ubastic / a2cat-vn

Target driven visual navigation using deep reinforcement learning implemented in Pytorch

Vision-based Navigation Using Deep Reinforcement Learning

Official implementation of A2CAT-VN. A reinforcement learning architecture capable of navigating an agent, e.g. a mobile robot, to a target given by an image. It extends the batched A2C algorithm with auxiliary tasks designed to improve visual navigation performance.

Paper    Web

Getting started

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

python -m venv a2catvn-env
source a2catvn-env/bin/activate

Start by cloning this repository and installing the dependencies:

git clone https://github.com/jkulhanek/a2cat-vn.git
cd a2cat-vn
pip install -r requirements.txt

For discrete AI2THOR experiments, you can speed up the loading of the dataset by downloading the pre-computed dataset:

mkdir -p ~/.cache/visual-navigation/datasets
for package in thor-cached-212 thor-cached-208 thor-cached-218 thor-cached-225 thor-cached-212-174 thor-cached-208-174 thor-cached-218-174 thor-cached-225-174; do
    curl -L -o ~/.cache/visual-navigation/datasets/$package.pkl https://data.ciirc.cvut.cz/public/projects/2019VisionBasedNavigation/resources/$package.pkl
done

NOTE: SUNCG dataset is not longer available and we cannot provide dataset samples.

Training

In order to start the training, run the following command:

python train.py {trainer}

where {trainer} is the name of the experiment and can be one of the following:

• thor-cached-auxiliary
• cthor-multigoal-auxiliary
• chouse-auxiliary-superviised (requires SUNCG dataset which is no longer publicly available!)
• chouse16-auxiliary (requires SUNCG dataset which is no longer publicly available!)

For chouse* experiments, you need to have House3D simulator installed and SUNCG dataset downloaded. We recommend using provided docker image.

Model checkpoints

Model checkpoints are available online here:
https://data.ciirc.cvut.cz/public/projects/2019VisionBasedNavigation/resources/model-checkpoints.tar.gz