Autonomous Car Progress

[daniel]

You are dismissing the most important point. Waymo has shown they have driverless L4. Achieving that is a huge accomplishment and a big milestone on the path to a generalized driverless solution (L5).

That's like landing a man on the Moon and saying "Nasa has not demonstrated jack when it comes to human space travel." What?!

Tesla has not demonstrated jack squat when it comes to a generalized driverless solution.

And Waymo is full driverless, no safety drivers, no remote operators.

They don't have a "safety driver" but they do have a "remote operator" who can send high-level commands on request. And they even only have this in the most ideal location, and only after producing (and maintaining up-to-date!) high-def local maps. This approach does not scale to general nation-wide driverless operation.

One aspect I found interesting is that Waymo, even with its more advanced driving policy, sensor suite and HD maps, still requires human "fleet response specialists" who can intervene remotely in case a driverless car gets stuck in a situation that it cannot resolve on its own. How is Tesla planning to resolve such situations in a potential robotaxi service using privately owned cars? If they have to hire their own specialists that will probably eat up all the profit for the owner.

You've pinpointed the crux of the matter: It's possible to operate "driverless with remote assistance" in perfect conditions after developing and while maintaining high-def maps. This does not scale even to the city level, much less national. And maintaining the high-def maps will require a lot of labor and computing power. Both Waymo and Tesla are very far from wide-scale Level 4.

I can see a few options:
1) Deploy driverless and if the cars get stuck, they get stuck. Maybe have the passenger take over? That does not seem like a very good option.
2) Require owners to be the safety drivers. Kinda defeats the purpose of driverless but could be a temporary solution until the software is good enough to go driverless.
3) Wait until the software is good enough that it does not get stuck before deploying ride-hailing. This seems like it would take too long.
4) Only deploy driverless robotaxis in a geofenced area like Waymo is doing. Hire some "fleet response specialists" that can monitor the fleet of Tesla robotaxis in the geofenced area.

1) You can't have taxis that are limited to taking only licensed drivers as passengers.

2) If the owner is a safety driver, it's not a robotaxi. It's an Uber knockoff.

3) This is the obvious solution.

4) This is a development/testing system. Perhaps a necessary step towards true Level 4 autonomy, but not really an indication of how close they are to either eliminating the remote operator or expanding into unmapped areas and regions with less than perfect weather.

Also note that a robotaxi is zero improvement over a regular taxi or Uber to the user: If a car comes to pick me up, I don't care whether it has a driver or not. What I want is my own car to be able to drive me while I nap in the back. Robotaxis are a step in the development of driverless cars. They are not the goal. (Unless all you care about is not having to listen to your taxi driver's choice of radio station.)

Well, at the moment it's science fiction anyway. I doubt that the cars that Tesla is currently building will ever be used as robotaxis.

Correct. Today's Teslas lack the sensors to ever be truly driverless. They do, however, have far and away the best Level 2 systems an ordinary consumer can buy today.

Google is quite a ways from covering the entire country (if that is even a goal). Here, for example, is the coverage in Austin, MN:
View attachment 597613

You are correct that Google Earth does not show Street-Level View for Austin, MN. However, your image appears to be Google Maps, not Google Earth. On Google Earth you can zoom in to see individual houses large and clear. I'm too lazy to take a screen shot, post it to a hosting site, and link here. But even in Google Maps you can zoom in to see more than adequate detail for navigation. Your image is a very wide view.

 

[Knightshade]

Also note that a robotaxi is zero improvement over a regular taxi or Uber to the user: If a car comes to pick me up, I don't care whether it has a driver or not. What I want is my own car to be able to drive me while I nap in the back. Robotaxis are a step in the development of driverless cars. They are not the goal. (Unless all you care about is not having to listen to your taxi driver's choice of radio station.)

FWIW, this is not correct if the RT is an EV that gets operating costs down to what some have speculated they could- as it'd suddenly potentially be cheaper to use an RT that your own car a fair amount of the time- something that's rarely true today with ICE cars that need to pay humans to drive them.

I don't buy into the idea RTs will entirely (or even majority) replace personally owned cars, but can easily see them replacing stuff like a 2nd or 3rd vehicle you only keep around for occasional need that easy/cheap access to RTs could replace for lower annual cost.

 

[bd7349]

Thought this was an interesting tweet thread from Elon on how the FSD chip works. I had no idea the two SoC’s ran different neural networks to utilize the full 144 TOPS that the FSD chip can do. For context, he was replying to this tweet from an ARK Invest analyst.

 

[Knightshade]

Thought this was an interesting tweet thread from Elon on how the FSD chip works. I had no idea the two SoC’s ran different neural networks to utilize the full 144 TOPS that the FSD chip can do. For context, he was replying to this tweet from an ARK Invest analyst.

