DAgger Diffusion Navigation: DAgger Boosted Diffusion Policy for Vision-Language Navigation

TODOs

• Release the evaluation code of the DifNav Model.
• Release the checkpoints of the DifNav Model for each scene.
• Release the online data augmentation code.
• Release the training data and code.

Requirements

1. Install Habitat simulator: follow instructions from ETPNav and VLN-CE.
2. Download the Matterport3D Scene Dataset (MP3D) from Matterport
3. Install the dependencies of the ETPNav and Nomad.
4. Download annotations and trained models from Google_Drive.
5. The data should be stored under the data folder with the following structure:

data-
├── scene_datasets
│ └── mp3d
├── datasets
│ └── R2R_VLNCE_v1-2_preprocessed_BERTidx
├── checkpoints
│ └── open_area.pth
└── ddppo-models

Evaluation

Evaluate each scene in our experiment(Open Area,Narrow Space,Stairs):

bash run_r2r/main.bash open_area eval 2333

Acknowledgement

• ETPNav: ETPNav implementation.
• Nomad: NoMaD implementation.