FSD Beta 10.69

[wknickless]

Did you report it when it went for the dirt? In the future, when you encounter very poorly marked roads, make sure you disengage before the turn and make the turn manually, then re-engage.

Put another way: tighten your grip on the wheel, overpowering FSD Beta's attempt to turn early. Only after the partial disengagement, hit the brake or stalk to fully disengage. Make the turn manually. Re-engage FSD Beta. Hit the report button.

This is a perfect example of a hard edge case that Tesla's auto labeler needs lots of video samples to learn from, so both (A) force disengagement and (B) report--early and often.

 

[thewishmaster]

A fun one from today. Going 35mph, 2 lane road, right lane is turn only, I’m in the left lane with a truck to my right. Car begins braking and I glance at the screen because the lane also splits just ahead and I’m wondering which lane it will pick - instead I see a “autopilot creeping for visibility” warning? Meanwhile we’re still traveling around 30mph at that point. No logical explanation as the next intersection was a stop light probably another 500 feet ahead

 

[Goose66]

+10 points for use of the word perspicacity.

"Well it's always in the last place you look..."

 

[AlanSubie4Life]

This is classic. A couple interventions here. First one might have been because of left turning traffic (it was just a slight recalculation of path leading to ridiculous wheel movement), but there was someone next to me turning so not sure where it thought it would go. It’s been doing very well on this turn. First failure in a while. Looks like it would have been fine to ride it out, but that would be silly since I can just drive myself.

Then: Tried to go on a red left turn arrow (visualization showed a brief flash of green, maybe? - yeah, it's true this was a tough one to see, lol), then once the light turned green it dinged at me. Lol. Which one do you believe??? Do they need to do green light detection fusion?

So useful.

 

[Goose66]

A fun one from today. Going 35mph, 2 lane road, right lane is turn only, I’m in the left lane with a truck to my right. Car begins braking and I glance at the screen because the lane also splits just ahead and I’m wondering which lane it will pick - instead I see a “autopilot creeping for visibility” warning? Meanwhile we’re still traveling around 30mph at that point. No logical explanation as the next intersection was a stop light probably another 500 feet ahead

The car was changing lanes on the Interstate today and all of a sudden a creep fence appeared on the screen with a message below, but it was gone in a 1/2 of a second and I couldn't read it. Car didn't phantom slow or anything. I did snapshot it.

 

[KLam]

I’ve commented on this before, RIGHT SIDE BIAS or FAR AWAY FROM MEDIAN when driving in urban areas at least.. does anyone else see this and or does anyone know is this an FSD interpretation issue, or camera placement or camera alignment issue? Would it be FRONT camera? I’ve tried recalibrating the cameras several times in the past 10 months since starting the beta, it’s always the same.

This is an image of the car driving, I let it drive for about 1/4 mile to settle and then came to a stop, got out to capture it’s offset from the middle yellow double lines. You can imagine what this is like for cars parked on the RIGHT or driving past cars parked on the right, it normally drives about 18 inches (or sometimes less!) from a parked car on the right side of the road. Too close IMHO and makes for a lot of ping ponging right, then left, then right to give at least SOME room to the cars on the ride side of the road.

View attachment 855721

When on NOA, do you see your car stay in the middle of the lane on the display? Do you also see your side camera views show equal distances from the lane markers?

 

[kabin]

Personal preference, guys. No way to prevent this. Not even a human can stop smoothly. We just need to get used to how FSD drives, we can't expect everyone to be happy with how it drives; these expectations are unreasonable.

No, anyone can stop smoothly. It's done all the time with and without software. And why would any EV design needlessly scrub-off energy in lieu of brake heat and dust? They don't. It's a shortcoming.

To be even remotely competitive for mainstream customers, the FSD team will eventually need to figure it out. It's all about system design, priorities, limited resources, and reasonable progress over time. Hopefully the first and last items aren't Tesla's biggest obstacles.

 

[tivoboy]

When on NOA, do you see your car stay in the middle of the lane on the display? Do you also see your side camera views show equal distances from the lane markers?

In NOA or FSD, the in car visualization SEEMS to think it’s closer to the center line than it actually is. When I’m about 12 inches away from the center line - as I tend to drive myself, the in car visualization seems to think I’ve essentially ON THE LINE (but I would caveat that by saying the tiny car and tiny lines -its hard to judge distances).

I’ve tried looking at distances in the side MIRRORS - of course that isn’t cameras. I’ll try and see what THEY are seeing - good tip.

 

[Mardak]

I see a “autopilot creeping for visibility” warning? Meanwhile we’re still traveling around 30mph at that point

I believe it's related to this new behavior:

• Enabled creeping for visibility at any intersection where objects might cross ego's path, regardless of presence of traffic controls.

I've had it incorrectly activate for through roads that intersect with "minor" highway on/off ramps that don't have traffic lights.

On the flip side, pretty sure this behavior now results in correct yielding to check for cross traffic at an intersection where the stop sign is placed very far from the stop line up a curve/hill. Previous FSD Beta would zoom right in believing it had right of way (completely ignoring the stop sign).

 

[Daniel in SD]

This is classic. A couple interventions here. First one might have been because of left turning traffic (it was just a slight recalculation of path leading to ridiculous wheel movement), but there was someone next to me turning so not sure where it thought it would go. It’s been doing very well on this corner. First failure in a while. Looks like it would have been fine to ride it out, but that would be silly since I can just drive myself.

