Tesla - where are my (HW2) auto rain sensing wipers?

[Daniellane]

You know when you're a Tesla enthusiast when.... you find yourself watching YouTube videos of windshield wiper activity.

In position 2, did you not expect more frequent wiping given the accumulation of raindrops? Iirc the release notes, position 1 was "medium" and position 2 was more... active, yes?

In any case, nice video.

In my limited testing today, the wipers are certainly operating differently than before when in positions 1 and 2. Can't yet say whether better than the old-style timed/consistent wiping, but it's not worse.

Feels perfect to me. It’s a delicate balance.
In my view, excessive wiping for a few drops that do not obscure vision as much as a wipe is counterproductive.
Seems like this feature has been misinterpreted a bit. To me, it’s as simple as thinking of position 1 as intermittent and position 2 is constant.
I have transcribed the release notes below for clarity:

“Want to turn them on intermittently when a medium amount of rain has accumulated? Choose Auto 1 option on the wiper lever”
Or choose the more sensitive Auto 2 option if you prefer wiping as soon as the vehicle senses just a light mist or drizzle, such as when you’re driving out of the garage on a foggy morning. To make sure the windshield is clear, the wiping speed will automatically adjust based on how much rain accumulates between each wipe.

 

[AnxietyRanger]

Feels perfect to me. It’s a delicate balance.
In my view, excessive wiping for a few drops that do not obscure vision as much as a wipe is counterproductive.
Seems like this feature has been misinterpreted a bit. To me, it’s as simple as thinking of position 1 as intermittent and position 2 is constant.
I have transcribed the release notes below for clarity:

“Want to turn them on intermittently when a medium amount of rain has accumulated? Choose Auto 1 option on the wiper lever”
Or choose the more sensitive Auto 2 option if you prefer wiping as soon as the vehicle senses just a light mist or drizzle, such as when you’re driving out of the garage on a foggy morning. To make sure the windshield is clear, the wiping speed will automatically adjust based on how much rain accumulates between each wipe.

That's not how they work, though. Auto 2 is not constant either. Both are operating only when the see rain and then stop, their sensitivity is just different. Try it in the dark and you'll see. The update notes seem more PR prose to me than an actual technical description.

I would actually find that a fairly useful distinction, though, to have what you mention... Either through the stalk on other settings: an option that only wipes when it "sees" something (with maybe a couple of sensitivity steps)... which by the way is how Auto 1 and Auto 2 operate now... and a third, which is not yet available, which is how a traditional rain sensing would work e.g. by engaging a suitable, constant wiping speed for a period of time. This would basically mean some of the time the wipers are engaged on a timer basically to keep the operation constant (just adjust the wiping speed based on rain detection "hits" frequency)...

Especially in dark I find the Auto 1 useless and Auto 2 too random in its wiping frequency. Some "Auto 3" that just kept wiping on a timer at a rain-dependent speed - and stopped only when no rain was detected for a distinct while - would be my favorite. It could still retain the single-swipe tendency for no rain situations by having some counter that measures how many single swipes there have been recently - and go into this constant wiping mode only after it has been a few in a row.

 

[MarcusMaximus]

That's not how they work, though. Auto 2 is not constant either. Both are operating only when the see rain and then stop, their sensitivity is just different. Try it in the dark and you'll see. The update notes seem more PR prose to me than an actual technical description.

I would actually find that a fairly useful distinction, though, to have what you mention... Either through the stalk on other settings: an option that only wipes when it "sees" something (with maybe a couple of sensitivity steps)... which by the way is how Auto 1 and Auto 2 operate now... and a third, which is not yet available, which is how a traditional rain sensing would work e.g. by engaging a suitable, constant wiping speed for a period of time. This would basically mean some of the time the wipers are engaged on a timer basically to keep the operation constant (just adjust the wiping speed based on rain detection "hits" frequency)...

Especially in dark I find the Auto 1 useless and Auto 2 too random in its wiping frequency. Some "Auto 3" that just kept wiping on a timer at a rain-dependent speed - and stopped only when no rain was detected for a distinct while - would be my favorite. It could still retain the single-swipe tendency for no rain situations by having some counter that measures how many single swipes there have been recently - and go into this constant wiping mode only after it has been a few in a row.

Speaking of using this in the dark, I mentioned this in another thread, but I noticed a bug in the network: if you’re driving up to a tunnel(or other sudden darkening of the area), even without any rain, it’ll trigger a wipe(at least on setting 2).

