Safety Score

[dduffey]

Seems perilous. How do you know you're pulling more than 0.2 and less than 0.4 Gs? Do you have a smartphone g-meter?

My neighborhood has a 2 mile general winding loop, so I can just circle the neighborhood a bit and take wider and tighter turns to always be turning. I am not being scientific to that level of measurement. I know the U-turn off ramp that penalized me, and that was the only one.

A single circle around the neighborhood dropped the aggressive turns dramatically and eventually back up to 100 overall score for the day.

My 7 day average is now 99, and I will need about 125 miles @ 100 to get me to 99.5 (100) overall.

 

[Knightshade]

Saw an interesting idea regarding the score that seems relevant to folks complaining about how it's unfair they get dinged if THEY are a super safe driver but they happen to live near a bunch of idiot drivers who cut them off and whatnot....(apologies if this was raised previously, thread approaches 40 pages)

Tesla isn't looking for the safest drivers.

They're looking for the lowest risk drivers. Which aren't necessarily the same thing.

It may be lower risk to select a merely pretty good driver who tends to be surrounded by largely other pretty good drivers....rather than an excellent driver surrounded by maniacs.

 

[stonelance]

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

Hills shouldn't matter? Deceleration is the same going up hill or down, assuming the same starting speed and stopping distance\time. That is assuming they are basing it off the actual velocity changes, because using an accelerometer to measure braking rate would be pretty silly.

 

[Frank99]

OK, so here's what has to happen to improve/maintain an excellent safety score.
There are 5 (Five) criteria listed on the Tesla page: Safety Score Beta

1. Forward Collision Warning. This is a simple counter - how many times would the warning have gone off if it was set on "Medium", regardless of where you actually have it set. This one is pretty simple - don't get an FCW. The only way to improve your FCW score once you get a ding or two is drive a lot, because it's incorporated as FCW / 1000 miles driven.

2. Hard Braking. This ones a bit more complex. It measures how much time you're braking by measuring the deceleration - anything over 0.1G is considered braking time. Hard braking (or "Extreme force") is anytime braking is over 0.3G. Note that full regen is around 0.2G, so anytime you press on the brake (assuming that your foot is off the accelerator so you're already at full regen) is likely to trigger a Hard Braking event. The Score is relative to (time spent braking over 0.3G) / (Time spent braking over 0.1G). The best way to recover from dings here is to spend a lot of time braking lightly - maybe get in the habit of going down the road accelerating up to speed, regen'ing down 10 mph, accelerating, regen'ing, etc. People behind you will hate your guts. The best way to keep from getting dinged is try to never touch the brake pedal - although I'm suspicious that I'm getting dinged just on regen.

3. Aggressive Turning. This is measured much like Hard Braking. Total time turning is measured by time turning at greater than 0.2G, Aggressive turning is measured as time turning at greater than 0.4G. The Score is relative to (Time spent turning over 0.4G) / (Time spent turning over 0.2G). There's a very narrow window here - turn too gently and you won't get credited with turning time, turn too aggressively and get dinged. Unfortunately, we don't have an easy crutch like we do with regen and braking to show us where the middle of the window is. The best way to recover from dings here is to find a big empty parking lot (say, at an abandoned shopping mall), load a G measurement app up on your phone, mark out a giant circle and figure out how fast you should drive around the circle to hit the middle of the window at 0.3G, then spend a half hour driving around the circle at that speed.

4. Unsafe Following. This is measured much like Hard Braking and Aggressive Turning. Total Following time is measured by the time you spend within 3 seconds of the car in front of you while going over 50 mph. Unsafe following is measured as the time you spend within 1 second of the car in front of you while going over 50 mph. Score is relative to (Time spent within 1 second) / (Time spent within 3 seconds). Once again, there's a narrow window - follow too far back and you don't get credit for following time. Too close and you get dinged. Your best bet for recovering from dings here is to get on a rural Interstate, follow a truck, set AP for a following distance of 3-4, and drive for 100 miles. Seems kinda stupid that following distance can be based on AP doing the work, but everything I've heard says that's the way it is.

