Wiki Tracking FSD Feature Complete

[EVNow]

Thanks for updating! I have one more suggestion. In their autonomy day, Tesla (Karpathy) said that Tesla's current strategy is not to explicitly recognize turn signals, but instead to holistically recognize "intent to turn". This is because sometimes people don't use the turn signal and sometimes people leave it on. So far they have implemented this for "cut-ins", or cars that are about to change into your Lane in front of you. So maybe it makes sense to replace turn signal recognition with "intent to change lanes" and "intent to turn". The first has already been implemented on ap2.

In fact, maybe there should be an entire section on "predict other drivers' behavior". This could include the above, plus the likelihood that stopped traffic at a crossroads will pull out in front of you (including 4-way stop signs), and also the intent to "let you in" on a lane change or "let you go first" on a narrow 2-way road or bridge that can only accommodate 1 car. I'm sure there are more examples that could be added.

I actually meant turn sign within traffic signals. Not the turn signal of a vehicle in front.

I think the predict other driver's behavior is a sub-feature under each of the features. So, it is kind of platform level feature. You need to be able to predict how the car that is coming in the opposite direction will behave before you can take a left turn, for eg. Or whether the car parked behind you is also trying to come out of the parking slot to be able to safely pull out of the parking.

So, just as they had to predict whether another car is cutting in to be able to do safe lane switches, or each of the features, they have to figure out the various obstructions that can happen on the way and be able to predict those obstructions.

 

[NateB]

I actually meant turn sign within traffic signals. Not the turn signal of a vehicle in front.

I think the predict other driver's behavior is a sub-feature under each of the features. So, it is kind of platform level feature. You need to be able to predict how the car that is coming in the opposite direction will behave before you can take a left turn, for eg. Or whether the car parked behind you is also trying to come out of the parking slot to be able to safely pull out of the parking.

So, just as they had to predict whether another car is cutting in to be able to do safe lane switches, or each of the features, they have to figure out the various obstructions that can happen on the way and be able to predict those obstructions.

Oh, that makes sense! Thanks for clarifying. You're right that most of these "behavior predictions" could fall under other features. I guess the only thing I'm thinking of that's not included in your list, then, is the ability to navigate constrictions in a two-way road that force alternating one-way traffic without a traffic light to guide you. Such constrictions are everywhere here in the UK. In the US they are less common, but you still can get them on narrow bridges or on small city/suburban roads with parked cars on the side. In the extreme case where both cars initially enter the constriction, one car actually has to back up and pull over. I have to do this maybe once every 100 miles of city/country driving. More normally, you just have to identify which car will enter the constriction first and the later car waits to enter the constriction. I think I have to do this once every 5 miles or less of city/country driving. This capability is definitely necessary for FSD in the UK, and also, in my opinion, Japan and lots of Europe.

 

[GeorgeSymonds]

A few others if I understand the table at the beginning

In various countries there's also navigating onto a roundabout (I think this is different to the left or right turn example), lane discipline around the roundabout (ie how do you pick the line and filter across to your exit and take that exit)

Dealing with exceptional events like emergency services where you yield and/or make way for priority traffic, in fact that could also be extended to any form of formal direction that is otherwise outside the scope

Single file roads as NateB suggests but also there's a pragmatic nature to giving way - ie a car towing would normally not be expected to reverse even if it was second.

A variation of the single lane roads, there's tight roads where depending on the approaching vehicle size and speed and yours you decide to either moderate your speed so you pass on a wide part or you decide whether you need to stop or continue.

But in practice each of the scenarios mentioned includes many many variations - each of them could have a completely different profile depending on the weather ie on snow you may be more cautious. Plus the layout of junctions can be different.

How do you deal with mapping errors and/or roadworks which temporarily or permanently change the layout?

 

[EVNow]

I have to do this maybe once every 100 miles of city/country driving. More normally, you just have to identify which car will enter the constriction first and the later car waits to enter the constriction. I think I have to do this once every 5 miles or less of city/country driving. This capability is definitely necessary for FSD in the UK, and also, in my opinion, Japan and lots of Europe.

Interesting. We don't have that here - its a rarer condition here. Usually happens in residential roads with parked cars on either side and very rarely would you ever reverse, since the narrow part is only one or two car lengths.

