Software Update 2018.10.4

[TaoJones]

Hi Tao, apologies if I am hashing something you already know. Yea, the Ultrasonic sensor is about as good as it is going to get. I think I might have said this somewhere else. Many people are assuming that if they see a flash on the display from the sensor, that it is the sensor sees that car on the side of the road 100 feet or more ahead. Not the case. It is not a direct one to one system. The forward radar and cameras are seeing the car and ALSO feeding that flash on the display

Thanks for this @FlyF4. Great to know the AP2 balloons are fed from multiple sources.

The new sensitivity is definitely notable, given the number of "Take the wheel" alerts now on a daily basis. With future updates hopefully we get the necessary tuning.

Good stuff.

 

[iandkirby]

Crazy Sensitive Lane Recognition - Wet Car Tracks Only

A few times today I was able to activate autopilot (2018.10.4) on a road with no lanes. Inner/Outer - nothing....but by the wet tracks created from the car in front. This lead to some interesting Russian roulette - but it did surprisingly well.

The problem more so lies when you live in a place like Vancouver - where it rains 99% of the time - and the road markings are total garbage in the rain, even in the day. Later, driving up a road with a centre lane, I was finding autopilot using the tracks of the car in front of me to override the faint yellow centre line. Causing me to drift into places I shouldn't have been and eventually off to the side of the road with parked cars.

I know you're not supposed to activate on side roads etc. etc....

Anywho - good times - getting there.

Ian.

 

[Scott-P]

In getting sick and tired of mine doing the same. I feel like I bought a Windows 95.

More like Windows ME... ;-)

 

[don123]

Buttery smooth (2018.10.4) is so exciting. Thanks team and Elon.

Now dreaming about:
1) Speed limit sign recognition
2) Buttery smooth, traffic aware, AP lane merge handling
3) Indicator triggered AP lane change on local roads
4) Everyone being on 10.4 or higher.

Dare we think about AP stop sign and traffic light handling?

So glad Tess now recognises distant vehicles stopped at a traffic light ... mostly. Dear wife's nervous twitch all gone.

 

[don123]

I'm pretty certain my loaner AP1 MS in September 2017 changed lanes on local roads if I merely indicated. Does that still work?

It was so nice not having to disengage AP, control the speed, change lanes and re-engage AP. Am I getting soft?

 

[Robwoodruff]

I'm pretty certain my loaner AP1 MS in September 2017 changed lanes on local roads if I merely indicated.

AP1 changes lanes on local roads, AP2 still does not.

 

[barjohn]

Observed two things today. The first, I was stopped at a stop sign and a bicyclist crossed in front of me riding his bike. Nothing displayed on the IC, it was as if he was invisible. Then at another stop sign pulling out to 4 lanes of traffic I noticed that cars crossing in front of me only occasionally displayed (as if they were in my lane) and then quickly vanished and often nothing displayed. The second thing was I noticed that when a car moved into my lane moving in the same direction, there was about a 500 msec delay between when the event happened and when it displayed. My car: 12/16 S 75D AP2 with 2018.10.4 Anyone else seeing similar behavior? Do AP2.5 cars have the same delay or does the faster computer eliminate the delay?

 

[MattMatt]

Observed two things today. The first, I was stopped at a stop sign and a bicyclist crossed in front of me riding his bike. Nothing displayed on the IC, it was as if he was invisible. Then at another stop sign pulling out to 4 lanes of traffic I noticed that cars crossing in front of me only occasionally displayed (as if they were in my lane) and then quickly vanished and often nothing displayed. The second thing was I noticed that when a car moved into my lane moving in the same direction, there was about a 500 msec delay between when the event happened and when it displayed. My car: 12/16 S 75D AP2 with 2018.10.4 Anyone else seeing similar behavior? Do AP2.5 cars have the same delay or does the faster computer eliminate the delay?

Normal stuff. Currently autopilot does not know how to recognize/classify/display the things your talking about (cars/bicycles going perpendicular to you) so seeing irregularities or funny car images is normal. Also the delay between AP determining if the car is in your lane is normal, it will also be slow to detect when they leave your lane.

