Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

FSD fails to detect children in the road

This site may earn commission on affiliate links.
I didn't say it was, and most humans are sensible cautious drivers. But not all. And that's the point, once ONE Tesla car is a decent driver they all are. Note most of the cars, all of them.
Sure Tesla may end up being good, but if someone doesn't establish performance standards to this technology I don't want to be on the road with FSD Skoda either (or whatever is the worst brand, someone is bound to be the worst).
 
  • Disagree
Reactions: ItsNotAboutTheMoney
The most it can possibly do is prevent every collision a human driver would have which is 1 per 2 million miles according to Tesla. Even if we assume it prevents all those collisions it still must cause only 1 collision per 2 million miles to be on par with a human.
It can do more. It can also prevent the near-accidents that were only avoided due to the reactions of the other cars. Those are not in the accident tally, but are situations where the driver did something bad.

I had a driver pull out in front of me from a side street, I saw it coming and could brake in time to avoid T-boning them. If FSD had been on that car and not done that, it wouldn't get credit on any current metrics.

Regarding statistics, if an accident involves multiple cars, did that count as multiple incidents or one? If FSD gets in 2 two car accidents for every 5 car accident it prevents, is that improvement ?
 
It can do more. It can also prevent the near-accidents that were only avoided due to the reactions of the other cars. Those are not in the accident tally, but are situations where the driver did something bad.

I had a driver pull out in front of me from a side street, I saw it coming and could brake in time to avoid T-boning them. If FSD had been on that car and not done that, it wouldn't get credit on any current metrics.

Regarding statistics, if an accident involves multiple cars, did that count as multiple incidents or one? If FSD gets in 2 two car accidents for every 5 car accident it prevents, is that improvement ?
While preventing near accidents is important for passenger mental health they are not a safety issue. I would gladly get in a near accident every 10k miles if it meant avoiding a severe accident every 2 million miles. Of course ideally I’d like the car to not crash and also not scare the crap out of me. I suspect the two issues are correlated so I doubt this will be an issue.
The number of severe collisions for a human driven Tesla is about 1 per 2 millions. The minimum number of collisions is zero (no such thing as a negative collision) therefore the number of severe collisions that can be prevented by FSD is 1 per 2 million miles.
Teslas statistics only include Teslas. If two Teslas crash into each other and both have active restraints activated (corerelates to > 12mph) then it will count as two collisions.
 
  • Like
Reactions: AlanSubie4Life
The number of severe collisions for a human driven Tesla is about 1 per 2 millions. The minimum number of collisions is zero (no such thing as a negative collision) therefore the number of severe collisions that can be prevented by FSD is 1 per 2 million miles.
Teslas statistics only include Teslas.
The correct word here is mitigated, not prevented. If the car can turn a serious accident into a fender-bender that I would argue that was a success (though perhaps less of one than avoiding it completely).
 
  • Like
Reactions: mongo
While preventing near accidents is important for passenger mental health they are not a safety issue. I would gladly get in a near accident every 10k miles if it meant avoiding a severe accident every 2 million miles. Of course ideally I’d like the car to not crash and also not scare the crap out of me. I suspect the two issues are correlated so I doubt this will be an issue.
The number of severe collisions for a human driven Tesla is about 1 per 2 millions. The minimum number of collisions is zero (no such thing as a negative collision) therefore the number of severe collisions that can be prevented by FSD is 1 per 2 million miles.
Teslas statistics only include Teslas. If two Teslas crash into each other and both have active restraints activated (corerelates to > 12mph) then it will count as two collisions.
Just so we're clear I'm refering to the following scenario:
Car A drifts into oncoming traffic
Oncoming traffic drives on shoulder to avoid accident.

Are you calling that "not a safety issue"?


Also, I'm referring to federal crash records. If Tesla counts a collision involving two Teslas as two incidents and NHTSA counts it as one (inclusive of any other vehicles), then the comparison is biased against Tesla.
 
Just so we're clear I'm refering to the following scenario:
Car A drifts into oncoming traffic
Oncoming traffic drives on shoulder to avoid accident.

Are you calling that "not a safety issue"?


Also, I'm referring to federal crash records. If Tesla counts a collision involving two Teslas as two incidents and NHTSA counts it as one (inclusive of any other vehicles), then the comparison is biased against Tesla.
Obviously that's a safety issue. Let's say that behavior caused 100% of all severe collisions meaning it occurs 1 per 2 million miles driven. The most collisions that FSD beta could possibly prevent would still be 1 per 2 million miles.

The NHTSA statistics are biased in favor of Tesla because the NHTSA is counting all police reported collisions (about 1 per 500k miles) and Tesla is only counting active restraint use. The two numbers just shouldn't be compared at all.

The correct word here is mitigated, not prevented. If the car can turn a serious accident into a fender-bender that I would argue that was a success (though perhaps less of one than avoiding it completely).
That's what I'm saying. The maximum number of severe collisions that can be mitigated or prevented is 1 per 2 million miles. I think it's helpful to talk about severe collisions because it's something we have a solid number for and it's what we care most about.

Now I suppose it might be possible for a perfect self-driving car surrounded by human drivers to eliminate more collisions by changing the behavior of the drivers around them. Of course what we see with Waymo and Cruise is that drivers are super annoyed by them and crashing while trying to pass them illegally. So I wouldn't bet on that and I certainly don't think it adds significantly to the safety calculation.
 
  • Like
Reactions: AlanSubie4Life
Obviously that's a safety issue. Let's say that behavior caused 100% of all severe collisions meaning it occurs 1 per 2 million miles driven. The most collisions that FSD beta could possibly prevent would still be 1 per 2 million miles.
I was thinking about this some more and if FSD beta could both avoid driving people off the road and not crash when people try to drive it off the road (when a human driver would have crashed) it could eliminate more than the average number of collisions. What percentage of collisions do not involve the driver who caused the collision? My guess is that it's not that many.
 
