Object Navigation in a 3D Rendered Home Environment with Multi-View Reinforcement Learning

CISC 856 Final Project Group 27

Example of a trained agent

To train Egocentric Agent:

python src/main.py --mode train_ego

To train Allocentric Agent:

python src/main.py --mode train_alo

To train Double Agent:

python src/main.py --mode train_double

To test a policy:

python src/main.py --mode test

To output saliency maps:

python visuals/saliency_map.py

The repository contains the following:

src contains the necessary python files used to train the algorithm. Follow the instructions above to run.
- agent.py contains DQN agent class
- network.py contrains all Neuarl Network class
- main.py contains a commandline program for our code
- solve.py contrains the traing and enviorment code for AI2-THOR
model_samples contains model checkpoints for the easy and hard conditions each trained on egocentric only, allocentric only and combined perspective.
visuals contains the necessary files to generate saliency maps using the model checkpoints