Model 3 - FSD Supervised 12.4.2 - Santa Clarita Test Loop - 2024.15.10 - 1 July 2024
Vložit
- čas přidán 30. 06. 2024
- This drive was doing well until we had a rapid succession of issues.
First issue was the car missed its navigation route due to not moving over soon enough.
After it missed its route it recalculated but missed that turn and stopped in the travel lane, thankfully with no one behind us while it recalculated again.
After we rerouted and got back on our path the car had trouble deciding which left turn lane to pick and I had to make a safety intervention as the car was on a collision course with the vehicle in front of me waiting to turn left.
Finally the next right turn it ran the red light it had lots of time to stop at.
So overall things went smoothly until they didn't. Shows how far things have come but still a ways to go. - Auta a dopravní prostředky
Eep lane selection is still its biggest issue thought it’d be fixed by now
Watched the whole clip and FSD is looking better every day. Thanks for posting this impressive demonstration of solid FSD
you sound like Joe Biden defending for his debate's performance..
22:11 you could see better then us when it was red vs yellow but the car display had it yellow at 14 mph right at the end of the painted arrow, if this is true it makes sense to run it to me but again you had a better view of when it actually turned red..
Thanks for posting your videos so fun to watch fsd progress.
Light turns yellow at the end of second 22:08. There is definitely enough time to stop, so this is not how I'd like to see the system to behave.
But hey, I'm from Europe, so what do I know 😅
@@zahlex im curious if it would have gone without the lead white car in front.
@@zahlexIt doesn’t react that way sit usually stops at yellows
@@blessguy5330 Usually is not often enough ... Thats the whole problem about AI systems, get them to decide for the correct option always.
Followed the typical CA driver ahead and ran thru the red light on a right turn. Very dangerous for pedestrian and bicyclists, and other cross-traffic.
Illegal yes. Dangerous no. It would have not gone forward if there was obstruction.
VERY DANGEROUS 😂😂
Safety intervention:
20:27
Ran red light:
22:09
4:35 actually it's exactly the same spot. It spots the right arrow and knows it's the wrong lane.
As explained before it has knowledge that there will be two left lanes and no knowledge how the 2 lanes will split up, so it keeps in the right lane until it knows more (because you're supposed to drive in the right lane if you don't pass).
So it did everything by the book, as there's no info how to proceed available before this point.
This is a good example of where map data is useful though. Local human drivers very often learn these things so they can plan ahead beyond how far they can see. Not a big deal that FSD can't do this, but I would like to see Tesla have better map data.
@@graemeesmith technically that's more of a recall thing, than a map thing. They could dynamically capture such details and if confirmed by a second pass and the cars get this as a hint what lies ahead to make better decisions.
The benefit of that is, that the round-trip is much shorter, as it would take a day or two until this information is gathered and available, rather than month for the next map update to arrive.
@@RubenKelevra That would be amzing but i don t see how they re gonna make it happen. Seems an impossible amount of data. Maybe i m wrong. I ve been thinking Tesla should develop their own maps. Is that what you re saying? Integrate a mental map in the neural network itself?
@@lucadellasciucca967 well it's kinda simple. Have the AI report issues in an area, and then overlay a fix of the map over the routing data for the next cars which come along.
Tesla has been talking about that for a long time, so it's on their todo list. But it's less explicit in their approach. They basically will stream the "future" to the car, so basically a look ahead of what other cars saw in terms of stationary objects when they drove by the last time.
This includes road marking, parked vehicles, closed road signs etc.
So it's like an additional camera view, just looking further ahead. This way the car can plan ahead better even on changing conditions, but also choose to ignore this information if the situation has changed and then report back the new information.
@@RubenKelevra That s good point. I m aware of that. Would definitely solve the map issue. I have a hell of a hard time immagining, the data pipeline etc I can tell it s phisically possible tho. Thank god i don t work at Tesla 😅
Thanks for the video and finding more defects than Dirty, Black or Whole Mars. Your earned my Subscription. The dithering between two lanes is extremely bad behavior and I've had close calls like yours before with V12.3.5. Tesla should be able to detect that condition in the AI engine and shock the AI out of this behavior. FSD should also automatically detect these conditions and report them back to Tesla automatically so mitigation measures can be developed. This sort of AI brain freeze is extremely bad behavior should be a top Tesla priority to stop. Nearly every brain freeze is heading toward an accident and FSD should be considered a Beta as long as any of this behavior still exists.
