Another tragic fatality with a semi in Florida. This time a Model 3

[MXWing]

AP is like guns, Human invention that has widely known good uses. But it occasionally kills people unintentionally.

I agree with you that AP is like guns. Though probably not in the manner you were expecting agreement.

The overwhelming majority of gun deaths is the result of an intentional act.

There's no child shooting his sibling unintentionally. A firearm was made available to him intentionally through a willful process of not securing it.

AP deaths are intentional failures from the driver because they failed to retain sufficient enough control of the vehicle to over ride it.

No one drives 1 second behind an identical car even though in theory you can decelerate as fast as they are decelerating.

That's why you need 2 seconds or more as lag and space cushion.

AP needs space cushion as well to safely operate.

 

[dgatwood]

I would be be curious to see an overpass that would cause a false positive with LIDAR or a semi trailer that would be invisible to it.

Overpasses don't cause false positives with RADAR, and semi trailers aren't invisible to it. What causes false positives is the computer making incorrect judgments about whether those things are actually blocking your lane or not. It's the software that is the problem, not the data source; providing data from a different source to similarly broken software is still going to produce similarly incorrect results.

 

[acoste]

Overpasses don't cause false positives with RADAR, and semi trailers aren't invisible to it. What causes false positives is the computer making incorrect judgments about whether those things are actually blocking your lane or not. It's the software that is the problem, not the data source; providing data from a different source to similarly broken software is still going to produce similarly incorrect results.

They actually do. Radars have somewhat ok horizontal resolution, but not very good vertical resolution. High resolution radar is needed here.

Here for example Continental proudly mentions that their new generation radars that entered in production in 2019 have good enough vertical resolution to solve this issue:
Fifth Radar Generation Meets Future Requirements for Automated Driving

"In addition, road limits such as curbstones as well as the height of objects like the tails of a traffic jam under a bridge are detected thanks to the sensors"

 

[Leafdriver333]

They actually do. Radars have somewhat ok horizontal resolution, but not very good vertical resolution. High resolution radar is needed here.

Here for example Continental proudly mentions that their new generation radars that entered in production in 2019 have good enough vertical resolution to solve this issue:
Fifth Radar Generation Meets Future Requirements for Automated Driving

"In addition, road limits such as curbstones as well as the height of objects like the tails of a traffic jam under a bridge are detected thanks to the sensors"

It sounds like they just gave a 2 year old boy a sandal to cross a river.
By Dec, the boy will get the other sandal.
When he turns 3, he will get a converse, and the other converse by 4 y.o.
At 6, he will finally get a rain boot.
When he turns 20, he will get a Speedo instead of a diaper.
At 55, he will get an oar.

etc...

 

[mongo]

They actually do. Radars have somewhat ok vertical resolution, but not very good horizontal resolution. High resolution radar is needed here.

Resolution is propotional to the antenna size, and for phased arrays, the number of elements controls steerability. If the antenna is wider than it is tall, the left/right resolution and steering will be better than up/down. This is the situation for the Tesla radars.
Undocumented | TeslaTap

 

[dgatwood]

They actually do. Radars have somewhat ok vertical resolution, but not very good horizontal resolution. High resolution radar is needed here.

And yet it would presumably be a the vertical resolution limitation that causes cars to drive under tractor trailers. So obviously that isn't nearly good enough, either. And yet, the RADAR hardware claims to be able to have an angular measurement accuracy of 0.3 degrees. Clearly there's more than a .3 degree-wide gap between a tractor trailer and the ground, even at a large distance. So how is that not detected?

I don't think for one minute that resolution could possibly be the issue. I think they're just being too aggressive at ignoring as ground clutter any large, flat objects that are moving at exactly the same speed as the ground, under the assumption that they cannot possibly be in the lane. And most of the time, they're right. Yet if they used that same software approach with LIDAR, they'd still get collisions, because in spite of having slightly more precise measurements, it would still look like you're driving towards a large, flat surface.

Mind you, having higher resolution *might* make the edges more obvious, and therefore *might* make neural nets more likely to detect them, but that assumes that the software that decides what to do based on the detection event doesn't just say, "That's a wall; it can't be in front of us," and ignore it.

