UAV Obstacle Avoidance and Landing on Moving Platform Based on Improved Proximal Policy Optimization Algorithm

Addressing the autonomous obstacle avoidance and landing control problem for unmanned aerial vehicles (UAV) in complex environments, this paper proposes a GRU-P30 algorithm that combines an improved Proximal Policy Optimization (PPO) algorithm and Gated Recurrent Unit (GRU) network. Firstly, we establish a UAV kinematic model and define the constraints for obstacle avoidance and landing. Subsequently, based on these constraints, we design the state and action spaces for the deep reinforcement learning algorithm. Additionally, considering the UAV's flight characteristics, we design a potential-based nested reward function to enhance the training efficiency of the policy network. Furthermore, considering the temporal memory aspect of the UAV's obstacle avoidance and landing phase, we introduce the GRU network structure and employ the P30 algorithm with improved clipping instead of the PPO algorithm, proposing the GRU-P30 algorithm to address the trajectory planning problem for UAV obstacle avoidance and following during landing. Finally, the performance of the algorithm was validated, and simulation results showed that the GRU-P30 algorithm had the highest landing accuracy, with an average time of about 34.407 steps.