It would make sense to me to have an independent, procedural thread running alongside the NN-based nav (backing it up) just to look for emergencies and avoid things like collisions. I can't imagine that would be too compute-expensive a fail safe.
-
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
AlanSubie4Life
Efficiency Obsessed Member
It’s almost like it’s not nothing but nets.
KArnold
Active Member
Think I just saw an epitaph with that written on it.
Breaking news:
www.notateslaapp.com
2024.2.2 Official Tesla Release Notes - Software Updates
Tesla software update 2024.2.2 includes Adaptive High Beams, Security Improvements.
www.notateslaapp.com
jebinc
Well-Known Member
For the past three years, it is always, "not quite ready," but will be a "brain-blow," soooooon.... Same 'ol, same 'ol story.
jebinc
Well-Known Member
It comes tonight at 1 am your local time!!
Except for @jebinc ... Elon just texted me and said he is pushing V12 to him asap. He wants him to be able to drive out of his neighborhood and make it out the gate with no disengagements..
I pissed myself, after reading your post!
KArnold
Active Member
One Improvement I'm looking for from v12: the ability to distinguish between light and heavy traffic, and to adjust decisions based on that.
For example, with UPLs (for which I keep a very tight rein in v11, and often avoid entirely), good human behavior is to pass up a small oncoming traffic gap, to wait for a more comfortable gap - that is either in sight already, or is expected due to the general traffic density at that particular location and time of day.
In my experience with v11, the ego car will often choose to make a marginally safe but unnecessary turn in front of an oncoming car. The choice to go would be perfectly acceptable and probably expected in heavy traffic, because everyone around knows you have to take those opportunities if you're ever going to complete the maneuver. However, in lighter traffic, the prudent and expected behavior is to pass up such a close-call turn and wait a little longer for the next large gap.
So many times I've approached a UPL or UPR (whethrr at a real traffic-light intersection or just at a residential road turn-in), and v11 will treat a lone oncoming car with the same safety-margin calculation that it would use in heavy traffic. Unnecessary, unexpected and unnerving for everyone. A good and comfortable human driver won't pull a maneuver like this, but simply wait for a clearer opportunity behind the oncoming car(s).
From everything we understand about v12 training and decision making, this kind of decision should be much improved. The clips from the teacher-drivers should intrinsically ally reflect the situational awareness that allows a tradeoff of go/no-go decisions vs. speed and distance of the hostile traffic.
It's not clear to me, however, that the exact decision threshold can easily be tied to a Chill/Average/Aggressive UI control. To do that would require much more extensive training data, with different aggressiveness ratings of teacher-drivers encountering very similar cases over and over, and then judging and classifying the aggressiveness so that the neural net could learn what kind of driver it's supposed to emulate based on the control setting at runtime.
(Does anyone know if the early v12 interface includes the aggressiveness control?)
For now, I agree with some of the arguments up-thread that say we should expect v12 to be trained on the cautious side - but specifically not to display the annoyingly cautious, hesitant kind of behavior that plagues v11 and actually makes it less safe when the cars are "negotiating" with other drivers and pedestrians. So far, the Whole Mars videos seem to bear this out, showing more of the confidence that's needed, without unsafe aggressiveness in dense SF driving. To me that's good news. But I haven't watched enough to say that there's noticeable Improvement in the UPL and UPR gap judgment (against higher speed traffic), which I'm also hoping for.
For example, with UPLs (for which I keep a very tight rein in v11, and often avoid entirely), good human behavior is to pass up a small oncoming traffic gap, to wait for a more comfortable gap - that is either in sight already, or is expected due to the general traffic density at that particular location and time of day.
In my experience with v11, the ego car will often choose to make a marginally safe but unnecessary turn in front of an oncoming car. The choice to go would be perfectly acceptable and probably expected in heavy traffic, because everyone around knows you have to take those opportunities if you're ever going to complete the maneuver. However, in lighter traffic, the prudent and expected behavior is to pass up such a close-call turn and wait a little longer for the next large gap.
So many times I've approached a UPL or UPR (whethrr at a real traffic-light intersection or just at a residential road turn-in), and v11 will treat a lone oncoming car with the same safety-margin calculation that it would use in heavy traffic. Unnecessary, unexpected and unnerving for everyone. A good and comfortable human driver won't pull a maneuver like this, but simply wait for a clearer opportunity behind the oncoming car(s).
