NVIDIA has released its OSMO platform under the Apache 2.0 license, making the source code available for public use. The system is designed to manage physical AI workflows by integrating training, simulation, and testing into a unified control plane.
The platform addresses the complexity of multi-tier AI development by treating data-center GPUs, workstation-class RTX hardware, and edge devices like Jetson AGX Thor as connected backends. Users register Kubernetes clusters via CLI to execute tasks.
OSMO enables users to define workflows based on platform requirements rather than specific cluster names. The system supports task prioritization, Jinja templating for parameterization, and automated resource management, including GPU borrowing across different pools.
Deployment is supported through Helm charts and containers available on NGC. Developers can also utilize a local quickstart option that runs the control plane on a workstation using KIND for initial testing and configuration.
