NVIDIA Open-Sources OSMO, Orchestrating Robot Training, Simulation and Edge Testing From a Single YAML

NVIDIA frames physical AI as a three-computer problem. Training runs on data-center GPUs such as GB200 or H100 clusters, simulation handles physics and sensor rendering on workstation-class RTX hardware, and deployment and hardware-in-the-loop testing run on edge devices such as Jetson AGX Thor, usually on premises. Each tier normally gets its own tooling, and the handoffs between them are where custom glue scripts accumulate. OSMO treats all three as backends of a single control plane. Each backend is a Kubernetes cluster registered through the command-line interface, and workflows never name a cluster; they name a platform such as gb200, rtx-pro-6000 or jetson-agx-thor, and OSMO routes the task to a pool that offers it.

The repository's canonical example chains three tasks by data dependency. A simulation task runs an Isaac Sim container on rtx-pro-6000; a train-policy task runs a PyTorch container on gb200 with eight GPUs, taking the simulation output as input; and an evaluate-thor task runs a ROS application on jetson-agx-thor, consuming the trained policy and writing results to a named dataset. Dependencies come from inputs, persistence from outputs, and placement from platform. The user guide covers serial and parallel task groups, Jinja templating for parameterized workflows, retry policies, and HIGH, NORMAL and LOW priorities with preemption and GPU borrowing across pools.

The same YAML can run on a laptop with Docker or KIND, or on EKS, AKS, GKE, on-premise and air-gapped clusters. Release 6.3.0 added a multi-provider deploy-k8s.sh script that provisions OSMO on Azure AKS, AWS EKS, microk8s or an existing cluster, with storage wiring for MinIO, Azure Blob, AWS S3 or a bring-your-own S3 endpoint. Developers can launch VS Code, Jupyter or SSH sessions on a remote GPU node, exec into running tasks, port-forward services and rsync files in both directions; version 6.3.0 added an osmo workflow rsync download command with a live progress bar.

OSMO uses the NVIDIA KAI Scheduler by default. Release 6.2.8 added NVLink topology-aware placement for multi-GPU tasks, and release 6.3.0 made exec_timeout and queue_timeout configurable per group, so a stalled simulation group no longer terminates sibling training groups. The project describes content-addressable datasets with deduplication that it says can cut storage by 10 to 100 times. The standalone osmo dataset CLI and the /datasets API were deprecated in 6.3.0 and are slated for removal in 6.4, with workflow-managed dataset outputs as the replacement.

On security and identity, OSMO has shipped an RBAC authorization sidecar, OAuth2 proxy integration with device-code login and identity-provider user mapping since 6.2.8. Release 6.3.0 added TLS termination at the Envoy gateway and cloud workload identity through Azure Workload Identity and AWS IRSA or Pod Identity, so services no longer mount storage keys as Kubernetes Secrets. Release 6.3.1 tightened the default osmo-user role to the default pool.

The repository also ships an AGENTS.md file, a skills directory and an MCP deployment guide. At GTC 2026, NVIDIA said OSMO integrates with Claude Code, OpenAI Codex and Cursor so coding agents can submit, monitor and debug pipelines. MarkTechPost reports that the orchestrator has been used with GR00T, Isaac Lab, Isaac Sim and Isaac ROS, and that Azure and Nebius integrations exist. The latest release listed is 6.3.1 from June 2026.