stopcrazypp
Well-Known Member
Every park assist feature out there has blind spots and a long list of limitations, so that doesn't mean it's instantly useless even if there are limitations to its accuracy in certain spots.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
FSD v12.x (end to end AI)
- Thread starter Buckminster
- Start date
-
- Tags
- #fixphantombraking autopilot chill is not chill elon is an ass fix fsd extreme late deceleration fsd does not commit to go fsd does not maintain proper speed fsd hard and late braking fsd jerky driving fsd late deceleration fsd stops too early too quickly fsd stops too far behind line fsd tailgating in city streets
stopcrazypp
Well-Known Member
I'm talking about even the existing ultrasonic sensors, which my current car has. They can miss thin poles that fit right between two sensors (I remember reading a thread where someone hit one due to this). They obviously don't detect side obstacles where there are no sensors. They also have a limited vertical range and will not reliably detect low objects like curbs unless you happened to park at a spot sloped downward toward it (this is the main feature I wish it had for my personal use). It's still useful for its intended purpose, which is to detect large solid objects like cars and walls to the front and rear bumpers.
In a similar sense, the 3D visualization will have limitations and blind spots, but can still be a useful feature.
mgs333
Member
Hmmm that line (... is
) sounds so so familiar...
diplomat33
Average guy who loves autonomous vehicles
I wonder how V12 will handle changing edge cases. For example, there is some construction on a road near my house. But the construction area keeps changing from week to week. One time, they closed the left lane. So traffic in the left lane had to follow the cones into the incoming traffic lane. Then they closed the right lane so that it was now traffic in the right lane that had to follow cones into the left lane. They had a temporary traffic light to control when traffic in each lane could go. Then they switched it so that it was the other lane that was closed. Current FSD beta could follow the temporary traffic light and could move over into the incoming lane to follow the cones but it would get "stuck" in the incoming lane and not move back into its proper lane. Then they opened both lanes so traffic was normal again. Then they had cones in the middle but both lanes were open. But they had a construction worker who would signal when it was your turn to go. This all in the span of a couple weeks. I feel like it might be hard to train end-to-end on these changing cases or will it be able to truly generalize and understand when to do what?
If V12 has been trained on the scenarios given, i.e. following construction cones into opposing lanes and following them back to the correct lanes, it should handle the situation. If not, then disengagements should help identify edge cases where more training is needed. I am especially interested in how V12 handles humans giving directions. That's something that FSD has never done.
The fact that a particular construction zone is changing should not be an issue so long as V12 knows how to handle the situation as it approaches it. Even if the construction area changes as you circle around the block for a second pass, I don't expect V12 to remember how things were set up on the first pass.
I fully expect there to be many situations that V12 does not initially handle correctly and these may be completely different than the current FSDb. It will be fun to explore these.
It's probably quite intentional that the general world model example video at CVPR showed how it can predict future frames of returning to the proper lane in a construction area. The livestream actually showed some initially incorrect behavior of staying in the center lane when exiting the construction area, but soon end-to-end figured out it should move back into its proper lane. Perhaps it was confused by its upcoming left turn and realized it shouldn't drive into oncoming traffic left turn lanes.
Good example of why Tesla should push for an L3 option as they work on L4/L5 for the future. Not so much because the cones and lanes change but because of the different hand gestures and styles different people use to control traffic. Sometimes it's police offers, other times it's construction workers. Then add how challenging it can be when coming up to emergency vehicles and people controlling traffic. Rather than trying to get all this right now just formally hand off FSD to the driver.
Who here thinks L4/L5 is happening in 2024?
diplomat33
Average guy who loves autonomous vehicles
This is precisely why car companies are taking a more incremental approach of L2 -> L3 -> L4. They recognize that L4/L5 is not solved yet and there are many remaining challenges. So they feel it makes sense to have the driver simply take over when the car encounters a case it cannot do yet. And then as more edge cases become solved, they can have the car handle more and more of the driving. I agree, I would love it if Tesla gave us some "hands off" or even limited "eyes off" in some cases. That way we could have get some value while we wait.
I think we need to distinguish between L4 and L5 because they can have very different ODDs and therefore very different chances of success. Remember that L4 can be any limited ODD. So L4 can be as small as a 5 square mile geofence or highway-only or every road in the US but only in day time and sunny weather. But L5 has to be the entire ODD of a typical human driver. A 5 sq mi geofence would be far easier than the entire human ODD. If the L4 ODD were small enough, Tesla probably could unsupervised L4 in 2024. Could Tesla do supervised L5 in 2024? Maybe since supervised FSD is much easier than unsupervised FSD. Could they do unsupervised L5 in 2025? Not a chance IMO.
AlanSubie4Life
Efficiency Obsessed Member
What? Can’t think of a meaningful ODD. Not even a parking lot. Not with current hardware, anyway.
L3 in a meaningful ODD (speeds over 50-60km/h) using camera-only ain't going to happen anytime soon. As long as Autopark and summon isn't L3 with manufacturer liability, I don't see anything Autopilot becoming L3. I don't think Tesla is anywhere near L3 in "city streets" nor on "limited access highways", which are the two partial ODD:s that are marketed.
Possibly they could aim for something similar to DrivePilot's "queue-chauffeur". But remember they need to design a hand-over process for that, and that takes months, so probably not 2024 even for that.
