Research on a SAC-Based Algorithm for UAV Autonomous Navigation and Obstacle Avoidance

Abstract

The traditional reinforcement learning algorithms face the problems of poor complex spatial adaptability and inefficient exploration in the autonomous UAV pathfinding and obstacle avoidance tasks in three-dimensional unknown environments. To this end, this paper combines the dynamically decaying -greedy exploration strategy, the dynamic learning rate adjustment mechanism and the offline strategy SAC (Soft Actor-Critic) algorithm based on the maximum entropy reinforcement learning framework, and proposes an autonomous pathfinding and obstacle avoidance algorithm based on SAC, aiming to enhance the obstacle avoidance ability of UAVs in complex environments. -Greedy exploration strategy promotes exploration in the early stage of training and exploitation in the later stage, which helps the intelligent body to find the globally optimal strategy in complex environments and avoids falling into the local optimum in the early stage; the dynamic learning rate adjustment mechanism can refine the strategy and improve the training stability and final performance. The results in the test environment show that this algorithm makes the UAV collision rate decrease significantly, and the timeout rate is close to 0, which effectively enhances the algorithm's adaptability and stability to the environment.