Research
Research & Numbers
The quantitative benchmarks cited across rigyd.com, with sources and methodology. Where a number is sourced externally, the source is named; where a number is derived from internal modelling, the assumptions are spelled out so you can re-run the calculation against your own rates.
· by Ugur Yekta
§1 · Asset-coverage gap
~1,000 vs 50,000+
Number of SimReady-grade assets readily available in NVIDIA’s curated library, versus the typical SKU count in a mid-sized industrial environment.
§2 · Synthetic + real data lift
+40%
Real-world policy performance when training combines synthetic and real data with physically accurate assets, versus synthetic-only or real-only baselines.
§3 · Scene-preparation cost reduction
~97%
Reduction in engineer time to prepare a single simulation-ready asset, manual workflow versus AI-automated workflow.
§4 · Per-project savings
~$370K
Estimated engineering-time savings on a 1,000-object simulation project when switching from manual annotation to AI-automated annotation.
§1 · Asset-coverage gap
~1,000 hand-validated assets vs 50,000+ catalog SKUs
NVIDIA Omniverse’s curated SimReady asset library offers approximately 1,000 hand-validated assets as of June 2026. A mid-sized warehouse, retail floor, or manufacturing line typically catalogs 10,000–100,000+ SKUs. The shortfall is the asset-coverage gap robotics teams have to close before simulation results transfer to their real environment.
Sources:
- NVIDIA Omniverse SimReady asset library (publicly listed catalog), ~1,000 assets as of June 2026.
- Industry SKU range based on standard supply-chain catalogs (typical mid-sized warehouse).
§2 · Synthetic + real data lift
+40% real-world performance with physics-accurate synthetic data
Robotics policies trained on a mix of physically accurate synthetic data and real captures outperform policies trained on either source alone by approximately 40% on real-world evaluation, when the synthetic assets carry calibrated physics (mass, friction, restitution, collision geometry). Generic visual meshes alone do not produce this lift — policies learn to exploit the unphysical artefacts of the simulator and fail to transfer.
Source:
- NVIDIA Project GR00T benchmark series, 2025.
§3 · Scene-preparation cost reduction
~97% reduction in engineer time per asset
Manual SimReady authoring takes about 4 engineer-hours per asset: geometry cleanup, convex collision decomposition, mass and inertia estimation from material density, friction coefficient assignment, validation in the target simulator, and the relevant USD / SDF / MJCF export work. AI-automated annotation compresses the same workflow to roughly 5 minutes per asset, dominated by upload, review, and validation rather than authoring.
Methodology (internal modelling):
- Engineer blended rate: $95/hour (typical robotics-simulation contractor rate, U.S., 2026).
- Manual per-asset cost: 4 × $95 = $380.
- AI per-asset cost: (5 / 60) × $95 ≈ $8.
- Reduction: ($380 − $8) / $380 ≈ 97.9% → rounded to ~97%.
Excluded: platform / software licence fees, which vary by tier and volume. The dominant cost in both workflows is engineer time, so the methodology focuses on that.
§4 · Per-project savings
~$370K saved on a 1,000-object simulation project
Engineering-time saving on a single 1,000-object simulation project when switching from manual SimReady annotation to AI-automated annotation, holding all other assumptions equal.
Methodology (internal modelling, builds on §3):
- Per-asset saving (from §3): $380 − $8 = $372.
- Project (1,000 assets): 1,000 × $372 = $372,000 → rounded to ~$370K.
Sensitivity: the model scales linearly with the engineer rate. At $75/hr the saving is ~$294K; at $125/hr it is ~$489K. Plug in your team’s effective rate to estimate your own saving.
Excluded: licensing, infrastructure, and downstream sim-compute costs — typically a small fraction of the engineer-time line item for production catalog projects.
Methodology summary
The manual-workflow figures (4 hours per asset, $95/hr blended rate) draw from interviews and published rates for robotics-simulation contractors. The AI-automated figures (5 minutes per asset) match the typical end-to-end time observed across real Rigyd 3D → SimReady submissions in 2026.
Third-party numbers (NVIDIA’s catalog size, the GR00T benchmark lift) are named in their respective sections. Internal-model numbers are labelled as such and every assumption is spelled out so any of the figures can be re-derived against your team’s actual rates.
How to cite this page
Suggested citation:
Rigyd Research, June 2026. https://rigyd.com/research
For a specific stat, cite the section number — e.g. Rigyd Research §3, June 2026.