Comment by Fricken | Hacker News Reader

Comment by Fricken | Hacker News Reader

Right now, with HW2 Tesla is in kind of an awkward no man's land where their hardware is overkill for basic ADAS, but nowhere near enough for any kind of autonomy that allows the driver to turn their attention away from the driving task.

By going to a custom ASIC with inferencing. I have little doubt about Musk's claim that it'll offer a 10x improvement over the current stripped down PX Drive2 variant for deep learning tasks. If I'm to decrypt Musk's branded terminology and take him at his word, then they'll be going from about 8 TOPS to 80 TOPs with that they'll be able to get even deeper into the no man's land that's even more overkill for basic ADAS but still not nearly enough to handle full self driving safely in all conditions, though maybe they'll get L3 highway out of it.

Nvidia's upcoming Pegasus board will supposedly do 320 TOPS, and their competitor, Intel/Mobileye's EyeQ5 will likely have comparable specs. These motherboards are designed to replace Robotaxi trunks brimming with ~$150k in liquid cooled compute that draw 3-4000 watts.

I ordered a Tesla Model 3 yesterday. When I see the "Full Self-Driving Capability" checkbox at checkout time (it's an option you can pay $3,000 for right now) I just have zero confidence that they can deliver on this. Almost every other aspect of the car is fantastic. But Tesla's delusions about achieving full self-driving with their current approach just make me lose some respect for them. It makes me worried that they might be deluded about other things too.

I use the word "deluded" above realizing that I might be completely wrong. If they pull it off I'll eat crow and give them props. But I just don't see it happening.

Are the model 3 hardware modular? For example, is it possible they’ll have you bring it to the dealer in a few years and get outfitted with these new chips?

Yes, Musk said that on the earnings call. But the hardware upgrade is not the hard part of self driving. At least they have stopped claiming that their current hardware is powerful enough for full self driving already, which was a ridiculous and worrying position for them to take.

> ridiculous and worrying

I think “fraudulent” would be more accurate; it was factually false, it was at best entirely unfounded if not actually knowingly false, it was material to decisions of people to purchase the self-driving add-on.

> At least they have stopped claiming that their current hardware is powerful enough for full self driving already

The Autopilot page [1] still states the following:

"All Tesla vehicles produced in our factory, including Model 3, have the hardware needed for full self-driving capability at a safety level substantially greater than that of a human driver."

[1] https://www.tesla.com/autopilot

If you consider “hardware” as sensor suite only (quite a stretch of an interpretation, I know), then it’s kind of true, especially if they will replace the brains at no cost once they have an actual use for them

This is a commonly missed point. You can buy the car today and be very happy with a fantastic car as it is. I did and I'm quite happy, you'll probably be very satisfied as well. I didn't even buy the EAP, as I don't do much highway driving and I think almost no one actually buys the FSD.

Having said that, I find it hilarious that people on here have such immensely strong opinions about something that doesn't exist. They'll go at length about how LIDAR is a hard requirement, about how many FLOPs you need, etc etc etc. Fact: FSD doesn't exist and no one has it. Nobody knows which approach will work and if more than one approach works no one knows the timelines or economics of the different approaches and very certainly no one knows the market impact of these. But people start thinking they know the future and then become very emotionally vested in their make-believe stories.

It's actually quite simple: companies sell you stuff, buy it if you want it/need it. Today Waymo sells me nothing, Cruise/GM sells me nothing, Tesla sells me a great BEV with Level 2 driving assistance. I got the car, but not the level 2 because it doesn't fit my driving needs. If they start selling level 3-4-5, I'll buy it if the price is right and I need it.

Waymo is very close to wide availability of self driving hire cars in Phoenix: https://www.bloomberg.com/news/features/2018-07-31/inside-th...

Yes, various companies are in varying states of readiness in various jurisdictions.

> Today Waymo sells me nothing

It's certainly not nothing

Only one of the companies you mentioned currently has cars on the road with self driving tech that can kill you.

Cadillac's CT6? Yeah, I speed away every time I see one.

But there's no need to pick on GM, they all have self-driving tech capable of killing you.

>>It makes me worried that they might be deluded about other things too.

Tesla will fail. But they will fail at 5% of their goals, and achieve 95% of that goals. It will still be called 'Failure'.

To avoid that 'Failure' tag, other companies won't even try.

