HW2.5 capabilities

[stopcrazypp]

I was specifically responding to @mrkisskiss who was claiming 3cm localization with 10kB/km map size. Apparently the claim is actually 3-5cm lateral (in other words, 5cm), 10cm longitudinal... further discussion below

"I don't know what Mobileye is doing, but that's not an HD map. They really claim to get 3cm localization without GPS using 10kB/km map data?"
I'm talking about this part. You have said similar things about other HD maps, claiming they are not HD maps when they are, under commonly accepted definition (which is extremely broad).

Useless in urban environments, degraded when skies are overcast, probably degraded when operating at vehicle speeds, and anyway that 30cm that you get in goldilocks conditions is an order of magnitude off for lane localization. But vehicle-based GPS is already way better than phone GPS, at least in good conditions (better antennas, consistent orientation with respect to ground and heading).

30cm refers to 95% probability. RMS will be 15 cm. Yes, you are right, vehicle based GPS is already way better, but this news suggests further improvements in that front.

If they only have landmarks every 200m, 3-5cm lateral/10cm longitudinal is aspirational and will only be achieved in goldilocks conditions (at least in the next couple of years). Real environments are punishing to systems like this. You could never build an L4 system that relied on this, and even an L3 system would need to have a very solid backup plan.

Also, going around a tight corner with +/- 10cm longitudinal accuracy is pretty iffy. In some cases you will be in the adjacent lane, or scraping the guardrail.​

The GPS positioning and HD map is not going to be the primary factor, it'll only aid the vision system (at least this is how Tesla/Mobileye does it). It only has to be good enough to get it in the ball park (which lane is the right one to take, where the approximate location of the traffic light/sign is etc) and the vision system will do the rest (even with no lane lines, there are contextual ways to tell where it is located; that's how we do it with our own brains). The inertial based (and similar) systems will also help in areas of poor or no signal.

For example, even though your GPS unit does not give you accurate information on lane positioning, you are in no danger of being in the adjacent lane or scraping a guard rail. Where more accurate positioning is useful is for example telling which lane you are in (leftmost/rightmost: useful for determining turn lanes or exit lanes).​

 

[Mobster]

1. Great thread. I too am an engineer (aerospace, retired after 40 years). I spent many hours in smoky rooms (yeah back when they ENCOURAGED it, with some cigs protecting you with asbestos filters, cough, I did not smoke so it was tough) in discussions much like this including how it would be NOT POSSIBLE (+dangerous plus liabilities) to bring back a primary booster rocket for re use PARTICULARLY if one was CrAzY enough to try and set it down on a platform. Doesn't Elon understand what is not possible? Mars?

2. To the poster who has embedded: "Like how many times do i have to be right before people start noticing?". I have done the calculations. It may be an non-determinant, or more specifically a non-determinant polynomial. However my results say "10 to the 50,000 power". You need to continue to post more "right" opinions. So far it is looking good.

3. Here is an incredibly non-engineering observation: My wonderful grandfather lived to 96. He had stories about covered wagons (yes there were lawyers and liability lawsuits then), his insight into the craziness of DC versus AC electricity (Tesla? a dangerous fool per Edison), and he would unfold a newspaper article written by the experts of the day warning of dangers to the human body when exceeding 40-45 mph and particularly the nutcase Henry Ford who had proven his idiocy by going bankrupt a number of times before, but persists in putting the incredibly dangerous "auto-mobile" in the hands of men everywhere and possibly even WOMEN (gasp). The article pointed out the stunning liability and the nonsense of the proposition. He most certainly would encounter a liability disaster ending his company and possibly ending up in jail. So to the proposition: "Are you saying current AP2/2.5 Model S/X/3 will never be Level 3 or above?" I, unlike others, do not have the definitive answer. I like the poster who deals in science, facts, and pointed out: "No pleasure in being wrong..." Um...I very respectfully disagree. Even though retired, I have embarked on a plethora (love that big word) of research projects. I have a lab full of strange equipment (re: "Frankenstein") and have moved FORWARD faster when "being wrong" brought me to a solution....lots of pleasure bordering on sheer joy. Nevertheless it is always fun to watch the "facts is facts, so there!" versus "...got a vision...goona move forward" ---- diverse folks. --- regarding the observation: ".....the move fast and break things strategy to self-driving might one day result in a mission killer incident. But even if we consider that so remote it is not worth mentioning, the risk of a major accident being realized could seriously set Tesla back..." Yes it could. I worked for a global "innovative" 100 year experienced aerospace company that could not accomplish anything NEAR (with a reasonable price) what Elon has accomplished in 9 years with SpaceX. We played it very safe. And even then we wanted money directly from the government to do ANY research (not energy credits or credits earned....cash on the barrel). However Tesla car development may be different. Maybe too much risk, wrong choice in AMD, bad decisions? Wait. Let's us start a car company and implement FSD! A lot of VERY specific answers here in this thread with great graphics. I'm in...who else?