I'd imagine that's only regarding the re-write.

IIRC Green had previously confirmed that until very end of 2019 node 2 ran nothing at all and since then has only been running a fully redundant copy of node 1 to take over in case of node 1 crashing.

So if they're now running different stuff, that's a significant change.

Tesla starts activating HW3's Autopilot Dual Redundancy in latest update


What's weird is it'd be problematic for L4/L5 if it REQUIRES both nodes running different stuff- because that kinda eliminates the whole redundancy thing.

 

[Eno Deb]

You've pinpointed the crux of the matter: It's possible to operate "driverless with remote assistance" in perfect conditions after developing and while maintaining high-def maps. This does not scale even to the city level, much less national. And maintaining the high-def maps will require a lot of labor and computing power. Both Waymo and Tesla are very far from wide-scale Level 4.

I think obtaining and maintaining some form of HD map at scale is not as difficult as some make it out to be. Google and Apple have proven that you can map streets at national scale with a relatively small fleet of dedicated cars, and Mobileye has started using crowdsourcing to build and maintain mapping data. This is probably a far easier problem to solve then attempting generalized sensor-only autonomy that works anywhere without map data. Even for humans it can be challenging to drive in unfamiliar environments, and AI technology today is far away from human capabilities in that regard.

 

[diplomat33]

You've pinpointed the crux of the matter: It's possible to operate "driverless with remote assistance" in perfect conditions after developing and while maintaining high-def maps. This does not scale even to the city level, much less national. And maintaining the high-def maps will require a lot of labor and computing power. Both Waymo and Tesla are very far from wide-scale Level 4.

Changes to the map are automatically detected by cars and uploaded and shared with the fleet. So a lot of the process of updating the maps is automatic. So I believe it is scalable to the city level at least.

"We’ve automated most of that process to ensure it’s efficient and scalable. Every time our cars detect changes on the road, they automatically upload the data, which gets shared with the rest of the fleet after, in some cases, being additionally checked by our mapping team."

Waypoint - The official Waymo blog: The Waymo Driver Handbook: How our highly-detailed maps help unlock new locations for autonomous driving

 

[heltok]

They don't have a "safety driver" but they do have a "remote operator" who can send high-level commands on request. And they even only have this in the most ideal location, and only after producing (and maintaining up-to-date!) high-def local maps. This approach does not scale to general nation-wide driverless operation.

I think the idea is to have the backup driver for when the car is uncertain. The input will then be used to train the algorithm so that the cars need fever interrupts over time. As they increase in scale they will decrease the need for backup drivers per vehicle.

Imo let’s give Waymo the credit they deserve. They have a very smart team and have been working on this problem for a long time. I know some of them, my engineering classmate who now works for them was the best student in our class of 150 people. Tesla’s approach might be better, future will tell, but until next week at least, Waymo is the leader.

 

[mikes_fsd]

Changes to the map are automatically detected by cars and uploaded and shared with the fleet. So a lot of the process of updating the maps is automatic. So I believe it is scalable to the city level at least.

"We’ve automated most of that process to ensure it’s efficient and scalable. Every time our cars detect changes on the road, they automatically upload the data, which gets shared with the rest of the fleet after, in some cases, being additionally checked by our mapping team."

Waypoint - The official Waymo blog: The Waymo Driver Handbook: How our highly-detailed maps help unlock new locations for autonomous driving

Except we have yet to see it scale in any sense (not even the same metro)!

 

[diplomat33]

Except we have yet to see it scale in any sense (not even the same metro)!

We will.

 

[mikes_fsd]

We will.

Just like I am still waiting on the Waymo One approval.
....

And seems the "open to the public" is even more limited than the original Waymo One pilot (the NDA based area was larger)
https://twitter.com/scottwww/status/1315707566547034112

I didn't realize this initially, but that's you arguing in that thread....

 

[diplomat33]

Just like I am still waiting on the Waymo One approval.
....

Stop trolling. You know full well that you are not in the service area. Waymo One is only for the Phoenix East Valley area.

And seems the "open to the public" is even more limited than the original Waymo One pilot (the NDA based area was larger)
https://twitter.com/scottwww/status/1315707566547034112

I didn't realize this initially, but that's you arguing in that thread....
View attachment 597790
View attachment 597791
View attachment 597792

Yes that is me because I am trying to correct misconceptions.

The NDA area was larger but it had safety drivers. Waymo is starting the public in the smaller area that is fully driverless and then the public will get to ride in the bigger area with safety drivers first and then fully driverless later:

"After giving people the opportunity to experience the fully autonomous rides, Waymo will begin to reintroduce safety drivers in the vehicles, which allows the vehicles to operate in the larger area than those with no safety driver.

