USD Environments for Synthetic Data

Start with authored geometry, materials and semantic labels for your synthetic-data workflow. Define the render outputs, label coverage and runtime checks your robotics pipeline needs.

Authored kitchen environment illustrating a synthetic-data workflow

Define the Outputs Your Pipeline Needs

Scene geometry and authored labels can support these render passes. The illustrations below explain output types; availability, schema and coverage depend on your chosen capture pipeline.

RGB RenderPhotorealistic renders for perception and visual policy training

Depth MapDepth outputs for spatial reasoning; check units and renderer settings

Surface NormalsPer-pixel surface orientation for geometry-aware model training

Semantic SegmentationClass-label outputs based on your scene taxonomy and capture pipeline

Instance SegmentationObject-level labels; confirm ID consistency across frames and viewpoints

Material IDPer-surface material classification for domain adaptation

Illustrative render passes. Agree output formats and alignment checks for your pipeline.

Representative 3D Asset

Explore an authored chair and its material references. The metadata panels illustrate inspection fields, not a validation report. Use each delivered file’s specifications for your integration.

Drag to rotate. Scroll to zoom.

Polygons—

Vertices—

Material Slots—

UV Channels1

Bounding Box—

Multiple Output Types From the Same Asset

Explore different views of an authored asset, from contextual imagery to material details. Record camera, lighting and environment settings when capturing outputs for your own training pipeline.

Lifestyle

Product placed in a realistic 3D environment with scene composition, material interaction, and contextual lighting.

3D environment lighting
Scene composition
Material interaction
Contextual placement

These views illustrate how geometry, materials and lighting contribute to rendered imagery. A visual preview does not confirm physics behavior.

Fully 3D Environments, Not HDRI Backplates

Inspect scene geometry, lighting and scale before capturing training data. The example below illustrates useful metadata fields; confirm the values in the USD package you use.

Wing Chair Sceneenv_042Drag to rotate. Scroll to zoom. Drag with two fingers to pan.

Lighting Metadata

TypeIntensityColor Temp

Sun Light9.05800K

Area (Key Light)40 W6000K

Area (Fill Light)10 W6000K

Spatial Properties

Floor PlaneXZ (Y-up)

World Up Axis+Y

UnitsCheck metersPerUnit in the delivered stage

Scale ReferenceReal-world dimensions

Confirm stage units, object dimensions and placement in your target simulator. Required measurement tolerances belong in the project scope.

What to Include in Your Scope

Use this checklist to define your content requirements. Library specifications describe available files; custom deliverables are agreed with the team before work begins.

3D Assets

Review geometry, UV mapping and supplied collision data. Check scale, orientation and any LOD requirements for the assets you choose.

Materials (PBR)

Inspect supplied maps, shader support and texture resolution. Base color, normal and roughness are common inputs; exact content varies by material.

Render Outputs

Define the views, cameras, lighting and output passes needed by your training pipeline. Agree any separately authored imagery with our team.

Configuration Graphs

Inspect parent-child relationships and material bindings in scene content. Agree any additional configuration rules or variant mappings as part of custom work.

Rich Metadata

Choose the dimensions, categories, labels and physics fields your pipeline needs. Inspect available properties instead of assuming one schema across the catalog.

Review Criteria

Define acceptance criteria for scene content and captured outputs. Additional ratings, annotations and review records require an agreed project scope.

Plan a Repeatable Data Workflow

A useful workflow records source content, configuration and checks. These steps help scope authoring and downstream capture; they are not a promise of an included dataset or managed render service.

Source Inputs

Agree the available references, dimensions and permissions for the objects and environments in your task.

3D Modeling and QC

Geometry normalization, topology cleanup, and UV mapping to production standards.

Canonical Object Model

Structured entity relationships linking products, variants, components, and materials.

Render Recipes

Configurable lighting, camera presets, scene compositions, and output format definitions.

Synthetic Generation

Run your configured capture pipeline and review the outputs for the tasks you plan to train.

Quality Control

Review validation findings and test the agreed requirements in your target runtime.

Versioned Releases

Record scene, simulator and capture settings alongside outputs so your team can reproduce a run.

