.48 feels like AP2 finally passed AP1

[AnxietyRanger]

Putting together EyeQ3's traffic light abilities, the icon imagery pulled from Teslas (with traffic light icons) and the combination of AP1/AP2+ codebases, it doesn't sound completely impossible that traffic light recognition and reaction might be coming even to AP1.

That said, what worries me AP1-wise, is the limited FoV of the camera. I would have always thought the AP1 Model X dual camera setup ("AP 1.5"?) would have been required...

 

[lunitiks]

Good point. I didn't have time to evaluate the distance from my camera to the traffic light.

All I noticed was that the light turned green, then my TACC-enabled car started moving, even though the car ahead of me stood still.

I'll be sure to try this again when possible and report back

 

[croman]

BTW, when my car fails to properly identify a car, it will surge ahead, so properly identifying and classifying cars is more than just reassurance that the system is functioning properly, it is a litmus test for actual proper function of the system. I'm not sure how often AP1 loses a car in front, but it happens with AP2 during autolane changes (probably because it cannot track adjacent traffic that well). I think Tesla is not showing adjacent traffic on AP2 because it would confirm how poorly it tracks cars as compared to AP1.

I've even had AP2 lose a tracked car at lights, like @lunitiks described, but it wasn't due to a light change. It just lost the car (visually and radar). One was a garbage truck (way back in March 2017 with a primitive AP2 fw). One was recently. The car kept blinking in and out and my car would surge forward and stop. I had to disengage AP but its unnerving because I can't readily predict when it will fail to notice a car (it had tracked that car for several miles before failing). My car chose the stopping distance between us, so its not like I engaged AP too close to the car and that's why it lost it.

So while AP2 performance is approaching AP1, I think the devil is in the details and I believe AP2 has some refinement before it can be as trustworthy as AP1.

I will disclose a very limited experience with AP1 (loaners and friends' cars). I've only owned AP2. So discount away but unless AP1 owners experience their system failing to recognize a vehicle in front and surging ahead (happens once a month or every 2k miles), then I think the comparison is valid. If it happens with AP1 and I just never noticed due to my limited experience, well, then we know AP2 is closer to AP1 than I realized.

 

[AnxietyRanger]

Good point. I didn't have time to evaluate the distance from my camera to the traffic light.

All I noticed was that the light turned green, then my TACC-enabled car started moving, even though the car ahead of me stood still.

I'll be sure to try this again when possible and report back

Could there have been a second car moving ahead of the car immediately in front of you?

 

[lunitiks]

Could there have been a second car moving ahead of the car immediately in front of you?

Man, I didn't think of that either. Could've been. I think it might've been.

Surely that might be the reason.

Trying to reproduce this might take some runs, since the cars ahead usually start rolling before the light turns green (at yellow). But I'll be sure to check it

 

[croman]

Man, I didn't think of that either. Could've been. I think it might've been.

Surely that might be the reason.

Trying to reproduce this might take some runs, since the cars ahead usually start rolling before the light turns green (at yellow). But I'll be sure to check it

AP2 waits a very long time before starting up in stop and go (and especially traffic lights). The gap gets huge and then it hunts the lead car down again (which is annoying to me).

Does AP1 usually get going without a large gap?

 

[chillaban]

AP2 waits a very long time before starting up in stop and go (and especially traffic lights). The gap gets huge and then it hunts the lead car down again (which is annoying to me).

Does AP1 usually get going without a large gap?

If you tap the gas a bit it will close the gap quicker. I think it’s meant to provide you some reaction time if the car is inappropriately inching forward like when it loses track of the lead car.

 

[lunitiks]

Does AP1 usually get going without a large gap?

No, it's normally way 'slower' to move than I'd do driving manually. Probably one of the reasons why I noticed to this incident.

 

[croman]

If you tap the gas a bit it will close the gap quicker. I think it’s meant to provide you some reaction time if the car is inappropriately inching forward like when it loses track of the lead car.

