In specialized missions such as rubble rescue and unstructured space exploration, aerial manipulators play a critical role. The core challenge lies in the inability of fixed-configuration aerial manipulators to adapt to dynamic environments and evolving task requirements. To address this, this study proposes an aerial modular manipulator (AMM) system based on standardized isomorphic modules. By varying module quantities and assembly configurations, this system achieves reconfigurable operational capabilities tailored to mission-specific demands. Introduce the virtual rotation axis to describe the assembly configuration relationship between modules and establish a unified kinematic model. To accommodate configuration changes arising from module reconfiguration and UAV coupling, neural network-based estimation of dynamic characteristics is implemented. An attitude extended state observer (AESO) is designed to estimate real-time attitude disturbances induced by manipulator motion and payload variations. Experimental validation demonstrates: During configuration switching, AMM exhibits a maximum attitude deviation \(\le 12^{\circ }\) (recovery time: 1 ± 0.21 s). Success rates in grasping tasks show progressive gradients: 93.8% for lightweight deformable objects, 87.5% and 93.8% for medium-load and heavy-load irregular objects respectively, 81.3% for heterogeneous eccentric-load objects and 75.0% for heavy-load irregular objects. These results verify the system’s comprehensive adaptability to geometric features, material properties, and dynamic behaviors within a 500 g payload range.