ContinuousAgree updates and delivery milestones with your team

Quality-ControlledDefine review criteria
for each deliverable

ExtensibleDiscuss additional content and output requirements

Schema and Data Structure

This illustrative schema shows how a pipeline can connect products, components, materials and review records. It is not the library API contract or a guarantee of fields included in every download.

ProductTop-level catalog entity with category, brand, and timestamps

FieldTypeDescription

idstringUnique product identifier

namestringDisplay name

categorystringHierarchical category path (e.g. furniture.kitchen.cabinetry)

brandstringManufacturer or brand name

created_attimestampRecord creation timestamp

updated_attimestampLast modification timestamp

Relationships

has many SKU/Varianthas many Component

sample_product.json

{
  "product": {
    "id": "prod_8f2a",
    "name": "Modern Kitchen Cabinet",
    "category": "furniture.kitchen.cabinetry",
    "variants": [
      {
        "sku": "MKC-WAL-36",
        "attributes": { "material": "walnut", "width": "914mm" },
        "dimensions": { "width": 914, "height": 1248, "depth": 904 }
      }
    ],
    "components": [
      {
        "name": "door_left",
        "type": "articulated",
        "joint": "revolute",
        "materials": ["walnut_pbr"]
      }
    ]
  }
}

Organise the Metadata You Need

The example below groups geometry, materials, lighting and camera settings. Agree required fields and export formats for your pipeline; actual asset metadata is specific to the delivered package.

GeometryTopology and scale

MaterialsBindings and maps

CameraCapture settings

LightingLight settings

EnvironmentScene state

ReferencesSource records

scene_metadata.json

{  "geometry": {    "vertex_count": 48200,    "face_count": 47800,    "bounding_box": [0.914, 1.248, 0.904],    "surface_area": 4.82  },  "materials": {    "slot": "door_panel",    "type": "PBR_metallic_roughness",    "roughness": 0.65,    "metallic": 0.0  },  "camera": {    "focal_length": 50,    "sensor_size": [36, 24],    "dof": { "f_stop": 2.8, "focus_distance": 2.1 }  },  "lighting": {    "key_light": { "type": "area", "intensity": 850, "temperature": 4200 },    "fill_light": { "type": "area", "intensity": 320, "temperature": 5000 }  }}

Regeneratable by Design

Keep source geometry and capture parameters together to make repeated renders easier to inspect. Recheck labels, collisions and task constraints when you change a scene or its configuration.

Source Asset

Camera

Environment

Configuration

Render A

Render B

Render C

Render N

Match Scene Content to Your Outputs

OutputScene geometryMaterial setupLabelsPipeline checks

RGB imagesRequiredRequiredTask-specificCamera and lighting

Depth mapsRequiredCheck rendererTask-specificUnits and clipping

Surface normalsRequiredCheck rendererTask-specificCoordinate space

Semantic segmentationRequiredTask-specificRequiredTaxonomy and coverage

Instance segmentationRequiredTask-specificObject identityID consistency

Material IDsRequiredMaterial bindingsTask-specificID-to-material mapping

Built for the Teams Training Physical AI

Perception Model Training

Use authored scenes to support RGB, depth and segmentation capture in your pipeline. Confirm annotation formats and label coverage for detection and depth tasks.

Object detectionInstance segDepth estimation

World Model Pre-training

Plan viewpoints and sequences for spatial reasoning. Validate camera settings and object identity across the frames your pipeline captures.

Multi-viewTemporal consistency3D understanding

VLA / Foundation Model Training

Start with scene content and inspect the physical interactions relevant to your task. Actions, trajectories and training labels are defined by your simulation pipeline.

Scene graphsAction annotationsSpatial reasoning

Vision Foundation Models

Explore authored scenes and product imagery for visual understanding. Review rights, image coverage and any additional metadata required for your intended use.

Image generationMultimodalPreference signals

Talk to Us About Your Data Pipeline

Tell us about your training pipeline and data requirements. We can discuss available content, required annotations and any custom authoring before agreeing the project scope.

Get in Touch