3 ms·

Tl;dr: The critical flaw is random braking in the mandatory crash avoidance system because they replaced radar with cameras + computer vision which has a lot of

by tpmx 5y ago

Tl;dr: The critical flaw is random braking in the mandatory crash avoidance system because they replaced radar with cameras + computer vision which has a lot of false positives in their current production software release. I'm curious: Is the reason cost, availability or ideology?

tziki 5y ago
It's widely thought to be due to the chip shortage. Tesla made numerous other adjustments due to chip shortage too.
adam_arthur 5y ago
I would guess a bit of both. In theory cameras should work fine, given that humans are pretty good at driving by just using our eyes. Of course vision will never be theoretically better than radar/laser. But should be good enough in the long run, and cheaper/more flexible. My question is, what level of resolution and processing power is necessary to interpret inputs accurately though? Just anecdotally, seems like autopilot is many years away from being close to autonomous. Self driving is an 80-20 kind of problem... The 20% of hard stuff seems completely unsolved by Tesla, judging by YouTube videos, reviews etc. Personally I think Google will beat them to the punch by many years
- FireBeyond 5y ago
  The cameras in a Tesla are not 576mp (which is the closest approximation or estimate of our eyes), and nor do they have 21 stops of dynamic range (same). And they also don't have 16+ years, minimum, of 24/7 "non-machine learning" to augment their judgment.
  - jefftk 5y ago
    That's not even counting the generations of evolution that went into pre-training the neural structure
  - dillondoyle 5y ago
    This source says they are only 1280×960 which ML reduces to 160x120 depth grid? Seems crazily low resolution to me and also somewhat amazing they can do what they can with that (which I think doesn't matter much as it's still scary if they can't avoid obvious concrete obstacles or do the opposite brake ing op says) https://twitter.com/greentheonly/status/1412597377228226562 https://twitter.com/greentheonly/status/1412597377228226562 https://www.reddit.com/r/teslamotors/comments/bgqekv/what_is_the_model_3s_camera_resolution/ https://www.reddit.com/r/teslamotors/comments/bgqekv/what_is...
    - muglug 5y ago
      Yeah, humans with eyesight that bad are not allowed on the road (for good reason).
      - stefan_ 5y ago
        Not a lot of those around at age 16, which is the last time anyone cares to check.
  - ravi-delia 5y ago
    Our eyes are only really high resolution in a tiny patch, and after some initial learning most of vision is a solved problem by the age of 3. That's not to say that Tesla is any closer to using cameras and machine learning to replicate human vision, just pointing out the comparison doesn't really make sense. There's no reason to think that 576mp cameras and 16 years of training are necessary or sufficient. We'll just have to wait and find out.
  - growt 5y ago
    I disagree with your second paragraph. Not all of your first 16 years are spent in driving situations (especially not 24h a day) but Tesla’s have driven many years in total and almost all of them can be applied to improve the AI running in all the cars.
    - adam_arthur 5y ago
      Anybody familiar with the design Tesla uses for self driving? Do they simply input video clips of common scenarios and what the correct output should be? Is it just a single neural net, or are there rules hard coded in? Multiple neural nets handling different things? I assume there are many approaches to automated driving that are effectively dead ends. But would be interesting if just having enough video clips/correct corresponding behavior was enough to build a fully functioning model.
      - belltaco 5y ago
        https://youtu.be/j0z4FweCy4M?t=2795 https://youtu.be/j0z4FweCy4M?t=2795
- kevin_thibedeau 5y ago
  A camera will never match radar range finding in adverse conditions. Those times where you'd like the safety features you're paying a premium for to actually work.
  - adam_arthur 5y ago
    That's true, though it may be acceptable if more economical. For most geographies, very poor visibility scenarios are quite rare. If it costs 10k more per car to support radar for example, then it's likely not worth it. If the cost difference is marginal, then I agree, may as well use whatever sensors are best.
fiddlerwoaroof 5y ago
I think what happened is they had a bunch of orders they couldn’t deliver because they couldn’t get shipments of the radar part in 2020. So, they decided to go vision-only for autopilot and now they’re scrambling to get to parity.

