Porsche Leads Shared AV Data Flywheel Across Volkswagen Group Brands

At NVIDIA’s GTC Berlin conference on October 21, 2026, Porsche, Volkswagen Group brands, and CARIAD announced they are jointly building a shared data flywheel for autonomous-vehicle development, with Porsche leading the initial implementation and ecosystem integration together with Voxel51, NVIDIA, and Microsoft.

The Volkswagen Group is leading development of a physical AI stack that links fleet data with simulation, training, and evaluation. Voxel51 said the group’s connected fleet generates data equivalent to 11,000 average drivers per day, and the companies expect the collaboration to reduce repeated physical test drives by up to 90% and to shorten AI regression test cycles by up to 85%.

Roles Across the Flywheel

The companies describe the collaboration as a reference architecture for automakers building next-generation advanced driving systems.

Volkswagen brings its connected fleet, real-world driving data, vehicle platform, and system requirements. Voxel51 provides the data layer, powering six workflows in the flywheel: multimodal search and retrieval, automated labeling, scenario-based curation, dataset balancing, training data selection, and model validation. NVIDIA contributes data retrieval, generation, and simulation technologies, including Cosmos Data Factory for preparing, augmenting, and evaluating training data, Omniverse NuRec libraries for reconstructing real-world drives, and open models and tools for testing driving systems in simulation. Microsoft Azure supplies the cloud and compute, with the workloads running at fleet scale on infrastructure the group already uses.

NVIDIA’s own newsroom release from the conference states that Porsche is leading development of a data flywheel workflow across Volkswagen Group brands through Voxel51, using Cosmos Data Factory and Omniverse NuRec on Azure.

NVIDIA automotive vice president Rishi Dhall said in the announcement that simulation is turning real-world driving experience into a continuous cycle of learning and improvement, and that Volkswagen Group can turn fleet data into high-fidelity simulation scenarios using Voxel51’s data platform together with Omniverse NuRec and Cosmos.

Audits, Generation Gates, and Reconstruction Results

Virtual validation of driving-software releases is one of the flywheel’s first applications for Porsche and the Volkswagen Group. Engineers render reconstructed camera inputs from actual drives and check the simulated outputs against real-world measurements. Voxel51 said the aim is to make testing repeatable, shorten feedback cycles, and reduce dependence on physical test drives.

Before any recording reaches GPU reconstruction, Voxel51 audits the raw log across roughly 80 checks covering sensor integrity, timing, calibration, trajectories, actors, and coordinate frames. A technical deep dive by Dan Gural, Voxel51’s head of physical AI partnerships, states that the audit reads MCAP files directly and runs in three tiers: a manifest pass of about a millisecond per drive, an episode pass of about a second, and a decoded pass of a few minutes per drive. Each check records whether it passed, failed, or could not run, under a rule Voxel51 calls “silence is not a pass,” so a missing input cannot be read as valid data.

Gural describes the loop in five stages: find, check, generate, reconstruct, and evaluate. For the find stage, the platform integrates NVIDIA Cosmos Dataset Search, with clips embedded by Cosmos-Embed1 and stored in Milvus. An agentic reranker then has a planner model write a rubric from the search intent, uses Cosmos Reason to score each clip against it, and re-ranks the list.

Gural reports that Voxel51 validated the audit against seven public datasets: nuScenes, Waymo, PandaSet, Argoverse 2, MAN TruckScenes, aiMotive, and Lyft Level 5. The audit found 25 confirmed issues in the published data, 17 of which were fixed before reconstruction. On synthetic-data generation with Cosmos Data Factory, the deep dive reports that stored gate evidence showed 12 sampling steps matched 35 on accept rate, cutting cost from 6.80 to 2.43 GPU-minutes per clip; that an attribute gate passed only 30 to 38% of generated clips; and that recalibrating the judge raised the semantic pass rate from 3% to 91% with no additional GPU time.

On reconstruction, the deep dive reports a median peak signal-to-noise ratio of 29.05 dB across 25 public nuScenes scenes processed on NVIDIA H100 GPUs. It cites one nuScenes night scene that scored 28.59 dB on that metric but recorded a structural similarity score of 0.56 and a learned perceptual similarity score of 0.61, which Gural uses to show that a single image-quality number can pass a visibly poor result. Every reconstruction run enters a human review queue, where engineers accept it for simulation, send it back for retraining, or reject it, and completed scenes are packaged as USDZ files ready for simulators, NVIDIA AlpaSim among them.

Azure Deployment and Ownership of the Loop

Each stage runs as a job on Microsoft Azure, where the Volkswagen Group manages its fleet data. Azure provides the storage and compute for data processing, reconstruction, and simulation while keeping those workloads inside the group’s existing security, identity, and regional governance model.

Voxel51 said a common data model lets driving data and scenarios created in one project carry forward across vehicle programs, software releases, and brands, with every new drive expanding the scenario base and reconstructed scenes available for reuse in future development and validation cycles.

Voxel51 co-founder and CEO Brian Moore wrote that an automaker can own its development loop without building every tool itself: “The key to really industry-leading performance is to harness the data that your company has access to. You need to own the development and iteration loop, even if you don’t build every layer of the stack yourself.” Voxel51 describes its platform as extensible, letting teams bring their own models, define custom scenario logic, and connect existing data and infrastructure without dictating how the driving system itself is built.

Voxel51 recounted that it began working with Porsche in 2024, when Porsche tech lead Tin Stribor Sohn described the data challenge as one requiring automated labeling to understand the data, scenario detection to find what is new or important, and curation to remove redundancy. Citing a McKinsey survey, Voxel51’s post also states that one in four consumers globally would switch brands for better self-driving features.

Separately at GTC Berlin, NVIDIA’s release states that Volkswagen Commercial Vehicles is presenting its autonomous ID. Buzz AD vehicle platform on the NVIDIA Hyperion computing platform to support multiple autonomous-driving systems.