5. Forced Autopilot Engagement. This is another simple one - either you pay attention to AP telling you to put your hands on the wheel, or you get dinged. Note that you only get one ding for each trip where Autopilot disengages due to lack of attention, no matter how many times it happens on a SINGLE trip. Tesla is a bit coy about what qualifies as a "trip", but you can look in the app and break down the scores by trip to get an idea. Simply putting the car in "park" then immediately shifting to drive is NOT sufficient to break a trip into two trips. The best way to recover from dings here is to make a lot of trips where you DO pay attention to the prompts, as the score is relative to (Number of trips with a ding) / (number of trips).

So, that's Frank's guide to gaming the Safety Score.

 

[AlanSubie4Life]

Deceleration is the same going up hill or down, assuming the same starting speed and stopping distance\time. That is assuming they are basing it off the actual velocity changes, because using an accelerometer to measure braking rate would be pretty silly.

Was discussed early in this thread. As far as I can tell, it uses the accelerometer (but I really can’t be sure). And it would make sense to do so, since it’s harder to stop going downhill. For a given wheel speed change, the tires have to work harder when driving downhill, so people who tend to try to stop at the same rate (which means more longitudinal g force and more work for the tires) going downhill may be less safe.
For safety, stops should be slower (wheel speed rate of change) when going downhill.

set AP for a following distance of 3-4, and drive for 100 miles. Seems kinda stupid that following distance can be based on AP doing the work, but everything I've heard says that's the way it is.

Do you have clear evidence that AP following distance time counts? The documentation clearly states it is not (neither positive results nor negative results).
I think you’d have to do what you say manually - or possibly using TACC (that’s less clear whether that counts as AP).

 

[Frank99]

>>> Do you have clear evidence that AP following distance time counts?
Excellent point. I was thinking about doing some experimenting tomorrow, maybe I'll figure that out.

 

[Sporty]

If passing on the right was a weight, then everyone in Washington state would have a score of 0...

I was thinking there are probably much better ways to measure attentiveness than what they are doing. For instance, instead of follow distance measure the time between a speed change of a car in front, and the reaction of the person in the Tesla. Most of the things they are measuring don't really relate to attentiveness at all.

We’re going to need a new thread for this. It’s because (a significant # of) drivers in wa don’t follow the rule of move to the right if you’re holding traffic up.

 

[Sporty]

I feel like my cognitive load has increased while driving with this safety score. Like I'm constantly trying to mentally project what my yaw rate is during turns and I'm fixated about what my follow distance is when I'm over 50+ mph. I'm not sure if that's good or bad for safety but it's something.

I’m also the first car at the stop lights now. First car everywhere actually - all the other drivers are behind me.

 

[Sporty]

My guess is that they will take all the 100’s and sort them by total miles driven. Then 99’s, etc.

The other question is will they be looking at 7-day rolling averages or your cumulative score over the 30 days. The first may be easier if you have one really bad day you want to drop off.

Will they weigh in and sort by location, or drive-type (highway, urban) before picking only 100s? What if the top 100s are all flat roads, highway, middle of the night. Are highway miles interesting to fsdbeta right now?

 

[jebinc]

Saw an interesting idea regarding the score that seems relevant to folks complaining about how it's unfair they get dinged if THEY are a super safe driver but they happen to live near a bunch of idiot drivers who cut them off and whatnot....(apologies if this was raised previously, thread approaches 40 pages)

Tesla isn't looking for the safest drivers.

They're looking for the lowest risk drivers. Which aren't necessarily the same thing.

It may be lower risk to select a merely pretty good driver who tends to be surrounded by largely other pretty good drivers....rather than an excellent driver surrounded by maniacs.

