Research on Three-Dimensional Autonomous Obstacle Avoidance Path Planning Methods for UAVs

With the advancement of science and technology and the rapid development of the socio-economy, unmanned aerial vehicles (UAVs) are playing an increasingly important role in daily economic activities. In civilian applications, UAVs can perform tasks such as high-voltage power line and oil pipeline inspections, as well as logistics transportation. In the military domain, they are widely used for logistics supply and delivery, battlefield situational monitoring, and reconnaissance in complex environments. Therefore, research on three-dimensional (3D) obstacle-avoidance autonomous path planning for UAVs has significant practical engineering applications. This paper proposes a deep reinforcement learning (DRL)–improved mayfly algorithm (IMA) for UAV autonomous obstacle-avoidance path planning. Considering the limitations of the mayfly algorithm (MA), such as the high randomness in initial population generation and slow convergence speed during iteration, Halton sequences and an adaptive Gaussian–Cauchy mutation strategy are introduced to balance the global exploration and local optimization capabilities of the MA. Furthermore, recognizing that the IMA still cannot overcome inherent drawbacks of swarm intelligence algorithms, such as the random selection of mutation strategy probability distributions and the lack of individual strategy optimization due to simultaneous group strategy optimization, this paper applies the deep deterministic policy gradient (DDPG) algorithm from DRL to update the population positions in the IMA. This further enhances the algorithm's optimization performance, convergence ability, and computational efficiency, thereby achieving autonomous obstacle-avoidance path planning for UAVs. The performance of the 3D UAV paths optimized by the traditional MA, the IMA, and the DRL–IMA is compared to validate the effectiveness of the proposed algorithm. The simulation's output indicates that the introduced DRL–IMA brings down the average fitness value to 0.05022 after 600 iterations and it takes just 72 generations to converge. This shows that the new method not only converges faster but also has better stability than both the old and modern MAs.