FSD v12.x (end to end AI)

[Knightshade]

So you are going on a long trip and you will be passing through several small towns. You are saying that you will research the local traffic laws of each one?

Random tiny town I'm driving straight through? No because I'm not doing any driving super-local laws are likely to matter for... but any place I'll actually be driving around and parking in? Yeah- I want to know at minimum what's legal for red lights and parking.

Humans can be relied upon to forget these signs, not notice them at all, or not understand what they mean. FSD doesn't have to be perfect any more than a human driver has to be perfect.

They do though- because humans ALSO know how to correct a mistake if they missed an important sign. FSD just sits at the road closed sign and does nothing (if it even stops for it- again V11 anyway there's lots of examples where it wasn't going to).

Forgetting a no passing sign would get the RT pulled over and possibly banned from the roads (since it would demonstrably be failing to obey all traffic laws- something required for >L2 vehicles in the states that allow them on the roads.

But also remember that I'm mainly talking about what it takes to start a robotaxi service in a single city. All this can be taken care of the same way it is done today with Waymo. With Waymo, there is some level of babysitting by a central command center.

This is 1000% the opposite of everything Tesla has ever said on how they want to do robotaxis though, as many have pointed out to you already

 

[spacecoin]

But we do currently trust CV in robotaxis already. Look at Waymo.

I know you said, "CV alone", but we truly are trusting computer vision in a Waymo. Yes, there are other sensors, but if CV fails people die.

A fundamental part of the problem is being able to trust the perception stack for safety critical applications. Not to mention the fact that Waymo's passive cameras in the 5th gen stack (from 2020) are a tad bit more/better/compete than what Tesla has on their cars, where they barely have 360 FoV

"With high-dynamic range and thermal stability over automotive temperature ranges, our vision system cameras are designed to capture more detail and provide sharper images in the toughest driving environments."

Then there is 360 Lidar up to 300 m range, imaging radars, microphones and HD-maps...

Introducing the 5th-generation Waymo Driver: Informed by experience, designed for scale, engineered to tackle more environments

At Waymo, we’re building the World's Most Experienced Driver - our combination of hardware, software, and compute that powers vehicles to safely get people and things where they’re going.

waymo.com

We don't trust unsupervised ML for almost anything unless there are a ton of safety layers and 5-10+ years of other engineering efforts around it (Waymo). Tesla has NOTHING by comparison, because the system isn't designed to be deployed without a driver in the loop.

 

[Therealm80]

Random question, I thought the dash cam records every drive every time however, when I went back to look at the video, it only records, two or three minutes of the beginning of my drives is there anyway to make it record the whole drive every time without pressing the horn?

 

[MP3Mike]

Random question, I thought the dash cam records every drive every time however, when I went back to look at the video, it only records, two or three minutes of the beginning of my drives is there anyway to make it record the whole drive every time without pressing the horn?

It does record everything, but it automatically deletes minutes as soon as they are an hour old. And no, there is no option to turn off the auto-delete feature.

 

[Therealm80]

It does record everything, but it automatically deletes minutes as soon as they are an hour old. And no, there is no option to turn off the auto-delete feature.

Wow that sucks ok thanks

 

[Usain]

This is 1000% the opposite of everything Tesla has ever said on how they want to do robotaxis though, as many have pointed out to you already

That's not really true. They have never said there will be no trial period in a single city. No matter how good Tesla might think FSD has become, they can't just turn on robotaxi and hope for the best.

So we can absolutely expect the first Tesla robotaxis to be geofenced and localized. There is just no other realistic way to get started.

 

[uscbucsfan]

That's not really true. They have never said there will be no trial period in a single city. No matter how good Tesla might think FSD has become, they can't just turn it on and hope for the best.

So we can absolutely expect the first Tesla robotaxis to be geofenced and localized. There is just no other realistic way to get started.

I think the confusion is between robotaxi and FSD. For FSD, Tesla has always said it will work everywhere and there will be no Geofencing, but Robotaxi is different.

People put them together because of the original idea that Robotaxi was coming to consumer vehicles. I don't think that is realistic or imminent, but a dedicated Robotaxi vehicle (Either new or Tesla owned current model) could absolutely be Geofenced.

