Elon: "Feature complete for full self driving this year"

[iTech]

Most likely yes since the goal is 2-3 months from now and they might be late.

Which isn't that bad given how the FSD computer isn't being used right now and probably won't for a while.

 

[pilotSteve]

It used to be 5k+3k. then 3j + 5k, now AP is included, but the car base price is 2k more + 5k for FSD. So it would be 8k jut for FSD on top of 2k more for the car or ~10k ish...

just remember..... 2016 EAP not the same feature set as 2019 EAP. Apples to oranges feature bundle. However EAP+FSD appears to encompass all original features.

 

[CK_Stuggi]

Well I saw a pretty cool looking video in October 2016. Elon said feature parity to AP1 would be there end of that year.
AP1 feature parity isn't there 30 months in (speed sign recognition is still missing), it was a horrible 1½ year downgrade with AP2, so...

I'll believe it when I see it. In my car.
Which, to this day, doesn't have NoA in Germany as the german M3s have. And the investors were not allowed to make videos. So.... Yeah, I'll wait.

With my experience of Tesla logistics, they better start shipping FSD units to Europe today if they consider installing them in 3-6 months.

Unless, of course, Europe is on the back burner, given NoA is only available in select markets and vehicles 6 months after the US rollout.

I bet I'll not see the upgrade before summer 2020 - if that.

 

[diplomat33]

A discussion in another thread got me thinking about what "feature complete" means. I know folks have different opinions on that. Some folks might have a list of 20 must-haves that they define as "feature complete" while someone else may have a list of 50 must-haves. But it occurred to me that another way to possibly think of "feature complete", and perhaps how Tesla is looking at it, is to think of Autopilot's Operational Design Domain (or ODD). ODD is essentially the driving domain that the self-driving car is designed to operate in. This does not mean that the self-driving car is perfect in its ODD. There may be edge cases in the ODD that the self-driving car still needs to solve. ODD just means that the car is designed to operate in that domain. For example, right now, Autopilot's ODD is highways and well marked roads. Intersections, city streets and roads with no lane markings are outside of Autopilot's ODD. So "feature complete" FSD would mean expanding Autopilot's ODD to include every domain (highways, city streets, intersections, parking lots and roads with no lane markings).

If you look at the FSD feature list on the website, it seems to confirm this way of looking at "feature complete":
- NOA on highways checks the highway domain
- Enhanced Summon checks the parking lot and probably the roads with no markings domain
- Automatic City Driving checks the city street and intersection domain
- Traffic Light and Stop Sign Recognition also checks the intersection domain

So all these FSD features, once complete, would cover the entire ODD needed for FSD. In other words, achieving "feature complete" means expanding AP's ODD to include every domain needed for FSD. And it is also worth noting that expanding AP's ODD to include every domain is a prerequisite for FSD. This is also consistent with what we know Tesla is working on. We know they are working on traffic light recognition, intersections, curbs and parking lots. That's because they are working to add intersections, parking lots and city streets to AP's ODD.

 

[rnortman]

Intersections, city streets and roads with no lane markings are outside of Autopilot's ODD.

Also, I'll point out, quite a few highways are still outside of the functioning ODD, particularly most of the highways I regularly drive on, where Autopilot/NOA routinely (dependably even) fails.

Anyway, by your definition many other companies are already feature complete. They already operate with human supervision in all of those domains. Tesla is late to the "feature complete" game by this definition, by several years, even if they hit that target this year. Which they won't.

 

[Thp3]

I agree. I’d add that Navigate-on-Autopilot (NOA) in all those domains is required for “feature complete”. Full self driving has to “know” where it is going to qualify for “feature complete”, not just capable to operate in the domain.

NOA is the important connection between driving domains. Right now it shuts off when you exit the expressway and there is no NOA in any of the other domains.

I’m really looking forward to NOA on enhanced summon, on two lane highways with stop signs and stop lights and eventually, I predict last, on city streets.

 

[diplomat33]

Anyway, by your definition many other companies are already feature complete. They already operate with human supervision in all of those domains. Tesla is late to the "feature complete" game by this definition, by several years, even if they hit that target this year. Which they won't.

