To address the core challenges in multi-UAV collaborative path planning-balancing computational efficiency and optimization accuracy for large-scale tasks, and coupling real-time response with constraint adaptation in complex scenarios- an improved GPU(graphics processing unit)-accelerated artificial hummingbird algorithm(AHA-GPU) was proposed. Firstly, the study constructed a multi-objective mathematical model integrating path length, task balance, and obstacle avoidance cost, with clear constraints including coverage integrity, UAV dynamics, collision and obstacle avoidance. Secondly, a hybrid exploration strategy combining territorial foraging, migratory foraging, and directed foraging was designed, and a constraint compensation mechanism was introduced to achieve synergy between obstacle avoidance and optimization. Finally, a CPU+GPU heterogeneous parallel framework based on CUDA(compute unified device architecture) was built, in order to improve computational efficiency through thread block optimization and shared memory scheduling. Simulation and real-machine experimental results show that the AHA-GPU algorithm achieves a 10~48x speedup in computation and reduces the total path length by 7%~13% compared with traditional CPU-based algorithms. In obstacle-containing environments, the algorithm reaches a 97.05% area coverage rate and an obstacle avoidance cost of 4.709 0. In 5 m×5 m indoor real-machine experiments, the obstacle avoidance success rate hits 100%. All indicators rank among the top, ensuring excellent solution quality while significantly improving performance. This algorithm effectively breaks through the multi-dimensional contradiction of "efficiency-accuracy-safety" in large-scale complex tasks, providing an efficient and reliable path planning solution for multi-UAV collaborative operations in complex environments.