Exploitation-Guided Exploration for Semantic Embodied Navigation

Broadly speaking this work is concerned with using a geometric policy that is very good at solving some portion of a task to guide exploration of a neural policy.

Setting up Anaconda

# If you haven't already git cloned then - 
# git clone git@github.com:Jbwasse2/XGX.git
# Create anaconda environment
conda create --name xgx python=3.7
# Install Habitat-Sim and Scikit-fmm
conda install habitat-sim=0.2.1 withbullet headless -c conda-forge -c aihabitat
conda install -c conda-forge scikit-fmm=2019.1.30
# Install Habitat-Lab
cd habitat-lab
pip install -r requirements.txt
python setup.py develop --all # install habitat and habitat_baselines
cd ..
# Install your torch with your version of cuda, we use cu113
# See https://pytorch.org/get-started/previous-versions/ for commands
# if your version of cuda does not match
pip install torch==1.10.0+cu113 torchvision==0.11.0+cu113 torchaudio==0.10.0 -f https://download.pytorch.org/whl/torch_stable.html
pip install -r requirements.txt

Setting up data and models

HM3D Dataset

We utilzie the HM3D-V1 dataset. Details of this dataset can be found here. This dataset should be placed into ./data/scene_dataset/ yielding for example data/scene_datasets/hm3d/val/00877-4ok3usBNeis/4ok3usBNeis.basis.glb

We also utilize the standard HM3D-V1 train/val splits. This can be found here. These splits should be placed into ./data/datasets/objectnav/ yielding as an example data/datasets/objectnav/hm3d/v1/val/content/4ok3usBNeis.json.gz.

Models

We retrain RedNet to the HM3D dataset. We also used XGX to retrain a CNN+RNN model.

Both of these models can be found here.

Put both of these models in ./models.

Evaluation

Your code base should look like this

 .
 ├── arguments.py
 ├── category_mapping.tsv
 ├── configs
 │   ├── experiments
 │   └── tasks
 ├── data
 │   ├── scene_dataset
 │   └── datasets
 ├── habitat-lab
 ├── habitat-sim
 ├── models
 │   ├── hm3d_rednet.pt
 │   └── XGX.pth
 ├── pirlnav
 ├── README.md
 ├── run.py
 └── src 

Now run

python run.py --exp-config ./configs/experiments/XGX.yaml --run-type eval

After running this we recorded the following results

Average episode reward: 0.7275
Average episode distance_to_goal: 2.4974
Average episode success: 0.7275
Average episode spl: 0.3613
Average episode softspl: 0.3906
Average episode sparse_reward: 1.8188
Average episode num_steps: 160.5845