Tesla works hard on unprotected left turns in FSD Beta 10.13, especially the Chuck Cook-style ones (Release Notes)

Tesla’s internal Autopilot testing team has started getting the next big FSD Beta update version 10.13 (2022.16.3.5). While non-employee Tesla vehicle owners registered on the automaker’s Early Access Program have yet to receive this new update as Elon Musk tweeted that it still needs a few tweaks.

However, the detailed FSD Beta 10.13 release notes have been leaked on Twitter (read below). These release notes contain a bundle of reporting by Tesla on the areas its Autopilot AI software team has worked on since version 10.12 was released about 2 months ago.

Tesla has solely focused on unprotected left turns (UPL) in this update as the UPL scenarios have been proving the most challenging for FSD Beta till now.

Another challenge has been roundabouts and Musk had previously said that Tesla will deeply focus on roundabouts in FSD Beta v10.13 but the release notes don’t mention this scenario. However, improvements in this area must have been integrated, we will know once the testing videos start coming out.

Chuck Cook-style Unprotected Left Turns

An interesting bit in these release notes is the mention of Chuck Cook who is a Tesla FSD Beta tester from Florida. Chuck has been testing a complex unprotected left turn that joins his neighborhood with the 6-lane Roosevelt Boulevard highway in Jacksonville, Florida.

It’s a fairly complex and potentially dangerous left turn that Elon Musk had previously mentioned in a tweet talking about the 10.13 update and the possible release of FSD Beta V11 in the near future.

10.13 goes to internal beta tomorrow, external next week. Should handle Chuck’s complex left turn.

Beta v11 hopefully end of next month, which just amounts to incorporating highway. Importance of “v11” has been reduced by all the 10.x releases. We’re already mostly at v11.
— Elon Musk (@elonmusk) July 14, 2022

The last time Chuck tested this unprotected left turn around 4 days ago (video below), FSD Beta 10.12.2 both passed and failed in multiple attempts.

The constant testing at the Chuck Chook left turn has provided Tesla with tons of testing data to teach Autopilot AI how to take unprotected left turns with the same pattern.

A fence at the end of the street partially blocks the view of the traffic that’s coming from the left and this makes this Chuck Cook UPL more difficult even for Tesla Autopilot — and even for a human driver.

A perfect creep forward is what is required to measure the amount of traffic coming from the left at around 50 – 60 mph (~80 – 100 km/h). Reading the FSD Beta 10.13 release notes, Tesla has improved the creeping movement to make these difficult left turns easy for Autopilot to complete.

Aerial view of the now famous Chuck Cook unprotected left turn in Jacksonville, Florida. Credits: Chuck Cook / Twitter.

In his latest test, when Chuck’s Tesla Model Y was waiting to take this specific left turn, an oncoming car on the left further blocked the left-side view in addition to the fence. It was enough for FSD Beta to take the wrong decision of moving forward, Chuck had to intervene and take control of the vehicle.

We will have to wait a bit for Chuck to try this tough unprotected left turn on FSD Beta 10.13 when he gets this over-the-air (OTA) software update. For now, let’s see how it works on FSD Beta 10.12.2.

Video: Chuck’s complex left turn and how FSD Beta is taking it until now.

FSD Beta 10.13 Release Notes

Improved decision-making for unprotected left turns using better estimation of ego’s interaction with other objects through the maneuver.
Improved stopping pose while yielding for crossing objects at “Chuck Cook style” unprotected left turns by utilizing the median safety regions.
Made speed profile more comfortable when creeping for visibility, to allow for smoother stops when protecting for potentially occluded objects.
Enabled creeping for visibility at any intersection where objects might cross ego’s path, regardless of the presence of traffic controls.
Improved lane position error by 5% and lane recall by 12% with a [obscured]
Improved lane position error of crossing and merging lanes by 22% by adding long-range skip connections and a more powerful trunk to the network architecture.
Improved pedestrian and bicyclist velocity error by 17%, especially when ego is making a turn, by improving the onboard trajectory estimation used as input to the neural network.
Improved animal detection recall by 34% and decreased false positives by 8% by doubling the size of the auto-labeled training set.
Improved detection recall of far away crossing vehicles by 4% by tuning the loss function used during training and improving label quality.
Improved the “is parked” attribute for vehicles by 5% by adding 20% more examples to the training set.
Upgraded the occupancy network to detect dynamic objects and improved performance by adding a video module, tuning the loss function, and adding 37k new clips to the training set.
Reduced false slowdowns around crosswalks by better classification of pedestrians and bicyclists as not intending to interact with ego.
Reduced false lane changes for cones or blockages by preferring gentle offsetting in-lane where appropriate.
Improved in-lane positioning on wide residential roads.
Improved object future path prediction in scenarios with a high yaw rate.
Improved speed limit sign accuracy on digital speed limits by 29%, on signs with difficult relevance by 23%, on 3-digit speeds by 39%, and on speed limit end signs by 62%. Neural network was trained with 84% more examples in the training set and with architectural changes which allocated more compute in the network head [obscured]