New Tesla software to offer 'full' autonomy, Musk says | Hacker News Reader

New Tesla software to offer 'full' autonomy, Musk says

yahoo.com

25 pointsuptown89 comments

> "To date, Autopilot resources have rightly focused entirely on safety. With V9, we will begin to enable full self-driving features," he said.

Running cars into medians at 70 mph is the result of "focus[ing] entirely on safety?" Now they're going to focus on other things, besides safety?

"Now, our cars will focus on Jersey barriers as well."

yeah like collision avoidance, which is not safety /s

n=1

The sample size is irrelevant. The NTSB report reveals that the Tesla crashed into the median not because there was a glitch, but because the design of the system fundamentally wasn't capable of avoiding it. Tesla can't stop in response to stationary objects (likely because it can't tell the background apart from real obstacles and enabling breaking would cause too many false positives). That design is only safe if you have a human in the loop to do what the computer cannot do.

What seems weird to me is that when the car would get closer to the object the probability that a crash would happen should increase, Tesla could not get a high enough crash probability at all, even from 1 meter or 1 millisecond before impact and continued to accelerate.

There was a good analysis of this elsewhere that effectively boiled down to how the tech being used (I want to say LIDAR, but it could have been lasers or cameras) had evolved. It's apparently really good at detecting how fast other objects are going, and really bad (or, at least, not designed well in its current state) at detecting stationary objects, since it's primarily been trained to ignore the millions of stationary objects all around it throughout a drive to be able to focus on moving objects (cars, pedestrians, bicyclists, etc).

Apparently, moving and stationary objects are treated differently, and both have to be intentionally trained, and that this was a case of seeing the stationary object but ignoring it, just like it does with most other stationary objects that aren't in the road. I'm not technical enough to go into specifics, but that's the gist from the article I saw-but-can-no-longer-find.

You're mistaken, that was about radar not LIDAR. Telsa cars have no LIDAR.

The radars in use have poor angular resolution meaning they can tell if something is moving toward or away from it, but cannot tell where that object is. Since driving down the road will necessarily involve many objects on the side of the road, but the radar apparently can't tell where they are because of it's poor angular resolution, the 'solution' was apparently to filter out everything moving towards the sensor at the car's current speed (e.g, all stationary objects.)

It's unclear to me what if any circumstances will cause the emergency breaking system in a Tesla to fire.

When is Tesla going to bite the bullet and issue a full recall so they can retrofit all existing Tesla cars with LIDAR?

They are still claiming on their website:

>"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."

...but without LIDAR that simply isn't true. Either the accident rate is going to accelerate, or their efforts to push out "full self-driving capability" will stall. And from what we've seen from this company so far, I'm afraid it's going to be the first. A recall to retrofit LIDAR is the only way they can save themselves from their marketing lie. They need to bite the bullet and fess up.

Tesla disagrees with you on this, they are still operating with the belief that cameras will eventually be superior and cheaper compared to LIDAR.

https://www.theverge.com/2018/2/7/16988628/elon-musk-lidar-s...

I know they disagree with me, they have multiple reasons to bury their heads in the sand. Admitting they were wrong would be a PR disaster, it would probably hurt Elon's ego, and a recall to retrofit LIDAR would probably be extremely costly.

But if you talk with the leading experts in the field of car automation, the research groups who are actually making it work to some degree today, they are all using LIDAR. Furthermore you have people like the CTO of Mobileye saying that Tesla was trying to push Mobileye's tech beyond what it's capable of (past tense was, because Mobileye apparently dropped Tesla over this, prompting Tesla to develop their own presumably equivalent non-LIDAR solution in house): https://arstechnica.com/cars/2016/09/tesla-dropped-by-mobile...

No. LIDAR has its place, but it certainly a tech that has seen its day. Integrated sensor arrays consisting of multiple cameras and image processors can build more accurate situational awareness pictures faster. And to imply that Tesla's automation assist is not working is ridiculous.

