FSD runs red light when have sun glare

[Vizca]

FSD (v11.7.2) runs red light when having sun glare to the traffic lights as shown in photo below. Basically, FSD sees the red as yellow.

The incident occurred on October-19-2024, 5:59PM CST, in Round Rock, Texas. My MY travels 45mph and ran on red lights which I can see it clearly on my eye-balls; I applied a had braking for avoiding a fetal collision. I replayed and reviewed the records multiple times, sent an email to “[email protected]” (with asking what I should do on reporting to his incident) and saw no engagement from Tesla.

My analysis: The FSD (either on HW cameras? or software?) sees red lights as yellow lights when having sun glare. If this problem is not properly addressed, there will be some fetal accidents from FSD. FSD users should/must be aware and prepared when sun light goes directly to the traffic lights (for preserving life); since Tesla FSD has not yet publicly notified this deadly critical issue to FSD Beta users (we just sacrificed our own life by paying for testing this new pioneer product?).

 

[Vizca]

Correction to typo “October-19-2024” to 2023.

 

[KArnold]

In the photo the lights look very yellow. Is this the only light you've observed this on?

 

[Vizca]

The above incident is my 2nd time in this year. The first incident (of FSD running on red light) occurred to me on Jan-12-2023, 11:42AM CST (at Palmer lane and Avery Ranch Blvd junction, Austin, TX 78717). My MYP traveled within the road speed limit of 55mph. The sunlight was shining into the traffic lights; and I can see red on my eye-balls (but the FSD probably sees as yellow?). It ran on red light; and I applied a hard braking and almost had a colliding accident as in this photo (that I snapped from replaying the record).

You can see how close I was to the running car in front; and the low-right cam showed my car passed the crosswalk.

My observation: The FDS cannot see the red light clearly when sunlight shines (directly or almost straight angle) into the traffic lights; and this is a dangerous flaw in FSD. Tesla FSD team should duplicate and accordingly solve this problem.

 

[Vizca]

In the photo the lights look very yellow. Is this the only light you've observed this on?

Texas traffic light orders (right to left): Green, Yellow, Red. You can see the most left is ON and it must be red … FSDb likely sees them as yellow and not stop.

 

[Vizca]

The above incident is my 2nd time in this year. The first incident (of FSD running on red light) occurred to me on Jan-12-2023, 11:42AM CST (at Palmer lane and Avery Ranch Blvd junction, Austin, TX 78717). My MYP traveled within the road speed limit of 55mph. The sunlight was shining into the traffic lights; and I can see red on my eye-balls (but the FSD probably sees as yellow?). It ran on red light; and I applied a hard braking and almost had a colliding accident as in this photo (that I snapped from replaying the record).
View attachment 994005
You can see how close I was to the running car in front; and the low-right cam showed my car passed the crosswalk.

My observation: The FDS cannot see the red light clearly when sunlight shines (directly or almost straight angle) into the traffic lights; and this is a dangerous flaw in FSD. Tesla FSD team should duplicate and accordingly solve this problem.

The FSDb version in the first incident is: 10.69.25.1 (2022.44.30.5)

 

[Supcom]

FSD (v11.7.2) runs red light when having sun glare to the traffic lights as shown in photo below. Basically, FSD sees the red as yellow.
View attachment 993892

The incident occurred on October-19-2024, 5:59PM CST, in Round Rock, Texas. My MY travels 45mph and ran on red lights which I can see it clearly on my eye-balls; I applied a had braking for avoiding a fetal collision. I replayed and reviewed the records multiple times, sent an email to “[email protected]” (with asking what I should do on reporting to his incident) and saw no engagement from Tesla.

My analysis: The FSD (either on HW cameras? or software?) sees red lights as yellow lights when having sun glare. If this problem is not properly addressed, there will be some fetal accidents from FSD. FSD users should/must be aware and prepared when sun light goes directly to the traffic lights (for preserving life); since Tesla FSD has not yet publicly notified this deadly critical issue to FSD Beta users (we just sacrificed our own life by paying for testing this new pioneer product?).

