This paper proposes a machine vision-guided flexible assembly system for industrial robots, focusing on resolving the precision-efficiency conflict in automotive engine docking. By integrating high-resolution industrial cameras with six-dimensional force sensors, and combining YOLOv3 object detection with nine-point calibration algorithms, the system achieves sub-pixel level component positioning and dynamic error compensation. Experimental results demonstrate that the system achieves 0.05mm positioning accuracy in engine crankshaft assembly, completes single-station operations within 8 seconds, and reduces defect rates by 20% compared to traditional processes.