Will Tesla be able to deliver FSD with HW3.0 and current Model 3 sensor suite, ever?

[MacsEV]

a well know expression.. “Perfect is the enemy of the good” applies here.
I agree that perfect FSD is likely not achievable any time soon..
the issue of insurance, laws etc is also a barrier..

I have done a few long drives ( 10 day round trips) and the current level of FSD, even just with HW2, does take a lot of stress off a 6-8 hrs drive..

Frankly, I’d be happy with focus on warning me ( much more effectively than is delivered now) that situational awareness is beyond the capabilities of the car.. give me 3 seconds warning and I can take over.. Current warnings are woefully inadequate

another discussion item to consider is that vehicles should be communicating with each other in order to achieve a more perfect FSD.

 

[run-the-joules]

I will take the range over self-driving any day. I do like my self-driving feature. But if Tesla was to offer me a car with the million-mile battery( not based on a million miles) and a range of close to 500 and above in lieu of self-driving, it would be an easy choice for me.

Lol, we're opposites. If they refreshed and the Model 3.1 had 10,000 mile range on a single charge but the autonomy never advanced past where Autopilot is now, and there was a competing car that was mostly similar but only had 250miles of range, but with L4+ autonomy? I'd change to them.

 

[Stv1der]

I'm in France and took delivery last week. Very disappointed. Stop signs, roundabouts, turn off at exit on autopilot nav..none working. Also, despite holding the steering wheel as if I'm strangling it, I'm required to move it slightly every 10 seconds. Uncomfortable and annoying.
The rest of the car is fine but I feel I've been blackmailed into enhanced autopilot

Here here. In the UK everything has been dumbed down, exactly as you say. Summon rarely works, smart summon is disabled. Self parking in perpendicular spaces is a joke and if there is parked car on the side of the road, it slams on the brakes and don’t get me started on push bikes! Don’t get me wrong I absolutely love this car but feel, like you, I’ve been cheated by the FSD upgrade!

 

[azfitz]

I think the real “tell” on FSD will be if they change the lease agreement on M3s, such that owners will be allowed to purchase them at the end of their lease. Currently all leased M3s are destined for Tesla’s autonomous FSD fleet. If true FSD doesn’t exist what will Tesla do with all these returns? Flood the market with low mileage M3s?

 

[Nirmaljal]

The more I see how slow Tesla is at delivering new FSD features, the more I think it will never happen. Assuming current HW3.0 and sensor suite.

I mean - Ok, stop signs. Great. In 5 years development timeline? How long it will be before Tesla releases unprotected left turns on intersections? Another 5 years?

Bottom line - my prediction is Tesla will never be able to release FSD, even in USA, on HW3.0 and current sensor suite.

Now I guess 99% of the forum would disagree with me. I’m curious - is there anyone else who agrees with my prediction?

Well the 5 year comparison is not accurate, it is only now that Tesla has started uncovering the true potential of HW3 with neural net (aka AI) learning algorithm. They had to do a complete software rewrite, and the effort is starting to bear fruit. I have had my M3 for over a year now, the progress that I have seen over this one year is remarkable. My drive to office is about 55 miles, 45 miles of it is highway driving on I35 in Texas. The M3 works like a charm, getting on and off ramps, merging in toll lanes, managing lanes in construction zones. Now for my office commute I have reached I point where I trust the car more than my own driving, as it is able to anticipate and react to drivers ahead suddenly breaking better than what I do. Well I would admit I do take over 2-3 times during the drive, not because I feel unsafe, but because I know that the autopilot is going to drive more conservatively and merge into a lane-change 2 miles before, whereas I would merge in the last 0.25 miles, like any other driver. I also take over sometimes for the pure thrill of driving fast.

For the past few weeks I have also tried the stop sign and intersection feature, again to a person who doesnt own or drive a Tesla, the feature might feel stupid that the car stops irrespective of whether the light is green or red. But based on my driving experience, I actually feel even more confident, as I am sure that the car is going to slow down, it ensures that I am engaged as a driver. I just give it a tap of the foot and it keep going. I know that I am teaching the car to learn and not in too distant future it will be able to do it itself.

