Safety Score

[tivoboy]

No. They have not specified anything about the camera - so they can't turn it on and spy you.

The only other considerations could be geography and actual model - but that would be for rollout and filtering.

The whole point of making a complicated safety score is that it captures all the things you care for and you can use that one score for decision making. Absolutely no point in adding hidden filters later resulting in bad press.

I read somewhere that for the FSD beta they were indeed using the internal camera to monitor driver attention and either ding them or disqualify them. It wasn’t speculation. I’ll see if I can dig that up.

 

[azred]

No, we don't. We need a score of 100 to get the FSD beta first (excluding those who already have it as part of the Early Access program). Although the timing is unclear, Tesla/Musk has said that the beta program (emphasizing those two words; see below) will be expanded starting with the highest-scoring drivers and moving down from there. Presumably there'll be some cutoff point, below which you'll have to wait for the general availability release, but that cutoff point is unlikely to be 100.

Remember, this scoring is to get into a beta-test program. The software is not complete; it's not what we've paid $3k to $10k (or $199/month) to get. It's beta-quality software that can kill people! Yes, it's cool new technology, but PLEASE keep in mind that it's also DANGEROUS new technology. Don't lose sight of that fact while competing to be the first in the door. I, too, signed up and have tried to keep a good driving score; but I'm re-thinking that as time has gone on. Some strategies for gaming the score are downright dangerous, and for what? Getting the earliest wide beta version a week before I otherwise might? It's just not worth it, IMHO.

Your warnings are not wrong but keep in mind most of us who want the Beta are routinely using the wide release on city streets and it is also far from perfect. For example, there is one spot in my commute where the car will try to move into the adjoining lane without warning. Another spot, the car suddenly slows and reports a speed limit that was removed a year ago when a construction project was completed. So while the beta may well be downright dangerous, so is the current wide release if not monitored.

Freeway usage is probably much safer than city streets but even on freeways the wide release software has significant challenges. I am amazed Tesla still hasn’t solved merge lanes for entering and exiting freeways. The car’s efforts to center itself in merge lanes is nuts and unsafe.

If I had an alternative to the gamed safety scores, I would probably use experience. For example anyone who has had autopilot/FSD more than a year has experienced a lot of bizarre events and needs no cautionary advice from you or Tesla. Giving the Beta to someone who has owned the car for a month but has gamed a 100 score makes no sense to me.

 

[JosiahLuscher]

I guess I have to decide what to do now…I have a score of "100" again (not really 100, but whatever...)

This score is an example of how the main score is not using the values on the same main score page; this set of values results in a score around 99.2. Instead, the mileage-weighted average score is used (as documented by Tesla). Furthermore, it looks like they are rounding the scores to the nearest integer before doing the mileage weighting (a big deal!):
My daily scores:
105 miles, 99 (98.51)
10 miles, 99 (99.05)
32 miles, 100 (100)
51 miles, 100 (99.63)
32 miles, 100 (100)

So, the formula is: ROUND( SUMOVERDAYS(DailyMileage*ROUND(DailyScore))/ TotalMileage )

There's no other way to get the result. The proof is: here are the various ways to do it, with my data:

View attachment 716037View attachment 716039

So basically I'm at 99.2 in reality, and the app is calculating 99.5 for the weighted average (exactly), and is rounding the display to 100. Got really lucky on that first day that I ended up with 99 rather than 98 - that would have really hurt (and in fact it does hurt my actual score).

If anyone sees an error here or some inconsistency, do let me know. I think the calculations are right.

So now we have to have a debate when they open it up to people by score, whether they are going to use the actual score, or the one with all this rounding error. I suppose I should attempt to get my actual score back up to 99.5 or higher.


Yes you can, but not too steep (below 10% grade would be best), and not too fast. Have to be very careful. but I

Please, how do you calculate a safety score? For example my first day I had all perfect except for braking which the app says 1.9. I travelled 104 miles, but if I understand correctly that only matters for overall score. The app calculated my score as 96. How is that calculation done? I swear I can do math, but appaently there is some context that Im missing. Please help, I'm so frustrated by my ignorance.

