FSD fails to detect children in the road

[AlanSubie4Life]

People run over those here... Of course they also run over pedestrian safety islands, ram into planters, etc...

Obviously, you have to look at percentages. I did say typically.

And…as I was the very first to observe based on the affidavit, we don’t know the success rate of FSD Beta in the Dan O’Dowd videos.

Subsequently some have claimed this to be an uncovered case which would not surprise me but I still don’t know the failure rate. FSD Beta is very inconsistent!

 

[Daniel in SD]

How did you compute that? The total driven miles in the US includes a lot of freeway miles. Since you don't often see toddlers playing on the interstate you need to look at deaths per non-freeway miles driven, or possibly exclude rural roads also. At that point, your human safety score is rather lower, though I would argue more realistic.

I commuted it by taking the vehicle miles traveled (VMT) in the United States per year and dividing by number of children pedestrians killed. Google search says 26% of VMT are on Interstate highways and child pedestrians sometimes do get killed on Interstates.
Don't get hung up whether its 1 per billion miles or 1 per 1.7 billion miles. The point is that these are very rare events meaning either children are almost never in the street or humans are actually very good at avoiding even the edge cases. So when we see an example that a human could easily avoid but a self-driving car cannot it needs to be fixed before that system can be taken out of beta if the goal is to achieve greater than human performance.

 

[AlanSubie4Life]

So when we see an example that a human could easily avoid but a self-driving car cannot it needs to be fixed before that system can be taken out of beta if the goal is to achieve greater than human performance.

Yes.

you need to look at deaths per non-freeway miles driven, or possibly exclude rural roads also. At that point, your human safety score is rather lower, though I would argue more realistic

ABC - Always Be Caving - whenever possible, lower the bar for FSD, accept the obvious faults in the beta, and shortchange Tesla and ignore their stated goals.

 

[alexgr]

Are we really making the argument that it is ok to hit small children, or that we need to settle for failed identifications in these cases, and blame the (very) irresponsible parents? It really kind of seems like it, even though that makes no sense. Maybe I am misunderstanding the argument here? Please help. All I know is I have had very young children playing in my cul-de-sac in the past and I am guessing I am not the only one. And even if the children are monitored, will there be time for the parent to run to the child and rescue it from the rolling death machine which ignores such detections?

Nonsense.
1) I bet all little children who run around your cul-de-sac are taller than 32".
2) I bet you have no cars running 40 mph in cul-de-sac.
3) Of course it is not a good idea to hit any person of any size.
4) All the O'Doubt experiment demonstrates is that at this time you still can hit a very small person with Tesla when on FSD if you try really really hard.

 

[MP3Mike]

4) All the O'Doubt experiment demonstrates is that at this time you still can hit a very small person with Tesla when on FSD if you try really really hard.

I would just add at this time "with the prior version of FSD". For all we know the FSD 10.69 resolves this issue. (Or maybe it doesn't.) But don't expect Dan to tell us if it is fixed, he will only say anything if he can engineer another "fail".

 

[AlanSubie4Life]

I would just add at this time "with the prior version of FSD". For all we know the FSD 10.69 resolves this issue. (Or maybe it doesn't.) But don't expect Dan to tell us if it is fixed, he will only say anything if he can engineer another "fail".

If the @Discoducky escape case is proven closed in 10.69, that’s great.

As long as people don’t go around claiming that 10.69 won’t hit children, I’ll be happy with the improvement. Of course, as you suggest, it will nearly certainly still be able to hit arbitrary child-like objects; it’s just a question of whether someone can engineer the situation (which will represent some real-world scenarios).

Anyway, the point is that there be consistent progress and we never accept as acceptable that FSD will fail in cases that are easy for a human to handle (those are the worst because a human might expect FSD to handle it). And ideally it also does well in situations where a human does not do well (I thought that was the point?).

We have to recognize that there are still many capabilities that are lacking or need refinement with FSD. It is nowhere near as capable as an attentive human driver yet. Which is why it is in beta with safety drivers, and why we are squarely responsible for driving at all times.

 

[Dan D.]

I would just add at this time "with the prior version of FSD". For all we know the FSD 10.69 resolves this issue. (Or maybe it doesn't.) But don't expect Dan to tell us if it is fixed, he will only say anything if he can engineer another "fail".

Maybe 10.69 is now 100x better at avoiding small mannequins in the road. That would be great.

What is it's 10x or 100x worse at that thing? How is the FSD beta driver supposed to know. Maybe it drives ULT's better at the expense of something else. I guess you'll figure it out after the first few mistakes. All part of the fun of beta testing.

 

[Dewg]

If the @Discoducky escape case is proven closed in 10.69, that’s great.

