Notes from GTC Berlin 2026

We are back from GTC Berlin. NVIDIA has started building the half of the synthetic data pipeline it used to leave to everyone else, the part that checks whether the data is any good. Three days, October 20 to 22, with Jensen Huang's keynote at the Tempodrom on the Wednesday. Before the event we listed the sessions we planned to cover, sorted by stage of the synthetic data pipeline. This is the follow-up, sorted the same way: what NVIDIA and its partners announced, and what it changes for anyone producing training data for robots, vehicles and inspection systems.

The headlines were not about data

Most of the keynote went to infrastructure and agents. DSX POD is a reference design for Vera Rubin AI factories that partners build inside modular containers; NVIDIA says it reduces deployment time "from six months to as little as 14 days" on a site whose power, cooling and networking are already in place. The OpenShell Reference Architecture sets a secure runtime boundary around autonomous agents, built on the open source OpenShell software. And on October 19, the day before the conference opened, Chancellor Merz, the Federal Ministry for Economic Affairs and Energy and the Federal Ministry for Research, Technology and Space signed a Joint Declaration of Intent with NVIDIA to advance "cutting edge AI models for physical AI." No funding figure came with the declaration. Huang also announced investments in four German startups, part of what NVIDIA calls a nearly $10 billion commitment across Europe.

The training data news was quieter. Most of it sits in a Cosmos post published on keynote day and in the press releases around it. Read together, they touch every stage of the pipeline, and the last stage more than any GTC before.

Building the scene

The robotics release came with data, not only models. The Isaac announcement puts a SimReady Central library of "more than 2,000 simulation-ready robotics and factory assets" on Hugging Face, with contributions from FANUC, Universal Robots, Franka Robotics and others, alongside a Human 1M dataset of human activities with synchronized three-view video and 3D hand tracking (the full set is open to Cosmos Coalition partners). Isaac Lab 3.0 is generally available. The hardware moment was Agile ONE SN, a reference humanoid from Agile Robots preconfigured with Isaac GR00T on Jetson Thor: 1.40 meters, 31 degrees of freedom, due in early 2027, meant to cut setup time from weeks to days.

On the factory side, the Factory Operations Blueprint (FOX) is now in early access on GitHub, a starting point for factory agents that connect to Cosmos and Nemotron as well as CAD, PLM, simulation software, and IT and operational technology systems. Schaeffler is integrating digital twins of its humanoid actuators with Isaac simulation frameworks.

A shared asset library is a real saving: in our own work, scene assembly is where the hours go. But it holds catalogue equipment. The parts, materials and camera calibration that make a scene yours still have to be built and checked against the real thing.

Generating variation

Cosmos Data Factory, in NVIDIA's words "the tools and skills for labeling, curating, augmenting and evaluating data at scale for training physical AI," now ships as five open blueprints. One curates and labels driving data, then adds synthetic examples of hard-to-capture scenarios. One varies object appearance. One turns still images into synthetic video of rare or unsafe events. One changes weather and lighting in existing footage while keeping the scene structure. And one, Defect Image Generation, "generates printed circuit board, metal and glass defect images for inspection models." The code is on GitHub under Apache 2.0, and NVIDIA lists the blueprints as available through Microsoft Azure, Nebius and CoreWeave as well.

The partner numbers that came with it, in the partners' own terms. Siemens used the Data Factory tools to generate defect images for circuit-board inspection at its Erlangen electronics factory and reports a 40% cut in the calibration time of its inspection processes. ANYbotics generated more than 1,500 synthetic video clips to teach its inspection robots to recognize liquid spills, and validated the output on Encord's platform. Porsche, building a shared driving-data flywheel across Volkswagen Group brands with Voxel51 on Azure, expects up to 90% fewer repeated physical test drives and up to 85% shorter development cycles; those two are expectations, not results. Ultralytics retrained YOLO with real images and targeted synthetic defects and reports better defect localization and recall, without publishing the numbers.

Glass on that list caught our attention. Our poster at GTC in San Jose in March was on windshield damage, and in Berlin NVIDIA said Mercedes-Benz is applying Cosmos 3 and the Metropolis video search and summarization blueprint to build agents for vehicle inspection and windshield assembly verification. In our windshield work, the damage was only half the problem. The other half was the glass: transparent, reflective, curved and laminated, with damage whose appearance changes with the light and the camera position, and labels that have to agree with what that camera actually sees. A defect blueprint is a good place to start on that. It does not solve it for you.

