FSD Beta 10.69

[generalmilk]

It’s not an apples-to-apples comparison. (Some of) the other numbers are not “usable” numbers. For example for the Model 3 RWD. That is a full capacity number.

Model Y 2022 latest results are closer to 81kWh I think (was for the Model 3 for the same pack). Can look it up from the test documents, it is public knowledge and well documented. Could be wrong but should be a nominal “82.1kWh” pack with more like 79-81 kWh initial full capacity. So usable around 77kWh max.

All I was saying. Anyway off-topic for FSD efficiency. We don’t care about absolute efficiency at all. Don’t even care about relative efficiency between different cars. Just relative - exact same vehicle, drive, etc. only one variable as much as possible.

Model Y LR has 79 useable, YP has 81. 77 is without bottom reserve. EPA test uses the bottom reserve (run to halt). Anyway, this is off topic. My point is 2022 Y should have same on the paper efficiency with 2020Y

 

[AlanSubie4Life]

I’m starting my experiment for my commute using the easiest route with most surface streets. There is a way with all surface streets but can try that later.

Another flawless drive home manually. Literally never moved my foot to the brake. Wasn’t even trying. 10Wh/mi better than expected; likely the standard slight onshore flow (will be similar tomorrow with FSD).

If FSD beats this tomorrow I will be very impressed! That would mean it has tricks up its sleeve!

Annoying the newish UI does not display fractional miles! Can’t calculate energy use accurately for such a short trip. Efficiency should be ok though. Distances basically identical; average of this morning and this evening is 234Wh/mi.

 

[FSDtester#1]

Another flawless drive home manually. Literally never moved my foot to the brake. Wasn’t even trying. 10Wh/mi better than expected; likely the standard slight onshore flow (will be similar tomorrow with FSD).

If FSD beats this tomorrow I will be very impressed! That would mean it has tricks up its sleeve!

Annoying the newish UI does not display fractional miles! Can’t calculate energy use accurately for such a short trip. Efficiency should be ok though. Distances basically identical; average of this morning and this evening is 234Wh/mi.

View attachment 862870

I have been wondering about this for a while... I think someone mentioned something about it before... does Fsd have any ability to increase regen??
That frigging regen bar can really go green under fast deceleration, or is that braking including regen and we are getting faked out?

 

[generalmilk]

I have been wondering about this for a while... I think someone mentioned something about it before... does Fsd have any ability to increase regen??
That frigging regen bar can really go green under fast deceleration, or is that braking including regen and we are getting faked out?

I don’t see FSD uses more regen. More often than not it uses more brakes than me and that can be seen clearly with latest energy bar update.

I did a similar test as @AlanSubie4Life before with AP and concluded that if I drive very carefully, I can be a tidbit more efficient than AP (310wh/mi vs 320wh/mi). 48 miles drive mainly on Highway doing 80 with some elevation increase.

AP is better at maintaining speed. I’m better at braking.

When I first got FSD beta on 10.2/10.3.1, the efficiency was bad due to all the harsh brakes. It was significantly improved since then.

 

[ironwill96]

I don’t see FSD uses more regen. More often than not it uses more brakes than me and that can be seen clearly with latest energy bar update.

I did a similar test as @AlanSubie4Life before with AP and concluded that if I drive very carefully, I can be a tidbit more efficient than AP (310wh/mi vs 320wh/mi). 48 miles drive mainly on Highway doing 80 with some elevation increase.

AP is better at maintaining speed. I’m better at braking.

When I first got FSD beta on 10.2/10.3.1, the efficiency was bad due to all the harsh brakes. It was significantly improved since then.

Fsd is currently terrible at predicting a smooth usage of regen braking to slow to a stop at precisely the right spot using only regen. All of us are capable of doing that after a month with our cars.

Hoping that’s just one of those minor things they’re not worried about yet but will fix later and not because they struggle with real depth perception.

 

[generalmilk]

Fsd is currently terrible at predicting a smooth usage of regen braking to slow to a stop at precisely the right spot using only regen. All of us are capable of doing that after a month with our cars.

Hoping that’s just one of those minor things they’re not worried about yet but will fix later and not because they struggle with real depth perception.

