HW3

[CarlK]

I saw the FSD tech pitch from Tesla and while it’s impressive it’s essentially trying to emulate a visual cortex with a neuronal network - the actual driving part doesn’t seem to be nearly that ambitious. It came across as “we’ll have an awesome vision system and the rest will follow”. We shall “see”.

Everything they need to do now is for vision and deep machine learning. The actual driving part is super easy. You only need three outputs from the NN to the car for two pedals and the steering wheel. That's actually way easier than how our brains needing to send out instructions to command the muscle system to do the driving.

 

[MountainJedi]

Ok, honestly, do y’all really think that the current product isn’t problematic from a sales point of view?

Tesla wants a significant percentage of the cost of the car for future FSD. And what they present today shows obviously erratic behavior.

You can ridicule this as “the AP core is totally different, why do you focus on the display” but the fact is that the car shows non existing objects that can damage trust in the underlying system. And I don’t care if this is due to <HW3 limitations. If they cannot make this better today, it should be removed.

How something like that can ever pass quality control is beyond me. Truly bizarre.
I guess some people just have low expectations.

Today when I merged manually from the leftmost lane to the adjacent lane, the display freaked out showing a frantically rotating red “truck” that was supposedly in that lane. There wasn’t one.

LOL. Please how can I disable that silly feature?!? I’d be pissed if I were in the FSD team at Tesla and whoever owns that crappy visualization damages the reputation of FSD.

Honestly, I didn't even notice it during my test drive. I didn't purchase FSD for a bunch of other reasons, but the dancing cars wasn't one of them. I suspect most people that make the decision to not purchase FSD do it for either financial or vaporware reasons, not because they think the display is some indication of a failure/fault in the system.

 

[Eno Deb]

While you might be right, history has shown that even the most knowledgeable technologist underestimate the rate at which technology progresses. Let's just hope the same is true with your guess

When it comes to autonomous driving a lot of smart people have vastly overestimated the rate of progress, and pumped billions of dollars into it with (so far) very little return.

 

[Silicon Desert]

When it comes to autonomous driving a lot of smart people have vastly overestimated the rate of progress, and pumped billions of dollars into it with (so far) very little return.

Yup, I guess their pocketbook was bigger than their brain.

 

[Octo]

Everything they need to do now is for vision and deep machine learning. The actual driving part is super easy. You only need three outputs from the NN to the car for two pedals and the steering wheel. That's actually way easier than how our brains needing to send out instructions to command the muscle system to do the driving.

Haha. Yeah, that’s why a toddler whose visual cortex is able to identify objects can drive a car.

Or my cat. She can also see things quite well and distinguish between cars and mice.

Wouldn’t want her behind the wheel though.

 

[dmurphy]

While you might be right, history has shown that even the most knowledgeable technologist underestimate the rate at which technology progresses. Let's just hope the same is true with your guess

My 2000 Pontiac had a cassette deck and cruise control, and I thought it was wonderful. It even had a DIGITAL odometer!

Now I drive a car with a tablet for an instrument panel, steers, accelerates, brakes, changes lanes and enters and exits highways for me automatically. All while it's recording a half dozen cameras around me, while streaming music via high speed cellular connectivity.

Heck, I bought my first GPS unit only a dozen years ago. Remember these?

From this: Magellan pops out RoadMate 2000 series
to this: Tesla Model 3 display confuses police officer

Amazing. This is why I can't even think about what Tesla's doing more than 3-4 years out from now.

 

[SammichLover]

I do not trust autopilot to work as advertised as I have had many "mishaps" that scare passengers,

I use it a lot and with over 35,000 miles now I've only had it do two 'dangerous' things that I needed to wrest it back from executing. Both of those were in areas that were well beyond what current use-case are, both were construction zones with a spaghetti of lines, old lines, scrape marks from removed lines and so on. Yes, I know that's out in the weeds for Autopilot, and yes I was handling it as such.

I have occasionally had it squawking at me because the lines are gone or way too far apart, and it wants help. It also still does some minor 'wiggle' during lane size/splitting transitions. I guess if you're new to it that might be unsettling?

 

[gilscales]

I use it a lot and with over 35,000 miles now I've only had it do two 'dangerous' things that I needed to wrest it back from executing. Both of those were in areas that were well beyond what current use-case are, both were construction zones with a spaghetti of lines, old lines, scrape marks from removed lines and so on. Yes, I know that's out in the weeds for Autopilot, and yes I was handling it as such.