I think you have an unrealistic understanding of the state of visual processing systems today.

The hardware simply isn't capable of the processing necessary to develop 360-degree situational awareness at the speeds that a vehicle would be driving.

On the contrary, LIDAR is only recently coming into its own:

https://www.bloomberg.com/news/articles/2017-01-08/alphabet-...

https://arstechnica.com/cars/2017/01/googles-waymo-invests-i...

What I'm 'implying' (saying outright I thought) is that Tesla hardware is unsuitable for the continued advance of automation functionality. What they're doing currently is already pushing their inadequate sensors to the limit.

Are Tesla cars equipped with enough processing power? I agree that with some good cameras a ton of processing power and a super AI you could avoid crashes, I mean Tesla hit a fire Truck, the cameras would have been enough to detect the obstacle but the AI or the hardware needed to process the camera output is years behind what it should be.

Well, that’s not what the article says:

“Musk reiterated his dislike of LIDAR and defended Tesla’s strategy of achieving “full autonomy” using only cameras, radar, and ultrasonic sensors.”

At this point, nobody really cares what Tesla (or specifically, Musk) thinks about the virtues of cameras vs LIDAR.

Autopilot's can't even sense stationary objects in front of the car. The most basic task of a self-driving system. Getting visual AI up to that point will require computing technology significantly more capable than what Tesla has on any of its vehicle--something on the order of two or three CPU/GPU generations more capable.

By the time the visual processing hardware is up to the task, the cost of a LIDAR system will no longer be the most expensive part of the car. Hell, there are already systems in pre-production that now cost less than $2000/sensor, before economies of scale.

They're going to say that until they go full-LIDAR having purchased a guarantee from major suppliers they can later vertically integrate.

Just like Apple did with so many things.

I really am of the belief he talked Tesla into a corner here. While LIDAR is bulkier it offers so much beyond RADAR and cameras that its not truly up for debate. I could ramble but this article Wired[1] does it properly

As for today's announcement I don't know what to say other than maybe they are trying to get in front of something they know is coming down from the NHTSA. I certainly hope they don't try to bluff them

https://www.wired.com/story/lidar-self-driving-cars-luminar-...

I believe Elon has stated it should not be necessary. However, he also tweetes if his compsny comes to find there is no other way without Lidar they will figure a way to upgrade the hardware at no additional cost to those who have prepaid for it.

There's no way to update the Model X for LIDAR. It's structurally incapable of supporting any of the existing LIDAR systems in production or currently in development. The same could arguably be said for the Model S, due to the front monopane windshield.

Even if these vehicles could be modified to support LIDAR systems, the modifications alone would likely cost upwards of $10,000 per vehicle, not including the cost of the LIDAR system (currently $20,000 for cheapest system offering 360 coverage). Those costs would bankrupt Tesla.

At some point, Musk needs to learn to keep his mouth shut. A sales person making these types of claims can be disregarded on the grounds that a sales person may not have the authority to bind the company. But if the CEO says the same things, customers who purchase cars based on the CEO's representations can definitely sue the company for failure to deliver promised goods.

Then all Musk needs to do is not allow Tesla to "come to find" out that LIDAR is necessary before the technology has advanced to such a degree that they can retrofit the Model X and to do so when the cost has come down sufficiently to afford it.

Which is what he's currently doing. When they do finally admit that LIDAR is necessary, Elon will play the "no body could have foreseen this situation and we were operating under the best knowledge at the time" card.

He's bullshit all the way down.

LIDAR is not magic. You can get full situational awareness with boring old video these days. LIDAR was just easier to characterize back in the day before gigahertz ASIC AD video processors.

LIDAR may be more expensive, but it has one huge advantage: it can identify stationary objects in 3D space.

