Things my wife said about Navigate on Autopilot tonight

[Knightshade]

The fact Elon tweets about it so much gives me the strong impression that this is not being done 'on the side'

Then your impression is demonstrably wrong.

All of these games were developed by people who don't work at Tesla.

Most of them by people who worked at Atari. Decades ago.

And the Atari Linux emulator came out in the 1990s. Also not written by people employed by Tesla to write it.

The racing game is a mobile phone game ported by the game developer, not Tesla to run on Linux (which again is what the Tesla runs)

Likewise the upcoming games like Fallout Shelter.


Tesla's "coding resources" going to this amount to "loading the already existing software someone who doesn't work for Tesla wrote and adding a button to launch it"

I feel sorry for all the folks who bought FSD all the way back in 2014 only now to hear their $5K (?) is not going to get them anything after all.

... what?

FSD didn't exist in 2014.

It didn't go on sale until October 2016.

For $2000 less than you just claimed.

Probably, but it was supposed to be be much faster hardware, even for the existing implementation.

No it's not- just the opposite. Elon specifically said HW3 was slightly worse right now due to basically running code written for 2.x hardware.

"run existing code faster" is potentially a terrible idea for reasons already described.

The benefits from HW3 will be when it's running native code actually intended to leverage the significantly greater power of the hardware- not when it's emulating slower HW.

 

[Knightshade]

I find that completely unacceptable.

I'm sorry you find "using your cars features as the manufacturer directs and intends" to be unacceptable.

I don't think that's a reasonable opinion but I suppose you're welcome to hold it.

Cruise control should work on any roadway. There are many long, straight roads where it is too easy to exceed the speed limit thus the reason I need to use it. Never have I seen any mfg state that cruise must ONLY be used on highways or limited access roads.

And neither does Tesla.

Again you appear to be confusing autosteer and TACC.

TACC says nothing about limited access roads- just to avoid its use on city streets.

In reading the manual, I like how Tesla says:

"Traffic-Aware Cruise Control is primarily
intended for driving on dry, straight roads,
such as highways and freeways. It should not
be used on city streets"

yet a few paragraphs later it says it can be engaged at speeds above 18MPH. Yeah, right. We have SO many highways where the speed limit is 20 or 30 mph.

Where did they say anything about the speed limit. That's a minimum listed there, not a max.

Andthere's tons of highways where traffic speed is often 20-30 for periods of time.

Often long periods. Often even slower. Some places call this "rush hour"

Which is one reason you can engage it below 18 as long as there's a car in front of you.

they should at least give the option to turn off TACC and revert to plain old cruise control

Now here I agree with you. That SHOULD be something you can toggle... (and obviously if turned off AP wouldn't work)

But I also think it should be an option you have to actively change at least 1 or 2 deep in a menu- because otherwise you'll start getting people whining about how they rear-ended someone because they "didn't realize" they were only using CC instead of TACC if it's in any way "easy" to be unsure which mode you picked.

 

[DrDabbles]

It really makes you wonder if we're all driving the same car.

We are. This is exactly the sort of behavior you'd expect from a neural network, in fact. Because they're dumb pattern recognition machines, you basically need to show them an example of every pattern from every angle in every condition before they work properly. Combine that with the fact that Tesla is intentionally not manipulating how the system behaves with code, and you've got a car that behaves extremely poorly in tons of situations until Tesla get around to training it.

And this doesn't even get into the effect of regressions when descending the plane toward global minimums.

maybe I'm just a jerky driver, but my gf finds it noticeably smoother with autopilot on

If AP is better than you, then you shouldn't be allowed to drive. That's as honest an opinion as I can possibly give.

The hardware is currently too slow to process all the information and seems to make abrupt corrections while driving.

This absolutely is not the problem at all. The hardware doesn't have as many features as it will need, but that doesn't mean it's too slow and reacting abruptly. That's just not at all how these cars work. Processing too slowly would mean literally the car wouldn't work at all.

 

[DopeGhoti]

Tesla's "coding resources" going to this amount to "loading the already existing software someone who doesn't work for Tesla wrote and adding a button to launch it"

I agree with your larger point, but to be fair, the effort on Tesla's part to implement the games on the MFD was not zero.

True, the Tesla runs Linux, and a Linux Atari emulator is already 'on the shelf', but there are several hurdles that Tesla would have to clear. Most of these provide ancillary benefits to those involved, and add libraries that could be put to use in other contexts.

Off the top of my head, they would have to implement:

Custom Controls - None of the Atari games were controlled with a touchscreen in their native habitat. They would have to have been reworked from their original trackballs and joysticks and buttons. But this would also be a safe, low-cost (in man-hours) means to train and develop code to read input on the MFD; possibly with eyes to increasing precision and responsiveness. Beach Buggy Racing also would have needed time to wire the original native mobile game to not use a touchscreen, but rather a steering wheel and pedals for input. Again, allowing for testing and development of an API to read the car's inputs (the aforementioned wheels and pedals).

Driving Lockouts - The games are presumably locked out from being in play (heh) while the car is not parked.

Other vehicular interfaces - The fireplace similarly is a low-cost way to test a means for the software to read from and write to the HVAC settings and audio system. "Emissions testing mode" is similar, but more explicitly is dealing with audio balance and fader settings.

All that said, it's exceedingly unlikely that the folks working on this sapped any time from the team marking up sample images or working on tuning of the Autosteer neural network.

I do enjoy Tesla's irreverent sense of humor ("emissions testing mode" indeed), but not everyone will. And that's okay. It's okay to not like things. But it's not reasonable to assert that it has cost Tesla zero man-hours to implement these diversions in the toybox. Nor is it (in my opinion) reasonable to assert that because someone spent a few minutes finding fart sounds, significant time was lost in developing the means for the car to read a "Stop" sign that happens to have "Hammertime" scrawled below the original markings.

 

[DrDabbles]

When it halts us suddenly on a freeway we sometimes notice that our cruise control setting on the dashboard briefly changes from 70 to 25 or 35 mph and then flips back to 70. Is this some navigation system error? It's scary and dangerous. We only drive our Chevy Bolt EV locally.

The best explanation I've been able to come up with for this is when traveling near off ramps or surface streets, the GPS location is placing you not on the freeway, so TACC limits itself to the surface street limit + 5MPH. The first time I experienced that bug, I was being followed in heavy rain by a large truck and I thought for sure I was going to be run over.

 

[TeeEmCee]

As a wise man once said:

To err is human, but to really foul things up requires a computer.

Taking existing code and just "make do go faster" could have all sorts of unintended consequences if not properly implemented. Possible bad results, off the top of my head: What used to be a smooth, sane lane change now careens the car to the side of the road because the steering wheel is being turned ten times more quickly than intended. What used to be a 30-second nag is now a 0.3 second nag. I'd be quite surprised if HW3 Teslas on the road today were not undreclocked or otherwise somehow self-limiting to match HW2.5 performance expectations.

Hehe, that’s not how software works. You’re assuming that the execution of maneuvers is written to be dependent on the processor clock. It isn’t. For example the command “apply -5 deg steering input for 1.5 seconds” will take equally few CPU cycles on any hw gen.

Anyway, what I was talking about is the speed of detection. That’s what’s critical and it’s nowhere good enough in Teslas (nor in any other vehicle currently on the market) and it can only be made better with more processing power. Having the computer become aware sooner of everything around (or becoming aware of MORE of the things around) is never undesirable.

 

[TeeEmCee]

We are. This is exactly the sort of behavior you'd expect from a neural network, in fact. Because they're dumb pattern recognition machines, you basically need to show them an example of every pattern from every angle in every condition before they work properly. Combine that with the fact that Tesla is intentionally not manipulating how the system behaves with code, and you've got a car that behaves extremely poorly in tons of situations until Tesla get around to training it.

And this doesn't even get into the effect of regressions when descending the plane toward global minimums.



If AP is better than you, then you shouldn't be allowed to drive. That's as honest an opinion as I can possibly give.



This absolutely is not the problem at all. The hardware doesn't have as many features as it will need, but that doesn't mean it's too slow and reacting abruptly. That's just not at all how these cars work. Processing too slowly would mean literally the car wouldn't work at all.

You comment shows good understanding of how these things work (I’m guessing from professional experience).

 

[DrDabbles]

You comment shows good understanding of how these things work (I’m guessing from professional experience).

Not directly. I'm a systems architect by day, which basically means I need to do deep research of hardware and software systems in relatively short order. So in my spare time I've been reading research papers about batteries and machine learning.

 

[TeeEmCee]

Not directly. I'm a systems architect by day, which basically means I need to do deep research of hardware and software systems in relatively short order. So in my spare time I've been reading research papers about batteries and machine learning.

