End-to-End Learning-Based Obstacle Avoidance for Fixed-Wing UAVs

In this paper, a deep reinforcement learning training framework based on attitude control is proposed for the obstacle avoidance problem of fixed-wing UAVs in 3D realistic scenes. The training framework includes an environment information feature extraction network and an obstacle avoidance strategy training network. In the training network module, attitude stabilization and altitude maintenance reward and punishment functions are designed for fixed-wing UAV attitude control. The learning of obstacle avoidance strategies is performed while the fixed-wing UAV maintains attitude stability and altitude stability. Finally, the method is validated in the joint simulation environment we built. The training method introduced in this paper and the built training framework provide simulation data and results for the fixed-wing UAV-aware avoidance strategy in a realistic environment, which provides a feasible solution for the deployment of the obstacle avoidance strategy network to real aircraft.