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

FSD Beta 10.6

This site may earn commission on affiliate links.
I saw the following posted on reddit:

FSD Beta v10.6 Release Notes

• ⁠Improved object detection network architecture for non-VRUs (e.g. cars, trucks, buses). 7% higher recall. 16% lower depth error, and 21% lower velocity error for crossing vehicles.
• ⁠New visibility network with 18.5% less mean relative error.
• ⁠New general static object network with 17% precision improvements in high curvature and nighttime cases.
• ⁠Improved stopping position at unprotected left turns while yielding to oncoming objects, using object predictions beyond the crossing point.
• ⁠Allow more room for longitudinal alignment during merges by incorporating modelling of merge region end.
• ⁠Improved comfort when offsetting for objects that are cutting out of your lane.

Update from Elon:
 
Maybe one day an improvement will read:
“Keeping car centered or to right of center lane markings while driving through curves when oncoming traffic is present”.
How about.... "XX% improved apex calculation. Smoothed apex algorithm allowing safe use of full lane width, when oncoming traffic is not present, to minimize lateral Gs on occupants during curved maneuvers"
 
Maybe one day an improvement will read:
“Keeping car centered or to right of center lane markings while driving through curves when oncoming traffic is present”.
Yes! And this brings me to the discussion I'm wanting to have - just don't know where to post. For most Beta tests I've been involved in (Office products, CA products, Win XP thru 11, Visual Studio, SQL Server, Exchange Server, etc.) there has been an email address, web forum, or other facility for testers to provide qualitative feedback with context and vote or otherwise support other's posts so that the vendor (I'm seeing a Microsoft trend here) can get a sense of what's important to the testers/consumers. I don't see how intervening or sending video is communicating any context or qualitative assessment to Tesla so that they can prioritize issues. There are three or four key issues in the current Beta (10.5) that, IMO, have to be addressed for safety (the cheating to the left line being one of my major ones) and lots of little annoyances that I am sure they will get to someday, but I don't see how Tesla will be able to distinguish between these with the current reporting options at our disposal. Is there a website or email address we are supposed to be reporting issues to? Are there guidelines from Tesla on when we should be reporting issues and/or how often to use the camera icon to send video? It almost seems like this is not a Beta test program at all and more like just a way to collect video for training of the neural networks for the problems that they internally perceive and that they are not really interested in the opinions of their Beta testers.
 
⁠New general static object network with 17% precision improvements in high curvature and nighttime cases
@EVNow your examples from Wiki - MASTER THREAD: Actual FSD Beta downloads and experiences are both curving and at night, so this improvement for reducing false positives might result in less braking for you. Interesting that there was no mention of recall improvements (reduction of false negatives) as maybe this detection was already good enough at not missing objects that Tesla is classifying, or maybe something like the generalized static object voxel detection will cover the "unknown."

⁠Improved object detection network architecture for non-VRUs (e.g. cars, trucks, buses). 7% higher recall. 16% lower depth error, and 21% lower velocity error for crossing vehicles.
Compared to this item that does improve recall, correctly detecting these vehicles (and not relying on a general detection) is probably more important to be able to assign additional attributes, e.g., position/depth and velocity, to plan for how it will move vs just knowing "something is there."
 
  • Like
Reactions: EVNow
Yes! And this brings me to the discussion I'm wanting to have - just don't know where to post. For most Beta tests I've been involved in (Office products, CA products, Win XP thru 11, Visual Studio, SQL Server, Exchange Server, etc.) there has been an email address, web forum, or other facility for testers to provide qualitative feedback with context and vote or otherwise support other's posts so that the vendor (I'm seeing a Microsoft trend here) can get a sense of what's important to the testers/consumers. I don't see how intervening or sending video is communicating any context or qualitative assessment to Tesla so that they can prioritize issues. There are three or four key issues in the current Beta (10.5) that, IMO, have to be addressed for safety (the cheating to the left line being one of my major ones) and lots of little annoyances that I am sure they will get to someday, but I don't see how Tesla will be able to distinguish between these with the current reporting options at our disposal. Is there a website or email address we are supposed to be reporting issues to? Are there guidelines from Tesla on when we should be reporting issues and/or how often to use the camera icon to send video? It almost seems like this is not a Beta test program at all and more like just a way to collect video for training of the neural networks for the problems that they internally perceive and that they are not really interested in the opinions of their Beta testers.

Just my opinion -
The external beta testers are really there to collect the corner cases. The engineers that work on FSD use it regularly; as do a host of internal beta testers. Chances are any generalish feedback you’d like to offer is already well-known to the team and reading some long-winded email from us about how we’d like better lane bias or curve handling isn’t an efficient use of their time. What they DO want to hear from you is how FSD failed at that super unique intersection near your house or on a weird blind turn on your commute, and that’s what the snapshot button is for.
 
The external beta testers are really there to collect the corner cases… What they DO want to hear from you is how FSD failed at that super unique intersection near your house
I've been in the public beta since 10.2 and extensively using the video snapshot button for problematic intersections such as those that are very large (e.g., SPUI - single-point urban interchange) or angled/shifted or limited visibility (e.g., cresting or sloped). I actually go through the intersection from multiple directions and capture a video even FSD handles it fine as the 8 cameras combined with previous snapshots have a better view of the whole intersection for auto-labeling. With enough experience with beta, one can generally get a sense of which intersections will confuse FSD especially with the visualizations showing wrong or blurry lines.

With the 10.5 update, almost all of the problematic intersections I've retested are now handled fine in that FSD Beta can now drive to the correct destination lane (instead of towards oncoming traffic or incorrect turn). The visualizations in general are more confident and show things from further away as well, and I would guess this was indeed from sending back videos of these corner cases.

Unclear if this means it required 6 weeks to fix (10.2 to 10.5) or that Tesla only recently got all the auto-labeling of clips from the fleet working. But hopefully this cycle continues and speeds up especially as FSD Beta population continues to grow to find these unique situations.
 
@EVNow your examples from Wiki - MASTER THREAD: Actual FSD Beta downloads and experiences are both curving and at night, so this improvement for reducing false positives might result in less braking for you. Interesting that there was no mention of recall improvements (reduction of false negatives) as maybe this detection was already good enough at not missing objects that Tesla is classifying, or maybe something like the generalized static object voxel detection will cover the "unknown."
My current hypothesis is it is not phantom braking I'm experiencing but slow down for the bends. I'll test in the morning again - and with just AP to see if there are any differences.

On the false negatives / recall - my guess is they had erred on the side of reducing false negatives to a bare minimum - since obviously crashes are a lor worse than phantom braking. Now with improvements in reduction of false positives, all they need to do is to maintain the low false negatives.

BTW, when they say "17% precision improvements" what are they talking about ?

Lets take an example. The current error rate is 20%. They claim "10% improvement". Are they saying the error rate has improved to 10% (20% - 10%) or are they saying it has improved by 10% of 20% (i.e. 2%) - so the new error rate is 18% ? I'm guessing it is the latter - given the large improvements.