Elon: FSD Beta tweets

[Goose66]

These posts are so funny. You're actually complaining that Tesla have not prioritised a goal of your choosing that they never had and never claimed to have.

'Tesla want something different than what I want, and it's their fault I bought their product so now I'm angry'.

First off, I don't know where you get "angry" from. But maybe believing in Elon's flights of fancy so whole heartedly as you do you've forgotten that there's another viewpoint out there - one that's driven by common sense. Many of these posts that you think are "so funny" may just be from folks who listen to a broad range of voices on the subject over the years and aren't simply Elon sycophants. You're so quick to say "Elon always said they were going straight to L5," "Elon specifically said in the past that he's not interested in L3," etc. First, I suggest anyone that has listened to Elon on the subject over the last 7 years cannot rationally claim to understand what Elon did or did not know about what FSD was going to be or when it was coming. I mean, he had said "autonomous driving was solved," "every Tesla manufactured after 2016 has all the hardware needed to be fully autonomous," "relying on LIDAR is a fools errand," etc. (I could go on and on). The guy has a horrible track record on understanding what the state of the art is in autonomous driving or where it is going.

On the other hand, many experts in the field all the way back in 2014 and earlier understood that there was a bifurcation in the autonomy efforts, with some working to push their ADAS systems into autonomy while others were building fully autonomous vehicles from the ground up. I think every logical person (including Tesla's legal department) would agree that Tesla is in the former group. And it is rather clear to most who have followed autonomous vehicle development that those in the former group would very likely pass through L3 - even if just on limited access highways and the like - on the way to higher levels of automation. It was simply common sense. So I don't find it "funny" at all to believe that many in 2016-2018 when purchasing FSD thought that at some point what they would have is an L3 highway driving system (hands-off NoA), even if it would eventually be followed by higher levels of automation.

Further, having driven with the latest automation products and "betas" that Tesla has to offer for over 5 years, having done work with other autonomous vehicle software/hardware providers, and having read extensively on the subject for nearly a decade, I am very comfortable in saying that had Tesla determined early on that a 4D radar and comprehensive mapping data would be beneficial, they could have evolved FSD (or NoA) into a fully L3 system by now - one that would have passed regulator scrutiny - and could have been the first to market with a true L3 system that covered 99% of the Country's limited access highway. And I would have liked that. As a feature of a consumer car this has real utility for me. It is a useful feature that reduces workload, stress, and fatigue on long trips and would improve highway safety. Taking me to work in the morning on surface streets? Why would I want that. I like driving my own car and the 15 minute commute through multiple twisty and turny intersections and back roads is the perfect place to drive it.

But instead, Tesla seems to be putting all their effort into creating a system for a driverless robotaxi - Elon's even said multiple times now that the company's financial future is dependent on it. Talking about a fool's errand. I don't hear ANY other voices in the industry suggesting that Tesla could possible have an L5 fully autonomous system with their current hardware. NOBODY believes it but Elon. And the one guy who could have made it happen if anybody could is long since gone. So you keep believing it if you want to, but the vast majority of us that have been in it for the long haul and who know are moving on.

 

[Mullermn]

First off, I don't know where you get "angry" from. But maybe believing in Elon's flights of fancy so whole heartedly as you do you've forgotten that there's another viewpoint out there - one that's driven by common sense. Many of these posts that you think are "so funny" may just be from folks who listen to a broad range of voices on the subject over the years and aren't simply Elon sycophants. You're so quick to say "Elon always said they were going straight to L5," "Elon specifically said in the past that he's not interested in L3," etc. First, I suggest anyone that has listened to Elon on the subject over the last 7 years cannot rationally claim to understand what Elon did or did not know about what FSD was going to be or when it was coming. I mean, he had said "autonomous driving was solved," "every Tesla manufactured after 2016 has all the hardware needed to be fully autonomous," "relying on LIDAR is a fools errand," etc. (I could go on and on). The guy has a horrible track record on understanding what the state of the art is in autonomous driving or where it is going.

