Autonomous Car Progress

[S4WRXTTCS]

The geofencing technical capability is already there I'm sure, defining all the roads and conditions is what I think would be insanely complex and then whatever implications associated with juggling multiple different branches of the software across all the different configurations etc -- I don't believe these are challenges other Robotaxi companies face.

Not sure how they would go about defining which municipal roads the system won't function on across an entire country and then maintain it. Unless we're talking going right down to the scale of current Robotaxi competitors and working on individual cities. But trying to do something Waymo-style across the US would be such a ridiculously massive task.

Is it even possible to strictly define an ODD without something like HD maps? Are the TomTom maps accurate enough to do Robotaxi-style geofencing while FSD draws everything in real time? Are the maps current enough?

If we look at how companies are positioned I'd say that GM is in the best position of selling a consumer level 4 systems which can work on both the interstate with properly mapped roads, and can work in cities that Cruise Fleet vehicles are operating in.

At this point they're not there yet on either front, but they're in a good position on both fronts to start bringing the two together.

What I'd like to see is Consumer L4 systems having gradually expanding geofenced roads along heavy traveled interstate roads. So every 6 months or so I'd see an increase in coverage area for the system I had.

The L4 system would have HD maps that defined where it could be. The Maps themselves would consist of maps that regularly got updated, but each car in the fleet would look for discrepancies between the HD maps and reality.

I expect this type of system to only work in areas with connectivity, and in areas that were economically viable. I'm not sure it makes much sense to have an L4 system out in the middle of nowhere that no one went to.

 

[stopcrazypp]

I question whether Lex Friedman can act impartially to Elon, and whether Lex is really aware of why the L4/L5 question is so important.

To me it seems like Tesla is using the promise of L5 like capabilities while knowingly not having the HW to achieve that. This allows them to buy time.

As a customer the answer to this question is vital.

If Elon acknowledges that the whole L5 thing was a fantasy, and is shifting focus to a geofencing L4 where possible it makes it seem realistic that I might get something.

If Elon doubles down on L5 and laughs off L3/L4 then I more likely to bail, and get something else that satisfices my wants outside of autonomy.

I can't blame any company in the US for failing to deliver on autonomous vehicles. As a nation we simply didn't do a good job of standardizing on technologies to help automate the driving task.

Like yesterday I was coming back from the coast, and a sign said the road was closed 26 miles ahead. The Tesla navigation didn't show anything so I tried Apple maps, and it desperately wanted me to turn around and go a different way. I then tried Google Maps, and it said there was an accident ahead but simply said there was a delay. I'm not sure why I didn't turn around. It was probably curiosity as to what the result would be, and the result was that I was an idiot. It seemed like they were allowing one lane of traffic to go through as it we'd move a bit, and then stop. But, then when I got further ahead about an hour later they were just turning cars around.

I bring that up because it shows how unready we are. I'm not sure how Apple Maps managed to get it right, and why Google Failed. I believe Tesla uses google for traffic information so I don't expect it to be right when Google isn't.

Navigation is comes before anything else.

Note these sorts of macro trip planning decisions are not part of self driving levels at all if you follow the SAE standard. As long as the car knows how to handle the turn around when it reaches it, it'll still qualify for L5. Being smarter in overall trip planning is not required.

 

[stopcrazypp]

The geofencing technical capability is already there I'm sure, defining all the roads and conditions is what I think would be insanely complex and then whatever implications associated with juggling multiple different branches of the software across all the different configurations etc -- I don't believe these are challenges other Robotaxi companies face.

Not sure how they would go about defining which municipal roads the system won't function on across an entire country and then maintain it. Unless we're talking going right down to the scale of current Robotaxi competitors and working on individual cities. But trying to do something Waymo-style across the US would be such a ridiculously massive task.

Is it even possible to strictly define an ODD without something like HD maps? Are the TomTom maps accurate enough to do Robotaxi-style geofencing while FSD draws everything in real time? Are the maps current enough?

Tesla already has geofencing in use in determining where to let you enable autosteer (there was a big deal made of this from the fatal accident in a residential neighborhood). But for FSD Beta it's not in use, as the idea was it should work in all public road types.

 

[S4WRXTTCS]

Note these sorts of macro trip planning decisions are not part of self driving levels at all if you follow the SAE standard. As long as the car knows how to handle the turn around when it reaches it, it'll still qualify for L5. Being smarter in overall trip planning is not required.

Yeah, its outside the scope of the SAE Levels.

Lots of the aspects of Autonomous driving are outside the scope of those levels.

My expectation is that Levels won't be mentioned as much as we get closer to actual autonomous driving. At some point I expect L3, and L5 to go by the way side.

