Sim-to-real gap
The sim-to-real gap describes mismatch between a simulated setting and the real setting where a system is used. The mismatch can involve visual appearance, physical behavior, timing, sensor noise, or task assumptions. It matters when learning or evaluation in simulation is used to predict real-world performance.
How it works
A simulator necessarily selects which details to represent. Transfer methods may vary simulated conditions, adapt models using real observations, or improve the simulator. Domain-randomization research illustrates training across varied visual conditions to support transfer in a particular task.
Evaluation still needs appropriate real-world evidence. A method that helps object recognition does not automatically solve control, contact dynamics, or every other source of mismatch.
Why it matters for licensing
Real and synthetic datasets can play different roles in testing whether a system generalizes. Documentation should identify capture or simulation settings and known gaps. A license should specify the actual data and rights involved; this definition does not describe a Rancher robotics offering.
Example
Fictional example: A robot vision model trained on simulated parts encounters reflective packaging and uneven lighting in a real test. The team measures those failures and revises its data and evaluation conditions before drawing broader conclusions.
Limitations and misconceptions
A realistic-looking simulation can still have incorrect dynamics or missing edge cases. More varied synthetic examples do not guarantee safe transfer. Testing must account for the consequences of failure in the intended setting.
Questions to ask
- Which visual, physical, timing, or sensor conditions differ?
- What independent real-world evaluation supports transfer claims?
- Which limitations remain outside the tested conditions?