FSD rewrite will go out on Oct 20 to limited beta

[diplomat33]

I already gave examples. Its not rocket science ...

For eg., SAE (or its US affiliate) can say, FSD Level X is when
- It works in top 100 counties, by population
- In all weather conditions, except when certain enumerated emergencies occur​

And as you have noted, standards on testing would be useful too - but without ODD standards, you can't test.

To put this in old software terms, what SAE has now done is given a "conceptual framework". The next step is to give a "functional specification". For those saying SAE doesn't tell "how to implement" - absolutely, functional specs never tell "how" - but only "what".

So basically, you want a classification system that ranks how good the FSD is, where "good" is defined but how big the ODD is? So FSD that works in 1 city is ranked lower than FSD that works in 2 countries. The idea being that the FSD that only works in 1 city is not as good because it can't handle as many driving situations as the FSD that works in 2 countries. Do I understand you correctly?

And you still need to define the differences between driver assist and autonomous.

 

[diplomat33]

We have many implementations of SAE level 1 and SAE level 2. Have those helped any manufacturer or driver to date? If not, then what’s the value in SAE3 or SAE 4?

The value is that SAE L3 and SAE L4 are very different systems. In simple terms, L3 means self-driving with a fall-back driver. L4 means self-driving with no fall-back driver. So there is a big difference.

We already see how this might work in practice. Traffic Jam Pilot says it can handle self-driving on highways at low speeds with slow traffic in adjacent lanes, basically stop-and-go traffic on the highway. When the traffic jam clears up and traffic speeds up, the system asks the human to take over. That's L3 since the human is the fall-back driver when conditions for FSD are no longer true. Waymo offers rides in cars with no safety driver, ie no fall-back driver. That's L4.

SAE 5 seems to be “Everything works, all the time, everywhere in any weather.” So I’m not seeing value there either since that’s likely physically impossible. Not even professional car drivers do that...

The SAE addresses this issue. SAE does not expect L5 to work in areas where a human cannot reasonably drive.

SAE J3016, section 5.6 p25 defines the ODD of L5 the following way:

“Unconditional/not ODD-specific” means that the ADS can operate the vehicle under all driver-manageable road conditions within its region of the world. This means, for example, that there are no design-based weather, time-of-day, or geographical restrictions on where and when the ADS can operate the vehicle. However, there may be conditions not manageable by a driver in which the ADS would also be unable to complete a given trip (e.g., white-out snow storm, flooded roads, glare ice, etc.) until or unless the adverse conditions clear. At the onset of such unmanageable conditions the ADS would perform the DDT fallback to achieve a minimal risk condition (e.g., by pulling over to the side of the road and waiting for the conditions to change)."

SAE J3016, section 8.8, p33 says "level 5 ADS must be capable of ‘operating the vehicle on-road anywhere that a typically skilled human driver can reasonably operate a conventional vehicle,’"

So there is that caveat that "everywhere" is only where "a typically skilled human driver can reasonably drive". So as you can see, L5, does not literally mean everywhere, all the time. There are some exceptions.

Also, the SAE allows for geofencing for L5 based on legal or business reasons that have nothing to do with the technology itself.

SAE J3016, section 8.8, p33 says:

"An ADS-equipped vehicle that is capable of operating a vehicle on all roads throughout the US, but, for legal or business reasons, cannot operate the vehicle across the border in Canada or Mexico can still be considered level 5, even if geo-fenced to operate only within the US. The rationale for this exception is that the geo-fenced limitation (i.e., US, only) is not due to limitations on the technological capability of the ADS, but rather is due to legal or business constraints, such as legal restrictions in Canada and Mexico/Central America that prohibit level 5 deployment, or the inability to make a business case for expansion to those markets."

 

[Thp3]

Again what was the value in SAE 1 or SAE 2? We can argue that complex early regulation in Europe slows, maybe halts, Autopilot development in Europe.
SAE levels IMO have not been seen to add value. Tesla Autopilot is level 2 or 3, but is that helpful?

Regulation should assist in defining useful features customers and users can benefit from. The different WIFI definitions have utility. The regulations on home insulation have utility in buying houses or heating them.

Where is the utility in the physical world of the early definitions of SAE levels? At my observation, it just give us something to argue with, it’s “almost” level 3, the geofence is too limited, etc.

I’m just not seeing the merit until we have driving systems to compare, and when we do, the SAE levels, like European regulations, may not help.

