4 ms·

They have a relatively small team when compared to Waymo, at least at the research/state-of-the-art level. Don't quote me on this but just look at the the diffe

They have a relatively small team when compared to Waymo, at least at the research/state-of-the-art level. Don't quote me on this but just look at the the difference between autopilot team size vs Waymo employees size who have a PhD in a related area. It'd at least 10x more ay Waymo, probably even more. Hardware and HD maps makes a lot of difference too. Planning and decision making always becomes easier when you have a perfect estimate of the environment and good predictions of other agents behaviors. Intent matters a lot more than just recognition/detection in this setting. To be fair I'd say Autopilot team is doing pretty good for the team of their size. They are no where close to Waymo, that's super clear, but if they focused all of their stack on making driver assistive tech rather than trying to push towards FSD they'd probably be better off.

Alex3917 3y ago
> They have a relatively small team when compared to Waymo, at least at the research/state-of-the-art level. Sure, but in every independent test, Tesla's smart driving capabilities get completely trounced by Subaru EyeSight. Does Subaru also have vastly more PhD researchers than Tesla?
- londons_explore 3y ago
  It looks like of the features both systems have, Subaru EyeSight performs better... But Tesla FSD has vastly more abilities (some of which it performs well at, some middling, and some poor).
- alsodumb 3y ago
  I'm not following the news on this, but what metrics do these independent tests use and what aspects do they cover? Are they driver assistance features?
- nradov 3y ago
  Subaru EyeSight works very well in my experience, but it's much less ambitious. It basically just does lane keep assist, automatic braking, and dynamic cruise control. It can't be used at all in heavy precipitation and will automatically deactivate. The cruise control also has an annoying tendency to unnecessarily slow down on the freeway when you're in the slow lane and the vehicle in front takes an exit ramp.
- porphyra 3y ago
  I'm sorry but there's no way that Subaru EyeSight can "trounce" FSD. Maybe it is slightly better in some scenarios than legacy Autopilot which came out many years ago and hasn't been updated since. Tesla FSD is vastly more capable and can handle city driving with traffic controls whereas EyeSight merely just follows a lane.
  - Alex3917 3y ago
    > I'm sorry but there's no way that Subaru EyeSight can "trounce" FSD. I mean that's what literally all of the data shows, e.g.: https://www.caranddriver.com/features/a24511826/safety-features-automatic-braking-system-tested-explained/ https://www.caranddriver.com/features/a24511826/safety-featu...
    - porphyra 3y ago
      AEB has nothing to do with self driving or Autopilot.
      - Alex3917 3y ago
        It's literally the most foundational part of self driving.
        porphyra 3y ago
        Not really. 1. the AEB may run on a different codepath than FSD 2. AEB needs to decide when it is appropriate to override the driver, e.g. if the driver is behaving as though they are paying attention, e.g. by repeatedly stepping on the accelerator, it is undesirable to override the driver's deliberate actions even if it appears that the driver is about to run into something. On the other hand, when the car is driving autonomously, it has full control to avoid running into anything at all. Also, instead of trusting unscientific tabloid news with inconsistent methodologies, why not use the standard assisted driving scores from official government tests, such as the Euro NCAP? Teslas score by far the highest. https://www.euroncap.com/en/ratings-rewards/assisted-driving-gradings/ https://www.euroncap.com/en/ratings-rewards/assisted-driving...
londons_explore 3y ago
I also see questionable engineering decisions. For example, some reverse engineering showed that there was an overlay map of problematic stop signs and road signals that was downloaded per-road on all navigation routes. Surprise surprise, as that data got out of date, the car would randomly slam the brakes on as the overlay data told it there must be a stop sign, when in fact it was a temporary sign from construction months ago. If you're aiming for end-to-end neural networks, you shouldn't have hard overlay facts like this. Ideally you get all the behaviour you need from curating training data, but if you can't do that the most you should have location specific info-vectors that nudge the network into making the right decision if needed. Such info-vectors can be used for all kinds of things, like "this city has aggressive drivers", "this street has a blind corner", "there's a dog at this house who loves to run under wheels", "this country drives on the other side of the road", etc.
- sheepshear 3y ago
  If the list is of objects they can't yet reliably detect, then how would they implement your suggestion to detect those objects? I'm sure they have a lower tolerance for false negatives, so of course the known problems must be hard-coded.
  - londons_explore 3y ago
    Lets imagine there is a stop sign that is half hidden by a bush and pretty faded, so the stop sign detection logic says "only 5% chance it's a stop sign". That in turn isn't enough to make the car stop. The hardcoding approach says "This is a sign. 100% sure.". The vector approach says "There's a hard to see stop sign around here, boost up the probability of anything stop-sign-ish a bunch". The difference functionally is that nothing in the real world is ever 100% certain. So you should never tell any bayesian machine (which a neural network effectively is) that anything is 100% true. The vector approach I outlined is far more general than the above though - it allows any behaviour of the car to be tweaked automatically or manually. location-specific vectors can be learned from data, and/or put in by operatives. The way the neural net trains, the meaning of a vector could 'evolve' too - for example, whenever a human puts in that there is a hidden stop sign, the neural net might learn that that means other human drivers might occasionally fail to see the sign and stop in those locations. Even though it had never witnessed a human failing to stop in this specific location, it has learnt that is part of the meaning of this vector.
etrautmann 3y ago
they don't use lidar, which seems like a strange choice at this point.

