STORM: Search-Guided Generative World Models for Robotic Manipulation

arXiv 25.12 2025 69.3 method

TLDR

STORM integrates diffusion-based action generation, video prediction, and MCTS for interpretable robotic manipulation planning, achieving state-of-the-art results on SimplerEnv.

Reasoning

The paper presents a novel framework that explicitly grounds planning in visual rollouts, with strong empirical results (51.0% success rate, 75% FVD reduction) on a simulation benchmark. However, it lacks real-world experiments and only evaluates on a single simulated environment, limiting generalizability claims.

Read-first score

Read-first score 69.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 58.

Recency 8%

86.7

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

Topical relevance 42%

82.9

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

Methodology quality 25%

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

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 157.

Keyword Scores

world model

generative world model

video world model

world simulator

world dynamics prediction

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Unifies diffusion-based action generation, conditional video prediction, and search-based planning into a single framework for spatio-temporal reasoning in robotic manipulation.
Grounds planning in explicit visual rollouts rather than abstract latent dynamics or language-based reasoning, enabling interpretable and foresight-driven decision-making.
Introduces reward-augmented video prediction that substantially improves spatio-temporal fidelity and task relevance, reducing Frechet Video Distance by over 75%.

Methodology

STORM employs a diffusion-based Vision-Language-Action (VLA) policy to propose diverse candidate actions, a generative video world model to simulate their visual and reward outcomes, and Monte Carlo Tree Search (MCTS) to selectively refine plans through lookahead evaluation. The framework is evaluated on the SimplerEnv manipulation benchmark against baselines such as CogACT.

Key Results

STORM achieves a new state-of-the-art average success rate of 51.0% on SimplerEnv, outperforming strong baselines like CogACT, and reduces Frechet Video Distance by over 75% through reward-augmented video prediction. It also exhibits robust re-planning and failure recovery behavior for long-horizon tasks.

Resources

Paper