If I were to guess Tesla will not probably not have all the ones I mentioned in the beginning in their list of features to get done for FC. There are probably a few more I should add (even though unlikely to be done by Tesla for FC) - like pulling over for an emergency vehicle.

 

[NateB]

Interesting. We don't have that here - its a rarer condition here. Usually happens in residential roads with parked cars on either side and very rarely would you ever reverse, since the narrow part is only one or two car lengths.

If I were to guess Tesla will not probably not have all the ones I mentioned in the beginning in their list of features to get done for FC. There are probably a few more I should add (even though unlikely to be done by Tesla for FC) - like pulling over for an emergency vehicle.

You're probably right. First "feature-complete" FSD will probably not have this capability nor the the more "advanced" capabilities on your list, but could still probably handle >95% of driving miles in the US without disengagement. I'm very curious how quickly this goes to 99% and then down the "march of nines". My intuition is something like a year per nine, but I have no idea. Falling asleep in car or car driving without a person would require something like >99.999% reliability per mile for crashes (1 accident every 100,000 miles?). US average is something like one accident every 165,000. Reliability for the car getting stuck and not knowing what to do (without crashing) could be somewhat lower, but people would still get pretty annoyed if it's too high. Maybe people would accept 99.9% per mile? Depends if Tesla can quickly intervene remotely to resolve the situation or not.

I'm also pretty sure the rest of the world will be pretty far behind the US in FSD rollout/reliability, but I'm really not sure how far behind. 1 year? 5 years? Of course it will depend at least partly on Tesla numbers in the country.

 

[EVNow]

In 2019.20 they have introduced time dependent speed limits, based on map data. So, one small step towards one of the features.
- Reduced speed based on time

 

[EVNow]

Nav On AP on city streets is available on current build of software available to all in the fleet - if you know how to enable it

Basically the autonomy day demo. But, still WIP.

Hacked Tesla Full Self Driving - Tesla Autonomy Day to San Mateo : teslamotors

ps : Looking at this, clearly Tesla's NOA on city streets is a couple of steps behind Cruise/Waymo. We'll see where they are by the end of the year when they plan to hit feature complete.​

 

[powertoold]

ps : Looking at this, clearly Tesla's NOA on city streets is a couple of steps behind Cruise/Waymo. We'll see where they are by the end of the year when they plan to hit feature complete.​

I believe that is on AP2, but yes, time will tell I'm as confident in a delay as Elon is in FSD soon

 

[Buckminster]

I guess the only thing I'm thinking of that's not included in your list, then, is the ability to navigate constrictions in a two-way road that force alternating one-way traffic without a traffic light to guide you. Such constrictions are everywhere here in the UK.

My assumption is that this will come later. What about this:
FSD legal in China 2022
FSD legal in US 2023
FSD legal in UK 2024

If constrictions are still a problem in 2024 (unlikely), the nav could advise:
1) Do you want to extend journey to avoid constriction
or
2) Take over driving at appropriate time
or
3) Journey cannot be conducted as constrictions are unavoidable and no driver present

 

[EVNow]

My assumption is that this will come later. What about this:
FSD legal in China 2022
FSD legal in US 2023
FSD legal in UK 2024

If constrictions are still a problem in 2024 (unlikely), the nav could advise:
1) Do you want to extend journey to avoid constriction
or
2) Take over driving at appropriate time
or
3) Journey cannot be conducted as constrictions are unavoidable and no driver present

I added the narrow driving space problem to the list (#9). I guess it happens often enough in the US - mostly due to parked cars on the sides.

If they declare feature complete by the end of the year, I think few of the listed features will still be not done. There are a number of these that have not been demonstrated or found in hacked cars. For eg., instead of reading the speed signs (by time !), they may just opt to use detailed maps with time based speed limits.

 

[heltok]

Nav On AP on city streets is available on current build of software available to all in the fleet - if you know how to enable it

Basically the autonomy day demo. But, still WIP.