Ah, I see you are spoiled by the newer version of FSD beta. Coming from 10.2, I’m super grateful on how well it uses regen now comparing to before. There is one update, can’t remember version number, with the better use of regen written into the release notes, that version made biggest difference, but it continuously improved in later updates.

It used to do harsh stop for last 10 to 15mph at stop sign or stop light, efficiency was the least you would worry about…

Now if on a flat road with no leading car or a smooth one, it can actually stop with just regen! Only when downhill, unclear stop line or with a rough driver ahead it uses brake to compensate. Looks to me it sticks with the deceleration curve too strictly, while us human not only have a better perception on environment but also a larger tolerance on following distance. Sometimes we can let the car move closer to the car in front in case of insufficient regen than expected.

 

[ironwill96]

Ah, I see you are spoiled by the newer version of FSD beta. Coming from 10.2, I’m super grateful on how well it uses regen now comparing to before. There is one update, can’t remember version number, with the better use of regen written into the release notes, that version made biggest difference, but it continuously improved in later updates.

It used to do harsh stop for last 10 to 15mph at stop sign or stop light, efficiency was the least you would worry about…

Now if on a flat road with no leading car or a smooth one, it can actually stop with just regen! Only when downhill, unclear stop line or with a rough driver ahead it uses brake to compensate. Looks to me it sticks with the deceleration curve too strictly, while us human not only have a better perception on environment but also a larger tolerance on following distance. Sometimes we can let the car move closer to the car in front in case of insufficient regen than expected.

Current stop sign behavior isn’t consistent - it almost seems like they’re testing different variables on stop signs with different people.

In my model X refresh it slams on the brakes 300 feet back from the stop sign and then approaches from there at 15-20mph before slowing to a full stop. I have to disengage every time I have anyone behind me as it’s bizarre behavior.

 

[aronth5]

Current stop sign behavior isn’t consistent - it almost seems like they’re testing different variables on stop signs with different people.

In my model X refresh it slams on the brakes 300 feet back from the stop sign and then approaches from there at 15-20mph before slowing to a full stop. I have to disengage every time I have anyone behind me as it’s bizarre behavior.

Just curious if you have any T intersections without a Stop sign on the street you're approaching the intersection? I suspect if yes braking behaves very differently (better) than roads where you have a Stop sign. At least that is the case for me.

 

[sandman1330]

A friend of mine disputes whether FSDb is this smart, but I’m kind of curious…

Like most, I’ve had it handle the same situations and intersections very differently at different times. For the AI experts here, is it possible it’s using a “trial and error” technique? Like, it’ll try one way, and if driver disengages, it records that as a fail and will do it different next time. If driver does not disengage, it considers what it just did as proper behaviour and uploads this for training.

Is this possible or am I giving it too much credit?

 

[clydeiii]

A friend of mine disputes whether FSDb is this smart, but I’m kind of curious…

Like most, I’ve had it handle the same situations and intersections very differently at different times. For the AI experts here, is it possible it’s using a “trial and error” technique? Like, it’ll try one way, and if driver disengages, it records that as a fail and will do it different next time. If driver does not disengage, it considers what it just did as proper behaviour and uploads this for training.

Is this possible or am I giving it too much credit?

Def not happening

 

[willow_hiller]

A friend of mine disputes whether FSDb is this smart, but I’m kind of curious…

Like most, I’ve had it handle the same situations and intersections very differently at different times. For the AI experts here, is it possible it’s using a “trial and error” technique? Like, it’ll try one way, and if driver disengages, it records that as a fail and will do it different next time. If driver does not disengage, it considers what it just did as proper behaviour and uploads this for training.

Is this possible or am I giving it too much credit?

Sort of. It's not doing the learning in the car, but Tesla is collecting disengagement data, and they will likely feed that into training the neural network. It's called "supervised learning" and you can read more about it in general here: Supervised learning - Wikipedia

 

[sleepydoc]

I have been wondering about this for a while... I think someone mentioned something about it before... does Fsd have any ability to increase regen??
That frigging regen bar can really go green under fast deceleration, or is that braking including regen and we are getting faked out?

I was going to say the same thing. When you drive manually it takes a second for full regen to kick in but with FSD it seems like the bar can go full green instantly. It does seem to use the friction brakes more but it’s usually for just a split second, so the actual effect may not be that great.