 

[AnxietyRanger]

Speaking of using this in the dark, I mentioned this in another thread, but I noticed a bug in the network: if you’re driving up to a tunnel(or other sudden darkening of the area), even without any rain, it’ll trigger a wipe(at least on setting 2).

Yes, it does for me too.

 

[Daniellane]

I know I’m stating the obvious here:
As it would seem with nearly all Tesla hardware as compared with other vehicles “fixed in stone” hardware, they have the opportunity to continually refine the software that controls it’s behavior. I’m guessing that the issues described above are the reason the engineers were not releasing it sooner. Seems like Elon forced there hand and it was released as “Beta” which may not work as intended perfectly in all conditions “yet” much like the “Easy Entry” feature.
Perhaps with constructive feedback from real world experiences of users, progress towards that end will be accelerated.

 

[jldf310]

ROFL! They were aiming to get in in the December 2016 release too. And that was according to their official statement on their website that they will bring AP2 mostly on par with AP1 by end of December, again 2016, so a year ago. Until it's deployed, it's on par with FSD, flying cars and Mars bound commuter Teslas.

has anyone else noticed that @whitex has gone silent in this conversation?

 

[lunitiks]

That's not how they work, though. Auto 2 is not constant either. Both are operating only when the see rain and then stop, their sensitivity is just different.

... which corraborates @jimmy_d's observation in the fisheye.PROTOTX that there is only one output from the wiper NN. so either "rain" or "no rain".

... which means that Auto 1 and Auto 2 could be just body control stuff (Auto 1 having some pre-programmed "delay" or something compared to Auto 2)

I know I’m stating the obvious here:
Seems like Elon forced there hand and it was released as “Beta” which may not work as intended perfectly in all conditions “yet” much like the “Easy Entry” feature.

This we will never know, but I totally agree this should and could be upgraded OTA

 

[jimmy_d]

... which corraborates @jimmy_d's observation in the fisheye.PROTOTX that there is only one output from the wiper NN. so either "rain" or "no rain".

... which means that Auto 1 and Auto 2 could be just body control stuff (Auto 1 having some pre-programmed "delay" or something compared to Auto 2)

This we will never know, but I totally agree this should and could be upgraded OTA

In an earlier version of the fisheye prototxt there was a 5 output classifier - so basically it gave you one of five categories. Each of those five outputs is a probability that what the NN is seeing fits into one of those categories, and all five probabilities will add up to one. This is a boilerplate way of making a network that gives you a one-to-five level for something, like maybe how hard it is raining.

In the current version - at least the latest one I've seen - it's a one output classifier which might just be 'probability that it is raining'. Or alternately it could be trained to be 'how hard is it raining' with, say, anything below 0.2 being 'probably not raining' and anything over 0.9 being 'raining cats and dogs'. If they do this second thing they they could get a wiper speed setting out of a single number by just picking different thresholds for different wiper speeds. Because probabilities aren't a good match for degree of intensity they aren't used this way a lot, but it doesn't seem like a stretch to maybe get two different wiper speeds plus 'not raining' out of a single number.

 

[lunitiks]

but it doesn't seem like a stretch to maybe get two different wiper speeds plus 'not raining' out of a single number.

I'm thinking the wiper speeds (actually, they're not wiper speeds) might be a delay setting / command to BCM.

If 1 is "raining", then actuate. If 0.. well.. don't

 

[lunitiks]

(output-wise)

 

[jimmy_d]

I'm thinking the wiper speeds (actually, they're not wiper speeds) might be a delay setting / command to BCM.

If 1 is "raining", then actuate. If 0.. well.. don't

That totally makes sense. It's a lot simpler than what I was thinking and it seems like it would work as well or even better.

Come to think of it - since the camera is taking many shots per second the system could just time the delay from the last wipe until the NN says 'i see rain' and use that as a proxy for how hard it is raining. If the delay gets short enough then you up the wiper speed, and if you don't see rain before the next swipe then you slow down or stop.

clever

 

[chillaban]

In an earlier version of the fisheye prototxt there was a 5 output classifier - so basically it gave you one of five categories. Each of those five outputs is a probability that what the NN is seeing fits into one of those categories, and all five probabilities will add up to one. This is a boilerplate way of making a network that gives you a one-to-five level for something, like maybe how hard it is raining.

In the current version - at least the latest one I've seen - it's a one output classifier which might just be 'probability that it is raining'. Or alternately it could be trained to be 'how hard is it raining' with, say, anything below 0.2 being 'probably not raining' and anything over 0.9 being 'raining cats and dogs'. If they do this second thing they they could get a wiper speed setting out of a single number by just picking different thresholds for different wiper speeds. Because probabilities aren't a good match for degree of intensity they aren't used this way a lot, but it doesn't seem like a stretch to maybe get two different wiper speeds plus 'not raining' out of a single number.

