Elon: "Feature complete for full self driving this year"

[mongo]

I don't think it is quite that simple. The drivable space is not always going to be a smooth flat surface like a road. In some cases, the drivable space might be a dirt road or even a grassy area that are definitely not smooth flat surfaces.

Yeah, that was an attempt at brevity. I live in Michigan in the country. My roads are not flat ( both the dirt and the paved) and if the car stopped for sticks, it would never make it up the driveway.

Point being, safe to drive on surfaces can be classified.
It's a corollary to a failure probability problem. Instead of adding all the different ways to succeed, just subtract the probability of failure from 1.

 

[clydeiii]

If you think image recognition "isn't AI", I think you need to read more about AI.

Also the notion that Tesla is where Google was 5-6 years ago simply isn't true.. They're doing completely different things, and you can't compare the two approaches that way. Tesla has billions of real-world miles, something Google only wishes it had access to.

 

[Bladerskb]

If you think image recognition "isn't AI", I think you need to read more about AI.

Also the notion that Tesla is where Google was 5-6 years ago simply isn't true.. They're doing completely different things, and you can't compare the two approaches that way. Tesla has billions of real-world miles, something Google only wishes it had access to.

No it doesn't. I wish people would actually do their own research and THINK things through rather than simply repeating what they heard.
Its not yet illegal to THINK yet people refuse to. There are several questions you can ask your and try to answer that will help you arrive to an actual factual conclusion yourself, that way your mind doesn't fight it because it came from someone else.

 

[clydeiii]

Ok I guess Tesla is lying here? Tesla Vehicle Safety Report

Tesla on Twitter

 

[mongo]

No it doesn't. I wish people would actually do their own research and THINK things through rather than simply repeating what they heard.
Its not yet illegal to THINK yet people refuse to. There are several questions you can ask your and try to answer that will help you arrive to an actual factual conclusion yourself, that way your mind doesn't fight it because it came from someone else.

Tesla is racking up over a billion miles of autopilot a year. So sure it might take 10 months (and dropping) to do a billion miles of testing, but they definitely have that available to them.
Tesla Autopilot nears 2 billion miles driven, Full-Self Driving continues to improve
Even more if they run testing in the background.

 

[clydeiii]

I can only tell you my personal Model 3 experience. I've owned since August 2018 and have about 30,000 miles on the car so far. I would estimate 25,000 have been on Autopilot (mostly Navigate-on-Autopilot). Every single disengagement is a data point for the neural network to potentially learn from. That's how you make an AI that can drive.

 

[Bladerskb]

Ok I guess Tesla is lying here? Tesla Vehicle Safety Report

Tesla on Twitter

You mean the same Tesla that said EAP's NOA and smart summon will be here by end of 2016 or the same Tesla that said driver-less Level 5 full autonomy will be here by Jan 2018?

Here are questions you should ask yourself.

1) How large is one second of the entire data available in the car? (For example 8 cameras, etc)

2) How large is one mile of the entire data available in the car? (For example 8 cameras, etc)

3) What data amount is Tesla actually collecting and uploading?

4) What data amount are self driving competitors actually collecting and uploading?

5) What data storage capacity would you need to host on all data?
​

Here is a few guides:

- Research by resident hacker verygreen

Twitter

green on Twitter

green on Twitter

green on Twitter

green on Twitter​

Gather independent verifiable data and do your own calculations. Again, think for yourself. I will try not to answer this for you because you will simply reject it on the sole basis that it came from me. So you have to see yourself generate it. Just like inception.

 

[mongo]

Why the heck would Tesla upload billions of miles where nothing interesting happened? Much more sensible to only upload problem spots based on the billion+ miles of AP real world testing (and background object of interest detection) they get every year.

 

[clydeiii]

It's true that Telsa is very bad at predicting when they'll be successful at the breakthrough required for FSD to be viable, but accusing them of lying about 1 billion miles driven on FSD is just ridiculous.

Here are questions you should ask yourself.

