Software Update 2018.39 4a3910f (plus other v9.0 early access builds)

[verygreen]

Access to the firmware bits gets you precisely nothing without hacking the car to install it. That has nothing to do with "very poor security practices".

not true. Zero hacking required to install the firmware once you have the binary.

 

[gearchruncher]

They've had stereo vision for over a year.

They do not have stereo vision. Even if they did, the cameras are only a few inches apart, which is much too short of a baseline to do anything useful more than a few feet away.

 

[NuclearPowered]

not true. Zero hacking required to install the firmware once you have the binary.

Go on...

I presume the binary is floating around online as well?

 

[SlicedBr3ad]

not true. Zero hacking required to install the firmware once you have the binary.

What would stop someone with that binary and steps from installing it? If person is Handy with software, would they be able to do it?

 

[MP3Mike]

not true. Zero hacking required to install the firmware once you have the binary.

So what if someone gives you the binary you just put it on a thumb drive and stick it in the USB port and the car recognizes it prompts you to upgrade?

 

[OPRCE]

That is correct, so far Autopilot has used mostly the radar for TACC and radar has this inherent limitation (at least the radar they're using, and automotive radar in general really).

Well it's a fundamental design flaw but it can in theory at least be overcome with the cameras. It's just really hard to overcome with the cameras reliably; it's a big problem. Now if they had lidar it would be much easier (still hard, but easier). They may need to upgrade the processing power even on non-FSD cars in order to get an adequate vision solution that can both stop for stopped cars and not have excessive false positives on signs and overpasses. Also they will clearly need a lot more time to develop the vision systems; there has not been much progress made on this recently. If anything I think we have regressed slightly on false positives; since the high-profile accidents (slamming into fire trucks) I think they have turned the dial to be more sensitive, leading to more false positives and phantom braking.

Cheers, that makes a lot of sense!

"It's just really hard to overcome with the cameras reliably" -- I am thinking it may indeed remain impossible to 100% preclude with the current sensor suite in certain lighting conditions, e.g. when driving straight into a low sun, which will effectively blind all the forward-facing cameras at once. NB that the Walter Huang smash and another very similar one in the same area occurred, according to this report, under such lighting:

Maybe a circular polarising filter over the lenses would help reduce the phenomenon, along with a recognition by the system that it has been thusly blinded leading to an early "Red Hands" warning to resume manual control?

Or maybe we should start lobbying already that particularly Tesla and other would-be manufacturers of AD systems must include LIDAR sensors in their systems in order to achieve demonstrable safety/reliability in their potentially lethal machines? Seems to me reasonable to at least make space for the LIDAR sensor behind the front bumper of all current Tesla models before too many have been built to ever retrofit them all.

Of note here is that Musk [while recently cruising through a few stop signs] has stated that “it’s important to emphasise it [Autopilot] will never be perfect, nothing in the real world is perfect. But I do think that long term, it can reduce accidents by a factor of 10. So there are 10 [read as 90%] fewer fatalities and tragedies and serious injuries. And that's a really huge difference."
Tesla CEO Elon Musk addresses autopilot system safety concerns: "We'll never be perfect"

Which sounds very much like a trite excuse for glossing over and accepting the ongoing fatal weaknesses of Autopilot, while "consoling" [read fobbing off] the bereaved/maimed with a mess of tendentious statistics juggled by his PR department.

Methinks a more relevant measure of Autonomous Driving systems than Musk's "quotient-of-accidents-per-distance-driven-compared-to-human-drivers" is the Shashua's concept of "demonstrable safety" ... if the machine does not cause [or allow avoidable] accidents within its usage domain then it is 100% safe, and if that usage is based on specific, commonly accepted rules, it can be shown from a reconstruction of the onboard sensor logs with a mathematical certainty in any given crash scenario to have performed 100% safely.

