The vision-only-versus-mapped debate is a religious war in autonomy, but Tesla’s side of it is stated flatly in a federal filing. The FY2023 10-K, filed January 2024 and indexed by EdgarBeast at sec.gov, says the company has "expertise in developing technologies, systems and software to enable self-driving vehicles using primarily vision-based technologies."
It goes further: "Our FSD Computer runs our neural networks in our vehicles," with additional hardware in development to use the field data captured by the fleet to "continually train and improve these neural networks." That is the full camera-only thesis in two sentences — perception from cameras, inference on board, improvement from fleet data.
The honest framing of the bet is that it is a cost-and-scale bet, not a faith. Cameras are cheap and every Tesla already has them, so a vision-only stack that works scales to millions of vehicles at marginal sensor cost. A mapped-LiDAR robotaxi has richer sensing per vehicle but a far heavier per-unit and per-city cost. The filing is choosing the scale economics and accepting the harder perception problem that comes with them.
What the 10-K does not claim is that vision-only has solved that perception problem. "Primarily vision-based" and "continually train and improve" are present-tense statements of approach, not declarations of achieved autonomy. The edge cases — low sun, occlusion, ambiguous geometry — are exactly where a camera-only stack is most stressed, and the filing’s language leaves room for that being unfinished work.
Before the next FSD demo persuades you either way, this is the document to read. The FY2023 sec.gov text is Tesla committing, in writing, to the camera-first path and the fleet-learning loop that has to make it work. Fair to both sensor camps means quoting the bet exactly as filed. Discovery via EdgarBeast.