Anthony Levandowski beats Tesla's Elon Musk to first Auto Cross country (3k Miles, 0 disengagement)

[Cheburashka]

"The car, a modified Toyota Prius, used only video cameras, computers and basic digital maps to make the cross-country trip.'

How does the computer turn the wheels, apply brakes etc?
Bolt-on hardware?

How did Waymo do it?
These are critical components that seem to never be part of media coverage.
What vehicles can adapt these systems, will they need this or that from the factory etc.

Same way as Openpilot.

If the car has active cruise control/lane keeping support out of the factory (which the Prius does), then injecting data to control steering and brakes/accel is pretty straightforward.

 

[TheTron]

For those saying this is no better than Autopilot, how many miles do you think Autopilot can do without a disengagement? In my experience it’s biggest failing on a highway is its lack of inference on the intent of other drivers. This results in Autopilot being too late to react to obvious developing situations. I’ve managed maybe 100 miles without disengagement, no way would it go 3000 even with a dozen attempts.

This demo must be a more advanced ADAS system on a highway than Autopilot is today. This is pretty disappointing for someone that bought into AP2 back in early 2017.

 

[Bladerskb]

Any AP2.5 Tesla owner can do the same thing today. Root your car, disable nags (or wedge nag defeating fruit onto steering wheel) and load the beta NoA that didn't require confirmation of lane changes. That's not a coast to coast autonomous drive.

So you can do it for 3,099 miles without disengagement? How come the people at
Navigate on Autopilot is Useless (2018.42.3)

Dont know about this they seem to have missed the memo

 

[diplomat33]

For those saying this is no better than Autopilot, how many miles do you think Autopilot can do without a disengagement? In my experience it’s biggest failing on a highway is its lack of inference on the intent of other drivers. This results in Autopilot being too late to react to obvious developing situations. I’ve managed maybe 100 miles without disengagement, no way would it go 3000 even with a dozen attempts.

This demo must be a more advanced ADAS system on a highway than Autopilot is today. This is pretty disappointing for someone that bought into AP2 back in early 2017.

I disagree. I think you and others are making a sampling error. That's where you use too small a sample and get a false result. Let's say you drive AP the same road 100 times. One trip, you might get 3 disengagements. One trip, you might get 30 disengagements. Another trip, you might get 0 disengagements, etc... If I just look at one trip, I will get the wrong idea about AP. If I just look at the trip with 30 disengagements, I will think that AP is worse than it really is. But if I look at the trip that had 0 disengagements, I will think AP is better than it really is. To get a true measurement of how good AP is, I would need to take an average of the disengagements for a big enough sample. Same goes for measuring how good Levandowski's system is. Remember that his first attempt failed. If we were to judge his system on the failed trip, we would conclude that his system is not very good at all. Of course, if we just look at the one trip that with 0 disengagements, we think his system is amazing. You only think AP is worse because you are comparing the worst example of AP with the best example of Levandowski's system. That's a false comparison!

If you want a better comparison, we should look at how many disengagements on AP over the millions of miles that Tesla owners have used AP. Divide the total number of AP disengagements by the total number of miles driven by owners on AP. See if the average is better or worse than 1 disengagement every 3000 miles to get a better sense of how AP compares to the coast to coast trip. And again,let's remember that Levandowski's system failed the first trip and on the next trip was able to do 3000 miles without a disengagement. But one 3000 mile trip is small compared to the millions of miles that AP has driven. How many disengagements would his system have over millions of miles compared to AP? That's what we should look at if we want to truly compare the two systems fairly. Personally, I think if we looked at a big enough sample of both AP and Levandowski's system, we would find that they are actually much closer to each other than we think.

 

[CarlK]

The fallacy of using closed demo to judge capability is you can make the run as many times as you want and announce to the world only the one you want people to know. Russia and China used to only announce successful rocket launchs afterwards until they became more confident of what they do. A credible "demo" needs to be pre-announced so it can be monitored by third party observers.

 

[Chaserr]

So you can do it for 3,099 miles without disengagement? How come the people at
Navigate on Autopilot is Useless (2018.42.3)

Dont know about this they seem to have missed the memo

Not a single system discussed in this thread can navigate for 3099 miles without a disengagement. Did you believe otherwise? Whoops! Since you didn't know any better,the video above and their write up shows they disengaged many, many times, roughly every 200-300 miles on average. I don't think there even was a memo, so if you were hoping for something like that to explain what this thread is about it explains why you've been so aggressively attacking anyone that mentions the assorted other systems that operate the same, while being ignorant and unaware of the how and why of why they work the same. Comprehension of a topic starts with understanding of the topic.

