NVIDIA Open-Sources OSMO: One YAML Orchestrates Physical AI Training, Simulation, and Robot Testing
Robot developers do not have one compute problem. They have three. A policy is trained on GB200 or H100 clusters, tested in Isaac Sim on RTX GPUs, then validated on a Jetson mounted inside a real robot. Each tier has its own cluster, its own scheduler, and its own glue scripts. NVIDIA OSMO is NVIDIA’s answer to that fragmentation: an open-source, Kubernetes-native workflow orchestrator that lets a team describe the whole pipeline in a single YAML file and run it across every tier without touching infrastructure code.
Deployable? Yes. OSMO is Apache-2.0 licensed, ships Helm charts and containers on NGC, and has a local quickstart that runs the full control plane on a workstation with KIND.
The Three Computer Problem
NVIDIA frames physical AI as a three computer problem. Training happens on data-center GPUs. Simulation, physics, and sensor rendering happen on workstation-class RTX hardware. Deployment and hardware-in-the-loop (HIL) testing happen on edge devices such as Jetson AGX Thor, usually on premises. Each tier usually gets its own tooling, and the handoffs are where custom scripts accumulate.
OSMO treats all three as backends of one control plane. Each backend is a Kubernetes cluster registered through the CLI. Workflows never name a cluster. They name a platform (for example gb200, rtx-pro-6000, or jetson-agx-thor) and OSMO routes the task to a pool that offers it.
What a Workflow Looks Like
The repository’s canonical example is three tasks chained by data:
simulationruns an Isaac Sim container onrtx-pro-6000train-policyruns a PyTorch container ongb200with 8 GPUs, taking the simulation task’s output as input- Validation on
jetson-agx-thorcompletes the loop, enabling hardware-in-the-loop testing directly from the same workflow definition
With OSMO, the entire pipeline—from simulation to training to robot testing—is orchestrated through a single YAML file, eliminating the need for separate cluster-specific scripts and manual handoffs. This approach simplifies reproducibility and accelerates the development of physical AI systems.
via MarkTechPost