Safety Score

[EVNow]

We're saying the same thing.

I guess I'm using common industry terms.

Basically, data prep is a big part of any data science venture. You acquire the data, cleanse it, transform it so that it becomes useable. Then you apply whatever models you have come up with (using clean data).

You can call cleansing & transformation part of the "algorithm" if you want. But the current formula they are using definitely calls for clean data - otherwise you would need a lot more parameters and a complex logic.

 

[AlanSubie4Life]

No way would the group be that small if it included people with 99.8% scores with one or two dings here and there.

I think seeing all the people posting here has led you to overestimate the driving skill of the average Tesla driver.

 

[EVNow]

I think seeing all the people posting here has led you to overestimate the driving skill of the average Tesla driver.

Actually - even here few have "real" perfect 100 i.e. with no infractions at all.

 

[AlanSubie4Life]

like take into account road gradient when looking for hard braking.

Don't count FCW in the garage.

Both of these things seem like they should count to me. Not once have I had an FCW in my garage and if I did I'd have to question how safely I am pulling into it.

Road gradient should most certainly be taken into account for hard braking since as currently implemented it's a measure of how hard the tires are working. Which, since they're the only thing keeping you in contact with the road, seems like a good thing. EDIT: I mean current implementation seems correct without need for "cleansing" on that particular metric.

 

[EVNow]

Road gradient should most certainly be taken into account for hard braking since as currently implemented it's a measure of how hard the tires are working. Which, since they're the only thing keeping you in contact with the road, seems like a good thing.

Do you mean road gradient should be taken in to account separately ?

Either way - its clear to me clubbing all the "hard braking" together masks what one is trying to measure - attentiveness / carefulness of the driver (which is correlated with accident risk). Basically the same person would brake harder on downhill than on flat roads.

 

[AlanSubie4Life]

Do you mean road gradient should be taken in to account separately ?

No, I mean it should be harder to get good hard braking scores on downhills than it is on uphills.

I'm saying it's currently implemented correctly.

This becomes relevant when your tires become incapable of providing 0.3g of stopping force (which is not rare as anyone who lives in Portland or Seattle in winter knows).

 

[momo3605]

Actually - even here few have "real" perfect 100 i.e. with no infractions at all.

How many? I don't think we have a poll for perfect 100s. But My guess is very very few people have perfect 100s. There's no way to really understand the thresholds until you get dinged by them a couple times lol. Like who would guess that tapping the brake lightly will ding you on hard braking? Once I figured that out, i stay off the brake pedal. But had to have that first ding to know.

 

[gearchruncher]

But the current formula they are using definitely calls for clean data - otherwise you would need a lot more parameters and a complex logic.

Correct, it clearly does. Which means it is totally irresponsible to release it running on real time data that is not clean, yet represent it as any kind of measure of driver safety. It's not a measure of this, it's just a lottery given you know the actual data will be full of false positives.

But what's the point of deriving an algorithm that only works on clean data if you have no way to feed it clean data?

Yet I guarantee this won't stop Elon from claiming FSD was only released to "the safest drivers" even though the analysis was run using dirty data into an algorithm that can only use clean data.

 

[EVNow]

No, I mean it should be harder to get good hard braking scores on downhills than it is on uphills.

I'm saying it's currently implemented correctly.

This becomes relevant when your tires become incapable of providing 0.3g of stopping force (which is not rare).

I don't think so - but either of us could be right.

Only way to figure out would be to separate that parameter out and run regressions.

 

[StealthLSU]

anyone know how many decimal places are rounded for score?

By my weighted calc for my score, I had a 96 day one and have been 99 for a while. I drove some today and I have my score at 99.4733, which still shows a 99 score.

So is it rounded at the tenths place? So I need a straight up 99.5 to get a 100 score?

 

[gearchruncher]

This becomes relevant when your tires become incapable of providing 0.3g of stopping force (which is not rare).

I'm not sure I'd call 0.3G "not rare" - that's basically only ice that can drop the coefficient of friction that low. Rain and even most snow doesn't do that. When was the last time you encountered less than 0.3G tire capability while driving in San Diego?