1) How large is one second of the entire data available in the car. (For example 8 cameras, etc)

2) How large is one mile of the entire data available in the car. (For example 8 cameras, etc)

3) What data amount is Tesla actually collecting and uploading.

4) What data amount are competitors actually collecting and uploading​

​

1) You need to make a lot of assumptions here, but I'll play along with you. Let's assume low-quality settings will use about 300MB per hour (cite: How much mobile data does streaming media use?). So 300MB/60m/60s = 0.083MB * 8 cameras = 0.6MB for 1 second of footage.

2) Ok so if you go back and read what I wrote I never said that Tesla has billions of hours of footage of real-world driving. That's ridiculous. They don't need that. They just need a way to figure out what are the most important seconds of footage, and that would involve human driver disengagements or any time the AI gets confused. If you can just send those back, you save yourself storing the entire billion hours because most of those hours are useless to the AI.

3) I would assume that they would want 5 seconds before and after disengagement, so let's say 6MB (based on the above calculation). The lowest quoted LTE upload on Verizon's website is 2 Mbps, so 24 seconds required to upload 10 seconds of footage. If you use the Intel figure of 40MB per second * 8 cameras * 10 seconds, that's 3.2GB (note it isn't clear from the Intel graphic that they are talking about 1 camera or 8 but I'll go with worst case scenario and assume 1), that does seem unfeasible to upload over LTE at the moment. You can do better though if you have the onboard computer tag the video and only draw it back on-demand. So if, say, you want disengagements that only occur on the highway in heavy traffic at 70+mph, and you can make the computer understand that that's what the footage is, you can only pull back the video when you need to train the AI on that type of footage.

4) Good question. Waymo probably collects everything, but Waymo has a much different and much smaller fleet of cars. GM, Ford, Toyota, Honda...they probably have...nothing? You might have to help me out here, I haven't paid too much attention to the competition in this regard. I just know what Waymo has millions of miles driven out in the real world, but only in certain geographically-bounded areas.

 

[clydeiii]

Also I love how the 'guide' you supplied is literally saying exactly what I wrote above: green on Twitter

"The primary approach is opportunistic and on-demand. Can quibble with details but I think sound. And fleet size is instrumental."

Even if we assume Tesla pulls back 1PB per month, that would still be very ingest-able given the correct infrastructure. A petabyte in 2020 is basically nothing.

Also from that "30mb/minute/cam is from the dashcam files saved to USB" = 4MB/sec, which is about what I came up with not knowing the capture size of the video on the dashcam, and with further compression and/or dropping the framerate, you can easily cut that size down to my 0.6MB.

 

[Bladerskb]

​

1) You need to make a lot of assumptions here, but I'll play along with you. Let's assume low-quality settings will use about 300MB per hour (cite: How much mobile data does streaming media use?). So 300MB/60m/60s = 0.083MB * 8 cameras = 0.6MB for 1 second of footage.

There are no assumptions. I gave you factual evidence with real numbers and instead of using that you used fabricated made up numbers.

REAL ACTUAL NUMBERS

"10 seconds of compressed video from all in-car cams takes anywhere from 100M to 300M."

So that's 10-30 MB per second.

2) Ok so if you go back and read what I wrote I never said that Tesla has billions of hours of footage of real-world driving. That's ridiculous.

Yes you did, you said they had billions of miles of data and compared it to others. Therefore they would need billions of miles of video. Unless they DONT have billions of miles of data.

Every SDC company collects ALL data from EVERY MILE. Every second because every single mile is interesting and important. These ain't lane keeping and adaptive cruise control systems on freeways alone like AP is. These are systems navigating the dense chaotic streets of SF, Jerusalem, etc. Every second matters. Every camera is recorded, every disengagement is recorded, every lidar/radar data is recorded and saved.

They don't need that. They just need a way to figure out what are the most important seconds of footage, and that would involve human driver disengagements or any time the AI gets confused. If you can just send those back, you save yourself storing the entire billion hours because most of those hours are useless to the AI.

As verygreen proved. They dont even send back every disengagement. Infact it will be a miracle if a disengagement triggered an upload. It is an extremely rare occurence. As green said "all the stars have to align".

