Is FSD a myth, never to become reality?

[NigelM]

Don't beat me up.....Since moving north from FL I've experienced some very different weather and road conditions and I'm becoming increasingly convinced that FSD is perhaps never going to be a reality for most of us.

• Road markings are critically important to FSD and I'm discovering that in the mid-Atlantic region there's plenty of roads where AP isn't even available.
• Currently, the car/camera often fails to see speed signs...this could be corrected via better SW but is there some inherent issue in sign differences?
• Even in light snow, road markings can become somewhat obscured and AP already becomes unavailable.
• With latest SW update in heavy rain yesterday, not only did AP become unavailable but I also got the dash message that blindspot detection was limited or unavailable.

AP is great, but it's still essentially a very sophisticated cruise control and Tesla SW improves all the time (except for #wipergate ) BUT will FSD ever become year-round available outside of the "sunshine" States? Discuss.

 

[ewoodrick]

We all understand that in Northern Virginia that snow is on the ground 364 days a year and than in that one day that it is not, there's not a single line on the road in the entire state.

So indeed, in your case, it will probably never come.

By the way, how do you ever manage to drive if you can never see the roads?

Or maybe you are exaggerating the conditions a little? And maybe you only have snow covered roads a few days of the year, before the plows make it out?

 

[NOLA_Mike]

Currently, the car/camera often fails to see speed signs...this could be corrected via better SW but is there some inherent issue in sign differences?

Nigel - I don't believe Tesla's current hardware/software (since the initial AutoPilot "1") is reading speed limit signs at all - it is simply using a stored database of speed limits. That will need to change before FSD because the database isn't always up to date (thus giving the impression that it is missing speed limit signs when the database and the actual posted speed limit are different) and also temporary speed limits (like construction zones) will need to be recognized.

 

[Tam]

...Currently, the car/camera often fails to see speed signs...

All public versions do not recognize speed signs. They have been relying on speed database and correlate that with GPS.

...Road markings are critically important to FSD...

Road markings are like training wheels for toddlers to learn to stand and walk. FSD should eventually be weaned off from road markings and should be able to drive with no road markings needed.

...heavy rain...

Radar should "see" through rain fine. TeslaVision will be trained on how to drive in the rain. But don't ask me when!

 

[VQTRVA]

Don't beat me up.....Since moving north from FL I've experienced some very different weather and road conditions and I'm becoming increasingly convinced that FSD is perhaps never going to be a reality for most of us.

• Road markings are critically important to FSD and I'm discovering that in the mid-Atlantic region there's plenty of roads where AP isn't even available.
• Currently, the car/camera often fails to see speed signs...this could be corrected via better SW but is there some inherent issue in sign differences?
• Even in light snow, road markings can become somewhat obscured and AP already becomes unavailable.
• With latest SW update in heavy rain yesterday, not only did AP become unavailable but I also got the dash message that blindspot detection was limited or unavailable.

AP is great, but it's still essentially a very sophisticated cruise control and Tesla SW improves all the time (except for #wipergate ) BUT will FSD ever become year-round available outside of the "sunshine" States? Discuss.

That's more a VDOT problem than a Tesla problem.

Having grown up & lived in NoVa most of my life, I know the area's hyper growth has taken the wear & tear to the roads up there up several levels.

So no navigational system can keep up until the rate of infrastructure changes or deterioration slows down in the DMV. Even a superior intellect human like me gets jittery & quizzical driving up there with all the changes to signage, local tribal road knowledge & huge shifts in road topography.
And driving in snow? Even i call in sick on that! Rain & blindspots? I turn my head & use my blinkers all the time.

That's just smarts not an engineering issue for Tesla.

 

[Fellsteruk]

It will happen just a matter of time for tech to enable it.

Humans drive in light snow when no white lines as could an AI when it gets advanced enough.

 

[NigelM]

By the way, how do you ever manage to drive if you can never see the roads?

We could wave off all discussion with flippancy, but my point is that current AP consistently can't deal with external variables; the human eye and brain allow us to deal with all sorts of scenarios and doesn't stop us driving e.g. on roads with zero markings or in adverse weather.

 

[jboy210]

We could wave off all discussion with flippancy, but my point is that current AP consistently can't deal with external variables; the human eye and brain allow us to deal with all sorts of scenarios and doesn't stop us driving e.g. on roads with zero markings or in adverse weather.

