Welcome to the Coverage Path Planning (CPP) Repository!

Authors

Peter Stratton
Email: pstratto@ucsd.edu, pstratt@umich.edu

Shreyas Arora
Email: sharora@ucsd.edu

Repo Overview

This repository contains an open source (gym-esque) grid world environment to enable researchers (or anyone interested in coverage) to develop and benchmark robotic coverage algorithms, with software infrastructure in place to train Deep Reinforcement Learning agents. In addition, we have implementations and benchmarks of several non-learning based staple robotic coverage algorithms: BSA, BA*, and Spiral STC.

Coverage Path Planning, or CPP, involves controlling a robot or group of robots to visit every open space in an environment. CPP can be applied in situations where robots need to cover large sweeps of area, such as search and rescue, environmental cleanup, and surveillance. It is known that the solution to the CPP problem corresponds to generating a Traveling Salesman Problem (TSP), a NP-Hard problem which entails finding a path through a graph such that every vertex is visited exactly once and no edge is used twice, through the environment's grid cells. path through the environment, which is a NP-hard problem. In addition, to be viable in environments where GPS is not available, the robots need to be able to use their on board sensors to inform their next decision.

By using this repository one will be able to:

• Quickly implement, run, and evaluate new RL/non RL based coverage algorithms
• Learn about existing coverage algorithms using our implementations
• Alter environment dynamics as needed
• Load maps to use for training and testing

Installation

Windows

First, install python by following this link and the instructions specified there: https://www.python.org/downloads/.

Next, install pip by first getting a copy of get-pip.py via the command line and then running get-pip.py:

curl https://bootstrap.pypa.io/get-pip.py -o get-pip.py
python3 get-pip.py

We want to use a virtual environment to isolate the required packages for the repo, so use this command to install venv, a package used to create and manage virtual environments:

pip install virtualenv

After that, we need to clone the repo and navigate to the right directory. To do this, run these commands:

git clone https://github.com/ExistentialRobotics/MARL-Coverage.git
cd MARL_Coverage

Now, we want to create a virtual environment. Run this command to create the virtual environment:

virtualenv coverage_env

Then, we need to activate the environment. Do this by running:

coverage_env\scripts\activate

Next, run the script easy_install_windows.bat to install all the required packages:

easy_install_windows

Last, deactivate the virtual environment:

deactivate

Linux

First, install python executing this in a terminal:

sudo apt-get install python3

Next, install pip by running:

sudo apt install python3-pip

We want to use a virtual environment to isolate the required packages for the repo, so use this command to install venv, a package used to create and manage virtual environments:

pip install virtualenv

After that, we need to clone the repo and navigate to the right directory. To do this, run these commands:

git clone https://github.com/ExistentialRobotics/MARL-Coverage.git
cd MARL_Coverage

Now, we want to create a virtual environment. Run this command to create the virtual environment:

virtualenv coverage_env

Then, we need to activate the environment. Do this by running:

source coverage_env/bin/activate

Next, run the script easy_install_linux.sh to install all the required packages:

bash ./easy_install_linux.sh

Last, deactivate the virtual environment:

deactivate

How to Run the Code

The process for running the code is to set up a folder which holds a config.json file that determines the environment, policy, sensor, and types of grid maps used for the experiment. Additionally, the config.json file sets assorted other hyperparameters related to the experiment, such as the number of training episodes (if an RL agent is trained), the neural network model (if an RL agent is trained), number of testing episodes, whether or not to record a video of the tests running, etc. In the top level directory of the cloned repo there is a folder called Example_Experiments, which contains examples of the config files used for different policy types in this repo.

First, activate the environment if it hasn't already been activated:

source coverage_env/bin/activate

Once a config file has been created, simply run:

python grid_rl_main.py [insert path to config.json file here]

For example, run this command to run an experiment using a STC policy:

python grid_rl_main.py ./Example_Experiments/Non_Learning/STC/Example/config.json

After the experiment has been completed, it will save the model weights for each checkpoint (if running RL), graphs of the average reward per episode for training and testing, recordings of the policy being ran during testing, and any terminal output that occured.

To create your own experiments, create a new folder that holds the config file of your experiment and run the above command, except using the path to your newly created config file. Feel free to check out and run the other example experiments for the different policies in this repo!