Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

FSD Beta 10.69

This site may earn commission on affiliate links.
correctly predict the future location of the vehicles coming from left since they will be traveling in a circle and straight line
Even not at a roundabout, I've noticed 10.69 getting confused about cross traffic that isn't directly to the left, e.g., coming around a curve or up/down a hill. Those are probably much more common than roundabouts at least in the US, so having perception correctly predict those moving vehicle behaviors will also help solve roundabouts even if roundabouts are not Tesla's current priority.
 
  • Like
Reactions: legendsk
In city-streets driving with 10.69.2, I estimate that I have to make a moderate intervention (e.g. tapping the accelerator or changing lanes) about once per minute, and a more serious intervention (disabling FSD, or FSD disables itself) about every 10 minutes. This 10 minutes will need to become 10 years to reach L4.
Or use a more standard method - miles per disengagement. Crowdsourced rate puts it at around 10 miles per disengagement. Highway AP/NOA is around 100 miles/disengagement. So, to get to "useful L2" - FSDb needs to become 10x better. To get to human level it needs to be 10,000 miles / disengagement or 1000x better.
 
Maybe you can explain why the accelerating slope of those particular miles driven isn’t ”illustrative of the accelerating pace of capability”?

Without glossing over the facts that:

1. The TYPE of driver has changed fairly dramatically (from paid internal tester with very high systems knowledge, to testers who demonstrated extreme test pilot capability in maintaining 100% on the arbitrary safety score, eventually to fairly average test-pilot capability in maintaining only 93% on the safety score).

2. The WILDLY increasing situational variety of roads and traffic patterns in all the variety of US/Can regions. (A critical challenge for generalized driving AI).

3. There has been one low-speed FSD Beta bollard-boop in all those millions of miles.

I’m sorry, but if you can’t see how that does illustrate accelerating pace of capability, you simply don’t understand what’s happening. 🤷‍♂️
When they release FSD beta later this year to all FSD purchasers the cumulative miles increase even faster! If that were actually a measure of capability why didn't they do that two years ago? Why not release FSD beta to everyone whether they purchased FSD or not? The value of robotaxis is so high that the slight loss of revenue now would be worth it.
I don't think the capability of FSD beta hasn't improved all that much in two years, it was already very capable in its first public release. It's the reliability that has improved, which is a much better metric for FSD performance.

Perhaps a metric of miles driven per user per day would be somewhat meaningful as that would give an indication of how valuable people find it as a "driver assist."
 
Even not at a roundabout, I've noticed 10.69 getting confused about cross traffic that isn't directly to the left, e.g., coming around a curve or up/down a hill. Those are probably much more common than roundabouts at least in the US, so having perception correctly predict those moving vehicle behaviors will also help solve roundabouts even if roundabouts are not Tesla's current priority.
Yes, just like CULT solved a number of problems, roundabouts will solve a number of related issues too.

My guess is currently the vehicle location prediction is probably independent of road geometry. They need to merge them - thats how people predict vehicle location. If the road is curved, the vehicle is probably going in a curved path too.

Ofcourse, need to watch out for vehicles misbehaving and going off the legal path.
 
  • Informative
Reactions: FSDtester#1
Highway AP/NOA is around 100 miles/disengagement.

I wager Highway AP is much better than 100 miles/disengagement. I suspect most of those disengagements are probably from lane changes (for normal AP people) and executing more assertive lane changes in moderate/heavy traffic (for NOA users). I suspect safety disengagement rate for AP could be in the thousands of miles/disengagement.
 
Probably the last report for a while on 69.2 from me.

So, today did the usual go-into-work-via-the-back-roads-and-no-interstates trip with the new 69.2 load. This is the leg where, typically, the "Ding!" button gets hit ten or twenty times.

Summary: About six times. Really, that's not bad.

