WMNav: Integrating Vision-Language Models into World Models for Object Goal Navigation

arXiv 25.3 2025 65.3 method

TLDR

WMNav integrates VLMs into world models for object goal navigation, predicting future states and using a Curiosity Value Map to improve success rate and exploration efficiency.

Reasoning

The paper presents a novel framework combining VLMs with world models for navigation, with strong empirical results on standard benchmarks (HM3D, MP3D). However, the abstract lacks detailed discussion of limitations and does not explicitly address several keyword concepts like generative or interactive world models.

Read-first score

Read-first score 65.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 30.

Recency 8%

86.7

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

Reproducibility 25%

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

Methodology quality 25%

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

Topical relevance 42%

42.9

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

Field roles

FrontierMethodology anchorReproducibility anchor

Rank sensitivity

Stability: volatile; rank range: 501.

Keyword Scores

world model

world dynamics prediction

model-based reinforcement learning world model

interactive world model

world simulator

generative world model

video world model

Deep Analysis

Innovations

Fully modular world model design for object goal navigation using Vision-Language Models (VLMs) that predicts future states to reduce risky interactions.
Online maintained Curiosity Value Map as part of the world model memory to provide dynamic configuration for navigation policy.
Human-like thinking process that alleviates model hallucination by making decisions based on feedback difference between the world model plan and observation.
Two-stage action proposer strategy: broad exploration followed by precise localization to boost efficiency.

Methodology

WMNav integrates Vision-Language Models into a world model framework for object goal navigation. It uses an online maintained Curiosity Value Map as memory and a two-stage action proposer strategy (broad exploration then precise localization). The system predicts future states and uses feedback difference between world model plan and observation to reduce hallucination. Evaluation is conducted on HM3D and MP3D datasets, comparing against zero-shot benchmarks using success rate (SR) and SPL.

Key Results

WMNav surpasses existing zero-shot benchmarks, achieving absolute improvements of +3.2% SR and +3.2% SPL on HM3D, and +13.5% SR and +1.1% SPL on MP3D.

Limitations

Model hallucination is alleviated but not fully eliminated; the framework still relies on the feedback difference mechanism to mitigate its impact.
The evaluation is limited to HM3D and MP3D datasets; generalization to other environments is not demonstrated.
The framework's reliance on VLM inference may introduce computational overhead, though not explicitly quantified in the abstract.

Resources

Paper Code