For those who have not yet seen it here is some chewful food for thought in a recent symposium from the latter, an expert authority who works to define and design such safe AD systems:

Prof. Shashua on Safety of Autonomous Vehicles at 2018 DATE Conference, Dresden, Germany

 

[GWord]

They do not have stereo vision. Even if they did, the cameras are only a few inches apart, which is much too short of a baseline to do anything useful more than a few feet away.

Thanks goodness our eyes are three feet apart or else we might not have depth perception.

 

[Silicon Desert]

Ran outside to my driveway in my boxers to connect my car to my mobile hostpot. I hope the neighbors didn't see me

BREAKING NEWS FLASH: Newport Beach police looking for a man facing indecent exposure charges by running around in his underwear. Identifying clues are that he has "MODEL X" embroidered on his underwear with a gull wing flap !

EDITED: Seriously though. I'm thinking there would be no need to connect to mobile hotspot unless you have poor or no cellular connections in your area. The car should revert to cellular if it lost your WiFi connection.

 

[gearchruncher]

Thanks goodness our eyes are three feet apart or else we might not have depth perception.

Human depth perception is completely gone at 30 feet. That's less than 1/3 of a second at 70 MPH. You don't use depth perception to drive a car. Humans use depth perception for fine tasks at close distances.

 

[Theflash95]

You may be right that FSD is several years away but I think you might be slightly off about the reasons why. Musk's tweet doesn't say that they are struggling to get drive on nav to work on limited highways. Actually, if you watched the videos, drive on nav seemed to work pretty well on limited highways. So the issue is not that they can't even get it right for limited highways. If FSD is years away, it is not, as you say, because they can't even get the limited cases right. The issue is that they want to get it right for more than just limited highways. If you look at Musk's tweet he says it is "Extremely difficult to achieve a general solution for self-driving that works well everywhere.". Note the emphasis in bold. They are looking to expand drive on nav to work everywhere. That is why it is so difficult. So if FSD is years away, it's because they can do the limited cases just fine but need to make it work for all those edge cases too.

If anything, Musk's tweet reveals his ambitious vision again. He is not content to release a self-driving feature in a very limited case like many other auto makers have done. He wants the self-driving feature that "works everywhere".

Musk's tweet also seems to imply, at least to me, what I have been thinking now for awhile. "drive on nav" is intended to be the bucket that all the self-driving features will go into. That is why he calls it a "general solution for self-driving". It is called "drive on nav" or "navigate on autopilot" because it is intended to be that "drive from A to B" that makes FSD work. So right now, it can do auto lane changes on the highway, but I suspect that Musk also wants it to do auto lane changes on other roads as well and eventually be able to make turns at intersections, stop at red lights etc..

The bottom line is that "navigate on autopilot" is a critical feature that has the potential to make or break "highway self-driving" so Tesla needs to get it right.

I know you're all gung ho about this FSD thing, but... drive on NAV from all the videos/notes I've seen was already purposefully limited to only certain highways... my guess the ones without crazy ramps/turns/markings... so... by limiting drive on NAV to what Tesla considers ideal highways they still are encountering issues and have pulled back the feature.

Here is the whole quote: "V9 moving to wide release now. Holding back Autopilot drive on navigation for a few more weeks of validation. Extremely difficult to achieve a general solution for self-driving that works well everywhere."

Again, sounds to me that even though Tesla limited drive on NAV to only certain highways they are still finding it difficult to reliably offer the feature in their limited scope. Remember, EAP can already handle keeping you at a set speed on the highway and changing lanes when you initiate it, it can even slow you down on an off ramp if you have just the TACC enabled (which definitely isn't perfect as is)... drive on NAV is expanding that a little by letting the car make lane changes for you (probably with user confirmation first) and take exit ramps also (I did see one user report that this was poorly executed as the car was suggesting to be in a left lane to avoid slow traffic, but too close to the exit ramp to reliably/safely get over for a right side exit). In the end we'll see how long "a few more weeks" of validation is and what we get when those weeks have gone by.

