Omniverse IoT Connectors
NVIDIA Omniverse · OmniUI · OmniGraph · USD · Python · MQTT · CSV
- Ingest modes
- CSV (scheduled) + MQTT (stream)
- Binding
- USD attributes + OmniGraph
- Exposed as
- Python variables + scene state
Problem
A USD scene is a static description. The plant it represents is not — its readings change every minute. Without a bridge, a digital twin is a diorama: accurate at the moment it was authored, wrong from then on.
Constraint
The people authoring scenes are researchers, not engine programmers. A connector that requires custom scripting for every new data source would not survive handover. It had to be configurable from the Omniverse UI and work for both scheduled file drops and live streams.
Decision
I built two ingest extensions for Omniverse:
- CSV ingest — polls a source on a schedule and converts response rows into USD attributes.
- MQTT ingest — subscribes to a live topic and updates the same attribute surface as messages arrive.
Both expose controls through OmniUI and write through OmniGraph, with geometry transforms applied so scene elements move or change state as the underlying data changes. The same values are available as Python variables, so the extension works for both visual authoring and scripted setups.
Verification
The CSV path was demoed against historical weather data from Open-Meteo. The same reading was confirmed in two places: as a Python variable and as a USD attribute on the scene element. That double exposure is deliberate — it makes the binding checkable without reading the scene graph by hand.
Result
The connectors became the reusable data layer for the PV plant twin. Live IoT data now reaches the USD model without rebuilding the scene.
The transferable lesson: the boundary between a data stream and a scene graph is a data-modelling problem, not a graphics problem. Get the attribute contract right and the visuals follow.