As Uncrewed Aircraft Systems (UAS) become more ubiquitous in urban airspace around the world, the need for reliable navigation and de-confliction technologies becomes paramount. In this paper, the authors improve the popular Deep Reinforcement Learning (RL) methods of Twin Delayed DDPG (TD3) and Proximal Policy Optimization (PPO) and propose two new integrated algorithms for de-confliction with single and multiple intruder UASs in different cases of fixed and variable altitudes. Based on the Actor-Critic method, new RL systems and reward functions are designed that enhance the training efficiency of the navigating UAS agent for the considered environment models. The simulation results show the capability of the trained agent to successfully navigate in a complex environment amid fixed and velocity obstacles. This research contributes to the development of autonomous navigation for UAS in urban airspace.
Intelligent Method for UAV Navigation and De-confliction --Powered by Multi-Agent Reinforcement Learning
06.06.2023
2309504 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
Low-Level De-confliction Problem? Checkmate!
Online Contents | 2007
Battlefield De-Confliction Sensor And Information System
NTRS | 1997
|Multi-Agent Navigation with Reinforcement Learning Enhanced Information Seeking
Deutsches Zentrum für Luft- und Raumfahrt (DLR) | 2022
|