Checking the result

This is the stage that moved most. NVIDIA released new Cosmos 3 Nano variants for simulation that "enable developers to test autonomous policies in repeatable, hard-to-capture, multiview multisensor scenarios," with code to train and evaluate your own policies for autonomous vehicles and bimanual robots before real-world testing. The Cosmos repository lists them by name, among them Cosmos3-Nano-Sim-Auto for closed-loop driving with multiple cameras and lidar and Cosmos3-Nano-Sim-Bimanual for closed-loop robot policies.

The Data Factory blueprints check their own output with Cosmos Evaluator, also Apache 2.0, which scores generated video for hallucination, object correspondence, requested attributes and physical plausibility. On the driving side, a new AlpaJudge model scores the reasoning in driving explanations, to "filter weak annotations and reduce manual review of training data."

The biggest item for us was on the Hyperion side. NVIDIA introduced Halos SAFE, the Safety Assurance Framework for Evaluation, with dSPACE, IVEX and TÜV Rheinland. It "connects simulation, software- and hardware-in-the-loop evaluation, on-road testing and fleet monitoring through defined scenarios, tests and measurable performance criteria" in order "to generate traceable safety evidence that supports deployment." Around it, Volvo Cars is adopting Hyperion and the Alpamayo Stack. Bolt selected the Alpamayo Stack for its European robotaxi program, starting with future Lucid vehicles built on Hyperion, and plans to gather driving data in Madrid in 2027, subject to regulatory approval. Alpamayo 2 Edge, a 2-billion-parameter reasoning model built on Cosmos, is designed to run at 10 Hz on DRIVE AGX Thor.

What changed since June

In June, writing about Cosmos 3, we said the generation half of a synthetic data pipeline had become dramatically cheaper and the validation half had barely moved. HUE, the benchmark that shipped with Cosmos 3, was the first sign of change. In Berlin, evaluation moved into the pipeline itself: the data factory scores its own output, a policy can be run closed-loop against a world-model simulator before a car or a robot leaves the lab, and the driving stack now has a framework whose stated purpose is traceable safety evidence. That is a real change, and we will use it.

What these tools check, though, is the generator and the policy against NVIDIA's own scenarios and models. Three questions stay with the team that ships the model.

Coverage. A defect blueprint generates defects. It does not know which defects your line produces, at what rates, under which lighting, or through which lens. Whether the generated distribution matches the one your system will meet in production is still measured against your own deployment.

Physical fidelity. Cosmos Evaluator's physical plausibility check is candid about its limits. Its documentation says the score "does not measure a specific physical quantity or guarantee that a video obeys physical laws." A generated image can pass it and still be wrong in the way that matters to your sensor: the refraction through laminated glass, the highlight on a machined part, the noise of a specific camera. Grounding that against the real thing is domain work.

Evidence. Which data, from which blueprint and checkpoint version, checked against what, with what result. Halos SAFE is the right shape for that record, and it is a driving program. For high-risk systems, the EU AI Act's Article 10 asks for that kind of record for training data whether the system drives or inspects, and the Isaac release named TÜV SÜD, Accenture and Reynolds & Moore as working to speed up certification for industrial physical AI. For an inspection model trained on generated glass defects, someone still has to define the scenarios and the measurable criteria, and keep the trail.

What we brought home

Each GTC this year has moved another stage of the pipeline into NVIDIA's open stack: the Physical AI Data Factory Blueprint in San Jose in March, a world model that reasons before it renders in Taipei in June, and in Berlin, evaluation wired into the data factory and the driving stack. We build on that stack, and what shipped in Berlin should make our generation and testing steps cheaper again. Showing that the data holds up outside the simulator, against the deployment domain and with the evidence kept, is still the part that decides whether a physical AI project ships. That is where datadoo works.

If you are trying the Defect Image Generation blueprint on glass or metal parts, or running a policy against the new simulation checkpoints and wondering how far to trust the result, write to hi@datadoo.ai. We would like to compare notes.