From Real Drives to Virtual Validation: Porsche's Workflow for Scalable AV Scenario Generation

Porsche is using an advanced workflow that integrates tools like FiftyOne, NVIDIA Omniverse NuRec, and NVIDIA Cosmos to transform raw sensor data into high-quality, reusable digital twins for autonomous vehicle (AV) training and validation. The process involves curating and auditing scenes to ensure data quality, reconstructing them into detailed digital environments, and expanding scenarios with novel views and annotations, significantly reducing costs associated with bad data training incidents. This approach allows for a tenfold increase in scenario diversity without additional fleet miles, focusing on creating valuable validation assets from curated data rather than amassing more raw footage. By leveraging this workflow, Porsche enhances its AV development by turning sensor captures into a robust library of scenarios that facilitate repeatable training, validation, and regression testing, thus streamlining and improving the quality of AV models.