AI algorithm for predicting and optimizing trajectory of massive UAV swarm

This paper explores the application of Artificial Intelligence (AI) techniques for generating the trajectories of massive swarm of Unmanned Aerial Vehicles (UAVs). The two main challenges addressed include accurately predicting the trajectories of UAVs and efficiently avoiding collisions between them, which we discuss in the two paragraphs below, respectively.

In previous works that did trajectory predictions, a Neural Network (NN) was used. The activation functions used were all standard like Sigmoid, Tanh, and ReLU, which resulted in low trajectory prediction accuracy. In this work, we apply application-oriented activation functions of Swish and Elliott that are known to be resilient to noisy data, which is common in UAV trajectory prediction. We also propose our new activation function, AdaptoSwelliGauss that is fusion of Swish, Elliott and a scaled and shifted Gaussian. This combination better captures the complexities of UAV trajectory prediction (noisy data as well as non-linear trajectory). The trajectory prediction accuracy obtained with our new activation function is three to four orders-of-magnitude better than that obtained from the standard activation functions.

In the UAV context, collision detection and avoidance of UAVs is of utmost importance. While there is a common standard for collision detection, collision avoidance can be done by multiple methods. The first method is by changing their trajectories and the second method is by changing their start times (called batching). The previous works on the trajectory change method were designed for small sets of UAVs. Applying these to our setup of massive UAVs leads to smooth but convoluted paths (including endless loops). On other hand, when we apply the batching method to our setup, then the number of batches is large delaying the launch of all UAVs. Therefore, in this paper we propose a novel collision avoidance strategy that combines a new trajectory change method with the batching method. This results in smooth, simple, and finite trajectory changes in the first method, and reduction-by-half in the number of batches in the second method.