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

[Knightshade]

My intention wasn't to direct attention to reasons why an operator may misuse autopilot, or even fail to read the manual (RTFM), but rather to identify why after three years Tesla still struggles with identifying this use-case

Because this use case is someplace the feature is explicitly not intended to work with AP. Ever.

Not sure how that's still unclear to anyone.


Now- this would be a case where FSD for city streets is expected to work once it's rolled out.

But that feature doesn't exist on customer cars yet.

And will require upgraded HW existing cars (more than a week or two old) don't have yet, and software that isn't published yet.

So we have no idea if that is already working or not. It might well be.

But that's not ever going to make a different feature set do something it's never ever been meant to do

 

[Knightshade]

Can someone provide a cookbook of how AP should be used? Maybe something like:

1. low speed stop-and-go creeping forward, say, 5-10mph
2. high speed cruising on freeways(assuming traffic permits), just like normal cruise control in non-AP/non-Tesla cars
3. no lane change anywhere
4. no turn(left or right) anywhere

So if one follows the above 4 rules, there will be no possibility of fatality, correct?

For autopilot as the feature is currently sold?


Much simpler list currently:

1. Limited access divided highways highways with no oncoming or cross traffic (including pedestrians)...in a single lane (since AP explicitly contains no features supporting other lanes)- and always with the driver paying attention and ultimately prepared to intervene if needed

That's it. (and still won't prevent freak accidents like a semi tire blowing off and slamming into you, or hydroplaning in bad weather, etc...)


EAP or "new" FSD is basically the same thing, minus the "in a single lane" bit since it does do lane changes (even automatically without any confirmation other than hands-on-wheel, when using NoA)

All of this is pretty clear in the owners manual.

Now- FSD will be adding more places later this year- likely requiring HW3 to do it- but covering city street type situations, stop lights, etc....we don't know the nitty gritty specifics yet, and it's not available to anybody right now (outside some testing folks)

 

[Zythryn]

Can someone provide a cookbook of how AP should be used? Maybe something like:

1. low speed stop-and-go creeping forward, say, 5-10mph
2. high speed cruising on freeways(assuming traffic permits), just like normal cruise control in non-AP/non-Tesla cars
3. no lane change anywhere
4. no turn(left or right) anywhere

So if one follows the above 4 rules, there will be no possibility of fatality, correct?

Incorrect. There will never be any level ofdriver assist, offered by any company, that will have 0% chance of fatal accidents.
If that is your expectations, you will be disappointed.

 

[diamond.g]

For autopilot as the feature is currently sold?


Much simpler list currently:

1. Limited access divided highways highways with no oncoming or cross traffic (including pedestrians)...in a single lane (since AP explicitly contains no features supporting other lanes)- and always with the driver paying attention and ultimately prepared to intervene if needed

That's it. (and still won't prevent freak accidents like a semi tire blowing off and slamming into you, or hydroplaning in bad weather, etc...)


EAP or "new" FSD is basically the same thing, minus the "in a single lane" bit since it does do lane changes (even automatically without any confirmation other than hands-on-wheel, when using NoA)

All of this is pretty clear in the owners manual.

Now- FSD will be adding more places later this year- likely requiring HW3 to do it- but covering city street type situations, stop lights, etc....we don't know the nitty gritty specifics yet, and it's not available to anybody right now (outside some testing folks)

I agree with what you are saying with a slight addendum. @verygreen has shown that the cars do see cross traffic. They just aren’t handling them, specifically trailers, properly at all. Which is why Tesla does not recommend usage of AP with cross traffic.

 

[GolanB]

Because this use case is someplace the feature is explicitly not intended to work with AP. Ever.

Not sure how that's still unclear to anyone.


Now- this would be a case where FSD for city streets is expected to work once it's rolled out.

But that feature doesn't exist on customer cars yet.

And will require upgraded HW existing cars (more than a week or two old) don't have yet, and software that isn't published yet.

So we have no idea if that is already working or not. It might well be.

But that's not ever going to make a different feature set do something it's never ever been meant to do

I think you're making some assumptions here, that I find both unlikely, and without a basis. Aside from the fact that Tesla adds features all the time that were not in the original design framework, there is also nothing to support the contention that HW 2.5 and below will never improve its operation under these conditions. Over time, we have seen drastic improvements going back to the first versions of Autopilot. We've seen hooks for this functionality already in development video captures.

