Do Reality-Gap-Reduction Methods Limit Simulation Data Scaling for Large Robotics Models?

Ascertain whether the methods and assumptions required to reduce the reality gap in synthetic data generation inherently constrain the scalability of simulation-based dataset collection for training large robotic models.

Background

Large behavior and foundation models for robotics increasingly depend on vast datasets; simulation offers a route to scale data generation. However, for synthetic data to be useful, its reality gap must be minimized through methods such as high-fidelity asset creation, accurate sensor modeling, and physics calibration.

The authors question whether these gap-reduction techniques—and the assumptions they entail—limit the achievable scale of simulated data, potentially constraining the training of large models.