zhangyuejoslin / SpC-NAV

Code and Data for Paper "Towards Navigation by Reasoning over Spatial Configurations"

Our paper is accepted in SpLU workshop of ACL2021.
paper: https://aclanthology.org/2021.splurobonlp-1.5/

Environment Installation

Flease follow https://github.com/airsplay/R2R-EnvDrop to set up the environment.

Run Agent

bash run/agent.bash 0

0 is the id of GPU. It will train the agent and save the snapshot under snap/agent/. Unseen success rate would be around 46%.

bash run/bt_envdrop.bash 0

0 is the id of GPU. It will load the pre-trained agent and run back translation with environmental dropout.

Spatial Configuration, Motion Indicator, Landmark Features

Modify the parameters in param.py to the corresponding path, and the following link is the corresponding features. https://drive.google.com/file/d/1Mf2yiP4mdTUwSn6r9aNiCjcsqLsoYRHM/view?usp=sharing