Firmware 8.1 - Autopilot HW2

[Tangible]

Thank you @2Cybers , for this very thorough report. Since we're not told much by Tesla about the inner workings of their systems, such observations and inferences can be very helpful.

My understanding has been that autosteer won't turn on at all until calibration is deemed to be complete, and that the very radical behavior you first experienced is not typical. What you're seeing now, is. Also, I don't think there's been any evidence that learning is something that happens to a particular vehicle running a particular software release, though some people have reported that repeat trips over the same route get better. I'm skeptical of these reports. I'm sure it's an accurate reflection of their experience, but you never really take the same trip twice: traffic and weather are always going to be variables.

My experience with service center people is that while they're skilled at repairing the hardware they know less about the software features than many owners who follow these forums. The company's secrecy extends to it's own employees, who are therefore not equipped to address these questions.

 

[alcibiades]

17.24.30

 

[KArnold]

All in all, I don't think what I am seeing in AutoSteer learning curve is a whole lot different

Uh, I'm confused. These are COMPUTERS. Can't Tesla take what your car has "learned" and load it into my new build? What is that each and every car has to learn independently of one another?

 

[Tangible]

Uh, I'm confused.

Justifiably so. Tesla likes to use high-falutin' terminology to describe ordinary things. I imagine the elevators in their office buildings are called Gravity Mitigation Systems.

"Learning" is the process by which each release is supposed to be incrementally better than its predecessor. The degree to which Tesla gathers data from our driving experience and applies it to the next release is unexplained. But I don't think your version of a given release is different from mine based on our different experiences.

 

[2Cybers]

Earlier today I got a text from the local Service Department. They had pushed a firmware update. I installed it and a huge improvement. Good enough I would never have complained to Tesla or posted here if things had started this way.

Is it perfect - no. If I look, it is not perfectly centered in the lane. The Land Departure warnings are not symmetrical - the right side warning happens a bit early. The left is on target. Lane change with turn indicators worked. The handling of interstate road exit ramps was fine. A bit of side to side wandering - but no ping-pong at all. No change in the above when I went through a local rain squall.

I went from 17.18.50 to 17.24.30

 

[ulrichw]

Uh, I'm confused. These are COMPUTERS. Can't Tesla take what your car has "learned" and load it into my new build? What is that each and every car has to learn independently of one another?

Justifiably so. Tesla likes to use high-falutin' terminology to describe ordinary things. I imagine the elevators in their office buildings are called Gravity Mitigation Systems.

"Learning" is the process by which each release is supposed to be incrementally better than its predecessor. The degree to which Tesla gathers data from our driving experience and applies it to the next release is unexplained. But I don't think your version of a given release is different from mine based on our different experiences.

Different opinion here: I think the "learning" they're talking about here is different and has to be individual.

Specifically, I believe the issue is that there are variances in both the cameras and the placement of the cameras due to manufacturing tolerances. For example, a given camera may have a slightly lower contrast or a slightly different color response than another.

This kind of thing can throw off decisions the autopilot makes and need to be corrected for. Figuring out what the right correction is may take time and exposure to a number of different situations - i.e., individual "learning."

So while you definitely can copy the learning from one car to another, a difference a human would perceive as minor in the detection hardware in the two cars could drastically influence how the autopilot reacts to a situation. The autopilot needs to first figure out what the differences in the hardware are before it can properly apply the learning from the other car.

Disclaimer: This is just a guess on my part - nonetheless I think it's a plausible explanation why two forms of learning are necessary: "Learning" that is generalized across Tesla's network of vehicle, and specific "learning" that adapts the general to the hardware on a particular car.

 

[Doctt]

Different opinion here: I think the "learning" they're talking about here is different and has to be individual.

Specifically, I believe the issue is that there are variances in both the cameras and the placement of the cameras due to manufacturing tolerances. For example, a given camera may have a slightly lower contrast or a slightly different color response than another.

This kind of thing can throw off decisions the autopilot makes and need to be corrected for. Figuring out what the right correction is may take time and exposure to a number of different situations - i.e., individual "learning."

So while you definitely can copy the learning from one car to another, a difference a human would perceive as minor in the detection hardware in the two cars could drastically influence how the autopilot reacts to a situation. The autopilot needs to first figure out what the differences in the hardware are before it can properly apply the learning from the other car.

