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

Software Update 2018.39.6 8377b4d

This site may earn commission on affiliate links.
And I noticed on my commute this morning that V9 has blind spots directly to the sides when at low speeds (stop & go), and I suspect that what really matters is low differential speeds. When cars are coming and going quickly I think it "remembers" the car in the next lane as it transitions from one camera to the next, even if the cameras lose track of it briefly, but when the speed difference is low the car gets stuck in the blind spot long enough to be forgotten. I saw it happen several times this morning as cars slowly passed me. This is something they need to fix before enabling uninitiated lane change.

I don't think that the car "forgot", it just does not know with great accuracy where next to you the car is when travelling at low speeds. But, it knows that there is a car somewhere next to you. Therefore I doubt that it would initiate a lane change. Its just that it can't show you where the car is at low speeds because the algorithm is not powerful enough to accurately place the car until it hits the next camera.

But since the moment my v9 installation completed, I've only seen the error message "Surround Vision limited - contact Tesla service" so I can't see anything and ALC is completely disabled. Doesn't work at all. So my autopilot capabilities actually decreased substantially since getting v9. Can't get into service for 3 weeks. Fun.
 
I do not think "learning" is the description I would use, but there does seam to be a "calibration" cycle.
I agree with you, just a speculation but don't think any active AI learning is going on but the car seems to be calibrating and adapting after calibration but not real artificial neural network learning going on. Maybe the programmers just look at data that is gathered and then readjusting their algorithm for future interactions with improved decision making tree. It would be nice if there were real AI learning but that would be much more advanced programming skills that we are still at infancy.
 
Can it/ is it actually doing this? I am not a coder so I genuinely don’t understand if it can “learn” and make improvements or is it simply following a decision tree the same time every time. Are we sure this isn’t a bit of believing what we want to believe? I promise I’m not trying to be a smart a$$.
Please educate me on if they software is truly learning over a few drive cycles.
I have been on V9 for several days now and feels the same to me.

From the Thread Below Post #272 and a Few others it seemed that the extra Cameras (v9 uses 8 vs v8.1 uses 3) are Calibrating for a few miles, so I think its more of a Calibration Period that makes the System better once it can use the Additional Cameras.

Software Update 2018.39 4a3910f (plus other v9.0 early access builds)

1b6f56c9-58cd-43f7-abff-aef917a3b0b2-jpeg.338606
 
From the Thread Below Post #272 and a Few others it seemed that the extra Cameras (v9 uses 8 vs v8.1 uses 3) are Calibrating for a few miles, so I think its more of a Calibration Period that makes the System better once it can use the Additional Cameras.
Yes but this calibration was associated with drive-on-navigation betas as the car would refuse to do the DON until they were calibrated. Now it's entirely possible that those cameras are calibrating anyway since it's the first time they've been activated on going to V9 but it doesn't enable a feature after calibration now. Perhaps this is for the best since DON will work immediately for everyone when it's finally released since it will have had weeks (months?) of camera calibration first.
 
There is one type of "challenging situation" where it's doing badly for me, though admittedly this is on roads where nobody has any business using autosteer -- on local roads where the pavement is in bad shape (potholes, patched potholes, cracks, etc) it occasionally freaks out.
An interesting thought is how much the lateral rotation of the vehicle when hitting potholes/uneven surfaces will affect the vision and confuse it too.
 
I don't think that the car "forgot", it just does not know with great accuracy where next to you the car is when travelling at low speeds. But, it knows that there is a car somewhere next to you. Therefore I doubt that it would initiate a lane change. Its just that it can't show you where the car is at low speeds because the algorithm is not powerful enough to accurately place the car until it hits the next camera.

But since the moment my v9 installation completed, I've only seen the error message "Surround Vision limited - contact Tesla service" so I can't see anything and ALC is completely disabled. Doesn't work at all. So my autopilot capabilities actually decreased substantially since getting v9. Can't get into service for 3 weeks. Fun.

Boy it sounds like one or more of your sensors are not working. I had Tesla test mine when it was in the shop and they all came back OK. Apparently it wasn't a very hard test. My issue was that the car isn't centered in the lane. I always thought it favored the left side of the road.

As noted above, I have noticed the avatars disappear when the the cars next to me are just ahead of the B pillar, but you can clearly see them with your eye.

Also the avatars are playing bumper cars all the time on the screen. They are jumpy.

I personally am not seeing any issues with 9 in the several hundred miles I've used it. I'm very happy with mine. I try to use it all the time on many roads as much as possible. I'm thinking it is learning some how some way.
 
Mine is solid. Of course, I predominately look out the windshield, side mirrors, and rearview mirror when driving. The car icons do occasionally wander but I use the realworld view of the windows and mirrors to understand my surroundings.

