.48 feels like AP2 finally passed AP1

[jimmy_d]

This thread is to collect good/bad use cases comparing AP2 firmware revision 48 to recent AP1 performance.

I upgraded from my AP1 S90D to an AP2 S100D a few months ago and since then have been through firmware versions .38, .40, .42, and now .48. As AP is the single most important feature to me I had timed my upgrade to minimize the window of crappy performance that AP2 was suffering relative to AP1 over the last year or so while still meeting certain other personal requirements. As such I've been very interested to compare how AP2 performs relative to AP1 and how that relative performance is changing over time.

There were a couple of serious deficiencies in AP1 that I found were corrected in AP2 when I first swapped over but it was also the case that AP2 had a couple of glaring deficiencies which made it unusable in a couple of important (to me) use cases that I regularly encounter. One of those deficiencies was mostly corrected as of .42 and now .48 seems to have solidified that improvement while also correcting the other one. This conclusion is kind of preliminary because I only have about 100 miles on .48 and haven't taken it on a road trip yet, but if the behavior I've seen so far holds up I expect it .48 will be the firmware revision that makes AP2 clearly better than AP1 for me in all my recommended use cases.

To be clear: I live in CA and use AP on urban freeways and on urban and rural limited access highways. My expectation of the system is that AP should work the large majority of the time in all those situations barring weird situations like inclement weather, road construction and so forth. And by 'work' I mean I can use it in the recommended fashion without being nervous about it's behavior. For me AP is a stress reducer and safety feature, and I find it to be wonderful in both of those aspects, though still not perfect.

Anyway, I think that as of .48 AP2 now works as well or better for me in all my use cases than AP1 did. The final use case that I was having problems with on AP2 (as of .42) was urban freeways with dense free flowing traffic and lots of curves. For me AP2 .42 was not doing an adequate job of holding it's lateral position in a lane on curves, occasionally even crossing lane boundaries. 48 seems to have fixed that, at least to the extent that it seems to be performing on par with AP1 as of firmware 42. I suspect 48 will still not be able to drive the entire grapevine (HW 5 crossing the San Gabriel mountains) without interventions, but then AP1 hasn't been able to do that for me yet either.

Anyone else see improvements in lane holding in 48?

What common and appropriate use cases are still failing in 48?

 

[S4WRXTTCS]

For me the real problem with AP1 is it's unpredictability, and a big part of why I don't use it with exception of TACC.

I can pass truck after truck without anything, and then suddenly truck lust
It can go past exit after exit just fine, and then swing for one with no reason
Sometimes it brakes (not massively), but enough to go WTF?

I still use it when the situation is so simple (like bumper to bumper, or middle of nowhere), but I don't tend to use it that often.

So I'm curious about what people think about AP2 at .48 who didn't think the lane-steering of AP1 was all that great.

 

[chillaban]

For me the real problem with AP1 is it's unpredictability, and a big part of why I don't use it with exception of TACC.

I can pass truck after truck without anything, and then suddenly truck lust
It can go past exit after exit just fine, and then swing for one with no reason
Sometimes it brakes (not massively), but enough to go WTF?

I still use it when the situation is so simple (like bumper to bumper, or middle of nowhere), but I don't tend to use it that often.

So I'm curious about what people think about AP2 at .48 who didn't think the lane-steering of AP1 was all that great.

Soooo I thought AP1 was good but not great. It had many many very predictable faults: Truck lust when driving next to long straight objects, losing lane markers if the car in front rides on them, a propensity to "monkey see monkey do" when the driver in front does something stupid despite blatantly clear lane lines to the contrary, etc.


AP2 actually does do a good job of addressing all of those points. It's been a long journey for AP2 to start feeling trustable -- AP2's faster computation time seems to lead to it applying a strong incorrect input at a split second and especially before .42, "be prepared to take over immediately at any time" was really the MO. But yeah, in 2017.42 and beyond, especially in 2017.48, I really do find AP2 to be at or above AP1's performance level. Especially if you frequently drive above 65mph.

Sure, we can quibble over the other aspects of AP1 vs AP2 (e.g. wipers, auto high beam effectiveness, cosmetic display of vehicle type / vehicles in other lanes), but quite honestly with regards to the last item, AP1 never seemed to do anything useful with the information it was displaying other than white cars. So I personally find that to be more or less cosmetic.