Thanks for this thread and all diverse (sometimes funny) posts. Great thread

 

[AnxietyRanger]

No pleasure in being wrong..." Um...I very respectfully disagree. Even though retired, I have embarked on a plethora (love that big word) of research projects. I have a lab full of strange equipment (re: "Frankenstein") and have moved FORWARD faster when "being wrong" brought me to a solution....lots of pleasure bordering on sheer joy.

Fair enough, a good point about different views to being wrong.

 

[jimmy_d]

Folks it must be said that it is reasonable to conclude that Mobileye is probably still well ahead of Tesla in terms of object recognition - artificial vision is all Mobileye does. It is all they have done in the past. EyeQ3 is almost four years old and they have been working on their next generation for a long time. Mobileye is run by a world expert in artificial vision - a guy with platinum academic credentials and a lot of real world experience as well in private industry. Amnon Shashua is laser focused on only one thing - artificial vision

If computer vision were like akin to designing internal combustion engines - a discipline which has seen nothing but incremental improvement for decades - then your sentiment might be appropriate given no other information. But computer vision is in the midst of a revolution the likes of which has rarely been seen in any discipline and the big bang moment for that revolution came in late 2012. Of course expertise prior to that date is not irrelevant but it has been mostly deprecated in favor of new techniques which are even now developing at an unprecedented speed. It's simply not possible to be 'too late' to the computer vision party in any meaningful sense. It's extremely likely that even the newest techniques in use today will be obsolete within a few years and they will likely rotate again a few years later.

At this moment in computer vision development expertise is nothing and execution is everything.

Additionally, in the critical new techniques there are very few people, perhaps a dozen, who have experience obtained prior to about 2014 which is substantial enough to give them an advantage in the coming years. To my knowledge none of those dozen-odd people has ever had any relationship with Mobileye. Based on the public records I've been able to find Mobileye had no staff dedicated to pure neural network vision systems as of late 2014 so it's not clear that they were even seriously studying the latest techniques until very recently.

Indeed Mobileye has computer vision expertise. And indeed they have put it to good use in developing automotive products. But I think it's far from clear that they have any substantial advantage today. I can't even find any trustworthy, independent evidence that they've assisted in the production of anything that is superior to AP1 as of Oct 2016 (the approximate time that they and Tesla parted ways). Indeed if you discount PR there is very little real world performance data about any product that has not been in the hands of the public for at least a year.

I think it's a little premature to be lauding any of the systems, especially Mobileye.

 

[Bladerskb]

Additionally, in the critical new techniques there are very few people, perhaps a dozen, who have experience obtained prior to about 2014 which is substantial enough to give them an advantage in the coming years. To my knowledge none of those dozen-odd people has ever had any relationship with Mobileye. Based on the public records I've been able to find Mobileye had no staff dedicated to pure neural network vision systems as of late 2014 so it's not clear that they were even seriously studying the latest techniques until very recently.

Indeed Mobileye has computer vision expertise. And indeed they have put it to good use in developing automotive products. But I think it's far from clear that they have any substantial advantage today. I can't even find any trustworthy, independent evidence that they've assisted in the production of anything that is superior to AP1 as of Oct 2016 (the approximate time that they and Tesla parted ways). Indeed if you discount PR there is very little real world performance data about any product that has not been in the hands of the public for at least a year.

