Why AP 2.0 Won't Be Here Soon, and It Won't Be What You Think It Is

[Canuck]

I was wrong about the 'more radars' but pretty good on the rest.

Braggart!

 

[HankLloydRight]

I just think about all the unpredictable things that happen on the highway that AP isn't coded to handle. I've posted about them before. I think 80% of the time AP2.0 would be measurably more safe than a human, which is great. But that last 20% beyond the envelope of what's normally expected, and where people won't be paying attention, will be measurably less safe than a human driver. So maybe it's a net positive in the end, who knows.

 

[MarkS22]

Rather than pick on those who were wrong, I'll toot my own horn. I posted the above back on Sept. 13 as post #78 in this thread.

But there was no additional radar. What a horrible prediction!

 

[ohmman]

You're right, I did say that. But it wasn't for AP 2.0. It was for a new browser.

Updated Broswer on its Way!

But you are correct in thinking I was equally skeptical about AP 2.0 coming out anytime soon. But all Tesla really did was a PR stunt to release the hardware for AP 2.0 or "Full Self-Driving Capability". Which really isn't that hard. It's the software that we won't see for how long? Oh, and AP 2.0 can't even do what AP 1.0 can do today!

Don't get me wrong, I love Tesla and everything they're doing. This is the third time I've guessed completely wrong on what they're releasing, so I've learned my lesson. But with the AP 2.0 announcement, there are A LOT of small print/weasel words to not really be on the hook to deliver anything other than some cameras and radar units on their cars.

I didn't actually know your position on this. I was just playing the stats.

 

[bob_p]

The AP 2.0 announcement was expected - and it makes sense they would announce that at the beginning of a quarter, after pushing hard to clear out as much inventory as possible at the end of the previous quarter.

Improved and more sensors and increased processing was expected.

And having this go into production immediately is what Tesla has done for all previous changes - they don't pre-announce changes - they announce the changes when they are in manufacturing, and then have to work with orders that are in the pipeline to transition those customers to the new pricing/features.

What was unexpected was Tesla's claim that AP 2.0 will be able to support full self driving. The expectation was that AP 2.0 was a major step forward, and that at least one more step might have been needed to achieve self-driving.

Does this mean there won't be AP 3.0 - no. This is an emerging technology area - with technology rapidly improving - and as volumes increase, prices will drop. AP 1.0 lasted 2 years. We shouldn't be surprised to see AP 3.0 in 2 to 3 years, with further sensor improvements, and possibly some price reduction. Though if AP 2.0 can really achieve self driving, then AP 3.0 will be a refinement, not a major feature upgrade.

Of course, Tesla may be overreaching on their claims of self driving with this hardware. The posted video looks pretty convincing, but there's a big difference in a staged demo - and implementing something that can operate reliably in real-world conditions - plus there are the regulator and liability hurdles (when the car is driving itself, isn't Tesla the "driver"?).

When the Model S was introduced - it was far ahead of the competition as a long range EV - and they've had 4 years in the market on their own. Chevy will change that soon with the Bolt.

AP 2.0 has the potential to reset the stage again. Tesla's strategy of deploying the hardware quickly, and then using the data collected from those cars to help refine the software is very clever - and again puts Tesla years ahead of their competition.

If Tesla really can achieve full self driving with AP 2.0 - and a 100D has 20% more range than our P85, that combination may be good enough to justify keeping our next Model S a very, very long time (which will make it easier to justify the extra $5,500 for AP 2.0).

 

[⚡️ELECTROMAN⚡️]

Ordered my MX today after speaking with the rep at the Tesla store who was "absolutely sure" that my X will include AP 2.0

Estimated delivery near end of October.

Is that turning out to be true?

 

[⚡️ELECTROMAN⚡️]

It's not rocket science to add a few more cameras and RADAR sensors. I don't think the next hardware upgrade will enable full autonomy, but it could be a big improvement over the current hardware. I'll be surprised if they don't start installing more sensors by the end of the year. Elon has recently said that the hardware exists and that full autonomy is coming sooner than you expect. I don't think a genius would say this unless he was planning on upgrading the AP hardware in the very near future.

Was I wrong about full autonomy? Only time will tell, but I'm guessing the new hardware won't be enough for level 5. Not saying I'm smarter than Elon, just no where near the optimist that he is.

 

[hollywood]

Is that turning out to be true?

The delivery part was not true, pushed into December.

 

[Johan]

There is a lot of speculation that Autopilot 2.0 is imminent--perhaps coming within a quarter or two. There's also a lot of speculation that Autopilot 2.0 will be capable of level 4 autonomy.

I'm going to put myself out there and say: Not a chance.

