Deep learning for autonomous aiming combat vehicle

Abstract

Over the last one decade, deep reinforcement learning (DRL) is set to transform the field of artificial intelligence (AI) and is a step toward developing autonomous systems that have a higher-level knowledge of their surroundings. Deep learning is currently allowing reinforcement learning (RL) to scale to previously unsolvable issues, such as learning to play video games straight from observation input like image pixels. In robotics, DRL algorithms are used to learn control strategies for robots directly from camera inputs and sensor data in the actual environment. However, most of the applications of DRL in video games are single-player games, and in robotics, it is mostly used to reproduce certain tasks with dynamically changing constraints. As such, this project used DRL for multiple autonomous combats aiming at robots to find the best strategy for the team in a dynamically changing environment. To further improve the performance of the robot, robotics navigation and motion planning algorithm were incorporated. The deployment of the robot agent is done on a simulation that is designed specifically for this project. To compare the DRL robot's performance to that of a rule-based algorithm created with a behavior tree, an AI decision-making algorithm for robotics, the DRL robot's performance is compared to that of a rule-based algorithm created with a behavior tree, an AI decision-making method for robotics.