4 ms·

Why are they removing these features from a car with a promise of patching them back in later? Shouldn’t that be ready to go before committing to a hardware cha

by tapoxi 5y ago

Why are they removing these features from a car with a promise of patching them back in later? Shouldn’t that be ready to go before committing to a hardware change?

akerl_ 5y ago
Because the actual change is driven by supply chain issues w/ radar gear, and the vision-is-better party line is just to make it sell better.
- TheSpiceIsLife 5y ago
  Can someone tell me if SpaceX's tail-landing rockets are vision-only systems? Street vehicle navigation seems at least as difficult as nailing a rocket landing.
  - agogdog 5y ago
    I'd argue that navigating streets is significantly harder than landing a rocket. Space X can land a rocket for example, but fully autonomous cars are still a ways off.
  - mdorazio 5y ago
    They are not. Cameras would not be all that useful. See [1] for an estimated system diagram - the main sensor stacks are based on AHRS and GPS. [1] https://www.reddit.com/r/spacex/comments/64ew8x/falcon_9_a_systems_overview/ https://www.reddit.com/r/spacex/comments/64ew8x/falcon_9_a_s...
  - shadowgovt 5y ago
    It's a bit of an apples-and-oranges comparison regarding "as hard as," but the answer to your question is that the rocket uses a combination of GPS for gross aiming followed by radar for fine-grained approach. Vision might be possible but presents challenges radar may not (including that as the rocket gets close to the pad, the back-blast form the exhaust obscures the pad).
  - gct 5y ago
    Not a chance, at the very least they have a hot shit inertial nav unit if not a full imaging landing radar
  - valine 5y ago
    Humans are proof that's its possible to safely drive a car with pure vision. Pretty sure no human has ever propulsively landed a rocket on earth using nothing but vision.
    - moojd 5y ago
      I don't think humans are a good benchmark for what we should expect from FSD.
      - valine 5y ago
        Not in the short term for sure. I've been pretty impressed with the FSD beta videos circulating youtube though. It's better than you would expect for a non-geofenced, vision based self driving system.
        wearywanderer 5y ago
        Better than I would have expected? I guess we didn't watch the same videos. I expected to see the system repeatedly trying to kill people and that's exactly what I saw in those videos.
        valine 5y ago
        I guess we haven't seen the same videos. I've watched many hours of FSD beta footage and if anything its excessively cautious. Not saying it never makes dangerous mistakes, but the mistakes I've seen it make seem solvable. For example it has trouble with unprotected left turns when there are trees or fences obscuring the view. That to me is understandable and could probably be fixed with an update to the vision stack, maybe train the object detector on cars partially obscured by trees.
        wearywanderer 5y ago
        From what I saw, it drives worse than an intoxicated teenager on the first week of their learner's permit.
- ggreer 5y ago
  I don't know about that. Musk and Karpathy have always been pretty gung-ho about solving self driving with nothing but cameras. We'll know for sure if they ever re-add radar to the Model 3/Y in the US. I think they won't. My main source of, "WTF are you doing, autopilot?" is phantom braking due to the radar incorrectly perceiving an overhead sign as an obstacle.
- azinman2 5y ago
  To me it's such a strange argument. Why wouldn't you want CV + additional sensors? CV has made all kinds of progress in current years, but it's certainly not fool-proof, doesn't do well in low-light or extremely high-light scenarios, can be slow, difficult to debounce between frames, subject to motion blur, difficult to get depth from, subject to optical illusions, etc. There are solutions that aren't full-blown LiDAR that can do a great job of depth estimation, collision detection, etc. Wouldn't you want more sensors in the mix? To me it's a really odd 'party line' to adhere to.
natch 5y ago
They test software changes with different beta fleets long before releasing them. So it could be that the feature is already there. Also there’s a lag time between manufacturing and delivery, so that gives them time to do a software update if needed. And it’s possible the cars manufactured this way will have the software to support vision only from day one.

by tapoxi 5y ago

Why are they removing these features from a car with a promise of patching them back in later? Shouldn’t that be ready to go before committing to a hardware change?

akerl_ 5y ago

Because the actual change is driven by supply chain issues w/ radar gear, and the vision-is-better party line is just to make it sell better.

TheSpiceIsLife 5y ago
Can someone tell me if SpaceX's tail-landing rockets are vision-only systems? Street vehicle navigation seems at least as difficult as nailing a rocket landing.
- agogdog 5y ago
  I'd argue that navigating streets is significantly harder than landing a rocket. Space X can land a rocket for example, but fully autonomous cars are still a ways off.
- mdorazio 5y ago
  They are not. Cameras would not be all that useful. See [1] for an estimated system diagram - the main sensor stacks are based on AHRS and GPS. [1] https://www.reddit.com/r/spacex/comments/64ew8x/falcon_9_a_systems_overview/ https://www.reddit.com/r/spacex/comments/64ew8x/falcon_9_a_s...
- shadowgovt 5y ago
  It's a bit of an apples-and-oranges comparison regarding "as hard as," but the answer to your question is that the rocket uses a combination of GPS for gross aiming followed by radar for fine-grained approach. Vision might be possible but presents challenges radar may not (including that as the rocket gets close to the pad, the back-blast form the exhaust obscures the pad).
- gct 5y ago
  Not a chance, at the very least they have a hot shit inertial nav unit if not a full imaging landing radar
- valine 5y ago
  Humans are proof that's its possible to safely drive a car with pure vision. Pretty sure no human has ever propulsively landed a rocket on earth using nothing but vision.
  - moojd 5y ago
    I don't think humans are a good benchmark for what we should expect from FSD.
    - valine 5y ago
      Not in the short term for sure. I've been pretty impressed with the FSD beta videos circulating youtube though. It's better than you would expect for a non-geofenced, vision based self driving system.
      - wearywanderer 5y ago
        Better than I would have expected? I guess we didn't watch the same videos. I expected to see the system repeatedly trying to kill people and that's exactly what I saw in those videos.
        valine 5y ago
        I guess we haven't seen the same videos. I've watched many hours of FSD beta footage and if anything its excessively cautious. Not saying it never makes dangerous mistakes, but the mistakes I've seen it make seem solvable. For example it has trouble with unprotected left turns when there are trees or fences obscuring the view. That to me is understandable and could probably be fixed with an update to the vision stack, maybe train the object detector on cars partially obscured by trees.
        wearywanderer 5y ago
        From what I saw, it drives worse than an intoxicated teenager on the first week of their learner's permit.
ggreer 5y ago
I don't know about that. Musk and Karpathy have always been pretty gung-ho about solving self driving with nothing but cameras. We'll know for sure if they ever re-add radar to the Model 3/Y in the US. I think they won't. My main source of, "WTF are you doing, autopilot?" is phantom braking due to the radar incorrectly perceiving an overhead sign as an obstacle.
azinman2 5y ago
To me it's such a strange argument. Why wouldn't you want CV + additional sensors? CV has made all kinds of progress in current years, but it's certainly not fool-proof, doesn't do well in low-light or extremely high-light scenarios, can be slow, difficult to debounce between frames, subject to motion blur, difficult to get depth from, subject to optical illusions, etc. There are solutions that aren't full-blown LiDAR that can do a great job of depth estimation, collision detection, etc. Wouldn't you want more sensors in the mix? To me it's a really odd 'party line' to adhere to.

natch 5y ago

They test software changes with different beta fleets long before releasing them. So it could be that the feature is already there. Also there’s a lag time between manufacturing and delivery, so that gives them time to do a software update if needed. And it’s possible the cars manufactured this way will have the software to support vision only from day one.