Case in point. There is no billions of miles of data, nor hundreds of millions. Not even 10 million miles of data. To address the #2 point, if tesla had billions of miles of data as you and others claim. A mile of data would be 1200-3600 MB (1.2 GB - 3.6 GB) per mile.

3) I would assume that they would want 5 seconds before and after disengagement, so let's say 6MB (based on the above calculation). The lowest quoted LTE upload on Verizon's website is 2 Mbps, so 24 seconds required to upload 10 seconds of footage. If you use the Intel figure of 40MB per second * 8 cameras * 10 seconds, that's 3.2GB (note it isn't clear from the Intel graphic that they are talking about 1 camera or 8 but I'll go with worst case scenario and assume 1), that does seem unfeasible to upload over LTE at the moment. You can do better though if you have the onboard computer tag the video and only draw it back on-demand. So if, say, you want disengagements that only occur on the highway in heavy traffic at 70+mph, and you can make the computer understand that that's what the footage is, you can only pull back the video when you need to train the AI on that type of footage.

This isnt the worst case scenario, this is the base case. 100-300MB for 10 seconds of footage. Also as verygreen has also proved, getting a disengagement to trigger a data collection is a miracle because of the amount of criteria needed to be matched. Also triggers are only uploaded to a very few amount of cars.

Not just that, because of the small amount of RAM. You can only hold 10 seconds of data (basically one trigger) in memory and data is only uploaded through WIFI because of how large it is.

So now you have a per trip...wait scratch that. More like per day data collection. You could go to work, after work go to the grocery store, to a ball game and if at any point you were very very lucky to trigger a collection due to a campaign or disengagement trigger. You can't upload it till you make it back home and get on WIFI. Plus you cant trigger anymore collection because the ram is full. So now you drove 100 miles that entire day and only have a 10 seconds footage.

Or let's say you took 2,000 mile road trip. Again if you were lucky enough to trigger collection during your road trip. At the end of your road trip you will only have 10 seconds of footage to upload.

4) Good question. Waymo probably collects everything, but Waymo has a much different and much smaller fleet of cars. GM, Ford, Toyota, Honda...they probably have...nothing? You might have to help me out here, I haven't paid too much attention to the competition in this regard. I just know what Waymo has millions of miles driven out in the real world, but only in certain geographically-bounded areas.

Every one collects everything, every second, every disengagement. So when waymo says they have 10 million miles of data, they actually have 10 million miles of data.

For tesla to have "billions of miles of data" that people claim, they would have to collect everything. If they collect everything they would be uploading a bare minimum of over 100-300 exabytes per year.

Dozens of times (10x-30x) the size of the entire YouTube database today and that's EVERY YEAR!

But we have already proven they cant even hold more than 10 seconds in ram making that impossible. Plus they need to be connected to WIFI to upload.

 

[clydeiii]

There are no assumptions. I gave you factual evidence with real numbers and instead of using that you used fabricated made up nonsense numbers. This is what in talking about.

REAL ACTUAL NUMBERS

"10 seconds of compressed video from all in-car cams takes anywhere from 100M to 300M."

So that's 10-30 MB per second.

There are tons of assumptions there. You assume you can't compress any further. You assume you keep all of the frames when transferring the foortage. Why do you make that assumption?

Every SDC company collects ALL data from EVERY MILE. every second because every single mile is interesting and important.

SDC? Smile Direct Club?!

Wrong. As verygreen proved. They dont send back every disengagement. Infact it will be a miracle if a disengagement triggered an upload. It is ab extremely rare occurence. As green said "all the stars have to align".

I never said they did. You only need to bring back disengagements that you care about. It is hard to divine what Tesla cares about any given moment.

This isnt thecworst case scenario, this is the base case. 100-300MB for 10 seconds of footage. Also as verygreen has also proved, getting a disengagement to trigger a collection is a miracle because of the amount of criteria needed.

verygreen can only guess at Tesla wants to bring back home at any point. He hacks client-side code, not server-side.

