Abstract: The cooperative inspection of multiple unmanned aerial vehicles (UAVs) in confined spaces, such as industrial plants, utility tunnels, and warehouse parks, presents significant challenges. These environments often feature narrow passages, frequent occlusions, and complicated layouts, resulting in congestion and close-range interaction conflicts among UAVs. Additionally, the instability of localization systems and the limited computational power available on UAVs further hinder their effective deployment in such environments. These conditions require novel solutions for safe and efficient UAV operation. To address these challenges, in this study, a cooperative deconfliction method, which integrates virtual-tube-based geometric layering with learning-enhanced control mechanisms, is proposed. In the geometric planning layer, a virtual tube is constructed using a parameterized centerline and variable radius, enabling the creation of allocable sub-channels. Continuous reference sub-trajectories are generated within the Frenet frame to ensure macroscopic spatial separation of the UAVs. This approach ensures that UAVs are kept within safe operational zones and prevents collisions using geometry-based constraints. Furthermore, in the learning-based control layer, a conflict-aware multi-agent proximal policy optimization (MA-PPO) controller with shared parameters is developed. This controller utilizes low-dimensional structured observations, which include task progress, lateral deviation, safety margin, and neighboring UAV information. The decision-making process is constrained to outputting a tangential-speed intention, which reduces the complexity of the control problem and ensures easier training. A pure pursuit method provides lateral geometric correction. When the inter-agent distance falls below a safety threshold, a potential-field-based repulsive velocity term is added to suppress close-range conflicts. To validate the proposed method, simulation experiments were conducted using a five-UAV maze scenario on the AirSim simulation platform. The experiments demonstrate that the UAVs successfully complete the mission without any collisions, with the minimum inter-agent distance consistently remaining above 0.6 m. Additionally, the average lateral tracking error decreased from 0.571 m to 0.201 m, representing a reduction of 64.8% when compared to a distributed rule-based control method. These results showcase the efficacy of the proposed method in terms of trajectory tracking accuracy and cooperative safety in confined spaces. Moreover, the proposed method was tested in a complex, inconsistent scenario, which was not included in the training environment. The results show that the method achieved zero-sample transfer and maintained zero collisions, proving the robustness and generalization ability of the learned policy under unknown geometric constraints. This demonstrates that the method is not only effective in the training environment but also capable of adapting to new, unforeseen scenarios. Overall, the results indicate that the proposed approach can simultaneously ensure safe cooperative passage and accurate trajectory tracking under low-sensing constraints. The integration of geometric constraints with learning-based decision-making provides a promising solution for multi-UAV cooperative missions in confined spaces. By addressing the challenges of low sensing and limited computational power, this method paves the way for the practical deployment of UAVs in complex and restricted environments.