I think it's a little premature to be lauding any of the systems, especially Mobileye.

This is just stupidly wrong. Why are u ppl allowed to spread such falsehood about a company? It should be a banneable offense!

24 mins

And lets just ignore gm supercruise, volvo pilot assist 2 and audi traffic jam pilot.

I mean whats wrong with u people? Why do u hate critical thinking?

 

[jimmy_d]

... but mostly they all are about offsetting weaknesses of computer vision...

Here the crux: "offsetting weaknesses of computer vision". Many of the arguments here regarding the fitness or otherwise of HW 2.5 and it's predecessor HW2 are premised on the notion that computer vision isn't going to get better very quickly.

So I can buy this argument: If all you have is low confidence systems then you need multiple of them, ideally with non-correlated failure modes. And even after you've done the hard work of making them play well together you still have a system that's not particularly great so it has to be used in constrained environments because you're just barely clearing the bar to usable under even good circumstances.

But the premise is wrong. Vision, and more importantly perception and prediction, are improving by leaps and bounds with no end in sight. Tesla is betting that these techniques will improve fast enough that they can fulfill the promise they made to those, apparently 30,000, of us who bought the FSD upgrade. I think they are very likely right about that, though of course predictions about future developments are open to error.

If they really believe in vision then you're not going to see them adding supplemental radar to the car. Improved processors maybe, and maybe better cameras if there's no downside. But that stuff will not be added because it's required, just because you might as well be using the best stuff that's within your cost envelope. If vision, perception, and prediction methods are really good then the cameras on HW2 are more than adequate to enable FSD. HW2.5 is nice, but it's only getting upgraded because the upgrades are basically free.

And on the other hand if vision, perception, and prediction aren't really good then all those redundant sensors don't get you much because the infrastructure of driving today is all geared towards making the most of vision - the only sense that human beings require to safely drive a car.

 

[rnortman]

If they really believe in vision then you're not going to see them adding supplemental radar to the car.

I'm skeptical of even achieving true L3 with the current hardware, but I'll say that they'll probably eventually (soon?) do pretty well at blind spot detection with the rear-facing cameras (the ones on the sides mostly, not the fisheye, though maybe that a little too), even for fast-approaching vehicles, without corner radar. It seems like they haven't even started on this yet (at least not in anything that's released) and actually it's sorta relatively easy I would guess... with one caveat: those camera lenses have to be clean. This may not work well if the lenses are dirty or overly wet. Corner radar mostly doesn't care about weather or nighttime, so point to radar there.

But compared to true L3+ driving, blind spot detection with vision only seems trivial.

 

[jimmy_d]

This is just stupidly wrong. Why are u ppl allowed to spread such falsehood about a company? It should be a banneable offense!

24 mins

And lets just ignore gm supercruise, volvo pilot assist 2 and audi traffic jam pilot.

I mean whats wrong with u people? Why do u hate critical thinking?

Of course Mobileye is using NN now. Nobody isn't using it now. But I haven't been able to find any evidence that they were using it, or planning to use it, when they developed the hardware that went into the original AP1. And if you want to offer counter evidence you're going to need to come up with something that dates from back then. I previously challenged you to provide a document authored by mobileye prior to 2014 which used the term 'neural network'. You chose to ignore that challenge. I'd still like to see it if you have it. It would make me feel better about having purchased AP1.

Nice video, my how things change. A couple of years ago I viewed a youtube video of Shashua arguing that deep learning wasn't needed for self driving cars, and perhaps wasn't the right way to go, and that in any case Mobileye's approach was superior. As a neural network guy I remember that I found that argument infuriating. That video seems to have been taken down because I can't find it anymore. I wonder why.

And as for supercruise, pilot assist, and traffic jam pilot: two of them exist only as demos so far so we'll have to see, and the customers who actually drive the third one don't seem to have a higher opinion of it than Tesla users have of autopilot. So for the purpose of demonstrating the ME has something better than what Tesla has fielded I'd have to say yes, we can ignore them for now.