Then chances are you share some of my uneasiness with the faith people put in systems such as this one.

 

[AlanSubie4Life]

is the speed of detection. That’s what’s critical and it’s nowhere good enough in Teslas (nor in any other vehicle currently on the market) and it can only be made better with more processing power.

In general probably true once you get many detection tasks operating at the same time, but clearly not the issue for the original topic in this thread. As I said, it is clearly detecting and responding plenty fast. Too fast (suddenly), in fact.

Of course, in terms of reliability of detection, additional features, etc., it’s entirely possible (likely) more cycles and higher resolution processing would be helpful.

But for the topic here, there are clearly no limits imposed by the existing hardware.

 

[DopeGhoti]

I just fervently hope that at no time does Tesla say anything about TACC, AP, or FSD being 'idiot-proof'. An even wiser man once said, "Make something idiot-proof, and some moron will come along and invent a better idiot."

 

[Fernand]

You’re essentially saying you’re driving the car while dancing around Autopilot’s limitations and failures. You consider that to be normal and expect others to do the same.
It’s not exactly normal though. Most people don’t consider driving using thumbwheels and stalks to be fun, safe or normal. .

Using the speed thumbwheel a few times on say a 200 mile trip isn't much of a "dancing around". The alternative is to e.g. enter those messy cloverleafs in manual, or just let the car enter them at legal GPS-posted speed. User's choice, it all works. My way is smoothest.

As to the stalk, I vastly prefer that to jerking the wheel or brakes to cancel AP. It also happens to be THE prescribed, the most normal and safest way, to drop out and get back into AS. Sorry if you think that's too much for "most people".

A lot of the complainers seem to come from an implied showboat expectation that they can turn on automation as they leave home and effectively go to sleep as the car barrels over hill and dale. I regret to inform them that no car on the road can do that as yet. What's toy-like and experimental is pretending you don't understanding that.

I do often use the car outside its stated limits, I see these limits gradually widen, and I'm fine with both the limits and the overall results, no stress, no whine. Just enjoying my car.

I can't fathom the people who pontificate about how Tesla must disable say TACC/AS on city boulevards. Has it ever occurred to them to NOT USE IT where they don't want it??? And not to drive off cliffs so they can bitch that the car let them do it?
.

 

[TeeEmCee]

In general probably true once you get many detection tasks operating at the same time, but clearly not the issue for the original topic in this thread. As I said, it is clearly detecting and responding plenty fast. Too fast (suddenly), in fact.

Of course, in terms of reliability of detection, additional features, etc., it’s entirely possible (likely) more cycles and higher resolution processing would be helpful.

But for the topic here, there are clearly no limits imposed by the existing hardware.

That may not be entirely true.

I understand what you’re saying, the system knows it needs to speed up, slow down, turn etc., it just needs to do it more gingerly.

So it already knows it needs to initiate the action, however, as it’s executing said action, it continues to evaluate the environment. If its evaluation abilities are crude, then it will go with the safer option, i.e. “do it quickly, while the available data is still valid”. With more processing power and better awareness, it could better quantify the surroundings and choose to perform a maneuver more slowly, having determined that there’s plenty of time to do so.

The current jerkiness is surely not there because the Tesla testers (both of them, hehe) do not recognize the need for smoothness. It’s likely there because that’ the best that the AP computer can do.

 

[TeeEmCee]

Using the speed thumbwheel a few times on say a 200 mile trip isn't much of a "dancing around". The alternative is to e.g. enter those messy cloverleafs in manual ...
.

Your words, I believe: using the speed max roller as a limiter, and dropping in and out of AP, sometimes several times in a minute ...

Yes, that is “dancing around”. As another poster alluded to, if you’re horrified by them “messy cloverleafs in manual”, it might be safer if you didn’t drive at all.

Apropos that, what’s with all the hatred for the act of driving? To unsophisticated rednecks like me, it’s one of the most enjoyable things to do, which is why I’m looking to these gizmos to enhance the pleasure of driving, not detract from it.

 

[AlanSubie4Life]

That may not be entirely true.

I understand what you’re saying, the system knows it needs to speed up, slow down, turn etc., it just needs to do it more gingerly.

So it already knows it needs to initiate the action, however, as it’s executing said action, it continues to evaluate the environment. If its evaluation abilities are crude, then it will go with the safer option, i.e. “do it quickly, while the available data is still valid”. With more processing power and better awareness, it could better quantify the surroundings and choose to perform a maneuver more slowly, having determined that there’s plenty of time to do so.

