NVIDIA Launches Physical AI Data Factory to Power Next-Gen Robotics and Autonomous Systems

NVIDIA announced the NVIDIA Physical AI Data Factory Blueprint at GTC today. This open reference architecture unifies how developers generate, augment, and evaluate training data for robotics and vision AI. The system aims to significantly reduce the costs and complexity associated with training physical AI systems at massive scales.

Scaling with Cosmos Foundation Models

The architecture relies on NVIDIA Cosmos foundation models to transform limited inputs into massive, diverse datasets. By generating synthetic data for rare edge cases, developers can simulate scenarios that are often dangerous or expensive to capture in the real world. This approach ensures that autonomous vehicles and robots are prepared for unpredictable environments before they ever hit the pavement. Cloud leaders including Microsoft Azure and Nebius are already providing the blueprint to their customers. Microsoft is integrating the toolchain with Azure IoT Operations and its new liquid-cooled hardware architectures. This allows developers to turn raw computing power into turnkey data production engines through agent-driven workflows.

Broad Industry Adoption

Industry giants such as Uber, Skild AI, and Teradyne Robotics have adopted the blueprint to accelerate their internal development pipelines. NVIDIA is also using the architecture to train Alpamayo, its own reasoning-based vision language model for autonomous driving. These collaborations demonstrate the broad utility of the system across the global physical AI ecosystem. Orchestration for these complex tasks is handled by NVIDIA OSMO, an open-source framework that manages distributed computing tasks. OSMO now integrates with AI coding agents to resolve infrastructure bottlenecks autonomously. This level of automation allows engineering teams to focus on model performance rather than manual data curation and resource management. The Physical AI Data Factory Blueprint is scheduled for release on GitHub in April. This release marks a significant step toward making high-fidelity simulation a standard part of AI development. It provides the foundation for the next generation of autonomous agents to interact safely and logically with the physical world.