To address the path planning problem for multiple unmanned vehicles in real-world environments,this paper proposes an algorithm design scheme under the Multi-Agent Soft Actor-Critic (MASAC) framework.To enhance the algorithm's performance,we propose three improvements.First,drawing inspiration from the Artificial Potential Field (APF) concept,we design a dense reward function based on potential shaping techniques to provide abundant,timely and effective feedback signals during the learning process,thereby significantly accelerating convergence.Second,the traditional experience replay buffer is modified by adopting a double-consecutive-frame technique.This approach incorporates two consecutive observation frames as unified units into the experience replay buffer,effectively capturing environmental dynamics and improving training stability.Third,a highly realistic dynamic obstacle environment is constructed using the Gazebo simulation platform,which provides diverse and challenging training samples,ensuring comprehensive learning and optimization under near-real conditions.Finally,the effectiveness of the proposed APF-MASAC algorithm is validated through ablation experiments and robustness tests.