Integrating convolutional and recurrent networks for synthetic-data-based PAV navigation

Abstract

Urban Air Mobility (UAM) promises transformative solutions for urban transportation by leveraging Personal Aerial Vehicles (PAVs) to reduce congestion, travel time, and environmental impact. Manual piloting in PAVs is discouraged due to the high traffic density of PAVs and their high operating speeds. Hence, PAVs should be equipped with an integrated Autonomous Navigation Control System (ANCS), which eliminates the need for the rider to be a skilled pilot. This paper presents an ANCS for PAVs, integrating a hybrid Convolutional Neural Network (CNN) and Gated Recurrent Unit (GRU)architecture to enable real-time obstacle avoidance and path planning. The system processes sequential RGB-D sensor inputs to predict 3D displacement vectors, ensuring safe manoeuvres around static and dynamic obstacles. A low-pass filtering technique smooths trajectories, enhancing passenger comfort and flight stability. Toward this end, an image data set was created by setting up an end-to-end simulator framework. This dataset includes various categories of RGB, Depth Perspective, Infrared, etc., images labelled with positional or orientation data taken. This dataset is used to train deep learning models to figure out new coordinates for the vehicle to generate the route.