The Zhiyuan Research Institute, in collaboration with multiple institutions, has launched Wujie · PhysCapBench—the first cross-domain benchmark for evaluating physical information in video descriptions. It covers three major domains: third-person scenes, first-person manipulation, and handheld gripper demonstrations. The benchmark comprises 1,500 video sequences, defines seven evaluation dimensions for physical information, and constructs verifiable reference physical descriptions grounded in video evidence, enabling unified measurement of the coverage and fidelity of physical information in video descriptions. Downstream experiments confirm that more complete and accurate physical descriptions can significantly improve video generation results and increase the success rate of robot policy learning. This provides a unified cross-domain reference for description strategy selection, annotation quality assessment, and training data optimization in the development of physical AI foundation models.