 

[mongo]

And yet it would presumably be a the vertical resolution limitation that causes cars to drive under tractor trailers. So obviously that isn't nearly good enough, either. And yet, the RADAR hardware claims to be able to have an angular measurement accuracy of 0.3 degrees. Clearly there's more than a .3 degree-wide gap between a tractor trailer and the ground, even at a large distance. So how is that not detected?

I don't think for one minute that resolution could possibly be the issue. I think they're just being too aggressive at ignoring as ground clutter any large, flat objects that are moving at exactly the same speed as the ground, under the assumption that they cannot possibly be in the lane. And most of the time, they're right. Yet if they used that same software approach with LIDAR, they'd still get collisions, because in spite of having slightly more precise measurements, it would still look like you're driving towards a large, flat surface.

Mind you, having higher resolution *might* make the edges more obvious, and therefore *might* make neural nets more likely to detect them, but that assumes that the software that decides what to do based on the detection event doesn't just say, "That's a wall; it can't be in front of us," and ignore it.

A flat vertical surface (side of a trailer) located above the radar is not going to have much of a return. Similar to not being able to see your reflection in a mirror that is above your eyes.

Vertical resolution would assist with identifying the exact height of an object, however, given that a car cannot control its height (and objects do not typically lack a base), as long as the beam covers the vehicle height, that is sufficient. Horizontal is more critcal for object detection, path planning and steering.

 

[verygreen]

and therefore *might* make neural nets more likely to detect them

they don't use NNs with radar inputs.

 

[acoste]

And yet it would presumably be a the vertical resolution limitation that causes cars to drive under tractor trailers. So obviously that isn't nearly good enough, either. And yet, the RADAR hardware claims to be able to have an angular measurement accuracy of 0.3 degrees. Clearly there's more than a .3 degree-wide gap between a tractor trailer and the ground, even at a large distance. So how is that not detected?

I don't think for one minute that resolution could possibly be the issue. I think they're just being too aggressive at ignoring as ground clutter any large, flat objects that are moving at exactly the same speed as the ground, under the assumption that they cannot possibly be in the lane. And most of the time, they're right. Yet if they used that same software approach with LIDAR, they'd still get collisions, because in spite of having slightly more precise measurements, it would still look like you're driving towards a large, flat surface.

Mind you, having higher resolution *might* make the edges more obvious, and therefore *might* make neural nets more likely to detect them, but that assumes that the software that decides what to do based on the detection event doesn't just say, "That's a wall; it can't be in front of us," and ignore it.

If you click on mongo's link above, it says Tesla uses this guy here: Continental Automotive
"Angular Accuracy: Horizontal: 0,2°; Vertical: 1,5

But I have my doubts about that 1.5 deg. Read about side lobes. The signal needs massive filtering. This is not a simple measurement.

Continental doesn't even mention vertical resolution in the spec: https://www.continental-automotive....d63/ARS408-21_datasheet_en_170707_V07.pdf.pdf


Anyway let's calculate with 1.5 deg.

When driving at 65mph the car drives 29m per second. It takes 50m to come to a full stop in good conditions.
1 second before reaching the object the resolution is 76cm
2 seconds before reaching the object the resolution is 1.5m (5 feet)

 

[mongo]

If you click on mongo's link above, it says Tesla uses this guy here: Continental Automotive
"Angular Accuracy: Horizontal: 0,2°; Vertical: 1,5

But I have my doubts about that 1.5 deg. Read about side lobes. The signal needs massive filtering. This is not a simple measurement.

Continental doesn't even mention vertical resolution in the spec: https://www.continental-automotive....d63/ARS408-21_datasheet_en_170707_V07.pdf.pdf


Anyway let's calculate with 1.5 deg.

When driving at 65mph the car drives 29m per second. It takes 50m to come to a full stop in good conditions.
1 second before reaching the object the resolution is 76cm
2 seconds before reaching the object the resolution is 1.5m (5 feet)

FWIW, the side lobes are not really an issue, a downward facing lobe would return too quickly to be valid, an upward lobe would not hit anything.

 

[dgatwood]

A flat vertical surface (side of a trailer) located above the radar is not going to have much of a return. Similar to not being able to see your reflection in a mirror that is above your eyes.

Then it also won't have much of a return from LIDAR. Either way you're bouncing something off a flat vertical surface.

