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

[diplomat33]

Actually that begs an interesting question. We keep talking about the race to L5, but can you call a vehicle L5 until it can drive on any public road? Including roads outside of the US?

Several companies are testing autonomous vehicles in single cities right now, and they might eventually run in several states, but who is going to be the first to achieve international L5? As far as I know, Tesla is the only company operating vehicles with autonomous features across multiple countries simultaneously.

The SAE does allow one exception to the L5 ODD where the ODD can be limited to a country and still be L5. So you can do FSD that works everywhere on US roads, and it would be L5.

Honestly, international L5 seems overly ambitious. Autonomous driving is hard enough. Let's achieve safe, reliable autonomous driving that works somewhere before we talk about achieving it everywhere in the world.

Also, international L5 seems a bit unnecessary. Just achieving safe L5 in the US would be enough to fundamentally transform transportation. And we can't actually drive anywhere in the world because of oceans. Sure, it would be the ultimate milestone to be able to sell autonomous cars anywhere. But driving in different countries is so different that you would need to tailor your autonomous driving to the specifics of each country. I doubt there will be a "one size fits all" FSD that works in every country.

 

[Knightshade]

Tesla would never release a level 3 feature. It's either level 5 or nothing, based on their approach. If it's good enough for level 3, then it's good enough for level 5. Based on my understanding.

That's definitely not correct.

L3 the car is explicitly expecting a human driver to be available on shortish notice (maybe 30 or 60 seconds) to take over, and is only expected to operate in certain places or types of roads.

The driver need not be actively monitoring all road conditions, but needs to be present and awake in the drivers seat to take over if the car exceeds its operational capabilities or domain.

L5 you don't even have to have a human in the car and is expected to work everywhere.


They're pretty widely far apart.

 

[willow_hiller]

Also, international L5 seems a bit unnecessary. Just achieving safe L5 in the US would be enough to fundamentally transform transportation.

Try telling that to all of the Canadians, Brits, and Europeans on this forum

It will be interesting to see how quickly Tesla releases Traffic Control for Canada and Europe. If they come within a few months of the US release of Traffic Control, I think it will show that Tesla has the ability to adapt their FSD to international situations relatively quickly.

 

[diplomat33]

Try telling that to all of the Canadians, Brits, and Europeans on this forum

It will be interesting to see how quickly Tesla releases Traffic Control for Canada and Europe. If they come within a few months of the US release of Traffic Control, I think it will show that Tesla has the ability to adapt their FSD to international situations relatively quickly.

I used US as an example because I live in the US. But I want FSD to come to Canada, England, Europeans, Chinese, etc... So Canada can get FSD that just works in Canada, Europe can get FSD that just works in Europe etc... but it will be a different FSD.

I am just saying that the FSD will probably need to be tailored for each country because each country is different. For example, in England, they travel on the opposite side of the road to American driving. So clearly a FSD that works in the US would need to be changed to work in England. So I don't think you can really do a "one size fits all" FSD that works in every country.

But maybe I am confused by your post. Are you asking for a single FSD that works everywhere in the world or merely asking when someone will deliver different FSD that works internationally?

 

[Knightshade]

I am just saying that the FSD will probably need to be tailored for each country because each country is different. For example, in England, they travel on the opposite side of the road to American driving. So clearly a FSD that works in the US would need to be changed to work in England. So I don't think you can really do a "one size fits all" FSD that works in every country.

Why?

The car knows where it is after all.

We already have FSD that behaves differently based on country (right now it doesn't stop for lights outside the US- but does inside the US.. and I believe the EU has some NOA stuff nerfed too right now).

No reason the car can't be smart enough to know it needs to change the sign recognition it's using for example if it gets moved to another country with different signs.


It's already going to need to do this- changing its driving rules based on location- even inside the US

Different states have different rules on making a right on red. In some cases it's different by city within a state.

A self driving car will need to know that and adjust behavior accordingly based on location (and time and day potentially- FSD in NYC will need to understand alternate side of the street parking for example)

 

[diplomat33]

Why?