A trained driver should also be sensitive to what is happening in the peripheral vision. Large or rapid motions in the instrument cluster definitely catch the eye and give the driver conflicting information about the surrounding traffic. This is a safety hazard.
 
A couple things that I have noticed...

Green traffic color gone. With my map now, only yellow and red traffic is rendered. Green is gone. One can assume that no yellow or red means traffic is OK, but it is t as clear because I don’t know if traffic is clear or if there is simply no data. Anyone see that too?

The “360” view shows car to the side, but not directly behind. Is that normal?
 
  • Informative
Reactions: quickstrike12
Many of us (myself included) thought v9 was really jumpy/jittery at first, and failed to take even simple corners. The surround view display was jumpy and the cars disappeared when going by certain camera views.

For me it took around 50 miles of assorted driving but then I could take it back to the exact same failure points and it would navigate them well.

So there's definitely some sort of calibration period, whether it is making sure all the cameras' overlapping views are stitched together properly, or it's your car's throttle/brake/steering response.

It's probably not learning in terms of the neural net changing though -- there's no evidence found of that.

39.6 = Late 2016 AP2.0 Telsa - so prob no dashcam yet for me I gathered (AP2.5+ req'd?), have noticed it to be more jittery on surface streets where as < 9.0 was smoother. Maybe I need to drive it more miles to calibrate. They still haven't figured out a couple of spots on my drive where the (scarily) car panics and thrashes almost into trees along the road (even though its a perfectly straight road - here's where AP 1.0 did a better job, no sudden jerky movements. Hopefully they can iron this out (I've reported them using the voice command and some problems spots are completely gone over the past year). The roads I have problems with shows one of the lane markers jumping left and right (probably it's not sensing the edge of the road correctly though it is striped on both sides (divided single lane country road with median). Freeways it's detecting the lines no jerky lines).

Overall very happy with 9.0 to see 1) cars on both sides (that's the surround view I'm guessing, vs. backup cam 360degree which would be awesome someday?), 2) red line when lane changing and car in blind spot.

ALC on surface streets - and lane changing in general I think they finally nailed it with smooth perfection.

The new UI is more intuitive in general, but I wish they had kept the cold weather submenu rather than having to go under climate (or keep in both locations). The buttons at the bottom make more sense rather than having to tap the screen at the top.

How do I get it to suggest lane change on autopilot to take the exit - I haven't found that option, or was that feature pulled from this release?
 
Last edited:
Cars only show up once they're very close in the rear.
True. Even people show up ... at 90 kph! :mad:

When towing a 5' x 10' utility trailer it also periodically shows up as a truck or car.

Occasionally the hand crank (for raising the front of the trailer when stopped) shows up as a little person! A person, evidently running, behind the Tesla at 90 kph! :)

Low priority software code suggestion: When the car is going faster than a runner could, at least don't show any people right behind the car. :rolleyes:
 
A couple things that I have noticed...

Green traffic color gone. With my map now, only yellow and red traffic is rendered. Green is gone. One can assume that no yellow or red means traffic is OK, but it is t as clear because I don’t know if traffic is clear or if there is simply no data. Anyone see that too?

The “360” view shows car to the side, but not directly behind. Is that normal?

Yes now that you mentioned that, I don't have green either, just yellow and red. I wondered what changed I thought the background had gotten greyer. Well it probably saves data and time to omit. I do think the Nav loaded faster this morning.
 
The green was removed intentionally to not mark all roads without traffic problems green. To me it's o.k. to have only the highligts of traffic problems in redish colors. When using my car for commuting(*) I do this in the earliest possible morning, so 90% of vienna's roads were rendered green. The map is way more clear now with only a few red marks.

(*) Usually I commute by bike as commuting in vienna by car really sucks.
 
I don't think that the car "forgot", it just does not know with great accuracy where next to you the car is when travelling at low speeds. But, it knows that there is a car somewhere next to you. Therefore I doubt that it would initiate a lane change. Its just that it can't show you where the car is at low speeds because the algorithm is not powerful enough to accurately place the car until it hits the next camera.

On what evidence do you base this remarkably trusting conclusion? The engineers developing this code certainly need to account for false positives, since clearly the vision system has false positives -- we've all seen them. A detection of a vehicle that then disappears for an extended period of time is highly likely to have been a false positive and should be discarded. Remembering every little glitch coming out of detection for all time would be a disaster. I am certain that they have some logic in their tracking system to forget tracked vehicles after some period of time if there are no further observations of that vehicle.

I am also nearly certain that the UI is willing to display vehicles when there is uncertainty about their position. I say this because I see vehicles dancing on my display constantly, including sometimes overlapping my own car's avatar, indicating significant uncertainty in their position -- and, I might add, significant lack of sophistication in the tracking and UI code!