L2 will be laughed at as:

"Remember the days when robots drove cars, but the human took all the responsibility?"
"Yeah, that was crazy. What were they smoking?"
"I heard it was pot. It was all the rage back then"

 

[Knightshade]

I agree. Except Elon has not talked about achieving "L2/ADAS", he's only talked about Tesla achieving L5.

...what?

Elon has talked about advancements where supervision is still required all the time (L2) many times. It's one of the steps he oft discussed in his "what is feature complete" explanations for example.

feature-complete, I mean, it’s the car able to drive from one’s house to work, most likely without interventions. So it will still be supervised, but it will be able to drive —

and

So feature-complete means it’s most likely able to do that without intervention, without human intervention, but it would still be supervised. And I’ve gone through this timeline before several times, but it is often misconstrued that there’s three major levels to autonomy. There’s the car being able to be autonomous, but requiring supervision and intervention at times. That’s feature complete. Then it doesn’t mean like every scenario, everywhere on earth, including every corner case, it just means most of the time.

 

[diplomat33]

...what?

Elon has talked about advancements where supervision is still required all the time (L2) many times. It's one of the steps he oft discussed in his "what is feature complete" explanations for example.

and

Elon also explained that supervision is temporary because FSD is still "Beta". Once FSD is good enough, they will remove supervision. The goal is still L5.

 

[diplomat33]

I like how Diess is subtly calling out Mobileye's bullsugar. Diess knows Tesla's approach is way ahead and superior, but Volkswagen has no choice:

Diess is doing nothing of the sort. He is simply asking a question about how computer vision can be improved with data and ML training, which Mobileye is already doing. If you continue watching Shashua's response, he explains that there are two parts to vision AI. There is the pattern recognition part, like classifying a dog, a car or a cyclist. That part can be improved with data and training. The second part of AI is understanding what a scene means, like understanding what lane goes with what traffic light or understanding a complex intersection. Yes, you can do that part with vision training as well but there are a lot of edge cases. Mobileye understands that HD maps can really help vision with that scene understanding part of AI. Maps basically augment vision, they don't replace vision. Shashua calls it "cloud knowledge" because you can collect valuable map data and share it with the entire fleet, making all your cars drive better. It would be silly not to take advantage of the benefits of HD maps to make your FSD more reliable. That's why Mobileye has vision-only FSD but augments it with HD maps, crowdsourced from their fleet.

You really need to take off your thick Tesla fan rose colored glasses.

 

[powertoold]

Check out FSD 10.9 respond to a hand signal. Game over for the competition. Tesla should hire the person narrating those 10 second Waymo highlights:

 

[diplomat33]

Check out FSD 10.9 respond to a hand signal. Game over for the competition. Tesla should hire the person narrating those 10 second Waymo highlights:

Stop with the "game over" nonsense. Many AVs (Waymo, Cruise etc) respond to hand signals. Tesla is not the first to do hand signals.

 

[S4WRXTTCS]

Stop with the "game over" nonsense. Many AVs (Waymo, Cruise etc) respond to hand signals. Tesla is not the first to do hand signals.

I'm not sure it did respond to the hand signal.

There are three possibilities.

It did respond the to hand signal, but isn't showing it. This is the impression the video gives.
It didn't pause for the pedestrian and was just doing what it would have done if the pedestrian has kept walking in the direction he was.
It simply lost the pedestrian and there is evidence of this as the person disappears from the display.

I don't really Musk ever mentioning hand signals. Knowing his humor he'd have it do something like fart on hand signal.

 

[powertoold]

I'm not sure it did respond to the hand signal.

There are three possibilities.

It did respond the to hand signal, but isn't showing it. This is the impression the video gives.
It didn't pause for the pedestrian and was just doing what it would have done if the pedestrian has kept walking in the direction he was.
It simply lost the pedestrian and there is evidence of this as the person disappears from the display.

I don't really Musk ever mentioning hand signals. Knowing his humor he'd have it do something like fart on hand signal.

It's more likely he just lightly pressed on the bottom of the pedal.

I just pointed that video out to show how blindly we can trust bullsugar from Waymo / Cruise / Mobileye. We have no idea what sorts of buttons they're pushing behind the scenes.

 

[powertoold]

I don't really Musk ever mentioning hand signals.

Elon has mentioned everything wrt FSD. You name it. UFOs landing. Fleet hack driving to Rhode Island. Potholes. James Bond.

https://twitter.com/x/status/1414435086548733952

 

[stopcrazypp]

It's more likely he just lightly pressed on the bottom of the pedal.

I just pointed that video out to show how blindly we can trust bullsugar from Waymo / Cruise / Mobileye. We have no idea what sorts of buttons they're pushing behind the scenes.