Besides, it should detect it long before it gets close enough to be meaningfully above the car. The RADAR has a maximum range (assuming I'm looking at the right part) of 160 meters (524 feet). At that distance, assuming the bottom of a truck is 58 inches off the ground, the bottom of that flat surface is only .528 degrees above horizontal.

And by the time it exceeds the top of the Tesla's field of view RADAR-wise (25 degrees above horizontal), the truck is 124 inches away, so if the car hasn't detected it by then, you're already screwed.

The real problem is that Tesla's software doesn't pay attention far enough out in front of the car. That's why it suddenly panic brakes for things that it could easily have anticipated, why it doesn't start turns soon enough to not jerk you around, etc. The unintentional convertible-making is just another symptom of that same software problem.

Edit: I see that we're looking at two different spec sheets. Apparently, mine was the original HW2 RADAR. The HW2.5/HW3 RADAR is even better.

 

[OPRCE]

Here for example Continental proudly mentions that their new generation radars that entered in production in 2019 have good enough vertical resolution to solve this issue:
Fifth Radar Generation Meets Future Requirements for Automated Driving

Looks like Conti have released their 5th Gen radar within the past few weeks, with >2x range and presumably better vertical resolution:
Continental Automotive

ARS540 is a high performance premium long range radar sensor which enables highly automated driving in combination with other technologies. It provides best radar performance in a state-of-the-art sensor size.

Benefits & Features

• Highly Automated Driving
• Direct and independent measurement of four dimensions (range, doppler, azimuth, elevation)
• 300 meter range
• Real height measurement
• Multi-hypothesis tracking for better prediction of high complexity scenarios
• Classification of traffic participants & infrastructure
• Radar-only VRU detection by means of micro-doppler technology
• Auto alignment
• Highest availability of all ADAS technologies
• Cyber Security

Technical Information

• Dimensions: 130 x 101 x 32 mm (w/o con.)
• Update rate: 60 to 70 ms
• Temp. Range: -40° to +85°
• Power Dissip.: ~16 W
• Operation Freq.: 76…77 GHz

 

[mongo]

Then it also won't have much of a return from LIDAR. Either way you're bouncing something off a flat vertical surface.

Besides, it should detect it long before it gets close enough to be meaningfully above the car. The RADAR has a maximum range (assuming I'm looking at the right part) of 160 meters (524 feet). At that distance, assuming the bottom of a truck is 58 inches off the ground, the bottom of that flat surface is only .528 degrees above horizontal.

And by the time it exceeds the top of the Tesla's field of view RADAR-wise (25 degrees above horizontal), the truck is 124 inches away, so if the car hasn't detected it by then, you're already screwed.

The real problem is that Tesla's software doesn't pay attention far enough out in front of the car. That's why it suddenly panic brakes for things that it could easily have anticipated, why it doesn't start turns soon enough to not jerk you around, etc. The unintentional convertible-making is just another symptom of that same software problem.

Edit: I see that we're looking at two different spec sheets. Apparently, mine was the original HW2 RADAR. The HW2.5/HW3 RADAR is even better.

Yes and no

Yes, lidar would have the same issue. No being further away does not help. Radar/ lidar needs to return to bounce back to it.
If the flat surface starts above the sensor, it will not reflect back to the radar regardless of the distance from it.
Back to the mirror analogy, find a vertical mirror, note where your eyes are when standing close, now back up. If the floor is flat and the mirror vertical, your eyes will stay in the same spot and you will never see your eyes in the part of the mirror above the original spot.

Similarly, a bumper mounted radar will not get a return off the side of a trailer unless the trailer is tilted such that the side is perpendicular to the beam. It might get a return from the undercarriage if that is exposed. A metal side skirt/ air deflector could make things worse in that case.

 

[Cloxxki]

Umm, no. The low standard is that of human drivers. Sorry dude, but we as a whole are pretty terrible.

Nothing is ever going to be perfect. That’s reality. Any automaker waiting for a perfected system will never release the system. That is pretty much why I give a Tesla a pass on this stuff. I get to enjoy cool features while Tesla improves the system. If you have a problem with the situation, then don’t use it.

And stop being so sensitive about the dislikes. Another fact of life, there will always be people who disagree with whatever position you take.

I was addressed with the question whether I am here to get as many dislikes as possible. I am now telling you what it means to disagree with common sense and how it portrays the Tesla community. To investors, drivers, other traffic participants. Bring a Tesla apologist on here doesn't even help Tesla...

