GAUGE: A Measurement-Grounded Benchmark for Physical Fidelity in Simulation Engines and Video World Models

GAUGE (a Measurement-Grounded benchmark for physical fidelity) addresses the fragmented evaluation of physics simulators and generative video world models. Existing tests rely heavily on perceptual similarity or human judgment and rarely pinpoint which physical principles or parameters are violated. GAUGE instead anchors evaluation in real-world trajectories and calibrated physical metadata, providing a diagnostic view of how numerical simulators and video models deviate from actual physics.

The benchmark spans 22 controlled task families covering rigid bodies, flexible cables, textiles, and volumetric deformable objects, with tasks designed around fundamental processes such as collision, friction, momentum transfer, oscillation, self-contact, and deformation. It includes uncertainty annotations and task-specific observables to support meaningful comparison. The authors benchmarked Isaac Sim, Genesis, and Newton across 14 task families using generalized trajectory errors, and evaluated 6 image-to-video models on 5 rigid-body tasks by checking physical-law consistency and temporal stability of inferred parameters.

Results show that no physics engine is uniformly faithful: the largest discrepancies appear in impulsive contact, rapid textile motion, and volumetric deformation. Video world models can produce trajectories with the expected equation form, but they recover incorrect accelerations, momentum transfer, and oscillation timing. GAUGE is intended as a foundation for building more physically faithful simulators and world models for embodied intelligence.