Yes indeed, by my definition, they are also "feature complete". Certainly Waymo is way beyond "feature complete" at this point. And yes, I think it would probably be fair to say that Tesla is late to getting to "feature complete". But my post was not claiming that Tesla is ahead or that other companies are not "feature complete" yet. I was merely making a point about a different way to define "feature complete".

 

[im.thatoneguy]

.yway, by your definition many other companies are already feature complete. They already operate with human supervision in all of those domains. Tesla is late to the "feature complete" game by this definition, by several years, even if they hit that target this year. Which they won't.

I think that's an accurate assessment of Tesla not a flaw in the definition. Arguably Waymo is feature complete. And the fact that waymo is so much further aheadbut still hasn't gone public with feature complete just highlights the gap between what Elon claims they will accomplish by the end of the year and where they need to be to remove the driver.

 

[diplomat33]

I agree. I’d add that Navigate-on-Autopilot (NOA) in all those domains is required for “feature complete”. Full self driving has to “know” where it is going to qualify for “feature complete”, not just capable to operate in the domain.

NOA is the important connection between driving domains. Right now it shuts off when you exit the expressway and there is no NOA in any of the other domains.

I’m really looking forward to NOA on enhanced summon, on two lane highways with stop signs and stop lights and eventually, I predict last, on city streets.

Yes. That is why NOA is so crucial and why Tesla puts NOA in the FSD column. NOA is the connective tissue of FSD. I am looking forward to the day when we get NOA on local roads because then we will have a NOA experience from start to finish.

 

[SOULPEDL]

Just a note to say that after about 3 recent upgrades, my EAP keeps getting more solid. Here's why I say feature complete this year.

Many subtleties and very hard to describe "confidence", but think of it as Deep Knowledge (where humans don't realize how they do things, more like a gut feel):

1. Less wandering when picking a lane or if it widens/narrows. Smoother transitions.
2. City street lane changing about a 99% successful (Chandler Az). I call it the Chicken-Out maneuver. This was ~75% just 2 months ago.
3. Taking the correct lane when exiting the freeway (need more examples to confirm this, maybe a fluke).
4. Merging has improved. Fits into traffic better.
5. After merging onto the freeway in Mad Max, it now just shoots over to the faster lane as if the decision is happening faster. Maybe the move was pre-assessed when it was merging.
6. Doesn't use the turn signal on freeway interchanges (was silly even though correct).
7. Before a freeway interchange, it allows more time in all lanes before moving far right (better now). Meanwhile, others in the past have complained (LA traffic) that it missed the turns and waits too long to get over. So maybe they're judging traffic density when deciding when to move right?

I did have another jump to a lane when the car in front blocks view of a new dotted line/new lane. This was causing the line to appear late so the car suddenly realizes it's in between lanes. This car definitely wants to be in a lane, but this posed no safety risk, and lanes were clear. Keep in mind, not a lot of data here compared to how many upgrades I've been getting lately.

I say feature complete year end, and drivers will also improve with these new co-pilot skills. We are learning how to work and drive WITH the car. It's not us or them, it's WE. And WE are getting better at doing driving this way. So FSD will be feature complete this year IMO, but some features will fail more than others (different environments and the last feature on Dec 31st), and no sleeping yet unless you do it unintentionally, in which case you'd be in pretty safe hands.

 

[diplomat33]

Just a note to say that after about 3 recent upgrades, my EAP keeps getting more solid.

Yes, I've noticed some real improvements in AP as well. Auto Steer is super solid now. Also:
1) AP is no longer confused by a parallel side road and goes perfectly straight now. in the past, AP used to veer a little like it was not sure if it should go straight or take the side road.
2) AP is no longer jerky when coming up to a stopped car at a red light. Before, AP would sometimes brake, lunge forward and then brake again. Now, it brakes smoothly way ahead of the stopped car.
3) I seem to be getting more intelligent "red hands on wheel" warning now when AP does lose confidence. For example, if I approach a road that is widening a bit where AP might wobble, AP handles it well, but if my hands are not on the wheel at the time, I get the "red hands on wheel" warning.

On the roads I drive, I feel like AP is close to "feature complete". If it could just handle traffic lights and turns at intersections, AP could handle my entire daily driving, with my supervision of course. As it is now, AP handles things very well. I only need to disengage AP at intersections if there is a red light with no lead car or if I need to make a turn or take an on ramp.