 

[barjohn]

Do you know whether this is true for AP2 and AP2.5?

 

[ChrML]

This release's NN must use more than pure lines to navigate. Perhaps it can see other objects such as rails too and fins a path between them.

Yesterday I drove this stretch. With fairly long parts of barely visible lines on just one side and some turns with no visible lines at all. Zero disengagement and it drove fairly well.

 

[chispas]

The second thing was I noticed that when a car moved into my lane moving in the same direction, there was about a 500 msec delay between when the event happened and when it displayed. My car: 12/16 S 75D AP2 with 2018.10.4 Anyone else seeing similar behavior? Do AP2.5 cars have the same delay or does the faster computer eliminate the delay?

This has been happening to me since I first got my Tesla in Dec 2017 (AP 2.5) if so now it works better because it happens for a couple of cars but not the ones following them

 

[dknisely]

The only complaint that I've had with 10.4 is that it seems to be very slow to acknowledge and respond to cars that cut me off (on both highway and local). I've had several scary takeovers required when I was trying very hard to not intervene until it became clear that a collision was imminent. No collision warning... No response whatsoever. It seems to take 5 seconds or more to recognize, and even when the car appears in the display, it doesn't respond. This problem is particularly bad in stop-and-go when a car wedges in front with everybody going very slowly (but moving). Cars merging from the right are a big problem at all speeds.

These are moving cars, so even radar should be registering them.

As far as picking up previously untracked stopped cars when I come upon them (e.g., stopped at a light), I think 10.4 is about the same reliability as previous releases or perhaps just slightly worse. I would say 90-95% are picked up, but perhaps 5% of the time, it would happily plow into the stopped car or truck.

 

[⚡️ELECTROMAN⚡️]

My car drove through Joshua Tree National Park very nicely today. I'm not experiencing any issues.

 

[J Roc]

This release's NN must use more than pure lines to navigate. Perhaps it can see other objects such as rails too and fins a path between them.

Yesterday I drove this stretch. With fairly long parts of barely visible lines on just one side and some turns with no visible lines at all. Zero disengagement and it drove fairly well.

View attachment 289832 View attachment 289833 View attachment 289834 View attachment 289835 View attachment 289836

I experienced something similar. I have a road that goes from well marked lines to no lines. It went 3 miles on the no line road on AP 2.5 with no issue. The road had curves, hills and dips. It seemed as if it tracked the right edge of the road and used a fixed with for the left.

 

[rnortman]

So I finally got 2018.10.4 and tested it out. I agree with others that this is a real improvement. But this is also the first really noticeable improvement in almost 6 months. What I find really shocking is how low our expectations have gotten for Autopilot -- the main improvements in this release are that it does not attempt to murder me when cresting hills, it's able to change lanes and stop when approaching stopped cars without giving me whiplash. (Stopping for stopped cars is still uncomfortably abrupt and a little scary, but no longer completely terrifying.) And everybody is just tickled pink about this!

We were supposed be here (AP1 parity or slightly better) a year ago. In 2018 we were supposed to have "full self driving", and the Tesla Network should be right around the corner -- where our cars were going to drop us off at work or the mall and then go around earning money for us in an autonomous ride sharing network. Anybody still think we're going to get even close to that in 2018? Or 2019?

Sorry to be a party pooper, I do still love my car, even when I hate it. I'm just trying to bring some perspective. This update is a welcome improvement, but it's still far from what Elon has promised, and honestly I don't think we're much closer to FSD on AP2/2.5 than we were a year ago. I'll take the less-murderous EAP, but let's not let Elon off the hook so easily.

BTW, this handles highway curves better but it's still a far cry from "Autosteer+". It still can't handle some of the curves on my daily commute, at least not at any speed that my fellow commuters would tolerate. And it's still basically worthless/dangerous on local roads.

 

[ChrML]

Yes, no doubt Tesla is delayed. It's fairly reasonable to think that Mr. Karpathy started pretty much at scratch when he started mid-2017 designing the architecture like it should've been designed from the beginning in winter 2016.

The previous framework before 10.4 was most likely just some light modification of a GoogleNet sample application to get started and get SOMETHING out there.