I should "color" this opinion, as the financial analysts like to put it, by stating that:

1. I am a huge Tesla fan
2. I am a huge Elon Musk fan
3. I am a fan of technology
4. I am a happy early Model S adopter (reservation January 2010, first S delivered Dec. 2012).

But this forum tends to be the king of manufacturing unrealistic expectations. I'll summarize this post right off the bat by saying the following: If you are putting off a Tesla purchase waiting for autonomous driving, you're wasting your time.

You will probably not see AP 2.0 hardware for another few quarters at the very earliest. And that's being optimistic.

You will not see AP 2.0 (what I am using to characterize level 4 autonomy) for about 5 years. And that's being VERY optimistic, even keeping in mind the blindingly fast pace Tesla is moving with this technology.

You will not see level 3 autonomy from Tesla for at least 2 years. And that's being optimistic.

Let me explain.

Autopilot has been out for a year now. While improvements have been remarkable, think about the basic highway driving scenarios that aren't handled yet.

1. Every time I crest a hill, my car dives for the left or right side of the road unless I'm following another car.
2. Every time the lane markings fade, the car drifts and I have to take over.
3. Every time I pass an entrance ramp with cars merging, I have to take control. AP does not handle sequencing itself with merging vehicles.
4. Every time I'm merging myself, I have to take control. The car cannot sequence itself to merge onto the highway.
5. The car does not automatically change lanes to maintain a target speed.
6. The car does not avoid large road obstructions.
7. The car does not move laterally to avoid parked cars on the shoulder, bicyclists, or pedestrians.
8. The car is not always clear about which lane a car ahead is in. Sometimes TACC slows for a car in an adjacent lane.
9. If a car cuts in front of you, the braking is later than most would consider to be comfortable.
10. The car cannot stop as smoothly or gradually as a human would (though something close to this may be coming in 8.0).

These are just a few scenarios in the simplest driving environment--on a highway. While some of these scenarios would clearly benefit from additional hardware, some of them should be perfectly doable with existing hardware: notably 1, 6, 7, 8, 9, and 10. There are more examples of improvements that can come with existing hardware. For example:

-Stoplight and stop sign recognition.
-Automatic speed limit adjustment.
-When following a vehicle, the car should maintain itself over the path that the lead car took over the ground. (This is useful when following a car on a road without lane markings, or when passing through an intersection without lane markings). Instead, it tends to somewhat "cut the corner" and head straight for the lead car, which could put you into the curb or an adjacent lane (or an adjacent car!)
-If following a car into an unmarked intersection and the lead car changes lanes, your car will follow it into the adjacent lane! This will greatly surprise the car next to you and could lead to a bad day.
-In stop and go traffic, if the lead car alternates between moving a few feet forward, then stopping...then moving forward, then stopping...then your car will annoyingly accelerate, brake, etc. Car should be able to recognize that the time average speed of the lead car is low, and glide along gradually at a very slow speed, using little energy or brakes and leading to a smooth slow ride.

These are additional scenarios that are perfectly achievable with existing hardware.

So over a year of autopilot, we've seen improvement in lane holding, smoothness of steering (and, to a lesser extent, braking). contrast, and other things. But there is a *LONG* way to go before we've exhausted AP1.0 hardware capabilities.

So while it's possible that, within a few quarters, Tesla could put out cars with hardware for full autonomy (or at least level 3) and then update them via software over time, I wouldn't expect fully autonomous hardware anytime soon. Why?

1. Achieving level 4 is clearly an interative process. Starting out, Tesla thought they could rely on cameras. After the Joshua Brown accident, they realized the limitation of using cameras as a primary sensor (I believe this caused Tesla to end the relationship with Mobileye) and changed their focus. This will probably happen again. For instance, Elon might be against LIDAR, but he may come around to it if radar processing doesn't turn out to work as well as he'd hoped.

2. Each sensor that's added makes the processing and software that much more involved and complicated.

3. Level 3 and certainly 4 will require an enormous amount of testing and validation. By Elon's benchmark, it has to be an order of magnitude safer than a human driver at the very least. Probably several years worth of testing data once they have a Level 4 system before Tesla says you can ride as a passenger while the car drives. But certainly even longer before the government says it's ok..

4. Building a bunch of cars with level 4 hardware and selling them to customers with a promise that they might be able to use them for level 4 driving 5 or more years down the road doesn't make sense. That would be a money-losing proposition for Tesla, unless there are enough foolish buyers out there to pay for the feature many years before it could potentially even be usable.

Yes, Elon mentioned the car being able to drive from NY to pick you up in LA in about 2 years. (Guess what? As awesome as he is, did anyone ever notice that he's a bit overly optimistic when it comes to time frames?)

Yes, what Tesla is doing is awesome.

Yes, Tesla's gathering FAR more data than *anyone* out there.