 

[S4WRXTTCS]

Of course Mobileye is using NN now. Nobody isn't using it now. But I haven't been able to find any evidence that they were using it, or planning to use it, when they developed the hardware that went into the original AP1.

Just as an FYI

MobileEye was the first to deploy a DNN on the road.

Source -
Exclusive: The Tesla AutoPilot - An In-Depth Look At The Technology Behind the Engineering Marvel - Page 4 of 6

As to this discussion I think its important to point out that the MobileEye solution in AP1 isn't really that great. It misidentifies speed signs all the time, and isn't trained on the sides of vehicles/etc. There is only a single camera, etc.

Heck the entire breakup between MobileEye and Tesla was due to Tesla trying to do much more with the HW than what MobileEye thought was safe. So they broke up with Tesla to save their brand from being tarnished.

We really don't know what element was the cause of the delays/issues with AP2. People assume it's the imaging aspects, but I have to wonder if it wasn't Radar related. The Radar changed in HW 2.5 and all the false braking events that we saw in the preliminary version of AP2 were due to things likely caused by radar returns, and Tesla's desire to actively react to them.

Oh, and I would recommend NOT challenging Bladerskb. It should be pretty evident that he has more time on his hands than any of us. He takes his Tesla hating hobby quite seriously.

 

[EinSV]

@Bladerskb,

Since you have resumed posting but have chosen not to respond to the posts on pages 47-49 of this thread by @calisnow and @buttershrimp that raise serious questions about your account, it would appear that their conclusions are correct. This is a pretty sorry state of affairs for TMC.

Bladerskb does not sound like the same person in post 932 as in any of his earlier posts - not even slightly the same - as whoever was using his handle earlier this year. Yet the post is not satire - it is simply straightforward technical explanation. But the voice is not that of a native English speaker. The sentence structure is different, the tone is different -> but the grammar, syntax and idiomatic missteps are the key giveaway here. He sounds like someone with a strong grasp of the technical side of autonomy, educated, with an academically sufficient working grasp of English that he has not bothered to completely polish because he is too busy doing - something else.

The bladerskb handle now looks like the work of a team of multiple people - and somebody on the team just slipped up.

What IS in common with the earlier and current posts of bladerskb is that both posters have a lot of knowledge of Mobileye and steadfastly defend it while making claims that Tesla cannot achieve certain autonomy levels on AP 2 hardware - strangely echoing what a GM executive has been widely quoted in the press saying recently. The timing is uncanny.

Why the different grammar all of a sudden? It appears someone took over his account - carelessly I might add because of the obvious non-native speaker giveaways - but why? If the account was taken over - why? Why now? Internal dispute over content? The primary spokesman get sick or is otherwise unavailable? Dispute over compensation? This is the closest to a smoking gun of a "team" account used to push out a particular viewpoint as TMC has ever had.

To the mods I am NOT accusing @Bladerskb of being a troll - this is not a troll post whatsoever. It is appears to be a team effort to push forth a particular viewpoint over an extended period of time.

 

[buttershrimp]

@Bladerskb,

Since you have resumed posting but have chosen not to respond to the posts on pages 47-49 of this thread by @calisnow and @buttershrimp that raise serious questions about your account, it would appear that their conclusions are correct. This is a pretty sorry state of affairs for TMC.

@EinSV, I don't think he can understand your post... let me translate it for him:

Поскольку вы возобновили публикацию, но решили не отвечать на сообщения на страницах 47-49 этой темы от @calisnow и @buttershrimp, которые поднимают серьезные вопросы о вашей учетной записи, похоже, что их выводы верны. Это довольно печальное состояние дел для TMC.

 

[S4WRXTTCS]

@Bladerskb,

Since you have resumed posting but have chosen not to respond to the posts on pages 47-49 of this thread by @calisnow and @buttershrimp that raise serious questions about your account, it would appear that their conclusions are correct. This is a pretty sorry state of affairs for TMC.

You asked this yesterday, and then Anxiety Ranger suggested that Bladerskb respond in a different thread specially made for thread de-railers. But, that made a moderator really pissed off, and this is what the mod had to say since you obviously didn't check that location.