The car knows where it is after all.

We already have FSD that behaves differently based on country (right now it doesn't stop for lights outside the US- but does inside the US.. and I believe the EU has some NOA stuff nerfed too right now).

No reason the car can't be smart enough to know it needs to change the sign recognition it's using for example if it gets moved to another country with different signs.


It's already going to need to do this- changing its driving rules based on location- even inside the US

Different states have different rules on making a right on red. In some cases it's different by city within a state.

A self driving car will need to know that and adjust behavior accordingly based on location (and time and day potentially- FSD in NYC will need to understand alternate side of the street parking for example)

Well yeah, but that's kinda the point. I am not saying it can't be done. I am just saying it will require different rules based on location. It's hard enough to get FSD right for one location. Doing international L5 will be exponentially harder.

@willow_hiller asked about international L5 and I am just trying to say that I think that might be premature. We don't even have reliable L4 yet so how about we focus on getting FSD right for one location and then for one country before we talk about doing L5 that works everywhere on the planet.

 

[willow_hiller]

@willow_hiller asked about international L5 and I am just trying to say that I think that might be premature

I am talking about a single provider of FSD. Not necessarily a single codebase or single neural network. But maybe a single codebase with an extensible neural network.

I don't think Waymo or Zoox necessarily have incentives to develop FSD for different countries. In the absence of Tesla, who would build a Canadian FSD? Or a British FSD?

 

[diplomat33]

I am talking about a single provider of FSD. Not necessarily a single codebase or single neural network. But maybe a single codebase with an extensible neural network.

I don't think Waymo or Zoox necessarily have incentives to develop FSD for different countries. In the absence of Tesla, who would build a Canadian FSD? Or a British FSD?

Thanks for the clarification.

Well, I am sure there are Canadian or British companies working on FSD too.

But I grant you that Tesla definitely has the potential of providing FSD internationally. Of course, Tesla needs to achieve FSD first. The benefit of companies like Waymo or Zoox is that they actually have FSD, even if it is limited to a location.

 

[daniel]

[...] And we can't actually drive anywhere in the world because of oceans. [...]

Oceans are not that much of an impediment. My car crossed more than 2,500 miles of ocean to get here. But I agree with your point: Not only the rules of the road, but also driving behavior, are very different in different places. A multi-country FSD would need a larger knowledge base of rules and customs.

[...]
Different states have different rules on making a right on red. In some cases it's different by city within a state.[...]

True. Sometimes turning right on red is generally allowed, but prohibited at specific intersections. The car would need to be able to read the signs. I think in some places it's prohibited at certain intersections at certain times of day.

 

[willow_hiller]

In speaking of EU FSD, it seems like Tesla has started retrofitting HW3 in the EU now: FSD computer upgrades for Model 3 are now available to be scheduled in Europe : teslamotors

They must be close to releasing some FSD features across the pond.

 

[linux-works]

I don't think Waymo or Zoox necessarily have incentives to develop FSD for different countries. In the absence of Tesla, who would build a Canadian FSD? Or a British FSD?

mobile eye.

I would bet on them, and they are not a 'taxi company' like others always like to argue back, about. waymo and cruise are, but M.E. is looking to be the single software source for the industry and they might actually be able to pull it off.

I know tesla is angry with that company, but that actually makes me more confident in M.E.

 

[diplomat33]

Speaking of Mobileye, they just posted a a few clips from one of their recent demos to illustrate some behaviors that their FSD can do.

The clips are informative because I think they illustrate a few key sub features of what I assume "autosteer on city streets" will need to do. So they illustrate a few features that Tesla is presumably working on as part of FSD "feature complete":

Unprotected left turn:

Navigating roundabouts:

Right turn onto narrow street:

A few observations:

This demo was done with cameras only (no lidar) and HD maps. I figure Mobileye probably uses the HD maps to help with localization. It should be noted that Mobileye intends to add lidar to the final product in order to increase the accuracy and reliability. While cameras-only can do a cool demo, Mobileye believes lidar is required to achieve the necessary 9's to actually deploy a driverless system to the public.

