NVIDIA introduces osmo: a unified workflow tool for AI robot development

NVIDIA has introduced OSMO, an open-source workflow orchestrator designed to streamline the development of AI-powered robots. The tool, built on Kubernetes, allows developers to describe the entire pipeline—from training to simulation to testing—in a single YAML file. This eliminates the need for separate infrastructure code across different stages of development. OSMO addresses the fragmented process of physical AI development, where training occurs on data-center GPUs, simulation on workstation-class RTX hardware, and deployment on edge devices like Jetson AGX Thor. By treating these stages as part of a unified control plane, OSMO simplifies the orchestration of tasks across different platforms. The system routes tasks based on platform names such as gb200, rtx-pro-6000, or jetson-agx-thor, without requiring explicit cluster references. The tool is Apache-2.0 licensed, includes Helm charts and containers on NGC, and offers a local quickstart for testing on a workstation. It supports serial and parallel task groups, Jinja templating, and priority-based scheduling with GPU resource sharing. NVIDIA highlights OSMO as a solution to the complexities of managing multiple tooling stacks in physical AI development.