Software Update 2018.10.4

[boaterva]

Maybe in this case we need a video of hill crest not working... because I know we’ve dozens of reports and dozens (literally) of videos of it working flawlessly over test cases where it failed before. It sure worked for me where it didn’t before. Erik posted one such today on their constant test route. So....

 

[ahkahn]

yeah, you can see the gun in one of her hands but not the fireball whisky in the other hand. It's supposed to look like this, though I don't know why.

Even she is celebrating how well Autopilot is performing!

 

[luckyj]

2. The major issue I’ve found with 10.4 is the tendency to center on wide lanes. When a lane drops off temporarily (ramp approach, etc) the car immediately centers in the double-wide space and then has to recenter on the single lane when the right side stripe re-appears. Those with the dotted line on the right over the gap work perfectly, of course.

Yes - noticed this a few times. I think it's only when there isn't a dashed line marking for the actual lane you are in. Seems that it doesn't adjust lane following at those merge areas based on knowing if you are entering vs were already on the road. It just tries to stay in the middle of observed lines, which in those cases is not really what you want to do.

 

[croman]

Does this update to anything to improve AP1? Don't bother replying with speculation.

No. This is new neural networks for AP2 and 2.5.

 

[Electroman]

Here is another video of 2018.10.4 driving much better than previous versions.

This is one of the best AP2 10-4 video out there. Most of the drive is done 5miles above the speed limit on two lane roads that do not have right shoulder or line marker on the right, with moderate to sharp turns and hills. It handled every one of them as if a human would do.

very impressive

 

[luckyj]

Maybe in this case we need a video of hill crest not working... because I know we’ve dozens of reports and dozens (literally) of videos of it working flawlessly over test cases where it failed before. It sure worked for me where it didn’t before. Erik posted one such today on their constant test route. So....

Yeah, I'm not saying the current release doesn't handle most crests quite well that it couldn't before. I've had the same experience. I'm just saying that doesn't mean it's "fixed", because fixed implies there are no or near zero failure cases. To believe that was accomplished in one release that's been out for three days and a good portion of the fleet doesn't have seems like a stretch. But heck, maybe they did. Tomorrow I should be able to drive through the same spot I had to take over today, I'll try to get some video.

 

[SlicedBr3ad]

Yeah, I'm not saying the current release doesn't handle most crests quite well that it couldn't before. I've had the same experience. I'm just saying that doesn't mean it's "fixed", because fixed implies there are no or near zero failure cases. To believe that was accomplished in one release that's been out for three days and a good portion of the fleet doesn't have seems like a stretch. But heck, maybe they did. Tomorrow I should be able to drive through the same spot I had to take over today, I'll try to get some video.

every video i've seen has passed crests without issues. really looking for evidence of single failure at least.

 

[Snowstorm]

I can add to the list on how the car (ap2.0) can recognize the travel lane even when it is not defined by lane markings. Today, I took the model s through a road I take everyday, one where the road is being widened so all traffic is routed to one side while they dig/pave the other, then reverse. It is a twisty path defined by cones, partial lane markings. I was watching it like hawk but ap handled the entire stretch without drama. It continued to recognize the drivable surface even if it was just a series of cones, and even handled the lane swap. AP would absolutely fail a few days ago.

 

[⚡️ELECTROMAN⚡️]

It failed, yet huge improvement overall. I don’t feel like I’m going cause someone to rear end me by stopping too abruptly anymore. Exited the highway two days in a row where it didn’t use to. Phantom breaking didn’t occur in same locations. Only strange decelerations occurred during lane change. Wouldn’t change lanes in traffic at slow speeds around 35 when there was an opening in traffic. Is this normal? Not a normal scenario for me. On the highway I attempted to change lanes when a car was coming up on me. It didn’t change lanes. It slowed down to around 40 from a setting of 64. When it finally changed lanes it stayed at around 40 with no one in front of me. I had to hit the accelerator after a few moments because there was a car coming up from behind at highway speeds. Very stable overall. On Wednesday I head out on a couple thousand mile road trip. Factory tour this Friday morning \0/. Good timing.

 

[boaterva]

It failed, yet huge improvement overall. I don’t feel like I’m going cause someone to rear end me by stopping too abruptly anymore. Exited the highway two days in a row where it didn’t use to. Phantom breaking didn’t occur in same locations. Only strange decelerations occurred during lane change. Wouldn’t change lanes in traffic at slow speeds around 35 when there was an opening in traffic. Is this normal? Not a normal scenario for me. On the highway I attempted to change lanes when a car was coming up on me. It didn’t change lanes. It slowed down to around 40 from a setting of 64. When it finally changed lanes it stayed at around 40 with no one in front of me. I had to hit the accelerator after a few moments because there was a car coming up from behind at highway speeds. Very stable overall. On Wednesday I head out on a couple thousand mile road trip. Factory tour this Friday morning \0/. Good timing.

Remember to change lanes automatically it has to show the neighboring lane and also broken lines meaning that it thinks an auto lane change is possible. Usually this is from being on a divided highway and no passing traffic recently.

 

[Snowstorm]

Any evidence it is using more cameras?

 

[conman]

Any evidence it is using more cameras?

Not according to the neural networks thread.

 

[croman]

Any evidence it is using more cameras?

Definitely not right now but soon. Blindspot monitoring is reportedly in the code for the repeater NNs.

 

[⚡️ELECTROMAN⚡️]

Remember to change lanes automatically it has to show the neighboring lane and also broken lines meaning that it thinks an auto lane change is possible. Usually this is from being on a divided highway and no passing traffic recently.

It’s a highway that it can normally change lanes on.

 

[MarcusMaximus]

I tried it today on the 17 freeway to/from Santa Cruz(for those unfamiliar with the Bay Area, it’s a windy mountain road going up the Santa Cruz mountains and then back down). Overall astoundingly good job. On the tighter turns it got a little too close to the inside of the turn for my liking(particularly when near the median wall), so I took over a couple times, but otherwise Autosteer took me the whole way. I’m pretty confident I could have just let it go the whole way without incident.

 

[nnick01]

I tried it today on the 17 freeway to/from Santa Cruz(for those unfamiliar with the Bay Area, it’s a windy mountain road going up the Santa Cruz mountains and then back down). Overall astoundingly good job. On the tighter turns it got a little too close to the inside of the turn for my liking(particularly when near the median wall), so I took over a couple times, but otherwise Autosteer took me the whole way. I’m pretty confident I could have just let it go the whole way without incident.

I did hw17 both ways today at 60mph. It did great. I didn't have to disengage except for changing lanes a couple of times (it won't change lanes on this hw) .

 

[buttershrimp]

Some other new observations from 10.4... active safety improvements.

 

[buttershrimp]

Definitely not right now but soon. Blindspot monitoring is reportedly in the code for the repeater NNs.

Something in the rear camera picked up the speeding car in my video... same as the motorcycle that buzzed me yesterday.

 

[PhaseWhite]

Some other new observations from 10.4... active safety improvements.

It's hard to tell from the video, did the car actively steer you to the left or did it just beep?

This is interesting, would ultrasonic sensors be able to do this or do we think the rear camera (or repeater) has a NN active?

 

[Snowstorm]

Something in the rear camera picked up the speeding car in my video... same as the motorcycle that buzzed me yesterday.

This would be a major finding if it works well. I find the rear camera too fished eyed to have any good distance viewing capability, and would be surprised if it can pick it up from that.

I asked my wife to buzz by with the minivan tomorrow to test this out but got a flat out rejection.