In Tesla's case, they will need accurate localization, 3D mapping of the road and intersections and detection of objects with just camera vision as well as very good driving policy to decide when it is safe to make the turn. So it is understandable why Tesla has not released "autosteer on city streets" yet. I think these clips do highlight why it is such a complex feature to do.

 

[mikes_fsd]

So, let's say you're approaching a green light, and the car starts to slow down, so you tap the accelerator to tell it the light is green, but then the light turns yellow while you're still far enough away to stop safely: Does the car recognize the yellow and initiate a stop again, or at this point do you have to completely take over and stop the car?

The car will start slowing down and come to a stop. Happened to me this weekend.

 

[Soda Popinski]

In Tesla's case, they will need accurate localization, 3D mapping of the road and intersections and detection of objects with just camera vision as well as very good driving policy to decide when it is safe to make the turn. So it is understandable why Tesla has not released "autosteer on city streets" yet. I think these clips do highlight why it is such a complex feature to do.

To be clear, MobilEye hasn't released any FSD either. This is all still prototype software, and extremely impressive and encouraging. Unfortunately for us curious about Tesla's FSD development, they haven't been particularly forthcoming on their progress or lack there of. We saw "Autosteer on city streets" back in Autonomy Day, so we know they've done some work on it.

 

[diplomat33]

Here is a good review of the latest AP/NOA in the UK:

At 14:45 mark, he gets into a really tight spot where he is sandwiched between traffic cones on the right and a semi truck on the left. Yikes.

At the start of the video, he drives manually through some tricky roundabouts. I think it is a good example of why Waymo and others use lidar and HD maps for complex city driving because you definitely need accurate and reliable localization and perception as well as the contextual information to help the car know what lane to be in and what to do. It will be interesting to see how Tesla handles this problem.

 

[daniel]

EAP (HW 2.5) now shows traffic cones on the screen (often when there's something else there entirely, and occasionally when there's nothing there at all) but I still don't have the courage to leave autosteer engaged when there are traffic cones.

I have the utmost respect for the courageous fools people who let autosteer (whether EAP or FSD) drive under conditions when there might be conflicting signals, such as traffic cones superseding painted lane lines.

 

[Webeevdrivers]

I wonder what hurdles (and with who) Tesla has to cross in getting the stop at stop lights activation in Canada. Does it need to be tested by MOT in Canada or is it held up on the Tesla side.

Will it be a separate software update or some magic keystroke on Tesla’s part to activate it?

 

[OPRCE]

In speaking of EU FSD, it seems like Tesla has started retrofitting HW3 in the EU now: FSD computer upgrades for Model 3 are now available to be scheduled in Europe : teslamotors

They must be close to releasing some FSD features across the pond.

Yes, I got my retrofit to HW3 on May 6th -- only differences so far are AP is noticeably smoother in acceleration/braking on motorway, IC visualisation of lights, cones & road-markings -- but stopping at lights is supposed (per Elon Tweet) to be activated in Q3.2020.

Phantom-braking for gantries still happens, possibly more often than on recent versions with HW2.5, but it ramps in softer, not the instantaneous snap to significant braking we know & love for its characteristic squirts of adrenaline out of a blue sky.

 

[diplomat33]

The upcoming 2020.20.5 update has improvements to GPS tracking:

Considering that Tesla uses GPS a lot for localization, I wonder if this improvement to GPS is related to upcoming FSD features? It would stand to reason that FSD features like the current "Traffic Light Control", and upcoming "autosteer on city streets" might benefit from better GPS localization.

 

[diplomat33]

Tesla has freeway NOA that works fairly well. Waymo doesn't work in my city.

I bet we'll get City NOA on Tesla before Waymo works in my city.

It is wrong to say that waymo does not work in your city. You have not tried a waymo to see if it works or not. You should say waymo is not deployed in your city yet. That is accurate. Deployment and working are different things. There are likely areas thar work but waymo has not deployed public ride hailing yet. Waymo could work in Seattle and just not deployed yet.