As long as people don’t go around claiming that 10.69 won’t hit children, I’ll be happy with the improvement. Of course, as you suggest, it will nearly certainly still be able to hit arbitrary child-like objects; it’s just a question of whether someone can engineer the situation (which will represent some real-world scenarios).

Anyway, the point is that there be consistent progress and we never accept as acceptable that FSD will fail in cases that are easy for a human to handle (those are the worst because a human might expect FSD to handle it). And ideally it also does well in situations where a human does not do well (I thought that was the point?).

We have to recognize that there are still many capabilities that are lacking or need refinement with FSD. It is nowhere near as capable as an attentive human driver yet. Which is why it is in beta with safety drivers, and why we are squarely responsible for driving at all times.

So many good points. People need to stop expecting FSD Beta to be flawless. If it was flawless it'd be released from Beta as a full L4 system. Waymo would be jealous.

Will it hit things? Yes
Will it be jerky making turns? Yes
Will it accelerate and decelerate too aggressively for people? Yes

But as Alan said - as long as we see progress, which we have, that's the important thing. FSD Beta has only been out for 22 months (for comparison Waymo has been testing for over 120 months). I've been in Beta for about 9 months, and I've seen significant improvement in my area (Southern Cali).

So do what you're supposed to do as testers. Drive safely, and attentively. Use Beta as much as you can. Intervene and disengage when necessary and report to Tesla.

 

[Daniel in SD]

But as Alan said - as long as we see progress, which we have, that's the important thing. FSD Beta has only been out for 22 months (for comparison Waymo has been testing for over 120 months). I've been in Beta for about 9 months, and I've seen significant improvement in my area (Southern Cali).

This is not quite fair. Tesla started testing FSD in October 2016. I'm sure there have been many rewrites since then!
From 2016:

 

[drtimhill]

ABC - Always Be Caving - whenever possible, lower the bar for FSD, accept the obvious faults in the beta, and shortchange Tesla and ignore their stated goals.

Nonsense. I'm merely pointing out that the death per miles driven quoted for humans doesnt reflect reality. It's a sad and gruesome truth that children are mostly killed on urban roads, NOT freeways and NOT rural roads. So if you want to rate human drivers you need to look at frequency in those conditions. I said nothing at all about what FSD does/does not do when compared against this adjusted benchmark.

Though it is worth noting that, right now, FSD is only deployed for non-freeway driving, so like it or not if you want an accurate comparison you DO need to compare it to the human rate for non-freeway driving. Once we get single-stack, and FSD is doing freeway driving too, then we can compare total FSD miles to total human miles.

At the end of the day, it all comes down to one thing, and only one thing. Is FSD measurably safer than the average human driver when driving realistically (that is, at typical driving speeds). All the other arguments about magical percentages, driving style and obscure failure modes is irrelevant. Answering this question is hard since it involves defining the measurement criteria both for the car and human drivers. But if the answer to the question is "no", then FSD is still beta (at best) or should be recalled (at worst), while if the answer is clearly "yes" then I dont see how you can argue that it is not safe for general release. I don't see how that is "ABC"?

 

[Daniel in SD]

Nonsense. I'm merely pointing out that the death per miles driven quoted for humans doesnt reflect reality. It's a sad and gruesome truth that children are mostly killed on urban roads, NOT freeways and NOT rural roads. So if you want to rate human drivers you need to look at frequency in those conditions. I said nothing at all about what FSD does/does not do when compared against this adjusted benchmark.

Please post a more accurate estimate based on the real numbers then. If you like you can assume that zero pedestrian deaths occur on freeways or rural roads (which isn't true of course!).
I think it will still be very rare but I look forward to your analysis.

 

[EVNow]

How did you compute that? The total driven miles in the US includes a lot of freeway miles. Since you don't often see toddlers playing on the interstate you need to look at deaths per non-freeway miles driven, or possibly exclude rural roads also. At that point, your human safety score is rather lower, though I would argue more realistic.

Look at the adjusted stats that Cruise uses that I’ve posted. That is just urban/sub-urban non-highway driving. Ofcourse they don’t have a separate kids fatality rate. We should look at L1 crash rates instead.

I consider using kids in these kinds of ads fairly despicable anyway.

 

[sleepydoc]

Please post a more accurate estimate based on the real numbers then. If you like you can assume that zero pedestrian deaths occur on freeways or rural roads (which isn't true of course!).
I think it will still be very rare but I look forward to your analysis.

You posted the original statistics and @drtimhill pointed out a major flaw in them. If you want to use them to make a point it's incumbent on you to show they are accurate or provide more accurate numbers.

 

[sleepydoc]

This thread has become incredibly tiresome - at this point it's just a series of posts by @Daniel in SD and @AlanSubie4Life mischaracterizing what other people have said or making exaggerations or slippery slope arguments to try and agree with a staged video that misrepresents the current state of Tesla's capabilities.

