At the Tesla AI Day 2022 event, the electric car maker revealed some key statistics about the Full Self Driving (FSD) tech that is currently still in the beta testing phase. The company divulged that the number of FSD beta testers has gone up from 2,000 last year to roughly 1,60,000 users in 2022, despite a few regulatory hiccups and incidents that raised questions about its safety.
Tesla still hasn’t provided a timeline for when the FSD package will formally exit the beta phase, but it doesn’t seem too far off. In a TED interview this year, Musk claimed that the FSD system, which now costs $15,000, will most likely be out by the end of 2022 for all customers. There are also plans for a global rollout by the end of this year, pending regulatory approval, of course.
“I think from a technical standpoint, FSD Beta should be possible for a rollout of FSD Beta worldwide by the end of this year. For a lot of countries, we need regulatory approval. so we somewhat gated by regulatory approval in other countries,” Musk said during the Q&A section.
Elon Musk says they should be able to roll out FSD Beta worldwide by the end of the year (A big update coming next month). #FSDBeta #AIDay @Tesla pic.twitter.com/aS7H6N340g
— techAU (@techAU) October 1, 2022
Talking about the FSD progress so far, the team claims to have expanded the training infrastructure by up to 40-60%, which has vastly improved the pace of development. The company trained over 75,000 neural network models in a span of one year, which comes down to about one model passing validation and testing every eight minutes.
Going a bit deeper into the infrastructure details, Tesla notes that it maintains a cache of over 14,000 GPUs for the training process. Nearly half a million videos flow through the system each day, and so far, 1.44 billion frames have been used for training the self-driving system. The entire training dataset, on the other hand, now has over 4.8 million clips.
Tesla has also switched to an advanced dataset labeling system for training the FSD system. It is claimed to have reduced 5 million hours of manual labeling to just 12 hours of automated cluster work.
From its image-centric roots in 2018, Tesla’s team has now moved to an approach that is all about 3D spatial data with high scalability, while bringing the labeling time down from an average of 533 hours per clip to less than 6 minutes.
The FSD experts described how, in particular situations where the camera produces a confusing image, the software makes use of supplementary frames to reconstruct the scene. You can tune in to the official video stream for more in-depth technical details about the FSD upgrades here.
With all the infrastructure upgrades happening behind the scenes and fresh feature additions, it is unclear if the price of FSD will again be hiked prior to a wider rollout. Musk has previously claimed that it is “basically currently ridiculously cheap, assuming FSD materializes, which it will.”