NVIDIA Open-Sources OSMO for Physical AI Workflows
NVIDIA releases OSMO under the Apache-2.0 license, orchestrating physical AI training, simulation, and robot testing using a single YAML file.
Stock photo for illustration only, not from the actual event
- NVIDIA open-sources OSMO under the Apache-2.0 license with Helm charts and NGC containers.
- Unifies the three-computer problem of physical AI into a single control plane.
- Workflows reference specific platforms rather than clusters for automated task routing.
- Supports serial, parallel task groups, retry policies, and priority-based scheduling.
NVIDIA has officially open-sourced OSMO under the Apache-2.0 license, shipping with Helm charts and containers on NGC, along with a local quickstart that runs the complete control plane on a workstation using KIND for immediate deployment.
NVIDIA frames physical AI as a three-computer problem. Training takes place on data-center GPUs, while simulation, physics, and sensor rendering run on workstation-class RTX hardware. Deployment and hardware-in-the-loop (HIL) testing occur on edge devices such as Jetson AGX Thor, typically on premises. Historically, each tier relied on its own tooling, leading to an accumulation of custom scripts during handoffs.
Consolidating the traditionally fragmented architecture of robotics and physical AI development into a unified Kubernetes control plane significantly reduces the complexity of managing workflows across vastly different hardware environments, ranging from massive data centers to edge devices.
OSMO treats all three tiers as backends of a single control plane. Each backend operates as a Kubernetes cluster registered through the CLI. Workflows never reference a cluster directly; instead, they name a platform such as gb200, rtx-pro-6000, or jetson-agx-thor, allowing OSMO to route tasks to a matching resource pool.
Stock photo for illustration only, not from the actual event
The repository's canonical example illustrates three tasks chained together by data. Dependencies are derived from inputs, persistence from outputs, and placement from the platform. The user guide also covers serial and parallel task groups, Jinja templating for parameterized workflows, retry policies, and HIGH, NORMAL, or LOW priorities featuring preemption and GPU borrowing across pools.
Users can trigger workflow execution to observe how OSMO schedules the README example task by task, and inspect specific actions by clicking on any tier or step number.
Interested developers can check out the GitHub repository, documentation, releases, cookbook, and the official NVIDIA OSMO page for further details.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment