Abstract: To address the low absolute positioning accuracy in heavy-duty industrial robots caused by significant joint flexibility, and the bottleneck where traditional linear stiffness models fail to precisely characterize the hysteresis and nonlinearities of joint transmission systems, a calibration method based on decoupled measurement and nonlinear stiffness pre-compensation is proposed. The method introduces a hierarchical observation strategy utilizing external active loading. By applying loads covering the full rated torque range, it achieves the numerical decoupling of J2 and J3 joint deflections, thereby directly acquiring independent nonlinear stiffness curves with clear physical interpretations. On this basis, calibration is performed using a strategy of "stiffness pre-compensation and geometric parameter identification". Experimental results demonstrate that this method effectively overcomes the challenges of parameter coupling and model mismatch inherent in traditional approaches. Under validation conditions with variable loads up to 300 kg, compared to traditional linear coupled calibration, the root mean square (RMS) error of the robot's absolute positioning accuracy is reduced from 1.159 mm to 0.735 mm, and the maximum error is reduced by approximately 40%. This significantly enhances the operational accuracy of heavy-duty robots under complex working conditions.