A Comprehensive Survey on World Models for Embodied AI

arXiv 25.10 2025 55.4 method

TLDR

A comprehensive survey proposing a unified framework and taxonomy for world models in embodied AI, covering functionality, temporal modeling, spatial representation, and open challenges.

Reasoning

Strengths: Provides a structured taxonomy and comprehensive overview of world models for embodied AI, including data resources, metrics, and quantitative comparisons. Weaknesses: As a survey, it does not introduce novel methods or experiments; the taxonomy may be subjective and open challenges are well-known.

Read-first score

Read-first score 55.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 52.

Recency 6%

86.7

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

Reproducibility 18%

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

Methodology quality 18%

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

Topical relevance 29%

74.3

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

Citation impact 18%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FrontierMethodology anchorReproducibility anchor

Rank sensitivity

Stability: volatile; rank range: 833.

Keyword Scores

world model

world simulator

world dynamics prediction

video world model

model-based reinforcement learning world model

generative world model

interactive world model

Deep Analysis

Innovations

Proposes a unified framework for world models in embodied AI, formalizing problem setting and learning objectives.
Introduces a three-axis taxonomy: Functionality (Decision-Coupled vs. General-Purpose), Temporal Modeling (Sequential Simulation and Inference vs. Global Difference Prediction), and Spatial Representation (Global Latent Vector, Token Feature Sequence, Spatial Latent Grid, Decomposed Rendering Representation).
Systematizes data resources and metrics across robotics, autonomous driving, and general video settings, covering pixel prediction quality, state-level understanding, and task performance.
Offers a quantitative comparison of state-of-the-art world models.
Distills key open challenges for the field.

Methodology

This survey conducts a comprehensive literature review of world models for embodied AI, formalizing the problem and learning objectives. It constructs a three-axis taxonomy to categorize existing approaches, systematizes available datasets and evaluation metrics, and performs a quantitative comparison of state-of-the-art models to identify trends and gaps.

Key Results

The survey provides a quantitative comparison of state-of-the-art world models and distills key open challenges, including the scarcity of unified datasets, the need for physical-consistency metrics over pixel fidelity, the trade-off between performance and real-time computational efficiency, and the difficulty of achieving long-horizon temporal consistency while mitigating error accumulation.

Limitations

Scarcity of unified datasets for training and evaluation.
Need for evaluation metrics that assess physical consistency over pixel fidelity.
Trade-off between model performance and computational efficiency required for real-time control.
Core modeling difficulty of achieving long-horizon temporal consistency while mitigating error accumulation.

Resources

Paper Project