Yufan Deng and colleagues present a repository for two key components in robotics video generation: RBench and RoVid-X. RBench is a specialized benchmark designed for evaluating video generation models in robotics, while RoVid-X is a dataset containing a million videos to support the training of these models. The authors highlight the shortcomings of existing video foundation models and propose avenues for enhancement, aiming to lay a solid groundwork for testing and training in physical AI. The repository includes detailed instructions for setting up the necessary environment, with specific Python packages and modules required for RBench. Users need to organize checkpoint files in a defined directory structure to run evaluations effectively. RBench achieves a Spearman correlation of 0.96 when comparing its evaluations of video generation models to human assessments, indicating its reliability. It evaluates 25 models across various categories, including open-source and commercial options, focusing on task-oriented and embodiment-specific dimensions. RoVid-X offers RGB, depth, and optical flow videos that facilitate real-world robotic interactions. The dataset aims to bridge gaps in training embodied video models. The authors are currently training a robotic video world model for practical applications and plan to release RoVid-X and RBench as open-source resources. Their research paper is already available, contributing to ongoing discussions in the field of embodied AI and video generation.