Recurrent VLN-BERT

Code of the CVPR 2021 Oral paper:
A Recurrent Vision-and-Language BERT for Navigation
Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez-Opazo, Stephen Gould

[Paper & Appendices]:new::new::new: [GitHub]

Prerequisites

Installation

Install the Matterport3D Simulator. Notice that this code uses the old version (v0.1) of the simulator, but you can easily change to the latest version which supports batches of agents and it is much more efficient.

Please find the versions of packages in our environment here.

Install the Pytorch-Transformers. In particular, we use this version (same as OSCAR) in our experiments.

Data Preparation

Please follow the instructions below to prepare the data in directories:

• MP3D navigability graphs: connectivity
  • Download the connectivity maps [23.8MB].
• R2R data: data
  • Download the R2R data [5.8MB].
• Augmented data: data/prevalent
  • Download the collected triplets in PREVALENT [1.5GB] (pre-processed for easy use).
• MP3D image features: img_features
  • Download the Scene features [4.2GB] (ResNet-152-Places365).

Initial OSCAR and PREVALENT weights

Please refer to vlnbert_init.py to set up the directories.

• Pre-trained OSCAR weights
  • Download the base-no-labels following this guide.
• Pre-trained PREVALENT weights
  • Download the pytorch_model.bin from here.

Trained Network Weights

• Recurrent-VLN-BERT: snap
  • Download the trained network weights [2.5GB] for our OSCAR-based and PREVALENT-based models.

R2R Navigation

Please read Peter Anderson's VLN paper for the R2R Navigation task.

Reproduce Testing Results

To replicate the performance reported in our paper, load the trained network weights and run validation:

bash run/test_agent.bash

You can simply switch between the OSCAR-based and the PREVALENT-based VLN models by changing the arguments vlnbert (oscar or prevalent) and load (trained model paths).

Training

Navigator

To train the network from scratch, simply run:

bash run/train_agent.bash

The trained Navigator will be saved under snap/.

Citation

If you use or discuss our Recurrent VLN-BERT, please cite our paper:

@article{hong2020recurrent,
  title={A Recurrent Vision-and-Language BERT for Navigation},
  author={Hong, Yicong and Wu, Qi and Qi, Yuankai and Rodriguez-Opazo, Cristian and Gould, Stephen},
  journal={arXiv preprint arXiv:2011.13922},
  year={2020}
}