Autonomous Exploration Under Uncertainty via Deep Reinforcement Learning on Graphs
This repository contains code for robot exploration under uncertainty that uses graph neural networks (GNNs) in conjunction with deep reinforcement learning (DRL), enabling decision-making over graphs containing exploration information to predict a robot’s optimal sensing action in belief space. A demonstration video can be found here.
Dependency
- Python 3
- PyTorch
- PyTorch Geometric
- gtsam (Georgia Tech Smoothing and Mapping library)
git clone -b emex --single-branch https://bitbucket.com/jinkunw/gtsam cd gtsam mkdir build && cd build cmake .. sudo make install - pybind11 (pybind11 — Seamless operability between C++11 and Python)
git clone https://github.com/pybind/pybind11.git cd pybind11 mkdir build && cd build cmake .. sudo make install
Compile
You can use the following commands to download and compile the package.
git clone https://github.com/RobustFieldAutonomyLab/DRL_graph_exploration.git
cd DRL_graph_exploration
mkdir build && cd build
cmake ..
make
Issues
There is an unsolved memory leak issue in the C++ code. So we use the python subprocess module to run the simulation training. The data in the process will be saved and reloaded every 10000 iterations.
How to Run?
- To run the saved policy:
cd DRL_graph_exploration/scripts python3 test.py - To show the average reward during the training:
cd DRL_graph_exploration/data tensorboard --logdir=torch_logs - To train your own policy:
cd DRL_graph_exploration/scripts python3 train.py
Cite
Please cite our paper if you use any of this code:
@inproceedings{ExplorGraphDRL2020,
title={Autonomous Exploration Under Uncertainty via Deep Reinforcement Learning on Graphs},
author={Chen, Fanfei and Martin, John D. and Huang, Yewei and Wang, Jinkun and Englot, Brendan},
booktitle={IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
year={2020},
}