⌁ Awesome World Model Hub Papers · Datasets · Projects
中文 GitHub
← Back to papers

Scalable Policy Evaluation with Video World Models

arXiv 25.11 2025 69 method, application

TLDR

Uses action-conditional video generation models as world models for scalable policy evaluation in robotic manipulation.

Reasoning

Strengths include leveraging pre-trained video models to reduce data requirements and addressing real-world testing costs. Weaknesses are the focus on evaluation only, not full simulation, and potential sim-to-real gaps.

Read-first score

Read-first score 69, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 50.

Methodology quality 25%
90

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=dataset,evaluation,experiment,metric

Recency 8%
86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Topical relevance 42%
71.4

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 25%
38

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=dataset

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 40.

Keyword Scores

video world model
10
world model
9
generative world model
8
world dynamics prediction
8
model-based reinforcement learning world model
6
world simulator
5
interactive world model
4

Deep Analysis

Innovations

  • Using action-conditional video generation models as scalable world models for policy evaluation in robotic manipulation.
  • Incorporating action conditioning into existing pre-trained video generation models to leverage internet-scale in-the-wild online videos.
  • Reducing the need for large paired video-action datasets by utilizing pre-trained video models and online video data.

Methodology

The paper proposes incorporating action conditioning into pre-trained video generation models to create world models for policy evaluation. It examines the effects of dataset diversity, pre-trained weights, and common failure cases, and evaluates policies using metrics such as policy ranking and correlation between actual and predicted policy values.

Key Results

Experiments demonstrate that the proposed approach offers a promising method for evaluating policies without real-world interactions, showing strong performance across metrics including policy ranking and correlation between actual and predicted policy values.

Limitations

  • The approach still requires some paired video-action data for conditioning, though less than traditional methods.
  • Common failure cases in video generation may affect evaluation accuracy and reliability.
  • Potential sim-to-real gap in physics and rendering remains a challenge, as video world models may not fully capture real-world dynamics.

Resources

Paper

Tags