member

    Joined:
    Jul 28, 2015
    Messages:
    270
    Location:
    Seattle, WA
    It'll be loaded in RAM for evaluation, anyway. The neural net itself is stateful, so you don't get a lot of advantage by keeping its topology in ROM.
     
  12. mn4az

    mn4az Member

    Joined:
    Aug 27, 2015
    Messages:
    151
    Location:
    USA
    Production version of the chip in 2018.

    http://www.prnewswire.com/news-releases/moving-closer-to-automated-driving-mobileye-unveils-eyeq4-system-on-chip-with-its-first-design-win-for-2018-300045242.html

    "The first design win for EyeQ4® in series production has been secured for a global premium European car manufacturer for production to start in early 2018. The EyeQ4® would be part of a scalable camera system starting from monocular processing for collision avoidance applications, in compliance with EU NCAP, US NHSTA and other regulatory requirements, up to trifocal camera configuration supporting high-end customer functions including semi-autonomous driving. The EyeQ4® would support fusion with radars and scanning-beam lasers in the high-end customer functions."

    Given that Tesla is "just" Lane Assist Right now (which is killer) I am very likely going to put my order on hold. With all the announcements by Mobileye during their recent briefing it leads me to believe that around the Model 3 reveal (if not sooner - I hope!) Tesla will announce the Model S will get the additional EyeQ3 chips and cameras to enable true semi-autonomous driving.
     
  13. Stoneymonster

    Stoneymonster Active Member

    Joined:
    Jan 8, 2013
    Messages:
    1,554
    Location:
    Aptos, Ca
    ROM as you are thinking of it has been largely replaced by Flash in nearly every application. Real masked ROMs are pretty rare outside the very high volume, low cost chip realm.
     
  14. Cosmacelf

    Cosmacelf Active Member

    Joined:
    Mar 6, 2013
    Messages:
    3,399
    Location:
    San Diego
    It is a lot of R&D for both Tesla and Mobileye to go from lane keeping to semi autonomous. Honestly, I would not put anything on hold waiting for it. It may very well take longer than expected.
     
  15. Cosmacelf

    Cosmacelf Active Member

    Joined:
    Mar 6, 2013
    Messages:
    3,399
    Location:
    San Diego
    I stand corrected. Thanks. Presumably flash downloaded to RAM for performance...
     
  16. mn4az

    mn4az Member

    Joined:
    Aug 27, 2015
    Messages:
    151
    Location:
    USA
    In this video on Amnon Shashua of Mobil Eye on March 23, 2015 said that, “By 2016 there’s going to be new launches by GM and Tesla as well.”

    https://youtu.be/0UzVBTgHqSQ?t=839

    Pair that up with the recent Mobileye quarterly briefing, the Citibank presentation by MobilEye's CEO and the fact that MS mules have been spotted in the wild with 8 cameras - I believe there is a very compelling case that additional Mobileye hardware (multiple EyeQ3s and cameras) is coming to Tesla cars in 2016. My hope is that during the M3 reveal, or MUCH sooner they say something to the effect of, "starting today new MS cars have the option to get the new hardware".
     
  17. Stoneymonster

    Stoneymonster Active Member

    Joined:
    Jan 8, 2013
    Messages:
    1,554
    Location:
    Aptos, Ca
    That's all well and good, but given their track record, there won't be new software to take advantage of it.
     
  18. mn4az

    mn4az Member

    Joined:
    Aug 27, 2015
    Messages:
    151
    Location:
    USA
    Given that I am going to buy and hold for 10+ years I am totally fine with the software lagging the hardware. :)
     
  19. jpet

    jpet Jan P.

    Joined:
    Aug 8, 2014
    Messages:
    3,288
    Location:
    Herne & Leuven, BE
    What about redundancy? Isn't that one of the hardware requirements in a fully autonomous vehicle?
     
  20. heltok

    heltok Member

    Joined:
    Aug 12, 2014
    Messages:
    12
    Location:
    Sweden
    For robustness purposes?
     

Share This Page