FSD 12.3.4 - ROBOTAXI IS NEAR!!! 😱 (Night Drive Tesla Model 3)
ฝัง
- เผยแพร่เมื่อ 30 พ.ค. 2024
- I will be testing out the FSD 12.3.4 with the Tesla Model 3 driving at night on local streets. Let's see how much better it is compared with FSD 12.3.3!
Correction: Should have said replacing 300k lines of code!
0:00 Introduction to FSD 12.3.4
0:15 FSD Driving Modes (Chill, Average, Assertive)
0:37 How to start FSD on Tesla
0:55 Navigating out of parking spot with FSD 12.3.4
1:50 Making a right turn on stop sign with FSD 12.3.4
3:59 Left turn on two way busy street with FSD 12.3.4
5:08 Curb hug test making a right turn with FSD 12.3.4
5:42 RoboTaxi getting near with FSD 12.3.4
8:00 Preparing for a right turn at lights with FSD 12.3.4
10:04 Avoiding a biker at a light with FSD 12.3.4
10:52 Unprotected left into a plaza with FSD 12.3.4
11:33 Navigating in a parking lot with FSD 12.3.4
11:52 Autoparking with FSD 12.3.4
Sharing my referral link for when you order your Tesla. You’ll get awards like 3 months of FSD (Supervised) when you take delivery. ts.la/kevin145437
Its not 300k lines of code that have been updated, its 300k lines of heuristic code that have been replaced by the neural network stack. Its even more impressive to think that 300k lines of code have been replaced by about 3k...
Yes thanks for the clarification!
Out of dozens of fsd drive videos, this is the first night drive! Kudos to you
Thanks!
That's because most daytime FSD videos have human interventions. Night driving would be even worse.
I agree. I have tested v12 for 2590 miles , each version fixed something and just gets better.
I completely feel as we speak now that the car is 99 percent better driver than most humans and we are just forced to supervise it due to regulations
The future is here folks… exciting times
Yeah it’s really exciting to be experiencing this development in AI!
You say that 12.3.4 is 99 percent better than most humans?
Well that is easy to test! Have FSD drive 100 miles from busy suburb to busy inner city, during morning or afternoon rush hour. And *without* human intervention. Then you will see that FSD is nothing like a human driver.
thanks for the research video! thumbs up
Hope it helped!
Kevin, it's worth noting that the discussed robo-taxi is expected to lack a steering wheel, eliminating the possibility of a robot or any entity taking control via a wheel. Consequently, the interior layout will likely be redesigned due to the absence of a steering wheel.
Yes great point!
I hope they bring the new vision auto park to USS cars soon! nice video!
Thanks
FSD 12.3.x is VERY good when compared to FSD 11.x or earlier. Is it close to ready for driving a 'robotaxi'? No, it's probably years away, plus a new car design is reaquired.
Kevin, you are driving in very easy conditions, on a new easy grid road network and light traffic. Driving around for a few minutes there is not a serious AV test. I know for a fact that the GoogleX robo car was at least as good as current FSD back in 2012, when it tested in Silicon Valley. it took them eleven years with at least three hardware upgrades, plus good custom maps and a remote operations system, to solve San Fran's streets. Waymo just passed ten million continuous driverless city miles without a serious crash.
FSD is not close to being ready for robotaxi. To get any valuable license to pull the driver, for a small test fleet in any state including Texas, Nevada, or Arizona (all light regulation states), it will have to be at least at good-human-driver levels of safety in busy city driving. That means able to drive around with no human for a million miles (about two lifetimes of driving) everywhere in a city like S.F. that it wants the license, with no serious failure. Any worse than that and a small test fleet, like fifty robocars, would have a car crashing every few days or months. That would shut down the program before it gets anywhere.
And for a moderate-sized commercial fleet of robotaxis, with say 10,000 robotaxis driving around a few cities, it will need to be very super-human at safety driving, meaning able to go at least ten million miles between failures, and likely even better than that. Otherwise there would be a bad crash far too often to keep a license and the public's trust. In a large robotaxi fleet of a million cars nationwide, it will have to be way over ten-million MTBF (mean time between failure).
I'm certain FSD currently wouldn't last one hour in a real city, like downtown San Fran, during busy hours on random taxi-like drives, regularly encountering all parts of the city, which is what real AV testing looks like. It might not even achieve fifteen minutes MTBF in that kind of urban driving. It's not remotely close to robotaxi levels of safety, and your easy tests do not address the real challenge of urban taxi driving. Drive around downtown S.F. or an equivalent big city for an hour at a time in busy conditions on video to give us a real idea of where FSD is at.
