How to create SimReady assets without manual modeling

A SimReady asset is a 3D model packaged with everything a simulator needs to use it physically: mass, friction, collision geometry, semantic labels, material identification, and a valid USD layer structure. Building one manually takes a skilled engineer about four hours. That’s the bottleneck preventing robotics teams from running simulations at the asset diversity their policies actually need.

This guide explains what makes an asset “ready,” why the manual workflow doesn’t scale, and how AI-driven asset preparation collapses the pipeline to about five minutes per upload.

Defining SimReady

NVIDIA’s SimReady spec is the closest thing the industry has to a normative definition. A SimReady asset must include:

• Physics: PhysicsRigidBodyAPI, PhysicsMassAPI, PhysicsCollisionAPI applied to the right prims with reasonable values.
• Semantics: object class labels via SemanticsAPI so perception models can train against the asset.
• Material physics: PhysicsMaterialAPI carrying friction, restitution, and density per material region.
• Collision geometry: a separate, simulator-friendly mesh — convex decomposition for dynamic objects, primitives for navigation obstacles. Almost never identical to the visual mesh.
• Validated USD structure: passes the NVIDIA SimReady validator (no inverted normals, no orphan prims, all schemas correctly applied).
• Material bindings: visual materials properly bound to mesh subsets so the asset renders correctly in Omniverse/Isaac Sim.

A high-poly hero mesh from a 3D scan is not SimReady. A low-poly placeholder with correct physics and a valid USD structure is. The “ready” qualifier refers to simulator compatibility, not visual quality.

Why manual SimReady authoring doesn’t scale

A typical robotics team needs SimReady assets at three different scales:

Project type	Asset count	Manual effort (4 hrs/asset)
Hero asset for a demo	5–10	20–40 hrs (1 engineer-week)
Manipulation policy training set	200–1,000	800–4,000 hrs (1–2 engineer-years)
Warehouse digital twin	10,000–50,000	40,000–200,000 hrs (entire teams for years)

The 4-hour-per-asset figure is from talking to dozens of robotics simulation engineers; it tracks with NVIDIA’s own public estimates of 4–10 hours per object for full SimReady authoring.

Manual workflows don’t fail catastrophically at small scale. They fail by simply being too slow. Project timelines that assume 4 hours per asset force you to:

• Train policies on 100 objects when 1,000 would generalize better
• Skip rare-object scenarios that matter for safety (debris, edge-case SKUs, accident scenes)
• Defer digital twin builds for months while assets are prepared
• Hire 3D artists into roles that should be machine-automated

NVIDIA’s curated SimReady library currently has about 1,000 assets. A real warehouse has 50,000+ SKUs. The gap is structural — and AI-driven preparation is how you close it.

The AI-automated pipeline

The technical components of automated SimReady generation:

1. Geometry ingestion and analysis

The system accepts standard 3D formats (.glb, .fbx, .obj) and runs:

• Mesh validation (manifold check, normal orientation, watertightness)
• Geometric analysis (bounding box, volume, principal axes, structural features)
• Multi-view rendering for downstream vision models

2. Material identification

A vision model trained on rendered views classifies surface materials per region. Modern systems handle materials at the level of “brushed aluminum vs. anodized aluminum vs. cast iron,” not just “metal.” Each identified material maps to:

• Density (kg/m³) — for mass computation
• Static friction coefficient — for contact dynamics
• Dynamic friction coefficient — for sliding and grasping
• Restitution — for impact behavior
• Visual properties — for material binding in the output USD

Calibration data comes from physical property databases like MatWeb and the MIT Handbook of Material Constants, cross-referenced with research benchmarks like NeRF2Physics (CVPR 2024).

3. Mass and inertia estimation

• Mass: volume × identified density, with corrections for hollowness inferred from geometric thickness. For composite objects (a tool with a steel head and rubber grip), per-region densities are weighted by per-region volumes.
• Center of mass: computed analytically from the inferred mass distribution. Not assumed at the geometric centroid, which is wrong for asymmetric objects.
• Inertia tensor: computed from the volumetric integral of the mass distribution. Not assumed to be diagonal or uniform.

These three values together determine how the object behaves under any combination of forces and torques during simulation.

4. Collision mesh generation

Visual meshes are the wrong input for physics. They’re high-poly, often non-convex, and frequently have small details (ornaments, surface bumps) that slow contact resolution without improving simulation quality.

The auto-pipeline runs:

• Convex decomposition via V-HACD, with parameters tuned per object class (manipulation-grade detail for graspables; coarser hulls for furniture).
• Primitive replacement for objects that fit basic shapes (boxes, cylinders, capsules) — much faster to simulate.
• Detail preservation for functional features (handles, slots, grasping points) — these need to be in the collision mesh even if visually small.

