Blog Musk Touts ‘Quantum Leap” in Full Self-Driving Performance

A “quantum leap” improvement is coming to Tesla’s Autopilot software in six to 10 weeks, Chief Executive Elon Musk said a tweet.

Musk called the new software a “fundamental architectural rewrite, not an incremental tweak.”

https://twitter.com/x/status/1294374864657162240

Musk said his personal car is running a “bleeding edge alpha build” of the software, which he also mentioned during Tesla’s Q2 earnings. 

“So it’s almost getting to the point where I can go from my house to work with no interventions, despite going through construction and widely varying situations,” Musk said on the earnings call. “So this is why I am very confident about full self-driving functionality being complete by the end of this year, is because I’m literally driving it.”

Tesla’s Full Self-Driving software has been slow to roll out against the company’s promises. Musk previously said a Tesla would drive from Los Angeles to New York using the Full Self Driving feature by the end of 2019. The company didn’t meet that goal. So, it will be interesting to see the state of Autopilot at the end of 2020.

 

[banned-66611]

Although these responses from Elon are always cryptic, this sort of dialogue between users and Elon adds to the enjoyment of owning AP/FSD. It's a major differentiator between Tesla and other FSD developers.

The enjoyment of being told, yet again, that the fix is right around the corner, honest!!

Musk is a serial liar, if he says something will be fixed soon you can assume that means not for a very very long time.

 

[Snuffysasa]

Makes sense, but even for simple features like staying within the lane, the Mobileye equipped cars are far behind Tesla..

People think lane keeping is simple. I know it seems simple, especially if the lane detection is being handed to you. But have you developed it and deployed it in a real vehicle at 80mph on highways? I have, and its not so simple.

I just think many people are underestimating the fact that Tesla has had full control of their autonomy stack for years. Even if Mobileye is collecting data, it's data from all sorts of sensors they have little control over.

I don't think this is very relevant.

 

[Snuffysasa]

I'll qualify my statement further:

If people brought up Mobileye's 2021 product stack and said "lets wait for 2021 and see what cars have it available for the consumer to use to see if it's better than Tesla" then I'd consider that neutral.

This is fair, and I will come back to this and explain why you will see even from a fair neutral perspective why this 2021 consumer vehicle (and other future consumer vehicles) will be way a head of Tesla FSD in 2021

 

[powertoold]

People think lane keeping is simple. I know it seems simple, especially if the lane detection is being handed to you. But have you developed it and deployed it in a real vehicle at 80mph on highways? I have, and its not so simple.

I agree with you. It's not easy to deploy autonomy features to consumers. That's why I keep saying many people underestimate what Tesla has achieved with AP in the last 4 years, especially the traffic control feature.

It's also why I point out that there doesn't seem to be any consumer-deployed Mobileye systems that are even close to Tesla's right now.

 

[Snuffysasa]

Do you have a source for that? All I see is that the SuperVision is capable of FSD (in terms of the sensors).

I will make a detailed post later of all of the consumer features and capabilities coming in this 2021 mobileye vehicle, that I do not expect Tesla to have in all 2021

 

[powertoold]

I will make a detailed post later of all of the consumer features and capabilities coming in this 2021 mobileye vehicle, that I do not expect Tesla to have in all 2021

You'll be right, since the future is all speculation. I'm just pointing out facts about the current state of consumer-deployed autonomy features. I'm actually curious to see if there're any currently available systems better than Tesla's in *any way*. If so, I'd like to see a video demo or stress test by an actual consumer.

 

[Snuffysasa]

I agree with you. It's not easy to deploy autonomy features to consumers. That's why I keep saying many people underestimate what Tesla has achieved with AP in the last 4 years, especially the traffic control feature.

It's also why I point out that there doesn't seem to be any consumer-deployed Mobileye systems that are even close to Tesla's right now.

You are not grasping, that this is the first and only time, that mobilye is making the product, and not an OEM. Mobileye has been testing real vehicles on the road for as long as Tesla, and knows that it takes to make a product.

 

[Snuffysasa]

Makes sense, but even for simple features like staying within the lane, the Mobileye equipped cars are far behind Tesla.

And not always true....

Nissan ProPilot 2 has much greater miles per disengagement than Tesla Autopilot on the highways it is allowed. (and that used 2015-2016 mobileye tech). Before Tesla AP2 was even out. (and we know how Tesla AP2 was when It launched in 2016....)

