deep-robust-robotnav

Environment Setup

Install required packages:

conda env create -f config/vis_env.yaml python=3.6.13
conda activate vis5

Install habitat sim

conda install -c aihabitat -c conda-forge bullet=2.88 habitat-sim=0.1.6 headless withbullet
pip install cupy # with your cuda version given by nvcc -V
conda install cudatoolkit
conda install cudnn

Setup habitat lab folder

Build habitat lab from source using the specified commit hash (the folder is already in the repository, no need to git clone or download).

cd habitat-api
python setup.py develop --all
cd ..

You may need to install the following libraries on Ubuntu if not already installed

sudo apt install freeglut3 freeglut3-dev

Get Gibson Dataset

wget https://dl.fbaipublicfiles.com/...
wget https://dl.fbaipublicfiles.com/...

Run Train / Eval:

Remember to activate conda environment before running command.

python run.py --exp-config config/habitat_config.yaml --run-type train

Experiment setup

• Same PPO settings everywhere
  • BATCH_SIZE = 32
  • Learning rate = 2.5e-4
• Train dataset, all maps in the gibson dataset starting with the letter 'A'
• Experiments
  1. Baseline midlevel
  2. Ground-truth map
  3. Mapper network form midlevel

TODO

Map vector in rnn hidden states Fix transform relating to map size (Titouan) Training with good subset of data?