 

[diplomat33]

Again what was the value in SAE 1 or SAE 2? We can argue that complex early regulation in Europe slows, maybe halts, Autopilot development in Europe.
SAE levels IMO have not been seen to add value. Tesla Autopilot is level 2 or 3, but is that helpful?

Regulation should assist in defining useful features customers and users can benefit from. The different WIFI definitions have utility. The regulations on home insulation have utility in buying houses or heating them.

Where is the utility in the physical world of the early definitions of SAE levels? At my observation, it just give us something to argue with, it’s “almost” level 3, the geofence is too limited, etc.

I’m just not seeing the merit until we have driving systems to compare, and when we do, the SAE levels, like European regulations, may not help.

The SAE levels won't tell you if one self-driving system drives better than another. The SAE levels won't tell you how smooth or how safe the self-driving is.

But they do have practical implications:

L1: the car can maintain speed or maintain distance with a lead car but the driver needs to do everything else.
L2: the car can control speed and steering but cannot react to everything so the driver still needs to pay attention to the road and be ready to take over at all times.
L3: the driver does not need to pay attention to the road unless the car asks them to take over. The car will give you advance notice to take over.
L4: No driver needed but the car can't self-drive everywhere. The human is a passenger.
L5: No driver needed. The human is a passenger.

To use your examples, if Tesla's AP is L2, that means that you have to pay attention and be ready to take over at all times. But if Tesla's AP is L3, it means that you don't need to pay attention unless the system asks you to take over. So L2 or L3 makes a difference.

Another example, Waymo is L4. So if you hail a ride on the app, that means that you can get in the back seat and do whatever you want. You are a passenger for the entire trip. But because it is L4, you know that the car can't take you everywhere.

if a company says their self-driving is L5 then that tells you that you can get in the back seat and the car can take you anywhere you want.

So I do think there is a lot of utility in knowing the level because they say a lot about your role in the car.

 

[Bladerskb]

Mobileye's lane choices are better because of human-edited HD maps.

Mobileye's map are crowd-sourced not human edited.

3) I think Tesla's predictions are better than both Waymo and Mobileye right now (see the fsd beta predict parked cars and which cars it should go around), and the predictions will only get much better, very soon (within a year). Meanwhile, Waymo will continue having to hand code a lot of their predictions (like what it means to be a parked car, because they lack the data).

It hilarious that the only one that actually Hand-codes alot of its prediction is Tesla.
https://twitter.com/greentheonly/status/1344289571937386498

FSD beta is actually mind blowing and impressive though. Tesla is making very fast progress towards "level 5" robotaxis, which many people didn't think was possible until 2030.

From what I've seen from the recent Mobileye video, they're behind Tesla.

So Mobileye demonstarted how thry are able to drive anywhere in Germany after shipping their car there, immediately after acouple days with no engineer and they credited that to their REM map which allows them to drive anywhere. Your analysis is that they are behind Tesla.

Mobileye AVs Can Go Anywhere in Germany | Intel Newsroom

What happened to 'its gameover for the competition' and L5 in 6 months which you said because you saw Influencers driving around the streets and having to take over every a-couple miles to avoid a crash?

Shouldn't those statements be credited also to mobileye whose system works in more market and countries than FSD Beta?
I thought being able to drive all over the country was unique to Tesla and what made them in your words "so far ahead, no one will catch up."

Yet mobileye just proved that is completely wrong. You claim one of the biggest problem is lane assignment yet in every video from Brandon and DirtyTesla. There are numerous disengagements to stop it from running red lights, crashing into parked cars, rear ending cars, changing lanes into cars in adjacent lane, running into debris on road including tree branches, running over curbs, cones, potholes, walls, etc. i could keep going.

Sadly, I think Tesla will be so far ahead, no one will catch up.

They are so ahead they still use c++ for prediction and now slowly moving it to NN when the rest of the industry already used NN.
https://twitter.com/greentheonly/status/1344289571937386498

True, but Karpathy and Elon consistently emphasize that any improvement in Autopilot, even minor, is a huge deal because it's country / continent / worldwide. That's a big deal. If Tesla widely releases FSD beta, likely for all USA owners, then consider the implications of that confidence... along with how many features are packed into the FSD beta.

Mobileye's systems are used in more markets and counties than Tesla.

