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

[Daniel in SD]

Imgur

What do you guys think of this intersection?

Scroll down to see all the images. I think the upcoming feature update would drive right into the intersection and then stop in the intersection... what do you think?

It depends how HD their maps are. I guess they'd have to using cameras to localize on the map which I doubt they're doing yet.
I'm looking forward to seeing what types of intersections trip it up.

 

[Snuffysasa]

It depends how HD their maps are. I guess they'd have to using cameras to localize on the map which I doubt they're doing yet.
I'm looking forward to seeing what types of intersections trip it up.

I mean I think we are all assuming they will use vision for the location to stop.... and it looks like it only sees the stop line on the far side of the intersection and doesn't detect a stop line on the correct side.

 

[Cheburashka]

This is on 2020.12

This is why a maps first approach is not safe nor reliable.

I assume it's labelled as a traffic light in the map.

 

[Daniel in SD]

I mean I think we are all assuming they will use vision for the location to stop.... and it looks like it only sees the stop line on the far side of the intersection and doesn't detect a stop line on the correct side.

oops I didn't scroll down. Yeah it definitely looks like it will stop on the other side of the intersection. Edge cases!

 

[Mardak]

I mean I think we are all assuming they will use vision for the location to stop.... and it looks like it only sees the stop line on the far side of the intersection and doesn't detect a stop line on the correct side.

Definitely uses vision to figure out where the stop lines are, but it's not only looking for lines to determine where to stop. I've gotten a false red light warning when passing a parking lot entrance that the NN thought was the intersection. Similarly in a residential area with no stop signs, Autopilot shows a stop line at a T intersection when there's no lines.

So for this specific Imgur example, Autopilot will probably stop too late given the current visualization, but Tesla can have a trigger to detect when the driver stops earlier than the detected line and do some math and reverse time to generate training data with the correct stopping location.

 

[banned-66611]

Third Row Tesla just tweeted this image that appears to be a preview of automatically stopping for red lights. He claims it is coming in the next update.

Third Row Tesla Podcast on Twitter

Stopping for red lights and stop signs is in Early Access. Here is video:
The Tesla Show on Twitter

We can see in the video that the car will ask the driver to press the accelerator or flip the stalk to resume driving when the light turns green. So there will driver input initially.

Wow. There are going to be a lot of red light running crashes in the near future as this lulls drivers into a false sense of security.

 

[mongo]

"I did test 2020.12.1 with new map, never panic with AP on while trying to hit new traffic light recently developed. Tried it 10 times and the car can see the traffic light but never gave me alert about it, while it always gave me alerts about old stop sign and traffic lights"
Mike Alani on Twitter

Given the new functionality is going through soft release (staged activation) , I don't see what your point is.
Also, 2012.12.1.1 has been rolling out...
Twitter

No they are using maps PERIOD and in your own words if you use maps for ANYTHING other than routes its not general.

Now they use it for smart summon, pot holes, intersections, stop signs & traffic lights control, etc. Map usage will only increase as AP's operating domain increases.

"Use for anything" != required

Given your signature as a senior engineer, I'm really surprised that you apparently have never had to deal with a specification document. If you were familiar with such things, you would understand the difference between these two statements:
A map of the area shall (must) exist to enable the FSD feature.
This system may have a map of the area.
http://www.ietf.org/rfc/rfc2119.txt

Is Tesla using maps? Yes
Is the map helpful for non-navigation skills at this point? Sure is
Would a map with transitory details (potholes) be helpful in the future? Yes
Does that mean this data is required? No
Can it function without maps? Yes
Can a car that requires maps ever drive the inital run on a new road? No
Can a car that requires maps therefore ever be L5? No

Infact i can probably find you saying worse quotes than that if i do some digging, such as if its not general then its "worthless", "useless", etc.

Good luck with that, because I haven't. In the arena of driving, I have not used worthless at all, and only used useless to describe classic wiper control and limiting AP to the posted speed limit. Oh, right, and in reply quoting you Neural Networks

Feel free to use the site search to check. I did, took 30 seconds. Use other terms if you want, I'm happy to correct a previous misstatement.
Again, needs/ requires vs can use/ may/ extends functionality

 

[diplomat33]

Does that mean this data is required? No
Can it function without maps? Yes

Clearly it can't really function reliably without maps or Tesla would not bother with maps.

Can a car that requires maps therefore ever be L5? No

What?! Of course it can. Just give it the maps first and then it will be L5. For example, create a map of the entire US and it would be L5 in the US.

 

[willow_hiller]

I know we've generally agreed that autonomous vehicles need to be better drivers than the average human, but I still think it's a useful frame of mind to think first about how humans solve certain problems. Regarding traffic lights around bends or over hills, people are also incapable of seeing the lights until moments before entering the intersection.

If you know a light is coming up (human memory being the equivalent of a map), the best you can do is slow down and be prepared to stop (very similar to the current auto-stopping behavior, actually).