Great suggestion! I will try it out in busier areas like downtown la sometime
Charlie, an *awesome* post. I am going to print this and share it with the Tesla cheerleaders at work.
FSD is nothing like a human driver!
There are plenty of other videos showing complex city driving on 12.3.3 with no problems whatsoever. FSD is of course a much more advanced system than Waymo and similar robotaxi systems which only work in small pre-mapped areas and under perfect conditions. Waymo and similar Robotaxi systems have 10-25 million km in AI training data, Tesla has around 5 billion km in training data. But with how FSD is developed to function as a level 5 system pretty much anywhere under any normal conditions, it has taken much more work and more time to perfect it. FSD 12 is driving in a much more human way than Waymo, DeepRoute, and other such systems, but working out the final small quirks has taken a long time. I think we are getting close to FSD being launched in an unsupervised/robotaxi mode, and I am pretty sure it will be a lot safer than human drivers, not to mention other comparable systems.
@@HiteshCeon you have no idea what you're talking about.
@@HiteshCeon _Tesla has around 5 billion km in training data._
Yes, and FSD v12 has been training for 90 days now. And remember, Tesla is no longer compute constrained. They have a huge data center, running 24 hours per day for months. Trillions of instructions per second.
And yet TODAY, the latest FSD is still making all of kinds of obvious and unforgivable mistakes. Parking is still broken and incomplete. Multiple videos on TH-cam show this to be true.
Full Autonomy? what a joke.
Can you dismiss the steering wheel nag using the scroll wheel on V12.3.4?
No you have to rotate the wheel
It was working on the v12.3.3 right?
How is it at night down a narrow country road with no street lights ?
If it’s dark it’ll turn on the high beam so that should help with visibility
Tesla doesnt want to come to a full complete stop at stop signs but lawmakers require it to come down to 0 miles and wait half a second or more before continuing
Good point!
You are right. "Rolling" a stop sign can get you a ticket and a large fine.
Supposedly its on HW3 so will try it but SUPER DOUBTFUL its remotely as GOOD AS ME ! LMFAO
Amazing how you barely touched the wheel even to park - just amazing.
Yeah it’s really coming along!
Sorry Elon but I will not pay for unfinished FDS 🤑Once its fully operational YES but not until
5:50 I think that's because of inside camera. I have MX without inside camera, didn't notice any difference in nagging.
Oh interesting
dd14b
FIRST MINUTE on VIDEO it FAILS !!!!!!! - LMFAO !!!!!!!!!!!
it will never be approved even for L3 ! mark my words
Says who? I take Waymo’s several times a week and those drives are 100% autonomous
@@aguzman222 Waymo vehicle is much more expensive with much better sensors than any Tesla.
Waymo only operates in a small zone, in two cities. FSD wants to operate all over America. And that is dangerous.
300k lines of human coding NOT needed anymore. THAT is the headline.
Yeah good one
Tesla have to be cautious to avoid accidents. Humans are generally poor drivers.
FYI - the visualisation has nothing to do with FSD now that it is all AI single stack..
I’ve never understood that. IMO the visualization was very accurate and wonder why fsd does not use it more to be more confident.
@@smoak888 Because the AI can no longer 'see' the visualisation. It just looks at the cameras and acts accordingly (according to patterns it has created based on videos of good driver behavior). This is how it should be - I would much prefer my driver look out the window in front of him than at the Tesla screen.. If they ever disagree with each other - I would pray that he follows what's going on out the window and not the screen..
The visualization is merely a fraction of what FSD sees. Look for the AI day video of the Tesla's au team. The amount of things the car sees is phenomenal compared to what is on the screen.
@@alexandreroy9071 I agree with you all. I was only commenting how accurate the visualization seems to be.
@@alexandreroy9071 to my understanding it no longer re-creates a 3d impression of its surroundings.. It merely looks at the video and acts according to its training data now.. this is what is meant by single stack FSD.. everything is AI now..
Failed Self Driving == No Lidar No Radar and No redundancy to handle Failed VISION ! get SERIOUS and STOPPING making CLICK. BAIT ok
OP said FSD 12.4 almost KILLED HIM ! and het says its rubbish ! get SERIOUS and stop making CLICK BAIT