FSD Beta 10.69

[willow_hiller]

Bad creep behavior

But I don't know why Chuck keeps calling the shared left turn lane a median. That lane is only for making left turns off the highway, and it's illegal to use it as a merging lane like he's expecting FSD to do. Yes, people use it for that purpose, but that doesn't mean it's a legal maneuver.

Here's an example where someone asked a police officer about the legality of using a shared left turn lane to merge into traffic: Answer Man: Can I drive in the center turn lane?

 

[nvx1977]

This early version of 10.69 is being extra cautious, so waits for a moderately big gap in traffic to cross. Upcoming releases will do better in heavy traffic.

This seems to imply that car not going for big gaps is not a perception problem but rather a hardcoded reduction in assertion until safety is vetted.

 

[diplomat33]

WOW. This was super dangerous.

 

[diplomat33]

But I don't know why Chuck keeps calling the shared left turn lane a median. That lane is only for making left turns off the highway, and it's illegal to use it as a merging lane like he's expecting FSD to do. Yes, people use it for that purpose, but that doesn't mean it's a legal maneuver.

Here's an example where someone asked a police officer about the legality of using a shared left turn lane to merge into traffic: Answer Man: Can I drive in the center turn lane?

Yeah, I agree. I have no idea why Chuck would think that a shared left turn lane should be used as a median.

 

[Mardak]

That lane is only for making left turns off the highway, and it's illegal to use it as a merging lane like he's expecting FSD to do

It is illegal to use center turn lanes for merging… in some states. The article you found I believe is from North Carolina, and here's one from Florida also saying it's illegal, so indeed Chuck is asking for what seems to be an illegal FSD feature for his state.

Some states like California allow some amount of distance to be traveled for entering and exiting the center turn lane: shall not be driven in that lane for more than 200 feet while preparing for and making the turn or while preparing to merge into the adjacent lanes of travel.

Others like Nevada specify different amounts: (b) A vehicle must not travel more than 200 feet in a center turn lane before making a left-hand turn from the highway. (c) A vehicle must not travel more than 50 feet in a center turn lane after making a left-hand turn onto the highway before merging with traffic.

 

[AlanSubie4Life]

WOW. This was super dangerous.

Behind the times, man.

 

[AlanSubie4Life]

This seems to imply that car not going for big gaps is not a perception problem but rather a hardcoded reduction in assertion until safety is vetted.

Which is counter to what we see, so I assume it means that they can dial up the crossing speed. It looks like it is mismatched to the go/no-go decision right now. Or something else is horribly wrong; hard to say.

In any case, looking forward to this after what we saw today. Maybe if it is 100% successful I’ll give the beer back (since that would likely bring it back above 90%)! Seems unlikely to occur but you never know, it’s just a sample, could get lucky/unlucky:

https://twitter.com/x/status/1561494630562242562

 

[JHCCAZ]

Which is counter to what we see, so I assume it means that they can dial up the crossing speed. It looks like it is mismatched to the go/no-go decision right now. Or something else is horribly wrong; hard to say.

This could be what's going on, and it would be a problem.

As a general proposition in development, if you optimize a system or process over time and tests, it's a huge risk to adjust things like input sensitivity, output gain, delay or bandwidth parameters without a nearly complete suite of re-qualification tests before release. And that's true even for relatively simpler and more linear systems.

It's dangerous to assume that dialing something up or down, though it may seem on the surface to be a more "conservative" setting, will actually have the desired result. The intent may be for a simply more "docile" system, trading away a little performance for more stability, but the reality can be unpleasantly surprising. I'm sure such risks are even greater with this highly nonlinear, somewhat opaque NN-based system.

We don't know, and I hate to second-guess these guys, but sometimes last-minute tweaks are slipped in under pressure of the schedule. I hope they're maintaining a strong culture of complete regression testing even for seemingly safe changes. The slow roll-out plan can help save things, but it's not to be used instead of a complete change-verification cycle.