Do this process enough number of times. And it will appear like Tesla is failing all the time, and yet achieving almost 100% more than other companies.

The problem here is in defining success in terms of binary outcomes of success and failure.

Oh, it's capable all right. Unless you actually try to use it anywhere that's not a perfectly spherical road in a vacuum, then it's also capable of killing you.

The interesting thing about the $3000 "upgrade" is that if you actually want to get it, I think you're better off taking the $3000 and investing it in Tesla.

Assuming autonomy is make or break for Tesla in the long term: if Tesla fails at it your stock will still be worth something (versus a useless upgrade). If Tesla succeeds you'll make more than enough to cover a retrofit.

It's hard to say for sure they are no where near full autonomy in their research although their current vehicles crashing into barriers and fire trucks obviously don't cut it.

I'm only commenting on what little there is to glean about the compute, but there are so many other red flags about their autonomous development strategy. Crippling per-unit cost constraints, No Lidar, SLAM-lite, ANN heavy, insufficient redundancy and they've got a major data collection problem because they're relying on what they can get over the air. Plus Musk scared away most of their their top engineers.

Serious question- why is LIDAR so important? Obviously it’s possible to drive without it (humans do it), so it’s conceivable that eventually machines will be able to do it as well... So I’m guessing your point is, that won’t happen anytime soon and LIDAR is the fastest way to market; can you confirm/expand/clarify?

The standard of safety for autonomous drivers will have to be far beyond parity with humans. Humans cause car accidents all the time. LIDAR gives unambiguous depth and probably much easier object segmentation, no complexity or reasoning required like computing from stereo images.

Good point, plus if you have to choose between having LIDAR and acceptable self-driving capability in 3 years, or no LIDAR and acceptable self-driving capability in 10 years, I think most people would prefer the first option.

Sure, but then you're still left with the unsolved problem of getting lidar quality depth maps out of video in all lighting and weather conditions.

Estimating distances robustly is very important for navigation. LIDAR is accurate at relevant distances and most importantly very robust against environmental effects. Without LIDAR you can use stereoscopic vision (like human with two eyes) but that is very demanding, not that accurate, and very error prone. Sure, human can do it somewhat well, but eye and brain are extremely complex things to implement (and it still takes years to learn to understand what you see).

Personally, I’d refuse to implement self-driving vehicle without at least a ”backup LIDAR” to check vision system results. Otherwise you are forced to assume stuff like ”things at stand still are either above the road, beside it, or just shadows”, causing crashes when there is suddenly a stopped car in front of you. (If you didn’t do that assumption, you would be dodging shadows and other clutter..)

[source: I’ve been researching vision algorithms in a related field.]

>> Without LIDAR you can use stereoscopic vision (like human with two eyes) but that is very demanding, not that accurate, and very error prone.

Because that isn't how people judge distances, at least not distances beyond a few feet in front of their faces. Plenty of people with only one eye do very well. It is a difficult problem because we use a variety of techniques and 'hardware' when estimating distances and speeds. Car companies are trying to do with one tool (ie lidar) something we do with many.

Yep, we do huge amount of assumptions to derive ”model of a world” from quite limited amount of data. And we do lots of mistakes without ever realising it. Fortunately, most of those mistakes are irrelevant, and safety margins let us correct most of the relevant mistakes. Rest become accidents.

I guess the same logic applies directly to self-driving cars as well..

> Without LIDAR you can use stereoscopic vision (like human with two eyes)

At any but very close distance, don't humans mostly use a combination of lighting cues, a priori knowledge of actual size vs. apparent size, motion parallax, and other flat-image cues instead of stereoscopy?

But, yeah, LIDAR cuts through all that, too.

You can use stereoscopic using the full width of the car as the baseline instead of eye distance. Humans have a pupillary distance of about 60mm which is good for about 10 meters (possibly much more: https://jov.arvojournals.org/article.aspx?articleid=2191614). A Model 3 is over 6 feet in width, so the pupillary distance is thus 30 times that of humans, and so should be good to about 300 meters, which is comparable to high end LIDARs (although the stereoscopic approach won't be as precise at those distances).

Anyway, if LIDAR becomes small and cheap, Tesla can just strap it on.

Waiting for someone to state that last point. Lidar is still not cheap enough.

That is until it rains and LIDAR falls flat on its face. LIDAR Is Great for training in perfect rainless conditions. For everything else we’ll have to use other tech, very likely camera based. Which is what tesla is doing.

