Intelligent Flight Controller Design for a Lightweight UAV Using Simplified Dynamics and Reinforcement Learning

The rise of lightweight unmanned aerial vehicles (UAVs) such as quadrotors has introduced new challenges in autonomous navigation and flight control, especially in environments with physical constraints and obstacles. In this study, a reinforcement learning-based flight control strategy is developed for a simplified ۲D quadrotor model navigating in a bounded environment with a static circular obstacle. The proposed approach leverages the Deep Deterministic Policy Gradient (DDPG) algorithm to learn continuous control actions without requiring a precise dynamic model. The state space is defined by four parameters: position and velocity in the x and y directions, while the action space includes two control accelerations. The environment incorporates a static obstacle centered at (۲.۵, ۲.۵) meters, and the target is fixed at (۵.۰, ۵.۰) meters. After ۱۰,۰۰۰ training steps, the agent successfully learns to navigate from the origin to the target within ۹۰ time steps, reducing its distance to less than ۰.۲ meters. The velocity components remain bounded within ±۱.۵ m/s, and the cumulative trajectory error remains below ۱۶۰ m.step. The learned policy results in a smooth, collision-free path that adapts dynamically to the environment structure. This research demonstrates that deep reinforcement learning can serve as an effective model-free solution for low-cost UAV navigation in constrained spaces. The proposed framework is highly scalable and can be extended to ۳D scenarios, dynamic environments, or multi-agent systems in future research. This work contributes a compact yet powerful approach to intelligent aerial control with minimal model dependency.

کلیدواژه ها:

Autonomous Flight Control ، Model-Free Reinforcement Learning ، UAV Navigation ، Obstacle Avoidance