This repository contains training and testing scripts for a decentralized CNN-based controller in a multi-robot coverage control scenario.

This CNN-based controller maps locally available information to control actions for each robot. Local information is represented in the form of a 3-channels grid-shaped object, where the first channel encodes information about the local density function, the second one the relative position of detected team-mates, and the third one the relative position of obstacles and boundaries.

The model processes the input information and returns the 2D velocity for the robot.

Data is collected running an expert algorithm, generating pairs of (images, velocities) pairs for the training algorithm to learn how to produce a suitable control input for the robot.

A pre-collected dataset containing 40000 pairs is available here:

The model can be tested on a simulated scenario with the desired number of robots and obstacles.