 

[Usain]

A fundamental part of the problem is being able to trust the perception stack for safety critical applications. Not to mention the fact that Waymo's passive cameras in the 5th gen stack (from 2020) are a tad bit more/better/compete than what Tesla has on their cars, where they barely have 360 FoV

"With high-dynamic range and thermal stability over automotive temperature ranges, our vision system cameras are designed to capture more detail and provide sharper images in the toughest driving environments."

Then there is 360 Lidar up to 300 m range, imaging radars, microphones and HD-maps...

Introducing the 5th-generation Waymo Driver: Informed by experience, designed for scale, engineered to tackle more environments

At Waymo, we’re building the World's Most Experienced Driver - our combination of hardware, software, and compute that powers vehicles to safely get people and things where they’re going.

waymo.com

We don't trust unsupervised ML for almost anything unless there are a ton of safety layers and 5-10+ years of other engineering efforts around it (Waymo). Tesla has NOTHING by comparison, because the system isn't designed to be deployed without a driver in the loop.

This is one point of disagreement. I expect Tesla's system to surpass the safety level of a human relatively soon after deployment in the first trial city. If it's a few times safer than a human, it's safe enough to trust.

So at first Tesla deploys with safety drivers and proves it's safe enough. Then they go driverless. This is no different from what Waymo did and I see no reason Tesla would not do the same.

I also fundamentally disagree with the statement that Tesla's "system isn't designed to be deployed without a driver in the loop." Autonomy has been the goal all along.

 

[powertoold]

The visualization framerate in LA seems to be much higher than in SF:

 

[spacecoin]

I also fundamentally disagree with the statement that Tesla's "system isn't designed to be deployed without a driver in the loop." Autonomy has been the goal all along.

In marketing, yes. In reality, no.

Marketing: Robotaxi Network 2020. Reality: Not autonomous even the the LVCC tunnels.

There is no redundancy (cabling, compute), no sensor cleaning except the front cameras, et c et c.. The Teslas on the road today will most likely never be autonomous in any ODD, because they need 1000-2000x better MTBF. And that's not happening in four years, looking at Cruise, Waymo as a point of reference.

 

[mongo]

I've said this many times on this forum, and here we go again:
If autonomy was solved by "more data" and "more compute" it would have been solved five years ago.

If it can't be solved by more compute and more data then you are saying there is no way to map a unique scene or group of scenes onto a set of optimal control outputs. Which is obviously not true as that is just a hash function.
Now, if you are claiming it can't be solved by a "reasonable" amount of compute/ data, that has the ability to be debated as regards "reasonable".

 

[Usain]

In marketing, yes. In reality, no.

Marketing: Robotaxi Network 2020. Reality: Not autonomous even the the LVCC tunnels.

There is no redundancy (cabling, compute), no sensor cleaning except the front cameras, et c et c.. The Teslas on the road today will most likely never be autonomous in any ODD, because they need 1000-2000x better MTBF. And that's not happening in four years, looking at Cruise, Waymo as a point of reference.

In reality, yes.

Again, if it's a few times safer than a human then Tesla will be ready for an initial rollout in the same way Waymo was rolled out.

Are the current vehicles good enough to become unsupervised robotaxis? Supervised, yes. Unsupervised? I don't think we know that yet.

But if the answer is that current vehicles require remote supervision to be robotaxis then that's OK. Tesla will learn the shortcomings and address them in future vehicles such as the Gen 3.

 

[powertoold]

With V12, Tesla has defeated the final boss of fsd. This is the approach that will eventually be good at driving in India; Ashok's dream will be reality ho ho ho

 

[Knightshade]

In reality, yes.

Again, if it's a few times safer than a human then Tesla will be ready for an initial rollout in the same way Waymo was rolled out.

Are the current vehicles good enough to become unsupervised robotaxis? Supervised, yes. Unsupervised? I don't think we know that yet.

We certainly know that it's not right now

If it were good enough to be unsupervised then they'd have already rolled out the supervised version to more than a tiny handful of people.