The inconsistent level of braking also probably doesn’t play too much of a roll in efficiency. Nominally slower deceleration is better because of increased power losses with higher currents but the actual difference may not be that great.

 

[Oil Freedom]

Model S and Model Y have about the same efficiency. On highway speed, S might be a bit more efficient due to lower height.

That being said, the efficiency of my Y is also a bit mystical with FSD beta. Before I enrolled last October, I averaged 293wh/mi. Since having beta, I used it whenever I could. Now my lifetime efficiency is 281wh/mi.

When I’m using AP/FSD, I let it drive at 85mph whenever it is possible. This used to be 80mph before the speed limit change, but I see no impact on efficiency. The last 2k miles including a road trip even averaged 261wh/mi…. It seems steadily decreasing.

I got my car on June 2021, so maybe it’s normal that the efficiency was worse for the first few thousands miles.

I wonder if using FSDb requires enough additional compute to slightly lower the wh/mi.

 

[Oil Freedom]

Hoping that’s just one of those minor things they’re not worried about yet but will fix later and not because they struggle with real depth perception.

Although FSDb's current occupancy network is supposed to provide accurate depth perception (and moving toward sufficient enough accuracy to even replace USS):

Maybe the non-smooth (but improving) regen use on approach to a stop could be due to too much variability/error in the range calculations during the approach. If the calculation becomes more accurate (or there's enough slop) as the stop draws nearer, this could result in the regen (and brake use) ping-ponging between too high/low as the rate of close vs. distance remaining jumps around due to the error factor.

I expect this to further improve, but it's probably not a high priority yet for Tesla.

 

[DanCar]

After 69.3, it's gonna be 3.1, 3.2, 3.3

And then 10.69.420 (eta Christmas)

I'll vote we just jump to 69.420, skip the 10, its cleaner.

 

[mjwojtukiewicz]

I got 10.69.2.3 the other day. It's worse then before. A simple Auto Navigate home yesterday was awful . The streets around here have problem if cars are on both sides of the street. The car did awful waiting on other cars to pass by. It almost hit a car on the right side of the street.
BTW I priced my car. 2019 M3SRP with FSD 30K miles. Caravana 28900, Carmax 32000 Tesla 30-34K. So FSD does NOT increase the value of the car.

 

[JulienW]

I'll vote we just jump to 69.420, skip the 10, its cleaner.

Kinda pointless since that is about the 25th of December SO in less than a week (Dec 31st) we will be getting the Full Self Driving L5 complete. Now that is some 420 for you.

 

[Dewg]

I got 10.69.2.3 the other day. It's worse then before. A simple Auto Navigate home yesterday was awful . The streets around here have problem if cars are on both sides of the street. The car did awful waiting on other cars to pass by. It almost hit a car on the right side of the street.
BTW I priced my car. 2019 M3SRP with FSD 30K miles. Caravana 28900, Carmax 32000 Tesla 30-34K. So FSD does NOT increase the value of the car.

You're looking at dealers who don't care about FSD, and Tesla will just remove it if you sell to them. You need to look at a private party sale if you want to get value from FSD, there are many private parties who would pay for FSD, especially if the car is in Beta.

 

[Mardak]

FSDb's current occupancy network is supposed to provide accurate depth perception

Seems to be pretty good already. The other night, FSD Beta 10.69.2.3 went around this garbage bag just fine, so of course I looped around to try it again but starting from right at the curb in front of the truck instead of approaching normally from more towards the center of the road:

When activating FSD Beta, it turned sharply to get past the bag no problem. It did seem to forget about the bag when I tried again but pulling up directly in front of it with only a little top bit visible from the main camera, so if there's object permanence, it's only for a few seconds before it forgets. Maybe 10.69.3 will have longer memory now that they know it's working fine for normal driving in preparation for replacing ultrasonic sensors?

 

[PACEMD]

Sort of. It's not doing the learning in the car, but Tesla is collecting disengagement data, and they will likely feed that into training the neural network. It's called "supervised learning" and you can read more about it in general here: Supervised learning - Wikipedia

I don't know, maybe it does. Mike keeps driving like it's drunk and I never even give it a breathalyzer........it just keeps right on and on........drunk driving. I'm gonna have to have a serious talk now........