Autonomous Landing of eVTOL Vehicles via Deep Q-Networks

摘要

View Video Presentation: https://doi.org/10.2514/6.2023-4499.vid Urban Air Mobility (UAM) promises a new mode of transportation system providing disruptive innovative operation safely and efficiently in a highly constrained urban environment. The UAM is attracting interest from a wide range of stakeholders, institutions, and businesses due to the technological advances in electric and distributed propulsion, which makes it easier to construct unique aircraft types with the capability for electric Vertical Take-off and Landing (eVTOL). The autonomous landing of an eVTOL vehicle to a ground vertiport needs highly automated systems, while in a piloted helicopter the landing is directed by the Air Traffic Control unit and pilot. In this paper, we propose a deep reinforcement learning-based method to locate the landing marker and land the eVTOL vehicle on it using only low-resolution images captured by a camera pointed downward. The proposed approach is based on Deep Q-Networks (DQNs) using a high-level control policy to navigate the eVTOL vehicle toward the landing marker. We have implemented different approaches, including DQNs and double DQNs algorithms. We trained and tested the vehicle in the Robotic Operating System (ROS) Gazebo simulation environment within a small-scale environmental setting. The performance of our model has a promising result to land the vehicle autonomously by following the generated policy throughout the network. The precise and safe landing of eVTOL vehicles is one of the key steps towards addressing conflict and delays during peak hours at the vertiport for future UAM applications.