Whether this will succeed is another question.

I find the "humans can do it with eyes" argument pretty weak. Humans also have brains, which are doing most of the heavy lifting. With something like LIDAR, you're shifting more of that heavy lifting into the sensors, so that you don't have to go as deep into trying emulate the human brain in processing. It's such an obvious point.

Yeah, a direct a distance measurement is much more straightforward than the integration of all the monocular and binocular depth cues human perception relies on.

Musk commented about LIDAR in a prior earnings call[1]. Tl;dr he believes that LIDAR is a short-term crutch that distracts from the real long-term challenge, which is the machine learning, not the sensors.

[1] https://www.youtube.com/watch?v=FaW85fIos64&t=21m40s

Tesla autopilot needs all the short term crutches it can get.

Supposed to be fully autonomous since, what, 18 months ago? How about that SF->NY demo?

He can say whatever he likes, the proof is in the pudding.

If they were anywhere near full autonomy that LA-to-NYC demo they keep promising would have already happened.

Elon mentioned on the Q2 conference call yesterday that specific demo is easily doable.

He also added that it would be sort of cheating, because that's a very defined route and he prefers it to be able to be much more dynamic. i.e. pick any two US cites and the car will drive you there.

There are a number of routes one can take to get from LA to NY. Elon is stalling because he knows that "Autopilot" isn't capable of making the trip.

>Elon is stalling because he knows that "Autopilot" isn't capable of making the trip.

I can only assume then that you have some kind of insider info to support your claim?

I'm not the one claiming that Autopilot is able to make a cross-country trip but delaying the test because it would be "too easy" to prove anything. If it were actually that easy, Tesla would have done it, because Musk has never passed up the chance for that sort of publicity.

Ergo, logically, if Tesla has not performed such a test, it is strong evidence that Autopilot is not currently capable of passing it.

Elon said the exact same thing in the Q4 2017 earnings call. And then said they would still do the demo in the next 3-6 months. It's been 6 months and this "easily doable" demo has still not been done. Tesla has shown zero evidence of their self driving program making any meaningful progress.

http://www.autonews.com/article/20180208/MOBILITY/180209770/...

>Tesla has shown zero evidence of their self driving program making any meaningful progress.

Google's autonomous efforts far precedes Tesla, yet they have not out in the market yet. I would say on the contrary, they made significant strides in a short amount of time even after a complete overhaul Mobileye (AP1) to an in-house solution (AP2/2.5) https://vimeo.com/192179727

Has Tesla demonstrated any progress whatsoever since that 3 minute demo video they posted over a year and a half ago? If they were really making amazing progress they would have shown it off. Tesla LOVES showing off their tech when it's actually working, and it is often very impressive.

Waymo has been operating a real actual passenger service for over a year now. They drive tens of thousands of miles a day. They're "not out in the market yet" only in the sense that they're not charging money for it yet, which they seem to plan on doing soon.

https://www.bloomberg.com/news/features/2018-07-31/inside-th...

Is "finding roads between two arbitrary cities" the hard part of self-driving?

I think that "using arbitrary roads" is a relatively hard part of self-driving.

That's navigation, and that's actually a solved problem. Actually following the roads...oh look, that's a nice set of traffic cones. What could that possibly mean? IDK, just continue through it.

Still not going to work for autonomy. Expect better ADAS and nothing else. The computer vision problems they need to solve to make autonomy work reliably are not solved yet, and I would be willing to bet good money that they won’t be able to get anywhere near L5 with the sensors they install today and without a LIDAR.

Are they hardware limited? It seems like the limiting factor is the convnet rather than the hardware...do they think it would be more safe with less latency?

What is Waymo using? Custom ASIC?

Waymo pretty top-secret about their hardware, so I'm not sure. I would venture to guess that if they don't have a custom ASIC yet they're working on one. I know they had Intel FPGAs and it's a safe bet they've integrated Google's tensor core architecture.

Wasn't a bunch of stuff declassified during their lawsuit with Uber?

A bunch of details about their lidar came out in the depositions, and lots of juicy gossip about their internal politics, power struggles and backroom dealings. I don't recall reading anything about their compute.

Waymo's advantages include LIDAR and extremely detailed maps, IIRC. If I had to guess, those maps probably help them distinguish between roadside objects that happen to be static right now and ones that are really not going anywhere.

Reply on news.ycombinator.com