For a visual processing system to accomplish the same task, it must first: identify objects in a frame. Identify objects in the next frame. Identify objects in the images taken from the other cameras. Compare the objects, taking into account positional differences arising from the speed of the vehicle, and triangulate their positions relative to the vehicle.

This is certainly doable (Microsoft Photosynth has been around for at least a decade), but each one of these tasks, and especially the last task, requires more processing power than Tesla has put into its vehicles.

Interestingly, it can be done even with a single sensor. We can map object’s 3D positions with just one eye. The brain uses the relative motion of various object in the scene due to forward motion induced parallax and other clues such as object size. It’s possible to drive with one eye covered, but might as well use multiple sensors if available.

See “Depth Perception” at [1].

[1] https://en.m.wikipedia.org/wiki/Depth_perception

Motion is not required either. There's a reason for a shallow DoF in big photographic lenses/sensors. Bokeh can be represented as a processed depth map.

Your brain is using something in the ballpark of a petaflop of computing power to do that sort of incredible stuff.

I have perfect vision in one eye and very poor vision in the other. While it's possible for me to drive without my glasses, I don't consider it safe. With one eye I can read signs, see pedestrians, cars, etc; but I find that the accuracy of my depth perception suffers. I estimate speed and distances with less precision. Try catching a baseball with one eye covered. It's possible but a lot harder.

“One eye”, in humans, has about 125 million sensors (120 million rods, 6 million cones) Calling that “one sensor” is cheating slightly.

But yes, you only need that second eye for stereo vision if you only have data from a single moment in time.

In theory, you can see perfect 3D with a single photoreceptor by scanning with it (giving up temporal resolution, a bit like a electron gun that can write a million pixels per second writes only 50 or 60 images a second on a tv)

By your logic, a 125megapixel camera arguably has 125 million sensors too?

Assuming it’s a color camera, at least double that :-)

I'm worried he's talking about going from what the NHTSA calls level 2 to level 3 autonomy[1], where the car is capable of navigating all the way to the destination but the driver has to remain alert for potential dangerous situations. That's probably an even worse idea in terms of people actually complying than the existing level 2 autonomy and I think Waymo is absolutely right in trying to skip straight to level 4/5 autonomy where you don't need a human paying constant attention.

[1]https://www.nhtsa.gov/technology-innovation/automated-vehicl...

We need to work with some real numbers.

In the US there are about 1.2 fatalities per 100 million miles driven [0].

About 19 months ago, Tesla said their cars logged 1.3 billion miles on autopilot [1].

I'd guess that autopilot number has at least doubled by now.

So if there have been far fewer than 26 people killed by autopilot, then it's already saving lives.

[0] https://en.wikipedia.org/wiki/Motor_vehicle_fatality_rate_in...

[1] https://electrek.co/2016/11/13/tesla-autopilot-billion-miles...

That's the number of miles with autopilot-equipped hardware. It's only 100 million miles with autopilot engaged (from the article you cite). Let's say its 200 million now. They're at two confirmed deaths, which is about what you'd expect.

To the extent the numbers are statistically significant, it's worse than you'd expect. The fatality rate for Telsa drivers is 1/4 the average. Since the overwhelming majority of miles on autopilot equipped hardware are driven with autopilot disabled (per your article), that means Tesla human drivers are a lot safer than the average. So you need 1 fatality per 300-400 million miles driven just to match human Tesla drivers.

This is a deceptive use of statistics because it's papering over all the nuance involved. Not all road miles are equivalent. Specifically, Tesla Autopilot only functions (or is meant to function) in a very narrow subset of all conceivable driving conditions. Many of the situations that Tesla Autopilot isn't meant for are the most dangerous conditions that unassisted humans drive in.

Furthermore the statistic doesn't break out the role of "Autopilot" features from the role of other safety features present in Tesla cars but not present in many other cars. More effective airbags, crumple zones, emergency braking(?) could all be lowering the fatality rate in Tesla cars enough to mask the "Tesla Autopilot" making driving more dangerous than if it weren't present.

