Physically accurate
synthetic data generation

Synthetic data is only as good as the simulation that produces it. Rigyd builds the physically accurate 3D asset layer, the foundation every synthetic dataset needs for training policies that actually transfer to real robots.

The problem

Why existing workflows fall short.

Synthetic data is bottlenecked by asset quality

Beautiful renders produced from objects with wrong mass or missing collision meshes produce training data that breaks policies on real hardware. Garbage physics in, garbage policies out.

Scale is limited by manual annotation

Synthetic data promises unlimited scale, but only if the asset-creation layer isn't the bottleneck. Hand-annotating physics caps dataset growth long before training needs it to.

Coverage requires object diversity

Robust policies need thousands of unique objects at training time. Most synthetic pipelines reuse a small asset pool, which quietly causes overfitting and poor out-of-distribution performance.

How Rigyd helps

AI-native infrastructure that automates the hard parts.

Physics layer for your rendering stack

Rigyd outputs drop into Isaac Sim Replicator, Omniverse, or custom synthetic-data pipelines, contributing the physics layer that makes rendered images and lidar scans trainable.

Bulk object generation

Convert entire 3D catalogs into physics-enabled assets. Enterprise API supports dataset pipelines with high throughput and continuous updates as new object classes are added.

Calibrated for transfer

Mass accuracy within 15-20% and friction within 0.1 coefficient, matching the variance ranges typical domain-randomization pipelines target for sim-to-real transfer.

40%

better real-world performance with physically accurate synthetic data

97%

cost reduction in asset preparation

1,000+

unique objects per dataset becomes practical

Generate synthetic data that actually transfers

Build your synthetic dataset on a foundation of physically accurate 3D assets.

Starts at $29/month. 30 credits included.

Frequently asked questions

Is synthetic data only useful with accurate physics?

Largely, yes. Rendering fancy images from objects with wrong mass or missing collision produces training data that breaks real-world policies. Garbage physics in, garbage policies out, even with pixel-perfect rendering.

Does Rigyd replace my rendering pipeline?

No, Rigyd provides the physics-enabled asset layer. Your Isaac Sim Replicator, Omniverse, or custom synthetic data pipeline still handles rendering; Rigyd makes the underlying assets transferable.

How does physics accuracy affect sim-to-real transfer rates?

Roughly 40% real-world performance improvement, per NVIDIA's GR00T N1 benchmark (2025), when synthetic datasets use physically accurate assets instead of default or randomized physics.

Does Rigyd integrate with NVIDIA Replicator and Omniverse Replicator?

Yes, directly. Replicator reads USD scenes with SemanticsAPI labels and PhysicsAPI schemas, both of which Rigyd populates by default. Per-object instance IDs and class labels generate pixel-perfect segmentation masks; physics ensures dropped, stacked, or perturbed objects settle into realistic configurations rather than floating. No additional annotation step needed before Replicator can produce labeled synthetic frames.

Can synthetic data from Rigyd-prepared assets replace real-world labeled data entirely?

Rarely 100%, most production deployments use 60-80% synthetic with 20-40% real fine-tuning. Pure-synthetic works for some perception tasks in known environments but rarely transfers cleanly to manipulation. The 20-40% real-world data does the final calibration; the 60-80% synthetic data does the scale, diversity, and edge-case coverage that real data can't economically reach.