 

[Dan D.]

This could be what's going on, and it would be a problem.

As a general proposition in development, if you optimize a system or process over time and tests, it's a huge risk to adjust things like input sensitivity, output gain, delay or bandwidth parameters without a nearly complete suite of re-qualification tests before release. And that's true even for relatively simpler and more linear systems.

It's dangerous to assume that dialing something up or down, though it may seem on the surface to be a more "conservative" setting, will actually have the desired result. The intent may be for a simply more "docile" system, trading away a little performance for more stability, but the reality can be unpleasantly surprising. I'm sure such risks are even greater with this highly nonlinear, somewhat opaque NN-based system.

We don't know, and I hate to second-guess these guys, but sometimes last-minute tweaks are slipped in under pressure of the schedule. I hope they're maintaining a strong culture of complete regression testing even for seemingly safe changes. The slow roll-out plan can help save things, but it's not to be used instead of a complete change-verification cycle.

Well, all we heard from Elon/Tesla was they were aiming to "solve" Chuck's turn before releasing 10.69, also words like getting to 100%, now it "waits for a moderately big gap in traffic to cross", things like that. So we are left somewhat assuming that they did what they said and have made it work, which is why they released it.

Did they say upon release that it did not achieve those goals? No, not really. So once again we are left guessing as to whether it's going to work, and being somewhat surprised at discovering that it's still launching itself directly into traffic sometimes. Just now it's doing it faster.

Bottom line as we all know is that you're still the one in control. But we are continually lead to believe it's now working, when it's demonstrably not working (100%). Just a little more clear pessimism from Elon and his no-disengagement cheering squad would make us less likely to have to look for the errors. They should admit the errors before we have to find them. If they don't know about the errors then they are irresponsible for releasing the software before fully testing it - the errors are not exactly difficult to find.

 

[jabloomf1230]

For those of you trying to jibe Elon's tweet about 1,000 people getting 10.69 with the tiny number of people getting it on TeslaFi, here's my answer. 10.69 has only been released to a few dozen "influencers". The rest must be Tesla employees.

 

[Ramphex]

TLDR, is this a shill only release or did anyone here get it too?

 

[jebinc]

Yep. I’ve had it with this. Way to treat your customers. Trade a car in that had it, pay double for “FSD” again, only to have to deal with the games for 7,000+miles of a 100 safety score and no beta. I actually found value in the beta, and used it all the time. Stock AP/NoA sucks.

Hear, hear.

 

[jebinc]

But 90% succes rate is terrible when driving a car. We would all be dead if we had that.

But the price is going to $15k!

 

[jebinc]

I had the same reaction. But we didn't see the before state. so while objectively FSD still fails, it's probably an improvement.

Shill report - it’s to be expected with Mars.

 

[ChiefToaster]

The influencers are only influencers because they get it first and we are all waiting in suspense for any news on updates. They are FSD early adopters and got in before safety scores existed. The rest are probably employees, but who knows. Like who has the one car in Pennsylvania that got 10.13? They’re not an influencer or we would’ve seen something online. I highly doubt employees are allowed to use TeslaFi, but that idk.

 

[Mardak]

Any idea on the diff between this and 10.13?

A few entries from 10.13 were deleted / didn't make the cut for the already long 10.69 release notes:

• Improved the “is parked” attribute for vehicles by 5% by adding 20% more examples to the training set.
• Reduced false lane changes for cones or blockages by preferring gentle offsetting in-lane where appropriate.
• Improved in-lane positioning on wide residential roads.

A couple made it unchanged:

• Made speed profile more comfortable when creeping for visibility, to allow for smoother stops when protecting for potentially occluded objects.
• Enabled creeping for visibility at any intersection where objects might cross ego’s path, regardless of presence of traffic controls.