But if the answer is that current vehicles require remote supervision to be robotaxis then that's OK. Tesla will learn the shortcomings and address them in future vehicles such as the Gen 3.

I think you misunderstand how remote supervision works.

Waymo can't "drive" the car remotely. When it gets stuck they can tell it, generally, how to get unstuck. (oh, something is blocking that road, turn around and take a different road-- they don't manually execute a u-turn by remotely driving the vehicle)

It's not "supervised" in the sense FSD is right now where a human has to actively intervene in real time during driving. That's why Teslas system is only L2 while Waymos is L4.

 

[powertoold]

I'm awestruck by V12. How does it generalize a behavior like this? This is typical LA/CA zip stopping behavior in locked up traffic:

 

[Usain]

We certainly know that it's not right now

If it were good enough to be unsupervised then they'd have already rolled out the supervised version to more than a tiny handful of people.





I think you misunderstand how remote supervision works.

Waymo can't "drive" the car remotely. When it gets stuck they can tell it, generally, how to get unstuck. (oh, something is blocking that road, turn around and take a different road-- they don't manually execute a u-turn by remotely driving the vehicle)

It's not "supervised" in the sense FSD is right now where a human has to actively intervene in real time during driving. That's why Teslas system is only L2 while Waymos is L4.

Yes, obviously it's not ready right now. V12 needs time to improve and then they have to begin a trial with human drivers. I've already said I don't expect that to happen for 1 to 3 years.

When it's time to take the human drivers out of the vehicles I expect Tesla will do remote supervision similar to Waymo. I understand completely.

 

[OxBrew]

With V12, Tesla has defeated the final boss of fsd. This is the approach that will eventually be good at driving in India; Ashok's dream will be reality ho ho ho

lol, FSD will never drive on crowded Indian roads.

The koolaid is flowing wildly in this thread. V12 will fail to live up to expectations, this and the other thread will quiet down and go back to mind numbingly boring circular discussions about pure speculation, and then V69 on HW69 will get all the fanbois het up again.

Rain, dark, mud on cameras, direct sun all still shut it down. I want what y'all are smoking.

 

[Daniel in SD]

I'm awestruck by V12. How does it generalize a behavior like this? This is typical LA/CA zip stopping behavior in locked up traffic:

That was pretty good. I had to watch 7 minute before it screwed up! (that round about was painful though)
Drives straight through in right turn only lane:

 

[Therealm80]

I’m wondering how it will handle all the roundabouts in Spain

 

[JB47394]

The visualization framerate in LA seems to be much higher than in SF:

V12 coming up to a stopped bus, choosing to go around, then not getting back into the rightmost lane for a coming right turn. It completed the maneuver, but only at the last second.

V12 letting a car in. That section also shows how V12's smoothness in traffic is almost entirely at the mercy of whoever is in front of it. If the car in front accelerates and brakes hard, so will V12. A human driver who is paying attention would start to smooth out that herky-jerky interaction, knowing that each start was just going to be another stop.

V12 entering a left turn lane, but refusing to cross a double yellow line at all to make the lane. Fortunately, V12 makes the green light. At the start, there is a little hitch where it begins to turn for the lane, but then decides that it can't make it, so it straightens out the wheel.

V12 alternating with traffic at an intersection. Perhaps someone can explain this, but I don't see a stop sign, yet V12 somehow knew to alternate.

V12 negotiating a residential road at 12 mph with cars on both sides and a woman standing beside her open car door. The woman didn't seem concerned.

V12 slowly managing a left turn at a tight traffic circle with quick cross traffic, followed by a quick trip straight through another circle without cross traffic.

V12 squeezing into a right turn lane, then going straight from that lane, including a quick merge into the straight lane, dodging an island.

V12 spotting construction in its lane and moving over in a timely fashion.

V12 making Omar nervous as it changes lanes to maneuver around stopped cars.

V12 completing an unprotected left on a four lane divided road. Note that it pulled into the road while a car was passing in the adjacent lane.

V12 making a sharp turn, slowing for speed bump, stop sign interaction, left turn, then pulling to the curb at the end of the drive.