README.md

metadata

task_categories:
  - robotics
license: mit
tags:
  - vision-and-language-navigation
  - vln
  - embodied-ai

StereoNav Dataset

This repository contains the training and evaluation datasets for StereoNav, as presented in the paper What Limits Vision-and-Language Navigation?.

Project Page | GitHub Repository

Dataset Description

StereoNav is a robust Vision-Language-Action framework designed to enhance real-world navigation consistency. It addresses gaps between synthetic training and physical execution using Target-Location Priors and stereo vision for enhanced depth awareness. This repository provides the pre-processed data used for the two-stage training pipeline described in the paper.

Data Structure

The pre-processed data should be placed under data/cache/ following the structure required by the training scripts:

data/cache/
└── train/
    ├── R2R/                  
    │   ├── images/
    │   ├── annotations_*.json
    │   ├── summary.json
    │   ├── baseline.txt
    │   └── intrinsics.txt
    ├── RxR/
    ├── ScaleVLN/
    └── dagger/
        ├── R2R/
        ├── RxR/
        ├── baseline.txt
        └── intrinsics.txt

Components:

• Stage 1 Data: Standard VLN datasets (R2R and RxR).
• Stage 2 Data: Joint training data including ScaleVLN and DAgger-collected trajectories for improved generalization and robustness.

Citation

If you find this dataset or the associated work useful, please consider citing:

@article{stereonav2026,
  title     = {What Limits Vision-and-Language Navigation?},
  author    = {Yunheng Wang and Yuetong Fang and Taowen Wang and Lusong Li and Kun Liu and Junzhe Xu and Zizhao Yuan and Yixiao Feng and Jiaxi Zhang and Wei Lu and Zecui Zeng and Renjing Xu},
  journal   = {arXiv preprint},
  year      = {2026},
}