FSD Beta Videos (and questions for FSD Beta drivers)

[Thp3]

We all don’t have FSD yet. Have you watched the other human drivers? I “see” a “error” every drive, often at least once a mile. I imagine that those people watching me also think I’m making mistakes at a similar rate.

I only see an accident Seasonally (~4 times a year, say 15,000 miles, other people’s cars).

So shouldn’t FSD simply be better most of the time that our average licensed driver?

That approach can be argued is how Autopilot with NAS and stoplights was deployed to the Tesla public. It’s better, safer, than driving without it most of the time. Yes, there is phantom braking, yes, some lane changes are iffy, yes Tesla gets close to curbs. But it’s measurably better than us normal drivers.

so FSD will be deployed to the Tesla Masses Real Soon Now. It doesn’t have to be so good that it’s truly autonomous.

I expect February/March 2021 to be general release time...

 

[Thp3]

They are not independent probabilities. The probability of driver A making a safety error is not independent to the probability of driver B making a safety error since drivers can see and react to what other drivers are doing. So, the probability of driver A making a safety error multiplied by the probability of driver B making a safety error is not the probability of an accident happening.

Good point, Bayesian probability applies here. It’s likely that the second driver is less likely to have an accident if the second driver sees the mistake of the first driver.

The upper limit is the same probability if there is nothing the second driver can do, eg whiteout on ice causes massive crashes. But the entire set of cars on the road does not crash. Eventually they slow down and stop. So the probability of an accident with the second driver does diminish eventually to zero.

So the probability of an accident of both the first and second driver is at the maximum the square of the first driver’s probability of an accident. Then it diminishes down to zero as the second driver can take increasing successful avoidance.

 

[TIppy]

They are not independent probabilities. The probability of driver A making a safety error is not independent to the probability of driver B making a safety error since drivers can see and react to what other drivers are doing. So, the probability of driver A making a safety error multiplied by the probability of driver B making a safety error is not the probability of an accident happening.

Dependence would mean that something driver A did would affect the probability of driver B. If driver A changes lanes into driver B in his blind spot, that doesn't change the odds that driver B is distracted. If he drives through a red light, that doesn't change the odds that the driver in the other direction will check for cross traffic before proceeding.
.

 

[aronth5]

Now I know why I probably didn't like my statistics class in high school!

 

[EVNow]

Dependence would mean that something driver A did would affect the probability of driver B. If driver A changes lanes into driver B in his blind spot, that doesn't change the odds that driver B is distracted. If he drives through a red light, that doesn't change the odds that the driver in the other direction will check for cross traffic before proceeding.
.

Dependence does not have to be absolute - some situations cause the other driver to behave differently but others don’t.

But i’ve an issue with that 150k miles per accident. It more frequent and i’ve explained that in prior posts. Small accidents are probably about 1 every 10k or so miles.

 

[diplomat33]

But i’ve an issue with that 150k miles per accident. It more frequent and i’ve explained that in prior posts. Small accidents are probably about 1 every 10k or so miles.

"Nationally, 6.1 million miles of driving by a good driver should result in about 40-60 events, most of which are small dings, 22-27 or which would involve an insurance claim, 12 which would get reported to police and 6 injury crashes."

Waymo Data Shows Superhuman Safety Record. They Should Deploy Today

If we count "small accidents" as small dings and insurance claims that comes to 42 out of 6.1M miles. Doing the math, 6.1M divided by 42 gives us 1 per 145k miles.

So small accidents would be around 1 per 145k miles. So I think the 1 per 150k miles is about right.

 

[buyleonard]

Wasn't there a separate thread for arguments about things like this???

 

[n.one.one]

FSD Beta 8.1 (OLD 10) Beta Renumbering - Modular Roundabout & Hooes Rd 4K - 11:43 - Dave Mac

 

[EVNow]

"Nationally, 6.1 million miles of driving by a good driver should result in about 40-60 events, most of which are small dings, 22-27 or which would involve an insurance claim, 12 which would get reported to police and 6 injury crashes."

Waymo Data Shows Superhuman Safety Record. They Should Deploy Today

If we count "small accidents" as small dings and insurance claims that comes to 42 out of 6.1M miles. Doing the math, 6.1M divided by 42 gives us 1 per 145k miles.

So small accidents would be around 1 per 145k miles. So I think the 1 per 150k miles is about right.

Something wrong there. I’ll look into it and post in a new thread.

150k miles is about 10 years of driving. The avg is not a ding per 10 years.

