AP1 -- red and yellow lights?

[tls]

I know the Mobileye marketing stuff touts traffic light detection and that this was originally expected to be an AP1 feature. I also know every time any of us have thought our cars independently braked for a traffic signal there's been some other explanation. But I'm wondering how sure we really are that this feature isn't, and never operates under any circumstances.

Case in point: about a week ago I was driving, AP engaged, on a 2-lane road with a speed limit that dropped successively from 55 to 45 to 30. As you'd expect the car slowed from 60 to 50 to 35. In the course of this it also slowed for one or two tight curves. No surprise there.

However, when I neared this intersection in the 30MPH zone, the car did a surprising thing. The light went yellow not long after it came into view (much further away than you'll see in the linked Street View image), and the car braked gently -- probably slowed me down to somewhere in the high 20s. The light went red, and the car braked quite a bit harder -- it nearly stopped before I took over. The change in braking effort exactly when the light went red probably surprised me most of all, though brake engagement exactly when the light went yellow was also a shocker.

Late Saturday afternoon, not much traffic -- none at all once I got into the 30MPH zone in town, and no pedestrians or cross-traffic. Certainly no overhead signs on this local road in this small town.

The only thing I can think of is that the cluster of road signs you can see on the right was mis-detected as a crossing or parked vehicle by the radar -- but I was centered in my lane and definitely not headed for the side of the road, so I wouldn't expect AP to slow so dramatically (nor in two distinct stages) for a parked car.

In fact, conditions were probably just about perfect for unambiguous, camera-based detection of the traffic light state. What I'm wondering is, are we 100% sure this feature doesn't exist at all or is completely disabled, or is there some possibility that under ideal conditions, signals can be detected with enough confidence that the car will actually brake for them, and that this has quietly been turned on in some recent update?

Software is 17.6.15.

 

[croman]

It is highly unlikely that AP1 will ever have this feature. For starters -- EAP costs 2x what AP cost and it won't have it. They will only include that feature in FSD.

Second, while your anecdotal evidence matches some other experiences, you can't ever rely on it because it is so sporadic, so what's the point? If it randomly stops at lights, its not really a great thing because its so unpredictable and we really need predictability in a driver assistance feature. So I'm not sure what to think.

Finally, the EyeQ3 chipset just can't do this. MobilEye both talks a big game but fails to back it up with action. They never really matched Tesla's ambitions with the hardware and their second gen stuff is weak as compared to Tesla's. I think Mobileye is too conservative to commercially realize some of the more ambitious stuff they've marketed.

 

[Sir Guacamolaf]

Nobody in front of you? Just making sure!

 

[JohnnyG]

Do you travel this route often? Can (will) you travel this route more often now? I agree with @croman for the most part. However, if you feel pretty confident in this, I would definitely suggest trying to reproduce it; especially with a dash cam. I would love for this not to be a fluke, but there's only one way to tell.....

 

[croman]

Do you travel this route often? Can (will) you travel this route more often now? I agree with @croman for the most part. However, if you feel pretty confident in this, I would definitely suggest trying to reproduce it; especially with a dash cam. I would love for this not to be a fluke, but there's only one way to tell.....

If it can be reproduced then it changes a lot about how useful it is. I've never heard of this happening other than at random instances where other factors could have been at play (for example, my car does detect cross traffic but it doesn't reliably slow because the cars move past the radar to quickly).

 

[BerTX]

I was wondering why I haven't experienced this. Then it hit me -- I'd never try to use AP in those conditions. Guess I'm a wimp.

 

[JohnnyG]

If it can be reproduced then it changes a lot about how useful it is. I've never heard of this happening other than at random instances where other factors could have been at play (for example, my car does detect cross traffic but it doesn't reliably slow because the cars move past the radar to quickly).

I'm totally with you on this.... I've experienced the same cross-traffic scenarios as you too. And I'll admit that I currently have my rose-colored optimist glasses on. But I haven't seen a report as confident and detail specific as this. So I figure that it couldn't hurt to try to reproduce. At worst it proves another anomaly. At best it opens a floodgate of questions. Honestly, I'm expecting anomaly, but I love a good experiment!

 

[tls]

Nobody in front of me. I'll be away for several weeks from now, unfortunately, but I can try to reproduce it under similar lighting and road conditions after that.

If what I observed was in fact not caused by the traffic signal, my money is on that cluster of 3 road signs, off to the right on the sidewalk, that you can see in the picture. If AP were programmed to stop for Bender lest he stagger into the road, I think we could easily see where there would be a radar signature that were a close match. As a "car" or a "pedestrian" though it seems a little questionable to me.

I guess the other thing to keep in mind is that traffic light placement and timing are dictated by visibility at or slightly above the speed limit at a given location; so if the AP was reacting to something very near the intersection but *not* the traffic signal, then it makes sense that it'd slow just after the signal came into view (when the light turned yellow) and brake harder as I approached the point-of-no-return for blowing through a red light (since that's how the yellow timing is usually set). But if it wasn't that stack of road signs, dunno what exactly right-at-the-intersection it could have been reacting to.

 

[chillaban]

Hmm, I use Autopilot on a traffic light expressway every day, and I've honestly never seen it slow down while approaching yellow or red lights. Maybe it does happen, maybe it's incredible coincidence.

