Tesla, Inc. (TSLA) has started rolling out the newest yet version of its self-driving software FSD Beta version 10.11. The automaker has provided extensive details on what the Tesla software development team has achieved with the release of this version.
One of the first Tesla owners reporting the availability of FSD Beta 10.11 was none other than the famous Whole Mars Catalog / Omar’s Twitter account. Omar is usually the one who gets the next FSD Beta update parallel to Tesla employees. He shared the release notes with the Tesla Community (below).
Currently, the pre-requisite to apply to be part of the FSD Beta testing team, a Tesla owner needs to have a safety score of at least 98.
Since last year Elon Musk shared a list of improvements Tesla had made in FSD Beta 9.2, the automaker now sends a similar list of items with every new release.
This time also, Tesla has shared some interesting progress and infrastructure upgrades that have come with the FSD Beta v10.11 release.
FSD Beta 10.11 Release Notes (And explanations)
Since this is a big update jumping from v10.10 to 10.11, we will need to discuss release notes items one by one.
Since most of the information Tesla shared in these release notes is highly technical and related to the most advanced AI technology available today. It is hard for even a guy like me with a background in software development to fully explain these.
However, till the point I understand this information, I will share it with you, and add if Elon or his team has any say on it specifically.
Bag of Points
Upgraded modeling of lane geometry from dense rasters (“bag of points”) to an autoregressive decoder that directly predicts and connects “vector space” lanes point by point using a transformer neural network. This enables us to predict crossing lanes, allows computationally cheaper and less error-prone post-processing, and paves the way for predicting many other signals and their relationships jointly and end-to-end.Tesla FSD Beta v10.11 release notes
I learned the difference between raster and vector graphics in the early 2000s when I was working on Adobe Photoshop and Adobe Flash for vector graphics-based web animation. In short, vector graphics are infinitely stretchable while raster has limits as they are based on pixels (referred to as Bage of Points here because of the AI machine learning algorithm involved).
This is a big change as Tesla Neural Net gets the ability to predict road lanes in an entirely new and expandable way.
Tesla AI Director and Autopilot Vision Team Lead Andrej Karpathy posted on Twitter that this particular point is his favorite in the FSD Beta 10.11 update.
“This enables us to predict crossing lanes, allows computationally cheaper and less error-prone post-processing, and paves the way for predicting many other signals and their relationships jointly and end-to-end,” Karpathy explained on Twitter.
“TLDR a GPT-like Transformer is now predicting the lanes and their connectivity. This “direct to vector space” framework allows predictions to be jointly coherent (due to sequential conditioning) and v easily used by planner (due to sparsity). Excellent work from the team,” he further added.
Refer to the video below for a deeper explanation by an expert.
Improved Turning and Merging
Use more accurate predictions of where vehicles are turning or merging to reduce unnecessary slowdowns for vehicles that will not cross our path.Tesla FSD Beta v10.11 release notes
In a previous testing video of FSD Beta v10.9, we witnessed some unnecessary slowdowns at an unprotected left turn. Tesla has most probably looked at this troublesome scenario closely. We will know if it is completely fixed as the new testing videos start coming in a few days (stay tuned).
Improved right of way understanding if the map is inaccurate or the car cannot follow the navigation. In particular, modeling intersection extends is now entirely based on network predictions and no longer uses map-based heuristics.
This improvement makes a positive step towards self-driving navigation where the maps are not available or map data is inaccurate. However, it is not clear how Tesla AI will decide if the map a car is currently following is inaccurate or not — perhaps, we will get an answer on this in future updates.
FSD Beta 10.4 was the first FSD Beta version that focused on vulnerable road user (VRU) detection and safety. Since then, VRUs are getting small updates.
In some scenarios (explained below), Tesla FSD Beta was falsely rendering pedestrians and bicycles, resulting in car braking. This false slowdown (phantom braking) is potentially a safety issue.
This is VRU-related phantom braking but there is another type of phantom braking that Tesla owners globally face when for example a Tesla car falsely predicts an oncoming truck.
In the FSD Beta 10.8 update, Tesla detailed its efforts to reduce phantom braking. Now VRU-based false slowdowns are being addressed in 10.11.
– Improved the precision of VRU detections by 44.9%, dramatically reducing spurious false-positive pedestrians and bicycles (especially around tar seams, skid marks, and raindrops). This was accomplished by increasing the data size of the next-gen autolabeler, training network parameters that were previously frozen, and modifying the network loss functions. We find that this decreases the incidence of VRU-related false slowdowns.
– Reduced the perdicted velocity error of very close-by motorcycles, scooters, wheelchairs, and pedestrians by 63.6%. To do this, we introduced a new dataset of simulated adversarial high-speed VRU interactions. This update improves autopilot control around fast-moving and cutting-in VRUs.Tesla FSD Beta v10.11 release notes
– Improved creeping profile with a higher jerk when creeping stance and ends.
-Improved control for nearby obstacles by predicting continuous distance to static geometry with a general static obstacle network.
-Reduced vehicle “parked” attribute error rate by 17%, achieved by increasing the dataset by 14%. Also improved brake light accuracy.
-Improved clear-to-go scenario velocity error by 5% and highway scenario velocity error by 10%, achieved by tuning loss function targeted at improving performance in difficult scenarios.
-Improved detection and control for open car doors.
Improved smoothness through turns by using an optimization-based approach to decide which road lines are irrelevant for control given lateral and longitudinal acceleration and jerk limits as well as vehicle kinematics.
-Improved stability of the FSD UI visualizations by optimizing the ethernet data transfer pipeline by 15%.
-Improved recall for vehicles directly behind ego, and improved precision for vehicle detection network.Tesla FSD Beta v10.11 release notes
It is particularly amazing that Tesla explicitly shares the % of improvements it has made in specific areas and scenarios.
The openness about the expansion of its AI-related physical infrastructure is also commendable and poses another challenge for the traditional automakers — can they be as open as Tesla, probably not in the foreseeable future.
- Tesla tests FSD V12 on Chuck Cook’s unprotected left turn (UPL) in Jacksonville, Florida
- FSD Beta 22.214.171.124 (2023.27.6) gets mostly positive reviews from testers, Tesla adds a new Data Sharing agreement prompt in the update
- Tesla update 2023.32.4 improves the clarity of Autopilot cameras, Model X Falcon Wing doors get new Service Mode release notes, more
- Tesla’s Dojo prediction by Morgan Stanley boosts TSLA stock price, report shows how critical the FSD supercomputer is for Musk’s tech empire
- Tesla FSD 12 wide release becomes imminent as Tesla removes ‘Autosteer on city streets’ feature from the Coming Soon list
- Elon Musk demonstrates Tesla FSD 12 in a live stream on X, not a single line of code is used to build this non-beta version of Autopilot
- Tesla adds the Traffic Flow speed matching feature for highways in FSD Beta v11.4.6 (video)
- Tesla starts rolling out FSD Beta 11.4.6 (2023.7.26) to non-employees, early testers report improvements in various scenarios