StradVision's AWS Cloud-Based Physical AI End-to-End Pipeline Acceleration Case

Advancing autonomous driving AI requires vast amounts of road data, but there are limitations to directly collecting rare data specific to certain regions, such as cows on Indian roads, or hazardous data such as accidents.

To solve this, StradVision introduced a Data Flywheel structure spanning data collection, training, and validation. This system consists of SVGenFlow (synthetic data generation), SVDeepFlow (deep learning training), and SVSimFlow (simulation validation).

On the infrastructure side, the company adopted a hybrid architecture combining its Pohang data center with AWS cloud. AWS Direct Connect links the two environments to securely transfer large volumes of data, and when large-scale computation is needed, various GPU instances such as Amazon EC2 P5en, G6e, and G6 are flexibly utilized.

In particular, SVGenFlow uses technology that naturally composites virtual objects into real road footage, generating complex scenarios—such as animals on Indian roads or large trailers—in bulk, contributing to improving the performance of the SVNet model.