So now you have a per trip...wait scratch that. You could go to work, after work go to the grocery store, to a ball game and if at any point you were very very lucky to trigger a collection due to a campaign or disengagement. You can't upload it till you make it back home and get on WIFI. Plus you cant trigger anymore collection because the ram is full. So now you drove 100 miles that entire day and only have a 10 seconds footage. Or let's say you took 2,000 mile road trip. Again if you were lucky enough to trigger collection during your road trip. At the end of your road trip you will only have 10 seconds of footage to upload.

If your point is that while Tesla has access to billions of miles on autopilot but only analyzes a fraction of that in any given development cycle, then we agree. I never claimed anything else.

 

[Bladerskb]

There are tons of assumptions there. You assume you can't compress any further. You assume you keep all of the frames when transferring the foortage. Why do you make that assumption?

Assumption... what in the world? Nothing is assumed. Verygreen has root access and a development car. He can pull the raw data and also he sees the actual files that are waiting to be sent up to Tesla's server after going through the compression process. He can actually see the compression code. He pulls up those files and looks at the bitrate. There's no guessing. None of this is guessing. He can see how large the ram is. This is absolute fact. 100%. Why do you hate facts? Why do you love fairytales?

"Well, the RAM on those nodes is finite. Storing huge buffers on EMMC all the time is impractical for obvious reasons. So today they store 8-10 seconds h265 per camera at 36fps. Tesla needs high quality high bitrate video. So they store 8-10 seconds in RAM and all 8 cams take up to 300M (I can probably look up per frame max) + 25 frames of raw per cam at 2.5M each + a bunch of other telemetry - it adds up fast."

I never said they did. You only need to bring back disengagements that you care about. It is hard to divine what Tesla cares about any given moment.

You can drive an entire month without triggering a disengagement based data collection. That's how rare it is.
That completely contradicts the "Tesla has 14 billions of miles of data while Waymo has 10 million miles of data".

Verygreen directly debunk this over 1,000 times.

The biggest Tesla investor that have been spreading complete BS who now claim that Tesla has "14 billion miles of data" absolute nonsense. He said "A lorry veered off the road and revealed a flock of sheep in rural England. A Tesla driver made a sharp brake. That's now in Tesla's training data set."

Verygreen quickly debunked this as he has several times.

"This is what people believe (and bet money on) and what's not happening."

verygreen can only guess at Tesla wants to bring back home at any point. He hacks client-side code, not server-side.

No server code is relevant here. Its the firmware on the car that does all the triggering, image/video conversion, compression and uploading.

Its amazing that you will claim that the only person outside of Tesla with the most knowledge and access to how Tesla system operates knows nothing and is only guessing. Incredible.

If your point is that while Tesla has access to billions of miles on autopilot but only analyzes a fraction of that in any given development cycle, then we agree. I never claimed anything else.

But I just proved that they don't have access to billions of miles of data.

Waymo has access to 20+ million miles of data. Its sitting in their server right now. Tesla doesn't have access to billions of miles of data.
Waymo has access to 20+ million miles of data. It allows them to do research and development. Such as:

or

Waymo Research

 

[mongo]

Waymo has access to 20+ million miles of data. Its sitting in their server right now.

The US only has 4 million miles worth of roads, and Waymo is not driving on most of them. How much duplication is there given their current geographic limitations?

Tesla doesn't have access to billions of miles of data

Teslas travel over a billion miles a year with AP active, and most of the fleet can analyze data with it off. Worst case, it takes Tesla 12 months to collect a billion miles of geographicly diverse real world performance data on a new version. (Which also allows for seasonal coverage).
With ~400k sales in 2020 and an average yearly mileage of 10,000 , the fleet data collection/ observation will grow by 4 billion miles/ year or a billion miles every 3 months.
By 2022 there will be a billion miles driven every month (not all AP).

Article on public release of a subset of Waymo's data:
Full Page Reload

 

[clydeiii]

tfw you're arguing past the other person on the internet

 

[banned-66611]

Two words: drivable surface

If you can detect a smoothish flat surface, you don't need to categorize every possible object/ obstacle/ hypothetical that could be in your path.

