3 ms·

It is illegal to turn right there without signaling though. > What I'm speculating happened here is the new obstacle detection failed and decided that the road

It is illegal to turn right there without signaling though. > What I'm speculating happened here is the new obstacle detection failed and decided that the road ahead was blocked Know what would be able to tell if the road ahead is blocked? Literally any kind of three dimensional sensing that isn't just single frame image object classification. It's a real shame that no one knows how to do that...

modeless 5y ago
Not signaling in advance for what the car likely considers to be an emergency maneuver is a pretty nitpicky criticism. The problem is the unnecessary emergency maneuver, not the lack of signaling. As for lidar, well lidar isn't perfect either. You might get the same behavior from a lidar car if a trash bag happened to be blown into the road ahead by the wind. All types of sensors rely on machine learning to classify their observations. Lidar isn't a shortcut that eliminates that machine learning dependency. Finally, calling what Tesla is doing "single frame image object classification" really betrays your bias here. That hasn't been an accurate description of the system for a long time.
- BugsJustFindMe 5y ago
  I don't care about lidar. What I want is _any_ system that starts from a foundational basis of establishing a geometric model of the environment to determine "flat ground" vs "thing" without first trying to classify its pixels and ignoring the things it can't identify. Humans do that without lidar just fine, but Tesla's failures consistently look like it doesn't even try to establish geometry with its cameras.
  - modeless 5y ago
    That's exactly what they're doing now, predicting depth without relying on object classification first. That's why it doesn't run into the pillars anymore. Obviously it needs a lot of work, but to me it seems plausible that the general approach can work. My biggest concerns are that the cameras aren't good enough and/or that there isn't enough compute on the car.
    - BugsJustFindMe 5y ago
      > predicting depth Better than predicting would be enough cameras to get a stereo baseline. But I'm glad at least that 6 years after launch they're finally maybe sort of thinking about it poorly.
      - modeless 5y ago
        Yes, they really ought to dramatically upgrade their cameras. No argument there. Their stereo baseline is too small and the cameras have different lenses so it's unclear if they can even get any stereo information at all. The problem for Tesla is that they already sold FSD on cars that have the existing camera setup. I think there's a good chance that they will get to a point where they realize their existing cameras aren't good enough and will have to either come up with retrofit kits or refund owners of existing cars with FSD.
        nikau 5y ago
        > think there's a good chance that they will get to a point where they realize their existing cameras aren't good enough I think it's less realize, and more being forced to acknowledge
- BugsJustFindMe 5y ago
  > Finally, calling what Tesla is doing "single frame image object classification" really betrays your bias here. That hasn't been an accurate description of the system for a long time. I might think differently if the moon hadn't turned into a traffic light a few months ago. But it did. That kind of failure only happens if you aren't sensing environment geometry. Many of their error scenarios very strongly mimic what you'd get if you looked at a single frame, identified things in it, and then used the vehicle's own motion to hallucinate those things moving by zero-update dead reckoning.
  - duskwuff 5y ago
    > That kind of failure only happens if you aren't sensing environment geometry. The truck carrying traffic lights is another fine example: https://twitter.com/FSD_in_6m/status/1400207129479352323 https://twitter.com/FSD_in_6m/status/1400207129479352323

modeless 5y ago

Not signaling in advance for what the car likely considers to be an emergency maneuver is a pretty nitpicky criticism. The problem is the unnecessary emergency maneuver, not the lack of signaling. As for lidar, well lidar isn't perfect either. You might get the same behavior from a lidar car if a trash bag happened to be blown into the road ahead by the wind. All types of sensors rely on machine learning to classify their observations. Lidar isn't a shortcut that eliminates that machine learning dependency. Finally, calling what Tesla is doing "single frame image object classification" really betrays your bias here. That hasn't been an accurate description of the system for a long time.

BugsJustFindMe 5y ago
I don't care about lidar. What I want is _any_ system that starts from a foundational basis of establishing a geometric model of the environment to determine "flat ground" vs "thing" without first trying to classify its pixels and ignoring the things it can't identify. Humans do that without lidar just fine, but Tesla's failures consistently look like it doesn't even try to establish geometry with its cameras.
- modeless 5y ago
  That's exactly what they're doing now, predicting depth without relying on object classification first. That's why it doesn't run into the pillars anymore. Obviously it needs a lot of work, but to me it seems plausible that the general approach can work. My biggest concerns are that the cameras aren't good enough and/or that there isn't enough compute on the car.
  - BugsJustFindMe 5y ago
    > predicting depth Better than predicting would be enough cameras to get a stereo baseline. But I'm glad at least that 6 years after launch they're finally maybe sort of thinking about it poorly.
    - modeless 5y ago
      Yes, they really ought to dramatically upgrade their cameras. No argument there. Their stereo baseline is too small and the cameras have different lenses so it's unclear if they can even get any stereo information at all. The problem for Tesla is that they already sold FSD on cars that have the existing camera setup. I think there's a good chance that they will get to a point where they realize their existing cameras aren't good enough and will have to either come up with retrofit kits or refund owners of existing cars with FSD.
      - nikau 5y ago
        > think there's a good chance that they will get to a point where they realize their existing cameras aren't good enough I think it's less realize, and more being forced to acknowledge
BugsJustFindMe 5y ago
> Finally, calling what Tesla is doing "single frame image object classification" really betrays your bias here. That hasn't been an accurate description of the system for a long time. I might think differently if the moon hadn't turned into a traffic light a few months ago. But it did. That kind of failure only happens if you aren't sensing environment geometry. Many of their error scenarios very strongly mimic what you'd get if you looked at a single frame, identified things in it, and then used the vehicle's own motion to hallucinate those things moving by zero-update dead reckoning.
- duskwuff 5y ago
  > That kind of failure only happens if you aren't sensing environment geometry. The truck carrying traffic lights is another fine example: https://twitter.com/FSD_in_6m/status/1400207129479352323 https://twitter.com/FSD_in_6m/status/1400207129479352323