Then: Tried to go on a red left turn arrow (visualization showed a brief flash of green, yeah, it's true this was a tough one to see, lol), then once the light turned green it dinged at me. Lol. Which one do you believe??? Do they need to do green light detection fusion?

So useful.

It clearly knew the light was about to change. It was perfectly safe to go, you just need to learn to trust FSD's superhuman perception. Sometimes the reaction time of the system is so fast it actually appears to perceive things before they happen.

 

[AlanSubie4Life]

It clearly knew the light was about to change. It was perfectly safe to go, you just need to learn to trust FSD's superhuman perception. Sometimes the reaction time of the system is so fast it actually appears to perceive things before they happen.

Oh, very interesting, I didn't know that. In future I'll just let it do its thing. The protected left arrow is meaningless here anyway, does not serve a purpose since you have to yield to oncoming traffic even after it turns green.

 

[sleepydoc]

Based on observations of most other drivers, signaling turns and lane changes must be optional, even discouraged

Unless you’re over 70 and driving a Buick or a Cadillac. Then you’re required to leave the turn signal on!

 

[sleepydoc]

That's a very good point. It can be dangerous to signal turns when they could be misinterpreted by another driver. Driving is hard!
They should only activate the rear turn signals early.

Look at the video where Omar almost got rear ended - the truck signaled properly but it was still confusing.

 

[Dewg]

Had my first big error today, on a big intersection that it's taken many many times before. Was in the left lane, needed to turn left, and instead of moving over into the two left turn lanes, it started to move, then aborted and stayed in the left lane, but then started to make the left turn from the left lane on a red arrow. Disengaged, proceeded straight through, and reported. Not sure what the failure was here. I was in front of everyone, no lead car ahead of me. The two left turn lanes were empty. I can't tell if it's a planner problem, where the planner aborted the lane change, or if it visually thought there was something in the turn lanes and didn't want to hit it. And then definitely the planner failed in the attempted left turn on red arrow from the wrong lane. Because of the final attempted left, I'm guessing the original issue was the planner as well.

Stay vigilant everyone, and be safe while testing.

 

[KArnold]

It clearly knew the light was about to change. It was perfectly safe to go, you just need to learn to trust FSD's superhuman perception. Sometimes the reaction time of the system is so fast it actually appears to perceive things before they happen.

Are you being serious or sarcastic? It's hard to tell from the video but it seems to me had it proceeded it would have turned directly into the path of that oncoming white SUV. Then of course, I don't know what his light was.

It just seems strange to me that a turn would even initiate past the stop line if the light is red - I don't care what it's about to do. Red means stop. No?

 

[AlanSubie4Life]

Here's a good stop from 10.69.2.2, on a downhill, where it typically struggles. I am not sure this has anything to do with the version. It used the brakes very slightly, but that's ok on a downhill like this (even though it's not necessary if you start slowing sooner - but that takes longer). Always good to avoid brake use, but it's more the consistency of the stopping power that is important, as discussed at great length. Watch the regen bar as usual.

Compare to the earlier example, same stop, on 10.69.2. You can clearly see in the failure case that the integrated value of the stopping power is similar; the peak power is much higher and is followed by a lot of lower stopping power.

Then of course, I don't know what his light was.

He had a yellow, turning to red. The car KNEW. Though typically people don't stop for that red light (it appears to be an optional stop, and people have to get to where they're going), so you just treat it as an unprotected left turn. One of these days I hope to stress test FSD with that scenario (that's why I engaged again here - was hoping to get lucky). Also...he was being sarcastic of course.

 

[Daniel in SD]

Are you being serious or sarcastic? It's hard to tell from the video but it seems to me had it proceeded it would have turned directly into the path of that oncoming white SUV. Then of course, I don't know what his light was.

It just seems strange to me that a turn would even initiate past the stop line if the light is red - I don't care what it's about to do. Red means stop. No?

The SUV stops!

Of course I was been sarcastic. I guess @AlanSubie4Life and I need to add a small /s to 50% of our posts.
It's fun to imagine that FSD knew the opposing light had turned red and recognized that the SUV had begun to slow.

 

[AndrewZ]

I've seen a few cases where FSD will try to proceed through red lights, the most recent example would have been from Dirty Tesla on 10.12 if I recall correctly and I believe it was another left turn lane.

This aligns with Tesla's previous comments about larger or more complex intersections being outside of the capabilities of the current OEDR. The system obviously knows to stop at red lights, but it seems to sometimes confuse lights controlling adjacent lanes for the light controlling its lane or something.


The green light chime afterwards is an amusing twist though

 

[PACEMD]

Looks like people above 80 may get 69.2.2 like Elon tweeted.

Really? We are going to have to wait until we're over 80 to get the updates? I may not even still be driving at that age, dang........

 

[AlanSubie4Life]

This aligns with Tesla's previous comments about larger or more complex intersections being outside of the capabilities of the current OEDR.

I don't see how this would be any different if there were a single lane going forward with one or two fewer traffic lights.

There's every reason to believe this will happen (sometimes! randomly! rarely?) on the simplest possible intersections with a protected left turn lane.

Looking at the video (unfortunately I didn't have the camera mounted closer and wasn't using the larger visualizations), I can't tell whether it was visualizing anything consistently in the lights. You can see green and red at times but some of the time it's ambiguous (nothing? green?). We'll never really know.