If you don't know a light is coming up, there are sometimes traffic signs that warn you of upcoming lights (which a camera may be able to read and act upon). But otherwise, there must be enough space for reasonable drivers to see the light and determine whether to stop. If a person can do it, a vision-only system must be able to do it as well. Even when I'm blinded by the sun, it's possible to use context clues to determine whether the light is green (are there other cars in the intersection? Is the oncoming direction stopped? Is the perpendicular direction stopped?).

 

[diplomat33]

I know we've generally agreed that autonomous vehicles need to be better drivers than the average human, but I still think it's a useful frame of mind to think first about how humans solve certain problems. Regarding traffic lights around bends or over hills, people are also incapable of seeing the lights until moments before entering the intersection.

If you know a light is coming up (human memory being the equivalent of a map), the best you can do is slow down and be prepared to stop (very similar to the current auto-stopping behavior, actually).

If you don't know a light is coming up, there are sometimes traffic signs that warn you of upcoming lights (which a camera may be able to read and act upon). But otherwise, there must be enough space for reasonable drivers to see the light and determine whether to stop. If a person can do it, a vision-only system must be able to do it as well. Even when I'm blinded by the sun, it's possible to use context clues to determine whether the light is green (are there other cars in the intersection? Is the oncoming direction stopped? Is the perpendicular direction stopped?).

You are right that humans use memory and context to compensate for situations when vision is not enough. And a lot of what you mention could probably be programmed into the autonomous car in the driving policy. But this is precisely where HD maps are really great. HD maps are like getting the answers to the test. Why not just give the car the "answers to the test"? In other words, just give the car a HD map that tells it in advance where stuff is that the vision can't see. So while safe driving policy is important and should absolutely be included, I feel like HD maps are a more reliable solution that trying to program "human intuition" into the car.

 

[willow_hiller]

You are right that humans use memory and context to compensate for situations when vision is not enough. And a lot of what you mention could probably be programmed into the autonomous car in the driving policy. But this is precisely where HD maps are really great. HD maps are like getting the answers to the test. Why not just give the car the "answers to the test"? In other words, just give the car a HD map that tells it in advance where stuff is that the vision can't see. So while safe driving policy is important and should absolutely be included, I feel like HD maps are a more reliable solution that trying to program "human intuition" into the car.

In speaking of safe driving policy, I wonder how far ahead the auto-stopping system is looking for vehicles. I've been in situations where the light is green, but there are cars from the perpendicular direction backed up into the intersection. It would be great if AP overrode driver input and stopped at the line anyway if it detects a fully blocked intersection.

 

[mongo]

Clearly it can't really function reliably without maps or Tesla would not bother with maps

The question is not, "what does it need for X level of performance now, while it is being trained?" it is "when trained, what does it need?".

Is it functional without maps (on an unmapped dirt road for instance).

What?! Of course it can. Just give it the maps first and then it will be L5. For example, create a map of the entire US and it would be L5 in the US.

You're skipping the line before:

Can a car that requires maps ever drive the inital run on a new road?

Regardless, you are still making the car L4, because it only works where there are maps.
Your L4 domain is mapped roads. L5 would be all roads.

You are right that humans use memory and context to compensate for situations when vision is not enough. And a lot of what you mention could probably be programmed into the autonomous car in the driving policy. But this is precisely where HD maps are really great. HD maps are like getting the answers to the test. Why not just give the car the "answers to the test"? In other words, just give the car a HD map that tells it in advance where stuff is that the vision can't see. So while safe driving policy is important and should absolutely be included, I feel like HD maps are a more reliable solution that trying to program "human intuition" into the car.

That only works if the test doesn't change. If you are relying on a crib sheet and the questions are different, you will fail. Hard. So your system needs to be able to determine the answers don't match the questions. If it can do that well enough to override the static map data, it doesn't need the static map data.

Maps are good for unsolvable warnings though, like where hidden driveway signs should be.

 

[diplomat33]

The question is not, "what does it need for X level of performance now, while it is being trained?" it is "when trained, what does it need?".

Is it functional without maps (on an unmapped dirt road for instance).

But is it functional without maps? Presumably, if it were functional without maps, Tesla would not use maps. But they are using maps.

Regardless, you are still making the car L4, because it only works where there are maps.
Your L4 domain is mapped roads. L5 would be all roads.

But if you pre-map all roads, then your domain would be all roads and then it would be L5. You just need to pre-map all roads first.

 

[mongo]

But is it functional without maps? Presumably, if it were functional without maps, Tesla would not use maps. But they are using maps.

Yes,it works without maps, but not at high enough reliability yet. People post their videos of Teslas APing down unmapped dirt roads.

But if you pre-map all roads, then your domain would be all roads and then it would be L5. You just need to pre-map all roads first.

Saying an car that requires maps is L5 presupposed complete mapping, now and forever. Which requires you instantaneously map all new construction. Else, those in new subdivisions cannot use AP.
It also presupposes instantaneous updates on road changes.

