Application of reinforcement learning for autonomous combat

• Main Author: Huang, Andrian
• Other Authors: Lap-Pui Chau
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Engineering::Electrical and electronic engineering::Control and instrumentation::Robotics
• Online Access: https://hdl.handle.net/10356/158211
• Institution: Nanyang Technological University

The RoboMaster University AI Challenge (RMUA) is an annual international robotics competition involving 2-versus-2 battle between autonomous robots armed with projectile launcher, where the goal is to cooperate with the ally robot and deplete the enemy robots’ health by shooting projectiles at them. Due to the nature of the competition, the robots must be able to autonomously perform all tasks pertaining to the competition. With the recent rise in popularity of deep learning, it seems compelling to apply deep reinforcement learning for such tasks, which are doable by traditional methods yet might be difficult to explicitly program or fine-tune. Despite this, the combat robots in RMUA predominantly still use traditional methods as they are tried-and-tested in the competition. Deep reinforcement learning should however still be able to bring benefit to the RMUA combat robot due to its ability to let agents learn without explicit programming, including for rather complex environments. With that in mind, this project seeks to explore the possibility of deep reinforcement learning in RMUA combat robot by applying it for a certain combat task: autonomous enemy aiming and tracking, a task proven difficult due to bullet drop, small hitbox on the enemy robot’s armour, and the bullet’s finite velocity. The result of this project suggests that deep reinforcement learning can be used as an alternative to classical methods, although it may not beat classical methods especially in a simulated environment. Deep reinforcement learning for this task has a lot of room for improvement and can potentially be combined with classical methods.