But if you extrapolate the improvements we've seen since 7.0 and project that out toward even Level 3 autonomy, you should be able to recognize it will be several years before you can chill out and watch your Tesla automatically merge with traffic and autonomously change lanes. And even longer before it will navigate an intersection and make a turn for you.

Yes, there are "marketing" and "research" videos out there showing what approaches Level 3 and 4 technology. But while you might see something near Level 3 or 4 autonomy in the video, that's a very narrow subset of the curveballs the world can throw at you.

I hate to burst any bubbles, but putting a Level 4 car out there in the real world is many orders of magnitude harder.

Long story short, if you're waiting for the "fully autonomous Tesla" before you put down your deposit, I recommend either buying now or moving on. It's going to be awhile.

It's going to be interesting as hell.

But it's going to be awhile.

Todd, what's your take on things now after the announcement? We know the hardware is going in now, so you were a little off on that prediction, but the question of when level 5 autonomy actually rolls out in the wild is still an open question. Did you change your estimates?

 

[bob_p]

Tesla believes the hardware is capable enough to implement full self driving. The sensors and processing power should be sufficient to work better than human vision - which is something they should have been able to test. So the claims the hardware is capable enough to support self driving are likely credible.

The hard part now is to implement the software to use all of that data, and do better than a human driver in ALL driving conditions. And for that, Tesla is going to use every driver buying a new Tesla as a "driving instructor", teaching the software how to drive. And as the software is refined, they can compare the driver's actions to what the software would have done, and verify how the software is performing compared to human drivers.

Allowing a car to fully self drive is going to require getting the software to work, getting regulations to allow it, and then addressing the pesky liability issue. If there's an accident when the car is driving without human control, who's responsible? Is it the owner (who may not even be in the car)? Or is it Tesla (who is effectively supplying the driver with the onboard software)? The regulatory and liability issues could take much longer to resolve than getting the software working.

What's likely is that we'll see self driving software available, but requiring human monitoring (like AP 1.0) - and that Tesla will require the driver to maintain control of the car - until the regulator and liability issues can be resolved.

And, that's OK - if it makes driving easier and safer.

 

[Todd Burch]

Todd, what's your take on things now after the announcement? We know the hardware is going in now, so you were a little off on that prediction, but the question of when level 5 autonomy actually rolls out in the wild is still an open question. Did you change your estimates?

The hardware came about a quarter or two sooner than I thought, although it makes sense to roll out the hardware before the software is ready to stop people from waiting to order--and it also gives them a large fleet to test on. It was starting to look obvious late in Q3 that now would be an ideal time.

I'm a bit concerned that this "bakes them in" to a hardware set before fully proving L5 autonomy. I would not be surprised if more hardware changes were needed for true L5.

As for when the software is ready for L5, I still think it's several years out, so I wouldn't change my prediction for that at this point. I think L4 will be about 2 years, but L5 another 2 after that, and probably another year or two before regulators see enough conclusive evidence via statistics to allow drivers to sleep.

I think it will still be quite awhile before the car can react to a construction flagger, or make other complex decisions, but basic driving functionality seems doable in about 2 years.

My big question is: what are the redundant systems? Seems like you'd need a second processing chip on an independent electrical system to be able to nap in the car. What happens iF the NVidia chip fails while you're sleeping in the back seat?

 

[electracity]

The hardware came about a quarter or two sooner than I thought, although it makes sense to roll out the hardware before the software is ready to stop people from waiting to order--and it also gives them a large fleet to test on. It was starting to look obvious late in Q3 that now would be an ideal time.

I'm a bit concerned that this "bakes them in" to a hardware set before fully proving L5 autonomy. I would not be surprised if more hardware changes were needed for true L5.

As for when the software is ready for L5, I still think it's several years out, so I wouldn't change my prediction for that at this point. I think L4 will be about 2 years, but L5 another 2 after that, and probably another year or two before regulators see enough conclusive evidence via statistics to allow drivers to sleep.

I think it will still be quite awhile before the car can react to a construction flagger, or make other complex decisions, but basic driving functionality seems doable in about 2 years.

My big question is: what are the redundant systems? Seems like you'd need a second processing chip on an independent electrical system to be able to nap in the car. What happens iF the NVidia chip fails while you're sleeping in the back seat?

The software can't be ready for level 5 until users have driven many miles. The alpha test drivers are Tesla employees. The beta test drivers are Tesla customers.

But many of the edge cases can't be solved by simlpy driving billions of miles. Tesla's autonomous driving lead guy explained this several months ago. So how Musk can claim that the 2.0 hardware is good enough is a mystery. How does he make that claim when the total solution is still undefined?

 

[MP3Mike]

I'm a bit concerned that this "bakes them in" to a hardware set before fully proving L5 autonomy. I would not be surprised if more hardware changes were needed for true L5.

