Reinforcement Learning for Autonomous Aircraft Avoidance

Effective collision avoidance strategy is crucial for the operation of any unmanned aerial vehicle. In order to maximise the safety and the effectiveness of the collision avoidance strategy, the strategy needs to solve for choosing the best action by taking account of any situation. In this paper, the traditional control method is replaced by a Reinforcement Learning (RL) method called Deep-Q-Network (DQN) and investigate the performance of DQN in aerial collision avoidance. This paper formulate the collision avoidance process as a Markov Decision Process (MDP). DQN will be trained in two simulated scenarios to approximate the best policy which will give us the best action for performing the collision avoidance. First simulation is head-to-head collision simulation following with head-to-head with a crossing aircraft simulation.