The IAWG World Generator has shipped a core generation system that takes a natural-language scene request all the way to a physically stable USD scene. It is more than a chat interface: scene intent, asset metadata, spatial solving, physics settling and deterministic composition are each built as a reusable, testable, deployable capability.
Walkthrough
IAWG world generator walkthrough
What has shipped
| Capability | What has shipped | Output |
|---|---|---|
| Natural-language scene generation | Turns a user request into a structured scene intent | An executable scene intent |
| Scene contract modelling | Defines objects, relations, constraints, placeholders and user parameters | A verifiable scene graph |
| Asset retrieval and matching | Finds candidate assets by category, description, tag and metadata | Resolved asset references |
| Placeholder-first workflow | Lets a missing object be expressed first and filled in from the asset library later | Traceable placeholders |
| Relation solving | Supports support surfaces, orientation, clearance and fixed world poses | Candidate object poses |
| Physics settling | Corrects interpenetration, floating objects and unstable placement | Settled poses |
| Deterministic composition | Instantiates assets under deterministic rules and writes the scene out | A USD scene |
| Run records | Keeps the state needed for replay, debugging and fault localisation | A replayable run record |
| Explicit failure | Returns a structured error at the validation, solving, settling and composition boundaries | A diagnosable failure result |
| Deployable shape | A containerised topology with a clear CPU/GPU boundary | A runnable deployment |
The generation pipeline
One sentence of natural language goes in; a physically stable, replayable, diagnosable USD world scene comes out:
- Submit a natural-language scene request.
- Turn it into a structured, verifiable scene intent.
- Retrieve and resolve object assets by category, description and tag; a missing object becomes a placeholder first and is filled in or generated once confirmed.
- Solve candidate layouts from the object relations and the asset dimensions.
- Settle the layout in physics, correcting interpenetration, floating objects and unstable placement.
- Instantiate assets under deterministic rules and write out the USD scene.
- Record the scene snapshot, the asset versions and the solving or error information, so a run can be replayed and a problem located.
Reliability by design
The system puts a "structured scene intent" between the natural language and the runtime execution: the scene is validated first, and only then enters solving, settling and composition, so the model never produces an uncontrolled result directly. The asset library ships against one geometric and semantic convention, which lets solving, physics settling and composition all share the same spatial semantics.
More important still isexplicit failure: the production path either returns a scene that has been settled in physics, or returns a structured error with its context. It never degrades silently when a critical stage is unavailable and hands back a scene that looks complete but is physically wrong.
Where it can be swapped out
Every critical stage is areplaceable module, decoupled behind a stable interface, so a project can swap the underlying implementation without disturbing the rest of the delivery:
| Replaceable component | What that buys you |
|---|---|
| Interaction entry point | Can sit behind a different frontend or conversational interface |
| Model provider | Can switch between LLM providers |
| Tool execution boundary | The tool set is pluggable and extends on demand |
| Asset catalogue | Can attach to a different asset library or retrieval backend |
| Pose solver | The layout solving implementation is replaceable |
| Physics settler | The physics settling implementation is replaceable |
| Asset generation backend | Can attach to a different asset generation approach |
What this delivers
IAWG has moved from a single "text generates a scene" trick to averifiable, replayable, extensibleworld generation system: natural language is pinned down by a structured intent, asset metadata connects semantic retrieval to spatial solving, layout solving is separated from physics correction, and a complete run record supports replay, debugging and fault localisation.