"Wow, how handy of everyone to conveniently create an entire THREAD that needs to go into snippiness.

For the record:

Calling someone a troll is a personal attack

Calling someone a fanboy is a personal attack.

Calling someone a short is a personal attack. Whether height or stock related.

Personal attacks are not permitted in this forum.

The overall negative tone is degrading the forum. Many of the culprits have contributed to this thread... you know who you are."

 

[calisnow]

@EinSV, I don't think he can understand your post... let me translate it for him:

Поскольку вы возобновили публикацию, но решили не отвечать на сообщения на страницах 47-49 этой темы от @calisnow и @buttershrimp, которые поднимают серьезные вопросы о вашей учетной записи, похоже, что их выводы верны. Это довольно печальное состояние дел для TMC.

OMG @buttershrimp wins the internet today!

 

[calisnow]

"I don't know what Mobileye is doing, but that's not an HD map. They really claim to get 3cm localization without GPS using 10kB/km map data?"

No I think they claim thirty centimeter localization not three. I remember watching the REM low bandwidth presentation Amnon gave and I believe he said thirty - roughly 15 inches - which he claims is plenty specific enough for safe self driving (seems logical to me because you combine that input with other sensors of course and arrive at even higher confidence).

 

[EinSV]

You asked this yesterday, and then Anxiety Ranger suggested that Bladerskb respond in a different thread specially made for thread de-railers. But, that made a moderator really pissed off, and this is what the mod had to say since you obviously didn't check that location.

"Wow, how handy of everyone to conveniently create an entire THREAD that needs to go into snippiness.

For the record:

Calling someone a troll is a personal attack

Calling someone a fanboy is a personal attack.

Calling someone a short is a personal attack. Whether height or stock related.

Personal attacks are not permitted in this forum.

The overall negative tone is degrading the forum. Many of the culprits have contributed to this thread... you know who you are."

@calisnow and @buttershrimp's posts raise serious questions about the misuse of this forum, which @Bladerskb has conspicuously avoided answering.

I think their posts warrant a direct answer from @Bladerskb, not more deflection.

 

[verygreen]

@EinSV, I don't think he can understand your post... let me translate it for him:

Поскольку вы возобновили публикацию, но решили не отвечать на сообщения на страницах 47-49 этой темы от @calisnow и @buttershrimp, которые поднимают серьезные вопросы о вашей учетной записи, похоже, что их выводы верны. Это довольно печальное состояние дел для TMC.

Nah, that's too rigid. Google Translate?

How about: В связи с твоим новым появлением в этой теме хотелось бы услышать и что-нибудь касательно подозрений @calisnow и @buttershrimp пару страниц назад.

That said, there are many other languages in use in Eastern Europe.

Full disclosure: Russian hackers are everywhere (though I am sure the local grammar experts have figured this already long ago).

 

[S4WRXTTCS]

@calisnow and @buttershrimp's posts raise serious questions about the misuse of this forum.

I think their posts warrant a direct answer from @Bladerskb, not more deflection.

Bladerskb is not obligated in any shape or form to respond to a witch-hunt.

Since when did we become Blade Runners?

Why the paranoia?

At the end of the day it's just technical details. The technical details are either right or wrong. He does add a lot to clear up misconceptions, and he provides a lot of insight into various technologies.

Now don't get me wrong. I'm not someone who even likes Bladerskb. He's the least friendly person on here. Heck I have enemies on here that are more friendly to me than he is. The guy/girl/group/whatever has no soul. I can joke around with lots of people on here, but not him. I give him a male pronoun because it seems to fit even though I have no idea.

I know it's human nature to wonder about someone, but I think you have to give someone the benefit of the doubt. To have trust in your own knowledge that you don't feel threatened by him.

We have to accept that we'll never know. Even if he does answer would we really even believe him?

I think it's best just to use him as a resource for into until he eventually finds himself a new home. I think he picked this place because it's likely the most happening place when it comes to speculation on autonomous cars. Ironically for the very reasons that he keeps telling us we're all wrong.

 

[jimmy_d]

Just as an FYI

MobileEye was the first to deploy a DNN on the road.