Counterfeit Detection (Part 1) - Tim Challies

A driveable surface is all well and good but what is it going to do when there is a tree branch fallen in the road that it doesn't recognize? The radar will ignore it because it's stationary, just like it ignores stationary fire trucks and rams into the back of them at high speed.

Have a look at some of Waymo's videos on YouTube. You can see what their cars can see, and it's everything. They need to be able to predict things like what pedestrians by the side of the road will do. Roads are a shared space, people can walk on them, bikes can ride on them etc. That's why all of the current systems limit themsleves to highways, the only roads that are dedicated to cars only and have things like trees trimmed back.

 

[banned-66611]

If you think image recognition "isn't AI", I think you need to read more about AI.

Also the notion that Tesla is where Google was 5-6 years ago simply isn't true.. They're doing completely different things, and you can't compare the two approaches that way. Tesla has billions of real-world miles, something Google only wishes it had access to.

It's not really AI, it's image recognition. Google started offering it many years ago, e.g. you can search your photos for "cat" and it will show all the ones with cats in them.

Nobody uses real AI that actually thinks about what it wants to do. They recognize images and lidar and GPS sensor data and then use an algorithm to plan a route.

Ok I guess Tesla is lying here? Tesla Vehicle Safety Report

Tesla on Twitter

Yes they are. Tesla is embarassed because they keep promising FSD but are doing zero FSD testing miles, so they fluff up their reports to claim all these "AP shadow mode" miles which in actual fact are meaningless.

At most they collect data on where AP makes radically different decisions to a human driver, or where it disengages, and then look at those areas to refine their algorithms. That was what Musk was talking about when he mentioned "corner cases".

 

[Mardak]

I'd add that they may achieve "feature complete" on the current generation of hardware but the current generation will always and forever be level 2

At least from Tesla's new support page Full Self-Driving Computer Installations, Tesla does seem to explicitly suggest that FSD computer (and existing sensors) will exceed "feature complete" (requiring supervision) thus achieving higher than level 2:

Will the FSD Computer make my car fully autonomous?
Not yet. All Tesla cars require active driver supervision and are not autonomous. With the FSD Computer, we expect to achieve a new level of autonomy as we gain billions of miles of experience using our features. The activation and use of these features are dependent on achieving reliability far in excess of human drivers, as well as regulatory approval, which may take longer in some jurisdictions.

 

[mt09]

I was thinking about how wheel torque driver monitoring isn't feasible for automatic driving on city streets/city NOA given the large turning degree required. How is a driver supposed to keep their hands on the wheel during a 90 deg right hand turn? Is there going to be an alert before any large degree turning actions "Comin' in hot, hands off!"
Consider the case of a driver coming up to a tight 90 deg intersection turn while holding the wheel and not paying attention. Autosteer starts to turn the wheel, driver isn't paying attention and disengages autosteer midturn. Driver refocuses and says oh *** and tries to correct. Highway wheel torque monitoring is pretty straight forward since the wheel degree doesn't go past +/- 45 deg.

 

[Daniel in SD]

I was thinking about how wheel torque driver monitoring isn't feasible for automatic driving on city streets/city NOA given the large turning degree required. How is a driver supposed to keep their hands on the wheel during a 90 deg right hand turn? Is there going to be an alert before any large degree turning actions "Comin' in hot, hands off!"
Consider the case of a driver coming up to a tight 90 deg intersection turn while holding the wheel and not paying attention. Autosteer starts to turn the wheel, driver isn't paying attention and disengages autosteer midturn. Driver refocuses and says oh *** and tries to correct. Highway wheel torque monitoring is pretty straight forward since the wheel degree doesn't go past +/- 45 deg.

Accidentally disengaging in the inside lane at the start of a double right hand turn would definitely be something you'd want to avoid. On the other hand what's the alternative? Are you really going to be able grab the wheel in time to avoid hitting a car that's only a couple feet away?