Elon: "Feature complete for full self driving this year"

[S4WRXTTCS]

This is precisely why I leased my M3. I don't think Tesla or anyone else knows enough at this point to truly understand what sort of hardware is going to be required to even reach the goal of true NOA in the city. I'm fairly confident that we'll see some rudimentary form of NOA in the city during 2020 but I'm not confident at all as to when we will see it being truly useful and/or close to how a human would drive.

I'd be SHOCKED if we don't see new hardware within the next few years that will be required to truly get good autopilot functionality in all scenarios.

I'll be shocked if the NHTSA allows for any kind of extensive NoA in the city with the lack of any kind of driver monitoring capability.

It's really going to come down Tesla either releasing a City NoA so terrible that no one would dare trust it or releasing a NoA that is so good that the human having brief moments of distraction won't be too big of a deal.

If it's somewhere in the middle, and incidences start to pile up like they have been lately for AP then it's all over. Until Tesla revamps the Driver Monitoring to switch to a hands free system.

I expect HW4 to have hands free driver monitoring similar to supercruise, rear/corner radars, and 360 degree downfacing parking cameras. Plus some improvements to the rear camera to make it better in the rain.

Those are the things HW3 is missing, and its really apparent in the glitches we've seen.

 

[boonedocks]

I'll be shocked if the NHTSA allows for any kind of extensive NoA in the city with the lack of any kind of driver monitoring capability.

It's really going to come down Tesla either releasing a City NoA so terrible that no one would dare trust it or releasing a NoA that is so good that the human having brief moments of distraction won't be too big of a deal.

If it's somewhere in the middle, and incidences start to pile up like they have been lately for AP then it's all over. Until Tesla revamps the Driver Monitoring to switch to a hands free system.

I expect HW4 to have hands free driver monitoring similar to supercruise, rear/corner radars, and 360 degree downfacing parking cameras. Plus some improvements to the rear camera to make it better in the rain.

Those are the things HW3 is missing, and its really apparent in the glitches we've seen.

well.....if CityNOA operates as poorly as HighwayNOA does for me today I won’t be using it. As is I can’t use NOA daily now without disengaging it repeatedly on the same spots every commute and then when it hits the 4 lane transition from HOV/PeachPass to the interchange change it is a true traffic hazard.

 

[diplomat33]

I expect HW4 to have hands free driver monitoring similar to supercruise, rear/corner radars, and 360 degree downfacing parking cameras. Plus some improvements to the rear camera to make it better in the rain.

A driver monitoring camera is a must-have in my opinion if you want to do hands-free driving.

Personally, I'd probably add front and rear lidar too to help the cameras with distance calculations and stopped objects. But I think your list makes a lot of sense as an incremental step that Tesla could take with HW4.

 

[daniel]

Yes, but we all know that at some point, if you want a real robotaxi which Tesla claims to want, then FSD requires that the car be able to actually stop on its own at a red light and then go again at a green light with no input from the driver. I am sure Elon has even talked about it at some point. So yeah, just flashing a warning when you are about a run a red is technically a response, but it can't the be the final response that Tesla is actually planning. It is only a preliminary response, a prerequisite for something more.

While it's obvious there'd eventually be more functionality we need to be clear of something here:

what was promised (in the most narrow sense) was already delivered with the warnings. ...

Tesla and Elon have made different promises at different times. In Post #2004, willow-hiller quoted Musk as promising that before the end of 2019, FSD-equipped Tesla cars would be able to drive you from home to work and back (implication: highway and city NoA) with driver supervision but only very rarely requiring driver intervention. There is no way this could happen without the ability to stop for red lights and go for green lights without driver intervention, as well as stopping for stop signs and then going when it is safe to do so, based on cross traffic and left-turning traffic.

He may have promised at some time that the car would "respond," and issuing a warning is certainly a response. But he also promised that NoA city and highway would only very rarely require driver intervention, before the end of 2019.

Clear promise, clearly broken.

OTOH, it's good that Tesla would rather break Elon's promises than release features that are not yet safe to use. I love EAP. I see no reason to pay for FSD now. Especially since it would likely be a year or two before I got the hardware, and if that has to be done at a service center, it would mean shipping my car to Oahu. I rather trade in my car if I'm still alive when real FSD becomes available.

