Scaling World Model for Hierarchical Manipulation Policies

arXiv 26.2 2026 49.5 method, application

TLDR

Hierarchical VLA framework using a world model as high-level planner to generate goal images, boosting OOD manipulation performance from 14% to 69%.

Reasoning

The paper presents a novel hierarchical approach that leverages a pre-trained world model for subgoal decomposition, showing strong empirical gains in out-of-distribution scenarios. However, the abstract lacks details on the world model architecture and the scope of real-world validation, limiting assessment of generalizability.

Read-first score

Read-first score 49.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 28.

Recency 8%

100

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

Methodology quality 25%

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

Topical relevance 42%

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%

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

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 476.

Keyword Scores

world model

generative world model

world dynamics prediction

world simulator

model-based reinforcement learning world model

interactive world model

video world model

Deep Analysis

Innovations

Hierarchical Vision-Language-Action framework combining a world model as high-level planner and a VLA as low-level executor
Use of a large-scale pre-trained world model to synthesize goal images for visual subgoal task decomposition (VISTA)
Significant improvement in out-of-distribution generalization by providing visually and physically grounded goal images to low-level policies

Methodology

The proposed hierarchical framework VISTA consists of a world model as the high-level planner that divides manipulation tasks into subtask sequences with goal images, and a VLA model as the low-level executor that follows textual and visual guidance to generate action sequences. The world model is pre-trained at scale, and the system is validated in massive out-of-distribution scenarios using the same-structured VLA with and without world model guidance.

Key Results

The performance of the same-structured VLA in novel scenarios improved from 14% to 69% when guided by the world model, and the method outperformed previous baselines with a clear margin, particularly in out-of-distribution scenarios.

Resources

Paper Project