One should not refer to a "manual" and then use words like "never" when they're dealing with Tesla's rapid pace of development.

I think its more likely that Tesla will eventually correct for these conditions and I was hoping to have a technical conversation about its merits and difficulties instead of a cyclical, repetitive and inside-the-box conversation about Tesla's manual.

 

[buyer123456]

For autopilot as the feature is currently sold?


Much simpler list currently:

1. Limited access divided highways highways with no oncoming or cross traffic (including pedestrians)...in a single lane (since AP explicitly contains no features supporting other lanes)- and always with the driver paying attention and ultimately prepared to intervene if needed

That's it. (and still won't prevent freak accidents like a semi tire blowing off and slamming into you, or hydroplaning in bad weather, etc...)


EAP or "new" FSD is basically the same thing, minus the "in a single lane" bit since it does do lane changes (even automatically without any confirmation other than hands-on-wheel, when using NoA)

All of this is pretty clear in the owners manual.

Now- FSD will be adding more places later this year- likely requiring HW3 to do it- but covering city street type situations, stop lights, etc....we don't know the nitty gritty specifics yet, and it's not available to anybody right now (outside some testing folks)

Got it. Maybe not paying for AP(for those trims that do not include AP as standard) will be OK then, given AP's current capabilities.

 

[whatthe2]

Got it. Maybe not paying for AP(for those trims that do not include AP as standard) will be OK then, given AP's current capabilities.

AP's current capabilities are great IMO. It continually improves and will only get better over time. It's one of the main reasons I bought my Tesla and I have no regrets.

 

[Knightshade]

View attachment 409995

I think you're making some assumptions here, that I find both unlikely, and without a basis.

Ok, let's see!

Aside from the fact that Tesla adds features all the time that were not in the original design framework, there is also nothing to support the contention that HW 2.5 and below will never improve its operation under these conditions.

Uh- no.

Tesla actually removed a ton of features when they switched their cheaper offering from EAP to AP.

They went exactly the opposite of the direction you claim- making it clear that AP is a limited system for use in limited cases- and if you want more you need to buy FSD.

Over time, we have seen drastic improvements going back to the first versions of Autopilot.

And during that entire time they repeatedly made clear that AP is not for use where there is cross traffic

They still do.

So far nothing I said appears to have been unfounded, at all... nor based on any assumptions at all, but on facts.

One should not refer to a "manual" and then use words like "never" when they're dealing with Tesla's rapid pace of development.

Why?

Especially when you consider Tesla went out of its way to narrow the functional scope and domain of AP with their changes a few months ago, it seems pretty clear they don't intend to ever offer advanced features in their basic system.

You want advanced, you buy FSD.

I think its more likely that Tesla will eventually correct for these conditions

There's nothing to correct for.

AP is not intended to be used with cross-traffic. At all. The manual is explicit about this.

FSD in the future is intended for this use so if you want that functionality, you need to buy FSD.

and I was hoping to have a technical conversation about its merits and difficulties instead of a cyclical, repetitive and inside-the-box conversation about Tesla's manual.

Again, the only bad assumptions here are yours.

You keep thinking this accident is a "failure" of autopilot.

It's not.

It's an idiot driver not understanding what AP is, or where it's actually intended to be used.

There's nothing for Tesla to "fix" with AP here, other than maybe putting the limits of AP in larger print I guess in the manual nobody seems to want to read.

I agree with what you are saying with a slight addendum. @verygreen has shown that the cars do see cross traffic.

Never said the system can't potentially "see" cross traffic.

I said the system is not intended to operating or react to it.

Hence why the manual tells you not to use it in such places.

I
They just aren’t handling them, specifically trailers, properly at all. Which is why Tesla does not recommend usage of AP with cross traffic.

Exactly.

This is a place the more advanced computer and NNs of HW3 and FSD will be relevant. If you want a car that'll offer advanced driver aids in those places, buy FSD.

Stop expecting them to "fix" something that is explicitly not broken in basic AP.



https://static.nhtsa.gov/odi/inv/2016/INCLA-PE16007-7876.PDF


that's the NHTSA report on the previous incident like this one.

