NVIDIA Open-Sources OSMO for Physical AI Workflows
NVIDIA releases OSMO, a YAML-driven orchestrator linking robot data, simulation, training, and hardware testing across mixed compute.
What happened
NVIDIA open-sourced OSMO, a workflow orchestration platform for physical AI and robotics. The release was reported at 08:04 UTC on September 14, within the edition window, and NVIDIA’s developer documentation describes the system as an open-source, agentic orchestrator for prompt-driven robotics development.
OSMO is designed to connect the separate computing environments typically used by robotics teams: large training clusters, simulation systems, and edge or hardware-in-the-loop devices. A workflow can be described in YAML, with OSMO managing datasets, dependencies, scheduling, and execution across cloud or on-premises infrastructure. NVIDIA says the platform can cover synthetic-data generation, distributed policy training, reinforcement learning, simulation evaluation, and tests on physical robots.
The platform’s distinctive feature is its attempt to make infrastructure itself legible to coding agents. OSMO includes agent context files and interfaces that allow an AI coding assistant to inspect workflows, query GPU capacity, monitor jobs, and reason about the development environment. That places orchestration between conventional cluster schedulers and the increasingly autonomous software agents used to build robot systems.
NVIDIA says OSMO grew out of internal workloads associated with projects including GR00T, Isaac Lab, Isaac Sim, and Isaac ROS, and that it has handled thousands of GPU-hours daily. Those claims indicate production lineage, but they are not independent performance measurements. Adoption will also depend on how well the system works beyond NVIDIA’s preferred stack and across varied Kubernetes and robotics environments.
Why it matters
Physical AI remains constrained by the messy handoff between data, simulation, training, and real hardware. An open orchestration layer could reduce repeated infrastructure work and make robot-development pipelines easier to reproduce. NVIDIA also gains a way to extend its software influence from model and simulation tools into the control plane where robotics projects are assembled.
Uncle Cat take
OSMO’s strategic value is not the YAML syntax; it is NVIDIA placing an agent-aware control layer above the expensive transition from Isaac simulation to real robot testing.