Disclaimer: This is just a guess on my part - nonetheless I think it's a plausible explanation why two forms of learning are necessary: "Learning" that is generalized across Tesla's network of vehicle, and specific "learning" that adapts the general to the hardware on a particular car.

You brought up a good point.

I am a new X owner. I got my car with 17.24.28. I found it work very well on except the first morning commute. The first morning was quite inconsistent. Heavy brake, sudden slow down, hug to the right size of the lane. Then after 100 miles, it got much much better and to the point i felt very comfortable using it on hwy 80% of the time of my long daily commute

Then I got the 17.24.30 2 days ago. Things turned ugly again. Same problems popped up. Then, again, after 50 miles. Work well again.

I suspect that every fw upgrade may recalibrate the camera or the system needs to relearn our habit

 

[verygreen]

I suspect that every fw upgrade may recalibrate the camera or the system needs to relearn our habit

The camera calibration process is ongoing, new firmware install does not wipe old stored data about it (besides 17.24.30 is almost the same as 17.24.28 wrt camera and image recognition anyway).

 

[Helmuth]

FW 17.17.17

I made a road trip today: 606 miles
visited 7 different Supercharger

AP2 use over 98% on the interstates
Needed to disengage AP ONCE
Disengagement was necessary because of heavy Florida downpour, the radar lost track of the car in front of me (heavy rain) and started to accelerate. Therefore I needed to disengage it, even though it still saw the lane markings (which was a surprise, I could barely see the lane).

I used auto lange change countless times without any issues!
I'm really happy how it performs on the interstates...

Not one phantom braking, no lane hunting (was centered or steered away from cars and trucks next to me), didn't want to take any exits, speed was set to 80mph most of the time, performed very well!

That was my experience from today!

 

[2Cybers]

I got my car with 17.24.28. I found it work very well on except the first morning commute. The first morning was quite inconsistent. Heavy brake, sudden slow down,

Are you aware that HW2 does not yet "read" speed limit signs? Rather the speed limit, as far as the car is concerned, depends on GPS readings and a database of posted speed limited downloaded to the car. There are limits on AutoSteet's speed - I believe it cannot be set more than 4 or 5 mph above the speed limit. So if you are trucking along say at 70 mph and move into an area the GPS database says is 55 mph, you will get a sudden slow down (or maybe some classed as "heavy braking'. You can determine if this is happening in Settings by having the "speed limit" displayed on your dashboard.

 

[Saghost]

Are you aware that HW2 does not yet "read" speed limit signs? Rather the speed limit, as far as the car is concerned, depends on GPS readings and a database of posted speed limited downloaded to the car. There are limits on AutoSteet's speed - I believe it cannot be set more than 4 or 5 mph above the speed limit. So if you are trucking along say at 70 mph and move into an area the GPS database says is 55 mph, you will get a sudden slow down (or maybe some classed as "heavy braking'. You can determine if this is happening in Settings by having the "speed limit" displayed on your dashboard.

Unless AP2 cars are different, the 5 mph over limitation only applies for what the car thinks are non-divided roads (there are some errors in this database) - it should never matter in the freeway.

When the car is worried about a radar return, it brakes harder than it does when it slows for a speed setting change, internal or external - and of course the new speed limitation will always be visible on the dash.

 

[_jal_]

Well, I'm still happy at 17.17.4, but the fact that so few AP2 cars have gotten an update after the 17.17.x builds sure makes me wonder what's going on in the AP department. I got updates every few weeks that incrementally made things better, and now it's been 9 weeks with nothing. Oh well. At least I can still wash my car.

 

[mitchellh3]

Well, I'm still happy at 17.17.4, but the fact that so few AP2 cars have gotten an update after the 17.17.x builds sure makes me wonder what's going on in the AP department. I got updates every few weeks that incrementally made things better, and now it's been 9 weeks with nothing. Oh well. At least I can still wash my car.

Same boat here, still on 17.17.4. And same thing: wondering what's going on with [E]AP.

Should we expect only perpendicular parking + display brightness this month? Or should we expect the next AP update to "go wide" to have more features? Also, the all too nebulous "silky smooth"...

I really want the to get to EAP features, but of course the speed at which parity alone is coming along is not inspiring. I know, I know, they're doing quite a lot in a short period of time. But as a buyer its still frustrating at times. I have a curvy driveway and summon is useless for me until "Summon+" or whatever they're calling it comes along. I'd love that.