 

[derotam]

Sorry this is so long and done from my phone.

After sleeping on this entire subject, here is where I land on this. All the technical jargon on this aside.

For about a year after getting my car and a few dead stopped fire engine accidents on highways is when this Amateur Tesla owner figured out this car will not be stopping, attempt to stop or maybe even not maneuvering around stopped objects in highways at higher speeds then 35mph or some threshold I am still trying to find that answer. I had gone a year not knowing this. I thought my car did everything. All the sheepish expressions when asking Tesla employee a question are all coming into view now in my memory.

Two years prior to buying and years after up until yesterday before I knew Jeremy was killed, I had written essays and recommended people to either buy a Tesla or seriously consider it even as their next car. To date I have zero referrals. There are 5 Teslas in my neighborhood and I know probably 3-4 others around my circle that own Tesla’s. Did I influence any of that? Either way I now almost feel it’s my duty as another human being to inform them completely that these cars are not going to save your life. In fact they might give you a false sense that they will save your life.

I am certainly from this point forward shutting my mouth to help this cause for fear I will have blood on my hands. What makes me think this way. If I had been able to sit down and just in a few minutes let Jeremy know some facts that he would have found enlightening, I’m sure being a software engineer he would have taken that info and at least investigated for himself. Would it have changed the outcome we have here, nobody knows.

I know for sure there are many aspects of at least of a Model S that cause undue loss of life. How about the BLUE Model S fire in Ft Lauderdale. People reported trying to open the car door. The handles were not or did not pop out upon impact. They sat and watched that person burn to death. Had they even broken the window (one Tesla employee suggested)and with all the heat and the rush: Do you think they would have found the special release handle located on the door we all know about up by the stationary wing window?
The one we all grab daily if your an owner of a model S. Think about it. It’s not easily understood.

My first long trip in my car, I showed the other “Emergency Release” to my daughter since she rides in the backseat. I wanted her to know in the event of a crash and the electronics (the back doors depend 100% on electricity) shut down how to get out in the event so as to perhaps save her life. She could barely do it after struggling to find it under her seat. So that’s it. Special releases for people to exit a badly damage or burning vehicle. In the model S that’s just the rear passengers have to know that. Now if any Model S owner just read this and you never knew this, well you need to get your manual out on your computer and read it from to back and do it again in another week.

So my point is this, this car, this entanglement of tech and not so tech, before a laymen, laywoman, person, child, uses it, rides in it, drives it, you would have to hold classes on the entire aspect of the car, then do it again, before you could safely say I informed, I educated, I have done my part to make sure everyone understands what to do in an Event, ie stopped fire engines, 90 degree facing semis, accidents, fire, on and on. In fact the Model S is the only car I have ever owned where you would have to do that. Dare I say there are other Tesla’s where it’s a requirement. I just have not read the manual and done the in-depth study on those vehicles like I have here. Still learning.

Now Jeremy was competitive. He was also a jealous personality. He was human. How do I know this. The camera quote in my previous posts. He like so many humans fed his ego on facts related to your behind on your tech. He was a software engineer after all. I cannot help but think he kept up with us through Facebook and those posts of my red Tesla Model S, well he was going to do one better with a Red Model 3. You see we are the marketing team for Tesla. The whole universe is looking at our cars. Last night I must have had at least 6-10 people I noticed at lights looking at my car. Those are the ones I noticed. Probably well over 1000 in the whole trip. These are special cars, they take special understanding. Even then I am not convinced they are that special any longer. I’m sorry for myself, others and the planet, because the dream of an electric car has been mine from a very early age. Unfortunately Tesla has taken it too far, too fast and made it so the human beings are not able to understand fully what they have gotten into here. Elon Musk wants to win, the cost for that is steep, the cause I get, the speed at which he is trying to do it in, that’s business.

I in no way blame myself for anything, I do though feel it important to educate. Perhaps I have found my next calling forward. Thanks for reading and safe Travels in whatever your driving. Have a great weekend.

I am only writing this because of your "shutting my mouth" comment as well as seemingly(to me) being conflicted in some of your feelings.

Look, I can understand where all of those thoughts are coming from but... then what we should just all live in a cave? Saying that you are going to keep your mouth shut because you feel bad hyping up or recommending a Tesla is a slippery slope in my opinion. Are you going to tell people to stay away from Toyota's because of unintended acceleration issues, don't buy a house with a big tree anywhere nearby...etc.

