Future of FSD after "traffic light & stop sign response" release

[Saghost]

I think that is still the goal.

In fact, the answer may have been under our noses the whole time:

I think that the FSD demo is what we can expect the MVP to be.

So ask yourself: what part of this FSD demo, does Tesla still need to deliver to the customer? Which is why I think "autosteer on city streets" refers to auto lane changes and making turns at intersections to follow a route. And yes, I still think, based on the demo, that we will get "City NOA". That is the last big, obvious, piece of "feature complete" needed to achieve that stated MVP. But it might come in stages. "autosteer on city streets" may be part 1 and "city NOA" might be part 2.

And we should remember that MVP is the minimum. It will probably not be able to handle downtown NYC in rush hour traffic without many disengagements:

I hope Tesla does not stop at MVP but continues to work to try to solve more complex driving cases like what we see in the NYC video.

There is the possibility though that the current hardware is not good enough for more complex FSD in which case, Tesla may declare FSD "good enough" once they achieve "feature complete" with no disengagements. It is also possible that, just like with the introduction of AP2, that Tesla at some point introduces another big sensor upgrade with more cameras or more radars etc... to give us FSD 2.0.

I don’t really see more cameras or more radars as particularly helpful for the NYC challenge.

Although there are a lot of things to worry about there, they all develop much slower than on the freeway.

I feel like most of the NYC challenge is figuring out how to respond to everyone else breaking the traffic laws around you without just sitting and waiting - less about environmental perception and more about allowable responses.

 

[powertoold]

FSD shouldn't be prioritized for super urban environments. NYC actually has a super low traffic fatality per capita since cars are moving slowly.

NYC (pop ~8 mil) traffic fatalities: ~200 ( Traffic Deaths in New York City Drop to 200, a Record Low )
Bay area (pop ~8 mil) fatalities: ~400 ( Fatalities from Crashes | Vital Signs )

 

[diplomat33]

FSD shouldn't be prioritized for super urban environments. NYC actually has a super low traffic fatality per capita since cars are moving slowly.

NYC (pop ~8 mil) traffic fatalities: ~200 ( Traffic Deaths in New York City Drop to 200, a Record Low )
Bay area (pop ~8 mil) fatalities: ~400 ( Fatalities from Crashes | Vital Signs )

The fatality rate might be lower yes, but they are the most complex and most challenging environments for autonomous driving. Once you solve FSD in super urban environments, solving FSD on the highway or in light urban areas is trivial. And if you want to be able to offer robotaxis in those environments, then you need to solve FSD in those environments regardless. That is why FSD companies like Waymo and Cruise are focused so heavily on developing their autonomous driving in those environments.

 

[powertoold]

Once you solve FSD in super urban environments, solving FSD on the highway or in light urban areas is trivial.

But as someone pointed out above, driving in super urban environments is less about driving and more about knowing which rules to break. I agree this would be helpful in some situations. However, the acceptable rules to break in NYC are likely different than in the Bay Area for example. FSD should be targeted towards the 95% in terms of land area, not the 5%.

In terms of financial reward and investment, this is also true.

 

[ItsNotAboutTheMoney]

The fatality rate might be lower yes, but they are the most complex and most challenging environments for autonomous driving. Once you solve FSD in super urban environments, solving FSD on the highway or in light urban areas is trivial. And if you want to be able to offer robotaxis in those environments, then you need to solve FSD in those environments regardless. That is why FSD companies like Waymo and Cruise are focused so heavily on developing their autonomous driving in those environments.

Forget autonomy, Tesla's approach is built on being able to sell it to individual consumers. There is value in reducing how much driving sucks. So, making it work well in challenging environments will increase sales of FSD.

 

[Knightshade]

We know that internally, at least as of the January 2020 earnings call, Tesla's Minimum Viable Product for FSD was autonomy that could travel from your home to your work without disengagements some of the time. That may have changed in the last 3 months, but I don't think Elon would give up on that goal that quickly.

Mind you- that "no disengagement at least SOME of the time" goal which is L3 at best and maybe never better than L2 was already walked back- twice- from previously stated goals...

Just a year earlier in January 2019 the stated goal was still "full self driving in almost all circumstances" with "no action required by the person in the drivers seat"- which is at least L4

And THAT was walked back from the previous 2016 promise that you could just summon the car without even having a human in it from the other side of the country....Which is unquestionably L5 (a goal he said they think they could reach in 2 years... which would be 2018- though to be fair he suggested that date might be "slightly optimistic")

 

[diplomat33]

Mind you- that "no disengagement at least SOME of the time" goal which is L3 at best and maybe never better than L2 was already walked back- twice- from previously stated goals...

Just a year earlier in January 2019 the stated goal was still "full self driving in almost all circumstances" with "no action required by the person in the drivers seat"- which is at least L4

And THAT was walked back from the previous 2016 promise that you could just summon the car without even having a human in it from the other side of the country....Which is unquestionably L5 (a goal he said they think they could reach in 2 years... which would be 2018- though to be fair he suggested that date might be "slightly optimistic")

In my personal opinion, I feel like the current "feature complete" FSD will be L2, not just because of disengagements, but because the driver will still need to actively monitor the environment for situations that the car cannot handle.

 

[Cheburashka]

From Tesla's website:

Autosteer: Assists in steering within a clearly marked lane, and uses traffic-aware cruise control

 

[diplomat33]

But as someone pointed out above, driving in super urban environments is less about driving and more about knowing which rules to break. I agree this would be helpful in some situations. However, the acceptable rules to break in NYC are likely different than in the Bay Area for example.

No, I don't believe it is just about knowing rules. You still need excellent and accurate perception. And in super urban environments, the perception tasks are more demanding. There are more objects to track and you need greater precision. So perception is harder in super urban environments. Of course, once you solve perception, you are right that there is a lot more planning and driving policy that also goes into driving in super urban environments. So it is both perception and driving policy that makes autonomous driving in super urban environment far more difficult than in other environments.

FSD should be targeted towards the 95% in terms of land area, not the 5%.

In terms of financial reward and investment, this is also true.

I disagree. FSD should focus on people, not land area. If you want to make money, you need to sell robotaxis to people or offer rides to people. Where are the highest concentrations of people that will need a robotaxi? Big cities. So it makes sense from a business perspective to focus on FSD in big cities.

 

[tinkererguy]

Seems my recent close look videos might be of interest to some folks reading along, in 4K 60fps. It iincludes a segment where a yellow signal light turns red as I approach, which is always an interesting "can I safely stop in time" calculation for humans or cars to make.

Long version:
Close 4K look at Tesla "Traffic Light and Stop Sign Control (Beta)" used safely on public roads

Short version (yellow turning red, stop sign):

 

[diplomat33]

I feel like most of the NYC challenge is figuring out how to respond to everyone else breaking the traffic laws around you without just sitting and waiting - less about environmental perception and more about allowable responses.

I disagree. Perception is also harder too compared to highway driving or light urban driving. There will be more cars, trucks, cyclists, motorcyclists, pedestrians etc to track in close proximity to your vehicle and their paths might be unpredictable. And you need greater accuracy because sometimes there is not a lot of margin for error. For example, you won't have a lot of margin for error when maneuvering in a narrow street where a delivery truck is partially blocking your path. So perception is also a lot harder.

But yes, driving policy will also be very challenging as well. Both perception and driving policy are more challenging in super urban environments.

 

[Knightshade]

In my personal opinion, I feel like the current "feature complete" FSD will be L2, not just because of disengagements, but because the driver will still need to actively monitor the environment for situations that the car cannot handle.

that's certainly how it's increasingly looking with Tesla walking back promised abilities... but it's awfully damn far from what the 2016 through early 2019 folks were explicitly promised when they spent their $ on it.