Elon’s recent tweet seems counter to that as he said safety scores of 100 will get the download first, 99 maybe a few days later. So, if you have a 90, you could be waiting and driving like a granny for months.

 

[dduffey]

Safety Score boost trip to Texas Gigafactory today.

 

[Padelford]

So, the safety score (beta) documentation says:
“The Safety Score formula ignores any events (not just Forward Collision Warnings) that occur while Autopilot is engaged and appropriately utilized.”
And:
“Only use the various Autopilot features as described in your Owner’s Manual.”
And the owner‘s manual says:
"Autosteer is intended for use on highways and limited-access roads with a fully attentive driver. When using Autosteer, hold the steering wheel and be mindful of road conditions and surrounding traffic. Do not use Autosteer on city streets, in construction zones, or in areas where bicyclists or pedestrians may be present."

Is it possible that @Knightshade has finally been proven right? Is Tesla penalizing people for using Autopilot on city streets? I'm seeing a lot of people getting low scores saying "I used Autopilot the whole drive!"

The other factor that people seem to be ignoring is that the hard braking, cornering, and unsafe distance metrics are all proportions. You need to increase the denominator by driving like a normal safe driver, if you drive like a grandma all the time the few times you're forced to brake hard or someone cutting in front of you is going to count a lot more.

I'm sure Tesla has the statistics to back up their formula. Obviously people trying to game the formula will reduce its predictive ability. Though maybe the ability to win the game is predictive of FSD Beta monitoring skills?

I had my worst day today score-wise, and all my dings occurred while I was using EAP on I-5. I’m going to stop using AP for the remainder of the week.

 

[jebinc]

I had my worst day today score-wise, and all my dings occurred while I was using EAP on I-5. I’m going to stop using AP for the remainder of the week.

Yeah, it blows. I gave up using it two days ago and have 100 scores ever since.

 

[FalconFour]

That's pretty nutty bordering on a parallel-universe insane, because I've pretty well demonstrated (as is documented) that AP usage masks all faults. Hell, today I had a quick-thinking moment where a stale left turn light turned yellow (no traffic, under the speed limit) and I either had to "aggressive turn" or "hard brake". "Chill" mode was smoothing my deceleration too much. I took the third option: with traffic light detection enabled, a quick double tap on the stick, AP snapped into action and laid on the brakes with a jolt! Landed perfectly at the line.

No ding. 100% for the drive.

Same with following. Any time I'm going over 50, the solution is to have AP on. Because AP is laggy as hell at acceleration, e.g. moving into an empty lane and lazily feathering the acceleration while cars approach from the rear: I give the pedal a good {GOOSE} (in other words I ****ing slam it to the floor while shouting like a Spanish football announcer without the "-allll") just to get it up to the set speed. No dings there either. 100%.

AP masks everything.

 

[FalconFour]

Hills shouldn't matter? Deceleration is the same going up hill or down, assuming the same starting speed and stopping distance\time. That is assuming they are basing it off the actual velocity changes, because using an accelerometer to measure braking rate would be pretty silly.

How do you think they measure turning G forces? I'd bet ya anything there's a simple accelerometer in there. They're so dirt cheap and in every phone, it'd cost pennies to just throw one into one of the ECUs.

 

[jebinc]

Yeah, it blows. I gave up using it two days ago and have 100 scores ever since.

Currently running a 99 average, with need of another 90 miles at 100 score to get average up to 99.5, which will be reported as 100.

 

[BitJam]

That's pretty nutty bordering on a parallel-universe insane, because I've pretty well demonstrated (as is documented) that AP usage masks all faults. Hell, today I had a quick-thinking moment where a stale left turn light turned yellow (no traffic, under the speed limit) and I either had to "aggressive turn" or "hard brake". "Chill" mode was smoothing my deceleration too much. I took the third option: with traffic light detection enabled, a quick double tap on the stick, AP snapped into action and laid on the brakes with a jolt! Landed perfectly at the line.

