Seeing the world in autopilot, part deux

[strangecosmos]

It's official: V9 (2018.40) will roll out to everybody on Monday, October 1st. Just confirmed over the phone with Tesla CS. : teslamotors

Would be awesome if true. Can't wait to see the hour-long YouTube drives.

 

[Pale_Rider]

Bleh, hopefully auto lane change isn’t gone, that’s what made it so impressive. Just telling me what lane I should be in and having me use the stalk is not nearly as interesting.=\

 

[strangecosmos]

First, Neural networks are only used in perception bullet point.

I believe jimmy_d is speculating that path planning in Autopilot v9 uses a neural network. Mobileye is using reinforcement learning for path planning, I think.

Furthermore, developing the trained NNs that will use for various tasks, is a very long challenging task for sure... It relies a lot more on hard working talented AI and computer vision engineers than it does rely on having vast amounts of data.

It seems like the best neural network architectures are all open source. Meaning the competitive advantage is not the architecture, but the training, and therefore the data.

And even more so... data that comes from the fleet is not the most helpful / most core to training the NNs...

Why not?

You're right they have, and this is awesome. However, Tesla vehicles are advanced driver assist. Albeit one of the coolest advanced driver assist systems out there, they do not replace the driver. (as you know)

Maybe when you say developing full autonomy you are referring to the ability to create a perception system that

• Makes perception errors that would result in an accident so infrequently that it is practical to use it in a system that can replace a human driver in some conditions
• and uses only 8 cameras, sonor, forward radar,

It sounds like you are saying you believe this system can be built? and the key to doing so is collecting lots and lots of fleet data? is that what you are suggesting?

That's Tesla's plan!

All Tesla Cars Being Produced Now Have Full Self-Driving Hardware

 

[Snuffysasa]

You are correct:
View attachment 339271
https://www.cadillac.com/content/da...18-cad-ct6-supercruise-personalization_v2.pdf

@verygreen - just for clarity: I have not driven v9.0, and it appears you have:

Does Tesla Autopilot v9.0:

-Steer to avoid any objects, including vehicles?

-Steer to merge the vehicle into the appropriate lane of traffic or to exit the freeway?

- Make lane changes?

I ask because GM outlines these limitations in their Supercruise owner guide. But, Tesla's autopilot page states these features are possible. Again, I do not know if you have driven Supercruise, so lets take GM's word.

Autopilot

View attachment 339284
View attachment 339285

You seem to be under the impression that having more features or more capabilities implies a more advanced system or indicated progress towards higher levels of autonomy or fully autonomous (replacing a driver).

@Bladerskb has been arguing that Tesla behind in comparison to others on progress towards higher levels of autonomy or having a system so reliable you can replace the driver.


Now first of all, I am not saying that that super cruise or other nonTesla adas systems are more advanced than Tesla AP nor am I saying that other OEMs have ADAS systems that are better or are something that consumers would choose over Tesla AP.

But I agree with @Bladerskb that Tesla is behind mobileye and other automakers on progress towards making a system reliable enough / advanced enough to replace a driver.


Consider this....

Which system is more advanced? a system that is capable of auto lane changing, highway interchanging, auto parking, stopping at intersections, understanding stop signs and traffic signals and waiting for its right of way, taking turns at intersections, running roundabouts, etc, etc, etc

or a system that is only capable of only single lane keeping, but is reliable enough that the human can take their eyes off the road and go to sleep?

I don't want to define "more advanced"... but I will say from a developers perspective.... I can develop the first system(capable of all those features) on my own using open source tools and no other team members in maybe a month or two.....

However, the 2nd system I could not do on my own... and it would take a large diverse team, and years of development and testing and more.

Obviously these are extreme examples. just trying to illustrate a point.


Tesla may be a leader in releasing sexy adas features (which I think are super fun and love to use them)...and will continue to do more of this... All of these features are driver assist, human has to keep their eye on the road and ready to take over 100% of the time... I would say 0% self driving, because for no time period does the autonomous system replace the driver.

