How much will original Autopilot improve

[Olle]

As an owner of a HW 1 car, I'm of course curious about how much it will improve with future software.
If it is self learning at no cost to Tesla, one would assume that it would improve. Has anyone seen evidence of self learning though? I have not. For example there is a road in my neighborhood where the car always selects the turn lane while driving in the straight ahead lane, blinkers off. When the car initiates the turn, I usually take over the wheel and steer the car back to the straight ahead lane, then turn on AP again. 8 months of doing this taught the car(s?) nothing.
So I wonder if, A; self learning refers to only non location based driving algoritms or B: self learning refers only to the AP2 neural network? Is there any learning and knowledge carry over from AP2 learning to AP1 perhaps?

If AP1 is then built on manual programming, the reasons for diverting coding resources from current and future products would be customer satisfaction and resale value. So is Tesla doing it? Or is it in fact self learning? Let the speculation begin...

 

[Derek Kessler]

Autopilot, in either implementation, is not self-learning. There's fleet learning, where the data is sent to Tesla's servers, processed, and distributed to all cars through updates and refined high-resolution maps, but no individual vehicle learning.

As for improvements to HW1 Autopilot? There are potentially a few new tricks it could get, as well as refined algorithms and maps to better handle the scenario you've described, but I expect that AP1 is at 90-95% of its full development. There might be more capability to unlock, but Tesla's going to be focused on AP2 cars going forward — it's a growing market, after all.

 

[TexLaw]

Based solely upon what we've heard from Tesla and Elon's tweets, I also get the impression that HW1 is pretty much fully developed. I would expect further refinement and, perhaps, some expansion of the features that already exist, but I would be surprised to see new features.

 

[doubleohwhat]

The only feature that I would really like to see added to AP1 that I think the HW would be capable of is traffic light and stop sign recognition. The mobileye chip that is in the AP1 cars is supposed to be capable of this feature.

 

[David29]

My guess is that there will be little if any improvement in AP1 autopilot function or performance. Tesla has too many software tasks and does not seem to put many resources on older (i.e., not current) cars. I wish it were otherwise -- AP1 leaves a lot to be desired in smoothness, accuracy, consistency. Still tries to take exits now and then, e.g. But they are under lots of pressure to deliver on AP2 from customers who paid lots of money for functions that do not yet work. So AP1 will take a literal backseat.

 

[Olle]

The only feature that I would really like to see added to AP1 that I think the HW would be capable of is traffic light and stop sign recognition. The mobileye chip that is in the AP1 cars is supposed to be capable of this feature.

That would be an awesome improvement.
How about summon the car to you on private property, not just forward or backward.
How about in-lane-adjustment so that we can move a couple of feet away from swaying 18 wheelers if the opposite side lane is empty? And all those landscapers' trailers with protruding wheels, rakes and weed eaters wobbling in and out of your lane as you pass them. Should be possible to first recognize truck in camera and then calculate relative speed + feel with sensors when it is adjacent.

 

[Olle]

My guess is that there will be little if any improvement in AP1 autopilot function or performance. Tesla has too many software tasks and does not seem to put many resources on older (i.e., not current) cars. I wish it were otherwise -- AP1 leaves a lot to be desired in smoothness, accuracy, consistency. Still tries to take exits now and then, e.g. But they are under lots of pressure to deliver on AP2 from customers who paid lots of money for functions that do not yet work. So AP1 will take a literal backseat.

Agreed that the priorities are elsewhere, especially AP2. But the question was partly if fleet learning (self earning?) was automatic so that the improvement wouldn't consume resources.

 

[David29]

Agreed that the priorities are elsewhere, especially AP2. But the question was partly if fleet learning (self earning?) was automatic so that the improvement wouldn't consume resources.

My understanding -- and it could be incomplete or incorrect -- was that the fleet learning was more about the roadways than about the car. The cars picked up data on details of exit ramp geometry, lane locations, stuff like that. So it was more about creating a super-precise map of the roadways than about the car itself.
That was what I understood.

 

[lunitiks]

Yes @David29 I think you're correct, and I expect to see that this "HD-mapping" data will help improve the accuracy of our AP1-vehicles.

Remember, AP1 is not solely vision. We've all experienced Autosteer on completely unmarked roads. This, I think, is one area where the fleet mapping comes in handy. I.e. more AP confidence and thereby smoothness

 

[Carl]

New features I would not expect, but HW1 did keep learning over the last months (even if slowly). Very noticeable in the speed it takes on ramps between highways for example (used to be the maximum allowed speed, and now it is generally something a lot more prudent and comfortable). It's not a fast learner obviously (frustrating on the last lap of my daily trip to home where I have to correct it every single time, since over a year) but if you compare May 2017 HW1 to May 2016 HW1 it is obviously better. I did Brussels-Rotterdam and back yesterday and literally never had to intervene on highways. That was not the case even four months ago.

 

[dhcp]

I think it's like any software or computer hardware - after a while, the software development stops beyond safety patching for something egregious. There wouldn't be much value in keeping an AP1 development team focused there just to help the existing or CPO market. There isn't much ROI on that. Remember this is a business, not a club or charity.

 

[commasign]

If Tesla could just improve the highway interchange handling I'll be a happy camper. Currently you have to time the time the turn signal perfectly to convince the car to take the exit ramp (or time it perfectly to convince the car to NOT take the exit ramp in some situations).

