Data-Centric Infrastructure for Physical AI

The workflow begins with the denoising of AV drive logs, which is fundamental to extracting the ground truth from data. This process enables both temporal scenario labelling and synthetic scenario design, allowing developers to accurately replicate specific scenarios or create controlled variations of them.

This foundational step is followed by data curation and warehouse exploration, where developers identify high-value driving segments to serve as the baseline for synthetic data generation. V&V engineers gain immediate visibility into the specific operational design domain (ODD) areas the system has encountered or tested across both virtual and real-world environments. The Foretellix solution organizes this data into a structured warehouse rather than a disorganized data lake.

Closing ODD Gaps with Synthetic Data Generation

Behavioral ODD coverage analysis and SDG are central to the Physical AI Toolchain. The toolchain identifies specific gaps in the ODD, allowing test engineers to design and edit new synthetic scenarios that directly fill the coverage gaps.

Safety and Scalability

Through the Foretellix Foretify scenario designer, integrated with NVIDIA Omniverse NuRec, engineers can modify actor behavior or add new actors in reconstructed scenes to analyze ODD completeness.  These scenes can be further enhanced and augmented with NVIDIA Cosmos.  Generating diverse scenarios at scale is key to validating that the stack is ready for safe deployment in complex physical environments.

Foretellix will be demonstrating the NVIDIA Alpamayo-based Reference Solution at CVPR at the NVIDIA Expo Theater on June 5th at 3:40 p.m. and at booth 826 during exhibition days.