MX HW2 detects overpasses as a car and hard breaks

[Pilot4Hire102]

Hi,

Recently on a road trip the car breaked hard in the middle of the highway just before an overpass, it also happen when I passed under some sort of metal structure used for signs.
We were lucky no car was behind us.

Anyone else with a similar experience?

This is how it works - tacc or autopilot use radar. Radar is great for many scenarios excep overpasses and overhead signs etc. to those the radar sees it as an oncoming collision.

As you continue to drive that route it becomes listed as safe passing .

Personally when I first got tacc active , it was really weird around any overpass on my normal commutes , but each time it was less careful - and now I drive straight through and it doesn't even change a thing.

So if it's your first time going through one , maybe you could take over if you're worried . Hw2 cars just hit the road late nov and dec and we are basically mapping the world for tesla as we drive .

If you drive my route near my home and work you'll not have this happen at all because my car has paved the way for you and anyone else for the future .

Kinda cool to know how much we are contributing actually :....

 

[Hans Allis]

I understand what you're saying, but it seems to contradict this snippet from the Tesla blog post called "Upgrading Autopilot: Seeing the world in radar":

When the data shows that false braking events would be rare, the car will begin mild braking using radar, even if the camera doesn't notice the object ahead. As the system confidence level rises, the braking force will gradually increase to full strength when it is approximately 99.99% certain of a collision.

It seems that Tesla's data already shows that false braking events would be rare and the braking we're experiencing is classified as "mild". Going from 120 km/h to 80 km/h in a short time doesn't constitute "mild braking" in my opinion though. I feel that they've either chosen a different approach from what was initially described or that they misinterpreted the data?

 

[disagree]

Have gone under a particular sign dozens of times with TACC. Last night as I went under it I got FCW and dropped more than "a brake tap" in speed. It's totally random when this garbage kicks on.

For Chicagoans here, it was the "These three lanes"-94E sign at the split with 90E by Montrose. Earlier in the same drive I got the 5mph brake tap with no warning from a sign around Fullerton.

This is ridiculous.

 

[Kant.Ing]

I believe the system now can dodge a random big bird. Cool.

 

[stefan.bc]

So if it's your first time going through one , maybe you could take over if you're worried . Hw2 cars just hit the road late nov and dec and we are basically mapping the world for tesla as we drive

I do not believe the system is learning when we are 1.5 months into AP2 and the car still has FWC on major highways. Unless I am the only one driving an AP2 in my area.

 

[disagree]

I do not believe the system is learning when we are 1.5 months into AP2 and the car still has FWC on major highways. Unless I am the only one driving an AP2 in my area.

People just refuse to listen.

Other drivers and I have done the highways in Chicago over and over, and we are still getting FCW/hard braking on these highways. It's completely unpredictable. I had never had a FCW in the spot where I did last night. It scared the living hell out of me.

There is no evidence of learning commonly taken routes.

There is evidence only of people wanting to believe that there is learning.

 

[acadapter]

Has anyone read the article that shows current AP2.0 cars are only using 1/8 cameras at this time?

Tesla Autopilot is currently using only 1 out of 8 cameras of the new hardware suite, still very much in ‘beta’

The system can quite possibly be learning, but feedback may not be getting back to the cars as of yet. This system is in beta and the 6-7 warnings everyone gets about it being in beta tell you this repeatedly. They also suggest if you feel as if the current release isn't something you can closely monitor, waiting until more miles have been fed into the system may be a better approach for you.

 

[stopcrazypp]

I do not believe the system is learning when we are 1.5 months into AP2 and the car still has FWC on major highways. Unless I am the only one driving an AP2 in my area.

People just refuse to listen.

Other drivers and I have done the highways in Chicago over and over, and we are still getting FCW/hard braking on these highways. It's completely unpredictable. I had never had a FCW in the spot where I did last night. It scared the living hell out of me.

There is no evidence of learning commonly taken routes.

There is evidence only of people wanting to believe that there is learning.

Learning does not work with only one or a few drivers driving the same route (even every day). It needs enough samples (from different drivers) to prune out bad data. Also, the way it works is the data will be processed by Tesla centrally and then if deemed valid, will be uploaded to cars in the next update. There is no evidence Tesla does realtime data updates, so only after a software updates will there be any changes. How many updates have there been since the AP2 update release? Probably not enough to prune out some of the less travelled highways (by less travelled, I mean by new Teslas, not in general).

 

[disagree]

Learning does not work with only one or a few drivers driving the same route (even every day). It needs enough samples (from different drivers) to prune out bad data. Also, the way it works is the data will be processed by Tesla centrally and then if deemed valid, will be uploaded to cars in the next update. There is no evidence Tesla does realtime data updates, so only after a software updates will there be any changes. How many updates have there been since the AP2 update release? Probably not enough to prune out some of the less travelled highways (by less travelled, I mean by new Teslas, not in general).

lol. 90/94 in chicago, bruh.

