Training the FSD system

[nvx1977]

Ok, so I let FSD do its thing. Clearly it needs a lot of polish. Some things I can give input to the system. For example if the car stops way too early for a stop sign, I can press the accelerator and supposedly "teach" it where to come to a stop (assuming Tesla is logging my actions). But if it takes a turn too forcefully, I can't retroactively "intervene" in any way except log the issue, which can be cumbersome, and at some point, it will feel like spamming them with too much information.

So I am curious: now that my car is part of the testing fleet, can I just not use FSD beta and just drive normally to train the car? I did not have to try hard to achieve the 100 safety score, and I feel like if Tesla just took 100% of my feedback in shadow mode, the NNs would learn much faster. I'm going to email the team and see if this is a viable option.

Also, I'm taking the beta testing seriously. My goal is to help get this system working well ASAP so everyone else can get the real deal, and if that means not using FSD beta is the more efficient way, so be it.

 

[ChimpledPot]

Training happens when they train the net in their servers and packaged into a release. Not in real time.

 

[nvx1977]

Training happens when they train the net in their servers and packaged into a release. Not in real time.

Yep I get that the car itself is not learning. I assume for now every intervention/disengagement is examined for labeling and training. I'm just curious if pure shadow mode training is a thing for FSD beta. Anyway I emailed them. will post if I hear back. I think it would be cool if you could push a "shadow mode" button for a complex intersection, drive it safely and properly, and have that data sent for labeling/training.

 

[dckiwi]

Yeah, they should have sent some guidance for reporting in that email. I feel bad if I report too many things on a single drive.

Anyone have any idea what is ‘report worthy’?

Some scenarios are no brainers but others aren’t - stopping short of an intersection, pausing or creeping far too long at an empty intersection, phantom braking, etc.

 

[jabloomf1230]

You can't report too many problems in a single drive. The car maintains a finite cache of reports per drive, which when that limit is exceeded, the newest report overwrites the oldest report. I won't speculate as to how many that is because it may have changed with this release. Chuck Cook mentions the reporting limitation in his videos. Basically, don't report every little glitch. Also, keep in mind that mapping errors can lead to problems. For example, the speed limits in my neighborhood are all wrong in the Tesla map database.

 

[Kingoglow]

Save you report button presses for unique driving situations and scenarios.

If you disengage FSD with the brake pedal or the steering wheel (forcibly taking over) the team at Tesla will automatically get a report of the data of what the car was doing.

Also when you intervene FSD by pressing the accelerator, the team at Tesla will also get those reports, but you wont be bumped out of FSD.

So intervene when you want to provide feedback on unsafe (too slow) acceleration. Disengage when the car will do something that will put you in danger or feels unsafe. Use your report button if you think that you come across a traffic situation that the car fails to understand, that is unique to your area (like unique traffic signs or complex intersections etc.).

 

[buz914]

My drive to work was much more exciting than I thought with FSD Beta. My biggest gripe is the phantom braking on clear roads(no oncoming cars or anything)? These sections of road used to be flawless on regular autopilot. Will see how it goes