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
Start a Discussionhttps://teslamotorsclub.com/tmc/tags/

Does anyone actually know how valuable large scale "fleet learning" really is??

Discussion in 'Model S' started by calisnow, Oct 23, 2016.

Tags:
  1. calisnow

    calisnow Active Member

    Joined:
    Oct 11, 2014
    Messages:
    2,570
    Location:
    Los Angeles
    There would be, it seems, two possibilities:

    1 - Large scale fleet learning of the kind Tesla is doing is the only way to iron out the corner cases and build a robust self driving system.

    2 - Large scale fleet learning is not necessary to build a robustly safe system.

    If #1 is the truth - then Tesla still has no competition and its lead is only growing. If #2 is the truth, then fleet learning's value is somewhat hyped up, and competing automakers can jump in the market over the next few years - going from zero autonomous cars to selling them off the shelf.

    If in fact #1 is the truth, as Musk claims, and the other automakers know it - then why in the world are we going into the third year of Tesla being the only automaker with a large scale fleet test/learning project of autonomous driving?

    The other makers could safely implement the exact same thing with zero liability risk if they just ran their fleets in a pure shadow mode - but they are not doing so. Why not? Instead we are seeing small scale deployments of test bed cars bristling with many more sensors than Tesla actually ships in the real world - what seem to amount to never-ending big dollar science projects.

    In terms of real world autonomous driving Tesla is competing with - itself.

    To play devil's advocate for a moment, let's remind ourselves that these other automakers also employ many phd artificial intelligence researchers. Let us presume these researchers are not stupid and that they are not living under a rock - and are thus aware of Tesla's fleet scale approach. Let's also assume these engineers could design and start shipping a fleet-wide learning system in actual vehicles which are for sale - today.

    So why aren't they doing it? Is it in fact, not necessary? If it isn't necessary then what, if anything, does that imply about Tesla's firm public claims that fleet learning is the only viable way to build out a robust neural network that is safe in the real world?
     
  2. Troy

    Troy Active Member

    Joined:
    Aug 24, 2015
    Messages:
    1,113
    Location:
    #2 Troy, Oct 23, 2016
    Last edited: Oct 23, 2016
    Hi. Start listening here from 27:02 to 32:35. Elon said other car companies have no chance without similar data.
     
    • Like x 1
  3. mblakele

    mblakele radial cross member

    Joined:
    Mar 7, 2016
    Messages:
    546
    Location:
    SF Bay Area
    What if other automakers are running in shadow mode? How would we know?

    But I suspect they aren't running in shadow mode, because they aren't thinking like software companies. They think in terms of COGS, so they want to charge for new hardware as soon as it goes into production. Even if they could see their way to deploying HAV-capable hardware quietly, they aren't set up to gather and analyze the massive amounts of data that this would generate.
     
    • Like x 1
  4. Lightmatrix

    Lightmatrix Member

    Joined:
    May 3, 2015
    Messages:
    183
    Location:
    Houston, TX


    That right there- "Even if they could see their way to deploying HAV-capable hardware quietly, they aren't set up to gather and analyze the massive amounts of data that this would generate."

    At the end of the day they are still car companies playing catch-up.
     
  5. bob_p

    bob_p Active Member

    Joined:
    Apr 5, 2012
    Messages:
    1,654
    Think about Tesla's software as a teenage driver.

    It takes time for a teenager to learn how to drive, often starting in a classroom and then transitioning to ride-alongs with another driver in the car (possibly with a second set of controls).

    And even after getting a license, it usually takes several years of driving experience to be proficient in driving in all conditions.

    The AP 2.0 software development will follow a similar path, except that every car being produced now will have the AP 2.0 capability, and Tesla will all of the new owners as "driving instructors", first providing examples of how cars should be driven (in all conditions) and then later for testing to verify AP 2.0 is driving at least as safely as a human driver.

    What Tesla is doing is very clever - effectively "crowdsourcing" AP 2.0 development - by getting a large number of cars on the road for training and testing the AP software.