 

[animorph]

Old news, but maybe there was more to the Chris Lattner departure than we know:

Tesla Autopilot: head of software Chris Lattner leaves, Tesla hires new head of AI and computer vision

 

[_jal_]

I guess what's weird is the tweet and email (not to me, but others) that there's a BIG upgrade to AP that is going to significantly increase performance. And I know AP1 users are reporting perceived improvements but Elon specifically said the new algorithm only runs on AP2 hardware so whatever AP1 owners are seeing isn't what Elon was tweeting about. And then... crickets. Oh, actually not crickets; we get news that the head of AP is no longer the head of AP. Hmmm.

Then you read Lattner's resume and even accounting for hubris (which I am unsure even needs to be accounted for because I don't know the guy) and it sure sounds like he changed a bunch of stuff in 6 months. But, shouldn't that part of development already been fairly well roadmapped? Makes me wonder if the new guy shows up and is all "Oh, no, this is all wrong! Let's do it this way..." and now we're two steps back.

I'm pretty much okay still with AP2, but it's been a while now. And had he not sent that tweet and instead just sent perpendicular parking and the goofy brightness thing, then at least I'd have some features. But he did send the tweet about significantly improved handling with AP and now I'm left wanting. It's just weird

 

[Cconway84]

I installed 17.26.76 on my 2016 hw2 S today after being stuck on 17.17.4 for some time. 40 highway miles do seem smoother on autopilot. It took a tight highway bend that semi trucks routinely crash at. I kept it at 68 and it was rock solid no seeking in the lane. On a few of my twisty rural 2 lane roads that are 35mph zones I used to have 6-8 failure points on curves in 1-2 miles. Now it was down to 1 and it gave me the red hands warning going into the curve where it failed. It slows down more preemptively going into curves on the 35mph roads and feels much more in control.

 

[mitchellh3]

Got 17.26.76 today also after being on 17.17.4 since release.

It definitely feels slightly better on curves and the accel/decel might be smoother but isn't dramatic enough to wow me in any real way.

The perpendicular parking is quite good if only for laziness reasons. Its still a bit awkward as other videos have shown due to it doing a 3-point entry (in most cases) into the spot. With AP1 I remember my spot being taken multiple times because the person behind me didn't realize I was about to park in the spot because no human being perpendicular parks like Autopilot. So that's still valid here.

What we really need is for Vision to be able to identify parking spots in front of the car (assuming the driver can verify, I realize the hardware can't know the spot is 100% usable, motorcycles might be hiding and such). If I could tell the car "yeah, parking in the next open spot in front me" and it did a much more human-normal reverse into the spot, it'd be great. Even nose-in would be fine. But hey.

I noticed that AP2 still struggles to confidently identify the HOV lane from the lane adjacent to it (to the right). For me, since 17.17.4 it has highly preferred the white line in the HOV line marker which makes it hug the left lane too hard and usually go over the bumps in the HOV marker. Basically an unusable lane for me on LA freeways. This seems like a [relatively] easy problem compared to the other problems AP has to solve.

Otherwise Autosteer feels pretty much like AS in 17.17.4. Not a lifechanging update for waiting for 7 weeks... but nice to see some changes.

@verygreen I didn't see in your post history... but if you have time: does 17.26.76 have new/modified NN models from 17.17.4? Or is this purely a control algo update and not a model update?

 

[verygreen]

I didn't see in your post history... but if you have time: does 17.26.76 have new/modified NN models from 17.17.4? Or is this purely a control algo update and not a model update?

There's a significant change in model (trained model size dropped in half ~29M -> ~12M, even smaller than 17.11.45 ~14M) and some kernels were renamed in the actual cuda code, though not sure if the logic there changed much.

 

[sdorn]

I recently got the update to 17.26.76 as well. I've only driven about 50 miles but didn't really notice any significant improvement over 17.17.17.

 

[ChrisZ]

I drove about 50 highway miles on 17.26.76 today and it was flawless. Well positioned in the lane, handled numerous interesting curves, regular lane changes, and exit ramps, all appropriately managed. I've only had my S100D for just over a week so only a little experience with the previous release but, at least on the highway I noticed smoother driving and good performance.