While, other automakers and mobileye, are less focused on sexy features that do not replace the driver. They are focused on bringing systems to market with much more limited applications and ODDs, but making systems that for these limited applications / use cases, the driver can take their hands/eyes off the road, and the system does replace the driver for certain use cases.... So this is semi-autonomous (>ADAS)


However, I am not saying these systems will be more desirable than sexy adas systems like Tesla AP, nor do I think they will outsell systems like Tesla AP. ITs just different goals by different types of companies. Personally, I am super excited by the progression of both types of companies

 

[Joel]

You seem to be under the impression that having more features or more capabilities implies a more advanced system or indicated progress towards higher levels of autonomy or fully autonomous (replacing a driver).

@Bladerskb has been arguing that Tesla behind in comparison to others on progress towards higher levels of autonomy or having a system so reliable you can replace the driver.


Now first of all, I am not saying that that super cruise or other nonTesla adas systems are more advanced than Tesla AP nor am I saying that other OEMs have ADAS systems that are better or are something that consumers would choose over Tesla AP.

But I agree with @Bladerskb that Tesla is behind mobileye and other automakers on progress towards making a system reliable enough / advanced enough to replace a driver.


Consider this....

Which system is more advanced? a system that is capable of auto lane changing, highway interchanging, auto parking, stopping at intersections, understanding stop signs and traffic signals and waiting for its right of way, taking turns at intersections, running roundabouts, etc, etc, etc

or a system that is only capable of only single lane keeping, but is reliable enough that the human can take their eyes off the road and go to sleep?

I don't want to define "more advanced"... but I will say from a developers perspective.... I can develop the first system(capable of all those features) on my own using open source tools and no other team members in maybe a month or two.....

However, the 2nd system I could not do on my own... and it would take a large diverse team, and years of development and testing and more.

Obviously these are extreme examples. just trying to illustrate a point.


Tesla may be a leader in releasing sexy adas features (which I think are super fun and love to use them)...and will continue to do more of this... All of these features are driver assist, human has to keep their eye on the road and ready to take over 100% of the time... I would say 0% self driving, because for no time period does the autonomous system replace the driver.

While, other automakers and mobileye, are less focused on sexy features that do not replace the driver. They are focused on bringing systems to market with much more limited applications and ODDs, but making systems that for these limited applications / use cases, the driver can take their hands/eyes off the road, and the system does replace the driver for certain use cases.... So this is semi-autonomous (>ADAS)


However, I am not saying these systems will be more desirable than sexy adas systems like Tesla AP, nor do I think they will outsell systems like Tesla AP. ITs just different goals by different types of companies. Personally, I am super excited by the progression of both types of companies

I agree with most if not all of that. There is a lot more Tesla can and should do for safety. I do not believe one system is better than the other. They have different objectives.

I believe Tesla does deserve a nod for finally delivering these convenience features even if they are 2 years late. More importantly, Tesla should add in car monitoring and other features to move these convenience features to be actual safety features.

 

[strangecosmos]

or a system that is only capable of only single lane keeping, but is reliable enough that the human can take their eyes off the road and go to sleep?

...While, other automakers and mobileye, are less focused on sexy features that do not replace the driver. They are focused on bringing systems to market with much more limited applications and ODDs, but making systems that for these limited applications / use cases, the driver can take their hands/eyes off the road, and the system does replace the driver for certain use cases.... So this is semi-autonomous (>ADAS)

Nobody yet has such a system in production...

Moreover, Tesla is developing such a system, just like Mobileye.

 

[Bladerskb]

Well, It was not doing that in the past. I did not get to a Cadillac dealership today and I don't think they are open on Sunday. We'll see if I manage to do it on Monday, but I have this distant memory they don't have any lane changes at all other than the ones you perform yourself?

Also what's "advanced lane change"? Because I don't think I do anything different than the car at this point.