Yup, unless Tesla is faking the improvements by releasing a crippled beta, it's game over. If they're actually able to improve the beta by using real world tester data this quickly, they've got an agile and effective stack that will hopefully allow them to widely deploy the FSD beta to all cars. Once the FSD beta is widely deployed, it's game over for sure.

By the time FSD beta is released to public (production) later in 2021 (well under 1,000 miles per disengagement), Lync & Co zero EV would have started deliveries (Sept) and would have Mobileye's supervision running in the cars. It will come with HANDS-FREE full highway system (like NOA) and you would be able to compare to see if its better, equal or worse than NOA.

It will also come with urban, suburban and rural driving that will most likely include features like:

1. traffic lights, yield and stop signs recognition and response
2. handling round about, traffic circles and T interactions
3. handle three and four way stops.
3. automatic lane changes
4. handling narrow/crowded streets
5. maneuvering around obstacles, on-road debris and stopped/parked cars.
6. Use full range of predictive driving meaning it will anticipate and stop for other actors (like peds on cross walk that's about to cross the street. or someone in drive way about to pull out.)
7. handling construction barriers/cones
8. Enter/exit highway

This is nothing like current AP, but rather the full feature set of FSD Beta but better, equal or worse (reliability, safety and comfort). This means it will have better or fewer disengagements. The only feature it wont have that FSD beta might have is that it probably won't make 90 degree turns for safety reasons. As regular drivers are not trained drivers. Tesla unlike others are much more cavalier on safety. FSD Beta might not even release with all the feature set.

I'm sure sooon enough Mobileye will post a video of their system in US. Seeing as the last time was Munich where their auto partners are HQ'ed. Next time would probably be Detroit and hopefully at CES.

Then we can compare its performance to this video which FSD beta struggled with

Anyways the countdown is almost down to 3 months till Tesla will have Level 5 with human level reliability (level 5 in 6 months as you said). Lets see how that prediction is working.

1) One of the fsd beta's main weaknesses is that it makes many wrong lane decisions.

Just needs better lane assignment it seems.

 

[ZeApelido]

@Bladerskb which models in production (or in future years) will have the full camera / radar ( not necessarily lidar) sensor suite needed for Mobileye's FSD equivalent?

E.g. I am curious how much data are they able to crowdsource from fully-sensored vehicles (not just front facing cameras for instance)

 

[powertoold]

Mobileye's map are crowd-sourced not human edited.

Doubt it, especially for the demo rides we've seen so far.

 

[Bladerskb]

Doubt it, especially for the demo rides we've seen so far.

Compare the Tel Aviv map that was hand crafted versus the Munich map that was crowdsourced. Notice the difference in quality of the map and how not clean it is compared to the Tel Aviv one.

Also you are directly calling Mobileye a liar in which you should report them to the SEC as they are defrauding their investors.

 

[Bladerskb]

@Bladerskb which models in production (or in future years) will have the full camera / radar ( not necessarily lidar) sensor suite needed for Mobileye's FSD equivalent?

E.g. I am curious how much data are they able to crowdsource from fully-sensored vehicles (not just front facing cameras for instance)

2021 lync & co Zero EV

 

[Buckminster]

Solved problem:

Omar did SF to LA FSD with zero interventions.

 

[powertoold]

Also you are directly calling Mobileye a liar in which you should report them to the SEC as they are defrauding their investors.

Nah, Mobileye isn't lying. You're just jumping to conclusions without evidence. Crowd-sourced doesn't necessarily mean humans didn't verify / polish / edit it, especially for the route taken in the recent German video.

 

[Bladerskb]

Nah, Mobileye isn't lying. You're just jumping to conclusions without evidence. Crowd-sourced doesn't necessarily mean humans didn't verify / polish / edit it, especially for the route taken in the recent German video.

It is you who is asserting that the map was specifically edited and polish just for that route.
The truth is quite obvious. By comparing both maps you can quickly tell one looks polished, perfect and the map is truncated at the edges of the demo route. While the other looks messy, with visible seams where all the stitching is going on at, splotches, protruding lines and wiggly lines. This goes in line with their recent statement that their car is driving on the world wide official REM map now and it showcases how robust their system is to mapping errors and inline with their recent marketing of "AV everywhere".