I've driven thousands of mile in Texas on FSDb since version 10.4 and never had the car attempt to run a red light or misidentify a red light as yellow. Not to say it doesn't happen. I've seen other posts showing the car misidentifying a green for cross traffic, but your example clearly is not one of those cases.

Interesting that the original post shows the sun low to your rear but the second has the sun at a high angle.

Are these issues repeatable at these intersections as opposed to a rare event?

 

[E90alex]

You should have noticed the car not slowing down for the light way earlier and applied the brakes yourself. FSD is nowhere near “full self driving” yet and requires the driver to fully pay attention. Extra attention even. Clearly you were not paying attention.

 

[Vizca]

You should have noticed the car not slowing down for the light way earlier and applied the brakes yourself. FSD is nowhere near “full self driving” yet and requires the driver to fully pay attention. Extra attention even. Clearly you were not paying attention.

First of all, a quick jump to a conclusion “Clearly you were not payed attention” is incorrect when I have evidence showing FSD runs on red light (when sunlight is in direct angle to traffic lights).

Let’s digest these scenarios:
(A) I gently applied brake for stopping car early: There is no proof of FSDb runs on red light.
(B) I not apply brake and let FSD runs on red light: I might lose my life and I would not do that.
(C) Let FSDb runs on red lights and abruptly stop it for avoiding collisions: The driver “is fully paying attention” and can proof that FSDb is having a fetal flaw (via reviewing the car cam records).

Now, let’s digest on what/where is the flaw? Can it be in one of these:
(A) Hardware: CPU? Cameras? My MYP is quite early production and delivered to me in Jul-2020. It uses ‘AMD Ryzen Embedded V1000C processor’ and the next year production, 2021 till now, use ‘AMD Ryzen Embedded Ryzen APU’. The cameras are also upgraded in 2022.
(B) Software: FSD? It has been improved rapidly in each release with some residual and new bad behaviors.
(C) The combo of HW and software (?): It’s very hard for sorting out without internal test/diagnose/debug from Tesla engineering.

I just established a data point on this issue with hope other Tesla owner who encountered same issue will add on more data points.

 

[Vizca]

I've driven thousands of mile in Texas on FSDb since version 10.4 and never had the car attempt to run a red light or misidentify a red light as yellow. Not to say it doesn't happen. I've seen other posts showing the car misidentifying a green for cross traffic, but your example clearly is not one of those cases.

Interesting that the original post shows the sun low to your rear but the second has the sun at a high angle.

Are these issues repeatable at these intersections as opposed to a rare event?

The issue is likely reproducible (since it happened to me twice already with very similar conditions). I decided to create a post with hope other owners with similar incidents will add more data points. (Info: I have used FSD daily for ~20 miles driving since Dec-24-2021.)

 

[Supcom]

The issue is likely reproducible (since it happened to me twice already with very similar conditions). I decided to create a post with hope other owners with similar incidents will add more data points. (Info: I have used FSD daily for ~20 miles driving since Dec-24-2021.)

The two photos you shared show very different relative sun positions. While the sun is shining onto the lens of the lights in both cases, the same would be true for about half of all traffic lights any time the sun is shining. If this were the cause, I would expect that we would all be seeing this problem daily. Hence, I suspect the issue might have some other root cause.

I've been using FSDb for exactly as long as you and have driven through countless intersections with the sun at all positions, yet have not had this issue. I'm not implying that your car didn't run the light. But I'm thinking that that it's not a simple sun angle issue. If it were, I would expect there to be many, many posts describing the same thing. Hence, my question on whether it is repeatable. By that, I meant, did it happen at the next light as well, or maybe you circled around to try the same intersection again?

 

[Vizca]

The two photos you shared show very different relative sun positions. While the sun is shining onto the lens of the lights in both cases, the same would be true for about half of all traffic lights any time the sun is shining. If this were the cause, I would expect that we would all be seeing this problem daily. Hence, I suspect the issue might have some other root cause.

I've been using FSDb for exactly as long as you and have driven through countless intersections with the sun at all positions, yet have not had this issue. I'm not implying that your car didn't run the light. But I'm thinking that that it's not a simple sun angle issue. If it were, I would expect there to be many, many posts describing the same thing. Hence, my question on whether it is repeatable. By that, I meant, did it happen at the next light as well, or maybe you circled around to try the same intersection again?