 

[diplomat33]

I'll make a prediction that Tesla will achieve some version of autonomous driving with driver supervision with HW3 but will stumble at improving the reliability enough to remove driver supervision. They will keep encountering driving cases, especially in city driving, that HW3 simply can't do, no matter how they try to tweak the NN with more machine learning. So they won't be able to remove driver supervision. Tesla will start realizing that they need higher resolution cameras, or an extra radar or an extra camera, etc... At that point, Tesla will start changing the sensors to solve those driving cases that HW3 can't do.

 

[verygreen]

In Post #2004, willow-hiller quoted Musk as promising

Here's your problem, you listen to what Musk promises. But in the order contract that you sign it's specifically stated that all outside promises are null and void and only things that are part of the contract are promised.

And that's been pretty much Tesla's stance in all cases I am aware of. "if it's not in the manual it was not promised" (like the on-ramp to off-ramp and meet you at doorstep summon on AP1)

 

[Daniel in SD]

"if it's not in the manual it was not promised"

What about the website? Seems like those promises would be binding.

 

[verygreen]

What about the website? Seems like those promises would be binding.

the order contract says that everything outside it is non-binding. so... Want to test it in court? Please let us know how it goes (after you clear mandatory arbitration first too of course)

 

[S4WRXTTCS]

Here's your problem, you listen to what Musk promises. But in the order contract that you sign it's specifically stated that all outside promises are null and void and only things that are part of the contract are promised.

And that's been pretty much Tesla's stance in all cases I am aware of. "if it's not in the manual it was not promised" (like the on-ramp to off-ramp and meet you at doorstep summon on AP1)

View attachment 499819

The problem with how it's stated is it also supersedes due bills, and Tesla loves using due bills.

I'm also skeptical that the contract would even go in-depth into what a given feature was supposed to do. Just like it didn't go in-depth with the HP of the P85 or lots of other things Tesla has been sued, and settled over.

EAP time frames is another one. I doubt the time frame was in the Contract, but it was on the website. So they got sued, and settled.

For me personally I think the contract, and the website are pretty darn binding within reason. Like I don't care about technical details like 4 active cameras with EAP, or 8 activate cameras. But, I certainly care about having Smart Summon with EAP.

I don't consider much of anything Elon says on twitter to be binding.

In fact I think Tesla should use the Tesla twitter more to state things that are actually promised. Where Elon's twitter can remain the fanciful fairytale thing it's always been. There needs to be a way for end customers to separate ideas Tesla is working on versus actual promises.

 

[Daniel in SD]

the order contract says that everything outside it is non-binding. so... Want to test it in court? Please let us know how it goes (after you clear mandatory arbitration first too of course)

I didn’t buy FSD. I’m sure some 2016 lessees are going to be testing it soon.

 

[verygreen]

The problem with how it's stated is it also supersedes due bills, and Tesla loves using due bills.

nope. Reportedly they stopped doing due bills some time ago.

 

[S4WRXTTCS]

nope. Reportedly they stopped doing due bills some time ago.

I'm just going off my experience which was back in September of 2018.

In any case that line is it pretty common in contracts as a way to protect both parties from some prior agreement that is null and void with the new agreement. Where you're supposed to have everything promised on the new agreement. My interpretation is it's near the top of the list of terms to have on a contract.

I'm not a lawyer, but I suspect there are limits to how it's to be interpreted and enforced.

It didn't help Tesla when they were sued over the HP claims of the P85.
It didn't help Tesla when it came to EAP taking too long, and having to settle it.

It's impossible to put ALL the promises made to a customer into the contract, and all the details of how something is supposed to perform.

The areas where it does likely work really well is for intended purpose of that line, and that's to protect against sales people making promises that aren't part of the agreement.

I'd love to hear an actual lawyers interpretation of though.

 

[verygreen]

I'm just going off my experience which was back in September of 2018.

I know it was big back then. But apparently not anymore.

 

[im.thatoneguy]

They're no longer promising Level 5.

I don't believe they ever promised "Level 5". Level 5 SAE™ is an abomination of a "Standard" and the SAE™ should be embarrassed that they still have it. At the remote chance that I look like an idiot in 20 years, I'm going to go ahead and predict that nobody will ever ship a Level 5 SAE™ product... ever.

Level 5 SAE™ means "anywhere a human can drive". Level 5 SAE™ means a self driving package can transit a narrow gravel mountain pass, pull into the town, navigate through a ferry terminal line, complete a transaction with the ticket counter and then find the right lane to park and wait in, then board the correct ferry, park... wait... get off the ferry, drive to the train station, buy a ticket for the train. Wait in the correct lane, get onto the correct rail car, park, wait, get off the train at the correct stop, drive offroad again through a rough 4x4 track to a mountain airstrip and drive onto an airplane that then flies it to another continent where it then exits the airplane and drives through 2' deep snow, using edge of road posts to orient to reach a mountain cabin.