A bunch had a lot of edits (and reordering within the list) and extra improvement for left turns and occupancy network:

• Improvedstopping pose while yielding for crossing objects at “Chuck Cook style” unprotected left turns with more appropriate speed profile when approaching and exiting median crossover regions, in the presence of high speed cross traffic ("Chuck Cook style" unprotected left turns). This was done by utilizingallowing optimizable initial jerk, to mimic the medianharsh pedal press by a human, when required to go in front of high speed objects. Also improved lateral profile approaching such safety regions. Improved decision makingregions to allow forunprotected left turns using better estimation of ego’spose that aligns well for exiting the region. Finally, improved interaction withother objects throughthat are entering or waiting inside the maneuver.median crossover region with better modeling of their future intent.
• Improved lane positiongeometry error of crossing and mergingego-relevant lanes by 22% by adding long-range skip connections and a more powerful trunk to the network architecture. Improved lane position error by 5%34% and lane recallcrossing lanes by 12%21% with a full vector lanesVector Lanes neural network update. Information bottlenecks in the network architecture were abatedeliminated by increasing the size of the per-camera feature extractors, video modules,and internals of the GPTautoregressive decoder, and by adding a hard attention modules.mechanism which greatly improved the fine position of lanes.
• Improvedpedestrian and bicyclist velocity error for pedestrians and bicyclists by 17%, especially when ego is making a turn, by improving the onboard trajectory estimation used as input to the neural network.
• Improvedanimal detection recall of animals by 34%and decreased false positives by 8% by doubling the size of the auto-labeled training set.
• Improveddetection recall of object detection, eliminating 26% of missing detections for far away crossing vehicles by4% by tuning the loss function used during training and improving label quality.
• Upgraded Occupancy Network to use video instead of images from single time step. This temporal context allows theoccupancy network to detect dynamic objectsbe robust to temporary occlusions and enables prediction of occupancy flow. Also, improved performance by adding a video module, tuning the loss function,ground truth with semantics-driven outlier rejection, hard example mining, and adding 37k new clips toincreasing the training set.dataset size by 2.4x.
• Reduced false slowdowns around crosswalks by better classificationnear crosswalks. This was done with improved understanding of pedestrianspedestrian and bicyclists as not intending to interact with ego.bicyclist intent based on their motion.

Most are new with 10.69:

• Added a new "deep lane guidance" module to the Vector Lanes neural network which fuses features extracted from the video streams with coarse map data, i.e. lane counts and lane connectivites. This architecture achieves a 44% lower error rate on lane topology compared to the previous model, enabling smoother control before lanes and their connectivities becomes visually apparent. This provides a way to make every Autopilot drive as good as someone driving their own commute, yet in a sufficiently general way that adapts for road changes.
• Improved overall driving smoothness, without sacrificing latency, through better modeling of system and actuation latency in trajectory planning. Trajectory planner now independently accounts for latency from steering commands to actual steering actuation, as well as acceleration and brake commands to actuation. This results in a trajectory that is a more accurate model of how the vehicle would drive. This allows better downstream controller tracking and smoothness while also allowing a more accurate response during harsh manevuers.
• Added control for arbitrary low-speed moving volumes from Occupancy Network. This also enables finer control for more precise object shapes that cannot be easily represented by a cuboid primitive. This required predicting velocity at every 3D voxel. We may now control for slow-moving UFOs.
• Upgraded to a new two-stage architecture to produce object kinematics (e.g. velocity, acceleration, yaw rate) where network compute is allocated O(objects) instead of O(space). This improved velocity estimates for far away crossing vehicles by 20%, while using one tenth of the compute.
• Increased smoothness for protected right turns by improving the association of traffic lights with slip lanes vs yield signs with slip lanes. This reduces false slowdowns when there are no relevant objects present and also improves yielding position when they are present.
• Improved accuracy of stopping position in critical scenarios with crossing objects, by allowing dynamic resolution in trajectory optimization to focus more on areas where finer control is essential.
• Increased recall of forking lanes by 36% by having topological tokens participate in the attention operations of the autoregressive decoder and by increasing the loss applied to fork tokens during training.
• Improved object future path prediction in scenarios with high yaw rate by incorporating yaw rate and lateral motion into the likelihood estimation. This helps with objects turning into or away from ego’s lane, especially in intersections or cut-in scenarios.
• Improved speed when entering highway by better handling of upcoming map speed changes, which increases the confidence of merging onto the highway.
• Reduced latency when starting from a stop by accounting for lead vehicle jerk.
• Enabled faster identification of red light runners by evaluating their current kinematic state against their expected braking profile.

