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

FSD will not work without this

This site may earn commission on affiliate links.
I believe there is a critical piece missing from FSD. I have not seen evidence of this existing in Beta videos.

To work correctly, FSD must estimate the speed and distance of other vehicles. Unprotected left turns and stop sign pullouts cannot work without this. I think it is doable, but I don’t think Tesla is there yet. It will require several frames of video and quick computations. Think about pulling out from a stop sign onto a four lane highway going left. You need to determine speed and distance of at least four other vehicles to make that maneuver.

My Model 3 currently does not do this. Which is why it will lane change in front of a semi going 20 mph faster than me and wait till the last minute to slow down for stopped traffic.

FSD is complicated. Tesla can get there. Don’t underestimate how complicated it is for programmers and the amount of data needed.
 
I believe there is a critical piece missing from FSD. I have not seen evidence of this existing in Beta videos.

To work correctly, FSD must estimate the speed and distance of other vehicles. Unprotected left turns and stop sign pullouts cannot work without this. I think it is doable, but I don’t think Tesla is there yet. It will require several frames of video and quick computations. Think about pulling out from a stop sign onto a four lane highway going left. You need to determine speed and distance of at least four other vehicles to make that maneuver.

My Model 3 currently does not do this. Which is why it will lane change in front of a semi going 20 mph faster than me and wait till the last minute to slow down for stopped traffic.

FSD is complicated. Tesla can get there. Don’t underestimate how complicated it is for programmers and the amount of data needed.
It does that
1629476013331.png
 
FSD must estimate the speed and distance of other vehicles
At AI Day, Karpathy showed how accurate vision-only distance and velocity are already:
vision accuracy.jpg


Green is radar, blue is the new vision-only video predictions, and orange is the old / single-frame behavior. At least in this example, the video predictions of depth are within 1 meter, so roughly 3% error, but more important is the velocity predictions are extremely accurate. The max velocity difference seems to be about 0.4m/s (less than 1mph), and it's smoother compared to old behavior.

Additionally, radar can be problematic in certain scenarios, so while it was quite smooth in the above example and in many situations, there can be significant noise from metallic trash or confusing radar returns from overhead signs and manhole covers. But vision/video approach should be much improved in those situations while still maintaining very high accuracy.

Hopefully this translates to accurate predictions with non-front cameras especially for the unprotected left turn example you gave where most likely the pillar and/or repeater cameras will be used. There's no readily available radar data from the fleet, but extensive offline scene recreation on Tesla servers from video should be able to position vehicles from these cameras with high accuracy with the benefit of looking forwards and backwards in time.
 
With all of that amazing tech and progress, i hope all that means Smart Summon will no longer run into poles and up onto curbs.
With point cloud vision I wonder how well FSD will detect small objects or very fine wires & chains. Do Smart Summon or Summon even have any updated code yet?

Here are some of my questions whether their system can detect and avoid these things:

Chain on pole just behind car
maxresdefault.jpg



Small object
guy-wire.jpg


Fine wires
23842592450_ac70fcd64d_b.jpg
warren-accident.jpg


Protruding object
vwch9d6skpz61.png
6384b4e0-65cb-11eb-9fdf-be4a3e21a459
 
  • Like
Reactions: 2101Guy