Foretellix and Voxel51 have announced a joint solution to optimize autonomous vehicle (AV) development. This integration combines Foretellix’s Physical AI toolchain with Voxel51’s visual AI data platform. The solution transforms real-world drive logs into high-fidelity 3D scenes and neural reconstructions for system validation.
Highlights
- End-to-End Integration: Combines data curation, audit, reconstruction, and scenario variation.
- High-Fidelity Simulation: Transforms real-world logs into 3D scenes for AI training.
- Data Quality Control: Automated detection of sensor misalignments and calibration errors.
- Scalable Validation: Generates synthetic data to fill gaps in operational design domains.
Optimizing Data for AI-Driven Architectures
AV developers are increasingly adopting end-to-end AI architectures, creating a demand for high-quality data. Real-world driving data is resource-intensive to collect and often lacks the specific complex events needed for robust validation.
Foretellix addresses this by utilizing neural reconstruction and synthetic sensor data generation. This process recreates real-world drives and introduces controlled variations within a simulation environment. This approach allows developers to fill critical gaps in real-world data coverage.
However, synthetic generation relies heavily on the quality of input data. Flawed reconstructions can result in significant engineering delays and wasted computational resources. Voxel51’s Physical AI Workbench mitigates this risk by ensuring simulations begin with accurate data.
The Joint Technical Workflow
The integration allows the Foretify Physical AI toolchain to connect with Voxel51’s FiftyOne Physical AI Workbench. This creates a production-grade workflow designed for rigorous AV stack testing. The process follows a specific sequence:
- Ingestion: The Foretify toolchain ingests real-world drive logs to evaluate operational design domain (ODD) coverage.
- Gap Identification: Foretellix identifies relevant snippets from large datasets to address specific coverage gaps using scenario-driven curation.
- Data Auditing: Voxel51 performs audit checks to detect pose calibrations, sensor misalignments, and coordinate conventions.
- Enrichment: Data is enriched with embeddings and scene understanding before processing with NVIDIA Omniverse NuRec 3D Gaussian splatting technology.
- Variation Generation: Foretellix generates controlled scenario variations and synthetic sensor data for closed-loop simulation.
- Inspection: The resulting datasets are analyzed within the FiftyOne environment for quality assurance.
- Final Validation: The Foretify toolchain re-analyzes the dataset to confirm that new variations have resolved the initial ODD gaps.
Enhancing Safety Through Data Integrity
The collaboration focuses on the reliability of the data underpinning Physical AI. Ziv Binyamini, CEO and Co-Founder of Foretellix, noted that safety is the foundation of Physical AI. He emphasized that the partnership unites real-world grounding with controllable scenario variation.
This integration aims to provide the automation required for next-generation autonomous systems. Brian Moore, Co-Founder and CEO of Voxel51, highlighted that data quality is mission-critical as data volumes explode. The joint solution helps engineering teams build systems with greater realism and efficiency.
About the Companies
Voxel51 is a visual AI data platform facilitating the development of reliable visual AI systems. Their flagship product, FiftyOne, allows organizations to visualize and curate visual data. Clients include LG Electronics and Microsoft.
Foretellix provides a Physical AI toolchain focused on the safety and validation of autonomous vehicles. The company enables OEMs to build AV technology with measurable safety and scale.
Sign up for our weekly emails to stay up to date on the latest news!