The current jerkiness is surely not there because the Tesla testers (both of them, hehe) do not recognize the need for smoothness. It’s likely there because that’ the best that the AP computer can do.

This does not make any sense to me. If the system were that slow I don’t think it would work at all.

What sort of processing frame rate are you suggesting here?

The most obvious and likely reason for the deficiency is that it has not been focused on specifically and there are complications. I could see them being concerned about compromising AEB performance when smoothing is applied. You want to be able to react suddenly when appropriate (and there car is demonstrably capable of that, though obviously not perfect, as is the case with any of those features).

 

[DrDabbles]

Then chances are you share some of my uneasiness with the faith people put in systems such as this one.

I am indeed uneasy with the lack of understanding and over reliance. Research papers like this worry me quite a bit, because it means these networks are extremely fragile and need to be made much more robust. Making them more robust may have development, timeline, and performance impacts on the neural nets we use in increasing numbers in daily life, let alone potential regressions.

Anyway, enjoy. Single-pixel adversarial network attacks against neural networks causing serious mis-identification. Adversarial Examples Are Not Bugs, They Are Features

 

[Mikelespipes]

Wife: "Why is it so jerky?"

Me: "Well, it can decelerate somewhat abruptly while on freeway interchanges..."

Wife: "It's not just around corners, it's all the time on the freeway; why do you use it if it is so bad?"

Me: "Well, I'm trying it out; I think maybe you notice less if you're driving"

Wife: "Well, I notice it. It's ok to try it out if you're the only one in the car, but it isn't cool to do it otherwise"

Me: ...

Wife: If you keep trying to use it, I'm going to stop riding in this car, and we're going to have to take my car*. It's not relaxing.

Me: <disengages Autopilot for remainder of trip>

<a couple minutes pass>

Wife: It's much smoother now. Did you turn it off?

Me: Yes, I did.

Wife: I'm much more relaxed now.


* My wife's car is a Chevrolet Spark EV. That tells you something.

You can't make this stuff up!

Here's hoping that Tesla discovers low pass filters soon.

Wife: "Why is it so jerky?"

Me: "Well, it can decelerate somewhat abruptly while on freeway interchanges..."

Wife: "It's not just around corners, it's all the time on the freeway; why do you use it if it is so bad?"

Me: "Well, I'm trying it out; I think maybe you notice less if you're driving"

Wife: "Well, I notice it. It's ok to try it out if you're the only one in the car, but it isn't cool to do it otherwise"

Me: ...

Wife: If you keep trying to use it, I'm going to stop riding in this car, and we're going to have to take my car*. It's not relaxing.

Me: <disengages Autopilot for remainder of trip>

<a couple minutes pass>

Wife: It's much smoother now. Did you turn it off?

Me: Yes, I did.

Wife: I'm much more relaxed now.


* My wife's car is a Chevrolet Spark EV. That tells you something.

You can't make this stuff up!

Here's hoping that Tesla discovers low pass filters soon.

This made me laugh so much because I experienced the "exact" same with my wife that, after following TMC for a while, I had to become a member and reply to this thread and thank "AlanSubie4Life" for making me laugh until I cried . I couldn't stop laughing at every sentence !!! So much like our experience !!! Thanks AlanSubie4Life !

 

[SteveBaker]

I get the sensation the computer is acting at its limit. For example, a car will cross the road to go on a side street. It's 100 yards away, by the time my car slows down the other car is completely off the road. In other words, my car did not anticipate the actions of the other car, acted far to soon but didn't clear the other car from its decision making des[ite that car being off the road. That says to me the computer is not able to keep pace with conditions.

Which computer do you have? The newest one shouldn't have any shortage of time - but the older ones are getting close to their limits.

 

[DopeGhoti]

but the older ones are getting close to their limit

According to the (certainly biased but hopefully still at least factual) presentation a little while ago, HW2.5 processes at 110 frames per second. Left unstated is whether input to the car's controls is at the same rate.

 

[wcorey]

It's the GPS Posted Speed Limit

I am not going to debate this point beyond saying that logically makes no sense. The correct terminology is geocoded speed limit. As I recall you simply said "posted speed limit". My understanding is the NVidia software, when the S used MobileEye (or whatever it was called) DID actually read posted speed limits and stop signs and traffic lights.