 

[daktari]

I have not tried .48 yet, as an error with a side repeater disabled autosteer, cruise etc completely. But AP2 .46 compared to AP1 (loaners) was lacking on lane tracking in more difficult conditions compared to AP1. Also, on a 100 km mostly local road, AP1 would drift out of lane 3 or 4 times on top of a crest, whereas AP2 would need apprx 30 take overs because of swerving on every slope up and down and not keeping in its lane in bends. AP1 was not perfect but more predictable.
Really hoping my experience is similar to yours when I get my car fixed.

 

[TaoJones]

Chill mode with .48 takes some of the rough edges off AP2. Still testing but so far so good and noticeably moreso with .48 than <.48.

AP2 use cases that still need work include but are not limited to:

1. Carpool lane lateral safe distance relative to no-shoulder (K-rail) barriers.

2. Any kind of elevation shift when approaching intersections. AP2 unreliable. AP1 wasn’t much better.

3. Lane (position) shifting - this would include truck lust (overtaking) and near lane head-on encounters (you and incoming traffic are in respective near (left in the States) lanes). You’d prefer that the car shift left in the former and shift right in the latter. My recollection from AP1 is that nudging the wheel was slightly more forgiving than AP2. In any case, what would be helpful here would be auto-nudge on steroids.

4. Low speed TACC management - this is one of AP2’s weakest areas relative to AP1 or otherwise. With AP1, you could deploy TACC down to 1mph and, whether AS was deployable or not, TACC stayed engaged. Now with AP2, *if* you can deploy TACC, then frequently when you try to add AS, you lose TACC. This is a problem when TACC provides braking/safe distance in close quarters.

While it seems implausible that the “silky smooth” behavior to which Elon referred ~8 months ago was in fact what became Chill here at year end, the combination of Chill and .48 is the closest to silky smooth that AP2 has been able to provide to date. You find acceleration a tad lacking when wanted but the other 95% of the time, it niiiiiiice (Ron White voice).

So far I’ve tested Chill/.48/Sport over maybe 150 miles and not yer through the local proving grounds (which include Portuguese Bend in both directions at both dawn and dusk). Presumably Chill/.48/Comfort would be yet silkier and yet smoother still. The ultimate currently-available solution for SoCal stop and go, perhaps.

 

[jimmy_d]

1. Carpool lane lateral safe distance relative to no-shoulder (K-rail) barriers.

Yes. Zero shoulder K-barrier lanes are uncomfortable. AP seems to be a lot more comfortable driving 2 feet from a K-barrier than I am. I can't decide if that's appropriate or not (it knows the distance to that barrier a lot better than I do, after all) but it sure makes me uncomfortable. I do it sometimes (just for testing) but I don't like it and I don't recommend it.

2. Any kind of elevation shift when approaching intersections. AP2 unreliable. AP1 wasn’t much better.

Yes. Elevation changes in general were weak as of .42 (I haven't had a chance to test them on .48 yet) and also on AP1. It seems like AP2 elevation change fails at 42 and earlier were related to general lane holding instability with elevation changes being one of the extreme cases. My experience with AP1 was rather that it really didn't like situations where it temporarily loses sight of the lane markers, for instance with crests in the road.

3. Lane (position) shifting - this would include truck lust (overtaking) and near lane head-on encounters (you and incoming traffic are in respective near (left in the States) lanes). You’d prefer that the car shift left in the former and shift right in the latter. My recollection from AP1 is that nudging the wheel was slightly more forgiving than AP2. In any case, what would be helpful here would be auto-nudge on steroids.

So this was the thing I found most improved between 42 and 48 - the car staying in the middle of the lane without wandering too much. Are you still experiencing to the same degree on 48?

4. Low speed TACC management - this is one of AP2’s weakest areas relative to AP1 or otherwise. With AP1, you could deploy TACC down to 1mph and, whether AS was deployable or not, TACC stayed engaged. Now with AP2, *if* you can deploy TACC, then frequently when you try to add AS, you lose TACC. This is a problem when TACC provides braking/safe distance in close quarters.

