· Research & technology· 日本語で読む
Volkswagen Group Builds Shared Platform Linking Driving Data to Simulation and Validation
Volkswagen Group is building a shared autonomous-vehicle (AV) data platform that links real-world driving data to simulation, training and evaluation. Porsche is leading the initial deployment, combining technology from Voxel51, NVIDIA and Microsoft. The company expects the platform to cut repeated real-world testing by up to 90% and AI regression testing time by up to 85%.
Key facts
- Volkswagen Group announced the construction of a shared AV data flywheel at NVIDIA GTC Berlin on Oct. 21, 2026.
- The company said the volume of data generated by Volkswagen Group’s connected-vehicle fleet is equivalent to that of about 11,000 average drivers.
- Porsche is leading the initial deployment and ecosystem integration.
- Voxel51 described a workflow that audits driving records against about 80 criteria.
- The company expects the initiative to cut repeated real-world testing by up to 90% and AI regression testing cycles by up to 85%.
What happened
Volkswagen Group announced Oct. 21, 2026, at NVIDIA GTC Berlin that it is building a shared autonomous-vehicle (AV) data platform linking real-world driving data to simulation, training and evaluation. Porsche is leading the initial deployment and ecosystem integration, with CARIAD, Voxel51, NVIDIA and Microsoft involved. The company said its connected-vehicle fleet generates a volume of data equivalent to that of about 11,000 average drivers.
Why it matters
Volkswagen Group outlined a system that links driving records collected on real roads to scenario extraction, reconstruction and virtual validation. Whether it can reduce repeated real-world testing will depend on the conditions under which the announced reduction of up to 90% can be achieved, and on how the quality of virtual validation is ensured.
Analysis
The main aim is not simply to create a place to store vehicle data, but to convert records of real-world driving into units that can be used in subsequent validation and training, creating a development loop. Even when driving records are collected, discrepancies in sensor timing or calibration can undermine the reliability of reconstructed scenes. The workflow’s roughly 80 checks before reconstruction suggests that quality control for simulation inputs, not just data volume, is built into the process.
The design connects Voxel51’s search and selection capabilities with NVIDIA’s reconstruction and data-generation tools and Azure’s processing infrastructure, and links model failures back to the original scenarios or records. In other words, the goal is to combine data collection, scene selection, virtual reproduction, software evaluation and selection of the next data set in a single workflow. Volkswagen Group says a common data model will allow scenarios developed for one program to be reused across vehicle programs, software updates and brands. If realized, a system for sharing validation results from individual vehicles across the group could become part of its competitive advantage.
[An article by this newspaper on NVIDIA presenting robotics research at CoRL on learning precision tasks through simulation](/en/news/nvidia-showcases-research-at-corl-on-robots-learning-precision-tasks-i-6e0769) covered learning precision manipulation by robots using simulation. The platform announced here focuses on validating autonomous-driving software. Although both use simulation, the former concerns learning task movements, while the latter concerns reconstructing real-world driving records and evaluating driving software. They should not be treated as the same implementation or technological achievement.
The structural focus is a division of labor in which Volkswagen Group provides vehicles, driving records and requirements, while connecting the specialized capabilities of its partners. Because the development loop includes record selection and quality checks rather than simply adding computing infrastructure, the extent to which data formats and evaluation criteria can be standardized across other brands and vehicle programs is likely to determine reusability. The announced reductions of up to 90% and up to 85% are estimates from the company; the announcement does not establish their scope, comparison conditions or achievement in actual operations. Going forward, it will be important to clarify the scope of the initial deployment, measured reductions, quality standards for reconstructed data and the handling of data shared across brands.
Details
Voxel51’s data platform handles searching and organizing driving records, extracting scenarios, creating data sets and validating models. NVIDIA provides tools including Cosmos Data Factory for preparing, augmenting and evaluating training data, and Omniverse NuRec for reconstructing real-world driving. Microsoft Azure provides the cloud and computing infrastructure needed for data processing and storage.
One initial use case is virtual validation that reconstructs real-world driving and compares the output of driving software with measured data. Voxel51 said it checks recorded data against about 80 criteria, including sensor alignment, timing, calibration and trajectories, then sends reconstruction results for quality evaluation. The company expects the initiative to cut repeated real-world testing by up to 90% and AI regression testing cycles by up to 85%.
Sources
Translated from the Japanese original: /news/volkswagen-group-soukou-deeta-wo-jidouunten-kaihatsu-ni-junk-faf641. The Japanese version is authoritative if they differ.