On the other hand, many experts in the field all the way back in 2014 and earlier understood that there was a bifurcation in the autonomy efforts, with some working to push their ADAS systems into autonomy while others were building fully autonomous vehicles from the ground up. I think every logical person (including Tesla's legal department) would agree that Tesla is in the former group. And it is rather clear to most who have followed autonomous vehicle development that those in the former group would very likely pass through L3 - even if just on limited access highways and the like - on the way to higher levels of automation. It was simply common sense. So I don't find it "funny" at all to believe that many in 2016-2018 when purchasing FSD thought that at some point what they would have is an L3 highway driving system (hands-off NoA), even if it would eventually be followed by higher levels of automation.

Further, having driven with the latest automation products and "betas" that Tesla has to offer for over 5 years, having done work with other autonomous vehicle software/hardware providers, and having read extensively on the subject for nearly a decade, I am very comfortable in saying that had Tesla determined early on that a 4D radar and comprehensive mapping data would be beneficial, they could have evolved FSD (or NoA) into a fully L3 system by now - one that would have passed regulator scrutiny - and could have been the first to market with a true L3 system that covered 99% of the Country's limited access highway. And I would have liked that. As a feature of a consumer car this has real utility for me. It is a useful feature that reduces workload, stress, and fatigue on long trips and would improve highway safety. Taking me to work in the morning on surface streets? Why would I want that. I like driving my own car and the 15 minute commute through multiple twisty and turny intersections and back roads is the perfect place to drive it.

But instead, Tesla seems to be putting all their effort into creating a system for a driverless robotaxi - Elon's even said multiple times now that the company's financial future is dependent on it. Talking about a fool's errand. I don't hear ANY other voices in the industry suggesting that Tesla could possible have an L5 fully autonomous system with their current hardware. NOBODY believes it but Elon. And the one guy who could have made it happen if anybody could is long since gone. So you keep believing it if you want to, but the vast majority of us that have been in it for the long haul and who know are moving on.

I mean you do sound pretty angry.

I'm not trying to be rude, but aren't you doing exactly what I said in my previous post? You have projected this objective on to Tesla when they've never had it (as far as know - I'm not claiming to be familiar with every tweet or presentation), and you're holding it against them that they haven't delivered it. You may well be right about everything you've said above, but at the end of the day if they choose to try and create one product over another, that's their choice, isn't it? Just like it's our choice which product we buy.

I would also observe that until such time as someone brings out the first generalised L5 system (by 'generalised' I mean an ODD of significant scope) then all bets remain off on what tech is or isn't required.

Regarding the hardware, was a 4D radar of significant quality available 'early on'? How much would adding LiDAR to consumer vehicles cost? Last estimate I read was that Waymo's current hardware costs them 10k per vehicle, down from many times that in their earlier platforms. How many people would have bought HW3-era (2019) vehicles if there was a 15k (conservative estimate) price premium for hardware to support a feature which, to restate, NOBODY has successfully implemented yet? How far would Tesla have gotten towards their mission ('We're building a world powered by solar energy, running on batteries and transported by electric vehicles.') without those sales?

 

[spacecoin]

I would also observe that until such time as someone brings out the first generalised L5 system (by 'generalised' I mean an ODD of significant scope) then all bets remain off on what tech is or isn't required.

This will likely not be in our lifetime. It's also extremely unlikely that these future cars would not to use multiple sensor modalities even if it would be technically viable in the future only to use passive cameras. Hardware costs are dropping by a factor of 10x every 3-5 years.

Regarding the hardware, was a 4D radar of significant quality available 'early on'? How much would adding LiDAR to consumer vehicles cost?

Forward facing mmW radar and Lidars are less than $150 (radar) - $750 (lidar) these days. In 5-10 years Lidars will likely be less than $100.
They compliment cameras by working more reliably in some environments (darkness, fog, smoke, rain), and they do not rely on machine learning to measure distances, and have centimeter level precision with 100% recall (never misses an object). These things matter if you want to build a product that is so safe that you can take on manufacturer liability for driving.

Last estimate I read was that Waymo's current hardware costs them 10k per vehicle, down from many times that in their earlier platforms. How many people would have bought HW3-era (2019) vehicles if there was a 15k (conservative estimate) price premium for hardware to support a feature which, to restate, NOBODY has successfully implemented yet?