First: Trying to get out of my local development. UPL onto a residential street. As per yesterday, crept up.. and then motored, verryy slloowwlly through. Better than yesterday, but got a Ding! as a result.

Second: Right turn off of residential onto two lane local road with stripes. It waited for traffic to clear - and then went for it. A bit slower than I do, but, still. Of course, the previous release could do this, too, but not with the same sense of purpose.

Third: NAV decided to turn left at the intersection of two local roads with stripes at a light. There's a left turn lane: Miracle, the car got into the correct lane without balking, unusual. Behind a couple of cars and, leading the parade, a dump truck. This intersection gets a left arrow then, after some traffic has cleared, a green light, where opposing traffic coming the other way has the right of way and one can go through if it's clear. I was more than mildly curious as to how 69.2 would handle it.

As expected, the dump truck was slow; the cars in front of me got through on the green and yellow, and the Tesla got stuck when the light went green for all comers. What it didn't do (and what the previous release would do) was try and drive straight into opposing traffic. It quivered a bit, but there was a lot of traffic. The last one went through on the red, leaving the M3 stuck in a red light situation in the middle of the intersection. Gassed it through and did a Ding! (That's two, now.) Had traffic been a bit less heavy, there wouldn't have been a problem. Or if the dump truck hadn't been there.

Next up: T-junction with a light, left and right turn lanes. No confusion, got into the correct lane on the right, turned and followed the parade, no jerking. Good.

Next up: Much more complex T-junction with heavy traffic in all three directions; the path is in the left lane for a left turn that goes into a dog-leg right turn lane. Dump truck got into the right lane along with a few other cars after the left; the Tesla went for the left lane, slowed down when it realized its error, at which point Ding! and took over. Without all the cars covering the lane markings, it might have made it. Three Dings!.

Right turn following ye fast-moving dump truck. No problems for three miles, kept up with traffic, handled straight-ahead stuff.

Next: Four-way intersection, two lanes in each direction on all four roads. The route has a protected left turn. Getting into the left lane hasn't been an issue; but the protected left turn has had, with the previous release, a very jerky response. Immediately after the left, there's an overhead blinking yellow for the day that the Fire Department decides to Set Forth which would drive the previous release nuts.

This time, smooth turn into the correct left lane. A bit of a slowdown under the yellow, but kept on going. No ding. (And still following that blame dump truck.)

Next up: The road that gives me the willies. Right turn off the major road (no problem, but less jerks in this release), onto a road that (a) has no markings, (b) goes up and down some steepish hills with crowns that one can't see behind, and (c) is narrow. Naturally, the dump truck went onto this road at 15 mph higher than the 25 mph limit (not unusual for an auto, but, jeez, that's a blame full-size dump truck!), thus scaring the heck out of opposing traffic.

The previous release thought that this was a Fine Road To Go Down The Middle Upon, Especially When Approaching The Crest of a Hill. 69.2.. followed the dump truck, who also thought that the middle of the road was a nice place to be. Of course, that guy is 10' up in the air up in the cab, and so has a view.

The 69.2 release followed the dump. When opposing traffic appeared, both the dump and the car edged over (safely) to the right. Until.. we hit the narrowest part of the road. The dump got way over for an opposing car and the 69.2 release seemed to recognize that it shouldn't have been born, drove towards the verge (which is branches and brush all the way), and came to a fast halt, staying there and quivering. Ding!

Weirdly enough, this is tons better than the previous release. Took over manual and got the rest of the way down off that hill and up to the next light, where stripes appeared. Much better for FSD-b, stripes.

Now, the good stuff: NAV wanted a left turn off this road onto various side streets. Handled with aplomb the left, and the right following, and smoothly, too. Previous release would have had at least two to three dings through there. And a long stretch on two laners with stripes until getting to a major, Jersey-barrier divided highway, two or three lanes in each direction, 50 mph+. Right turned onto that from a light, including getting into the right lane before the right, all good. Previous release would have been two or four dings through here, this had none, with no disengagements.