    Unless other manufacturers take a similar path - and deploy their self driving hardware in cars years before the self driving feature will be ready, then they'll be at a significant disadvantage, because they'll have to use test vehicles or a smaller fleet of cars to do the learning and testing.

    A case might even be made with Tesla, that they should be giving away the AP 2.0 hardware - because of the value they are getting in using the cars for AP learning & testing.

    Wait... They are doing that - every car has the AP 2.0 hardware, and if the owner doesn't activate it now, it can be activated later - which is a benefit Tesla is providing in exchange for using those cars as AP 2.0 trainers...
     
    • Like x 1
  6. Thumper

    Thumper Member

    Joined:
    May 6, 2011
    Messages:
    669
    Location:
    Corvallis, OR
    The missing link for other car companies is management, not engineering expertise. No other company is willing to absorb the abuse from the financial community that a couple of years of expensive R & D for autonomy requires. Tesla has been working on this for a while and the payoff is near.
     
    • Like x 3
  7. Max*

    Max* Not Banned

    Joined:
    Apr 8, 2015
    Messages:
    6,310
    Location:
    NoVa
    This thread is filled with Tesla fanboyism.

    Many other large car companies have autonomous divisions working on their autonomous cars. They just don't go flaunting it, like Tesla does. They also don't release a product that's 90% complete, like Tesla does.

    As of now, it seems that Tesla is ahead of the game leading to L4/L5, but who actually gets there first? Time will tell.



    As for the whole fleet learning, I feel it was spun into a lot more than it's capable of. But I hope I'm wrong.
     
    • Like x 2
    • Disagree x 1
  8. gearchruncher

    gearchruncher Member

    Joined:
    Sep 20, 2016
    Messages:
    445
    Location:
    Seattle, WA
    Tesla spends less than 0.5B a year on R&D. VW spends over $15B. VW could re-direct 3% of their R&D budget and outspend Tesla (on all of Tesla's R&D, not just AP). I don't think this is an R&D money issue.
     
  9. deonb

    deonb Supporting Member

    Joined:
    Mar 4, 2013
    Messages:
    3,462
    Location:
    Redmond, WA
    So true! They release it when it's 40% complete.
     
    • Like x 1
  10. Max*

    Max* Not Banned

    Joined:
    Apr 8, 2015
    Messages:
    6,310
    Location:
    NoVa
    Can you provide some examples?
     
  11. deonb

    deonb Supporting Member

    Joined:
    Mar 4, 2013
    Messages:
    3,462
    Location:
    Redmond, WA
    • Like x 1
  12. Max*

    Max* Not Banned

    Joined:
    Apr 8, 2015
    Messages:
    6,310
    Location:
    NoVa
    You're right, my comment was more in a general sense of software releases*. But when it comes to Level 2 autonomous driving, Tesla is ahead of the same, I don't think anyone doubts that (and yes, I've seen that article before)


    *When it comes to the "less important" stuff like media, nav, etc. Tesla often releases things when they're good enough, but not complete. Whereas the other manufacturers release more polished items.
     
  13. deonb

    deonb Supporting Member

    Joined:
    Mar 4, 2013
    Messages:
    3,462
    Location:
    Redmond, WA
    Yes and no. (Well, yes on the music).

    For example comparing the Tesla Nav with my GMC IntelliLink. So yes, the GMC has waypoints and a bunch of other stuff.

    However, the GMC has the worst voice input system in the world, bar none. I've had MUCH better systems in the early 1990's. It takes me about 15 minutes to enter an address - and when you enter it you have to switch to the main console to finish it with 3 steps, which requires you to take your eyes of the road. And even if you have a passenger (which the car knows BTW.), that passenger can't enter an address - even though it would be MUCH safer for a passenger to just use the keyboard than having the driver distracted with 10 minutes of voice control.

    The Tesla voice recognition works > 95% on the first attempt, and for the rest, my wife can enter it using the keyboard. So yeah, it technically can't do everything that the GMC can do, but the parts that really matter work. The GMC is just a long feature list of things that don't properly work.

    I'd call that a 40% implementation on the GMC part, even though it's technically more "complete".
     
