Looks like the same "FSD Beta v12" release notes from a few weeks ago for 2023.38.10. But presumably they made improvements to bump it to "FSD Beta v12.1" as well as updating to 2023.44.30.10, so that employee vehicles already on holiday update could get this version. The release notes seem to be remotely hosted anyway, so theoretically customer vehicles could update to this same version if there aren't issues found with wider employee testing to slowly rollout even more.
-
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
ODD limitation; close vicinity of a parking space? I don’t think Tesla ever will take on liability for even Autopark for existing vehicles
Thanks for the example. I'm not sure if FSD Beta has had that type of control logic specifically for roundabouts as it seems to be more of a general driving behavior to check for intersecting paths of relevant cars continuously in addition to all the other things to consider like lanes. The complexity comes from all the other things to simultaneously control for like a roundabout-looking intersection but actually has traffic lights or a pedestrian in the center or construction cones are on the edge, etc. Or even more basic, the vehicles in the roundabout suddenly stop, so proceeding still needs to avoid hitting the new lead vehicle.
If a neural network was trained to handle all the potential behaviors needed at a roundabout, it seems like it might as well be trained to handle every other intersection as well non-intersections. But I suppose if there was special roundabout-specific control logic and this control neural network was only trained for roundabouts, that portion of software 1.0 could switch over to 2.0, and the rest of control could follow the pattern and incrementally get to end-to-end modular control
Perhaps Tesla did look into incrementally increasing scope of a control network vs a complete control network for every situation vs single end-to-end network and decided that it would be best to go with single.
Hah. 
TeslaFi is on top of this one -- already labeled 2023.44.30.10 as FSD Beta 12.1 without any pending installs. TeslaInfo already has a 2017 Model 3 on that version, so we might be able to notice when Tesla expands to their widest employee testing group on these services. If they follow their 2-week pattern, I suppose it could be end of the first week of new year?
TeslaFi is on top of this one -- already labeled 2023.44.30.10 as FSD Beta 12.1 without any pending installs. TeslaInfo already has a 2017 Model 3 on that version, so we might be able to notice when Tesla expands to their widest employee testing group on these services. If they follow their 2-week pattern, I suppose it could be end of the first week of new year?
Todd Burch
14-Year Member
The smart way to do it—and they way I suspect they’re actually doing it—is to send the recording through voice recognition, then look for keywords. For example, “speed bump”.
Then, suddenly you have 50,000 immediate examples of the car failing to slow down for a speed bump. Then run those clips through the auto-labeler (or if early in the process, label manually as needed). Review the auto-labeled results by a human and adjust as necessary. Would be a great way to get a lot of pertinent clips quickly.
Last edited:
powertoold
Active Member
Yes, I believe this is the case. A software 2.0 strategy is actually compute inefficient for inference. Karpathy has talked about FSD being a balancing act of inference compute since 2020, where he mentioned that different fsd teams would be rationed some portion of the compute for the 1000+ outputs from HW3. Over time, it follows that you'd need more efficient architectures as you consume more heuristics with NNs.
diplomat33
Average guy who loves autonomous vehicles
Anyone notice that the release notes for V12 say that it upgrades "city streets driving"? This seems to imply to me that the new end-to-end stack is only for city streets right now and therefore V12 still uses the "V11 stack" for highway driving.
Tesla has been very cautious in modifying Navigate on Autopilot. Likely due it its level of performance in that domain. Methodology may be to test new stuff in City which has more stressors, and only migrate to NoA when its solid (and after running in parallel with NoA).
willow_hiller
Well-Known Member
Teslascope did answer the question. I think it's a possibility, but their understanding is that this is just draft text and not really indicative of what we'll eventually see as public release notes:
diplomat33
Average guy who loves autonomous vehicles
Is this a broadly agreed on community consensus?
I really thought they were trying to end-to-end one single neural network where it takes in video (and other sensor inputs) and spits out a path and velocity plan.
A previous post makes the point the if they're reusing the existing perception layer they could be building V12 more incrementally.
Another downside I see is that separate layers for perception and planning implies a connection between them in the middle (think "public API" in computer engineering terminology). I think people have used the term "Vector Space" for the output from the perception stack (locations of vehicles, locations of VRUs, lane lines, driveable space, signs, signals, metadata about these objects, etc).
The downside I see is these are all human picked and curated concepts. By building planning on top of these the planner can only know about the concepts that human decided to build into the perception layer. Combining all of it into one neural net, end to end, eliminates the need to even pick what stuff is represented in the intermediary "Vector Space", or how it's represented.
willow_hiller
Well-Known Member
It's still possible to connect intermediate layers of one module to another, prior to the high-dimensional vector space being reduced down to the human-interpretable concepts.
The only requirement I see for a neural network to be truly end to end is that the back propagation during training is done in one continuous step. Tesla could hold the current final layer of the perception stack constant and separate, for visualization purposes, and train the rest of the network end to end.
But throwing out the training efficiently of actual end to end for the sake of visualizations would be throwing the baby out with the bathwater.
I see, sort of. This is where my lack of ML experience isn't helping...
How sort of information is passed around in this architecture? What are the inputs and outputs of "layered" NNs like this? Are those inputs and outputs completely non-human readable? How do you train it then?
How sort of information is passed around in this architecture? What are the inputs and outputs of "layered" NNs like this? Are those inputs and outputs completely non-human readable? How do you train it then?
willow_hiller
Well-Known Member
If you have a couple hours and are interested in learning, this video from Karpathy is really useful for understanding the core mechanics of modern NNs:
It also comes with a GitHub repository: GitHub - karpathy/ng-video-lecture
I can't speak for C++ implementations, but the inputs and outputs in Python are just a special type of numerical matrix, with defined dimensions. The architecture that defines the connections between the weights are modules, and the training is a process of initializing random weights, running the inputs through them to achieve a predicted output, and then updating the values of the weights from back to front based on how the predicted output compares to the desired output.
In a discrete human understandable NN, you feed data in one end and gave human labels at the other. Video -> {lanes, cars, pedestrians} that kind of thing. So it might go video, objects, path planning.
However, that intermediate labeling step throws out all the additional data in the feed and only leaves the next NN an N label vector of probabilities/values. Maybe there is a low confidence pedestrian object that is actually something important.
Connecting all the NN together without the downsampled intermediate stages let's it pull more information into the final output.
Result is less/no human friendly quantization layers.
If doing OCR.
NN1: circle, arc, line detection
NN2 or C++: based on NN1, determine letters
NN3 or C++: based on NN2 outputs, determine word
vs
NN: determine word
sleepydoc
Well-Known Member
My experience with AP and, more recently, FSD on the highway has been excellent. The only time I’ve had to take over for FSD is because it misses exits, and those are primarily cloverleaf exits which are somewhat difficult anyway because there’s such a short merge period. If you restrict it to Straight interstate driving with no exits then it absolutely could be ready.
I don’t think it’s ready for all highway driving but definitely some limited use scenarios, and even those would be huge.
Edit: my other minor gripe is it doesn’t reliably merge back over to the right lane after passing unless there’s a car on your tail. That should be a trivial fix for them to make, though.
If you feel it’s ready, why not try it with a blindfold and see if it still feels ready?
I love my AP but I also understand that I am driving.
Similar threads
- Article
