Today we're launching Lucitra Validate — an API-first platform that scores synthetic training datasets for coverage completeness, physics plausibility, and sim-to-real transfer confidence. If you're building robotics or autonomous systems that train on synthetic data, Validate tells you whether your dataset is ready before you spend compute on training.
What Lucitra Validate Does
You upload a synthetic dataset from your simulation tool — NVIDIA Isaac Sim, Omniverse, Unreal Engine, or any pipeline that produces labeled images. Lucitra Validate returns a structured report with scores across multiple quality dimensions:
- Coverage score — How thoroughly does the dataset represent the target domain? Are there gaps in viewpoint distribution, lighting variation, or object diversity?
- Annotation accuracy — Do bounding boxes, segmentation masks, and depth maps align with the rendered scene geometry?
- Physics plausibility — Are object placements, collisions, and material properties physically consistent?
- Distribution analysis — Are class frequencies, pose distributions, and scene configurations balanced for your use case?
- Sim-to-real transfer confidence — An aggregate score estimating how well a model trained on this dataset will generalize to real-world sensor data.
Each score includes specific recommendations: which object classes need more variation, which scenes have annotation errors, and which distribution gaps to address.
Four Ways to Use It
API
The REST API is the foundation. Upload a dataset, get back a validation report with scores and recommendations. Designed for integration into automated pipelines.
curl -X POST https://api.lucitra.com/v1/validations \
-H "Authorization: Bearer luci_your_key" \
-F "dataset=@warehouse-v3.tar.gz" \
-F "config=@validation-config.json"
CLI
The lucitra CLI wraps the API for terminal workflows. Install via npm, validate datasets from the command line, and integrate into CI/CD scripts.
npx lucitra validate ./datasets/warehouse-v3/ \
--config validation.yaml \
--fail-under 0.85
The --fail-under flag makes the CLI exit with a non-zero code if the overall score drops below your threshold — turning validation into a pipeline gate.
Dashboard
The web dashboard visualizes validation results. Compare runs over time, drill into per-category scores, and share reports with your team. Built for the engineer who wants to understand why a score changed between dataset versions.
Isaac Sim Plugin
For teams using NVIDIA Isaac Sim, the Omniverse extension validates directly from within the simulation environment. Select a scene, run validation, and see results without leaving the editor.
Architecture
Lucitra Validate is built on a straightforward stack:
- API: FastAPI (Python) running on Google Cloud Run with GPU-accelerated inference
- Analysis models: Qwen2.5-VL for scene understanding, Depth Anything v2 for geometric validation
- Storage: Google Cloud Storage for datasets and reports
- Dashboard: Next.js with TypeScript
The API processes datasets asynchronously. For a typical dataset (10,000 images with annotations), validation completes in under 10 minutes. Results are stored and accessible via the API or dashboard indefinitely.
What's Next
We're focused on three areas:
- More simulation tools — Adding direct integrations for Unreal Engine and custom USD pipelines
- Custom scoring models — Let teams define domain-specific validation criteria
- Comparison baselines — Track dataset quality across versions and alert on regressions
Lucitra Validate is available now. Start with the free tier (5 validation runs per month) or contact us for enterprise access.