Interesting. One of the things I liked the most about AP2 relative to AP1 when I first got it was that it's behavior in the rightmost lane of a busy highway was better. With AP1 I had to disengage because it didn't seem to manage a safe following distance when cars were frequently merging or exiting the lane. In AP2 I don't have this problem. I attributed that to TACC being better at low/moderate speeds in complex situations. Though if I understand your objection it's more about not being able to engage TACC in situations that you would like to rather than how TACC behaves at low speed - is that correct?

 

[jimmy_d]

It's been a long journey for AP2 to start feeling trustable --

I didn't get a chance to use AP2 until to recently, but I had AP1 from the start. Over the two-odd years that I drove on AP1 it improved tremendously and, at least so far, it seems like AP2 is going through a similar evolution. In the beginning AP1 seemed miraculous and I think I gave it a lot of credit for being able to save me any effort at all, but by the time I got into an AP2 car my standards had risen to the point where things I forgave in early AP1 I found to be intolerable in AP2. I'm hoping that period is finally coming to an end.

 

[chillaban]

I didn't get a chance to use AP2 until to recently, but I had AP1 from the start. Over the two-odd years that I drove on AP1 it improved tremendously and, at least so far, it seems like AP2 is going through a similar evolution. In the beginning AP1 seemed miraculous and I think I gave it a lot of credit for being able to save me any effort at all, but by the time I got into an AP2 car my standards had risen to the point where things I forgave in early AP1 I found to be intolerable in AP2. I'm hoping that period is finally coming to an end.

I've told this story before:

My first drive with AP1 was right when Autosteer was enabled (7.0?), and long story short, we actually ended up nearly sideways on 280 and the salesperson screamed like a girl.

I ridiculed Autosteer for almost the next year. The next time I tried it was almost a year later with a late build of 7.1 and at that time, the same hardware blew me away in how well it worked relative to my Audi's ACC + lane holding system. I bought a Model S70D 2 weeks later and parked my Audi lease for almost 9 months collecting dust.

During my time with AP1, it started off fantastic, and 8.0 and beyond only made it better. But incrementally so, compared to when I joined.


So yeah, I guess I joined AP1 at such a great time that all I could see were the limitations of a well-polished system. I joined AP2 much earlier and got to see it evolve from just a demoware quality slow speed lane holding system all the way to its current state in 2017.48, which is rather commendable.


My overall opinion is still that Tesla in one year deployed the hardware and developed the software that largely performs what EyeQ3 does, which is the product of around a decade of experience from the domain-leading automotive supplier combined with 2-3 years of systems integration work on Tesla's part. That's rather impressive. But still, none of that excuses from a good business practices standpoint, the way AP2 was marketed and the promises made with ambiguous timelines to sell the feature.

 

[jimmy_d]

So I'm curious about what people think about AP2 at .48 who didn't think the lane-steering of AP1 was all that great.

Would be nice to hear other people weigh in on this, but just to flesh out my own:

AP1 seemed to hold lanes quite well as long as the visibility was ok and the lane markers were clear. It could coast through a section of road where the lane markers disappeared briefly but it couldn't deal with ambiguous lane markers. For instance, the HOV lanes in LA have a white lane line on the inside with a double yellow on the outside. If you drive in the lane adjacent to the HOV lane AP1 will center the car between the white line on your left and the white line on your right, which means you are often driving on top of the double yellow (which is REALLY annoying). In this case AP1 is ignoring the double yellow in favor of the white line. AP1 always makes this mistake and AP2 never makes this mistake.

AP2 also detects lane boundaries accurately in a whole host of complicated situations where AP1 fails - like intermittent curbs or other pavement boundaries, unusually wide lanes, faded white markers on white concrete, temporary lane marker tabs on newly constructed but not yet painted roads and so forth. In short it's weaker with clear and unambiguous lane boundaries but stronger otherwise. So AP1 fails when you have clear lane markers that suddenly do something weird, but if they are clear, unambiguous, and consistently in the field of view it will be rock solid in the center of the lane.

The big failing of AP2 up to .42 is that it wanders around the lane a lot more than AP1 and it is slow to correct itself. So if you hit a series of curves, especially an S bend you're often hitting the second curve just as the car is overcorrecting for missing the first one and you wander out of your lane - which is a huge fail IMO. In urban traffic where you have a lot of curves on a multilane highway you have to be concerned with how your car's movement might be interpreted by the drivers around you. In that situation I find that I had to disengage 42 a lot in order to not impact other drivers. So far I'm not finding that with 48 - it seems to wander less and correct faster. I'll have a lot more data in a few weeks but I found the improvement to be so substantial that I wanted to hear from other drivers.