I see no reason to continue, you've admitted you thought one of these systems could operate 3099 miles without a disengagement and that's factually incorrect so I don't think any further discussion will change that erroneous starting point.

 

[Bladerskb]

Sorry, I misunderstood your question because it was not very clear. Well yes, if I get the FSD option on my car and I download the V10 Alpha version that claims to be FSD but I still need to take over, of course I will happily admit that I was wrong. That's not a big deal to me. I am going to be honest when I am wrong.

With all due respect, I don't believe you one bit, from my 2+ years discussing here I have yet to see one Tesla proponent that is honest. The arguments are always done in bad faith with changing criteria and goal posts on the fly to absolve Tesla from any responsibility. So no, in 6 months you will be coming up with another excuse/reason of why you weren't wrong and contradicting your very own statements.

Remember you said this;

So when Nav on AP is able to handle auto lane changes without driver input and thus is able to handle large parts of highway driving on its own without driver input and notifies the driver when it will need to disengage, I would think that comes pretty darn close to Level 3 Autonomy.

Yet when introduced to another system that is 100x better than NOA the response is "just a lane keeping assist", "cool parlot trick".
Again all i want is for you to see the different criteria you use for AP vs other systems. Before yesterday, it would have been a sin if you called what AP did just "lane keeping". A system 100x better than NOA? "phh another lane keeping assist".

Well yes, if I get the FSD option on my car and I download the V10 Alpha version that claims to be FSD but I still need to take over, of course I will happily admit that I was wrong.

Notice i didn't say "claim to be FSD". But I said the version (which they say will be v10 alpha) that the FSD cross country demo is done on.
Hopefully this isn't you trying to already create wiggle room.

EDIT; If you doubt my statement of not one single Tesla proponent (notice i didn't say "fan", because there are fans who are actually honest even opponents who are honest). But I haven't seen an honest proponent.

Charity bets: $100/$1K/$10K -> EAP/FSD feature release by 12/31, 1/31, 3/31

 

[diplomat33]

With all due respect, I don't believe you one bit, from my 2+ years discussing here I have yet to see one Tesla proponent that is honest. The arguments are always done in bad faith with changing criteria and goal posts on the fly to absolve Tesla from any responsibility. So no, in 6 months you will be coming up with another excuse/reason of why you weren't wrong and contradicting your very own statements.

Sounds like you are prejudging me.

Remember you said this;

Yet when introduced to another system that is 100x better than NOA the response is "just a lane keeping assist", "cool parlot trick".
Again all i want is for you to see the different criteria you use for AP vs other systems. Before yesterday, you would have been a sin if you called what AP did just "lane keeping". A system 100x better than NOA? "phh another lane keeping assist".

First, please, please, please stop saying that Levandowski's system is 100x better than NOA as if it is settled fact. It's not. It is your opinion, one that I strongly disagree with.

Second, I am not using different criteria. The guardian article itself describes Levandowski's the system this way:

"Pronto.AI will not be selling Levandowski’s new technology in a self-driving vehicle, nor using it for passenger cars at all. Instead, it will form the basis of an advanced driver assistance system (ADAS) called Copilot, offering lane keeping, cruise control and collision avoidance for commercial semi-trucks."

So, on one hand, we have AP which is a sophisticated ADAS that can do a lot more than lane keeping and that Tesla is actively working towards Full Self-Driving and on the other hand, we have a system that is not 100x better than NOA, that won't be in self-driving cars, won't even be in passenger car and is described as just offering lane keeping, cruise control and collision avoidance. There is a big difference between the two systems. That's why I described NOA as almost being L3 and why I described Levandowski's system as just lane keeping and a parlor trick. because that's basically what they are! When you rig a desktop computer and some cameras to a car and after a few failed attempts manage a coast to coast trip with no disengagements, yeah, it's a parlor trick. EAP and NOA is a serious system that is installed in tens of thousands of cars actually driving on roads every day. Very different!

Also, if Levandowski's system is 100x better than NOA why isn't he planning on putting it in self-driving cars but instead just wants to offer lane keeping and cruise control and collision avoidance in trucks? Surely a system 100x better than NOA should be in every self-driving car on the road. Imagine how much better those cars would be compared to NOA!!! So yeah, I called it just lane keeping because that's apparently all he wants to use it for, according to the article itself.

