Yiding Qiu*, Anwesan Pal* and Henrik I. Christensen

Paper | Website | Video

PyTorch implementation of our CoRL 2020 paper Learning hierarchical relationships for object-goal navigation in AI2-THOR environment. This implementation is modified based on SAVN.

1. The code has been implemented and tested on Ubuntu 18.04, python 3.6, PyTorch 0.6 and CUDA 10.1
2. (Recommended) Create a virtual environment using virtualenv or conda:

virtualenv mjolnir_env --python=python3.6
source mjolnir_env/bin/activate

conda create -n mjolnir_env python=3.6
conda activate mjolnir_env

3. For the rest of dependencies, please run pip install -r requirements.txt --ignore-installed
4. Clone the repository as:

    git clone https://github.com/cassieqiuyd/MJOLNIR.git

Note: Upon running any code the first time, the AI2-THOR 3D scenes will be downloaded (~500MB) to your home directory.

The offline data can be found here.

"data.zip" (~5 GB) contains everything needed for evalution. Please unzip it and put it into the MJOLNIR folder.

For training, please also download "train.zip" (~9 GB), and put all "Floorplan" folders into MJOLNIR/data/thor_v1_offline_data

Note: the knowledge graph is already provided in the data set downloaded from above. The following instructions show how these files are generated.

1. Run the following script to generate the list of iThor v1.0.1 objects.

python kg_prep/object_list.py

This will generate a .txt file in MJOLNIR/kg_prep/kg_data/.

2. Download the Visual Genome relationships file.

wget https://visualgenome.org/static/data/dataset/relationships.json.zip -P kg_prep/kg_data/
unzip kg_prep/kg_data/relationships.json.zip -d kg_prep/kg_data/ && rm -rf kg_prep/kg_data/relationships.json.zip

3. Download the relationship aliases.

wget https://visualgenome.org/static/data/dataset/relationship_alias.txt -P kg_prep/kg_data/

4. Run the relationship denoising step which will prune the object relationships of Visual-Genome dataset for the objects in iThor.

python kg_prep/kg_denoising.py

This will generate several files containing the statistics of the knowledge graph.

5. Generate adjacency matrix.

python kg_prep/make_adjacency.py

This will generate the adjacency matrix file adjmat.dat file inside MJOLNIR/kg_prep/kg_data/. This is the same file which can be found inside MJOLNIR/data/gcn folder downloaded in the Data section.

Note: If you want the adjacency matrix to be weighted, pass the --weighted argument to the above command line.

Note: if you are not using gpu, you can remove the argument --gpu-ids 0

python main.py --eval \
    --test_or_val test \
    --episode_type TestValEpisode \
    --load_model pretrained_models/mjolnir_o_pretrain.dat \
    --model MJOLNIR_O \
    --results_json mjolnir_o.json \
    --gpu-ids 0

This should also generate a log of the actions taken by the agents (required for visualization)

To print the evaluation results,

cat mjolnir_o.json 

If you did evaluation, the action log should be generated.

cd visualization
python visualization.py --actionList ../saved_action_mjolnir_o_test.log

python main.py \
    --title mjolnir_train \
    --model MJOLNIR_O \
    --gpu-ids 0\
    --workers 8
    --vis False
    --save-model-dir trained_models

Other model options are "SAVN" and "GCN"

python full_eval.py \
    --title mjolnir \
    --model MJOLNIR_O \
    --results_json mjolnir_o.json \
    --gpu-ids 0
    --save-model-dir trained_models
    
cat mjolnir_o.json

Please cite our work if you found this research useful for your work:

@article{qiu2020learning,
  title={Learning hierarchical relationships for object-goal navigation},
  author={Qiu, Yiding and Pal, Anwesan and Christensen, Henrik I},
  journal={arXiv preprint arXiv:2003.06749},
  year={2020}
}