README.md

Diagnose_VLN

This repository contains the code and data for our paper: Diagnosing Vision-and-language Navigation: What Really Matters.

Directory Structure

We cover three VLN datasets and nine agents in our study.

• data_processing/
  • process_instructions/: scripts to prepare/download instructions
  • Matterport3DSimulator: a copy of Matterport simulator, and scripts to prepare/download R2R/RxR image features
• r2r/: Data and code for experiments on Room-to-room (R2R) for indoor VLN
  • data/
  • model/
    • R2R-EnvDrop
    • FAST
    • Recurrent-VLN-BERT
    • PREVALENT_R2R
• rxr/: Data and code for experiments on Room-across-room (RxR) for indoor VLN
  • data/
  • model/
    • CLIP-ViL-VLN
    • VLN-HAMT
• touchdown/: Data and code for experiments on Touchdown for outdoor VLN
  • data/
  • model/
    • RCONCAT
    • ARC
    • VLN-Transformer

Installation

git clone --recursive https://github.com/VegB/Diagnose_VLN

We describe the detailed environment setup for each model in the corresponding directory. For instance, guidance to setup R2R-EnvDrop can be found here.

Data Preparation

• Prepare R2R data: link
• Prepare RxR data: link
• Prepare Touchdown data: link

Acknowledgements

We thank the authors for Matterport3DSimulator, R2R-EnvDrop, FAST, Recurrent-VLN-BERT, PREVALENT_R2R, CLIP-ViL-VLN, VLN-HAMT, RCONCAT, ARC, VLN-Transformer for sharing their code!