Tesla used this clever workaround to train FSD for China's roads
Yesterday at 03:22 PM
An initial version of what is effectively Tesla’s Full Self-Driving (FSD) has started going live in China this week, and some early users are commenting on how well it’s performing in the country—despite only being available in North America previously.
In response to some of these reports, Elon Musk has revealed how the system has so quickly adapted to China’s road signs and traffic laws, noting that the company’s emphasis on video-based training makes the FSD-based system able to learn so quickly.
On Tuesday, users on X commented on how the FSD system had seemingly begun learning about driving in China ahead of time, and with “virtually no additional software development costs,” as Ray4Tesla points out. In response to him and James Douma, Musk said that Tesla deployed a pretty straightforward solution for pre-training FSD on China’s roads, utilizing videos like how its systems in North America are trained.
“We just used publicly available video on the Internet of roads and signs in China and used that to train in sim,” Musk says.
READ MORE ON FSD SUPERVISED: Tesla’s FSD rollout in Mexico is a bigger deal than it seems
Crucially, the process was similar to how Tesla’s camera-based, “vision-only” system works for FSD Supervised in North America, but instead using a simulator in China ahead of the access to real-world video. Now, with the service starting to roll out (despite not being termed FSD just yet due to some data privacy laws in China), Tesla should soon or already be able to start utilizing real-time footage to train the model, building on top of the simulated experiences grabbed from online.
Back in 2023, ahead of the launch of FSD beta v12, Musk and FSD engineers pointed out that the then-forthcoming version would play a key role in solving autonomy—especially due to its move toward a neural network that would be trained on video instead of on a “rules-based approach” and human-written code.
“FSD (Supervised) v12 upgrades the city-streets driving stack to a single end-to-end neural network trained on millions of video clips, replacing over 300k lines of explicit C++ code,” Tesla wrote in the release notes for the version.
Now well into v13 and having dropped the beta moniker nearly a year ago, Tesla has continued to tout its non-Lidar, camera-only based system, with executives regularly highlighting FSD’s scalability. While first reviews in China seem generally positive, we can definitely expect to see how the system fairs in new markets in the coming months, as multiple companies race toward commercial autonomy.
What are your thoughts? Let me know at zach@teslarati.com, find me on X at @zacharyvisconti, or send us tips at tips@teslarati.com.
Elon Musk discusses Tesla FSD rollout challenges for Europe and China
Need accessories for your Tesla? Check out the Teslarati Marketplace:
- https://shop.teslarati.com/collections/tesla-cybertruck-accessories
- https://shop.teslarati.com/collections/tesla-model-y-accessories
- https://shop.teslarati.com/collections/tesla-model-3-accessories
The post Tesla used this clever workaround to train FSD for China’s roads appeared first on TESLARATI.
