A Reinforcement Learning Agent for UAV Control: Mathematical Foundations, Implementation, and Human-vs-AI Benchmarking

Neelesh, Mungoli (2025) A Reinforcement Learning Agent for UAV Control: Mathematical Foundations, Implementation, and Human-vs-AI Benchmarking. International Journal of Trend in Scientific Research and Development, 9 (2). pp. 242-254. ISSN 2456-6470

Text ijtsrd76308.pdf Download (1MB)

Abstract

In this work, we propose a novel deep reinforcement learning (DRL) agent architecture for fully autonomous UAV control that fuses real-time sensor fusion with advanced multi-objective reward shaping to achieve robust flight dynamics under varied environmental conditions. We begin by defining the system’s decision-making process as a partially observable Markov decision process (POMDP), wherein the UAV’s state space encapsulates high-dimensional sensor inputs, including LIDAR point clouds, inertial measurement unit (IMU) data, and geospatial telemetry, while the agent’s action space is composed of continuous motor velocity commands. Our learning algorithm employs a hierarchical policy gradient method with parallelizable sub-policies dedicated to tasks such as obstacle avoidance, trajectory planning, and energy conservation. Each sub-policy is trained using a variant of proximal policy optimization (PPO) that is adapted to dynamic flight constraints through Lagrangian relaxation techniques and enforced via real-time on-policy updates.

Item Type:	Article
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Divisions:	Postgraduate > Master's of Islamic Education
Depositing User:	Journal Editor
Date Deposited:	15 Mar 2025 11:56
Last Modified:	15 Mar 2025 11:56
URI:	http://eprints.umsida.ac.id/id/eprint/15834

Actions (login required)

View Item