Abstract: Aiming at the problems such as uncertain pose and stacking occlusion of industrial parts in robotic grasping scene, a monocular pose estimation method is proposed based on multi-module neural network and genetic algorithm, to realize a step-by-step optimization from the detection of industrial parts to their positioning in 2D plane, and to the all-round stereo matching. Firstly, the industrial parts are identified and their location area is segmented by neural network, and the L-shaped boundary is constructed by combining the predicted center position, so as to obtain the local effective model of industrial parts projection. Then, the edge information in the area of industrial parts is extracted to generate the chamfer distance function which is delimited by direction angles, and the matching function is constructed combined with the shape of the partially effective model to adapt to the occlusion. Lastly, the strategy of large-scale search and small-scale optimization is adopted, and the quick convergence of 6D pose is realized by genetic algorithm. Moreover, a dataset of the industrial parts with ArUco markers is constructed for experimental verification. Results show that the proposed method can estimate the pose of the industrial parts in about 0.5 s. In the observation distance of 420 mm, the lateral translation error can be controlled within 1 mm, the average rotation angle error can be controlled within 2°. Experimental comparison shows that the proposed method can effectively deal with the accurate estimation of industrial part pose in complex environment and improve the working efficiency of robot.