Simulation Fidelity

How closely a simulation matches the real world, in both the accuracy of its physics and the realism of its appearance.

Simulation fidelity is usually split into two parts: physical fidelity, meaning how accurately dynamics, contact, friction, deformable objects, and motor characteristics are computed; and visual or sensor fidelity, meaning how closely rendered images, depth, and lidar readings match a real sensor. Higher fidelity means a smaller sim-to-real gap, but it also costs more compute. A 2021 PNAS review by Choi and colleagues points out that game engines aim for “looking plausible” rather than being precise, and that practical use often requires compromises such as reducing solver iterations or substituting a rigid ground for a deformable one. In practice this is usually paired with domain randomization and system identification; work such as SIMPLER has also shown that policy evaluation doesn't necessarily require a fully realistic digital twin.

ExampleSIMPLER didn't reconstruct a complete digital twin — it only composited a real background onto a green screen, aligned textures, and identified control parameters, which was enough to make its simulated evaluation scores correlate strongly with real-robot results.

Also called

Physical Fidelity, Visual Fidelity

Related

Sim-to-Real Gap (Reality Gap) · Physics Engine · Rendering · Sensor Simulation · Digital Twin · Domain Randomization

Sources

Choi et al., On the use of simulation in robotics: Opportunities, challenges, and suggestions for moving forward (PNAS 2021)
Evaluating Real-World Robot Manipulation Policies in Simulation (SIMPLER, arXiv 2405.05941)