Tesla used a transformer in the latest FSD update

Tesla released the 10.11 update for the Full Self-Driving (FSD) system.

FSD Beta 10.11 release notes. Fave item:
"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 neural network-transformer." https://t.co/Z6PpYrNiA1

— Andrej Karpathy (@karpathy) March 14, 2022

According to the company’s head of AI, Andrej Karpathy, one of the significant changes in FSD was the modelling of lane geometry from dense rasters (a set of points) to an autoregressive decoder. This allows the system to directly predict and connect lanes in ‘vector space’ point by point using a neural network-transformer.

"This enables us to forecast lane intersections, provides cheaper, computationally efficient and less error-prone post-processing, and also paves the way for joint and end-to-end forecasting of many other signals and their interrelations," wrote Karpathy.

Other changes in the update include fewer false slowdowns, improved understanding of lanes with an imprecise map, increased accuracy in detecting pedestrians and cyclists. Also improved smoothness of turns.

Earlier in January, California authorities said they intended to ban Tesla from testing Autopilot in the state without a permit.

In September 2021, researchers found that drivers with FSD enabled are more likely to lose vigilance on the road.

In the same month, the driverless Tesla drove through the centre of Kyiv thanks to an autopilot that leaked online.

Follow ForkLog AI on Telegram: ForkLog AI — all the news from the world of AI!