Evaluating Robot Policies in a World Model

arXiv 25.5 2025 69.4 method, system

TLDR

WorldGym uses an action-conditioned video world model to evaluate robot policies, showing high correlation with real-world success rates.

Reasoning

The paper introduces a novel world-model-based evaluation environment that correlates well with real-world performance, enabling safe and reproducible policy testing. However, it acknowledges limitations in object interaction realism and susceptibility to adversarial facades.

Read-first score

Read-first score 69.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 60.

Recency 8%

86.7

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

Topical relevance 42%

85.7

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=evaluation

Reproducibility 25%

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

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 214.

Keyword Scores

world model

video world model

generative world model

interactive world model

world simulator

world dynamics prediction

model-based reinforcement learning world model

Deep Analysis

Innovations

Proposes WorldGym, an autoregressive action-conditioned video generation model as a proxy for real-world robot policy evaluation
Uses a vision-language model to provide rewards during Monte Carlo rollouts in the world model
Demonstrates that policy success rates in the world model highly correlate with real-world success rates
Shows that WorldGym preserves relative policy rankings across different policy versions, sizes, and training checkpoints
Enables efficient evaluation of robot policies' generalization ability on novel tasks and environments using only a single start frame

Methodology

WorldGym is an autoregressive, action-conditioned video generation model that serves as a proxy environment. Policies are evaluated via Monte Carlo rollouts in the world model, with a vision-language model providing rewards. The approach uses only initial frames from real robots and evaluates a set of VLA-based real-robot policies.

Key Results

Policy success rates within the world model highly correlate with real-world success rates, and WorldGym preserves relative policy rankings across different policy versions, sizes, and training checkpoints.

Limitations

Generating highly realistic object interaction remains challenging
Modern VLA-based robot policies still struggle to distinguish object shapes
Policies can become distracted by adversarial facades of objects

Resources

Paper Demo