Tesla has given an update on its self-driving roadmap. It confirmed that its FSD v13 update has been delayed into next month and it makes more promises.
Last month, Tesla released its first ‘AI roadmap’, which consisted of a more detailed plan about updates and new features it plans to push through the (Supervised) Full Self-Driving (FSD) program.
It was a welcomed way for us to track progress better. Still, we have already noted several problems with it, such as the fact that Tesla uses “miles between necessary disengagement” as a metric to track progress, and yet, it refuses to share any miles between disengagement data.
Furthermore, Tesla claimed that it completed all its September goals on the AI roadmap despite the biggest one being v12.5, achieving “~3x improvement miles between necessary interventions,” and we have seen no evidence of that happening.
In fact, crowdsourced data, the best available data since Tesla refuses to share any, shows that v12.5 is actually a regression compared to v12.3, the last widely released FSD update:
That’s based on over 40,000 miles of v12.5 data.
In the original roadmap, Tesla stated these goals for October:
- Unpark, Park and Reverse in FSD
- v13 with ~6x improved miles between necessary interventions
With October coming to a close today, Tesla has released an update. The automaker says that it accomplished this on its AI roadmap this month:
- End-to-end on highway has shipped to ~50k customers with v12.5.6.1
- Cybertruck build that improves responsiveness
- Successful We, Robot event with 50 autonomous Teslas safely transporting over 2,000 passengers
“End-to-End network on highway”, which constitutes using neural net controls for highway driving instead of just city streets, was supposed to happen back in September, but instead, Tesla has only been able to push it to a limited number of customers in October.
As for what’s coming next, Tesla now says that “end-to-end highway driving” is coming next week, but only for Tesla owners with HW4:
“Full rollout of end-to-end highway driving to all AI4 users, targeted for early next week, including enhancements in stop smoothness, less annoying bad weather notifications, and other safety improvements.”
For now, HW3 owners are stuck with this:
Improved v12.5.x models for AI3 city driving
Tesla also notes that “Actually Smart Summon will be released to Europe, China and other regions of the world” without a clear timeline.
Finally, Tesla comes back to v13, which the automaker now claims will include these improvements:
- 36 Hz, full-resolution AI4 video inputs
- Native AI4 inputs and neural network architectures
- 3x model size scaling
- 3x model context length scaling
- 4.2x data scaling
- 5x training compute scaling (enabled by the Cortex training cluster)
- Much improved reward predictions for collision avoidance, following traffic controls, navigation, etc.
- Efficient representation of maps and navigation inputs
- Audio inputs for better handling of emergency vehicles
- Redesigned controller for smoother, more accurate tracking – Integrated unpark, reverse, and park capabilities
- Support for destination options including pulling over, parking in a spot, driveway, or garage
- Improved camera cleaning and handling of camera occlusions
Tesla added about v13 in its updated AI roadmap:
We have integrated several of these improvements and are already seeing a 4x increase in miles between necessary interventions compared to v12.5.4.
Interestingly, Tesla originally said it would be a “~6x improvement in miles between necessary interventions” in September and during its earnings call just last week, Elon Musk said it would be a “5-6x improvement.”
As for the new timeline for v13, Tesla is now targeting a wide release at the end of November:
This lays the foundation for the v13 series, and we are targeting to ship v13.0 to internal customers by the end of this week. Most of the remaining items are independently validated and will be integrated over November in a series of point releases. We are targeting a wide release with v13.3 with most of the above improvements for AI4 vehicles around Thanksgiving!
As you can see, it’s also again only for HW4 owners.
Electrek’s Take
My key takeaways here are: HW3 owners are screwed. Even though there was some progress with Elon finally admitting that HW3 might not be enough for unsupervised self-driving, we are still far from an actual resolution.
Doing a 2nd month long free test of FSD. Seems worse. Some disengagement reasons: Exceeding speed limit several times. 80 kmph zone to 50 zone, still 20 k over passing the start of the 50 zone. Left the roadway on rural 80k zone left turn. Did not see or react to construction worker hold a stop sign on two occasions. Failed to move ahead when roadway clear. This was at a T junction from stem of the T, turning right. Doesn’t stay mid lane. Often tires touch center line markings. Other non-disengagement issues Nova Scotia roads: Single lane highway changes to 2 lane highway. Car take left lane always. Doesn’t know difference between speed limit sign saying 80 ahead (using the arrow, not words) and 80.
HW3 development is now falling months behind HW4 without even a clear timeline for catching up, and Tesla is only talking about maybe having to develop a retrofittable computer for HW3.
As for the pace of improvement, v13 is now at least a full month behind schedule, and we don’t know if it will actually result in a meaningful improvement in miles between disengagement. Even if it does, Tesla needs about several more updates that bring order of magnitudes improvements in miles between disengagement.
Considering the fact that v13 is now scheduled to come 8 months after v12 and 4 months after v12.5, the pace of improvement looks to be nowhere near what Tesla needs it to be to achieve its stated goal of 600,000 miles between critical disengagement by Q2 2025.
At this stage, it remains a pipe dream, in my opinion, which is based on the best data available. I will change my mind as more data comes in.
