Tesla autopilot HW3

[rnortman]

Who is that smart guy who disagree with Lex Fridman and Levandowski and without even bother to give the arument of why? @rnortman you have just showed your true color as another worthless troll. Sorry that I wasted my time on you. Off you go to my ignore list too.

I have answered your points previously and you continue to make the same arguments. And I thought you had already ignored me?

 

[turnem]

@turnem Most of your concerns are valid but they were pretty much all answered in the investor event. The deep learning is not only powerful enought to recognize speed signs, stop lights etc but could even drive on snow covered roads and making right predition of curves we can not see past (human drivers actually have this capability). The most significant one we first heard of is the Tesla system can observe small movement of cars and pedestrians to predict their actions (changine lanes of walk off the curb). Again it is using the fleet learning to it in the shadow mode even if the AP is not turned one. The AI will just make preditions and then compare them to what actually happen to train and fine tune the machine. It's super smart and powerful approach and probably solved the last missing link of self driving cars. That's what Levandowski aluded to in his short interview I linked above.

Very cool. Thanks for the summary! I knew that they were using machine learning but I wasn't sure if they were specifically going to attempt to read all road signs (not just speed limit signs) and/or car blinkers and tail lights.

I'm curious to see the improvements along the way!

 

[CarlK]

That is not quite correct. There's a very interesting video from Mobileye demonstrating a camera-only (NO LIDAR) vehicle merging "assertively" into traffic.
The video is from May 2018, citing implementation with major automakers in July (no idea if that happened, if it's providing the buyers with ADAS or is just helping data collection)(relevant section starts at 30 min in)

And, yes, I've followed the development and the sensor approach with keen interest. And I really hope all the claims eventually come true. It's a fun ride (mostly) to see features delivered. But the track record of announcement is a desaster and once you bump your car into a wall because the most recent update broke front ultrasonics, the fun stops (happened about a year ago, to very few users, me among them).
Hence, AP/FSD announcements by Tesla are like electric car announcements by German automakers - looks good in PowerPoint, I'll believe it when it's delivered.

We really should ban video from that liar on this board. This guy from Mobileye had been giving misleading presentations for years trying to con people with what it did not have or does not know. The answer to your question is NO that inlpementation has never happened and it never will.

I can't blame you for your skeptism although I always have faith in Elon. He never BS even though many times people don't understand what he says because it's so ahead of our undertanding (Lidar is a crutch for example). If you have time to watch, and really absorb it with an open mind, what was said in the investor event plus comments from people who had been to the event and took test rides I think you will change your thinking too.

 

[iTech]

We really should ban video from that liar on this board.

Seems a tad harsh to ban any mention of Elon on this forum, no?

 

[CK_Stuggi]

I've watched it all. In it's entirety. But, aside from the fact they've abandoned HD maps and local references, they have shown a preprogrammed self driving loop. Yes, they've shown advanced summon. And NoA. Functions already "out in the wild".
But other than the words of people at the investor event, who are people with a vested interest in Tesla, what more has Tesla actually demonstrated that goes beyond what's shown in the Mobileye video?

Personally I'd be very careful to call someone a liar who was able to sell his company to Intel at a hefty price - cause Intel certainly does thorough due diligence. Others call Musk a liar, because of all those poor predictions. I'm not, I like to think he's always delivered on everything - except autopilot(hopefully: yet).

Why didn't Tesla at least offer any crumb of info as to which Kick-Ass features are right around the corner to their customers? Why were all investors under NDA? Why did they shuffle around the definition of FSD prior to the event?

It really is wait and see.
My bets are on Tesla, as I said.
I'm just saying the track record and what they've demonstrated as "under development", from an objective standpoint, is not unprecedented, and aside from my personal hope in their approach there is little evidence to believe they can achieve what they claim, feature complete FSD end of 2019 (that would be for the US and Canada, of course).

I sure like to believe it. I just don't claim to have evidence.