 

[BitJam]

I definitely do not get dinged while using AP. Also, can anyone confirm they dropped the radar on radar cars yet? Is there a way to tell, isn’t the top speed on AP limited to 90mph on vision cars?

The last I heard (August 9th) the top AP speed for vision-only cars is 80 mph. I believe the top AP speed for cars with radar is 90 mph.

I wonder why speeding wouldn't be one of the metrics for the safety score?

My guess is that speeding did not correlate well with collision probability. One problem with using speeding as a metric is that the car sometimes does not know the speed limit. Another problem is that in some places everyone routinely speeds and driving at the speed limit could be less safe. I'm confident that if speeding had a high correlation then it would have been used.

No. They have not specified anything about the camera - so they can't turn it on and spy you.

IIUC the internal camera is already used to detect inattentiveness. If so then it is already being used on the score indirectly because it could trigger a forced disengagement.

IF Tesla, after you consent, could have just pulled your car logs and derived the safety score, why subject us all to a week or two of gaming the score, when unbiased data is readily available? It kinda lends credence to the idea that they are looking for the best gamers.... the people going way out of their way to get FSD might be the ones most desperate not to lose the privilege, and therefore will be safe.

One reason Tesla would not use past telemetry is many people would think it unfair and would want a chance to have Tesla ignore the past and only monitor them when they are on good behavior. Ignoring the past is the path of least complaint.

 

[tivoboy]

Not part of safety rating and can't leave the car. So, obviously can't be used for filtering.

More importantly, there is absolutely no reason why Tesla would hide this if they were going to use it.

Just because the images don’t leave the car, doesn’t mean that the in car ANALYSIS of the images doesn’t leave the car. If the car is interpreting the images as “driver not attentive”. Or “driver using both hands on cell phone” and giving in car alerts for such, the are almost certainly going to be sending those alerts data up to the cloud for safety score analysIs. The in car alerts will be just another example/need to reboot Hal quickly so it doesn’t get sent up to the mother ship.

 

[EVNow]

If so then it is already being used on the score indirectly because it could trigger a forced disengagement.

Thats true. But - definitely not a separate filter Tesla would be using "secretly".

Just because the images don’t leave the car, doesn’t mean that the in car ANALYSIS of the images doesn’t leave the car. If the car is interpreting the images as “driver not attentive”. Or “driver using both hands on cell phone” and giving in car alerts for such, the are almost certainly going to be sending those alerts data up to the cloud for safety score analysIs. The in car alerts will be just another example/need to reboot Hal quickly so it doesn’t get sent up to the mother ship.

See above. Either its already a part of safety score (like above) or its not. If its not part of safety score already - it won't be used secretly by Tesla to filter.

BTW, once someone gets Beta, I expect part of that will be camera monitoring like it is now with current testers. That will come with the agreement when we download beta and start using it.

 

[Darmie]

I guess I should clarify. If you need to come to rest while descending that slope, you will need to experience more than 0.3 g's of deceleration. And if the slope is greater than 30%, then even a constant (smooth) braking force to maintain a constant speed would result in more than 0.3g. This is an example of where breaking is "smooth" ie. constant, yet would be interpreted as hard braking. My issue was that it is not possible in all cases, not just a specifically a 30% grade. Having lived in Pittsburgh, and having lived on a 27% grade, there are plenty of places where driving legally, that is, not impeding traffic by creeping at 1mph, there is no way to have no point of the deceleration be less than 0.3g.

Right, and all this for FDS after paying $10,000, This safety score is becoming a joke. I'm getting dinged for aggressive turning too but yet, never spill coffee during any of my commutes.

 

[Daniel in SD]

I wonder why speeding wouldn't be one of the metrics for the safety score?

Try speeding without triggering hard braking or aggressive turning or FCWs. It’s probably correlated enough with those metrics that it doesn‘t provide much predictive value.

Note that unsafe following distance has almost zero effect on the Safety Score, probably because it is so correlated with hard braking.

 

[BitJam]

That's true. But - [internal camera footage] definitely not a separate filter Tesla would be using "secretly".

Agreed. But there is no need for it to be included again since people will already get dinged heavily if AP is disengaged due to inattentiveness. Keep your eyes on the road, your hands upon the wheel ...

