Quadcopter Guidance Law Design using Deep Reinforcement Learning

This paper examines the quadcopter-target interception problem and proposes a deep reinforcement learning-based approach to solve this problem. The quadcopter-target interception problem is formulated by constructing an Markov Decision Process (MDP) which consist of states, possible actions, transition probabilities and real-valued reward function. The relative position, velocity and angle information between the quadcopter and the target is used when the agent selects the appropriate actions to intercept the target. Permissible acceleration commands are defined as the action space and closing velocity is used in the definition of real-valued reward function. The proposed algorithm is compared with the True Proportional Navigation (TPN) and Model Predictive Control (MPC) algorithms. Numerical simulation results confirm that proposed approach is a suitable solution for the quadcopter guidance problem.