GAUGE Benchmark Evaluates Physical Fidelity in Simulation Engines and Video World Models

A new benchmark named GAUGE has been developed by researchers to assess the physical accuracy of simulation engines and generative video world models. This benchmark, discussed in a paper on arXiv (2608.05948), aims to create a standardized method for evaluating how accurately these systems mimic real-world physics. GAUGE includes 22 controlled task families that involve rigid bodies, flexible cables, textiles, and volumetric deformable objects. These tasks utilize real-world trajectories and are accompanied by calibrated physical metadata, uncertainty annotations, and task-specific observables. By focusing on fundamental physical processes such as collision, GAUGE intends to reveal which physical principles or parameters are compromised by simulators and world models. Traditional evaluations often depend on perceptual similarity or human assessments, which provide limited diagnostic insights. GAUGE aims to address this issue by allowing for the simultaneous evaluation of numerical simulators and generative models. This research is significant for embodied intelligence, where physics engines support extensive training and evaluation, as well as for the emerging domain of video world models functioning as implicit simulators. The paper can be accessed at https://arxiv.org/abs/2608.05948.