Abstract: For multi-UAV cooperative inspection in confined spaces such as industrial plants, utility tunnels, and warehouse parks, narrow passages and frequent occlusions often lead to congestion and close-range interaction conflicts, while unstable localization and limited onboard computing further hinder real-time deployment. To address these challenges, this paper proposes a multi-UAV cooperative deconfliction method that integrates virtual-tube-based geometric layering with learning enhancement. In the geometric planning layer, a virtual tube parameterized by a centerline and a variable radius is constructed, and allocable sub-channels and continuous reference sub-trajectories are generated in the Frenet frame to achieve macroscopic spatial separation. In the learning layer, a conflict-aware Multi-Agent Proximal Policy Optimization controller with parameter sharing (MA-PPO-PS) is developed, which employs a low-dimensional structured observation composed of progress, lateral deviation, safety margin, and neighboring-agent information, constrains the policy output to a tangential-speed intention, and uses Pure Pursuit (PP) for lateral geometric correction; when the inter-agent distance falls below a safety threshold, a potential-field-type repulsive velocity term is superimposed to suppress close-range conflicts. Five-UAV maze-scenario experiments on the AirSim simulation platform demonstrate that no collision occurs during flight and the minimum inter-agent distance remains above 0.6 m; under the same task completion rate, the proposed method reduces lateral tracking error compared with a distributed rule-based controller. These results indicate that the proposed approach can simultaneously ensure safe cooperative passage and accurate trajectory tracking under low-sensing constraints.