Notice i didn't say "claim to be FSD". But I said the version (which they say will be v10 alpha) that the FSD cross country demo is done on. Hopefully this isn't you trying to already create wiggle room.

No I am not trying to create wiggle room. Let's see what V10A actually does and whether it achieves FSD or not.

 

[Bladerskb]

So, on one hand, we have AP which is a sophisticated ADAS that can do a lot more than lane keeping

What makes AP a sophiscated ADAS...What does AP NOA do other than lane keeping?

 

[diplomat33]

What does AP NOA does other than lane keeping?

- Auto lane changes
- auto exit taking
- highway transitions
- lane change suggestions based on speed of rear and surrounding traffic
- yielding to merging traffic

Also, EAP has millions of miles of real world driving miles in thousands of cars on the road every day. That makes it a legit ADAS!

 

[Bladerskb]

- Auto lane changes
- auto exit taking
- highway transitions
- lane change suggestions based on speed of rear and surrounding traffic
- yielding to merging traffic

You see what am talking about? the different criteria, goal posts and bizarre contradictions.
Co-Pilot does all of that 100x better. (3,099 cross country miles without disengagement)

Co-pilot does auto lane change WITHOUT driver initiating turn signal or driver confirmation
Co-Pilot takes Ramp exits to transition to another highway by itself
Co-Pilot does highway transitions by itself
Co-pilot DOES'NT suggest lane changes it just TAKES THEM
Co-pilot yields to merging traffic.

Also, EAP has millions of miles of real world driving miles in thousands of cars on the road every day. That makes it a legit ADAS!

Openpilot has almost 10 million miles of real world driving miles in one thousand cars. You are literally just creating criterias on the fly to fit your narrative just like i predicted. Just like you will do after the FSD cross country drive happens and the subsequent software update fails to live up to it.

I have nothing against you but this is your playbook.

 

[diplomat33]

You see what am talking about? the different criteria, goal posts and bizarre contradictions.

No, I don't see what you are talking about. I just stated facts about what NOA can do. How is that a contradiction? I guess you think I am somehow misrepresenting NOA's capabilities to make it look better than Co-Pilot??? I am not.

Co-pilot does auto lane change WITHOUT driver initiating turn signal or driver confirmation

NOA does auto lane change with driver confirmation but will be able to do it without driver confirmation soon.

Co-Pilot takes Ramp exits to transition to another highway by itself
Co-Pilot does highway transitions by itself

NOA does both of these now.

Co-pilot DOES'NT suggest lane changes it just TAKES THEM

NOA will do this soon.

Co-pilot yields to merging traffic.

NOA does this too.

So, NOA does 3 out of 5 already now and will do the other 2 soon.

Openpilot has almost 10 million miles of real world driving miles in one thousand cars. You are literally just creating criterias on the fly to fit your narrative just like i predicted.

I have no idea what you are talking about. What criteria am I making up? Again, I just stated facts about how EAP has million of miles. In fact, according to Tesla's report this year, EAP has 1.2 billion miles of real world driving. That's a fact. I am not making up criteria.

Just like you will do after the FSD cross country drive happens. I have nothing against you but this is your playbook.

You are projecting what you think a "Tesla fanboy" will do unto me.

 

[Bladerskb]

That's why I described NOA as almost being L3 and why I described Levandowski's system as just lane keeping and a parlor trick. because that's basically what they are!

What makes AP a sophiscated ADAS...What does AP NOA do other than lane keeping?

- Auto lane changes
- auto exit taking
- highway transitions
- lane change suggestions based on speed of rear and surrounding traffic
- yielding to merging traffic

Co-pilot does all of that by itself without driver input and is 100x better in performance.

So, NOA does 3 out of 5 already now and will do the other 2 soon.

Your Warped Logic...

 

[thewishmaster]

lol I like how one or two demo videos is sufficient to convince anyone that one system being demoed is orders of magnitude better than another. It's a cool trick and a nice video but it's way too early to judge capabilities. Example: the pronto.ai car almost rear-ended a semi truck when merging onto a freeway early on in the video.

 

[⚡️ELECTROMAN⚡️]

Your Warped Logic...

Who cares? Just state your opinion and then go get a life. No need to keep beating a dead horse.