Tesla fatality rates should be compared against equivalent class cars when driven by humans under the conditions which Tesla Autopilot are meant for.

> Not all road miles are equivalent

Great point.

I'm looking for numbers and found this story [0] that says the fatality rate on rural roads is 2.5x urban highways.

How many miles are driven on highways and how many fatalities? Does one human driver die per 100M, 200M, 300M miles?

Should the standard for autonomous driving be 100%, 50%, or 10% of the human-driver standard before we agree that it's beneficial?

[0] https://www.npr.org/2009/11/29/120716625/the-deadliest-roads...

Also one still needs to include the impact bad wheather conditions, where Tesla's AP is not to be used either. Then there's the impact of "incompetent" drivers on the statistic. As a toy example, consider a world where competent drivers cause 1 death per 100mm miles. AP causes 2 deaths per 100mm miles. However, 20% of all human drivers are driving while drunk and/or sleepy and cause 10 deaths per 100mm miles. This gives an average of 2.8 deaths per 100mm miles. Clearly, AP has a better average but you should only use it while you're drunk.

Tesla's Autopilot miles are highway miles, but Tesla is comparing those miles to fatalities across all types of roads--freeways, highways, normal streets, dirt roads, etc. On the safest type of road (i.e., freeways, which have the fewest obstacles or impediments to driving), Tesla is no better than the national average over all types of roads.

And that's horrific. It means that Tesla isn't saving lives, it's causing more people to die than would have died if they hadn't been using Autopilot. Worse, as Tesla's many competitors have demonstrated, is that self-driving cars can be safer than human-drven cars. Just not the way Tesla does it.

Well, this is going to get interesting.

Blaming the customer might have worked for regular AutoPilot because of that legal footnote that AutoPilot-the-product is it not intended to be used as autopilot-the-concept, but good luck blaming the customer for a car that was in "full autonomy" mode.

Edit: the timing of this announcement is so very premature, I believe there might be another motive to it. Perhaps Tesla wants to negotiate from a stronger position in the next financing round.

"...With V9, we will begin to enable full self-driving features," he said.

"...begin to enable..." - bit of a jump to "full autonomy"

Begin to enable*

*Terms and conditions still apply, so it's totally your fault if your get into a crash while in the new "fully self-driving" mode.

"full self-driving features"

What does that mean? The statement reads like it was carefully crafted to mean whatever Elon needs it to mean when it's time for it to mean something.

What's the point? Even if it were to be true, there's no regulation and approval for it to be used on public roads. Seems like another PR stunt from tesla to induce hype.

And it's working. I've seen so many tweets and articles "Tesla announces fully autonomous cars" that I'm genuinely worried about what would happen if this announcement is actually fulfilled.

If it is true that Tesla will - by design - not try to stop in front of a stationary obstacle that should not be on the road, I don't understand how come Tesla autopilot and similar self-driving techniques are not immediately banned. Every driver knows stationary obstacles can and will appear on the road.

I'm a big Tesla fan (I have a Model 3 preorder) but this seems like a really dumb move in light of recent press coverage.

To make it clear: the headline is completely wrong. Musk only announced a step towards autonomy which could be anything, but certainly not full autonomy.

It could be as simple as extended self parking.

From a software engineering perspective it is an interesting announcement. With all the obvious limitations of the current autopilot, Tesla must have been working on a very different package for full autonomy - we might see a first glimpse of it with that release.

Exactly! It is absurd how people hear Elon Musk say, "we will begin to enable full self-driving features..." and then conclude that Elon Musk was talking about full self-driving features. They are all just Tesla haters!

He said “full self driving features”, not “full self driving”. With the word “features” he refers to components of a system that will offer self driving, but not the whole system yet. These features could be better self parking, or traffic sign recognition, or the long talked about feature of changing between freeways autonomous.

