Tesla now using deep neural net for auto wipers (2019.40)

[mtndrew1]

This version is an improvement for me. Not perfect, but far better than they’ve ever been.

We had some truly torrential rain on Sunday night and they performed admirably.

 

[banned-66611]

I imagine that in order for the neural network to learn, it will need to start with a lower activation threshold than before. There's positive reinforcement in the form of pressing the button to get it to wipe more often, but there's no negative reinforcement to punish the algorithm for wiping too frequently. This form of learning requires that it starts off wiping less than you would like, and will slowly learn to increase the frequency to your liking.

It's 2020 and our cars have regressed to a child like state where they need to learn what rain is again.

 

[croman]

I still get dry wipes in garage and tunnels and sometimes randomly without cause. Unacceptable to drop cheap reliable part in favor of unproven not even prototyped concept of NN based wipers....

 

[jdicecco]

My wipers are much improved. I have seen further improvement with latest release 5.7

 

[murc585]

Does not handle salty run off from vehicles in front of me in winter at all. Pretty much useless for me until winter is over.

 

[willow_hiller]

Does not handle salty run off from vehicles in front of me in winter at all. Pretty much useless for me until winter is over.

Would a traditional IR sensor handle salt spray? I think the neural network has the potential to be better than a simple sensor, it's just not there yet.

 

[navguy12]

Does not handle salty run off from vehicles in front of me in winter at all. Pretty much useless for me until winter is over.

At least when the voice commands decide to function, one can now tell the wipers to turn on (speeds 1,2,3 or 4) without staring down at the lower left corner of the screen.

 

[banned-66611]

I still get dry wipes in garage and tunnels and sometimes randomly without cause. Unacceptable to drop cheap reliable part in favor of unproven not even prototyped concept of NN based wipers....

I guess they were like "guys we have to have robotaxi working next year, we gotta at least be able to wipe the windows with a NN!"

 

[Can37]

Would a traditional IR sensor handle salt spray? I think the neural network has the potential to be better than a simple sensor, it's just not there yet.

I have an e-Golf with a normal sensor, the wipers just work, salt spray, changes in lighting, misty rain all work close to perfectly. My 3 messes up in all these cases, no wipes, manic wiping, random wiping on a dry day.

NN and cameras is Telsa's hammer so they treat all problems as nails. Dumb product planning, they are spending a crap tonne of money on a fully solved problem with a very cheap solution.

 

[motocoder]

It's that time of year when the Pacific Northwest US gets a lot of different rain conditions. I have to say, this "deep rain" is just not getting the job done. If I were designing a system, I would have at least one camera with a shallow depth of field focused on the windshield, and maybe a second camera focused at infinity to use as another input. I suspect Tesla is just using a fisheye camera with a very deep DOF, focused at infinity. I don't think any neural net will be able to do a sufficient job with that as an input. Doesn't matter how much training data you have.

 

[arghx7]

It's that time of year when the Pacific Northwest US gets a lot of different rain conditions. I have to say, this "deep rain" is just not getting the job done. If I were designing a system, I would have at least one camera with a shallow depth of field focused on the windshield, and maybe a second camera focused at infinity to use as another input. I suspect Tesla is just using a fisheye camera with a very deep DOF, focused at infinity. I don't think any neural net will be able to do a sufficient job with that as an input. Doesn't matter how much training data you have.

They could use the other cameras to assist, but would probably need HW3 to handle all the extra image processing.

 

[Sporty]

I wish there was a way to keep the wiper UI from auto-dismissing. I routinely have to turn the auto feature off and I’m always adjusting the setting. (I’m in the Seattle area now in the raining season). It would take less attention if the UI could be pinned on and I could just tap new setting when driving around town. Currently I have to look twice, and the second look has to be right after the first look.

For the NN implementation: 1. it routinely wipes too much/fast (especially if the rain is being variable). 2. I don’t think it’s using the current speed as one of the inputs.

 

[banned-66611]

They could use the other cameras to assist, but would probably need HW3 to handle all the extra image processing.

The utter insanity of using a more powerful computer and more cameras and more energy just to replace a reliable $5 rain sensor.

 

[arghx7]

Is it really only 5 dollars though?

 

[mongo]

Is it really only 5 dollars though?

Even if it were (and ignoring logistics, warranty, installation), that would have been over $1.5 million to date, and when GF3 and GF4 are fully on line, over $5 million a year. The equivalent bottom line impact of over 300 additional vehicle sales.
And I doubt the rain net added substantially to the the cost of the autopilot computer which would exist anyway.
Nor would a standard rain sensor provide coverage for the tricamera unit. Just like your passenger may ask for a wipe while your side is clear.
And existing sensors aren't perfect either, but mine will never get better.

 

[banned-66611]

Even if it were (and ignoring logistics, warranty, installation), that would have been over $1.5 million to date, and when GF3 and GF4 are fully on line, over $5 million a year.

How much money do you think they pissed away on developing a NN for rain detection and then spending a year trying to make it actually work? Just so far I mean, obviously no-one knows how much it will cost if and when they finally get it working.

 

[mongo]

How much money do you think they pissed away on developing a NN for rain detection and then spending a year trying to make it actually work? Just so far I mean, obviously no-one knows how much it will cost if and when they finally get it working.

You are comparing a low number of engineers (sub 5) for a limited amount of time against a cost adder on every car they every produce.
$5 million per year and increasing vs $300k * 5 engineers * 2 years = $3 million max development cost.

 

[navguy12]

How much money do you think they pissed away on developing a NN for rain detection and then spending a year trying to make it actually work? Just so far I mean, obviously no-one knows how much it will cost if and when they finally get it working.

Maybe the work to make the auto wipers smart is providing lessons learned for the autopilot system.

If that is the case, it's an investment in how to get the machine learning thing sorted out sooner than later.

 

[holmgang]

You are comparing a low number of engineers (sub 5) for a limited amount of time against a cost adder on every car they every produce.
$5 million per year and increasing vs $300k * 5 engineers * 2 years = $3 million max development cost.

at claimed 5$ per sensor, tesla is producing 1,000,000 vehicles per year to incur that cost?

and once the software is developed theres no cost for its continued maintenance?

 

[mongo]

at claimed 5$ per sensor, tesla is producing 1,000,000 vehicles per year to incur that cost?

and once the software is developed theres no cost for its continued maintenance?

Pretty much. Fremont produced 120k cars last quarter, that would be $600k at $5 each of $2.4 million a year without GF3 and GF4.
AP will have a team already, less than one full time equivalent needed for maintenance. Call it $300k a year.