World Models for Robotic Manipulation: A Survey

arXiv 2026 64 survey

TLDR

A survey of world models for robotic manipulation, defining them as action-conditioned predictive systems and organizing approaches by representation, prediction-action coupling, and usage pipeline.

Reasoning

The paper provides a comprehensive taxonomy and review of world models in robotic manipulation, clearly defining the scope and distinguishing different families. However, as a survey, it lacks novel experiments or real-world validation, and the breadth may obscure specific design trade-offs.

Read-first score

Read-first score 64, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 51.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Citation impact 18%

89.8

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.89803948

Methodology quality 18%

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

Topical relevance 29%

72.9

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 18%

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

Citation velocity 12%

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

Field roles

FoundationFrontierBridgeMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 396.

Keyword Scores

world model

world dynamics prediction

generative world model

video world model

world simulator

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Operational definition of world model as action-conditioned predictive system
Organization of existing work into five representation families
Functional taxonomy separating integrated prediction-action models from explicit predictive planners
Characterization of infrastructure roles (synthetic experience generation, candidate filtering, search-based evaluation, learned environments, outcome verification)
Mapping of roles across pretraining, post-training, and inference adaptation
Review of 34 manipulation datasets and synthesis of evaluation protocols

Methodology

The authors conduct a systematic survey of world models for robotic manipulation, organizing the literature by three guiding questions: what future representation is predicted, how prediction is connected to action, and when prediction is used. They develop a functional taxonomy, characterize infrastructure roles, review 34 datasets, and synthesize evaluation protocols for predictive fidelity, task performance, and simulator reliability.

Key Results

The survey reveals that world models are evolving from task-specific dynamics predictors into predictive infrastructure for robot learning, and identifies critical open challenges including contact modeling, hallucination control, action alignment, and benchmarking under closed-loop use.

Limitations

Contact modeling remains an open challenge
Hallucination control is not yet solved
Action alignment between prediction and control is difficult
Benchmarking under closed-loop use is underdeveloped

Resources

Paper