This repository contains the official implementation of paper Visual Navigation in Real-World Indoor Environments Using End-to-End Deep Reinforcement Learning.

Paper    Web    Demo

Before getting started, ensure, that you have Python 3.6+ ready. We recommend activating a new virtual environment for the repository:

python -m venv robot-visual-navigation-env
source robot-visual-navigation-env/bin/activate

Start by cloning this repository and installing the dependencies:

git clone https://github.com/jkulhanek/robot-visual-navigation.git
cd robot-visual-navigation
pip install -r requirements.txt
cd python

For DMHouse package, we recommend starting with Ubuntu 18+ and installing dependencies as follows:

apt-get install libsdl2-dev libosmesa6-dev gettext g++ unzip zip curl gnupg libstdc++6

You can download the pre-trained models from: https://data.ciirc.cvut.cz/public/projects/2021RealWorldNavigation/checkpoints/dmhouse-models.tar.gz https://data.ciirc.cvut.cz/public/projects/2021RealWorldNavigation/checkpoints/turtlebot-models.tar.gz

Download the pre-trained models using the following commands:

mkdir -p ~/.cache/robot-visual-navigation/models

# Download DMHouse models
curl -L https://data.ciirc.cvut.cz/public/projects/2021RealWorldNavigation/checkpoints/dmhouse-models.tar.gz | tar -xz -C ~/.cache/robot-visual-navigation/models

# Download real-world dataset models
curl -L https://data.ciirc.cvut.cz/public/projects/2021RealWorldNavigation/checkpoints/turtlebot-models.tar.gz | tar -xz -C ~/.cache/robot-visual-navigation/models

# Download real-world dataset
mkdir -p ~/.cache/robot-visual-navigation/datasets
curl -L -o ~/.cache/robot-visual-navigation/datasets/turtle_room_compiled.hdf5 https://data.ciirc.cvut.cz/public/projects/2021RealWorldNavigation/datasets/turtle_room_compiled.hdf5

Run the evaluation on the DMHouse simulator to verify that everything is working correctly:

python evaluate_dmhouse.py dmhouse --num-episodes 100

Similarly for the real-world dataset:

python evaluate_turtlebot.py turtlebot --num-episodes 100

Alternatively, you can also use other agents as described in the Training section.

Start the training by running ./train.py <trainer>, where trainer is the experiment you want to run. Available experiments are the following:

• dmhouse: our method (A2CAT-VN) trained with the dmhouse simulator
• dmhouse-unreal: UNREAL trained with the dmhouse simulator
• dmhouse-a2c: PAAC trained with the dmhouse simulator
• turtlebot: our method (A2CAT-VN) fine-tuned on the real-world dataset
• turtlebot-unreal: UNREAL fine-tuned on the real-world dataset
• turtlebot-a2c: PAAC fine-tuned on the real-world dataset
• turtlebot-noprior: our method (A2CAT-VN) trained on the real-world dataset; the model is trained from scretch
• turtlebot-unreal-noprior: UNREAL trained on the real-world dataset; the model is trained from scretch
• turtlebot-a2c-noprior: PAAC trained on the real-world dataset; the model is trained from scretch

All model checkpoints are available online:
https://data.ciirc.cvut.cz/public/projects/2021RealWorldNavigation/checkpoints

Please use the following citation:

@article{kulhanek2021visual,
  title={Visual navigation in real-world indoor environments using end-to-end deep reinforcement learning},
  author={Kulh{\'a}nek, Jon{\'a}{\v{s}} and Derner, Erik and Babu{\v{s}}ka, Robert},
  journal={IEEE Robotics and Automation Letters},
  volume={6},
  number={3},
  pages={4345--4352},
  year={2021},
  publisher={IEEE}
}