That's never going to happen. Nobody is ever going to bother implementing or testing half of those tasks because they aren't worth the development time for the 1 in a billion road trip that no vehicle has ever actually taken (but a human driver could execute). We'll get Level 4 SAE™ sometime in the next decade but Level 5 SAE™ is dumb it's just a catch-all for "Effectively Perfect".

Even if there was a desire for fully autonomous cars to handle rail and ferry loading/unloading... we're going to need a way to bill the cars autonomously and communicate with them on where they need to go. That means we'll probably develop autonomous vehicle lanes that streamline the process so that Autonomous Vehicles never need to learn the vagueries that even as a human with general intelligence I can often barely comprehend without talking. Society will adapt to autonomous cars faster than autonomous cars adapt to society. The most difficult edge cases that human drivers can handle, will be filed off and smoothed over before it becomes necessary for engineers to actually train cars to handle them. It'll be like finding someone who knows how to make stone knives. It would be like teaching an AI to drive a stick-shift. Plenty of human drivers can do it, but I highly doubt we'll ever have an autonomous vehicle with a manual transmission.

(And yes, I've gotten close to that road trip in the past. I didn't get onto an airplane but I did cross two 4x4 mountain passes that had wooden planks for bridges and boarded a teeny tiny little 8 car ferry, so it's not an impossible route). The correct response from an autonomous car is "Find a different route."

 

[S4WRXTTCS]

I don't believe they ever promised "Level 5". Level 5 SAE™ is an abomination of a "Standard" and the SAE™ should be embarrassed that they still have it. At the remote chance that I look like an idiot in 20 years, I'm going to go ahead and predict that nobody will ever ship a Level 5 SAE™ product... ever.

Ha, Apparently I'm not the only one that mocks SAE for the Level 5. They probably mock themselves, and they'd probably redo the entire thing if they had a do-over.

Level 3, and Level 5 would probably disappear and get replaced with something else.

Anyways the only one I'm aware of that made the L5 promise was Elon. Which makes me question whether he really understands the SAE Levels.

 

[Az_Rael]

I don't believe they ever promised "Level 5"

It is funny, when you go back and read what they did promise when I bought, I am not sure it is even Level 4 and maybe not Level 3 when I read with a more Tesla critical eye today.

It says the car will take trips with no action from the driver, but notably doesn’t say the driver won’t need to pay attention during those trips. It only says that you won’t have to intervene during a trip at a level twice as safe as the average human. This might explain Elon’s odd comment in an investor call that “they already released highway FSD” in reference to NoA. If NoA plus city NoA gets good enough, it might meet the below statement (trip with no intervention) while still being a Level 2 system where you have to be ready to intervene at any time. Granted, the below does imply hands free operation, but they always have the “regulation” card to pull out for that part.

 

[Mardak]

I'll make a prediction that Tesla will achieve some version of autonomous driving with driver supervision with HW3 but will stumble at improving the reliability enough to remove driver supervision. They will keep encountering driving cases, especially in city driving, that HW3 simply can't do, no matter how they try to tweak the NN with more machine learning.

It seems a bit premature to predict what HW3 can't do as Tesla has previously only deployed NNs and driving behavior that are limited by HW2/.5 capabilities, and this behavior was using only 5% load on HW3 without redundancy. If Tesla wants to maintain the same 20% load buffer as now while adding redundancy, the NN on HW3 could handle 700% more computations.

One "simple" way to use the additional compute is to "add more layers" where if we say double the size of the network for double the compute load, and the accuracy of the network increases without needing to write smarter / complex custom code. The trade-off is that this requires even more data to avoid over-fitting, however this plays into Tesla's strengths of potentially gathering a lot of training data automatically.

Another use of extra compute is more outputs, e.g., identify more objects or predict more behaviors. Given that verygreen videos with the old hardware already shows quite a bit of complex output, e.g., driveable space, bounding boxes with type/position/distance/velocity, lane lines, path prediction; doubling outputs would be similar to adding a completely new feature as good as existing highway Autopilot. Perhaps Karpathy showing off 3d reconstruction from video is something they're looking to do from vehicles in real-time.

