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

FSD Beta Videos (and questions for FSD Beta drivers)

This site may earn commission on affiliate links.
#652
JulienW said:
In all likelihood we are about to enter a slower release phase since the teams have started taking time off for the holidays and end of year.
Click to expand...
Well... they released multiple updates right before and immediately after Thanksgiving. Including one late on Sunday so... I think they are pressing hard for a wider December release.
 
  • Like
Reactions: pilotSteve, bd7349, aronth5 and 3 others
#654
powertoold said:
What is this magic we're seeing?
Click to expand...
Quite likely a good portion of the 2020.44.15.3 -> .4 update is from an improved neural network as opposed to writing new code for new functionality. Our car with 2020.44.15 has been uploading an average of about 1.5GB per day and up to 3.6GB. The snapshots seem to be in about 400MB chunks, which might mean the local storage can only save up to 9 all-camera clips between Wi-Fi uploads.

There are probably parts of the neural network training data that doesn't require humans to label, e.g., those that take main camera line predictions to label lines for other cameras as the car moves closer. So the data engine just collects data and retrains the network and verifies tests pass all automatically without much developer involvement. Then deploying the updated neural network is packaged, signed and deployed to FSD beta testers resulting in much improved behavior as it should have more confident predictions even for a given intersection it failed before, e.g., the path to take for a left turn.
 
  • Like
  • Informative
Reactions: pilotSteve, bd7349, powertoold and 1 other person
#658
The significant roundabout behavior improvement with 2020.44.15.3 might be from better software 1.0 code path selection. My guess is that the neural network makes several path predictions at all times with various likelihood scores, and for mini-roundabouts the highest probability path is to go straight as they can look like a regular 4-way intersection that happens to have a small circle in the middle. Potentially a fix would be to skip predicted paths that would cross a circle/road-edge, and this code didn't need to exist before as predicted paths wouldn't drive into curbs in most cases anyway (or predictions for curbs in intersections weren't good enough). Longer term, the neural network would be trained to have a lower score for going straight through a mini roundabout, but having the path skipping code is a reasonable safeguard.

More generally, the neural network making several path predictions is likely used to have software 1.0 code pick a path for turning through an intersection. Normally the highest score would be to go straight, so that path would need to be skipped if navigation is making a turn or needs to switch lanes. Similarly, we can see the visualized path jump around sometimes as the predicted path with the highest score keeps switching to others with similar scores. Tesla provided in their Q3 slidedeck a visualization showing more of the internals of multiple predicted paths:

multiple paths.png
 
You must log in or register to reply here.

Similar threads

S
Making left/right turns during FSD on city streets
Replies
6
Views
191
AI, Autopilot, & Autonomous/FSD
sradoff
S
B
FSD 11.4.9, First time use, subscribed to get v12 Beta
Replies
12
Views
1K
AI, Autopilot, & Autonomous/FSD
Tethlaaaa
Tethlaaaa
J
Does location matter for FSD performance?
Replies
8
Views
952
AI, Autopilot, & Autonomous/FSD
enemji
enemji
V
FSD Neural Network Training question
Replies
2
Views
283
AI, Autopilot, & Autonomous/FSD
zroger73
zroger73
F
I got HW 4.0 but NO FSD Beta
Replies
269
Views
30K
AI, Autopilot, & Autonomous/FSD
freek4iphone
F
Top
Back
Blog
New
Forum list
Marketplace
Menu