Hacked Tesla Full Self Driving - Tesla Autonomy Day to San Mateo : teslamotors

ps : Looking at this, clearly Tesla's NOA on city streets is a couple of steps behind Cruise/Waymo. We'll see where they are by the end of the year when they plan to hit feature complete.​

My take is that this is development control system with production perception system compared to Cruise showing development control and perception. I assume that development perception will be a step change improvement, as AKNET_V9 looks to be 10x bigger than current neural networks. Still behind Waymo but rapidly catching up with Waymo is slow to move from development into mass scale.

 

[EVNow]

My take is that this is development control system with production perception system compared to Cruise showing development control and perception. I assume that development perception will be a step change improvement, as AKNET_V9 looks to be 10x bigger than current neural networks. Still behind Waymo but rapidly catching up with Waymo is slow to move from development into mass scale.

Not sure what you mean by "Cruise showing development control and perception.". They are running on FSD in San Francisco all the time - compared to Waymo which is running in safe suburbs of Phoenix. At this point I'd think Cruise is further along than Tesla on City NOA.

 

[heltok]

Not sure what you mean by "Cruise showing development control and perception.". They are running on FSD in San Francisco all the time - compared to Waymo which is running in safe suburbs of Phoenix. At this point I'd think Cruise is further along than Tesla on City NOA.

Are you arguing that the software cruise is running is available to public? Can I buy a car that has it?

 

[EVNow]

Are you arguing that the software cruise is running is available to public? Can I buy a car that has it?

Neither can you buy a waymo car. But they are not in that business. They aim to create robotaxis.

 

[heltok]

Neither can you buy a waymo car. But they are not in that business. They aim to create robotaxis.

But you can buy a ride with them. However they have selected a limited area to deploy suggesting they still believe their product is in development rather than in production. So semantics about waymo is less clear, thus I choose not to argue about them.

 

[diplomat33]

Nav On AP on city streets is available on current build of software available to all in the fleet - if you know how to enable it

Basically the autonomy day demo. But, still WIP.

Hacked Tesla Full Self Driving - Tesla Autonomy Day to San Mateo : teslamotors

ps : Looking at this, clearly Tesla's NOA on city streets is a couple of steps behind Cruise/Waymo. We'll see where they are by the end of the year when they plan to hit feature complete.​

Informative. But I think we need to remember that hacking to enable a feature on the current build is far from representative of what the final product will be. After all, there is a good reason why Tesla has not already enabled NOA on city streets: it's not ready yet for the public.

 

[EVNow]

Informative. But I think we need to remember that hacking to enable a feature on the current build is far from representative of what the final product will be. After all, there is a good reason why Tesla has not already enabled NOA on city streets: it's not ready yet for the public.

It roughly shows the current state (since these cars don't have HW3) in terms of quality - but they also show the features that are being developed, which is the main point of this thread.

 

[diplomat33]

It roughly shows the current state (since these cars don't have HW3) in terms of quality - but they also show the features that are being developed, which is the main point of this thread.

Does it really? The hackers essentially cheated by turning on NOA on city streets on the current AP2 software which is not designed for NOA on city streets whereas the real NOA on city streets that Tesla will give us when they release it, will be on AP3 and will most likely include a bigger NN. So it shows us the current state of the AP2 software, sure, but does not tell us about what Tesla has in development when they actually release the feature to the public. For example, the video is not going to show us anything that Tesla has in development that has not been uploaded to AP2 cars yet.

 

[EVNow]

Does it really? The hackers essentially cheated by turning on NOA on city streets on the current AP2 software which is not designed for NOA on city streets whereas the real NOA on city streets that Tesla will give us when they release it, will be on AP3 and will most likely include a bigger NN. So it shows us the current state of the AP2 software, sure, but does not tell us about what Tesla has in development when they actually release the feature to the public. For example, the video is not going to show us anything that Tesla has in development that has not been uploaded to AP2 cars yet.

It seems they have only one main branch where they are developing FSD features - but they are turned off unless you are in dev mode. That is what @verygreen is essentially saying.

The fact that all the features shown in the autonomy demo can be recreated by hackers, I think proves the point.

 

[EVNow]

But you can buy a ride with them. However they have selected a limited area to deploy suggesting they still believe their product is in development rather than in production. So semantics about waymo is less clear, thus I choose not to argue about them.

Ofcourse no one has FSD in "production" as such. Given limited information we are trying to figure out the current state of art.