Never is a long time, and Artificial Neural Networks (ANNs) are in their infancy. The head of Tesla's AI group, Andrej Kaparthy, is quite brilliant and has made many interesting discoveries in the areas of both recognition and cognition. And even an idiot like me can create models that do better at vision tasks that untrained humans, and match skilled humans in some tasks currently performed by human doctors.

Finally, consider that Machine Learning systems can learn by watching humans. So if a human can learn to drive in the snow by following other car's tracks when external markings are not present, so can the ANN.

 

[BAven8068]

Not only are the covering of road markings an issue, but so is the covering of the radar by snow. Just yesterday it was snowing in Chicago, I had about 2 miles of highway driving with AP before it was all gone due to the front bumper being covered in a thin layer of snow.

I assume they can't heat this area due to the possibility of damaging to plastic bumper over time. If that's the case the radar needs to be moved from behind the bumper and placed somewhere else where the actual surface of the radar can be heated. But they have already made all these cars that are supposedly ready for "FSD" with the radar unheated and behind the bumper.

Yes I know that in the future vision alone may be able to do the job of the radar, but you're going to still need a back up to that vision especially in snowy conditions where the snow could diminish the vision systems effectives, so now you have a snow cover radar as a back up in that situation, not idea.

 

[NigelM]

That's more a VDOT problem than a Tesla problem.

Well then it's also a Maryland, DC, WV and even NC problem. FSD vehicles will have to deal with road conditions, it's never going to be possible to have (decentralized) infrastructure adjust to suit the technology.

That's just smarts not an engineering issue for Tesla.

Surely meeting the variable conditions is an engineering issue for anyone who wants to offer true FSD?

 

[1375mlm]

The reason I think that the new timeline—1 year—Is different & now believable, is because it’s Tesla issued

Now correct me if I’m wrong...

Tesla has NEVER issued a timeline for FSD. No hints, no partial timelines, nothing. Not on the configurator, not on blog posts, and not through their tweets.

Notice I’m not talking about EAP, which they did offer a timeline of, and were sued over it. That’s the last and only timeline for AP they issued.

And also notice I’m not talking of Elon’s tweets/interviews... just Tesla as a car company talking about its own products.

This is the first time Tesla itself has spoken about a time goal related to Autopilot Full Self Driving, leading me to believe that this time they know it for sure or are even being a bit conservative (haha), because the last time they missed the AP deadline they got sued.

Not to say that it will be fully functioning and thought out by the end of the year... it will probably be quite rudimentary with tons of hesitation and close calls if left to make its own decisions. Most likely also with nag, and jerky acceleration/braking.

But Tesla will be able to call it “feature complete” In the sense that ideally, it can handle all driving situations.

Also Elon reiterated his confidence in the system during a recent interview, going as far as to say that not only would they be FSD, but they would be able to function as driverless Ubers in the near future.

 

[VQTRVA]

Well then it's also a Maryland, DC, WV and even NC problem. FSD vehicles will have to deal with road conditions, it's never going to be possible to have (decentralized) infrastructure adjust to suit the technology.

Surely meeting the variable conditions is an engineering issue for anyone who wants to offer true FSD?

I still believe humans the most skilled & intelligent drivers & even then we fall short due to changing conditions AND we learn constantly through doing. So falling short as human due to changing conditions are normal.
So when AMZ is having a room full of people listening into our conversations to help its ML & AI efforts & Alexa operate under very narrow parameters, I can't imagine people thinking FSD is as easy when the rules are constantly changing much much more.

FSD will come, but don't put a timeline when you aren't prepare to consider our own shortcomings.

 

[widodh]

We could wave off all discussion with flippancy, but my point is that current AP consistently can't deal with external variables; the human eye and brain allow us to deal with all sorts of scenarios and doesn't stop us driving e.g. on roads with zero markings or in adverse weather.

This!

I emphasize the "Full" in FSD. I live in Europe where we usually have very good roads, but, I go to some places in Germany a few times a year where I have to drive through the woods over a dirt road to reach my destination.

I need to use my consciousness to make judgements and this has nothing to do with driving. I know my place in my surroundings.

The same applies to various situations on the road where you need to bend the rules slightly to continue your way or to make the best out of a situation.

99% FSD, sure, but that's not "Full". We'll get very close, but I don't believe a robot will ever perform in ALL situations.

 

[jboy210]

I still believe humans the most skilled & intelligent drivers & even then we fall short due to changing conditions AND we learn constantly through doing. So falling short as human due to changing conditions are normal.
So when AMZ is having a room full of people listening into our conversations to help its ML & AI efforts & Alexa operate under very narrow parameters, I can't imagine people thinking FSD is as easy when the rules are constantly changing much much more.