 

[CarlK]

Intel was just dumb and desparate because it's really behind Nvidia in this lucrative arena. That Amnon was a great liar , OK salesman helped too. Mark my words the truth will come out very soon for Intel. It's not THAT dumb.

We are starting to see Tesla really meant what it said. Remember it hired Pete Bannon along with Jim Keller, two of the most respected chip designers in the industry, more than three years ago to work on this project. And it put sensor suit in every car it makes whether you pay for the option or not. Everything has been long planned for it's not just dumb luck. No problem with you having your doubts. Again time will tell.

 

[verygreen]

Reading road signs. Like reading a sharp turn sign and knowing to slow down a little.

Maps cover that usecase. It works today. It actually contains data about what's the safe speed there, not some "slow down some" info.

Reading blinkers on cars in front of you

This is unreliable. People forget to turn them off, people forget to turn them on. AK gave an example at the autonomy day how they use the car "body language" to infer if they want to merge. This is a lot more robust. People do it too.

NOBODY other than Tesla is using vision+deep learning

Check out Subaru vision only system (eyesight).

But, aside from the fact they've abandoned HD maps

They did not.

 

[turnem]

Maps cover that usecase. It works today. It actually contains data about what's the safe speed there, not some "slow down some" info.


This is unreliable. People forget to turn them off, people forget to turn them on. AK gave an example at the autonomy day how they use the car "body language" to infer if they want to merge. This is a lot more robust. People do it too.


Check out Subaru vision only system (eyesight).


They did not.

Responding to your point on maps... I disagree completely. Map speed limits are not accurate enough and there is nowhere close to enough information on a map to tell the car that it needs to slow down for an upcoming sharp turn. WAY too many variables there which is precisely why we have road sings to warn us humans of suggested speeds for upcoming sections of the road. I would think Tesla should attempt to use this same information to improve autopilot driving to be more human like. What about speed humps on city streets? Is that supposed to be contained in a map? Try putting your car on TAAC on a 35mph road that uses speed humps.

 

[verygreen]

ap speed limits are not accurate enough and there is nowhere close to enough information on a map to tell the car that it needs to slow down for an upcoming sharp turn

let me refer you here: Tesla Autopilot maps so oyou can see what's actually is available on the maps (the speed limit information is incorrect, but that's besides the point).

Tesla tracks "fleet speed" so if you drive some place slowly - They take notice.

 

[J1mbo]

Except AP2.0 does not recognize traffic signs at all yet. It is all database.

Do you honestly think there is there anyone on here who didn't know this already?

The system in the i3 is so unreliable that I ignore it. It might be able to read matrix signs in variable speed limit areas, which AP 2 cannot do, but I cannot afford to trust it.

 

[electronblue]

They did not.

Are you saying Elon lied on the 22nd?

 

[turnem]

let me refer you here: Tesla Autopilot maps so oyou can see what's actually is available on the maps (the speed limit information is incorrect, but that's besides the point).

Tesla tracks "fleet speed" so if you drive some place slowly - They take notice.

If that's their sole angle of attack on this issue then they are going to be out of date before then even update their database more often than not. Isn't the more dynamic, real-time option to read the same data that's provided to human drivers and adjust accordingly? If you want to supplement that with a stale database then that's great but even if they're using their own fleet to constantly update their database it's not going to be robust enough to be reliable.

 

[Matias]

If that's their sole angle of attack on this issue then they are going to be out of date before then even update their database more often than not. Isn't the more dynamic, real-time option to read the same data that's provided to human drivers and adjust accordingly? If you want to supplement that with a stale database then that's great but even if they're using their own fleet to constantly update their database it's not going to be robust enough to be reliable.

My AP1 slows nicely before tight turns based on map info.

 

[mongo]

I think you have drifted off of the original point:

Like reading a sharp turn sign and knowing to slow down a little.

If that's their sole angle of attack on this issue then they are going to be out of date before then even update their database more often than not. Isn't the more dynamic, real-time option to read the same data that's provided to human drivers and adjust accordingly? If you want to supplement that with a stale database then that's great but even if they're using their own fleet to constantly update their database it's not going to be robust enough to be reliable.