 

[AlanSubie4Life]

10.69 has only been released to a few dozen "influencers". The rest must be Tesla employees.

and being somewhat surprised at discovering that it's still launching itself directly into traffic sometimes.

The lesson here, of course, since we can’t monetize the sweet early release YouTube clicks, is to convert future releases into free craft beers by making predictable bets based on your insider knowledge of current FSD performance.

 

[micvite]

But I don't know why Chuck keeps calling the shared left turn lane a median. That lane is only for making left turns off the highway, and it's illegal to use it as a merging lane like he's expecting FSD to do. Yes, people use it for that purpose, but that doesn't mean it's a legal maneuver.

Here's an example where someone asked a police officer about the legality of using a shared left turn lane to merge into traffic: Answer Man: Can I drive in the center turn lane?

To all the geniuses and lawyers and experts on believing what the media has to say; here is the official DOT document from FLORIDA explaining what MEDIANS are and what they do. In the image specifically it says TO PROTECT PEOPLE MAKING LEFT TURNS aka that means you're expected to stop there but hey, yall know better than the DOT right? If you're gonna Google random bs at least up your googling skills or you won't be able to face the tesla keyboard warriors.

 

[Dan D.]

To all the geniuses and lawyers and experts on believing what the media has to say; here is the official DOT document from FLORIDA explaining what MEDIANS are and what they do. In the image specifically it says TO PROTECT PEOPLE MAKING LEFT TURNS aka that means you're expected to stop there but hey, yall know better than the DOT right? If you're gonna Google random bs at least up your googling skills or you won't be able to face the tesla keyboard warriors.

There are two different kinds of median being discussed here, as far as I can figure it. Where a physical raised median exists (aka Chuck's Turn) cars tend to want to wait there to merge if they only got halfway across the road. In your DOT image Exhibit 77 it's saying good design should make the physical median wide enough to protect the car, or alternately they should create medians that prohibit turning by design. If the median already exists then it has to be used as is. Chuck's Turn is an example of this.

The other kind of median is the shared painted center turn lane that Chuck is showing in his later video. This median has no raised physical barrier and is the one that people are saying Tesla (correctly) seemed not to be attempting to use as a merge lane, and Chuck is mistakenly saying Tesla should use it to merge, and maybe also to stop in. This lane isn't legally supposed to be used as a merge lane, despite that people sometimes do so. Maybe some States allow it to be used for 50ft to 200ft or so but Tesla has to drive according to local laws everywhere. They have to make a generalized solution, or a State-specific solution (which would be much more difficult to keep track of). Thus in the generalized solution Tesla shouldn't be using it to go halfway and sit there waiting to merge. I haven't watched enough videos to see if Tesla got it right on purpose by not using the center turn lane, or if it just happened to do it this time and sometimes uses the center turn lane to merge from. Anyway, they are not supposed to use it.

Florida DOT is also saying that "All 7 lane (6-lane roadways with a two-way center turn lane) roadway sections should be given the highest priority for retrofit." presumably because of their high risk (1.3.7 Retrofit Multi-lane Multilane Roadways with Center Turn Lanes)

I hope my interpretation is correct. That's how I see it.

 

[mhan00]

Thats so interesting. In California you can use center turn lanes to start and to complete left turns. I had no idea that other states don’t allow you to use it to complete left turns.