Abstract
Multi-UAV formation navigation in narrow passages requires coordinated collision avoidance and adaptive formation adjustment under geometric constraints. This paper proposes a leader–follower framework based on task-decomposed hierarchical Soft Actor–Critic (SAC) training. The leader is first trained to establish a stable navigation reference, after which its policy is fixed for follower training. A heterogeneous GRU–CNN (G-C) perception network processes proprioceptive and LiDAR observations, while distance-dependent formation-reward weighting enables progressive formation compression and recovery without predefined switching rules. Experiments on four narrow-passage topologies included in training show that the proposed method achieves 100% task success with mean formation-tracking errors of 0.2734–0.4263 m in static environments, and 90–100% task success with errors of 0.3838–0.5374 m in dynamic environments. Comparisons with hierarchical Proximal Policy Optimization (PPO) and curriculum-assisted multi-agent reinforcement learning baselines demonstrate a more favorable balance between task success and formation-tracking accuracy, particularly under moving-obstacle interference. These results demonstrate the effectiveness of the proposed framework within the evaluated scenarios.