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

WMPO: World Model-based Policy Optimization for Vision-Language-Action Models

arXiv 25.11 2025 63.3 method, application

TLDR

WMPO uses a pixel-based world model to enable on-policy reinforcement learning for VLA robots without real environment interaction, improving sample efficiency and performance.

Reasoning

The paper introduces a novel framework that combines world models with on-policy GRPO for VLA robots, showing strong results in simulation and real-world settings. However, the abstract lacks details on world model architecture and potential limitations of relying on pretrained VLA features.

Read-first score

Read-first score 63.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 48.

Recency 8%
86.7

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

Methodology quality 25%
80

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

Topical relevance 42%
68.6

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%
30

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: 170.

Keyword Scores

world model
10
model-based reinforcement learning world model
10
world dynamics prediction
8
generative world model
7
world simulator
6
video world model
5
interactive world model
2

Deep Analysis

Innovations

  • World-model-based policy optimization for Vision-Language-Action (VLA) models that enables on-policy RL without real environment interaction
  • Pixel-based world model predictions aligned with pretrained VLA features to generate imagined trajectories
  • Application of on-policy GRPO (Group Relative Policy Optimization) for VLA RL, providing stronger performance than off-policy methods

Methodology

WMPO trains a pixel-based world model that predicts future observations and rewards, aligning imagined trajectories with features from a pretrained VLA model. The policy is then optimized using on-policy GRPO entirely within the world model, avoiding real-robot interactions. The framework is evaluated in both simulation and real-robot settings using standard manipulation benchmarks.

Key Results

WMPO substantially improves sample efficiency and achieves stronger overall performance compared to baselines. It also exhibits emergent self-correction behaviors and demonstrates robust generalization and lifelong learning capabilities.

Limitations

  • World model inaccuracies may lead to policy exploitation of model errors, potentially causing sim-to-real gaps
  • Dependence on pretrained VLA features may limit applicability to domains where such features are not available
  • On-policy GRPO may require careful hyperparameter tuning and could be computationally expensive due to pixel-level predictions

Resources

Paper

Tags