Rethinking Video Generation Model for the Embodied World

TL;DR

The authors tackle the challenge of generating videos that faithfully reflect embodied robotic interactions by introducing RBench, a benchmark that jointly evaluates task completion and visual fidelity across five tasks and four embodiments, with automated MLLM‑based metrics and physics‑aware subcriteria. They validate RBench on 25 diverse models, revealing substantial gaps in physical realism and suggesting that current video foundation models are transitioning toward physical intelligence rather than mere visual quality. To address data bottlenecks, they introduce RoVid‑X, a four‑stage pipeline that yields roughly 4 million annotated robotic videos spanning thousands of tasks, enabling scalable training and evaluation for embodied video models. Together, RBench and RoVid‑X provide an integrated evaluation and data ecosystem that accelerates progress toward robust, physically grounded embodied AI. The work also demonstrates a strong human–machine alignment in assessment (ρ ≈ 0.96) and outlines future directions toward executable action extraction via inverse dynamics and more physically grounded evaluation metrics.

Abstract

Video generation models have significantly advanced embodied intelligence, unlocking new possibilities for generating diverse robot data that capture perception, reasoning, and action in the physical world. However, synthesizing high-quality videos that accurately reflect real-world robotic interactions remains challenging, and the lack of a standardized benchmark limits fair comparisons and progress. To address this gap, we introduce a comprehensive robotics benchmark, RBench, designed to evaluate robot-oriented video generation across five task domains and four distinct embodiments. It assesses both task-level correctness and visual fidelity through reproducible sub-metrics, including structural consistency, physical plausibility, and action completeness. Evaluation of 25 representative models highlights significant deficiencies in generating physically realistic robot behaviors. Furthermore, the benchmark achieves a Spearman correlation coefficient of 0.96 with human evaluations, validating its effectiveness. While RBench provides the necessary lens to identify these deficiencies, achieving physical realism requires moving beyond evaluation to address the critical shortage of high-quality training data. Driven by these insights, we introduce a refined four-stage data pipeline, resulting in RoVid-X, the largest open-source robotic dataset for video generation with 4 million annotated video clips, covering thousands of tasks and enriched with comprehensive physical property annotations. Collectively, this synergistic ecosystem of evaluation and data establishes a robust foundation for rigorous assessment and scalable training of video models, accelerating the evolution of embodied AI toward general intelligence.

Figures (23)

• Figure 1: Overview of the comprehensive robotics benchmark and dataset for video generation. Top: We present RBench that includes the embodiment-based evaluation set and automated evaluation metrics. Our evaluation results on 25 video models show a high level of agreement with subjective human assessments. Bottom: We introduce a large-scale high-quality robotic dataset (RoVid-X) specifically designed for training video generation models, with data sourced from internet videos and open-source embodied videos.
• Figure 2: Qualitative illustration of failure modes captured by RBench. Unlike conventional metrics that focus primarily on pixel-level fidelity, RBench provides a granular evaluation across multiple dimensions, including physical plausibility and task-level consistency. These results highlight persistent challenges in robotic video generation, such as structural distortion, floating components, and key action omission, which are accurately identified by our proposed sub-metrics. More cases are shown in the Appendix \ref{['subsec:evaluation_metrics']}.
• Figure 3: Statistics in RBench. The benchmark covers diverse tasks, object categories, and environments, demonstrating the high quality and comprehensiveness of the evaluation set, highlighting its high applicability to a wide range of robotic video generation scenarios.
• Figure 4: Overview of RoVid-X Construction and Descriptive Statistics. (a) shows the four-stage pipeline for constructing the RoVid-X. (b) presents descriptive statistics, covering frame intervals, skill distribution, and common objects, highlighting the dataset's diversity and suitability for robotic task training and video generation.
• Figure 5: Qualitative comparison across representative tasks. We visualize the generated results for three representative tasks: Visual Reasoning, Long-horizon Planning, and Spatial Relationship, across six models. Each row displays temporally sampled frames from the same generated video, with captions below indicating the corresponding task instruction. More cases are shown in the Appendix.