I did not circle around for retest the issue and went straight to home for reporting issue to Tesla. At home I replayed the recording multiple times, snapped some photos, emailed to ‘[email protected]’, got no engagement/explanation from Tesla. I believe retesting the issue should be in Tesla boundary and should not in the FSD user’s side.

 

[Vizca]

Issue: Tesla FSDb sees the OFF railroad light as red light when sun shines straight to the light.

In some mornings when having strong sunshiny, my FSD reacts oddly to the OFF railroad light with slowing down and attempting to stop (and I have to press acceleration for unjamming traffic). In this photo, FSD sees the OFF railway light as active red. We can tell the light state from eyeballs (it is slightly red from sunlight reflecting, but not bright/shiny red from inside out) and the barrier is in straight up position). The FSD creates the red line on screen indicating that it is seeing the red railway light.

Does any FSD users see this issue?

 

[Dewg]

Hmmmm. That doesn't look like FSD Beta. Even while not in use, Beta shows the full visualisations around it, and shows a grey tether projected in front of it. This looks like the old AP code, which does stop, even for green lights without confirmation from the driver.

 

[eli_]

Hmmmm. That doesn't look like FSD Beta. Even while not in use, Beta shows the full visualisations around it, and shows a grey tether projected in front of it. This looks like the old AP code, which does stop, even for green lights without confirmation from the driver.

The driver is in TACC mode, not even Autosteer is enabled. So yeah, even if they have FSD beta, it's being forced to use the old stack.

 

[Supcom]

The driver is in TACC mode, not even Autosteer is enabled. So yeah, even if they have FSD beta, it's being forced to use the old stack.

It has been over a year, but I have had this issue occur using FSD beta. It would not surprise me to see recent versions still suffer from the problem if the sun is in the right position. Given that FSD beta does not seem to look for railroad crossing lights specifically, the sun shining off the lens could easily be perceived as a solid red traffic light.

 

[Vizca]

Today( Sunday, Dec-10-2023, 8:30am) I have chance for testing the FSD misjudged the red light from the direct sunlight into the railroad-crossing light. The traffic is clear and I wanted to see the complete behavior of FSD on this scenario: First, it slowed down rapidly from 40mph (Speed Limit of this highway segment) to 0mph or full stop in front of the white line (Same behavior on Stop sign). Then, it slowly moved and gradually gained speed back to 40mph.

The main point here proved that the FSD misjudged on traffic light under sunlight influence. Tesla FSD engineering must test this scenario and improve FSD accordingly (else, fetal collision will happen).

 

[Dewg]

The only way cars can detect signal lights is camera (visual). Radar and LIDAR can't see them. I'm not sure how it will be able to filter out the red light reflecting off the lens of the signal.

 

[Supcom]

The only way cars can detect signal lights is camera (visual). Radar and LIDAR can't see them. I'm not sure how it will be able to filter out the red light reflecting off the lens of the signal.

If V11 were to continue, FSDb would need to be trained to recognize a railroad crossing signal separately from the basic traffic light signals. Hopefully, with V12 training, the system will have been fed many examples of railroad crossings, so will not stop unless the lights are actually flashing and the gates are down.

 

[Supcom]

View attachment 998141
Today( Sunday, Dec-10-2023, 8:30am) I have chance for testing the FSD misjudged the red light from the direct sunlight into the railroad-crossing light. The traffic is clear and I wanted to see the complete behavior of FSD on this scenario: First, it slowed down rapidly from 40mph (Speed Limit of this highway segment) to 0mph or full stop in front of the white line (Same behavior on Stop sign). Then, it slowly moved and gradually gained speed back to 40mph.

The main point here proved that the FSD misjudged on traffic light under sunlight influence. Tesla FSD engineering must test this scenario and improve FSD accordingly (else, fetal collision will happen).

I doubt that a fetal (or fatal) collision will occur in this scenario unless you fail to wake up and push on the accelerator.