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

Safe Planning and Policy Optimization via World Model Learning

arXiv 25.6 2025 50.7 method

TLDR

A model-based RL framework jointly optimizes safety and performance using adaptive planning, implicit world models, and dynamic safety thresholds.

Reasoning

The paper presents a novel approach to safe RL by addressing world model errors and objective mismatch, with strong empirical results on safety-critical tasks. However, the abstract lacks details on the specific environments and baselines, and the implicit world model concept is not fully explained.

Read-first score

Read-first score 50.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 31.

Recency 8%
86.7

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

Methodology quality 25%
70

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

Topical relevance 42%
44.3

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

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: 366.

Keyword Scores

world model
10
model-based reinforcement learning world model
9
world dynamics prediction
5
world simulator
2
generative world model
2
interactive world model
2
video world model
1

Deep Analysis

Innovations

  • Joint optimization of task performance and safety in a model-based RL framework
  • Adaptive mechanism that dynamically switches between model-based planning and direct policy execution to address world model errors
  • Implicit world model to resolve the objective mismatch problem of traditional model-based approaches
  • Dynamic safety thresholds that adapt to the agent's evolving capabilities

Methodology

The proposed framework is a model-based RL approach that learns a world model and jointly optimizes task performance and safety. It incorporates an adaptive mechanism to switch between model-based planning and direct policy execution, uses an implicit world model to resolve objective mismatch, and employs dynamic safety thresholds that adapt to the agent's capabilities. The method is evaluated on diverse safety-critical continuous control tasks against non-adaptive and existing methods.

Key Results

The framework achieves significant improvements over non-adaptive methods, optimizing safety and performance simultaneously rather than merely meeting minimum safety requirements. It outperforms existing methods on diverse safety-critical continuous control tasks.

Limitations

  • The adaptive switching mechanism and dynamic thresholds may introduce additional complexity and require careful tuning
  • Evaluation is limited to continuous control tasks; generalization to discrete or other domains is not demonstrated
  • The method still relies on world model learning, and severe model inaccuracies could potentially degrade performance despite the adaptive mechanism

Resources

Paper

Tags