 

[phaduman]

....
Case in point: about a week ago I was driving, AP engaged, on a 2-lane road with a speed limit that dropped successively from 55 to 45 to 30. As you'd expect the car slowed from 60 to 50 to 35. In the course of this it also slowed for one or two tight curves. No surprise there.

I haven't even experienced the red text experience you normally have, above. Once my car as AP enabled (AP1 here), say even in freeway, if the speed markings drop down e.g. from 65 to 55, my car shows the slower speed on the dashboard, but doesn't slow down on its own. I have to drop the speed manually. Not sure I am missing any configuration somewhere? I have set up +5 in the setting.

 

[Max*]

If Tesla thought this feature might come true, Elon would be on Twitter telling us how well he treats his customers and how AP1 will get an update "soon" with red light detection months before he can get it working 1 in a 100 times.

 

[Max*]

I haven't even experienced the red text experience you normally have, above. Once my car as AP enabled (AP1 here), say even in freeway, if the speed markings drop down e.g. from 65 to 55, my car shows the slower speed on the dashboard, but doesn't slow down on its own. I have to drop the speed manually. Not sure I am missing any configuration somewhere? I have set up +5 in the setting.

This "automatic" lowering is only on undivided roads, where AP1 can be set to +5mph. It doesn't take place on divided freeways.

 

[Max*]

And one more thing, if you mentioned just the red light detection, I'd say technically possible, though unlikely.

To my knowledge, Mobileye uses a grayscale channel and a red channel. Sure, yellow has some red in it, but I'm not sure how reliably it'll be able to differentiate a yellow light being on vs. off.

 

[BerTX]

And one more thing, if you mentioned just the red light detection, I'd say technically possible, though unlikely.

To my knowledge, Mobileye uses a grayscale channel and a red channel. Sure, yellow has some red in it, but I'm not sure how reliably it'll be able to differentiate a yellow light being on vs. off.

I had this same thought. Then I wondered if it could go by positioning on the yellow light -- is it always in the middle? Or maybe the grayscale can pick up yellow, but not reliably tell the difference between red and green...

 

[green1]

If Tesla thought this feature might come true, Elon would be on Twitter telling us how well he treats his customers and how AP1 will get an update "soon" with red light detection months before he can get it working 1 in a 100 times.

It *WAS* part of the original promises, so Elon *did* brag about it already.

This is also a stated feature of the Mobileye chip which they have demonstrated working. it's just the Tesla implementation that doesn't.

that said, I can say with no hesitation, that AP1 does not now, nor is it ever likely to in the future, have this feature. In fact they started to play with it a bit in 7.1, but later gave up (the variable to control it was introduced in either 7.0 or 7.1 firmwares, but didn't do anything, then later the variable was removed altogether and the feature was quietly dropped)

 

[Max*]

I had this same thought. Then I wondered if it could go by positioning on the yellow light -- is it always in the middle? Or maybe the grayscale can pick up yellow, but not reliably tell the difference between red and green...

Grayscale picks up intensity. So it could pick up the intensity of a traffic light being on, assuming it knows that red is always on top, yellow is always in the middle, and green is always on the bottom. But I have my doubts they'd just use this without a color detection to confirm. Though I could be wrong.

 

[Max*]

It *WAS* part of the original promises, so Elon *did* brag about it already.

I realize that, which is why I said what I said.

 

[JohnnyG]

The photocells in a camera sensors only pickup light intensity. The sensor itself can't distinguish color, only intensity. The way a camera sees in color is by placing filters over each photocell, and then measuring the intensity of that specific wavelength at that specific photocell. I highly doubt the system would be looking specifically for the color red (although it could use a red filter as a secondary validation), It would most likely be looking for the higher intensity of light coming from one of the three circles/squares in the traffic light. In the USA, traffic lights are always Red, Yellow, Green; either top to bottom, or left to right. Now, to complicate things, There are traffic lights with more than three circles/squares, and I don't really think there is a specific standard on that. I've seen different configurations of traffic lights with more than three lights. But, I do believe that the first two are always Red - Yellow, and then after that is where I've seen it differ with green arrows, or even a fourth light which is another red light....

 

[green1]

The photocells in a camera sensors only pickup light intensity. The sensor itself can't distinguish color, only intensity. The way a camera sees in color is by placing filters over each photocell, and then measuring the intensity of that specific wavelength at that specific photocell. I highly doubt the system would be looking specifically for the color red (although it could use a red filter as a secondary validation), It would most likely be looking for the higher intensity of light coming from one of the three circles/squares in the traffic light. In the USA, traffic lights are always Red, Yellow, Green; either top to bottom, or left to right. Now, to complicate things, There are traffic lights with more than three circles/squares, and I don't really think there is a specific standard on that. I've seen different configurations of traffic lights with more than three lights. But, I do believe that the first two are always Red - Yellow, and then after that is where I've seen it differ with green arrows, or even a fourth light which is another red light....

I don't know how mobileye does it for certain, only that I know that they do. But if I had to bet, I'd guess that position is how mobileye does it.

 

[kort677]

AFAIK
the only way AP1 would stop for a red light is if there was a car already stopped for the light in your lane. the AP1 sees the car ahead of you, not the traffic signal.