 

[stopcrazypp]

lol. 90/94 in chicago, bruh.

Unfortunately updated state breakdowns are not available, so hard to tell how many new Tesla sales have been there since AP2 has been released, but then it can not be ruled out either that there are special circumstances related to overpass designs there that the system fails to whitelist.

But just because one area still has problems does not prove there is no fleet learning going on. The person upthread provided an example that suggests fleet learning from varying behavior over time.

 

[disagree]

Unfortunately updated state breakdowns are not available, so hard to tell how many new Tesla sales have been there since AP2 has been released, but then it can not be ruled out either that there are special circumstances related to overpass designs there that the system fails to whitelist.

But just because one area still has problems does not prove there is no fleet learning going on. The person upthread provided an example that suggests fleet learning from varying behavior over time.

The truth is out there.

 

[Pilot4Hire102]

People just refuse to listen.

Other drivers and I have done the highways in Chicago over and over, and we are still getting FCW/hard braking on these highways. It's completely unpredictable. I had never had a FCW in the spot where I did last night. It scared the living hell out of me.

There is no evidence of learning commonly taken routes.

There is evidence only of people wanting to believe that there is learning.

What " evidence do you need " ? Because my comment along with personal experience is obviously a lie because you personally haven't experienced the same thing ( or so you suggest ) .

If a tree falls and You're not there it doesn't make sound either does it ?

 

[disagree]

What " evidence do you need " ? Because my comment along with personal experience is obviously a lie because you personally haven't experienced the same thing ( or so you suggest ) .

If a tree falls and You're not there it doesn't make sound either does it ?

Of course it makes a sound when it hits the ground in spot X! It's just that the next time a tree falls, if it falls in spot Y, we don't say that spot X had learned how to dodge falling trees.

I believe that you believe that your anecdote is evidence that AP is "learning." I don't think you're lying. I think you're wrong. The system is not learning; the system is scattershot.

 

[cybergates]

why is TACC so hard for tesla. My 2007 infiniti performed flawlessly with its TACC or my sisters 2014 lexus. Never heard of an unexpected slowdown in those cars - why is this so hard for tesla to figure out. really no excuse (AP2 17.17.17)

I think these guys need to popup the disclaimer every time the car is sorted until the "beta" glitches are sorted out.

 

[Topher]

Never heard of an unexpected slowdown in those cars

There are two types of problems here: False positives (rare mistakes on frequent occurrences) and false negatives (mistakes on extremely rare occurrences). They are linked, and false positives can be reduced, at the expense of false negatives (without a lot of actual hard work) . Knowing the false positive rate does not tell you much about the false negative rate.

Thank you kindly.

 

[chillaban]

why is TACC so hard for tesla. My 2007 infiniti performed flawlessly with its TACC or my sisters 2014 lexus. Never heard of an unexpected slowdown in those cars - why is this so hard for tesla to figure out. really no excuse (AP2 17.17.17)

I think these guys need to popup the disclaimer every time the car is sorted until the "beta" glitches are sorted out.

There are two types of problems here: False positives (rare mistakes on frequent occurrences) and false negatives (mistakes on extremely rare occurrences). They are linked, and false positives can be reduced, at the expense of false negatives (without a lot of actual hard work) . Knowing the false positive rate does not tell you much about the false negative rate.

Thank you kindly.

Bingo. Most cars have an ACC system that assumes the driver knows what he's doing, and if it ever doubts anything, it assumes the driver is in control and to keep maintaining speed. In other words, it's more than happy to chop your head off or clip a car that just barely inched into your lane, or assume around a sharp curve that any cars it sees are not in your lane.

You tend to have more cases where you needed to brake and disengage ACC because it was going to ram into something.

TACC in AP1 was basically the same way, but for whatever reason, in AP2, it seems to be annoying paranoid and assume the worst instead of the best.

 

[stefan.bc]

TACC in AP1 was basically the same way, but for whatever reason, in AP2, it seems to be annoying paranoid and assume the worst instead of the best.

Why didn't they allowed drivers to disable it and downgrade to a regular CC? How hard is it to add a button to disable TACC, similar to how AS2 works? Or disable TACC when AS is disabled.

Current TACC should not have been forced on us without the additional layer of GPS data.

 

[AnxietyRanger]

FWIW, I have gotten maybe 5-10 smaller braking events before overpasses and one hard on AP2. It happens often enough to not make me very confident around overpasses.

 

[ScottieLV]

We have 2017AP2 MX 90D and drove it from Vegas to Stockton, CA through Bakersfiled up 99 and back on I-5 freeway. During our trip I only used cruise control. We had at least 9 occurrences of the cruise control braking hard without reason. All but one we were going under overpasses, but one time on I-5 it braked hard with nothing around us. If someone were following close behind it could cause a severe rear end collision. We love our MX but this needs to be remedied soon before someone gets killed or injured. Just sayin'