High speed Lane change on highway is piss easy. Again i showed you a vid from audi in 2016. Also remember that Delphi completed a cross country drive autonomously in 2015.

But you want to see what advanced lane change looks like?

13m 50s

I also haven't seen video of interchange merging. But here's the message i wrote to @Snuffysasa Last year,

And about automatic lane change, it will be disappointing because...It won't change lanes in low speed/traffic jams and it won't be doing processing such as... is there obstruction/stopped car/barrier/lane closure upcoming in the lane that i want to turn into?

Basically it wont monitor the environment and will be prone to too many mistakes to be useful

Also as you mentioned freeway transition. when merging into a new freeway and also when changing lane while on the ramp and having to interact with other cars in order to get to the lane to merge into the new freeway. all these things wont be possible.

its literally a glorified toy that you have to monitor moment by moment.

All the things i said above will become apparent once the flood of videos and experiences start coming out.

 

[strangecosmos]

Bladerskb, are you comparing demo videos to actual production features? In that case, check out this demo video:

More impressive than anything Mobileye has in production!

I think it only makes sense to compare demos to demos, and production features to production features. Not demos to production features.

 

[Bladerskb]

Bladerskb, are you comparing demo videos to actual production features? In that case, check out this demo video:

More impressive than anything Mobileye has in production!

I think it only makes sense to compare demos to demos, and production features to production features. Not demos to production features.

did you even watch the video and timelapse i provided?

 

[BigD0g]

Bleh, hopefully auto lane change isn’t gone, that’s what made it so impressive. Just telling me what lane I should be in and having me use the stalk is not nearly as interesting.=\

It's most definitely gone, at least for now....

 

[Bladerskb]

They are limited by data. Most of them don’t have enough “production” miles driven to get a true estimate of real world failure rates. In particular you need billions of miles to get a statistically significant estimate of miles per mortality. Simulation is not enough. Simulation is doomed to succeed.

Mobileye's Amon talks about it here, using a more advanced Monte Carlo reinforcement learning technique similar to alpha go zero and how its data independent therefore will not require human data which @strangecosmos is suggesting. It will generate all its own data and all possible path/response/outcome in simulation.

3mins 50 secs

 

[verygreen]

But you want to see what advanced lane change looks like?

ah, so "madmax mode"? that's already included. In a production build, no less.

All the things i said above will become apparent once the flood of videos and experiences start coming out.

Nope. Reportedly they yanked it out. though I guess people would still try by allowing the car to do it after they signal their allowance anyway (the autopilot game of chicken?)

The most amazing thing is I, you and anybody else (with a Tesla car of a certain vintage) might run this TODAY, and it works. Not some testing car after a bunch of NDAs and whatnot.

 

[S4WRXTTCS]

It's most definitely gone, at least for now....

In what version.

From a different thread.

"The second tech I spoke with also apologized for the fact that the rooted-Tesla community downloaded the software early and leaked features without having been officially sent the latest firmware."

So we really have no idea what's in the latest firmware.

 

[verygreen]

So we really have no idea what's in the latest firmware.

Well, if history is any guide, the "latest firmware" was not even built yet.

 

[Bladerskb]

ah, so "madmax mode"? that's already included. In a production build, no less.

uhh no, not even close...
mad max is a joke compared to what's required in the real world.
Show me Tesla AP 9 doing the above.

 

[strangecosmos]

did you even watch the video and timelapse i provided?

Yeah, it's a demo video. Not a production system. Don't compare demo videos to production systems. That doesn't make sense.

Autopilot v9 is a production system. A few non-early access customers already have it. According to rumour, it's going to be pushed to the fleet more widely starting on Monday. There are about 225,000 HW2 cars out there. This is a wide-scale production release.

We will see how v9.0 does, and we will see how it improves over time. Seems unclear right now whether autonomous lane changes will be part of the wide release, but if they aren't in the initial v9.0 update, we will still see them eventually.