NHTSA’s examination did not identify any defects in the design or performance of the AEB or Autopilot systems of the subject vehicles nor any incidents in which the systems did not perform as designed

The report later points out:

The Tesla owner’s manual contains the following warnings: 1) “Autosteer is intended for use only on highways and limited-access roads with a fully attentive driver...... According to Tesla, Autosteer is designed for use on highways that have a center divider and clear lane markings.....The driver is responsible for deciding when the road type and other conditions are appropriate for system activation

Working. As. Designed.

Driver. Error.

 

[StellarRat]

The issue is people who don't read the manual using AP in places it's explicitly not intended to be used.

AP does not handle cross-traffic well because it's not intended to handle that situation

You can't fix stupid. Even after 3 years.

Yep. I don't use it in cross traffic situations. That's why I don't use it in town. I think using it on a road where it is high speed traffic with crossing traffic is fairly suicidal.

 

[Az_Rael]

Now- FSD will be adding more places later this year- likely requiring HW3 to do it- but covering city street type situations, stop lights, etc....we don't know the nitty gritty specifics yet, and it's not available to anybody right now (outside some testing folks)

What do you think of the new beta stoplight warning for Autosteer that recently released? That one is interesting. According to the manual no one ever should encounter that situation (Autosteer with stoplights) and yet Tesla released a beta form of a warning for it.

I feel like things like that are why people use AP in cross traffic situations. They perceive Tesla as endorsing it despite it not being in the manual. I mean, the manual hasn’t been updated yet for the new (lesser) maintenance requirements, so why would people assume it is up to date for everything else?

 

[Knightshade]

What do you think of the new beta stoplight warning for Autosteer that recently released? That one is interesting. According to the manual no one ever should encounter that situation (Autosteer with stoplights) and yet Tesla released a beta form of a warning for it.

Yup, it's somewhat weird...but for one thing...

Some places actually have stoplights for on/off ramps of highways... which would be a place EAP/FSD (and autosteer via NoA) are explicitly intended to work.... so I suppose it's possible that would be a final-polish NoA feature...and as a bonus they can get a decent amount of info for the eventual HW3/in-city NN.

I feel like things like that are why people use AP in cross traffic situations. They perceive Tesla as endorsing it despite it not being in the manual.

I'm not sure "Well, it says it may or may not notice if I'm gonna run a red light-but won't do anything about it even if I am going to- so I guess it's ok to use it with cross-traffic despite it literally telling me to never use it then!" is a compelling argument...

 

[Az_Rael]

I'm not sure "Well, it says it may or may not notice if I'm gonna run a red light-but won't do anything about it even if I am going to- so I guess it's ok to use it with cross-traffic despite it literally telling me to never use it then!" is a compelling argument...

I agree, but based on the number of YouTube videos of people “testing” it by deliberately trying to run red lights, I am guessing it is enough of an argument for some people.

 

[Swampgator]

Ok, let's see!





Uh- no.

Tesla actually removed a ton of features when they switched their cheaper offering from EAP to AP.

They went exactly the opposite of the direction you claim- making it clear that AP is a limited system for use in limited cases- and if you want more you need to buy FSD.



And during that entire time they repeatedly made clear that AP is not for use where there is cross traffic

They still do.

So far nothing I said appears to have been unfounded, at all... nor based on any assumptions at all, but on facts.




Why?

Especially when you consider Tesla went out of its way to narrow the functional scope and domain of AP with their changes a few months ago, it seems pretty clear they don't intend to ever offer advanced features in their basic system.

You want advanced, you buy FSD.




There's nothing to correct for.

AP is not intended to be used with cross-traffic. At all. The manual is explicit about this.

FSD in the future is intended for this use so if you want that functionality, you need to buy FSD.




Again, the only bad assumptions here are yours.

You keep thinking this accident is a "failure" of autopilot.

It's not.

It's an idiot driver not understanding what AP is, or where it's actually intended to be used.

There's nothing for Tesla to "fix" with AP here, other than maybe putting the limits of AP in larger print I guess in the manual nobody seems to want to read.




Never said the system can't potentially "see" cross traffic.

