Dynamic Obstacle Avoidance of Fixed-Wing Aircraft in Final Phase via Reinforcement Learning

A fixed-wing aircraft can be in the final phase of a potential collision with a noncooperative dynamic obstacle (e.g., a drone) because of the limited sensing range. In the final phase of a potential collision, the performance of the existing obstacle avoidance approaches that do not take into account the bounded and nonisotropic maneuver capability and dynamic and aerodynamic characteristics of a fixed-wing aircraft is limited. To enhance the performance of fixed-wing aircraft in the final phase of a potential collision, this study develops a hierarchical reinforcement learning (RL)-based obstacle avoidance strategy. The RL-based obstacle avoidance strategy learns a high-level RL-based navigator that provides a velocity vector to avoid an obstacle and maintain the altitude, course, and airspeed of an aircraft as possible. The high-level RL-based navigator is combined with a low-level controller to guide and control the aircraft to avoid obstacles. To evaluate the RL-based obstacle avoidance strategy, the strategy is compared with a 3-D velocity obstacle (3DVO)-based obstacle avoidance strategy based on addressing dynamic obstacle avoidance problems of fixed-wing aircraft in a flight simulator. Experimental results show that for an aircraft with a sensing range of 1000 m, the RL-based obstacle avoidance strategy can achieve a success ratio of 92% in obstacle avoidance, whereas the 3DVO-based obstacle avoidance strategy can only achieve a success ratio of 60%.