I feel that the money that I paid for FSD is worth it, as my mental state after my commute is totally relaxed now. My family tells me that I am really much more pleasant after returning from work. Now that is totally worth the 7K.

Also I feel feature-complete FSD is a misnomer. Are we as humans feature-complete in driving? Do we make no mistakes? I feel the saying applies perfectly here "that perfection is the enemy of progress". Full autonomy is a distant dream, I totally agree. But well for my daily commute I have two unprotected left turns on intersections for my whole commute, I am more than happy to do those myself and let the car deal with the rest.

 

[Freakyguy666]

Well the 5 year comparison is not accurate, it is only now that Tesla has started uncovering the true potential of HW3 with neural net (aka AI) learning algorithm. They had to do a complete software rewrite, and the effort is starting to bear fruit. I have had my M3 for over a year now, the progress that I have seen over this one year is remarkable. My drive to office is about 55 miles, 45 miles of it is highway driving on I35 in Texas. The M3 works like a charm, getting on and off ramps, merging in toll lanes, managing lanes in construction zones. Now for my office commute I have reached I point where I trust the car more than my own driving, as it is able to anticipate and react to drivers ahead suddenly breaking better than what I do. Well I would admit I do take over 2-3 times during the drive, not because I feel unsafe, but because I know that the autopilot is going to drive more conservatively and merge into a lane-change 2 miles before, whereas I would merge in the last 0.25 miles, like any other driver. I also take over sometimes for the pure thrill of driving fast.

For the past few weeks I have also tried the stop sign and intersection feature, again to a person who doesnt own or drive a Tesla, the feature might feel stupid that the car stops irrespective of whether the light is green or red. But based on my driving experience, I actually feel even more confident, as I am sure that the car is going to slow down, it ensures that I am engaged as a driver. I just give it a tap of the foot and it keep going. I know that I am teaching the car to learn and not in too distant future it will be able to do it itself.

I feel that the money that I paid for FSD is worth it, as my mental state after my commute is totally relaxed now. My family tells me that I am really much more pleasant after returning from work. Now that is totally worth the 7K.

Also I feel feature-complete FSD is a misnomer. Are we as humans feature-complete in driving? Do we make no mistakes? I feel the saying applies perfectly here "that perfection is the enemy of progress". Full autonomy is a distant dream, I totally agree. But well for my daily commute I have two unprotected left turns on intersections for my whole commute, I am more than happy to do those myself and let the car deal with the rest.

what you described can be achieved with Enhanced AP. No need for FSD.

 

[Webeevdrivers]

Here here. In the UK everything has been dumbed down, exactly as you say. Summon rarely works, smart summon is disabled. Self parking in perpendicular spaces is a joke and if there is parked car on the side of the road, it slams on the brakes and don’t get me started on push bikes! Don’t get me wrong I absolutely love this car but feel, like you, I’ve been cheated by the FSD upgrade!

That kinda sucks. Here in Canada our car does exactly what the published aggreement said it would do. No surprises. Looking forward to when it stops at stop signs and lights etc. But none of that was in the Tesla Statement that I read and agreed to. I’d be interested in seeing what the UK documents say.

Also here you don’t have to have a death grip on the steering wheel or shake it every 10 seconds. Just a slight amount of torque or the occasional volume control flick does it. Different regions, different sensitivity settings I guess.

 

[smatthew]

I believe we are still a ways off. My up to date M3 with 3.0 cannot make a decent 90 degree right hand turn to follow A/P directions on a secondary road. It isn't ready for city FSD driving!

Tesla has never said your car can take a right hand turn yet. Occasionally autopilot will do it, but it's not an advertised feature at this time.

 

[Nirmaljal]

what you described can be achieved with Enhanced AP. No need for FSD.