I said the system is not intended to operating or react to it.

Hence why the manual tells you not to use it in such places.



Exactly.

This is a place the more advanced computer and NNs of HW3 and FSD will be relevant. If you want a car that'll offer advanced driver aids in those places, buy FSD.

Stop expecting them to "fix" something that is explicitly not broken in basic AP.



https://static.nhtsa.gov/odi/inv/2016/INCLA-PE16007-7876.PDF


that's the NHTSA report on the previous incident like this one.



The report later points out:




Working. As. Designed.

Driver. Error.

Good post. I would be careful about calling someone you do not know an idiot.
One of my issues with EAP and FSD as currently deployed is that these systems encourage "testing" which then encourages false confidence, and then leads to skills atrophy.
I don't know of a better way to approach this wicked problem, but to call someone like Walter Huang an idiot (you did not here btw, but same implication) may be missing some major points.

 

[Daniel in SD]

Last year Elon said this and I tend to agree. I bet it is the people who think they know more about Autopilot's capabilities that run into trouble.

“When there is a serious accident it is almost always, in fact maybe always, the case that it is an experienced user, and the issue is more one of complacency,” Musk said. “They just get too used to it. That tends to be more of an issue. It’s not a lack of understanding of what Autopilot can do. It’s [drivers] thinking they know more about Autopilot than they do.”

I'm seeing a lot of people "testing" Autopilot and then assuming that because it worked X times that will behave exactly the same way the next time.

 

[Knightshade]

Good post. I would be careful about calling someone you do not know an idiot.
One of my issues with EAP and FSD as currently deployed is that these systems encourage "testing" which then encourages false confidence, and then leads to skills atrophy.
I don't know of a better way to approach this wicked problem, but to call someone like Walter Huang an idiot (you did not here btw, but same implication) may be missing some major points.

To be clear, FWIW Huang was the one guy who died on AP actually using it someplace it was supposed to work

So it's the one case where it demonstrated an actual failure of the system to operate someplace it was meant to.

THAT said- he still had, based on recorded data, plenty of time to have intervened and either braked or steered the car safely if he'd been paying attention as the system also requires- so I'm still going to call that driver error in the end (and I expect the lawsuits will work out that way too).... but I certainly don't put him in the same class of feature misuse as someone who uses the system in a place Tesla repeatedly informs you it's explicitly not intended to be used at all

 

[Kilotango74]

Not on a X (the car in the article).

That’s the point. Tesla has put substandard control measures in their vehicles to ensure driver compliance. They are easily fooled an ineffective.

 

[dgatwood]

Radio signals can bounce a lot, but they don't just scatter/ spread.

If you were correct, RADAR would not work for literally any purpose. In nearly every situation you can plausibly think of, the vast majority of surfaces will NOT be in a position to reflect a signal directly back towards the receiver. The only reason RADAR works at all is because when that signal hits, the signal does scatter to some degree. If it just reflected straight or nearly straight, it would approximately never work.

Think about an ATC tower, for example. It is aiming a signal at an aircraft, and nearly the entire surface of the plane is pointing at an angle that would reflect the signal away from the receiver. Yet we know that works despite only a fraction of a percent of the area being lined up to reflect in the right direction, if that.

Similarly, when police use RADAR to clock a car, there are approximately NEVER any surfaces that aren't slanted in a direction that would deflect the signal away from them. Yet because a car is mostly metal, the skin of the car acts like a sort of isotropic radiator, reflecting a portion of the signal in every direction.

Per Wikipedia, the things that affect how visible an object is to RADAR include:

• Material
• Size (both absolute and relative)
• Incident angle (emphasis mine)
• Reflected angle, which depends on the incident angle (emphasis mine)
• Polarization relative to the target's orientation

You're right that a truck presents a much weaker signal than if it were at ground level, but that doesn't mean you won't be able to see it at all with RADAR. If you can't see it at all — particularly at a large distance, where the angle of incidence is small — something is very, very wrong.

 

[OPRCE]

In a Reddit article posted last month (Autopilot doesn't detect a truck partially in lane :( : teslamotors), a Tesla Driver had been following a Truck engaged in a lane-shift. The Tesla hit the truck before it had a chance to exit the lane completely. Of note is the fact that the Tesla vehicle was approaching an overpass.

