Interactive World Simulator for Robot Policy Training and Evaluation

arXiv 2026 75.9 method, system, application

TLDR

Presents Interactive World Simulator, a fast consistency-model-based world model for robot policy training, enabling stable long-horizon simulation and real-world comparable performance.

Reasoning

The paper introduces a novel framework for building interactive world models using consistency models, achieving fast and stable simulation. It demonstrates strong real-world evaluation and correlation between simulated and real performance, but lacks details on limitations and comparisons to other world model approaches.

Read-first score

Read-first score 75.9, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 64.

Recency 8%

100

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

Topical relevance 42%

91.4

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

Methodology quality 25%

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

Reproducibility 25%

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

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 56.

Keyword Scores

world model

world simulator

interactive world model

generative world model

video world model

world dynamics prediction

model-based reinforcement learning world model

Deep Analysis

Innovations

Leverages consistency models for image decoding and latent-space dynamics prediction, enabling fast and stable simulation.
Interactive World Simulator framework builds interactive world models from a moderate-sized robot interaction dataset.
Supports stable long-horizon interactions for more than 10 minutes at 15 FPS on a single GPU.
Enables scalable demonstration collection purely inside the world model to train state-of-the-art imitation policies.
Demonstrates that policies trained on world-model-generated data perform comparably to those trained on the same amount of real-world data.
Shows strong correlation between simulated and real-world policy performance, enabling faithful, reproducible evaluation.

Methodology

The framework uses consistency models for both image decoding and latent-space dynamics prediction, trained on a moderate-sized robot interaction dataset. It generates demonstrations solely within the learned world model to train imitation policies. Real-world evaluation covers tasks with rigid objects, deformable objects, object piles, and their interactions.

Key Results

The learned world models achieve interaction-consistent pixel-level predictions and sustain stable long-horizon interactions over 10 minutes at 15 FPS on a single RTX 4090 GPU. Policies trained on world-model data perform comparably to real-data-trained policies, and simulated performance strongly correlates with real-world performance across diverse tasks.

Resources

Paper