I know, I will tap the accelerator to ensure a prompt reaction at lights. I don't mind doing that and I hope that the stoplight/sign recognition will operate similarly (like the HOLD for pedestrians that trigger sonar).

I just wanted an AP1 perspective on the gap issue as it has varied across AP2 builds (some being more or less responsive to traffic clearing in front and some allowing larger and some smaller initial gaps).

 

[lunitiks]

AP1 is no better at reacting 'normally fast' to a car ahead that's still and starts moving.

OTOH, one big and clear improvement on AP1 since I got the car, is that it won't slow too much down when a car ahead takes an exit. In the beginning, my car would almost come to a halt when that happened, but now it's got a much 'smoother understanding' of what the car ahead is actually doing. But that's old news, really.

 

[chillaban]

AP1 is no better at reacting 'normally fast' to a car ahead that's still and starts moving.

OTOH, one big and clear improvement on AP1 since I got the car, is that it won't slow too much down when a car ahead takes an exit. In the beginning, my car would almost come to a halt when that happened, but now it's got a much 'smoother understanding' of what the car ahead is actually doing. But that's old news, really.

Yeah, I completely agree with both points. AP1 seems equally cautious about starting back up again. Realistically speaking, we're all on our phones or whatever when the car is stopped with TACC/AP engaged, and when the car starts rolling again is when we lazily look up at what's going on. From a liability/risk standpoint, seems like one of the more likely cases where AP/TACC could cause a crash.

I agree with AP1 and lane assignments. AP2 has been getting better at reacting to cars that move out of your lane, but IMO it's still way too conservative about that. Sometimes it will copy the deceleration of a car on a nearby offramp for seconds before snapping out of it. AP1 seemed to be much more willing to say "ok car is moving out of my lane, I can start accelerating towards it even if it disrespects the TACC set distance"

 

[Joe F]

Let's put it this way. If the only difference was rain-sensing, I'd agree. (That said, it is interesting to note that rain-sensing now apparently is an "AP feature", from AP2 onwards, so it is not surprising people are watching the topic on Autonomous Vehicles forum.)



I am passing no judgement on "so what". I am interested in the objective status of the situation. Speed sign recognition is a genuinely important AP feature related to speed matching. If people are putting out the idea that AP2 has reached AP1 parity, I am expressing my opinion on that. It is useful for people considering purchase decisions to know the status.

Both AP2 and AP1 use a speed sign database, like most speed sign recognitions in the last decade. But AP1 also has actual speed sign recognition to improve the results, again like most speed sign recognitions in the last decade. AP2 does not. Add to that more limited auto lane changing and less visual IDing in AP2. That simply means AP2 has not reached feature parity yet with AP1.

"So what" is wholly subjective. I am merely interested in the objective status of what's what.

But it's not a difference: Rain-sensing is not an AP feature of AP1. Certainly not a feature of AP2, but if they ever release it, will have to use a camera based solution given the lack of a dedicated sensor. For me, it's a non issue as I rarely use my wipers anyhow given proper windshield treatment.

Can you cite a Tesla link or reference for your statement it's an "AP feature" from AP2 onwards? I've not been able to find anything on Tesla which speaks to this statement. Thanks in advance.

Speed matching refers to matching the speed of a followed vehicle, as defined, at least in the written description of AP. Certainly in a FSD case it needs to address and respect posted speed limits, if/when we get there. While the current attempts at not exceeding posted speed limits are nice to have, people posting here for both AP1 and AP2 have noted mistakes for both. For the case of AP2, it's well documented, and now with the great work several have made towards understanding the map tiles, it's obvious as to why AP2 fails. The upside is, there's a possibility of resolving the map errors. I'll agree though, AP2 needs to have proper speed sign capability. However, and I really hate to point this out, the use case for AS is highway use, where speeds don't often change as they do for roads the pass through towns where the speed limit can change every other minute.

I've never owned or driven an AP1 Tesla. What happens in areas where the posted speed changes due to construction or another temporary reason? Does AP1 react accordingly? I'm sure you may not have the answer given you've never owned an AP1 Tesla, but would welcome others findings.

