Neural Network Based Algorithm for Multi-UAV Coverage Path Planning

This paper proposes a method to tackle the Coverage Path Planning (CPP) problem for a fleet of AI-driven UAVs while accounting for congestion, collision avoidance, and efficiency of the path. The algorithm relies on a mixed-use of decentralized Artificial Neural Networks (ANN) and the A* pathfinder. Each UAV has elementary cognitive skills to sense information about the nearby environment which are then fed as an input to the network. The neural network creates a correlation between the current state of a UAV and the best action to take at each time step. The exploration strategy is stored in a labeled database for the single-UAV case; the neural network learns and replicates it in the multi-UAV case, being able to generalize the acquired skills over new maps not contained in the database. The training session imitates human priors in a multi-class classification, which completely bypasses common drawbacks such as the need for large databases or high computational resources. The case study focuses on complex urban areas, for which the grid resolution of the traditional approaches can't model the problem with sufficient accuracy.