<![CDATA[The ability to avoid obstacles is a subsystem of the navigation system that must be owned by an autonomous robot such as the Unmanned Aerial Vehicle (UAV). This ability governs the process of observing environmental data, making decisions and commands to control movement. Implementation of the avoiding obstacles with conventional methods tends to be static and unable to adapt. Therefore, Implementing this with an artificial intelligence (AI)-based systems will enhance the adaptivity. Reinforcement Learning (RL) is one of the AI or Machine Learning methods that has a characteristic in terms of direct learning in the environment. This study proposes an RL simulation analysis using the Dueling Double Deep Q-Learning (D3QN) method on the UAV to avoid obstacles. The simulation is run on the Airsim simulator and the UAV is a quadcopter. The perceptual configuration of the quadcopter is using a monocular camera, ultrasonic sensor and collision sensor. This study examines the comparative analysis of thresholds of 1 meter and 2 meters, batch sizes 32 and 128. The test results show that the thresholds of 2 meters and batch size 128 provide a more stable visualization plot and converge in less than 70 episode.]]>
Copyrights © 2022
