NVIDIA's Cosmos 3 Cuts Robot Test Drives by 90% With Synthetic Data

NVIDIA expanded its Cosmos platform at GTC Berlin with tools for generating training data, adapting world models, simulating autonomous systems and serving models in production. Cosmos targets physical AI systems, including robots, autonomous vehicles and industrial inspection equipment that must interpret sensor data and act in the real world.

World models predict how a scene will change in response to actions and sensor inputs. A driving system can use one to anticipate traffic movement, while a warehouse robot can test a planned maneuver before executing it. NVIDIA says Cosmos models now rank among the top five on six open-model leaderboards. The broader update connects raw video, synthetic data, model training, simulation and deployment in a shared toolchain.

Five blueprints tackle data bottlenecks

Cosmos Data Factory provides GPU-accelerated tools for labeling, curating, augmenting and evaluating multimodal sensor data. Its five cloud-native blueprints address common physical AI workloads:

The blueprints are available through GitHub and cloud providers Azure, Nebius and CoreWeave. A Volkswagen Group project led by Porsche uses Cosmos Data Factory with Voxel51 on Microsoft Azure to build a shared autonomous-vehicle data flywheel. NVIDIA says the group expects the system to reduce repeated physical test drives by as much as 90% and shorten development cycles by as much as 85%.

ANYbotics generated more than 1,500 synthetic video clips and validated them with Encord before training industrial inspection robots to detect liquid spills and other facility changes. Ultralytics retrained its YOLO detector on real images and synthetic defects, reporting better localization and recall.

Cosmos 3 moves toward deployment

NVIDIA also expanded the Cosmos Framework and released Cosmos 3. Developers can adapt compatible open models, including GR00T, Alpamayo, Qwen and Pi0.5, through fine-tuning and reinforcement learning. They can then apply distillation, which transfers behavior from a larger model to a smaller one, and quantization, which reduces numerical precision to lower memory and compute requirements.

MicroAGI reported that it post-trained Cosmos 3 on videos of people manipulating objects, then distilled the 32-billion-parameter reasoning model into an 8-billion-parameter Nano model without reducing accuracy on its evaluation. The smaller model lowers the resources required for inference on robots. Roboflow offers a graphical workflow for labeling and post-training Cosmos 3 models, with deployment to its cloud or on-premises NVIDIA GPUs.

Simulation closes the test loop

New Cosmos 3 Nano variants target repeatable simulations of multiview, multisensor scenarios that are dangerous, rare or expensive to capture. NVIDIA provides recipes for closed-loop evaluation of autonomous-driving stacks and bimanual robots. In a closed-loop test, each model action changes the simulated environment and influences the next observation, exposing errors that static datasets can miss.

Updated Cosmos 3 NIM microservices package the models behind deployable inference APIs. NVIDIA reports up to twice the inference speed without a quality reduction in its validated configurations, along with subsecond alerts for continuous video analysis. The Metropolis Video Search and Summarization 3.3 blueprint can process 40% more video streams on the same GPU, according to the company.

ADNOC uses Fogsphere’s MistIQ platform, built with Cosmos and Metropolis VSS, to detect safety hazards across 65 sites in real time. The deployment shows how the components can support continuous industrial monitoring, although performance will depend on the target hardware, video resolution and model configuration.

FOX connects models to factories

The FOX early access release provides a scaffold for building always-on factory manager agents and specialist agents for inspection, root-cause analysis and fleet management. FOX connects Cosmos and Nemotron models with computer-aided design, product lifecycle management and operational technology systems.

Siemens is integrating FOX-based agents into its Intelligence Center X enterprise system and estimates potential cost savings of 20%. Its Process Preparation X agents use Cosmos Data Factory to generate defect images for circuit-board inspection. NVIDIA says Siemens reduced calibration time for inspection processes by 40% at its electronics factory in Erlangen.

Mercedes-Benz is applying Cosmos 3 and Metropolis VSS to vehicle inspection and windshield assembly verification. Audi is combining an Omniverse battery-production digital twin with Agile Robots’ mobile humanoids to validate logistics workflows in Dassault Systèmes’ DELMIA software.

One pipeline, fewer handoffs

Physical AI development depends on representative sensor data, coverage of rare events and safe testing before deployment. Cosmos now places data curation, synthetic generation, model adaptation, compression, simulation and serving within one modular stack. That structure can reduce the custom integration work required to move from recorded video to a deployed robot, vehicle or inspection system.

Teams evaluating the stack will still need to test its claims against their own workloads:

• Data fit: Measure whether synthetic examples improve performance on real sensor data, especially for rare events and unusual operating conditions.
• Simulation fidelity: Check whether simulated physics, sensors and agent interactions reproduce the failures seen in production.
• Serving economics: Benchmark latency, throughput, memory use and quality on the GPUs intended for deployment.
• Integration cost: Assess how the blueprints connect with existing annotation systems, data stores, simulators and operational software.
• Benchmark scope: Review the tasks, datasets and configurations behind NVIDIA’s leaderboard and performance claims before comparing models.