My Tesla Learned to Autopark in my Driveway (Video)

[anon]

I thought the whole point of neural networks was to learn from "experience"??

Tesla's overall system learns - each car sends relevant information back to Tesla's servers, and that helps to improve the neural network in future firmware releases (or other software updates?) for all cars. However, each individual car does not learn to improve its own ability to drive - this only improves by the release of a software update later.

 

[jgmartin]

They do. They don’t all do it in real-time though. In the vast majority of cases (including this one) that’s not actually what you want at all.

 

[davewill]

I thought the whole point of neural networks was to learn from "experience"??

Sort of. What you do is train them with lots of samples of the data and results you want from the NN. After you train, then you validate the newly trained model on a separate set of test data to be sure that you are getting the results and accuracy that you need for the task. For something like automotive, you then have to do a TON of real world testing. THEN you can deploy the new model.

Nobody in their right mind would let owners train their own cars and trust the system to actually improve. Number one, NNs aren't that trustworthy. Number two, you might be a bad driver who would give the autopilot bad inputs (Garbage in, garbage out).

 

[nvx1977]

Here is the way to think about Tesla's fleet "learning" in layman's terms:

There is a master brain at Tesla HQ that does all the learning. The cars collect specific situations that are deemed worthy of a learning moment for that master brain. Those learning moments are initially triggered by a set of preprogrammed situations (common one being a user-disengagement of AP). Then humans at Tesla review what cars send and further screens out situations not useful for learning. What remains is then presented to the master brain, and the brain makes a call based on its current intelligence. If it gets it right, it is TOLD it got it right. This allows the brain to reinforce that correct decision. If it gets it wrong, it is told it got it wrong. The brain then de-emphases that decision for the next time a similar situation is presented.

Once Tesla sees that the master brain is handling certain situations at a high degree of correctness, they then clone the existing intelligence of that brain and send it to the fleet via firmware update. Now all the cars with the new firmware have the same intelligence as the master brain.

The individual cars in the fleet do not have the capability to train themselves. When we as users see our cars repeatedly fail at a task, we can apply corrective action as much as we like, but the car will never get better from those actions alone. Our only hope is that our corrective measure is sent to Tesla, and there they deem it worthy of training the master brain to get better at that situation. Often times this is not easy because there's some other situation out in the world that is training the brain to react improperly to your situation.

 

[StellarRat]

OK, it's more like an ant colony. A collective intelligence with the queen (Tesla server) passing down the new best practices to the cars (the worker ants.)

 

[FloridaGary]

So for the back story, every night I back into my driveway in the exact same manner. After a few weeks of doing this, and seemingly after an update to 2019.12, the car began offering to auto park into my driveway at times. I didnt let it do it out of fear it would hit the curb but tonight I let it and it was FLAWLESS.

The car isn't even supposed to offer auto park if it doesnt have 2 other vehicles to park between, correct? I believe my car has learned my action.

It is HW3 2019.16.3 FYI. Unfortunately I didnt get the entire process on film because it didnt hit me to record at first.

Love your car's name....Major Debt. lol