Code release for the CoRL 2019 paper: HRL4IN: Hierarchical Reinforcement Learning for Interactive Navigation with Mobile Manipulators

The PPO implementation is partially adopted from pytorch-a2c-ppo-acktr-gail and habitat-api

Dependency

• torch==1.1.0
• torchvision==0.2.2

Installation

1. Install iGibson with the archive/hrl4in tag.

cd $HOME
git clone https://github.com/StanfordVL/iGibson.git
cd $HOME/iGibson
git checkout archive/hrl4in
git submodule init
git submodule update
pip install -e .

2. Download iGibson assets.

wget https://storage.googleapis.com/gibson_scenes/assets_igibson.tar.gz -O /tmp/assets_igibson.tar.gz
tar -zxf /tmp/assets_igibson.tar.gz --directory $HOME/iGibson/gibson2
rm /tmp/assets_igibson.tar.gz

3. Install HRL4IN

cd $HOME
git clone --recursive https://github.com/ChengshuLi/HRL4IN.git 
cd $HOME/HRL4IN
pip install -e .

4. Copy the updated JR URDF file from this repo to iGibson's asset folder

cp $HOME/HRL4IN/hrl4in/envs/gibson/jr2_kinova.urdf $HOME/iGibson/gibson2/assets/models/jr2_urdf/jr2_kinova.urdf

Usage

Train in ToyEnv

./run_train_toy_env.sh

Train in iGibson

./run_train_gibson.sh

Eval in ToyEnv

./run_eval_toy_env.sh

Eval in iGibson

./run_eval_gibson.sh

Citation

If you use this code, please cite it as:

@article{li2019hrl4in,
  title={HRL4IN: Hierarchical Reinforcement Learning for Interactive Navigation with Mobile Manipulators},
  author={Li, Chengshu and Xia, Fei and Martin-Martin, Roberto and Savarese, Silvio},
  journal={arXiv preprint arXiv:1910.11432},
  year={2019}
}