Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

FSD Error Reporting/Feedback Best Practices

This site may earn commission on affiliate links.
Does anyone know how to best send feedback to Tesla via the video button on FSD drives? I know they've said it's helpful to them if you note the time, the circumstances, write a bit about what happened, etc., but I believe this is impractical for most people, so assume you can only hit the video record button. In that case, what's the optimal way to provide feedback?

I have been "reporting" the same behavior on the same road for a while. Do you think there is value in this, or do you think once or twice for each situation is enough? Do you report when FSD picks a wrong lane? When it phantom brakes? When navigation disagrees with where the actual roads are?

What feedback has Tesla given us about what helps them best?
 
Some of the early beta tester YouTube videos have mentioned that Tesla told them the car only has storage space to save roughly 5-6 snapshots before uploading the data after the drive (I assume only on Wifi). They also said it wasn't a hard number of clips, so I'm guessing that means there is logic trying to decide how much video history to save in the clip based on what happened before the button was pressed, or the clips are compressed to a variable compression ratio depending on video content, or the storage is also used for something else that sometimes needs more or less space.

However, Tesla has been conspicuously silent on answers to these questions. We don't have any reliable info to tell us if the 5-6 snapshot limit even still applies (maybe snapshots are shorter now and take up less space?).

My speculation is that Tesla's silence indicates one or more of these things is happening (in order of how likely I think each is, most likely at the top):
  1. They need a good random and un-biased data set to train the neural nets, to avoid over-fitting on scenarios we think Tesla wants us to save, so they don't want to bias the data collection by giving any guidance
  2. They are already getting what they need, so why bother spending time to write up a communication to us all when they could spend that time continuing to make the software better
  3. There is some IP or privacy law liability concern with Tesla asking customers to send certain types of videos
  4. The button is a placebo. It was deployed to make use feel like we're helping, but it doesn't do anything
I'd tell you how I treat the button, but just in case #1 is correct above, I don't want to introduce bias into their data set.

As for emails, my philosophy is to send emails about snapshots only when I think it might not be obvious why I took the snapshot, or if it was a dangerous failure mode (running red lights, trying to pass a stopped school bus at full speed, etc.). When something like that happens, I take a mental note of the time and location, and jot it down in a notes app on my phone as soon as I finish the drive, then later I'll send the email whenever I get around to it. With maybe 15,000 or so people in the beta program now, I'm sure they get tons of emails, so I try to keep my emails clear, concise, and well formatted (I list the Date, Time, Location, and if a snapshot was saved at the top of the email). I usually show the location both by longitude/latitude coordinates and nearest address, because I'm not sure which they can use more easily.
 
  • Like
Reactions: beachmiles
The optimal way for Tesla is to do what they asked. It is a beta that I am participating in not the other way around. So I participate the way asked.

I press button. Make note of time. Send email with Vin, date, time, and with slight snippet(like one sentence) of incident. Have it already in a pre populated email to the team. I send two to three emails per drive post drive. Some of it is the same thing over and over. Angles and conditions are never the same. If it is too much of a burden for some I guess one could wait. Most here seem like they want to participate.

Tesla doesn't give feedback unless it is in the notes or via Elon tweet. Anything else is pure speculation. I gave up a long time ago trying to guess at what they wanted, and why they wanted it.

Does Tesla do anything with my emails? I don't know. I do see significant uploads on the three routers I visit many times a day that are in GB's.
 
  • Like
Reactions: Krock65
Given the 10.5 release notes explicitly mentions that auto-labeling has been taking place, I plan to use the snapshot button more liberally, and I won't hesitate to send multiple incidents at the same spot. There's still that conventional wisdom floating out there that the car can only handle 5 or so snapshots. So be it. If I log too many and only 5 get sent, no big deal. In the long run, most everything you encounter will get sent.


4. The button is a placebo. It was deployed to make use feel like we're helping, but it doesn't do anything
Those of us that monitor network traffic know that huge amounts of data are sent when you use the snapshot feature. So it's certainly not a placebo from the car's perspective. Again with auto-labeling confirmed, odds are better that the video data you send can provide value, esp when a human doesn't have to sift through all the snapshots coming in (although I think that is still happening to some degree, along with auto-labeling).
 
We don't have any reliable info to tell us if the 5-6 snapshot limit even still applies (maybe snapshots are shorter now and take up less space?)
My guess is the video snapshot button records the last 30 seconds to upload later at home. This would allow for camera views before, during and after an incident as Autopilot team has previously recommended waiting 10 seconds before pushing the button.

Karpathy said at CVPR this year the initial Tesla Vision release trained on 1 million 10-second clips. I seem to recall a lot of upload earlier this year as multiples of ~400MB data bursts. With Tesla curated triggers, they can customize the saved clips to match what they specifically need -- here only 10 seconds, but FSD Beta snapshot button probably needs to be a bit more general purpose. AI Day presentation updated the radar-removal number to 2.5 million clips probably across 4 months of releases.

10.3 release notes mentioned +25k clips for vehicle semantics and +7k clips for static objects. 10.4 notes started mentioning autolabeling but no clips count. 10.5 notes say +15k clips for static objects but more importantly 165k clips for static world. So indeed it seems like Tesla is increasing capabilities to process clips, and this is likely also in parallel preparing FSD Beta 11 with autolabeling clips of moving objects.
 
I very recently (late September 2022) got the FSD beta update. I don't remember seeing anything from Tesla about suggested ways to report issues with FSD. So far I've been using it quite a bit in my wife's Model 3 and under some pretty wonky driving conditions here in Austin, TX, but there have definitely been a few quirks. I'm still using the voice command "bug report" feature but that may just be a black hole. Was there something in the latest FSD release notes that I just didn't pay attention to that has the "how to submit bug reports" or did they stop including that?
 
I very recently (late September 2022) got the FSD beta update. I don't remember seeing anything from Tesla about suggested ways to report issues with FSD. So far I've been using it quite a bit in my wife's Model 3 and under some pretty wonky driving conditions here in Austin, TX, but there have definitely been a few quirks. I'm still using the voice command "bug report" feature but that may just be a black hole. Was there something in the latest FSD release notes that I just didn't pay attention to that has the "how to submit bug reports" or did they stop including that?

The best way to report a problem is to disengage via steering wheel or brake.

Otherwise, if you want to, you can press the camera icon on the top middle of your screen. It'll say "autopilot clip recorded" or something like that.