The sim-to-real gap is a physics gap
When simulation objects have wrong mass, friction, or collision geometry, trained policies fail on real hardware. The gap isn't the simulator, it's the data.
3D Assets for Better Sim-to-Real Transfer
The #1 cause of sim-to-real failure is inaccurate simulation data. Rigyd generates 3D assets with calibrated physics properties so policies trained in simulation actually work on real robots.
Updated 2026-05-08
The problem
Why existing workflows fall short.
Domain randomization needs a good baseline
Randomizing around inaccurate physics values produces worse policies, not better ones. Domain randomization works best when centered on realistic parameters.
Real-world validation is expensive
Each sim-to-real iteration costs hardware time, engineer hours, and potential damage to expensive robots. Getting it right in simulation saves real-world cycles.
How Rigyd helps
Automate the hard parts.
Calibrated physics baselines
Rigyd estimates mass within 15-20% of measured values and friction within 0.1 coefficient, providing accurate baselines for domain randomization.
Material-based estimation
AI identifies per-region materials (ceramic, aluminum, rubber) and maps them to a calibrated physical properties database. No guesswork.
Validated against research benchmarks
Physics estimation approach validated against NeRF2Physics (CVPR 2024) and GaussianProperty methodologies for mass and friction accuracy.
40%
better real-world performance with physically accurate training data
15-20%
mass accuracy vs measured values
0.1
friction coefficient accuracy
FAQ
Frequently asked questions
What is the primary cause of sim-to-real failure?
Inaccurate simulation physics. When simulated objects have wrong mass, friction, or collision geometry, trained policies fail on real hardware. The gap isn't the simulator, it's the asset data.
How much does physically accurate data improve real-world performance?
Roughly 40%, per NVIDIA's GR00T N1 benchmark (2025), when synthetic training data uses calibrated physics. Rigyd provides mass accuracy within 15-20% and friction within 0.1 of measured values.
Does Rigyd replace domain randomization?
No, Rigyd gives domain randomization better baselines. DR works best when centered on realistic physics values; random variation around accurate starting points produces more robust policies than uniformly random physics.
What real-world performance gap remains after training with Rigyd assets?
For most manipulation tasks, the remaining sim-to-real gap drops to roughly 15-25% of baseline accuracy when training combines Rigyd's calibrated physics with domain randomization. Precision tasks (peg-in-hole sub-millimeter clearance, dexterous in-hand reorientation) still benefit from real-world fine-tuning on 5-15% of training data. The dominant residual is camera-real gap, not physics-real gap.
Can Rigyd assets be used with system identification (sysID) pipelines?
Yes. Rigyd outputs are starting points; sysID pipelines can refine mass, friction, and inertia values against real-robot trajectories and write them back into the USD as overrides in a layer. Because USD composition is layered, sysID refinements don't modify the source mesh, they stack on top, and you can roll back, A/B test, or version-control physics overrides independently of geometry.
Close the sim-to-real gap with better data
Start with physically accurate assets and let domain randomization do the rest.