 

[SalisburySam]

Right, and all this for FDS after paying $10,000, This safety score is becoming a joke. I'm getting dinged for aggressive turning too but yet, never spill coffee during any of my commutes.

Agree. From a driver or passenger perspective, not spilling coffee is indeed a far better metric than those being scored.

 

[insaneoctane]

I wonder why speeding wouldn't be one of the metrics for the safety score?

It should be for insurance, but not babysitting FSD beta....

 

[JackSparrowV]

SumProduct (Weight Average) in Excel or Google Sheets would seem to work.
Here is an example:

Just do the math.... Do a weighted daily average (each day's miles times score for that day divided by 100). Add all of the days weighted scores up. Call that "X". Now just add up all the daily miles (unweighted), call that "Y". When X/Y = .995 or higher, you score will report 100. If you are under 100 and want to know how many 100 miles you need to get your average up to .995, or 99.5, just use goal seek in excel to figure that for you. Likewise, you can just do the brute force method and add the same number of miles to each "X" and "Y" until you get a .99.5 result. Enjoy, Homie's!

Gotcha! Thanks!

 

[Watts_Up]

I wonder if the Scoring is activated as soon as you click on the 'Button' or only when FSD is on?

 

[insaneoctane]

I'm admittedly wavering. I told my wife she can only be a passenger in the Tesla for 1 week (she'll DESTROY the hard braking score). Now it's 2 weeks(?) and this game is already old. Haven't told her it's another week. You guys might have less competition soon...

 

[AlanSubie4Life]

Please, how do you calculate a safety score? For example my first day I had all perfect except for braking which the app says 1.9

115.38-22.53*0.683*1.127^1.9 = 96

 

[dduffey]

So, I had something strange happen. Yesterday I was an overall 13.8% unsafe following score (for all days, not just yesterday).

Today I did a single trip, 76 (mostly autopilot) miles @ 100 score (all individual measures are 0 for the trip and for the day), but my overall unsafe following across all days INCREASED to 14.6%

I don't think it impacted my overall score, but it seems very strange that it does impact the combined "unsafe following" score as if that reflects AP unsafe following, but in the day/trip details it does not show it there.

Hmm.. Anyway, shooting for 100!

 

[Knightshade]

Not part of safety rating and can't leave the car. So, obviously can't be used for filtering.

That's not correct though.

The first consent box you have to check to sign up for the beta does not restrict the data collection to ONLY safety-score data.

And to be clear in the internal monitoring-
Camera IMAGES don't leave the vehicle. They could certainly report # of "inattentiveness found" results if either you enabled data sharing in the first place- or if you press the button and consent to data collection.


That doesn't mean they're doing it- but they have certainly been given permission to.

And we know for sure they're doing it, and have kicked people out, among the folks actually on the beta.

I definitely do not get dinged while using AP. Also, can anyone confirm they dropped the radar on radar cars yet? Is there a way to tell, isn’t the top speed on AP limited to 90mph on vision cars?

For wide-release SW, cars with radar still use it.

The fact they get options for higher max speed and shorter follow distance vs vision only cars further confirms that.


I'm not sure if radar cars on the beta still use radar or not-- I would SUSPECT they do on highways at least since that's still the legacy code, dunno about in the city.

 

[Darmie]

115.38-22.53*0.683*1.127^1.9 = 96

Yep verified by Google.

 

[BitJam]

I wonder if the Scoring is activated as soon as you click on the 'Button' or only when FSD is on?

Most of the scoring happens when FSD is not enabled. When FSD is enabled then you should only get dinged if you get forced out of AP/FSD due to inattentiveness. Some people are getting dinged when AP violates the safety score rules but other people aren't. We don't know why.

 

[EVNow]

The first consent box you have to check to sign up for the beta does not restrict the data collection to ONLY safety-score data.

Sure.

The way this whole subject came up was - OP speculated that they will use some hidden filters apart from the safety score. I'm just contending that, that is not likely (esp. with camera data). Using camera data in a way not explicitly stated will always get bad publicity, besides there is zero reason why Tesla wouldn't disclose it if they were planning to use the camera data in any other way.