Nvidia open-sources the workflow tool it uses to train its robots
single source · 1 articles · release · confidence: medium · first seen 2026-09-14 08:04 UTC
Nvidia has open-sourced OSMO, the workflow scheduler it uses internally for its robot work, under Apache-2.0; the current release is 6.3.1. A single YAML file describes training, simulation and hardware-in-the-loop jobs — tests that run real robot hardware against simulated parts — and OSMO routes each to the right machine, from GB200 datacentre clusters down to Jetson AGX Thor boards on the robot. Nvidia names Project GR00T, Isaac Lab and Isaac Sim as internal users. The post gives no performance figures and no outside adoption.
What this means for you
If your robotics stack already runs Isaac Sim or Isaac Lab, this replaces some of the glue you wrote between training jobs and on-robot tests, and it is free to use. If you do not, there is nothing here to evaluate: no benchmarks, no hosted service, no support commitment.
Key facts
- ·OSMO is released under the Apache-2.0 licence, with 6.3.1 the latest release. source
- ·A single YAML file defines training, simulation and hardware-in-the-loop tasks. source
- ·OSMO routes each task to a compute tier, spanning GB200 clusters and Jetson AGX Thor devices. source
- ·Nvidia uses OSMO internally for Project GR00T, Isaac Lab and Isaac Sim. source
- ·OSMO is Kubernetes-native. source
What the sources say
- MarkTechPost — Short release note covering the licence, version number and hardware tiers the tool targets.
Sources
The original reporting. Follow these — they did the work.