Could the single output be some metric of the "amount" of rain buildup as well? The documentation seems to imply the two sensitivity settings relate to the amount of accumulation between wipes.

 

[MarcusMaximus]

Because probabilities aren't a good match for degree of intensity

I'd note that depends on the application and classifier in question. At my last company, we used the evidence(not quite straight probability, but with the same meaning) as a proxy for intensity and it worked quite well for that purpose.

 

[lunitiks]

That totally makes sense. It's a lot simpler than what I was thinking and it seems like it would work as well or even better.

Come to think of it - since the camera is taking many shots per second the system could just time the delay from the last wipe until the NN says 'i see rain' and use that as a proxy for how hard it is raining. If the delay gets short enough then you up the wiper speed, and if you don't see rain before the next swipe then you slow down or stop.

clever

Aligns with the fact that we haven't seen the auto wipers speed up. Same lazy speed no matter what, only a perceived different "sensitivity". While the rain sensor setup on AP1 has an actual speed setting/command. (Sometimes these babes go absolutely crazy on me)

 

[chillaban]

Aligns with the fact that we haven't seen the auto wipers speed up. Same lazy speed no matter what, only a perceived different "sensitivity". While the rain sensor setup on AP1 has an actual speed setting/command. (Sometimes these babes go absolutely crazy on me)

Yeah, IMO that's "missing software" so to speak. I think the neural net is probably doing the right set of responsibilities by measuring whether there's rain accumulation on the windshield. I think a control loop belongs on the AP2 ECU that looks for how quickly rain accumulates between wipes and uses that to pick a wiping speed.

 

[lunitiks]

actually the word "speed" is becoming quite confusing now (english not my tounge)

 

[J1mbo]

Sounds like the NN is trying to replicate the output of the AP1/pre-AP rain sensor....

If that is the case, and the BCM is actually controlling the wiper speed, then the result of the NN sensing rain should be exactly the same as the dedicated sensor in AP sensing rain? Does AP1 only have 2 settings? I can't remember what settings my pre-AP car had.

This approach makes a lot of sense in a KISS kind of way, it means that they "only" need to refine the NN to improve performance (assuming the BCM is up to the job of setting wiper speed accurately).

 

[chillaban]

Sounds like the NN is trying to replicate the output of the AP1/pre-AP rain sensor....

If that is the case, and the BCM is actually controlling the wiper speed, then the result of the NN sensing rain should be exactly the same as the dedicated sensor in AP sensing rain? Does AP1 only have 2 settings? I can't remember what settings my pre-AP car had.

This approach makes a lot of sense in a KISS kind of way, it means that they "only" need to refine the NN to improve performance.

AP1 did only have two auto settings. The wiper stalk did not change between AP1 and AP2.

The key difference in behavior is:
AP1's automatic wiper controller seems to output a fixed wiping speed, ranging from one of the intermittent settings all the way to the fast wipe. There's hysteresis in which setting it chooses, but you can get out a stopwatch and find that without big changes in rain, the number of seconds between wipes is consistent.

AP2's automatic wiper controller seems to output single fixed wipes whenever it feels like it based off rain accumulation. Wipe, pause, wipe, wipe wipe, pause pause pause….

I'm not convinced one behavior is necessarily better than the other. It's just the AP1 way is pretty much how every car I've had has done it. AP2's way is… interesting. I don't find it objectionable, but it does feel unfamiliar.

 

[lunitiks]

Does AP1 only have 2 settings?

Yes.

But in both settings, AP1 does not only wipe at different intervals (depending on rain volume), but the actual speed of the wiper blades are not constant - they can suddenly speed up like crazy, then slow down, then go nuts again... In both settings

 

[chillaban]

Yes.

But in both settings, AP1 does not only wipe at different intervals (depending on rain volume), but the actual speed of the wiper blades are not constant - they can suddenly speed up like crazy, then slow down, then go nuts again... In both settings

FWIW, the speed change is when AP1's wiper chooses the "II" position rather than "I" (e.g. continuous turbo wiping speed).

My dad worked on wiper motors for 20+ years, from what I remember, the fastest wiping setting actually does make the motor run faster, but it's at the expense of the longevity of the motor and wiper arm link.

Funniest thing is, the car I drove when I was in high school had a wiper system from his company, and he would always yell at me whenever I put the wiper in the fast setting. Talk about believing in what you worked on