What percentage of collisions do not involve the driver who caused the collision? My guess is that it's not that many.
I'd say it was a decent percentage. Especially on freeways a common scenario is someone makes a mistake (say, moving into a lane when it was occupied) causing other drivers to swerve and lose control, hitting others cars or barriers etc. Often the original car is off in the distance by that time. Not many people know how to handle cars in these kinds of emergencies, resulting in a magnifying effect as a small swerve by driver A makes driver B do a huge swerve and driver C spin out of control.
 
  • Like
Reactions: mongo
I was thinking about this some more and if FSD beta could both avoid driving people off the road and not crash when people try to drive it off the road (when a human driver would have crashed) it could eliminate more than the average number of collisions. What percentage of collisions do not involve the driver who caused the collision? My guess is that it's not that many.
So in this hypothetical, you’re saying that for these drivers who cause all the collisions without being involved, FSD could not reduce severe collision rates at all, since there are none, while for the rest of the drivers they would have more than the average number of severe collisions prevented? So the weighted average of collisions prevented would be the average number of collisions?

This would only make a big difference in the reduction in accidents for a given driver if most of the drivers who cause accidents are not involved in the accident (and for them there would be no benefit of course). Otherwise it doesn’t change much. If this happens 20% of the time even, it hardly matters.
 
So in this hypothetical, you’re saying that for these drivers who cause all the collisions without being involved, FSD could not reduce severe collision rates at all, since there are none, while for the rest of the drivers they would have more than the average number of severe collisions prevented? So the weighted average of collisions prevented would be the average number of collisions?

This would only make a big difference in the reduction in accidents for a given driver if most of the drivers who cause accidents are not involved in the accident (and for them there would be no benefit of course). Otherwise it doesn’t change much. If this happens 20% of the time even, it hardly matters.
I have no idea what you're saying. haha.
I'm saying that a driver using this perfect version of FSD would have zero collisions. So the 1 per 2 million miles of collisions that would have occurred had they been driving manually would be prevented (they would never run into anything or anyone else and no one would ever run into them!). Not only that they would also not cause any of the collisions that a human driver would have caused but not been involved in. Therefore perfect FSD could actually reduce collisions by more than 1 per 2 million miles.
I'd say it was a decent percentage. Especially on freeways a common scenario is someone makes a mistake (say, moving into a lane when it was occupied) causing other drivers to swerve and lose control, hitting others cars or barriers etc. Often the original car is off in the distance by that time. Not many people know how to handle cars in these kinds of emergencies, resulting in a magnifying effect as a small swerve by driver A makes driver B do a huge swerve and driver C spin out of control.
But of course we have an upper bound on how common it is. We know that 100% or less of collisions are of this type. So, let's assume that 100% of collision are of this type. The best FSD could do is prevent 100% of this type of collision caused by the driver and avoid 100% of this type of collision where the driver is the victim. Therefore the maximum benefit would be reducing the number of collisions by double the expected number of collisions for manual driving. In reality I doubt it makes any difference and in fact it probably works against FSD since my interpretation of Waymo and Cruise data is that avoiding not at fault collisions is harder than avoiding at fault collisions*. Also, Waymo and Cruise appear to change driver's behavior for the worse so they're probably causing collisions that way too (people also report being honked at while using FSD beta, road rage is also not good for preventing collisions).

Anyway, it seems best to be significantly greater than human safety as I doubt you'll ever be able to get your error bars tight enough on this absurd analysis to deploy if it's 1% better.

*for example it may prevent 50% of this type of collision where it would be at fault but get in more than enough collisions where a human driver would have successfully swerved to offset that and net more total collisions.
 
Last edited:
  • Funny
Reactions: AlanSubie4Life
Not only that they would also not cause any of the collisions that a human driver would have caused but not been involved in.
Ah yes, of course, most vehicles are not Teslas. Anyway my comments applied to a situation where every vehicle has this feature.

I was just commenting that if the overall average is 1 per 2 million miles and this is a real phenomenon (and presumably only some drivers have this sort of tendency), then the collision rate for some drivers is probably lower than others. And it wouldn’t help drivers who only cause accidents but are not involved in accidents. Of course it would help the other drivers so they would see a larger reduction (since they would both not be involved in accidents that they cause and would not suffer due to the hypothetical accident-causing-but-not-involved drivers).
 
The number of severe collisions for a human driven Tesla is about 1 per 2 millions. The minimum number of collisions is zero (no such thing as a negative collision) therefore the number of severe collisions that can be prevented by FSD is 1 per 2 million miles.
I think it is misleading to go to integers when the denominator can change. For example it can go from 1 per 2 million to 1 per 2.5 million. Doesn't have to go to zero or negative to be an improvement. You can also use 0.5 collisions per 2 million miles (this is common in other collisions stats where they fix the number of miles and vary only the number of collisions).
 
  • Like
Reactions: Daniel in SD
I think it is misleading to go to integers when the denominator can change. For example it can go from 1 per 2 million to 1 per 2.5 million. Doesn't have to go to zero or negative to be an improvement. You can also use 0.5 collisions per 2 million miles (this is common in other collisions stats where they fix the number of miles and vary only the number of collisions).
I absolutely agree when you're comparing two numbers. However when stating a single number it feels more intuitive to think of it as the average Tesla driver driving manually 2 million miles per severe collision. Like how miles per gallon is more useful for a car you own but gallons per 100 miles is more useful when comparing efficiency.
In my Tesla I think about miles per rated mile not rated miles per mile.:p