FSD will come, but don't put a timeline when you aren't prepare to consider our own shortcomings.

Unfortunately human were never designed to drive vehicle and we have the accidents rates to prove it.

40,000 traffic deaths/year in the US and over 90% caused by human error. If FSD or even EAP could reduce that by 25% by preventing accidents when we fall asleep, are distracted, drive too fast for conditions, don't look before changing lanes, etc, that would be 10,000 lives saved.

Also remember AI systems are constantly learning and evolving. Any AI system is part of framework that provides new data and retrains models. This is similar to humans learning from new experiences and can occur at much faster rates.

 

[azred]

Don't beat me up.....Since moving north from FL I've experienced some very different weather and road conditions and I'm becoming increasingly convinced that FSD is perhaps never going to be a reality for most of us...
...
AP is great, but it's still essentially a very sophisticated cruise control and Tesla SW improves all the time (except for #wipergate ) BUT will FSD ever become year-round available outside of the "sunshine" States? Discuss.

You may be onto something except for the “most of us” comment since there are many more owners in the Sun Belt than the Rust Belt. Perhaps AP4 or 5 will accommodate salted streets. In the meantime, get a heated garage so your door handles don’t freeze. Or buy a Detroit product and support your fellow Rust Belters.

 

[DCGOO]

All public versions do not recognize speed signs. They have been relying on speed database and correlate that with GPS.

Road markings are like training wheels for toddlers to learn to stand and walk. FSD should eventually be weaned off from road markings and should be able to drive with no road markings needed.

Radar should "see" through rain fine. TeslaVision will be trained on how to drive in the rain. But don't ask me when!

I don't think that is correct. AutoPilot and FSD will always be based on visual cues. There is only a single radar, out the front which is basically only used for speed (TACC). Lane keeping and all other FSD features rely upon the cameras. it does not appear that any effort is being made to change that approach. If you cannot see the road well enough to drive, Autopilot/FSD cannot either, in fact probably less so.

Actually the old AP1 read street signs just fine. I think there was an IP dispute over the techniques used to read signs, when the relationship between Tesla and Mobileeye went sideways.

 

[Tam]

...Actually the old AP1 read...

Agreed on AP1 Speed Signs recognition.

...will always be based on visual cues...

Tesla now can drive on its own very slowly within 150 feet of its owner on a parking lot with no lane markers needed.

In the future, it could transfer that skill to roads with no lane markers such as when they are covered by snow or on a dirt road...

I am explaining why "never" is an overstatement and future possibilities are there to be solved but don't ask me when

 

[NigelM]

Or buy a Detroit product and support your fellow Rust Belters.

Helpful addition to the conversation....

 

[dfwjensen]

Unfortunately human were never designed to drive vehicle and we have the accidents rates to prove it.

40,000 traffic deaths/year in the US and over 90% caused by human error. If FSD or even EAP could reduce that by 25% by preventing accidents when we fall asleep, are distracted, drive too fast for conditions, don't look before changing lanes, etc, that would be 10,000 lives saved.

Also remember AI systems are constantly learning and evolving. Any AI system is part of framework that provides new data and retrains models. This is similar to humans learning from new experiences and can occur at much faster rates.

As humans, we all have to learn from our own experiences, and become better drivers as we learn and experience various conditions as we travel. But each new driver must go through this learning experience. The advantage of an AI system is that all these learning experiences are shared with the whole network, and the AI never has to "start over".

 

[jboy210]

As humans, we all have to learn from our own experiences, and become better drivers as we learn and experience various conditions as we travel. But each new driver must go through this learning experience. The advantage of an AI system is that all these learning experiences are shared with the whole network, and the AI never has to "start over".

This.

We learn through experience and by guidance. AI systems do also, but can shortcut that learning experience by having all the experienced drop into their NN on day one. The individual versions of the system running in each car can then gain a bit of experience in it's individual situations and have that feedback incorporated into a new version of the NN model. This updated model then becomes the basis for the next generation, and the process repeats again.

Another thing I am studying is the concept of learning tasks via simulation. This will let a NN model learn without the every having to experience a task in the real world. In some cases they are able to do this at a high enough level that we can train a model in simulation to throw a ball 10 meters into a hole. And when the trained model is dropped into to a robot and the robot can hit the hole on it's first try, and the second, and .. the nth, even as the target and robot move.