Source -
Exclusive: The Tesla AutoPilot - An In-Depth Look At The Technology Behind the Engineering Marvel - Page 4 of 6

As to this discussion I think its important to point out that the MobileEye solution in AP1 isn't really that great. It misidentifies speed signs all the time, and isn't trained on the sides of vehicles/etc. There is only a single camera, etc.

Heck the entire breakup between MobileEye and Tesla was due to Tesla trying to do much more with the HW than what MobileEye thought was safe. So they broke up with Tesla to save their brand from being tarnished.

We really don't know what element was the cause of the delays/issues with AP2. People assume it's the imaging aspects, but I have to wonder if it wasn't Radar related. The Radar changed in HW 2.5 and all the false braking events that we saw in the preliminary version of AP2 were due to things likely caused by radar returns, and Tesla's desire to actively react to them.

Oh, and I would recommend NOT challenging Bladerskb. It should be pretty evident that he has more time on his hands than any of us. He takes his Tesla hating hobby quite seriously.

It may well be that AP1 included a DNN (which in the research community means discrete neural network but which is often repurposed to mean deep neural network in the lay press) and it's possible that Mobileye was using backprop to train some of the layers. I've seen other bloggers claim that as well, but I've recently been treating it as hearsay lacking something more concrete. In late 2015 it was my assumption that AP1 included an NN and was using full training but various sources suggested that most of the lower layer kernels were being hand built and that most of the feature extraction was being done heuristically. This is not what most people at the time would have considered deep learning. My own analysis of public available architectural data found no NN specific features in anything for AP1 and estimates of the hardware compute capabilities suggested that it was not capable of running any modern CNN as of 2014. I could find no authoritative statements from either Mobileye or Tesla on this matter, but at the time Mobileye was publicly suggesting that deep networks were not appropriate to the application. Earlier in this thread I posted a recent video where Mobileye's Shashua mentions offhand that EyeQ3 takes 6 seconds to run a single frame of AlexNet. His point in that statement is to suggest that pure deep learning networks are impractical because they are too compute intensive. And certainly that was a valid point when considering the capabilities of 2014 era embedded vision processors, which is what the EyeQ3 appears to be. Certainly hand coded networks can outperform trained systems in terms of efficiency. But hand built networks have many drawbacks not the least of which is they lag trained networks in terms of accuracy on all recent public benchmarks.

My point here is not to fault Mobileye. As I've mentioned before I have no complaint about their products, or their development approach. My objective is to counter the perception that the capabilities of AP1 circa 2016 are somehow unlikely to be exceeded by AP2, and part of that is to put AP1's actual capabilities in perspective. There is a wealth of new methodology that can be employed by HW2 which cannot be used in AP1. In time AP2 will greatly exceed the capabilities of AP1 and I think it's quite likely that eventually FSD will become possible in HW2 equipped vehicles.

As for our boisterous peer, I have no problem with him. If he becomes motivated to bring something to the table that illuminates our shared interest then he can say whatever he wants as far as I am concerned. I can separate wheat from chaff. I really would like to see some good information on what Mobileye put into AP1 and I'll be happy to see it no matter who brings it or why.

Though of course, if the consensus on the thread is that I should desist then of course I'll bow to the community.

 

[lunitiks]

Hey guys in one week it's the 1yr anniversary of AP2.0! We should plan something big!

 

[AnxietyRanger]

But computer vision is in the midst of a revolution the likes of which has rarely been seen in any discipline and the big bang moment for that revolution came in late 2012.

True, but 2012 was five half a decade ago, though. The likes of Google and MobilEye has been leading this space for those five years - deep-learning included.

To suggest Tesla will bypass them in what appears to a significantly shorter time, is to believe Tesla has something up their sleeve that will allow them to bypass all that progress in what appears to be a significantly shorter time...

That suggestion is, of course, the fleet. But to simply label the existing players as stalwarts of an old-world would be doing the truth a disservice. The truth is, everyone is doing deep learning and we have pretty much no proof of Tesla's leadership in this area at all yet.

Now, that doesn't mean there isn't potential. If one has faith, one can believe many things.