 

[AnxietyRanger]

Can you cite a Tesla link or reference for your statement it's an "AP feature" from AP2 onwards? I've not been able to find anything on Tesla which speaks to this statement. Thanks in advance.

Well @verygreen says rain-sensing is being handled by a NN inside APE - and FSD regulations apparently require rain-sensing to be handled somehow. This is mostly what I mean by rain sensing inside AP, that it technically is there and that it is required.

But as said, if this was just about rain-sensing, I'd let it go and call feature parity. Speed-signs recognition, certain lane changing and visual identification are more AP related feature lacks.

 

[AnxietyRanger]

I've never owned or driven an AP1 Tesla. What happens in areas where the posted speed changes due to construction or another temporary reason? Does AP1 react accordingly? I'm sure you may not have the answer given you've never owned an AP1 Tesla, but would welcome others findings.

I welcome others chiming in, but as far as I know, AP1 will limit its speed accordingly.

 

[WannabeOwner]

the use case for AS is highway use, where speeds don't often change as they do for roads the pass through towns

That's not the case here. We have traffic flow speed limits on gantry above all lanes (i.e. no chance that truck in adjacent lane blocks the sign) when traffic is heavy. They can change on successive gantries a few hundred yards apart, although in practice once you get one change the next several gantries are likely to post the same limit. They are backed up by speed enforcement cameras and fines. Not reacting to them is not an option ... and obviously a static database is no use either.

 

[chillaban]

I welcome others chiming in, but as far as I know, AP1 will limit its speed accordingly.

AP1 will read speed limit signs and take that to be more truthful than the maps. This is good for most cases, but bad in some cases:

For example, in my area they put the signs sideways or upside down on the ground when construction zones are inactive. The AP1 vision algorithm doesn't seem orientation-sensitive so it will randomly read signs dangling around that no human would consider reasonable.

It also sometimes reads yellow advisory signs as speed limits which can be annoying, or when the "55mph if you're a truck" sign is separate from the 65mph sign, it'll read that. And sometimes it will hilariously misread things like Route 85 as 85mph.


If the mapped speed limits were kept more up to date I would honestly like those better than sign reading.

 

[AnxietyRanger]

If the mapped speed limits were kept more up to date I would honestly like those better than sign reading.

I'm not sure about the U.S., but in Europe we have electronic speed-signs, though. No database will help with them.

That said, I'm not saying the database is useless. If it was a good database, it might be useful in some cases. What is bad is that the new database is much worse than the old (pre-mid-2017 or so). It too often shows too big speed limits, which is not good...

 

[chillaban]

I'm not sure about the U.S., but in Europe we have electronic speed-signs, though. No database will help with them.

That said, I'm not saying the database is useless. If it was a good database, it might be useful in some cases. What is bad is that the new database is much worse than the old (pre-mid-2017 or so). It too often shows too big speed limits, which is not good...

That's a good point — very few parts of the US have dynamic speed limits. And in general, US enforcement of speed limits is frankly a joke compared to the EU. I shouldn't jinx this but I literally completed a 1200 mile trip going a minimum of 15% over the speed limit the whole way and there was maybe 1 opportunity where I could have gotten a ticket. In Europe I presume by the time I get home my mailbox would be stuffed full of photo tickets.

I do think that some balance of the two approaches seems right. But right now neither AP1 nor AP2's approach are very smart.

 

[Joe F]

I'm not sure about the U.S., but in Europe we have electronic speed-signs, though. No database will help with them.

That said, I'm not saying the database is useless. If it was a good database, it might be useful in some cases. What is bad is that the new database is much worse than the old (pre-mid-2017 or so). It too often shows too big speed limits, which is not good...

Mostly fixed side of the road speed signs here. There are some electronic I've seen, they too are on the side of the road. I do not ever recall seeing overhead speed signs, at least not on any highways I've been on.

And I agree, the speed limit database as it stands now is crap.

 

[AnxietyRanger]

And I agree, the speed limit database as it stands now is crap.

I forget, did we know for sure who was the old database provider and the new?