This repo consists of Tensorflow 1.x code to train an agent to navigate in the Habitat environment (https://arxiv.org/abs/1904.01201). 


0. Resources: 

Habitat paper:
https://arxiv.org/abs/1904.01201

Habitat-Sim:
https://github.com/facebookresearch/habitat-sim

Habitat-Api
https://github.com/facebookresearch/habitat-api


NOTE: this code was tested in Tensorflow 1.8, Cuda 9.0, Ubuntu 18.04


1. Download Gibson data files:

Download the Gibson Quantico files and place them in some directory called data/
data/ should contain the two sub dirs:

datasets
scene_datasets
 
Either data/ should be present in the location of this code directory, otherwise create a symbolic link and place it here.
We will focus on the Quantico environment in the Gibson dataset. You can create a similar directory structure and place only Quantico in it; alternatively, if you prefer to use any other Gibson environment, or even multiple such Gibson environments, place them appropriately in this directory.


2. Set the path in the yaml file:

Go to: pointnav_gibson_kb.yaml and set the following path as appropriate:
DATA_PATH: /home/kb/habitat_experiments/data/datasets/pointnav/gibson/v1/{split}/{split}.json.gz


3. Collect raw RGB and Depth images by running the following:

cd collect_imgs/
mkdir imgs
mkdir imgs/depth
mkdir imgs/rgb
python collect_imgs2.py

This will save the raw RGB and Depth images by randomly spawning the agent and rotating it 360 degrees to get a full view of the surroundings.


4. Train the VAEs

cd vae/

Open train_vae.py and set: 
img_train = 'rgb' 
python train_vae.py
This will train the VAE for RGB images - this takes a few hours

Then, open train_vae.py and set:
img_train = 'depth'
python train_vae.py
This trains the VAE for Depth images - this also takes a few hours


5. To train the PPO-LSTM agent, go to the code main directory (which is one "../" from the VAE directory)

Open hab1.py and set the path to where the yaml is (it is the code base directory if you place the yaml as done so in this repo):   
pwd = "/home/kb/habitat_experiments/code/vae_128/t1/"

For a clean training start from scratch (i.e., not loading a previous ckpt file), set the following at the top of hab1.py:
load_ckpt = False
ep_start = 0

Train the Reinforcement Learning agent by typing:

python hab1.py

This can take a few days to train as it is a very intensive run!
See the attached SPL plot - the blue is the best performing agent with a success rate of over 90% and a mean SPL of about 0.77.


6. After the training is complete, to test the agent, open hab1_test.py and set the path to the yaml file:

pwd = "/home/kb/habitat_experiments/code/vae_128/t1/"

To test the agent, run this file:

python hab1_test.py



POTENTIAL ERRORS

1. When this code was developed, the actions were stored in SimulatorActions, which has since been moved to HabitatSimActions by the developers of Habitat
To fix this, import the following instead of SimulatorActions:
from habitat.sims.habitat_simulator.actions import HabitatSimActions
Also replace every occurrence of 'SimulatorActions' with 'HabitatSimActions'

2. pointgoal: this code was developed using a Habitat version where state['pointgoal'] was a list of 3 entries, i.e., 3D, with the middle entry being useless. In the newer Habitat, the developers have made this 2D. 
Fix this by commenting out pointgoal = [pointgoal[0], pointgoal[2]]:
pointgoal = s["pointgoal"]; #pointgoal = [pointgoal[0], pointgoal[2]]




TO DO:

Also provided is the Astar code for the same problem to plan path in a top-down map. Plan is to couple RL with Astar.