Safe trajectories from local information for coverage control in non-convex environments

<h2>Description</h2> <p>This dataset contains 3-channels grid-based representations of local information individually retrieved by robots in a team, tasked with a coverage control operation. Data collection was performed running 50 episodes of a coverage control mission with a team of 16 robots controlled by a theoretically proven safe expert controller.</p> <h3>Features</h3> <p>Features encode local information in a 3-channels&nbsp;<em>64 x 64</em> image, corresponding to the discretized sensing region of the robot. The first channel encodes the local likelihood density, the second one the position of team-mates, and the third channel contains the position of obstacles and boundaries.&nbsp;</p> <p><em>imgs{i}.npy</em> files contain data collected over each episode in the form of a [<em>S, N, C, W,</em> W] <em>numpy </em>array, where S is the number of steps of that episode, N is the number of robots, C = 3 is the number of channels, and W = 64 is the size of the image.&nbsp;</p> <h3>Labels</h3> <p>Labels contain the 2D velocity calculated by the expert controller, which is theoretically proven to guarantee collision avoidance.&nbsp;</p> <p><em>vels{i}.npy&nbsp;</em>files contain data collected over each episode in the form of a [<em>S, N, 2</em>]&nbsp;<em>numpy</em> array, associated to the corresponding feature.</p> <p>&nbsp;</p> <h2>Training and Testing</h2> <p>Code for training and testing a CNN-based model mapping local information to 2D velocity is available at <a href="https://github.com/ARSControl/cnn_coverage.git">https://github.com/ARSControl/cnn_coverage.git</a>.</p> <p>&nbsp;</p> <h2>Contact Information</h2> <p>If you are interested in any further information, please contact&nbsp;<a href="mailto:mattia.catellani@unimore.it">mattia.catellani@unimore.it</a>.</p> <p>&nbsp;</p>