No ding. 100% for the drive.

Same with following. Any time I'm going over 50, the solution is to have AP on. Because AP is laggy as hell at acceleration, e.g. moving into an empty lane and lazily feathering the acceleration while cars approach from the rear: I give the pedal a good {GOOSE} (in other words I ****ing slam it to the floor while shouting like a Spanish football announcer without the "-allll") just to get it up to the set speed. No dings there either. 100%.

AP masks everything.

This is interesting. I wonder why some people get dinged on AP and others do not. Is everyone using the same 2021.32.22 firmware? Maybe there's a difference between radar and no-radar cars. Maybe some of the autopilot or driving settings affect if you get dinged on AP or not. Someone who gets dinged a lot on AP should compare settings, firmware, etc with FalconFour.

 

[jebinc]

This is interesting. I wonder why some people get dinged on AP and others do not. Is everyone using the same 2021.32.22 firmware? Maybe there's a difference between radar and no-radar cars. Maybe some of the autopilot or driving settings affect if you get dinged on AP or not. Someone who gets dinged a lot on AP should compare settings, firmware, etc with FalconFour.

Vision car vs radar car is an interesting thought…

 

[FalconFour]

This is interesting. I wonder why some people get dinged on AP and others do not. Is everyone using the same 2021.32.22 firmware? Maybe there's a difference between radar and no-radar cars. Maybe some of the autopilot or driving settings affect if you get dinged on AP or not. Someone who gets dinged a lot on AP should compare settings, firmware, etc with FalconFour.

Mine's a 2018 LR RWD, HW3 upgraded in Feb 2020. Also 2021.32.22, was batched together with the big dump that came at 3:40am Saturday morning.

My biggest, most loud complaint about Autopilot right now is its incredibly lethargic response to external stimuli. It's so laggy, I am genuinely concerned that it'll ram into stopped cars some day - and it's a fairly big mental load to balance everything. But my experience has 100% been that "Autopilot masks everything" - the more I use AP, the more a 100% score I'll get. I'm still at 99 average. Desperately hoping to finally get FSD Beta after numerous attempts over the months. It didn't take much change to driving to eliminate other factors, but "following" is ridiculous; I wonder if something is actually affecting my car that makes it so glacially laggy to respond.

Second most annoying gripe is its tendency to brake for any kind of flashing yellow light on a sign, or accidentally notice a train-traffic signal as my own signal - today I even had it brake out of nowhere while a yellow light flashed somewhere up on screen! Traffic light detection is madness, and I suspect they stopped developing it after FSD Beta started development on an all new stack. But... having light detection enabled masks any kind of hard-braking to catch a light, so I keep it enabled for peace of mind.

I never really hear much from other people about this maddeningly laggy response to anything acceleration- or braking-related on AP. It used to be, prior to the "no-radar push", that AP would be pretty well glued to the car in front. I could even tell if a car I'm following was a stick-shift, feeling the clutch lag their acceleration and thus mine as well. Now, not so much. AP accelerates like it's asleep, 2mph per second acceleration. And when I roll down the speed limit as I see stopped cars ahead, it's like "yeah that's nice" and decelerates at 2mph/sec until it finally notices them, and grinds on the brakes to stop in time. I'm legit weirded out that folks aren't yelling about this.

 

[AlanSubie4Life]

I'm legit weirded out that folks aren't yelling about this.

Fifth stage of Kübler-Ross.

Man, that seems like you're tempting fate for a FCW, which are painful.

2.5-3 seconds is fine (and normal/ok per driver’s manual). I don’t do it on a line of traffic, just solitary vehicles, to rack up the denominator. Plenty of time and sight lines to keep an eye on any reasons for sudden braking ahead. I never get FCWs on the freeway anyway, unless I’m very aggressively accelerating, and changing lanes concurrently, which is not something that is happening these days!