Next: Exit ramp to a sharp right turn, followed by a near-180-degree turn-and-merge onto a bridge over the divided highway. Previous release would come to a quivering halt, seeing approaching traffic on the turn-and-merge, not to mention a crosswalk. It did slow down for the crosswalk, but did keep on going and successfully merged with heavy traffic on the left. A ding! for the slowdown, but that was it.

So, this road goes over a bridge, splits into four lanes, two of which go left, two of which go straight, all of which curve sharply to the right together to the next light. Of the right two lanes, the rightmost has right-turning traffic as well as straight ahead; after the light, the two forward-going lanes merge to one. Correct lane for humans, if one is going straight: The left of the two straight lanes. The previous release always picked the right lane and would, well, get kinda stuck sometimes.

This time on 69.2 the car picked the correct straight-ahead lane, went on through, merged with all the traffic, and kept on going. No interventions and no dings. That's a first. Usually do two to four dings going through this spot.

Next up: FSD-b's nemesis. The road does twisty-turney following a stream then, at the last, goes up a steep, short hill, splitting from one lane to three. The two right lanes are straight ahead, but the rightmost is both for right turns and, eventually, an exit-only onto an interstate. The left lane is left turn only. The correct lane for straight ahead is in the middle. There's a big sign visible before the hill detailing all this and there are arrows painted on the various lanes plus stripes, but none of stripes and paints are visible until one crests the hill. Previous release would head straight for the left most lane, realize it was stuck, and come to a quivering halt. Sorry to say, but 69.2 is exactly the same. One of these days the vision will be able to see and interpret road signs, but this is not those days. Ding! At least, after manhandling the car into the correct lane, it stayed there. Previous release would try to cut left, sometimes, to go around traffic.

Finally: Up just past the interstate, the right lane is blocked off by a Jersey Barrier for construction. Why it takes 9 months to finish, I'll never know. But the previous release, until a recent map update, kept on trying to drive straight into the end of the Jersey Barrier. Well, it wouldn't - it would come to a halt, stuck, instead. Then the maps got changed and the previous release wouldn't do that. Glad to report that the 69.2 is at least as good as the previous.

So: Total of 6 dings on this trip. That's not bad: Normally, it's 15-20. Turns are much smoother, especially lefts, and it tends to move with a bit more confidence, most places. It does, sometimes, tends to move ssslloowwllyy through a left turn, just begging for some fast bugger to come up and t-bone the car; but it probably (probably!) wouldn't try the turn if there was a fast-moving bugger out there. It has eyes/cameras.

So, on a single day, there'd be, on the previous release, a total of fifteen to thirty Record! button presses. This day, that'll drop to eleven.

Definitely not passing a DMV driving test yet, but the progress is unmistakable. Still making mistakes, yes, but it's better.

Previous poster's comment about the last 10% taking 90% of the time.. Don't know. Sometimes there really are magic bullets in software. But don't knock the progress.
 
Definitely not passing a DMV driving test yet, but the progress is unmistakable. Still making mistakes, yes, but it's better.
I chuckled at this one because I imagined some 16 year-old in his dad's Tesla going for his drivers test and trying to pass off FSD Beta and hope the instructor doesn't notice.
 
I wager Highway AP is much better than 100 miles/disengagement. I suspect most of those disengagements are probably from lane changes (for normal AP people) and executing more assertive lane changes in moderate/heavy traffic (for NOA users). I suspect safety disengagement rate for AP could be in the thousands of miles/disengagement.
Possible … but just like with FSD, drivers need to feel safe with the driving.

I don’t use NOA because I don’t like it’s lane changes, for eg. I’m sure a lot of people use NOA and hardly intervene. Just like Omar hardly intervenes ;)

BTW, I’ve not looked closely to crowdsourced AP data, I’ll when I get a chance.
 