This means we should begin seeing incremental improvements from pretty much now and on I hope, now that their architecture is right (or so they say).

 

[⚡️ELECTROMAN⚡️]

So I finally got 2018.10.4 and tested it out. I agree with others that this is a real improvement. But this is also the first really noticeable improvement in almost 6 months. What I find really shocking is how low our expectations have gotten for Autopilot -- the main improvements in this release are that it does not attempt to murder me when cresting hills, it's able to change lanes and stop when approaching stopped cars without giving me whiplash. (Stopping for stopped cars is still uncomfortably abrupt and a little scary, but no longer completely terrifying.) And everybody is just tickled pink about this!

We were supposed be here (AP1 parity or slightly better) a year ago. In 2018 we were supposed to have "full self driving", and the Tesla Network should be right around the corner -- where our cars were going to drop us off at work or the mall and then go around earning money for us in an autonomous ride sharing network. Anybody still think we're going to get even close to that in 2018? Or 2019?

Sorry to be a party pooper, I do still love my car, even when I hate it. I'm just trying to bring some perspective. This update is a welcome improvement, but it's still far from what Elon has promised, and honestly I don't think we're much closer to FSD on AP2/2.5 than we were a year ago. I'll take the less-murderous EAP, but let's not let Elon off the hook so easily.

BTW, this handles highway curves better but it's still a far cry from "Autosteer+". It still can't handle some of the curves on my daily commute, at least not at any speed that my fellow commuters would tolerate. And it's still basically worthless/dangerous on local roads.

I agree with what you are saying, but I don’t share your pessimism. I think Tesla is working very hard and taking great risks to bring us what they’ve delivered to us so far. It’s not about letting anyone off the hook.

 

[Bebop]

My car drove through Joshua Tree National Park very nicely today. I'm not experiencing any issues.


View attachment 289851

Car looks great. I have the same one minus the grey wheels.

 

[barjohn]

My biggest gripe with the EAP software is the failure of Tesla to fully inform us with each release what new capabilities it is supposed to have and what know issues it has resolved. Instead I see a steady stream of videos where drivers are conducting what could be dangerous tests to see what has changed or to find the limitations. If we knew what it was supposed to be able to do, we could test for that capability and the feedback to Tesla would be more meaningful. Either the capability works as designed or it fails under the (specified by tester) conditions. Another improvement would be to display on the IC what the camera is seeing. This could be a rendering but wouldn't you like to know if the camera system is seeing the car coming up on you in the lane adjacent? Wouldn't you like to know it sees the motorcycle rider straddling the lane and coming up behind you? How about the pedestrian crossing the street? If we were cognizant that the camera was not seeing something we are seeing we would know not to trust its action. Does it see the stopped car in front of us? Is it going to apply the brakes before we hit it? For now we guess and hope that the answer is yes but we have no visual indication to confirm our hope.

 

[mongo]

My biggest gripe with the EAP software is the failure of Tesla to fully inform us with each release what new capabilities it is supposed to have and what know issues it has resolved. Instead I see a steady stream of videos where drivers are conducting what could be dangerous tests to see what has changed or to find the limitations. If we knew what it was supposed to be able to do, we could test for that capability and the feedback to Tesla would be more meaningful. Either the capability works as designed or it fails under the (specified by tester) conditions. Another improvement would be to display on the IC what the camera is seeing. This could be a rendering but wouldn't you like to know if the camera system is seeing the car coming up on you in the lane adjacent? Wouldn't you like to know it sees the motorcycle rider straddling the lane and coming up behind you? How about the pedestrian crossing the street? If we were cognizant that the camera was not seeing something we are seeing we would know not to trust its action. Does it see the stopped car in front of us? Is it going to apply the brakes before we hit it? For now we guess and hope that the answer is yes but we have no visual indication to confirm our hope.

I understand the desire, your reasoning is why I'm very rarely a passenger (or is it that people refuse to drive with me in the car? )
However, having such a distraction would decrease from overall safety. Even if it sees the object, it may not do what you predict it should.

If the car is not doing what you want, intervene, don't test its limits. If you think it should slow for a pedestrian, slow.