It's likely impossible to get to autonomy on existing cars in any wide ODD in the coming five years, or possibly ever.

 

[Goose66]

I mean you do sound pretty angry.

One man’s anger is another man’s clarity.

Am I disappointed? Yes. But it’s not like I drove my Tesla down to Elon’s tiny house at Boca Chica and did donuts in his yard. Elon and Tesla and SpaceX are doing crazy, wonderful things. But you must admit dude is certainly eclectic and can be compulsive and childlike in his decision making at times, and in this I think he’s indulged by the folks around him. Can he create a robotaxi with his current platform? I am extremely confident that the answer is no, but anything is possible. However, a wise automaker would have not have missed the chance to be first to deliver a real and useful automation system to their customers, increasing sales and solidifying customer loyalty. And that’s what L3 NoA would have done, IMHO.

 

[Mullermn]

However, a wise automaker would have not have missed the chance to be first to deliver a real and useful automation system to their customers, increasing sales and solidifying customer loyalty. And that’s what L3 NoA would have done, IMHO.

On a personal level, I agree with you entirely. I'm in the UK where FSD is only just in very early, semi-secret testing and I don't see it getting to the level it is in the US for some time as our road system is quite a lot more complicated than the US. I think they could do a L3 option for the same roads that NoA already covers fairly easily and that would be immediately useful to me.

 

[Mullermn]

This will likely not be in our lifetime. It's also extremely unlikely that these future cars would not to use multiple sensor modalities even if it would be technically viable in the future only to use passive cameras. Hardware costs are dropping by a factor of 10x every 3-5 years.

I agree, once you have self driving cars then adding more sensors on to them to make them more capable is an easy decision, even if it carries a modest extra cost. I do think that's different than getting consumers to agree to pay for those sensors in advance (in fact, many years in advance) of the time when they're actually useful though.

Plus, a lot gets made of the improved weather behaviour of LiDAR and radar on this forum, but since neither can detect visual markings it's not obvious to me that a car so equipped would actually be able to drive any better if it's cameras were blinded by weather, especially given the roads are almost entirely populated by vehicles driven by humans who don't have either.

Forward facing mmW radar and Lidars are less than $150 (radar) - $750 (lidar) these days.

I'm not sure that purely forward facing sensors bring much to Tesla's objective if they want full FSD. You can't pull out of a turning in to traffic if you can't see to your side.

This article (Waymo Reveals More Details On New Sensor Hardware That Could Indicate Imminent Mass Production) from 2020 says:

Costs​

When it comes to the exact cost of the hardware suite, Waymo continues to hold back, pointing out once again that the cost of LiDARs has been reduced by 90 percent. With a Velodyne LiDAR, as Waymo was using it just a few years ago before the in-house development of the hardware was started, one had to reckon with costs of 75,000 US dollars. It is estimated that a current sensor stack, as used by most off-the-shelf self-propelled technology developers, costs between $150,000 and $250,000 US dollars. For Waymo, the cost of an entire 5th generation stack could already be down to between $50,000 and $100,000.

I did also find this 2021 article (Waymo CEO Says Tesla Is Not A Competitor, Gives Estimated Cost Of Autonomous Vehicles) where the CEO of Waymo said it was more like $7,500 but also says that "If we equip a Chrysler Pacifica Van or a Jaguar I-Pace with our sensors and computers, it costs no more than a moderately equipped Mercedes S-Class". Given that the van seems to retail about $50,000 and the S class about $180,000 I'm not sure how those sums work.

In 5-10 years Lidars will likely be less than $100.

I can't say that you're wrong, but if so that's a hell of an advancement from what the article says, and it doesn't help Tesla circa 2018 and before when HW1-3 was being put together.

I'd also note that while Waymos look kinda stylish and neatly put together in their own way, there no way on Earth a random consumer is going to choose to buy a car that looks like the Tardis materialised in its sunroof when, again, that hardware doesn't currently deliver self driving. Remember that perhaps Tesla's defining achievement was taking cars that were previously ugly, slow and so limited in range as to be useless and turning them in to something people actually wanted to own.

Std disclaimer: None of the above is a criticism of Waymo's progress, they're simply the best comparison for the points under discussion.