Sharp turns don't just show up or go away. The map tile data includes this information, there is zero need to read the warning sign each time a car passes it.

 

[turnem]

I think you have drifted off of the original point:





Sharp turns don't just show up or go away. The map tile data includes this information, there is zero need to read the warning sign each time a car passes it.

Then why are there countless YouTube videos of cars barreling into tight turns? That question sounds more negative than I mean for it to. I'm not trying to pick a fight - I'm trying to learn so please don't read that with a negative tone if possible.

I just know from my admittedly limited experience that the car rarely slows down for upcoming tight turns. I agree with you that if the road is on the map that the data must exist. I'm just wondering if it's not ENOUGH data to tell the car to slow down. So either AP isn't yet factoring this in on a consistent basis or it needs more information to work reliably.

 

[turnem]

My AP1 slows nicely before tight turns based on map info.

Interesting. Mine doesn't. Lots of videos on YT show roughly what I experience.

 

[CarlK]

@Evergreen

If that's their sole angle of attack on this issue then they are going to be out of date before then even update their database more often than not. Isn't the more dynamic, real-time option to read the same data that's provided to human drivers and adjust accordingly? If you want to supplement that with a stale database then that's great but even if they're using their own fleet to constantly update their database it's not going to be robust enough to be reliable.

Yes road conditions change constantly. That's why Elon said HD map is too brittle to be relied on. Map will not tell you if an accident was just happened. It's better if your NN was trained to recognize what has just happened. BTW that's one of the tail cases Elon and Karpathy said they need a lot driving miles to get enough training data.

 

[mongo]

Then why are there countless YouTube videos of cars barreling into tight turns? That question sounds more negative than I mean for it to. I'm not trying to pick a fight - I'm trying to learn so please don't read that with a negative tone if possible.

I just know from my admittedly limited experience that the car rarely slows down for upcoming tight turns. I agree with you that if the road is on the map that the data must exist. I'm just wondering if it's not ENOUGH data to tell the car to slow down. So either AP isn't yet factoring this in on a consistent basis or it needs more information to work reliably.

Note that this is not talking about interpreting the map and deciding if a turn is sharp. This is meta data for that section of road that applies a speed adjustment.

For non-varying conditions:
If the map is missing data, then sign reading is important. If the map data is correct, sign reading is redundant.
For the general condition:
Real time (sign reading) data is better. Even better is correctly detecting a tuning road and adjust speed without a sign.

 

[turnem]

Note that this is not talking about interpreting the map and deciding if a turn is sharp. This is meta data for that section of road that applies a speed adjustment.

For non-varying conditions:
If the map is missing data, then sign reading is important. If the map data is correct, sign reading is redundant.
For the general condition:
Real time (sign reading) data is better. Even better is correctly detecting a tuning road and adjust speed without a sign.

I agree on all of this. But now I'm confused because it seems like we've come full circle. So now we are both saying that sign reading PLUS map data is better than just relying on map data correct?

All I know is that currently, my car rarely slows properly for sharp turns on city streets. There's one by my house that has been there as long as I've been alive and it's a white knuckle affair if I let the car try it on Autopilot. But that's not the only one. In fact I'm not sure I've seen it slow down appropriately for any turn yet.

 

[mongo]

I agree on all of this. But now I'm confused because it seems like we've come full circle. So now we are both saying that sign reading PLUS map data is better than just relying on map data correct?

All I know is that currently, my car rarely slows properly for sharp turns on city streets. There's one by my house that has been there as long as I've been alive and it's a white knuckle affair if I let the car try it on Autopilot. But that's not the only one. In fact I'm not sure I've seen it slow down appropriately for any turn yet.

My point in commenting was that for items like sharp turns maps are sufficient, if they have the data.

I like live data more than maps. In your case, the map tile does not have the needed data, nor is the system anticipating the turn. With a good fleet self updating map system it would still not need signs after the first time through (of any Tesla). However, there is always a first time, and conditions change (construction) so relying on maps is brittle.