 

[TheTron]

I disagree. I think you and others are making a sampling error. That's where you use too small a sample and get a false result. Let's say you drive AP the same road 100 times. One trip, you might get 3 disengagements. One trip, you might get 30 disengagements. Another trip, you might get 0 disengagements, etc... If I just look at one trip, I will get the wrong idea about AP. If I just look at the trip with 30 disengagements, I will think that AP is worse than it really is. But if I look at the trip that had 0 disengagements, I will think AP is better than it really is. To get a true measurement of how good AP is, I would need to take an average of the disengagements for a big enough sample. Same goes for measuring how good Levandowski's system is. Remember that his first attempt failed. If we were to judge his system on the failed trip, we would conclude that his system is not very good at all. Of course, if we just look at the one trip that with 0 disengagements, we think his system is amazing. You only think AP is worse because you are comparing the worst example of AP with the best example of Levandowski's system. That's a false comparison!

If you want a better comparison, we should look at how many disengagements on AP over the millions of miles that Tesla owners have used AP. Divide the total number of AP disengagements by the total number of miles driven by owners on AP. See if the average is better or worse than 1 disengagement every 3000 miles to get a better sense of how AP compares to the coast to coast trip. And again,let's remember that Levandowski's system failed the first trip and on the next trip was able to do 3000 miles without a disengagement. But one 3000 mile trip is small compared to the millions of miles that AP has driven. How many disengagements would his system have over millions of miles compared to AP? That's what we should look at if we want to truly compare the two systems fairly. Personally, I think if we looked at a big enough sample of both AP and Levandowski's system, we would find that they are actually much closer to each other than we think.

Well I’m no stats expert but I’ve driven tens of thousands of miles on autopilot and I’d say my average forced disengagement rate is about once every 50 miles. That’s on a sample size of about 20,000 autopilot miles. That’s enough data for me to know to a decent confidence that it would take many, many attempts to go 3000 without a disengagement. I think my experiences of reliability are in line with other experiences I’ve read on this forum.

This guy does 3000 miles in two attempts. Maybe he got really lucky, but it would take a lot of luck to have a system as bad as Autopilot and manage that trip in two attempts.

If Autopilot could replicate this so readily I would expect Tesla to release a similar video of NOA going cross country with no disengagement’s, expect to charge. Failing that, given the fan base of owners Tesla have, one of us could demonstrate the same. Maybe we could set up a prize fund for the first person to demonstrate it!

 

[MP3Mike]

This guy does 3000 miles in two attempts. Maybe he got really lucky, but it would take a lot of luck to have a system as bad as Autopilot and manage that trip in two attempts.

Actually it was at least three attempts.

 

[TheTron]

Actually it was at least three attempts.

I stand corrected. Still find it difficult to imagine a Tesla coming anywhere near that with the current firmware.

 

[CarlK]

I stand corrected. Still find it difficult to imagine a Tesla coming anywhere near that with the current firmware.

By current firmware do you mean Autopilot or Tesla's still unreleased FSD software? Elon said a while ago Tesla's FSD software can already do coast to coast run but it does not want to game the system. Gaming the system I believe means running a pre-mapped route with practice runs probably just what Levandowski has done. A true capability is you tell the car at the driveway where, including anywhere in the other coast, you want to go and it will take you there. That's what every Tesla will need to have. Once again you can't compare someone's "demo" to Tesla's cars that everyone could have one in hand and scrutinize its capability. There will be no engineers there to help you to plan and run every trip.

 

[electronblue]

By current firmware do you mean Autopilot or Tesla's still unreleased FSD software? Elon said a while ago Tesla's FSD software can already do coast to coast run but it does not want to game the system. Gaming the system I believe means running a pre-mapped route with practice runs probably just what Levandowski has done. A true capability is you tell the car at the driveway where, including anywhere in the other coast, you want to go and it will take you there. That's what every Tesla will need to have. Once again you can't compare someone's "demo" to Tesla's cars that everyone could have one in hand and scrutinize its capability. There will be no engineers there to help you to plan and run every trip.

There was one rumour on Reddit or TMC or such early this year that Tesla indeed was attempting the drive with freshly painted Supercharger exits and all that, but it was too brittle. I recall the rumor indeed said it was pretty much hardcoded but it still didn’t fully work out. I think it is at least a possibility they were trying to ”game it”, but even that did not work out.

Just a rumour, but I do recall reading about and afterwards Elon’s comments made sense...