VibeusNVIDIA has open-sourced OSMO, a Kubernetes-native orchestrator for... - Vibeus
NVIDIA has open-sourced OSMO, a Kubernetes-native orchestrator for physical AI training, simulation, and robot testing. Its workflow is configured in one YAML file.
Robot pipelines can use three compute tiers: GB200/H100 for training, RTX PRO 6000 for simulation, and Jetson AGX Thor for hardware-in-the-loop tests. OSMO treats them as backends of one control plane, so tasks name a platform instead of a cluster.
The README example chains Isaac Sim, PyTorch training on eight GPUs, and a ROS evaluation on Jetson, with results written to a named dataset. Platform selects placement, inputs define order, and outputs provide persistence.
The same file runs locally on KIND and on EKS, AKS, GKE, on-premises, or air-gapped clusters. Release 6.3.0 added a multi-provider deploy script with MinIO, Azure Blob, or S3 storage.
For production, OSMO uses NVIDIA KAI Scheduler by default, NVLink topology-aware placement, per-group timeouts, an RBAC sidecar, OAuth2 login, TLS at the gateway, and cloud workload identity. NVIDIA says it powers Project GR00T, Isaac Lab, Isaac Sim, and Isaac ROS internally.