If they don't have 360 sensing I think it's hard to remove human supervision even if the ODD is "close vicinity of a parking space in daylight and zero precipitation".
My guess is that Tesla remains an L2 for all existing vehicles forever in all but the simplest ODD. A great L2 with the possibility of a "queue-chauffeur" in 2025 is my honest best guess. By 2026 at the latest they will add more sensing and camera cleaning if they have unsupervised ambitions.
Where I live AP is barely usable 5 months of the year due to wet/slushy roads, darkness and snow.
Last edited:
KArnold
Active Member
I still wonder if an L4 with ODD of 1) Interstates or limited interstates, 2) no construction zones, 3) no precipitation, and 4) daytime only isn't far away.
I drove from Cleveland to Columbus (160 miles) without intervention today. Did have a few lane changes. Granted only first/last few miles were not a well marked interstate. I'd love to know how many safety-related interventions happen under these circumstances. Then again, Midwest drivers are different from left/right coast drInters as well.
I'm not suggesting it was the most efficient or wouldn't PO someone by hanging out in the left lane - I'm suggesting I felt safe.
You guys seeing many safety issues with FSD on daytime, dry, no construction interstates?
I drove from Cleveland to Columbus (160 miles) without intervention today. Did have a few lane changes. Granted only first/last few miles were not a well marked interstate. I'd love to know how many safety-related interventions happen under these circumstances. Then again, Midwest drivers are different from left/right coast drInters as well.
I'm not suggesting it was the most efficient or wouldn't PO someone by hanging out in the left lane - I'm suggesting I felt safe.
You guys seeing many safety issues with FSD on daytime, dry, no construction interstates?
...
For me it's been pretty safe in those scenarios, even with light precip and/or at night. I think the bigger issue with such an ODD is how you safely switch out of it when the conditions are no longer met. You don't want it doing a safety stop on the shoulder or whatever, but you need the human to acknowledge a transition back to L2 fairly quickly, and to actually pay attention from there forward. There could a lot of confusion issues.
AlanSubie4Life
Efficiency Obsessed Member
I don’t wonder. It is not happening with current hardware. These minor caveats alone demonstrate that it is not possible. Remember construction zones can appear out of nowhere at any time: they can’t be mapped. What about CHP running a traffic break? That’s no easier and can happen suddenly and there’s no time for a graceful transition to a minimum risk condition which is required for L4. It needs to be able to handle a remarkably difficult and highly unpredictable set of emergent conditions to be able to manage fair-weather L4 on the interstate.
I doubt L3 is possible either, for the same reason. Challenging conditions can occur in seconds.
Don’t get me wrong: I’d love to be wrong. But I do not intend to hold out hope and delude myself when the evidence is so compelling that it will not happen.
I'm inclined to agree with - there's a lot of hard driving situations that the car can't give a driver 30 seconds* of lead time to switch back to switch back to L2 operation.
Even the fair weather part of this could be tricky. Does the car check the weather forecast and only engage if there's no precipitation and weather is warm enough to prevent any risk of icy conditions?
All of this is probably why Mercedes in their limited L3 capabilities heavily limit max speed.
*I'm not sure what reasonable handoff times are but 30 seconds seems reasonable for a person to go from zero situational awareness to ready to take over.
Even the fair weather part of this could be tricky. Does the car check the weather forecast and only engage if there's no precipitation and weather is warm enough to prevent any risk of icy conditions?
All of this is probably why Mercedes in their limited L3 capabilities heavily limit max speed.
*I'm not sure what reasonable handoff times are but 30 seconds seems reasonable for a person to go from zero situational awareness to ready to take over.
KArnold
Active Member
The rules are somewhat vague and I'm starting to better appreciate a more clear standard. Especially with Elon who plays the rules very loosely with his "release now, fix later" approach. I don't believe the existing rules talk much about how an L4 vehicle reacts when leaving an ODD. Some articles say the vehicle needs to "stop" - obviously not practical if in the middle if a freeway construction zone. I do not see "graceful" or a timeframe in the standard but perhaps I am missing it. And "no construction zones" can be a definable ODD as can weather.
So perhaps by letter of the standards L4 with such an ODD could be done safely soonish and perhaps even safely depending on the accident rates we don't have. But practically speaking if an L4 car must react to construction, obstructions, and practical transitions when leaving ODD beyond simply stopping - then I agree it's nowhere close. An L4 car doesn't need a human within ODD areas - if "no rain" is part of that ODD maybe that model would work in the Sahara only.
Even safely pulling over in the next 30 seconds will not work if your ODD includes light rain or construction zones - imagine those headlines.Sorry, probably the wrong comnent for this thread anyway. Just curious.
Last edited:
Yeah the timing of the transition is everything. I could imagine someone coming up with such an ODD that worked along the lines of: when the rain/construction/whatever is imminent, it does some beeps/warnings and gives maybe ~5-10 seconds to take over gracefully while it's already beginning to slow and prepare to pull off the road or whatever, but probably within 15s at most it's gonna have to really stop if the driver hasn't taken over. This is maybe enough for L3+ (no steering nags, no "pay attention to road") which still requires a capable driver be "ready to take over". You could still be reading a book or doomscrolling your phone or whatever, but be able to take over in those conditions and in that short time window. That's very different than a robotaxi, or the ability to crawl in the back and go to sleep. Basically the ODD's definition would have to include driver readiness (sleeping/absent driver is not ok. driver reading a book is ok).
Similar threads