The output is a collision mesh that’s accurate enough for contact-rich manipulation but fast enough for million-step RL training.

5. Semantic labeling

Object class labels (box, pallet, tool, bottle, vehicle) are assigned automatically based on the same vision pipeline that identifies materials. Labels are written into the USD output via SemanticsAPI, ready for perception model training, scene understanding, and synthetic data pipelines that need pixel-level segmentation maps.

6. USD authoring

The pipeline emits OpenUSD with:

• PhysicsRigidBodyAPI on dynamic prims
• PhysicsMassAPI carrying mass, COM, and inertia
• PhysicsCollisionAPI referencing the auto-generated collision mesh
• PhysicsMaterialAPI with material physics
• SemanticsAPI with class labels
• Material bindings for visual rendering
• A clean prim tree that passes the NVIDIA SimReady validator

Format conversion to MJCF (for MuJoCo), URDF (for Gazebo/ROS 2), or FBX with physics metadata (for Unity Articulation Bodies) is available on demand from the same OpenUSD source of truth.

7. Validation

Every output runs through automated checks:

• USD prim hierarchy is well-formed
• All applied schemas have required attributes
• No inverted normals or non-manifold geometry
• No interpenetration between collision sub-meshes
• Mass and inertia values fall within physically plausible ranges
• Friction coefficients fall within material-specific calibration bounds

Anything that fails validation is flagged for human review rather than shipped silently.

Comparing manual and AI workflows

Same job, both pipelines, side by side:

Stage	Manual	AI-automated
Geometry cleanup	15–45 min (Blender)	<30 s
Material identification	Manual lookup, often guessed	Vision model, per-region
Mass estimation	Volume × assumed density	Volume × identified density, with hollowness corrections
Inertia tensor	Assumed uniform	Computed from inferred distribution
Collision mesh	30–90 min V-HACD tuning	Auto-tuned per object class
Semantic labels	Manual tagging	Auto-classified
USD authoring	30–90 min hand-written schemas	Direct emission
Validation	20–45 min in-sim testing	Automated checks
Total	~4 hours	~5 minutes

For a 1,000-object project, this is the difference between 4,000 engineer-hours (~$370K at a $90/hr blended rate) and a single afternoon of automated processing.

Step-by-step: producing a SimReady asset with Rigyd

1. Upload a .glb, .fbx, or .obj file via the Rigyd dashboard or API. Multi-file batches are supported on Enterprise plans.
2. AI estimates physics — typically 3–6 minutes per asset depending on geometry complexity. The dashboard shows progress and a preview as analysis completes.
3. Review and override — inspect the estimated mass, friction, COM, and collision-mesh visualization. Override any value if you have measured ground truth (catalog mass, lab-measured friction, etc.).
4. Export as OpenUSD (default) or convert on-the-fly to MJCF, URDF, or FBX. Output passes the NVIDIA SimReady validator out of the box.
5. Drop into your simulator — physics components populate automatically on import; no further authoring needed.

Where AI estimation is and isn’t enough

AI is enough when:

• You’re building general-purpose training environments (warehouses, homes, offices, factories, urban scenes).
• Your downstream pipeline uses domain randomization with typical variance ranges (mass ±20%, friction ±0.1).
• You need scale: hundreds to tens of thousands of objects.
• The cost of a 10–20% physics error is small compared to the cost of not having the asset at all.

Manual still wins when:

• A single asset is mission-critical and you have access to lab measurement (a specific surgical instrument, a flight-control component).
• The geometry is articulated or soft-body in ways the auto-pipeline doesn’t yet support (specific cable dynamics, deformable plates).
• The material is exotic and not in the calibration database.

For everything else — and especially for the long tail of warehouse SKUs, household objects, and edge-case scenarios — automated SimReady generation has become table stakes for production simulation teams.

The economic argument

A simulation team’s effective output is roughly: (asset diversity) × (training time) × (algorithm quality) × (real-world validation cycles).

Manual asset authoring caps the first term well below where it should be. A team with $100K of asset budget at 4 hours per asset can produce ~270 SimReady objects. The same team using AI-driven preparation can produce 100,000+ for a fraction of the cost. The compounding effect on policy generalization is significant — most reinforcement learning research that scales to real-world deployment depends on diverse environments more than on architectural innovation.

If you’re standing up a new simulation environment in 2026, the practical sequence is:

1. Use the manual workflow for 5–10 hero assets to learn your simulator’s quirks and accuracy needs.
2. Switch to AI-automated preparation for everything else.
3. Validate end-to-end with domain randomization rather than chasing 1% physics accuracy on individual assets.

The 4-hours-per-asset bottleneck has become a choice, not a constraint. Robotics teams that recognize this earlier ship policies that generalize better, sooner.