If you want to stress that people should maybe point out some limitations with the system, stress that people still need to pay 100% attention because the system may do something, or not do something unexpectedly, then great, I'm all for that...but saying that you are going to keep your mouth shut is almost just as bad.

I do not intend to talk bad about the deceased, but you are talking as if Jeremy didn't do anything wrong and that it is Tesla's fault. None of us know exactly what happened, but a lot of us know very well how someone may turn on AP specifically so they can fish something off the floor or wherever else that got dropped. There has to still be some personal responsibility and that is actually dictated and demanded with the system right now. Is there seemingly an issue with the scenario that occurred, yes. Same with the firetruck incidents(other issues with this though) Just because it will "let you" do something doesn't mean you should. I could have a car that will drive perfectly straight on a perfectly straight road...should I be comfortable taking my hands off the wheel, no.

I didn't want to click the disagree flag because your heart is in the right place, but I do think some of your ideas are a little too rigid.

Sorry for your loss.

 

[derotam]

This
Tesla driver gets license suspended after drunkenly falling asleep on Autopilot

And this
Tesla on Autopilot drove 7 miles with sleeping drunk driver, police say
And this
Tesla on Autopilot drove 7 miles with sleeping drunk driver, police say
And plenty of more examples should not be possible but Tesla doesn’t implement better controls relying instead on a blurb in the owners manual stating the driver must be in control.

And yet, that was the safest drunk driver out there. If he had gotten into any other car what would the result have been?

 

[Msjulie]

This is veering off topic a little - but I ofter wonder if there won't be, perhaps for an interim period?, highway lanes for smart cars only.. where the cars behavior is predictable because it's not a person and maybe even the cars talk to one another.. for long range cross country trips, this makes some kind of sense to me.. but I also remember that we really probably won't those kind of limited access roadways as we won't want to spend the resources... it also kinda reminds me of trains somehow..

 

[acoste]

Yes and no

Yes, lidar would have the same issue. No being further away does not help. Radar/ lidar needs to return to bounce back to it.
If the flat surface starts above the sensor, it will not reflect back to the radar regardless of the distance from it.
Back to the mirror analogy, find a vertical mirror, note where your eyes are when standing close, now back up. If the floor is flat and the mirror vertical, your eyes will stay in the same spot and you will never see your eyes in the part of the mirror above the original spot.

Similarly, a bumper mounted radar will not get a return off the side of a trailer unless the trailer is tilted such that the side is perpendicular to the beam. It might get a return from the undercarriage if that is exposed. A metal side skirt/ air deflector could make things worse in that case.

No, lidar doesn't have the same issue unless the truck is covered by a perfect mirror.

The frequency of the lidar and the radar are far away and showing different behaviors. Since lidar is very close to visible light, better think of its capabilities as driving at night with the high beams on. If I can see an object, lidar can see it as well.

Since the wavelength of the radar is around 4mm, any surface roughness larger than that will give reflections. Bolt heads, edges. But the reflected energy is small.

 

[mongo]

No, lidar doesn't have the same issue unless the truck is covered by a perfect mirror.

The frequency of the lidar and the radar are far away and showing different behaviors. Since lidar is very close to visible light, better think of its capabilities as driving at night with the high beams on. If I can see an object, lidar can see it as well.

Since the wavelength of the radar is around 4mm, any surface roughness larger than that will give reflections. Bolt heads, edges. But the reflected energy is small.

Yah, the frequency does help with getting reflections. Headlights are a bit different due to the size and angle variation of the original beam. Lidar is similar to bouncing a laser pointer off an object.

 

[Oldschool496]

This is of great concern, especially after reading about this accident, versus a similar one a few years ago, the AP system has been completely overhauled(different hardware, different software) .... So does it mean 3 years of Tesla R&D still cannot solve the same problem?

AP was not designed to be FSD. Unfortunately many confuse the two, and in this whole case and other similar to it it might be found that those driving treated AP like FSD or certainly used it as a temporary measure to take care of spilled coffee, a bee in the passenger compartment, or get something from the backseat real quick?

Its not designed for evading stopped vehicles in your lane or cross traffic. FSD will take that role perhaps someday when fully tested and approved.