RoboPhys-3D tests world models in 50 tasks

RoboPhys-3D evaluates embodied world models across 50 manipulation tasks and 25,000 multi-view ground-truth videos, forcing generated and reference videos through the same 3D reconstruction pipeline. The paper reports Cosmos 3 at a RoboPhyscore of 0.6330, or 92.7% of ground truth, while state- and execution-level metrics exposed failures missed by perceptual and vision-language judges. Robot teams should add execution-grounded checks before trusting video quality, building on the archive’s case for physical-AI scale.