  14. Joe F

    Joe F Member

    Joined:
    Sep 19, 2016
    Messages:
    363
    Location:
    Willow Grove, PA
    So who here believes an "average" driver training AP is a scary thought? For that matter, who here believes they are an average driver?

    I'm surprised by some of the comments RE the radar on 8+ now seeing 2 vehicles ahead. Really, two? Is that the twice as safe as an average driver that has been stated as the target goal?

    I highly doubt anyone here fixates on one or two cars in front of them. To me that's not a safe defensive driving posture. While it might be a life saver in some cases, is it something you'd consider as being a better driver than you? I don't. To me it's not enough to compete with the situational awareness that I'm sure all here employ without hesitation.

    To me, if I'm not scanning everything as far as I can see down the road, I'm not doing a good job behind the wheel. There are times I'm slowing down and on the brake before several cars in front of me are, and I'm sure others here fall into that same defensive style too.

    In 50 years of driving I've seen a lot of crazy things on the roads that are going to take more than I fear is being taking into consideration by all who are working to improve safety. Sure hope I'm wrong.

    V2V communications in addition with AP might be an answer, but legacy vehicles will take a long time to cycle out I'm sure.

    Still want an S with AP2 as I feel it's going to be the best stopgap we'll see for a while, but I would not trust it with my life and that of my friends and family to the extent I've read on various threads here. Sleep between points A and B? Could be a long sleep. :(
     
  15. thimel

    thimel Member

    Joined:
    Feb 27, 2015
    Messages:
    413
    I'm dubious that shadow mode is sufficient to test self driving software. It seems to me that many miles with an expert driver ready to take over are necessary to test it.
    Hitting the brakes to avoid hitting an object is a case where I understand using shadow mode to test self driving. If the driver hits the brakes and shadow mode did not, that info along with all the sensor data can be sent to the mothership for examination and improvement.
    However, consider changing lanes to pass a slow car. There are many possible times to do this. The fact that self driving software in shadow mode and the real driver choose to do it at different times is not an error worthy of sending to the mothership. I don't see a way to test automatic lane changes without making that aspect of the self driving software active and having the driver ready to take over if the lane change would be dangerous. This is just one example. I'm sure there are many others.
    Can anyone think of a way to use shadow mode to test whether the self driving software properly knows when it is safe to change lanes?
     
  16. Lightmatrix

    Lightmatrix Member

    Joined:
    May 3, 2015
    Messages:
    183
    Location:
    Houston, TX

    Auto Makers Struggle With High-Tech Dashboard Screens


    They cannot even nail down the infotainment systems of vehicles, are you sure that these autonomous systems are anywhere near as advanced? I cannot answer that.
     
    • Like x 1
  17. jspayneii

    jspayneii Member

    Joined:
    Jul 16, 2016
    Messages:
    28
    Location:
    Saint Petersburg, FL
    From a machine learning perspective, the key is data. The more data you have, the better the algorithm will learn. It also helps to handle cases that the software hasn't seen before.

    I don't know why the existing automakers are not doing it, other than they are not worried about Tesla. In the grand scheme of things, even though Tesla has the hardware for level 5, they don't have the software to support it. That will take time.
     
  18. bob_p

    bob_p Active Member

    Joined:
    Apr 5, 2012
    Messages:
    1,654
    A report from Rand concluded that hundreds of millions or billions of miles of driving data may be required to adequately verify the self driving software is safe, which they also concluded would never be possible with small fleets of test cars.

    If Tesla has 100,000 AP 2.0 cars on the road (which could be possible by late next year when Model 3 goes into production), each driving (on the average 50 miles per day) - that's 5 million miles of data - every day. As Tesla continues to improve the self driving software, they can roll it out to their cars - have the software shadow driver's actions - and quickly build up data verifying how well the software is doing compared to the human drivers.

    Even driving 24x7 test fleets will never be large enough to get anywhere close to enough real-world testing to validate the software.
     
    • Like x 1
  19. Max*

    Max* Not Banned

    Joined:
    Apr 8, 2015
    Messages:
    6,310
    Location:
    NoVa
    That's not true.

    It highly depends on the type of data you have. More is not always better.
     

Share This Page