Exactly! Only Tesla Haters would think that "full self-driving features" would refer to features that enable fully autonomous driving... every common sense consumer out there will clearly understand that when Elon Musk says full self-driving features he clearly is referring to parallel parking helpers and such. All these Tesla Haters would have us think that Elon has a habit of going around raising unreasonable expectations through grand pronouncements when he's clearly just trying to tell consumers that soon they won't have to worry about parallel parking their cars!

When does this level of rhetoric approach criminal negligence? From the previous HN discussions about Tesla’s,s running into stationary objects, the problem seems like a fundamental design issue not likely to be fixed with software. This has already killed people. What happens when some poor person relies on Musk’s ‘full autonomy’ and gets killed?

Elon Musk is the Elizabeth Holmes of self driving cars.

The headline misquotes Musk. He did not announce full autonomy.

That's exactly what he announced:

> With V9, we will begin to enable full self-driving features

The caveat ("begin to enable") doesn't change things. You might be at the beginning of pregnancy, but you're fully pregnant.

> The caveat ("begin to enable") doesn't change things. You might be at the beginning of pregnancy, but you're fully pregnant.

That would be correct if the Musk quote said they would "begin to enable" full autonomy. That assertion is distinguishable from beginning to enable features related to autonomy.

You're changing the meaning by adding the word "related." Say Microsoft stated it would "will begin to enable full 8K features" on the XBox One. That would mean that something would be in 8K--maybe just for movies but not games. It wouldn't mean "features related to 8K" but not a actually being able to display 8K video anywhere.

Your interpretation seems to be that if the car is ever in a "fully autonomous" state, then it's "fully autonomous". Thus if a "fully autonomous feature" is enabled (such as on the highway), that's full autonomy even if that feature is disabled automatically in other situations (such as in the city). Some people would interpret "fully autonomous" to mean end-to-end autonomy, sort of like enabling a "full encryption" feature on a chat application doesn't make chats on that application "fully encrypted".

[EDIT: Given the medium and proposition, the appropriate interpretation seems to be the narrower one.]

That is my interpretation, yes. "Fully self-driving features" means the car is self-driving at least some of the time. So a fully autonomous mode that's only active on the highway in good weather counts, but not modes where the human still has to pay attention, or features that only support but do not actually enable full autonomy at least some of the time.

Even limited full-autonomy seems totally out of the reach of Tesla given what we know to date.

Whether or not rayiner is correct in his interpretation, I think it's pretty clear his interpretation is not an outlandish one. By that I mean many other people in the general public will share his interpretation, and unlike rayiner many of those people will be blissfully unaware of the problems with automation Tesla has been having.

What this means is that a large number of people are going to be deceived by statements like this from Musk, even if you think what Musk is saying is technically true, because they interpreted/misinterpreted Musk's statement the same way rayiner did.

If say, Tesla would now be able to parallel park all the time without human intervention or oversight, would that not be considered a fully autonomous feature?

So long as it could park without any human oversight, yes, I'd consider that to fit Musk's statement. But will Tesla be even there by August?

Yes, they will be. Q2 is going to be negative again, but Tesla has enough cash left for that. From all announcements it looks as if Tesla has made all investments needed into the Model 3 production to reach the 5k cars/week production goal at the end of the month. At that rate Tesla should be slightly profitable.

No, I mean, will Tesla be able to have something it'll be able to advertise as "full self parking" by August? That doesn't run over stationary dogs/children?

That's a valid question/concern but that we're addressing the correct interpretation of Musk's statement.

If Microsoft said they'd begin to enable 8K on the Xbox One and released a feature that let you view 8K images (with the assumption that more features will come in the future), that'd be completely reasonable.

If Tesla said they'd begin to enable self-driving functionality and released a feature that you complete some part of that like parking for you, or bringing your car up front (with the assumption that more features will come in the future), that'd also be completely reasonable.

