Abstract: In order to solve the problems of the traditional arc welding robot, such as the lack of autonomous planning of complex weld curve welding path, the low level of intelligence and automation, this paper introduces machine vision into the arc welding robot system, and proposes an algorithm for weld trajectory recognition and path generation. Firstly, the image of the welding object is collected by the industrial computer through the camera, and the image preprocessing operations such as image graying, histogram equalization and threshold segmentation are carried out; Secondly, the edge lines of weld pixels are extracted from the preprocessed image by Canny operator, and a weld centerline extraction algorithm is designed; Finally, an adaptive polynomial curve fitting algorithm is proposed, and the pixel coordinates of the fitting curve are transformed into robot coordinates through the coordinate transformation model, and the coordinate data is sent to the arc welding robot. The experimental results show that the improved arc welding robot system can independently identify the complex weld trajectory and generate the welding path, guide the arc welding robot to complete the welding operation, and the maximum error value is not more than 0.30mm, which meets the requirements of welding accuracy, and has certain application and reference value.