I don’t use NOA because I don’t like it’s lane changes, for eg. I’m sure a lot of people use NOA and hardly intervene.
I don't let it suggest lane changes or do auto lane change, but i do leave it on for interchanges and exits, handling off-ramps. I'm actually much more happy with off-ramp performance since it switches to FSDb, so it can handle sharper curves now and adjust speed correctly.
 
To properly handle roundabouts FSDb has to
- understand entering roundabouts is like an unprotected right, with an implicit yield sign
- understand which lane to use in multi-lane roundabouts depending on the exit
- understand we have right of the way when it comes to vehicles coming from the right
- correctly predict the future location of the vehicles coming from left since they will be traveling in a circle and straight line

Roundabouts don't have any of these logic. The stop / not stop behavior is probably a flag they set about how to handle yield sign.
I've had 10.69.2 slow and roll through a roundabout entry yield sign as well as (more commonly) stop at the yield sign even when no traffic is in the roundabout. So, I don't think there's a simple setting. More likely, it isn't making the determination that the way is clear in time.
 
When they release FSD beta later this year to all FSD purchasers the cumulative miles increase even faster! If that were actually a measure of capability why didn't they do that two years ago? Why not release FSD beta to everyone whether they purchased FSD or not? The value of robotaxis is so high that the slight loss of revenue now would be worth it.
I don't think the capability of FSD beta hasn't improved all that much in two years, it was already very capable in its first public release. It's the reliability that has improved, which is a much better metric for FSD performance.

Perhaps a metric of miles driven per user per day would be somewhat meaningful as that would give an indication of how valuable people find it as a "driver assist."
A better measure would be miles per disengagement.
 
  • Like
Reactions: Daniel in SD
A better measure would be miles per disengagement.
An even better metric is miles per necessary disengagement. That only becomes an issue once disengagements get much more infrequent than they are now though. And only Tesla will have that measurement by running counterfactual simulations. And then of course you have to weight the simulated consequences, hitting a pedestrian is much worse than hitting a curb.
 
Even not at a roundabout, I've noticed 10.69 getting confused about cross traffic that isn't directly to the left, e.g., coming around a curve or up/down a hill. Those are probably much more common than roundabouts at least in the US, so having perception correctly predict those moving vehicle behaviors will also help solve roundabouts even if roundabouts are not Tesla's current priority.
Lack of a front camera at a low height limits Tesla's ability to Look left and right without creeping ahead a lot. The location of the backward looking side camera's fixed behind of even the driver seat area is a big disadvantage.
 
Lack of a front camera at a low height limits Tesla's ability to Look left and right without creeping ahead a lot. The location of the backward looking side camera's fixed behind of even the driver seat area is a big disadvantage.
Unless there's an occlusion, the angle at which you enter a roundabout usually puts traffic in the roundabout well within the FOV of the B-pillar camera. Mount a GoPro above the B-pillar camera and you'll see.
 
  • Like
Reactions: EVNow
Has anybody else stopped to think that allowing FSD Beta to complete a drive with errors that the system ended up correcting but for any other human driver would have been an error or even “kind of” unsafe is not the right thing to do?

If you don’t disengage when FSD does something wrong but not deadly you are telling the system / program that it did it right. Meaning no changes necessary to that drive.

Those errors and “close calls” are never reported as such if you don’t disengage and send a snapshot of what happened. Just kind of odd to read here and watch videos of people saying great job / no disengagements when a lot of those drives had obvious errors no human would have made.
 
+1 for disengagement per mile, and further separated by highways (NoA) and surface streets (Autosteer on City Streets).

In Elon-speak, when he says it's, e.g., 99% there, he means 99% "of the miles you drive," as from the video. That means if you drive a mile straight down a road and then disengage FSD because it made a bad lane selection in a left turn, he's probably only counting the 40-50 foot arc of the turn as the failure of FSD. Accordingly, FSD successfully completed 99.25% of that drive.