Foretellix and Voxel51 today announced a partnership that brings scalable neural reconstruction and simulation ready data pipelines to the fast shifting world of AI powered autonomous driving. The collaboration pairs Foretellix’s Physical AI toolchain with Voxel51’s FiftyOne Physical AI Workbench, creating what the companies describe as a “production grade, end to end” workflow that begins with real world drive logs and ends with high fidelity reconstructions, synthetic sensor data, and measurable improvements in AV safety coverage.
It begins inside Foretellix’s Foretify platform, which combs through vast real world drive logs to identify the blind spots in an AV’s operational design domain. Once those gaps surface, Foretellix isolates the relevant snippets and hands them to Voxel51. FiftyOne runs a battery of checks across pose calibration, sensor alignment, coordinate conventions, and annotation quality.
With the dataset stabilized, Voxel51 enriches it, adding embeddings, scene structure, and metadata that make it far more interpretable, before passing everything into NVIDIA’s Omniverse NuRec library for Gaussian Splatting reconstruction.
Foretellix then picks the world back up and begins bending it. Scenario variations layer on top of the reconstructed scenes, producing synthetic sensor data that can be run through closed loop evaluations or used to expand long tail coverage during training. The loop closes inside FiftyOne, where teams can inspect, debug, and compare real and synthetic data before Foretify reassesses ODD coverage to confirm that the original gaps have been filled.
Both companies describe the partnership as a response to the growing reality of Physical AI development. AI models cannot be safer than the data they’re trained on. The line between real world capture and synthetic variation is disappearing. Reconstruction pipelines like NuRec, and radiance field representations like gaussian splatting, increasingly serve as the connective tissue between the world we record and the worlds we need to simulate.