 

[David29]

I don't take many long rides, but this past weekend I drove from the Boston area to Quechee, VT and back, a round trip of 320 miles or so, the vast majority on Interstate highways. I used my AP1 Autopilot feature as much as I could, which means for virtually all of the Interstate travel. On the one hand, I was able to use it and it was mostly satisfactory, and that is helpful. On the other hand, it still wanders on the slightest crests, and tends to over-correct. Passengers notice it sometimes, because it is not the way a person would drive. What continues to puzzle me is how the instrument cluster shows the lines and that the car is moving closer to one than the other, or even over the line, and yet the system can't keep the car in the middle even though it "knows" the car is off-center. (Plus I have a memory that it was better in the past, but not certain my recollection is correct.)
I also don't like the way every lane change has a moment when the car hesitates, just as it gets into the new lane, almost as if it were down-shifting or peeking around a corner before accelerating again. Again, disconcerting and not like a human. It happens just about every time I change lanes. (And I do not recall seeing mention of this by other people.)
The only other problem I recall was one attempt to take an exit rather than to stay in a lane, rather disconcerting. (This might be in that category of self-learning discussed above.)
Otherwise I was glad to have it. Five+ hours of driving is a lot for me, and I feel safer having the AP keeping an eye on things for me.

 

[sorka]

New features I would not expect, but HW1 did keep learning over the last months (even if slowly). Very noticeable in the speed it takes on ramps between highways for example (used to be the maximum allowed speed, and now it is generally something a lot more prudent and comfortable). It's not a fast learner obviously (frustrating on the last lap of my daily trip to home where I have to correct it every single time, since over a year) but if you compare May 2017 HW1 to May 2016 HW1 it is obviously better. I did Brussels-Rotterdam and back yesterday and literally never had to intervene on highways. That was not the case even four months ago.

Confused. Does AP1 navigate to freeway interchange ramps automatically based on the nav route????

 

[eschummer]

@David29 my observations match yours. It used to be much better at holding the center of the lane, and it was also better at smooth lane changes without any noticeable changes in speed. Something broke in the last iteration. I'm almost to the point of just not using it at all on my commute. I hope they have the hart to fix this for us HW1 owners. After all, the car is just about a year old...

 

[GasKilla]

Confused. Does AP1 navigate to freeway interchange ramps automatically based on the nav route????

That was a promise at some point, yet to be implemented for HW1 or HW2. I've never been able to get my AP to take an exit ramp by using the turn signal.

 

[sorka]

That was a promise at some point, yet to be implemented for HW1 or HW2. I've never been able to get my AP to take an exit ramp by using the turn signal.

That's why I was confused. The post I responded to seemed to claim that it already works this way which would have been big news to me.

 

[Carl]

Confused. Does AP1 navigate to freeway interchange ramps automatically based on the nav route????

No. But if at an interchange for example the two left lanes go south and the right lanes west onto another highway, you won't have to deactivate AP to manage the interchange.

 

[suraj1194]

I got my car in July/August 2016. Here's what I recall of the changes in AP1 since then:

* Indicator driven lane changes were more like veering across lanes. They are now smooth as if piloted expertly.
* Exit ramps used to cause brain fade with the AP trying to veer one way or the other, then give up and bail with the audio alert. Not anymore. It notices off ramps and avoids the hunting, sticking on course.
* Ditto for two lanes merging into one. It used to hunt left and right while trying to capture the new lane boundaries. Not anymore.
* Freeway ramps curvature seems to be taken into account reliably now. Earlier it would keep at the preset speed too long, rather than slow to a judicious speed for the ramp before speeding back up. It does so now.
* When the car in front kept me from going at preset speed and then it changed lanes opening up my lane, the car used to bolt forward like a jackrabbit (sort of). Not anymore. The acceleration is smooth, starting out slow and building momentum, rather than jerk forward.
* The indecision between 'follow the lane' vs 'follow the car in front' that used to happen when the car in front switched lanes, does not happen anymore.
* Cresting hills and bottoming underpasses used to cause AP to become indecisive and even bail out. This is noticeably less apparent now.

Things I'd like to see improve:

* When riding much faster in the carpool lane around a curvature with most cars in the adjacent lanes way slower, AP / AEB gets confused and applies brakes thinking a collision is imminent. The algorithm could better track the lane rather than act due to cars across the lane in adjacent lanes.

Feature addition suggestions
* Maybe there's already one and I don't know it, but some way to change multiple lanes without turning the turn signal off and back on.
* Recognize stop lights and stop signs. Already a capability of Mobileye EyeQ3 hardware. Just not being handled by the software.

I'll add more as and when I remember.

 

[eschummer]

I got my car in July/August 2016. Here's what I recall of the changes in AP1 since then:

* Indicator driven lane changes were more like veering across lanes. They are now smooth as if piloted expertly.
* Exit ramps used to cause brain fade with the AP trying to veer one way or the other, then give up and bail with the audio alert. Not anymore. It notices off ramps and avoids the hunting, sticking on course.
* Ditto for two lanes merging into one. It used to hunt left and right while trying to capture the new lane boundaries. Not anymore.
* Freeway ramps curvature seems to be taken into account reliably now. Earlier it would keep at the preset speed too long, rather than slow to a judicious speed for the ramp before speeding back up. It does so now.
* When the car in front kept me from going at preset speed and then it changed lanes opening up my lane, the car used to bolt forward like a jackrabbit (sort of). Not anymore. The acceleration is smooth, starting out slow and building momentum, rather than jerk forward.
* The indecision between 'follow the lane' vs 'follow the car in front' that used to happen when the car in front switched lanes, does not happen anymore.
* Cresting hills and bottoming underpasses used to cause AP to become indecisive and even bail out. This is noticeably less apparent now.

I'll add more as and when I remember.

Interesting Suraj. What does yours do when someone cuts in in front of you? Even though it would often be sufficient to just ease off the accelerator a bit, mine slams on the brake and slows down to where the car behind has to brake hard was well. Then it takes for ever to get back up to speed , opening an even wider gap in front, thus inviting more cars to cut in, etc. In those situations I tend to just turn it off. It didn't used to do that before the last iteration.