A Recipe for Efficient Sim-to-Real Transfer in Manipulation with Online Imitation-Pretrained World Models

arXiv 25.10 2025 46.4 method, application

TLDR

A sim-to-real framework using world models for online imitation pretraining and offline finetuning improves manipulation transfer with limited real-world data.

Reasoning

The paper clearly addresses a practical problem (limited real-world data) and proposes a novel combination of online imitation pretraining with world models, showing significant empirical gains in both sim-to-sim and sim-to-real transfers. However, the abstract lacks details on the world model architecture and the specific real-world experimental setup, limiting full assessment of methodology.

Read-first score

Read-first score 46.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 28.

Recency 8%

86.7

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

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=none

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 375.

Keyword Scores

world model

world simulator

world dynamics prediction

model-based reinforcement learning world model

interactive world model

generative world model

video world model

Deep Analysis

Innovations

Combining online imitation pretraining with offline finetuning in a world model framework for sim-to-real transfer
Using online interactions to alleviate data coverage limitations of offline imitation learning methods
Achieving significant improvement in success rates (≥31.7% sim-to-sim, ≥23.3% sim-to-real) over offline baselines

Methodology

The proposed sim-to-real framework is based on world models and combines online imitation pretraining using a robot simulator with offline finetuning on limited real-world expert data. Online interactions during pretraining improve data coverage and robustness, reducing performance degradation during finetuning and enhancing domain transfer.

Key Results

The method improves success rates by at least 31.7% in sim-to-sim transfer and 23.3% in sim-to-real transfer compared to existing offline imitation learning baselines.

Resources

Paper