And so far the above is focusing on increasing neural network "software 2.0" capabilities, but there's plenty of features that can be initially implemented with traditional "software 1.0" like how Navigate on Autopilot was introduced (and still behaves to some people's dissatisfaction, e.g., relying on map data for yielding instead of visually). Elon Musk at Autonomy Day seemed to describe experiencing already implemented FSD capability for making a right turn that clearly hasn't been deployed to users yet:

So if you’re turning onto a road that's got a lot of a lot of high-speed traffic, you can just do what a person does: just gradually turn a little bit, don't go fully into the road, let the cameras see what's going on, and if things look good and the rear cameras don't show any oncoming traffic — off you go. And if it looks sketchy, you can just pull back a little bit just like a person The behaviors start to become remarkably lifelike. It's quite eerie actually. The car just starts behaving like a person.

 

[diplomat33]

Level 5 SAE™ means "anywhere a human can drive". Level 5 SAE™ means a self driving package can transit a narrow gravel mountain pass, pull into the town, navigate through a ferry terminal line, complete a transaction with the ticket counter and then find the right lane to park and wait in, then board the correct ferry, park... wait... get off the ferry, drive to the train station, buy a ticket for the train. Wait in the correct lane, get onto the correct rail car, park, wait, get off the train at the correct stop, drive offroad again through a rough 4x4 track to a mountain airstrip and drive onto an airplane that then flies it to another continent where it then exits the airplane and drives through 2' deep snow, using edge of road posts to orient to reach a mountain cabin.

That's not what L5 is. For one, buying tickets is not a DDT so it is not required for any autonomous car to do. Second, I don't believe L5 includes off road driving.

 

[diplomat33]

It seems a bit premature to predict what HW3 can't do as Tesla has previously only deployed NNs and driving behavior that are limited by HW2/.5 capabilities, and this behavior was using only 5% load on HW3 without redundancy. If Tesla wants to maintain the same 20% load buffer as now while adding redundancy, the NN on HW3 could handle 700% more computations.

One "simple" way to use the additional compute is to "add more layers" where if we say double the size of the network for double the compute load, and the accuracy of the network increases without needing to write smarter / complex custom code. The trade-off is that this requires even more data to avoid over-fitting, however this plays into Tesla's strengths of potentially gathering a lot of training data automatically.

Another use of extra compute is more outputs, e.g., identify more objects or predict more behaviors. Given that verygreen videos with the old hardware already shows quite a bit of complex output, e.g., driveable space, bounding boxes with type/position/distance/velocity, lane lines, path prediction; doubling outputs would be similar to adding a completely new feature as good as existing highway Autopilot. Perhaps Karpathy showing off 3d reconstruction from video is something they're looking to do from vehicles in real-time.

And so far the above is focusing on increasing neural network "software 2.0" capabilities, but there's plenty of features that can be initially implemented with traditional "software 1.0" like how Navigate on Autopilot was introduced (and still behaves to some people's dissatisfaction, e.g., relying on map data for yielding instead of visually). Elon Musk at Autonomy Day seemed to describe experiencing already implemented FSD capability for making a right turn that clearly hasn't been deployed to users yet:

I have no doubt that Tesla can add a lot of capabilities with software and will. I just suspect that there might be some cases that require additional hardware, not to do, but to do well. Remember, I am not talking about merely doing a feature but about doing the feature well enough that the driver is not required. I am sure there are a lot of features that Tesla will figure out a way to do on the HW3 hardware. That is why I said that HW3 will probably do some form of autonomous driving. The question is can it do it reliably well enough that the driver is not required.

And just look at the changes from AP2 to AP2.5 and from AP2.5 to AP3. Tesla has already made tweaks to the hardware. So I think it is evident that Tesla will change the hardware as needed to get to FSD.

I would add that Tesla seems to be taking a minimalist approach to FSD hardware. They are starting with the minimum amount of hardware, trying to squeeze as much FSD as they can out it, and only changing the hardware when it is absolutely strictly needed. This approach makes it harder to get to autonomous driving but it has the advantage of keeping costs down and when you do achieve safe autonomous driving, you will already have it on the cheapest, most efficient set up.

 

[willow_hiller]

For all of the healthy skepticism of Tesla's optimism around the HW3 hardware achieving any higher level of autonomy, a thought did occur to me this morning.

If you want to buy a car today that has a possibility of being autonomous at some point in the future, I think Tesla is the only option. It may not be a big possibility, but given how inflexible so many other car operating systems are, Tesla is the only one that offers any sort of feasible upgrade path.