But I recognize that my roads may be a bit easier and there are owners who don't have such good experiences with AP on the roads they drive on. Obviously, the tricky part will be getting AP good enough for different conditions, not just the easier ones.

 

[SOULPEDL]

Yes, I've noticed some real improvements in AP as well. Auto Steer is super solid now. Also:
1) AP is no longer confused by a parallel side road and goes perfectly straight now. in the past, AP used to veer a little like it was not sure if it should go straight or take the side road.
2) AP is no longer jerky when coming up to a stopped car at a red light. Before, AP would sometimes brake, lunge forward and then brake again. Now, it brakes smoothly way ahead of the stopped car.
3) I seem to be getting more intelligent "red hands on wheel" warning now when AP does lose confidence. For example, if I approach a road that is widening a bit where AP might wobble, AP handles it well, but if my hands are not on the wheel at the time, I get the "red hands on wheel" warning.

On the roads I drive, I feel like AP is close to "feature complete". If it could just handle traffic lights and turns at intersections, AP could handle my entire daily driving, with my supervision of course. As it is now, AP handles things very well. I only need to disengage AP at intersections if there is a red light with no lead car or if I need to make a turn or take an on ramp.

But I recognize that my roads may be a bit easier and there are owners who don't have such good experiences with AP on the roads they drive on. Obviously, the tricky part will be getting AP good enough for different conditions, not just the easier ones.

Now that you mention it... ditto to all of those here as well.

FYI, I spend most of my time in the investor thread and see a clear connection between FSD and share price. As shareholder since 2015 with 90-100% of my retirement in TSLA, in full disclosure, I could be biased in my assessment. So I just had a conversation with myself regarding this... no change, we're good.

 

[Az_Rael]

Less wandering when picking a lane or if it widens/narrows. Smoother transitions.

Weird, in my commute, this has gotten worse in the latest updates not better. AP seems to lightly seesaw in the lane on a regular basis now where it didn't before. It also hugs the left lane line way more than it used to. Makes for some white knuckle moments when the guy in the fast lane is hugging his right lane line and AP is hugging the left lane line. This is all on straight well marked freeway.

 

[Mo City]

Weird, in my commute, this has gotten worse in the latest updates not better. AP seems to lightly seesaw in the lane on a regular basis now where it didn't before. It also hugs the left lane line way more than it used to. Makes for some white knuckle moments when the guy in the fast lane is hugging his right lane line and AP is hugging the left lane line. This is all on straight well marked freeway.

Lane positioning bothers me more than anything else about AP on the highway. It should sense walls (usually on the left) and other vehicles on both sides and adjust lane position accordingly instead of just centering based on the painted white lines.

Passing trucks is often white-knuckle time for me. All AP needs to do is adjust 2-4 inches to the left and all would be good.

 

[Thp3]

Lane positioning bothers me more than anything else about AP on the highway. It should sense walls (usually on the left) and other vehicles on both sides and adjust lane position accordingly instead of just centering based on the painted white lines.

Passing trucks is often white-knuckle time for me. All AP needs to do is adjust 2-4 inches to the left and all would be good.

I really think that is a limit on the 2.5 neural net. With a faster bigger computer the neural net will be able too incorporate more options in lane positioning. I think we simply have to wait for HW3 and a bigger neural net.

 

[AlanSubie4Life]

I really think that is a limit on the 2.5 neural net. With a faster bigger computer the neural net will be able too incorporate more options in lane positioning. I think we simply have to wait for HW3 and a bigger neural net.

I don’t understand what this would have to do with the neural net. The neural net already has identified all relevant objects for this. It seems extremely low computational overhead to change the driving behavior.

I think it is just due to being beta and performing such steering tasks smoothly...which seems to be a general issue...

 

[EVNow]

I really think that is a limit on the 2.5 neural net. With a faster bigger computer the neural net will be able too incorporate more options in lane positioning. I think we simply have to wait for HW3 and a bigger neural net.

No - this is a driving policy thing. Not NN.