by tpmx 5y ago

tziki 5y ago

It's widely thought to be due to the chip shortage. Tesla made numerous other adjustments due to chip shortage too.

adam_arthur 5y ago

I would guess a bit of both. In theory cameras should work fine, given that humans are pretty good at driving by just using our eyes. Of course vision will never be theoretically better than radar/laser. But should be good enough in the long run, and cheaper/more flexible. My question is, what level of resolution and processing power is necessary to interpret inputs accurately though? Just anecdotally, seems like autopilot is many years away from being close to autonomous. Self driving is an 80-20 kind of problem... The 20% of hard stuff seems completely unsolved by Tesla, judging by YouTube videos, reviews etc. Personally I think Google will beat them to the punch by many years

FireBeyond 5y ago
The cameras in a Tesla are not 576mp (which is the closest approximation or estimate of our eyes), and nor do they have 21 stops of dynamic range (same). And they also don't have 16+ years, minimum, of 24/7 "non-machine learning" to augment their judgment.
- jefftk 5y ago
  That's not even counting the generations of evolution that went into pre-training the neural structure
- dillondoyle 5y ago
  This source says they are only 1280×960 which ML reduces to 160x120 depth grid? Seems crazily low resolution to me and also somewhat amazing they can do what they can with that (which I think doesn't matter much as it's still scary if they can't avoid obvious concrete obstacles or do the opposite brake ing op says) https://twitter.com/greentheonly/status/1412597377228226562 https://twitter.com/greentheonly/status/1412597377228226562 https://www.reddit.com/r/teslamotors/comments/bgqekv/what_is_the_model_3s_camera_resolution/ https://www.reddit.com/r/teslamotors/comments/bgqekv/what_is...
  - muglug 5y ago
    Yeah, humans with eyesight that bad are not allowed on the road (for good reason).
    - stefan_ 5y ago
      Not a lot of those around at age 16, which is the last time anyone cares to check.
- ravi-delia 5y ago
  Our eyes are only really high resolution in a tiny patch, and after some initial learning most of vision is a solved problem by the age of 3. That's not to say that Tesla is any closer to using cameras and machine learning to replicate human vision, just pointing out the comparison doesn't really make sense. There's no reason to think that 576mp cameras and 16 years of training are necessary or sufficient. We'll just have to wait and find out.
- growt 5y ago
  I disagree with your second paragraph. Not all of your first 16 years are spent in driving situations (especially not 24h a day) but Tesla’s have driven many years in total and almost all of them can be applied to improve the AI running in all the cars.
  - adam_arthur 5y ago
    Anybody familiar with the design Tesla uses for self driving? Do they simply input video clips of common scenarios and what the correct output should be? Is it just a single neural net, or are there rules hard coded in? Multiple neural nets handling different things? I assume there are many approaches to automated driving that are effectively dead ends. But would be interesting if just having enough video clips/correct corresponding behavior was enough to build a fully functioning model.
    - belltaco 5y ago
      https://youtu.be/j0z4FweCy4M?t=2795 https://youtu.be/j0z4FweCy4M?t=2795
kevin_thibedeau 5y ago
A camera will never match radar range finding in adverse conditions. Those times where you'd like the safety features you're paying a premium for to actually work.
- adam_arthur 5y ago
  That's true, though it may be acceptable if more economical. For most geographies, very poor visibility scenarios are quite rare. If it costs 10k more per car to support radar for example, then it's likely not worth it. If the cost difference is marginal, then I agree, may as well use whatever sensors are best.

fiddlerwoaroof 5y ago

I think what happened is they had a bunch of orders they couldn’t deliver because they couldn’t get shipments of the radar part in 2020. So, they decided to go vision-only for autopilot and now they’re scrambling to get to parity.