If I understood correctly the way the ADAS maps work when marking a location "radar do not brake" (as is common near overpasses) we should still be relying on visual input from cameras and ultrasonic sensors.

This begs the question:

Does anyone know whether or not the Tesla can brake without input from its Radar (meaning, camera, and ultrasonics alone, or in combination with one another?)

1. I think the answer is a "very conditional yes" -- From 50kmh I've seen AP brake to a halt in daylight behind a car waiting at a temporary roadworks traffic light, but at the exact same spot/time fail to brake if there was no car already waiting, which without intervention would have meant colliding into a 1.2m high barricade of chevron-painted wooden planks perpendicular across the lane before some kind of tank-trap. My conclusion from that was the wood was not recognised visually whereas the rear of stationary car was and neither returned anything useful on radar as both stationary.

2. It also depends on your speed: at 125kmh I've had it not brake for a stationary car in my lane on the motorway in perfect daylight conditions, though visible from 200m out (see sig). The difference could be due to the processing limits of HW2.5 having been reached.

3. In the above Reddit truck situation, even though at low speed, the system may be insufficiently trained to visually recognise somewhat-unusual-looking objects up close and angled.

4. It seems @verygreen verified in the same Reddit thread that this particular "overpass is marked as "radar do not brake" section on the adas maps", meaning that [in the US at least] it relies on GPS to identify when not to brake, which, apart from being an arguably more dangerous "solution" than the phantom-braking issue it is supposed to ameliorate, also introduces another possible point of failure: I've seen Tesla's GPS (and Waze on the phone simultaneously) can sometimes be off by up to 2.4km on clear and open roads, i.e. in worst case scenario the system will ram an unrecognised slow-moving truck that far before or after the overpass, not just within radar range of it.

5. Can anyone confirm that the "radar do not brake" bodge is not in use in Europe? (I've seen plenty of phantom-braking in and around tunnels, but less frequently in recent updates)

 

[VT_EE]

If you were correct, RADAR would not work for literally any purpose. In nearly every situation you can plausibly think of, the vast majority of surfaces will NOT be in a position to reflect a signal directly back towards the receiver. The only reason RADAR works at all is because when that signal hits, the signal does scatter to some degree. If it just reflected straight or nearly straight, it would approximately never work.

Think about an ATC tower, for example. It is aiming a signal at an aircraft, and nearly the entire surface of the plane is pointing at an angle that would reflect the signal away from the receiver. Yet we know that works despite only a fraction of a percent of the area being lined up to reflect in the right direction, if that.

Similarly, when police use RADAR to clock a car, there are approximately NEVER any surfaces that aren't slanted in a direction that would deflect the signal away from them. Yet because a car is mostly metal, the skin of the car acts like a sort of isotropic radiator, reflecting a portion of the signal in every direction.

Per Wikipedia, the things that affect how visible an object is to RADAR include:

• Material
• Size (both absolute and relative)
• Incident angle (emphasis mine)
• Reflected angle, which depends on the incident angle (emphasis mine)
• Polarization relative to the target's orientation

You're right that a truck presents a much weaker signal than if it were at ground level, but that doesn't mean you won't be able to see it at all with RADAR. If you can't see it at all — particularly at a large distance, where the angle of incidence is small — something is very, very wrong.

It’s not a question of whether it can see the truck. It certainly will see some amount of energy reflected. It’s a question of how much confidence the Tesla has that radar signature is actually a truck and not a bridge or background noise. You can’t initiate hard brakes at highway speed unless you have a high confidence there is actually an object in front of you.

 

[mongo]

To be clear, FWIW Huang was the one guy who died on AP actually using it someplace it was supposed to work

So it's the one case where it demonstrated an actual failure of the system to operate someplace it was meant to.

THAT said- he still had, based on recorded data, plenty of time to have intervened and either braked or steered the car safely if he'd been paying attention as the system also requires- so I'm still going to call that driver error in the end (and I expect the lawsuits will work out that way too).... but I certainly don't put him in the same class of feature misuse as someone who uses the system in a place Tesla repeatedly informs you it's explicitly not intended to be used at all

He also reportedly had complained about the cars behavior at that exact location.