Well I agree that it is called enhanced AP and that FSD includes Smart Summon among many other features. We can argue about labels and keep peeling at these names that Tesla keeps coming up with. But the underlying framework move to neural net learning changes the game. Based on such neural net learning the algorithm is better able to label and mark scenarios and therefore learn to handle complex driving scenarios. Throw millions of scenarios at it and it will learn to deal with these, be it left turns, or two-way, four-way stop signs. That is my take based on experience of driving the car. It just keeps improving. I have never been excited at owning any other car more than a month out, but for M3 it feel like I feel is love just yesterday.
I think a major challenge for Tesla before rolling out true FSD, even if it is ready, is the lack of legal/legislative framework. That would be a bigger challenge than the technology being ready.

 

[Actaeon]

"I mean - Ok, stop signs. Great. In 5 years development timeline? How long it will be before Tesla releases unprotected left turns on intersections? Another 5 years?
Bottom line - my prediction is Tesla will never be able to release FSD, even in USA, on HW3.0 and current sensor suite.
Now I guess 99% of the forum would disagree with me. I’m curious - is there anyone else who agrees with my prediction?"

Couldn't agree more. And who are all these Tesla Kool-Ade drinkers?

And after ALL of this "development," we finally get stop light recognition. With HERKY-JERKY stops?! What's up with that?!

 

[Actaeon]

Tesla is building a package for FSD and waiting for legislation etc to catch up.

Is Tesla really WAITING for legislation!!! The product should be SO good and SO advanced that Tesla should DRIVE the legislation. NOT the other way around.

I'm not falling for the "waiting for legislation" excuse for Tesla failing (badly and repeatedly) to perform. Class action lawsuit for deceptive advertising, anyone?

 

[LisaW]

I concur with the opinion that FSD is not going to happen any time soon, but for a different reason. The sensor suite has dozens or hundreds problems with edge cases, but that's not the real impediment.

1. The car relies on GPS to know where it is, and GPS is only accurate to around 30 feet / 10 meters. That's not accurate enough. The car does not know exactly where it is.

2. The car relies on the map to know where it's supposed to be, and I've personally seen hundreds of map errors where the roads on the map don't match the actual roads. Most of this is in construction areas, but construction areas are a reality -- they exist, and they're not infrequent. The maps don't get updated often enough, in some cases months or years. That's not good enough, you need the map accurately updated in a matter of hours. Therefore, the car doesn't know exactly where it is supposed to be.

If the car doesn't know either where it is or where it is supposed to be, it is impossible for it to do anywhere near full self driving, no matter how good the sensor suite or the algorithms are. You could have the best algorithm and sensor suite on the planet, but if the car thinks it's on the service road instead of the freeway's right-hand lane due to a map or GPS error, you're doomed. (And I've seen both cases, repeatedly).

I see no easy way to solve either issue. There are some hard ways to solve both issues, however, and Elon does like to do things the hard way, so who knows?

Maybe when starlink is working well at top capacity that would make a difference...would our cars be online via starlink? Maybe...
Lisa

 

[DS369]

Read this and tell me how it changes your view: The End of Starsky Robotics

From the article:
"In fact, the better your model, the harder it is to find robust data sets of novel edge cases. Additionally, the better your model, the more accurate the data you need to improve it. Rather than seeing exponential improvements in the quality of AI performance (a la Moore’s Law), we’re instead seeing exponential increases in the cost to improve AI systems — supervised ML seems to follow an S-Curve."

They clearly ran into lack of data problem. That's why Tesla is ahead, that's why Tesla will this race. Tesla will be the only one running it.

He also talks about Teleop, which will be a real possibility when Skynet... sorry, Starlink will become operational. A real possibility, cause right now, there are still too many dead zones, even in NA.

