This directory contains the Subgoal Prediction model that is trained on Matterport3D to predict the location of adjacent graph nodes in the Matterport graph from the 360 image and a 270 degree flat laser scan.
The laser scan for each pano has been generated in advance using code from the laser_scan directory.
First set up some symlinks to the Matterport3D dataset and the Matterport3D simulator. From the top-level directory run:
ln -s <MATTERPORT3D_DATA_DIR> actions/data
ln -s <MATTERPORT3D_SIMULATOR> Matterport3DSimulator
Note that the MATTERPORT3D_DATA_DIR must contain generated laser scans (see the laser_scan subdirectory).
Example command to train and validate the model (from the top-level directory):
cd actions
python3 train.py