From everything we understand about v12 training and decision making, this kind of decision should be much improved. The clips from the teacher-drivers should intrinsically ally reflect the situational awareness that allows a tradeoff of go/no-go decisions vs. speed and distance of the hostile traffic.
It's not clear to me, however, that the exact decision threshold can easily be tied to a Chill/Average/Aggressive UI control. To do that would require much more extensive training data, with different aggressiveness ratings of teacher-drivers encountering very similar cases over and over, and then judging and classifying the aggressiveness so that the neural net could learn what kind of driver it's supposed to emulate based on the control setting at runtime.
(Does anyone know if the early v12 interface includes the aggressiveness control?)
For now, I agree with some of the arguments up-thread that say we should expect v12 to be trained on the cautious side - but specifically not to display the annoyingly cautious, hesitant kind of behavior that plagues v11 and actually makes it less safe when the cars are "negotiating" with other drivers and pedestrians. So far, the Whole Mars videos seem to bear this out, showing more of the confidence that's needed, without unsafe aggressiveness in dense SF driving. To me that's good news. But I haven't watched enough to say that there's noticeable Improvement in the UPL and UPR gap judgment (against higher speed traffic), which I'm also hoping for.
Is that version 44.30.12 which contains V12? Why is it going to Europe?
KArnold
Active Member
That might be possible from the videos they have. Let's say they have videos for 1b UPL's. Measure all, throw out the top/bottom 10%, cut remaing time-to-complete-turn into 3 categories connected to driver aggressiveness buttons. I know, it's likely down the road but from what I see in AI we'll have aggressiveness profiles of some kind.
Personally I'd love to upload my driving videos to the mothership nightly (which I probably do anyway) in exchange for an aggressiveness profile specific to me - i.e. on this road I drive limit plus 5, but on that road it's always limit -5 due to lots of kids. I know, that one might be a pipe dream.
From what I've seen in Omar's videos, unless Tesla makes some big improvements on hesitance at stop signs, you will be drinking for free for some time yet.
Agreed, but I think we're both saying it may not happen in the initial deployments. I'm saying it requires enough data to do that classification, and I'm assuming that right now they're trying just to get to enough training data to deploy a safe AI driver. As you say, the step beyond is likely down the road.
But, there may be some features of the model that make it easier for the AI to judge aggressiveness and therefore adjust Its behavior at user request. Maybe tied into all the various confidence judgments that we're pretty sure are already part of the decision networks.
I've long wanted that kind of customization. Starting perhaps with an interface that allows us to submit our own map corrections and hints on our own familiar routes. But understandably, I think Tesla is trying to do all this learning from Fleet telemetry and not from any level of offline user programming.
Eventually this could lead to route planning and behavioral settings tied to individual operator behavior - but it feels pretty far "down the road"!
Todd Burch
14-Year Member
No, it has nothing to do with FSD v12. v12's not going wide for at least a few weeks.
willow_hiller
Well-Known Member
Again, you're failing to distinguish between the part of FSD Beta that is currently trained via large amounts of data (perception), and the bit that is largely hand-coded (planning and control).
In all my testing of FSD Beta, very few cases of bad behavior were caused by the system not perceiving the environment. I'm reasonably confident that if you only gave a driver the FSD Beta visuals, they could operate the car much better than V11 does now.
You said "FSD training hasn't been TSLA's strong suit," but it's always been their strongest suit. The planning and control has always been the weakest, and now V12 largely removes that weak link in the chain.
My meaning was that FSD has always been about expanding the features to secondary roads - autosteer on city streets. So, when V12, the seeming superior system, is out among customers, we should see more use of FSD on secondary roads. That should drive down the miles per accident figure because secondary road accidents are three times as common, per mile, as highway accidents.
It had nothing to do with speaking for Tesla and everything about an expectation of what will happen to the safety numbers when V12 is in common use.
AlanSubie4Life
Efficiency Obsessed Member
Yes, we’ll possibly see reduction in safety as measured by miles traveled per accident.
It is kind of crazy that they don’t break out City Streets stats separately.
But it is in keeping with them publishing no actual usable data on safety at all.
yerEVan
Active Member
Similar threads
- Article
