Unmanned Aerial Vehicle (UAV) swarms have been extensively employed in search, rescue, environmental monitoring and surveillance. However, existing multi‐UAV planning methods often struggle to adapt to rapidly changing, unknown environments or require large computational resources. To address these challenges, a novel UAV’s path planning method is proposed that integrates an improved rotational artificial potential field (IRAPF) with a multi‐agent proximal policy optimization (MAPPO) algorithm under a centralized‐training and decentralized‐execution paradigm. In IRAPF, a traditional Euclidean distance-based repulsion is replaced with a superellipse-norm potential to generate shape conforming repulsive gradients around obstacles, and introduce a dynamically selected rotational force whose tangential direction aligns with the goal vector, thereby ensuring continuous obstacle circumvention and eliminating local minima. The IRAPF factors are normalized to a continuous space, and deep reinforcement learning (DRL) is used in the continuous space to optimize the IRAPF parameters to improve the accuracy and flexibility of path planning. In the scenario, simulations and comparisons are performed to demonstrate the advancement of proposed method with success rate, convergence average reward, and average path length indicators.