LOViS

Code for LOViS

This is the code for the paper ''LOViS: Learning Orientation and Visual Signals for Vision and Language Navigation''. The paper is accepted by COLING 2022.

Environment Installation

Please refer to Env-Dropout to install the Matterport3D simulators, download Room-to-room dataset and install the python environments.

Pre-training

Pre-training code is in pre-train file. You can download the pre-training weights Also, please refer to PREVALENT to download the pre-training data.

Run Agent

bash run/train_agent.bash 0

0 is the gpu id.

Test Agent

bash run/test_agent.bash 0

Best Results

The following is our best val_unseen result. Download Weights

Env name: val_unseen, nav_error: 3.5419, oracle_error: 2.2834, steps: 5.9506, lengths: 11.4193, success_rate: 0.6692, oracle_rate: 0.7361, spl: 0.6176