"But for, I don't know, we'll see, the next several weeks, perhaps a month or more, every ride, 100% of rides with Waymo One, will be fully driverless," Krafcik said."

Ready to ride in a driverless car? Waymo vans going public

 

[mikes_fsd]

George Hots of comma ai does seems to think you're the one with the "misconceptions" (if the timestamp does not come through, start at 20:55 )

 

[diplomat33]

George Hots of comma ai does seems to think you're the one with the "misconceptions" (if the timestamp does not come through, start at 20:55 )

Seriously?

Your source is Hotz? He has no credibility. He's a hacker. You believe his word about Waymo?

 

[mikes_fsd]

Seriously?

Your source is Hotz? He has no credibility. He's a hacker. You believe his word about Waymo?

Seriously, that's your rebuttal?

From what I've seen him "hack" he has way more understanding of the problem than you!

 

[Knightshade]

Yes that is me because I am trying to correct misconceptions.

The NDA area was larger but it had safety drivers. Waymo is starting the public in the smaller area that is fully driverless and then the public will get to ride in the bigger area with safety drivers first and then fully driverless later:

If you think what you just described qualifies as "efficient and scalable" in any remotely sane sense, I think you might be the one with misconceptions.

Also if the safety drivers is a step to removing the safety drivers, why did they have to shrink the area from the one they previously had the safety drivers in?

 

[diplomat33]

From what I've seen him "hack" he has way more understanding of the problem than you!

No, he does not. You just like what he says because he agrees with Elon. Compared to companies like Waymo, he knows very little about autonomous driving.

 

[Bitdepth]

 

[AlanSubie4Life]

Thanks for the video! Looking forward to a lot more (in their entirety rather than spliced together).

TOTALLY off-topic: Can't believe the guy asked "I don't know why we're wearing masks, we're the only ones in the car." He definitely should have cut that part of the video.

This is the (one of the many) reason(s) this goes on and on. Everyone should know the answer to that question!

 

[daniel]

I think the idea is to have the backup driver for when the car is uncertain. The input will then be used to train the algorithm so that the cars need fever interrupts over time. As they increase in scale they will decrease the need for backup drivers per vehicle.

Imo let’s give Waymo the credit they deserve. They have a very smart team and have been working on this problem for a long time. I know some of them, my engineering classmate who now works for them was the best student in our class of 150 people. Tesla’s approach might be better, future will tell, but until next week at least, Waymo is the leader.

Absolutely, they deserve a mega-ton of credit for what they've accomplished. It is mind-boggling. But it's still just the first baby step.

I was addressing claims that they have basically solved autonomous driving. The majority of the work is still ahead of them if the goal is a consumer car that does not need a remote operator and can operate on any public road and under any conditions that a cautious typical driver would drive in. When I lived in North Dakota I did not drive in a snowstorm and I did not drive after a snowstorm until the snowplows had come through, so I would not expect a Level 4 or 5 car to do so either. But expanding into less-than-perfect weather, and expanding throughout the nation from Phoenix East Valley is not a trivial matter and they have a very long way to go.

OTOH, Tesla is not going to get to Level 4 without some serious upgrades to the sensor suite.

Some folks are making a big deal out of taking the driver out of the robotaxi. I regard that as more of a publicity stunt than a real milestone, considering how limited are the conditions and area involved.

But to repeat, I am impressed as all get-out by what they've done so far. And yet, what Tesla has done (Level 2 lane-keeping and speed control) has impacted me personally in making my driving safer and more enjoyable.

 

[mikes_fsd]

OTOH, Tesla is not going to get to Level 4 without some serious upgrades to the sensor suite.

To anyone on this thread & forum, that keeps spewing this $#!t out.
Put your money where your mouth is.

I am willing to bet 1 share for every calendar year that the sensor suite on Tesla's will stay essentially the same.
Sensor suite being, cameras, radar ultrasonics (maybe an exterior mic)
The housing will change.
The camera versions will change.
The radar and ultrasonic sensors will be upgraded to whatever latest incarnations are on the market.
There will not be new sensors added in addition to the current sensor suite. And definitely no Lidar!


How it works.
Each year (until December 31, 2025)
On December 31st if changes to the sensor suite have been made to any production vehicle at Tesla, I transfer the closing price of 1 Tesla share from the last trading day of that year.
If on December 31st, there have not been major changes to the the sensor suite, you transfer the closing price of 1 Tesla share from the last trading day of that year.

This offer is open until October 31, 2020 and the long running bet will end December 31, 2025
You have a chance to "earn" 6 Tesla shares if you are right.

Who will take me on this bet?