@AlanSubie4Life & @Daniel in SD : I'll just say that there's a lot of space between the extremes that you portray.

 

[Daniel in SD]

You posted the original statistics and @drtimhill pointed out a major flaw in them. If you want to use them to make a point it's incumbent on you to show they are accurate or provide more accurate numbers.

But @drtimhill said that freeway miles would make a significant difference and they don't so his first claim of a major flaw was wrong. I should have rounded down from 1 per 1.7 billion to 1 per billion miles or whatever. I'm sorry for overestimating but it was only because it was the most simple analysis to show that such collisions are rare.
I take it that everyone would agree with me if it turned out to be 1 per 1.7 billion but not if it were 1 per billion?

This thread has become incredibly tiresome - at this point it's just a series of posts by @Daniel in SD and @AlanSubie4Life mischaracterizing what other people have said or making exaggerations or slippery slope arguments to try and agree with a staged video that misrepresents the current state of Tesla's capabilities.

@AlanSubie4Life & @Daniel in SD : I'll just say that there's a lot of space between the extremes that you portray.

With FSD beta doesn't run over children on one side and FSD beta indiscriminately runs over children on the other I would say I'm in the middle.
Actually much closer to FSD beta doesn't run over children if we're going by percentage.
If we're going by required performance to go out of beta I'm much closer to O'Dowd's view.

 

[AlanSubie4Life]

agree with a staged video that misrepresents the current state of Tesla's capabilities.

Keep in mind I was the first one in this forum (to my knowledge) to look at the full video and post that it seemed contrived, and point out that FSD reacted to the obstacles in all three cases. This was well before it appeared in the media (it took a couple days and a few false claims before the media realized this video existed).

So talk about mischaracterizing! I think the video (both the ad and the “full” video) is terribly produced and terribly uniformative. If Dan O’Dowd actually wanted to help reduce the risk of something bad happening, he would explain exactly how he managed to defeat FSD in this case, and report it to Tesla as well.

In the end I just don’t want people to misrepresent the capabilities of FSD or think it is better than it is. That’s dangerous. I have no idea how much this is actually happening, but the responses to the video (on Twitter and here and YouTube), which I realize suffer from massive self-selection bias, to me show a very poor grasp of reality.

 

[drtimhill]

In the end I just don’t want people to misrepresent the capabilities of FSD or think it is better than it is. That’s dangerous. I have no idea how much this is actually happening, but the responses to the video (on Twitter and here and YouTube), which I realize suffer from massive self-selection bias, to me show a very poor grasp of reality.

The reality, of course, is that no-one wants the car to hurt anyone; adult or child, pedestrian or cyclist. But you have also posted here opinions wanting the car to be super-human at driving, which reminds me of the old adage "perfect is the enemy of good". If you push for something at one far end of the spectrum, you are bound to get people responding that you have gone too far, and you then seem to interpret that as people being Tesla shills or fanboys, when often they are just pointing out that a lower bar can still be good. As I've already noted here, what matters is that the car is better then average human drivers. You might WANT it to be far better (and so do I), or meet some criteria that you find personally satisfying, but purely from a safety standpoint the bar is average humans. What else could it be?

 

[Daniel in SD]

The reality, of course, is that no-one wants the car to hurt anyone; adult or child, pedestrian or cyclist. But you have also posted here opinions wanting the car to be super-human at driving, which reminds me of the old adage "perfect is the enemy of good". If you push for something at one far end of the spectrum, you are bound to get people responding that you have gone too far, and you then seem to interpret that as people being Tesla shills or fanboys, when often they are just pointing out that a lower bar can still be good. As I've already noted here, what matters is that the car is better then average human drivers. You might WANT it to be far better (and so do I), or meet some criteria that you find personally satisfying, but purely from a safety standpoint the bar is average humans. What else could it be?

Can you quantify how good FSD beta is relative to the average human driver?
I'm all in favor of taking FSD out of beta once it achieves a statistically significant safety advantage over the human average (20% safer?).

 

[Electroman]

This thread has become incredibly tiresome - at this point it's just a series of posts by @Daniel in SD and @AlanSubie4Life mischaracterizing what other people have said or making exaggerations or slippery slope arguments to try and agree with a staged video that misrepresents the current state of Tesla's capabilities.

@AlanSubie4Life & @Daniel in SD : I'll just say that there's a lot of space between the extremes that you portray.

I have stopped following this thread primarily because of this Alan character. (moderator edit)

 

[alexgr]

I have stopped following this thread primarily because of this Alan character. (moderator edit)

(moderator edit) …he is into energy consumption as far as I know. Difference of opinions is good, otherwise - see Russia.