Path Planning of UAV Using Step-wise Deep Q-learning Algorithm

With the increasing application of the Unmanned Aerial Vehicle(UAV) technology, the path planning of UAV is becoming increasing important, However, with the increasing complexity of UAV applications, the application scenario is always complex, crowded with dense obstacles, open, and dynamic. In this paper, we dedicate to deep reinforcement learning algorithms for autonomous obstacle avoidance and navigation of UAV. The navigation problem is considered as a target-driven MDP problem, in which UAV takes its next action conditioned on both its current observation and the destination, Additionally, DRL algorithm with sparse rewards is hard to convergence, we introduce a step-wise dynamic relative goal method to extract the common feature between different navigation targets.