Ideally, we would get autonomous lane changes on Monday, so we can see what the system is like in Mad Max mode. Maybe it's not aggressive enough for Israeli freeways, but it's only intended for U.S. use at the moment. Different aggressiveness settings might be needed for different countries.

verygreen tried v9 in Tennessee (or wherever they are) and said the car was able to autonomously negotiate lane changes by nudging into the next lane and getting other cars to make room. That's pretty cool for a system in production right now.

 

[verygreen]

mad max is a joke compared to what's required in the real world.
Show me Tesla AP 9 doing the above.

and you experienced the madmax mode exactly where to be sure?

It's a bit hard to arrange the same setup as in the demo video because: it's probably illegal to fly drones over the highway (I don't keep track of regulations), I don't have a drone to do it anyway, I don't control the traffic, I don't live in LA or any other such place with terrible traffic and so on. But I suspect it will work decently well in that scenario. And most importantly it's possible to try and verify that theory for regular people like you and me.

 

[Bladerskb]

Yeah, it's a demo video. Not a production system. Don't compare demo videos to production systems. That doesn't make sense.

Its NOT a demo video or a system. You clearly did not watch it after i replied a second time for you to watch it.
Its a drone footage of cars navigating highway. Its hard to take you seriously when you don't even look up what you are responding to.

Ideally, we would get autonomous lane changes on Monday, so we can see what the system is like in Mad Max mode. Maybe it's not aggressive enough for Israeli freeways, but it's only intended for U.S. use at the moment. Different aggressiveness settings might be needed for different countries.

Its basic lane change and none of the above in the video will be happening. that drone footage is not exclusive to Israel. That happens all over the US. you see that all the time during low-speed traffic jams.

If you actually think a tesla in AP V9 in a 5 lane freeway in a traffic jam, will go from the first lane to the last exit lane in under 100 meters or the things in the vid then i want to have whatever you are sipping!

AP V9 as I said last year and 2 years ago WILL be a glorified high speed lane change system period!
I have never been wrong, my biggest nemesis @stopcrazypp learned the hard way when he tried to contend with me and lost on every count.
He still hasn't shown up to face the music after his final vow that Tesla will have FSD (complete/almost complete) by end of 2018. All my statements are based on facts and level headed rationale.

 

[croman]

and you experienced the madmax mode exactly where to be sure?

It's a bit hard to arrange the same setup as in the demo video because: it's probably illegal to fly drones over the highway (I don't keep track of regulations), I don't have a drone to do it anyway, I don't control the traffic, I don't live in LA or any other such place with terrible traffic and so on. But I suspect it will work decently well in that scenario. And most importantly it's possible to try and verify that theory for regular people like you and me.

You can get a cheap FAA license. Interstates might have other regulations prohibiting drone flight but as long as you don't go above a certain altitude and have the $10 FAA drone license, you could maybe also make a badass driving video.

 

[strangecosmos]

Its NOT a demo video or a system. You clearly did not watch it after i replied a second time for you to watch it.
Its a drone footage of cars navigating highway. Its hard to take you seriously when you don't even look up what you are responding to.

My apologies. I've watched most of the Amnon Shashua videos on YouTube including that one, and I assumed the drone footage was the lead-in to the Israeli highway path planning demo. I didn't understand the point you were trying to make.

Yes, some lane changes are much harder than others. We will have concrete evidence once v9 is widely released and people test it in Mad Max mode. Even if the driver has to pull the stalk to confirm the lane change, presumably the car will still autonomously negotiate the lane change as verygreen described.

It is impossible to say conclusively which lane changes Mad Max mode is and isn't capable of until we see it tested more broadly out in the wild. If v9 is indeed widely released on Monday, hopefully we will see somebody on YouTube attempting a low-speed lane change in Mad Max mode.

How certain are you that this is impossible? Would you change your signature to "I was wrong!" if someone can demonstrate a low-speed lane change is possible with v9?