And even in some places, SuperCruise 1 which has (2012-2013 mobileye tech), has simply more stable performance and greater miles per disengagements on the highways it is allowed (maybe not as much today though, Tesla may have passed them now)

But SuperCruise 2nd gen, then can take lane changes, highway interchanges, and more, Is even better than Tesla Autopilot at certain things (clearly off highway it stands no chance)... but for a lot of core important things go a highway pilot. SuperCruise 2 is better than Tesla Autopilot, and these things are: handling vehicles in adjacent lane that are slightly leaning in ego lane / close range vehicle cut ins, when the ego vehicle tries to take a lane change and tracking vehicles in behind the ego vehicle that. have a changing velocity and changing lanes, and overall speed and steering control when taking sharp highway interchanges (my Tesla will perform undesirably on these all the time).... And this system uses 2015-2016 mobileye tech (as well as some homegrown GM tech). I am not saying SC2 is better than Tesla Autopilot though, just that there are certain tasks that is is.

And these aren't the only systems, there many more coming.

But the main point I am making now is that there is a gap of about 5 years... in an automaker product cycle... the latest OEM vehicles coming out today use 5 year old perception and other technology. This is why Tesla is now significantly ahead of other ADAS systems available.

However, like @diplomat33 said, this is exactly why this news is different. This is now for the first time, where that 5 year gap is closed. And for the first time mobileye will deliver the whole product, and update it with over the air updates, keeping it current.

It will have address to address / full FSD (in hands free / eyes on mode).... yes its possible it may not have all of those things at day 1... however, at day 1 it will still have many capabilities greater than what Tesla FSD will have in fall of 2021, and I will detail those later.

 

[powertoold]

however, at day 1 it will still have many capabilities greater than what Tesla FSD will have in fall of 2021, and I will detail those later.

I can't argue with that. It may be true, but like I've mentioned before, based on the 10 articles I've browsed on SuperVision in the Concept Zero car, I've yet to see any mention of a traffic control feature. Without a traffic control feature in late 2021, they are far behind Tesla imo.

 

[Snuffysasa]

Great, show us some stress tests, like on a mountain road or something with sharper turns (there are 1000+ of these videos for AP)

So they would probably fail to Tesla in these conditions.... in general Tesla is far ahead of other OEMs in challenging conditions...

SuperCruise would not allow you to use the system in a condition like this.... and that's because other OEM's minimum performance threshold is far greater than Tesla's

Other OEMs are more focused... on the simple conditions but getting even more reliability in these areas.

Do you think they are implying level 3? If so, why don't they just say level 3?

I've heard about this before, but I was referring to the SuperVision system graphic above, where it says hands free driving. Do you think or know if this is level 3?

No mobileye told me definitely no L3 planned for this car, it must be missing some necessary hardware redundancies (even aside from radar / lidar)

 

[Snuffysasa]

I can't argue with that. It may be true, but like I've mentioned before, based on the 10 articles I've browsed on SuperVision in the Concept Zero car, I've yet to see any mention of a traffic control feature. Without a traffic control feature in late 2021, they are far behind Tesla imo.

that's reasonable, If I had to guess, I'd bet at least detection and alert of stop lights is there at day 1... and stopping at them where REM maps are there would be enabled at day 1.


but! but!... even if there is no traffic light control feature once so ever.... I don't think that even remotely implies they are far behind Tesla, and Ill explain why in a while later.

 

[Bladerskb]

Are the two linked videos showing examples with cities which have been per-scanned for use with high definition maps?

What is the power consumption of this system and can it work without high definition maps?

Mobileye does crowdsourced HD mapping from millions of commercial cars. Their map is called REM maps. They have most/all of US, Europe, Japan and China basically mapped.
Their latest chip EyeQ5 is 24 TOPS at 10 watts.

 

[Snuffysasa]

Self-driving is still an arena where no products have really materialized on the market yet.

self driving? no. But this thread hasn't really been about that. It's been about advanced ADAS features.

For MobileEye it's about convincing they're the leaders by giving impressive demos. Common denominator is that neither system is ready, or there would be self-driving cars today.

Again, not talking about self driving cars, but ADAS that creates a great consumer experience and reduces accidents on the road.

Although, mobileye is making actual self driving cars, and they will deploy them around Jan 2022 assuming pandemic is behind us. (and that is not a date, that just keeps getting pushed back, that is the first and only date they gave us)

 

[Bladerskb]

I can’t find much info on the numbers, but very few car models actually have autosteer and tacc. Plus, for these models, you have to pay extra for both. I’d say less than 1 million cars have both enabled (and also use mobileeye).

Cars with Autopilot in 2020 - AutoPilot Review

This is clearly not true. Just by looking at the luxury market alone. You need to look up how many millions of new cars are sold each year.

 

[Mardak]

I don't think so, seems like such a minor fix for a whole rewrite, but then again, like I said, cryptic lol

If he was replying to the "intersection… large box area," I would think it's definitely waiting on the rewrite as Karpathy's presentations show the birds eye view identifying lanes and direction of traffic for the whole intersection, and this needs to make use of more than just the main camera.

And even for the "minor" problem of adjacent lines, the 4D rewrite combining all cameras at the neural network level will likely significantly improve this lane perception problem as currently the side repeater view uses software 1.0 heuristics to coordinate with main camera outputs to decide when a lane is adjacent or not. For those who have experienced unnecessary lane change aborts, this will probably improve quite a bit as the software should stop thinking the adjacent lane instantly disappeared when it looks like the camera outputs disagree.

