Temporal Logic Guidance for Action-Only Diffusion Policies with World Models

arXiv 2026 54.1 method

TLDR

A method using a learned world model to guide action-only diffusion policies via STL robustness gradients, reducing constraint violations from 80% to 4% on a robot task.

Reasoning

The paper presents a novel integration of world models with STL-guided diffusion, achieving strong empirical results on a single benchmark task. However, the evaluation is limited to one task and the method's scalability to complex constraints is only discussed, not demonstrated.

Read-first score

Read-first score 54.1, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 33.

Recency 6%

100

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

Citation impact 18%

94.8

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.94825964

Methodology quality 18%

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

Topical relevance 29%

47.1

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=none

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: 402.

Keyword Scores

world model

world dynamics prediction

world simulator

model-based reinforcement learning world model

generative world model

interactive world model

video world model

Deep Analysis

Innovations

Guidance method for action-only diffusion policies using a separate learned world model to enable differentiable evaluation of STL robustness
Gradient injection from STL robustness into the diffusion process to steer behavior toward constraint satisfaction without retraining
Maintaining task performance while significantly reducing constraint violations compared to baseline methods

Methodology

The method uses an action-only diffusion policy and a separate learned world model to compute differentiable Signal Temporal Logic (STL) robustness. The gradient of the STL robustness with respect to the action sequence is injected into the diffusion process during inference, steering behavior toward constraint satisfaction without requiring retraining. Evaluation is performed on the Can Transport task from Robomimic, measuring task success and constraint violation rates.

Key Results

On the Can Transport task, the method maintains 100% task success while reducing constraint violations from over 80% for baseline methods to 4%.

Limitations

Requires a separate learned world model, which may introduce additional complexity and potential inaccuracies
Currently only tested on a single task (Can Transport) from Robomimic
Extensions are discussed for improved robustness and handling of more complex constraints, indicating current limitations in these areas

Resources

Paper