 

[fvanson]

I see some challenges with current hardware which is important for global FSD implementation; current camera’s are unable to deal with blinding due to low sun. This applies to a great portion of the 360 degrees the camera’s are covering. As a human, you can put an object like your hand or a mirror or whatever to cover the sun so that you can make something out of the image, but I do not believe the current camera system is able to perform such a thing. Maybe an upgraded camera system which can blind of an extremely bright object like the sun, like a kind of LCD matrix before the lens, is able to resolve this problem.
When I then look to the region I live in being Europe, there are practical issues. Most prominent one is traffic lights. Where as in North America, traffic lights are at the opposite side of the street you want to pass, in Europe the traffic lights are hanging above or slightly past the stop line (so that the same side of the street which you want to pass). This means that when you are the first car and are standing at the stop line, there is a good chance that the front cameras are not able to see the traffic light. Now there is technology to get the traffic light signal, as modern traffic lights are able to provide this information via an internet connection (this is already demo’d in radar / speed check software like FlitsMeister), but this will become an issue in Europe. And stopping way before the traffic light / stop line has the disadvantage that you do not cross the detection loops in the road to signal the traffic light system that a vehicle is waiting for a red light.
Also the deviating road signs in different countries of Europe, the differences in road lining for exits (which is now already a kind of kamikaze exercise in Belgium as they sometimes abrupt exit road lines, road sign recognition (like that you are on a road where you need to give priority to the vehicles on the road you want to cross on turn onto, difference in how priority has to be given on for instance roundabouts across Europe, how to deal with the multiple connected roundabouts in the UK, and what more.

So I have especially for Europe my worries that my Model 3 will ever become 100% FSD during the 8 years I plan to own it. Maybe it will happen for the US, as that is Tesla’s home market. For Europe I hope that the capability further increases so that it becomes a better and more capable advanced driver assistance system for on the highway as there I use it predominantly - especially now that maximum highway speeds are reduced to 100 km/h (from 130 km/h).

 

[royleeepp]

Wow, surprised that so many people actually agree, at least to some degree.

A few related thoughts:

1. Making FSD a subscription as opposed to an option will allow Tesla to release FSD in limited areas, which will allow staged deployment. Think about it. If you buy it as an option, it has to work ANYWHERE in the country you drive. If you sell it as a subscription, Tesla can geofence certain areas, and people will be fine with it (as they will be able to cancel subscription once moved from specific geofenced and well tested area).

2. Tesla is looking for some "shortcuts" to make it work. Let me explain. Some years ago, Tesla blamed other auto-manufacturers for using shortcuts in a form of Lidar. Here is the exact quote (Elon Musk: “Anyone relying on lidar is doomed.” Experts: Maybe not): "Lidar is really a shortcut," added Tesla AI guru Andrej Karpathy. "It sidesteps the fundamental problems of visual recognition that is necessary for autonomy. It gives a false sense of progress, and is ultimately a crutch."

It definitely makes sense, if your goal is to advance specialized AI. However, if you want to deliver FSD as fast as possible, you want to take ALL shortcuts you can have, because of you don't know when AI will be that good so no shortcuts will be needed. You can deliver your product faster with shortcuts, and then remove shortcuts when you don't need them.

Now, Tesla is taking its own shortcut - using maps and related info. If you truly want "no shortcuts" solution, as Andrej argued, you should make Tesla drive without maps at all, at least they should NOT be mandatory for Tesla to drive. Which is clearly not the case. Earlier released smart summon heavily relates on maps of parking lots. And the recent FSD stop signs update will not even activate unless you updated maps, so it knows all the info it needs about intersections, etc. Basically, Tesla can't even detect intersection by itself. It has to know that the intersection is there, and only knowing that it can then try to identify stop signs/traffic signals status. Obviously, Tesla can't trust its software to even detect intersections, so it has to use the shortcut.

All in all, I think Tesla realized it needs some shortcuts, which I'm fine with (maps). It will give us the product faster (at least in US). But don't expect any breakthroughs or something even close to ability for Tesla to drive in areas like NYC. Again, that would be difficult to admit that shortcuts are Ok and accept Lidar (I don't see anything wrong in allowing my car to have better senses than I do), so that means using maps and possibly geofencing once FSD subscription is deployed. Subscription will remove lots of liability from Tesla, by not having the "FSD delivery promise" permanent. And we, current FSD owners, will quite likely get something from Tesla for our patience.

