Reference-Free Assessment of Physical Consistency in World Model-based Video Generation

The paper presents a novel reference-free methodology for assessing the physical consistency of videos generated by world models, utilizing DROID-SLAM and SEA-RAFT to quantify inconsistencies without relying on human evaluations or ground-truth references. This approach has demonstrated an over 8% improvement in task success rates, effectively bridging the gap between simulated and real-world performance. The method's ability to provide spatio-temporal localization of artifacts is significant for practitioners aiming to enhance the fidelity of video generation in robotic simulations.