Development and evaluation of an AI-based approach for path planning of aerial vehicles

Abstract

To ensure safe helicopter operations during close proximity missions under degraded visual conditions, assistance systems are necessary. Obstacle-avoiding path planning algorithms are particularly useful in such scenarios. However, conventional planning setups have limitations for such constraint planning problems, making the integration of methods from the field of Artificial Intelligence (AI) a promising starting point for improvement. This thesis aims to develop and evaluate various concepts for integrating Machine Learning (ML) into conventional path planning algorithms for aerial vehicles. The two most promising concepts were implemented and thoroughly tested to assess their effectiveness. The first concept provides support to a conventional, sampling-based planner by providing it with favorable samples. It achieves those by computing a biased sampling distribution based on the obstacle scenery, start, and goal state using convolutional layers in the form of a U-NET. The second approach aims to replace the conventional planner entirely with Reinforcement Learning agents, utilizing the Neuroevolution of Augmenting Topologies (NEAT) algorithm to train Neural Networks to perform control tasks. The study focuses on non-holonomic path planning in three dimensions, considering helicopter kinematics. Experiments were conducted on both concepts to evaluate their performance, focusing on path length, computation time, and success rate of the planning setup. The results showed that both concepts were valid, with the informed sampling approach demonstrating a significant reduction in average path cost for optimizing planner and acting as an optimization technique for non-optimizing planners. Furthermore, the NEAT algorithm exhibited the ability to take over the task of helicopter control, successfully finding paths to spatial positions within a minimum of time and with a high success rate. Although the suggested algorithms hold promise, further research is necessary for their practical deployment. Overall, the findings presented in this thesis highlight the potential of integrating Machine Learning into conventional path planning algorithms for aerial vehicles.