NVIDIA Robot Blueprints

NVIDIA is packaging robot training into an open “data factory” blueprint that turns cloud compute into pipelines for collecting, generating and grading robot data. (nvidianews.nvidia.com) The company announced the Physical AI Data Factory Blueprint on March 16, 2026, at GTC, and said Microsoft Azure and Nebius are the first cloud providers integrating it. NVIDIA said the reference architecture handles data processing, curation, synthetic data generation, reinforcement learning and evaluation for robotics, vision systems and autonomous vehicles. (nvidianews.nvidia.com) NVIDIA named FieldAI, Hexagon Robotics, Linker Vision, Milestone Systems, RoboForce, Skild AI, Teradyne Robotics and Uber as early users. In a parallel announcement the same day, NVIDIA said Azure developers including FieldAI, Teradyne Robotics and Hexagon Robotics, and Nebius customer RoboForce, are using the setup for synthetic data generation. (investor.nvidia.com, investor.nvidia.com) Robot learning runs on examples, the way a driving student needs miles on the road, but real robot data is slow and expensive to gather. NVIDIA’s pitch is to mix internet video, simulated environments and smaller amounts of real robot footage so developers can train faster before testing on actual machines. (nvidia.com, blogs.nvidia.com) That is the bottleneck NVIDIA is trying to address. The company said the real problem is not only a lack of data, but fragmented pipelines for processing, simulating and evaluating it, which is why the blueprint is framed as a factory rather than a single model. (blogs.nvidia.com, nvidianews.nvidia.com) The technical idea is layered training. NVIDIA’s recent robotics stack starts with world models that generate video-like simulations, then uses tools such as Isaac GR00T and Isaac Lab to turn those scenes into robot actions and test them in simulation before deployment. (nvidianews.nvidia.com, nvidianews.nvidia.com) Uber’s role points to one source of scale: existing driving video. NVIDIA said rich driving datasets from Uber, combined with Cosmos models and DGX Cloud, can help autonomous vehicle partners build training data and models more efficiently. (blogs.nvidia.com) Hexagon and Skild show the same strategy moving into industrial robots and general-purpose robot “brains.” NVIDIA has said Hexagon is building on its robotics platform, while Skild says robotics needs billions of training samples even though a single high-quality real-world demonstration can take minutes to collect. (nvidianews.nvidia.com, nvidia.com) NVIDIA said the data factory blueprint is expected to be available on GitHub in April 2026. If that release lands on schedule, the next test is whether developers can use the same pipeline to narrow the gap between robots that work in simulation and robots that keep working on factory floors, roads and warehouses. (nvidianews.nvidia.com, blogs.nvidia.com)