That may be why they upped the prices for AP so much. To cover replacing cameras and/or the GPU if necessary to get to the L5 promised.

My big question is: what are the redundant systems? Seems like you'd need a second processing chip on an independent electrical system to be able to nap in the car. What happens iF the NVidia chip fails while you're sleeping in the back seat?

What does the car do when the battery/motor fail? The car turns on the flashers and comes to a stop. Or it gets a flat tire?

Redundancy is certainly an option and I am sure some has been built in, but how much is anyone's guess at this point.

 

[electracity]

That may be why they upped the prices for AP so much. To cover replacing cameras and/or the GPU if necessary to get to the L5 promised.

I think it is a steal at $8K. Look at what modest infotainment upgrades cost in luxury brands. Tesla is adding a ton of capability for a modest price.

 

[Johan]

The hardware came about a quarter or two sooner than I thought, although it makes sense to roll out the hardware before the software is ready to stop people from waiting to order--and it also gives them a large fleet to test on. It was starting to look obvious late in Q3 that now would be an ideal time.

I'm a bit concerned that this "bakes them in" to a hardware set before fully proving L5 autonomy. I would not be surprised if more hardware changes were needed for true L5.

As for when the software is ready for L5, I still think it's several years out, so I wouldn't change my prediction for that at this point. I think L4 will be about 2 years, but L5 another 2 after that, and probably another year or two before regulators see enough conclusive evidence via statistics to allow drivers to sleep.

I think it will still be quite awhile before the car can react to a construction flagger, or make other complex decisions, but basic driving functionality seems doable in about 2 years.

My big question is: what are the redundant systems? Seems like you'd need a second processing chip on an independent electrical system to be able to nap in the car. What happens iF the NVidia chip fails while you're sleeping in the back seat?

I agree with your realistic estimates. I also think we could define some important milestones such as when can you legally sleep while the car drives and when can the car drive without anyone in it?

Not sure about the redundancy concerns. In cars today, that are drive by wire, there's only one system for steering. And apart from the parking brake there is no second braking system. There's only one gas pedal, only one transmission etc etc. There are many potential points of catastrophic failure in current cars that have no redundancy and we accept that as long as the failure rate is very low.

 

[⚡️ELECTROMAN⚡️]

Maybe they are just looking at the $3,000 full autonomy cost as an interest free loan, if they don't pull it off with the newest hardware. At some point they may have to pay it back if full autonomy isn't achieved.

 

[MarkS22]

That may be why they upped the prices for AP so much. To cover replacing cameras and/or the GPU if necessary to get to the L5 promised.



What does the car do when the battery/motor fail? The car turns on the flashers and comes to a stop. Or it gets a flat tire?

Redundancy is certainly an option and I am sure some has been built in, but how much is anyone's guess at this point.

My biggest concern is putting all three forward cameras in a single location. That's the failure point. A single wet leaf can knock it out. A human has the ability to lean around windshield obstructions.

I would've expected redundant locations (like top-left and top-right corners of the windshield) for both safety, better imaging of the forward view, and seeing around forward obstacles.

Even if the radar is enough to pull you over, it doesn't seem optimal.

 

[MP3Mike]

My biggest concern is putting all three forward cameras in a single location. That's the failure point. A single wet leaf can knock it out. A human has the ability to lean around windshield obstructions.

What you don't think it will be able to trigger the wipers to clear the windshield?

 

[LargeHamCollider]

My biggest concern is putting all three forward cameras in a single location. That's the failure point. A single wet leaf can knock it out. A human has the ability to lean around windshield obstructions.

I would've expected redundant locations (like top-left and top-right corners of the windshield) for both safety, better imaging of the forward view, and seeing around forward obstacles.

Even if the radar is enough to pull you over, it doesn't seem optimal.

I was surprised by the camera placement too, but GPS, ultrasonics, radar, and remaining cameras would be able to get you to the side of the road safely in most cases if the wipers can't take the leaf off. And I do think the cameras were placed where they are in large part so that they'd be covered by the wipers.

Edit: Mike beat me to it.

 

[Todd Burch]

My biggest concern is putting all three forward cameras in a single location. That's the failure point. A single wet leaf can knock it out. A human has the ability to lean around windshield obstructions.

I would've expected redundant locations (like top-left and top-right corners of the windshield) for both safety, better imaging of the forward view, and seeing around forward obstacles.

Even if the radar is enough to pull you over, it doesn't seem optimal.

Valid concern, although I can't think of a time an obstruction like that stuck to my windshield at any appreciable speed, so it's probably a very rare event.

I was surprised that Tesla's gone back to vision systems instead of radar or other ranging equipment like lidar. I wonder how it will handle direct sunlight? The cameras must be blinded at certain sun positions, unless they can depend on other regions of the electromagnetic spectrum in those instances (infrared, for example).