 

[spacecoin]

I agree, once you have self driving cars then adding more sensors on to them to make them more capable is an easy decision, even if it carries a modest extra cost. I do think that's different than getting consumers to agree to pay for those sensors in advance (in fact, many years in advance) of the time when they're actually useful though.

Plus, a lot gets made of the improved weather behaviour of LiDAR and radar on this forum, but since neither can detect visual markings it's not obvious to me that a car so equipped would actually be able to drive any better if it's cameras were blinded by weather, especially given the roads are almost entirely populated by vehicles driven by humans who don't have either.

I'm not sure that purely forward facing sensors bring much to Tesla's objective if they want full FSD. You can't pull out of a turning in to traffic if you can't see to your side.

By forward facing I mean single direction, solid state, rather than 360 deg rotating.

There is a lot happening in the sensor space. Waymo's Drago Angelouv talked about cameras with built in active and passive lighting that are in practice both a lidar and a camera in an interview a while back. He didn't share much detail.

 

[Goose66]

By forward facing I mean single direction, solid state, rather than 360 deg rotating.

There is a lot happening in the sensor space. Waymo's Drago Angelouv talked about cameras with built in active and passive lighting that are in practice both a lidar and a camera in an interview a while back. He didn't share much detail.

The 4D radars utilized by NVIDIA’s Drive platform when integrated with cameras can be used to create impressive occupancy networks with similar resolution and accuracy to LIDAR but without the expense, and without the latency and errors inherent in a vision only occupancy networks like Tesla.

BTW, Tesla has been talking about 4D radar at least since 2019 along with their own patent filings in the technology and I’m pretty sure that’s what you are looking at in the front bumper of HW4 Model S and Xs.

 

[eli_]

The 4D radars utilized by NVIDIA’s Drive platform when integrated with cameras can be used to create impressive occupancy networks with similar resolution and accuracy to LIDAR but without the expense, and without the latency and errors inherent in a vision only occupancy networks like Tesla.

BTW, Tesla has been talking about 4D radar at least since 2019 along with their own patent filings in the technology and I’m pretty sure that’s what you are looking at in the front bumper of HW4 Model S and Xs.

There's a 2023 paper that uses a transformer-based encoder to get similar performance to segmentation models that use LIDAR, using only RGB vision: https://arxiv.org/pdf/2302.07817.pdf

Also I learned that everyone is bad at doing segmentation for bicycles, low scores for both LIDAR and vision-based approaches.

 

[Bladerskb]

Tesla's radar is nothing like Waymo & Mobileye's Ultra Resolution 4D Imaging Radar. The difference is staggering.

Here's an indepth look:

Tearing-down Tesla’s in-house radar design – Why did they bother? | Ghost

A radar engineer’s analysis of the new Tesla HW4 radar.

www.ghostautonomy.com

Quick comparison: You have Huawei's 4D imaging radar with 288 virtual channels.

Mobileye’s Ultra Resolution 4D Imaging Radars with 2,304 virtual channels and ~500k points per second. (0.5 angular res, 2 vertical res).

Arbe radar 2,304 virtual channels (1 angular res, 1.7 vertical res).

Fisker Ocean is using Uhnder's radar with 192 virtual channels.

Tesla's New Radar has 48 virtual channels.

 

[kabin]

Tesla's radar is nothing like Waymo & Mobileye's Ultra Resolution 4D Imaging Radar. The difference is staggering.

Here's an indepth look:

Tearing-down Tesla’s in-house radar design – Why did they bother? | Ghost

A radar engineer’s analysis of the new Tesla HW4 radar.

www.ghostautonomy.com

Quick comparison: You have Huawei's 4D imaging radar with 288 virtual channels.

Mobileye’s Ultra Resolution 4D Imaging Radars with 2,304 virtual channels and ~500k points per second. (0.5 angular res, 2 vertical res).

Arbe radar 2,304 virtual channels (1 angular res, 1.7 vertical res).

Fisker Ocean is using Uhnder's radar with 192 virtual channels.

Tesla's New Radar has 48 virtual channels..

It makes sense to have different types of sensors. Too bad that sounds like another half-hearted radar effort. It will be interesting to see if they tout its capabilities.