I have occasionally had it squawking at me because the lines are gone or way too far apart, and it wants help. It also still does some minor 'wiggle' during lane size/splitting transitions. I guess if you're new to it that might be unsettling?

Strange in that many miles you have only had this happen twice, it has been 40 or 50 times if I had to guess that passengers were uncomfortable with the cars actions in autopilot, I never worried but each time i'm talking about I did correct steering to the point of disengaging autopilot, passengers freak out a little but it was always in control, I am talking WAY more times than not on clearly marked freeways so not sure why you have not experienced this? maybe you don't get the traffic congestion in "Yup" we get here in the greater Los Angeles area, haha!

 

[SammichLover]

Strange in that many miles you have only had this happen twice, it has been 40 or 50 times if I had to guess that passengers were uncomfortable with the cars actions in autopilot, I never worried but each time i'm talking about I did correct steering to the point of disengaging autopilot, passengers freak out a little but it was always in control, I am talking WAY more times than not on clearly marked freeways so not sure why you have not experienced this? maybe you don't get the traffic congestion in "Yup" we get here in the greater Los Angeles area, haha!

Got some video of what you're talking about?

P.S. Houston only wishes they didn't have "70mph, see the whites of their eyes" congestion. Although I don't drive in that that much, I suppose. Not daily anyway.

 

[M109Rider]

Ok, honestly, do y’all really think that the current product isn’t problematic from a sales point of view?

Tesla wants a significant percentage of the cost of the car for future FSD. And what they present today shows obviously erratic behavior.

You can ridicule this as “the AP core is totally different, why do you focus on the display” but the fact is that the car shows non existing objects that can damage trust in the underlying system. And I don’t care if this is due to <HW3 limitations. If they cannot make this better today, it should be removed.

How something like that can ever pass quality control is beyond me. Truly bizarre.
I guess some people just have low expectations.

Today when I merged manually from the leftmost lane to the adjacent lane, the display freaked out showing a frantically rotating red “truck” that was supposedly in that lane. There wasn’t one.

LOL. Please how can I disable that silly feature?!? I’d be pissed if I were in the FSD team at Tesla and whoever owns that crappy visualization damages the reputation of FSD.

I think Elon has vision that maybe you don’t see.
He’s engineered a product years ahead of everyone else, and it is a work in progress.

You know that when you bought it, you know it, because of the disclaimer when you turned it on, and you have and likely are still using it.
Tesla feels real life usage is the best way and likely a more cost effective way to get to the end goal.
He’s ahead of everyone else, so just maybe he has something here. ..

You seem to feel technology needs to be perfect before rolling it out.
That would be great, but has and likely never will happen in any industry.

How do you think FSD should get to the end goal. I would like to know how you suggest Tesla and others should do this better to get there.?

 

[superbaka]

I can solve the dancing cars on your display with a few lines of smoothing code. That part is not a big deal. I come from decades of experience in the computer graphics industry where we literally created the fore-bearer of all this vision tracking stuff back on Silicon Graphics and Sun Microsystems workstations. As you increase sampling frequency, you increase noise. Can apply various smoothing algos, FFT etc. Same thing happens when you record audio or picture or even your iPhone's GPS tracks. Apple could show your GPS position bouncing all over the place if they wanted. Just because they smooth it, doesn't mean its functioning better, and vice versa.

How does AP perform when you drive? Thats all that matters. The degree of granularity in the representation doesn't. The rate of change I've observed in AP in the past year has given me confidence it will continue to improve. Having said that, it always works best when your eyes are on the road.

 

[dingyibvs]

From what I've read, autopilot features are always running in "shadow mode" if they're not currently in use. Tesla is able to gather this data and look at what the car would have done in certain situations had autopilot actually been engaged. So the system as a whole is constantly learning from data it collects at all times whether AP is active or not.

There is no way that's happening, not the way most people would think anyway. That is a MASSIVE amount of data, and there's no way they can all be transmitted let alone processed by Tesla.