Yeah the only real issue for me is the near collision with lane dancing. Others could've been done by human and are not dangerous, even when stopped in the wrong spot, a human could've done the same mistake and it's not sure that with someone behind it would've done the same thing.
Please continue testing the speed bumps. Is any progress being made on this frequent problem area for FSD?
Yep, wrong-lane-anxiety 😅
I would prefer the FsD not stay in the slow lane, or the far right lane unless its part of its route
Monkey see FSD do.
On the right turn the driver ahead also rolled through the red light and FSD tends to copy the driver ahead of it.
It looks like both cars ahead rolled through, but definitely the car directly ahead of FSD did not stop.
It slowed down to about 11 mph, - I'm a pretty cautious driver so I would have probably rolled through about 4 or 5 mph.
When the car is waiting too long to get into a lane for an upcoming turn, could you manually turn on the turn signal to have the car change lanes? I can't quite remember if it works like that, but if it does I imagine that will count as an intervention for Tesla and they might fix it sooner.
May be you use wipers for clear visuals👓🕶
Im always last to get the new versions since im in nyc. I think tesla is scared to release to nyc because fsd true test is here.
What are your settings as far as chill, average, aggressive? Minimal lane changes?
@James Locke: Is there a setting so that when you engage in FSD, to disable the voice navigation? (I am a new Tesla Owner and still learning)
James i think a lot if your criticisms are down to understanding the local driving conditions which FSD cant have right now. Where you've said it's not getting in to lanes early enough, your basing that on experience of being blocked from the lane change by other traffic. Anyone driving in the area for the first time would likely make the same mistakes.
It might be that FSD can never work wothout this local knowledge. Most people drive the same roads everyday, but put people in a new city, navigating by their phone, and they make similar mistakes.
It seems that FSD is capable of maneuvering in real-time to follow a path to a destination. However, it does not "remember" the path. This may be why it makes lane changes at 0.2 miles from a turn rather than earlier as the driver would. Could this be what is called "depth" in AI?
20:25 is interesting - intervening was definitely the right call. It looks like the path planner was alternating between stopping and staying in its lane, and changing lanes and accelerating. Neither path that it was alternating between was on a collision course, but by jumping back and forth it was averaging them out. Curious to know if it would have stopped or committed to changing lanes once it itself close enough that it has to pick one or the other.
It’s situations like these where I’m dying to know if the Tesla would allow itself to collide or stop itself, even though both supposed paths were clear.
Wow looks like almost perfect drive for a robotaxi.
The human driver had to intervene to prevent a *collision* at 20:26 .
Definitely not ready for robotaxi. And Tesla robotaxi is 4 years late anyway.
Primitive basic driving and “FSD” makes unforced errors.
What will happen during non-standard situations or unexpected obstacles?
Do you have it set to aggressive mode?
@@davidhite1868 there's no such mode
@@RubenKelevra did they eliminate it in this version? It's in the 12.3 version.
20:29 Certain that it would have hit? Wonder if it would not have.
lol, I don't think so, but everybody would have disengaged panically.
We are dealing with a simpleminded FSD that makes mistakes every day, on challenging drives.
Of course the human disengaged. There was no other option.
Timestamps in the description of the critical points would be much appreciated. I don't have time to scrub through 40 minutes of video searching.
Late deceleration forcing friction brake usage still happening. 0:40, 1:43, 3:57, 4:49, 12:53, 14:20
First
I think fsd should not compete with human drivers because they are the reason for many accidents!. Let's fsd do its things for safety reasons
Look up how many accidents a typical human driver has, over their lifetime. Humans are very good drivers !
Do your research, invest in Tesla.
Seems FSD hit a wall. Every fixed bug is met with 1 or 2 equally bad new bugs.
Time to accept the truth. Tesla's FSD isn't happening.
Will be fixed soon🎉🎉🎉
You will always be just a month away from the next version that will fix everything.
They've hit walls before. And they've passed them. The question is how tall the wall is, and if it's possible for them to climb it. And then when.
So please explain to me why all the videos except 1 that I saw stated v12.4.1 was an improvement? Even this OP has said v12.4.2 is an improvement. FSD has already happened by the way. I use it every day for every drive and hate driving an ICE car. (i.e. car rental) If this was a robotaxi drive it would have completed the drive safely.
A 39min drive with one disengagement is actually not that bad and definitely a big improvement outside S.F. area.
To much relying on ML, using no logic looks terrible forever
FSD cannot think or reason or adapt like a human. It will be terrible forever, the design is flawed.