 

[ChrML]

self driving? no. But this thread hasn't really

Really? The thread title says otherwise.

Although, mobileye is making actual self driving cars, and they will deploy them around Jan 2022 assuming pandemic is behind us. (and that is not a date, that just keeps getting pushed back, that is the first and only date they gave us)

They might succeed, or they might not, or they might deploy it to a few cars in a geofenced area to check of the "System deployed"- milestone for the investors.

A date that far ahead suggests it's a goal, or an aspirational target date. It's pretty much impossible to pick a firm date that far ahead unless they've already solved all the issues. And if they had, they should deploy it now. I've worked long enough in the software biz to know that when a system is said to be tested/validated atm, it usually means the team is in the middle of writing code for an MVP.

In Tesla's case validating something means they're just about starting to write the code, but they're kinda extreme (2016 AP2.0 anyone?).

 

[Snuffysasa]

Really? The thread title says otherwise.

Title of the thread says Tesla Full Self driving which refers to Tesla's ADAS package known as "FSD" and that is what we have been discussing.

A date that far ahead suggests it's a goal, or an aspirational target date. It's pretty much impossible to pick a firm date that far ahead unless they've already solved all the issues. And if they had, they should deploy it now. I've worked long enough in the software biz to know that when a system is said to be tested/validated atm, it usually means the team is in the middle of writing code for an MVP.

That date is 15 months ahead... and additionally that date hasn't slipped at all since they announced it years ago... actually it has been moved up. It's not just a goal. Could it end up getting pushed back? of course

or they might deploy it to a few cars in a geofenced area to check of the "System deployed"- milestone for the investors.

This is exactly what they'll do..... and this is the bread and butter of what self driving cars are.... anything else is pipe dream for this decade... why change the subject to something like that.

and change few cars... to a few thousand cars

In Tesla's case validating something means they're just about starting to write the code, but they're kinda extreme (2016 AP2.0 anyone?).

Not sure what you are are saying here.

 

[Bladerskb]

I don't know what point I'm missing. This new ME system is marketed as "L2/L2+". No where does it mention traffic light feature.

I'm going to say there's no traffic light feature planned for consumer release right now.

You do realize that the "+" implies additionally other features.

I can't argue with that. It may be true, but like I've mentioned before, based on the 10 articles I've browsed on SuperVision in the Concept Zero car, I've yet to see any mention of a traffic control feature. Without a traffic control feature in late 2021, they are far behind Tesla imo.

Look Mobileye has supported traffic lights/stop signs since EyeQ3 (2014) and have been pushing it and collecting map data on all stop line, stop signs, traffic lights, traffic signs, road markings/signs in the entire world since the launch of EyeQ4 (Late 2017).

All BMW from Late 2018 has this capability but BMW and their Tier 1 is just too slow, conservative and incompetent to write and release a control algorithm that will use the outputs of the NN. As it was Mobileye's expectation that BMW would have this feature and more out by now. Because everything is already done on their part (Vision, HD Map). BMW even made a video of traffic light/ sign response but haven't released it. This is a company that is just getting OTA to work after struggling. So you can't count on the trad OEM to get software right. This is why Mobileye doing the entire thing end to end is a game changer. You don't have to rely on the trad OEM and how long they drag their feet.

To think that the upcoming EyeQ5 with mobileye's end to end / full software stack and suite won't support traffic light/stop sign recognition and response is ludicrous.

https://i.imgur.com/Vg3p4zi.png

 

[powertoold]

Man, you guys are hilarious. I can't find a single video or piece of evidence that these BMW traffic control systems have been deployed to consumers. All the videos or information I see is marketing and/or marketing demos.

I'd help a lot if you guys can provide videos of actual consumers showing off these features.

Once once once again again again, UNDER DEVELOPMENT:

BMW's next-gen adaptive cruise control stops at traffic lights - Roadshow

 

[Snuffysasa]

Man, you guys are hilarious. I can't find a single video or piece of evidence that these BMW traffic control systems have been deployed to consumers. All the videos or information I see is marketing and/or marketing demos.

I'd help a lot if you guys can provide videos of actual consumers showing off these features.

Once once once again again again, UNDER DEVELOPMENT:

BMW's next-gen adaptive cruise control stops at traffic lights - Roadshow

We have told you multiple times now, that this is on the OEM and the tier1. And yes they are slow and incompetent when it comes to releasing features to consumers, and yes much behind Tesla. And now we are telling you that this is the first and only time where mobileye is actually taking on this role to actually release these features to customers.

So you keep giving examples and comparing to old Mobileye systems, where the OEM / tier1 was in charge of feature deployment, which is of course entirely irrelevant.