Application of the CR-PPO Algorithm in UAV Flight Action Control within Complex Gate-Passing Environments

With the rapid development of Reinforcement Learning (RL) in the field of intelligent control, vision-based Deep Reinforcement Learning (DRL) methods have shown broad prospects for application in autonomous UAV flight control. This study constructs a three-dimensional virtual environment on the AirSim simulation platform and proposes an improved UAV flight control algorithm, Clipped and Reward-optimized PPO with Channel Attention (CR-PPO), which is a PPO-based variant that optimizes the reward function and integrates a Channel Attention (CA) mechanism. The method first extends the standard Proximal Policy Optimization (PPO) algorithm by incorporating adaptive clipping range adjustment, entropy scheduling, and KL-divergence-based early stopping to form the Clipped PPO (C-PPO) algorithm to enhance training stability and convergence speed. On this basis, a Convolutional Neural Network (CNN) perception module with CA and a finely designed reward function are integrated to form the complete CR-PPO framework. The proposed perception-decision system takes image data as input and achieves end-to-end autonomous control without reliance on physical state information. Experimental results show that CR-PPO significantly outperforms DQN, A2C, standard PPO, and C-PPO in terms of training efficiency, policy stability, and task completion rate, while also demonstrating strong generalization and deployment potential. This work provides an effective technical pathway and a feasible validation basis for intelligent UAV control in complex environments.