NVIDIA Open-Sources OSMO: One YAML Orchestrates Physical AI Training, Simulation, and Robot Testing

Details

Robot development spans three compute tiers: data-center GPUs for policy training, RTX workstations for physics simulation, and edge devices like Jetson AGX Thor for hardware-in-the-loop (HIL) testing. Each tier historically required its own tooling and handoff scripts. NVIDIA OSMO addresses this by treating all three as backends of a single Kubernetes control plane.\n\nOSMO is Apache-2.0 licensed, ships Helm charts and containers on NGC, and supports local quickstart via KIND. Workflows specify a platform (e.g., 'gb200', 'rtx-pro-6000', 'jetson-agx-thor') rather than a cluster name; OSMO routes tasks to matching pools. The canonical example chains simulation → policy training → robot evaluation, with data flowing between tasks as named datasets.\n\nVersion 6.3.0 adds multi-provider deployment scripts for Azure AKS, AWS EKS, microk8s, and on-premise or air-gapped clusters, with storage wiring for MinIO, Azure Blob, and AWS S3. Interactive development features let engineers launch VS Code or Jupyter on remote GPU nodes and rsync files bidirectionally. NVIDIA's KAI Scheduler handles GPU allocation with NVLink topology-aware placement for multi-GPU tasks.\n\nFor teams building physical AI products, OSMO could standardize the workflow layer the way Kubernetes standardized container orchestration—with the significant advantage of NVIDIA's hardware ecosystem backing.