Measure the gap. Then close it.
Simulation gives you volume; reality gives you truth. Agentuor holds both under one ontology so the gap between them becomes a metric you can track and act on — per condition, not on average.
The challenge
Teams usually know their sim-to-real gap exists but not where it lives. Simulated and real data sit in different tools with different schemas, so comparisons happen in ad-hoc notebooks. Domain randomization is tuned by intuition. When a policy that aced simulation fails on hardware, nobody can point to the lighting, texture, or dynamics condition responsible.
Built for this data
Shared ontology
One label schema and one evaluation suite across simulated and real episodes.
Gap metrics per slice
Performance delta by lighting, texture, object class, dynamics, and embodiment.
Randomization guidance
Agents identify conditions where sim coverage is high but real performance lags, pointing to where randomization is insufficient.
Paired episodes
Link simulated and real recordings of the same task for direct comparison.
Divergence tracing
Jump from a gap metric to the real episodes driving it and their annotation history.
Real-data prioritization
Recommendations for which real scenarios to collect to validate simulation most efficiently.
Data and labels
- Simulation rollouts with ground-truth labels and randomization parameters
- Real episodes with human-verified labels under the same ontology
- Task and scene metadata aligned across domains
- Paired-episode links
- Evaluation suites frozen across both domains
Typical workflow
Import both domains
Sim and real data ingested under one schema with domain tags.
Align
Metadata mapped so conditions are comparable across domains.
Evaluate
Policies scored on both; gaps computed per slice.
Act
Randomization and collection recommendations turned into tasks.
Outcomes teams work toward
Common questions
Which simulators are supported?
Any simulator that exports labeled rollouts; importers exist for common robotics simulators.
Can sim labels be trusted as ground truth?
They are treated as ground truth by default, with agents flagging cases where sim labels conflict with real conventions.
Does Agentuor run simulations?
No. It ingests, organizes, and evaluates simulation output.
Show us an episode. We'll show you the gaps.
Start with a walkthrough on your own data.