But it's almost like the algorithm should understand weather and traction, not just ding everyone for 0.3G because there is a 0.1% chance you were on ice at that moment. I'd imagine safe driving is tied to being highly sensitive to your situation, not driving like you are always in the worst case situation humans have ever encountered.

 

[scottf200]

He said the number of 100s will increase by Friday. That sounds like people can improve their scores which would then round to 100. How else can the number of 100s increase otherwise

I think by people opting in for the first time tomorrow (Fri), taking a short drive, getting 100 and parking. /s haha

 

[AlanSubie4Life]

anyone know how many decimal places are rounded for score?

By my weighted calc for my score, I had a 96 day one and have been 99 for a while. I drove some today and I have my score at 99.4733, which still shows a 99 score.

So is it rounded at the tenths place? So I need a straight up 99.5 to get a 100 score?

Each individual day is first rounded, and then the mileage weighted average is calculated, and then rounded.

There are a bunch of spreadsheets posted here, you can plug in your data.

So if you had x miles on the first day, you need 7x miles at a score of 100 to get back to "100" (99.5).

Obviously have to adjust accordingly for any other non 100 scores. Just use the spreadsheets.

 

[Ramphex]

For funsies someone should record the G data for AP braking, I feel like sometimes it does some careless braking that’s way worse than any of my regen braking that causes me to have dings

 

[AlanSubie4Life]

When was the last time you encountered less than 0.3G tire capability while driving in San Diego?

I've not encountered it in San Diego but I have encountered in in Mammoth Lakes. Note that I added the details of what I was talking about before you had a chance to respond. I've also driven my car to both Portland and Seattle (not in winter yet though).

But in any case it doesn't matter - the traction limit could be 0.6g for all I care - the point is that the tires are working harder when you're stopping on a downhill, and it's important to capture that, and I think it's very likely that people who have higher braking forces (the force measured by the score) due to stopping on hills have higher accident rates. Excursions above a threshold of 0.3g are very likely correlated with excursions above a threshold of 0.7g (or whatever). The threshold they chose doesn't really matter that much.

I can say that my stops have been very very safe for the last two weeks. Just have to pretend you're driving on ice as someone mentioned earlier. I've also broken my very bad habit of stopping quite abruptly for yellow lights (which isn't very safe, actually).

 

[EVNow]

For funsies someone should record the G data for AP braking, I feel like sometimes it does some careless braking that’s way worse than any of my regen braking that causes me to have dings

Will definitely do after Saturday. I'm fairly certain that on some of the hills we have here, the only way to stop in between / at the bottom in a reasonable braking distance is to apply more than 0.3g of force.

 

[thewishmaster]

I feel like the more I drive the worse it is because it opens up more opportunities for “hard braking” or “following too closely” those are the only sections I get strikes in no matter what I do driving manually, pretty much have to drive on AP all the time.

If AP got ranked for the type of braking it does, the AP would never qualify for FSD beta

Hard braking has been dinging me a bunch on my regular short trips because the first section is always downhill and it takes empty streets and a lot of luck to keep their definition of extreme hard braking to close to 0. I’ve stopped using AP on the straight well-marked roads here because its braking is the opposite of cautious haha

 

[EVNow]

I think it's very likely that people who have higher braking forces (the force measured by the score) due to stopping on hills have higher accident rates.

The point is - this should be ascertained using data rather than just making assumptions.

 

[EVNow]

There is still another day for Elon to back it off even more.... Don't short Elon

The way these things work ... here is what I guess happened.
- Elon / Karpathy wanted to roll out 1k/day
- But the team wasn't confident (may be after the simulation tests / data coming in from testers)
- The team convinced Elon / Karpathy to go slow after the first batch

 

[_jal_]

For funsies someone should record the G data for AP braking, I feel like sometimes it does some careless braking that’s way worse than any of my regen braking that causes me to have dings

For sure! I rarely have to use the brake, but today AP in stop n go traffic dropped the brake pedal to the floor almost every third stop. Panic at 5MPH!