 

[diplomat33]

Yes,it works without maps, but not at high enough reliability yet. People post their videos of Teslas APing down unmapped dirt roads.

Thanks. But if the reliability is not high enough, then it is not good enough. It needs to work with high enough reliability in order to do autonomous driving without driver supervision.

Waymo cars also work without maps too. But Waymo does not let the cars drive autonomously without maps because it would not be reliable enough.

Saying an car that requires maps is L5 presupposed complete mapping, now and forever. Which requires you instantaneously map all new construction. Else, those in new subdivisions cannot use AP.
It also presupposes instantaneous updates on road changes.

So? There is a new method of building HD maps with commercial satellites. It should be doable to keep maps up to date. Plus, you can have your fleet update maps when they encounter something. So the first car that encounters a new construction zone could update a new map to the cloud for the entire fleet to download. With a large enough fleet, every car would be helping to keep the maps updated for the entire fleet.

 

[Bladerskb]

Given the new functionality is going through soft release (staged activation) , I don't see what your point is.
Also, 2012.12.1.1 has been rolling out...
Twitter

You keep changing the goal post. Stick to one thing. The traffic light/sign stop system DOES NOT work without a map PERIOD.

"Use for anything" != required

Now you are again changing the goal post and EVEN then the traffic light/sign stop system NEEDS the Map.
It doesn't work without it, ITS A REQUIREMENT. As pointed out above and you just dismissed.

Given your signature as a senior engineer, I'm really surprised that you apparently have never had to deal with a specification document. If you were familiar with such things, you would understand the difference between these two statements:
A map of the area shall (must) exist to enable the FSD feature.
This system may have a map of the area.
http://www.ietf.org/rfc/rfc2119.txt

The top is right though. You NEED one of the latest maps with the Traffic light/stop sign data in order to enable the traffic light/stop sign feature. Also if you reach an area that isn't mapped the feature WON'T work!

Is Tesla using maps? Yes

You previously said Tesla weren't using maps. Now all of a sudden its okay?

Is the map helpful for non-navigation skills at this point? Sure is
Would a map with transitory details (potholes) be helpful in the future? Yes

Again you and others argued for months that using maps for anything other than routes were a crutch..

Does that mean this data is required? No
Can it function without maps? Yes

Both smart summon feature and the traffic light / stop sign feature NEEDS the map to function. This has already been proven and supported by verygreen.

"Manual has comprehensive explanation of all the stopping stuff. And yes. It's all GPS based with just some vision assist."

"I read it as "Once we decided to stop based on map data, we use vision to determine where to stop looking for a stop line or an intersection"
green on Twitter

And here is what you said about that previously:

If you ignore your vision data because it doesn't match your HD map, then the map is the primary data source

Moreover this has already BEEN proven. You disregard it just like anything. This whole discussion is pointless because you keep refusing all facts and evidence.

A newly developed Traffic light that is NOT in the map is detected using vision but is ignored because its not in the map. So no control algorithm or warning is initiated.

"I did test 2020.12.1 with new map, never panic with AP on while trying to hit new traffic light recently developed. Tried it 10 times and the car can see the traffic light but never gave me alert about it, while it always gave me alerts about old stop sign and traffic lights"
Mike Alani on Twitter

Another Mapped traffic light that conflicts with vision
Victor on Twitter

 

[diplomat33]

Interesting edge case: the car phantom braked because it mistook red lights on the back of a truck as a bunch of traffic lights. Derek Teay on Twitter

 

[ramonneke]

I get that but the car should not stop at all if the light is green. It will cause you to get rear ended.

It will not stop, because the driver will slam the throttle.

 

[mongo]

Thanks. But if the reliability is not high enough, then it is not good enough. It needs to work with high enough reliability in order to do autonomous driving without driver supervision.

Thanks for what?
Tesla is still in development, so of course it is not as reliable as it will be.

Waymo cars also work without maps too. But Waymo does not let the cars drive autonomously without maps because it would not be reliable enough.

OK, but does Waymo have a plan / goal to ever not require maps?

So? There is a new method of building HD maps with commercial satellites. It should be doable to keep maps up to date. Plus, you can have your fleet update maps when they encounter something. So the first car that encounters a new construction zone could update a new map to the cloud for the entire fleet to download. With a large enough fleet, every car would be helping to keep the maps updated for the entire fleet.

If the first car can drive it without a map, then mapping is not required. If the map is required to be accurate, then that first drive in a construction zone can never happen.

Lame analogy:
A CPU can run without branch prediction
A CPU can run faster with branch prediction (maps as hints)
A CPU cannot run in a useful manner if the branch prediction is fixed (fixed map as primary source)

 

[ramonneke]

The risk is that AP needs to recognize all lights. If it doesn't recognize one as a light and it happens to be red, you just ran a red light.

Which..... would happen now too anyway so no difference there....