 

[DanCar]

Just more hype like quantum leap.

https://twitter.com/x/status/1671369154673123329

Quote:
Penny2x: The next few versions of FSD will be like watching the future unfold before our eyes.
Elon: It will feel like magic

 

[jeremymc7]

Quote:
Penny2x: The next few versions of FSD will be like watching the future unfold before our eyes.
Elon: It will feel like magic

Sure Musk. Sure.

 

[BobbyTables]

Just more hype like quantum leap.

https://twitter.com/x/status/1671369154673123329

Quote:
Penny2x: The next few versions of FSD will be like watching the future unfold before our eyes.
Elon: It will feel like magic

It took…18 months to go from version 10 to 11?

So 3 more versions is 4.5 years

Yeah! I’d hope they would progress over longer than the time it takes to transform a middle school student into a college freshman.

The world will also be very different then.

 

[Mullermn]

Tesla's radar is nothing like Waymo & Mobileye's Ultra Resolution 4D Imaging Radar. The difference is staggering.

Here's an indepth look:

Tearing-down Tesla’s in-house radar design – Why did they bother? | Ghost

A radar engineer’s analysis of the new Tesla HW4 radar.

www.ghostautonomy.com

Quick comparison: You have Huawei's 4D imaging radar with 288 virtual channels.

Mobileye’s Ultra Resolution 4D Imaging Radars with 2,304 virtual channels and ~500k points per second. (0.5 angular res, 2 vertical res).

Arbe radar 2,304 virtual channels (1 angular res, 1.7 vertical res).

Fisker Ocean is using Uhnder's radar with 192 virtual channels.

Tesla's New Radar has 48 virtual channels.

Sounds like what you’re saying is that despite the hooting on this forum, Tesla haven’t completely flipped on their previous stance about radar not being necessary since they’re bringing it in a more limited capacity than the others.

Given that they have more money to burn than the others it seems a safe bet that they could have put the same level of radar in if they wanted to.

 

[JB47394]

Given that they have more money to burn than the others it seems a safe bet that they could have put the same level of radar in if they wanted to.

I disagree. Tesla is extraordinarily tight with a buck. They put in the absolute minimum that they thought would work. When they thought that they could do without radar, it was gone. When they thought they could do without ultrasonic sensors, those were gone. "The best part is no part." If radar is making a comeback, then the unit they've chosen must have significant bang for the buck. It may simply be a patch to address something that they just can't get vision to do, as opposed to trying to merge a more capable radar unit into their system. For example, it may only be for Automatic Emergency Braking and not for the FSDb system. Perhaps AEB is the root of all or most of the phantom braking issues and this is the only way they've come up with to solve that problem.

Remember too that Tesla has to install the sensor suite in almost 2 million cars this year, the vast majority of which are priced between $40,000 and $51,000.

 

[Goose66]

It's hard to figure out what one of these 4D radar unit costs. Some articles I read say they are roughly the same price per unit as the conventional 2D radar unit, and others suggest they are 10x more expensive - more in line with a low-end LIDAR unit. With Tesla manufacturing their own, they can't be that expensive, right?

 

[uscbucsfan]

Remember too that Tesla has to install the sensor suite in almost 2 million cars this year, the vast majority of which are priced between $40,000 and $51,000.

It's still unknown what they will be doing with radar seeing that the only cars getting it are the S/X currently.

 

[BobbyTables]

It's still unknown what they will be doing with radar seeing that the only cars getting it are the S/X currently.

hopefully this is just a test and the real ones will come out later with better resolution?

If what we see on new s and x cars is hw4, it seems that tesla is still targeting L2, albeit with stressful party tricks where it can jerk around a roundabout and successfully take UPLs half the time.

 

[uscbucsfan]

hopefully this is just a test and the real ones will come out later with better resolution?

If what we see on new s and x cars is hw4, it seems that tesla is still targeting L2, albeit with stressful party tricks where it can jerk around a roundabout and successfully take UPLs half the time.

It's not an "if", it's confirmed to be HW4.

I don't see any indication that Tesla is just targeting level 2, just a long road with the equipment and plan they've put into place. Whether they can or not acheive level 5 with the selected path is another story.