This doesn't seem like a good idea to me at all. It's another point of failure to add into the system. There can always be lag even on a fast data connection. What if you're driving in an area with spotty or no cellular data coverage? What if there's a server outage? The current system is able to react to potential dangerous situations in a fraction of a second. I wouldn't want to extend that reaction time by adding in the delay that would come with transmitting the current environment around the car to some central server and then waiting for a response back on how to handle the current situation. This would also make for a rather attractive target to some malicious actors. Want to take down autopilot across the entire Tesla fleet? Just DDoS their servers.

Oh believe me, it's gonna happen, just a matter of time. 5G's revolutionary nature is not in cell phones, but exactly this type of application. It enables in many situations what currently can only be done with a wired connection. Ultra low latency is also one of the hallmarks of 5G, a few milliseconds as of now, compared to 100+ms even for the fastest humans, and certainly not a limiting factor in reaction time in traffic. AP as it stands now or even a few years back is enough to safely stop the vehicle if a massive system outage happens. Such a black swan event will still result in traffic paralysis, but the benefits the other 99.99+% of the time will make it a worthwhile commercial endeavor for people.

 

[Octo]

I think Elon has vision that maybe you don’t see.
He’s engineered a product years ahead of everyone else, and it is a work in progress.

You know that when you bought it, you know it, because of the disclaimer when you turned it on, and you have and likely are still using it.
Tesla feels real life usage is the best way and likely a more cost effective way to get to the end goal.
He’s ahead of everyone else, so just maybe he has something here. ..

You seem to feel technology needs to be perfect before rolling it out.
That would be great, but has and likely never will happen in any industry.

How do you think FSD should get to the end goal. I would like to know how you suggest Tesla and others should do this better to get there.?

I like his vision and I’m personally convinced that the camera approach (vs Lidar) is good for the simple fact that humans can drive with a pair of eyes and articulating head. Without Lidar.

The Tesla way of using a neural net and multiple cameras for machine vision seems a great solution based on evolution of several species in the biological world.

What I don’t like is releasing half baked products like the visualization in my Model 3 because I’m afraid that this will sour the public on self driving vehicles. There’s already plenty of skepticism against EVs and if THE best, leading EV manufacturer releases that crappy visualization that makes it look like it’s broken - make no mistake, that’s how this comes across to non Tesla fans - it has the potential to damage Tesla’s reputation in that field.

Maybe I have judgmental friends but I can see their eyeroll when they see the stupid erratic toy cars (and often weird lanes) on my Model 3 display.

 

[PoitNarf]

There is no way that's happening, not the way most people would think anyway. That is a MASSIVE amount of data, and there's no way they can all be transmitted let alone processed by Tesla.

I'm just going by what I've read and what Tesla said during their Autonomy Day a few months back. They don't need to have every data point of every drive sent back to them. They work on specific corner cases and then query the fleet for data on specific types of conditions and then vehicles that have matching data send it back to Tesla.

 

[turnem]

I'm pretty impressed by the current iteration of Autopilot. I use it on more than 80% of my daily suburban commute (surface streets) and it works very well.

The biggest failure that I see right now is that it cannot seem to handle going through an intersection that bends slightly to the left. It does fine when going slightly to the right or straight but a slight bend to the left confuses it greatly. When it picks up the left hand lane marker on the other side of the intersection it assumes it's the right hand lane marker and wants to pull your sharply into the left lane (assuming a four lane road and I'm in the right lane - it doesn't do this if I'm in the left most lane on my side of the road thankfully because that would put me into oncoming traffic).

If it weren't for this issue I'd only have to engage to make stops at stop lights and stop signs where no cars are in front of me and to make turns onto new streets.

It does very well on interstates and divided highways in my opinion. I rarely have to engage here.

 

[Kognos]

I can solve the dancing cars on your display with a few lines of smoothing code. That part is not a big deal. I come from decades of experience in the computer graphics industry where we literally created the fore-bearer of all this vision tracking stuff back on Silicon Graphics and Sun Microsystems workstations. As you increase sampling frequency, you increase noise. Can apply various smoothing algos, FFT etc. Same thing happens when you record audio or picture or even your iPhone's GPS tracks. Apple could show your GPS position bouncing all over the place if they wanted. Just because they smooth it, doesn't mean its functioning better, and vice versa.

I respect your experience, first off - but it speaks volumes that this technology exists, they're aware of it, and the entire fleet is laughing (and not in the friendly, jokey way) about the dancing cars and if it were so simple to fix, it should have been fixed by now.

