AwareVLN: Reasoning with Self-awareness for Vision-Language Navigation

Forum topic · 小凯 · 2026-05-23

Summary

AwareVLN is a new vision-language navigation (VLN) framework that equips navigation models with self-aware reasoning, enabling them to understand agent state and task progress in a fully end-to-end, data-driven manner. Proposed by Wenxuan Guo, Xiuwei Xu, and Yichen Liu (arXiv:2505.17383, May 2025), the work addresses two limitations of existing approaches: end-to-end VLM-based methods lack explicit, interpretable understanding of the relationships between agent, instruction, and scene, while explicit map-building approaches require extra 3D sensors and hinder large-scale vision-language pretraining. AwareVLN introduces two key innovations: (1) a structured reasoning module that cultivates spatial and task-oriented self-awareness, and (2) an automatic data engine with progress partitioning for efficient training. Experiments across multiple datasets in the Habitat simulator show that AwareVLN significantly outperforms prior state-of-the-art VLN methods.

Paper Overview

Field: Computer Vision (CV) Authors: Wenxuan Guo, Xiuwei Xu, Yichen Liu Published: 2025-05-23 arXiv: 2505.17383

Background

Vision-Language Navigation (VLN) requires an agent to ground language instructions into its own motion within visual environments. Existing approaches face a trade-off:

End-to-end VLM-based methods leverage the reasoning capabilities of vision-language models for action prediction, but often lack explicit, interpretable understanding of the relationships between the agent, the instruction, and the scene.
Explicit scene-map construction for heuristic planning is intuitively appealing, but relies on additional 3D sensors and hinders large-scale vision-language pretraining.

Method

AwareVLN bridges this gap with a framework that equips navigation models with self-aware reasoning, allowing them to understand agent state and task progress in a fully end-to-end and data-driven manner. The approach has two key innovations:

1. A structured reasoning module that cultivates spatial and task-oriented self-awareness. 2. An automatic data engine with progress partitioning for efficient training.

Results

Extensive experiments on multiple datasets in the Habitat simulator demonstrate that AwareVLN significantly outperforms previous state-of-the-art vision-language navigation methods.