Alex3917 3y ago

> They have a relatively small team when compared to Waymo, at least at the research/state-of-the-art level. Sure, but in every independent test, Tesla's smart driving capabilities get completely trounced by Subaru EyeSight. Does Subaru also have vastly more PhD researchers than Tesla?

londons_explore 3y ago
It looks like of the features both systems have, Subaru EyeSight performs better... But Tesla FSD has vastly more abilities (some of which it performs well at, some middling, and some poor).
alsodumb 3y ago
I'm not following the news on this, but what metrics do these independent tests use and what aspects do they cover? Are they driver assistance features?
nradov 3y ago
Subaru EyeSight works very well in my experience, but it's much less ambitious. It basically just does lane keep assist, automatic braking, and dynamic cruise control. It can't be used at all in heavy precipitation and will automatically deactivate. The cruise control also has an annoying tendency to unnecessarily slow down on the freeway when you're in the slow lane and the vehicle in front takes an exit ramp.
porphyra 3y ago
I'm sorry but there's no way that Subaru EyeSight can "trounce" FSD. Maybe it is slightly better in some scenarios than legacy Autopilot which came out many years ago and hasn't been updated since. Tesla FSD is vastly more capable and can handle city driving with traffic controls whereas EyeSight merely just follows a lane.
- Alex3917 3y ago
  > I'm sorry but there's no way that Subaru EyeSight can "trounce" FSD. I mean that's what literally all of the data shows, e.g.: https://www.caranddriver.com/features/a24511826/safety-features-automatic-braking-system-tested-explained/ https://www.caranddriver.com/features/a24511826/safety-featu...
  - porphyra 3y ago
    AEB has nothing to do with self driving or Autopilot.
    - Alex3917 3y ago
      It's literally the most foundational part of self driving.
      - porphyra 3y ago
        Not really. 1. the AEB may run on a different codepath than FSD 2. AEB needs to decide when it is appropriate to override the driver, e.g. if the driver is behaving as though they are paying attention, e.g. by repeatedly stepping on the accelerator, it is undesirable to override the driver's deliberate actions even if it appears that the driver is about to run into something. On the other hand, when the car is driving autonomously, it has full control to avoid running into anything at all. Also, instead of trusting unscientific tabloid news with inconsistent methodologies, why not use the standard assisted driving scores from official government tests, such as the Euro NCAP? Teslas score by far the highest. https://www.euroncap.com/en/ratings-rewards/assisted-driving-gradings/ https://www.euroncap.com/en/ratings-rewards/assisted-driving...

londons_explore 3y ago

I also see questionable engineering decisions. For example, some reverse engineering showed that there was an overlay map of problematic stop signs and road signals that was downloaded per-road on all navigation routes. Surprise surprise, as that data got out of date, the car would randomly slam the brakes on as the overlay data told it there must be a stop sign, when in fact it was a temporary sign from construction months ago. If you're aiming for end-to-end neural networks, you shouldn't have hard overlay facts like this. Ideally you get all the behaviour you need from curating training data, but if you can't do that the most you should have location specific info-vectors that nudge the network into making the right decision if needed. Such info-vectors can be used for all kinds of things, like "this city has aggressive drivers", "this street has a blind corner", "there's a dog at this house who loves to run under wheels", "this country drives on the other side of the road", etc.

sheepshear 3y ago
If the list is of objects they can't yet reliably detect, then how would they implement your suggestion to detect those objects? I'm sure they have a lower tolerance for false negatives, so of course the known problems must be hard-coded.
- londons_explore 3y ago
  Lets imagine there is a stop sign that is half hidden by a bush and pretty faded, so the stop sign detection logic says "only 5% chance it's a stop sign". That in turn isn't enough to make the car stop. The hardcoding approach says "This is a sign. 100% sure.". The vector approach says "There's a hard to see stop sign around here, boost up the probability of anything stop-sign-ish a bunch". The difference functionally is that nothing in the real world is ever 100% certain. So you should never tell any bayesian machine (which a neural network effectively is) that anything is 100% true. The vector approach I outlined is far more general than the above though - it allows any behaviour of the car to be tweaked automatically or manually. location-specific vectors can be learned from data, and/or put in by operatives. The way the neural net trains, the meaning of a vector could 'evolve' too - for example, whenever a human puts in that there is a hidden stop sign, the neural net might learn that that means other human drivers might occasionally fail to see the sign and stop in those locations. Even though it had never witnessed a human failing to stop in this specific location, it has learnt that is part of the meaning of this vector.

etrautmann 3y ago

they don't use lidar, which seems like a strange choice at this point.