Nvidia has expanded its Cosmos platform for physical AI with new data-generation blueprints, training tools and simulation models, announced around its GTC Berlin keynote this week. According to Nvidia, the update gives builders of world models, which help robots and autonomous vehicles understand and simulate the physical world, a pipeline that covers data curation, model adaptation, policy testing and deployment. The company says Cosmos models rank in the top five on six open-model leaderboards. As with most vendor announcements, the performance and savings figures below come from Nvidia and its partners and have not been independently tested by us.

Cosmos Data Factory: build better training data

Training a world model starts with large amounts of multimodal, multisensory data, and collecting it in the real world is slow and sometimes dangerous. Nvidia’s Cosmos Data Factory brings together tools for labeling, curating, augmenting and evaluating that data. Its video curation component, Cosmos Curator, which is GPU-accelerated, now supports multimodal and multisensory inputs and processing blocks.

Data Factory now offers five cloud-native blueprints, available through GitHub, Microsoft Azure, Nebius and CoreWeave:

  1. Autonomous vehicle dataset generation: curates and labels driving data, then adds synthetic examples of hard-to-capture scenarios.
  2. Image attribute augmentation: varies object appearance to improve image search.
  3. Event video generation: turns still images into synthetic videos of rare or unsafe events.
  4. Defect image generation: produces printed circuit board, metal and glass defect images for inspection models.
  5. Video data augmentation: changes weather and lighting in existing footage while preserving scene structure.

Cosmos Framework and Cosmos 3

For training and adaptation, Nvidia extended the Cosmos Framework, an open and modular set of training and optimization tools, and packaged its recipes as agent skills for more open models. Developers and AI agents can now adapt open models such as GR00T, Alpamayo, Qwen, Pi0.5 and other compatible Hugging Face models through fine-tuning and reinforcement learning, then optimize them with distillation and quantization before deploying with high-performance inference. Alongside it, Nvidia introduced Cosmos 3.

As an example, Nvidia says robotics research lab MicroAGI post-trained Cosmos 3 on videos of humans manipulating objects and distilled a 32-billion-parameter reasoner into an 8-billion-parameter Nano model with no loss in accuracy. The company also says developers can label data and post-train Cosmos 3 on Roboflow and deploy to Roboflow-hosted or on-premises Nvidia GPUs.

Testing policies before the real world

New Cosmos 3 Nano variants for simulation let developers test autonomous policies in repeatable, multiview, multisensor scenarios that are hard to capture on real roads or factory floors. Accompanying code lets developers adapt the models to train and evaluate their own policies for autonomous vehicles and bimanual robots in interactive simulations before real-world testing. Updated Cosmos 3 NIM microservices deliver up to twice the inference speed with no quality loss in validated configurations, according to Nvidia, and continuous video analysis enables sub-second alerts. A new Metropolis Blueprint for video search and summarization, version 3.3, is said to let teams monitor 40 percent more video streams on the same GPU.

Early adopters and claimed results

Nvidia names several users, with results that are the companies’ own figures:

  • Porsche is leading a shared autonomous-vehicle data flywheel across Volkswagen Group brands using Cosmos Data Factory through Voxel51 on Azure, expected to reduce repeated physical test drives by up to 90 percent and shorten development cycles by up to 85 percent. These are expected, not measured, savings.
  • ANYbotics generated over 1,500 synthetic video clips to train inspection robots to recognize liquid spills and other changes between facility visits.
  • Ultralytics retrained its YOLO model with real images and targeted synthetic defects, improving defect localization and recall.
  • ADNOC uses Fogsphere’s MistIQ platform, built with Cosmos, to detect safety hazards in real time across 65 sites.

Why it matters

World models are a core ingredient of “physical AI,” the idea that robots and vehicles can learn from simulated experience instead of slow, costly real-world trials. Synthetic data helps when the events you care about, such as a spill or a near collision, are rare. The risk is that models trained mostly on generated data inherit its gaps, so teams still need real-world validation. Nvidia’s strategy is also clear: an open, modular stack that runs best on its GPUs and cloud partners, which strengthens its position as robotics and autonomous driving mature. For context on how robots learn, see our explainers on vision-language-action models and learning from demonstrations.

What to watch next

  • Independent evaluations of Cosmos 3 Nano for closed-loop policy testing.
  • Evidence that synthetic-data pipelines improve real-world robot and vehicle performance, not only benchmark scores.
  • Measured, not expected, reductions in test-drive miles from partners like Porsche.
  • How open models such as GR00T and Alpamayo perform when adapted with the new Cosmos Framework skills.
  • Competing world-model platforms from other labs and chipmakers.

Key takeaways

  • Nvidia added five Cosmos Data Factory blueprints for curating and generating physical-AI training data.
  • Cosmos Framework tools and recipes now work as agent skills across more open models, alongside Cosmos 3.
  • Cosmos 3 Nano variants enable repeatable simulation to test robot and vehicle policies.
  • NIM microservices claim up to 2x faster inference in validated configurations.
  • Partner results, such as Porsche’s up-to-90% fewer test drives, are expected or vendor-reported figures.

Frequently asked questions

What is a world model?

It is a model that learns how the physical world behaves, so a robot or vehicle can predict outcomes and be tested in simulation. Our glossary has a definition.

Is Cosmos open source?

Nvidia describes the platform as fully modular and open, with tools on GitHub and models you can adapt. Check each model’s license before commercial use.

Does synthetic data replace real-world testing?

No. It reduces the need for some real-world miles and helps with rare events, but teams still validate with real data and physical tests.