"All streets in #Israel - Mobileye's home base - have been mapped in HD detail. Our test AVs can drive anywhere in the country - including my own new, fully self-driving test #AV..."

https://twitter.com/AmnonShashua/status/1341480316431724545

"Mobileye AVs Can Go Anywhere in Germany"
"The video demonstrates the company’s unmatched ability to drive AVs “everywhere”
"Because Mobileye has invested in a scalable, sustainable approach to mapping, the company was able to drive its cars autonomously on highways and urban roads – pretty much everywhere – within just a few days of delivering the AV to Munich."

Mobileye AVs Can Go Anywhere in Germany | Intel Newsroom

Mobileye also said that what you see IS supervision not their L4 vehicle. This is what you expect with supervision. Running on uncertified map everywhere, with full capability of L4 while just requiring driver attention.
Amnon will go into more on CES 2021 (jan 12).

With tesla you said 'game over', 'no one else can catch up', level 5 in '6 months' (now almost 3).
With mobileye you criticize.
With Waymo you try to point out a 3 second ultra zoomed in clip to claim it struggles with path planning.
When there are thousands of instances of FSD beta attempting to crash/cause accident. When numerous instances in almost each video and yet you totally dismiss it as lane assignment/selection problem remaining.

Do you not see your bias? Don't let you love for Tesla cloud your judgment to stop you from enjoying and appreciating tech.

Perfect L4 certified polished, human edited Map

Crowdsourced map from the worldwide road book with no special editing or polishing other than the typical production fully automated creation/updating process

 

[powertoold]

It is you who is asserting that the map was specifically edited and polish just for that route.
The truth is quite obvious.

...

typical production fully automated creation/updating process

Doesn't matter if it's obvious to you. You're making an assertion without evidence (that it's FULLY automated HD map generation).

To me, it's obvious all those maps in the videos (past and recent) have been human edited.

 

[emmz0r]

Doesn't matter if it's obvious to you. You're making an assertion without evidence (that it's FULLY automated HD map generation).

To me, it's obvious all those maps in the videos (past and recent) have been human edited.

So what if they are edited? as long as it works.

 

[EVNow]

So what if they are edited? as long as it works.

Then, they are not that scalable.

 

[emmz0r]

Then, they are not that scalable.

In the end of the day I am just interested that it works where I drive personally, not the plains of Serengeti
I agree that not relying on crutches is the correct approach, but I am scared that it will take a long time before it's ready.
It's good that competitors take other approaches so you can A/B the best solution. You will have quicker gains in the short run with detailed maps, but Tesla's solution will ultimately be the best where the map is a supplement to keep it on route / have reference points to vision.

 

[Bladerskb]

Doesn't matter if it's obvious to you. You're making an assertion without evidence (that it's FULLY automated HD map generation).

To me, it's obvious all those maps in the videos (past and recent) have been human edited.

"REM technology is actively generating data about more than 15 million kilometers of roads daily to build the map and keep it up to date. "

Now this picture below is from CES 2020. Which says fully automated process at all speeds by 2021. Remind you that fully automated process means there isn't a human going through the hundreds of millions of miles to edit them as you claim. As Amnon likes to say, "Its literally a push of a button no human editing" (paraphrasing)

Its now 1 year later, I'm sure in 9 days Mobileye will update us on how all the Munich map was created and how often its updated. I already know but I will just wait for it to be said publicly.

 

[willow_hiller]

"REM technology is actively generating data about more than 15 million kilometers of roads daily to build the map and keep it up to date. "

"ERROR: please wait for your neighborhood's daily LIDAR scan before enabling autonomous modes."

 

[diplomat33]

"ERROR: please wait for your neighborhood's daily LIDAR scan before enabling autonomous modes."

Mobileye does not use lidar to create HD maps.

 

[KArnold]

"REM for Autonomous Driving...Maps created in a fully automated process TODAY..."

Heck, we have the cameras and communications - so why is it impossible to get a map correction? The above implies maps are somehow created in real-time by real drivers but maybe I'm reading too much into that statement. Although I wish they said "maintained" instead of "created".

There has been an exit on the Tesla map that NAV simply refuses to take for much of this year (it does use them to enter the freeway). Exit appears on map but both directions it will not use it. Of course this is the exit I need. It used to work perfectly. Using the NAV-suggested exit is .4 miles longer. Map is 2020.48.12628, SW is 2020.48.12.1 (and holding).

In retrospect mine is probably not a "maps" issue but real-time/custom updates of maps would be incredible.