Regarding shortcuts, I think maps are a temporary crutch until the AI can learn on its own. When you drive to someplace new, you drive more cautiously until you learn the nuances of the route, e.g. unexpected stop sign just around a curve, etc. The AI should approach the route just like a human and “learn” the route, but right now I don’t think it can just yet, hence the maps.

Also, I think the Tesla team is small and the number if edge cases is overwhelming, so I think the approach they are implementing now is to automatically identify (shadow AI says “stay left” but driver actually does a “stay right” situation, etc), auto label that situation and collect more situations like that automatically for training. Then with a million cars on the road collecting this data and automatically transforming it into learning data, the system will build itself to almost perfection without too much Tesla team interaction.

I hope they are successful.

 

[KeithT]

The more I see how slow Tesla is at delivering new FSD features, the more I think it will never happen. Assuming current HW3.0 and sensor suite.

I mean - Ok, stop signs. Great. In 5 years development timeline? How long it will be before Tesla releases unprotected left turns on intersections? Another 5 years?

Bottom line - my prediction is Tesla will never be able to release FSD, even in USA, on HW3.0 and current sensor suite.

Now I guess 99% of the forum would disagree with me. I’m curious - is there anyone else who agrees with my prediction?

Charles Dickens is honoured as popularizing "never say never" in The Pickwick Papers. The main point seems to be the perceived number of edge cases that have to be solved and that that could take forever!

But it is not the case that FSD has to be perfect just that it is an order of magnitude better than the average homo sapiens. Regulators are not going to bar FSD if it is demonstrable that it cuts the accident rate to a small fraction of current.

On the AI side I see no reason to assume that the sensor suite is lacking. The software can create the equivalent of a Lidar image at a fraction of the cost and already Autopilot is better than me at driving on a highway mainly because it doesn't get distracted.

And AI development gets to tipping points when progress becomes exponential. Patience.

 

[Mo City]

So I have especially for Europe my worries that my Model 3 will ever become 100% FSD during the 8 years I plan to own it. Maybe it will happen for the US, as that is Tesla’s home market. For Europe I hope that the capability further increases so that it becomes a better and more capable advanced driver assistance system for on the highway as there I use it predominantly - especially now that maximum highway speeds are reduced to 100 km/h (from 130 km/h).

"100% FSD" is never going to happen, anywhere in the world. Nobody will develop autonomy to a level where the vehicle can handle all circumstances on all roads. That is a complete pipe dream.

 

[smatthew]

"I mean - Ok, stop signs. Great. In 5 years development timeline? How long it will be before Tesla releases unprotected left turns on intersections? Another 5 years?
Bottom line - my prediction is Tesla will never be able to release FSD, even in USA, on HW3.0 and current sensor suite.
Now I guess 99% of the forum would disagree with me. I’m curious - is there anyone else who agrees with my prediction?"

Couldn't agree more. And who are all these Tesla Kool-Ade drinkers?

And after ALL of this "development," we finally get stop light recognition. With HERKY-JERKY stops?! What's up with that?!

You have herky jerky stops? Maybe you need to stop talking bad about your car while you're in it. Mine stops super smoothly.

 

[smatthew]

"100% FSD" is never going to happen, anywhere in the world. Nobody will develop autonomy to a level where the vehicle can handle all circumstances on all roads. That is a complete pipe dream.

Dude - you can't handle all circumstances on all roads!

 

[Sail049]

The more I see how slow Tesla is at delivering new FSD features, the more I think it will never happen. Assuming current HW3.0 and sensor suite.

Now I guess 99% of the forum would disagree with me. I’m curious - is there anyone else who agrees with my prediction?

It depends on your viewpoint of “never”. I am 67 years old, and I would put money on it that my Model 3 will “never” allow me to sit in the back seat and take me where I want to go. That’s why I didn’t give Elon a $6000 interest free loan when I bought the car. Now thirty-somethings may see it, so “never” may not be true in that case. IMHO, summon, lane changes, and TL and stop signs still isn’t worth $6000, never mind the $7000 it is now, and more likely in the near future. And EAP doesn’t work well on secondary roads on top of it all. I’m happy with it standard for use on long trips, free. Now that was a good deal.