Inverse reinforcement learning for autonomous navigation via differentiable semantic mapping and planning

[Project page] [Paper]

Installation

We assume that the following commands are run in a python virtual environment.

1. Install gym-minigrid from this fork

git clone git@github.com:tianyudwang/Minigrid.git
cd Minigrid
pip install -e .
cd ..

2. Download pybind11

git clone git@github.com:pybind/pybind11.git

3. Build the A* c++ code

cd astar_cpp
mkdir build
cd build
cmake ..
make

Return to project page and link

cd ../..
export PYTHONPATH=./astar_cpp/lib:${PYTHONPATH}

4. Generate dataset

python3 scripts/expert_policy.py --grid_size 16

The generated trajectories are under demonstration/ folder

5. Run training

python3 scripts/train.py --grid_size 16

The tensorboard logs with training metrics are in the logs/ folder and can be opened with

tensorboard --logdir logs

6. Run testing To run testing with pretrained models in the trained_models/ folder

python3 scripts/test.py --grid_size 16

This shows the success rate of rolling out the trained policy at each state.

Here are some visualizations of test episodes in 64 by 64 maps

Bibtex

@article{Wang2021sirl,
         author = {Wang, Tianyu and Dhiman, Vikas and Atanasov, Nikolay},
         title = {Inverse Reinforcement Learning for Autonomous Navigation 
                  via Differentiable Semantic Mapping and Planning},
         journal={arXiv preprint arXiv:2101.00186},
         year={2021}
         }