3 ms·

> You can't simulate sensors with high enough fidelity to be meaningful. You can if you invest in good simulators. See: https://waymo.com/blog/2021/06/Simulati

by ra7 3y ago

> You can't simulate sensors with high enough fidelity to be meaningful. You can if you invest in good simulators. See: https://waymo.com/blog/2021/06/SimulationCity.html https://waymo.com/blog/2021/06/SimulationCity.html > One way we ensure statistical realism in our simulated world is by creating realistic conditions for our autonomous driving technology to experience. So, if the Waymo Driver is going through spring showers in Detroit at sunset in SimulationCity, we can recreate raindrops on our sensors and even simulate other minute details such as the dimming light and solar glare. If they can simulate rain drops, it's definitely possible to simulate traffic lights. They are a basic scenario for a self driving car. What this video highlights is the extreme lack of validation culture within Tesla FSD.

lolinder 3y ago
My point isn't that you can't simulate it at all, my point is that at some point you have to take it out of the simulation and it will behave differently in the real world than it did in the simulated one, because your simulation will never be good enough to account for everything. We can't expect people to never test self-driving cars on public, populated roads unless we are just acknowledging that we don't believe self-driving cars should be a thing, because I sure as hell am not comfortable with the idea of cars going straight from simulation to production.
- bluefishinit 3y ago
  > I sure as hell am not comfortable with the idea of cars going straight from simulation to production. This is what you're advocating for by saying people should test these things on public roads. That is production, none of the other drivers have consented to be part of the "test".
- ra7 3y ago
  FSD makes basic errors, the kind that can easily be identified in simulation rather than requiring 450k cars on the road. They don't have to go straight from simulation to production either. That's why all the serious players have closed course testing and safety driver supervised testing.
- Veserv 3y ago
  If your simulation and regression testing is so bad that it can not be used to reliably validate 1 in 10,000 mile failure modes, then you are not ready to validate on public streets with untrained consumers. Tesla FSD routinely fails at 1 in 100 mile safety-critical failure modes. The simulation does not need to be perfect, it just needs to be 100x better than what they have and FSD needs to be 100x better before they should be allowed to inflict it on general consumers. If they want to test with trained safety drivers in a robust and well-controlled real world environment to identify failure modes that need to be addressed, that is fine. But non-criminal usage by consumers is at least a factor of 100x away.

by ra7 3y ago

lolinder 3y ago

My point isn't that you can't simulate it at all, my point is that at some point you have to take it out of the simulation and it will behave differently in the real world than it did in the simulated one, because your simulation will never be good enough to account for everything. We can't expect people to never test self-driving cars on public, populated roads unless we are just acknowledging that we don't believe self-driving cars should be a thing, because I sure as hell am not comfortable with the idea of cars going straight from simulation to production.

bluefishinit 3y ago
> I sure as hell am not comfortable with the idea of cars going straight from simulation to production. This is what you're advocating for by saying people should test these things on public roads. That is production, none of the other drivers have consented to be part of the "test".
ra7 3y ago
FSD makes basic errors, the kind that can easily be identified in simulation rather than requiring 450k cars on the road. They don't have to go straight from simulation to production either. That's why all the serious players have closed course testing and safety driver supervised testing.
Veserv 3y ago
If your simulation and regression testing is so bad that it can not be used to reliably validate 1 in 10,000 mile failure modes, then you are not ready to validate on public streets with untrained consumers. Tesla FSD routinely fails at 1 in 100 mile safety-critical failure modes. The simulation does not need to be perfect, it just needs to be 100x better than what they have and FSD needs to be 100x better before they should be allowed to inflict it on general consumers. If they want to test with trained safety drivers in a robust and well-controlled real world environment to identify failure modes that need to be addressed, that is fine. But non-criminal usage by consumers is at least a factor of 100x away.