They need to add heuristics like
- If there is a big vehicle or wall on one side (and nothing or small vehicle on the other) shift x inches to the other side
- If there is a cyclist or construction workers on one side shift x inches to the other side
- If the lane is expanding but contracting a hundred feet later (because of merging lanes, for eg), don't try to center. Just maintain same distance from the side that is not changing.

 

[M3BlueGeorgia]

A discussion in another thread got me thinking about what "feature complete" means. I know folks have different opinions on that. Some folks might have a list of 20 must-haves that they define as "feature complete" while someone else may have a list of 50 must-haves. But it occurred to me that another way to possibly think of "feature complete", and perhaps how Tesla is looking at it, is to think of Autopilot's Operational Design Domain (or ODD). ODD is essentially the driving domain that the self-driving car is designed to operate in. This does not mean that the self-driving car is perfect in its ODD. There may be edge cases in the ODD that the self-driving car still needs to solve. ODD just means that the car is designed to operate in that domain. For example, right now, Autopilot's ODD is highways and well marked roads. Intersections, city streets and roads with no lane markings are outside of Autopilot's ODD. So "feature complete" FSD would mean expanding Autopilot's ODD to include every domain (highways, city streets, intersections, parking lots and roads with no lane markings).

If you look at the FSD feature list on the website, it seems to confirm this way of looking at "feature complete":
- NOA on highways checks the highway domain
- Enhanced Summon checks the parking lot and probably the roads with no markings domain
- Automatic City Driving checks the city street and intersection domain
- Traffic Light and Stop Sign Recognition also checks the intersection domain

So all these FSD features, once complete, would cover the entire ODD needed for FSD. In other words, achieving "feature complete" means expanding AP's ODD to include every domain needed for FSD. And it is also worth noting that expanding AP's ODD to include every domain is a prerequisite for FSD. This is also consistent with what we know Tesla is working on. We know they are working on traffic light recognition, intersections, curbs and parking lots. That's because they are working to add intersections, parking lots and city streets to AP's ODD.

From a software development perspective, "Feature Complete" is when all the software development groups, teams and sub-groups claim they've implemented the specifications and have done some unit testing, and maybe a little system testing.

It usually signals the start of a development death march to turn a mass of semi-functional software into something that can be shipped.
The problems to be addressed are usually:
- Interoperability
- partially implemented features
- stuff missed out of the specification, therefore never implemented
- misunderstanding of requirements, therefore the wrong capability implemented
- extremely delicate software

Interestingly enough Tesla is using Agile development, and "feature complete" is something that Agile development eliminates.

Based on the (predictable) problems Tesla have experienced in just getting the narrow subsets of FSD to work properly, they are a long way from having full, proper FSD. Probably 2 to 3 years.
I do wish they'd focus on progressively rolling out advanced features and improving the ones they have released.

Note: The delayed rollout of Advanced Summon is an excellent example of the problems trying to get a narrow FSD subset working.
Also the inconsistent performance of NoA's lane change recommendations is another example of a narrow but critical part of FSD not working well.

 

[EVNow]

ODD just means that the car is designed to operate in that domain. For example, right now, Autopilot's ODD is highways and well marked roads. Intersections, city streets and roads with no lane markings are outside of Autopilot's ODD. So "feature complete" FSD would mean expanding Autopilot's ODD to include every domain (highways, city streets, intersections, parking lots and roads with no lane markings).

Yes, in Lex's interview Musk agreed with Lex when he said FSD is NOA on freeway + City NOA. They are clearly now working on city features (as Musk has talked about and we see features added to the prod code but not enabled).

But I hope, they have as the feature list to complete the NHTSA list. That will do 2 things
- Legitimately claim FSD complete (and recognize FSD revenue)
- Make them comparable to Waymo

Both will help make Tesla an unattractive target of tock price manipulation, which will reduce negative media attacks greatly.

 

[EVNow]

Interestingly enough Tesla is using Agile development, and "feature complete" is something that Agile development eliminates.

Since classic agile development is based on small teams with no external dependencies, all large companies implement agile in different ways. Also, there is a particular problem with FSD being a novel project with very high accuracy needs. So, I'm guessing Tesla has decided to have a FC milestone that completes all the features they are targeting - but with a lower quality standard than actually needed for FSD deployment.