 

[chillaban]

The biggest issues I'm seeing in .42 (can't speak to .48 quite yet) are:

(1) In sharp sweeping curves it will wander way too much near the edge of the lane or even outside of the lane. The screen clearly shows it knows it's off-center but the car does not correct itself until much after exiting the curve. This seems to have improved on .48 but I don't think it's completely gone….

(2) When hitting bumps, it seems to suddenly destabilize AP2's lane recognition, sometimes resulting in unwanted swerving.

(3) When tracking certain curves combined with hills, AP2 has trouble. But AP1 had a lot of trouble in those situations too, and it's largely a toss up whether AP1 or AP2 is better at doing this at the moment.

 

[WannabeOwner]

situations where AP1 fails - like intermittent curbs or other pavement boundaries

AP1 is fine on divided highways here - they have good road markings. On undivided I only engage AP1 if the road has a line on outside of lane, but plenty or roads around here have a well defined curb, or just a (e.g. grass) verge. I don't engage AP1 on either of those ...

... makes me wonder what is different between AP1 and AP2 in that regard. Is it possible that MobileEye software could not cope with curbs instead of lines? or is the AP2 extra hardware the winner here?

Seems unlikely (to me) that MobileEye had not already achieved that ability (or astounding that Tesla has done it in a short time-frame ...) so maybe MobileEye had that ability (along with being able to identify lots of different street signs) but just not "turned on" in AP1 ...for some reason ... if so: can I have that in AP1 for Christmas please?!

 

[Blup85]

5k on ap1, 150 miles on ap2 loaner MX 100d.

They do different things better.. that's the best thing I can describe it as..
Ap2 doesn't wander as much cresting hills as my ap1 car does, but i ap1 totally does silky smoothe reverse curves (80 up to tahoe) way better than ap2. I also don't like the way ap2 hugs curves next to vehicles and concrete barriers..
My prefer ed nice is still ap1, but i love 20th he MX 100d I got as a loaner.
Almost have the wife sold on selling the rati for a 90 d MX.

 

[Nosken]

Over 20,000 using AP1 and over 15,000 using AP2 here. Out of almost 80,000 total miles. Those of us with earlier deliveries of AP1 cars had the luxury of learning the wonderful TACC before we got to learn AutoPilot. For me, learning each separately, produced less angst, and caused less leg cramps as my ever ready, foot hovered over the brake pedal. I leased my original Model S since I new full self driving and the Model E were inevitable.
AP1 had teething issues in the beginning, but those of us that started using it in October 2015 were experiencing a new and different experience, so the early "irregularities" in performance were masked by the new experience. When I left AP1 in February 2017 it had refined nicely. Gone were the days of diving for many exits, gone was the "drunk driving" while hunting for a lane. There was still the desire for the car to turn abruptly when going over rises in the roadway. There was the reassurance that the radar was seeing the other cars, motorcycles and trucks being displayed in the lane of travel and adjacent lanes in the binnacle.
I was so intrigued with AP2 and the wonderful autonomous driving video posted in late 2016, I decided to get out of my P85D lease a year early and purchased an S90D 12/31/16, just when unlimited Supercharging was to go away. I ordered FSD, that has been shown to be premature.
AP2-Ouch, I was back to the days of Ping Pong drunk driving and my cars attraction to random freeway exits.
AP2 has not evolved as quickly as anyone wanted, but I am encouraged by how "Silky Smooth" it has become. I think that the small, incremental improvements we experience, mask how much better it has become. If you notice, there are less complaints about auto high beams. That "feature" was abysmal when introduced. Now it is at the least, acceptable. OK, rain sensing wipers....well we won’t talk about that
My car is less attracted to freeway off ramps and truck lust is gone.
My takeaway is that AP progress is tough, and I’m sure some of the hardest engineering is dealing with the mitigation of false positives. We all experience the awakening, when our car starts to brake for over crossings and overhead signs, but I marvel at the complexity of what is really going on here.
We have all learned to anticipate the problem areas we encounter on the roadway, as more sensors get integrated, this will lessen.
One of the main reasons I bought my Tesla was safety. While crash testing is superb for Tesla, crash avoidance is preferred. AP1 cars didn’t have forward collision avoidance initially, but was added with an OTA update, and was further enhanced with the development of differential radar, where the car in front of the car in front was detected. This alone is remarkable, but hardly gets acknowledged. Radar has saved me from rear ending someone once. Thank you!
As I get older, to assure my independence, I want FSD to arrive by the time I am to old to safely drive.