I don't understand the jump between "beginning to enable full self-driving features" and turning on full autonomy in any near future. Obviously, Tesla has been releasing functionality piecemeal when they can; see: "features".

Some features of self driving, not the whole package. Eg. a better parking assistant.

It seems like there's quite a difference in self driving technology between highspeed and lowspeed ability. Generally low speed seems to be pretty safe. But apparently all high speed systems have the "blind spot" of having to disregard stationary objects in front of them due to false positives from overhead highway signs, flyovers etc.

So when there's a broken car in your lane on the freeway the autodriver will just slam right into it.

I doubt the public will accept any system that has this kind of inherent flaw because it seems just insane to a human driver. This will probably have to be fixed before highway driving is approved but city driving, with a speed limit, might happen much sooner.

Wake me up when I can tell my car to go get serviced and come back by 5 so _I_ can drive home.

Does he mean full-autonomy as in: "enter a destination and be done with it"? That would be an amazing jump forward from the current system.

That would be pretty amazing, considering that the hardware in a Tesla is incapable of detecting and avoiding stationary obstacles.

Nope, better highway merging of lanes.

Maybe in the event of a crash it can automatically issue a press release blaming the driver?

It will probably be catching up to the Audi A8's traffic jam pilot. I.e. full self driving under certain limited situations.

"Full self-driving in limited situations" - do you even see the contradiction?

I don't think that is a contradiction. It means that there are some situations (probably traffic jams) where the car can fully drive itself, instead of merely assisting you as you drive the car.

If there's a need for a human to take over (e.g. traffic jam dispersed and the car can no longer drive autonomously), that's not full autonomy (Level 5), that's weasel-autonomy (or, more charitably, Level 4): "autonomous, autonomous, auton-gotcha, not autonomous, that block straight ahead is your problem now." That does sound familiar, doesn't it.

In the last 12 months it’s become clear that Musk relies on the same defensive self-aggrandizing bluster that makes Trump so ridiculous. Quite a quick trip down from demigod for him.

There's always been an element of this to Musk's marketing, but it's gotten much more obvious lately, alright.

Given that Musk and his companies have accomplished some amazing things. Did you see the simultaneous rocket landings?

So 'self aggrandizing bluster' may not be the phrase the you're looking for. How about 'low cost, highly effective guerilla marketing' based on past performance?

> "low cost, highly effective guerilla marketing' based on past performance?"

I'm not sure that quite covers it either, since at least in the case of SpaceX he had a sizable number people rooting for him back when his rockets were still blowing up every time. Back in the Falcon 1 days.

I was one of them. I think rockets are really fucking cool. I think a lot of the leeway and fandom Elon has gotten comes from a hardcore contingent of rocket enthusiasts and electric car enthusiasts who really want to see him succeed because they were already passionate about that domain before Elon entered the picture.

In other words, wishful thinking > actual dead people, because they're not supposed to have died, therefore did not happen. Great Kool-Aid, mmmm.

Sure, his accomplishments are immense. But that doesn’t change the fact that he shipped an incomplete Autopilot and then tried to blame the guy who died using it.

It's not good to think of anyone as a demigod but there's a big difference between Elon and Trump. Elon very much has a problem with defensiveness and he tweets too much. Sometimes he's also wrong about things. But I'll give him credit for being both honest and idealistic which Trump certainly isn't.

>New Tesla software to offer 'full' suicide, Musk says

Why bother with autonomy crap? Just focus on making cheap electric cars that can save the planet.

because, as history has shown, that isn't a core competency of Tesla.

How Odd

Yahoo insist that for me to read this page they must store cookies on my machine [0]

However when I load the page in lynx and choose "neVer accept cookies", the content loads just fine [1]

[0] https://i.imgur.com/OIcmAby.png [1] https://i.imgur.com/ZdtpCxC.png

Noscript and/or umatrix kills that kind of crap.

Reply on news.ycombinator.com