Waymo for their part have said they don't have direct remote control their vehicles (as in direct steering/accelerator/brake input) and all they can give are macro instructions which the car has to figure out how to execute. This has resulted in some embarrassing incidents, but so far there is no sign that they have any mechanism to do tricks (although there are some crazy conspiracy theories on the web with no supporting evidence).

Cruise is just beginning to operate without safety drivers and they have declined to say if they have remote control abilities when they were asked, so it's unknown at the moment. Although personally I have seen them operating locally for long enough (and seen the improvement in how the cars react throughout the years), that I don't have reason to doubt the maturity of their system.

Mobileye seems to be in the early days of their L4 system and when it launches it will have safety drivers (and presumably will for a while), so I don't think it'll really matter yet.

 

[powertoold]

Waymo for their part have said they don't have direct remote control their vehicles (as in direct steering/accelerator/brake input) and all they can give are macro instructions which the car has to figure out how to execute.

Yup, they can still press a button that tells the car to proceed or not when it's unsure. Currently, Waymos are always under direct remote supervision, afaik.

Cruise "declined to say if they have remote control abilities when they were asked, so it's unknown at the moment."

Cruise definitely has extensive remote capabilities. I've posted a video in the past where Voigt mentioned them.

 

[EVNow]

I just pointed that video out to show how blindly we can trust bullsugar from Waymo / Cruise / Mobileye. We have no idea what sorts of buttons they're pushing behind the scenes.

I've been thinking about this.

Just see all the videos Mars puts out. 1x, raw videos. Only thing he can be doing without telling us is pressing accelerator.

But, he can drive an hour without disengagement in San Francisco. Some of the videos are when the roads are quite crowded too. If FSD Beta can be used by an amateur to drive for an hour without disengagement, just imagine what billion dollar companies can do. We should be careful about Waymo/Cruise etc putting out "raw" videos - those videos may not tell the whole story - just as Mars 1 hour no disengagement video doesn't tell the whole story either. Neither of those videos are "fake" or fraudulent in anyway - just that there is so much more to FSD than an hour of raw video in one or two places.

PS : Only way to assess the state of FSD of a company is to have it be available to a lot of people who can experience for themselves what it is like. They should be able to take and put out videos without NDAs on youtube.

 

[daktari]

I'm not sure it did respond to the hand signal.

There are three possibilities.

It did respond the to hand signal, but isn't showing it. This is the impression the video gives.
It didn't pause for the pedestrian and was just doing what it would have done if the pedestrian has kept walking in the direction he was.
It simply lost the pedestrian and there is evidence of this as the person disappears from the display.

I don't really Musk ever mentioning hand signals. Knowing his humor he'd have it do something like fart on hand signal.

Yes, it seems to have right of way since there is not a crossing for pedeatrians. The pedestrian stops and car goes.

If it wanted to yield for the pedestrian, it should have stayed at the stop line without entering the intersection while waiting for the ped to cross.

 

[S4WRXTTCS]

Elon has mentioned everything wrt FSD. You name it. UFOs landing. Fleet hack driving to Rhode Island. Potholes. James Bond.

https://twitter.com/x/status/1414435086548733952

I laughed because this was legit funny, and not some weird sarcastic laugh.

What drives me nuts about Elon is everything he says is "coming soon"

Like you'd think hand gestures would get a mention in the release notes.

Ambulance/police lights is easy to know its there because the car tells you when you're in AP.

Turn signals it does a piss poor job of because there is no indication that it's recognizing it. Now obviously FSD Beta wouldn't work (or work even worse) with out it, but there is no visual representation of it.

 

[S4WRXTTCS]

I've been thinking about this.

Just see all the videos Mars puts out. 1x, raw videos. Only thing he can be doing without telling us is pressing accelerator.

But, he can drive an hour without disengagement in San Francisco. Some of the videos are when the roads are quite crowded too. If FSD Beta can be used by an amateur to drive for an hour without disengagement, just imagine what billion dollar companies can do. We should be careful about Waymo/Cruise etc putting out "raw" videos - those videos may not tell the whole story - just as Mars 1 hour no disengagement video doesn't tell the whole story either. Neither of those videos are "fake" or fraudulent in anyway - just that there is so much more to FSD than an hour of raw video in one or two places.

PS : Only way to assess the state of FSD of a company is to have it be available to a lot of people who can experience for themselves what it is like. They should be able to take and put out videos without NDAs on youtube.

We certainly see him scrolling the speed.

In fact FSD Beta should be called "Scroll speed control"

 

[EVNow]

We certainly see him scrolling the speed.

In fact FSD Beta should be called "Scroll speed control"

Thats true. But most of the time he is making it more than speed limit.

Though, with Cruise and others we have no idea what they can control remotely ....

 

[diplomat33]