NVIDIA open-sources OSMO: a single YAML Kubernetes orchestrator routing physical AI pipelines across GB200 training clusters, RTX simulation nodes, and Jetson edge devices.
September 14, 2026
brightray analysis
Summary
NVIDIA has open-sourced OSMO (Apache-2.0), a Kubernetes-native workflow orchestrator that solves the 'three computer problem' in robotics: training on data-center GPUs, simulation on RTX workstations, and HIL validation on Jetson edge devices all expressed in one YAML file. OSMO routes tasks by named platform rather than cluster, uses the KAI Scheduler with NVLink-aware placement, and ships with RBAC, OAuth2, and TLS. Version 6.3.1 adds multi-cloud provisioning, cloud workload identity, and per-group timeouts; the standalone dataset CLI is deprecated in 6.3 and removed in 6.4.
Why it matters
- Single YAML orchestrates training (GB200/H100), simulation (RTX), and edge HIL (Jetson AGX Thor) across heterogeneous Kubernetes clusters — eliminating per-tier glue scripts.
- KAI Scheduler default with NVLink topology-aware placement and GPU borrowing across pools; per-group exec/queue timeouts prevent stalled tasks from killing sibling groups.
- 6.3.0 adds cloud workload identity (Azure Workload Identity, AWS IRSA) so no storage keys are mounted as Kubernetes Secrets — significant security improvement for production deployments.
- Agent integration ships with AGENTS.md, skills directory, and MCP guide; Claude Code, OpenAI Codex, and Cursor can submit, monitor, and debug pipelines directly.
- Dataset CLI deprecated in 6.3 and removed in 6.4 — teams using the standalone osmo dataset commands must migrate to workflow-managed dataset outputs.
Community notes—
No notes yet — be the first.
See every signal in the Feed