<p>The limited load capacity of single-arm manipulators restricts their ability to assemble large and heavy components such as pipes. The human-robot collaborative assembly (HRCA) of dual-arm manipulators enhances the load capacity and stiffness of robots, while combining with human flexibility, making it a very promising assembly solution. Therefore, this study specifically addresses the challenges of HRCA of dual-arm manipulators, mainly including effectively integrating human motion intentions, controlled components, and dual-arm robots to achieve force decoupling and force/position collaborative control under constraints. Specifically, a four-level control framework is firstly proposed. A human-level Gaussian process regression (GPR) algorithm was introduced to estimate unknown human motion intentions. The object-level external admittance controller and end-effector-level radial basis function neural-network-based variable admittance controller are proposed to regulate external and internal forces, thereby improving the compliance of HRCA systems and ensuring stable internal force tracking performance. Additionally, a joint-level finite-time nonlinear extended state disturbance observer (FTNESDO) is developed to detect collisions and protect humans from harm. Lyapunov stability analysis confirms that the observer can converge within a finite-time. The experimental results demonstrate that the GPR can effectively estimate unknown human motion intentions and minimize human effort. The two-level admittance controller enhances system compliance and reduces internal force tracking errors. FTNESDO exhibits superior response performance when compared to traditional disturbance observers. A series of experiments focusing on the HRCA of dual-arm manipulators have showcased the effectiveness and superiority of the four-level control framework.</p>