Hey, hopefully I'm wrong and EAP is perfected before we know it and FSD comes hot on the heels of that, but I'm not seeing it. I see very slow (understandably... you could injure/kill someone or damage the vehicle with a wrong move) incremental updates to autonomous driving and I'm fine with that. Like I've said before Tesla shouldn't have sold FSD and instead waited until it was ready... now people have them under the gun to get this system developed and it's not easy.... or should I say: "Extremely difficult to achieve a general solution for self-driving that works well everywhere."

 

[ucmndd]

And still only going to 2.5. Tesla really likes spitting in the faces of 2.0 customers. They aren't even updating 2.0 to 36.2.

Remember that the composition of Tesla’s fleet is very heavily tilted to AP 2.5 at this point. The tiny sample you’re seeing by looking at Teslafi might just be representative of that fact.

My AP2 car has been on 36.2 for weeks.

 

[diplomat33]

Again, sounds to me that even though Tesla limited drive on NAV to only certain highways they are still finding it difficult to reliably offer the feature in their limited scope.

Except the quote you have from Musk literally says the exact opposite. It says that they cannot reliably offer the feature everywhere.

 

[jdchi]

Was he able to update to 39.5 for you after pushing 36 first?

No, because of the timing. He could only stage 36 and then I had to wait for it to download and then install. After he left, I sat in my garage waiting for the download to start. After 15 minutes I gave up and drove to work so I can afford my Tesla.

 

[nemSoma]

not true. Zero hacking required to install the firmware once you have the binary.

That depends on what you call hacking. Yeah sure you don't need root, but you still need to do some computer work to give it to the car.

 

[mitchellh3]

Remember that the composition of Tesla’s fleet is very heavily tilted to AP 2.5 at this point. The tiny sample you’re seeing by looking at Teslafi might just be representative of that fact.

My AP2 car has been on 36.2 for weeks.

If you look only at the AP2 samples on Teslafi, 36.2 is still a low %.

I'm AP2 and still on 32.2, have been for nearly a couple months. No sign of 36.2 or v9 in sight.

 

[OPRCE]

They've had stereo vision for over a year. Only AP1 seems to slam into firetrucks. 2.0 lusts for gorepoints because it couldn't properly identify navigable road paths. 2.0 is improving, 1.0 seems stagnant.

My car has all the latest hardware and it tried pretty damned hard in August past to run me headlong into fully stopped 2-lane traffic on the Autobahn [ as described here, Phantom braking, etc. overflow from 2018.32.2 thread ], so AP2.5 certainly still had Firetruck Super-Destruction mode in v.2018.28.5. Don't know [ but severely doubt ] if this situation has improved much in the more recent software updates? Problem being it is quite rare and risky in natural highway driving to properly test this danger scenario.

 

[croman]

My car has all the latest hardware and it tried pretty damned hard in August past to run me headlong into fully stopped 2-lane traffic on the Autobahn [ as described here, Phantom braking, etc. overflow from 2018.32.2 thread ], so AP2.5 certainly still had Firetruck Super-Destruction mode in v.2018.28.5. Don't know [ but severely doubt ] if this situation has improved much in the more recent software updates? Problem being it is quite rare and risky in natural highway driving to properly test this danger scenario.

2.5 uses a different radar than 1.0 and 2.0. it is still being fine tuned by Tesla.

 

[rnortman]

2.5 uses a different radar than 1.0 and 2.0. it is still being fine tuned by Tesla.

More than a year after they started delivering cars with this hardware, they are still tuning it for basic safety functionality... but don't worry, FSD is right around the corner!

 

[verygreen]

Go on...

I presume the binary is floating around online as well?

I cannot help you with the binaries.

The procedure to install is simple.
1. plug your computer into the in-car network
2. contact updater service on the car to ask it to install a new firmware from url you control
3. once the yellow clock icon signifies the update was staged - install whenever.

 

[Pale_Rider]

“2. contact updater service on the car to ask it to install a new firmware from url you control”

*presses voice control button* “hello updater service, please install update”

Lol, made me laugh anyway. I’m also heavily drugged for a nasty sinus inspection, so that may have something to do with my humor being off. Lol