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

should Tesla increase our EAP feedback?

Discussion in 'Model S' started by JRMW, Jan 28, 2017.

  1. JRMW

    JRMW Member

    Joined:
    Nov 13, 2016
    Messages:
    255
    Location:
    Minneapolis
    So I've been using TACC and Autosteer.
    I know little to nothing about Fleet learning and the Tesla AI.

    But it seems to me that it would be a great idea to us (Beta testers) to have more feedback.
    Today, Autosteer/TACC slammed on the brakes for no obvious reason.

    It's done this twice when a large truck (like Ford F-250) passed me in an adjacent lane. But today I couldn't see any reason for it

    Not sure how The Fleet would know that's an error. By the time I would have adjusted EAP had already hit the accelerator

    Seems like there should be an easy button to push to report it. (I know there is some voice command but don't know how to use it)

    Even better... too bad we can't watch the video of our car in action (like the FSD videos that Tesla placed) so that we could try to figure out what it saw.

    Seems like there are a lot of us who are interested who could give real time human feedback, which could improve AP2 much faster.
     
  2. lunitiks

    lunitiks Memba

    Joined:
    Nov 19, 2016
    Messages:
    1,025
    Location:
    Norway
    False positive, no need to report manually. Car gets better. TTT
     
  3. calisnow

    calisnow Active Member

    Joined:
    Oct 11, 2014
    Messages:
    2,217
    Location:
    Los Angeles
    How does the car and Tesla know a braking event is a false positive and learn from it? Trained learning is a basic techniques with neural nets.
     
  4. lunitiks

    lunitiks Memba

    Joined:
    Nov 19, 2016
    Messages:
    1,025
    Location:
    Norway
    I dont think that the car or Tesla 'learns' that this exact braking event should not have happened. Instead, though thousands of miles in shadow or active mode, the system will eventually 'learn' that no human brakes in these kind of situations. Similar occurances happen enough times, the system will feel it's safe to keep same speed and not decelerate.

    It's kind of like those overhead passes / signs - TACC will initially think they're possible obstacles and slows down for safety's sake. Over time, system will 'whitelist' it. (Not exactly comparable, since the latter example is a stationary object in a certain place that I think the fleet mapping will take care of. But you get the idea.)
     
  5. josher32

    josher32 Member

    Joined:
    Oct 29, 2015
    Messages:
    172
    Location:
    West Linn OR
    #5 josher32, Jan 29, 2017
    Last edited: Jan 29, 2017
    My understanding is the car will remember the location where it thought there was an object to brake for but in fact there was no object and the next time not brake. I know for me my car has braked for a few freeway overhead signs and then later I went through same area without any braking. Below is a helpful description of how the radar uses fleet learning to geocode a database to prevent these false positive braking events.

    Upgrading Autopilot: Seeing the World in Radar
     
    • Like x 1
  6. Joe F

    Joe F Member

    Joined:
    Sep 19, 2016
    Messages:
    307
    Location:
    Willow Grove, PA
    How does that explain the severe level of braking when it happens for a split second, if TACC/AP thought it was a real threat or obstacle? It can't be that it's relying on human intervention to teach it because it's too short in duration to even react to before the "event" is over.

    Unless the microphone picks up the expletives from the driver... ;)
     
    • Like x 1
  7. lunitiks

    lunitiks Memba

    Joined:
    Nov 19, 2016
    Messages:
    1,025
    Location:
    Norway
    Like I said, those false positives will be quite literally wiped out over time, as the mothership (Nvidia DGX-1, I think) gets enough experience from the fleet, telling it that the car shouldn't brake it those types of situations or places
     
  8. lunitiks

    lunitiks Memba

    Joined:
    Nov 19, 2016
    Messages:
    1,025
    Location:
    Norway
    BTW, we should appreciate the fact that TACC brakes more often than it should, instead of NOT braking when it's supposed to. In an autonomous system, the latter is generally much, much worse as it potentially means the car's sensors dont pick up relevant input data (blindness). False positives can of course be devastating, but between the two false negatives is the worst
     
    • Like x 1
  9. Joe F

    Joe F Member

    Joined:
    Sep 19, 2016
    Messages:
    307
    Location:
    Willow Grove, PA
    Unless you're the only one within hundreds of miles traveling the same roads. Small fleet. Great if you're in a highly concentrated Tesla area.

    Time to get out and drive I guess. Got lots of mapping to do. Just have to do it without AP features turned on for sanity's sake. :eek:

    Thinking now a color coded map indicating which roads have been mapped would be a good idea. Would help with trusting that the system isn't going to cause chaos behind me when it brakes hard for no apparent reason to those tailgating.
     
  10. lunitiks

    lunitiks Memba

    Joined:
    Nov 19, 2016
    Messages:
    1,025
    Location:
    Norway
    Good point. I'd really like to have access to Teslas fleet learning map. Probably top secret
     

Share This Page