Obstacle avoidance and local path planning based on reinforcement learning for hexacopter drone in simulated environments

This article details the development, training, and application of an artificial intelligence (AI) agent designed for autonomous navigation tasks in a hexacopter drone. The focus is on two key subtasks: obstacle avoidance and local path planning. Using the Gazebo physics simulator, a simulated environment based on an oil palm plantation was created. The drone, controlled by an LQG controller for flight and trajectory management, was integrated with an AI agent trained using the D3QN (Dueling Double Deep Q-Network) reinforcement learning algorithm. This setup allowed the drone to autonomously determine destination coordinates and navigate safely through the environment using LIDAR sensors for real-time obstacle detection. The results demonstrate that the D3QN agent effectively reaches target positions while avoiding obstacles, showcasing the efficacy of reinforcement learning in training local path planners for drones in complex scenarios. These findings underline the potential of using simulation-based reinforcement learning for enhancing drone navigation and operational efficiency.