CarlK
Active Member
I’m pretty sure all autonomous vehicle testing on public roads must be reported to the DMV.
Not production cars used by customers I would think. Otherwise we will all need to report to DMV when Tesla turns on the FSD sometime this or next year. Either way Tesla reported zero miles in CA last year to DMV.
I don't think Elon's words about the fleet being trained by billions of miles driving is wrong at all.
Look at it this way: Training the network is not about feeding some Matrix-style computer with 1000's of terrabytes of plain well marked roads. That data is useless. Even if you had the ability and capacity to actually feed everything into the network, the results wouldn't be good because basically any uncommon scenario would be blurred out by all that common data.
What you are interested in is the edge scenarios. Every time the driver did something not predicted, every time you see something not recognized, generally every time you disengaged the system. Edge scenarios happen frequently in the beginning, but with every iteration of the NN it happens more rarely. At some point maybe you have just 1 abnormal situation every 1000 km, but to get that abnormal situation you still have to drive those 999 km first. You still count as 1000 km fleet experience as improving/training the system, but you just send the useful data needed for the next iteration. Not the flawless data, or data you don't need until a future iteration.
It's like a carpenter with 10 years work experience. If he mounts drywalls for 30 days he's not learning anything new until a difficult problem he needs to handle shows up. He finds a solution and next time maybe it takes 60 days for next problem because he has more experience. In the end despite low learning rate (transferred data), he still has 10 years of carpenting experience. What you'd do is ofcourse let the carpenter get a more challenging project, which is why Tesla also gradually increases difficulty rate of which rare events to actually ask the fleet for detailed info about.
It's absolutely correct but people just interpret it wrong. That's how it usually was. People who are ignorant is in no position to judge others.
In Karpathy's talks yesterday and last year he said they got normal cases pretty much down and do not need anything more for them. It's only edge cases that they still need more data to train the NN. Some situations of how these worked that he mentioned are:
-- Tesla will send request to cars for data of edge cases it needs. Cars will automatically send data to Tesla whenever they see those cases and ignore the rest.
-- Whenever there is a driver intervention it will take it as a special case and will send to Tesla for engineers to look at. Engineers will feed whatever needed back to the neural net to improve it.
-- Neural net in Tesla cars has the ability to observe minor movements (body languages) of outside cars and pedestrians to predict their intentions. These are always run in shadow mode even when the AP is not turned on and there is no action taken with the prediction. However the computer will compare the eventual true movement of cars and pedestrians to what computer has predicted and send data to the server for further training of the NN. This again is the first time I've heard anyone doing this and is an extremely powerful tool. Tesla can do this better than anyone because of all those Tesla cars driving in the real world.
So even that Tesla does not need all those data from billions of miles cars driven the more miles cars has driven it will still provide more those needed data of interest. Maybe another few billion or tens of billion miles later Tesla will acquire enough data of all needed edge cases. Other companies that are only running few millions of test car miles just don't have the ability to match this.
Last edited: