Synthetic dataset generation and deep learning-based real-time navigation for PAVs in urban airspaces

Abstract

A Personal Aerial Vehicle (PAV) is a compact aircraft designed to carry one to four passengers, functioning much like a personal flying car. Due to the high traffic density expected in urban air mobility and the high operating speeds of PAVs, manual piloting is generally discouraged for safety and scalability reasons. Consequently, the development of an integrated Autonomous Navigation and Control System (ANCS) becomes crucial to ensure reliable and efficient operation without requiring the rider to be an experienced pilot. Such systems can be realized using machine learning techniques that rely heavily on high-quality datasets during the training phase. Towards this objective, we developed a simulation-based framework to generate a diverse and rich image dataset capturing various flight scenarios. The dataset includes multiple image modalities such as RGB, depth, and infrared, each labeled with associated positional and orientation data, enabling the system to understand the spatial layout and dynamic conditions of the environment. This data is essential for training deep learning models capable of learning navigation behavior in a supervised setting. At the heart of our approach lies a Convolutional Neural Network (CNN) architecture that processes visual data and predicts optimal navigation paths by estimating new positional coordinates for the vehicle. The CNN learns to associate visual cues with spatial decisions, enabling the PAV to autonomously adapt to di↵erent flight conditions. We present our dataset creation methodology and the proposed CNN-based model architecture aimed at realizing the ANCS. Experimental results in a simulated environment demonstrate that the trained model is capable of safely navigating the PAV through complex urban scenarios, showing promising potential for real-world deployment in autonomous air transportation systems.