You understand the behavior given your background. I understand it philosophically coming from a similar background. But my wife gets in the car and is immediately freaked out that it looks like another car is slicing into the back end of our car, over and over again. If you get enough bad information, humanity engineers us to start ignoring and distrusting what we thought was "true" which is - the visual representation of cars is buggers. And we're supposed to trust this system? This coded behavior creates distrust.

I'm not in the business to make excuses on behalf of tesla, except where it's literally true. Sure the panel gaps are real, but I can make a call and get those fixed, for example. The dancing cars? .... it's pretty crazy bad at stop lights. Almost a joke.

So I completely sympathize with the poster who chose not to get FSD because of the dancing cars. When a car is run by technology, and that technology is doing something bad, AND when someone says "it should be simple to fix" ... why isn't it?

I'd like to think the answer actually IS NOT as easy as you suggest, because I would like to have more faith in the Tesla software. Not to be argumentative, but we all need to trust this software, literally with our lives.

 

[Malbers]

My comment was from the view of a customer who Tesla tried to pitch FSD for $6000.

If the underlying system is much better in detection of objects as you claim, why is the visualization, arguably magnitudes simpler to program than a freaking vision system for FSD, so terrible?

At best it’s a marketing disaster because people like me who see dancing cars where non are in reality, won’t fork over $6000. If the AP core is truly better than what the car shows on its screen, then a fix should be a 1 week SW project for an enthusiastic intern, right?

I suggest fixing this ASAP if they want to upsell this feature. Because outside of hardcore engineering fans like you seem to be it’s a ridiculous marketing boondoggle.

Agreed. Dancing cars is bad PR, and AFAIK, smoothing out the image of the dancing cars shouldn't be a difficult proposition.

Maybe the executive shake-up was a good thing if they weren't able to recognize an easy win like fixing the dancing cars.

 

[Knightshade]

Heck, I bought my first GPS unit only a dozen years ago. Remember these?

From this: Magellan pops out RoadMate 2000 series
to this: Tesla Model 3 display confuses police officer

Amazing. This is why I can't even think about what Tesla's doing more than 3-4 years out from now.

Hopefully in 3-4 years they'll have figured out how to allow you to add a waypoint in the nav system.

Like those GPS units were already doing 15 years ago.


My Teslas inability to do so seems a lot more embarrassing a "simple thing it somehow still can't do" than the dancing cars thing.

The biggest failure that I see right now is that it cannot seem to handle going through an intersection that bends slightly to the left. It does fine when going slightly to the right or straight but a slight bend to the left confuses it greatly.

It's not a failure- it's using the system someplace it is explicitly not intended to be used

The manual is pretty clear on that.

The fact it often does seem to work in places it's not at all meant to doesn't mean it's a failure when it suddenly doesn't work there.

It does very well on interstates and divided highways in my opinion.

Which is the primary place the system is actually meant to be used.

 

[SammichLover]

I'm just going by what I've read and what Tesla said during their Autonomy Day a few months back. They don't need to have every data point of every drive sent back to them. They work on specific corner cases and then query the fleet for data on specific types of conditions and then vehicles that have matching data send it back to Tesla.

I have take issue with the concept overall, its inherent limitations for passive field testing. However its ability to gather certain types of pertinent data for use in human driven development of the NN isn't part of that problem. They do seem to be able to effectively gather snapshots of certain classes of events from fleet wide, which somewhat is able to sidestep the bottleneck of data capture and reporting bandwidth that a single vehicle is capable of.

 

[gilscales]

Got some video of what you're talking about?

P.S. Houston only wishes they didn't have "70mph, see the whites of their eyes" congestion. Although I don't drive in that that much, I suppose. Not daily anyway.

No I don't have any video just a lot of instances where I start to use autopilot and get a little disgusted with its lack of finesse or the heavy phantom braking slowing the car down 30 mph or more quickly or the unintended incomplete lane change attempts or the hugging of the center divider placed to close to the lane on so many freeways around here or the constant brake checks when no one is entering my lane or the (insert various problems here)
Its not that I do not like autopilot as I do use it and like it when I can use it for extended periods but I find I do not get as much use out of it as I had expected I would and I hope soon it will improve for me.

FWIW, this is the best car I have ever owned and beats the hell out of many more expensive cars I have had in the past and I cannot imagine owning anything other than a Tesla for a long, long time as I feel it will be many years before anyone ever challenges them in a meaningful way.