At the time of the first big walk-back of promises (Feb/March 2019) I suggested it was happening in part to limit the # of owners they'd eventually owe FSD refunds to having promised, but been unable to deliver, L4/L5 driving to- that prediction continues to look pretty solid so far.

 

[Saghost]

No, I don't believe it is just about knowing rules. You still need excellent and accurate perception. And in super urban environments, the perception tasks are more demanding. There are more objects to track and you need greater precision. So perception is harder in super urban environments. Of course, once you solve perception, you are right that there is a lot more planning and driving policy that also goes into driving in super urban environments. So it is both perception and driving policy that makes autonomous driving in super urban environment far more difficult than in other environments.



I disagree. FSD should focus on people, not land area. If you want to make money, you need to sell robotaxis to people or offer rides to people. Where are the highest concentrations of people that will need a robotaxi? Big cities. So it makes sense from a business perspective to focus on FSD in big cities.

I don’t think I agree that perception is harder in the superurban. The requirements are different, yes. More things to track, from more angles. Possibly more precision.

But it all comes at you much slower, so a lot more frames and processor cycles to work with, even on a per object basis.

Between that and the rules differences, it is almost a different problem altogether.

What I was expecting Tesla to do is finalize highway driving, then expand into rural, suburban, and finally urban driving.

This release is clearly a big step into rural and suburban driving, but they haven’t given us level 3 freeway driving yet.

 

[diplomat33]

Here's a pretty impressive FSD drive in Las Vegas by Zoox. It definitely gives you a sense of some of the city driving problems that Tesla will need to solve in order to do "city FSD":

 

[mikes_fsd]

Here's a pretty impressive FSD drive in Las Vegas by Zoox. It definitely gives you a sense of some of the city driving problems that Tesla will need to solve in order to do "city FSD":

Ahh, yes the same company that stole IP from Tesla...
Maybe what you're seeing is only a sliver of what Tesla has developed in house, after all Tesla did accuse them of stealing Autopilot related code as well.

 

[diplomat33]

Ahh, yes the same company that stole IP from Tesla...
Maybe what you're seeing is only a sliver of what Tesla has developed in house, after all Tesla did accuse them of stealing Autopilot related code as well.

Zoox did not steal AP code. Tesla alleged that Zoox stole information and trade secrets related to inventory and logistics. Nothing to do with AP.

"Tesla lawyers filed a lawsuit in March last year against four former employees and Zoox, alleging the employees stole proprietary information and trade secrets for developing warehousing, logistics and inventory control operations."
U.S. self-driving car startup Zoox agrees to settle lawsuit with Tesla

 

[Knightshade]

Yeah, it's XPeng that is using the stolen AP source code- totally different company-

Former Tesla employee admits uploading Autopilot source code to his iCloud

 

[diplomat33]

In the Q1 earnings statement that was just released, Tesla confirms what we expected, that "traffic light response" will remove the confirmation on green once more data is collected:

"Autopilot & Full Self Driving(FSD)
We enabled stop sign and traffic light recognition and braking for our Early Access Program users at the end of Q1 and to the wider public in April 2020. As we have done with previous releases of major new features, drivers will be required to confirm their attention in order to continue. Once enough real-world data is collected, the system will become more capable and our vehicles will continue driving through intersections without a confirmation"
https://ir.tesla.com/static-files/c1723af4-ffda-4881-ae12-b6f3c972b795

 

[mikes_fsd]

Did people actually expect the confirmation to be a standard feature going forward?
I mean, I am glad they added that comment, but I don't see why people would be excited about it...

Oh and my bad on XPeng vs Zoox AP code, Zoox just stole supply chain IP.

 

[Knightshade]

Did people actually expect the confirmation to be a standard feature going forward?
I mean, I am glad they added that comment, but I don't see why people would be excited about it...

Ditto.


Would've been more interested in them commenting on the apparent dialing back expectations on the one remaining listed as undelivered FSD feature.