 

[chillaban]

2017.48 to Vegas: so far 400 miles and 1 safety related disengagement: hit a bump before an uphill curve in Tehachapi pass and the car steered the opposite direction of the curve. Caught it before departing the lane.

Overall this is a higher level of performance than ap1, and I’ve done this trip 7 times in AP1 late last year to early this year.

And my speed has been set at 80 or so this whole time.

 

[skitown]

The theme that AP2/AP1 are basically at parity (or AP2 is now better) - albeit with differences, is definitely the new normal. Good job helping get us to this promised-land Buttershrimp. .48 feels like AP2 finally passed AP1

 

[buttershrimp]

The theme that AP2/AP1 are basically at parity (or AP2 is now better) - albeit with differences, is definitely the new normal. Good job helping get us to this promised-land Buttershrimp. .48 feels like AP2 finally passed AP1

Did someone say Buttershrimp? First new video from 17.48

 

[jimmy_d]

Did someone say Buttershrimp? First new video from 17.48

I've been playing around with .48 on similar rush hour surface street drives here in LA - not recommended behavior of course, but curious about it's capabilities. So far it has managed everything fine - except of course that it doesn't automatically stop at lights/signs so you have to intervene there if there's no helpful lead car to follow. It'll do 8 miles of surface street stop and go with no interventions as long as the intersection timing works out right.

And of course I'm listening to J-pop.

 

[TaoJones]

So this was the thing I found most improved between 42 and 48 - the car staying in the middle of the lane without wandering too much. Are you still experiencing to the same degree on 48?

It’s the fixed middle position that’s the problem. Wandering or no, and yes, that’s been a problem during AP2, the inability to shift right for oncoming and shift left for overtaking increases risk - both in terms of safety and for rock chips - especially with the refresh.

...I attributed that to TACC being better at low/moderate speeds in complex situations. Though if I understand your objection it's more about not being able to engage TACC in situations that you would like to rather than how TACC behaves at low speed - is that correct?

Using the simplest case of stop and go same lane traffic, I would prefer that AP2 work as did AP1 and certainly not worse than AP1 as it does now at low(est) speed. I don’t recommend reproducing the use case but if you happen to experience it, it’ll be at low speed - again, after engaging TACC, and trying to engage AS, that attempt will both not engage AS *and* will cancel TACC. And that’s bad as it takes out your braking and buffer.

Something new:

Here’s a new one experienced just now exiting the 105W onto Crenshaw in the right lane. Car slowed down to accommodate the mild curve once exiting until I signaled to get into the left lane in anticipation of a left turn ahead. Car made the lane change automagically which was fine, but then immediately attempted to accelerate to 70mph, completely ignoring the fact that there was a curve, let alone a lower limit. Not Good. Disengagement ensued.

Coming down Crenshaw of course, the usual disengagements aplenty.

Chill mode continues to impress. Early morning across a collection of LA’s finest poorly-marked freeways (hello, 91, hello 605), tracking was good with .48 and *smoother* with Chill mode on than ANY AP2 to date. Yep - AP1-Level smoothness, especially in lane changing.

What do you lose? Top end acceleration. But as long as you remember that, it’s not horrible.

Another glimmer of hope from this morning that I’ve yet to reproduce: When straying from my lane on a 7-lane surface street, the red hands of death appeared. That was refreshing.

Overall I’d describe AP2 at this time as... brittle, with potential. Conversely, AP1 provided and continues to provide confidence across the board. AP2... not so much.

Another fellow put it well: TACC is decent - but AS remains underwhelming.

 

[avoigt]

I guess so!

 

[AnxietyRanger]

The theme that AP2/AP1 are basically at parity (or AP2 is now better) - albeit with differences, is definitely the new normal. Good job helping get us to this promised-land Buttershrimp. .48 feels like AP2 finally passed AP1

Let's talk when we have speed sign recognition, vehicle identification display and auto-wipers in AP2... 'mmkay?