 

[Mardak]

The car does still stop for the stop sign though. I'm guessing it used the map data to know the stop sign was there.

Yup, this is the same map-based stopping behavior as we have in general release Autopilot. Here's the mapped stop sign: Node: 2267854160 | OpenStreetMap

Sometimes relying on the map data is useful, but other times it can cause FSD beta to swerve switching lanes within an intersection:

Looks like the incorrect lanes=2 data makes the path planner think there's only one destination lane across the intersection, so it biases to the left maybe assuming the "extra" lane on the right is for bikes?

 

[TIppy]

Dependence does not have to be absolute - some situations cause the other driver to behave differently but others don’t.

I suspect the correlation between the probabilities of A and B is quite low, so the product is a good approximation. In my original post, I wasn't trying to come to a precise value. The point was that individual (fsd) error rates can be much higher than the ultimate accident rate. Now, I do agree that fsd shouldn't be in actual accidents more frequently than 1 in 150K miles, but it can make mistakes at a much higher rate than that.

 

[momo3605]

"Nationally, 6.1 million miles of driving by a good driver should result in about 40-60 events, most of which are small dings, 22-27 or which would involve an insurance claim, 12 which would get reported to police and 6 injury crashes."

Waymo Data Shows Superhuman Safety Record. They Should Deploy Today

If we count "small accidents" as small dings and insurance claims that comes to 42 out of 6.1M miles. Doing the math, 6.1M divided by 42 gives us 1 per 145k miles.

So small accidents would be around 1 per 145k miles. So I think the 1 per 150k miles is about right.

Have you ever seen an average 2-3 year old Tesla on the road? They tend to have all 4 wheels curb rashed and a couple of door dings at least. LOL

This may just be anecdotal evidence, but i still disagree with the average of 1 ding per 150,000 miles. I think that's completely wrong. I think it must be 1 police reported crash per 150,000 miles

Checks out based on the data here:
Rates of Motor Vehicle Crashes, Injuries and Deaths in Relation to Driver Age, United States, 2014-2015 - AAA Foundation

 

[n.one.one]

Tesla FSD Beta 10.1 With Stephanie! Bigger Visualizations and Not Much Else - 16:24 - Dirty Tesla
** This is actually 8.1 under the new numbering system. Dirty Tesla explains in a comment: "Hey everyone! Been a bit busy so it took me a few days to get this one out. I have a FSD Beta 8.1 video recorded (I'll explain more about the versioning numbers in that video...) so stay tuned!" **

 

[Huskyf]

"Nationally, 6.1 million miles of driving by a good driver should result in about 40-60 events, most of which are small dings, 22-27 or which would involve an insurance claim, 12 which would get reported to police and 6 injury crashes."

Waymo Data Shows Superhuman Safety Record. They Should Deploy Today

If we count "small accidents" as small dings and insurance claims that comes to 42 out of 6.1M miles. Doing the math, 6.1M divided by 42 gives us 1 per 145k miles.

So small accidents would be around 1 per 145k miles. So I think the 1 per 150k miles is about right.

It a old new from octobre 2020

 

[n.one.one]

Beta FSD - Test Loop 1 - Build 8.1 - 2020.48.35.7 - 27 Jan 2021 - 10:21 - James Locke

 

[Mardak]

I was wondering how do human drivers usually know to ignore these lights in the middle of two streets to avoid suddenly slamming on the brakes? Looks like FSD beta correctly ignored the two red lights on the other side of the barrier but got confused assuming the "light to the right of your lane" was for you:

 

[n.one.one]

I was wondering how do human drivers usually know to ignore these lights in the middle of two streets to avoid suddenly slamming on the brakes? Looks like FSD beta correctly ignored the two red lights on the other side of the barrier but got confused assuming the "light to the right of your lane" was for you:

View attachment 631225

Wow...forget FSD, that's really poor traffic design for humans (as you 'sort of' said)!

 

[n.one.one]

Beta FSD - Test Loop 2 - Build 8.1 - 2020.48.35.7 - 27 Jan 2021 - 4:13 - James Locke

 

[diplomat33]

In the Earnings Call just now, Elon said that there is "almost 1000" people in the FSD Beta group. AFAIK, I think that is the first time, we've gotten an official number on the testers in FSD Beta.

 

[EVNow]

In the Earnings Call just now, Elon said that there is "almost 1000" people in the FSD Beta group. AFAIK, I think that is the first time, we've gotten an official number on the testers in FSD Beta.

External or internal too ?