Nvidia Cosmos Adds Synthetic Data Factory
Nvidia on Oct. 21 expanded its Cosmos world-model platform at GTC Berlin with Cosmos Data Factory, five cloud blueprints for synthetic training data, agent-skill training recipes and faster Cosmos 3 NIM microservices. A Porsche-led data pipeline built with Voxel51 is expected to cut repeated physical test drives by up to 90%, Nvidia said.
Nvidia on Oct. 21 expanded Cosmos, its family of open world models for robots and self-driving cars, with a new Cosmos Data Factory that labels, curates and generates training data. The company unveiled the tools at its GTC Berlin conference and said a Porsche-led data pipeline using them is expected to cut repeated physical test drives by up to 90%.
Key takeaways
- Cosmos Data Factory ships with five cloud blueprints that create synthetic data for rare or unsafe events, from road hazards to circuit-board defects.
- Training, fine-tuning and deployment recipes are now packaged as agent skills that also work with non-Nvidia open models such as Qwen and Pi0.5.
- Porsche, ANYbotics, Ultralytics and ADNOC were named as early users. The efficiency figures come from Nvidia and its partners and have not been independently verified.
What is Nvidia Cosmos Data Factory?
Physical AI systems struggle with cases they rarely see, and filming enough of them is slow and sometimes dangerous. Cosmos Data Factory is a set of tools for labeling, curating, augmenting and evaluating training data, according to the Nvidia blog post announcing it.
Its GPU-accelerated Cosmos Curator now also handles multisensor inputs. The five blueprints are available on GitHub, Microsoft Azure, Nebius and CoreWeave.
| Blueprint | What it does |
|---|---|
| Autonomous Vehicle Dataset Generation | Curates and labels driving data, then adds synthetic hard-to-capture scenarios |
| Image Attribute Augmentation | Varies how objects look to improve image search |
| Event Video Generation | Turns still images into synthetic videos of rare or unsafe events |
| Defect Image Generation | Creates circuit-board, metal and glass defect images for inspection models |
| Video Data Augmentation | Changes weather and lighting in footage while keeping scene structure |
Agent skills, faster inference and simulation
Nvidia also repackaged its Cosmos Framework recipes for fine-tuning, reinforcement learning, distillation and quantization as agent skills. The framework supports Nvidia's GR00T and Alpamayo models alongside Qwen, Pi0.5 and other compatible Hugging Face models.
New Cosmos 3 Nano simulation variants let developers test driving and two-armed robot policies in repeatable, multicamera scenarios before real-world trials. Updated Cosmos 3 NIM microservices deliver up to twice the inference speed with no quality loss in validated configurations, Nvidia said. Its Metropolis video search and summarization blueprint, version 3.3, can monitor 40% more video streams on the same GPU.
Porsche's 90% test-drive claim
Porsche, working with data platform Voxel51 on Microsoft Azure, is building a shared autonomous-driving data pipeline across Volkswagen Group brands. Nvidia said it is expected to reduce repeated physical test drives by up to 90% and shorten development cycles by up to 85%.
Unite.AI reported further detail: the connected fleet produces data equivalent to about 11,000 average drivers a day, and tuning cut synthetic clip generation from 6.80 to 2.43 GPU-minutes per clip.
Who else is using Cosmos
ANYbotics generated more than 1,500 synthetic clips to train inspection robots to spot liquid spills. MicroAGI distilled a 32-billion-parameter Cosmos reasoner into an 8-billion-parameter Nano model with no loss in accuracy, Nvidia said. Energy company ADNOC uses a Cosmos-based safety platform from Fogsphere across 65 sites.
Context: the race for world models
World models, which simulate how scenes change over time, are a crowded field, and general-purpose models are also being tested physically, as in our report on how GPT-6 Astra finished a real cone course at 0.94 mph. Nvidia says Cosmos now ranks in the top five on six open model leaderboards.
What's next
The key test is whether synthetic data holds up on real roads and factory floors. Nvidia presents Porsche's figures as expected gains, not measured results, and has not published independent benchmarks for the new blueprints. The Nano variants are on Hugging Face and the NIM microservices on Nvidia's NGC catalog.
FAQ
Is Cosmos Data Factory free to use?
The blueprints and Cosmos Framework are published on GitHub. Running them requires GPU compute, either your own or on clouds such as Azure, Nebius or CoreWeave. Nvidia did not list pricing in its announcement.
Does Cosmos only work with Nvidia models?
No. Nvidia says the framework supports Qwen, Pi0.5 and other compatible Hugging Face models alongside its own GR00T and Alpamayo.
Has Porsche cut test drives by 90%?
Nvidia did not report a measured result. It describes "up to 90%" fewer repeated physical test drives as an expected outcome of the shared data pipeline.
References
- NVIDIA Cosmos Opens New Tools to Train, Evaluate and Deploy World Models — NVIDIA Blog (Oct. 21, 2026)
- Porsche Leads Shared AV Data Flywheel Across Volkswagen Group Brands — Unite.AI (Oct. 21, 2026)
- A World Model Community Guide to NVIDIA GTC Berlin 2026 — NVIDIA Developer Forums
Editor responsible for AI Newsway. Writes about AI, developer tools, cloud and productivity technology.
Comments 0
Comments are currently disabled.